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,
91 const struct cl_page_slice *slice, struct cl_io *io)
93 struct osc_page *opg = cl2osc_page(slice);
97 osc_page_transfer_get(opg, "transfer\0cache");
98 result = osc_queue_async_io(env, io, opg);
100 osc_page_transfer_put(env, opg);
102 osc_page_transfer_add(env, opg, CRT_WRITE);
107 void osc_index2policy(union ldlm_policy_data *policy,
108 const struct cl_object *obj, pgoff_t start, pgoff_t end)
110 memset(policy, 0, sizeof *policy);
111 policy->l_extent.start = cl_offset(obj, start);
112 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
115 static const char *osc_list(struct list_head *head)
117 return list_empty(head) ? "-" : "+";
120 static inline s64 osc_submit_duration(struct osc_page *opg)
122 if (ktime_to_ns(opg->ops_submit_time) == 0)
125 return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
128 static int osc_page_print(const struct lu_env *env,
129 const struct cl_page_slice *slice,
130 void *cookie, lu_printer_t printer)
132 struct osc_page *opg = cl2osc_page(slice);
133 struct osc_async_page *oap = &opg->ops_oap;
134 struct osc_object *obj = cl2osc(slice->cpl_obj);
135 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
137 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
138 "1< %#x %d %u %s %s > "
139 "2< %lld %u %u %#x %#x | %p %p %p > "
141 "4< %d %d %d %lu %s | %s %s %s %s > "
142 "5< %s %s %s %s | %d %s | %d %s %s>\n",
145 oap->oap_magic, oap->oap_cmd,
146 oap->oap_interrupted,
147 osc_list(&oap->oap_pending_item),
148 osc_list(&oap->oap_rpc_item),
150 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
151 oap->oap_async_flags, oap->oap_brw_flags,
152 oap->oap_request, oap->oap_cli, obj,
154 opg->ops_transfer_pinned,
155 osc_submit_duration(opg), opg->ops_srvlock,
157 cli->cl_r_in_flight, cli->cl_w_in_flight,
158 cli->cl_max_rpcs_in_flight,
160 osc_list(&cli->cl_cache_waiters),
161 osc_list(&cli->cl_loi_ready_list),
162 osc_list(&cli->cl_loi_hp_ready_list),
163 osc_list(&cli->cl_loi_write_list),
164 osc_list(&cli->cl_loi_read_list),
166 osc_list(&obj->oo_ready_item),
167 osc_list(&obj->oo_hp_ready_item),
168 osc_list(&obj->oo_write_item),
169 osc_list(&obj->oo_read_item),
170 atomic_read(&obj->oo_nr_reads),
171 osc_list(&obj->oo_reading_exts),
172 atomic_read(&obj->oo_nr_writes),
173 osc_list(&obj->oo_hp_exts),
174 osc_list(&obj->oo_urgent_exts));
177 static void osc_page_delete(const struct lu_env *env,
178 const struct cl_page_slice *slice)
180 struct osc_page *opg = cl2osc_page(slice);
181 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
185 CDEBUG(D_TRACE, "%p\n", opg);
186 osc_page_transfer_put(env, opg);
187 rc = osc_teardown_async_page(env, obj, opg);
189 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
190 "Trying to teardown failed: %d\n", rc);
194 osc_lru_del(osc_cli(obj), opg);
196 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
199 spin_lock(&obj->oo_tree_lock);
200 if (opg->ops_intree) {
201 value = radix_tree_delete(&obj->oo_tree,
208 spin_unlock(&obj->oo_tree_lock);
210 LASSERT(ergo(value != NULL, value == opg));
216 static void osc_page_clip(const struct lu_env *env,
217 const struct cl_page_slice *slice,
220 struct osc_page *opg = cl2osc_page(slice);
221 struct osc_async_page *oap = &opg->ops_oap;
223 opg->ops_from = from;
225 spin_lock(&oap->oap_lock);
226 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
227 spin_unlock(&oap->oap_lock);
230 static int osc_page_cancel(const struct lu_env *env,
231 const struct cl_page_slice *slice)
233 struct osc_page *opg = cl2osc_page(slice);
236 /* Check if the transferring against this page
237 * is completed, or not even queued. */
238 if (opg->ops_transfer_pinned)
239 /* FIXME: may not be interrupted.. */
240 rc = osc_cancel_async_page(env, opg);
