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) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_page for OSC layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
38 #define DEBUG_SUBSYSTEM S_OSC
39 #include <lustre_osc.h>
41 #include "osc_internal.h"
43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
46 struct osc_page *opg);
55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
57 struct cl_page *page = opg->ops_cl.cpl_page;
59 LASSERT(!opg->ops_transfer_pinned);
61 lu_ref_add_atomic(&page->cp_reference, label, page);
62 opg->ops_transfer_pinned = 1;
65 static void osc_page_transfer_put(const struct lu_env *env,
68 struct cl_page *page = opg->ops_cl.cpl_page;
70 if (opg->ops_transfer_pinned) {
71 opg->ops_transfer_pinned = 0;
72 lu_ref_del(&page->cp_reference, "transfer", page);
73 cl_page_put(env, page);
78 * This is called once for every page when it is submitted for a transfer
79 * either opportunistic (osc_page_cache_add()), or immediate
80 * (osc_page_submit()).
82 static void osc_page_transfer_add(const struct lu_env *env,
83 struct osc_page *opg, enum cl_req_type crt)
85 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
87 /* ops_lru and ops_inflight share the same field, so take it from LRU
88 * first and then use it as inflight. */
89 osc_lru_use(osc_cli(obj), opg);
92 int osc_page_cache_add(const struct lu_env *env,
93 const struct cl_page_slice *slice, struct cl_io *io)
95 struct osc_page *opg = cl2osc_page(slice);
99 osc_page_transfer_get(opg, "transfer\0cache");
100 result = osc_queue_async_io(env, io, opg);
102 osc_page_transfer_put(env, opg);
104 osc_page_transfer_add(env, opg, CRT_WRITE);
109 void osc_index2policy(union ldlm_policy_data *policy,
110 const struct cl_object *obj, pgoff_t start, pgoff_t end)
112 memset(policy, 0, sizeof *policy);
113 policy->l_extent.start = cl_offset(obj, start);
114 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
117 static const char *osc_list(struct list_head *head)
119 return list_empty(head) ? "-" : "+";
122 static inline s64 osc_submit_duration(struct osc_page *opg)
124 if (ktime_to_ns(opg->ops_submit_time) == 0)
127 return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
130 static int osc_page_print(const struct lu_env *env,
131 const struct cl_page_slice *slice,
132 void *cookie, lu_printer_t printer)
134 struct osc_page *opg = cl2osc_page(slice);
135 struct osc_async_page *oap = &opg->ops_oap;
136 struct osc_object *obj = cl2osc(slice->cpl_obj);
137 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
139 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
140 "1< %#x %d %u %s %s > "
141 "2< %lld %u %u %#x %#x | %p %p %p > "
143 "4< %d %d %d %lu %s | %s %s %s %s > "
144 "5< %s %s %s %s | %d %s | %d %s %s>\n",
147 oap->oap_magic, oap->oap_cmd,
148 oap->oap_interrupted,
149 osc_list(&oap->oap_pending_item),
150 osc_list(&oap->oap_rpc_item),
152 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
153 oap->oap_async_flags, oap->oap_brw_flags,
154 oap->oap_request, oap->oap_cli, obj,
156 opg->ops_transfer_pinned,
157 osc_submit_duration(opg), opg->ops_srvlock,
159 cli->cl_r_in_flight, cli->cl_w_in_flight,
160 cli->cl_max_rpcs_in_flight,
162 osc_list(&cli->cl_cache_waiters),
163 osc_list(&cli->cl_loi_ready_list),
164 osc_list(&cli->cl_loi_hp_ready_list),
165 osc_list(&cli->cl_loi_write_list),
166 osc_list(&cli->cl_loi_read_list),
168 osc_list(&obj->oo_ready_item),
169 osc_list(&obj->oo_hp_ready_item),
170 osc_list(&obj->oo_write_item),
171 osc_list(&obj->oo_read_item),
172 atomic_read(&obj->oo_nr_reads),
173 osc_list(&obj->oo_reading_exts),
174 atomic_read(&obj->oo_nr_writes),
175 osc_list(&obj->oo_hp_exts),
176 osc_list(&obj->oo_urgent_exts));
179 static void osc_page_delete(const struct lu_env *env,
180 const struct cl_page_slice *slice)
182 struct osc_page *opg = cl2osc_page(slice);
183 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
187 CDEBUG(D_TRACE, "%p\n", opg);
188 osc_page_transfer_put(env, opg);
189 rc = osc_teardown_async_page(env, obj, opg);
191 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
192 "Trying to teardown failed: %d\n", rc);
196 osc_lru_del(osc_cli(obj), opg);
