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 = cl_offset(obj, start);
110 policy->l_extent.end = cl_offset(obj, end + 1) - 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",
130 oap->oap_magic, oap->oap_cmd,
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 /* This isn't really necessary for transient pages, but we also don't
211 * call clip on transient pages often, so it's OK.
213 spin_lock(&oap->oap_lock);
214 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
215 spin_unlock(&oap->oap_lock);
218 static int osc_page_flush(const struct lu_env *env,
219 const struct cl_page_slice *slice,
222 struct osc_page *opg = cl2osc_page(slice);
225 rc = osc_flush_async_page(env, io, opg);
229 static void osc_page_touch(const struct lu_env *env,
230 const struct cl_page_slice *slice, size_t to)
232 struct osc_page *opg = cl2osc_page(slice);
233 struct cl_object *obj = osc2cl(osc_page_object(opg));
235 osc_page_touch_at(env, obj, osc_index(opg), to);
238 static const struct cl_page_operations osc_page_ops = {
239 .cpo_print = osc_page_print,
240 .cpo_delete = osc_page_delete,
241 .cpo_clip = osc_page_clip,
242 .cpo_flush = osc_page_flush,
243 .cpo_page_touch = osc_page_touch,
246 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
247 struct cl_page *cl_page, pgoff_t index)
249 struct osc_object *osc = cl2osc(obj);
250 struct osc_page *opg = cl_object_page_slice(obj, cl_page);
251 struct osc_io *oio = osc_env_io(env);
255 opg->ops_to = PAGE_SIZE - 1;
257 INIT_LIST_HEAD(&opg->ops_lru);
259 result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
263 opg->ops_srvlock = osc_io_srvlock(oio);
264 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
266 /* reserve an LRU space for this page */
267 if (cl_page->cp_type == CPT_CACHEABLE) {
268 result = osc_lru_alloc(env, osc_cli(osc), opg);
270 result = radix_tree_preload(GFP_NOFS);
272 spin_lock(&osc->oo_tree_lock);
273 result = radix_tree_insert(&osc->oo_tree,
279 spin_unlock(&osc->oo_tree_lock);
281 radix_tree_preload_end();
288 EXPORT_SYMBOL(osc_page_init);
291 * Helper function called by osc_io_submit() for every page in an immediate
292 * transfer (i.e., transferred synchronously).
294 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
295 enum cl_req_type crt, int brw_flags)
297 struct osc_io *oio = osc_env_io(env);
298 struct osc_async_page *oap = &opg->ops_oap;
300 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
301 "magic 0x%x\n", oap, oap->oap_magic);
302 LASSERT(oap->oap_async_flags & ASYNC_READY);
303 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
305 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
306 oap->oap_page_off = opg->ops_from;
307 oap->oap_count = opg->ops_to - opg->ops_from + 1;
308 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
310 if (oio->oi_cap_sys_resource) {
311 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
312 oap->oap_cmd |= OBD_BRW_SYS_RESOURCE;
315 osc_page_transfer_get(opg, "transfer\0imm");
316 osc_page_transfer_add(env, opg, crt);
319 /* --------------- LRU page management ------------------ */
321 /* OSC is a natural place to manage LRU pages as applications are specialized
322 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
323 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
324 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
325 * for free LRU slots - this will be very bad so the algorithm requires each
326 * OSC to free slots voluntarily to maintain a reasonable number of free slots
330 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
333 * LRU pages are freed in batch mode. OSC should at least free this
334 * number of pages to avoid running out of LRU slots.
336 static inline int lru_shrink_min(struct client_obd *cli)
338 return cli->cl_max_pages_per_rpc * 2;
342 * free this number at most otherwise it will take too long time to finsih.
344 static inline int lru_shrink_max(struct client_obd *cli)
346 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
350 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
351 * we should free slots aggressively. In this way, slots are freed in a steady
352 * step to maintain fairness among OSCs.
354 * Return how many LRU pages should be freed.
