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_object *osc,
51 struct osc_async_page *oap, int sent, int rc);
52 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
54 static int osc_refresh_count(const struct lu_env *env, struct osc_object *osc,
55 struct osc_async_page *oap, int cmd);
56 static int osc_io_unplug_async(const struct lu_env *env,
57 struct client_obd *cli, struct osc_object *osc);
58 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
59 unsigned int lost_grant, unsigned int dirty_grant);
61 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
62 const char *func, int line);
63 #define osc_extent_tree_dump(mask, obj) \
64 osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
66 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
73 /* ------------------ osc extent ------------------ */
74 static inline char *ext_flags(struct osc_extent *ext, char *flags)
77 *buf++ = ext->oe_rw ? 'r' : 'w';
78 if (!RB_EMPTY_NODE(&ext->oe_node))
90 if (ext->oe_trunc_pending)
92 if (ext->oe_fsync_wait)
98 #define EXTSTR "[%lu -> %lu/%lu]"
99 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
100 static const char *const oes_strings[] = {
101 "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
103 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
104 static struct cfs_debug_limit_state cdls; \
105 struct osc_extent *__ext = (extent); \
108 __CDEBUG_WITH_LOC(file, func, line, mask, &cdls, \
109 "extent %p@{" EXTSTR ", " \
110 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \
111 /* ----- extent part 0 ----- */ \
112 __ext, EXTPARA(__ext), \
113 /* ----- part 1 ----- */ \
114 kref_read(&__ext->oe_refc), \
115 atomic_read(&__ext->oe_users), \
116 list_empty_marker(&__ext->oe_link), \
117 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
119 /* ----- part 2 ----- */ \
120 __ext->oe_grants, __ext->oe_nr_pages, \
121 list_empty_marker(&__ext->oe_pages), \
122 waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \
123 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner, \
124 /* ----- part 4 ----- */ \
126 if (mask == D_ERROR && __ext->oe_dlmlock != NULL) \
127 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext); \
129 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext); \
132 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...) \
133 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__, \
134 mask, ext, fmt, ## __VA_ARGS__)
137 #define EASSERTF(expr, ext, fmt, args...) do { \
139 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args); \
140 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj); \
146 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
148 static inline struct osc_extent *rb_extent(struct rb_node *n)
150 return rb_entry_safe(n, struct osc_extent, oe_node);
153 static inline struct osc_extent *next_extent(struct osc_extent *ext)
158 LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
159 return rb_extent(rb_next(&ext->oe_node));
162 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
167 LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
168 return rb_extent(rb_prev(&ext->oe_node));
171 static inline struct osc_extent *first_extent(struct osc_object *obj)
173 return rb_extent(rb_first(&obj->oo_root));
176 /* object must be locked by caller. */
177 static int osc_extent_sanity_check0(struct osc_extent *ext,
178 const char *func, const int line)
180 struct osc_object *obj = ext->oe_obj;
181 struct osc_async_page *oap;
185 assert_osc_object_is_locked(obj);
187 if (ext->oe_state >= OES_STATE_MAX)
190 if (kref_read(&ext->oe_refc) <= 0)
193 if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
196 switch (ext->oe_state) {
198 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
203 if (atomic_read(&ext->oe_users) == 0)
207 if (ext->oe_fsync_wait && !ext->oe_urgent)
211 if (ext->oe_grants == 0)
213 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
217 if (atomic_read(&ext->oe_users) > 0)
221 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
224 if (ext->oe_sync && ext->oe_grants > 0)
227 if (ext->oe_dlmlock != NULL &&
228 ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
229 !ldlm_is_failed(ext->oe_dlmlock)) {
230 struct ldlm_extent *extent;
232 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
233 if (!(extent->start <= ext->oe_start << PAGE_SHIFT &&
234 extent->end >= ext->oe_max_end << PAGE_SHIFT))
237 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
241 if (ext->oe_nr_pages > ext->oe_mppr)
244 /* Do not verify page list if extent is in RPC. This is because an
245 * in-RPC extent is supposed to be exclusively accessible w/o lock. */
246 if (ext->oe_state > OES_CACHE)
253 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
254 pgoff_t index = osc_index(oap2osc(oap));
256 if (index > ext->oe_end || index < ext->oe_start)
259 if (page_count != ext->oe_nr_pages)
264 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
265 "sanity check %p failed: rc = %d\n",
270 #define sanity_check_nolock(ext) \
271 osc_extent_sanity_check0(ext, __func__, __LINE__)
273 #define sanity_check(ext) ({ \
275 osc_object_lock((ext)->oe_obj); \
276 __res = sanity_check_nolock(ext); \
277 osc_object_unlock((ext)->oe_obj); \
282 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
284 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
288 * sanity check - to make sure there is no overlapped extent in the tree.
290 static int osc_extent_is_overlapped(struct osc_object *obj,
291 struct osc_extent *ext)
293 struct osc_extent *tmp;
295 assert_osc_object_is_locked(obj);
300 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
303 if (overlapped(tmp, ext))
309 static void osc_extent_state_set(struct osc_extent *ext, int state)
311 assert_osc_object_is_locked(ext->oe_obj);
312 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
314 /* Never try to sanity check a state changing extent :-) */
315 /* LASSERT(sanity_check_nolock(ext) == 0); */
317 /* TODO: validate the state machine */
318 smp_store_release(&ext->oe_state, state);
319 wake_up(&ext->oe_waitq);
322 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
324 struct osc_extent *ext;
326 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
330 RB_CLEAR_NODE(&ext->oe_node);
332 cl_object_get(osc2cl(obj));
333 kref_init(&ext->oe_refc);
334 atomic_set(&ext->oe_users, 0);
335 INIT_LIST_HEAD(&ext->oe_link);
336 ext->oe_state = OES_INV;
337 INIT_LIST_HEAD(&ext->oe_pages);
338 init_waitqueue_head(&ext->oe_waitq);
339 ext->oe_dlmlock = NULL;
344 static void osc_extent_free(struct kref *kref)
346 struct osc_extent *ext = container_of(kref, struct osc_extent,
349 LASSERT(list_empty(&ext->oe_link));
350 LASSERT(atomic_read(&ext->oe_users) == 0);
351 LASSERT(ext->oe_state == OES_INV);
352 LASSERT(RB_EMPTY_NODE(&ext->oe_node));
354 if (ext->oe_dlmlock) {
355 lu_ref_del(&ext->oe_dlmlock->l_reference,
357 LDLM_LOCK_RELEASE(ext->oe_dlmlock);
358 ext->oe_dlmlock = NULL;
361 /* If/When cl_object_put drops the need for 'env',
362 * this code can be enabled, and matching code in
363 * osc_extent_put removed.
365 cl_object_put(osc2cl(ext->oe_obj));
367 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
371 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
373 LASSERT(kref_read(&ext->oe_refc) >= 0);
374 kref_get(&ext->oe_refc);
378 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
380 LASSERT(kref_read(&ext->oe_refc) > 0);
381 if (kref_put(&ext->oe_refc, osc_extent_free)) {
382 /* This should be in osc_extent_free(), but
383 * while we need to pass 'env' it cannot be.
385 cl_object_put(env, osc2cl(ext->oe_obj));
387 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
392 * osc_extent_put_trust() is a special version of osc_extent_put() when
393 * it's known that the caller is not the last user. This is to address the
394 * problem of lacking of lu_env ;-).
396 static void osc_extent_put_trust(struct osc_extent *ext)
398 LASSERT(kref_read(&ext->oe_refc) > 1);
399 assert_osc_object_is_locked(ext->oe_obj);
400 osc_extent_put(NULL, ext);
404 * Return the extent which includes pgoff @index, or return the greatest
405 * previous extent in the tree.
407 static struct osc_extent *osc_extent_search(struct osc_object *obj,
410 struct rb_node *n = obj->oo_root.rb_node;
411 struct osc_extent *tmp, *p = NULL;
413 assert_osc_object_is_locked(obj);
416 if (index < tmp->oe_start) {
418 } else if (index > tmp->oe_end) {
429 * Return the extent covering @index, otherwise return NULL.
430 * caller must have held object lock.
432 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
435 struct osc_extent *ext;
437 ext = osc_extent_search(obj, index);
438 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
439 return osc_extent_get(ext);
443 /* caller must have held object lock. */
444 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
446 struct rb_node **n = &obj->oo_root.rb_node;
447 struct rb_node *parent = NULL;
448 struct osc_extent *tmp;
450 LASSERT(RB_EMPTY_NODE(&ext->oe_node));
451 LASSERT(ext->oe_obj == obj);
452 assert_osc_object_is_locked(obj);
457 if (ext->oe_end < tmp->oe_start)
459 else if (ext->oe_start > tmp->oe_end)
462 EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
464 rb_link_node(&ext->oe_node, parent, n);
465 rb_insert_color(&ext->oe_node, &obj->oo_root);
469 /* caller must have held object lock. */
470 static void osc_extent_erase(struct osc_extent *ext)
472 struct osc_object *obj = ext->oe_obj;
473 assert_osc_object_is_locked(obj);
474 if (!RB_EMPTY_NODE(&ext->oe_node)) {
475 rb_erase(&ext->oe_node, &obj->oo_root);
476 RB_CLEAR_NODE(&ext->oe_node);
477 /* rbtree held a refcount */
478 osc_extent_put_trust(ext);
482 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
484 struct osc_object *obj = ext->oe_obj;
486 assert_osc_object_is_locked(obj);
487 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
488 if (ext->oe_state == OES_CACHE) {
489 osc_extent_state_set(ext, OES_ACTIVE);
490 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
492 atomic_inc(&ext->oe_users);
493 list_del_init(&ext->oe_link);
494 return osc_extent_get(ext);
497 static void __osc_extent_remove(struct osc_extent *ext)
499 assert_osc_object_is_locked(ext->oe_obj);
500 LASSERT(list_empty(&ext->oe_pages));
501 osc_extent_erase(ext);
502 list_del_init(&ext->oe_link);
503 osc_extent_state_set(ext, OES_INV);
504 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
507 static void osc_extent_remove(struct osc_extent *ext)
509 struct osc_object *obj = ext->oe_obj;
511 osc_object_lock(obj);
512 __osc_extent_remove(ext);
513 osc_object_unlock(obj);
517 * This function is used to merge extents to get better performance. It checks
518 * if @cur and @victim are contiguous at block level.