241 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
245 static int osc_page_flush(const struct lu_env *env,
246 const struct cl_page_slice *slice,
249 struct osc_page *opg = cl2osc_page(slice);
252 rc = osc_flush_async_page(env, io, opg);
256 static void osc_page_touch(const struct lu_env *env,
257 const struct cl_page_slice *slice, size_t to)
259 struct osc_page *opg = cl2osc_page(slice);
260 struct cl_object *obj = opg->ops_cl.cpl_obj;
262 osc_page_touch_at(env, obj, osc_index(opg), to);
265 static const struct cl_page_operations osc_page_ops = {
266 .cpo_print = osc_page_print,
267 .cpo_delete = osc_page_delete,
268 .cpo_clip = osc_page_clip,
269 .cpo_cancel = osc_page_cancel,
270 .cpo_flush = osc_page_flush,
271 .cpo_page_touch = osc_page_touch,
274 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
275 struct cl_page *page, pgoff_t index)
277 struct osc_object *osc = cl2osc(obj);
278 struct osc_page *opg = cl_object_page_slice(obj, page);
279 struct osc_io *oio = osc_env_io(env);
283 opg->ops_to = PAGE_SIZE;
285 INIT_LIST_HEAD(&opg->ops_lru);
287 result = osc_prep_async_page(osc, opg, page->cp_vmpage,
288 cl_offset(obj, index));
292 opg->ops_srvlock = osc_io_srvlock(oio);
293 cl_page_slice_add(page, &opg->ops_cl, obj, index,
297 /* reserve an LRU space for this page */
298 if (page->cp_type == CPT_CACHEABLE) {
299 result = osc_lru_alloc(env, osc_cli(osc), opg);
301 result = radix_tree_preload(GFP_NOFS);
303 spin_lock(&osc->oo_tree_lock);
304 result = radix_tree_insert(&osc->oo_tree,
310 spin_unlock(&osc->oo_tree_lock);
312 radix_tree_preload_end();
319 EXPORT_SYMBOL(osc_page_init);
322 * Helper function called by osc_io_submit() for every page in an immediate
323 * transfer (i.e., transferred synchronously).
325 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
326 enum cl_req_type crt, int brw_flags)
328 struct osc_io *oio = osc_env_io(env);
329 struct osc_async_page *oap = &opg->ops_oap;
331 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
332 "magic 0x%x\n", oap, oap->oap_magic);
333 LASSERT(oap->oap_async_flags & ASYNC_READY);
334 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
336 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
337 oap->oap_page_off = opg->ops_from;
338 oap->oap_count = opg->ops_to - opg->ops_from;
339 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
341 if (oio->oi_cap_sys_resource) {
342 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
343 oap->oap_cmd |= OBD_BRW_NOQUOTA;
346 opg->ops_submit_time = ktime_get();
347 osc_page_transfer_get(opg, "transfer\0imm");
348 osc_page_transfer_add(env, opg, crt);
351 /* --------------- LRU page management ------------------ */
353 /* OSC is a natural place to manage LRU pages as applications are specialized
354 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
355 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
356 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
357 * for free LRU slots - this will be very bad so the algorithm requires each
358 * OSC to free slots voluntarily to maintain a reasonable number of free slots
362 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
365 * LRU pages are freed in batch mode. OSC should at least free this
366 * number of pages to avoid running out of LRU slots.
368 static inline int lru_shrink_min(struct client_obd *cli)
370 return cli->cl_max_pages_per_rpc * 2;
374 * free this number at most otherwise it will take too long time to finsih.
376 static inline int lru_shrink_max(struct client_obd *cli)
378 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
382 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
383 * we should free slots aggressively. In this way, slots are freed in a steady
384 * step to maintain fairness among OSCs.
386 * Return how many LRU pages should be freed.
