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 = cl2osc(opg->ops_cl.cpl_obj);
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 inline s64 osc_submit_duration(struct osc_page *opg)
115 if (ktime_to_ns(opg->ops_submit_time) == 0)
118 return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
121 static int osc_page_print(const struct lu_env *env,
122 const struct cl_page_slice *slice,
123 void *cookie, lu_printer_t printer)
125 struct osc_page *opg = cl2osc_page(slice);
126 struct osc_async_page *oap = &opg->ops_oap;
127 struct osc_object *obj = cl2osc(slice->cpl_obj);
128 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
130 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
132 "2< %lld %u %u %#x %#x | %p %p %p > "
134 "4< %d %d %d %lu %c | %c %c %c %c > "
135 "5< %c %c %c %c | %d %c | %d %c %c>\n",
138 oap->oap_magic, oap->oap_cmd,
139 list_empty_marker(&oap->oap_pending_item),
140 list_empty_marker(&oap->oap_rpc_item),
142 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
143 oap->oap_async_flags, oap->oap_brw_flags,
144 oap->oap_request, oap->oap_cli, obj,
146 opg->ops_transfer_pinned,
147 osc_submit_duration(opg), opg->ops_srvlock,
149 cli->cl_r_in_flight, cli->cl_w_in_flight,
150 cli->cl_max_rpcs_in_flight,
152 waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
153 list_empty_marker(&cli->cl_loi_ready_list),
154 list_empty_marker(&cli->cl_loi_hp_ready_list),
155 list_empty_marker(&cli->cl_loi_write_list),
156 list_empty_marker(&cli->cl_loi_read_list),
158 list_empty_marker(&obj->oo_ready_item),
159 list_empty_marker(&obj->oo_hp_ready_item),
160 list_empty_marker(&obj->oo_write_item),
161 list_empty_marker(&obj->oo_read_item),
162 atomic_read(&obj->oo_nr_reads),
163 list_empty_marker(&obj->oo_reading_exts),
164 atomic_read(&obj->oo_nr_writes),
165 list_empty_marker(&obj->oo_hp_exts),
166 list_empty_marker(&obj->oo_urgent_exts));
169 static void osc_page_delete(const struct lu_env *env,
170 const struct cl_page_slice *slice)
172 struct osc_page *opg = cl2osc_page(slice);
173 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
177 CDEBUG(D_TRACE, "%p\n", opg);
178 osc_page_transfer_put(env, opg);
179 rc = osc_teardown_async_page(env, obj, opg);
181 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
182 "Trying to teardown failed: %d\n", rc);
186 osc_lru_del(osc_cli(obj), opg);
188 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
191 spin_lock(&obj->oo_tree_lock);
192 if (opg->ops_intree) {
193 value = radix_tree_delete(&obj->oo_tree,
200 spin_unlock(&obj->oo_tree_lock);
202 LASSERT(ergo(value != NULL, value == opg));
208 static void osc_page_clip(const struct lu_env *env,
209 const struct cl_page_slice *slice,
212 struct osc_page *opg = cl2osc_page(slice);
213 struct osc_async_page *oap = &opg->ops_oap;
215 opg->ops_from = from;
216 /* argument @to is exclusive, but @ops_to is inclusive */
217 opg->ops_to = to - 1;
218 /* This isn't really necessary for transient pages, but we also don't
219 * call clip on transient pages often, so it's OK.
