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, 2016, 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
40 #include "osc_cl_internal.h"
42 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
43 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
44 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
45 struct osc_page *opg);
54 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
56 struct cl_page *page = opg->ops_cl.cpl_page;
58 LASSERT(!opg->ops_transfer_pinned);
60 lu_ref_add_atomic(&page->cp_reference, label, page);
61 opg->ops_transfer_pinned = 1;
64 static void osc_page_transfer_put(const struct lu_env *env,
67 struct cl_page *page = opg->ops_cl.cpl_page;
69 if (opg->ops_transfer_pinned) {
70 opg->ops_transfer_pinned = 0;
71 lu_ref_del(&page->cp_reference, "transfer", page);
72 cl_page_put(env, page);
77 * This is called once for every page when it is submitted for a transfer
78 * either opportunistic (osc_page_cache_add()), or immediate
79 * (osc_page_submit()).
81 static void osc_page_transfer_add(const struct lu_env *env,
82 struct osc_page *opg, enum cl_req_type crt)
84 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
86 /* ops_lru and ops_inflight share the same field, so take it from LRU
87 * first and then use it as inflight. */
88 osc_lru_use(osc_cli(obj), opg);
91 int osc_page_cache_add(const struct lu_env *env,
92 const struct cl_page_slice *slice, struct cl_io *io)
94 struct osc_page *opg = cl2osc_page(slice);
98 osc_page_transfer_get(opg, "transfer\0cache");
99 result = osc_queue_async_io(env, io, opg);
101 osc_page_transfer_put(env, opg);
103 osc_page_transfer_add(env, opg, CRT_WRITE);
108 void osc_index2policy(union ldlm_policy_data *policy,
109 const struct cl_object *obj, pgoff_t start, pgoff_t end)
111 memset(policy, 0, sizeof *policy);
112 policy->l_extent.start = cl_offset(obj, start);
113 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
116 static const char *osc_list(struct list_head *head)
118 return list_empty(head) ? "-" : "+";
121 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
123 if (opg->ops_submit_time == 0)
126 return (cfs_time_current() - opg->ops_submit_time);
129 static int osc_page_print(const struct lu_env *env,
130 const struct cl_page_slice *slice,
131 void *cookie, lu_printer_t printer)
133 struct osc_page *opg = cl2osc_page(slice);
134 struct osc_async_page *oap = &opg->ops_oap;
135 struct osc_object *obj = cl2osc(slice->cpl_obj);
136 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
138 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
139 "1< %#x %d %u %s %s > "
140 "2< %lld %u %u %#x %#x | %p %p %p > "
142 "4< %d %d %d %lu %s | %s %s %s %s > "
143 "5< %s %s %s %s | %d %s | %d %s %s>\n",
146 oap->oap_magic, oap->oap_cmd,
147 oap->oap_interrupted,
148 osc_list(&oap->oap_pending_item),
149 osc_list(&oap->oap_rpc_item),
151 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
152 oap->oap_async_flags, oap->oap_brw_flags,
153 oap->oap_request, oap->oap_cli, obj,
155 opg->ops_transfer_pinned,
156 osc_submit_duration(opg), opg->ops_srvlock,
158 cli->cl_r_in_flight, cli->cl_w_in_flight,
159 cli->cl_max_rpcs_in_flight,
161 osc_list(&cli->cl_cache_waiters),
162 osc_list(&cli->cl_loi_ready_list),
163 osc_list(&cli->cl_loi_hp_ready_list),
164 osc_list(&cli->cl_loi_write_list),
165 osc_list(&cli->cl_loi_read_list),
167 osc_list(&obj->oo_ready_item),
168 osc_list(&obj->oo_hp_ready_item),
169 osc_list(&obj->oo_write_item),
170 osc_list(&obj->oo_read_item),
171 atomic_read(&obj->oo_nr_reads),
172 osc_list(&obj->oo_reading_exts),
173 atomic_read(&obj->oo_nr_writes),
174 osc_list(&obj->oo_hp_exts),
175 osc_list(&obj->oo_urgent_exts));
178 static void osc_page_delete(const struct lu_env *env,
179 const struct cl_page_slice *slice)
181 struct osc_page *opg = cl2osc_page(slice);
182 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
186 CDEBUG(D_TRACE, "%p\n", opg);
187 osc_page_transfer_put(env, opg);
188 rc = osc_teardown_async_page(env, obj, opg);
