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/
32 * osc cache management.
34 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include <lustre_osc.h>
40 #include <lustre_dlm.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';
77 if (!RB_EMPTY_NODE(&ext->oe_node))
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 *const 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 kref_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)
149 return rb_entry_safe(n, struct osc_extent, oe_node);
152 static inline struct osc_extent *next_extent(struct osc_extent *ext)
157 LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
158 return rb_extent(rb_next(&ext->oe_node));
161 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
166 LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
167 return rb_extent(rb_prev(&ext->oe_node));
170 static inline struct osc_extent *first_extent(struct osc_object *obj)
172 return rb_extent(rb_first(&obj->oo_root));
175 /* object must be locked by caller. */
176 static int osc_extent_sanity_check0(struct osc_extent *ext,
177 const char *func, const int line)
179 struct osc_object *obj = ext->oe_obj;
180 struct osc_async_page *oap;
184 assert_osc_object_is_locked(obj);
186 if (ext->oe_state >= OES_STATE_MAX)
189 if (kref_read(&ext->oe_refc) <= 0)
192 if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
195 switch (ext->oe_state) {
197 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
202 if (atomic_read(&ext->oe_users) == 0)
206 if (ext->oe_fsync_wait && !ext->oe_urgent)
210 if (ext->oe_grants == 0)
212 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
216 if (atomic_read(&ext->oe_users) > 0)
220 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
223 if (ext->oe_sync && ext->oe_grants > 0)
226 if (ext->oe_dlmlock != NULL &&
227 ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
228 !ldlm_is_failed(ext->oe_dlmlock)) {
229 struct ldlm_extent *extent;
231 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
232 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
233 extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end)))
236 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
240 if (ext->oe_nr_pages > ext->oe_mppr)
243 /* Do not verify page list if extent is in RPC. This is because an
244 * in-RPC extent is supposed to be exclusively accessible w/o lock. */
245 if (ext->oe_state > OES_CACHE)
252 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
253 pgoff_t index = osc_index(oap2osc(oap));
255 if (index > ext->oe_end || index < ext->oe_start)
258 if (page_count != ext->oe_nr_pages)
263 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
264 "sanity check %p failed: rc = %d\n",
269 #define sanity_check_nolock(ext) \
270 osc_extent_sanity_check0(ext, __func__, __LINE__)
272 #define sanity_check(ext) ({ \
274 osc_object_lock((ext)->oe_obj); \
275 __res = sanity_check_nolock(ext); \
276 osc_object_unlock((ext)->oe_obj); \
281 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
283 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
287 * sanity check - to make sure there is no overlapped extent in the tree.
289 static int osc_extent_is_overlapped(struct osc_object *obj,
290 struct osc_extent *ext)
292 struct osc_extent *tmp;
294 assert_osc_object_is_locked(obj);
299 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
302 if (overlapped(tmp, ext))
308 static void osc_extent_state_set(struct osc_extent *ext, int state)
310 assert_osc_object_is_locked(ext->oe_obj);
311 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
313 /* Never try to sanity check a state changing extent :-) */
314 /* LASSERT(sanity_check_nolock(ext) == 0); */
316 /* TODO: validate the state machine */
317 smp_store_release(&ext->oe_state, state);
318 wake_up(&ext->oe_waitq);
321 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
323 struct osc_extent *ext;
325 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
329 RB_CLEAR_NODE(&ext->oe_node);
331 cl_object_get(osc2cl(obj));
332 kref_init(&ext->oe_refc);
333 atomic_set(&ext->oe_users, 0);
334 INIT_LIST_HEAD(&ext->oe_link);
335 ext->oe_state = OES_INV;
336 INIT_LIST_HEAD(&ext->oe_pages);
337 init_waitqueue_head(&ext->oe_waitq);
338 ext->oe_dlmlock = NULL;
343 static void osc_extent_free(struct kref *kref)
345 struct osc_extent *ext = container_of(kref, struct osc_extent,
348 LASSERT(list_empty(&ext->oe_link));
349 LASSERT(atomic_read(&ext->oe_users) == 0);
350 LASSERT(ext->oe_state == OES_INV);
351 LASSERT(RB_EMPTY_NODE(&ext->oe_node));
353 if (ext->oe_dlmlock) {
354 lu_ref_del(&ext->oe_dlmlock->l_reference,
356 LDLM_LOCK_PUT(ext->oe_dlmlock);
357 ext->oe_dlmlock = NULL;
360 /* If/When cl_object_put drops the need for 'env',
361 * this code can be enabled, and matching code in
362 * osc_extent_put removed.
364 cl_object_put(osc2cl(ext->oe_obj));
366 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
370 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
372 LASSERT(kref_read(&ext->oe_refc) >= 0);
373 kref_get(&ext->oe_refc);
377 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
379 LASSERT(kref_read(&ext->oe_refc) > 0);
380 if (kref_put(&ext->oe_refc, osc_extent_free)) {
381 /* This should be in osc_extent_free(), but
382 * while we need to pass 'env' it cannot be.
384 cl_object_put(env, osc2cl(ext->oe_obj));
386 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
391 * osc_extent_put_trust() is a special version of osc_extent_put() when
392 * it's known that the caller is not the last user. This is to address the
393 * problem of lacking of lu_env ;-).
395 static void osc_extent_put_trust(struct osc_extent *ext)
397 LASSERT(kref_read(&ext->oe_refc) > 1);
398 assert_osc_object_is_locked(ext->oe_obj);
399 osc_extent_put(NULL, ext);
403 * Return the extent which includes pgoff @index, or return the greatest
404 * previous extent in the tree.
406 static struct osc_extent *osc_extent_search(struct osc_object *obj,
409 struct rb_node *n = obj->oo_root.rb_node;
410 struct osc_extent *tmp, *p = NULL;
412 assert_osc_object_is_locked(obj);
415 if (index < tmp->oe_start) {
417 } else if (index > tmp->oe_end) {
428 * Return the extent covering @index, otherwise return NULL.
429 * caller must have held object lock.
431 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
434 struct osc_extent *ext;
436 ext = osc_extent_search(obj, index);
437 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
438 return osc_extent_get(ext);
442 /* caller must have held object lock. */
443 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
445 struct rb_node **n = &obj->oo_root.rb_node;
446 struct rb_node *parent = NULL;
447 struct osc_extent *tmp;
449 LASSERT(RB_EMPTY_NODE(&ext->oe_node));
450 LASSERT(ext->oe_obj == obj);
451 assert_osc_object_is_locked(obj);
456 if (ext->oe_end < tmp->oe_start)
458 else if (ext->oe_start > tmp->oe_end)
461 EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
463 rb_link_node(&ext->oe_node, parent, n);
464 rb_insert_color(&ext->oe_node, &obj->oo_root);
468 /* caller must have held object lock. */
469 static void osc_extent_erase(struct osc_extent *ext)
471 struct osc_object *obj = ext->oe_obj;
472 assert_osc_object_is_locked(obj);
473 if (!RB_EMPTY_NODE(&ext->oe_node)) {
474 rb_erase(&ext->oe_node, &obj->oo_root);
475 RB_CLEAR_NODE(&ext->oe_node);
476 /* rbtree held a refcount */
477 osc_extent_put_trust(ext);
481 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
483 struct osc_object *obj = ext->oe_obj;
485 assert_osc_object_is_locked(obj);
486 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
487 if (ext->oe_state == OES_CACHE) {
488 osc_extent_state_set(ext, OES_ACTIVE);
489 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
491 atomic_inc(&ext->oe_users);
492 list_del_init(&ext->oe_link);
493 return osc_extent_get(ext);
496 static void __osc_extent_remove(struct osc_extent *ext)
498 assert_osc_object_is_locked(ext->oe_obj);
499 LASSERT(list_empty(&ext->oe_pages));
500 osc_extent_erase(ext);
501 list_del_init(&ext->oe_link);
502 osc_extent_state_set(ext, OES_INV);
503 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
506 static void osc_extent_remove(struct osc_extent *ext)
508 struct osc_object *obj = ext->oe_obj;
510 osc_object_lock(obj);
511 __osc_extent_remove(ext);
512 osc_object_unlock(obj);
516 * This function is used to merge extents to get better performance. It checks
517 * if @cur and @victim are contiguous at block level.
519 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
520 struct osc_extent *victim)
522 struct osc_object *obj = cur->oe_obj;
523 struct client_obd *cli = osc_cli(obj);
528 LASSERT(cur->oe_state == OES_CACHE);
529 assert_osc_object_is_locked(obj);
533 if (victim->oe_state != OES_INV &&
534 (victim->oe_state != OES_CACHE || victim->oe_fsync_wait))
537 if (cur->oe_max_end != victim->oe_max_end)
541 * In the rare case max_pages_per_rpc (mppr) is changed, don't
542 * merge extents until after old ones have been sent, or the
543 * "extents are aligned to RPCs" checks are unhappy.
545 if (cur->oe_mppr != victim->oe_mppr)
548 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
549 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
550 chunk_start = cur->oe_start >> ppc_bits;
551 chunk_end = cur->oe_end >> ppc_bits;
552 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
553 chunk_end + 1 != victim->oe_start >> ppc_bits)
556 /* overall extent size should not exceed the max supported limit
557 * reported by the server */
558 if (cur->oe_end - cur->oe_start + 1 +
559 victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
562 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
564 cur->oe_start = min(cur->oe_start, victim->oe_start);
565 cur->oe_end = max(cur->oe_end, victim->oe_end);
566 /* per-extent tax should be accounted only once for the whole extent */
567 cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax;
568 cur->oe_nr_pages += victim->oe_nr_pages;
569 /* only the following bits are needed to merge */
570 cur->oe_urgent |= victim->oe_urgent;
571 cur->oe_memalloc |= victim->oe_memalloc;
572 list_splice_init(&victim->oe_pages, &cur->oe_pages);
573 victim->oe_nr_pages = 0;
575 osc_extent_get(victim);
576 __osc_extent_remove(victim);
577 osc_extent_put(env, victim);
579 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
584 * Drop user count of osc_extent, and unplug IO asynchronously.