356 static int osc_cache_too_much(struct client_obd *cli)
358 struct cl_client_cache *cache = cli->cl_cache;
359 long pages = atomic_long_read(&cli->cl_lru_in_list);
360 unsigned long budget;
362 LASSERT(cache != NULL);
363 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
365 /* if it's going to run out LRU slots, we should free some, but not
366 * too much to maintain faireness among OSCs. */
367 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
369 return lru_shrink_max(cli);
370 else if (pages >= budget / 2)
371 return lru_shrink_min(cli);
373 time64_t duration = ktime_get_real_seconds();
376 /* knock out pages by duration of no IO activity */
377 duration -= cli->cl_lru_last_used;
379 * The difference shouldn't be more than 70 years
380 * so we can safely case to a long. Round to
381 * approximately 1 minute.
383 timediff = (long)(duration >> 6);
384 if (timediff > 0 && pages >= budget / timediff)
385 return lru_shrink_min(cli);
390 int lru_queue_work(const struct lu_env *env, void *data)
392 struct client_obd *cli = data;
395 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
396 count = osc_cache_too_much(cli);
398 int rc = osc_lru_shrink(env, cli, count, false);
400 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
401 cli_name(cli), rc, count);
403 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
404 ptlrpcd_queue_work(cli->cl_lru_work);
411 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
414 struct osc_async_page *oap;
417 list_for_each_entry(oap, plist, oap_pending_item) {
418 struct osc_page *opg = oap2osc_page(oap);
420 if (!opg->ops_in_lru)
424 LASSERT(list_empty(&opg->ops_lru));
425 list_add(&opg->ops_lru, &lru);
429 spin_lock(&cli->cl_lru_list_lock);
430 list_splice_tail(&lru, &cli->cl_lru_list);
431 atomic_long_sub(npages, &cli->cl_lru_busy);
432 atomic_long_add(npages, &cli->cl_lru_in_list);
433 cli->cl_lru_last_used = ktime_get_real_seconds();
434 spin_unlock(&cli->cl_lru_list_lock);
436 if (waitqueue_active(&osc_lru_waitq))
437 (void)ptlrpcd_queue_work(cli->cl_lru_work);
441 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
443 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
444 list_del_init(&opg->ops_lru);
445 atomic_long_dec(&cli->cl_lru_in_list);
449 * Page is being destroyed. The page may be not in LRU list, if the transfer
450 * has never finished(error occurred).
452 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
454 if (opg->ops_in_lru) {
455 spin_lock(&cli->cl_lru_list_lock);
456 if (!list_empty(&opg->ops_lru)) {
457 __osc_lru_del(cli, opg);
459 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
460 atomic_long_dec(&cli->cl_lru_busy);
462 spin_unlock(&cli->cl_lru_list_lock);
464 atomic_long_inc(cli->cl_lru_left);
465 /* this is a great place to release more LRU pages if
466 * this osc occupies too many LRU pages and kernel is
467 * stealing one of them. */
468 if (osc_cache_too_much(cli)) {
469 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
470 (void)ptlrpcd_queue_work(cli->cl_lru_work);
472 wake_up(&osc_lru_waitq);
474 LASSERT(list_empty(&opg->ops_lru));
479 * Delete page from LRU list for redirty.
481 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
483 /* If page is being transferred for the first time,
484 * ops_lru should be empty */
485 if (opg->ops_in_lru) {
486 if (list_empty(&opg->ops_lru))
488 spin_lock(&cli->cl_lru_list_lock);
489 if (!list_empty(&opg->ops_lru)) {
490 __osc_lru_del(cli, opg);
491 atomic_long_inc(&cli->cl_lru_busy);
493 spin_unlock(&cli->cl_lru_list_lock);
497 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
498 struct cl_page **pvec, int max_index)
500 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
503 ll_pagevec_init(pagevec, 0);
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_pagevec_put(env, page, pagevec);
515 pagevec_release(pagevec);
519 * Check if a cl_page can be released, i.e, it's not being used.
521 * If unstable account is turned on, bulk transfer may hold one refcount
522 * for recovery so we need to check vmpage refcount as well; otherwise,
523 * even we can destroy cl_page but the corresponding vmpage can't be reused.