190 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
191 "Trying to teardown failed: %d\n", rc);
195 osc_lru_del(osc_cli(obj), opg);
197 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
200 spin_lock(&obj->oo_tree_lock);
201 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
204 spin_unlock(&obj->oo_tree_lock);
206 LASSERT(ergo(value != NULL, value == opg));
212 static void osc_page_clip(const struct lu_env *env,
213 const struct cl_page_slice *slice,
216 struct osc_page *opg = cl2osc_page(slice);
217 struct osc_async_page *oap = &opg->ops_oap;
219 opg->ops_from = from;
221 spin_lock(&oap->oap_lock);
222 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
223 spin_unlock(&oap->oap_lock);
226 static int osc_page_cancel(const struct lu_env *env,
227 const struct cl_page_slice *slice)
229 struct osc_page *opg = cl2osc_page(slice);
232 /* Check if the transferring against this page
233 * is completed, or not even queued. */
234 if (opg->ops_transfer_pinned)
235 /* FIXME: may not be interrupted.. */
236 rc = osc_cancel_async_page(env, opg);
237 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
241 static int osc_page_flush(const struct lu_env *env,
242 const struct cl_page_slice *slice,
245 struct osc_page *opg = cl2osc_page(slice);
248 rc = osc_flush_async_page(env, io, opg);
252 static const struct cl_page_operations osc_page_ops = {
253 .cpo_print = osc_page_print,
254 .cpo_delete = osc_page_delete,
255 .cpo_clip = osc_page_clip,
256 .cpo_cancel = osc_page_cancel,
257 .cpo_flush = osc_page_flush
260 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
261 struct cl_page *page, pgoff_t index)
263 struct osc_object *osc = cl2osc(obj);
264 struct osc_page *opg = cl_object_page_slice(obj, page);
268 opg->ops_to = PAGE_SIZE;
270 result = osc_prep_async_page(osc, opg, page->cp_vmpage,
271 cl_offset(obj, index));
273 struct osc_io *oio = osc_env_io(env);
274 opg->ops_srvlock = osc_io_srvlock(oio);
275 cl_page_slice_add(page, &opg->ops_cl, obj, index,
278 INIT_LIST_HEAD(&opg->ops_lru);
280 /* reserve an LRU space for this page */
281 if (page->cp_type == CPT_CACHEABLE && result == 0) {
282 result = osc_lru_alloc(env, osc_cli(osc), opg);
284 spin_lock(&osc->oo_tree_lock);
285 result = radix_tree_insert(&osc->oo_tree, index, opg);
288 spin_unlock(&osc->oo_tree_lock);
289 LASSERT(result == 0);
297 * Helper function called by osc_io_submit() for every page in an immediate
298 * transfer (i.e., transferred synchronously).
300 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
301 enum cl_req_type crt, int brw_flags)
303 struct osc_async_page *oap = &opg->ops_oap;
305 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
306 "magic 0x%x\n", oap, oap->oap_magic);
307 LASSERT(oap->oap_async_flags & ASYNC_READY);
308 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
310 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
311 oap->oap_page_off = opg->ops_from;
312 oap->oap_count = opg->ops_to - opg->ops_from;
313 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
315 if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
316 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
317 oap->oap_cmd |= OBD_BRW_NOQUOTA;
320 opg->ops_submit_time = cfs_time_current();
321 osc_page_transfer_get(opg, "transfer\0imm");
322 osc_page_transfer_add(env, opg, crt);
325 /* --------------- LRU page management ------------------ */
327 /* OSC is a natural place to manage LRU pages as applications are specialized
328 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
329 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
330 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
331 * for free LRU slots - this will be very bad so the algorithm requires each
332 * OSC to free slots voluntarily to maintain a reasonable number of free slots
336 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
339 * LRU pages are freed in batch mode. OSC should at least free this
340 * number of pages to avoid running out of LRU slots.
