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_object *osc,
90 struct osc_page *opg, struct cl_io *io,
96 osc_page_transfer_get(opg, "transfer\0cache");
97 result = osc_queue_async_io(env, io, osc, opg, cb);
99 osc_page_transfer_put(env, opg);
101 osc_page_transfer_add(env, opg, CRT_WRITE);
106 void osc_index2policy(union ldlm_policy_data *policy,
107 const struct cl_object *obj, pgoff_t start, pgoff_t end)
109 memset(policy, 0, sizeof *policy);
110 policy->l_extent.start = start << PAGE_SHIFT;
111 policy->l_extent.end = ((end + 1) << PAGE_SHIFT) - 1;
114 static int osc_page_print(const struct lu_env *env,
115 const struct cl_page_slice *slice,
116 void *cookie, lu_printer_t printer)
118 struct osc_page *opg = cl2osc_page(slice);
119 struct osc_async_page *oap = &opg->ops_oap;
120 struct osc_object *obj = osc_page_object(opg);
121 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
123 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
125 "2< %lld %u %u %#x %#x | %p %p > "
127 "4< %d %d %d %lu %c | %c %c %c %c > "
128 "5< %c %c %c %c | %d %c | %d %c %c>\n",
132 list_empty_marker(&oap->oap_pending_item),
133 list_empty_marker(&oap->oap_rpc_item),
135 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
136 oap->oap_async_flags, oap->oap_brw_flags,
139 opg->ops_transfer_pinned,
142 cli->cl_r_in_flight, cli->cl_w_in_flight,
143 cli->cl_max_rpcs_in_flight,
145 waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
146 list_empty_marker(&cli->cl_loi_ready_list),
147 list_empty_marker(&cli->cl_loi_hp_ready_list),
148 list_empty_marker(&cli->cl_loi_write_list),
149 list_empty_marker(&cli->cl_loi_read_list),
151 list_empty_marker(&obj->oo_ready_item),
152 list_empty_marker(&obj->oo_hp_ready_item),
153 list_empty_marker(&obj->oo_write_item),
154 list_empty_marker(&obj->oo_read_item),
155 atomic_read(&obj->oo_nr_reads),
156 list_empty_marker(&obj->oo_reading_exts),
157 atomic_read(&obj->oo_nr_writes),
158 list_empty_marker(&obj->oo_hp_exts),
159 list_empty_marker(&obj->oo_urgent_exts));
162 static void osc_page_delete(const struct lu_env *env,
163 const struct cl_page_slice *slice)
165 struct osc_page *opg = cl2osc_page(slice);
166 struct osc_object *obj = osc_page_object(opg);
170 CDEBUG(D_TRACE, "%p\n", opg);
171 osc_page_transfer_put(env, opg);
173 if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
176 rc = osc_teardown_async_page(env, obj, opg);
178 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
179 "Trying to teardown failed: %d\n", rc);
183 osc_lru_del(osc_cli(obj), opg);
185 spin_lock(&obj->oo_tree_lock);
186 if (opg->ops_intree) {
187 value = radix_tree_delete(&obj->oo_tree,
194 spin_unlock(&obj->oo_tree_lock);
196 LASSERT(ergo(value != NULL, value == opg));
202 static void osc_page_clip(const struct lu_env *env,
203 const struct cl_page_slice *slice,
206 struct osc_page *opg = cl2osc_page(slice);
207 struct osc_async_page *oap = &opg->ops_oap;
209 CDEBUG(D_CACHE, "from %d, to %d\n", from, to);
211 opg->ops_from = from;
212 /* argument @to is exclusive, but @ops_to is inclusive */
213 opg->ops_to = to - 1;
214 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
217 static int osc_page_flush(const struct lu_env *env,
218 const struct cl_page_slice *slice,
221 struct osc_page *opg = cl2osc_page(slice);
224 rc = osc_flush_async_page(env, io, opg);
228 static void osc_page_touch(const struct lu_env *env,
229 const struct cl_page_slice *slice, size_t to)
231 struct osc_page *opg = cl2osc_page(slice);
232 struct cl_object *obj = osc2cl(osc_page_object(opg));
234 osc_page_touch_at(env, obj, osc_index(opg), to);
237 static const struct cl_page_operations osc_transient_page_ops = {
238 .cpo_print = osc_page_print,
