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 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_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
108 static struct cfs_debug_limit_state cdls; \
109 struct osc_extent *__ext = (extent); \
112 __CDEBUG_WITH_LOC(file, func, line, mask, &cdls, \
113 "extent %p@{" EXTSTR ", " \
114 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \
115 /* ----- extent part 0 ----- */ \
116 __ext, EXTPARA(__ext), \
117 /* ----- part 1 ----- */ \
118 atomic_read(&__ext->oe_refc), \
119 atomic_read(&__ext->oe_users), \
120 list_empty_marker(&__ext->oe_link), \
121 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
123 /* ----- part 2 ----- */ \
124 __ext->oe_grants, __ext->oe_nr_pages, \
125 list_empty_marker(&__ext->oe_pages), \
126 waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \
127 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner, \
128 /* ----- part 4 ----- */ \
130 if (mask == D_ERROR && __ext->oe_dlmlock != NULL) \
131 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext); \
133 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext); \
136 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...) \
137 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__, \
138 mask, ext, fmt, ## __VA_ARGS__)
141 #define EASSERTF(expr, ext, fmt, args...) do { \
143 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args); \
144 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj); \
150 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
152 static inline struct osc_extent *rb_extent(struct rb_node *n)
157 return container_of(n, struct osc_extent, oe_node);
160 static inline struct osc_extent *next_extent(struct osc_extent *ext)
165 LASSERT(ext->oe_intree);
166 return rb_extent(rb_next(&ext->oe_node));
169 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
174 LASSERT(ext->oe_intree);
175 return rb_extent(rb_prev(&ext->oe_node));
178 static inline struct osc_extent *first_extent(struct osc_object *obj)
180 return rb_extent(rb_first(&obj->oo_root));
183 /* object must be locked by caller. */
184 static int osc_extent_sanity_check0(struct osc_extent *ext,
185 const char *func, const int line)
187 struct osc_object *obj = ext->oe_obj;
188 struct osc_async_page *oap;
192 if (!osc_object_is_locked(obj))
195 if (ext->oe_state >= OES_STATE_MAX)
198 if (atomic_read(&ext->oe_refc) <= 0)
201 if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
204 switch (ext->oe_state) {
206 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
211 if (atomic_read(&ext->oe_users) == 0)
215 if (ext->oe_fsync_wait && !ext->oe_urgent)
219 if (ext->oe_grants == 0)
221 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
225 if (atomic_read(&ext->oe_users) > 0)
229 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
232 if (ext->oe_sync && ext->oe_grants > 0)
235 if (ext->oe_dlmlock != NULL &&
236 ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
237 !ldlm_is_failed(ext->oe_dlmlock)) {
238 struct ldlm_extent *extent;
240 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
241 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
242 extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end)))
245 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
249 if (ext->oe_nr_pages > ext->oe_mppr)
252 /* Do not verify page list if extent is in RPC. This is because an
253 * in-RPC extent is supposed to be exclusively accessible w/o lock. */
254 if (ext->oe_state > OES_CACHE)
261 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
262 pgoff_t index = osc_index(oap2osc(oap));
264 if (index > ext->oe_end || index < ext->oe_start)
267 if (page_count != ext->oe_nr_pages)
272 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
273 "sanity check %p failed: rc = %d\n",
278 #define sanity_check_nolock(ext) \
279 osc_extent_sanity_check0(ext, __func__, __LINE__)
281 #define sanity_check(ext) ({ \
283 osc_object_lock((ext)->oe_obj); \
284 __res = sanity_check_nolock(ext); \
285 osc_object_unlock((ext)->oe_obj); \
291 * sanity check - to make sure there is no overlapped extent in the tree.
293 static int osc_extent_is_overlapped(struct osc_object *obj,
294 struct osc_extent *ext)
296 struct osc_extent *tmp;
298 LASSERT(osc_object_is_locked(obj));
303 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
306 if (tmp->oe_end >= ext->oe_start &&
307 tmp->oe_start <= ext->oe_end)
313 static void osc_extent_state_set(struct osc_extent *ext, int state)
315 LASSERT(osc_object_is_locked(ext->oe_obj));
316 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
318 /* Never try to sanity check a state changing extent :-) */
319 /* LASSERT(sanity_check_nolock(ext) == 0); */
321 /* TODO: validate the state machine */
322 ext->oe_state = state;
323 wake_up_all(&ext->oe_waitq);
326 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
328 struct osc_extent *ext;
330 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
334 RB_CLEAR_NODE(&ext->oe_node);
336 cl_object_get(osc2cl(obj));
337 atomic_set(&ext->oe_refc, 1);
338 atomic_set(&ext->oe_users, 0);
339 INIT_LIST_HEAD(&ext->oe_link);
340 ext->oe_state = OES_INV;
341 INIT_LIST_HEAD(&ext->oe_pages);
342 init_waitqueue_head(&ext->oe_waitq);
343 ext->oe_dlmlock = NULL;
348 static void osc_extent_free(struct osc_extent *ext)
350 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
353 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
355 LASSERT(atomic_read(&ext->oe_refc) >= 0);
356 atomic_inc(&ext->oe_refc);
360 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
362 LASSERT(atomic_read(&ext->oe_refc) > 0);
363 if (atomic_dec_and_test(&ext->oe_refc)) {
364 LASSERT(list_empty(&ext->oe_link));
365 LASSERT(atomic_read(&ext->oe_users) == 0);
366 LASSERT(ext->oe_state == OES_INV);
367 LASSERT(!ext->oe_intree);
369 if (ext->oe_dlmlock != NULL) {
370 lu_ref_add(&ext->oe_dlmlock->l_reference,
372 LDLM_LOCK_PUT(ext->oe_dlmlock);
373 ext->oe_dlmlock = NULL;
375 cl_object_put(env, osc2cl(ext->oe_obj));
376 osc_extent_free(ext);
381 * osc_extent_put_trust() is a special version of osc_extent_put() when
382 * it's known that the caller is not the last user. This is to address the
383 * problem of lacking of lu_env ;-).
385 static void osc_extent_put_trust(struct osc_extent *ext)
387 LASSERT(atomic_read(&ext->oe_refc) > 1);
388 LASSERT(osc_object_is_locked(ext->oe_obj));
389 atomic_dec(&ext->oe_refc);
393 * Return the extent which includes pgoff @index, or return the greatest
394 * previous extent in the tree.
396 static struct osc_extent *osc_extent_search(struct osc_object *obj,
399 struct rb_node *n = obj->oo_root.rb_node;
400 struct osc_extent *tmp, *p = NULL;
402 LASSERT(osc_object_is_locked(obj));
405 if (index < tmp->oe_start) {
407 } else if (index > tmp->oe_end) {
418 * Return the extent covering @index, otherwise return NULL.
419 * caller must have held object lock.
421 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
424 struct osc_extent *ext;
426 ext = osc_extent_search(obj, index);
427 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
428 return osc_extent_get(ext);
432 /* caller must have held object lock. */
433 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
435 struct rb_node **n = &obj->oo_root.rb_node;
436 struct rb_node *parent = NULL;
437 struct osc_extent *tmp;
439 LASSERT(ext->oe_intree == 0);
440 LASSERT(ext->oe_obj == obj);
441 LASSERT(osc_object_is_locked(obj));
446 if (ext->oe_end < tmp->oe_start)
448 else if (ext->oe_start > tmp->oe_end)
451 EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
453 rb_link_node(&ext->oe_node, parent, n);
454 rb_insert_color(&ext->oe_node, &obj->oo_root);
459 /* caller must have held object lock. */
460 static void osc_extent_erase(struct osc_extent *ext)
462 struct osc_object *obj = ext->oe_obj;
463 LASSERT(osc_object_is_locked(obj));
464 if (ext->oe_intree) {
465 rb_erase(&ext->oe_node, &obj->oo_root);
467 /* rbtree held a refcount */
468 osc_extent_put_trust(ext);
472 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
474 struct osc_object *obj = ext->oe_obj;
476 LASSERT(osc_object_is_locked(obj));
477 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
478 if (ext->oe_state == OES_CACHE) {
479 osc_extent_state_set(ext, OES_ACTIVE);
480 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
482 atomic_inc(&ext->oe_users);
483 list_del_init(&ext->oe_link);
484 return osc_extent_get(ext);
487 static void __osc_extent_remove(struct osc_extent *ext)
489 LASSERT(osc_object_is_locked(ext->oe_obj));
490 LASSERT(list_empty(&ext->oe_pages));
491 osc_extent_erase(ext);
492 list_del_init(&ext->oe_link);
493 osc_extent_state_set(ext, OES_INV);
494 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
497 static void osc_extent_remove(struct osc_extent *ext)
499 struct osc_object *obj = ext->oe_obj;
501 osc_object_lock(obj);
502 __osc_extent_remove(ext);
503 osc_object_unlock(obj);
507 * This function is used to merge extents to get better performance. It checks
508 * if @cur and @victim are contiguous at block level.
510 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
511 struct osc_extent *victim)
513 struct osc_object *obj = cur->oe_obj;
514 struct client_obd *cli = osc_cli(obj);
519 LASSERT(cur->oe_state == OES_CACHE);
520 LASSERT(osc_object_is_locked(obj));
524 if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
527 if (cur->oe_max_end != victim->oe_max_end)
530 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
531 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
532 chunk_start = cur->oe_start >> ppc_bits;
533 chunk_end = cur->oe_end >> ppc_bits;
534 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
535 chunk_end + 1 != victim->oe_start >> ppc_bits)
538 /* overall extent size should not exceed the max supported limit
539 * reported by the server */
540 if (cur->oe_end - cur->oe_start + 1 +
541 victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
544 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
546 cur->oe_start = min(cur->oe_start, victim->oe_start);
547 cur->oe_end = max(cur->oe_end, victim->oe_end);
548 /* per-extent tax should be accounted only once for the whole extent */
549 cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax;
550 cur->oe_nr_pages += victim->oe_nr_pages;
551 /* only the following bits are needed to merge */
552 cur->oe_urgent |= victim->oe_urgent;
553 cur->oe_memalloc |= victim->oe_memalloc;
554 list_splice_init(&victim->oe_pages, &cur->oe_pages);
555 list_del_init(&victim->oe_link);
556 victim->oe_nr_pages = 0;
558 osc_extent_get(victim);
559 __osc_extent_remove(victim);
560 osc_extent_put(env, victim);
562 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
567 * Drop user count of osc_extent, and unplug IO asynchronously.