525 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
527 if (cl_page_in_use_noref(page))
530 if (cli->cl_cache->ccc_unstable_check) {
531 struct page *vmpage = cl_page_vmpage(page);
533 /* this check is racy because the vmpage is not locked, but
534 * that's OK - the code which does the actual page release
535 * checks this again before releasing
537 * vmpage have two known users: cl_page and VM page cache
539 if (vmpage_in_use(vmpage, 0))
546 * Drop @target of pages from LRU at most.
548 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
549 long target, bool force)
552 struct cl_object *clobj = NULL;
553 struct cl_page **pvec;
554 struct osc_page *opg;
561 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
562 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
565 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
566 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
568 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
571 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
572 atomic_dec(&cli->cl_lru_shrinkers);
576 atomic_inc(&cli->cl_lru_shrinkers);
579 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
580 io = osc_env_thread_io(env);
582 spin_lock(&cli->cl_lru_list_lock);
584 cli->cl_lru_reclaim++;
585 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
586 while (!list_empty(&cli->cl_lru_list)) {
587 struct cl_page *page;
588 bool will_free = false;
590 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
596 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
598 page = opg->ops_cl.cpl_page;
599 if (lru_page_busy(cli, page)) {
600 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
604 LASSERT(page->cp_obj != NULL);
605 if (clobj != page->cp_obj) {
606 struct cl_object *tmp = page->cp_obj;
609 spin_unlock(&cli->cl_lru_list_lock);
612 discard_pagevec(env, io, pvec, index);
616 cl_object_put(env, clobj);
622 io->ci_ignore_layout = 1;
623 rc = cl_io_init(env, io, CIT_MISC, clobj);
625 spin_lock(&cli->cl_lru_list_lock);
634 if (cl_page_own_try(env, io, page) == 0) {
635 if (!lru_page_busy(cli, page)) {
636 /* remove it from lru list earlier to avoid
638 __osc_lru_del(cli, opg);
639 opg->ops_in_lru = 0; /* will be discarded */
644 cl_page_disown(env, io, page);
649 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
653 /* Don't discard and free the page with cl_lru_list held */
654 pvec[index++] = page;
655 if (unlikely(index == OTI_PVEC_SIZE)) {
656 spin_unlock(&cli->cl_lru_list_lock);
657 discard_pagevec(env, io, pvec, index);
660 spin_lock(&cli->cl_lru_list_lock);
663 if (++count >= target)
666 spin_unlock(&cli->cl_lru_list_lock);
669 discard_pagevec(env, io, pvec, index);
672 cl_object_put(env, clobj);
675 atomic_dec(&cli->cl_lru_shrinkers);
677 atomic_long_add(count, cli->cl_lru_left);
678 wake_up(&osc_lru_waitq);
680 RETURN(count > 0 ? count : rc);
682 EXPORT_SYMBOL(osc_lru_shrink);
685 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
686 * \@npages of LRU slots. For performance consideration, it's better to drop
687 * LRU pages in batch. Therefore, the actual number is adjusted at least
690 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
693 struct cl_client_cache *cache = cli->cl_cache;
694 struct client_obd *scan;
700 LASSERT(cache != NULL);
702 env = cl_env_get(&refcheck);
706 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
707 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
708 cli_name(cli), npages);
709 rc = osc_lru_shrink(env, cli, npages, true);
711 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
712 cli_name(cli), rc, npages);
713 if (osc_cache_too_much(cli) > 0)
714 ptlrpcd_queue_work(cli->cl_lru_work);
720 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
721 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
722 atomic_long_read(&cli->cl_lru_busy), npages);
724 /* Reclaim LRU slots from other client_obd as it can't free enough
725 * from its own. This should rarely happen. */
726 spin_lock(&cache->ccc_lru_lock);
727 LASSERT(!list_empty(&cache->ccc_lru));
729 cache->ccc_lru_shrinkers++;
730 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
732 max_scans = atomic_read(&cache->ccc_users) - 2;
733 while (--max_scans > 0 &&
734 (scan = list_first_entry_or_null(&cache->ccc_lru,
736 cl_lru_osc)) != NULL) {
737 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
738 cli_name(scan), scan,
739 atomic_long_read(&scan->cl_lru_in_list),
740 atomic_long_read(&scan->cl_lru_busy));
742 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
743 if (osc_cache_too_much(scan) > 0) {
744 spin_unlock(&cache->ccc_lru_lock);
746 rc = osc_lru_shrink(env, scan, npages, true);
747 spin_lock(&cache->ccc_lru_lock);
754 spin_unlock(&cache->ccc_lru_lock);
757 cl_env_put(env, &refcheck);
758 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
759 cli_name(cli), cli, rc);
764 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
766 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
767 * Only in the case that the LRU slots are in extreme shortage, it should
768 * have reserved enough slots for an IO.