520 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
521 struct osc_extent *victim)
523 struct osc_object *obj = cur->oe_obj;
524 struct client_obd *cli = osc_cli(obj);
529 LASSERT(cur->oe_state == OES_CACHE);
530 assert_osc_object_is_locked(obj);
534 if (victim->oe_state != OES_INV &&
535 (victim->oe_state != OES_CACHE || victim->oe_fsync_wait))
538 if (cur->oe_max_end != victim->oe_max_end)
542 * In the rare case max_pages_per_rpc (mppr) is changed, don't
543 * merge extents until after old ones have been sent, or the
544 * "extents are aligned to RPCs" checks are unhappy.
546 if (cur->oe_mppr != victim->oe_mppr)
549 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
550 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
551 chunk_start = cur->oe_start >> ppc_bits;
552 chunk_end = cur->oe_end >> ppc_bits;
553 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
554 chunk_end + 1 != victim->oe_start >> ppc_bits)
557 /* overall extent size should not exceed the max supported limit
558 * reported by the server */
559 if (cur->oe_end - cur->oe_start + 1 +
560 victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
563 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
565 cur->oe_start = min(cur->oe_start, victim->oe_start);
566 cur->oe_end = max(cur->oe_end, victim->oe_end);
567 /* per-extent tax should be accounted only once for the whole extent */
568 cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax;
569 cur->oe_nr_pages += victim->oe_nr_pages;
570 /* only the following bits are needed to merge */
571 cur->oe_urgent |= victim->oe_urgent;
572 cur->oe_memalloc |= victim->oe_memalloc;
573 list_splice_init(&victim->oe_pages, &cur->oe_pages);
574 victim->oe_nr_pages = 0;
576 osc_extent_get(victim);
577 __osc_extent_remove(victim);
578 osc_extent_put(env, victim);
580 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
585 * Drop user count of osc_extent, and unplug IO asynchronously.
587 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
589 struct osc_object *obj = ext->oe_obj;
590 struct client_obd *cli = osc_cli(obj);
593 LASSERT(atomic_read(&ext->oe_users) > 0);
594 LASSERT(sanity_check(ext) == 0);
595 LASSERT(ext->oe_grants > 0);
597 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
598 LASSERT(ext->oe_state == OES_ACTIVE);
599 if (ext->oe_trunc_pending) {
600 /* a truncate process is waiting for this extent.
601 * This may happen due to a race, check
602 * osc_cache_truncate_start(). */
603 osc_extent_state_set(ext, OES_TRUNC);
604 ext->oe_trunc_pending = 0;
605 osc_object_unlock(obj);
609 osc_extent_state_set(ext, OES_CACHE);
610 osc_update_pending(obj, OBD_BRW_WRITE,
613 /* try to merge the previous and next extent. */
614 if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
615 grant += cli->cl_grant_extent_tax;
616 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
617 grant += cli->cl_grant_extent_tax;
620 list_move_tail(&ext->oe_link,
622 else if (ext->oe_urgent)
623 list_move_tail(&ext->oe_link,
624 &obj->oo_urgent_exts);
625 else if (ext->oe_nr_pages == ext->oe_mppr) {
626 list_move_tail(&ext->oe_link,
629 osc_object_unlock(obj);
631 osc_unreserve_grant(cli, 0, grant);
634 osc_io_unplug_async(env, cli, obj);
636 osc_extent_put(env, ext);
642 * Find or create an extent which includes @index, core function to manage
645 static struct osc_extent *osc_extent_find(const struct lu_env *env,
646 struct osc_object *obj, pgoff_t index,
647 unsigned int *grants)
649 struct client_obd *cli = osc_cli(obj);
650 struct osc_lock *olck;
651 struct cl_lock_descr *descr;
652 struct osc_extent *cur;
653 struct osc_extent *ext;
654 struct osc_extent *conflict = NULL;
655 struct osc_extent *found = NULL;
658 unsigned int max_pages; /* max_pages_per_rpc */
659 unsigned int chunksize;
660 int ppc_bits; /* pages per chunk bits */
665 cur = osc_extent_alloc(obj);
667 RETURN(ERR_PTR(-ENOMEM));
669 olck = osc_env_io(env)->oi_write_osclock;
670 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
671 LASSERT(olck->ols_state == OLS_GRANTED);
673 descr = &olck->ols_cl.cls_lock->cll_descr;
674 LASSERT(descr->cld_mode >= CLM_WRITE);
676 LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
677 "chunkbits: %u\n", cli->cl_chunkbits);
678 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
679 chunk_mask = ~((1 << ppc_bits) - 1);
680 chunksize = 1 << cli->cl_chunkbits;
681 chunk = index >> ppc_bits;
683 /* align end to RPC edge. */
684 max_pages = cli->cl_max_pages_per_rpc;
685 if ((max_pages & ~chunk_mask) != 0) {
686 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
687 max_pages, cli->cl_chunkbits, chunk_mask);
688 RETURN(ERR_PTR(-EINVAL));
690 max_end = index - (index % max_pages) + max_pages - 1;
691 max_end = min_t(pgoff_t, max_end, descr->cld_end);
693 /* initialize new extent by parameters so far */
694 cur->oe_max_end = max_end;
695 cur->oe_start = index & chunk_mask;
696 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
697 if (cur->oe_start < descr->cld_start)
698 cur->oe_start = descr->cld_start;
699 if (cur->oe_end > max_end)
700 cur->oe_end = max_end;
701 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
702 cur->oe_mppr = max_pages;
703 if (olck->ols_dlmlock != NULL) {
704 LASSERT(olck->ols_hold);
705 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
706 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
709 /* grants has been allocated by caller */
710 LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
711 "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
712 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
716 osc_object_lock(obj);
717 ext = osc_extent_search(obj, cur->oe_start);
719 ext = first_extent(obj);
720 for (; ext; ext = next_extent(ext)) {
721 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
722 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
724 LASSERT(sanity_check_nolock(ext) == 0);
725 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
728 /* if covering by different locks, no chance to match */
729 if (olck->ols_dlmlock != ext->oe_dlmlock) {
730 EASSERTF(!overlapped(ext, cur), ext,
731 EXTSTR"\n", EXTPARA(cur));
736 /* discontiguous chunks? */
737 if (chunk + 1 < ext_chk_start)
740 /* ok, from now on, ext and cur have these attrs:
741 * 1. covered by the same lock
742 * 2. contiguous at chunk level or overlapping. */
744 if (overlapped(ext, cur)) {
745 /* cur is the minimum unit, so overlapping means
747 EASSERTF((ext->oe_start <= cur->oe_start &&
748 ext->oe_end >= cur->oe_end),
749 ext, EXTSTR"\n", EXTPARA(cur));
751 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
752 /* for simplicity, we wait for this extent to
753 * finish before going forward. */
754 conflict = osc_extent_get(ext);
758 found = osc_extent_hold(ext);
762 /* non-overlapped extent */
763 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
764 /* we can't do anything for a non OES_CACHE extent, or
765 * if there is someone waiting for this extent to be
766 * flushed, try next one. */
769 if (osc_extent_merge(env, ext, cur) == 0) {
770 LASSERT(*grants >= chunksize);
771 *grants -= chunksize;
774 * Try to merge with the next one too because we
775 * might have just filled in a gap.
777 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
778 /* we can save extent tax from next extent */
779 *grants += cli->cl_grant_extent_tax;
781 found = osc_extent_hold(ext);
786 osc_extent_tree_dump(D_CACHE, obj);
788 LASSERT(conflict == NULL);
789 if (!IS_ERR(found)) {
790 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
791 OSC_EXTENT_DUMP(D_CACHE, found,
792 "found caching ext for %lu.\n", index);
794 } else if (conflict == NULL) {
795 /* create a new extent */
796 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
797 LASSERT(*grants >= cur->oe_grants);
798 *grants -= cur->oe_grants;
800 cur->oe_state = OES_CACHE;
801 found = osc_extent_hold(cur);
802 osc_extent_insert(obj, cur);
803 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
804 index, descr->cld_end);
806 osc_object_unlock(obj);
808 if (conflict != NULL) {
809 LASSERT(found == NULL);
811 /* waiting for IO to finish. Please notice that it's impossible
812 * to be an OES_TRUNC extent. */
813 rc = osc_extent_wait(env, conflict, OES_INV);
814 osc_extent_put(env, conflict);
817 GOTO(out, found = ERR_PTR(rc));
824 osc_extent_put(env, cur);
829 * Called when IO is finished to an extent.
831 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
834 struct client_obd *cli = osc_cli(ext->oe_obj);
835 struct osc_object *osc = ext->oe_obj;
836 struct osc_async_page *oap;
837 struct osc_async_page *tmp;
838 int nr_pages = ext->oe_nr_pages;
840 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
845 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
847 ext->oe_rc = rc ?: ext->oe_nr_pages;
848 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
850 /* dio pages do not go in the LRU */
852 osc_lru_add_batch(cli, &ext->oe_pages);
854 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
856 list_del_init(&oap->oap_rpc_item);
857 list_del_init(&oap->oap_pending_item);
858 if (last_off <= oap->oap_obj_off) {
859 last_off = oap->oap_obj_off;
860 last_count = oap->oap_count;
864 osc_ap_completion(env, cli, osc, oap, sent, rc);
866 EASSERT(ext->oe_nr_pages == 0, ext);
869 lost_grant = ext->oe_grants;
870 } else if (cli->cl_ocd_grant_param == 0 &&
871 blocksize < PAGE_SIZE &&
872 last_count != PAGE_SIZE) {
873 /* For short writes without OBD_CONNECT_GRANT support, we
874 * shouldn't count parts of pages that span a whole chunk on
875 * the OST side, or our accounting goes wrong. Should match
876 * the code in tgt_grant_check.