569 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
571 struct osc_object *obj = ext->oe_obj;
572 struct client_obd *cli = osc_cli(obj);
576 LASSERT(atomic_read(&ext->oe_users) > 0);
577 LASSERT(sanity_check(ext) == 0);
578 LASSERT(ext->oe_grants > 0);
580 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
581 LASSERT(ext->oe_state == OES_ACTIVE);
582 if (ext->oe_trunc_pending) {
583 /* a truncate process is waiting for this extent.
584 * This may happen due to a race, check
585 * osc_cache_truncate_start(). */
586 osc_extent_state_set(ext, OES_TRUNC);
587 ext->oe_trunc_pending = 0;
588 osc_object_unlock(obj);
592 osc_extent_state_set(ext, OES_CACHE);
593 osc_update_pending(obj, OBD_BRW_WRITE,
596 /* try to merge the previous and next extent. */
597 if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
598 grant += cli->cl_grant_extent_tax;
599 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
600 grant += cli->cl_grant_extent_tax;
603 list_move_tail(&ext->oe_link,
604 &obj->oo_urgent_exts);
605 else if (ext->oe_nr_pages == ext->oe_mppr) {
606 list_move_tail(&ext->oe_link,
609 osc_object_unlock(obj);
611 osc_unreserve_grant(cli, 0, grant);
614 osc_io_unplug_async(env, cli, obj);
616 osc_extent_put(env, ext);
620 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
622 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
626 * Find or create an extent which includes @index, core function to manage
629 static struct osc_extent *osc_extent_find(const struct lu_env *env,
630 struct osc_object *obj, pgoff_t index,
631 unsigned int *grants)
633 struct client_obd *cli = osc_cli(obj);
634 struct osc_lock *olck;
635 struct cl_lock_descr *descr;
636 struct osc_extent *cur;
637 struct osc_extent *ext;
638 struct osc_extent *conflict = NULL;
639 struct osc_extent *found = NULL;
642 unsigned int max_pages; /* max_pages_per_rpc */
643 unsigned int chunksize;
644 int ppc_bits; /* pages per chunk bits */
649 cur = osc_extent_alloc(obj);
651 RETURN(ERR_PTR(-ENOMEM));
653 olck = osc_env_io(env)->oi_write_osclock;
654 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
655 LASSERT(olck->ols_state == OLS_GRANTED);
657 descr = &olck->ols_cl.cls_lock->cll_descr;
658 LASSERT(descr->cld_mode >= CLM_WRITE);
660 LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
661 "chunkbits: %u\n", cli->cl_chunkbits);
662 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
663 chunk_mask = ~((1 << ppc_bits) - 1);
664 chunksize = 1 << cli->cl_chunkbits;
665 chunk = index >> ppc_bits;
667 /* align end to RPC edge. */
668 max_pages = cli->cl_max_pages_per_rpc;
669 if ((max_pages & ~chunk_mask) != 0) {
670 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
671 max_pages, cli->cl_chunkbits, chunk_mask);
672 RETURN(ERR_PTR(-EINVAL));
674 max_end = index - (index % max_pages) + max_pages - 1;
675 max_end = min_t(pgoff_t, max_end, descr->cld_end);
677 /* initialize new extent by parameters so far */
678 cur->oe_max_end = max_end;
679 cur->oe_start = index & chunk_mask;
680 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
681 if (cur->oe_start < descr->cld_start)
682 cur->oe_start = descr->cld_start;
683 if (cur->oe_end > max_end)
684 cur->oe_end = max_end;
686 cur->oe_mppr = max_pages;
687 if (olck->ols_dlmlock != NULL) {
688 LASSERT(olck->ols_hold);
689 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
690 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
693 /* grants has been allocated by caller */
694 LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
695 "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
696 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
700 osc_object_lock(obj);
701 ext = osc_extent_search(obj, cur->oe_start);
703 ext = first_extent(obj);
704 while (ext != NULL) {
705 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
706 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
708 LASSERT(sanity_check_nolock(ext) == 0);
709 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
712 /* if covering by different locks, no chance to match */
713 if (olck->ols_dlmlock != ext->oe_dlmlock) {
714 EASSERTF(!overlapped(ext, cur), ext,
715 EXTSTR"\n", EXTPARA(cur));
717 ext = next_extent(ext);
721 /* discontiguous chunks? */
722 if (chunk + 1 < ext_chk_start) {
723 ext = next_extent(ext);
727 /* ok, from now on, ext and cur have these attrs:
728 * 1. covered by the same lock
729 * 2. contiguous at chunk level or overlapping. */
731 if (overlapped(ext, cur)) {
732 /* cur is the minimum unit, so overlapping means
734 EASSERTF((ext->oe_start <= cur->oe_start &&
735 ext->oe_end >= cur->oe_end),
736 ext, EXTSTR"\n", EXTPARA(cur));
738 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
739 /* for simplicity, we wait for this extent to
740 * finish before going forward. */
741 conflict = osc_extent_get(ext);
745 found = osc_extent_hold(ext);
749 /* non-overlapped extent */
750 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
751 /* we can't do anything for a non OES_CACHE extent, or
752 * if there is someone waiting for this extent to be
753 * flushed, try next one. */
754 ext = next_extent(ext);
758 /* check if they belong to the same rpc slot before trying to
759 * merge. the extents are not overlapped and contiguous at
760 * chunk level to get here. */
761 if (ext->oe_max_end != max_end) {
762 /* if they don't belong to the same RPC slot or
763 * max_pages_per_rpc has ever changed, do not merge. */
764 ext = next_extent(ext);
768 /* check whether maximum extent size will be hit */
769 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
770 cli->cl_max_extent_pages) {
771 ext = next_extent(ext);
775 /* it's required that an extent must be contiguous at chunk
776 * level so that we know the whole extent is covered by grant
777 * (the pages in the extent are NOT required to be contiguous).
778 * Otherwise, it will be too much difficult to know which
779 * chunks have grants allocated. */
781 /* try to do front merge - extend ext's start */
782 if (chunk + 1 == ext_chk_start) {
783 /* ext must be chunk size aligned */
784 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
786 /* pull ext's start back to cover cur */
787 ext->oe_start = cur->oe_start;
788 ext->oe_grants += chunksize;
789 LASSERT(*grants >= chunksize);
790 *grants -= chunksize;
792 found = osc_extent_hold(ext);
793 } else if (chunk == ext_chk_end + 1) {
795 ext->oe_end = cur->oe_end;
796 ext->oe_grants += chunksize;
797 LASSERT(*grants >= chunksize);
798 *grants -= chunksize;
800 /* try to merge with the next one because we just fill
802 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
803 /* we can save extent tax from next extent */
804 *grants += cli->cl_grant_extent_tax;
806 found = osc_extent_hold(ext);
811 ext = next_extent(ext);
814 osc_extent_tree_dump(D_CACHE, obj);
816 LASSERT(conflict == NULL);
817 if (!IS_ERR(found)) {
818 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
819 OSC_EXTENT_DUMP(D_CACHE, found,
820 "found caching ext for %lu.\n", index);
822 } else if (conflict == NULL) {
823 /* create a new extent */
824 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
825 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
826 LASSERT(*grants >= cur->oe_grants);
827 *grants -= cur->oe_grants;
829 cur->oe_state = OES_CACHE;
830 found = osc_extent_hold(cur);
831 osc_extent_insert(obj, cur);
832 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
833 index, descr->cld_end);
835 osc_object_unlock(obj);
837 if (conflict != NULL) {
838 LASSERT(found == NULL);
840 /* waiting for IO to finish. Please notice that it's impossible
841 * to be an OES_TRUNC extent. */
842 rc = osc_extent_wait(env, conflict, OES_INV);
843 osc_extent_put(env, conflict);
846 GOTO(out, found = ERR_PTR(rc));
853 osc_extent_put(env, cur);
858 * Called when IO is finished to an extent.
860 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
863 struct client_obd *cli = osc_cli(ext->oe_obj);
864 struct osc_async_page *oap;
865 struct osc_async_page *tmp;
866 int nr_pages = ext->oe_nr_pages;
868 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
873 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
875 ext->oe_rc = rc ?: ext->oe_nr_pages;
876 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
878 osc_lru_add_batch(cli, &ext->oe_pages);
879 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
881 list_del_init(&oap->oap_rpc_item);
882 list_del_init(&oap->oap_pending_item);
883 if (last_off <= oap->oap_obj_off) {
884 last_off = oap->oap_obj_off;
885 last_count = oap->oap_count;
889 osc_ap_completion(env, cli, oap, sent, rc);
891 EASSERT(ext->oe_nr_pages == 0, ext);
894 lost_grant = ext->oe_grants;
895 } else if (blocksize < PAGE_SIZE &&
896 last_count != PAGE_SIZE) {
897 /* For short writes we shouldn't count parts of pages that
898 * span a whole chunk on the OST side, or our accounting goes
899 * wrong. Should match the code in filter_grant_check. */
900 int offset = last_off & ~PAGE_MASK;
901 int count = last_count + (offset & (blocksize - 1));
902 int end = (offset + last_count) & (blocksize - 1);
904 count += blocksize - end;
906 lost_grant = PAGE_SIZE - count;
908 if (ext->oe_grants > 0)
909 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
911 osc_extent_remove(ext);
912 /* put the refcount for RPC */
913 osc_extent_put(env, ext);
917 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
921 osc_object_lock(ext->oe_obj);
922 ret = ext->oe_state == state;
923 osc_object_unlock(ext->oe_obj);
929 * Wait for the extent's state to become @state.
931 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
932 enum osc_extent_state state)
934 struct osc_object *obj = ext->oe_obj;
935 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
936 LWI_ON_SIGNAL_NOOP, NULL);
940 osc_object_lock(obj);
941 LASSERT(sanity_check_nolock(ext) == 0);
942 /* `Kick' this extent only if the caller is waiting for it to be
944 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
945 if (ext->oe_state == OES_ACTIVE) {
947 } else if (ext->oe_state == OES_CACHE) {
949 osc_extent_hold(ext);
953 osc_object_unlock(obj);
955 osc_extent_release(env, ext);
957 /* wait for the extent until its state becomes @state */
958 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
959 if (rc == -ETIMEDOUT) {
960 OSC_EXTENT_DUMP(D_ERROR, ext,
961 "%s: wait ext to %u timedout, recovery in progress?\n",
962 cli_name(osc_cli(obj)), state);
964 lwi = LWI_INTR(NULL, NULL);
965 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
968 if (rc == 0 && ext->oe_rc < 0)
974 * Discard pages with index greater than @size. If @ext is overlapped with
975 * @size, then partial truncate happens.
