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);
273 cl_page->cp_osc_index = index;
275 /* reserve an LRU space for this page */
276 if (cl_page->cp_type == CPT_CACHEABLE) {
277 result = osc_lru_alloc(env, osc_cli(osc), opg);
279 result = radix_tree_preload(GFP_NOFS);
281 spin_lock(&osc->oo_tree_lock);
282 result = radix_tree_insert(&osc->oo_tree,
288 spin_unlock(&osc->oo_tree_lock);
290 radix_tree_preload_end();
297 EXPORT_SYMBOL(osc_page_init);
300 * Helper function called by osc_io_submit() for every page in an immediate
301 * transfer (i.e., transferred synchronously).
303 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
304 enum cl_req_type crt, int brw_flags, ktime_t submit_time)
306 struct osc_io *oio = osc_env_io(env);
307 struct osc_async_page *oap = &opg->ops_oap;
309 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
310 "magic 0x%x\n", oap, oap->oap_magic);
311 LASSERT(oap->oap_async_flags & ASYNC_READY);
312 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
314 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
315 oap->oap_page_off = opg->ops_from;
316 oap->oap_count = opg->ops_to - opg->ops_from + 1;
317 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
319 if (oio->oi_cap_sys_resource) {
320 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
321 oap->oap_cmd |= OBD_BRW_NOQUOTA;
324 opg->ops_submit_time = submit_time;
325 osc_page_transfer_get(opg, "transfer\0imm");
326 osc_page_transfer_add(env, opg, crt);
329 /* --------------- LRU page management ------------------ */
331 /* OSC is a natural place to manage LRU pages as applications are specialized
332 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
333 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
334 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
335 * for free LRU slots - this will be very bad so the algorithm requires each
336 * OSC to free slots voluntarily to maintain a reasonable number of free slots
340 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
343 * LRU pages are freed in batch mode. OSC should at least free this
344 * number of pages to avoid running out of LRU slots.
346 static inline int lru_shrink_min(struct client_obd *cli)
348 return cli->cl_max_pages_per_rpc * 2;
352 * free this number at most otherwise it will take too long time to finsih.
354 static inline int lru_shrink_max(struct client_obd *cli)
356 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
360 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
361 * we should free slots aggressively. In this way, slots are freed in a steady
362 * step to maintain fairness among OSCs.
364 * Return how many LRU pages should be freed.
366 static int osc_cache_too_much(struct client_obd *cli)
368 struct cl_client_cache *cache = cli->cl_cache;
369 long pages = atomic_long_read(&cli->cl_lru_in_list);
370 unsigned long budget;
372 LASSERT(cache != NULL);
373 budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
375 /* if it's going to run out LRU slots, we should free some, but not
376 * too much to maintain faireness among OSCs. */
377 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
379 return lru_shrink_max(cli);
380 else if (pages >= budget / 2)
381 return lru_shrink_min(cli);
383 time64_t duration = ktime_get_real_seconds();
386 /* knock out pages by duration of no IO activity */
387 duration -= cli->cl_lru_last_used;
389 * The difference shouldn't be more than 70 years
390 * so we can safely case to a long. Round to
391 * approximately 1 minute.