878 int offset = last_off & ~PAGE_MASK;
879 int count = last_count + (offset & (blocksize - 1));
880 int end = (offset + last_count) & (blocksize - 1);
882 count += blocksize - end;
884 lost_grant = PAGE_SIZE - count;
886 if (ext->oe_grants > 0)
887 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
889 osc_extent_remove(ext);
890 /* put the refcount for RPC */
891 osc_extent_put(env, ext);
896 * Wait for the extent's state to become @state.
898 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
899 enum osc_extent_state state)
901 struct osc_object *obj = ext->oe_obj;
905 osc_object_lock(obj);
906 LASSERT(sanity_check_nolock(ext) == 0);
907 /* `Kick' this extent only if the caller is waiting for it to be
909 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
910 if (ext->oe_state == OES_ACTIVE) {
912 } else if (ext->oe_state == OES_CACHE) {
914 osc_extent_hold(ext);
918 osc_object_unlock(obj);
920 osc_extent_release(env, ext);
922 /* wait for the extent until its state becomes @state */
923 rc = wait_event_idle_timeout(ext->oe_waitq,
924 smp_load_acquire(&ext->oe_state) == state,
925 cfs_time_seconds(600));
927 OSC_EXTENT_DUMP(D_ERROR, ext,
928 "%s: wait ext to %u timedout, recovery in progress?\n",
929 cli_name(osc_cli(obj)), state);
931 wait_event_idle(ext->oe_waitq,
932 smp_load_acquire(&ext->oe_state) == state);
942 * Discard pages with index greater than @size. If @ext is overlapped with
943 * @size, then partial truncate happens.
945 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
950 struct osc_object *obj = ext->oe_obj;
951 struct client_obd *cli = osc_cli(obj);
952 struct osc_async_page *oap;
953 struct osc_async_page *tmp;
954 struct folio_batch *fbatch;
955 int pages_in_chunk = 0;
956 int ppc_bits = cli->cl_chunkbits -
958 __u64 trunc_chunk = trunc_index >> ppc_bits;
965 LASSERT(sanity_check(ext) == 0);
966 LASSERT(ext->oe_state == OES_TRUNC);
967 LASSERT(!ext->oe_urgent);
969 /* Request new lu_env.
970 * We can't use that env from osc_cache_truncate_start() because
971 * it's from lov_io_sub and not fully initialized. */
972 env = cl_env_get(&refcheck);
974 RETURN(PTR_ERR(env));
976 io = osc_env_thread_io(env);
977 io->ci_obj = cl_object_top(osc2cl(obj));
978 io->ci_ignore_layout = 1;
979 fbatch = &osc_env_info(env)->oti_fbatch;
980 ll_folio_batch_init(fbatch, 0);
981 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
985 /* discard all pages with index greater than trunc_index */
986 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
988 pgoff_t index = osc_index(oap2osc(oap));
989 struct cl_page *page = oap2cl_page(oap);
991 LASSERT(list_empty(&oap->oap_rpc_item));
993 /* only discard the pages with their index greater than
994 * trunc_index, and ... */
995 if (index < trunc_index ||
996 (index == trunc_index && partial)) {
997 /* accounting how many pages remaining in the chunk
998 * so that we can calculate grants correctly. */
999 if (index >> ppc_bits == trunc_chunk)
1004 list_del_init(&oap->oap_pending_item);
1007 lu_ref_add(&page->cp_reference, "truncate", current);
1009 if (cl_page_own(env, io, page) == 0) {
1010 cl_page_discard(env, io, page);
1011 cl_page_disown(env, io, page);
1013 LASSERT(page->cp_state == CPS_FREEING);
1017 lu_ref_del(&page->cp_reference, "truncate", current);
1018 cl_batch_put(env, page, fbatch);
1023 folio_batch_release(fbatch);
1025 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1026 ext->oe_nr_pages == 0),
1027 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1029 osc_object_lock(obj);
1030 if (ext->oe_nr_pages == 0) {
1031 LASSERT(pages_in_chunk == 0);
1032 grants = ext->oe_grants;
1034 } else { /* calculate how many grants we can free */
1035 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1039 /* if there is no pages in this chunk, we can also free grants
1040 * for the last chunk */
1041 if (pages_in_chunk == 0) {
1042 /* if this is the 1st chunk and no pages in this chunk,
1043 * ext->oe_nr_pages must be zero, so we should be in
1044 * the other if-clause. */
1045 LASSERT(trunc_chunk > 0);
1050 /* this is what we can free from this extent */
1051 grants = chunks << cli->cl_chunkbits;
1052 ext->oe_grants -= grants;
1053 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1054 ext->oe_end = min(last_index, ext->oe_max_end);
1055 LASSERT(ext->oe_end >= ext->oe_start);
1056 LASSERT(ext->oe_grants > 0);
1058 osc_object_unlock(obj);
1060 if (grants > 0 || nr_pages > 0)
1061 osc_free_grant(cli, nr_pages, grants, grants);
1064 cl_io_fini(env, io);
1065 cl_env_put(env, &refcheck);
1070 * This function is used to make the extent prepared for transfer.
1071 * A race with flusing page - ll_writepage() has to be handled cautiously.
1073 static int osc_extent_make_ready(const struct lu_env *env,
1074 struct osc_extent *ext)
1076 struct osc_async_page *oap;
1077 struct osc_async_page *last = NULL;
1078 struct osc_object *obj = ext->oe_obj;
1079 unsigned int page_count = 0;
1083 /* we're going to grab page lock, so object lock must not be taken. */
1084 LASSERT(sanity_check(ext) == 0);
1085 /* in locking state, any process should not touch this extent. */
1086 EASSERT(ext->oe_state == OES_LOCKING, ext);
1087 EASSERT(ext->oe_owner != NULL, ext);
1089 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1091 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1093 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1096 /* checking ASYNC_READY is race safe */
1097 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1100 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1103 oap->oap_async_flags |= ASYNC_READY;
1106 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1109 LASSERTF(0, "unknown return code: %d\n", rc);
1113 LASSERT(page_count == ext->oe_nr_pages);
1114 LASSERT(last != NULL);
1115 /* the last page is the only one we need to refresh its count by
1116 * the size of file. */
1117 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1118 int last_oap_count = osc_refresh_count(env, obj, last,
1120 LASSERTF(last_oap_count > 0,
1121 "last_oap_count %d\n", last_oap_count);
1122 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1123 last->oap_count = last_oap_count;
1124 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1127 /* for the rest of pages, we don't need to call osf_refresh_count()
1128 * because it's known they are not the last page */
1129 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1130 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1131 oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1132 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
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 cl_page *page = oap2cl_page(oap);
1259 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1263 result = cl_page_make_ready(env, page, CRT_WRITE);
1268 static int osc_refresh_count(const struct lu_env *env, struct osc_object *osc,
1269 struct osc_async_page *oap, int cmd)
1271 struct osc_page *opg = oap2osc_page(oap);
1272 pgoff_t index = osc_index(oap2osc(oap));
1273 struct cl_object *obj = osc2cl(osc);
1274 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1278 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1279 LASSERT(!(cmd & OBD_BRW_READ));
1280 LASSERT(opg != NULL);
1282 cl_object_attr_lock(obj);
1283 result = cl_object_attr_get(env, obj, attr);
1284 cl_object_attr_unlock(obj);
1287 kms = attr->cat_kms;
1288 if (index << PAGE_SHIFT >= kms)
1289 /* catch race with truncate */
1291 else if ((index + 1) << PAGE_SHIFT > kms)
1292 /* catch sub-page write at end of file */
1293 return kms & ~PAGE_MASK;
1298 static int osc_completion(const struct lu_env *env, struct osc_object *osc,
1299 struct osc_async_page *oap, int cmd, int rc)
1301 struct osc_page *opg = oap2osc_page(oap);
1302 struct cl_page *page = oap2cl_page(oap);
1303 enum cl_req_type crt;
1308 cmd &= ~OBD_BRW_NOQUOTA;
1309 if (page->cp_type != CPT_TRANSIENT) {
1310 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1311 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1312 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1313 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1314 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 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1324 if (rc == 0 && srvlock) {
1325 struct lu_device *ld = osc->oo_cl.co_lu.lo_dev;
1326 struct osc_stats *stats = &lu2osc_dev(ld)->osc_stats;
1327 size_t bytes = oap->oap_count;
1329 if (crt == CRT_READ)
1330 stats->os_lockless_reads += bytes;
1332 stats->os_lockless_writes += bytes;
1336 * This has to be the last operation with the page, as locks are
1337 * released in cl_page_completion() and nothing except for the
1338 * reference counter protects page from concurrent reclaim.
1340 lu_ref_del(&page->cp_reference, "transfer", page);
1342 cl_page_completion(env, page, crt, rc);
1343 cl_page_put(env, page);
1348 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \
1349 struct client_obd *__tmp = (cli); \
1350 CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1351 "dropped: %ld avail: %ld, dirty_grant: %ld, " \
1352 "reserved: %ld, flight: %d } lru {in list: %ld, " \
1353 "left: %ld, waiters: %d }" fmt "\n", \
1355 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1356 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1357 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1358 __tmp->cl_dirty_grant, \
1359 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1360 atomic_long_read(&__tmp->cl_lru_in_list), \
1361 atomic_long_read(&__tmp->cl_lru_busy), \
1362 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1365 /* caller must hold loi_list_lock */
1366 static void osc_consume_write_grant(struct client_obd *cli,
1367 struct brw_page *pga)
1369 assert_spin_locked(&cli->cl_loi_list_lock);
1370 LASSERT(!(pga->bp_flag & OBD_BRW_FROM_GRANT));
1371 cli->cl_dirty_pages++;
1372 pga->bp_flag |= OBD_BRW_FROM_GRANT;
1373 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1374 PAGE_SIZE, pga, pga->bp_page);
1377 /* the companion to osc_consume_write_grant, called when a brw has completed.
