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 osc_lru_use(osc_cli(obj), opg);
90 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
91 struct cl_io *io, cl_commit_cbt cb)
96 osc_page_transfer_get(opg, "transfer\0cache");
97 result = osc_queue_async_io(env, io, opg, cb);
99 osc_page_transfer_put(env, opg);
101 osc_page_transfer_add(env, opg, CRT_WRITE);
106 void osc_index2policy(union ldlm_policy_data *policy,
107 const struct cl_object *obj, pgoff_t start, pgoff_t end)
109 memset(policy, 0, sizeof *policy);
110 policy->l_extent.start = cl_offset(obj, start);
111 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
114 static inline s64 osc_submit_duration(struct osc_page *opg)
116 if (ktime_to_ns(opg->ops_submit_time) == 0)
119 return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
122 static int osc_page_print(const struct lu_env *env,
123 const struct cl_page_slice *slice,
124 void *cookie, lu_printer_t printer)
126 struct osc_page *opg = cl2osc_page(slice);
127 struct osc_async_page *oap = &opg->ops_oap;
128 struct osc_object *obj = cl2osc(slice->cpl_obj);
129 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
131 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
133 "2< %lld %u %u %#x %#x | %p %p %p > "
135 "4< %d %d %d %lu %c | %c %c %c %c > "
136 "5< %c %c %c %c | %d %c | %d %c %c>\n",
139 oap->oap_magic, oap->oap_cmd,
140 list_empty_marker(&oap->oap_pending_item),
141 list_empty_marker(&oap->oap_rpc_item),
143 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
144 oap->oap_async_flags, oap->oap_brw_flags,
145 oap->oap_request, oap->oap_cli, obj,
147 opg->ops_transfer_pinned,
148 osc_submit_duration(opg), opg->ops_srvlock,
150 cli->cl_r_in_flight, cli->cl_w_in_flight,
151 cli->cl_max_rpcs_in_flight,
153 waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
154 list_empty_marker(&cli->cl_loi_ready_list),
155 list_empty_marker(&cli->cl_loi_hp_ready_list),
156 list_empty_marker(&cli->cl_loi_write_list),
157 list_empty_marker(&cli->cl_loi_read_list),
159 list_empty_marker(&obj->oo_ready_item),
160 list_empty_marker(&obj->oo_hp_ready_item),
161 list_empty_marker(&obj->oo_write_item),
162 list_empty_marker(&obj->oo_read_item),
163 atomic_read(&obj->oo_nr_reads),
164 list_empty_marker(&obj->oo_reading_exts),
165 atomic_read(&obj->oo_nr_writes),
166 list_empty_marker(&obj->oo_hp_exts),
167 list_empty_marker(&obj->oo_urgent_exts));
170 static void osc_page_delete(const struct lu_env *env,
171 const struct cl_page_slice *slice)
173 struct osc_page *opg = cl2osc_page(slice);
174 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
178 CDEBUG(D_TRACE, "%p\n", opg);
179 osc_page_transfer_put(env, opg);
180 rc = osc_teardown_async_page(env, obj, opg);
182 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
183 "Trying to teardown failed: %d\n", rc);
187 osc_lru_del(osc_cli(obj), opg);
189 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
192 spin_lock(&obj->oo_tree_lock);
193 if (opg->ops_intree) {
194 value = radix_tree_delete(&obj->oo_tree,
201 spin_unlock(&obj->oo_tree_lock);
203 LASSERT(ergo(value != NULL, value == opg));
209 static void osc_page_clip(const struct lu_env *env,
210 const struct cl_page_slice *slice,
213 struct osc_page *opg = cl2osc_page(slice);
214 struct osc_async_page *oap = &opg->ops_oap;
216 opg->ops_from = from;
217 /* argument @to is exclusive, but @ops_to is inclusive */
218 opg->ops_to = to - 1;
219 spin_lock(&oap->oap_lock);
220 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
221 spin_unlock(&oap->oap_lock);
224 static int osc_page_flush(const struct lu_env *env,
225 const struct cl_page_slice *slice,
228 struct osc_page *opg = cl2osc_page(slice);
231 rc = osc_flush_async_page(env, io, opg);
235 static void osc_page_touch(const struct lu_env *env,
236 const struct cl_page_slice *slice, size_t to)
238 struct osc_page *opg = cl2osc_page(slice);
239 struct cl_object *obj = opg->ops_cl.cpl_obj;
241 osc_page_touch_at(env, obj, osc_index(opg), to);
244 static const struct cl_page_operations osc_page_ops = {
245 .cpo_print = osc_page_print,