977 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
982 struct osc_object *obj = ext->oe_obj;
983 struct client_obd *cli = osc_cli(obj);
984 struct osc_async_page *oap;
985 struct osc_async_page *tmp;
986 struct pagevec *pvec;
987 int pages_in_chunk = 0;
988 int ppc_bits = cli->cl_chunkbits -
990 __u64 trunc_chunk = trunc_index >> ppc_bits;
997 LASSERT(sanity_check(ext) == 0);
998 LASSERT(ext->oe_state == OES_TRUNC);
999 LASSERT(!ext->oe_urgent);
1001 /* Request new lu_env.
1002 * We can't use that env from osc_cache_truncate_start() because
1003 * it's from lov_io_sub and not fully initialized. */
1004 env = cl_env_get(&refcheck);
1006 RETURN(PTR_ERR(env));
1008 io = osc_env_thread_io(env);
1009 io->ci_obj = cl_object_top(osc2cl(obj));
1010 io->ci_ignore_layout = 1;
1011 pvec = &osc_env_info(env)->oti_pagevec;
1012 ll_pagevec_init(pvec, 0);
1013 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1017 /* discard all pages with index greater than trunc_index */
1018 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1020 pgoff_t index = osc_index(oap2osc(oap));
1021 struct cl_page *page = oap2cl_page(oap);
1023 LASSERT(list_empty(&oap->oap_rpc_item));
1025 /* only discard the pages with their index greater than
1026 * trunc_index, and ... */
1027 if (index < trunc_index ||
1028 (index == trunc_index && partial)) {
1029 /* accounting how many pages remaining in the chunk
1030 * so that we can calculate grants correctly. */
1031 if (index >> ppc_bits == trunc_chunk)
1036 list_del_init(&oap->oap_pending_item);
1039 lu_ref_add(&page->cp_reference, "truncate", current);
1041 if (cl_page_own(env, io, page) == 0) {
1042 cl_page_discard(env, io, page);
1043 cl_page_disown(env, io, page);
1045 LASSERT(page->cp_state == CPS_FREEING);
1049 lu_ref_del(&page->cp_reference, "truncate", current);
1050 cl_pagevec_put(env, page, pvec);
1055 pagevec_release(pvec);
1057 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1058 ext->oe_nr_pages == 0),
1059 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1061 osc_object_lock(obj);
1062 if (ext->oe_nr_pages == 0) {
1063 LASSERT(pages_in_chunk == 0);
1064 grants = ext->oe_grants;
1066 } else { /* calculate how many grants we can free */
1067 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1071 /* if there is no pages in this chunk, we can also free grants
1072 * for the last chunk */
1073 if (pages_in_chunk == 0) {
1074 /* if this is the 1st chunk and no pages in this chunk,
1075 * ext->oe_nr_pages must be zero, so we should be in
1076 * the other if-clause. */
1077 LASSERT(trunc_chunk > 0);
1082 /* this is what we can free from this extent */
1083 grants = chunks << cli->cl_chunkbits;
1084 ext->oe_grants -= grants;
1085 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1086 ext->oe_end = min(last_index, ext->oe_max_end);
1087 LASSERT(ext->oe_end >= ext->oe_start);
1088 LASSERT(ext->oe_grants > 0);
1090 osc_object_unlock(obj);
1092 if (grants > 0 || nr_pages > 0)
1093 osc_free_grant(cli, nr_pages, grants, grants);
1096 cl_io_fini(env, io);
1097 cl_env_put(env, &refcheck);
1102 * This function is used to make the extent prepared for transfer.
1103 * A race with flusing page - ll_writepage() has to be handled cautiously.
1105 static int osc_extent_make_ready(const struct lu_env *env,
1106 struct osc_extent *ext)
1108 struct osc_async_page *oap;
1109 struct osc_async_page *last = NULL;
1110 struct osc_object *obj = ext->oe_obj;
1111 unsigned int page_count = 0;
1115 /* we're going to grab page lock, so object lock must not be taken. */
1116 LASSERT(sanity_check(ext) == 0);
1117 /* in locking state, any process should not touch this extent. */
1118 EASSERT(ext->oe_state == OES_LOCKING, ext);
1119 EASSERT(ext->oe_owner != NULL, ext);
1121 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1123 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1125 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1128 /* checking ASYNC_READY is race safe */
1129 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1132 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1135 spin_lock(&oap->oap_lock);
1136 oap->oap_async_flags |= ASYNC_READY;
1137 spin_unlock(&oap->oap_lock);
1140 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1143 LASSERTF(0, "unknown return code: %d\n", rc);
1147 LASSERT(page_count == ext->oe_nr_pages);
1148 LASSERT(last != NULL);
1149 /* the last page is the only one we need to refresh its count by
1150 * the size of file. */
1151 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1152 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1153 LASSERT(last_oap_count > 0);
1154 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1155 last->oap_count = last_oap_count;
1156 spin_lock(&last->oap_lock);
1157 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1158 spin_unlock(&last->oap_lock);
1161 /* for the rest of pages, we don't need to call osf_refresh_count()
1162 * because it's known they are not the last page */
1163 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1164 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1165 oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1166 spin_lock(&oap->oap_lock);
1167 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1168 spin_unlock(&oap->oap_lock);
1172 osc_object_lock(obj);
1173 osc_extent_state_set(ext, OES_RPC);
1174 osc_object_unlock(obj);
1175 /* get a refcount for RPC. */
1176 osc_extent_get(ext);
1182 * Quick and simple version of osc_extent_find(). This function is frequently
1183 * called to expand the extent for the same IO. To expand the extent, the
1184 * page index must be in the same or next chunk of ext->oe_end.
1186 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1187 unsigned int *grants)
1189 struct osc_object *obj = ext->oe_obj;
1190 struct client_obd *cli = osc_cli(obj);
1191 struct osc_extent *next;
1192 int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1193 pgoff_t chunk = index >> ppc_bits;
1196 unsigned int chunksize = 1 << cli->cl_chunkbits;
1200 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1201 osc_object_lock(obj);
1202 LASSERT(sanity_check_nolock(ext) == 0);
1203 end_chunk = ext->oe_end >> ppc_bits;
1204 if (chunk > end_chunk + 1)
1205 GOTO(out, rc = -ERANGE);
1207 if (end_chunk >= chunk)
1210 LASSERT(end_chunk + 1 == chunk);
1212 /* try to expand this extent to cover @index */
1213 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1215 /* don't go over the maximum extent size reported by server */
1216 if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1217 GOTO(out, rc = -ERANGE);
1219 next = next_extent(ext);
1220 if (next != NULL && next->oe_start <= end_index)
1221 /* complex mode - overlapped with the next extent,
1222 * this case will be handled by osc_extent_find() */
1223 GOTO(out, rc = -EAGAIN);
1225 ext->oe_end = end_index;
1226 ext->oe_grants += chunksize;
1227 LASSERT(*grants >= chunksize);
1228 *grants -= chunksize;
1229 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1230 "overlapped after expanding for %lu.\n", index);
1234 osc_object_unlock(obj);
1238 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1239 const char *func, int line)
1241 struct osc_extent *ext;
1244 if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1247 CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1248 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1250 /* osc_object_lock(obj); */
1252 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1253 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1256 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1257 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1260 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1261 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1264 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1265 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1266 /* osc_object_unlock(obj); */
1269 /* ------------------ osc extent end ------------------ */
1271 static inline int osc_is_ready(struct osc_object *osc)
1273 return !list_empty(&osc->oo_ready_item) ||
1274 !list_empty(&osc->oo_hp_ready_item);
1277 #define OSC_IO_DEBUG(OSC, STR, args...) \
1278 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1279 (OSC), osc_is_ready(OSC), \
1280 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1281 list_empty_marker(&(OSC)->oo_ready_item), \
1282 atomic_read(&(OSC)->oo_nr_writes), \
1283 list_empty_marker(&(OSC)->oo_hp_exts), \
1284 list_empty_marker(&(OSC)->oo_urgent_exts), \
1285 atomic_read(&(OSC)->oo_nr_reads), \
1286 list_empty_marker(&(OSC)->oo_reading_exts), \
1289 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1292 struct osc_page *opg = oap2osc_page(oap);
1293 struct cl_page *page = oap2cl_page(oap);
1296 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1299 result = cl_page_make_ready(env, page, CRT_WRITE);
1301 opg->ops_submit_time = ktime_get();
1305 static int osc_refresh_count(const struct lu_env *env,
1306 struct osc_async_page *oap, int cmd)
1308 struct osc_page *opg = oap2osc_page(oap);
1309 pgoff_t index = osc_index(oap2osc(oap));
1310 struct cl_object *obj;
1311 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1315 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1316 LASSERT(!(cmd & OBD_BRW_READ));
1317 LASSERT(opg != NULL);
1318 obj = opg->ops_cl.cpl_obj;
1320 cl_object_attr_lock(obj);
1321 result = cl_object_attr_get(env, obj, attr);
1322 cl_object_attr_unlock(obj);
1325 kms = attr->cat_kms;
1326 if (cl_offset(obj, index) >= kms)
1327 /* catch race with truncate */
1329 else if (cl_offset(obj, index + 1) > kms)
1330 /* catch sub-page write at end of file */
1331 return kms % PAGE_SIZE;
1336 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1339 struct osc_page *opg = oap2osc_page(oap);
1340 struct cl_page *page = oap2cl_page(oap);
1341 enum cl_req_type crt;
1346 cmd &= ~OBD_BRW_NOQUOTA;
1347 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1348 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1349 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1350 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1351 LASSERT(opg->ops_transfer_pinned);
1353 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1354 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1355 opg->ops_transfer_pinned = 0;
1357 opg->ops_submit_time = ktime_set(0, 0);
1358 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1361 if (rc == 0 && srvlock) {
1362 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1363 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1364 size_t bytes = oap->oap_count;
1366 if (crt == CRT_READ)
1367 stats->os_lockless_reads += bytes;
1369 stats->os_lockless_writes += bytes;
1373 * This has to be the last operation with the page, as locks are
1374 * released in cl_page_completion() and nothing except for the
1375 * reference counter protects page from concurrent reclaim.
