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/
31 * Implementation of cl_page for OSC layer.
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
37 #define DEBUG_SUBSYSTEM S_OSC
38 #include <lustre_osc.h>
40 #include "osc_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 = osc_page_object(opg);
86 osc_lru_use(osc_cli(obj), opg);
89 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
90 struct cl_io *io, cl_commit_cbt cb)
95 osc_page_transfer_get(opg, "transfer\0cache");
96 result = osc_queue_async_io(env, io, opg, cb);
98 osc_page_transfer_put(env, opg);
100 osc_page_transfer_add(env, opg, CRT_WRITE);
105 void osc_index2policy(union ldlm_policy_data *policy,
106 const struct cl_object *obj, pgoff_t start, pgoff_t end)
108 memset(policy, 0, sizeof *policy);
109 policy->l_extent.start = start << PAGE_SHIFT;
110 policy->l_extent.end = ((end + 1) << PAGE_SHIFT) - 1;
113 static int osc_page_print(const struct lu_env *env,
114 const struct cl_page_slice *slice,
115 void *cookie, lu_printer_t printer)
117 struct osc_page *opg = cl2osc_page(slice);
118 struct osc_async_page *oap = &opg->ops_oap;
119 struct osc_object *obj = osc_page_object(opg);
120 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
122 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
124 "2< %lld %u %u %#x %#x | %p %p %p > "
126 "4< %d %d %d %lu %c | %c %c %c %c > "
127 "5< %c %c %c %c | %d %c | %d %c %c>\n",
131 list_empty_marker(&oap->oap_pending_item),
132 list_empty_marker(&oap->oap_rpc_item),
134 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
135 oap->oap_async_flags, oap->oap_brw_flags,
136 oap->oap_request, cli, obj,
138 opg->ops_transfer_pinned,
141 cli->cl_r_in_flight, cli->cl_w_in_flight,
142 cli->cl_max_rpcs_in_flight,
144 waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
145 list_empty_marker(&cli->cl_loi_ready_list),
146 list_empty_marker(&cli->cl_loi_hp_ready_list),
147 list_empty_marker(&cli->cl_loi_write_list),
148 list_empty_marker(&cli->cl_loi_read_list),
150 list_empty_marker(&obj->oo_ready_item),
151 list_empty_marker(&obj->oo_hp_ready_item),
152 list_empty_marker(&obj->oo_write_item),
153 list_empty_marker(&obj->oo_read_item),
154 atomic_read(&obj->oo_nr_reads),
155 list_empty_marker(&obj->oo_reading_exts),
156 atomic_read(&obj->oo_nr_writes),
157 list_empty_marker(&obj->oo_hp_exts),
158 list_empty_marker(&obj->oo_urgent_exts));
161 static void osc_page_delete(const struct lu_env *env,
162 const struct cl_page_slice *slice)
164 struct osc_page *opg = cl2osc_page(slice);
165 struct osc_object *obj = osc_page_object(opg);
169 CDEBUG(D_TRACE, "%p\n", opg);
170 osc_page_transfer_put(env, opg);
171 rc = osc_teardown_async_page(env, obj, opg);
173 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
174 "Trying to teardown failed: %d\n", rc);
178 osc_lru_del(osc_cli(obj), opg);
180 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
183 spin_lock(&obj->oo_tree_lock);
184 if (opg->ops_intree) {
185 value = radix_tree_delete(&obj->oo_tree,
192 spin_unlock(&obj->oo_tree_lock);
194 LASSERT(ergo(value != NULL, value == opg));
200 static void osc_page_clip(const struct lu_env *env,
201 const struct cl_page_slice *slice,
204 struct osc_page *opg = cl2osc_page(slice);
205 struct osc_async_page *oap = &opg->ops_oap;
207 opg->ops_from = from;
208 /* argument @to is exclusive, but @ops_to is inclusive */
209 opg->ops_to = to - 1;
210 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
213 static int osc_page_flush(const struct lu_env *env,
214 const struct cl_page_slice *slice,
217 struct osc_page *opg = cl2osc_page(slice);
220 rc = osc_flush_async_page(env, io, opg);
224 static void osc_page_touch(const struct lu_env *env,
225 const struct cl_page_slice *slice, size_t to)
227 struct osc_page *opg = cl2osc_page(slice);
228 struct cl_object *obj = osc2cl(osc_page_object(opg));
230 osc_page_touch_at(env, obj, osc_index(opg), to);