393 timediff = (long)(duration >> 6);
394 if (timediff > 0 && pages >= budget / timediff)
395 return lru_shrink_min(cli);
400 int lru_queue_work(const struct lu_env *env, void *data)
402 struct client_obd *cli = data;
405 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
406 count = osc_cache_too_much(cli);
408 int rc = osc_lru_shrink(env, cli, count, false);
410 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
411 cli_name(cli), rc, count);
413 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
414 ptlrpcd_queue_work(cli->cl_lru_work);
421 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
424 struct osc_async_page *oap;
427 list_for_each_entry(oap, plist, oap_pending_item) {
428 struct osc_page *opg = oap2osc_page(oap);
430 if (!opg->ops_in_lru)
434 LASSERT(list_empty(&opg->ops_lru));
435 list_add(&opg->ops_lru, &lru);
439 spin_lock(&cli->cl_lru_list_lock);
440 list_splice_tail(&lru, &cli->cl_lru_list);
441 atomic_long_sub(npages, &cli->cl_lru_busy);
442 atomic_long_add(npages, &cli->cl_lru_in_list);
443 cli->cl_lru_last_used = ktime_get_real_seconds();
444 spin_unlock(&cli->cl_lru_list_lock);
446 if (waitqueue_active(&osc_lru_waitq))
447 (void)ptlrpcd_queue_work(cli->cl_lru_work);
451 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
453 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
454 list_del_init(&opg->ops_lru);
455 atomic_long_dec(&cli->cl_lru_in_list);
459 * Page is being destroyed. The page may be not in LRU list, if the transfer
460 * has never finished(error occurred).
462 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
464 if (opg->ops_in_lru) {
465 spin_lock(&cli->cl_lru_list_lock);
466 if (!list_empty(&opg->ops_lru)) {
467 __osc_lru_del(cli, opg);
469 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
470 atomic_long_dec(&cli->cl_lru_busy);
472 spin_unlock(&cli->cl_lru_list_lock);
474 atomic_long_inc(cli->cl_lru_left);
475 /* this is a great place to release more LRU pages if
476 * this osc occupies too many LRU pages and kernel is
477 * stealing one of them. */
478 if (osc_cache_too_much(cli)) {
479 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
480 (void)ptlrpcd_queue_work(cli->cl_lru_work);
482 wake_up(&osc_lru_waitq);
484 LASSERT(list_empty(&opg->ops_lru));
489 * Delete page from LRU list for redirty.
491 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
493 /* If page is being transferred for the first time,
494 * ops_lru should be empty */
495 if (opg->ops_in_lru) {
496 if (list_empty(&opg->ops_lru))
498 spin_lock(&cli->cl_lru_list_lock);
499 if (!list_empty(&opg->ops_lru)) {
500 __osc_lru_del(cli, opg);
501 atomic_long_inc(&cli->cl_lru_busy);
503 spin_unlock(&cli->cl_lru_list_lock);
507 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
508 struct cl_page **pvec, int max_index)
510 struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
513 ll_pagevec_init(pagevec, 0);
514 for (i = 0; i < max_index; i++) {
515 struct cl_page *page = pvec[i];
517 LASSERT(cl_page_is_owned(page, io));
518 cl_page_delete(env, page);
519 cl_page_discard(env, io, page);
520 cl_page_disown(env, io, page);
521 cl_pagevec_put(env, page, pagevec);
525 pagevec_release(pagevec);
529 * Check if a cl_page can be released, i.e, it's not being used.
531 * If unstable account is turned on, bulk transfer may hold one refcount
532 * for recovery so we need to check vmpage refcount as well; otherwise,
533 * even we can destroy cl_page but the corresponding vmpage can't be reused.
535 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
537 if (cl_page_in_use_noref(page))
540 if (cli->cl_cache->ccc_unstable_check) {
541 struct page *vmpage = cl_page_vmpage(page);
543 /* vmpage have two known users: cl_page and VM page cache */
544 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
551 * Drop @target of pages from LRU at most.