586 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
588 struct osc_object *obj = ext->oe_obj;
589 struct client_obd *cli = osc_cli(obj);
592 LASSERT(atomic_read(&ext->oe_users) > 0);
593 LASSERT(sanity_check(ext) == 0);
594 LASSERT(ext->oe_grants > 0);
596 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
597 LASSERT(ext->oe_state == OES_ACTIVE);
598 if (ext->oe_trunc_pending) {
599 /* a truncate process is waiting for this extent.
600 * This may happen due to a race, check
601 * osc_cache_truncate_start(). */
602 osc_extent_state_set(ext, OES_TRUNC);
603 ext->oe_trunc_pending = 0;
604 osc_object_unlock(obj);
608 osc_extent_state_set(ext, OES_CACHE);
609 osc_update_pending(obj, OBD_BRW_WRITE,
612 /* try to merge the previous and next extent. */
613 if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
614 grant += cli->cl_grant_extent_tax;
615 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
616 grant += cli->cl_grant_extent_tax;
619 list_move_tail(&ext->oe_link,
621 else if (ext->oe_urgent)
622 list_move_tail(&ext->oe_link,
623 &obj->oo_urgent_exts);
624 else if (ext->oe_nr_pages == ext->oe_mppr) {
625 list_move_tail(&ext->oe_link,
628 osc_object_unlock(obj);
630 osc_unreserve_grant(cli, 0, grant);
633 osc_io_unplug_async(env, cli, obj);
635 osc_extent_put(env, ext);
641 * Find or create an extent which includes @index, core function to manage
644 static struct osc_extent *osc_extent_find(const struct lu_env *env,
645 struct osc_object *obj, pgoff_t index,
646 unsigned int *grants)
648 struct client_obd *cli = osc_cli(obj);
649 struct osc_lock *olck;
650 struct cl_lock_descr *descr;
651 struct osc_extent *cur;
652 struct osc_extent *ext;
653 struct osc_extent *conflict = NULL;
654 struct osc_extent *found = NULL;
657 unsigned int max_pages; /* max_pages_per_rpc */
658 unsigned int chunksize;
659 int ppc_bits; /* pages per chunk bits */
664 cur = osc_extent_alloc(obj);
666 RETURN(ERR_PTR(-ENOMEM));
668 olck = osc_env_io(env)->oi_write_osclock;
669 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
670 LASSERT(olck->ols_state == OLS_GRANTED);
672 descr = &olck->ols_cl.cls_lock->cll_descr;
673 LASSERT(descr->cld_mode >= CLM_WRITE);
675 LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
676 "chunkbits: %u\n", cli->cl_chunkbits);
677 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
678 chunk_mask = ~((1 << ppc_bits) - 1);
679 chunksize = 1 << cli->cl_chunkbits;
680 chunk = index >> ppc_bits;
682 /* align end to RPC edge. */
683 max_pages = cli->cl_max_pages_per_rpc;
684 if ((max_pages & ~chunk_mask) != 0) {
685 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
686 max_pages, cli->cl_chunkbits, chunk_mask);
687 RETURN(ERR_PTR(-EINVAL));
689 max_end = index - (index % max_pages) + max_pages - 1;
690 max_end = min_t(pgoff_t, max_end, descr->cld_end);
692 /* initialize new extent by parameters so far */
693 cur->oe_max_end = max_end;
694 cur->oe_start = index & chunk_mask;
695 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
696 if (cur->oe_start < descr->cld_start)
697 cur->oe_start = descr->cld_start;
698 if (cur->oe_end > max_end)
699 cur->oe_end = max_end;
700 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
701 cur->oe_mppr = max_pages;
702 if (olck->ols_dlmlock != NULL) {
703 LASSERT(olck->ols_hold);
704 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
705 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
708 /* grants has been allocated by caller */
709 LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
710 "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
711 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
715 osc_object_lock(obj);
716 ext = osc_extent_search(obj, cur->oe_start);
718 ext = first_extent(obj);
719 for (; ext; ext = next_extent(ext)) {
720 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
721 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
723 LASSERT(sanity_check_nolock(ext) == 0);
724 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
727 /* if covering by different locks, no chance to match */
728 if (olck->ols_dlmlock != ext->oe_dlmlock) {
729 EASSERTF(!overlapped(ext, cur), ext,
730 EXTSTR"\n", EXTPARA(cur));
735 /* discontiguous chunks? */
736 if (chunk + 1 < ext_chk_start)
739 /* ok, from now on, ext and cur have these attrs:
740 * 1. covered by the same lock
741 * 2. contiguous at chunk level or overlapping. */
743 if (overlapped(ext, cur)) {
744 /* cur is the minimum unit, so overlapping means
746 EASSERTF((ext->oe_start <= cur->oe_start &&
747 ext->oe_end >= cur->oe_end),
748 ext, EXTSTR"\n", EXTPARA(cur));
750 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
751 /* for simplicity, we wait for this extent to
752 * finish before going forward. */
753 conflict = osc_extent_get(ext);
757 found = osc_extent_hold(ext);
761 /* non-overlapped extent */
762 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
763 /* we can't do anything for a non OES_CACHE extent, or
764 * if there is someone waiting for this extent to be
765 * flushed, try next one. */
768 if (osc_extent_merge(env, ext, cur) == 0) {
769 LASSERT(*grants >= chunksize);
770 *grants -= chunksize;
773 * Try to merge with the next one too because we
774 * might have just filled in a gap.
776 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
777 /* we can save extent tax from next extent */
778 *grants += cli->cl_grant_extent_tax;
780 found = osc_extent_hold(ext);
785 osc_extent_tree_dump(D_CACHE, obj);
787 LASSERT(conflict == NULL);
788 if (!IS_ERR(found)) {
789 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
790 OSC_EXTENT_DUMP(D_CACHE, found,
791 "found caching ext for %lu.\n", index);
793 } else if (conflict == NULL) {
794 /* create a new extent */
795 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
796 LASSERT(*grants >= cur->oe_grants);
797 *grants -= cur->oe_grants;
799 cur->oe_state = OES_CACHE;
800 found = osc_extent_hold(cur);
801 osc_extent_insert(obj, cur);
802 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
803 index, descr->cld_end);
805 osc_object_unlock(obj);
807 if (conflict != NULL) {
808 LASSERT(found == NULL);
810 /* waiting for IO to finish. Please notice that it's impossible
811 * to be an OES_TRUNC extent. */
812 rc = osc_extent_wait(env, conflict, OES_INV);
813 osc_extent_put(env, conflict);
816 GOTO(out, found = ERR_PTR(rc));
823 osc_extent_put(env, cur);
828 * Called when IO is finished to an extent.
830 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
833 struct client_obd *cli = osc_cli(ext->oe_obj);
834 struct osc_async_page *oap;
835 struct osc_async_page *tmp;
836 int nr_pages = ext->oe_nr_pages;
838 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
843 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
845 ext->oe_rc = rc ?: ext->oe_nr_pages;
846 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
848 osc_lru_add_batch(cli, &ext->oe_pages);
849 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
851 list_del_init(&oap->oap_rpc_item);
852 list_del_init(&oap->oap_pending_item);
853 if (last_off <= oap->oap_obj_off) {
854 last_off = oap->oap_obj_off;
855 last_count = oap->oap_count;
859 osc_ap_completion(env, cli, oap, sent, rc);
861 EASSERT(ext->oe_nr_pages == 0, ext);
864 lost_grant = ext->oe_grants;
865 } else if (cli->cl_ocd_grant_param == 0 &&
866 blocksize < PAGE_SIZE &&
867 last_count != PAGE_SIZE) {
868 /* For short writes without OBD_CONNECT_GRANT support, we
869 * shouldn't count parts of pages that span a whole chunk on
870 * the OST side, or our accounting goes wrong. Should match
871 * the code in tgt_grant_check.
873 int offset = last_off & ~PAGE_MASK;
874 int count = last_count + (offset & (blocksize - 1));
875 int end = (offset + last_count) & (blocksize - 1);
877 count += blocksize - end;
879 lost_grant = PAGE_SIZE - count;
881 if (ext->oe_grants > 0)
882 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
884 osc_extent_remove(ext);
885 /* put the refcount for RPC */
886 osc_extent_put(env, ext);
891 * Wait for the extent's state to become @state.
893 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
894 enum osc_extent_state state)
896 struct osc_object *obj = ext->oe_obj;
900 osc_object_lock(obj);
901 LASSERT(sanity_check_nolock(ext) == 0);
902 /* `Kick' this extent only if the caller is waiting for it to be
904 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
905 if (ext->oe_state == OES_ACTIVE) {
907 } else if (ext->oe_state == OES_CACHE) {
909 osc_extent_hold(ext);
913 osc_object_unlock(obj);
915 osc_extent_release(env, ext);
917 /* wait for the extent until its state becomes @state */
918 rc = wait_event_idle_timeout(ext->oe_waitq,
919 smp_load_acquire(&ext->oe_state) == state,
920 cfs_time_seconds(600));
922 OSC_EXTENT_DUMP(D_ERROR, ext,
923 "%s: wait ext to %u timedout, recovery in progress?\n",
924 cli_name(osc_cli(obj)), state);
926 wait_event_idle(ext->oe_waitq,
927 smp_load_acquire(&ext->oe_state) == state);
937 * Discard pages with index greater than @size. If @ext is overlapped with
938 * @size, then partial truncate happens.
940 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
945 struct osc_object *obj = ext->oe_obj;
946 struct client_obd *cli = osc_cli(obj);
947 struct osc_async_page *oap;
948 struct osc_async_page *tmp;
949 struct pagevec *pvec;
950 int pages_in_chunk = 0;
951 int ppc_bits = cli->cl_chunkbits -
953 __u64 trunc_chunk = trunc_index >> ppc_bits;
960 LASSERT(sanity_check(ext) == 0);
961 LASSERT(ext->oe_state == OES_TRUNC);
962 LASSERT(!ext->oe_urgent);
964 /* Request new lu_env.