1377 lu_ref_del(&page->cp_reference, "transfer", page);
1379 cl_page_completion(env, page, crt, rc);
1380 cl_page_put(env, page);
1385 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \
1386 struct client_obd *__tmp = (cli); \
1387 CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1388 "dropped: %ld avail: %ld, dirty_grant: %ld, " \
1389 "reserved: %ld, flight: %d } lru {in list: %ld, " \
1390 "left: %ld, waiters: %d }" fmt "\n", \
1392 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1393 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1394 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1395 __tmp->cl_dirty_grant, \
1396 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1397 atomic_long_read(&__tmp->cl_lru_in_list), \
1398 atomic_long_read(&__tmp->cl_lru_busy), \
1399 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1402 /* caller must hold loi_list_lock */
1403 static void osc_consume_write_grant(struct client_obd *cli,
1404 struct brw_page *pga)
1406 assert_spin_locked(&cli->cl_loi_list_lock);
1407 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1408 cli->cl_dirty_pages++;
1409 pga->flag |= OBD_BRW_FROM_GRANT;
1410 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1411 PAGE_SIZE, pga, pga->pg);
1412 osc_update_next_shrink(cli);
1415 /* the companion to osc_consume_write_grant, called when a brw has completed.
1416 * must be called with the loi lock held. */
1417 static void osc_release_write_grant(struct client_obd *cli,
1418 struct brw_page *pga)
1422 assert_spin_locked(&cli->cl_loi_list_lock);
1423 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1428 pga->flag &= ~OBD_BRW_FROM_GRANT;
1429 atomic_long_dec(&obd_dirty_pages);
1430 cli->cl_dirty_pages--;
1431 if (pga->flag & OBD_BRW_NOCACHE) {
1432 pga->flag &= ~OBD_BRW_NOCACHE;
1433 atomic_long_dec(&obd_dirty_transit_pages);
1434 cli->cl_dirty_transit--;
1440 * To avoid sleeping with object lock held, it's good for us allocate enough
1441 * grants before entering into critical section.
1443 * client_obd_list_lock held by caller
1445 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1449 if (cli->cl_avail_grant >= bytes) {
1450 cli->cl_avail_grant -= bytes;
1451 cli->cl_reserved_grant += bytes;
1457 static void __osc_unreserve_grant(struct client_obd *cli,
1458 unsigned int reserved, unsigned int unused)
1460 /* it's quite normal for us to get more grant than reserved.
1461 * Thinking about a case that two extents merged by adding a new
1462 * chunk, we can save one extent tax. If extent tax is greater than
1463 * one chunk, we can save more grant by adding a new chunk */
1464 cli->cl_reserved_grant -= reserved;
1465 if (unused > reserved) {
1466 cli->cl_avail_grant += reserved;
1467 cli->cl_lost_grant += unused - reserved;
1468 cli->cl_dirty_grant -= unused - reserved;
1470 cli->cl_avail_grant += unused;
1471 cli->cl_dirty_grant += reserved - unused;
1475 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1476 unsigned int reserved,
1477 unsigned int unused)
1479 __osc_unreserve_grant(cli, reserved, unused);
1481 osc_wake_cache_waiters(cli);
1484 static void osc_unreserve_grant(struct client_obd *cli,
1485 unsigned int reserved, unsigned int unused)
1487 spin_lock(&cli->cl_loi_list_lock);
1488 osc_unreserve_grant_nolock(cli, reserved, unused);
1489 spin_unlock(&cli->cl_loi_list_lock);
1493 * Free grant after IO is finished or canceled.
1495 * @lost_grant is used to remember how many grants we have allocated but not
1496 * used, we should return these grants to OST. There're two cases where grants
1499 * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1500 * written. In this case OST may use less chunks to serve this partial
1501 * write. OSTs don't actually know the page size on the client side. so
1502 * clients have to calculate lost grant by the blocksize on the OST.
1503 * See filter_grant_check() for details.
1505 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1506 unsigned int lost_grant, unsigned int dirty_grant)
1508 unsigned long grant;
1510 grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1512 spin_lock(&cli->cl_loi_list_lock);
1513 atomic_long_sub(nr_pages, &obd_dirty_pages);
1514 cli->cl_dirty_pages -= nr_pages;
1515 cli->cl_lost_grant += lost_grant;
1516 cli->cl_dirty_grant -= dirty_grant;
1517 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1518 /* borrow some grant from truncate to avoid the case that
1519 * truncate uses up all avail grant */
1520 cli->cl_lost_grant -= grant;
1521 cli->cl_avail_grant += grant;
1523 osc_wake_cache_waiters(cli);
1524 spin_unlock(&cli->cl_loi_list_lock);
1525 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1526 lost_grant, cli->cl_lost_grant,
1527 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1528 cli->cl_dirty_grant);
1532 * The companion to osc_enter_cache(), called when @oap is no longer part of
1533 * the dirty accounting due to error.
1535 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1537 spin_lock(&cli->cl_loi_list_lock);
1538 osc_release_write_grant(cli, &oap->oap_brw_page);
1539 spin_unlock(&cli->cl_loi_list_lock);
1543 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1546 static int osc_enter_cache_try(struct client_obd *cli,
1547 struct osc_async_page *oap,
1548 int bytes, int transient)
1552 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1554 rc = osc_reserve_grant(cli, bytes);
1558 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1559 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1560 obd_max_dirty_pages) {
1561 osc_consume_write_grant(cli, &oap->oap_brw_page);
1563 cli->cl_dirty_transit++;
1564 atomic_long_inc(&obd_dirty_transit_pages);
1565 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1570 atomic_long_dec(&obd_dirty_pages);
1572 __osc_unreserve_grant(cli, bytes, bytes);
1578 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1581 spin_lock(&cli->cl_loi_list_lock);
1582 rc = list_empty(&ocw->ocw_entry);
1583 spin_unlock(&cli->cl_loi_list_lock);
1588 * The main entry to reserve dirty page accounting. Usually the grant reserved
1589 * in this function will be freed in bulk in osc_free_grant() unless it fails
1590 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1592 * The process will be put into sleep if it's already run out of grant.
1594 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1595 struct osc_async_page *oap, int bytes)
1597 struct osc_object *osc = oap->oap_obj;
1598 struct lov_oinfo *loi = osc->oo_oinfo;
1599 struct osc_cache_waiter ocw;
1600 struct l_wait_info lwi;
1604 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1605 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1607 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1609 spin_lock(&cli->cl_loi_list_lock);
1611 /* force the caller to try sync io. this can jump the list
1612 * of queued writes and create a discontiguous rpc stream */
1613 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1614 cli->cl_dirty_max_pages == 0 ||
1615 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1616 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1617 GOTO(out, rc = -EDQUOT);
1620 /* Hopefully normal case - cache space and write credits available */
1621 if (list_empty(&cli->cl_cache_waiters) &&
1622 osc_enter_cache_try(cli, oap, bytes, 0)) {
1623 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1627 /* We can get here for two reasons: too many dirty pages in cache, or
1628 * run out of grants. In both cases we should write dirty pages out.
1629 * Adding a cache waiter will trigger urgent write-out no matter what
1631 * The exiting condition is no avail grants and no dirty pages caching,
1632 * that really means there is no space on the OST. */
1633 init_waitqueue_head(&ocw.ocw_waitq);
1635 ocw.ocw_grant = bytes;
1636 while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1637 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1639 spin_unlock(&cli->cl_loi_list_lock);
1641 osc_io_unplug_async(env, cli, NULL);
1643 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1644 cli_name(cli), &ocw, oap);
1646 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1648 spin_lock(&cli->cl_loi_list_lock);
1651 /* l_wait_event is interrupted by signal or timed out */
1652 list_del_init(&ocw.ocw_entry);
1655 LASSERT(list_empty(&ocw.ocw_entry));
1660 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1668 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1671 OSC_DUMP_GRANT(D_CACHE, cli,
1672 "timeout, fall back to sync i/o\n");
1673 osc_extent_tree_dump(D_CACHE, osc);
1674 /* fall back to synchronous I/O */
1678 /* Ensures restartability - LU-3581 */
1679 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1683 OSC_DUMP_GRANT(D_CACHE, cli,
1684 "no grant space, fall back to sync i/o\n");
1687 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1688 "due to %d, fall back to sync i/o\n",
1689 cli_name(cli), &ocw, rc);
1694 spin_unlock(&cli->cl_loi_list_lock);
1698 /* caller must hold loi_list_lock */
1699 void osc_wake_cache_waiters(struct client_obd *cli)
1701 struct list_head *l, *tmp;
1702 struct osc_cache_waiter *ocw;
1705 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1706 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1708 ocw->ocw_rc = -EDQUOT;
1710 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1713 if (ocw->ocw_rc == 0 ||
1714 !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1715 list_del_init(&ocw->ocw_entry);
1716 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1717 "%ld, %d\n", ocw, ocw->ocw_oap,
1718 cli->cl_avail_grant, ocw->ocw_rc);
1720 wake_up(&ocw->ocw_waitq);
1726 EXPORT_SYMBOL(osc_wake_cache_waiters);
1728 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1730 int hprpc = !!list_empty(&osc->oo_hp_exts);
1731 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1734 /* This maintains the lists of pending pages to read/write for a given object
1735 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1736 * to quickly find objects that are ready to send an RPC. */
1737 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1740 int invalid_import = 0;
1743 /* if we have an invalid import we want to drain the queued pages
1744 * by forcing them through rpcs that immediately fail and complete
1745 * the pages. recovery relies on this to empty the queued pages
1746 * before canceling the locks and evicting down the llite pages */
1747 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1750 if (cmd & OBD_BRW_WRITE) {
1751 if (atomic_read(&osc->oo_nr_writes) == 0)
1753 if (invalid_import) {
1754 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1757 if (!list_empty(&osc->oo_hp_exts)) {
1758 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1761 if (!list_empty(&osc->oo_urgent_exts)) {
1762 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1765 /* trigger a write rpc stream as long as there are dirtiers