239 .cpo_delete = osc_page_delete,
240 .cpo_clip = osc_page_clip,
243 static const struct cl_page_operations osc_page_ops = {
244 .cpo_print = osc_page_print,
245 .cpo_delete = osc_page_delete,
246 .cpo_clip = osc_page_clip,
247 .cpo_flush = osc_page_flush,
248 .cpo_page_touch = osc_page_touch,
251 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
252 struct cl_page *cl_page, pgoff_t index)
254 struct osc_object *osc = cl2osc(obj);
255 struct osc_page *opg = cl_object_page_slice(obj, cl_page);
256 struct osc_io *oio = osc_env_io(env);
260 opg->ops_to = PAGE_SIZE - 1;
262 INIT_LIST_HEAD(&opg->ops_lru);
264 result = osc_prep_async_page(osc, opg, cl_page, index << PAGE_SHIFT);
268 opg->ops_srvlock = osc_io_srvlock(oio);
270 if (cl_page->cp_type == CPT_TRANSIENT) {
271 cl_page_slice_add(cl_page, &opg->ops_cl, obj,
272 &osc_transient_page_ops);
273 } else if (cl_page->cp_type == CPT_CACHEABLE) {
274 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
275 /* reserve an LRU space for this page */
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)
305 struct osc_io *oio = osc_env_io(env);
306 struct osc_async_page *oap = &opg->ops_oap;
308 LASSERT(oap->oap_async_flags & ASYNC_READY);
309 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
311 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
312 oap->oap_page_off = opg->ops_from;
313 oap->oap_count = opg->ops_to - opg->ops_from + 1;
314 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
316 if (oio->oi_cap_sys_resource)
317 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
319 osc_page_transfer_get(opg, "transfer\0imm");
320 osc_page_transfer_add(env, opg, crt);
323 /* --------------- LRU page management ------------------ */
325 /* OSC is a natural place to manage LRU pages as applications are specialized
326 * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
327 * occupy more LRU slots. On the other hand, we should avoid using up all LRU
328 * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
329 * for free LRU slots - this will be very bad so the algorithm requires each
330 * OSC to free slots voluntarily to maintain a reasonable number of free slots
334 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
337 * LRU pages are freed in batch mode. OSC should at least free this
338 * number of pages to avoid running out of LRU slots.
340 static inline int lru_shrink_min(struct client_obd *cli)
342 return cli->cl_max_pages_per_rpc * 2;
346 * free this number at most otherwise it will take too long time to finsih.
348 static inline int lru_shrink_max(struct client_obd *cli)
350 return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
354 * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
355 * we should free slots aggressively. In this way, slots are freed in a steady
356 * step to maintain fairness among OSCs.
358 * Return how many LRU pages should be freed.
360 static int osc_cache_too_much(struct client_obd *cli)
362 struct cl_client_cache *cache = cli->cl_cache;
363 long pages = atomic_long_read(&cli->cl_lru_in_list);
364 unsigned long budget;
366 LASSERT(cache != NULL);
367 budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
369 /* if it's going to run out LRU slots, we should free some, but not
370 * too much to maintain faireness among OSCs. */
371 if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
373 return lru_shrink_max(cli);
374 else if (pages >= budget / 2)
375 return lru_shrink_min(cli);
377 time64_t duration = ktime_get_real_seconds();
380 /* knock out pages by duration of no IO activity */
381 duration -= cli->cl_lru_last_used;
383 * The difference shouldn't be more than 70 years
384 * so we can safely case to a long. Round to
385 * approximately 1 minute.