388 static int osc_cache_too_much(struct client_obd *cli)
390 struct cl_client_cache *cache = cli->cl_cache;
391 long pages = atomic_long_read(&cli->cl_lru_in_list);
392 unsigned long budget;
394 LASSERT(cache != NULL);
395 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
397 /* if it's going to run out LRU slots, we should free some, but not
398 * too much to maintain faireness among OSCs. */
399 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
401 return lru_shrink_max(cli);
402 else if (pages >= budget / 2)
403 return lru_shrink_min(cli);
405 time64_t duration = ktime_get_real_seconds();
408 /* knock out pages by duration of no IO activity */
409 duration -= cli->cl_lru_last_used;
411 * The difference shouldn't be more than 70 years
412 * so we can safely case to a long. Round to
413 * approximately 1 minute.
415 timediff = (long)(duration >> 6);
416 if (timediff > 0 && pages >= budget / timediff)
417 return lru_shrink_min(cli);
422 int lru_queue_work(const struct lu_env *env, void *data)
424 struct client_obd *cli = data;
427 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
428 count = osc_cache_too_much(cli);
430 int rc = osc_lru_shrink(env, cli, count, false);
432 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
433 cli_name(cli), rc, count);
435 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
436 ptlrpcd_queue_work(cli->cl_lru_work);
443 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
445 struct list_head lru = LIST_HEAD_INIT(lru);
446 struct osc_async_page *oap;
449 list_for_each_entry(oap, plist, oap_pending_item) {
450 struct osc_page *opg = oap2osc_page(oap);
452 if (!opg->ops_in_lru)
456 LASSERT(list_empty(&opg->ops_lru));
457 list_add(&opg->ops_lru, &lru);
461 spin_lock(&cli->cl_lru_list_lock);
462 list_splice_tail(&lru, &cli->cl_lru_list);
463 atomic_long_sub(npages, &cli->cl_lru_busy);
464 atomic_long_add(npages, &cli->cl_lru_in_list);
465 cli->cl_lru_last_used = ktime_get_real_seconds();
466 spin_unlock(&cli->cl_lru_list_lock);
468 if (waitqueue_active(&osc_lru_waitq))
469 (void)ptlrpcd_queue_work(cli->cl_lru_work);
473 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
475 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
476 list_del_init(&opg->ops_lru);
477 atomic_long_dec(&cli->cl_lru_in_list);
481 * Page is being destroyed. The page may be not in LRU list, if the transfer
482 * has never finished(error occurred).
484 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
486 if (opg->ops_in_lru) {
487 spin_lock(&cli->cl_lru_list_lock);
488 if (!list_empty(&opg->ops_lru)) {
489 __osc_lru_del(cli, opg);
491 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
492 atomic_long_dec(&cli->cl_lru_busy);
494 spin_unlock(&cli->cl_lru_list_lock);
496 atomic_long_inc(cli->cl_lru_left);
497 /* this is a great place to release more LRU pages if
498 * this osc occupies too many LRU pages and kernel is
499 * stealing one of them. */
500 if (osc_cache_too_much(cli)) {
501 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
502 (void)ptlrpcd_queue_work(cli->cl_lru_work);
504 wake_up(&osc_lru_waitq);
506 LASSERT(list_empty(&opg->ops_lru));
511 * Delete page from LRU list for redirty.
513 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
515 /* If page is being transferred for the first time,
516 * ops_lru should be empty */
517 if (opg->ops_in_lru) {
518 if (list_empty(&opg->ops_lru))
520 spin_lock(&cli->cl_lru_list_lock);
521 if (!list_empty(&opg->ops_lru)) {
522 __osc_lru_del(cli, opg);
523 atomic_long_inc(&cli->cl_lru_busy);
525 spin_unlock(&cli->cl_lru_list_lock);
529 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
530 struct cl_page **pvec, int max_index)
532 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
535 ll_pagevec_init(pagevec, 0);
536 for (i = 0; i < max_index; i++) {
537 struct cl_page *page = pvec[i];
539 LASSERT(cl_page_is_owned(page, io));
540 cl_page_delete(env, page);
541 cl_page_discard(env, io, page);
542 cl_page_disown(env, io, page);
543 cl_pagevec_put(env, page, pagevec);
547 pagevec_release(pagevec);
551 * Check if a cl_page can be released, i.e, it's not being used.
