X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosc%2Fosc_page.c;h=ae7d01b526829d1dccc338c298f5ac025f6d4f5a;hb=9821754235e2417868dbaeb987b33e22c517346f;hp=f07f4622ca65f75bf8a06ede0112ef5de03dbfb2;hpb=7cb2e4b642b51b42f9f3a448251774929073e17c;p=fs%2Flustre-release.git diff --git a/lustre/osc/osc_page.c b/lustre/osc/osc_page.c index f07f462..ae7d01b 100644 --- a/lustre/osc/osc_page.c +++ b/lustre/osc/osc_page.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2014, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -40,128 +36,21 @@ */ #define DEBUG_SUBSYSTEM S_OSC +#include -#include "osc_cl_internal.h" +#include "osc_internal.h" static void osc_lru_del(struct client_obd *cli, struct osc_page *opg); static void osc_lru_use(struct client_obd *cli, struct osc_page *opg); -static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj, - struct osc_page *opg); +static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli, + struct osc_page *opg); /** \addtogroup osc * @{ */ /* - * Comment out osc_page_protected because it may sleep inside the - * the client_obd_list_lock. - * client_obd_list_lock -> osc_ap_completion -> osc_completion -> - * -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base - * -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep. - */ -#if 0 -static int osc_page_is_dlocked(const struct lu_env *env, - const struct osc_page *opg, - enum cl_lock_mode mode, int pending, int unref) -{ - struct cl_page *page; - struct osc_object *obj; - struct osc_thread_info *info; - struct ldlm_res_id *resname; - struct lustre_handle *lockh; - ldlm_policy_data_t *policy; - ldlm_mode_t dlmmode; - __u64 flags; - - might_sleep(); - - info = osc_env_info(env); - resname = &info->oti_resname; - policy = &info->oti_policy; - lockh = &info->oti_handle; - page = opg->ops_cl.cpl_page; - obj = cl2osc(opg->ops_cl.cpl_obj); - - flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED; - if (pending) - flags |= LDLM_FL_CBPENDING; - - dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW; - osc_lock_build_res(env, obj, resname); - osc_index2policy(policy, page->cp_obj, osc_index(opg), osc_index(opg)); - return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy, - dlmmode, &flags, NULL, lockh, unref); -} - -/** - * Checks an invariant that a page in the cache is covered by a lock, as - * needed. - */ -static int osc_page_protected(const struct lu_env *env, - const struct osc_page *opg, - enum cl_lock_mode mode, int unref) -{ - struct cl_object_header *hdr; - struct cl_lock *scan; - struct cl_page *page; - struct cl_lock_descr *descr; - int result; - - LINVRNT(!opg->ops_temp); - - page = opg->ops_cl.cpl_page; - if (page->cp_owner != NULL && - cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER) - /* - * If IO is done without locks (liblustre, or lloop), lock is - * not required. - */ - result = 1; - else - /* otherwise check for a DLM lock */ - result = osc_page_is_dlocked(env, opg, mode, 1, unref); - if (result == 0) { - /* maybe this page is a part of a lockless io? */ - hdr = cl_object_header(opg->ops_cl.cpl_obj); - descr = &osc_env_info(env)->oti_descr; - descr->cld_mode = mode; - descr->cld_start = osc_index(opg); - descr->cld_end = osc_index(opg); - spin_lock(&hdr->coh_lock_guard); - list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) { - /* - * Lock-less sub-lock has to be either in HELD state - * (when io is actively going on), or in CACHED state, - * when top-lock is being unlocked: - * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse(). - */ - if ((scan->cll_state == CLS_HELD || - scan->cll_state == CLS_CACHED) && - cl_lock_ext_match(&scan->cll_descr, descr)) { - struct osc_lock *olck; - - olck = osc_lock_at(scan); - result = osc_lock_is_lockless(olck); - break; - } - } - spin_unlock(&hdr->coh_lock_guard); - } - return result; -} -#else -static int osc_page_protected(const struct lu_env *env, - const struct osc_page *opg, - enum cl_lock_mode mode, int unref) -{ - return 1; -} -#endif - -/***************************************************************************** - * * Page operations. - * */ static void osc_page_transfer_get(struct osc_page *opg, const char *label) { @@ -195,27 +84,17 @@ static void osc_page_transfer_add(const struct lu_env *env, { struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); - /* ops_lru