246 .cpo_delete = osc_page_delete,
247 .cpo_clip = osc_page_clip,
248 .cpo_flush = osc_page_flush,
249 .cpo_page_touch = osc_page_touch,
252 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
253 struct cl_page *cl_page, pgoff_t index)
255 struct osc_object *osc = cl2osc(obj);
256 struct osc_page *opg = cl_object_page_slice(obj, cl_page);
257 struct osc_io *oio = osc_env_io(env);
261 opg->ops_to = PAGE_SIZE - 1;
263 INIT_LIST_HEAD(&opg->ops_lru);
265 result = osc_prep_async_page(osc, opg, cl_page->cp_vmpage,
266 cl_offset(obj, index));
270 opg->ops_srvlock = osc_io_srvlock(oio);
271 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
272 cl_page->cp_osc_index = index;
274 /* reserve an LRU space for this page */
275 if (cl_page->cp_type == CPT_CACHEABLE) {
276 result = osc_lru_alloc(env, osc_cli(osc), opg);
278 result = radix_tree_preload(GFP_NOFS);
280 spin_lock(&osc->oo_tree_lock);
281 result = radix_tree_insert(&osc->oo_tree,
287 spin_unlock(&osc->oo_tree_lock);
289 radix_tree_preload_end();
296 EXPORT_SYMBOL(osc_page_init);
299 * Helper function called by osc_io_submit() for every page in an immediate
300 * transfer (i.e., transferred synchronously).
302 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
303 enum cl_req_type crt, int brw_flags)
305 struct osc_io *oio = osc_env_io(env);
306 struct osc_async_page *oap = &opg->ops_oap;
308 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
309 "magic 0x%x\n", oap, oap->oap_magic);
310 LASSERT(oap->oap_async_flags & ASYNC_READY);
311 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
313 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
314 oap->oap_page_off = opg->ops_from;
315 oap->oap_count = opg->ops_to - opg->ops_from + 1;
316 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
318 if (oio->oi_cap_sys_resource) {
319 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
320 oap->oap_cmd |= OBD_BRW_NOQUOTA;
323 opg->ops_submit_time = ktime_get();
324 osc_page_transfer_get(opg, "transfer\0imm");
325 osc_page_transfer_add(env, opg, crt);
328 /* --------------- LRU page management ------------------ */
330 /* OSC is a natural place to manage LRU pages as applications are specialized
331 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
332 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
333 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
334 * for free LRU slots - this will be very bad so the algorithm requires each
335 * OSC to free slots voluntarily to maintain a reasonable number of free slots
339 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
342 * LRU pages are freed in batch mode. OSC should at least free this
343 * number of pages to avoid running out of LRU slots.
345 static inline int lru_shrink_min(struct client_obd *cli)
347 return cli->cl_max_pages_per_rpc * 2;
351 * free this number at most otherwise it will take too long time to finsih.
353 static inline int lru_shrink_max(struct client_obd *cli)
355 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
359 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
360 * we should free slots aggressively. In this way, slots are freed in a steady
361 * step to maintain fairness among OSCs.
363 * Return how many LRU pages should be freed.
365 static int osc_cache_too_much(struct client_obd *cli)
367 struct cl_client_cache *cache = cli->cl_cache;
368 long pages = atomic_long_read(&cli->cl_lru_in_list);
369 unsigned long budget;
371 LASSERT(cache != NULL);
372 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
374 /* if it's going to run out LRU slots, we should free some, but not
375 * too much to maintain faireness among OSCs. */
376 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
378 return lru_shrink_max(cli);
379 else if (pages >= budget / 2)
380 return lru_shrink_min(cli);
382 time64_t duration = ktime_get_real_seconds();
385 /* knock out pages by duration of no IO activity */
386 duration -= cli->cl_lru_last_used;
388 * The difference shouldn't be more than 70 years
389 * so we can safely case to a long. Round to
390 * approximately 1 minute.