965 * We can't use that env from osc_cache_truncate_start() because
966 * it's from lov_io_sub and not fully initialized. */
967 env = cl_env_get(&refcheck);
969 RETURN(PTR_ERR(env));
971 io = osc_env_thread_io(env);
972 io->ci_obj = cl_object_top(osc2cl(obj));
973 io->ci_ignore_layout = 1;
974 pvec = &osc_env_info(env)->oti_pagevec;
975 ll_pagevec_init(pvec, 0);
976 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
980 /* discard all pages with index greater than trunc_index */
981 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
983 pgoff_t index = osc_index(oap2osc(oap));
984 struct cl_page *page = oap2cl_page(oap);
986 LASSERT(list_empty(&oap->oap_rpc_item));
988 /* only discard the pages with their index greater than
989 * trunc_index, and ... */
990 if (index < trunc_index ||
991 (index == trunc_index && partial)) {
992 /* accounting how many pages remaining in the chunk
993 * so that we can calculate grants correctly. */
994 if (index >> ppc_bits == trunc_chunk)
999 list_del_init(&oap->oap_pending_item);
1002 lu_ref_add(&page->cp_reference, "truncate", current);
1004 if (cl_page_own(env, io, page) == 0) {
1005 cl_page_discard(env, io, page);
1006 cl_page_disown(env, io, page);
1008 LASSERT(page->cp_state == CPS_FREEING);
1012 lu_ref_del(&page->cp_reference, "truncate", current);
1013 cl_pagevec_put(env, page, pvec);
1018 pagevec_release(pvec);
1020 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1021 ext->oe_nr_pages == 0),
1022 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1024 osc_object_lock(obj);
1025 if (ext->oe_nr_pages == 0) {
1026 LASSERT(pages_in_chunk == 0);
1027 grants = ext->oe_grants;
1029 } else { /* calculate how many grants we can free */
1030 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1034 /* if there is no pages in this chunk, we can also free grants
1035 * for the last chunk */
1036 if (pages_in_chunk == 0) {
1037 /* if this is the 1st chunk and no pages in this chunk,
1038 * ext->oe_nr_pages must be zero, so we should be in
1039 * the other if-clause. */
1040 LASSERT(trunc_chunk > 0);
1045 /* this is what we can free from this extent */
1046 grants = chunks << cli->cl_chunkbits;
1047 ext->oe_grants -= grants;
1048 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1049 ext->oe_end = min(last_index, ext->oe_max_end);
1050 LASSERT(ext->oe_end >= ext->oe_start);
1051 LASSERT(ext->oe_grants > 0);
1053 osc_object_unlock(obj);
1055 if (grants > 0 || nr_pages > 0)
1056 osc_free_grant(cli, nr_pages, grants, grants);
1059 cl_io_fini(env, io);
1060 cl_env_put(env, &refcheck);
1065 * This function is used to make the extent prepared for transfer.
1066 * A race with flusing page - ll_writepage() has to be handled cautiously.
1068 static int osc_extent_make_ready(const struct lu_env *env,
1069 struct osc_extent *ext)
1071 struct osc_async_page *oap;
1072 struct osc_async_page *last = NULL;
1073 struct osc_object *obj = ext->oe_obj;
1074 unsigned int page_count = 0;
1078 /* we're going to grab page lock, so object lock must not be taken. */
1079 LASSERT(sanity_check(ext) == 0);
1080 /* in locking state, any process should not touch this extent. */
1081 EASSERT(ext->oe_state == OES_LOCKING, ext);
1082 EASSERT(ext->oe_owner != NULL, ext);
1084 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1086 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1088 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1091 /* checking ASYNC_READY is race safe */
1092 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1095 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1098 spin_lock(&oap->oap_lock);
1099 oap->oap_async_flags |= ASYNC_READY;
1100 spin_unlock(&oap->oap_lock);
1103 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1106 LASSERTF(0, "unknown return code: %d\n", rc);
1110 LASSERT(page_count == ext->oe_nr_pages);
1111 LASSERT(last != NULL);
1112 /* the last page is the only one we need to refresh its count by
1113 * the size of file. */
1114 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1115 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1116 LASSERTF(last_oap_count > 0,
1117 "last_oap_count %d\n", last_oap_count);
1118 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1119 last->oap_count = last_oap_count;
1120 spin_lock(&last->oap_lock);
1121 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1122 spin_unlock(&last->oap_lock);
1125 /* for the rest of pages, we don't need to call osf_refresh_count()
1126 * because it's known they are not the last page */
1127 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1128 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1129 oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1130 spin_lock(&oap->oap_lock);
1131 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1132 spin_unlock(&oap->oap_lock);
1136 osc_object_lock(obj);
1137 osc_extent_state_set(ext, OES_RPC);
1138 osc_object_unlock(obj);
1139 /* get a refcount for RPC. */
1140 osc_extent_get(ext);
1146 * Quick and simple version of osc_extent_find(). This function is frequently
1147 * called to expand the extent for the same IO. To expand the extent, the
1148 * page index must be in the same or next chunk of ext->oe_end.
1150 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1151 unsigned int *grants)
1153 struct osc_object *obj = ext->oe_obj;
1154 struct client_obd *cli = osc_cli(obj);
1155 struct osc_extent *next;
1156 int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1157 pgoff_t chunk = index >> ppc_bits;
1160 unsigned int chunksize = 1 << cli->cl_chunkbits;
1164 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1165 osc_object_lock(obj);
1166 LASSERT(sanity_check_nolock(ext) == 0);
1167 end_chunk = ext->oe_end >> ppc_bits;
1168 if (chunk > end_chunk + 1)
1169 GOTO(out, rc = -ERANGE);
1171 if (end_chunk >= chunk)
1174 LASSERT(end_chunk + 1 == chunk);
1176 /* try to expand this extent to cover @index */
1177 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1179 /* don't go over the maximum extent size reported by server */
1180 if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1181 GOTO(out, rc = -ERANGE);
1183 next = next_extent(ext);
1184 if (next != NULL && next->oe_start <= end_index)
1185 /* complex mode - overlapped with the next extent,
1186 * this case will be handled by osc_extent_find() */
1187 GOTO(out, rc = -EAGAIN);
1189 ext->oe_end = end_index;
1190 ext->oe_grants += chunksize;
1191 LASSERT(*grants >= chunksize);
1192 *grants -= chunksize;
1193 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1194 "overlapped after expanding for %lu.\n", index);
1198 osc_object_unlock(obj);
1202 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1203 const char *func, int line)
1205 struct osc_extent *ext;
1208 if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1211 CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1212 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1214 /* osc_object_lock(obj); */
1216 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1217 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1220 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1221 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1224 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1225 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1228 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1229 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1230 /* osc_object_unlock(obj); */
1233 /* ------------------ osc extent end ------------------ */
1235 static inline int osc_is_ready(struct osc_object *osc)
1237 return !list_empty(&osc->oo_ready_item) ||
1238 !list_empty(&osc->oo_hp_ready_item);
1241 #define OSC_IO_DEBUG(OSC, STR, args...) \
1242 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1243 (OSC), osc_is_ready(OSC), \
1244 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1245 list_empty_marker(&(OSC)->oo_ready_item), \
1246 atomic_read(&(OSC)->oo_nr_writes), \
1247 list_empty_marker(&(OSC)->oo_hp_exts), \
1248 list_empty_marker(&(OSC)->oo_urgent_exts), \
1249 atomic_read(&(OSC)->oo_nr_reads), \
1250 list_empty_marker(&(OSC)->oo_reading_exts), \
1253 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1256 struct osc_page *opg = oap2osc_page(oap);
1257 struct cl_page *page = oap2cl_page(oap);
1260 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1263 result = cl_page_make_ready(env, page, CRT_WRITE);
1265 opg->ops_submit_time = ktime_get();
1269 static int osc_refresh_count(const struct lu_env *env,
1270 struct osc_async_page *oap, int cmd)
1272 struct osc_page *opg = oap2osc_page(oap);
1273 pgoff_t index = osc_index(oap2osc(oap));
1274 struct cl_object *obj;
1275 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1279 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1280 LASSERT(!(cmd & OBD_BRW_READ));
1281 LASSERT(opg != NULL);
1282 obj = opg->ops_cl.cpl_obj;
1284 cl_object_attr_lock(obj);
1285 result = cl_object_attr_get(env, obj, attr);
1286 cl_object_attr_unlock(obj);
1289 kms = attr->cat_kms;
1290 if (cl_offset(obj, index) >= kms)
1291 /* catch race with truncate */
1293 else if (cl_offset(obj, index + 1) > kms)
1294 /* catch sub-page write at end of file */
1295 return kms & ~PAGE_MASK;
1300 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1303 struct osc_page *opg = oap2osc_page(oap);
1304 struct cl_page *page = oap2cl_page(oap);
1305 enum cl_req_type crt;
1310 cmd &= ~OBD_BRW_NOQUOTA;
1311 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1312 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1313 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1314 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1315 LASSERT(opg->ops_transfer_pinned);
1317 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1318 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1319 opg->ops_transfer_pinned = 0;
1321 opg->ops_submit_time = ktime_set(0, 0);
1322 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1325 if (rc == 0 && srvlock) {
1326 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1327 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1328 size_t bytes = oap->oap_count;
1330 if (crt == CRT_READ)
1331 stats->os_lockless_reads += bytes;
1333 stats->os_lockless_writes += bytes;
1337 * This has to be the last operation with the page, as locks are
1338 * released in cl_page_completion() and nothing except for the
1339 * reference counter protects page from concurrent reclaim.
1341 lu_ref_del(&page->cp_reference, "transfer", page);
1343 cl_page_completion(env, page, crt, rc);
1344 cl_page_put(env, page);
1349 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \
1350 struct client_obd *__tmp = (cli); \
1351 CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1352 "dropped: %ld avail: %ld, dirty_grant: %ld, " \
1353 "reserved: %ld, flight: %d } lru {in list: %ld, " \
1354 "left: %ld, waiters: %d }" fmt "\n", \
1356 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1357 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1358 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1359 __tmp->cl_dirty_grant, \
1360 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1361 atomic_long_read(&__tmp->cl_lru_in_list), \
1362 atomic_long_read(&__tmp->cl_lru_busy), \
1363 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1366 /* caller must hold loi_list_lock */
1367 static void osc_consume_write_grant(struct client_obd *cli,
1368 struct brw_page *pga)
1370 assert_spin_locked(&cli->cl_loi_list_lock);
1371 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1372 cli->cl_dirty_pages++;
1373 pga->flag |= OBD_BRW_FROM_GRANT;
1374 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1375 PAGE_SIZE, pga, pga->pg);
1378 /* the companion to osc_consume_write_grant, called when a brw has completed.
