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);
461 * Helper function called by osc_io_submit() for every page in an immediate
462 * transfer (i.e., transferred synchronously).
464 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
465 enum cl_req_type crt, int brw_flags)
467 struct osc_async_page *oap = &opg->ops_oap;
468 struct osc_object *obj = oap->oap_obj;
470 LINVRNT(osc_page_protected(env, opg,
471 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
473 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
474 "magic 0x%x\n", oap, oap->oap_magic);
475 LASSERT(oap->oap_async_flags & ASYNC_READY);
476 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
478 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
479 oap->oap_page_off = opg->ops_from;
480 oap->oap_count = opg->ops_to - opg->ops_from;
481 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
483 if (!client_is_remote(osc_export(obj)) &&
484 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
485 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
486 oap->oap_cmd |= OBD_BRW_NOQUOTA;
489 opg->ops_submit_time = cfs_time_current();
490 osc_page_transfer_get(opg, "transfer\0imm");
491 osc_page_transfer_add(env, opg, crt);
494 /* --------------- LRU page management ------------------ */
496 /* OSC is a natural place to manage LRU pages as applications are specialized
497 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
498 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
499 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
500 * for free LRU slots - this will be very bad so the algorithm requires each
501 * OSC to free slots voluntarily to maintain a reasonable number of free slots
505 static CFS_DECL_WAITQ(osc_lru_waitq);
506 /* LRU pages are freed in batch mode. OSC should at least free this
507 * number of pages to avoid running out of LRU budget, and.. */
508 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
509 /* free this number at most otherwise it will take too long time to finsih. */
510 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
512 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
513 * we should free slots aggressively. In this way, slots are freed in a steady
514 * step to maintain fairness among OSCs.
516 * Return how many LRU pages should be freed. */
517 static int osc_cache_too_much(struct client_obd *cli)
519 struct cl_client_cache *cache = cli->cl_cache;
520 int pages = atomic_read(&cli->cl_lru_in_list);
521 unsigned long budget;
523 budget = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
525 /* if it's going to run out LRU slots, we should free some, but not
526 * too much to maintain faireness among OSCs. */
527 if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
529 return lru_shrink_max;
530 else if (pages >= budget / 2)
531 return lru_shrink_min;
532 } else if (pages >= budget * 2)
533 return lru_shrink_min;
537 int lru_queue_work(const struct lu_env *env, void *data)
539 struct client_obd *cli = data;
541 CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
543 if (osc_cache_too_much(cli))
544 osc_lru_shrink(env, cli, lru_shrink_max, true);
549 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
551 struct list_head lru = LIST_HEAD_INIT(lru);
552 struct osc_async_page *oap;
555 list_for_each_entry(oap, plist, oap_pending_item) {
556 struct osc_page *opg = oap2osc_page(oap);
558 if (!opg->ops_in_lru)
562 LASSERT(list_empty(&opg->ops_lru));
563 list_add(&opg->ops_lru, &lru);
567 client_obd_list_lock(&cli->cl_lru_list_lock);
568 list_splice_tail(&lru, &cli->cl_lru_list);
569 atomic_sub(npages, &cli->cl_lru_busy);
570 atomic_add(npages, &cli->cl_lru_in_list);
571 client_obd_list_unlock(&cli->cl_lru_list_lock);
573 /* XXX: May set force to be true for better performance */
574 if (osc_cache_too_much(cli))
575 (void)ptlrpcd_queue_work(cli->cl_lru_work);
579 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
581 LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
582 list_del_init(&opg->ops_lru);
583 atomic_dec(&cli->cl_lru_in_list);
587 * Page is being destroyed. The page may be not in LRU list, if the transfer
588 * has never finished(error occurred).
590 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
592 if (opg->ops_in_lru) {
593 client_obd_list_lock(&cli->cl_lru_list_lock);
594 if (!list_empty(&opg->ops_lru)) {
595 __osc_lru_del(cli, opg);
597 LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
598 atomic_dec(&cli->cl_lru_busy);
600 client_obd_list_unlock(&cli->cl_lru_list_lock);
602 atomic_inc(cli->cl_lru_left);
603 /* this is a great place to release more LRU pages if
604 * this osc occupies too many LRU pages and kernel is
605 * stealing one of them. */
606 if (!memory_pressure_get())
607 (void)ptlrpcd_queue_work(cli->cl_lru_work);
608 wake_up(&osc_lru_waitq);
610 LASSERT(list_empty(&opg->ops_lru));
615 * Delete page from LRUlist for redirty.