392 timediff = (long)(duration >> 6);
393 if (timediff > 0 && pages >= budget / timediff)
394 return lru_shrink_min(cli);
399 int lru_queue_work(const struct lu_env *env, void *data)
401 struct client_obd *cli = data;
404 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
405 count = osc_cache_too_much(cli);
407 int rc = osc_lru_shrink(env, cli, count, false);
409 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
410 cli_name(cli), rc, count);
412 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
413 ptlrpcd_queue_work(cli->cl_lru_work);
420 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
423 struct osc_async_page *oap;
426 list_for_each_entry(oap, plist, oap_pending_item) {
427 struct osc_page *opg = oap2osc_page(oap);
429 if (!opg->ops_in_lru)
433 LASSERT(list_empty(&opg->ops_lru));
434 list_add(&opg->ops_lru, &lru);
438 spin_lock(&cli->cl_lru_list_lock);
439 list_splice_tail(&lru, &cli->cl_lru_list);
440 atomic_long_sub(npages, &cli->cl_lru_busy);
441 atomic_long_add(npages, &cli->cl_lru_in_list);
442 cli->cl_lru_last_used = ktime_get_real_seconds();
443 spin_unlock(&cli->cl_lru_list_lock);
445 if (waitqueue_active(&osc_lru_waitq))
446 (void)ptlrpcd_queue_work(cli->cl_lru_work);
450 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
452 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
453 list_del_init(&opg->ops_lru);
454 atomic_long_dec(&cli->cl_lru_in_list);
458 * Page is being destroyed. The page may be not in LRU list, if the transfer
459 * has never finished(error occurred).
461 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
463 if (opg->ops_in_lru) {
464 spin_lock(&cli->cl_lru_list_lock);
465 if (!list_empty(&opg->ops_lru)) {
466 __osc_lru_del(cli, opg);
468 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
469 atomic_long_dec(&cli->cl_lru_busy);
471 spin_unlock(&cli->cl_lru_list_lock);
473 atomic_long_inc(cli->cl_lru_left);
474 /* this is a great place to release more LRU pages if
475 * this osc occupies too many LRU pages and kernel is
476 * stealing one of them. */
477 if (osc_cache_too_much(cli)) {
478 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
479 (void)ptlrpcd_queue_work(cli->cl_lru_work);
481 wake_up(&osc_lru_waitq);
483 LASSERT(list_empty(&opg->ops_lru));
488 * Delete page from LRU list for redirty.
490 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
492 /* If page is being transferred for the first time,
493 * ops_lru should be empty */
494 if (opg->ops_in_lru) {
495 if (list_empty(&opg->ops_lru))
497 spin_lock(&cli->cl_lru_list_lock);
498 if (!list_empty(&opg->ops_lru)) {
499 __osc_lru_del(cli, opg);
500 atomic_long_inc(&cli->cl_lru_busy);
502 spin_unlock(&cli->cl_lru_list_lock);
506 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
507 struct cl_page **pvec, int max_index)
509 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
512 ll_pagevec_init(pagevec, 0);
513 for (i = 0; i < max_index; i++) {
514 struct cl_page *page = pvec[i];
516 LASSERT(cl_page_is_owned(page, io));
517 cl_page_delete(env, page);
518 cl_page_discard(env, io, page);
519 cl_page_disown(env, io, page);
520 cl_pagevec_put(env, page, pagevec);
524 pagevec_release(pagevec);
528 * Check if a cl_page can be released, i.e, it's not being used.
530 * If unstable account is turned on, bulk transfer may hold one refcount
531 * for recovery so we need to check vmpage refcount as well; otherwise,
532 * even we can destroy cl_page but the corresponding vmpage can't be reused.