198 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
201 spin_lock(&obj->oo_tree_lock);
202 if (opg->ops_intree) {
203 value = radix_tree_delete(&obj->oo_tree,
210 spin_unlock(&obj->oo_tree_lock);
212 LASSERT(ergo(value != NULL, value == opg));
218 static void osc_page_clip(const struct lu_env *env,
219 const struct cl_page_slice *slice,
222 struct osc_page *opg = cl2osc_page(slice);
223 struct osc_async_page *oap = &opg->ops_oap;
225 opg->ops_from = from;
227 spin_lock(&oap->oap_lock);
228 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
229 spin_unlock(&oap->oap_lock);
232 static int osc_page_cancel(const struct lu_env *env,
233 const struct cl_page_slice *slice)
235 struct osc_page *opg = cl2osc_page(slice);
238 /* Check if the transferring against this page
239 * is completed, or not even queued. */
240 if (opg->ops_transfer_pinned)
241 /* FIXME: may not be interrupted.. */
242 rc = osc_cancel_async_page(env, opg);
243 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
247 static int osc_page_flush(const struct lu_env *env,
248 const struct cl_page_slice *slice,
251 struct osc_page *opg = cl2osc_page(slice);
254 rc = osc_flush_async_page(env, io, opg);
258 static void osc_page_touch(const struct lu_env *env,
259 const struct cl_page_slice *slice, size_t to)
261 struct osc_page *opg = cl2osc_page(slice);
262 struct cl_object *obj = opg->ops_cl.cpl_obj;
264 osc_page_touch_at(env, obj, osc_index(opg), to);
267 static const struct cl_page_operations osc_page_ops = {
268 .cpo_print = osc_page_print,
269 .cpo_delete = osc_page_delete,
270 .cpo_clip = osc_page_clip,
271 .cpo_cancel = osc_page_cancel,
272 .cpo_flush = osc_page_flush,
273 .cpo_page_touch = osc_page_touch,
276 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
277 struct cl_page *page, pgoff_t index)
279 struct osc_object *osc = cl2osc(obj);
280 struct osc_page *opg = cl_object_page_slice(obj, page);
281 struct osc_io *oio = osc_env_io(env);
285 opg->ops_to = PAGE_SIZE;
287 INIT_LIST_HEAD(&opg->ops_lru);
289 result = osc_prep_async_page(osc, opg, page->cp_vmpage,
290 cl_offset(obj, index));
294 opg->ops_srvlock = osc_io_srvlock(oio);
295 cl_page_slice_add(page, &opg->ops_cl, obj, index,
299 /* reserve an LRU space for this page */
300 if (page->cp_type == CPT_CACHEABLE) {
301 result = osc_lru_alloc(env, osc_cli(osc), opg);
303 result = radix_tree_preload(GFP_NOFS);
305 spin_lock(&osc->oo_tree_lock);
306 result = radix_tree_insert(&osc->oo_tree,
312 spin_unlock(&osc->oo_tree_lock);
314 radix_tree_preload_end();
321 EXPORT_SYMBOL(osc_page_init);
324 * Helper function called by osc_io_submit() for every page in an immediate
325 * transfer (i.e., transferred synchronously).
327 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
328 enum cl_req_type crt, int brw_flags)
330 struct osc_async_page *oap = &opg->ops_oap;
332 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
333 "magic 0x%x\n", oap, oap->oap_magic);
334 LASSERT(oap->oap_async_flags & ASYNC_READY);
335 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
337 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
338 oap->oap_page_off = opg->ops_from;
339 oap->oap_count = opg->ops_to - opg->ops_from;
340 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
342 if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
343 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
344 oap->oap_cmd |= OBD_BRW_NOQUOTA;
347 opg->ops_submit_time = ktime_get();
348 osc_page_transfer_get(opg, "transfer\0imm");
349 osc_page_transfer_add(env, opg, crt);
352 /* --------------- LRU page management ------------------ */
354 /* OSC is a natural place to manage LRU pages as applications are specialized
355 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
356 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
357 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
358 * for free LRU slots - this will be very bad so the algorithm requires each
359 * OSC to free slots voluntarily to maintain a reasonable number of free slots
363 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
366 * LRU pages are freed in batch mode. OSC should at least free this
367 * number of pages to avoid running out of LRU slots.