and ops_inflight share the same field, so take it from LRU - * first and then use it as inflight. */ osc_lru_use(osc_cli(obj), opg); - - spin_lock(&obj->oo_seatbelt); - list_add(&opg->ops_inflight, &obj->oo_inflight[crt]); - opg->ops_submitter = current; - spin_unlock(&obj->oo_seatbelt); } -int osc_page_cache_add(const struct lu_env *env, - const struct cl_page_slice *slice, struct cl_io *io) +int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg, + struct cl_io *io, cl_commit_cbt cb) { - struct osc_page *opg = cl2osc_page(slice); int result; ENTRY; - LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0)); - osc_page_transfer_get(opg, "transfer\0cache"); - result = osc_queue_async_io(env, io, opg); + result = osc_queue_async_io(env, io, opg, cb); if (result != 0) osc_page_transfer_put(env, opg); else @@ -224,89 +103,80 @@ int osc_page_cache_add(const struct lu_env *env, RETURN(result); } -void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj, - pgoff_t start, pgoff_t end) +void osc_index2policy(union ldlm_policy_data *policy, + const struct cl_object *obj, pgoff_t start, pgoff_t end) { - memset(policy, 0, sizeof *policy); - policy->l_extent.start = cl_offset(obj, start); - policy->l_extent.end = cl_offset(obj, end + 1) - 1; + memset(policy, 0, sizeof *policy); + policy->l_extent.start = cl_offset(obj, start); + policy->l_extent.end = cl_offset(obj, end + 1) - 1; } -static const char *osc_list(struct list_head *head) +static inline s64 osc_submit_duration(struct osc_page *opg) { - return list_empty(head) ? "-" : "+"; -} + if (ktime_to_ns(opg->ops_submit_time) == 0) + return 0; -static inline cfs_time_t osc_submit_duration(struct osc_page *opg) -{ - if (opg->ops_submit_time == 0) - return 0; - - return (cfs_time_current() - opg->ops_submit_time); + return ktime_ms_delta(ktime_get(), opg->ops_submit_time); } static int osc_page_print(const struct lu_env *env, - const struct cl_page_slice *slice, - void *cookie, lu_printer_t printer) + const struct cl_page_slice *slice, + void *cookie, lu_printer_t printer) { - struct osc_page *opg = cl2osc_page(slice); - struct osc_async_page *oap = &opg->ops_oap; - struct osc_object *obj = cl2osc(slice->cpl_obj); - struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli; + struct osc_page *opg = cl2osc_page(slice); + struct osc_async_page *oap = &opg->ops_oap; + struct osc_object *obj = cl2osc(slice->cpl_obj); + struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli; return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: " - "1< %#x %d %u %s %s > " - "2< "LPD64" %u %u %#x %#x | %p %p %p > " - "3< %s %p %d %lu %d > " - "4< %d %d %d %lu %s | %s %s %s %s > " - "5< %s %s %s %s | %d %s | %d %s %s>\n", + "1< %#x %d %c %c > " + "2< %lld %u %u %#x %#x | %p %p %p > " + "3< %d %lld %d > " + "4< %d %d %d %lu %c | %c %c %c %c > " + "5< %c %c %c %c | %d %c | %d %c %c>\n", opg, osc_index(opg), - /* 1 */ - oap->oap_magic, oap->oap_cmd, - oap->oap_interrupted, - osc_list(&oap->oap_pending_item), - osc_list(&oap->oap_rpc_item), - /* 2 */ - oap->oap_obj_off, oap->oap_page_off, oap->oap_count, - oap->oap_async_flags, oap->oap_brw_flags, + /* 1 */ + oap->oap_magic, oap->oap_cmd, + list_empty_marker(&oap->oap_pending_item), + list_empty_marker(&oap->oap_rpc_item), + /* 2 */ + oap->oap_obj_off, oap->oap_page_off, oap->oap_count, + oap->oap_async_flags, oap->oap_brw_flags, oap->oap_request, oap->oap_cli, obj, - /* 3 */ - osc_list(&opg->ops_inflight), - opg->ops_submitter, opg->ops_transfer_pinned, - osc_submit_duration(opg), opg->ops_srvlock, - /* 4 */ - cli->cl_r_in_flight, cli->cl_w_in_flight, - cli->cl_max_rpcs_in_flight, - cli->cl_avail_grant, - osc_list(&cli->cl_cache_waiters), - osc_list(&cli->cl_loi_ready_list), - osc_list(&cli->cl_loi_hp_ready_list), - osc_list(&cli->cl_loi_write_list), - osc_list(&cli->cl_loi_read_list), - /* 5 */ - osc_list(&obj->oo_ready_item), - osc_list(&obj->oo_hp_ready_item), - osc_list(&obj->oo_write_item), - osc_list(&obj->oo_read_item), + /* 3 */ + opg->ops_transfer_pinned, + osc_submit_duration(opg), opg->ops_srvlock, + /* 4 */ + cli->cl_r_in_flight, cli->cl_w_in_flight, + cli->cl_max_rpcs_in_flight, + cli->cl_avail_grant, + waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-', + list_empty_marker(&cli->cl_loi_ready_list), + list_empty_marker(&cli->cl_loi_hp_ready_list), + list_empty_marker(&cli->cl_loi_write_list), + list_empty_marker(&cli->cl_loi_read_list), + /* 