1766 * waiting for space. as they're waiting, they're not going to
1767 * create more pages to coalesce with what's waiting.. */
1768 if (!list_empty(&cli->cl_cache_waiters)) {
1769 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1772 if (!list_empty(&osc->oo_full_exts)) {
1773 CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1777 if (atomic_read(&osc->oo_nr_reads) == 0)
1779 if (invalid_import) {
1780 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1783 /* all read are urgent. */
1784 if (!list_empty(&osc->oo_reading_exts))
1791 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1793 struct client_obd *cli = osc_cli(obj);
1794 if (cmd & OBD_BRW_WRITE) {
1795 atomic_add(delta, &obj->oo_nr_writes);
1796 atomic_add(delta, &cli->cl_pending_w_pages);
1797 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1799 atomic_add(delta, &obj->oo_nr_reads);
1800 atomic_add(delta, &cli->cl_pending_r_pages);
1801 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1803 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1806 static int osc_makes_hprpc(struct osc_object *obj)
1808 return !list_empty(&obj->oo_hp_exts);
1811 static void on_list(struct list_head *item, struct list_head *list,
1814 if (list_empty(item) && should_be_on)
1815 list_add_tail(item, list);
1816 else if (!list_empty(item) && !should_be_on)
1817 list_del_init(item);
1820 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1821 * can find pages to build into rpcs quickly */
1822 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1824 if (osc_makes_hprpc(osc)) {
1826 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1827 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1829 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1830 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1831 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1832 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1835 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1836 atomic_read(&osc->oo_nr_writes) > 0);
1838 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1839 atomic_read(&osc->oo_nr_reads) > 0);
1841 return osc_is_ready(osc);
1844 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1848 spin_lock(&cli->cl_loi_list_lock);
1849 is_ready = __osc_list_maint(cli, osc);
1850 spin_unlock(&cli->cl_loi_list_lock);
1855 /* this is trying to propogate async writeback errors back up to the
1856 * application. As an async write fails we record the error code for later if
1857 * the app does an fsync. As long as errors persist we force future rpcs to be
1858 * sync so that the app can get a sync error and break the cycle of queueing
1859 * pages for which writeback will fail. */
1860 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1867 ar->ar_force_sync = 1;
1868 ar->ar_min_xid = ptlrpc_sample_next_xid();
1873 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1874 ar->ar_force_sync = 0;
1877 /* this must be called holding the loi list lock to give coverage to exit_cache,
1878 * async_flag maintenance, and oap_request */
1879 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1880 struct osc_async_page *oap, int sent, int rc)
1882 struct osc_object *osc = oap->oap_obj;
1883 struct lov_oinfo *loi = osc->oo_oinfo;
1887 if (oap->oap_request != NULL) {
1888 xid = ptlrpc_req_xid(oap->oap_request);
1889 ptlrpc_req_finished(oap->oap_request);
1890 oap->oap_request = NULL;
1893 /* As the transfer for this page is being done, clear the flags */
1894 spin_lock(&oap->oap_lock);
1895 oap->oap_async_flags = 0;
1896 spin_unlock(&oap->oap_lock);
1897 oap->oap_interrupted = 0;
1899 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1900 spin_lock(&cli->cl_loi_list_lock);
1901 osc_process_ar(&cli->cl_ar, xid, rc);
1902 osc_process_ar(&loi->loi_ar, xid, rc);
1903 spin_unlock(&cli->cl_loi_list_lock);
1906 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1908 CERROR("completion on oap %p obj %p returns %d.\n",
1914 struct extent_rpc_data {
1915 struct list_head *erd_rpc_list;
1916 unsigned int erd_page_count;
1917 unsigned int erd_max_pages;
1918 unsigned int erd_max_chunks;
1919 unsigned int erd_max_extents;
1922 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1924 struct client_obd *cli = osc_cli(ext->oe_obj);
1925 unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1927 return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1931 * Try to add extent to one RPC. We need to think about the following things:
1932 * - # of pages must not be over max_pages_per_rpc
1933 * - extent must be compatible with previous ones
1935 static int try_to_add_extent_for_io(struct client_obd *cli,
1936 struct osc_extent *ext,
1937 struct extent_rpc_data *data)
1939 struct osc_extent *tmp;
1940 unsigned int chunk_count;
1941 struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1942 struct osc_async_page,
1946 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1948 OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1950 if (data->erd_max_extents == 0)
1953 chunk_count = osc_extent_chunks(ext);
1954 EASSERTF(data->erd_page_count != 0 ||
1955 chunk_count <= data->erd_max_chunks, ext,
1956 "The first extent to be fit in a RPC contains %u chunks, "
1957 "which is over the limit %u.\n", chunk_count,
1958 data->erd_max_chunks);
1959 if (chunk_count > data->erd_max_chunks)
1962 data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1963 EASSERTF(data->erd_page_count != 0 ||
1964 ext->oe_nr_pages <= data->erd_max_pages, ext,
1965 "The first extent to be fit in a RPC contains %u pages, "
1966 "which is over the limit %u.\n", ext->oe_nr_pages,
1967 data->erd_max_pages);
1968 if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1971 list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1972 struct osc_async_page *oap2;
1973 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1975 EASSERT(tmp->oe_owner == current, tmp);
1977 if (overlapped(tmp, ext)) {
1978 OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1982 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1983 CDEBUG(D_CACHE, "Do not permit different types of IO "
1988 if (tmp->oe_srvlock != ext->oe_srvlock ||
1989 !tmp->oe_grants != !ext->oe_grants ||
1990 tmp->oe_ndelay != ext->oe_ndelay ||
1991 tmp->oe_no_merge || ext->oe_no_merge)
1994 /* remove break for strict check */
1998 data->erd_max_extents--;
1999 data->erd_max_chunks -= chunk_count;
2000 data->erd_page_count += ext->oe_nr_pages;
2001 list_move_tail(&ext->oe_link, data->erd_rpc_list);
2002 ext->oe_owner = current;
2007 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2008 * get_write_extent() takes all appropriate extents in atomic.
2010 * The following policy is used to collect extents for IO:
2011 * 1. Add as many HP extents as possible;
2012 * 2. Add the first urgent extent in urgent extent list and take it out of
2014 * 3. Add subsequent extents of this urgent extent;
2015 * 4. If urgent list is not empty, goto 2;
2016 * 5. Traverse the extent tree from the 1st extent;
2017 * 6. Above steps exit if there is no space in this RPC.
2019 static unsigned int get_write_extents(struct osc_object *obj,
2020 struct list_head *rpclist)
2022 struct client_obd *cli = osc_cli(obj);
2023 struct osc_extent *ext;
2024 struct extent_rpc_data data = {
2025 .erd_rpc_list = rpclist,
2026 .erd_page_count = 0,
2027 .erd_max_pages = cli->cl_max_pages_per_rpc,
2028 .erd_max_chunks = osc_max_write_chunks(cli),
2029 .erd_max_extents = 256,
2032 LASSERT(osc_object_is_locked(obj));
2033 while (!list_empty(&obj->oo_hp_exts)) {
2034 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2036 LASSERT(ext->oe_state == OES_CACHE);
2037 if (!try_to_add_extent_for_io(cli, ext, &data))
2038 return data.erd_page_count;
2039 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2041 if (data.erd_page_count == data.erd_max_pages)
2042 return data.erd_page_count;
2044 while (!list_empty(&obj->oo_urgent_exts)) {
2045 ext = list_entry(obj->oo_urgent_exts.next,
2046 struct osc_extent, oe_link);
2047 if (!try_to_add_extent_for_io(cli, ext, &data))
2048 return data.erd_page_count;
2050 if (data.erd_page_count == data.erd_max_pages)
2051 return data.erd_page_count;
2053 /* One key difference between full extents and other extents: full
2054 * extents can usually only be added if the rpclist was empty, so if we
2055 * can't add one, we continue on to trying to add normal extents. This
2056 * is so we don't miss adding extra extents to an RPC containing high
2057 * priority or urgent extents. */
2058 while (!list_empty(&obj->oo_full_exts)) {
2059 ext = list_entry(obj->oo_full_exts.next,
2060 struct osc_extent, oe_link);
2061 if (!try_to_add_extent_for_io(cli, ext, &data))
2064 if (data.erd_page_count == data.erd_max_pages)
2065 return data.erd_page_count;
2067 ext = first_extent(obj);
2068 while (ext != NULL) {
2069 if ((ext->oe_state != OES_CACHE) ||
2070 /* this extent may be already in current rpclist */
2071 (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2072 ext = next_extent(ext);
2076 if (!try_to_add_extent_for_io(cli, ext, &data))
2077 return data.erd_page_count;
2079 ext = next_extent(ext);
2081 return data.erd_page_count;
2085 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2086 struct osc_object *osc)
2089 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2090 struct osc_extent *ext;
2091 struct osc_extent *tmp;
2092 struct osc_extent *first = NULL;
2093 unsigned int page_count = 0;
2098 LASSERT(osc_object_is_locked(osc));
2100 page_count = get_write_extents(osc, &rpclist);
2101 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2103 if (list_empty(&rpclist))
2106 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2108 list_for_each_entry(ext, &rpclist, oe_link) {
2109 LASSERT(ext->oe_state == OES_CACHE ||
2110 ext->oe_state == OES_LOCK_DONE);
2111 if (ext->oe_state == OES_CACHE)
2112 osc_extent_state_set(ext, OES_LOCKING);
2114 osc_extent_state_set(ext, OES_RPC);
2117 /* we're going to grab page lock, so release object lock because
2118 * lock order is page lock -> object lock. */
2119 osc_object_unlock(osc);
2121 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2122 if (ext->oe_state == OES_LOCKING) {
2123 rc = osc_extent_make_ready(env, ext);
2124 if (unlikely(rc < 0)) {
2125 list_del_init(&ext->oe_link);
2126 osc_extent_finish(env, ext, 0, rc);
2130 if (first == NULL) {
2132 srvlock = ext->oe_srvlock;
2134 LASSERT(srvlock == ext->oe_srvlock);
2138 if (!list_empty(&rpclist)) {
2139 LASSERT(page_count > 0);
2140 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2141 LASSERT(list_empty(&rpclist));
2144 osc_object_lock(osc);
2149 * prepare pages for ASYNC io and put pages in send queue.
2151 * \param cmd OBD_BRW_* macroses
2152 * \param lop pending pages
2154 * \return zero if no page added to send queue.
2155 * \return 1 if pages successfully added to send queue.
2156 * \return negative on errors.