534 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
536 if (cl_page_in_use_noref(page))
539 if (cli->cl_cache->ccc_unstable_check) {
540 struct page *vmpage = cl_page_vmpage(page);
542 /* vmpage have two known users: cl_page and VM page cache */
543 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
550 * Drop @target of pages from LRU at most.
552 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
553 long target, bool force)
556 struct cl_object *clobj = NULL;
557 struct cl_page **pvec;
558 struct osc_page *opg;
565 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
566 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
569 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
570 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
572 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
575 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
576 atomic_dec(&cli->cl_lru_shrinkers);
580 atomic_inc(&cli->cl_lru_shrinkers);
583 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
584 io = osc_env_thread_io(env);
586 spin_lock(&cli->cl_lru_list_lock);
588 cli->cl_lru_reclaim++;
589 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
590 while (!list_empty(&cli->cl_lru_list)) {
591 struct cl_page *page;
592 bool will_free = false;
594 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
600 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
602 page = opg->ops_cl.cpl_page;
603 if (lru_page_busy(cli, page)) {
604 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
608 LASSERT(page->cp_obj != NULL);
609 if (clobj != page->cp_obj) {
610 struct cl_object *tmp = page->cp_obj;
613 spin_unlock(&cli->cl_lru_list_lock);
616 discard_pagevec(env, io, pvec, index);
620 cl_object_put(env, clobj);
626 io->ci_ignore_layout = 1;
627 rc = cl_io_init(env, io, CIT_MISC, clobj);
629 spin_lock(&cli->cl_lru_list_lock);
638 if (cl_page_own_try(env, io, page) == 0) {
639 if (!lru_page_busy(cli, page)) {
640 /* remove it from lru list earlier to avoid
642 __osc_lru_del(cli, opg);
643 opg->ops_in_lru = 0; /* will be discarded */
648 cl_page_disown(env, io, page);
653 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
657 /* Don't discard and free the page with cl_lru_list held */
658 pvec[index++] = page;
659 if (unlikely(index == OTI_PVEC_SIZE)) {
660 spin_unlock(&cli->cl_lru_list_lock);
661 discard_pagevec(env, io, pvec, index);
664 spin_lock(&cli->cl_lru_list_lock);
667 if (++count >= target)
670 spin_unlock(&cli->cl_lru_list_lock);
673 discard_pagevec(env, io, pvec, index);
676 cl_object_put(env, clobj);
679 atomic_dec(&cli->cl_lru_shrinkers);
681 atomic_long_add(count, cli->cl_lru_left);
682 wake_up_all(&osc_lru_waitq);
684 RETURN(count > 0 ? count : rc);
686 EXPORT_SYMBOL(osc_lru_shrink);
689 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
690 * \@npages of LRU slots. For performance consideration, it's better to drop
691 * LRU pages in batch. Therefore, the actual number is adjusted at least
694 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
697 struct cl_client_cache *cache = cli->cl_cache;
703 LASSERT(cache != NULL);
705 env = cl_env_get(&refcheck);
709 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
710 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
711 cli_name(cli), npages);
712 rc = osc_lru_shrink(env, cli, npages, true);
714 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
715 cli_name(cli), rc, npages);
716 if (osc_cache_too_much(cli) > 0)
717 ptlrpcd_queue_work(cli->cl_lru_work);
723 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
724 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
725 atomic_long_read(&cli->cl_lru_busy), npages);
727 /* Reclaim LRU slots from other client_obd as it can't free enough
728 * from its own. This should rarely happen. */
729 spin_lock(&cache->ccc_lru_lock);
730 LASSERT(!list_empty(&cache->ccc_lru));
732 cache->ccc_lru_shrinkers++;
733 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
735 max_scans = atomic_read(&cache->ccc_users) - 2;
736 while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
737 cli = list_entry(cache->ccc_lru.next, struct client_obd,
740 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
742 atomic_long_read(&cli->cl_lru_in_list),
743 atomic_long_read(&cli->cl_lru_busy));
745 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
746 if (osc_cache_too_much(cli) > 0) {
747 spin_unlock(&cache->ccc_lru_lock);
749 rc = osc_lru_shrink(env, cli, npages, true);
750 spin_lock(&cache->ccc_lru_lock);
757 spin_unlock(&cache->ccc_lru_lock);
760 cl_env_put(env, &refcheck);
761 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
762 cli_name(cli), cli, rc);
767 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
769 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
770 * Only in the case that the LRU slots are in extreme shortage, it should
771 * have reserved enough slots for an IO.