369 static inline int lru_shrink_min(struct client_obd *cli)
371 return cli->cl_max_pages_per_rpc * 2;
375 * free this number at most otherwise it will take too long time to finsih.
377 static inline int lru_shrink_max(struct client_obd *cli)
379 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
383 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
384 * we should free slots aggressively. In this way, slots are freed in a steady
385 * step to maintain fairness among OSCs.
387 * Return how many LRU pages should be freed.
389 static int osc_cache_too_much(struct client_obd *cli)
391 struct cl_client_cache *cache = cli->cl_cache;
392 long pages = atomic_long_read(&cli->cl_lru_in_list);
393 unsigned long budget;
395 LASSERT(cache != NULL);
396 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
398 /* if it's going to run out LRU slots, we should free some, but not
399 * too much to maintain faireness among OSCs. */
400 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
402 return lru_shrink_max(cli);
403 else if (pages >= budget / 2)
404 return lru_shrink_min(cli);
406 time64_t duration = ktime_get_real_seconds();
409 /* knock out pages by duration of no IO activity */
410 duration -= cli->cl_lru_last_used;
412 * The difference shouldn't be more than 70 years
413 * so we can safely case to a long. Round to
414 * approximately 1 minute.
416 timediff = (long)(duration >> 6);
417 if (timediff > 0 && pages >= budget / timediff)
418 return lru_shrink_min(cli);
423 int lru_queue_work(const struct lu_env *env, void *data)
425 struct client_obd *cli = data;
428 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
429 count = osc_cache_too_much(cli);
431 int rc = osc_lru_shrink(env, cli, count, false);
433 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
434 cli_name(cli), rc, count);
436 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
437 ptlrpcd_queue_work(cli->cl_lru_work);
444 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
446 struct list_head lru = LIST_HEAD_INIT(lru);
447 struct osc_async_page *oap;
450 list_for_each_entry(oap, plist, oap_pending_item) {
451 struct osc_page *opg = oap2osc_page(oap);
453 if (!opg->ops_in_lru)
457 LASSERT(list_empty(&opg->ops_lru));
458 list_add(&opg->ops_lru, &lru);
462 spin_lock(&cli->cl_lru_list_lock);
463 list_splice_tail(&lru, &cli->cl_lru_list);
464 atomic_long_sub(npages, &cli->cl_lru_busy);
465 atomic_long_add(npages, &cli->cl_lru_in_list);
466 cli->cl_lru_last_used = ktime_get_real_seconds();
467 spin_unlock(&cli->cl_lru_list_lock);
469 if (waitqueue_active(&osc_lru_waitq))
470 (void)ptlrpcd_queue_work(cli->cl_lru_work);
474 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
476 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
477 list_del_init(&opg->ops_lru);
478 atomic_long_dec(&cli->cl_lru_in_list);
482 * Page is being destroyed. The page may be not in LRU list, if the transfer
483 * has never finished(error occurred).
485 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
487 if (opg->ops_in_lru) {
488 spin_lock(&cli->cl_lru_list_lock);
489 if (!list_empty(&opg->ops_lru)) {
490 __osc_lru_del(cli, opg);
492 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
493 atomic_long_dec(&cli->cl_lru_busy);
495 spin_unlock(&cli->cl_lru_list_lock);
497 atomic_long_inc(cli->cl_lru_left);
498 /* this is a great place to release more LRU pages if
499 * this osc occupies too many LRU pages and kernel is
500 * stealing one of them. */
501 if (osc_cache_too_much(cli)) {
502 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
503 (void)ptlrpcd_queue_work(cli->cl_lru_work);
505 wake_up(&osc_lru_waitq);
507 LASSERT(list_empty(&opg->ops_lru));
512 * Delete page from LRU list for redirty.