1379 * must be called with the loi lock held. */
1380 static void osc_release_write_grant(struct client_obd *cli,
1381 struct brw_page *pga)
1385 assert_spin_locked(&cli->cl_loi_list_lock);
1386 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1391 pga->flag &= ~OBD_BRW_FROM_GRANT;
1392 atomic_long_dec(&obd_dirty_pages);
1393 cli->cl_dirty_pages--;
1398 * To avoid sleeping with object lock held, it's good for us allocate enough
1399 * grants before entering into critical section.
1401 * client_obd_list_lock held by caller
1403 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1407 if (cli->cl_avail_grant >= bytes) {
1408 cli->cl_avail_grant -= bytes;
1409 cli->cl_reserved_grant += bytes;
1415 static void __osc_unreserve_grant(struct client_obd *cli,
1416 unsigned int reserved, unsigned int unused)
1418 /* it's quite normal for us to get more grant than reserved.
1419 * Thinking about a case that two extents merged by adding a new
1420 * chunk, we can save one extent tax. If extent tax is greater than
1421 * one chunk, we can save more grant by adding a new chunk */
1422 cli->cl_reserved_grant -= reserved;
1423 if (unused > reserved) {
1424 cli->cl_avail_grant += reserved;
1425 cli->cl_lost_grant += unused - reserved;
1426 cli->cl_dirty_grant -= unused - reserved;
1428 cli->cl_avail_grant += unused;
1429 cli->cl_dirty_grant += reserved - unused;
1433 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1434 unsigned int reserved,
1435 unsigned int unused)
1437 __osc_unreserve_grant(cli, reserved, unused);
1439 osc_wake_cache_waiters(cli);
1442 static void osc_unreserve_grant(struct client_obd *cli,
1443 unsigned int reserved, unsigned int unused)
1445 spin_lock(&cli->cl_loi_list_lock);
1446 osc_unreserve_grant_nolock(cli, reserved, unused);
1447 spin_unlock(&cli->cl_loi_list_lock);
1451 * Free grant after IO is finished or canceled.
1453 * @lost_grant is used to remember how many grants we have allocated but not
1454 * used, we should return these grants to OST. There're two cases where grants
1457 * 2. Without OBD_CONNECT_GRANT support and blocksize at OST is less than
1458 * PAGE_SIZE and a partial page was written. In this case OST may use less
1459 * chunks to serve this partial write. OSTs don't actually know the page
1460 * size on the client side. so clients have to calculate lost grant by the
1461 * blocksize on the OST. See tgt_grant_check() for details.
1463 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1464 unsigned int lost_grant, unsigned int dirty_grant)
1466 unsigned long grant;
1468 grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1470 spin_lock(&cli->cl_loi_list_lock);
1471 atomic_long_sub(nr_pages, &obd_dirty_pages);
1472 cli->cl_dirty_pages -= nr_pages;
1473 cli->cl_lost_grant += lost_grant;
1474 cli->cl_dirty_grant -= dirty_grant;
1475 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1476 /* borrow some grant from truncate to avoid the case that
1477 * truncate uses up all avail grant */
1478 cli->cl_lost_grant -= grant;
1479 cli->cl_avail_grant += grant;
1481 osc_wake_cache_waiters(cli);
1482 spin_unlock(&cli->cl_loi_list_lock);
1483 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1484 lost_grant, cli->cl_lost_grant,
1485 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1486 cli->cl_dirty_grant);
1490 * The companion to osc_enter_cache(), called when @oap is no longer part of
1491 * the dirty accounting due to error.
1493 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1495 spin_lock(&cli->cl_loi_list_lock);
1496 osc_release_write_grant(cli, &oap->oap_brw_page);
1497 spin_unlock(&cli->cl_loi_list_lock);
1501 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1504 static int osc_enter_cache_try(struct client_obd *cli,
1505 struct osc_async_page *oap,
1510 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1512 rc = osc_reserve_grant(cli, bytes);
1516 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1517 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1518 obd_max_dirty_pages) {
1519 osc_consume_write_grant(cli, &oap->oap_brw_page);
1523 atomic_long_dec(&obd_dirty_pages);
1525 __osc_unreserve_grant(cli, bytes, bytes);
1531 /* Following two inlines exist to pass code fragments
1532 * to wait_event_idle_exclusive_timeout_cmd(). Passing
1533 * code fragments as macro args can look confusing, so
1534 * we provide inlines to encapsulate them.
1536 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1537 struct client_obd *cli,
1538 struct osc_async_page *oap)
1540 spin_unlock(&cli->cl_loi_list_lock);
1541 osc_io_unplug_async(env, cli, NULL);
1543 "%s: sleeping for cache space for %p\n",
1544 cli_name(cli), oap);
1547 static inline void cli_lock_after_unplug(struct client_obd *cli)
1549 spin_lock(&cli->cl_loi_list_lock);
1552 * The main entry to reserve dirty page accounting. Usually the grant reserved
1553 * in this function will be freed in bulk in osc_free_grant() unless it fails
1554 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1556 * The process will be put into sleep if it's already run out of grant.
1558 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1559 struct osc_async_page *oap, int bytes)
1561 struct osc_object *osc = oap->oap_obj;
1562 struct lov_oinfo *loi = osc->oo_oinfo;
1565 bool entered = false;
1566 /* We cannot wait for a long time here since we are holding ldlm lock
1567 * across the actual IO. If no requests complete fast (e.g. due to
1568 * overloaded OST that takes a long time to process everything, we'd
1569 * get evicted if we wait for a normal obd_timeout or some such.
1570 * So we try to wait half the time it would take the client to be
1571 * evicted by server which is half obd_timeout when AT is off
1572 * or at least ldlm_enqueue_min with AT on.
1574 unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1575 ldlm_enqueue_min / 2);
1579 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1581 spin_lock(&cli->cl_loi_list_lock);
1583 /* force the caller to try sync io. this can jump the list
1584 * of queued writes and create a discontiguous rpc stream */
1585 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1586 cli->cl_dirty_max_pages == 0 ||
1587 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1588 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1589 GOTO(out, rc = -EDQUOT);
1593 * We can wait here for two reasons: too many dirty pages in cache, or
1594 * run out of grants. In both cases we should write dirty pages out.
1595 * Adding a cache waiter will trigger urgent write-out no matter what
1597 * The exiting condition (other than success) is no avail grants
1598 * and no dirty pages caching, that really means there is no space
1601 remain = wait_event_idle_exclusive_timeout_cmd(
1602 cli->cl_cache_waiters,
1603 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1604 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1606 cli_unlock_and_unplug(env, cli, oap),
1607 cli_lock_after_unplug(cli));
1610 if (remain == timeout)
1611 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1613 OSC_DUMP_GRANT(D_CACHE, cli,
1614 "finally got grant space\n");
1615 wake_up(&cli->cl_cache_waiters);
1617 } else if (remain == 0) {
1618 OSC_DUMP_GRANT(D_CACHE, cli,
1619 "timeout, fall back to sync i/o\n");
1620 osc_extent_tree_dump(D_CACHE, osc);
1621 /* fall back to synchronous I/O */
1623 OSC_DUMP_GRANT(D_CACHE, cli,
1624 "no grant space, fall back to sync i/o\n");
1625 wake_up_all(&cli->cl_cache_waiters);
1629 spin_unlock(&cli->cl_loi_list_lock);
1633 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1635 int hprpc = !!list_empty(&osc->oo_hp_exts);
1636 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1639 /* This maintains the lists of pending pages to read/write for a given object
1640 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1641 * to quickly find objects that are ready to send an RPC. */
1642 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1645 int invalid_import = 0;
1648 /* if we have an invalid import we want to drain the queued pages
1649 * by forcing them through rpcs that immediately fail and complete
1650 * the pages. recovery relies on this to empty the queued pages
1651 * before canceling the locks and evicting down the llite pages */
1652 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1655 if (cmd & OBD_BRW_WRITE) {
1656 if (atomic_read(&osc->oo_nr_writes) == 0)
1658 if (invalid_import) {
1659 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1662 if (!list_empty(&osc->oo_hp_exts)) {
1663 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1666 if (!list_empty(&osc->oo_urgent_exts)) {
1667 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1670 /* trigger a write rpc stream as long as there are dirtiers
1671 * waiting for space. as they're waiting, they're not going to
1672 * create more pages to coalesce with what's waiting..