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_flush(const struct lu_env *env,
227 const struct cl_page_slice *slice,
230 struct osc_page *opg = cl2osc_page(slice);
233 rc = osc_flush_async_page(env, io, opg);
237 static void osc_page_touch(const struct lu_env *env,
238 const struct cl_page_slice *slice, size_t to)
240 struct osc_page *opg = cl2osc_page(slice);
241 struct cl_object *obj = opg->ops_cl.cpl_obj;
243 osc_page_touch_at(env, obj, osc_index(opg), to);
246 static const struct cl_page_operations osc_page_ops = {
247 .cpo_print = osc_page_print,
248 .cpo_delete = osc_page_delete,
249 .cpo_clip = osc_page_clip,
250 .cpo_flush = osc_page_flush,
251 .cpo_page_touch = osc_page_touch,
254 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
255 struct cl_page *cl_page, pgoff_t index)
257 struct osc_object *osc = cl2osc(obj);
258 struct osc_page *opg = cl_object_page_slice(obj, cl_page);
259 struct osc_io *oio = osc_env_io(env);
263 opg->ops_to = PAGE_SIZE - 1;
265 INIT_LIST_HEAD(&opg->ops_lru);
267 result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
271 opg->ops_srvlock = osc_io_srvlock(oio);
272 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
274 /* reserve an LRU space for this page */
275 if (cl_page->cp_type == CPT_CACHEABLE) {
276 result = osc_lru_alloc(env, osc_cli(osc), opg);
278 result = radix_tree_preload(GFP_NOFS);
280 spin_lock(&osc->oo_tree_lock);
281 result = radix_tree_insert(&osc->oo_tree,
287 spin_unlock(&osc->oo_tree_lock);
289 radix_tree_preload_end();
296 EXPORT_SYMBOL(osc_page_init);
299 * Helper function called by osc_io_submit() for every page in an immediate
300 * transfer (i.e., transferred synchronously).
302 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
303 enum cl_req_type crt, int brw_flags, ktime_t submit_time)
305 struct osc_io *oio = osc_env_io(env);
306 struct osc_async_page *oap = &opg->ops_oap;
308 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
309 "magic 0x%x\n", oap, oap->oap_magic);
310 LASSERT(oap->oap_async_flags & ASYNC_READY);
311 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
313 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
314 oap->oap_page_off = opg->ops_from;
315 oap->oap_count = opg->ops_to - opg->ops_from + 1;
316 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
318 if (oio->oi_cap_sys_resource) {
319 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
320 oap->oap_cmd |= OBD_BRW_SYS_RESOURCE;
323 opg->ops_submit_time = submit_time;
324 osc_page_transfer_get(opg, "transfer\0imm");
325 osc_page_transfer_add(env, opg, crt);
328 /* --------------- LRU page management ------------------ */
330 /* OSC is a natural place to manage LRU pages as applications are specialized
331 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
332 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
333 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
334 * for free LRU slots - this will be very bad so the algorithm requires each
335 * OSC to free slots voluntarily to maintain a reasonable number of free slots
339 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
342 * LRU pages are freed in batch mode. OSC should at least free this
343 * number of pages to avoid running out of LRU slots.
345 static inline int lru_shrink_min(struct client_obd *cli)
347 return cli->cl_max_pages_per_rpc * 2;
351 * free this number at most otherwise it will take too long time to finsih.
353 static inline int lru_shrink_max(struct client_obd *cli)
355 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
359 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
360 * we should free slots aggressively. In this way, slots are freed in a steady
361 * step to maintain fairness among OSCs.
363 * Return how many LRU pages should be freed.
365 static int osc_cache_too_much(struct client_obd *cli)
367 struct cl_client_cache *cache = cli->cl_cache;
368 long pages = atomic_long_read(&cli->cl_lru_in_list);
369 unsigned long budget;
371 LASSERT(cache != NULL);
372 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
374 /* if it's going to run out LRU slots, we should free some, but not
375 * too much to maintain faireness among OSCs. */
376 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
378 return lru_shrink_max(cli);
379 else if (pages >= budget / 2)
380 return lru_shrink_min(cli);
382 time64_t duration = ktime_get_real_seconds();
385 /* knock out pages by duration of no IO activity */
386 duration -= cli->cl_lru_last_used;
388 * The difference shouldn't be more than 70 years
389 * so we can safely case to a long. Round to
390 * approximately 1 minute.