514 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
516 /* If page is being transferred for the first time,
517 * ops_lru should be empty */
518 if (opg->ops_in_lru) {
519 spin_lock(&cli->cl_lru_list_lock);
520 if (!list_empty(&opg->ops_lru)) {
521 __osc_lru_del(cli, opg);
522 atomic_long_inc(&cli->cl_lru_busy);
524 spin_unlock(&cli->cl_lru_list_lock);
528 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
529 struct cl_page **pvec, int max_index)
531 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
534 ll_pagevec_init(pagevec, 0);
535 for (i = 0; i < max_index; i++) {
536 struct cl_page *page = pvec[i];
538 LASSERT(cl_page_is_owned(page, io));
539 cl_page_delete(env, page);
540 cl_page_discard(env, io, page);
541 cl_page_disown(env, io, page);
542 cl_pagevec_put(env, page, pagevec);
546 pagevec_release(pagevec);
550 * Check if a cl_page can be released, i.e, it's not being used.
552 * If unstable account is turned on, bulk transfer may hold one refcount
553 * for recovery so we need to check vmpage refcount as well; otherwise,
554 * even we can destroy cl_page but the corresponding vmpage can't be reused.
556 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
558 if (cl_page_in_use_noref(page))
561 if (cli->cl_cache->ccc_unstable_check) {
562 struct page *vmpage = cl_page_vmpage(page);
564 /* vmpage have two known users: cl_page and VM page cache */
565 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
572 * Drop @target of pages from LRU at most.
574 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
575 long target, bool force)
578 struct cl_object *clobj = NULL;
579 struct cl_page **pvec;
580 struct osc_page *opg;
587 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
588 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
591 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
592 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
594 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
597 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
598 atomic_dec(&cli->cl_lru_shrinkers);
602 atomic_inc(&cli->cl_lru_shrinkers);
605 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
606 io = osc_env_thread_io(env);
608 spin_lock(&cli->cl_lru_list_lock);
610 cli->cl_lru_reclaim++;
611 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
612 while (!list_empty(&cli->cl_lru_list)) {
613 struct cl_page *page;
614 bool will_free = false;
616 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
622 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
624 page = opg->ops_cl.cpl_page;
625 if (lru_page_busy(cli, page)) {
626 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
630 LASSERT(page->cp_obj != NULL);
631 if (clobj != page->cp_obj) {
632 struct cl_object *tmp = page->cp_obj;
635 spin_unlock(&cli->cl_lru_list_lock);
638 discard_pagevec(env, io, pvec, index);
642 cl_object_put(env, clobj);
648 io->ci_ignore_layout = 1;
649 rc = cl_io_init(env, io, CIT_MISC, clobj);
651 spin_lock(&cli->cl_lru_list_lock);
660 if (cl_page_own_try(env, io, page) == 0) {
661 if (!lru_page_busy(cli, page)) {
662 /* remove it from lru list earlier to avoid
664 __osc_lru_del(cli, opg);
665 opg->ops_in_lru = 0; /* will be discarded */
670 cl_page_disown(env, io, page);
675 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
679 /* Don't discard and free the page with cl_lru_list held */
680 pvec[index++] = page;
681 if (unlikely(index == OTI_PVEC_SIZE)) {
682 spin_unlock(&cli->cl_lru_list_lock);
683 discard_pagevec(env, io, pvec, index);
686 spin_lock(&cli->cl_lru_list_lock);
689 if (++count >= target)
692 spin_unlock(&cli->cl_lru_list_lock);
695 discard_pagevec(env, io, pvec, index);
698 cl_object_put(env, clobj);
701 atomic_dec(&cli->cl_lru_shrinkers);
703 atomic_long_add(count, cli->cl_lru_left);
704 wake_up_all(&osc_lru_waitq);
706 RETURN(count > 0 ? count : rc);
708 EXPORT_SYMBOL(osc_lru_shrink);
711 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
712 * \@npages of LRU slots. For performance consideration, it's better to drop
713 * LRU pages in batch. Therefore, the actual number is adjusted at least
716 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
719 struct cl_client_cache *cache = cli->cl_cache;
725 LASSERT(cache != NULL);
727 env = cl_env_get(&refcheck);
731 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
732 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
733 cli_name(cli), npages);
734 rc = osc_lru_shrink(env, cli, npages, true);
736 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
737 cli_name(cli), rc, npages);
738 if (osc_cache_too_much(cli) > 0)
739 ptlrpcd_queue_work(cli->cl_lru_work);
745 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
746 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
747 atomic_long_read(&cli->cl_lru_busy), npages);
749 /* Reclaim LRU slots from other client_obd as it can't free enough
750 * from its own. This should rarely happen. */
751 spin_lock(&cache->ccc_lru_lock);
752 LASSERT(!list_empty(&cache->ccc_lru));
754 cache->ccc_lru_shrinkers++;
755 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
757 max_scans = atomic_read(&cache->ccc_users) - 2;
758 while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
759 cli = list_entry(cache->ccc_lru.next, struct client_obd,
762 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
764 atomic_long_read(&cli->cl_lru_in_list),
765 atomic_long_read(&cli->cl_lru_busy));
767 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
768 if (osc_cache_too_much(cli) > 0) {
769 spin_unlock(&cache->ccc_lru_lock);
771 rc = osc_lru_shrink(env, cli, npages, true);
772 spin_lock(&cache->ccc_lru_lock);
779 spin_unlock(&cache->ccc_lru_lock);
782 cl_env_put(env, &refcheck);
783 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
784 cli_name(cli), cli, rc);
789 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
791 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
792 * Only in the case that the LRU slots are in extreme shortage, it should
793 * have reserved enough slots for an IO.