1378 * must be called with the loi lock held. */
1379 static void osc_release_write_grant(struct client_obd *cli,
1380 struct brw_page *pga)
1384 assert_spin_locked(&cli->cl_loi_list_lock);
1385 if (!(pga->bp_flag & OBD_BRW_FROM_GRANT)) {
1390 pga->bp_flag &= ~OBD_BRW_FROM_GRANT;
1391 atomic_long_dec(&obd_dirty_pages);
1392 cli->cl_dirty_pages--;
1397 * To avoid sleeping with object lock held, it's good for us allocate enough
1398 * grants before entering into critical section.
1400 * client_obd_list_lock held by caller
1402 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1406 if (cli->cl_avail_grant >= bytes) {
1407 cli->cl_avail_grant -= bytes;
1408 cli->cl_reserved_grant += bytes;
1414 static void __osc_unreserve_grant(struct client_obd *cli,
1415 unsigned int reserved, unsigned int unused)
1417 /* it's quite normal for us to get more grant than reserved.
1418 * Thinking about a case that two extents merged by adding a new
1419 * chunk, we can save one extent tax. If extent tax is greater than
1420 * one chunk, we can save more grant by adding a new chunk */
1421 cli->cl_reserved_grant -= reserved;
1422 if (unused > reserved) {
1423 cli->cl_avail_grant += reserved;
1424 cli->cl_lost_grant += unused - reserved;
1425 cli->cl_dirty_grant -= unused - reserved;
1427 cli->cl_avail_grant += unused;
1428 cli->cl_dirty_grant += reserved - unused;
1432 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1433 unsigned int reserved,
1434 unsigned int unused)
1436 __osc_unreserve_grant(cli, reserved, unused);
1438 osc_wake_cache_waiters(cli);
1441 static void osc_unreserve_grant(struct client_obd *cli,
1442 unsigned int reserved, unsigned int unused)
1444 spin_lock(&cli->cl_loi_list_lock);
1445 osc_unreserve_grant_nolock(cli, reserved, unused);
1446 spin_unlock(&cli->cl_loi_list_lock);
1450 * Free grant after IO is finished or canceled.
1452 * @lost_grant is used to remember how many grants we have allocated but not
1453 * used, we should return these grants to OST. There're two cases where grants
1456 * 2. Without OBD_CONNECT_GRANT support and blocksize at OST is less than
1457 * PAGE_SIZE and a partial page was written. In this case OST may use less
1458 * chunks to serve this partial write. OSTs don't actually know the page
1459 * size on the client side. so clients have to calculate lost grant by the
1460 * blocksize on the OST. See tgt_grant_check() for details.
1462 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1463 unsigned int lost_grant, unsigned int dirty_grant)
1465 unsigned long grant;
1467 grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1469 spin_lock(&cli->cl_loi_list_lock);
1470 atomic_long_sub(nr_pages, &obd_dirty_pages);
1471 cli->cl_dirty_pages -= nr_pages;
1472 cli->cl_lost_grant += lost_grant;
1473 cli->cl_dirty_grant -= dirty_grant;
1474 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1475 /* borrow some grant from truncate to avoid the case that
1476 * truncate uses up all avail grant */
1477 cli->cl_lost_grant -= grant;
1478 cli->cl_avail_grant += grant;
1480 osc_wake_cache_waiters(cli);
1481 spin_unlock(&cli->cl_loi_list_lock);
1482 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1483 lost_grant, cli->cl_lost_grant,
1484 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1485 cli->cl_dirty_grant);
1489 * The companion to osc_enter_cache(), called when @oap is no longer part of
1490 * the dirty accounting due to error.
1492 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1494 spin_lock(&cli->cl_loi_list_lock);
1495 osc_release_write_grant(cli, &oap->oap_brw_page);
1496 spin_unlock(&cli->cl_loi_list_lock);
1500 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1503 static int osc_enter_cache_try(struct client_obd *cli,
1504 struct osc_async_page *oap,
1509 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1511 rc = osc_reserve_grant(cli, bytes);
1515 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1516 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1517 obd_max_dirty_pages) {
1518 osc_consume_write_grant(cli, &oap->oap_brw_page);
1522 atomic_long_dec(&obd_dirty_pages);
1524 __osc_unreserve_grant(cli, bytes, bytes);
1530 /* Following two inlines exist to pass code fragments
1531 * to wait_event_idle_exclusive_timeout_cmd(). Passing
1532 * code fragments as macro args can look confusing, so
1533 * we provide inlines to encapsulate them.
1535 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1536 struct client_obd *cli,
1537 struct osc_async_page *oap)
1539 spin_unlock(&cli->cl_loi_list_lock);
1540 osc_io_unplug_async(env, cli, NULL);
1542 "%s: sleeping for cache space for %p\n",
1543 cli_name(cli), oap);
1546 static inline void cli_lock_after_unplug(struct client_obd *cli)
1548 spin_lock(&cli->cl_loi_list_lock);
1551 * The main entry to reserve dirty page accounting. Usually the grant reserved
1552 * in this function will be freed in bulk in osc_free_grant() unless it fails
1553 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1555 * The process will be put into sleep if it's already run out of grant.
1557 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1558 struct osc_object *osc, struct osc_async_page *oap,
1561 struct lov_oinfo *loi = osc->oo_oinfo;
1564 bool entered = false;
1565 struct obd_device *obd = cli->cl_import->imp_obd;
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 =
1575 cfs_time_seconds(obd_at_off(obd) ?
1577 obd_get_ldlm_enqueue_min(obd) / 2);
1581 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1583 spin_lock(&cli->cl_loi_list_lock);
1585 /* force the caller to try sync io. this can jump the list
1586 * of queued writes and create a discontiguous rpc stream */
1587 if (CFS_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1588 cli->cl_dirty_max_pages == 0 ||
1589 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1590 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1591 GOTO(out, rc = -EDQUOT);
1595 * We can wait here for two reasons: too many dirty pages in cache, or
1596 * run out of grants. In both cases we should write dirty pages out.
1597 * Adding a cache waiter will trigger urgent write-out no matter what
1599 * The exiting condition (other than success) is no avail grants
1600 * and no dirty pages caching, that really means there is no space
1603 remain = wait_event_idle_exclusive_timeout_cmd(
1604 cli->cl_cache_waiters,
1605 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1606 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1608 cli_unlock_and_unplug(env, cli, oap),
1609 cli_lock_after_unplug(cli));
1612 if (remain == timeout)
1613 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1615 OSC_DUMP_GRANT(D_CACHE, cli,
1616 "finally got grant space\n");
1617 wake_up(&cli->cl_cache_waiters);
1619 } else if (remain == 0) {
1620 OSC_DUMP_GRANT(D_CACHE, cli,
1621 "timeout, fall back to sync i/o\n");
1622 osc_extent_tree_dump(D_CACHE, osc);
1623 /* fall back to synchronous I/O */
1625 OSC_DUMP_GRANT(D_CACHE, cli,
1626 "no grant space, fall back to sync i/o\n");
1627 wake_up_all(&cli->cl_cache_waiters);
1631 spin_unlock(&cli->cl_loi_list_lock);
1635 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1637 int hprpc = !!list_empty(&osc->oo_hp_exts);
1638 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1641 /* This maintains the lists of pending pages to read/write for a given object
1642 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1643 * to quickly find objects that are ready to send an RPC. */
1644 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1647 int invalid_import = 0;
1650 /* if we have an invalid import we want to drain the queued pages
1651 * by forcing them through rpcs that immediately fail and complete
1652 * the pages. recovery relies on this to empty the queued pages
1653 * before canceling the locks and evicting down the llite pages */
1654 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1657 if (cmd & OBD_BRW_WRITE) {
1658 if (atomic_read(&osc->oo_nr_writes) == 0)
1660 if (invalid_import) {
1661 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1664 if (!list_empty(&osc->oo_hp_exts)) {
1665 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1668 if (!list_empty(&osc->oo_urgent_exts)) {
1669 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1672 /* trigger a write rpc stream as long as there are dirtiers
1673 * waiting for space. as they're waiting, they're not going to
1674 * create more pages to coalesce with what's waiting..