342 static inline int lru_shrink_min(struct client_obd *cli)
344 return cli->cl_max_pages_per_rpc * 2;
348 * free this number at most otherwise it will take too long time to finsih.
350 static inline int lru_shrink_max(struct client_obd *cli)
352 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
356 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
357 * we should free slots aggressively. In this way, slots are freed in a steady
358 * step to maintain fairness among OSCs.
360 * Return how many LRU pages should be freed.
362 static int osc_cache_too_much(struct client_obd *cli)
364 struct cl_client_cache *cache = cli->cl_cache;
365 long pages = atomic_long_read(&cli->cl_lru_in_list);
366 unsigned long budget;
368 LASSERT(cache != NULL);
369 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
371 /* if it's going to run out LRU slots, we should free some, but not
372 * too much to maintain faireness among OSCs. */
373 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
375 return lru_shrink_max(cli);
376 else if (pages >= budget / 2)
377 return lru_shrink_min(cli);
379 time64_t duration = ktime_get_real_seconds();
382 /* knock out pages by duration of no IO activity */
383 duration -= cli->cl_lru_last_used;
385 * The difference shouldn't be more than 70 years
386 * so we can safely case to a long. Round to
387 * approximately 1 minute.
389 timediff = (long)(duration >> 6);
390 if (timediff > 0 && pages >= budget / timediff)
391 return lru_shrink_min(cli);
396 int lru_queue_work(const struct lu_env *env, void *data)
398 struct client_obd *cli = data;
401 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
402 count = osc_cache_too_much(cli);
404 int rc = osc_lru_shrink(env, cli, count, false);
406 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
407 cli_name(cli), rc, count);
409 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
410 ptlrpcd_queue_work(cli->cl_lru_work);
417 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
419 struct list_head lru = LIST_HEAD_INIT(lru);
420 struct osc_async_page *oap;
423 list_for_each_entry(oap, plist, oap_pending_item) {
424 struct osc_page *opg = oap2osc_page(oap);
426 if (!opg->ops_in_lru)
430 LASSERT(list_empty(&opg->ops_lru));
431 list_add(&opg->ops_lru, &lru);
435 spin_lock(&cli->cl_lru_list_lock);
436 list_splice_tail(&lru, &cli->cl_lru_list);
437 atomic_long_sub(npages, &cli->cl_lru_busy);
438 atomic_long_add(npages, &cli->cl_lru_in_list);
439 cli->cl_lru_last_used = ktime_get_real_seconds();
440 spin_unlock(&cli->cl_lru_list_lock);
442 if (waitqueue_active(&osc_lru_waitq))
443 (void)ptlrpcd_queue_work(cli->cl_lru_work);
447 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
449 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
450 list_del_init(&opg->ops_lru);
451 atomic_long_dec(&cli->cl_lru_in_list);
455 * Page is being destroyed. The page may be not in LRU list, if the transfer
456 * has never finished(error occurred).
458 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
460 if (opg->ops_in_lru) {
461 spin_lock(&cli->cl_lru_list_lock);
462 if (!list_empty(&opg->ops_lru)) {
463 __osc_lru_del(cli, opg);
465 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
466 atomic_long_dec(&cli->cl_lru_busy);
468 spin_unlock(&cli->cl_lru_list_lock);
470 atomic_long_inc(cli->cl_lru_left);
471 /* this is a great place to release more LRU pages if
472 * this osc occupies too many LRU pages and kernel is
473 * stealing one of them. */
474 if (osc_cache_too_much(cli)) {
475 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
476 (void)ptlrpcd_queue_work(cli->cl_lru_work);
478 wake_up(&osc_lru_waitq);
480 LASSERT(list_empty(&opg->ops_lru));
485 * Delete page from LRUlist for redirty.