553 * If unstable account is turned on, bulk transfer may hold one refcount
554 * for recovery so we need to check vmpage refcount as well; otherwise,
555 * even we can destroy cl_page but the corresponding vmpage can't be reused.
557 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
559 if (cl_page_in_use_noref(page))
562 if (cli->cl_cache->ccc_unstable_check) {
563 struct page *vmpage = cl_page_vmpage(page);
565 /* vmpage have two known users: cl_page and VM page cache */
566 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
573 * Drop @target of pages from LRU at most.
575 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
576 long target, bool force)
579 struct cl_object *clobj = NULL;
580 struct cl_page **pvec;
581 struct osc_page *opg;
588 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
589 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
592 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
593 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
595 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
598 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
599 atomic_dec(&cli->cl_lru_shrinkers);
603 atomic_inc(&cli->cl_lru_shrinkers);
606 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
607 io = osc_env_thread_io(env);
609 spin_lock(&cli->cl_lru_list_lock);
611 cli->cl_lru_reclaim++;
612 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
613 while (!list_empty(&cli->cl_lru_list)) {
614 struct cl_page *page;
615 bool will_free = false;
617 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
623 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
625 page = opg->ops_cl.cpl_page;
626 if (lru_page_busy(cli, page)) {
627 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
631 LASSERT(page->cp_obj != NULL);
632 if (clobj != page->cp_obj) {
633 struct cl_object *tmp = page->cp_obj;
636 spin_unlock(&cli->cl_lru_list_lock);
639 discard_pagevec(env, io, pvec, index);
643 cl_object_put(env, clobj);
649 io->ci_ignore_layout = 1;
650 rc = cl_io_init(env, io, CIT_MISC, clobj);
652 spin_lock(&cli->cl_lru_list_lock);
661 if (cl_page_own_try(env, io, page) == 0) {
662 if (!lru_page_busy(cli, page)) {
663 /* remove it from lru list earlier to avoid
665 __osc_lru_del(cli, opg);
666 opg->ops_in_lru = 0; /* will be discarded */
671 cl_page_disown(env, io, page);
676 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
680 /* Don't discard and free the page with cl_lru_list held */
681 pvec[index++] = page;
682 if (unlikely(index == OTI_PVEC_SIZE)) {
683 spin_unlock(&cli->cl_lru_list_lock);
684 discard_pagevec(env, io, pvec, index);
687 spin_lock(&cli->cl_lru_list_lock);
690 if (++count >= target)
693 spin_unlock(&cli->cl_lru_list_lock);
696 discard_pagevec(env, io, pvec, index);
699 cl_object_put(env, clobj);
702 atomic_dec(&cli->cl_lru_shrinkers);
704 atomic_long_add(count, cli->cl_lru_left);
705 wake_up_all(&osc_lru_waitq);
707 RETURN(count > 0 ? count : rc);
709 EXPORT_SYMBOL(osc_lru_shrink);
712 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
713 * \@npages of LRU slots. For performance consideration, it's better to drop
714 * LRU pages in batch. Therefore, the actual number is adjusted at least
717 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
720 struct cl_client_cache *cache = cli->cl_cache;
726 LASSERT(cache != NULL);
728 env = cl_env_get(&refcheck);
732 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
733 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
734 cli_name(cli), npages);
735 rc = osc_lru_shrink(env, cli, npages, true);
737 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
738 cli_name(cli), rc, npages);
739 if (osc_cache_too_much(cli) > 0)
740 ptlrpcd_queue_work(cli->cl_lru_work);
746 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
747 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
748 atomic_long_read(&cli->cl_lru_busy), npages);
750 /* Reclaim LRU slots from other client_obd as it can't free enough
751 * from its own. This should rarely happen. */
752 spin_lock(&cache->ccc_lru_lock);
753 LASSERT(!list_empty(&cache->ccc_lru));
755 cache->ccc_lru_shrinkers++;
756 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
758 max_scans = atomic_read(&cache->ccc_users) - 2;
759 while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
760 cli = list_entry(cache->ccc_lru.next, struct client_obd,
763 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
765 atomic_long_read(&cli->cl_lru_in_list),
766 atomic_long_read(&cli->cl_lru_busy));
768 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
769 if (osc_cache_too_much(cli) > 0) {
770 spin_unlock(&cache->ccc_lru_lock);
772 rc = osc_lru_shrink(env, cli, npages, true);
773 spin_lock(&cache->ccc_lru_lock);
780 spin_unlock(&cache->ccc_lru_lock);
783 cl_env_put(env, &refcheck);
784 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
785 cli_name(cli), cli, rc);
790 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
792 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
793 * Only in the case that the LRU slots are in extreme shortage, it should
794 * have reserved enough slots for an IO.