1676 if (waitqueue_active(&cli->cl_cache_waiters)) {
1677 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1680 if (!list_empty(&osc->oo_full_exts)) {
1681 CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1685 if (atomic_read(&osc->oo_nr_reads) == 0)
1687 if (invalid_import) {
1688 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1691 /* all read are urgent. */
1692 if (!list_empty(&osc->oo_reading_exts))
1699 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1701 struct client_obd *cli = osc_cli(obj);
1702 if (cmd & OBD_BRW_WRITE) {
1703 atomic_add(delta, &obj->oo_nr_writes);
1704 atomic_add(delta, &cli->cl_pending_w_pages);
1705 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1707 atomic_add(delta, &obj->oo_nr_reads);
1708 atomic_add(delta, &cli->cl_pending_r_pages);
1709 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1711 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1714 static int osc_makes_hprpc(struct osc_object *obj)
1716 return !list_empty(&obj->oo_hp_exts);
1719 static void on_list(struct list_head *item, struct list_head *list,
1722 if (list_empty(item) && should_be_on)
1723 list_add_tail(item, list);
1724 else if (!list_empty(item) && !should_be_on)
1725 list_del_init(item);
1728 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1729 * can find pages to build into rpcs quickly */
1730 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1732 if (osc_makes_hprpc(osc)) {
1734 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1735 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1737 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1738 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1739 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1740 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1743 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1744 atomic_read(&osc->oo_nr_writes) > 0);
1746 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1747 atomic_read(&osc->oo_nr_reads) > 0);
1749 return osc_is_ready(osc);
1752 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1756 spin_lock(&cli->cl_loi_list_lock);
1757 is_ready = __osc_list_maint(cli, osc);
1758 spin_unlock(&cli->cl_loi_list_lock);
1763 /* this must be called holding the loi list lock to give coverage to exit_cache,
1764 * async_flag maintenance
1766 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1767 struct osc_object *osc,
1768 struct osc_async_page *oap, int sent, int rc)
1772 /* As the transfer for this page is being done, clear the flags */
1773 oap->oap_async_flags = 0;
1775 rc = osc_completion(env, osc, oap, oap->oap_cmd, rc);
1777 CERROR("completion on oap %p obj %p returns %d.\n",
1783 struct extent_rpc_data {
1784 struct list_head *erd_rpc_list;
1785 unsigned int erd_page_count;
1786 unsigned int erd_max_pages;
1787 unsigned int erd_max_chunks;
1788 unsigned int erd_max_extents;
1791 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1793 struct client_obd *cli = osc_cli(ext->oe_obj);
1794 unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1796 return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1800 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1802 if (ext->oe_no_merge || in_rpc->oe_no_merge)
1805 if (ext->oe_srvlock != in_rpc->oe_srvlock)
1808 if (ext->oe_ndelay != in_rpc->oe_ndelay)
1811 if (!ext->oe_grants != !in_rpc->oe_grants)
1814 if (ext->oe_dio != in_rpc->oe_dio)
1817 /* It's possible to have overlap on DIO */
1818 if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1821 if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1828 * Try to add extent to one RPC. We need to think about the following things:
1829 * - # of pages must not be over max_pages_per_rpc
1830 * - extent must be compatible with previous ones
1832 static int try_to_add_extent_for_io(struct client_obd *cli,
1833 struct osc_extent *ext,
1834 struct extent_rpc_data *data)
1836 struct osc_extent *tmp;
1837 unsigned int chunk_count;
1840 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1842 OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1844 if (data->erd_max_extents == 0)
1847 chunk_count = osc_extent_chunks(ext);
1848 EASSERTF(data->erd_page_count != 0 ||
1849 chunk_count <= data->erd_max_chunks, ext,
1850 "The first extent to be fit in a RPC contains %u chunks, "
1851 "which is over the limit %u.\n", chunk_count,
1852 data->erd_max_chunks);
1853 if (chunk_count > data->erd_max_chunks)
1856 data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1857 EASSERTF(data->erd_page_count != 0 ||
1858 ext->oe_nr_pages <= data->erd_max_pages, ext,
1859 "The first extent to be fit in a RPC contains %u pages, "
1860 "which is over the limit %u.\n", ext->oe_nr_pages,
1861 data->erd_max_pages);
1862 if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1865 list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1866 EASSERT(tmp->oe_owner == current, tmp);
1868 if (!can_merge(ext, tmp))
1872 data->erd_max_extents--;
1873 data->erd_max_chunks -= chunk_count;
1874 data->erd_page_count += ext->oe_nr_pages;
1875 list_move_tail(&ext->oe_link, data->erd_rpc_list);
1876 ext->oe_owner = current;
1881 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1882 * get_write_extent() takes all appropriate extents in atomic.
1884 * The following policy is used to collect extents for IO:
1885 * 1. Add as many HP extents as possible;
1886 * 2. Add the first urgent extent in urgent extent list and take it out of
1888 * 3. Add subsequent extents of this urgent extent;
1889 * 4. If urgent list is not empty, goto 2;
1890 * 5. Traverse the extent tree from the 1st extent;
1891 * 6. Above steps exit if there is no space in this RPC.
1893 static unsigned int get_write_extents(struct osc_object *obj,
1894 struct list_head *rpclist)
1896 struct client_obd *cli = osc_cli(obj);
1897 struct osc_extent *ext;
1898 struct extent_rpc_data data = {
1899 .erd_rpc_list = rpclist,
1900 .erd_page_count = 0,
1901 .erd_max_pages = cli->cl_max_pages_per_rpc,
1902 .erd_max_chunks = osc_max_write_chunks(cli),
1903 .erd_max_extents = 256,
1906 assert_osc_object_is_locked(obj);
1907 while ((ext = list_first_entry_or_null(&obj->oo_hp_exts,
1909 oe_link)) != NULL) {
1910 if (!try_to_add_extent_for_io(cli, ext, &data))
1911 return data.erd_page_count;
1912 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1914 if (data.erd_page_count == data.erd_max_pages)
1915 return data.erd_page_count;
1917 while ((ext = list_first_entry_or_null(&obj->oo_urgent_exts,
1919 oe_link)) != NULL) {
1920 if (!try_to_add_extent_for_io(cli, ext, &data))
1921 return data.erd_page_count;
1923 if (data.erd_page_count == data.erd_max_pages)
1924 return data.erd_page_count;
1926 /* One key difference between full extents and other extents: full
1927 * extents can usually only be added if the rpclist was empty, so if we
1928 * can't add one, we continue on to trying to add normal extents. This
1929 * is so we don't miss adding extra extents to an RPC containing high
1930 * priority or urgent extents.
1932 while ((ext = list_first_entry_or_null(&obj->oo_full_exts,
1934 oe_link)) != NULL) {
1935 if (!try_to_add_extent_for_io(cli, ext, &data))
1938 if (data.erd_page_count == data.erd_max_pages)
1939 return data.erd_page_count;
1941 for (ext = first_extent(obj);
1943 ext = next_extent(ext)) {
1944 if ((ext->oe_state != OES_CACHE) ||
1945 /* this extent may be already in current rpclist */
1946 (!list_empty(&ext->oe_link) && ext->oe_owner))
1949 if (!try_to_add_extent_for_io(cli, ext, &data))
1950 return data.erd_page_count;
1952 return data.erd_page_count;
1956 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1957 struct osc_object *osc)
1961 struct osc_extent *ext;
1962 struct osc_extent *tmp;
1963 struct osc_extent *first = NULL;
1964 unsigned int page_count = 0;
1969 assert_osc_object_is_locked(osc);
1971 page_count = get_write_extents(osc, &rpclist);
1972 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1974 if (list_empty(&rpclist))
1977 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1979 list_for_each_entry(ext, &rpclist, oe_link) {
1980 LASSERT(ext->oe_state == OES_CACHE ||
1981 ext->oe_state == OES_LOCK_DONE);
1982 if (ext->oe_state == OES_CACHE)
1983 osc_extent_state_set(ext, OES_LOCKING);
1985 osc_extent_state_set(ext, OES_RPC);
1988 /* we're going to grab page lock, so release object lock because
1989 * lock order is page lock -> object lock. */
1990 osc_object_unlock(osc);
1992 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
1993 if (ext->oe_state == OES_LOCKING) {
1994 rc = osc_extent_make_ready(env, ext);
1995 if (unlikely(rc < 0)) {
1996 list_del_init(&ext->oe_link);
1997 osc_extent_finish(env, ext, 0, rc);
2001 if (first == NULL) {
2003 srvlock = ext->oe_srvlock;
2005 LASSERT(srvlock == ext->oe_srvlock);
2009 if (!list_empty(&rpclist)) {
2010 LASSERT(page_count > 0);
2011 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2012 LASSERT(list_empty(&rpclist));
2015 osc_object_lock(osc);
2020 * prepare pages for ASYNC io and put pages in send queue.
2022 * \param cmd OBD_BRW_* macroses
2023 * \param lop pending pages
2025 * \return zero if no page added to send queue.
2026 * \return 1 if pages successfully added to send queue.
2027 * \return negative on errors.
2030 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2031 struct osc_object *osc)
2034 struct osc_extent *ext;
2035 struct osc_extent *next;
2037 struct extent_rpc_data data = {
2038 .erd_rpc_list = &rpclist,
2039 .erd_page_count = 0,
2040 .erd_max_pages = cli->cl_max_pages_per_rpc,
2041 .erd_max_chunks = UINT_MAX,
2042 .erd_max_extents = UINT_MAX,
2047 assert_osc_object_is_locked(osc);
2048 list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2049 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2050 if (!try_to_add_extent_for_io(cli, ext, &data))
2052 osc_extent_state_set(ext, OES_RPC);
2053 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2055 LASSERT(data.erd_page_count <= data.erd_max_pages);
2057 osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2059 if (!list_empty(&rpclist)) {
2060 osc_object_unlock(osc);
2062 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2063 LASSERT(list_empty(&rpclist));
2065 osc_object_lock(osc);
2070 #define list_to_obj(list, item) ({ \
2071 struct list_head *__tmp = (list)->next; \
2072 list_del_init(__tmp); \
2073 list_entry(__tmp, struct osc_object, oo_##item); \
2076 /* This is called by osc_check_rpcs() to find which objects have pages that
2077 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2078 static struct osc_object *osc_next_obj(struct client_obd *cli)
2082 /* First return objects that have blocked locks so that they
2083 * will be flushed quickly and other clients can get the lock,
2084 * then objects which have pages ready to be stuffed into RPCs */
2085 if (!list_empty(&cli->cl_loi_hp_ready_list))
2086 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2087 if (!list_empty(&cli->cl_loi_ready_list))
2088 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2090 /* then if we have cache waiters, return all objects with queued
2091 * writes. This is especially important when many small files
2092 * have filled up the cache and not been fired into rpcs because
2093 * they don't pass the nr_pending/object threshhold
2095 if (waitqueue_active(&cli->cl_cache_waiters) &&
2096 !list_empty(&cli->cl_loi_write_list))
2097 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2099 /* then return all queued objects when we have an invalid import
2100 * so that they get flushed */
2101 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2102 if (!list_empty(&cli->cl_loi_write_list))
2103 RETURN(list_to_obj(&cli->cl_loi_write_list,
2105 if (!list_empty(&cli->cl_loi_read_list))
2106 RETURN(list_to_obj(&cli->cl_loi_read_list,
2112 /* called with the loi list lock held */
2113 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2114 __must_hold(&cli->cl_loi_list_lock)
2116 struct osc_object *osc;
2120 while ((osc = osc_next_obj(cli)) != NULL) {
2121 struct cl_object *obj = osc2cl(osc);
2122 struct lu_ref_link link;
2124 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2126 /* even if we have reached our max in flight RPCs, we still
2127 * allow all high-priority RPCs through to prevent their
2128 * starvation and leading to server evicting us for not
2129 * writing out pages in a timely manner LU-13131 */
2130 if (osc_max_rpc_in_flight(cli, osc) &&
2131 list_empty(&osc->oo_hp_exts)) {
2132 __osc_list_maint(cli, osc);
2137 spin_unlock(&cli->cl_loi_list_lock);
2138 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2140 /* attempt some read/write balancing by alternating between
2141 * reads and writes in an object. The makes_rpc checks here
2142 * would be redundant if we were getting read/write work items
2143 * instead of objects. we don't want send_oap_rpc to drain a
2144 * partial read pending queue when we're given this object to
2145 * do io on writes while there are cache waiters */
2146 osc_object_lock(osc);
2147 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2148 rc = osc_send_write_rpc(env, cli, osc);
2150 CERROR("Write request failed with %d\n", rc);
2152 /* osc_send_write_rpc failed, mostly because of
2155 * It can't break here, because if:
2156 * - a page was submitted by osc_io_submit, so
2158 * - no request in flight
2159 * - no subsequent request
2160 * The system will be in live-lock state,
2161 * because there is no chance to call
2162 * osc_io_unplug() and osc_check_rpcs() any
2163 * more. pdflush can't help in this case,
2164 * because it might be blocked at grabbing
2165 * the page lock as we mentioned.