795 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
796 struct osc_page *opg)
798 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
799 struct osc_io *oio = osc_env_io(env);
804 if (cli->cl_cache == NULL) /* shall not be in LRU */
807 if (oio->oi_lru_reserved > 0) {
808 --oio->oi_lru_reserved;
812 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
813 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
814 /* run out of LRU spaces, try to drop some by itself */
815 rc = osc_lru_reclaim(cli, 1);
822 rc = l_wait_event(osc_lru_waitq,
823 atomic_long_read(cli->cl_lru_left) > 0,
831 atomic_long_inc(&cli->cl_lru_busy);
840 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
842 * The benefit of doing this is to reduce contention against atomic counter
843 * cl_lru_left by changing it from per-page access to per-IO access.
845 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
847 unsigned long reserved = 0;
848 unsigned long max_pages;
851 /* reserve a full RPC window at most to avoid that a thread accidentally
852 * consumes too many LRU slots */
853 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
854 if (npages > max_pages)
857 c = atomic_long_read(cli->cl_lru_left);
858 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
859 c = atomic_long_read(cli->cl_lru_left);
860 while (c >= npages) {
861 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
865 c = atomic_long_read(cli->cl_lru_left);
867 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
868 /* If there aren't enough pages in the per-OSC LRU then
869 * wake up the LRU thread to try and clear out space, so
870 * we don't block if pages are being dirtied quickly. */
871 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
872 cli_name(cli), atomic_long_read(cli->cl_lru_left),
874 (void)ptlrpcd_queue_work(cli->cl_lru_work);
881 * osc_lru_unreserve() is called to unreserve LRU slots.
883 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
884 * reasons such as page already existing or I/O error. Those reserved slots
885 * should be freed by calling this function.
887 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
889 atomic_long_add(npages, cli->cl_lru_left);
890 wake_up_all(&osc_lru_waitq);
894 * Atomic operations are expensive. We accumulate the accounting for the
895 * same page zone to get better performance.
896 * In practice this can work pretty good because the pages in the same RPC
897 * are likely from the same page zone.
899 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
900 struct osc_brw_async_args *aa,
909 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
910 page_count = desc->bd_iov_count;
912 page_count = aa->aa_page_count;
915 for (i = 0; i < page_count; i++) {
918 pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
920 pz = page_zone(aa->aa_ppga[i]->pg);
922 if (likely(pz == zone)) {
928 mod_zone_page_state(zone, NR_UNSTABLE_NFS,
936 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
939 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
940 struct osc_brw_async_args *aa)
942 unstable_page_accounting(desc, aa, 1);
945 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
946 struct osc_brw_async_args *aa)
948 unstable_page_accounting(desc, aa, -1);
952 * Performs "unstable" page accounting. This function balances the
953 * increment operations performed in osc_inc_unstable_pages. It is
954 * registered as the RPC request callback, and is executed when the
955 * bulk RPC is committed on the server. Thus at this point, the pages
956 * involved in the bulk transfer are no longer considered unstable.
958 * If this function is called, the request should have been committed
959 * or req:rq_unstable must have been set; it implies that the unstable
960 * statistic have been added.
