4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_page for OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
42 #define DEBUG_SUBSYSTEM S_OSC
44 #include "osc_cl_internal.h"
46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
48 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
49 struct osc_page *opg);
56 * Comment out osc_page_protected because it may sleep inside the
57 * the client_obd_list_lock.
58 * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
59 * -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
60 * -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
63 static int osc_page_is_dlocked(const struct lu_env *env,
64 const struct osc_page *opg,
65 enum cl_lock_mode mode, int pending, int unref)
68 struct osc_object *obj;
69 struct osc_thread_info *info;
70 struct ldlm_res_id *resname;
71 struct lustre_handle *lockh;
72 ldlm_policy_data_t *policy;
78 info = osc_env_info(env);
79 resname = &info->oti_resname;
80 policy = &info->oti_policy;
81 lockh = &info->oti_handle;
82 page = opg->ops_cl.cpl_page;
83 obj = cl2osc(opg->ops_cl.cpl_obj);
85 flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
87 flags |= LDLM_FL_CBPENDING;
89 dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
90 osc_lock_build_res(env, obj, resname);
91 osc_index2policy(policy, page->cp_obj, osc_index(opg), osc_index(opg));
92 return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
93 dlmmode, &flags, NULL, lockh, unref);
97 * Checks an invariant that a page in the cache is covered by a lock, as
100 static int osc_page_protected(const struct lu_env *env,
101 const struct osc_page *opg,
102 enum cl_lock_mode mode, int unref)
104 struct cl_object_header *hdr;
105 struct cl_lock *scan;
106 struct cl_page *page;
107 struct cl_lock_descr *descr;
110 LINVRNT(!opg->ops_temp);
112 page = opg->ops_cl.cpl_page;
113 if (page->cp_owner != NULL &&
114 cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
116 * If IO is done without locks (liblustre, or lloop), lock is
121 /* otherwise check for a DLM lock */
122 result = osc_page_is_dlocked(env, opg, mode, 1, unref);
124 /* maybe this page is a part of a lockless io? */
125 hdr = cl_object_header(opg->ops_cl.cpl_obj);
126 descr = &osc_env_info(env)->oti_descr;
127 descr->cld_mode = mode;
128 descr->cld_start = osc_index(opg);
129 descr->cld_end = osc_index(opg);
130 spin_lock(&hdr->coh_lock_guard);
131 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
133 * Lock-less sub-lock has to be either in HELD state
134 * (when io is actively going on), or in CACHED state,
135 * when top-lock is being unlocked:
136 * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
138 if ((scan->cll_state == CLS_HELD ||
139 scan->cll_state == CLS_CACHED) &&
140 cl_lock_ext_match(&scan->cll_descr, descr)) {
141 struct osc_lock *olck;
143 olck = osc_lock_at(scan);
144 result = osc_lock_is_lockless(olck);
148 spin_unlock(&hdr->coh_lock_guard);
153 static int osc_page_protected(const struct lu_env *env,
154 const struct osc_page *opg,
155 enum cl_lock_mode mode, int unref)
161 /*****************************************************************************
166 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
168 struct cl_page *page = opg->ops_cl.cpl_page;
170 LASSERT(!opg->ops_transfer_pinned);
172 lu_ref_add_atomic(&page->cp_reference, label, page);
173 opg->ops_transfer_pinned = 1;
176 static void osc_page_transfer_put(const struct lu_env *env,
177 struct osc_page *opg)
179 struct cl_page *page = opg->ops_cl.cpl_page;
181 if (opg->ops_transfer_pinned) {
182 opg->ops_transfer_pinned = 0;
183 lu_ref_del(&page->cp_reference, "transfer", page);
184 cl_page_put(env, page);
189 * This is called once for every page when it is submitted for a transfer
190 * either opportunistic (osc_page_cache_add()), or immediate
191 * (osc_page_submit()).