487 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
489 /* If page is being transferred for the first time,
490 * ops_lru should be empty */
491 if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
492 spin_lock(&cli->cl_lru_list_lock);
493 __osc_lru_del(cli, opg);
494 spin_unlock(&cli->cl_lru_list_lock);
495 atomic_long_inc(&cli->cl_lru_busy);
499 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
500 struct cl_page **pvec, int max_index)
504 for (i = 0; i < max_index; i++) {
505 struct cl_page *page = pvec[i];
507 LASSERT(cl_page_is_owned(page, io));
508 cl_page_delete(env, page);
509 cl_page_discard(env, io, page);
510 cl_page_disown(env, io, page);
511 cl_page_put(env, page);
518 * Check if a cl_page can be released, i.e, it's not being used.
520 * If unstable account is turned on, bulk transfer may hold one refcount
521 * for recovery so we need to check vmpage refcount as well; otherwise,
522 * even we can destroy cl_page but the corresponding vmpage can't be reused.
524 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
526 if (cl_page_in_use_noref(page))
529 if (cli->cl_cache->ccc_unstable_check) {
530 struct page *vmpage = cl_page_vmpage(page);
532 /* vmpage have two known users: cl_page and VM page cache */
533 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
540 * Drop @target of pages from LRU at most.
542 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
543 long target, bool force)
546 struct cl_object *clobj = NULL;
547 struct cl_page **pvec;
548 struct osc_page *opg;
555 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
556 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
559 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
560 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
562 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
565 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
566 atomic_dec(&cli->cl_lru_shrinkers);
570 atomic_inc(&cli->cl_lru_shrinkers);
573 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
574 io = &osc_env_info(env)->oti_io;
576 spin_lock(&cli->cl_lru_list_lock);
578 cli->cl_lru_reclaim++;
579 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
580 while (!list_empty(&cli->cl_lru_list)) {
581 struct cl_page *page;
582 bool will_free = false;
584 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
590 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
592 page = opg->ops_cl.cpl_page;
593 if (lru_page_busy(cli, page)) {
594 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
598 LASSERT(page->cp_obj != NULL);
599 if (clobj != page->cp_obj) {
600 struct cl_object *tmp = page->cp_obj;
603 spin_unlock(&cli->cl_lru_list_lock);
606 discard_pagevec(env, io, pvec, index);
610 cl_object_put(env, clobj);
616 io->ci_ignore_layout = 1;
617 rc = cl_io_init(env, io, CIT_MISC, clobj);
619 spin_lock(&cli->cl_lru_list_lock);
628 if (cl_page_own_try(env, io, page) == 0) {
629 if (!lru_page_busy(cli, page)) {
630 /* remove it from lru list earlier to avoid
632 __osc_lru_del(cli, opg);
633 opg->ops_in_lru = 0; /* will be discarded */
638 cl_page_disown(env, io, page);
643 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
647 /* Don't discard and free the page with cl_lru_list held */
648 pvec[index++] = page;
649 if (unlikely(index == OTI_PVEC_SIZE)) {
650 spin_unlock(&cli->cl_lru_list_lock);
651 discard_pagevec(env, io, pvec, index);
654 spin_lock(&cli->cl_lru_list_lock);
657 if (++count >= target)
660 spin_unlock(&cli->cl_lru_list_lock);
663 discard_pagevec(env, io, pvec, index);
666 cl_object_put(env, clobj);
669 atomic_dec(&cli->cl_lru_shrinkers);
671 atomic_long_add(count, cli->cl_lru_left);
672 wake_up_all(&osc_lru_waitq);
674 RETURN(count > 0 ? count : rc);
678 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
679 * \@npages of LRU slots. For performance consideration, it's better to drop
680 * LRU pages in batch. Therefore, the actual number is adjusted at least
683 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
686 struct cl_client_cache *cache = cli->cl_cache;
692 LASSERT(cache != NULL);
694 env = cl_env_get(&refcheck);
698 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
699 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
700 cli_name(cli), npages);
701 rc = osc_lru_shrink(env, cli, npages, true);
703 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
704 cli_name(cli), rc, npages);
705 if (osc_cache_too_much(cli) > 0)
706 ptlrpcd_queue_work(cli->cl_lru_work);
712 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
713 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
714 atomic_long_read(&cli->cl_lru_busy), npages);
716 /* Reclaim LRU slots from other client_obd as it can't free enough
717 * from its own. This should rarely happen. */
718 spin_lock(&cache->ccc_lru_lock);
719 LASSERT(!list_empty(&cache->ccc_lru));
721 cache->ccc_lru_shrinkers++;
722 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
724 max_scans = atomic_read(&cache->ccc_users) - 2;
725 while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
726 cli = list_entry(cache->ccc_lru.next, struct client_obd,
729 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
731 atomic_long_read(&cli->cl_lru_in_list),
732 atomic_long_read(&cli->cl_lru_busy));
734 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
735 if (osc_cache_too_much(cli) > 0) {
736 spin_unlock(&cache->ccc_lru_lock);
738 rc = osc_lru_shrink(env, cli, npages, true);
739 spin_lock(&cache->ccc_lru_lock);
746 spin_unlock(&cache->ccc_lru_lock);
749 cl_env_put(env, &refcheck);
750 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
751 cli_name(cli), cli, rc);
756 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
758 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
759 * Only in the case that the LRU slots are in extreme shortage, it should
760 * have reserved enough slots for an IO.