392 timediff = (long)(duration >> 6);
393 if (timediff > 0 && pages >= budget / timediff)
394 return lru_shrink_min(cli);
399 int lru_queue_work(const struct lu_env *env, void *data)
401 struct client_obd *cli = data;
404 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
405 count = osc_cache_too_much(cli);
407 int rc = osc_lru_shrink(env, cli, count, false);
409 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
410 cli_name(cli), rc, count);
412 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
413 ptlrpcd_queue_work(cli->cl_lru_work);
420 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
423 struct osc_async_page *oap;
426 list_for_each_entry(oap, plist, oap_pending_item) {
427 struct osc_page *opg = oap2osc_page(oap);
429 if (!opg->ops_in_lru)
433 LASSERT(list_empty(&opg->ops_lru));
434 list_add(&opg->ops_lru, &lru);
438 spin_lock(&cli->cl_lru_list_lock);
439 list_splice_tail(&lru, &cli->cl_lru_list);
440 atomic_long_sub(npages, &cli->cl_lru_busy);
441 atomic_long_add(npages, &cli->cl_lru_in_list);
442 cli->cl_lru_last_used = ktime_get_real_seconds();
443 spin_unlock(&cli->cl_lru_list_lock);
445 if (waitqueue_active(&osc_lru_waitq))
446 (void)ptlrpcd_queue_work(cli->cl_lru_work);
450 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
452 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
453 list_del_init(&opg->ops_lru);
454 atomic_long_dec(&cli->cl_lru_in_list);
458 * Page is being destroyed. The page may be not in LRU list, if the transfer
459 * has never finished(error occurred).
461 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
463 if (opg->ops_in_lru) {
464 spin_lock(&cli->cl_lru_list_lock);
465 if (!list_empty(&opg->ops_lru)) {
466 __osc_lru_del(cli, opg);
468 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
469 atomic_long_dec(&cli->cl_lru_busy);
471 spin_unlock(&cli->cl_lru_list_lock);
473 atomic_long_inc(cli->cl_lru_left);
474 /* this is a great place to release more LRU pages if
475 * this osc occupies too many LRU pages and kernel is
476 * stealing one of them. */
477 if (osc_cache_too_much(cli)) {
478 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
479 (void)ptlrpcd_queue_work(cli->cl_lru_work);
481 wake_up(&osc_lru_waitq);
483 LASSERT(list_empty(&opg->ops_lru));
488 * Delete page from LRU list for redirty.
490 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
492 /* If page is being transferred for the first time,
493 * ops_lru should be empty */
494 if (opg->ops_in_lru) {
495 if (list_empty(&opg->ops_lru))
497 spin_lock(&cli->cl_lru_list_lock);
498 if (!list_empty(&opg->ops_lru)) {
499 __osc_lru_del(cli, opg);
500 atomic_long_inc(&cli->cl_lru_busy);
502 spin_unlock(&cli->cl_lru_list_lock);
506 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
507 struct cl_page **pvec, int max_index)
509 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
512 ll_pagevec_init(pagevec, 0);
513 for (i = 0; i < max_index; i++) {
514 struct cl_page *page = pvec[i];
516 LASSERT(cl_page_is_owned(page, io));
517 cl_page_delete(env, page);
518 cl_page_discard(env, io, page);
519 cl_page_disown(env, io, page);
520 cl_pagevec_put(env, page, pagevec);
524 pagevec_release(pagevec);
528 * Check if a cl_page can be released, i.e, it's not being used.
530 * If unstable account is turned on, bulk transfer may hold one refcount
531 * for recovery so we need to check vmpage refcount as well; otherwise,
532 * even we can destroy cl_page but the corresponding vmpage can't be reused.
534 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
536 if (cl_page_in_use_noref(page))
539 if (cli->cl_cache->ccc_unstable_check) {
540 struct page *vmpage = cl_page_vmpage(page);
542 /* this check is racy because the vmpage is not locked, but
543 * that's OK - the code which does the actual page release
544 * checks this again before releasing
546 * vmpage have two known users: cl_page and VM page cache
548 if (vmpage_in_use(vmpage, 0))
555 * Drop @target of pages from LRU at most.