796 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
797 struct osc_page *opg)
799 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
800 struct osc_io *oio = osc_env_io(env);
805 if (cli->cl_cache == NULL) /* shall not be in LRU */
808 if (oio->oi_lru_reserved > 0) {
809 --oio->oi_lru_reserved;
813 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
814 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
815 /* run out of LRU spaces, try to drop some by itself */
816 rc = osc_lru_reclaim(cli, 1);
823 rc = l_wait_event(osc_lru_waitq,
824 atomic_long_read(cli->cl_lru_left) > 0,
832 atomic_long_inc(&cli->cl_lru_busy);
841 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
843 * The benefit of doing this is to reduce contention against atomic counter
844 * cl_lru_left by changing it from per-page access to per-IO access.
846 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
848 unsigned long reserved = 0;
849 unsigned long max_pages;
852 /* reserve a full RPC window at most to avoid that a thread accidentally
853 * consumes too many LRU slots */
854 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
855 if (npages > max_pages)
858 c = atomic_long_read(cli->cl_lru_left);
859 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
860 c = atomic_long_read(cli->cl_lru_left);
861 while (c >= npages) {
862 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
866 c = atomic_long_read(cli->cl_lru_left);
868 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
869 /* If there aren't enough pages in the per-OSC LRU then
870 * wake up the LRU thread to try and clear out space, so
871 * we don't block if pages are being dirtied quickly. */
872 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
873 cli_name(cli), atomic_long_read(cli->cl_lru_left),
875 (void)ptlrpcd_queue_work(cli->cl_lru_work);
882 * osc_lru_unreserve() is called to unreserve LRU slots.
884 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
885 * reasons such as page already existing or I/O error. Those reserved slots
886 * should be freed by calling this function.
888 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
890 atomic_long_add(npages, cli->cl_lru_left);
891 wake_up_all(&osc_lru_waitq);
895 * Atomic operations are expensive. We accumulate the accounting for the
896 * same page zone to get better performance.
897 * In practice this can work pretty good because the pages in the same RPC
898 * are likely from the same page zone.
900 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
901 struct osc_brw_async_args *aa,
910 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
911 page_count = desc->bd_iov_count;
913 page_count = aa->aa_page_count;
916 for (i = 0; i < page_count; i++) {
919 pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
921 pz = page_zone(aa->aa_ppga[i]->pg);
923 if (likely(pz == zone)) {
929 mod_zone_page_state(zone, NR_UNSTABLE_NFS,
937 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
940 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
941 struct osc_brw_async_args *aa)
943 unstable_page_accounting(desc, aa, 1);
946 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
947 struct osc_brw_async_args *aa)
949 unstable_page_accounting(desc, aa, -1);
953 * Performs "unstable" page accounting. This function balances the
954 * increment operations performed in osc_inc_unstable_pages. It is
955 * registered as the RPC request callback, and is executed when the
956 * bulk RPC is committed on the server. Thus at this point, the pages
957 * involved in the bulk transfer are no longer considered unstable.
959 * If this function is called, the request should have been committed
960 * or req:rq_unstable must have been set; it implies that the unstable
961 * statistic have been added.