2159 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2160 struct osc_object *osc)
2163 struct osc_extent *ext;
2164 struct osc_extent *next;
2165 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2166 struct extent_rpc_data data = {
2167 .erd_rpc_list = &rpclist,
2168 .erd_page_count = 0,
2169 .erd_max_pages = cli->cl_max_pages_per_rpc,
2170 .erd_max_chunks = UINT_MAX,
2171 .erd_max_extents = UINT_MAX,
2176 LASSERT(osc_object_is_locked(osc));
2177 list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2178 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2179 if (!try_to_add_extent_for_io(cli, ext, &data))
2181 osc_extent_state_set(ext, OES_RPC);
2182 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2184 LASSERT(data.erd_page_count <= data.erd_max_pages);
2186 osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2188 if (!list_empty(&rpclist)) {
2189 osc_object_unlock(osc);
2191 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2192 LASSERT(list_empty(&rpclist));
2194 osc_object_lock(osc);
2199 #define list_to_obj(list, item) ({ \
2200 struct list_head *__tmp = (list)->next; \
2201 list_del_init(__tmp); \
2202 list_entry(__tmp, struct osc_object, oo_##item); \
2205 /* This is called by osc_check_rpcs() to find which objects have pages that
2206 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2207 static struct osc_object *osc_next_obj(struct client_obd *cli)
2211 /* First return objects that have blocked locks so that they
2212 * will be flushed quickly and other clients can get the lock,
2213 * then objects which have pages ready to be stuffed into RPCs */
2214 if (!list_empty(&cli->cl_loi_hp_ready_list))
2215 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2216 if (!list_empty(&cli->cl_loi_ready_list))
2217 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2219 /* then if we have cache waiters, return all objects with queued
2220 * writes. This is especially important when many small files
2221 * have filled up the cache and not been fired into rpcs because
2222 * they don't pass the nr_pending/object threshhold */
2223 if (!list_empty(&cli->cl_cache_waiters) &&
2224 !list_empty(&cli->cl_loi_write_list))
2225 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2227 /* then return all queued objects when we have an invalid import
2228 * so that they get flushed */
2229 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2230 if (!list_empty(&cli->cl_loi_write_list))
2231 RETURN(list_to_obj(&cli->cl_loi_write_list,
2233 if (!list_empty(&cli->cl_loi_read_list))
2234 RETURN(list_to_obj(&cli->cl_loi_read_list,
2240 /* called with the loi list lock held */
2241 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2242 __must_hold(&cli->cl_loi_list_lock)
2244 struct osc_object *osc;
2248 while ((osc = osc_next_obj(cli)) != NULL) {
2249 struct cl_object *obj = osc2cl(osc);
2250 struct lu_ref_link link;
2252 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2254 if (osc_max_rpc_in_flight(cli, osc)) {
2255 __osc_list_maint(cli, osc);
2260 spin_unlock(&cli->cl_loi_list_lock);
2261 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2263 /* attempt some read/write balancing by alternating between
2264 * reads and writes in an object. The makes_rpc checks here
2265 * would be redundant if we were getting read/write work items
2266 * instead of objects. we don't want send_oap_rpc to drain a
2267 * partial read pending queue when we're given this object to
2268 * do io on writes while there are cache waiters */
2269 osc_object_lock(osc);
2270 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2271 rc = osc_send_write_rpc(env, cli, osc);
2273 CERROR("Write request failed with %d\n", rc);
2275 /* osc_send_write_rpc failed, mostly because of
2278 * It can't break here, because if:
2279 * - a page was submitted by osc_io_submit, so
2281 * - no request in flight
2282 * - no subsequent request
2283 * The system will be in live-lock state,
2284 * because there is no chance to call
2285 * osc_io_unplug() and osc_check_rpcs() any
2286 * more. pdflush can't help in this case,
2287 * because it might be blocked at grabbing
2288 * the page lock as we mentioned.
2290 * Anyway, continue to drain pages. */
2294 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2295 rc = osc_send_read_rpc(env, cli, osc);
2297 CERROR("Read request failed with %d\n", rc);
2299 osc_object_unlock(osc);
2301 osc_list_maint(cli, osc);
2302 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2303 cl_object_put(env, obj);
2305 spin_lock(&cli->cl_loi_list_lock);
2309 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2310 struct osc_object *osc, int async)
2314 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2318 spin_lock(&cli->cl_loi_list_lock);
2319 osc_check_rpcs(env, cli);
2320 spin_unlock(&cli->cl_loi_list_lock);
2322 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2323 LASSERT(cli->cl_writeback_work != NULL);
2324 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2328 EXPORT_SYMBOL(osc_io_unplug0);
2330 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2331 struct page *page, loff_t offset)
2333 struct obd_export *exp = osc_export(osc);
2334 struct osc_async_page *oap = &ops->ops_oap;
2338 return cfs_size_round(sizeof(*oap));
2340 oap->oap_magic = OAP_MAGIC;
2341 oap->oap_cli = &exp->exp_obd->u.cli;
2344 oap->oap_page = page;
2345 oap->oap_obj_off = offset;
2346 LASSERT(!(offset & ~PAGE_MASK));
2348 INIT_LIST_HEAD(&oap->oap_pending_item);
2349 INIT_LIST_HEAD(&oap->oap_rpc_item);
2351 spin_lock_init(&oap->oap_lock);
2352 CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2353 oap, page, oap->oap_obj_off);
2356 EXPORT_SYMBOL(osc_prep_async_page);
2358 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2359 struct osc_page *ops)
2361 struct osc_io *oio = osc_env_io(env);
2362 struct osc_extent *ext = NULL;
2363 struct osc_async_page *oap = &ops->ops_oap;
2364 struct client_obd *cli = oap->oap_cli;
2365 struct osc_object *osc = oap->oap_obj;
2368 unsigned int grants = 0;
2369 u32 brw_flags = OBD_BRW_ASYNC;
2370 int cmd = OBD_BRW_WRITE;
2371 int need_release = 0;
2375 if (oap->oap_magic != OAP_MAGIC)
2378 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2381 if (!list_empty(&oap->oap_pending_item) ||
2382 !list_empty(&oap->oap_rpc_item))
2385 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2386 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2387 if (oio->oi_cap_sys_resource) {
2388 brw_flags |= OBD_BRW_NOQUOTA;
2389 cmd |= OBD_BRW_NOQUOTA;
2392 /* check if the file's owner/group is over quota */
2393 if (!(cmd & OBD_BRW_NOQUOTA)) {
2394 struct cl_object *obj;
2395 struct cl_attr *attr;
2396 unsigned int qid[LL_MAXQUOTAS];
2398 obj = cl_object_top(&osc->oo_cl);
2399 attr = &osc_env_info(env)->oti_attr;
2401 cl_object_attr_lock(obj);
2402 rc = cl_object_attr_get(env, obj, attr);
2403 cl_object_attr_unlock(obj);
2405 qid[USRQUOTA] = attr->cat_uid;
2406 qid[GRPQUOTA] = attr->cat_gid;
2407 qid[PRJQUOTA] = attr->cat_projid;
2408 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2415 oap->oap_page_off = ops->ops_from;
2416 oap->oap_count = ops->ops_to - ops->ops_from;
2417 /* No need to hold a lock here,
2418 * since this page is not in any list yet. */
2419 oap->oap_async_flags = 0;
2420 oap->oap_brw_flags = brw_flags;
2422 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2423 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2425 index = osc_index(oap2osc(oap));
2427 /* Add this page into extent by the following steps:
2428 * 1. if there exists an active extent for this IO, mostly this page
2429 * can be added to the active extent and sometimes we need to
2430 * expand extent to accomodate this page;
2431 * 2. otherwise, a new extent will be allocated. */
2433 ext = oio->oi_active;
2434 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2435 /* one chunk plus extent overhead must be enough to write this
2437 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2438 if (ext->oe_end >= index)
2441 /* it doesn't need any grant to dirty this page */
2442 spin_lock(&cli->cl_loi_list_lock);
2443 rc = osc_enter_cache_try(cli, oap, grants, 0);
2444 if (rc == 0) { /* try failed */
2447 } else if (ext->oe_end < index) {
2449 /* try to expand this extent */
2450 rc = osc_extent_expand(ext, index, &tmp);
2453 /* don't free reserved grant */
2455 OSC_EXTENT_DUMP(D_CACHE, ext,
2456 "expanded for %lu.\n", index);
2457 osc_unreserve_grant_nolock(cli, grants, tmp);
2461 spin_unlock(&cli->cl_loi_list_lock);
2463 } else if (ext != NULL) {
2464 /* index is located outside of active extent */
2468 osc_extent_release(env, ext);
2469 oio->oi_active = NULL;
2474 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2476 /* try to find new extent to cover this page */
2477 LASSERT(oio->oi_active == NULL);
2478 /* we may have allocated grant for this page if we failed
2479 * to expand the previous active extent. */
2480 LASSERT(ergo(grants > 0, grants >= tmp));
2484 /* we haven't allocated grant for this page. */
2485 rc = osc_enter_cache(env, cli, oap, tmp);
2492 ext = osc_extent_find(env, osc, index, &tmp);
2494 LASSERT(tmp == grants);
2495 osc_exit_cache(cli, oap);
2499 oio->oi_active = ext;
2503 osc_unreserve_grant(cli, grants, tmp);
2506 LASSERT(ergo(rc == 0, ext != NULL));
2508 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2509 ext, "index = %lu.\n", index);
2510 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2512 osc_object_lock(osc);
2513 if (ext->oe_nr_pages == 0)
2514 ext->oe_srvlock = ops->ops_srvlock;
2516 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2518 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2519 osc_object_unlock(osc);
2521 if (!ext->oe_layout_version)
2522 ext->oe_layout_version = io->ci_layout_version;
2528 int osc_teardown_async_page(const struct lu_env *env,
2529 struct osc_object *obj, struct osc_page *ops)
2531 struct osc_async_page *oap = &ops->ops_oap;
2535 LASSERT(oap->oap_magic == OAP_MAGIC);
2537 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2538 oap, ops, osc_index(oap2osc(oap)));
2540 if (!list_empty(&oap->oap_rpc_item)) {
2541 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2543 } else if (!list_empty(&oap->oap_pending_item)) {
2544 struct osc_extent *ext = NULL;
2546 osc_object_lock(obj);
2547 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2548 osc_object_unlock(obj);
2549 /* only truncated pages are allowed to be taken out.
2550 * See osc_extent_truncate() and osc_cache_truncate_start()
2552 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2553 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2554 osc_index(oap2osc(oap)));
2558 osc_extent_put(env, ext);
2564 * This is called when a page is picked up by kernel to write out.