762 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
763 struct osc_page *opg)
765 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
766 struct osc_io *oio = osc_env_io(env);
770 if (cli->cl_cache == NULL) /* shall not be in LRU */
773 if (oio->oi_lru_reserved > 0) {
774 --oio->oi_lru_reserved;
778 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
779 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
780 /* run out of LRU spaces, try to drop some by itself */
781 rc = osc_lru_reclaim(cli, 1);
788 rc = l_wait_event(osc_lru_waitq,
789 atomic_long_read(cli->cl_lru_left) > 0,
797 atomic_long_inc(&cli->cl_lru_busy);
806 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
808 * The benefit of doing this is to reduce contention against atomic counter
809 * cl_lru_left by changing it from per-page access to per-IO access.
811 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
813 unsigned long reserved = 0;
814 unsigned long max_pages;
817 /* reserve a full RPC window at most to avoid that a thread accidentally
818 * consumes too many LRU slots */
819 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
820 if (npages > max_pages)
823 c = atomic_long_read(cli->cl_lru_left);
824 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
825 c = atomic_long_read(cli->cl_lru_left);
826 while (c >= npages) {
827 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
831 c = atomic_long_read(cli->cl_lru_left);
833 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
834 /* If there aren't enough pages in the per-OSC LRU then
835 * wake up the LRU thread to try and clear out space, so
836 * we don't block if pages are being dirtied quickly. */
837 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
838 cli_name(cli), atomic_long_read(cli->cl_lru_left),
840 (void)ptlrpcd_queue_work(cli->cl_lru_work);
847 * osc_lru_unreserve() is called to unreserve LRU slots.
849 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
850 * reasons such as page already existing or I/O error. Those reserved slots
851 * should be freed by calling this function.
853 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
855 atomic_long_add(npages, cli->cl_lru_left);
856 wake_up_all(&osc_lru_waitq);
860 * Atomic operations are expensive. We accumulate the accounting for the
861 * same page zone to get better performance.
862 * In practice this can work pretty good because the pages in the same RPC
863 * are likely from the same page zone.
865 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
868 int page_count = desc->bd_iov_count;
873 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
875 for (i = 0; i < page_count; i++) {
876 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
878 if (likely(pz == zone)) {
884 mod_zone_page_state(zone, NR_UNSTABLE_NFS,
892 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
895 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
897 unstable_page_accounting(desc, 1);
900 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
902 unstable_page_accounting(desc, -1);
906 * Performs "unstable" page accounting. This function balances the
907 * increment operations performed in osc_inc_unstable_pages. It is
908 * registered as the RPC request callback, and is executed when the
909 * bulk RPC is committed on the server. Thus at this point, the pages
910 * involved in the bulk transfer are no longer considered unstable.
912 * If this function is called, the request should have been committed
913 * or req:rq_unstable must have been set; it implies that the unstable
914 * statistic have been added.