557 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
558 long target, bool force)
561 struct cl_object *clobj = NULL;
562 struct cl_page **pvec;
563 struct osc_page *opg;
570 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
571 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
574 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
575 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
577 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
580 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
581 atomic_dec(&cli->cl_lru_shrinkers);
585 atomic_inc(&cli->cl_lru_shrinkers);
588 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
589 io = osc_env_thread_io(env);
591 spin_lock(&cli->cl_lru_list_lock);
593 cli->cl_lru_reclaim++;
594 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
595 while (!list_empty(&cli->cl_lru_list)) {
596 struct cl_page *page;
597 bool will_free = false;
599 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
605 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
607 page = opg->ops_cl.cpl_page;
608 if (lru_page_busy(cli, page)) {
609 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
613 LASSERT(page->cp_obj != NULL);
614 if (clobj != page->cp_obj) {
615 struct cl_object *tmp = page->cp_obj;
618 spin_unlock(&cli->cl_lru_list_lock);
621 discard_pagevec(env, io, pvec, index);
625 cl_object_put(env, clobj);
631 io->ci_ignore_layout = 1;
632 rc = cl_io_init(env, io, CIT_MISC, clobj);
634 spin_lock(&cli->cl_lru_list_lock);
643 if (cl_page_own_try(env, io, page) == 0) {
644 if (!lru_page_busy(cli, page)) {
645 /* remove it from lru list earlier to avoid
647 __osc_lru_del(cli, opg);
648 opg->ops_in_lru = 0; /* will be discarded */
653 cl_page_disown(env, io, page);
658 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
662 /* Don't discard and free the page with cl_lru_list held */
663 pvec[index++] = page;
664 if (unlikely(index == OTI_PVEC_SIZE)) {
665 spin_unlock(&cli->cl_lru_list_lock);
666 discard_pagevec(env, io, pvec, index);
669 spin_lock(&cli->cl_lru_list_lock);
672 if (++count >= target)
675 spin_unlock(&cli->cl_lru_list_lock);
678 discard_pagevec(env, io, pvec, index);
681 cl_object_put(env, clobj);
684 atomic_dec(&cli->cl_lru_shrinkers);
686 atomic_long_add(count, cli->cl_lru_left);
687 wake_up(&osc_lru_waitq);
689 RETURN(count > 0 ? count : rc);
691 EXPORT_SYMBOL(osc_lru_shrink);
694 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
695 * \@npages of LRU slots. For performance consideration, it's better to drop
696 * LRU pages in batch. Therefore, the actual number is adjusted at least
699 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
702 struct cl_client_cache *cache = cli->cl_cache;
703 struct client_obd *scan;
709 LASSERT(cache != NULL);
711 env = cl_env_get(&refcheck);
715 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
716 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
717 cli_name(cli), npages);
718 rc = osc_lru_shrink(env, cli, npages, true);
720 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
721 cli_name(cli), rc, npages);
722 if (osc_cache_too_much(cli) > 0)
723 ptlrpcd_queue_work(cli->cl_lru_work);
729 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
730 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
731 atomic_long_read(&cli->cl_lru_busy), npages);
733 /* Reclaim LRU slots from other client_obd as it can't free enough
734 * from its own. This should rarely happen. */
735 spin_lock(&cache->ccc_lru_lock);
736 LASSERT(!list_empty(&cache->ccc_lru));
738 cache->ccc_lru_shrinkers++;
739 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
741 max_scans = atomic_read(&cache->ccc_users) - 2;
742 while (--max_scans > 0 &&
743 (scan = list_first_entry_or_null(&cache->ccc_lru,
745 cl_lru_osc)) != NULL) {
746 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
747 cli_name(scan), scan,
748 atomic_long_read(&scan->cl_lru_in_list),
749 atomic_long_read(&scan->cl_lru_busy));
751 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
752 if (osc_cache_too_much(scan) > 0) {
753 spin_unlock(&cache->ccc_lru_lock);
755 rc = osc_lru_shrink(env, scan, npages, true);
756 spin_lock(&cache->ccc_lru_lock);
763 spin_unlock(&cache->ccc_lru_lock);
766 cl_env_put(env, &refcheck);
767 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
768 cli_name(cli), cli, rc);
773 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
775 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
776 * Only in the case that the LRU slots are in extreme shortage, it should
777 * have reserved enough slots for an IO.