387 timediff = (long)(duration >> 6);
388 if (timediff > 0 && pages >= budget / timediff)
389 return lru_shrink_min(cli);
394 int lru_queue_work(const struct lu_env *env, void *data)
396 struct client_obd *cli = data;
399 CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
400 count = osc_cache_too_much(cli);
402 int rc = osc_lru_shrink(env, cli, count, false);
404 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
405 cli_name(cli), rc, count);
407 CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
408 ptlrpcd_queue_work(cli->cl_lru_work);
415 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
418 struct osc_async_page *oap;
421 list_for_each_entry(oap, plist, oap_pending_item) {
422 struct osc_page *opg = oap2osc_page(oap);
424 if (!opg->ops_in_lru)
428 LASSERT(list_empty(&opg->ops_lru));
429 list_add(&opg->ops_lru, &lru);
433 spin_lock(&cli->cl_lru_list_lock);
434 list_splice_tail(&lru, &cli->cl_lru_list);
435 atomic_long_sub(npages, &cli->cl_lru_busy);
436 atomic_long_add(npages, &cli->cl_lru_in_list);
437 cli->cl_lru_last_used = ktime_get_real_seconds();
438 spin_unlock(&cli->cl_lru_list_lock);
440 if (waitqueue_active(&osc_lru_waitq))
441 (void)ptlrpcd_queue_work(cli->cl_lru_work);
445 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
447 LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
448 list_del_init(&opg->ops_lru);
449 atomic_long_dec(&cli->cl_lru_in_list);
453 * Page is being destroyed. The page may be not in LRU list, if the transfer
454 * has never finished(error occurred).
456 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
458 if (opg->ops_in_lru) {
459 spin_lock(&cli->cl_lru_list_lock);
460 if (!list_empty(&opg->ops_lru)) {
461 __osc_lru_del(cli, opg);
463 LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
464 atomic_long_dec(&cli->cl_lru_busy);
466 spin_unlock(&cli->cl_lru_list_lock);
468 atomic_long_inc(cli->cl_lru_left);
469 /* this is a great place to release more LRU pages if
470 * this osc occupies too many LRU pages and kernel is
471 * stealing one of them. */
472 if (osc_cache_too_much(cli)) {
473 CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
474 (void)ptlrpcd_queue_work(cli->cl_lru_work);
476 wake_up(&osc_lru_waitq);
478 LASSERT(list_empty(&opg->ops_lru));
483 * Delete page from LRU list for redirty.
485 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
487 /* If page is being transferred for the first time,
488 * ops_lru should be empty */
489 if (opg->ops_in_lru) {
490 if (list_empty(&opg->ops_lru))
492 spin_lock(&cli->cl_lru_list_lock);
493 if (!list_empty(&opg->ops_lru)) {
494 __osc_lru_del(cli, opg);
495 atomic_long_inc(&cli->cl_lru_busy);
497 spin_unlock(&cli->cl_lru_list_lock);
501 static void discard_cl_pages(const struct lu_env *env, struct cl_io *io,
502 struct cl_page **pvec, int max_index)
504 struct folio_batch *fbatch = &osc_env_info(env)->oti_fbatch;
507 ll_folio_batch_init(fbatch, 0);
508 for (i = 0; i < max_index; i++) {
509 struct cl_page *page = pvec[i];
511 LASSERT(page->cp_type != CPT_TRANSIENT);
512 LASSERT(cl_page_is_owned(page, io));
513 cl_page_discard(env, io, page);
514 cl_page_disown(env, io, page);
515 cl_batch_put(env, page, fbatch);
519 folio_batch_release(fbatch);
523 * Check if a cl_page can be released, i.e, it's not being used.
525 * If unstable account is turned on, bulk transfer may hold one refcount
526 * for recovery so we need to check vmpage refcount as well; otherwise,
527 * even we can destroy cl_page but the corresponding vmpage can't be reused.
529 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
531 if (cl_page_in_use_noref(page))
534 if (cli->cl_cache->ccc_unstable_check) {
535 struct page *vmpage = cl_page_vmpage(page);
537 /* vmpage have two known users: cl_page and VM page cache */
538 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
545 * Drop @target of pages from LRU at most.