233 static const struct cl_page_operations osc_page_ops = {
234 .cpo_print = osc_page_print,
235 .cpo_delete = osc_page_delete,
236 .cpo_clip = osc_page_clip,
237 .cpo_flush = osc_page_flush,
238 .cpo_page_touch = osc_page_touch,
241 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
242 struct cl_page *cl_page, pgoff_t index)
244 struct osc_object *osc = cl2osc(obj);
245 struct osc_page *opg = cl_object_page_slice(obj, cl_page);
246 struct osc_io *oio = osc_env_io(env);
250 opg->ops_to = PAGE_SIZE - 1;
252 INIT_LIST_HEAD(&opg->ops_lru);
254 result = osc_prep_async_page(osc, opg, cl_page, index << PAGE_SHIFT);
258 opg->ops_srvlock = osc_io_srvlock(oio);
259 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
261 /* reserve an LRU space for this page */
262 if (cl_page->cp_type == CPT_CACHEABLE) {
263 result = osc_lru_alloc(env, osc_cli(osc), opg);
265 result = radix_tree_preload(GFP_NOFS);
267 spin_lock(&osc->oo_tree_lock);
268 result = radix_tree_insert(&osc->oo_tree,
274 spin_unlock(&osc->oo_tree_lock);
276 radix_tree_preload_end();
283 EXPORT_SYMBOL(osc_page_init);
286 * Helper function called by osc_io_submit() for every page in an immediate
287 * transfer (i.e., transferred synchronously).
289 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
290 enum cl_req_type crt, int brw_flags)
292 struct osc_io *oio = osc_env_io(env);
293 struct osc_async_page *oap = &opg->ops_oap;
295 LASSERT(oap->oap_async_flags & ASYNC_READY);
296 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
298 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
299 oap->oap_page_off = opg->ops_from;
300 oap->oap_count = opg->ops_to - opg->ops_from + 1;
301 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
303 if (oio->oi_cap_sys_resource)
304 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
306 osc_page_transfer_get(opg, "transfer\0imm");
307 osc_page_transfer_add(env, opg, crt);
310 /* --------------- LRU page management ------------------ */
312 /* OSC is a natural place to manage LRU pages as applications are specialized
313 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
314 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
315 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
316 * for free LRU slots - this will be very bad so the algorithm requires each
317 * OSC to free slots voluntarily to maintain a reasonable number of free slots
321 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
324 * LRU pages are freed in batch mode. OSC should at least free this
325 * number of pages to avoid running out of LRU slots.
327 static inline int lru_shrink_min(struct client_obd *cli)
329 return cli->cl_max_pages_per_rpc * 2;
333 * free this number at most otherwise it will take too long time to finsih.
335 static inline int lru_shrink_max(struct client_obd *cli)
337 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
341 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
342 * we should free slots aggressively. In this way, slots are freed in a steady
343 * step to maintain fairness among OSCs.
345 * Return how many LRU pages should be freed.
347 static int osc_cache_too_much(struct client_obd *cli)
349 struct cl_client_cache *cache = cli->cl_cache;
350 long pages = atomic_long_read(&cli->cl_lru_in_list);
351 unsigned long budget;
353 LASSERT(cache != NULL);
354 budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
356 /* if it's going to run out LRU slots, we should free some, but not
357 * too much to maintain faireness among OSCs. */
358 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
360 return lru_shrink_max(cli);
361 else if (pages >= budget / 2)
362 return lru_shrink_min(cli);
364 time64_t duration = ktime_get_real_seconds();
367 /* knock out pages by duration of no IO activity */
368 duration -= cli->cl_lru_last_used;
370 * The difference shouldn't be more than 70 years
371 * so we can safely case to a long. Round to
372 * approximately 1 minute.