773 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
774 struct osc_page *opg)
776 struct osc_io *oio = osc_env_io(env);
781 if (cli->cl_cache == NULL) /* shall not be in LRU */
784 if (oio->oi_lru_reserved > 0) {
785 --oio->oi_lru_reserved;
789 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
790 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
791 /* run out of LRU spaces, try to drop some by itself */
792 rc = osc_lru_reclaim(cli, 1);
799 rc = l_wait_event_abortable(
801 atomic_long_read(cli->cl_lru_left) > 0);
810 atomic_long_inc(&cli->cl_lru_busy);
819 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
821 * The benefit of doing this is to reduce contention against atomic counter
822 * cl_lru_left by changing it from per-page access to per-IO access.
824 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
826 unsigned long reserved = 0;
827 unsigned long max_pages;
830 /* reserve a full RPC window at most to avoid that a thread accidentally
831 * consumes too many LRU slots */
832 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
833 if (npages > max_pages)
836 c = atomic_long_read(cli->cl_lru_left);
837 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
838 c = atomic_long_read(cli->cl_lru_left);
839 while (c >= npages) {
840 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
844 c = atomic_long_read(cli->cl_lru_left);
846 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
847 /* If there aren't enough pages in the per-OSC LRU then
848 * wake up the LRU thread to try and clear out space, so
849 * we don't block if pages are being dirtied quickly. */
850 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
851 cli_name(cli), atomic_long_read(cli->cl_lru_left),
853 (void)ptlrpcd_queue_work(cli->cl_lru_work);
860 * osc_lru_unreserve() is called to unreserve LRU slots.
862 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
863 * reasons such as page already existing or I/O error. Those reserved slots
864 * should be freed by calling this function.
866 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
868 atomic_long_add(npages, cli->cl_lru_left);
869 wake_up_all(&osc_lru_waitq);
873 * Atomic operations are expensive. We accumulate the accounting for the
874 * same page zone to get better performance.
875 * In practice this can work pretty good because the pages in the same RPC
876 * are likely from the same page zone.
878 #ifdef HAVE_NR_UNSTABLE_NFS
879 /* Old kernels use a separate counter for unstable pages,
880 * newer kernels treat them like any other writeback.
882 #define NR_WRITEBACK NR_UNSTABLE_NFS
885 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
886 struct osc_brw_async_args *aa,
895 page_count = desc->bd_iov_count;
897 page_count = aa->aa_page_count;
900 for (i = 0; i < page_count; i++) {
903 pz = page_zone(desc->bd_vec[i].bv_page);
905 pz = page_zone(aa->aa_ppga[i]->pg);
907 if (likely(pz == zone)) {
913 mod_zone_page_state(zone, NR_WRITEBACK,
921 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
924 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
925 struct osc_brw_async_args *aa)
927 unstable_page_accounting(desc, aa, 1);
930 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
931 struct osc_brw_async_args *aa)
933 unstable_page_accounting(desc, aa, -1);
937 * Performs "unstable" page accounting. This function balances the
938 * increment operations performed in osc_inc_unstable_pages. It is
939 * registered as the RPC request callback, and is executed when the
940 * bulk RPC is committed on the server. Thus at this point, the pages
941 * involved in the bulk transfer are no longer considered unstable.
943 * If this function is called, the request should have been committed
944 * or req:rq_unstable must have been set; it implies that the unstable
945 * statistic have been added.