2566 * We should find out the corresponding extent and add the whole extent
2567 * into urgent list. The extent may be being truncated or used, handle it
2570 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2571 struct osc_page *ops)
2573 struct osc_extent *ext = NULL;
2574 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2575 struct cl_page *cp = ops->ops_cl.cpl_page;
2576 pgoff_t index = osc_index(ops);
2577 struct osc_async_page *oap = &ops->ops_oap;
2578 bool unplug = false;
2582 osc_object_lock(obj);
2583 ext = osc_extent_lookup(obj, index);
2585 osc_extent_tree_dump(D_ERROR, obj);
2586 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2589 switch (ext->oe_state) {
2592 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2596 /* If we know this extent is being written out, we should abort
2597 * so that the writer can make this page ready. Otherwise, there
2598 * exists a deadlock problem because other process can wait for
2599 * page writeback bit holding page lock; and meanwhile in
2600 * vvp_page_make_ready(), we need to grab page lock before
2601 * really sending the RPC. */
2603 /* race with truncate, page will be redirtied */
2605 /* The extent is active so we need to abort and let the caller
2606 * re-dirty the page. If we continued on here, and we were the
2607 * one making the extent active, we could deadlock waiting for
2608 * the page writeback to clear but it won't because the extent
2609 * is active and won't be written out. */
2610 GOTO(out, rc = -EAGAIN);
2615 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2619 spin_lock(&oap->oap_lock);
2620 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2621 spin_unlock(&oap->oap_lock);
2623 if (memory_pressure_get())
2624 ext->oe_memalloc = 1;
2627 if (ext->oe_state == OES_CACHE) {
2628 OSC_EXTENT_DUMP(D_CACHE, ext,
2629 "flush page %p make it urgent.\n", oap);
2630 if (list_empty(&ext->oe_link))
2631 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2638 osc_object_unlock(obj);
2639 osc_extent_put(env, ext);
2641 osc_io_unplug_async(env, osc_cli(obj), obj);
2646 * this is called when a sync waiter receives an interruption. Its job is to
2647 * get the caller woken as soon as possible. If its page hasn't been put in an
2648 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2649 * desiring interruption which will forcefully complete the rpc once the rpc
2652 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2654 struct osc_async_page *oap = &ops->ops_oap;
2655 struct osc_object *obj = oap->oap_obj;
2656 struct client_obd *cli = osc_cli(obj);
2657 struct osc_extent *ext;
2658 struct osc_extent *found = NULL;
2659 struct list_head *plist;
2660 pgoff_t index = osc_index(ops);
2665 LASSERT(!oap->oap_interrupted);
2666 oap->oap_interrupted = 1;
2668 /* Find out the caching extent */
2669 osc_object_lock(obj);
2670 if (oap->oap_cmd & OBD_BRW_WRITE) {
2671 plist = &obj->oo_urgent_exts;
2672 cmd = OBD_BRW_WRITE;
2674 plist = &obj->oo_reading_exts;
2677 list_for_each_entry(ext, plist, oe_link) {
2678 if (ext->oe_start <= index && ext->oe_end >= index) {
2679 LASSERT(ext->oe_state == OES_LOCK_DONE);
2680 /* For OES_LOCK_DONE state extent, it has already held
2681 * a refcount for RPC. */
2682 found = osc_extent_get(ext);
2686 if (found != NULL) {
2687 list_del_init(&found->oe_link);
2688 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2689 osc_object_unlock(obj);
2691 osc_extent_finish(env, found, 0, -EINTR);
2692 osc_extent_put(env, found);
2695 osc_object_unlock(obj);
2696 /* ok, it's been put in an rpc. only one oap gets a request
2698 if (oap->oap_request != NULL) {
2699 ptlrpc_mark_interrupted(oap->oap_request);
2700 ptlrpcd_wake(oap->oap_request);
2701 ptlrpc_req_finished(oap->oap_request);
2702 oap->oap_request = NULL;
2706 osc_list_maint(cli, obj);
2710 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2711 struct osc_object *obj, struct list_head *list,
2714 struct client_obd *cli = osc_cli(obj);
2715 struct osc_extent *ext;
2716 struct osc_async_page *oap;
2718 int mppr = cli->cl_max_pages_per_rpc;
2719 bool can_merge = true;
2720 pgoff_t start = CL_PAGE_EOF;
2724 list_for_each_entry(oap, list, oap_pending_item) {
2725 struct osc_page *opg = oap2osc_page(oap);
2726 pgoff_t index = osc_index(opg);
2733 mppr <<= (page_count > mppr);
2735 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2739 ext = osc_extent_alloc(obj);
2741 struct osc_async_page *tmp;
2743 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2744 list_del_init(&oap->oap_pending_item);
2745 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2750 ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2752 ext->oe_no_merge = !can_merge;
2754 ext->oe_start = start;
2755 ext->oe_end = ext->oe_max_end = end;
2757 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2758 ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2759 ext->oe_nr_pages = page_count;
2760 ext->oe_mppr = mppr;
2761 list_splice_init(list, &ext->oe_pages);
2762 ext->oe_layout_version = io->ci_layout_version;
2764 osc_object_lock(obj);
2765 /* Reuse the initial refcount for RPC, don't drop it */
2766 osc_extent_state_set(ext, OES_LOCK_DONE);
2767 if (!ext->oe_rw) { /* write */
2768 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2769 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2771 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2772 osc_update_pending(obj, OBD_BRW_READ, page_count);
2774 osc_object_unlock(obj);
2776 osc_io_unplug_async(env, cli, obj);
2781 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2783 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2784 __u64 size, struct osc_extent **extp)
2786 struct client_obd *cli = osc_cli(obj);
2787 struct osc_extent *ext;
2788 struct osc_extent *waiting = NULL;
2790 struct list_head list = LIST_HEAD_INIT(list);
2795 /* pages with index greater or equal to index will be truncated. */
2796 index = cl_index(osc2cl(obj), size);
2797 partial = size > cl_offset(osc2cl(obj), index);
2800 osc_object_lock(obj);
2801 ext = osc_extent_search(obj, index);
2803 ext = first_extent(obj);
2804 else if (ext->oe_end < index)
2805 ext = next_extent(ext);
2806 while (ext != NULL) {
2807 EASSERT(ext->oe_state != OES_TRUNC, ext);
2809 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2810 /* if ext is in urgent state, it means there must exist
2811 * a page already having been flushed by write_page().
2812 * We have to wait for this extent because we can't
2813 * truncate that page. */
2814 OSC_EXTENT_DUMP(D_CACHE, ext,
2815 "waiting for busy extent\n");
2816 waiting = osc_extent_get(ext);
2820 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2822 osc_extent_get(ext);
2823 if (ext->oe_state == OES_ACTIVE) {
2824 /* though we grab inode mutex for write path, but we
2825 * release it before releasing extent(in osc_io_end()),
2826 * so there is a race window that an extent is still
2827 * in OES_ACTIVE when truncate starts. */
2828 LASSERT(!ext->oe_trunc_pending);
2829 ext->oe_trunc_pending = 1;
2831 EASSERT(ext->oe_state == OES_CACHE, ext);
2832 osc_extent_state_set(ext, OES_TRUNC);
2833 osc_update_pending(obj, OBD_BRW_WRITE,
2836 /* This extent could be on the full extents list, that's OK */
2837 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2838 if (!list_empty(&ext->oe_link))
2839 list_move_tail(&ext->oe_link, &list);
2841 list_add_tail(&ext->oe_link, &list);
2843 ext = next_extent(ext);
2845 osc_object_unlock(obj);
2847 osc_list_maint(cli, obj);
2849 while (!list_empty(&list)) {
2852 ext = list_entry(list.next, struct osc_extent, oe_link);
2853 list_del_init(&ext->oe_link);
2855 /* extent may be in OES_ACTIVE state because inode mutex
2856 * is released before osc_io_end() in file write case */
2857 if (ext->oe_state != OES_TRUNC)
2858 osc_extent_wait(env, ext, OES_TRUNC);
2860 rc = osc_extent_truncate(ext, index, partial);
2865 OSC_EXTENT_DUMP(D_ERROR, ext,
2866 "truncate error %d\n", rc);
2867 } else if (ext->oe_nr_pages == 0) {
2868 osc_extent_remove(ext);
2870 /* this must be an overlapped extent which means only
2871 * part of pages in this extent have been truncated.
2873 EASSERTF(ext->oe_start <= index, ext,
2874 "trunc index = %lu/%d.\n", index, partial);
2875 /* fix index to skip this partially truncated extent */
2876 index = ext->oe_end + 1;
2879 /* we need to hold this extent in OES_TRUNC state so
2880 * that no writeback will happen. This is to avoid
2882 * Only partial truncate can reach here, if @size is
2883 * not zero, the caller should provide a valid @extp. */
2884 LASSERT(*extp == NULL);
2885 *extp = osc_extent_get(ext);
2886 OSC_EXTENT_DUMP(D_CACHE, ext,
2887 "trunc at %llu\n", size);
2889 osc_extent_put(env, ext);
2891 if (waiting != NULL) {
2894 /* ignore the result of osc_extent_wait the write initiator
2895 * should take care of it. */
2896 rc = osc_extent_wait(env, waiting, OES_INV);
2898 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2900 osc_extent_put(env, waiting);
2906 EXPORT_SYMBOL(osc_cache_truncate_start);
2909 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2911 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2914 struct osc_object *obj = ext->oe_obj;
2915 bool unplug = false;
2917 EASSERT(ext->oe_nr_pages > 0, ext);
2918 EASSERT(ext->oe_state == OES_TRUNC, ext);
2919 EASSERT(!ext->oe_urgent, ext);
2921 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2922 osc_object_lock(obj);
2923 osc_extent_state_set(ext, OES_CACHE);
2924 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2926 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2929 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2930 osc_object_unlock(obj);
2931 osc_extent_put(env, ext);
2934 osc_io_unplug_async(env, osc_cli(obj), obj);
2939 * Wait for extents in a specific range to be written out.
2940 * The caller must have called osc_cache_writeback_range() to issue IO
2941 * otherwise it will take a long time for this function to finish.
2943 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2944 * nobody else can dirty this range of file while we're waiting for
2945 * extents to be written.