1674 if (waitqueue_active(&cli->cl_cache_waiters)) {
1675 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1678 if (!list_empty(&osc->oo_full_exts)) {
1679 CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1683 if (atomic_read(&osc->oo_nr_reads) == 0)
1685 if (invalid_import) {
1686 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1689 /* all read are urgent. */
1690 if (!list_empty(&osc->oo_reading_exts))
1697 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1699 struct client_obd *cli = osc_cli(obj);
1700 if (cmd & OBD_BRW_WRITE) {
1701 atomic_add(delta, &obj->oo_nr_writes);
1702 atomic_add(delta, &cli->cl_pending_w_pages);
1703 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1705 atomic_add(delta, &obj->oo_nr_reads);
1706 atomic_add(delta, &cli->cl_pending_r_pages);
1707 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1709 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1712 static int osc_makes_hprpc(struct osc_object *obj)
1714 return !list_empty(&obj->oo_hp_exts);
1717 static void on_list(struct list_head *item, struct list_head *list,
1720 if (list_empty(item) && should_be_on)
1721 list_add_tail(item, list);
1722 else if (!list_empty(item) && !should_be_on)
1723 list_del_init(item);
1726 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1727 * can find pages to build into rpcs quickly */
1728 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1730 if (osc_makes_hprpc(osc)) {
1732 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1733 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1735 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1736 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1737 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1738 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1741 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1742 atomic_read(&osc->oo_nr_writes) > 0);
1744 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1745 atomic_read(&osc->oo_nr_reads) > 0);
1747 return osc_is_ready(osc);
1750 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1754 spin_lock(&cli->cl_loi_list_lock);
1755 is_ready = __osc_list_maint(cli, osc);
1756 spin_unlock(&cli->cl_loi_list_lock);
1761 /* this is trying to propogate async writeback errors back up to the
1762 * application. As an async write fails we record the error code for later if
1763 * the app does an fsync. As long as errors persist we force future rpcs to be
1764 * sync so that the app can get a sync error and break the cycle of queueing
1765 * pages for which writeback will fail. */
1766 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1773 ar->ar_force_sync = 1;
1774 ar->ar_min_xid = ptlrpc_sample_next_xid();
1779 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1780 ar->ar_force_sync = 0;
1783 /* this must be called holding the loi list lock to give coverage to exit_cache,
1784 * async_flag maintenance, and oap_request */
1785 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1786 struct osc_async_page *oap, int sent, int rc)
1788 struct osc_object *osc = oap->oap_obj;
1789 struct lov_oinfo *loi = osc->oo_oinfo;
1793 if (oap->oap_request != NULL) {
1794 xid = ptlrpc_req_xid(oap->oap_request);
1795 ptlrpc_req_finished(oap->oap_request);
1796 oap->oap_request = NULL;
1799 /* As the transfer for this page is being done, clear the flags */
1800 spin_lock(&oap->oap_lock);
1801 oap->oap_async_flags = 0;
1802 spin_unlock(&oap->oap_lock);
1804 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1805 spin_lock(&cli->cl_loi_list_lock);
1806 osc_process_ar(&cli->cl_ar, xid, rc);
1807 osc_process_ar(&loi->loi_ar, xid, rc);
1808 spin_unlock(&cli->cl_loi_list_lock);
1811 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1813 CERROR("completion on oap %p obj %p returns %d.\n",
1819 struct extent_rpc_data {
1820 struct list_head *erd_rpc_list;
1821 unsigned int erd_page_count;
1822 unsigned int erd_max_pages;
1823 unsigned int erd_max_chunks;
1824 unsigned int erd_max_extents;
1827 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1829 struct client_obd *cli = osc_cli(ext->oe_obj);
1830 unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1832 return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1836 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1838 if (ext->oe_no_merge || in_rpc->oe_no_merge)
1841 if (ext->oe_srvlock != in_rpc->oe_srvlock)
1844 if (ext->oe_ndelay != in_rpc->oe_ndelay)
1847 if (!ext->oe_grants != !in_rpc->oe_grants)
1850 if (ext->oe_dio != in_rpc->oe_dio)
1853 /* It's possible to have overlap on DIO */
1854 if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1857 if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1864 * Try to add extent to one RPC. We need to think about the following things:
1865 * - # of pages must not be over max_pages_per_rpc
1866 * - extent must be compatible with previous ones
1868 static int try_to_add_extent_for_io(struct client_obd *cli,
1869 struct osc_extent *ext,
1870 struct extent_rpc_data *data)
1872 struct osc_extent *tmp;
1873 unsigned int chunk_count;
1876 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1878 OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1880 if (data->erd_max_extents == 0)
1883 chunk_count = osc_extent_chunks(ext);
1884 EASSERTF(data->erd_page_count != 0 ||
1885 chunk_count <= data->erd_max_chunks, ext,
1886 "The first extent to be fit in a RPC contains %u chunks, "
1887 "which is over the limit %u.\n", chunk_count,
1888 data->erd_max_chunks);
1889 if (chunk_count > data->erd_max_chunks)
1892 data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1893 EASSERTF(data->erd_page_count != 0 ||
1894 ext->oe_nr_pages <= data->erd_max_pages, ext,
1895 "The first extent to be fit in a RPC contains %u pages, "
1896 "which is over the limit %u.\n", ext->oe_nr_pages,
1897 data->erd_max_pages);
1898 if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1901 list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1902 EASSERT(tmp->oe_owner == current, tmp);
1904 if (!can_merge(ext, tmp))
1908 data->erd_max_extents--;
1909 data->erd_max_chunks -= chunk_count;
1910 data->erd_page_count += ext->oe_nr_pages;
1911 list_move_tail(&ext->oe_link, data->erd_rpc_list);
1912 ext->oe_owner = current;
1917 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1918 * get_write_extent() takes all appropriate extents in atomic.
1920 * The following policy is used to collect extents for IO:
1921 * 1. Add as many HP extents as possible;
1922 * 2. Add the first urgent extent in urgent extent list and take it out of
1924 * 3. Add subsequent extents of this urgent extent;
1925 * 4. If urgent list is not empty, goto 2;
1926 * 5. Traverse the extent tree from the 1st extent;
1927 * 6. Above steps exit if there is no space in this RPC.
1929 static unsigned int get_write_extents(struct osc_object *obj,
1930 struct list_head *rpclist)
1932 struct client_obd *cli = osc_cli(obj);
1933 struct osc_extent *ext;
1934 struct extent_rpc_data data = {
1935 .erd_rpc_list = rpclist,
1936 .erd_page_count = 0,
1937 .erd_max_pages = cli->cl_max_pages_per_rpc,
1938 .erd_max_chunks = osc_max_write_chunks(cli),
1939 .erd_max_extents = 256,
1942 assert_osc_object_is_locked(obj);
1943 while ((ext = list_first_entry_or_null(&obj->oo_hp_exts,
1945 oe_link)) != NULL) {
1946 if (!try_to_add_extent_for_io(cli, ext, &data))
1947 return data.erd_page_count;
1948 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1950 if (data.erd_page_count == data.erd_max_pages)
1951 return data.erd_page_count;
1953 while ((ext = list_first_entry_or_null(&obj->oo_urgent_exts,
1955 oe_link)) != NULL) {
1956 if (!try_to_add_extent_for_io(cli, ext, &data))
1957 return data.erd_page_count;
1959 if (data.erd_page_count == data.erd_max_pages)
1960 return data.erd_page_count;
1962 /* One key difference between full extents and other extents: full
1963 * extents can usually only be added if the rpclist was empty, so if we
1964 * can't add one, we continue on to trying to add normal extents. This
1965 * is so we don't miss adding extra extents to an RPC containing high
1966 * priority or urgent extents.
1968 while ((ext = list_first_entry_or_null(&obj->oo_full_exts,
1970 oe_link)) != NULL) {
1971 if (!try_to_add_extent_for_io(cli, ext, &data))
1974 if (data.erd_page_count == data.erd_max_pages)
1975 return data.erd_page_count;
1977 for (ext = first_extent(obj);
1979 ext = next_extent(ext)) {
1980 if ((ext->oe_state != OES_CACHE) ||
1981 /* this extent may be already in current rpclist */
1982 (!list_empty(&ext->oe_link) && ext->oe_owner))
1985 if (!try_to_add_extent_for_io(cli, ext, &data))
1986 return data.erd_page_count;
1988 return data.erd_page_count;
1992 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1993 struct osc_object *osc)
1997 struct osc_extent *ext;
1998 struct osc_extent *tmp;
1999 struct osc_extent *first = NULL;
2000 unsigned int page_count = 0;
2005 assert_osc_object_is_locked(osc);
2007 page_count = get_write_extents(osc, &rpclist);
2008 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2010 if (list_empty(&rpclist))
2013 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2015 list_for_each_entry(ext, &rpclist, oe_link) {
2016 LASSERT(ext->oe_state == OES_CACHE ||
2017 ext->oe_state == OES_LOCK_DONE);
2018 if (ext->oe_state == OES_CACHE)
2019 osc_extent_state_set(ext, OES_LOCKING);
2021 osc_extent_state_set(ext, OES_RPC);
2024 /* we're going to grab page lock, so release object lock because
2025 * lock order is page lock -> object lock. */
2026 osc_object_unlock(osc);
2028 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2029 if (ext->oe_state == OES_LOCKING) {
2030 rc = osc_extent_make_ready(env, ext);
2031 if (unlikely(rc < 0)) {
2032 list_del_init(&ext->oe_link);
2033 osc_extent_finish(env, ext, 0, rc);
2037 if (first == NULL) {
2039 srvlock = ext->oe_srvlock;
2041 LASSERT(srvlock == ext->oe_srvlock);
2045 if (!list_empty(&rpclist)) {
2046 LASSERT(page_count > 0);
2047 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2048 LASSERT(list_empty(&rpclist));
2051 osc_object_lock(osc);
2056 * prepare pages for ASYNC io and put pages in send queue.
2058 * \param cmd OBD_BRW_* macroses
2059 * \param lop pending pages
2061 * \return zero if no page added to send queue.
2062 * \return 1 if pages successfully added to send queue.
2063 * \return negative on errors.
2066 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2067 struct osc_object *osc)
2070 struct osc_extent *ext;
2071 struct osc_extent *next;
2073 struct extent_rpc_data data = {
2074 .erd_rpc_list = &rpclist,
2075 .erd_page_count = 0,
2076 .erd_max_pages = cli->cl_max_pages_per_rpc,
2077 .erd_max_chunks = UINT_MAX,
2078 .erd_max_extents = UINT_MAX,
2083 assert_osc_object_is_locked(osc);
2084 list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2085 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2086 if (!try_to_add_extent_for_io(cli, ext, &data))
2088 osc_extent_state_set(ext, OES_RPC);
2089 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2091 LASSERT(data.erd_page_count <= data.erd_max_pages);
2093 osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2095 if (!list_empty(&rpclist)) {
2096 osc_object_unlock(osc);
2098 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2099 LASSERT(list_empty(&rpclist));
2101 osc_object_lock(osc);
2106 #define list_to_obj(list, item) ({ \
2107 struct list_head *__tmp = (list)->next; \
2108 list_del_init(__tmp); \
2109 list_entry(__tmp, struct osc_object, oo_##item); \
2112 /* This is called by osc_check_rpcs() to find which objects have pages that
2113 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2114 static struct osc_object *osc_next_obj(struct client_obd *cli)
2118 /* First return objects that have blocked locks so that they
2119 * will be flushed quickly and other clients can get the lock,
2120 * then objects which have pages ready to be stuffed into RPCs */
2121 if (!list_empty(&cli->cl_loi_hp_ready_list))
2122 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2123 if (!list_empty(&cli->cl_loi_ready_list))
2124 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2126 /* then if we have cache waiters, return all objects with queued
2127 * writes. This is especially important when many small files
2128 * have filled up the cache and not been fired into rpcs because
2129 * they don't pass the nr_pending/object threshhold
2131 if (waitqueue_active(&cli->cl_cache_waiters) &&
2132 !list_empty(&cli->cl_loi_write_list))
2133 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2135 /* then return all queued objects when we have an invalid import
2136 * so that they get flushed */
2137 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2138 if (!list_empty(&cli->cl_loi_write_list))
2139 RETURN(list_to_obj(&cli->cl_loi_write_list,
2141 if (!list_empty(&cli->cl_loi_read_list))
2142 RETURN(list_to_obj(&cli->cl_loi_read_list,
2148 /* called with the loi list lock held */
2149 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2150 __must_hold(&cli->cl_loi_list_lock)
2152 struct osc_object *osc;
2156 while ((osc = osc_next_obj(cli)) != NULL) {
2157 struct cl_object *obj = osc2cl(osc);
2158 struct lu_ref_link link;
2160 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2162 /* even if we have reached our max in flight RPCs, we still
2163 * allow all high-priority RPCs through to prevent their
2164 * starvation and leading to server evicting us for not
2165 * writing out pages in a timely manner LU-13131 */
2166 if (osc_max_rpc_in_flight(cli, osc) &&
2167 list_empty(&osc->oo_hp_exts)) {
2168 __osc_list_maint(cli, osc);
2173 spin_unlock(&cli->cl_loi_list_lock);
2174 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2176 /* attempt some read/write balancing by alternating between
2177 * reads and writes in an object. The makes_rpc checks here
2178 * would be redundant if we were getting read/write work items
2179 * instead of objects. we don't want send_oap_rpc to drain a
2180 * partial read pending queue when we're given this object to
2181 * do io on writes while there are cache waiters */
2182 osc_object_lock(osc);
2183 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2184 rc = osc_send_write_rpc(env, cli, osc);
2186 CERROR("Write request failed with %d\n", rc);
2188 /* osc_send_write_rpc failed, mostly because of
2191 * It can't break here, because if:
2192 * - a page was submitted by osc_io_submit, so
2194 * - no request in flight
2195 * - no subsequent request
2196 * The system will be in live-lock state,
2197 * because there is no chance to call
2198 * osc_io_unplug() and osc_check_rpcs() any
2199 * more. pdflush can't help in this case,
2200 * because it might be blocked at grabbing
2201 * the page lock as we mentioned.