779 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
780 struct osc_page *opg)
782 struct osc_io *oio = osc_env_io(env);
787 if (cli->cl_cache == NULL) /* shall not be in LRU */
790 if (oio->oi_lru_reserved > 0) {
791 --oio->oi_lru_reserved;
795 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
796 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
797 /* run out of LRU spaces, try to drop some by itself */
798 rc = osc_lru_reclaim(cli, 1);
803 /* IO issued by readahead, don't try hard */
804 if (oio->oi_is_readahead) {
805 if (atomic_long_read(cli->cl_lru_left) > 0)
812 rc = l_wait_event_abortable(
814 atomic_long_read(cli->cl_lru_left) > 0);
823 atomic_long_inc(&cli->cl_lru_busy);
832 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
834 * The benefit of doing this is to reduce contention against atomic counter
835 * cl_lru_left by changing it from per-page access to per-IO access.
837 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
839 unsigned long reserved = 0;
840 unsigned long max_pages;
845 c = atomic_long_read(cli->cl_lru_left);
846 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
847 c = atomic_long_read(cli->cl_lru_left);
851 * Trigger writeback in the hope some LRU slot could
854 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
859 while (c >= npages) {
860 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
864 c = atomic_long_read(cli->cl_lru_left);
867 if (reserved != npages) {
869 rc = l_wait_event_abortable(
871 atomic_long_read(cli->cl_lru_left) > 0);
875 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
876 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
877 /* If there aren't enough pages in the per-OSC LRU then
878 * wake up the LRU thread to try and clear out space, so
879 * we don't block if pages are being dirtied quickly. */
880 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
881 cli_name(cli), atomic_long_read(cli->cl_lru_left),
883 (void)ptlrpcd_queue_work(cli->cl_lru_work);
890 * osc_lru_unreserve() is called to unreserve LRU slots.
892 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
893 * reasons such as page already existing or I/O error. Those reserved slots
894 * should be freed by calling this function.
896 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
898 atomic_long_add(npages, cli->cl_lru_left);
899 wake_up(&osc_lru_waitq);
903 * Atomic operations are expensive. We accumulate the accounting for the
904 * same page zone to get better performance.
905 * In practice this can work pretty good because the pages in the same RPC
906 * are likely from the same page zone.
908 #ifdef HAVE_NR_UNSTABLE_NFS
909 /* Old kernels use a separate counter for unstable pages,
910 * newer kernels treat them like any other writeback.
912 #define NR_WRITEBACK NR_UNSTABLE_NFS
915 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
916 struct osc_brw_async_args *aa,
925 page_count = desc->bd_iov_count;
927 page_count = aa->aa_page_count;
930 for (i = 0; i < page_count; i++) {
933 pz = page_zone(desc->bd_vec[i].bv_page);
935 pz = page_zone(aa->aa_ppga[i]->pg);
937 if (likely(pz == zone)) {
943 mod_zone_page_state(zone, NR_WRITEBACK,
951 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
954 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
955 struct osc_brw_async_args *aa)
957 unstable_page_accounting(desc, aa, 1);
960 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
961 struct osc_brw_async_args *aa)
963 unstable_page_accounting(desc, aa, -1);
967 * Performs "unstable" page accounting. This function balances the
968 * increment operations performed in osc_inc_unstable_pages. It is
969 * registered as the RPC request callback, and is executed when the
970 * bulk RPC is committed on the server. Thus at this point, the pages
971 * involved in the bulk transfer are no longer considered unstable.
973 * If this function is called, the request should have been committed
974 * or req:rq_unstable must have been set; it implies that the unstable
975 * statistic have been added.