374 timediff = (long)(duration >> 6);
375 if (timediff > 0 && pages >= budget / timediff)
376 return lru_shrink_min(cli);
381 int lru_queue_work(const struct lu_env *env, void *data)
383 struct client_obd *cli = data;
386 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
387 count = osc_cache_too_much(cli);
389 int rc = osc_lru_shrink(env, cli, count, false);
391 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
392 cli_name(cli), rc, count);
394 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
395 ptlrpcd_queue_work(cli->cl_lru_work);
402 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
405 struct osc_async_page *oap;
408 list_for_each_entry(oap, plist, oap_pending_item) {
409 struct osc_page *opg = oap2osc_page(oap);
411 if (!opg->ops_in_lru)
415 LASSERT(list_empty(&opg->ops_lru));
416 list_add(&opg->ops_lru, &lru);
420 spin_lock(&cli->cl_lru_list_lock);
421 list_splice_tail(&lru, &cli->cl_lru_list);
422 atomic_long_sub(npages, &cli->cl_lru_busy);
423 atomic_long_add(npages, &cli->cl_lru_in_list);
424 cli->cl_lru_last_used = ktime_get_real_seconds();
425 spin_unlock(&cli->cl_lru_list_lock);
427 if (waitqueue_active(&osc_lru_waitq))
428 (void)ptlrpcd_queue_work(cli->cl_lru_work);
432 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
434 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
435 list_del_init(&opg->ops_lru);
436 atomic_long_dec(&cli->cl_lru_in_list);
440 * Page is being destroyed. The page may be not in LRU list, if the transfer
441 * has never finished(error occurred).
443 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
445 if (opg->ops_in_lru) {
446 spin_lock(&cli->cl_lru_list_lock);
447 if (!list_empty(&opg->ops_lru)) {
448 __osc_lru_del(cli, opg);
450 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
451 atomic_long_dec(&cli->cl_lru_busy);
453 spin_unlock(&cli->cl_lru_list_lock);
455 atomic_long_inc(cli->cl_lru_left);
456 /* this is a great place to release more LRU pages if
457 * this osc occupies too many LRU pages and kernel is
458 * stealing one of them. */
459 if (osc_cache_too_much(cli)) {
460 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
461 (void)ptlrpcd_queue_work(cli->cl_lru_work);
463 wake_up(&osc_lru_waitq);
465 LASSERT(list_empty(&opg->ops_lru));
470 * Delete page from LRU list for redirty.
472 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
474 /* If page is being transferred for the first time,
475 * ops_lru should be empty */
476 if (opg->ops_in_lru) {
477 if (list_empty(&opg->ops_lru))
479 spin_lock(&cli->cl_lru_list_lock);
480 if (!list_empty(&opg->ops_lru)) {
481 __osc_lru_del(cli, opg);
482 atomic_long_inc(&cli->cl_lru_busy);
484 spin_unlock(&cli->cl_lru_list_lock);
488 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
489 struct cl_page **pvec, int max_index)
491 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
494 ll_pagevec_init(pagevec, 0);
495 for (i = 0; i < max_index; i++) {
496 struct cl_page *page = pvec[i];
498 LASSERT(cl_page_is_owned(page, io));
499 cl_page_delete(env, page);
500 cl_page_discard(env, io, page);
501 cl_page_disown(env, io, page);
502 cl_pagevec_put(env, page, pagevec);
506 pagevec_release(pagevec);
510 * Check if a cl_page can be released, i.e, it's not being used.
512 * If unstable account is turned on, bulk transfer may hold one refcount
513 * for recovery so we need to check vmpage refcount as well; otherwise,
514 * even we can destroy cl_page but the corresponding vmpage can't be reused.
516 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
518 if (cl_page_in_use_noref(page))
521 if (cli->cl_cache->ccc_unstable_check) {
522 struct page *vmpage = cl_page_vmpage(page);
524 /* vmpage have two known users: cl_page and VM page cache */
525 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
532 * Drop @target of pages from LRU at most.