617 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
619 /* If page is being transfered for the first time,
620 * ops_lru should be empty */
621 if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
622 client_obd_list_lock(&cli->cl_lru_list_lock);
623 __osc_lru_del(cli, opg);
624 client_obd_list_unlock(&cli->cl_lru_list_lock);
625 atomic_inc(&cli->cl_lru_busy);
629 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
630 struct cl_page **pvec, int max_index)
634 for (i = 0; i < max_index; i++) {
635 struct cl_page *page = pvec[i];
637 LASSERT(cl_page_is_owned(page, io));
638 cl_page_discard(env, io, page);
639 cl_page_disown(env, io, page);
640 cl_page_put(env, page);
647 * Check if a cl_page can be released, i.e, it's not being used.
649 * If unstable account is turned on, bulk transfer may hold one refcount
650 * for recovery so we need to check vmpage refcount as well; otherwise,
651 * even we can destroy cl_page but the corresponding vmpage can't be reused.
653 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
655 if (cl_page_in_use_noref(page))
658 if (cli->cl_cache->ccc_unstable_check) {
659 struct page *vmpage = cl_page_vmpage(page);
661 /* vmpage have two known users: cl_page and VM page cache */
662 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
669 * Drop @target of pages from LRU at most.
671 int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
672 int target, bool force)
675 struct cl_object *clobj = NULL;
676 struct cl_page **pvec;
677 struct osc_page *opg;
684 LASSERT(atomic_read(&cli->cl_lru_in_list) >= 0);
685 if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
689 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
692 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
693 atomic_dec(&cli->cl_lru_shrinkers);
697 atomic_inc(&cli->cl_lru_shrinkers);
700 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
701 io = &osc_env_info(env)->oti_io;
703 client_obd_list_lock(&cli->cl_lru_list_lock);
704 maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list));
705 while (!list_empty(&cli->cl_lru_list)) {
706 struct cl_page *page;
707 bool will_free = false;
712 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
714 page = opg->ops_cl.cpl_page;
715 if (lru_page_busy(cli, page)) {
716 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
720 LASSERT(page->cp_obj != NULL);
721 if (clobj != page->cp_obj) {
722 struct cl_object *tmp = page->cp_obj;
725 client_obd_list_unlock(&cli->cl_lru_list_lock);
728 discard_pagevec(env, io, pvec, index);
732 cl_object_put(env, clobj);
738 io->ci_ignore_layout = 1;
739 rc = cl_io_init(env, io, CIT_MISC, clobj);
741 client_obd_list_lock(&cli->cl_lru_list_lock);
750 if (cl_page_own_try(env, io, page) == 0) {
751 if (!lru_page_busy(cli, page)) {
752 /* remove it from lru list earlier to avoid
754 __osc_lru_del(cli, opg);
755 opg->ops_in_lru = 0; /* will be discarded */
760 cl_page_disown(env, io, page);
765 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
769 /* Don't discard and free the page with cl_lru_list held */
770 pvec[index++] = page;
771 if (unlikely(index == OTI_PVEC_SIZE)) {
772 client_obd_list_unlock(&cli->cl_lru_list_lock);
773 discard_pagevec(env, io, pvec, index);
776 client_obd_list_lock(&cli->cl_lru_list_lock);
779 if (++count >= target)
782 client_obd_list_unlock(&cli->cl_lru_list_lock);
785 discard_pagevec(env, io, pvec, index);
788 cl_object_put(env, clobj);
791 atomic_dec(&cli->cl_lru_shrinkers);
793 atomic_add(count, cli->cl_lru_left);
794 wake_up_all(&osc_lru_waitq);
796 RETURN(count > 0 ? count : rc);
799 static inline int max_to_shrink(struct client_obd *cli)
801 return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
804 int osc_lru_reclaim(struct client_obd *cli)
806 struct cl_env_nest nest;
808 struct cl_client_cache *cache = cli->cl_cache;
813 LASSERT(cache != NULL);
814 LASSERT(!list_empty(&cache->ccc_lru));
816 env = cl_env_nested_get(&nest);
820 rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
825 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
826 cli->cl_import->imp_obd->obd_name, rc, cli);
830 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
831 cli->cl_import->imp_obd->obd_name, cli,
832 atomic_read(&cli->cl_lru_in_list),
833 atomic_read(&cli->cl_lru_busy));
835 /* Reclaim LRU slots from other client_obd as it can't free enough
836 * from its own. This should rarely happen. */
837 spin_lock(&cache->ccc_lru_lock);
838 cache->ccc_lru_shrinkers++;
839 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
841 max_scans = atomic_read(&cache->ccc_users);
842 while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
843 cli = list_entry(cache->ccc_lru.next, struct client_obd,
846 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
847 cli->cl_import->imp_obd->obd_name, cli,
848 atomic_read(&cli->cl_lru_in_list),
849 atomic_read(&cli->cl_lru_busy));
851 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
852 if (osc_cache_too_much(cli) > 0) {
853 spin_unlock(&cache->ccc_lru_lock);
855 rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
857 spin_lock(&cache->ccc_lru_lock);
862 spin_unlock(&cache->ccc_lru_lock);
865 cl_env_nested_put(&nest, env);
866 CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
867 cli->cl_import->imp_obd->obd_name, cli, rc);
872 * osc_lru_reserve() is called to reserve an LRU slot for a cl_page.