2203 * Anyway, continue to drain pages. */
2207 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2208 rc = osc_send_read_rpc(env, cli, osc);
2210 CERROR("Read request failed with %d\n", rc);
2212 osc_object_unlock(osc);
2214 osc_list_maint(cli, osc);
2215 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2216 cl_object_put(env, obj);
2218 spin_lock(&cli->cl_loi_list_lock);
2222 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2223 struct osc_object *osc, int async)
2227 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2231 spin_lock(&cli->cl_loi_list_lock);
2232 osc_check_rpcs(env, cli);
2233 spin_unlock(&cli->cl_loi_list_lock);
2235 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2236 LASSERT(cli->cl_writeback_work != NULL);
2237 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2241 EXPORT_SYMBOL(osc_io_unplug0);
2243 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2244 struct cl_page *page, loff_t offset)
2246 struct obd_export *exp = osc_export(osc);
2247 struct osc_async_page *oap = &ops->ops_oap;
2248 struct page *vmpage = page->cp_vmpage;
2252 return cfs_size_round(sizeof(*oap));
2254 oap->oap_magic = OAP_MAGIC;
2255 oap->oap_cli = &exp->exp_obd->u.cli;
2258 oap->oap_page = vmpage;
2259 oap->oap_obj_off = offset;
2260 LASSERT(!(offset & ~PAGE_MASK));
2262 /* Count of transient (direct i/o) pages is always stable by the time
2263 * they're submitted. Setting this here lets us avoid calling
2264 * cl_page_clip later to set this.
2266 if (page->cp_type == CPT_TRANSIENT)
2267 oap->oap_async_flags |= ASYNC_COUNT_STABLE|ASYNC_URGENT|
2270 INIT_LIST_HEAD(&oap->oap_pending_item);
2271 INIT_LIST_HEAD(&oap->oap_rpc_item);
2273 spin_lock_init(&oap->oap_lock);
2274 CDEBUG(D_INFO, "oap %p vmpage %p obj off %llu\n",
2275 oap, vmpage, oap->oap_obj_off);
2278 EXPORT_SYMBOL(osc_prep_async_page);
2280 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2281 struct osc_page *ops, cl_commit_cbt cb)
2283 struct osc_io *oio = osc_env_io(env);
2284 struct osc_extent *ext = NULL;
2285 struct osc_async_page *oap = &ops->ops_oap;
2286 struct client_obd *cli = oap->oap_cli;
2287 struct osc_object *osc = oap->oap_obj;
2288 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
2291 unsigned int grants = 0;
2292 u32 brw_flags = OBD_BRW_ASYNC;
2293 int cmd = OBD_BRW_WRITE;
2294 int need_release = 0;
2298 if (oap->oap_magic != OAP_MAGIC)
2301 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2304 if (!list_empty(&oap->oap_pending_item) ||
2305 !list_empty(&oap->oap_rpc_item))
2308 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2309 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2310 if (io->ci_noquota) {
2311 brw_flags |= OBD_BRW_NOQUOTA;
2312 cmd |= OBD_BRW_NOQUOTA;
2315 if (oio->oi_cap_sys_resource) {
2316 brw_flags |= OBD_BRW_SYS_RESOURCE;
2317 cmd |= OBD_BRW_SYS_RESOURCE;
2320 /* check if the file's owner/group is over quota */
2321 /* do not check for root without root squash, because in this case
2322 * we should bypass quota
2324 if ((!oio->oi_cap_sys_resource ||
2325 cli->cl_root_squash) &&
2327 struct cl_object *obj;
2328 struct cl_attr *attr;
2329 unsigned int qid[LL_MAXQUOTAS];
2331 obj = cl_object_top(&osc->oo_cl);
2332 attr = &osc_env_info(env)->oti_attr;
2334 cl_object_attr_lock(obj);
2335 rc = cl_object_attr_get(env, obj, attr);
2336 cl_object_attr_unlock(obj);
2338 qid[USRQUOTA] = attr->cat_uid;
2339 qid[GRPQUOTA] = attr->cat_gid;
2340 qid[PRJQUOTA] = attr->cat_projid;
2341 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2348 oap->oap_page_off = ops->ops_from;
2349 oap->oap_count = ops->ops_to - ops->ops_from + 1;
2350 /* No need to hold a lock here,
2351 * since this page is not in any list yet. */
2352 oap->oap_async_flags = 0;
2353 oap->oap_brw_flags = brw_flags;
2355 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2356 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2358 index = osc_index(oap2osc(oap));
2360 /* Add this page into extent by the following steps:
2361 * 1. if there exists an active extent for this IO, mostly this page
2362 * can be added to the active extent and sometimes we need to
2363 * expand extent to accomodate this page;
2364 * 2. otherwise, a new extent will be allocated. */
2366 ext = oio->oi_active;
2367 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2368 /* one chunk plus extent overhead must be enough to write this
2370 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2371 if (ext->oe_end >= index)
2374 /* it doesn't need any grant to dirty this page */
2375 spin_lock(&cli->cl_loi_list_lock);
2376 rc = osc_enter_cache_try(cli, oap, grants);
2377 if (rc == 0) { /* try failed */
2380 } else if (ext->oe_end < index) {
2382 /* try to expand this extent */
2383 rc = osc_extent_expand(ext, index, &tmp);
2386 /* don't free reserved grant */
2388 OSC_EXTENT_DUMP(D_CACHE, ext,
2389 "expanded for %lu.\n", index);
2390 osc_unreserve_grant_nolock(cli, grants, tmp);
2394 spin_unlock(&cli->cl_loi_list_lock);
2396 } else if (ext != NULL) {
2397 /* index is located outside of active extent */
2401 osc_extent_release(env, ext);
2402 oio->oi_active = NULL;
2407 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2409 /* try to find new extent to cover this page */
2410 LASSERT(oio->oi_active == NULL);
2411 /* we may have allocated grant for this page if we failed
2412 * to expand the previous active extent. */
2413 LASSERT(ergo(grants > 0, grants >= tmp));
2417 /* We must not hold a page lock while we do osc_enter_cache()
2418 * or osc_extent_find(), so we must mark dirty & unlock
2419 * any pages in the write commit pagevec. */
2420 if (pagevec_count(pvec)) {
2422 pagevec_reinit(pvec);
2426 rc = osc_enter_cache(env, cli, oap, tmp);
2433 ext = osc_extent_find(env, osc, index, &tmp);
2435 LASSERT(tmp == grants);
2436 osc_exit_cache(cli, oap);
2440 oio->oi_active = ext;
2444 osc_unreserve_grant(cli, grants, tmp);
2447 LASSERT(ergo(rc == 0, ext != NULL));
2449 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2450 ext, "index = %lu.\n", index);
2451 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2453 osc_object_lock(osc);
2454 if (ext->oe_nr_pages == 0)
2455 ext->oe_srvlock = ops->ops_srvlock;
2457 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2459 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2460 osc_object_unlock(osc);
2462 if (!ext->oe_layout_version)
2463 ext->oe_layout_version = io->ci_layout_version;
2469 int osc_teardown_async_page(const struct lu_env *env,
2470 struct osc_object *obj, struct osc_page *ops)
2472 struct osc_async_page *oap = &ops->ops_oap;
2476 LASSERT(oap->oap_magic == OAP_MAGIC);
2478 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2479 oap, ops, osc_index(oap2osc(oap)));
2481 if (!list_empty(&oap->oap_rpc_item)) {
2482 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2484 } else if (!list_empty(&oap->oap_pending_item)) {
2485 struct osc_extent *ext = NULL;
2487 osc_object_lock(obj);
2488 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2489 osc_object_unlock(obj);
2490 /* only truncated pages are allowed to be taken out.
2491 * See osc_extent_truncate() and osc_cache_truncate_start()
2493 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2494 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2495 osc_index(oap2osc(oap)));
2499 osc_extent_put(env, ext);
2505 * This is called when a page is picked up by kernel to write out.