534 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
535 long target, bool force)
538 struct cl_object *clobj = NULL;
539 struct cl_page **pvec;
540 struct osc_page *opg;
547 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
548 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
551 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
552 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
554 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
557 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
558 atomic_dec(&cli->cl_lru_shrinkers);
562 atomic_inc(&cli->cl_lru_shrinkers);
565 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
566 io = osc_env_thread_io(env);
568 spin_lock(&cli->cl_lru_list_lock);
570 cli->cl_lru_reclaim++;
571 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
572 while (!list_empty(&cli->cl_lru_list)) {
573 struct cl_page *page;
574 bool will_free = false;
576 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
582 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
584 page = opg->ops_cl.cpl_page;
585 if (lru_page_busy(cli, page)) {
586 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
590 LASSERT(page->cp_obj != NULL);
591 if (clobj != page->cp_obj) {
592 struct cl_object *tmp = page->cp_obj;
595 spin_unlock(&cli->cl_lru_list_lock);
598 discard_pagevec(env, io, pvec, index);
602 cl_object_put(env, clobj);
608 io->ci_ignore_layout = 1;
609 rc = cl_io_init(env, io, CIT_MISC, clobj);
611 spin_lock(&cli->cl_lru_list_lock);
620 if (cl_page_own_try(env, io, page) == 0) {
621 if (!lru_page_busy(cli, page)) {
622 /* remove it from lru list earlier to avoid
624 __osc_lru_del(cli, opg);
625 opg->ops_in_lru = 0; /* will be discarded */
630 cl_page_disown(env, io, page);
635 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
639 /* Don't discard and free the page with cl_lru_list held */
640 pvec[index++] = page;
641 if (unlikely(index == OTI_PVEC_SIZE)) {
642 spin_unlock(&cli->cl_lru_list_lock);
643 discard_pagevec(env, io, pvec, index);
646 spin_lock(&cli->cl_lru_list_lock);
649 if (++count >= target)
652 spin_unlock(&cli->cl_lru_list_lock);
655 discard_pagevec(env, io, pvec, index);
658 cl_object_put(env, clobj);
661 atomic_dec(&cli->cl_lru_shrinkers);
663 atomic_long_add(count, cli->cl_lru_left);
664 wake_up(&osc_lru_waitq);
666 RETURN(count > 0 ? count : rc);
668 EXPORT_SYMBOL(osc_lru_shrink);
671 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
672 * \@npages of LRU slots. For performance consideration, it's better to drop
673 * LRU pages in batch. Therefore, the actual number is adjusted at least
676 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
679 struct cl_client_cache *cache = cli->cl_cache;
680 struct client_obd *scan;
686 LASSERT(cache != NULL);
688 env = cl_env_get(&refcheck);
692 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
693 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
694 cli_name(cli), npages);
695 rc = osc_lru_shrink(env, cli, npages, true);
697 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
698 cli_name(cli), rc, npages);
699 if (osc_cache_too_much(cli) > 0)
700 ptlrpcd_queue_work(cli->cl_lru_work);
706 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
707 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
708 atomic_long_read(&cli->cl_lru_busy), npages);
710 /* Reclaim LRU slots from other client_obd as it can't free enough
711 * from its own. This should rarely happen. */
712 spin_lock(&cache->ccc_lru_lock);
713 LASSERT(!list_empty(&cache->ccc_lru));
715 cache->ccc_lru_shrinkers++;
716 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
718 max_scans = refcount_read(&cache->ccc_users) - 2;
719 while (--max_scans > 0 &&
720 (scan = list_first_entry_or_null(&cache->ccc_lru,
722 cl_lru_osc)) != NULL) {
723 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
724 cli_name(scan), scan,
725 atomic_long_read(&scan->cl_lru_in_list),
726 atomic_long_read(&scan->cl_lru_busy));
728 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
729 if (osc_cache_too_much(scan) > 0) {
730 spin_unlock(&cache->ccc_lru_lock);
732 rc = osc_lru_shrink(env, scan, npages, true);
733 spin_lock(&cache->ccc_lru_lock);
740 spin_unlock(&cache->ccc_lru_lock);
743 cl_env_put(env, &refcheck);
744 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
745 cli_name(cli), cli, rc);
750 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
752 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
753 * Only in the case that the LRU slots are in extreme shortage, it should
754 * have reserved enough slots for an IO.