874 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
875 * Only in the case that the LRU slots are in extreme shortage, it should
876 * have reserved enough slots for an IO.
878 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
879 struct osc_page *opg)
881 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
882 struct osc_io *oio = osc_env_io(env);
883 struct client_obd *cli = osc_cli(obj);
887 if (cli->cl_cache == NULL) /* shall not be in LRU */
890 if (oio->oi_lru_reserved > 0) {
891 --oio->oi_lru_reserved;
895 LASSERT(atomic_read(cli->cl_lru_left) >= 0);
896 while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) {
898 /* run out of LRU spaces, try to drop some by itself */
899 rc = osc_lru_reclaim(cli);
906 rc = l_wait_event(osc_lru_waitq,
907 atomic_read(cli->cl_lru_left) > 0,
915 atomic_inc(&cli->cl_lru_busy);
924 * Atomic operations are expensive. We accumulate the accounting for the
925 * same page zone to get better performance.
926 * In practice this can work pretty good because the pages in the same RPC
927 * are likely from the same page zone.
929 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
932 obd_count page_count = desc->bd_iov_count;
937 for (i = 0; i < page_count; i++) {
938 void *pz = page_zone(desc->bd_iov[i].kiov_page);
940 if (likely(pz == zone)) {
946 mod_zone_page_state(zone, NR_UNSTABLE_NFS,
954 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
957 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
959 unstable_page_accounting(desc, 1);
962 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
964 unstable_page_accounting(desc, -1);
968 * Performs "unstable" page accounting. This function balances the
969 * increment operations performed in osc_inc_unstable_pages. It is
970 * registered as the RPC request callback, and is executed when the
971 * bulk RPC is committed on the server. Thus at this point, the pages
972 * involved in the bulk transfer are no longer considered unstable.
974 * If this function is called, the request should have been committed
975 * or req:rq_unstable must have been set; it implies that the unstable
976 * statistic have been added.
978 void osc_dec_unstable_pages(struct ptlrpc_request *req)
980 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
981 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
982 int page_count = desc->bd_iov_count;
985 LASSERT(page_count >= 0);
986 dec_unstable_page_accounting(desc);
988 unstable_count = atomic_sub_return(page_count, &cli->cl_unstable_count);
989 LASSERT(unstable_count >= 0);
991 unstable_count = atomic_sub_return(page_count,
992 &cli->cl_cache->ccc_unstable_nr);
993 LASSERT(unstable_count >= 0);
994 if (unstable_count == 0)
995 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
997 if (osc_cache_too_much(cli))
998 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1002 * "unstable" page accounting. See: osc_dec_unstable_pages.
1004 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1006 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1007 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1008 obd_count page_count = desc->bd_iov_count;
1010 /* No unstable page tracking */
1011 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1014 add_unstable_page_accounting(desc);
1015 atomic_add(page_count, &cli->cl_unstable_count);
1016 atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1018 /* If the request has already been committed (i.e. brw_commit
1019 * called via rq_commit_cb), we need to undo the unstable page
1020 * increments we just performed because rq_commit_cb wont be
1022 spin_lock(&req->rq_lock);
1023 if (unlikely(req->rq_committed)) {
1024 spin_unlock(&req->rq_lock);
1026 osc_dec_unstable_pages(req);
1028 req->rq_unstable = 1;
1029 spin_unlock(&req->rq_lock);
1034 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1035 * This function will be called by every BRW RPC so it's critical
1036 * to make this function fast.
1038 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1040 long unstable_nr, osc_unstable_count;
1042 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1043 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1046 osc_unstable_count = atomic_read(&cli->cl_unstable_count);
1047 unstable_nr = atomic_read(&cli->cl_cache->ccc_unstable_nr);
1050 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1051 cli->cl_import->imp_obd->obd_name, cli,
1052 unstable_nr, osc_unstable_count);
1054 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1055 * has one full RPC window of unstable pages, it's a good chance
1056 * to piggyback a SOFT_SYNC flag.
1057 * Please notice that the OST won't take immediate response for the
1058 * SOFT_SYNC request so active OSCs will have more chance to carry
1059 * the flag, this is reasonable. */
1060 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1061 osc_unstable_count > cli->cl_max_pages_per_rpc *
1062 cli->cl_max_rpcs_in_flight;