916 void osc_dec_unstable_pages(struct ptlrpc_request *req)
918 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
919 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
920 int page_count = desc->bd_iov_count;
923 LASSERT(page_count >= 0);
924 dec_unstable_page_accounting(desc);
926 unstable_count = atomic_long_sub_return(page_count,
927 &cli->cl_unstable_count);
928 LASSERT(unstable_count >= 0);
930 unstable_count = atomic_long_sub_return(page_count,
931 &cli->cl_cache->ccc_unstable_nr);
932 LASSERT(unstable_count >= 0);
933 if (unstable_count == 0)
934 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
936 if (waitqueue_active(&osc_lru_waitq))
937 (void)ptlrpcd_queue_work(cli->cl_lru_work);
941 * "unstable" page accounting. See: osc_dec_unstable_pages.
943 void osc_inc_unstable_pages(struct ptlrpc_request *req)
945 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
946 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
947 long page_count = desc->bd_iov_count;
949 /* No unstable page tracking */
950 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
953 add_unstable_page_accounting(desc);
954 atomic_long_add(page_count, &cli->cl_unstable_count);
955 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
957 /* If the request has already been committed (i.e. brw_commit
958 * called via rq_commit_cb), we need to undo the unstable page
959 * increments we just performed because rq_commit_cb wont be
961 spin_lock(&req->rq_lock);
962 if (unlikely(req->rq_committed)) {
963 spin_unlock(&req->rq_lock);
965 osc_dec_unstable_pages(req);
967 req->rq_unstable = 1;
968 spin_unlock(&req->rq_lock);
973 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
974 * This function will be called by every BRW RPC so it's critical
975 * to make this function fast.
977 bool osc_over_unstable_soft_limit(struct client_obd *cli)
979 long unstable_nr, osc_unstable_count;
981 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
982 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
985 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
986 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
989 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
990 cli_name(cli), cli, unstable_nr, osc_unstable_count);
992 /* If the LRU slots are in shortage - 25% remaining AND this OSC
993 * has one full RPC window of unstable pages, it's a good chance
994 * to piggyback a SOFT_SYNC flag.
995 * Please notice that the OST won't take immediate response for the
996 * SOFT_SYNC request so active OSCs will have more chance to carry
997 * the flag, this is reasonable. */
998 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
999 osc_unstable_count > cli->cl_max_pages_per_rpc *
1000 cli->cl_max_rpcs_in_flight;
1004 * Return how many LRU pages in the cache of all OSC devices
1006 * \retval return # of cached LRU pages times reclaimation tendency
1007 * \retval SHRINK_STOP if it cannot do any scanning in this time
1009 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1010 struct shrink_control *sc)
1012 struct client_obd *cli;
1013 unsigned long cached = 0;
1015 spin_lock(&osc_shrink_lock);
1016 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1017 cached += atomic_long_read(&cli->cl_lru_in_list);
1018 spin_unlock(&osc_shrink_lock);
1020 return (cached * sysctl_vfs_cache_pressure) / 100;
1024 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1026 * \retval number of cached LRU pages reclaimed
1027 * \retval SHRINK_STOP if it cannot do any scanning in this time
1029 * Linux kernel will loop calling this shrinker scan routine with
1030 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1032 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1033 * to scan and try to reclaim LRU pages, just return 0 and
1034 * osc_cache_shrink_count() will report the LRU page number.
1036 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1037 struct shrink_control *sc)
1039 struct client_obd *cli;
1040 struct client_obd *stop_anchor = NULL;
1046 if (sc->nr_to_scan == 0)
1049 if (!(sc->gfp_mask & __GFP_FS))
1052 env = cl_env_get(&refcheck);
1056 spin_lock(&osc_shrink_lock);
1057 while (!list_empty(&osc_shrink_list)) {
1058 cli = list_entry(osc_shrink_list.next, struct client_obd,
1061 if (stop_anchor == NULL)
1063 else if (cli == stop_anchor)
1066 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1067 spin_unlock(&osc_shrink_lock);
1069 /* shrink no more than max_pages_per_rpc for an OSC */
1070 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1071 cli->cl_max_pages_per_rpc ?
1072 cli->cl_max_pages_per_rpc :
1073 sc->nr_to_scan - shrank, true);
1077 if (shrank >= sc->nr_to_scan)
1080 spin_lock(&osc_shrink_lock);
1082 spin_unlock(&osc_shrink_lock);
1085 cl_env_put(env, &refcheck);