2167 * Anyway, continue to drain pages. */
2171 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2172 rc = osc_send_read_rpc(env, cli, osc);
2174 CERROR("Read request failed with %d\n", rc);
2176 osc_object_unlock(osc);
2178 osc_list_maint(cli, osc);
2179 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2180 cl_object_put(env, obj);
2182 spin_lock(&cli->cl_loi_list_lock);
2187 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2188 struct osc_object *osc, int async)
2192 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2196 spin_lock(&cli->cl_loi_list_lock);
2197 osc_check_rpcs(env, cli);
2198 spin_unlock(&cli->cl_loi_list_lock);
2200 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2201 LASSERT(cli->cl_writeback_work != NULL);
2202 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2206 EXPORT_SYMBOL(osc_io_unplug0);
2208 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2209 struct cl_page *page, loff_t offset)
2211 struct osc_async_page *oap = &ops->ops_oap;
2215 return round_up(sizeof(*oap), 8);
2218 oap->oap_page = page->cp_vmpage;
2219 oap->oap_obj_off = offset;
2220 LASSERT(!(offset & ~PAGE_MASK));
2222 /* Count of transient (direct i/o) pages is always stable by the time
2223 * they're submitted. Setting this here lets us avoid calling
2224 * cl_page_clip later to set this.
2226 if (page->cp_type == CPT_TRANSIENT)
2227 oap->oap_async_flags |= ASYNC_COUNT_STABLE|ASYNC_URGENT|
2230 INIT_LIST_HEAD(&oap->oap_pending_item);
2231 INIT_LIST_HEAD(&oap->oap_rpc_item);
2233 CDEBUG(D_INFO, "oap %p vmpage %p obj off %llu\n",
2234 oap, oap->oap_page, oap->oap_obj_off);
2237 EXPORT_SYMBOL(osc_prep_async_page);
2239 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2240 struct osc_object *osc, struct osc_page *ops,
2243 struct osc_io *oio = osc_env_io(env);
2244 struct osc_extent *ext = NULL;
2245 struct osc_async_page *oap = &ops->ops_oap;
2246 struct client_obd *cli = osc_cli(osc);
2247 struct folio_batch *fbatch = &osc_env_info(env)->oti_fbatch;
2250 unsigned int grants = 0;
2251 u32 brw_flags = OBD_BRW_ASYNC;
2252 int cmd = OBD_BRW_WRITE;
2253 int need_release = 0;
2257 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2260 if (!list_empty(&oap->oap_pending_item) ||
2261 !list_empty(&oap->oap_rpc_item))
2264 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2265 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2266 if (io->ci_noquota) {
2267 brw_flags |= OBD_BRW_NOQUOTA;
2268 cmd |= OBD_BRW_NOQUOTA;
2271 if (oio->oi_cap_sys_resource) {
2272 brw_flags |= OBD_BRW_SYS_RESOURCE;
2273 cmd |= OBD_BRW_SYS_RESOURCE;
2276 /* check if the file's owner/group is over quota */
2277 /* do not check for root without root squash, because in this case
2278 * we should bypass quota
2280 if ((!oio->oi_cap_sys_resource ||
2281 cli->cl_root_squash || cli->cl_root_prjquota) &&
2283 struct cl_object *obj;
2284 struct cl_attr *attr;
2285 unsigned int qid[LL_MAXQUOTAS];
2287 obj = cl_object_top(&osc->oo_cl);
2288 attr = &osc_env_info(env)->oti_attr;
2290 cl_object_attr_lock(obj);
2291 rc = cl_object_attr_get(env, obj, attr);
2292 cl_object_attr_unlock(obj);
2294 qid[USRQUOTA] = attr->cat_uid;
2295 qid[GRPQUOTA] = attr->cat_gid;
2296 qid[PRJQUOTA] = attr->cat_projid;
2297 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2304 oap->oap_page_off = ops->ops_from;
2305 oap->oap_count = ops->ops_to - ops->ops_from + 1;
2306 /* No need to hold a lock here,
2307 * since this page is not in any list yet. */
2308 oap->oap_async_flags = 0;
2309 oap->oap_brw_flags = brw_flags;
2311 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2312 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2314 index = osc_index(oap2osc(oap));
2316 /* Add this page into extent by the following steps:
2317 * 1. if there exists an active extent for this IO, mostly this page
2318 * can be added to the active extent and sometimes we need to
2319 * expand extent to accomodate this page;
2320 * 2. otherwise, a new extent will be allocated. */
2322 ext = oio->oi_active;
2323 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2324 /* one chunk plus extent overhead must be enough to write this
2326 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2327 if (ext->oe_end >= index)
2330 /* it doesn't need any grant to dirty this page */
2331 spin_lock(&cli->cl_loi_list_lock);
2332 rc = osc_enter_cache_try(cli, oap, grants);
2333 if (rc == 0) { /* try failed */
2336 } else if (ext->oe_end < index) {
2338 /* try to expand this extent */
2339 rc = osc_extent_expand(ext, index, &tmp);
2342 /* don't free reserved grant */
2344 OSC_EXTENT_DUMP(D_CACHE, ext,
2345 "expanded for %lu.\n", index);
2346 osc_unreserve_grant_nolock(cli, grants, tmp);
2350 spin_unlock(&cli->cl_loi_list_lock);
2352 } else if (ext != NULL) {
2353 /* index is located outside of active extent */
2357 osc_extent_release(env, ext);
2358 oio->oi_active = NULL;
2363 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2365 /* try to find new extent to cover this page */
2366 LASSERT(oio->oi_active == NULL);
2367 /* we may have allocated grant for this page if we failed
2368 * to expand the previous active extent. */
2369 LASSERT(ergo(grants > 0, grants >= tmp));
2373 /* We must not hold a page lock while we do osc_enter_cache()
2374 * or osc_extent_find(), so we must mark dirty & unlock
2375 * any pages in the write commit folio_batch.
2377 if (folio_batch_count(fbatch)) {
2378 cb(env, io, fbatch);
2379 folio_batch_reinit(fbatch);
2383 rc = osc_enter_cache(env, cli, osc, oap, tmp);
2390 ext = osc_extent_find(env, osc, index, &tmp);
2392 LASSERT(tmp == grants);
2393 osc_exit_cache(cli, oap);
2397 oio->oi_active = ext;
2401 osc_unreserve_grant(cli, grants, tmp);
2404 LASSERT(ergo(rc == 0, ext != NULL));
2406 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2407 ext, "index = %lu.\n", index);
2408 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2410 osc_object_lock(osc);
2411 if (ext->oe_nr_pages == 0)
2412 ext->oe_srvlock = ops->ops_srvlock;
2414 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2416 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2417 osc_object_unlock(osc);
2419 if (!ext->oe_layout_version)
2420 ext->oe_layout_version = io->ci_layout_version;
2426 int osc_teardown_async_page(const struct lu_env *env,
2427 struct osc_object *obj, struct osc_page *ops)
2429 struct osc_async_page *oap = &ops->ops_oap;
2433 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2434 oap, ops, osc_index(oap2osc(oap)));
2436 if (!list_empty(&oap->oap_rpc_item)) {
2437 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2439 } else if (!list_empty(&oap->oap_pending_item)) {
2440 struct osc_extent *ext = NULL;
2442 osc_object_lock(obj);
2443 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2444 osc_object_unlock(obj);
2445 /* only truncated pages are allowed to be taken out.
2446 * See osc_extent_truncate() and osc_cache_truncate_start()
2448 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2449 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2450 osc_index(oap2osc(oap)));
2454 osc_extent_put(env, ext);
2460 * This is called when a page is picked up by kernel to write out.