2507 * We should find out the corresponding extent and add the whole extent
2508 * into urgent list. The extent may be being truncated or used, handle it
2511 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2512 struct osc_page *ops)
2514 struct osc_extent *ext = NULL;
2515 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2516 struct cl_page *cp = ops->ops_cl.cpl_page;
2517 pgoff_t index = osc_index(ops);
2518 struct osc_async_page *oap = &ops->ops_oap;
2519 bool unplug = false;
2523 osc_object_lock(obj);
2524 ext = osc_extent_lookup(obj, index);
2526 osc_extent_tree_dump(D_ERROR, obj);
2527 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2530 switch (ext->oe_state) {
2533 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2537 /* If we know this extent is being written out, we should abort
2538 * so that the writer can make this page ready. Otherwise, there
2539 * exists a deadlock problem because other process can wait for
2540 * page writeback bit holding page lock; and meanwhile in
2541 * vvp_page_make_ready(), we need to grab page lock before
2542 * really sending the RPC. */
2544 /* race with truncate, page will be redirtied */
2546 /* The extent is active so we need to abort and let the caller
2547 * re-dirty the page. If we continued on here, and we were the
2548 * one making the extent active, we could deadlock waiting for
2549 * the page writeback to clear but it won't because the extent
2550 * is active and won't be written out. */
2551 GOTO(out, rc = -EAGAIN);
2556 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2560 spin_lock(&oap->oap_lock);
2561 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2562 spin_unlock(&oap->oap_lock);
2564 if (current->flags & PF_MEMALLOC)
2565 ext->oe_memalloc = 1;
2568 if (ext->oe_state == OES_CACHE) {
2569 OSC_EXTENT_DUMP(D_CACHE, ext,
2570 "flush page %p make it urgent.\n", oap);
2571 if (list_empty(&ext->oe_link))
2572 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2579 osc_object_unlock(obj);
2580 osc_extent_put(env, ext);
2582 osc_io_unplug_async(env, osc_cli(obj), obj);
2586 int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io,
2587 struct osc_object *obj, struct list_head *list,
2590 struct client_obd *cli = osc_cli(obj);
2591 struct osc_extent *ext;
2592 struct osc_async_page *oap;
2594 int mppr = cli->cl_max_pages_per_rpc;
2595 bool can_merge = true;
2596 pgoff_t start = CL_PAGE_EOF;
2600 list_for_each_entry(oap, list, oap_pending_item) {
2601 struct osc_page *opg = oap2osc_page(oap);
2602 pgoff_t index = osc_index(opg);
2609 mppr <<= (page_count > mppr);
2611 if (unlikely(opg->ops_from > 0 ||
2612 opg->ops_to < PAGE_SIZE - 1))
2616 ext = osc_extent_alloc(obj);
2618 struct osc_async_page *tmp;
2620 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2621 list_del_init(&oap->oap_pending_item);
2622 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2627 ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2629 ext->oe_no_merge = !can_merge;
2631 ext->oe_start = start;
2632 ext->oe_end = ext->oe_max_end = end;
2634 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2635 ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2636 ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2637 if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2641 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2642 grants = cli->cl_grant_extent_tax;
2643 grants += (1 << cli->cl_chunkbits) *
2644 ((page_count + ppc - 1) / ppc);
2646 CDEBUG(D_CACHE, "requesting %d bytes grant\n", grants);
2647 spin_lock(&cli->cl_loi_list_lock);
2648 if (osc_reserve_grant(cli, grants) == 0) {
2649 list_for_each_entry(oap, list, oap_pending_item) {
2650 osc_consume_write_grant(cli,
2651 &oap->oap_brw_page);
2653 atomic_long_add(page_count, &obd_dirty_pages);
2654 osc_unreserve_grant_nolock(cli, grants, 0);
2655 ext->oe_grants = grants;
2657 /* We cannot report ENOSPC correctly if we do parallel
2658 * DIO (async RPC submission), so turn off parallel dio
2659 * if there is not sufficient grant available. This
2660 * makes individual RPCs synchronous.
2662 io->ci_parallel_dio = false;
2664 "not enough grant available, switching to sync for this i/o\n");
2666 spin_unlock(&cli->cl_loi_list_lock);
2667 osc_update_next_shrink(cli);
2670 ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2671 ext->oe_nr_pages = page_count;
2672 ext->oe_mppr = mppr;
2673 list_splice_init(list, &ext->oe_pages);
2674 ext->oe_layout_version = io->ci_layout_version;
2676 osc_object_lock(obj);
2677 /* Reuse the initial refcount for RPC, don't drop it */
2678 osc_extent_state_set(ext, OES_LOCK_DONE);
2679 if (!ext->oe_rw) { /* write */
2680 if (!ext->oe_srvlock && !ext->oe_dio) {
2681 /* The most likely case here is from lack of grants
2682 * so we are either out of quota or out of space.
2683 * Since this means we are holding locks across
2684 * potentially multi-striped IO, we must send out
2685 * everything out instantly to avoid prolonged
2686 * waits resulting in lock eviction (likely since
2687 * the extended wait in osc_cache_enter() did not
2688 * yield any additional grant due to a timeout.
2691 list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2693 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2695 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2697 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2698 osc_update_pending(obj, OBD_BRW_READ, page_count);
2700 osc_object_unlock(obj);
2702 osc_io_unplug_async(env, cli, obj);
2707 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2709 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2710 __u64 size, struct osc_extent **extp)
2712 struct client_obd *cli = osc_cli(obj);
2713 struct osc_extent *ext;
2714 struct osc_extent *waiting = NULL;
2721 /* pages with index greater or equal to index will be truncated. */
2722 index = cl_index(osc2cl(obj), size);
2723 partial = size > cl_offset(osc2cl(obj), index);
2726 osc_object_lock(obj);
2727 ext = osc_extent_search(obj, index);
2729 ext = first_extent(obj);
2730 else if (ext->oe_end < index)
2731 ext = next_extent(ext);
2732 while (ext != NULL) {
2733 EASSERT(ext->oe_state != OES_TRUNC, ext);
2735 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2736 /* if ext is in urgent state, it means there must exist
2737 * a page already having been flushed by write_page().
2738 * We have to wait for this extent because we can't
2739 * truncate that page. */
2740 OSC_EXTENT_DUMP(D_CACHE, ext,
2741 "waiting for busy extent\n");
2742 waiting = osc_extent_get(ext);
2746 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2748 osc_extent_get(ext);
2749 if (ext->oe_state == OES_ACTIVE) {
2750 /* though we grab inode mutex for write path, but we
2751 * release it before releasing extent(in osc_io_end()),
2752 * so there is a race window that an extent is still
2753 * in OES_ACTIVE when truncate starts. */
2754 LASSERT(!ext->oe_trunc_pending);
2755 ext->oe_trunc_pending = 1;
2757 EASSERT(ext->oe_state == OES_CACHE, ext);
2758 osc_extent_state_set(ext, OES_TRUNC);
2759 osc_update_pending(obj, OBD_BRW_WRITE,
2762 /* This extent could be on the full extents list, that's OK */
2763 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2764 if (!list_empty(&ext->oe_link))
2765 list_move_tail(&ext->oe_link, &list);
2767 list_add_tail(&ext->oe_link, &list);
2769 ext = next_extent(ext);
2771 osc_object_unlock(obj);
2773 osc_list_maint(cli, obj);
2775 while ((ext = list_first_entry_or_null(&list,
2777 oe_link)) != NULL) {
2780 list_del_init(&ext->oe_link);
2782 /* extent may be in OES_ACTIVE state because inode mutex
2783 * is released before osc_io_end() in file write case */
2784 if (ext->oe_state != OES_TRUNC)
2785 osc_extent_wait(env, ext, OES_TRUNC);
2787 rc = osc_extent_truncate(ext, index, partial);
2792 OSC_EXTENT_DUMP(D_ERROR, ext,
2793 "truncate error %d\n", rc);
2794 } else if (ext->oe_nr_pages == 0) {
2795 osc_extent_remove(ext);
2797 /* this must be an overlapped extent which means only
2798 * part of pages in this extent have been truncated.
2800 EASSERTF(ext->oe_start <= index, ext,
2801 "trunc index = %lu/%d.\n", index, partial);
2802 /* fix index to skip this partially truncated extent */
2803 index = ext->oe_end + 1;
2806 /* we need to hold this extent in OES_TRUNC state so
2807 * that no writeback will happen. This is to avoid
2809 * Only partial truncate can reach here, if @size is
2810 * not zero, the caller should provide a valid @extp. */
2811 LASSERT(*extp == NULL);
2812 *extp = osc_extent_get(ext);
2813 OSC_EXTENT_DUMP(D_CACHE, ext,
2814 "trunc at %llu\n", size);
2816 osc_extent_put(env, ext);
2818 if (waiting != NULL) {
2821 /* ignore the result of osc_extent_wait the write initiator
2822 * should take care of it. */
2823 rc = osc_extent_wait(env, waiting, OES_INV);
2825 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2827 osc_extent_put(env, waiting);
2833 EXPORT_SYMBOL(osc_cache_truncate_start);
2836 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2838 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2841 struct osc_object *obj = ext->oe_obj;
2842 bool unplug = false;
2844 EASSERT(ext->oe_nr_pages > 0, ext);
2845 EASSERT(ext->oe_state == OES_TRUNC, ext);
2846 EASSERT(!ext->oe_urgent, ext);
2848 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2849 osc_object_lock(obj);
2850 osc_extent_state_set(ext, OES_CACHE);
2851 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2853 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2856 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2857 osc_object_unlock(obj);
2858 osc_extent_put(env, ext);
2861 osc_io_unplug_async(env, osc_cli(obj), obj);
2866 * Wait for extents in a specific range to be written out.
2867 * The caller must have called osc_cache_writeback_range() to issue IO
2868 * otherwise it will take a long time for this function to finish.
2870 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2871 * nobody else can dirty this range of file while we're waiting for
2872 * extents to be written.
2874 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2875 pgoff_t start, pgoff_t end)
2877 struct osc_extent *ext;
2878 pgoff_t index = start;
2883 osc_object_lock(obj);
2884 ext = osc_extent_search(obj, index);
2886 ext = first_extent(obj);
2887 else if (ext->oe_end < index)
2888 ext = next_extent(ext);
2889 while (ext != NULL) {
2892 if (ext->oe_start > end)
2895 if (!ext->oe_fsync_wait) {
2896 ext = next_extent(ext);
2900 EASSERT(ergo(ext->oe_state == OES_CACHE,
2901 ext->oe_hp || ext->oe_urgent), ext);
2902 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2903 !ext->oe_hp && ext->oe_urgent), ext);
2905 index = ext->oe_end + 1;
2906 osc_extent_get(ext);
2907 osc_object_unlock(obj);
2909 rc = osc_extent_wait(env, ext, OES_INV);
2912 osc_extent_put(env, ext);
2915 osc_object_unlock(obj);
2917 OSC_IO_DEBUG(obj, "sync file range.\n");
2920 EXPORT_SYMBOL(osc_cache_wait_range);
2923 * Called to write out a range of osc object.