553 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
554 long target, bool force)
557 struct cl_object *clobj = NULL;
558 struct cl_page **pvec;
559 struct osc_page *opg;
566 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
567 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
570 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
571 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
573 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
576 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
577 atomic_dec(&cli->cl_lru_shrinkers);
581 atomic_inc(&cli->cl_lru_shrinkers);
584 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
585 io = osc_env_thread_io(env);
587 spin_lock(&cli->cl_lru_list_lock);
589 cli->cl_lru_reclaim++;
590 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
591 while (!list_empty(&cli->cl_lru_list)) {
592 struct cl_page *page;
593 bool will_free = false;
595 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
601 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
603 page = opg->ops_cl.cpl_page;
604 if (lru_page_busy(cli, page)) {
605 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
609 LASSERT(page->cp_obj != NULL);
610 if (clobj != page->cp_obj) {
611 struct cl_object *tmp = page->cp_obj;
614 spin_unlock(&cli->cl_lru_list_lock);
617 discard_pagevec(env, io, pvec, index);
621 cl_object_put(env, clobj);
627 io->ci_ignore_layout = 1;
628 rc = cl_io_init(env, io, CIT_MISC, clobj);
630 spin_lock(&cli->cl_lru_list_lock);
639 if (cl_page_own_try(env, io, page) == 0) {
640 if (!lru_page_busy(cli, page)) {
641 /* remove it from lru list earlier to avoid
643 __osc_lru_del(cli, opg);
644 opg->ops_in_lru = 0; /* will be discarded */
649 cl_page_disown(env, io, page);
654 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
658 /* Don't discard and free the page with cl_lru_list held */
659 pvec[index++] = page;
660 if (unlikely(index == OTI_PVEC_SIZE)) {
661 spin_unlock(&cli->cl_lru_list_lock);
662 discard_pagevec(env, io, pvec, index);
665 spin_lock(&cli->cl_lru_list_lock);
668 if (++count >= target)
671 spin_unlock(&cli->cl_lru_list_lock);
674 discard_pagevec(env, io, pvec, index);
677 cl_object_put(env, clobj);
680 atomic_dec(&cli->cl_lru_shrinkers);
682 atomic_long_add(count, cli->cl_lru_left);
683 wake_up(&osc_lru_waitq);
685 RETURN(count > 0 ? count : rc);
687 EXPORT_SYMBOL(osc_lru_shrink);
690 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
691 * \@npages of LRU slots. For performance consideration, it's better to drop
692 * LRU pages in batch. Therefore, the actual number is adjusted at least
695 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
698 struct cl_client_cache *cache = cli->cl_cache;
704 LASSERT(cache != NULL);
706 env = cl_env_get(&refcheck);
710 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
711 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
712 cli_name(cli), npages);
713 rc = osc_lru_shrink(env, cli, npages, true);
715 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
716 cli_name(cli), rc, npages);
717 if (osc_cache_too_much(cli) > 0)
718 ptlrpcd_queue_work(cli->cl_lru_work);
724 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
725 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
726 atomic_long_read(&cli->cl_lru_busy), npages);
728 /* Reclaim LRU slots from other client_obd as it can't free enough
729 * from its own. This should rarely happen. */
730 spin_lock(&cache->ccc_lru_lock);
731 LASSERT(!list_empty(&cache->ccc_lru));
733 cache->ccc_lru_shrinkers++;
734 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
736 max_scans = atomic_read(&cache->ccc_users) - 2;
737 while (--max_scans > 0 &&
738 (cli = list_first_entry_or_null(&cache->ccc_lru,
740 cl_lru_osc)) != NULL) {
741 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
743 atomic_long_read(&cli->cl_lru_in_list),
744 atomic_long_read(&cli->cl_lru_busy));
746 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
747 if (osc_cache_too_much(cli) > 0) {
748 spin_unlock(&cache->ccc_lru_lock);
750 rc = osc_lru_shrink(env, cli, npages, true);
751 spin_lock(&cache->ccc_lru_lock);
758 spin_unlock(&cache->ccc_lru_lock);
761 cl_env_put(env, &refcheck);
762 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
763 cli_name(cli), cli, rc);
768 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
770 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
771 * Only in the case that the LRU slots are in extreme shortage, it should
772 * have reserved enough slots for an IO.