547 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
548 long target, bool force)
551 struct cl_object *clobj = NULL;
552 struct cl_page **pvec;
553 struct osc_page *opg;
560 LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
561 if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
564 CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
565 cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
567 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
570 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
571 atomic_dec(&cli->cl_lru_shrinkers);
575 atomic_inc(&cli->cl_lru_shrinkers);
578 pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
579 io = osc_env_thread_io(env);
581 spin_lock(&cli->cl_lru_list_lock);
583 cli->cl_lru_reclaim++;
584 maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
585 while (!list_empty(&cli->cl_lru_list)) {
586 struct cl_page *page;
587 bool will_free = false;
589 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
595 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
597 page = opg->ops_cl.cpl_page;
598 if (lru_page_busy(cli, page)) {
599 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
603 LASSERT(page->cp_obj != NULL);
604 if (clobj != page->cp_obj) {
605 struct cl_object *tmp = page->cp_obj;
608 spin_unlock(&cli->cl_lru_list_lock);
611 discard_cl_pages(env, io, pvec, index);
615 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_cl_pages(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_cl_pages(env, io, pvec, index);
672 cl_object_put(env, clobj);
676 atomic_dec(&cli->cl_lru_shrinkers);
678 atomic_long_add(count, cli->cl_lru_left);
679 wake_up(&osc_lru_waitq);
681 RETURN(count > 0 ? count : rc);
683 EXPORT_SYMBOL(osc_lru_shrink);
686 * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
687 * \@npages of LRU slots. For performance consideration, it's better to drop
688 * LRU pages in batch. Therefore, the actual number is adjusted at least
691 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
694 struct cl_client_cache *cache = cli->cl_cache;
695 struct client_obd *scan;
701 LASSERT(cache != NULL);
703 env = cl_env_get(&refcheck);
707 npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
708 CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
709 cli_name(cli), npages);
710 rc = osc_lru_shrink(env, cli, npages, true);
712 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
713 cli_name(cli), rc, npages);
714 if (osc_cache_too_much(cli) > 0)
715 ptlrpcd_queue_work(cli->cl_lru_work);
721 CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
722 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
723 atomic_long_read(&cli->cl_lru_busy), npages);
725 /* Reclaim LRU slots from other client_obd as it can't free enough
726 * from its own. This should rarely happen. */
727 spin_lock(&cache->ccc_lru_lock);
728 LASSERT(!list_empty(&cache->ccc_lru));
730 cache->ccc_lru_shrinkers++;
731 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
733 max_scans = refcount_read(&cache->ccc_users) - 2;
734 while (--max_scans > 0 &&
735 (scan = list_first_entry_or_null(&cache->ccc_lru,
737 cl_lru_osc)) != NULL) {
738 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
739 cli_name(scan), scan,
740 atomic_long_read(&scan->cl_lru_in_list),
741 atomic_long_read(&scan->cl_lru_busy));
743 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
744 if (osc_cache_too_much(scan) > 0) {
745 spin_unlock(&cache->ccc_lru_lock);
747 rc = osc_lru_shrink(env, scan, npages, true);
748 spin_lock(&cache->ccc_lru_lock);
755 spin_unlock(&cache->ccc_lru_lock);
758 cl_env_put(env, &refcheck);
759 CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
760 cli_name(cli), cli, rc);
765 * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
767 * Usually the LRU slots are reserved in osc_io_iter_rw_init().
768 * Only in the case that the LRU slots are in extreme shortage, it should
769 * have reserved enough slots for an IO.
771 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
772 struct osc_page *opg)
774 struct osc_io *oio = osc_env_io(env);
779 if (cli->cl_cache == NULL) /* shall not be in LRU */
782 if (oio->oi_lru_reserved > 0) {
783 --oio->oi_lru_reserved;
787 LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
788 while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
789 /* run out of LRU spaces, try to drop some by itself */
790 rc = osc_lru_reclaim(cli, 1);
795 /* IO issued by readahead, don't try hard */
796 if (oio->oi_is_readahead) {
797 if (atomic_long_read(cli->cl_lru_left) > 0)
804 rc = l_wait_event_abortable(
806 atomic_long_read(cli->cl_lru_left) > 0);
815 atomic_long_inc(&cli->cl_lru_busy);
824 * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
826 * The benefit of doing this is to reduce contention against atomic counter
827 * cl_lru_left by changing it from per-page access to per-IO access.