962 void osc_dec_unstable_pages(struct ptlrpc_request *req)
964 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
965 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
966 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
971 page_count = desc->bd_iov_count;
973 page_count = aa->aa_page_count;
975 LASSERT(page_count >= 0);
977 dec_unstable_page_accounting(desc, aa);
979 unstable_count = atomic_long_sub_return(page_count,
980 &cli->cl_unstable_count);
981 LASSERT(unstable_count >= 0);
983 unstable_count = atomic_long_sub_return(page_count,
984 &cli->cl_cache->ccc_unstable_nr);
985 LASSERT(unstable_count >= 0);
986 if (unstable_count == 0)
987 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
989 if (waitqueue_active(&osc_lru_waitq))
990 (void)ptlrpcd_queue_work(cli->cl_lru_work);
994 * "unstable" page accounting. See: osc_dec_unstable_pages.
996 void osc_inc_unstable_pages(struct ptlrpc_request *req)
998 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
999 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1000 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1003 /* No unstable page tracking */
1004 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1008 page_count = desc->bd_iov_count;
1010 page_count = aa->aa_page_count;
1012 add_unstable_page_accounting(desc, aa);
1013 atomic_long_add(page_count, &cli->cl_unstable_count);
1014 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1016 /* If the request has already been committed (i.e. brw_commit
1017 * called via rq_commit_cb), we need to undo the unstable page
1018 * increments we just performed because rq_commit_cb wont be
1020 spin_lock(&req->rq_lock);
1021 if (unlikely(req->rq_committed)) {
1022 spin_unlock(&req->rq_lock);
1024 osc_dec_unstable_pages(req);
1026 req->rq_unstable = 1;
1027 spin_unlock(&req->rq_lock);
1032 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1033 * This function will be called by every BRW RPC so it's critical
1034 * to make this function fast.
1036 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1038 long unstable_nr, osc_unstable_count;
1040 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1041 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1044 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1045 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1048 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1049 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1051 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1052 * has one full RPC window of unstable pages, it's a good chance
1053 * to piggyback a SOFT_SYNC flag.
1054 * Please notice that the OST won't take immediate response for the
1055 * SOFT_SYNC request so active OSCs will have more chance to carry
1056 * the flag, this is reasonable. */
1057 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1058 osc_unstable_count > cli->cl_max_pages_per_rpc *
1059 cli->cl_max_rpcs_in_flight;
1063 * Return how many LRU pages in the cache of all OSC devices
1065 * \retval return # of cached LRU pages times reclaimation tendency
1066 * \retval SHRINK_STOP if it cannot do any scanning in this time
1068 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1069 struct shrink_control *sc)
1071 struct client_obd *cli;
1072 unsigned long cached = 0;
1074 spin_lock(&osc_shrink_lock);
1075 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1076 cached += atomic_long_read(&cli->cl_lru_in_list);
1077 spin_unlock(&osc_shrink_lock);
1079 return (cached * sysctl_vfs_cache_pressure) / 100;
1083 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1085 * \retval number of cached LRU pages reclaimed
1086 * \retval SHRINK_STOP if it cannot do any scanning in this time
1088 * Linux kernel will loop calling this shrinker scan routine with
1089 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1091 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1092 * to scan and try to reclaim LRU pages, just return 0 and
1093 * osc_cache_shrink_count() will report the LRU page number.
1095 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1096 struct shrink_control *sc)
1098 struct client_obd *cli;
1099 struct client_obd *stop_anchor = NULL;
1105 if (sc->nr_to_scan == 0)
1108 if (!(sc->gfp_mask & __GFP_FS))
1111 env = cl_env_get(&refcheck);
1115 spin_lock(&osc_shrink_lock);
1116 while (!list_empty(&osc_shrink_list)) {
1117 cli = list_entry(osc_shrink_list.next, struct client_obd,
1120 if (stop_anchor == NULL)
1122 else if (cli == stop_anchor)
1125 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1126 spin_unlock(&osc_shrink_lock);
1128 /* shrink no more than max_pages_per_rpc for an OSC */
1129 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1130 cli->cl_max_pages_per_rpc ?
1131 cli->cl_max_pages_per_rpc :
1132 sc->nr_to_scan - shrank, true);
1136 if (shrank >= sc->nr_to_scan)
1139 spin_lock(&osc_shrink_lock);
1141 spin_unlock(&osc_shrink_lock);
1144 cl_env_put(env, &refcheck);