774 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
775 struct osc_page *opg)
777 struct osc_io *oio = osc_env_io(env);
782 if (cli->cl_cache == NULL) /* shall not be in LRU */
785 if (oio->oi_lru_reserved > 0) {
786 --oio->oi_lru_reserved;
790 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
791 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
792 /* run out of LRU spaces, try to drop some by itself */
793 rc = osc_lru_reclaim(cli, 1);
798 /* IO issued by readahead, don't try hard */
799 if (oio->oi_is_readahead) {
800 if (atomic_long_read(cli->cl_lru_left) > 0)
807 rc = l_wait_event_abortable(
809 atomic_long_read(cli->cl_lru_left) > 0);
818 atomic_long_inc(&cli->cl_lru_busy);
827 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
829 * The benefit of doing this is to reduce contention against atomic counter
830 * cl_lru_left by changing it from per-page access to per-IO access.
832 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
834 unsigned long reserved = 0;
835 unsigned long max_pages;
840 c = atomic_long_read(cli->cl_lru_left);
841 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
842 c = atomic_long_read(cli->cl_lru_left);
846 * Trigger writeback in the hope some LRU slot could
849 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
854 while (c >= npages) {
855 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
859 c = atomic_long_read(cli->cl_lru_left);
862 if (reserved != npages) {
864 rc = l_wait_event_abortable(
866 atomic_long_read(cli->cl_lru_left) > 0);
870 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
871 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
872 /* If there aren't enough pages in the per-OSC LRU then
873 * wake up the LRU thread to try and clear out space, so
874 * we don't block if pages are being dirtied quickly. */
875 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
876 cli_name(cli), atomic_long_read(cli->cl_lru_left),
878 (void)ptlrpcd_queue_work(cli->cl_lru_work);
885 * osc_lru_unreserve() is called to unreserve LRU slots.
887 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
888 * reasons such as page already existing or I/O error. Those reserved slots
889 * should be freed by calling this function.
891 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
893 atomic_long_add(npages, cli->cl_lru_left);
894 wake_up(&osc_lru_waitq);
898 * Atomic operations are expensive. We accumulate the accounting for the
899 * same page zone to get better performance.
900 * In practice this can work pretty good because the pages in the same RPC
901 * are likely from the same page zone.
903 #ifdef HAVE_NR_UNSTABLE_NFS
904 /* Old kernels use a separate counter for unstable pages,
905 * newer kernels treat them like any other writeback.
907 #define NR_WRITEBACK NR_UNSTABLE_NFS
910 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
911 struct osc_brw_async_args *aa,
920 page_count = desc->bd_iov_count;
922 page_count = aa->aa_page_count;
925 for (i = 0; i < page_count; i++) {
928 pz = page_zone(desc->bd_vec[i].bv_page);
930 pz = page_zone(aa->aa_ppga[i]->pg);
932 if (likely(pz == zone)) {
938 mod_zone_page_state(zone, NR_WRITEBACK,
946 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
949 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
950 struct osc_brw_async_args *aa)
952 unstable_page_accounting(desc, aa, 1);
955 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
956 struct osc_brw_async_args *aa)
958 unstable_page_accounting(desc, aa, -1);
962 * Performs "unstable" page accounting. This function balances the
963 * increment operations performed in osc_inc_unstable_pages. It is
964 * registered as the RPC request callback, and is executed when the
965 * bulk RPC is committed on the server. Thus at this point, the pages
966 * involved in the bulk transfer are no longer considered unstable.
968 * If this function is called, the request should have been committed
969 * or req:rq_unstable must have been set; it implies that the unstable
970 * statistic have been added.
972 void osc_dec_unstable_pages(struct ptlrpc_request *req)
974 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
975 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
976 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
981 page_count = desc->bd_iov_count;
983 page_count = aa->aa_page_count;
985 LASSERT(page_count >= 0);
987 dec_unstable_page_accounting(desc, aa);
989 unstable_count = atomic_long_sub_return(page_count,
990 &cli->cl_unstable_count);
991 LASSERT(unstable_count >= 0);
993 unstable_count = atomic_long_sub_return(page_count,
994 &cli->cl_cache->ccc_unstable_nr);
995 LASSERT(unstable_count >= 0);
996 if (unstable_count == 0)
997 wake_up(&cli->cl_cache->ccc_unstable_waitq);
999 if (waitqueue_active(&osc_lru_waitq))
1000 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1004 * "unstable" page accounting. See: osc_dec_unstable_pages.