5 */ + list_empty_marker(&obj->oo_ready_item), + list_empty_marker(&obj->oo_hp_ready_item), + list_empty_marker(&obj->oo_write_item), + list_empty_marker(&obj->oo_read_item), atomic_read(&obj->oo_nr_reads), - osc_list(&obj->oo_reading_exts), + list_empty_marker(&obj->oo_reading_exts), atomic_read(&obj->oo_nr_writes), - osc_list(&obj->oo_hp_exts), - osc_list(&obj->oo_urgent_exts)); + list_empty_marker(&obj->oo_hp_exts), + list_empty_marker(&obj->oo_urgent_exts)); } static void osc_page_delete(const struct lu_env *env, - const struct cl_page_slice *slice) + const struct cl_page_slice *slice) { struct osc_page *opg = cl2osc_page(slice); struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); - int rc; - - LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1)); + int rc; - ENTRY; - CDEBUG(D_TRACE, "%p\n", opg); - osc_page_transfer_put(env, opg); + ENTRY; + CDEBUG(D_TRACE, "%p\n", opg); + osc_page_transfer_put(env, opg); rc = osc_teardown_async_page(env, obj, opg); if (rc) { CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page, @@ -314,23 +184,20 @@ static void osc_page_delete(const struct lu_env *env, LASSERT(0); } - spin_lock(&obj->oo_seatbelt); - if (opg->ops_submitter != NULL) { - LASSERT(!list_empty(&opg->ops_inflight)); - list_del_init(&opg->ops_inflight); - opg->ops_submitter = NULL; - } - spin_unlock(&obj->oo_seatbelt); - osc_lru_del(osc_cli(obj), opg); if (slice->cpl_page->cp_type == CPT_CACHEABLE) { - void *value; + void *value = NULL; spin_lock(&obj->oo_tree_lock); - value = radix_tree_delete(&obj->oo_tree, osc_index(opg)); - if (value != NULL) - --obj->oo_npages; + if (opg->ops_intree) { + value = radix_tree_delete(&obj->oo_tree, + osc_index(opg)); + if (value != NULL) { + --obj->oo_npages; + opg->ops_intree = 0; + } + } spin_unlock(&obj->oo_tree_lock); LASSERT(ergo(value != NULL, value == opg)); @@ -343,35 +210,17 @@ static void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice, int from, int to) { - struct osc_page *opg = cl2osc_page(slice); - struct osc_async_page *oap = &opg->ops_oap; - - LINVRNT(osc_page_protected(env, opg, CLM_READ, 0)); + struct osc_page *opg = cl2osc_page(slice); + struct osc_async_page *oap = &opg->ops_oap; - opg->ops_from = from; - opg->ops_to = to; + opg->ops_from = from; + /* argument @to is exclusive, but @ops_to is inclusive */ + opg->ops_to = to - 1; spin_lock(&oap->oap_lock); oap->oap_async_flags |= ASYNC_COUNT_STABLE; spin_unlock(&oap->oap_lock); } -static int osc_page_cancel(const struct lu_env *env, - const struct cl_page_slice *slice) -{ - struct osc_page *opg = cl2osc_page(slice); - int rc = 0; - - LINVRNT(osc_page_protected(env, opg, CLM_READ, 0)); - - /* Check if the transferring against this page - * is completed, or not even queued. */ - if (opg->ops_transfer_pinned) - /* FIXME: may not be interrupted.. */ - rc = osc_cancel_async_page(env, opg); - LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0)); - return rc; -} - static int osc_page_flush(const struct lu_env *env, const struct cl_page_slice *slice, struct cl_io *io) @@ -383,59 +232,68 @@ static int osc_page_flush(const struct lu_env *env, RETURN(rc); } +static void osc_page_touch(const struct lu_env *env, + const struct cl_page_slice *slice, size_t to) +{ + struct osc_page *opg = cl2osc_page(slice); + struct cl_object *obj = opg->ops_cl.cpl_obj; + + osc_page_touch_at(env, obj, osc_index(opg), to); +} + static const struct cl_page_operations osc_page_ops = { .cpo_print = osc_page_print, .cpo_delete = osc_page_delete, .cpo_clip = osc_page_clip, - .cpo_cancel = osc_page_cancel, - .cpo_flush = osc_page_flush + .cpo_flush = osc_page_flush, + .cpo_page_touch = osc_page_touch, }; int osc_page_init(const struct lu_env *env, struct cl_object *obj, - struct cl_page *page, pgoff_t index) + struct cl_page *cl_page, pgoff_t index) { struct osc_object *osc = cl2osc(obj); - struct osc_page *opg = cl_object_page_slice(obj, page); + struct osc_page *opg = cl_object_page_slice(obj, cl_page); + struct osc_io *oio = osc_env_io(env); int result; opg->ops_from = 0; - opg->ops_to = PAGE_CACHE_SIZE; + opg->ops_to = PAGE_SIZE - 1; - result = osc_prep_async_page(osc, opg, page->cp_vmpage, - cl_offset(obj, index)); - if (result == 0) { - struct osc_io *oio = osc_env_io(env); - opg->ops_srvlock = osc_io_srvlock(oio); - cl_page_slice_add(page, &opg->ops_cl, obj, index, - &osc_page_ops); - } - /* - * Cannot assert osc_page_protected() here as read-ahead - * creates temporary pages outside of a lock. - */ -#ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK - opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1); -#endif - /* ops_inflight and ops_lru are the same field, but it doesn't - * hurt to initialize it twice :-) */ - INIT_LIST_HEAD(&opg->ops_inflight); INIT_LIST_HEAD(&opg->ops_lru); + result = osc_prep_async_page(osc, opg, cl_page->cp_vmpage, + cl_offset(obj, index)); + if (result != 0) + return result; + + opg->ops_srvlock = osc_io_srvlock(oio); + cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops); + cl_page->cp_osc_index = index; + /* reserve an LRU space for this page */ - if (page->cp_type == CPT_CACHEABLE && result == 0) { - result = osc_lru_reserve(env, osc, opg); + if (cl_page->cp_type == CPT_CACHEABLE) { + result = osc_lru_alloc(env, osc_cli(osc), opg); if (result == 0) { - spin_lock(&osc->oo_tree_lock); - result = radix_tree_insert(&osc->oo_tree, index, opg); - if (result == 0) - ++osc->oo_npages; - spin_unlock(&osc->oo_tree_lock); - LASSERT(result == 0); + result = radix_tree_preload(GFP_NOFS); + if (result == 0) { + spin_lock(&osc->oo_tree_lock); + result = radix_tree_insert(&osc->oo_tree, + index, opg); + if (result == 0) { + ++osc->oo_npages; + opg->ops_intree = 1; + } + spin_unlock(&osc->oo_tree_lock); + + radix_tree_preload_end(); + } } } return result; } +EXPORT_SYMBOL(osc_page_init); /** * Helper function called by osc_io_submit() for every page in an immediate @@ -444,29 +302,25 @@ int osc_page_init(const struct lu_env *env, struct cl_object *obj, void osc_page_submit(const struct lu_env *env, struct osc_page *opg, enum cl_req_type crt, int brw_flags) { + struct osc_io *oio = osc_env_io(env); struct osc_async_page *oap = &opg->ops_oap; - struct osc_object *obj = oap->oap_obj; - - LINVRNT(osc_page_protected(env, opg, - crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1)); LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, " "magic 0x%x\n", oap, oap->oap_magic); LASSERT(oap->oap_async_flags & ASYNC_READY); LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE); - oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ; - oap->oap_page_off = opg->ops_from; - oap->oap_count = opg->ops_to - opg->ops_from; + oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ; + oap->oap_page_off = opg->ops_from; + oap->oap_count = opg->ops_to - opg->ops_from + 1; oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags; - if (!client_is_remote(osc_export(obj)) && - cfs_capable(CFS_CAP_SYS_RESOURCE)) { + if (oio->oi_cap_sys_resource) { oap->oap_brw_flags |= OBD_BRW_NOQUOTA; oap->oap_cmd |= OBD_BRW_NOQUOTA; } - opg->ops_submit_time = cfs_time_current(); + opg->ops_submit_time = ktime_get(); osc_page_transfer_get(opg, "transfer\0imm"); osc_page_transfer_add(env, opg, crt); } @@ -483,17 +337,31 @@ void osc_page_submit(const struct lu_env *env, struct osc_page *opg, */ static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq); -/* LRU pages are freed in batch mode. OSC should at least free this - * number of pages to avoid running out of LRU budget, and.. */ -static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */ -/* free this number at most otherwise it will take too long time to finsih. */ -static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */ -/* Check if we can free LRU slots from this OSC. If there exists LRU waiters, +/** + * LRU pages are freed in batch mode. OSC should at least free this + * number of pages to avoid running out of LRU slots. + */ +static inline int lru_shrink_min(struct client_obd *cli) +{ + return cli->cl_max_pages_per_rpc * 2; +} + +/** + * free this number at most otherwise it will take too long time to finsih. + */ +static inline int lru_shrink_max(struct client_obd *cli) +{ + return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight; +} + +/** + * Check if we can free LRU slots from this OSC. If there exists LRU waiters, * we should free slots aggressively. In this way, slots are freed in a steady * step to