770 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
771 struct osc_page *opg)
773 struct osc_io *oio = osc_env_io(env);
778 if (cli->cl_cache == NULL) /* shall not be in LRU */
781 if (oio->oi_lru_reserved > 0) {
782 --oio->oi_lru_reserved;
786 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
787 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
788 /* run out of LRU spaces, try to drop some by itself */
789 rc = osc_lru_reclaim(cli, 1);
794 /* IO issued by readahead, don't try hard */
795 if (oio->oi_is_readahead) {
796 if (atomic_long_read(cli->cl_lru_left) > 0)
803 rc = l_wait_event_abortable(
805 atomic_long_read(cli->cl_lru_left) > 0);
814 atomic_long_inc(&cli->cl_lru_busy);
823 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
825 * The benefit of doing this is to reduce contention against atomic counter
826 * cl_lru_left by changing it from per-page access to per-IO access.
828 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
830 unsigned long reserved = 0;
831 unsigned long 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);
842 * Trigger writeback in the hope some LRU slot could
845 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
850 while (c >= npages) {
851 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
855 c = atomic_long_read(cli->cl_lru_left);
858 if (reserved != npages) {
860 rc = l_wait_event_abortable(
862 atomic_long_read(cli->cl_lru_left) > 0);
866 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
867 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
868 /* If there aren't enough pages in the per-OSC LRU then
869 * wake up the LRU thread to try and clear out space, so
870 * we don't block if pages are being dirtied quickly. */
871 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
872 cli_name(cli), atomic_long_read(cli->cl_lru_left),
874 (void)ptlrpcd_queue_work(cli->cl_lru_work);
881 * osc_lru_unreserve() is called to unreserve LRU slots.
883 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
884 * reasons such as page already existing or I/O error. Those reserved slots
885 * should be freed by calling this function.
887 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
889 atomic_long_add(npages, cli->cl_lru_left);
890 wake_up(&osc_lru_waitq);
894 * Atomic operations are expensive. We accumulate the accounting for the
895 * same page zone to get better performance.
896 * In practice this can work pretty good because the pages in the same RPC
897 * are likely from the same page zone.
899 #ifdef HAVE_NR_UNSTABLE_NFS
900 /* Old kernels use a separate counter for unstable pages,
901 * newer kernels treat them like any other writeback.
903 #define NR_WRITEBACK NR_UNSTABLE_NFS
906 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
907 struct osc_brw_async_args *aa,
916 page_count = desc->bd_iov_count;
918 page_count = aa->aa_page_count;
921 for (i = 0; i < page_count; i++) {
924 pz = page_zone(desc->bd_vec[i].bv_page);
926 pz = page_zone(aa->aa_ppga[i]->pg);
928 if (likely(pz == zone)) {
934 mod_zone_page_state(zone, NR_WRITEBACK,
942 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
945 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
946 struct osc_brw_async_args *aa)
948 unstable_page_accounting(desc, aa, 1);
951 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
952 struct osc_brw_async_args *aa)
954 unstable_page_accounting(desc, aa, -1);
958 * Performs "unstable" page accounting. This function balances the
959 * increment operations performed in osc_inc_unstable_pages. It is
960 * registered as the RPC request callback, and is executed when the
961 * bulk RPC is committed on the server. Thus at this point, the pages
962 * involved in the bulk transfer are no longer considered unstable.
964 * If this function is called, the request should have been committed
965 * or req:rq_unstable must have been set; it implies that the unstable
966 * statistic have been added.