1006 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1008 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1009 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1010 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1013 /* No unstable page tracking */
1014 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1018 page_count = desc->bd_iov_count;
1020 page_count = aa->aa_page_count;
1022 add_unstable_page_accounting(desc, aa);
1023 atomic_long_add(page_count, &cli->cl_unstable_count);
1024 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1026 /* If the request has already been committed (i.e. brw_commit
1027 * called via rq_commit_cb), we need to undo the unstable page
1028 * increments we just performed because rq_commit_cb wont be
1030 spin_lock(&req->rq_lock);
1031 if (unlikely(req->rq_committed)) {
1032 spin_unlock(&req->rq_lock);
1034 osc_dec_unstable_pages(req);
1036 req->rq_unstable = 1;
1037 spin_unlock(&req->rq_lock);
1042 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1043 * This function will be called by every BRW RPC so it's critical
1044 * to make this function fast.
1046 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1048 long unstable_nr, osc_unstable_count;
1050 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1051 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1054 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1055 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1058 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1059 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1061 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1062 * has one full RPC window of unstable pages, it's a good chance
1063 * to piggyback a SOFT_SYNC flag.
1064 * Please notice that the OST won't take immediate response for the
1065 * SOFT_SYNC request so active OSCs will have more chance to carry
1066 * the flag, this is reasonable. */
1067 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1068 osc_unstable_count > cli->cl_max_pages_per_rpc *
1069 cli->cl_max_rpcs_in_flight;
1073 * Return how many LRU pages in the cache of all OSC devices
1075 * \retval return # of cached LRU pages times reclaimation tendency
1076 * \retval SHRINK_STOP if it cannot do any scanning in this time
1078 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1079 struct shrink_control *sc)
1081 struct client_obd *cli;
1082 unsigned long cached = 0;
1084 spin_lock(&osc_shrink_lock);
1085 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1086 cached += atomic_long_read(&cli->cl_lru_in_list);
1087 spin_unlock(&osc_shrink_lock);
1089 return (cached * sysctl_vfs_cache_pressure) / 100;
1093 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1095 * \retval number of cached LRU pages reclaimed
1096 * \retval SHRINK_STOP if it cannot do any scanning in this time
1098 * Linux kernel will loop calling this shrinker scan routine with
1099 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1101 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1102 * to scan and try to reclaim LRU pages, just return 0 and
1103 * osc_cache_shrink_count() will report the LRU page number.
1105 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1106 struct shrink_control *sc)
1108 struct client_obd *cli;
1109 struct client_obd *stop_anchor = NULL;
1115 if (sc->nr_to_scan == 0)
1118 if (!(sc->gfp_mask & __GFP_FS))
1121 env = cl_env_get(&refcheck);
1125 spin_lock(&osc_shrink_lock);
1126 while ((cli = list_first_entry_or_null(&osc_shrink_list,
1128 cl_shrink_list)) != NULL) {
1129 if (stop_anchor == NULL)
1131 else if (cli == stop_anchor)
1134 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1135 spin_unlock(&osc_shrink_lock);
1137 /* shrink no more than max_pages_per_rpc for an OSC */
1138 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1139 cli->cl_max_pages_per_rpc ?
1140 cli->cl_max_pages_per_rpc :
1141 sc->nr_to_scan - shrank, true);
1145 if (shrank >= sc->nr_to_scan)
1148 spin_lock(&osc_shrink_lock);
1150 spin_unlock(&osc_shrink_lock);
1153 cl_env_put(env, &refcheck);