193 static void osc_page_transfer_add(const struct lu_env *env,
194 struct osc_page *opg, enum cl_req_type crt)
196 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
198 /* ops_lru and ops_inflight share the same field, so take it from LRU
199 * first and then use it as inflight. */
200 osc_lru_use(osc_cli(obj), opg);
202 spin_lock(&obj->oo_seatbelt);
203 list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
204 opg->ops_submitter = current;
205 spin_unlock(&obj->oo_seatbelt);
208 int osc_page_cache_add(const struct lu_env *env,
209 const struct cl_page_slice *slice, struct cl_io *io)
211 struct osc_page *opg = cl2osc_page(slice);
215 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
217 osc_page_transfer_get(opg, "transfer\0cache");
218 result = osc_queue_async_io(env, io, opg);
220 osc_page_transfer_put(env, opg);
222 osc_page_transfer_add(env, opg, CRT_WRITE);
227 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
228 pgoff_t start, pgoff_t end)
230 memset(policy, 0, sizeof *policy);
231 policy->l_extent.start = cl_offset(obj, start);
232 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
235 static int osc_page_is_under_lock(const struct lu_env *env,
236 const struct cl_page_slice *slice,
237 struct cl_io *unused, pgoff_t *max_index)
239 struct osc_page *opg = cl2osc_page(slice);
240 struct cl_lock *lock;
241 int result = -ENODATA;
245 lock = cl_lock_at_pgoff(env, slice->cpl_obj, osc_index(opg),
248 *max_index = lock->cll_descr.cld_end;
249 cl_lock_put(env, lock);
255 static const char *osc_list(struct list_head *head)
257 return list_empty(head) ? "-" : "+";
260 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
262 if (opg->ops_submit_time == 0)
265 return (cfs_time_current() - opg->ops_submit_time);
268 static int osc_page_print(const struct lu_env *env,
269 const struct cl_page_slice *slice,
270 void *cookie, lu_printer_t printer)
272 struct osc_page *opg = cl2osc_page(slice);
273 struct osc_async_page *oap = &opg->ops_oap;
274 struct osc_object *obj = cl2osc(slice->cpl_obj);
275 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
277 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
278 "1< %#x %d %u %s %s > "
279 "2< "LPU64" %u %u %#x %#x | %p %p %p > "
280 "3< %s %p %d %lu %d > "
281 "4< %d %d %d %lu %s | %s %s %s %s > "
282 "5< %s %s %s %s | %d %s | %d %s %s>\n",
285 oap->oap_magic, oap->oap_cmd,
286 oap->oap_interrupted,
287 osc_list(&oap->oap_pending_item),
288 osc_list(&oap->oap_rpc_item),
290 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
291 oap->oap_async_flags, oap->oap_brw_flags,
292 oap->oap_request, oap->oap_cli, obj,
294 osc_list(&opg->ops_inflight),
295 opg->ops_submitter, opg->ops_transfer_pinned,
296 osc_submit_duration(opg), opg->ops_srvlock,
298 cli->cl_r_in_flight, cli->cl_w_in_flight,
299 cli->cl_max_rpcs_in_flight,
301 osc_list(&cli->cl_cache_waiters),
302 osc_list(&cli->cl_loi_ready_list),
303 osc_list(&cli->cl_loi_hp_ready_list),
304 osc_list(&cli->cl_loi_write_list),
305 osc_list(&cli->cl_loi_read_list),
307 osc_list(&obj->oo_ready_item),
308 osc_list(&obj->oo_hp_ready_item),
309 osc_list(&obj->oo_write_item),
310 osc_list(&obj->oo_read_item),
311 atomic_read(&obj->oo_nr_reads),
312 osc_list(&obj->oo_reading_exts),
313 atomic_read(&obj->oo_nr_writes),
314 osc_list(&obj->oo_hp_exts),
315 osc_list(&obj->oo_urgent_exts));
318 static void osc_page_delete(const struct lu_env *env,
319 const struct cl_page_slice *slice)
321 struct osc_page *opg = cl2osc_page(slice);
322 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
325 LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
328 CDEBUG(D_TRACE, "%p\n", opg);
329 osc_page_transfer_put(env, opg);
330 rc = osc_teardown_async_page(env, obj, opg);
332 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
333 "Trying to teardown failed: %d\n", rc);
337 spin_lock(&obj->oo_seatbelt);
338 if (opg->ops_submitter != NULL) {
339 LASSERT(!list_empty(&opg->ops_inflight));
340 list_del_init(&opg->ops_inflight);
341 opg->ops_submitter = NULL;
343 spin_unlock(&obj->oo_seatbelt);
345 osc_lru_del(osc_cli(obj), opg);
347 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
350 spin_lock(&obj->oo_tree_lock);
351 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
354 spin_unlock(&obj->oo_tree_lock);
356 LASSERT(ergo(value != NULL, value == opg));
362 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
365 struct osc_page *opg = cl2osc_page(slice);
366 struct osc_async_page *oap = &opg->ops_oap;
368 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
370 opg->ops_from = from;
372 spin_lock(&oap->oap_lock);
373 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
374 spin_unlock(&oap->oap_lock);
377 static int osc_page_cancel(const struct lu_env *env,
378 const struct cl_page_slice *slice)
380 struct osc_page *opg = cl2osc_page(slice);
383 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
385 /* Check if the transferring against this page
386 * is completed, or not even queued. */
387 if (opg->ops_transfer_pinned)