756 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
757 struct osc_page *opg)
759 struct osc_io *oio = osc_env_io(env);
764 if (cli->cl_cache == NULL) /* shall not be in LRU */
767 if (oio->oi_lru_reserved > 0) {
768 --oio->oi_lru_reserved;
772 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
773 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
774 /* run out of LRU spaces, try to drop some by itself */
775 rc = osc_lru_reclaim(cli, 1);
780 /* IO issued by readahead, don't try hard */
781 if (oio->oi_is_readahead) {
782 if (atomic_long_read(cli->cl_lru_left) > 0)
789 rc = l_wait_event_abortable(
791 atomic_long_read(cli->cl_lru_left) > 0);
800 atomic_long_inc(&cli->cl_lru_busy);
809 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
811 * The benefit of doing this is to reduce contention against atomic counter
812 * cl_lru_left by changing it from per-page access to per-IO access.
814 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
816 unsigned long reserved = 0;
817 unsigned long max_pages;
822 c = atomic_long_read(cli->cl_lru_left);
823 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
824 c = atomic_long_read(cli->cl_lru_left);
828 * Trigger writeback in the hope some LRU slot could
831 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
836 while (c >= npages) {
837 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
841 c = atomic_long_read(cli->cl_lru_left);
844 if (reserved != npages) {
846 rc = l_wait_event_abortable(
848 atomic_long_read(cli->cl_lru_left) > 0);
852 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
853 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
854 /* If there aren't enough pages in the per-OSC LRU then
855 * wake up the LRU thread to try and clear out space, so
856 * we don't block if pages are being dirtied quickly. */
857 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
858 cli_name(cli), atomic_long_read(cli->cl_lru_left),
860 (void)ptlrpcd_queue_work(cli->cl_lru_work);
867 * osc_lru_unreserve() is called to unreserve LRU slots.
869 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
870 * reasons such as page already existing or I/O error. Those reserved slots
871 * should be freed by calling this function.
873 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
875 atomic_long_add(npages, cli->cl_lru_left);
876 wake_up(&osc_lru_waitq);
880 * Atomic operations are expensive. We accumulate the accounting for the
881 * same page zone to get better performance.
882 * In practice this can work pretty good because the pages in the same RPC
883 * are likely from the same page zone.
885 #ifdef HAVE_NR_UNSTABLE_NFS
886 /* Old kernels use a separate counter for unstable pages,
887 * newer kernels treat them like any other writeback.
889 #define NR_WRITEBACK NR_UNSTABLE_NFS
892 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
893 struct osc_brw_async_args *aa,
902 page_count = desc->bd_iov_count;
904 page_count = aa->aa_page_count;
907 for (i = 0; i < page_count; i++) {
910 pz = page_zone(desc->bd_vec[i].bv_page);
912 pz = page_zone(aa->aa_ppga[i]->pg);
914 if (likely(pz == zone)) {
920 mod_zone_page_state(zone, NR_WRITEBACK,
928 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
931 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
932 struct osc_brw_async_args *aa)
934 unstable_page_accounting(desc, aa, 1);
937 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
938 struct osc_brw_async_args *aa)
940 unstable_page_accounting(desc, aa, -1);
944 * Performs "unstable" page accounting. This function balances the
945 * increment operations performed in osc_inc_unstable_pages. It is
946 * registered as the RPC request callback, and is executed when the
947 * bulk RPC is committed on the server. Thus at this point, the pages
948 * involved in the bulk transfer are no longer considered unstable.
950 * If this function is called, the request should have been committed
951 * or req:rq_unstable must have been set; it implies that the unstable
952 * statistic have been added.