829 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
831 unsigned long reserved = 0;
832 unsigned long max_pages;
837 c = atomic_long_read(cli->cl_lru_left);
838 if (c < npages && osc_lru_reclaim(cli, npages) > 0)
839 c = atomic_long_read(cli->cl_lru_left);
843 * Trigger writeback in the hope some LRU slot could
846 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
851 while (c >= npages) {
852 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
856 c = atomic_long_read(cli->cl_lru_left);
859 if (reserved != npages) {
861 rc = l_wait_event_abortable(
863 atomic_long_read(cli->cl_lru_left) > 0);
867 max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
868 if (atomic_long_read(cli->cl_lru_left) < max_pages) {
869 /* If there aren't enough pages in the per-OSC LRU then
870 * wake up the LRU thread to try and clear out space, so
871 * we don't block if pages are being dirtied quickly. */
872 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
873 cli_name(cli), atomic_long_read(cli->cl_lru_left),
875 (void)ptlrpcd_queue_work(cli->cl_lru_work);
882 * osc_lru_unreserve() is called to unreserve LRU slots.
884 * LRU slots reserved by osc_lru_reserve() may have entries left due to several
885 * reasons such as page already existing or I/O error. Those reserved slots
886 * should be freed by calling this function.
888 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
890 atomic_long_add(npages, cli->cl_lru_left);
891 wake_up(&osc_lru_waitq);
895 * Atomic operations are expensive. We accumulate the accounting for the
896 * same page zone to get better performance.
897 * In practice this can work pretty good because the pages in the same RPC
898 * are likely from the same page zone.
900 #ifdef HAVE_NR_UNSTABLE_NFS
901 /* Old kernels use a separate counter for unstable pages,
902 * newer kernels treat them like any other writeback.
903 * (see Linux commit: v5.7-467-g8d92890bd6b8)
905 #define NR_ZONE_WRITE_PENDING ((enum zone_stat_item)NR_UNSTABLE_NFS)
906 #elif !defined(HAVE_NR_ZONE_WRITE_PENDING)
907 #define NR_ZONE_WRITE_PENDING ((enum zone_stat_item)NR_WRITEBACK)
910 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
920 page_count = desc->bd_iov_count;
922 CDEBUG(D_PAGE, "%s %d unstable pages\n",
923 factor == 1 ? "adding" : "removing", page_count);
925 for (i = 0; i < page_count; i++) {
926 void *pz = page_zone(desc->bd_vec[i].bv_page);
928 if (likely(pz == zone)) {
934 mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
942 mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
948 static inline void add_unstable_pages(struct ptlrpc_bulk_desc *desc)
950 unstable_page_accounting(desc, 1);
953 static inline void dec_unstable_pages(struct ptlrpc_bulk_desc *desc)
955 unstable_page_accounting(desc, -1);
959 * Performs "unstable" page accounting. This function balances the
960 * increment operations performed in osc_inc_unstable_pages. It is
961 * registered as the RPC request callback, and is executed when the
962 * bulk RPC is committed on the server. Thus at this point, the pages
963 * involved in the bulk transfer are no longer considered unstable.
965 * If this function is called, the request should have been committed
966 * or req:rq_unstable must have been set; it implies that the unstable
967 * statistic have been added.
969 void osc_dec_unstable_pages(struct ptlrpc_request *req)
971 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
972 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
976 /* no desc means short io, which doesn't have separate unstable pages,
977 * it's just using space inside the RPC itself
982 page_count = desc->bd_iov_count;
984 LASSERT(page_count >= 0);
986 dec_unstable_pages(desc);
988 unstable_count = atomic_long_sub_return(page_count,
989 &cli->cl_unstable_count);
990 LASSERT(unstable_count >= 0);
992 unstable_count = atomic_long_sub_return(page_count,
993 &cli->cl_cache->ccc_unstable_nr);
994 LASSERT(unstable_count >= 0);
996 if (waitqueue_active(&osc_lru_waitq))
997 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1001 * "unstable" page accounting. See: osc_dec_unstable_pages.