388 /* FIXME: may not be interrupted.. */
389 rc = osc_cancel_async_page(env, opg);
390 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
394 static int osc_page_flush(const struct lu_env *env,
395 const struct cl_page_slice *slice,
398 struct osc_page *opg = cl2osc_page(slice);
401 rc = osc_flush_async_page(env, io, opg);
405 static const struct cl_page_operations osc_page_ops = {
406 .cpo_print = osc_page_print,
407 .cpo_delete = osc_page_delete,
408 .cpo_is_under_lock = osc_page_is_under_lock,
409 .cpo_clip = osc_page_clip,
410 .cpo_cancel = osc_page_cancel,
411 .cpo_flush = osc_page_flush
414 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
415 struct cl_page *page, pgoff_t index)
417 struct osc_object *osc = cl2osc(obj);
418 struct osc_page *opg = cl_object_page_slice(obj, page);
422 opg->ops_to = PAGE_CACHE_SIZE;
424 result = osc_prep_async_page(osc, opg, page->cp_vmpage,
425 cl_offset(obj, index));
427 struct osc_io *oio = osc_env_io(env);
428 opg->ops_srvlock = osc_io_srvlock(oio);
429 cl_page_slice_add(page, &opg->ops_cl, obj, index,
433 * Cannot assert osc_page_protected() here as read-ahead
434 * creates temporary pages outside of a lock.
436 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
437 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
439 /* ops_inflight and ops_lru are the same field, but it doesn't
440 * hurt to initialize it twice :-) */
441 INIT_LIST_HEAD(&opg->ops_inflight);
442 INIT_LIST_HEAD(&opg->ops_lru);
444 /* reserve an LRU space for this page */
445 if (page->cp_type == CPT_CACHEABLE && result == 0) {
446 result = osc_lru_reserve(env, osc, opg);
448 spin_lock(&osc->oo_tree_lock);
449 result = radix_tree_insert(&osc->oo_tree, index, opg);
452 spin_unlock(&osc->oo_tree_lock);
453 LASSERT(result == 0);
460 int osc_over_unstable_soft_limit(struct client_obd *cli)
462 long obd_upages, obd_dpages, osc_upages;
464 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
468 obd_upages = atomic_read(&obd_unstable_pages);
469 obd_dpages = atomic_read(&obd_dirty_pages);
471 osc_upages = atomic_read(&cli->cl_unstable_count);
473 /* obd_max_dirty_pages is the max number of (dirty + unstable)
474 * pages allowed at any given time. To simulate an unstable page
475 * only limit, we subtract the current number of dirty pages
476 * from this max. This difference is roughly the amount of pages
477 * currently available for unstable pages. Thus, the soft limit
478 * is half of that difference. Check osc_upages to ensure we don't
479 * set SOFT_SYNC for OSCs without any outstanding unstable pages. */
480 return osc_upages != 0 &&
481 obd_upages >= (obd_max_dirty_pages - obd_dpages) / 2;
485 * Helper function called by osc_io_submit() for every page in an immediate
486 * transfer (i.e., transferred synchronously).
488 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
489 enum cl_req_type crt, int brw_flags)
491 struct osc_async_page *oap = &opg->ops_oap;
492 struct osc_object *obj = oap->oap_obj;
494 LINVRNT(osc_page_protected(env, opg,
495 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
497 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
498 "magic 0x%x\n", oap, oap->oap_magic);
499 LASSERT(oap->oap_async_flags & ASYNC_READY);
500 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
502 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
503 oap->oap_page_off = opg->ops_from;
504 oap->oap_count = opg->ops_to - opg->ops_from;
505 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
507 if (osc_over_unstable_soft_limit(oap->oap_cli))
508 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
510 if (!client_is_remote(osc_export(obj)) &&
511 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
512 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
513 oap->oap_cmd |= OBD_BRW_NOQUOTA;
516 opg->ops_submit_time = cfs_time_current();
517 osc_page_transfer_get(opg, "transfer\0imm");
518 osc_page_transfer_add(env, opg, crt);
521 /* --------------- LRU page management ------------------ */
523 /* OSC is a natural place to manage LRU pages as applications are specialized
524 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
525 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
526 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
527 * for free LRU slots - this will be very bad so the algorithm requires each
528 * OSC to free slots voluntarily to maintain a reasonable number of free slots
532 static CFS_DECL_WAITQ(osc_lru_waitq);
533 /* LRU pages are freed in batch mode. OSC should at least free this
534 * number of pages to avoid running out of LRU budget, and.. */
535 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
536 /* free this number at most otherwise it will take too long time to finsih. */
537 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
539 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
540 * we should free slots aggressively. In this way, slots are freed in a steady
541 * step to maintain fairness among OSCs.