2947 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2948 pgoff_t start, pgoff_t end)
2950 struct osc_extent *ext;
2951 pgoff_t index = start;
2956 osc_object_lock(obj);
2957 ext = osc_extent_search(obj, index);
2959 ext = first_extent(obj);
2960 else if (ext->oe_end < index)
2961 ext = next_extent(ext);
2962 while (ext != NULL) {
2965 if (ext->oe_start > end)
2968 if (!ext->oe_fsync_wait) {
2969 ext = next_extent(ext);
2973 EASSERT(ergo(ext->oe_state == OES_CACHE,
2974 ext->oe_hp || ext->oe_urgent), ext);
2975 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2976 !ext->oe_hp && ext->oe_urgent), ext);
2978 index = ext->oe_end + 1;
2979 osc_extent_get(ext);
2980 osc_object_unlock(obj);
2982 rc = osc_extent_wait(env, ext, OES_INV);
2985 osc_extent_put(env, ext);
2988 osc_object_unlock(obj);
2990 OSC_IO_DEBUG(obj, "sync file range.\n");
2993 EXPORT_SYMBOL(osc_cache_wait_range);
2996 * Called to write out a range of osc object.
2998 * @hp : should be set this is caused by lock cancel;
2999 * @discard: is set if dirty pages should be dropped - file will be deleted or
3000 * truncated, this implies there is no partially discarding extents.
3002 * Return how many pages will be issued, or error code if error occurred.
3004 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3005 pgoff_t start, pgoff_t end, int hp, int discard)
3007 struct osc_extent *ext;
3008 struct list_head discard_list = LIST_HEAD_INIT(discard_list);
3009 bool unplug = false;
3013 osc_object_lock(obj);
3014 ext = osc_extent_search(obj, start);
3016 ext = first_extent(obj);
3017 else if (ext->oe_end < start)
3018 ext = next_extent(ext);
3019 while (ext != NULL) {
3020 if (ext->oe_start > end)
3023 ext->oe_fsync_wait = 1;
3024 switch (ext->oe_state) {
3026 result += ext->oe_nr_pages;
3028 struct list_head *list = NULL;
3030 EASSERT(!ext->oe_hp, ext);
3032 list = &obj->oo_hp_exts;
3033 } else if (!ext->oe_urgent) {
3035 list = &obj->oo_urgent_exts;
3038 list_move_tail(&ext->oe_link, list);
3041 struct client_obd *cli = osc_cli(obj);
3042 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
3043 pgoff_t align_by = (1 << pcc_bits);
3044 pgoff_t a_start = round_down(start, align_by);
3045 pgoff_t a_end = round_up(end, align_by);
3049 a_end = CL_PAGE_EOF;
3050 /* the only discarder is lock cancelling, so
3051 * [start, end], aligned by chunk size, must
3052 * contain this extent */
3053 LASSERTF(ext->oe_start >= a_start &&
3054 ext->oe_end <= a_end,
3055 "ext [%lu, %lu] reg [%lu, %lu] "
3056 "orig [%lu %lu] align %lu bits "
3057 "%d\n", ext->oe_start, ext->oe_end,
3058 a_start, a_end, start, end,
3059 align_by, pcc_bits);
3060 osc_extent_state_set(ext, OES_LOCKING);
3061 ext->oe_owner = current;
3062 list_move_tail(&ext->oe_link,
3064 osc_update_pending(obj, OBD_BRW_WRITE,
3069 /* It's pretty bad to wait for ACTIVE extents, because
3070 * we don't know how long we will wait for it to be
3071 * flushed since it may be blocked at awaiting more
3072 * grants. We do this for the correctness of fsync. */
3073 LASSERT(hp == 0 && discard == 0);
3077 /* this extent is being truncated, can't do anything
3078 * for it now. it will be set to urgent after truncate
3079 * is finished in osc_cache_truncate_end(). */
3083 ext = next_extent(ext);
3085 osc_object_unlock(obj);
3087 LASSERT(ergo(!discard, list_empty(&discard_list)));
3088 if (!list_empty(&discard_list)) {
3089 struct osc_extent *tmp;
3092 osc_list_maint(osc_cli(obj), obj);
3093 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3094 list_del_init(&ext->oe_link);
3095 EASSERT(ext->oe_state == OES_LOCKING, ext);
3097 /* Discard caching pages. We don't actually write this
3098 * extent out but we complete it as if we did. */
3099 rc = osc_extent_make_ready(env, ext);
3100 if (unlikely(rc < 0)) {
3101 OSC_EXTENT_DUMP(D_ERROR, ext,
3102 "make_ready returned %d\n", rc);
3107 /* finish the extent as if the pages were sent */
3108 osc_extent_finish(env, ext, 0, 0);
3113 osc_io_unplug(env, osc_cli(obj), obj);
3115 if (hp || discard) {
3117 rc = osc_cache_wait_range(env, obj, start, end);
3118 if (result >= 0 && rc < 0)
3122 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3125 EXPORT_SYMBOL(osc_cache_writeback_range);
3128 * Returns a list of pages by a given [start, end] of \a obj.
3130 * \param resched If not NULL, then we give up before hogging CPU for too
3131 * long and set *resched = 1, in that case caller should implement a retry
3134 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3135 * crucial in the face of [offset, EOF] locks.
3137 * Return at least one page in @queue unless there is no covered page.
3139 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3140 struct osc_object *osc, pgoff_t start, pgoff_t end,
3141 osc_page_gang_cbt cb, void *cbdata)
3143 struct osc_page *ops;
3144 struct pagevec *pagevec;
3150 int res = CLP_GANG_OKAY;
3151 bool tree_lock = true;
3155 pvec = osc_env_info(env)->oti_pvec;
3156 pagevec = &osc_env_info(env)->oti_pagevec;
3157 ll_pagevec_init(pagevec, 0);
3158 spin_lock(&osc->oo_tree_lock);
3159 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3160 idx, OTI_PVEC_SIZE)) > 0) {
3161 struct cl_page *page;
3162 bool end_of_region = false;
3164 for (i = 0, j = 0; i < nr; ++i) {
3168 idx = osc_index(ops);
3170 end_of_region = true;
3174 page = ops->ops_cl.cpl_page;
3175 LASSERT(page->cp_type == CPT_CACHEABLE);
3176 if (page->cp_state == CPS_FREEING)
3180 lu_ref_add_atomic(&page->cp_reference,
3181 "gang_lookup", current);
3187 * Here a delicate locking dance is performed. Current thread
3188 * holds a reference to a page, but has to own it before it
3189 * can be placed into queue. Owning implies waiting, so
3190 * radix-tree lock is to be released. After a wait one has to
3191 * check that pages weren't truncated (cl_page_own() returns
3192 * error in the latter case).
3194 spin_unlock(&osc->oo_tree_lock);
3197 for (i = 0; i < j; ++i) {
3199 if (res == CLP_GANG_OKAY)
3200 res = (*cb)(env, io, ops, cbdata);
3202 page = ops->ops_cl.cpl_page;
3203 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3204 cl_pagevec_put(env, page, pagevec);
3206 pagevec_release(pagevec);
3208 if (nr < OTI_PVEC_SIZE || end_of_region)
3211 if (res == CLP_GANG_OKAY && need_resched())
3212 res = CLP_GANG_RESCHED;
3213 if (res != CLP_GANG_OKAY)
3216 spin_lock(&osc->oo_tree_lock);
3220 spin_unlock(&osc->oo_tree_lock);
3223 EXPORT_SYMBOL(osc_page_gang_lookup);
3226 * Check if page @page is covered by an extra lock or discard it.
3228 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3229 struct osc_page *ops, void *cbdata)
3231 struct osc_thread_info *info = osc_env_info(env);
3232 struct osc_object *osc = cbdata;
3235 index = osc_index(ops);
3236 if (index >= info->oti_fn_index) {
3237 struct ldlm_lock *tmp;
3238 struct cl_page *page = ops->ops_cl.cpl_page;
3240 /* refresh non-overlapped index */
3241 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3242 OSC_DAP_FL_TEST_LOCK);
3244 __u64 end = tmp->l_policy_data.l_extent.end;
3245 /* Cache the first-non-overlapped index so as to skip
3246 * all pages within [index, oti_fn_index). This is safe
3247 * because if tmp lock is canceled, it will discard
3249 info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3250 if (end == OBD_OBJECT_EOF)
3251 info->oti_fn_index = CL_PAGE_EOF;
3253 } else if (cl_page_own(env, io, page) == 0) {
3254 /* discard the page */
3255 cl_page_discard(env, io, page);
3256 cl_page_disown(env, io, page);
3258 LASSERT(page->cp_state == CPS_FREEING);
3262 info->oti_next_index = index + 1;
3263 return CLP_GANG_OKAY;
3266 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3267 struct osc_page *ops, void *cbdata)
3269 struct osc_thread_info *info = osc_env_info(env);
3270 struct cl_page *page = ops->ops_cl.cpl_page;
3272 /* page is top page. */
3273 info->oti_next_index = osc_index(ops) + 1;
3274 if (cl_page_own(env, io, page) == 0) {
3275 if (!ergo(page->cp_type == CPT_CACHEABLE,
3276 !PageDirty(cl_page_vmpage(page))))
3277 CL_PAGE_DEBUG(D_ERROR, env, page,
3278 "discard dirty page?\n");
3280 /* discard the page */
3281 cl_page_discard(env, io, page);
3282 cl_page_disown(env, io, page);
3284 LASSERT(page->cp_state == CPS_FREEING);
3287 return CLP_GANG_OKAY;
3289 EXPORT_SYMBOL(osc_discard_cb);
3292 * Discard pages protected by the given lock. This function traverses radix
3293 * tree to find all covering pages and discard them. If a page is being covered
3294 * by other locks, it should remain in cache.
3296 * If error happens on any step, the process continues anyway (the reasoning
3297 * behind this being that lock cancellation cannot be delayed indefinitely).
3299 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3300 pgoff_t start, pgoff_t end, bool discard)
3302 struct osc_thread_info *info = osc_env_info(env);
3303 struct cl_io *io = osc_env_thread_io(env);
3304 osc_page_gang_cbt cb;
3310 io->ci_obj = cl_object_top(osc2cl(osc));
3311 io->ci_ignore_layout = 1;
3312 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3316 cb = discard ? osc_discard_cb : check_and_discard_cb;
3317 info->oti_fn_index = info->oti_next_index = start;
3319 res = osc_page_gang_lookup(env, io, osc,
3320 info->oti_next_index, end, cb, osc);
3321 if (info->oti_next_index > end)
3324 if (res == CLP_GANG_RESCHED)
3326 } while (res != CLP_GANG_OKAY);
3328 cl_io_fini(env, io);