947 void osc_dec_unstable_pages(struct ptlrpc_request *req)
949 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
950 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
951 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
956 page_count = desc->bd_iov_count;
958 page_count = aa->aa_page_count;
960 LASSERT(page_count >= 0);
962 dec_unstable_page_accounting(desc, aa);
964 unstable_count = atomic_long_sub_return(page_count,
965 &cli->cl_unstable_count);
966 LASSERT(unstable_count >= 0);
968 unstable_count = atomic_long_sub_return(page_count,
969 &cli->cl_cache->ccc_unstable_nr);
970 LASSERT(unstable_count >= 0);
971 if (unstable_count == 0)
972 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
974 if (waitqueue_active(&osc_lru_waitq))
975 (void)ptlrpcd_queue_work(cli->cl_lru_work);
979 * "unstable" page accounting. See: osc_dec_unstable_pages.
981 void osc_inc_unstable_pages(struct ptlrpc_request *req)
983 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
984 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
985 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
988 /* No unstable page tracking */
989 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
993 page_count = desc->bd_iov_count;
995 page_count = aa->aa_page_count;
997 add_unstable_page_accounting(desc, aa);
998 atomic_long_add(page_count, &cli->cl_unstable_count);
999 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1001 /* If the request has already been committed (i.e. brw_commit
1002 * called via rq_commit_cb), we need to undo the unstable page
1003 * increments we just performed because rq_commit_cb wont be
1005 spin_lock(&req->rq_lock);
1006 if (unlikely(req->rq_committed)) {
1007 spin_unlock(&req->rq_lock);
1009 osc_dec_unstable_pages(req);
1011 req->rq_unstable = 1;
1012 spin_unlock(&req->rq_lock);
1017 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1018 * This function will be called by every BRW RPC so it's critical
1019 * to make this function fast.
1021 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1023 long unstable_nr, osc_unstable_count;
1025 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1026 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1029 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1030 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1033 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1034 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1036 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1037 * has one full RPC window of unstable pages, it's a good chance
1038 * to piggyback a SOFT_SYNC flag.
1039 * Please notice that the OST won't take immediate response for the
1040 * SOFT_SYNC request so active OSCs will have more chance to carry
1041 * the flag, this is reasonable. */
1042 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1043 osc_unstable_count > cli->cl_max_pages_per_rpc *
1044 cli->cl_max_rpcs_in_flight;
1048 * Return how many LRU pages in the cache of all OSC devices
1050 * \retval return # of cached LRU pages times reclaimation tendency
1051 * \retval SHRINK_STOP if it cannot do any scanning in this time
1053 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1054 struct shrink_control *sc)
1056 struct client_obd *cli;
1057 unsigned long cached = 0;
1059 spin_lock(&osc_shrink_lock);
1060 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1061 cached += atomic_long_read(&cli->cl_lru_in_list);
1062 spin_unlock(&osc_shrink_lock);
1064 return (cached * sysctl_vfs_cache_pressure) / 100;
1068 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1070 * \retval number of cached LRU pages reclaimed
1071 * \retval SHRINK_STOP if it cannot do any scanning in this time
1073 * Linux kernel will loop calling this shrinker scan routine with
1074 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1076 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1077 * to scan and try to reclaim LRU pages, just return 0 and
1078 * osc_cache_shrink_count() will report the LRU page number.
1080 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1081 struct shrink_control *sc)
1083 struct client_obd *cli;
1084 struct client_obd *stop_anchor = NULL;
1090 if (sc->nr_to_scan == 0)
1093 if (!(sc->gfp_mask & __GFP_FS))
1096 env = cl_env_get(&refcheck);
1100 spin_lock(&osc_shrink_lock);
1101 while (!list_empty(&osc_shrink_list)) {
1102 cli = list_entry(osc_shrink_list.next, struct client_obd,
1105 if (stop_anchor == NULL)
1107 else if (cli == stop_anchor)
1110 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1111 spin_unlock(&osc_shrink_lock);
1113 /* shrink no more than max_pages_per_rpc for an OSC */
1114 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1115 cli->cl_max_pages_per_rpc ?
1116 cli->cl_max_pages_per_rpc :
1117 sc->nr_to_scan - shrank, true);
1121 if (shrank >= sc->nr_to_scan)
1124 spin_lock(&osc_shrink_lock);
1126 spin_unlock(&osc_shrink_lock);
1129 cl_env_put(env, &refcheck);