543 * Return how many LRU pages should be freed. */
544 static int osc_cache_too_much(struct client_obd *cli)
546 struct cl_client_cache *cache = cli->cl_cache;
547 int pages = atomic_read(&cli->cl_lru_in_list);
548 unsigned long budget;
550 budget = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
552 /* if it's going to run out LRU slots, we should free some, but not
553 * too much to maintain faireness among OSCs. */
554 if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
556 return lru_shrink_max;
557 else if (pages >= budget / 2)
558 return lru_shrink_min;
559 } else if (pages >= budget * 2)
560 return lru_shrink_min;
564 int lru_queue_work(const struct lu_env *env, void *data)
566 struct client_obd *cli = data;
568 CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
570 if (osc_cache_too_much(cli))
571 osc_lru_shrink(env, cli, lru_shrink_max, true);
576 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
578 struct list_head lru = LIST_HEAD_INIT(lru);
579 struct osc_async_page *oap;
582 list_for_each_entry(oap, plist, oap_pending_item) {
583 struct osc_page *opg = oap2osc_page(oap);
585 if (!opg->ops_in_lru)
589 LASSERT(list_empty(&opg->ops_lru));
590 list_add(&opg->ops_lru, &lru);
594 client_obd_list_lock(&cli->cl_lru_list_lock);
595 list_splice_tail(&lru, &cli->cl_lru_list);
596 atomic_sub(npages, &cli->cl_lru_busy);
597 atomic_add(npages, &cli->cl_lru_in_list);
598 client_obd_list_unlock(&cli->cl_lru_list_lock);
600 /* XXX: May set force to be true for better performance */
601 if (osc_cache_too_much(cli))
602 (void)ptlrpcd_queue_work(cli->cl_lru_work);
606 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
608 LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
609 list_del_init(&opg->ops_lru);
610 atomic_dec(&cli->cl_lru_in_list);
614 * Page is being destroyed. The page may be not in LRU list, if the transfer
615 * has never finished(error occurred).
617 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
619 if (opg->ops_in_lru) {
620 client_obd_list_lock(&cli->cl_lru_list_lock);
621 if (!list_empty(&opg->ops_lru)) {
622 __osc_lru_del(cli, opg);
624 LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
625 atomic_dec(&cli->cl_lru_busy);
627 client_obd_list_unlock(&cli->cl_lru_list_lock);
629 atomic_inc(cli->cl_lru_left);
630 /* this is a great place to release more LRU pages if
631 * this osc occupies too many LRU pages and kernel is
632 * stealing one of them. */
633 if (!memory_pressure_get())
634 (void)ptlrpcd_queue_work(cli->cl_lru_work);
635 wake_up(&osc_lru_waitq);
637 LASSERT(list_empty(&opg->ops_lru));
642 * Delete page from LRUlist for redirty.
644 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
646 /* If page is being transfered for the first time,
647 * ops_lru should be empty */
648 if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
649 client_obd_list_lock(&cli->cl_lru_list_lock);
650 __osc_lru_del(cli, opg);
651 client_obd_list_unlock(&cli->cl_lru_list_lock);
652 atomic_inc(&cli->cl_lru_busy);
656 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
657 struct cl_page **pvec, int max_index)
661 for (i = 0; i < max_index; i++) {
662 struct cl_page *page = pvec[i];
664 LASSERT(cl_page_is_owned(page, io));
665 cl_page_discard(env, io, page);
666 cl_page_disown(env, io, page);
667 cl_page_put(env, page);
674 * Drop @target of pages from LRU at most.