963 void osc_dec_unstable_pages(struct ptlrpc_request *req)
965 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
966 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
967 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
972 page_count = desc->bd_iov_count;
974 page_count = aa->aa_page_count;
976 LASSERT(page_count >= 0);
978 dec_unstable_page_accounting(desc, aa);
980 unstable_count = atomic_long_sub_return(page_count,
981 &cli->cl_unstable_count);
982 LASSERT(unstable_count >= 0);
984 unstable_count = atomic_long_sub_return(page_count,
985 &cli->cl_cache->ccc_unstable_nr);
986 LASSERT(unstable_count >= 0);
987 if (unstable_count == 0)
988 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
990 if (waitqueue_active(&osc_lru_waitq))
991 (void)ptlrpcd_queue_work(cli->cl_lru_work);
995 * "unstable" page accounting. See: osc_dec_unstable_pages.
997 void osc_inc_unstable_pages(struct ptlrpc_request *req)
999 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1000 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1001 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1004 /* No unstable page tracking */
1005 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1009 page_count = desc->bd_iov_count;
1011 page_count = aa->aa_page_count;
1013 add_unstable_page_accounting(desc, aa);
1014 atomic_long_add(page_count, &cli->cl_unstable_count);
1015 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1017 /* If the request has already been committed (i.e. brw_commit
1018 * called via rq_commit_cb), we need to undo the unstable page
1019 * increments we just performed because rq_commit_cb wont be
1021 spin_lock(&req->rq_lock);
1022 if (unlikely(req->rq_committed)) {
1023 spin_unlock(&req->rq_lock);
1025 osc_dec_unstable_pages(req);
1027 req->rq_unstable = 1;
1028 spin_unlock(&req->rq_lock);
1033 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1034 * This function will be called by every BRW RPC so it's critical
1035 * to make this function fast.
1037 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1039 long unstable_nr, osc_unstable_count;
1041 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1042 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1045 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1046 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1049 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1050 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1052 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1053 * has one full RPC window of unstable pages, it's a good chance
1054 * to piggyback a SOFT_SYNC flag.
1055 * Please notice that the OST won't take immediate response for the
1056 * SOFT_SYNC request so active OSCs will have more chance to carry
1057 * the flag, this is reasonable. */
1058 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1059 osc_unstable_count > cli->cl_max_pages_per_rpc *
1060 cli->cl_max_rpcs_in_flight;
1064 * Return how many LRU pages in the cache of all OSC devices
1066 * \retval return # of cached LRU pages times reclaimation tendency
1067 * \retval SHRINK_STOP if it cannot do any scanning in this time
1069 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1070 struct shrink_control *sc)
1072 struct client_obd *cli;
1073 unsigned long cached = 0;
1075 spin_lock(&osc_shrink_lock);
1076 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1077 cached += atomic_long_read(&cli->cl_lru_in_list);
1078 spin_unlock(&osc_shrink_lock);
1080 return (cached * sysctl_vfs_cache_pressure) / 100;
1084 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1086 * \retval number of cached LRU pages reclaimed
1087 * \retval SHRINK_STOP if it cannot do any scanning in this time
1089 * Linux kernel will loop calling this shrinker scan routine with
1090 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1092 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1093 * to scan and try to reclaim LRU pages, just return 0 and
1094 * osc_cache_shrink_count() will report the LRU page number.
1096 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1097 struct shrink_control *sc)
1099 struct client_obd *cli;
1100 struct client_obd *stop_anchor = NULL;
1106 if (sc->nr_to_scan == 0)
1109 if (!(sc->gfp_mask & __GFP_FS))
1112 env = cl_env_get(&refcheck);
1116 spin_lock(&osc_shrink_lock);
1117 while (!list_empty(&osc_shrink_list)) {
1118 cli = list_entry(osc_shrink_list.next, struct client_obd,
1121 if (stop_anchor == NULL)
1123 else if (cli == stop_anchor)
1126 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1127 spin_unlock(&osc_shrink_lock);
1129 /* shrink no more than max_pages_per_rpc for an OSC */
1130 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1131 cli->cl_max_pages_per_rpc ?
1132 cli->cl_max_pages_per_rpc :
1133 sc->nr_to_scan - shrank, true);
1137 if (shrank >= sc->nr_to_scan)
1140 spin_lock(&osc_shrink_lock);
1142 spin_unlock(&osc_shrink_lock);
1145 cl_env_put(env, &refcheck);