2925 * @hp : should be set this is caused by lock cancel;
2926 * @discard: is set if dirty pages should be dropped - file will be deleted or
2927 * truncated, this implies there is no partially discarding extents.
2929 * Return how many pages will be issued, or error code if error occurred.
2931 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2932 pgoff_t start, pgoff_t end, int hp, int discard)
2934 struct osc_extent *ext;
2935 LIST_HEAD(discard_list);
2936 bool unplug = false;
2940 osc_object_lock(obj);
2941 ext = osc_extent_search(obj, start);
2943 ext = first_extent(obj);
2944 else if (ext->oe_end < start)
2945 ext = next_extent(ext);
2946 while (ext != NULL) {
2947 if (ext->oe_start > end)
2950 ext->oe_fsync_wait = 1;
2951 switch (ext->oe_state) {
2953 result += ext->oe_nr_pages;
2955 struct list_head *list = NULL;
2957 EASSERT(!ext->oe_hp, ext);
2959 list = &obj->oo_hp_exts;
2960 } else if (!ext->oe_urgent && !ext->oe_hp) {
2962 list = &obj->oo_urgent_exts;
2965 list_move_tail(&ext->oe_link, list);
2968 struct client_obd *cli = osc_cli(obj);
2969 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2970 pgoff_t align_by = (1 << pcc_bits);
2971 pgoff_t a_start = round_down(start, align_by);
2972 pgoff_t a_end = round_up(end, align_by);
2976 a_end = CL_PAGE_EOF;
2977 /* the only discarder is lock cancelling, so
2978 * [start, end], aligned by chunk size, must
2979 * contain this extent */
2980 LASSERTF(ext->oe_start >= a_start &&
2981 ext->oe_end <= a_end,
2982 "ext [%lu, %lu] reg [%lu, %lu] "
2983 "orig [%lu %lu] align %lu bits "
2984 "%d\n", ext->oe_start, ext->oe_end,
2985 a_start, a_end, start, end,
2986 align_by, pcc_bits);
2987 osc_extent_state_set(ext, OES_LOCKING);
2988 ext->oe_owner = current;
2989 list_move_tail(&ext->oe_link,
2991 osc_update_pending(obj, OBD_BRW_WRITE,
2996 /* It's pretty bad to wait for ACTIVE extents, because
2997 * we don't know how long we will wait for it to be
2998 * flushed since it may be blocked at awaiting more
2999 * grants. We do this for the correctness of fsync. */
3000 LASSERT(hp == 0 && discard == 0);
3004 /* this extent is being truncated, can't do anything
3005 * for it now. it will be set to urgent after truncate
3006 * is finished in osc_cache_truncate_end(). */
3010 ext = next_extent(ext);
3012 osc_object_unlock(obj);
3014 LASSERT(ergo(!discard, list_empty(&discard_list)));
3015 if (!list_empty(&discard_list)) {
3016 struct osc_extent *tmp;
3019 osc_list_maint(osc_cli(obj), obj);
3020 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3021 list_del_init(&ext->oe_link);
3022 EASSERT(ext->oe_state == OES_LOCKING, ext);
3024 /* Discard caching pages. We don't actually write this
3025 * extent out but we complete it as if we did. */
3026 rc = osc_extent_make_ready(env, ext);
3027 if (unlikely(rc < 0)) {
3028 OSC_EXTENT_DUMP(D_ERROR, ext,
3029 "make_ready returned %d\n", rc);
3034 /* finish the extent as if the pages were sent */
3035 osc_extent_finish(env, ext, 0, 0);
3040 osc_io_unplug(env, osc_cli(obj), obj);
3042 if (hp || discard) {
3044 rc = osc_cache_wait_range(env, obj, start, end);
3045 if (result >= 0 && rc < 0)
3049 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3052 EXPORT_SYMBOL(osc_cache_writeback_range);
3055 * Returns a list of pages by a given [start, end] of \a obj.
3057 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3058 * crucial in the face of [offset, EOF] locks.
3060 * Return at least one page in @queue unless there is no covered page.
3062 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3063 struct osc_object *osc, pgoff_t start, pgoff_t end,
3064 osc_page_gang_cbt cb, void *cbdata)
3066 struct osc_page *ops;
3067 struct pagevec *pagevec;
3074 bool tree_lock = true;
3078 pvec = osc_env_info(env)->oti_pvec;
3079 pagevec = &osc_env_info(env)->oti_pagevec;
3080 ll_pagevec_init(pagevec, 0);
3081 spin_lock(&osc->oo_tree_lock);
3082 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3083 idx, OTI_PVEC_SIZE)) > 0) {
3084 struct cl_page *page;
3085 bool end_of_region = false;
3087 for (i = 0, j = 0; i < nr; ++i) {
3091 idx = osc_index(ops);
3093 end_of_region = true;
3097 page = ops->ops_cl.cpl_page;
3098 LASSERT(page->cp_type == CPT_CACHEABLE);
3099 if (page->cp_state == CPS_FREEING)
3103 lu_ref_add_atomic(&page->cp_reference,
3104 "gang_lookup", current);
3110 * Here a delicate locking dance is performed. Current thread
3111 * holds a reference to a page, but has to own it before it
3112 * can be placed into queue. Owning implies waiting, so
3113 * radix-tree lock is to be released. After a wait one has to
3114 * check that pages weren't truncated (cl_page_own() returns
3115 * error in the latter case).
3117 spin_unlock(&osc->oo_tree_lock);
3120 res = (*cb)(env, io, pvec, j, cbdata);
3122 for (i = 0; i < j; ++i) {
3124 page = ops->ops_cl.cpl_page;
3125 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3126 cl_pagevec_put(env, page, pagevec);
3128 pagevec_release(pagevec);
3130 if (nr < OTI_PVEC_SIZE || end_of_region)
3136 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SLOW_PAGE_EVICT,
3137 cfs_fail_val ?: 20);
3139 if (io->ci_type == CIT_MISC &&
3140 io->u.ci_misc.lm_next_rpc_time &&
3141 ktime_get_seconds() > io->u.ci_misc.lm_next_rpc_time) {
3142 osc_send_empty_rpc(osc, idx << PAGE_SHIFT);
3143 io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3144 5 * obd_timeout / 16;
3150 spin_lock(&osc->oo_tree_lock);
3154 spin_unlock(&osc->oo_tree_lock);
3157 EXPORT_SYMBOL(osc_page_gang_lookup);
3160 * Check if page @page is covered by an extra lock or discard it.
3162 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3163 void **pvec, int count, void *cbdata)
3165 struct osc_thread_info *info = osc_env_info(env);
3166 struct osc_object *osc = cbdata;
3169 for (i = 0; i < count; i++) {
3170 struct osc_page *ops = pvec[i];
3171 struct cl_page *page = ops->ops_cl.cpl_page;
3172 pgoff_t index = osc_index(ops);
3173 bool discard = false;
3175 /* negative lock caching */
3176 if (index < info->oti_ng_index) {
3178 } else if (index >= info->oti_fn_index) {
3179 struct ldlm_lock *tmp;
3180 /* refresh non-overlapped index */
3181 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3182 OSC_DAP_FL_TEST_LOCK |
3187 tmp->l_policy_data.l_extent.end;
3189 tmp->l_policy_data.l_extent.start;
3191 /* no lock covering this page */
3192 if (index < cl_index(osc2cl(osc), start)) {
3193 /* no lock at @index,
3194 * first lock at @start
3196 info->oti_ng_index =
3197 cl_index(osc2cl(osc), start);
3200 /* Cache the first-non-overlapped
3201 * index so as to skip all pages
3202 * within [index, oti_fn_index).
3203 * This is safe because if tmp lock
3204 * is canceled, it will discard these
3207 info->oti_fn_index =
3208 cl_index(osc2cl(osc), end + 1);
3209 if (end == OBD_OBJECT_EOF)
3210 info->oti_fn_index =
3215 info->oti_ng_index = CL_PAGE_EOF;
3221 if (cl_page_own(env, io, page) == 0) {
3222 cl_page_discard(env, io, page);
3223 cl_page_disown(env, io, page);
3225 LASSERT(page->cp_state == CPS_FREEING);
3229 info->oti_next_index = index + 1;
3234 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3235 void **pvec, int count, void *cbdata)
3237 struct osc_thread_info *info = osc_env_info(env);
3240 for (i = 0; i < count; i++) {
3241 struct osc_page *ops = pvec[i];
3242 struct cl_page *page = ops->ops_cl.cpl_page;
3244 /* page is top page. */
3245 info->oti_next_index = osc_index(ops) + 1;
3246 if (cl_page_own(env, io, page) == 0) {
3247 if (!ergo(page->cp_type == CPT_CACHEABLE,
3248 !PageDirty(cl_page_vmpage(page))))
3249 CL_PAGE_DEBUG(D_ERROR, env, page,
3250 "discard dirty page?\n");
3252 /* discard the page */
3253 cl_page_discard(env, io, page);
3254 cl_page_disown(env, io, page);
3256 LASSERT(page->cp_state == CPS_FREEING);
3262 EXPORT_SYMBOL(osc_discard_cb);
3265 * Discard pages protected by the given lock. This function traverses radix
3266 * tree to find all covering pages and discard them. If a page is being covered
3267 * by other locks, it should remain in cache.
3269 * If error happens on any step, the process continues anyway (the reasoning
3270 * behind this being that lock cancellation cannot be delayed indefinitely).
3272 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3273 pgoff_t start, pgoff_t end, bool discard)
3275 struct osc_thread_info *info = osc_env_info(env);
3276 struct cl_io *io = osc_env_thread_io(env);
3277 osc_page_gang_cbt cb;
3282 io->ci_obj = cl_object_top(osc2cl(osc));
3283 io->ci_ignore_layout = 1;
3284 io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3285 5 * obd_timeout / 16;
3286 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3290 cb = discard ? osc_discard_cb : check_and_discard_cb;
3291 info->oti_fn_index = info->oti_next_index = start;
3292 info->oti_ng_index = 0;
3294 osc_page_gang_lookup(env, io, osc,
3295 info->oti_next_index, end, cb, osc);
3297 cl_io_fini(env, io);