954 void osc_dec_unstable_pages(struct ptlrpc_request *req)
956 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
957 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
958 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
963 page_count = desc->bd_iov_count;
965 page_count = aa->aa_page_count;
967 LASSERT(page_count >= 0);
969 dec_unstable_page_accounting(desc, aa);
971 unstable_count = atomic_long_sub_return(page_count,
972 &cli->cl_unstable_count);
973 LASSERT(unstable_count >= 0);
975 unstable_count = atomic_long_sub_return(page_count,
976 &cli->cl_cache->ccc_unstable_nr);
977 LASSERT(unstable_count >= 0);
978 if (unstable_count == 0)
979 wake_up(&cli->cl_cache->ccc_unstable_waitq);
981 if (waitqueue_active(&osc_lru_waitq))
982 (void)ptlrpcd_queue_work(cli->cl_lru_work);
986 * "unstable" page accounting. See: osc_dec_unstable_pages.
988 void osc_inc_unstable_pages(struct ptlrpc_request *req)
990 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
991 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
992 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
995 /* No unstable page tracking */
996 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1000 page_count = desc->bd_iov_count;
1002 page_count = aa->aa_page_count;
1004 add_unstable_page_accounting(desc, aa);
1005 atomic_long_add(page_count, &cli->cl_unstable_count);
1006 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1008 /* If the request has already been committed (i.e. brw_commit
1009 * called via rq_commit_cb), we need to undo the unstable page
1010 * increments we just performed because rq_commit_cb wont be
1012 spin_lock(&req->rq_lock);
1013 if (unlikely(req->rq_committed)) {
1014 spin_unlock(&req->rq_lock);
1016 osc_dec_unstable_pages(req);
1018 req->rq_unstable = 1;
1019 spin_unlock(&req->rq_lock);
1024 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1025 * This function will be called by every BRW RPC so it's critical
1026 * to make this function fast.
1028 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1030 long unstable_nr, osc_unstable_count;
1032 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1033 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1036 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1037 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1040 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1041 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1043 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1044 * has one full RPC window of unstable pages, it's a good chance
1045 * to piggyback a SOFT_SYNC flag.
1046 * Please notice that the OST won't take immediate response for the
1047 * SOFT_SYNC request so active OSCs will have more chance to carry
1048 * the flag, this is reasonable. */
1049 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1050 osc_unstable_count > cli->cl_max_pages_per_rpc *
1051 cli->cl_max_rpcs_in_flight;
1055 * Return how many LRU pages in the cache of all OSC devices
1057 * \retval return # of cached LRU pages times reclaimation tendency
1058 * \retval SHRINK_STOP if it cannot do any scanning in this time
1060 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1061 struct shrink_control *sc)
1063 struct client_obd *cli;
1064 unsigned long cached = 0;
1066 spin_lock(&osc_shrink_lock);
1067 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1068 cached += atomic_long_read(&cli->cl_lru_in_list);
1069 spin_unlock(&osc_shrink_lock);
1071 return (cached * sysctl_vfs_cache_pressure) / 100;
1075 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1077 * \retval number of cached LRU pages reclaimed
1078 * \retval SHRINK_STOP if it cannot do any scanning in this time
1080 * Linux kernel will loop calling this shrinker scan routine with
1081 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1083 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1084 * to scan and try to reclaim LRU pages, just return 0 and
1085 * osc_cache_shrink_count() will report the LRU page number.
1087 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1088 struct shrink_control *sc)
1090 struct client_obd *cli;
1091 struct client_obd *stop_anchor = NULL;
1097 if (sc->nr_to_scan == 0)
1100 if (!(sc->gfp_mask & __GFP_FS))
1103 env = cl_env_get(&refcheck);
1107 spin_lock(&osc_shrink_lock);
1108 while ((cli = list_first_entry_or_null(&osc_shrink_list,
1110 cl_shrink_list)) != NULL) {
1111 if (stop_anchor == NULL)
1113 else if (cli == stop_anchor)
1116 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1117 spin_unlock(&osc_shrink_lock);
1119 /* shrink no more than max_pages_per_rpc for an OSC */
1120 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1121 cli->cl_max_pages_per_rpc ?
1122 cli->cl_max_pages_per_rpc :
1123 sc->nr_to_scan - shrank, true);
1127 if (shrank >= sc->nr_to_scan)
1130 spin_lock(&osc_shrink_lock);
1132 spin_unlock(&osc_shrink_lock);
1135 cl_env_put(env, &refcheck);