977 void osc_dec_unstable_pages(struct ptlrpc_request *req)
979 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
980 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
981 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
986 page_count = desc->bd_iov_count;
988 page_count = aa->aa_page_count;
990 LASSERT(page_count >= 0);
992 dec_unstable_page_accounting(desc, aa);
994 unstable_count = atomic_long_sub_return(page_count,
995 &cli->cl_unstable_count);
996 LASSERT(unstable_count >= 0);
998 unstable_count = atomic_long_sub_return(page_count,
999 &cli->cl_cache->ccc_unstable_nr);
1000 LASSERT(unstable_count >= 0);
1001 if (unstable_count == 0)
1002 wake_up(&cli->cl_cache->ccc_unstable_waitq);
1004 if (waitqueue_active(&osc_lru_waitq))
1005 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1009 * "unstable" page accounting. See: osc_dec_unstable_pages.
1011 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1013 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1014 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1015 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1018 /* No unstable page tracking */
1019 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1023 page_count = desc->bd_iov_count;
1025 page_count = aa->aa_page_count;
1027 add_unstable_page_accounting(desc, aa);
1028 atomic_long_add(page_count, &cli->cl_unstable_count);
1029 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1031 /* If the request has already been committed (i.e. brw_commit
1032 * called via rq_commit_cb), we need to undo the unstable page
1033 * increments we just performed because rq_commit_cb wont be
1035 spin_lock(&req->rq_lock);
1036 if (unlikely(req->rq_committed)) {
1037 spin_unlock(&req->rq_lock);
1039 osc_dec_unstable_pages(req);
1041 req->rq_unstable = 1;
1042 spin_unlock(&req->rq_lock);
1047 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1048 * This function will be called by every BRW RPC so it's critical
1049 * to make this function fast.
1051 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1053 long unstable_nr, osc_unstable_count;
1055 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1056 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1059 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1060 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1063 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1064 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1066 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1067 * has one full RPC window of unstable pages, it's a good chance
1068 * to piggyback a SOFT_SYNC flag.
1069 * Please notice that the OST won't take immediate response for the
1070 * SOFT_SYNC request so active OSCs will have more chance to carry
1071 * the flag, this is reasonable. */
1072 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1073 osc_unstable_count > cli->cl_max_pages_per_rpc *
1074 cli->cl_max_rpcs_in_flight;
1078 * Return how many LRU pages in the cache of all OSC devices
1080 * \retval return # of cached LRU pages times reclaimation tendency
1081 * \retval SHRINK_STOP if it cannot do any scanning in this time
1083 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1084 struct shrink_control *sc)
1086 struct client_obd *cli;
1087 unsigned long cached = 0;
1089 spin_lock(&osc_shrink_lock);
1090 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1091 cached += atomic_long_read(&cli->cl_lru_in_list);
1092 spin_unlock(&osc_shrink_lock);
1094 return (cached * sysctl_vfs_cache_pressure) / 100;
1098 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1100 * \retval number of cached LRU pages reclaimed
1101 * \retval SHRINK_STOP if it cannot do any scanning in this time
1103 * Linux kernel will loop calling this shrinker scan routine with
1104 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1106 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1107 * to scan and try to reclaim LRU pages, just return 0 and
1108 * osc_cache_shrink_count() will report the LRU page number.
1110 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1111 struct shrink_control *sc)
1113 struct client_obd *cli;
1114 struct client_obd *stop_anchor = NULL;
1120 if (sc->nr_to_scan == 0)
1123 if (!(sc->gfp_mask & __GFP_FS))
1126 env = cl_env_get(&refcheck);
1130 spin_lock(&osc_shrink_lock);
1131 while ((cli = list_first_entry_or_null(&osc_shrink_list,
1133 cl_shrink_list)) != NULL) {
1134 if (stop_anchor == NULL)
1136 else if (cli == stop_anchor)
1139 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1140 spin_unlock(&osc_shrink_lock);
1142 /* shrink no more than max_pages_per_rpc for an OSC */
1143 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1144 cli->cl_max_pages_per_rpc ?
1145 cli->cl_max_pages_per_rpc :
1146 sc->nr_to_scan - shrank, true);
1150 if (shrank >= sc->nr_to_scan)
1153 spin_lock(&osc_shrink_lock);
1155 spin_unlock(&osc_shrink_lock);
1158 cl_env_put(env, &refcheck);