maintain fairness among OSCs. * - * Return how many LRU pages should be freed. */ + * Return how many LRU pages should be freed. + */ static int osc_cache_too_much(struct client_obd *cli) { struct cl_client_cache *cache = cli->cl_cache; @@ -505,15 +373,25 @@ static int osc_cache_too_much(struct client_obd *cli) /* if it's going to run out LRU slots, we should free some, but not * too much to maintain faireness among OSCs. */ - if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) { + if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) { if (pages >= budget) - return lru_shrink_max; + return lru_shrink_max(cli); else if (pages >= budget / 2) - return lru_shrink_min; -#if 0 - } else if (pages >= budget * 2) - return lru_shrink_min; -#endif + return lru_shrink_min(cli); + } else { + time64_t duration = ktime_get_real_seconds(); + long timediff; + + /* knock out pages by duration of no IO activity */ + duration -= cli->cl_lru_last_used; + /* + * The difference shouldn't be more than 70 years + * so we can safely case to a long. Round to + * approximately 1 minute. + */ + timediff = (long)(duration >> 6); + if (timediff > 0 && pages >= budget / timediff) + return lru_shrink_min(cli); } return 0; } @@ -521,18 +399,27 @@ static int osc_cache_too_much(struct client_obd *cli) int lru_queue_work(const struct lu_env *env, void *data) { struct client_obd *cli = data; + int count; - CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli); + CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli)); + count = osc_cache_too_much(cli); + if (count > 0) { + int rc = osc_lru_shrink(env, cli, count, false); - if (osc_cache_too_much(cli)) - osc_lru_shrink(env, cli, lru_shrink_max, true); + CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n", + cli_name(cli), rc, count); + if (rc >= count) { + CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli)); + ptlrpcd_queue_work(cli->cl_lru_work); + } + } RETURN(0); } void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist) { - struct list_head lru = LIST_HEAD_INIT(lru); + LIST_HEAD(lru); struct osc_async_page *oap; long npages = 0; @@ -552,10 +439,10 @@ void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist) list_splice_tail(&lru, &cli->cl_lru_list); atomic_long_sub(npages, &cli->cl_lru_busy); atomic_long_add(npages, &cli->cl_lru_in_list); + cli->cl_lru_last_used = ktime_get_real_seconds(); spin_unlock(&cli->cl_lru_list_lock); - /* XXX: May set force to be true for better performance */ - if (osc_cache_too_much(cli)) + if (waitqueue_active(&osc_lru_waitq)) (void)ptlrpcd_queue_work(cli->cl_lru_work); } } @@ -587,8 +474,10 @@ static void osc_lru_del(struct client_obd *cli, struct osc_page *opg) /* this is a great place to release more LRU pages if * this osc occupies too many LRU pages and kernel is * stealing one of them. */ - if (!memory_pressure_get()) + if (osc_cache_too_much(cli)) { + CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli)); (void)ptlrpcd_queue_work(cli->cl_lru_work); + } wake_up(&osc_lru_waitq); } else { LASSERT(list_empty(&opg->ops_lru)); @@ -596,35 +485,43 @@ static void osc_lru_del(struct client_obd *cli, struct osc_page *opg) } /** - * Delete page from LRUlist for redirty. + * Delete page from LRU list for redirty. */ static void osc_lru_use(struct client_obd *cli, struct osc_page *opg) { /* If page is being transferred for the first time, * ops_lru should be empty */ - if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) { + if (opg->ops_in_lru) { + if (list_empty(&opg->ops_lru)) + return; spin_lock(&cli->cl_lru_list_lock); - __osc_lru_del(cli, opg); + if (!list_empty(&opg->ops_lru)) { + __osc_lru_del(cli, opg); + atomic_long_inc(&cli->cl_lru_busy); + } spin_unlock(&cli->cl_lru_list_lock); - atomic_long_inc(&cli->cl_lru_busy); } } static void discard_pagevec(const struct lu_env *env, struct cl_io *io, struct cl_page **pvec, int max_index) { - int i; + struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec; + int i; - for (i = 0; i < max_index; i++) { - struct cl_page *page = pvec[i]; + ll_pagevec_init(pagevec, 0); + for (i = 0; i < max_index; i++) { + struct cl_page *page = pvec[i]; LASSERT(cl_page_is_owned(page, io)); + cl_page_delete(env, page); cl_page_discard(env, io, page); cl_page_disown(env, io, page); - cl_page_put(env, page); + cl_pagevec_put(env, page, pagevec); - pvec[i] = NULL; - } + pvec[i] = NULL; + } + pagevec_release(pagevec); } /** @@ -669,6 +566,8 @@ long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli, if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0) RETURN(0); + CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n", + cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force); if (!force) { if (atomic_read(&cli->cl_lru_shrinkers) > 0) RETURN(-EBUSY); @@ -682,14 +581,19 @@ long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli, } pvec = (struct cl_page **)osc_env_info(env)->oti_pvec; - io = &osc_env_info(env)->oti_io; + io = osc_env_thread_io(env); spin_lock(&cli->cl_lru_list_lock); + if (force) + cli->cl_lru_reclaim++; maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list)); while (!list_empty(&cli->cl_lru_list)) { struct cl_page *page; bool will_free = false; + if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1) + break; + if (--maxscan < 0) break; @@ -779,36 +683,46 @@ long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli, } RETURN(count > 0 ? count : rc); } +EXPORT_SYMBOL(osc_lru_shrink); -long osc_lru_reclaim(struct client_obd *cli) +/** + * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least + * \@npages of LRU slots. For performance consideration, it's better to drop + * LRU pages in batch. Therefore, the actual number is adjusted at least + * max_pages_per_rpc. + */ +static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages) { - struct cl_env_nest nest; struct lu_env *env; struct cl_client_cache *cache = cli->cl_cache; - long rc = 0; int max_scans; + __u16 refcheck; + long rc = 0; ENTRY; LASSERT(cache != NULL); - env = cl_env_nested_get(&nest); + env = cl_env_get(&refcheck); if (IS_ERR(env)) RETURN(rc); - rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false); - if (rc != 0) { - if (rc == -EBUSY) - rc = 0; - - CDEBUG(D_CACHE, "%s: Free %ld pages from own LRU: %p.\n", - cli->cl_import->imp_obd->obd_name, rc, cli); + npages = max_t(int, npages, cli->cl_max_pages_per_rpc); + CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n", + cli_name(cli), npages); + rc = osc_lru_shrink(env, cli, npages, true); + if (rc >= npages) { + CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n", + cli_name(cli), rc, npages); + if (osc_cache_too_much(cli) > 0) + ptlrpcd_queue_work(cli->cl_lru_work); GOTO(out, rc); + } else if (rc > 0) { + npages -= rc; } - CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld, busy: %ld.\n", - cli->cl_import->imp_obd->obd_name, cli, - atomic_long_read(&cli->cl_lru_in_list), - atomic_long_read(&cli->cl_lru_busy)); + CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n", + cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list), + atomic_long_read(&cli->cl_lru_busy), npages); /* Reclaim LRU slots from other client_obd as it can't free enough * from its own. This should rarely happen. */ @@ -824,7 +738,7 @@ long osc_lru_reclaim(struct client_obd *cli) cl_lru_osc); CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n", - cli->cl_import->imp_obd->obd_name, cli, + cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list), atomic_long_read(&cli->cl_lru_busy)); @@ -832,36 +746,36 @@ long osc_lru_reclaim(struct client_obd *cli) if (osc_cache_too_much(cli) > 0) { spin_unlock(&cache->ccc_lru_lock); - rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), - true); + rc = osc_lru_shrink(env, cli, npages, true); spin_lock(&cache->ccc_lru_lock); - if (rc != 0) + if (rc >= npages) break; + if (rc > 0) + npages -= rc; } } spin_unlock(&cache->ccc_lru_lock); out: - cl_env_nested_put(&nest, env); + cl_env_put(env, &refcheck); CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n", - cli->cl_import->imp_obd->obd_name, cli, rc); + cli_name(cli), cli, rc); return rc; } /** - * osc_lru_reserve() is called to reserve an LRU slot for a cl_page. + * osc_lru_alloc() is called to allocate an LRU slot for a cl_page. * * Usually the LRU slots are reserved in osc_io_iter_rw_init(). * Only in the case that the LRU slots are in extreme shortage, it should * have reserved enough slots for an IO. */ -static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj, - struct osc_page *opg) +static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli, + struct osc_page *opg) { - struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); struct osc_io *oio = osc_env_io(env); - struct client_obd *cli = osc_cli(obj); int rc = 0; + ENTRY; if (cli->cl_cache == NULL) /* shall not be in LRU */ @@ -874,20 +788,21 @@ static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj, LASSERT(atomic_long_read(cli->cl_lru_left) >= 0); while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) { - /* run out of LRU spaces, try to drop some by itself */ - rc = osc_lru_reclaim(cli); + rc = osc_lru_reclaim(cli, 1); if (rc < 0) break; if (rc > 0) continue; cond_resched(); - rc = l_wait_event(osc_lru_waitq, - atomic_long_read(cli->cl_lru_left) > 0, - &lwi); - if (rc < 0) + rc = l_wait_event_abortable( + osc_lru_waitq, + atomic_long_read(cli->cl_lru_left) > 0); + if (rc < 0) { + rc = -EINTR; break; + } } out: @@ -901,23 +816,86 @@ out: } /** + * osc_lru_reserve() is called to reserve enough LRU slots for I/O. + * + * The benefit of doing this is to reduce contention against atomic counter + * cl_lru_left by changing it from per-page access to per-IO access. + */ +unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages) +{ + unsigned long reserved = 0; + unsigned long max_pages; + unsigned long c; + + /* reserve a full RPC window at most to avoid that a thread accidentally + * consumes too many LRU slots */ + max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight; + if (npages > max_pages) + npages = max_pages; + + c = atomic_long_read(cli->cl_lru_left); + if (c < npages && osc_lru_reclaim(cli, npages) > 0) + c = atomic_long_read(cli->cl_lru_left); + while (c >= npages) { + if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) { + reserved = npages; + break; + } + c = atomic_long_read(cli->cl_lru_left); + } + if (atomic_long_read(cli->cl_lru_left) < max_pages) { + /* If there aren't enough pages in the per-OSC LRU then + * wake up the LRU thread to try and clear out space, so + * we don't block if pages are being dirtied quickly. */ + CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n", + cli_name(cli), atomic_long_read(cli->cl_lru_left), + max_pages); + (void)ptlrpcd_queue_work(cli->cl_lru_work); + } + + return reserved; +} + +/** + * osc_lru_unreserve() is called to unreserve LRU slots. + * + * LRU slots reserved by osc_lru_reserve() may have entries left due to several + * reasons such as page already existing or I/O error. Those reserved slots + * should be freed by calling this function. + */ +void osc_lru_unreserve(struct client_obd *cli, unsigned long npages) +{ + atomic_long_add(npages, cli->cl_lru_left); + wake_up_all(&osc_lru_waitq); +} + +/** * Atomic operations are expensive. We accumulate the accounting for the * same page zone to get better performance. * In practice this can work pretty good because the pages in the same RPC * are likely from the same page zone. */ static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc, + struct osc_brw_async_args *aa, int factor) { - int page_count = desc->bd_iov_count; + int page_count; void *zone = NULL; int count = 0; int i; - LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type)); + if (desc != NULL) { + page_count = desc->bd_iov_count; + } else { + page_count = aa->aa_page_count; + } for (i = 0; i < page_count; i++) { - void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page); + void *pz; + if (desc) + pz = page_zone(desc->bd_vec[i].bv_page); + else + pz = page_zone(aa->aa_ppga[i]->pg); if (likely(pz == zone)) { ++count; @@ -936,14 +914,16 @@ static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc, mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count); } -static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc) +static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc, + struct osc_brw_async_args *aa) { - unstable_page_accounting(desc, 1); + unstable_page_accounting(desc, aa, 1); } -static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc) +static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc, + struct osc_brw_async_args *aa) { - unstable_page_accounting(desc, -1); + unstable_page_accounting(desc, aa, -1); } /** @@ -960,12 +940,19 @@ static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc) void osc_dec_unstable_pages(struct ptlrpc_request *req) { struct ptlrpc_bulk_desc *desc = req->rq_bulk; + struct osc_brw_async_args *aa = (void *)&req->rq_async_args; struct client_obd *cli = &req->rq_import->imp_obd->u.cli; - int page_count = desc->bd_iov_count; + int page_count; long unstable_count; + if (desc) + page_count = desc->bd_iov_count; + else + page_count = aa->aa_page_count; + LASSERT(page_count >= 0); - dec_unstable_page_accounting(desc); + + dec_unstable_page_accounting(desc, aa); unstable_count = atomic_long_sub_return(page_count, &cli->cl_unstable_count); @@ -977,7 +964,7 @@ void osc_dec_unstable_pages(struct ptlrpc_request *req) if (unstable_count == 0) wake_up_all(&cli->cl_cache->ccc_unstable_waitq); - if (osc_cache_too_much(cli)) + if (waitqueue_active(&osc_lru_waitq)) (void)ptlrpcd_queue_work(cli->cl_lru_work); } @@ -987,14 +974,20 @@ void osc_dec_unstable_pages(struct ptlrpc_request *req) void osc_inc_unstable_pages(struct ptlrpc_request *req) { struct ptlrpc_bulk_desc *desc = req->rq_bulk; + struct osc_brw_async_args *aa = (void *)&req->rq_async_args; struct client_obd *cli = &req->rq_import->imp_obd->u.cli; - long page_count = desc->bd_iov_count; + long page_count; /* No unstable page tracking */ if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check) return; - add_unstable_page_accounting(desc); + if (desc) + page_count = desc->bd_iov_count; + else + page_count = aa->aa_page_count; + + add_unstable_page_accounting(desc, aa); atomic_long_add(page_count, &cli->cl_unstable_count); atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr); @@ -1031,8 +1024,7 @@ bool osc_over_unstable_soft_limit(struct client_obd *cli) CDEBUG(D_CACHE, "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n", - cli->cl_import->imp_obd->obd_name, cli, - unstable_nr, osc_unstable_count); + cli_name(cli), cli, unstable_nr, osc_unstable_count); /* If the LRU slots are in shortage - 25% remaining AND this OSC * has one full RPC window of unstable pages, it's a good chance @@ -1045,4 +1037,91 @@ bool osc_over_unstable_soft_limit(struct client_obd *cli) cli->cl_max_rpcs_in_flight; } +/** + * Return how many LRU pages in the cache of all OSC devices + * + * \retval return # of cached LRU pages times reclaimation tendency + * \retval SHRINK_STOP if it cannot do any scanning in this time + */ +unsigned long osc_cache_shrink_count(struct shrinker *sk, + struct shrink_control *sc) +{ + struct client_obd *cli; + unsigned long cached = 0; + + spin_lock(&osc_shrink_lock); + list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list) + cached += atomic_long_read(&cli->cl_lru_in_list); + spin_unlock(&osc_shrink_lock); + + return (cached * sysctl_vfs_cache_pressure) / 100; +} + +/** + * Scan and try to reclaim sc->nr_to_scan cached LRU pages + * + * \retval number of cached LRU pages reclaimed + * \retval SHRINK_STOP if it cannot do any scanning in this time + * + * Linux kernel will loop calling this shrinker scan routine with + * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory. + * + * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need + * to scan and try to reclaim LRU pages, just return 0 and + * osc_cache_shrink_count() will report the LRU page number. + */ +unsigned long osc_cache_shrink_scan(struct shrinker *sk, + struct shrink_control *sc) +{ + struct client_obd *cli; + struct client_obd *stop_anchor = NULL; + struct lu_env *env; + long shrank = 0; + int rc; + __u16 refcheck; + + if (sc->nr_to_scan == 0) + return 0; + + if (!(sc->gfp_mask & __GFP_FS)) + return SHRINK_STOP; + + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + return SHRINK_STOP; + + spin_lock(&osc_shrink_lock); + while (!list_empty(&osc_shrink_list)) { + cli = list_entry(osc_shrink_list.next, struct client_obd, + cl_shrink_list); + + if (stop_anchor == NULL) + stop_anchor = cli; + else if (cli == stop_anchor) + break; + + list_move_tail(&cli->cl_shrink_list, &osc_shrink_list); + spin_unlock(&osc_shrink_lock); + + /* shrink no more than max_pages_per_rpc for an OSC */ + rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) > + cli->cl_max_pages_per_rpc ? + cli->cl_max_pages_per_rpc : + sc->nr_to_scan - shrank, true); + if (rc > 0) + shrank += rc; + + if (shrank >= sc->nr_to_scan) + goto out; + + spin_lock(&osc_shrink_lock); + } + spin_unlock(&osc_shrink_lock); + +out: + cl_env_put(env, &refcheck); + + return shrank; +} + /** @} osc */