676 int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
677 int target, bool force)
680 struct cl_object *clobj = NULL;
681 struct cl_page **pvec;
682 struct osc_page *opg;
689 LASSERT(atomic_read(&cli->cl_lru_in_list) >= 0);
690 if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
694 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
697 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
698 atomic_dec(&cli->cl_lru_shrinkers);
702 atomic_inc(&cli->cl_lru_shrinkers);
705 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
706 io = &osc_env_info(env)->oti_io;
708 client_obd_list_lock(&cli->cl_lru_list_lock);
709 maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list));
710 while (!list_empty(&cli->cl_lru_list)) {
711 struct cl_page *page;
712 bool will_free = false;
717 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
719 page = opg->ops_cl.cpl_page;
720 if (cl_page_in_use_noref(page)) {
721 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
725 LASSERT(page->cp_obj != NULL);
726 if (clobj != page->cp_obj) {
727 struct cl_object *tmp = page->cp_obj;
730 client_obd_list_unlock(&cli->cl_lru_list_lock);
733 discard_pagevec(env, io, pvec, index);
737 cl_object_put(env, clobj);
743 io->ci_ignore_layout = 1;
744 rc = cl_io_init(env, io, CIT_MISC, clobj);
746 client_obd_list_lock(&cli->cl_lru_list_lock);
755 if (cl_page_own_try(env, io, page) == 0) {
756 if (!cl_page_in_use_noref(page)) {
757 /* remove it from lru list earlier to avoid
759 __osc_lru_del(cli, opg);
760 opg->ops_in_lru = 0; /* will be discarded */
765 cl_page_disown(env, io, page);
770 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
774 /* Don't discard and free the page with cl_lru_list held */
775 pvec[index++] = page;
776 if (unlikely(index == OTI_PVEC_SIZE)) {
777 client_obd_list_unlock(&cli->cl_lru_list_lock);
778 discard_pagevec(env, io, pvec, index);
781 client_obd_list_lock(&cli->cl_lru_list_lock);
784 if (++count >= target)
787 client_obd_list_unlock(&cli->cl_lru_list_lock);
790 discard_pagevec(env, io, pvec, index);
793 cl_object_put(env, clobj);
796 atomic_dec(&cli->cl_lru_shrinkers);
798 atomic_add(count, cli->cl_lru_left);
799 wake_up_all(&osc_lru_waitq);
801 RETURN(count > 0 ? count : rc);
804 static inline int max_to_shrink(struct client_obd *cli)
806 return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
809 int osc_lru_reclaim(struct client_obd *cli)
811 struct cl_env_nest nest;
813 struct cl_client_cache *cache = cli->cl_cache;
818 LASSERT(cache != NULL);
819 LASSERT(!list_empty(&cache->ccc_lru));
821 env = cl_env_nested_get(&nest);
825 rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
830 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
831 cli->cl_import->imp_obd->obd_name, rc, cli);
835 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
836 cli->cl_import->imp_obd->obd_name, cli,
837 atomic_read(&cli->cl_lru_in_list),
838 atomic_read(&cli->cl_lru_busy));
840 /* Reclaim LRU slots from other client_obd as it can't free enough
841 * from its own. This should rarely happen. */
842 spin_lock(&cache->ccc_lru_lock);
843 cache->ccc_lru_shrinkers++;
844 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
846 max_scans = atomic_read(&cache->ccc_users);
847 while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
848 cli = list_entry(cache->ccc_lru.next, struct client_obd,
851 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
852 cli->cl_import->imp_obd->obd_name, cli,
853 atomic_read(&cli->cl_lru_in_list),
854 atomic_read(&cli->cl_lru_busy));
856 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
857 if (osc_cache_too_much(cli) > 0) {
858 spin_unlock(&cache->ccc_lru_lock);
860 rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
862 spin_lock(&cache->ccc_lru_lock);
867 spin_unlock(&cache->ccc_lru_lock);
870 cl_env_nested_put(&nest, env);
871 CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
872 cli->cl_import->imp_obd->obd_name, cli, rc);
876 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
877 struct osc_page *opg)
879 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
880 struct osc_io *oio = osc_env_io(env);
881 struct client_obd *cli = osc_cli(obj);
885 if (cli->cl_cache == NULL) /* shall not be in LRU */
888 if (oio->oi_lru_reserved > 0) {
889 --oio->oi_lru_reserved;
893 LASSERT(atomic_read(cli->cl_lru_left) >= 0);
894 while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) {
896 /* run out of LRU spaces, try to drop some by itself */
897 rc = osc_lru_reclaim(cli);
904 rc = l_wait_event(osc_lru_waitq,
905 atomic_read(cli->cl_lru_left) > 0,
913 atomic_inc(&cli->cl_lru_busy);