2462 * We should find out the corresponding extent and add the whole extent
2463 * into urgent list. The extent may be being truncated or used, handle it
2466 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2467 struct osc_page *ops)
2469 struct osc_extent *ext = NULL;
2470 struct osc_object *obj = osc_page_object(ops);
2471 struct cl_page *cp = ops->ops_cl.cpl_page;
2472 pgoff_t index = osc_index(ops);
2473 struct osc_async_page *oap = &ops->ops_oap;
2474 bool unplug = false;
2478 osc_object_lock(obj);
2479 ext = osc_extent_lookup(obj, index);
2481 osc_extent_tree_dump(D_ERROR, obj);
2482 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2485 switch (ext->oe_state) {
2488 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2492 /* If we know this extent is being written out, we should abort
2493 * so that the writer can make this page ready. Otherwise, there
2494 * exists a deadlock problem because other process can wait for
2495 * page writeback bit holding page lock; and meanwhile in
2496 * vvp_page_make_ready(), we need to grab page lock before
2497 * really sending the RPC. */
2499 /* race with truncate, page will be redirtied */
2501 /* The extent is active so we need to abort and let the caller
2502 * re-dirty the page. If we continued on here, and we were the
2503 * one making the extent active, we could deadlock waiting for
2504 * the page writeback to clear but it won't because the extent
2505 * is active and won't be written out. */
2506 GOTO(out, rc = -EAGAIN);
2511 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2515 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2517 if (current->flags & PF_MEMALLOC)
2518 ext->oe_memalloc = 1;
2521 if (ext->oe_state == OES_CACHE) {
2522 OSC_EXTENT_DUMP(D_CACHE, ext,
2523 "flush page %p make it urgent.\n", oap);
2524 if (list_empty(&ext->oe_link))
2525 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2532 osc_object_unlock(obj);
2533 osc_extent_put(env, ext);
2535 osc_io_unplug_async(env, osc_cli(obj), obj);
2539 int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io,
2540 struct osc_object *obj, struct list_head *list,
2543 struct osc_io *oio = osc_env_io(env);
2544 struct client_obd *cli = osc_cli(obj);
2545 struct osc_extent *ext;
2546 struct osc_async_page *oap;
2548 int mppr = cli->cl_max_pages_per_rpc;
2549 bool can_merge = true;
2550 pgoff_t start = CL_PAGE_EOF;
2552 struct osc_lock *oscl;
2555 list_for_each_entry(oap, list, oap_pending_item) {
2556 struct osc_page *opg = oap2osc_page(oap);
2557 pgoff_t index = osc_index(opg);
2564 mppr <<= (page_count > mppr);
2566 if (unlikely(opg->ops_from > 0 ||
2567 opg->ops_to < PAGE_SIZE - 1))
2571 ext = osc_extent_alloc(obj);
2573 struct osc_async_page *tmp;
2575 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2576 list_del_init(&oap->oap_pending_item);
2577 osc_ap_completion(env, cli, obj, oap, 0, -ENOMEM);
2582 ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2584 ext->oe_no_merge = !can_merge;
2586 ext->oe_start = start;
2587 ext->oe_end = ext->oe_max_end = end;
2589 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2590 ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2591 ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2592 oscl = oio->oi_write_osclock ? : oio->oi_read_osclock;
2593 if (oscl && oscl->ols_dlmlock != NULL) {
2594 ext->oe_dlmlock = LDLM_LOCK_GET(oscl->ols_dlmlock);
2595 lu_ref_add(&ext->oe_dlmlock->l_reference, "osc_extent", ext);
2597 if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2601 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2602 grants = cli->cl_grant_extent_tax;
2603 grants += (1 << cli->cl_chunkbits) *
2604 ((page_count + ppc - 1) / ppc);
2606 CDEBUG(D_CACHE, "requesting %d bytes grant\n", grants);
2607 spin_lock(&cli->cl_loi_list_lock);
2608 if (osc_reserve_grant(cli, grants) == 0) {
2609 list_for_each_entry(oap, list, oap_pending_item) {
2610 osc_consume_write_grant(cli,
2611 &oap->oap_brw_page);
2613 atomic_long_add(page_count, &obd_dirty_pages);
2614 osc_unreserve_grant_nolock(cli, grants, 0);
2615 ext->oe_grants = grants;
2617 /* We cannot report ENOSPC correctly if we do parallel
2618 * DIO (async RPC submission), so turn off parallel dio
2619 * if there is not sufficient grant available. This
2620 * makes individual RPCs synchronous.
2622 io->ci_parallel_dio = false;
2624 "not enough grant available, switching to sync for this i/o\n");
2626 spin_unlock(&cli->cl_loi_list_lock);
2627 osc_update_next_shrink(cli);
2630 ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2631 ext->oe_nr_pages = page_count;
2632 ext->oe_mppr = mppr;
2633 list_splice_init(list, &ext->oe_pages);
2634 ext->oe_layout_version = io->ci_layout_version;
2636 osc_object_lock(obj);
2637 /* Reuse the initial refcount for RPC, don't drop it */
2638 osc_extent_state_set(ext, OES_LOCK_DONE);
2639 if (!ext->oe_rw) { /* write */
2640 if (!ext->oe_srvlock && !ext->oe_dio) {
2641 /* The most likely case here is from lack of grants
2642 * so we are either out of quota or out of space.
2643 * Since this means we are holding locks across
2644 * potentially multi-striped IO, we must send out
2645 * everything out instantly to avoid prolonged
2646 * waits resulting in lock eviction (likely since
2647 * the extended wait in osc_cache_enter() did not
2648 * yield any additional grant due to a timeout.
2651 list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2653 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2655 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2657 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2658 osc_update_pending(obj, OBD_BRW_READ, page_count);
2660 osc_object_unlock(obj);
2662 osc_io_unplug_async(env, cli, obj);
2667 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2669 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2670 __u64 size, struct osc_extent **extp)
2672 struct client_obd *cli = osc_cli(obj);
2673 struct osc_extent *ext;
2674 struct osc_extent *waiting = NULL;
2681 /* pages with index greater or equal to index will be truncated. */
2682 index = size >> PAGE_SHIFT;
2683 partial = size > (index << PAGE_SHIFT);
2686 osc_object_lock(obj);
2687 ext = osc_extent_search(obj, index);
2689 ext = first_extent(obj);
2690 else if (ext->oe_end < index)
2691 ext = next_extent(ext);
2692 while (ext != NULL) {
2693 EASSERT(ext->oe_state != OES_TRUNC, ext);
2695 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2696 /* if ext is in urgent state, it means there must exist
2697 * a page already having been flushed by write_page().
2698 * We have to wait for this extent because we can't
2699 * truncate that page. */
2700 OSC_EXTENT_DUMP(D_CACHE, ext,
2701 "waiting for busy extent\n");
2702 waiting = osc_extent_get(ext);
2706 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2708 osc_extent_get(ext);
2709 if (ext->oe_state == OES_ACTIVE) {
2710 /* though we grab inode mutex for write path, but we
2711 * release it before releasing extent(in osc_io_end()),
2712 * so there is a race window that an extent is still
2713 * in OES_ACTIVE when truncate starts. */
2714 LASSERT(!ext->oe_trunc_pending);
2715 ext->oe_trunc_pending = 1;
2717 EASSERT(ext->oe_state == OES_CACHE, ext);
2718 osc_extent_state_set(ext, OES_TRUNC);
2719 osc_update_pending(obj, OBD_BRW_WRITE,
2722 /* This extent could be on the full extents list, that's OK */
2723 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2724 if (!list_empty(&ext->oe_link))
2725 list_move_tail(&ext->oe_link, &list);
2727 list_add_tail(&ext->oe_link, &list);
2729 ext = next_extent(ext);
2731 osc_object_unlock(obj);
2733 osc_list_maint(cli, obj);
2735 while ((ext = list_first_entry_or_null(&list,
2737 oe_link)) != NULL) {
2740 list_del_init(&ext->oe_link);
2742 /* extent may be in OES_ACTIVE state because inode mutex
2743 * is released before osc_io_end() in file write case */
2744 if (ext->oe_state != OES_TRUNC)
2745 osc_extent_wait(env, ext, OES_TRUNC);
2747 rc = osc_extent_truncate(ext, index, partial);
2752 OSC_EXTENT_DUMP(D_ERROR, ext,
2753 "truncate error %d\n", rc);
2754 } else if (ext->oe_nr_pages == 0) {
2755 osc_extent_remove(ext);
2757 /* this must be an overlapped extent which means only
2758 * part of pages in this extent have been truncated.
2760 EASSERTF(ext->oe_start <= index, ext,
2761 "trunc index = %lu/%d.\n", index, partial);
2762 /* fix index to skip this partially truncated extent */
2763 index = ext->oe_end + 1;
2766 /* we need to hold this extent in OES_TRUNC state so
2767 * that no writeback will happen. This is to avoid
2769 * Only partial truncate can reach here, if @size is
2770 * not zero, the caller should provide a valid @extp. */
2771 LASSERT(*extp == NULL);
2772 *extp = osc_extent_get(ext);
2773 OSC_EXTENT_DUMP(D_CACHE, ext,
2774 "trunc at %llu\n", size);
2776 osc_extent_put(env, ext);
2778 if (waiting != NULL) {
2781 /* ignore the result of osc_extent_wait the write initiator
2782 * should take care of it. */
2783 rc = osc_extent_wait(env, waiting, OES_INV);
2785 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2787 osc_extent_put(env, waiting);
2793 EXPORT_SYMBOL(osc_cache_truncate_start);
2796 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2798 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2801 struct osc_object *obj = ext->oe_obj;
2802 bool unplug = false;
2804 EASSERT(ext->oe_nr_pages > 0, ext);
2805 EASSERT(ext->oe_state == OES_TRUNC, ext);
2806 EASSERT(!ext->oe_urgent, ext);
2808 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2809 osc_object_lock(obj);
2810 osc_extent_state_set(ext, OES_CACHE);
2811 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2813 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2816 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2817 osc_object_unlock(obj);
2818 osc_extent_put(env, ext);
2821 osc_io_unplug_async(env, osc_cli(obj), obj);
2826 * Wait for extents in a specific range to be written out.
2827 * The caller must have called osc_cache_writeback_range() to issue IO
2828 * otherwise it will take a long time for this function to finish.
2830 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2831 * nobody else can dirty this range of file while we're waiting for
2832 * extents to be written.
2834 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2835 pgoff_t start, pgoff_t end)
2837 struct osc_extent *ext;
2838 pgoff_t index = start;
2843 osc_object_lock(obj);
2844 ext = osc_extent_search(obj, index);
2846 ext = first_extent(obj);
2847 else if (ext->oe_end < index)
2848 ext = next_extent(ext);
2849 while (ext != NULL) {
2852 if (ext->oe_start > end)
2855 if (!ext->oe_fsync_wait) {
2856 ext = next_extent(ext);
2860 EASSERT(ergo(ext->oe_state == OES_CACHE,
2861 ext->oe_hp || ext->oe_urgent), ext);
2862 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2863 !ext->oe_hp && ext->oe_urgent), ext);
2865 index = ext->oe_end + 1;
2866 osc_extent_get(ext);
2867 osc_object_unlock(obj);
2869 rc = osc_extent_wait(env, ext, OES_INV);
2872 osc_extent_put(env, ext);
2875 osc_object_unlock(obj);
2877 OSC_IO_DEBUG(obj, "sync file range.\n");
2880 EXPORT_SYMBOL(osc_cache_wait_range);
2883 * Called to write out a range of osc object.
2885 * @hp : should be set this is caused by lock cancel;
2886 * @discard: is set if dirty pages should be dropped - file will be deleted or
2887 * truncated, this implies there is no partially discarding extents.