1003 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1005 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
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)
1013 /* no desc means short io, which doesn't have separate unstable pages,
1014 * it's just using space inside the RPC itself
1019 page_count = desc->bd_iov_count;
1021 add_unstable_pages(desc);
1022 atomic_long_add(page_count, &cli->cl_unstable_count);
1023 atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1025 /* If the request has already been committed (i.e. brw_commit
1026 * called via rq_commit_cb), we need to undo the unstable page
1027 * increments we just performed because rq_commit_cb wont be
1029 spin_lock(&req->rq_lock);
1030 if (unlikely(req->rq_committed)) {
1031 spin_unlock(&req->rq_lock);
1033 osc_dec_unstable_pages(req);
1035 req->rq_unstable = 1;
1036 spin_unlock(&req->rq_lock);
1041 * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1042 * This function will be called by every BRW RPC so it's critical
1043 * to make this function fast.
1045 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1047 long unstable_nr, osc_unstable_count;
1049 /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1050 if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1053 osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1054 unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1057 "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1058 cli_name(cli), cli, unstable_nr, osc_unstable_count);
1060 /* If the LRU slots are in shortage - 25% remaining AND this OSC
1061 * has one full RPC window of unstable pages, it's a good chance
1062 * to piggyback a SOFT_SYNC flag.
1063 * Please notice that the OST won't take immediate response for the
1064 * SOFT_SYNC request so active OSCs will have more chance to carry
1065 * the flag, this is reasonable. */
1066 return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1067 osc_unstable_count > cli->cl_max_pages_per_rpc *
1068 cli->cl_max_rpcs_in_flight;
1072 * Return how many LRU pages in the cache of all OSC devices
1074 * \retval return # of cached LRU pages times reclaimation tendency
1075 * \retval SHRINK_STOP if it cannot do any scanning in this time
1077 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1078 struct shrink_control *sc)
1080 struct client_obd *cli;
1081 unsigned long cached = 0;
1083 if (!osc_page_cache_shrink_enabled)
1086 spin_lock(&osc_shrink_lock);
1087 list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1088 cached += atomic_long_read(&cli->cl_lru_in_list);
1089 spin_unlock(&osc_shrink_lock);
1091 return (cached * sysctl_vfs_cache_pressure) / 100;
1095 * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1097 * \retval number of cached LRU pages reclaimed
1098 * \retval SHRINK_STOP if it cannot do any scanning in this time
1100 * Linux kernel will loop calling this shrinker scan routine with
1101 * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1103 * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1104 * to scan and try to reclaim LRU pages, just return 0 and
1105 * osc_cache_shrink_count() will report the LRU page number.
1107 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1108 struct shrink_control *sc)
1110 struct client_obd *cli;
1111 struct client_obd *stop_anchor = NULL;
1117 if (sc->nr_to_scan == 0)
1120 if (!(sc->gfp_mask & __GFP_FS))
1123 env = cl_env_get(&refcheck);
1127 spin_lock(&osc_shrink_lock);
1128 while ((cli = list_first_entry_or_null(&osc_shrink_list,
1130 cl_shrink_list)) != NULL) {
1131 if (stop_anchor == NULL)
1133 else if (cli == stop_anchor)
1136 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1137 spin_unlock(&osc_shrink_lock);
1139 /* shrink no more than max_pages_per_rpc for an OSC */
1140 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1141 cli->cl_max_pages_per_rpc ?
1142 cli->cl_max_pages_per_rpc :
1143 sc->nr_to_scan - shrank, true);
1147 if (shrank >= sc->nr_to_scan)
1150 spin_lock(&osc_shrink_lock);
1152 spin_unlock(&osc_shrink_lock);
1155 cl_env_put(env, &refcheck);