968 void osc_dec_unstable_pages(struct ptlrpc_request *req)
970 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
971 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
972 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
977 page_count = desc->bd_iov_count;
979 page_count = aa->aa_page_count;
981 LASSERT(page_count >= 0);
983 dec_unstable_page_accounting(desc, aa);
985 unstable_count = atomic_long_sub_return(page_count,
986 &cli->cl_unstable_count);
987 LASSERT(unstable_count >= 0);
989 unstable_count = atomic_long_sub_return(page_count,
990 &cli->cl_cache->ccc_unstable_nr);
991 LASSERT(unstable_count >= 0);
992 if (unstable_count == 0)
993 wake_up(&cli->cl_cache->ccc_unstable_waitq);
995 if (waitqueue_active(&osc_lru_waitq))
996 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1000 * "unstable" page accounting. See: osc_dec_unstable_pages.
1002 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1004 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1005 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1006 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1009 /* No unstable page tracking */
1010 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1014 page_count = desc->bd_iov_count;
1016 page_count = aa->aa_page_count;
1018 add_unstable_page_accounting(desc, aa);
1019 atomic_long_add(page_count, &cli->cl_unstable_count);
1020 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1022 /* If the request has already been committed (i.e. brw_commit
1023 * called via rq_commit_cb), we need to undo the unstable page
1024 * increments we just performed because rq_commit_cb wont be
1026 spin_lock(&req->rq_lock);
1027 if (unlikely(req->rq_committed)) {
1028 spin_unlock(&req->rq_lock);
1030 osc_dec_unstable_pages(req);
1032 req->rq_unstable = 1;
1033 spin_unlock(&req->rq_lock);
1038 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1039 * This function will be called by every BRW RPC so it's critical
1040 * to make this function fast.
1042 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1044 long unstable_nr, osc_unstable_count;
1046 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1047 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1050 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1051 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1054 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1055 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1057 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1058 * has one full RPC window of unstable pages, it's a good chance
1059 * to piggyback a SOFT_SYNC flag.
1060 * Please notice that the OST won't take immediate response for the
1061 * SOFT_SYNC request so active OSCs will have more chance to carry
1062 * the flag, this is reasonable. */
1063 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1064 osc_unstable_count > cli->cl_max_pages_per_rpc *
1065 cli->cl_max_rpcs_in_flight;
1069 * Return how many LRU pages in the cache of all OSC devices
1071 * \retval return # of cached LRU pages times reclaimation tendency
1072 * \retval SHRINK_STOP if it cannot do any scanning in this time
1074 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1075 struct shrink_control *sc)
1077 struct client_obd *cli;
1078 unsigned long cached = 0;
1080 spin_lock(&osc_shrink_lock);
1081 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1082 cached += atomic_long_read(&cli->cl_lru_in_list);
1083 spin_unlock(&osc_shrink_lock);
1085 return (cached * sysctl_vfs_cache_pressure) / 100;
1089 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1091 * \retval number of cached LRU pages reclaimed
1092 * \retval SHRINK_STOP if it cannot do any scanning in this time
1094 * Linux kernel will loop calling this shrinker scan routine with
1095 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1097 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1098 * to scan and try to reclaim LRU pages, just return 0 and
1099 * osc_cache_shrink_count() will report the LRU page number.
1101 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1102 struct shrink_control *sc)
1104 struct client_obd *cli;
1105 struct client_obd *stop_anchor = NULL;
1111 if (sc->nr_to_scan == 0)
1114 if (!(sc->gfp_mask & __GFP_FS))
1117 env = cl_env_get(&refcheck);
1121 spin_lock(&osc_shrink_lock);
1122 while ((cli = list_first_entry_or_null(&osc_shrink_list,
1124 cl_shrink_list)) != NULL) {
1125 if (stop_anchor == NULL)
1127 else if (cli == stop_anchor)
1130 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1131 spin_unlock(&osc_shrink_lock);
1133 /* shrink no more than max_pages_per_rpc for an OSC */
1134 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1135 cli->cl_max_pages_per_rpc ?
1136 cli->cl_max_pages_per_rpc :
1137 sc->nr_to_scan - shrank, true);
1141 if (shrank >= sc->nr_to_scan)
1144 spin_lock(&osc_shrink_lock);
1146 spin_unlock(&osc_shrink_lock);
1149 cl_env_put(env, &refcheck);