2889 * Return how many pages will be issued, or error code if error occurred.
2891 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2892 pgoff_t start, pgoff_t end, int hp, int discard)
2894 struct osc_extent *ext;
2895 LIST_HEAD(discard_list);
2896 bool unplug = false;
2900 osc_object_lock(obj);
2901 ext = osc_extent_search(obj, start);
2903 ext = first_extent(obj);
2904 else if (ext->oe_end < start)
2905 ext = next_extent(ext);
2906 while (ext != NULL) {
2907 if (ext->oe_start > end)
2910 ext->oe_fsync_wait = 1;
2911 switch (ext->oe_state) {
2913 result += ext->oe_nr_pages;
2915 struct list_head *list = NULL;
2917 EASSERT(!ext->oe_hp, ext);
2919 list = &obj->oo_hp_exts;
2920 } else if (!ext->oe_urgent && !ext->oe_hp) {
2922 list = &obj->oo_urgent_exts;
2925 list_move_tail(&ext->oe_link, list);
2928 struct client_obd *cli = osc_cli(obj);
2929 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2930 pgoff_t align_by = (1 << pcc_bits);
2931 pgoff_t a_start = round_down(start, align_by);
2932 pgoff_t a_end = round_up(end, align_by);
2936 a_end = CL_PAGE_EOF;
2937 /* the only discarder is lock cancelling, so
2938 * [start, end], aligned by chunk size, must
2939 * contain this extent */
2940 LASSERTF(ext->oe_start >= a_start &&
2941 ext->oe_end <= a_end,
2942 "ext [%lu, %lu] reg [%lu, %lu] "
2943 "orig [%lu %lu] align %lu bits "
2944 "%d\n", ext->oe_start, ext->oe_end,
2945 a_start, a_end, start, end,
2946 align_by, pcc_bits);
2947 osc_extent_state_set(ext, OES_LOCKING);
2948 ext->oe_owner = current;
2949 list_move_tail(&ext->oe_link,
2951 osc_update_pending(obj, OBD_BRW_WRITE,
2956 /* It's pretty bad to wait for ACTIVE extents, because
2957 * we don't know how long we will wait for it to be
2958 * flushed since it may be blocked at awaiting more
2959 * grants. We do this for the correctness of fsync. */
2960 LASSERT(hp == 0 && discard == 0);
2964 /* this extent is being truncated, can't do anything
2965 * for it now. it will be set to urgent after truncate
2966 * is finished in osc_cache_truncate_end(). */
2970 ext = next_extent(ext);
2972 osc_object_unlock(obj);
2974 LASSERT(ergo(!discard, list_empty(&discard_list)));
2975 if (!list_empty(&discard_list)) {
2976 struct osc_extent *tmp;
2979 osc_list_maint(osc_cli(obj), obj);
2980 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2981 list_del_init(&ext->oe_link);
2982 EASSERT(ext->oe_state == OES_LOCKING, ext);
2984 /* Discard caching pages. We don't actually write this
2985 * extent out but we complete it as if we did. */
2986 rc = osc_extent_make_ready(env, ext);
2987 if (unlikely(rc < 0)) {
2988 OSC_EXTENT_DUMP(D_ERROR, ext,
2989 "make_ready returned %d\n", rc);
2994 /* finish the extent as if the pages were sent */
2995 osc_extent_finish(env, ext, 0, 0);
3000 osc_io_unplug(env, osc_cli(obj), obj);
3002 if (hp || discard) {
3004 rc = osc_cache_wait_range(env, obj, start, end);
3005 if (result >= 0 && rc < 0)
3009 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3012 EXPORT_SYMBOL(osc_cache_writeback_range);
3015 * Returns a list of pages by a given [start, end] of \a obj.
3017 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3018 * crucial in the face of [offset, EOF] locks.
3020 * Return at least one page in @queue unless there is no covered page.
3022 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3023 struct osc_object *osc, pgoff_t start, pgoff_t end,
3024 osc_page_gang_cbt cb, void *cbdata)
3026 struct osc_page *ops;
3027 struct folio_batch *fbatch;
3034 bool tree_lock = true;
3038 pvec = osc_env_info(env)->oti_pvec;
3039 fbatch = &osc_env_info(env)->oti_fbatch;
3040 ll_folio_batch_init(fbatch, 0);
3041 spin_lock(&osc->oo_tree_lock);
3042 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3043 idx, OTI_PVEC_SIZE)) > 0) {
3044 struct cl_page *page;
3045 bool end_of_region = false;
3047 for (i = 0, j = 0; i < nr; ++i) {
3051 idx = osc_index(ops);
3053 end_of_region = true;
3057 page = ops->ops_cl.cpl_page;
3058 LASSERT(page->cp_type == CPT_CACHEABLE);
3059 if (page->cp_state == CPS_FREEING)
3063 lu_ref_add_atomic(&page->cp_reference,
3064 "gang_lookup", current);
3070 * Here a delicate locking dance is performed. Current thread
3071 * holds a reference to a page, but has to own it before it
3072 * can be placed into queue. Owning implies waiting, so
3073 * radix-tree lock is to be released. After a wait one has to
3074 * check that pages weren't truncated (cl_page_own() returns
3075 * error in the latter case).
3077 spin_unlock(&osc->oo_tree_lock);
3080 res = (*cb)(env, io, pvec, j, cbdata);
3082 for (i = 0; i < j; ++i) {
3084 page = ops->ops_cl.cpl_page;
3085 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3086 cl_batch_put(env, page, fbatch);
3088 folio_batch_release(fbatch);
3090 if (nr < OTI_PVEC_SIZE || end_of_region)
3096 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_SLOW_PAGE_EVICT,
3097 cfs_fail_val ?: 20);
3099 if (io->ci_type == CIT_MISC &&
3100 io->u.ci_misc.lm_next_rpc_time &&
3101 ktime_get_seconds() > io->u.ci_misc.lm_next_rpc_time) {
3102 osc_send_empty_rpc(osc, idx << PAGE_SHIFT);
3103 io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3104 5 * obd_timeout / 16;
3110 spin_lock(&osc->oo_tree_lock);
3114 spin_unlock(&osc->oo_tree_lock);
3117 EXPORT_SYMBOL(osc_page_gang_lookup);
3120 * Check if page @page is covered by an extra lock or discard it.
3122 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3123 void **pvec, int count, void *cbdata)
3125 struct osc_thread_info *info = osc_env_info(env);
3126 struct osc_object *osc = cbdata;
3129 for (i = 0; i < count; i++) {
3130 struct osc_page *ops = pvec[i];
3131 struct cl_page *page = ops->ops_cl.cpl_page;
3132 pgoff_t index = osc_index(ops);
3133 bool discard = false;
3135 /* negative lock caching */
3136 if (index < info->oti_ng_index) {
3138 } else if (index >= info->oti_fn_index) {
3139 struct ldlm_lock *tmp;
3140 /* refresh non-overlapped index */
3141 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3142 OSC_DAP_FL_TEST_LOCK |
3147 tmp->l_policy_data.l_extent.end;
3149 tmp->l_policy_data.l_extent.start;
3151 /* no lock covering this page */
3152 if (index < start >> PAGE_SHIFT) {
3153 /* no lock at @index,
3154 * first lock at @start
3156 info->oti_ng_index =
3157 start >> PAGE_SHIFT;
3160 /* Cache the first-non-overlapped
3161 * index so as to skip all pages
3162 * within [index, oti_fn_index).
3163 * This is safe because if tmp lock
3164 * is canceled, it will discard these
3167 info->oti_fn_index =
3168 (end + 1) >> PAGE_SHIFT;
3169 if (end == OBD_OBJECT_EOF)
3170 info->oti_fn_index =
3175 info->oti_ng_index = CL_PAGE_EOF;
3181 if (cl_page_own(env, io, page) == 0) {
3182 cl_page_discard(env, io, page);
3183 cl_page_disown(env, io, page);
3185 LASSERT(page->cp_state == CPS_FREEING);
3189 info->oti_next_index = index + 1;
3194 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3195 void **pvec, int count, void *cbdata)
3197 struct osc_thread_info *info = osc_env_info(env);
3200 for (i = 0; i < count; i++) {
3201 struct osc_page *ops = pvec[i];
3202 struct cl_page *page = ops->ops_cl.cpl_page;
3204 /* page is top page. */
3205 info->oti_next_index = osc_index(ops) + 1;
3206 if (cl_page_own(env, io, page) == 0) {
3207 if (!ergo(page->cp_type == CPT_CACHEABLE,
3208 !PageDirty(cl_page_vmpage(page))))
3209 CL_PAGE_DEBUG(D_ERROR, env, page,
3210 "discard dirty page?\n");
3212 /* discard the page */
3213 cl_page_discard(env, io, page);
3214 cl_page_disown(env, io, page);
3216 LASSERT(page->cp_state == CPS_FREEING);
3222 EXPORT_SYMBOL(osc_discard_cb);
3225 * Discard pages protected by the given lock. This function traverses radix
3226 * tree to find all covering pages and discard them. If a page is being covered
3227 * by other locks, it should remain in cache.
3229 * If error happens on any step, the process continues anyway (the reasoning
3230 * behind this being that lock cancellation cannot be delayed indefinitely).
3232 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3233 pgoff_t start, pgoff_t end, bool discard)
3235 struct osc_thread_info *info = osc_env_info(env);
3236 struct cl_io *io = osc_env_thread_io(env);
3237 osc_page_gang_cbt cb;
3242 io->ci_obj = cl_object_top(osc2cl(osc));
3243 io->ci_ignore_layout = 1;
3244 io->ci_invalidate_page_cache = 1;
3245 io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3246 5 * obd_timeout / 16;
3247 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3251 cb = discard ? osc_discard_cb : check_and_discard_cb;
3252 info->oti_fn_index = info->oti_next_index = start;
3253 info->oti_ng_index = 0;
3255 osc_page_gang_lookup(env, io, osc,
3256 info->oti_next_index, end, cb, osc);
3258 cl_io_fini(env, io);