X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_request.c;h=5d718c972adeceb678dfc7f3acb007871ff4ae2d;hb=e70798ccca271c77a96494285c39f5141f6de8e0;hp=56c48ff35a6ca90f4cd1d2f0a5f9bded4dd0c265;hpb=c1f6b32958c799412c830f35f8d16ed7275407ea;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_request.c b/lustre/ldlm/ldlm_request.c index 56c48ff..5d718c9 100644 --- a/lustre/ldlm/ldlm_request.c +++ b/lustre/ldlm/ldlm_request.c @@ -1,25 +1,37 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * Copyright (C) 2002, 2003 Cluster File Systems, Inc. + * GPL HEADER START * - * This file is part of the Lustre file system, http://www.lustre.org - * Lustre is a trademark of Cluster File Systems, Inc. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * You may have signed or agreed to another license before downloading - * this software. If so, you are bound by the terms and conditions - * of that agreement, and the following does not apply to you. See the - * LICENSE file included with this distribution for more information. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * If you did not agree to a different license, then this copy of Lustre - * is open source software; you can redistribute it and/or modify it - * under the terms of version 2 of the GNU General Public License as - * published by the Free Software Foundation. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * In either case, Lustre is distributed in the hope that it will be - * useful, but WITHOUT ANY WARRANTY; without even the implied warranty - * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * license text for more details. + * 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. + * + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. */ #define DEBUG_SUBSYSTEM S_LDLM @@ -34,6 +46,13 @@ #include "ldlm_internal.h" +int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; +CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644, + "lock enqueue timeout minimum"); + +/* in client side, whether the cached locks will be canceled before replay */ +unsigned int ldlm_cancel_unused_locks_before_replay = 1; + static void interrupted_completion_wait(void *data) { } @@ -62,15 +81,16 @@ int ldlm_expired_completion_wait(void *data) RETURN(0); LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", " - CFS_DURATION_T"s ago); not entering recovery in " + CFS_DURATION_T"s ago); not entering recovery in " "server code, just going back to sleep", - lock->l_enqueued_time.tv_sec, - cfs_time_current_sec() - lock->l_enqueued_time.tv_sec); + lock->l_last_activity, + cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity)); if (cfs_time_after(cfs_time_current(), next_dump)) { last_dump = next_dump; next_dump = cfs_time_shift(300); ldlm_namespace_dump(D_DLMTRACE, - lock->l_resource->lr_namespace); + ldlm_lock_to_ns(lock)); if (last_dump == 0) libcfs_debug_dumplog(); } @@ -82,27 +102,102 @@ int ldlm_expired_completion_wait(void *data) ptlrpc_fail_import(imp, lwd->lwd_conn_cnt); LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", " CFS_DURATION_T"s ago), entering recovery for %s@%s", - lock->l_enqueued_time.tv_sec, - cfs_time_current_sec() - lock->l_enqueued_time.tv_sec, + lock->l_last_activity, + cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity), obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid); RETURN(0); } -static int is_granted_or_cancelled(struct ldlm_lock *lock) +/* We use the same basis for both server side and client side functions + from a single node. */ +int ldlm_get_enq_timeout(struct ldlm_lock *lock) { - int ret = 0; + int timeout = at_get(ldlm_lock_to_ns_at(lock)); + if (AT_OFF) + return obd_timeout / 2; + /* Since these are non-updating timeouts, we should be conservative. + It would be nice to have some kind of "early reply" mechanism for + lock callbacks too... */ + timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */ + return max(timeout, ldlm_enqueue_min); +} +EXPORT_SYMBOL(ldlm_get_enq_timeout); - lock_res_and_lock(lock); - if (((lock->l_req_mode == lock->l_granted_mode) && - !(lock->l_flags & LDLM_FL_CP_REQD)) || - (lock->l_flags & LDLM_FL_FAILED)) - ret = 1; - unlock_res_and_lock(lock); +/** + * Helper function for ldlm_completion_ast(), updating timings when lock is + * actually granted. + */ +static int ldlm_completion_tail(struct ldlm_lock *lock) +{ + long delay; + int result; + + if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) { + LDLM_DEBUG(lock, "client-side enqueue: destroyed"); + result = -EIO; + } else { + delay = cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity); + LDLM_DEBUG(lock, "client-side enqueue: granted after " + CFS_DURATION_T"s", delay); + + /* Update our time estimate */ + at_measured(ldlm_lock_to_ns_at(lock), + delay); + result = 0; + } + return result; +} + +/** + * Implementation of ->l_completion_ast() for a client, that doesn't wait + * until lock is granted. Suitable for locks enqueued through ptlrpcd, of + * other threads that cannot block for long. + */ +int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data) +{ + ENTRY; + + if (flags == LDLM_FL_WAIT_NOREPROC) { + LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); + RETURN(0); + } + + if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | + LDLM_FL_BLOCK_CONV))) { + cfs_waitq_signal(&lock->l_waitq); + RETURN(ldlm_completion_tail(lock)); + } - return ret; + LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " + "going forward"); + ldlm_lock_dump(D_OTHER, lock, 0); + ldlm_reprocess_all(lock->l_resource); + RETURN(0); } +/** + * Client side LDLM "completion" AST. This is called in several cases: + * + * - when a reply to an ENQUEUE rpc is received from the server + * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at + * this point (determined by flags); + * + * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has + * been granted; + * + * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock + * gets correct lvb; + * + * - to force all locks when resource is destroyed (cleanup_resource()); + * + * - during lock conversion (not used currently). + * + * If lock is not granted in the first case, this function waits until second + * or penultimate cases happen in some other thread. + * + */ int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data) { /* XXX ALLOCATE - 160 bytes */ @@ -110,6 +205,7 @@ int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data) struct obd_device *obd; struct obd_import *imp = NULL; struct l_wait_info lwi; + __u32 timeout; int rc = 0; ENTRY; @@ -127,15 +223,20 @@ int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data) LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " "sleeping"); ldlm_lock_dump(D_OTHER, lock, 0); - ldlm_reprocess_all(lock->l_resource); noreproc: obd = class_exp2obd(lock->l_conn_export); /* if this is a local lock, then there is no import */ - if (obd != NULL) + if (obd != NULL) { imp = obd->u.cli.cl_import; + } + + /* Wait a long time for enqueue - server may have to callback a + lock from another client. Server will evict the other client if it + doesn't respond reasonably, and then give us the lock. */ + timeout = ldlm_get_enq_timeout(lock) * 2; lwd.lwd_lock = lock; @@ -143,23 +244,26 @@ noreproc: LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT"); lwi = LWI_INTR(interrupted_completion_wait, &lwd); } else { - lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), + lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout), ldlm_expired_completion_wait, interrupted_completion_wait, &lwd); } if (imp != NULL) { - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); lwd.lwd_conn_cnt = imp->imp_conn_cnt; - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); } - /* Go to sleep until the lock is granted or cancelled. */ - rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi); - - if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) { - LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed"); - RETURN(-EIO); + if (ns_is_client(ldlm_lock_to_ns(lock)) && + OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST, + OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) { + lock->l_flags |= LDLM_FL_FAIL_LOC; + rc = -EINTR; + } else { + /* Go to sleep until the lock is granted or cancelled. */ + rc = l_wait_event(lock->l_waitq, + is_granted_or_cancelled(lock), &lwi); } if (rc) { @@ -168,35 +272,25 @@ noreproc: RETURN(rc); } - LDLM_DEBUG(lock, "client-side enqueue waking up: granted"); - RETURN(0); + RETURN(ldlm_completion_tail(lock)); } -/* - * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs. +/** + * A helper to build a blocking ast function + * + * Perform a common operation for blocking asts: + * defferred lock cancellation. + * + * \param lock the lock blocking or canceling ast was called on + * \retval 0 + * \see mdt_blocking_ast + * \see ldlm_blocking_ast */ -int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, - void *data, int flag) +int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock) { int do_ast; ENTRY; - if (flag == LDLM_CB_CANCELING) { - /* Don't need to do anything here. */ - RETURN(0); - } - - lock_res_and_lock(lock); - /* Get this: if ldlm_blocking_ast is racing with intent_policy, such - * that ldlm_blocking_ast is called just before intent_policy method - * takes the ns_lock, then by the time we get the lock, we might not - * be the correct blocking function anymore. So check, and return - * early, if so. */ - if (lock->l_blocking_ast != ldlm_blocking_ast) { - unlock_res_and_lock(lock); - RETURN(0); - } - lock->l_flags |= LDLM_FL_CBPENDING; do_ast = (!lock->l_readers && !lock->l_writers); unlock_res_and_lock(lock); @@ -217,6 +311,42 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, RETURN(0); } +/** + * Server blocking AST + * + * ->l_blocking_ast() callback for LDLM locks acquired by server-side + * OBDs. + * + * \param lock the lock which blocks a request or cancelling lock + * \param desc unused + * \param data unused + * \param flag indicates whether this cancelling or blocking callback + * \retval 0 + * \see ldlm_blocking_ast_nocheck + */ +int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, + void *data, int flag) +{ + ENTRY; + + if (flag == LDLM_CB_CANCELING) { + /* Don't need to do anything here. */ + RETURN(0); + } + + lock_res_and_lock(lock); + /* Get this: if ldlm_blocking_ast is racing with intent_policy, such + * that ldlm_blocking_ast is called just before intent_policy method + * takes the lr_lock, then by the time we get the lock, we might not + * be the correct blocking function anymore. So check, and return + * early, if so. */ + if (lock->l_blocking_ast != ldlm_blocking_ast) { + unlock_res_and_lock(lock); + RETURN(0); + } + RETURN(ldlm_blocking_ast_nocheck(lock)); +} + /* * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See * comment in filter_intent_policy() on why you may need this. @@ -251,11 +381,16 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, ldlm_blocking_callback blocking, ldlm_completion_callback completion, ldlm_glimpse_callback glimpse, - void *data, __u32 lvb_len, void *lvb_swabber, + void *data, __u32 lvb_len, + const __u64 *client_cookie, struct lustre_handle *lockh) { struct ldlm_lock *lock; int err; + const struct ldlm_callback_suite cbs = { .lcs_completion = completion, + .lcs_blocking = blocking, + .lcs_glimpse = glimpse, + }; ENTRY; LASSERT(!(*flags & LDLM_FL_REPLAY)); @@ -264,22 +399,23 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, LBUG(); } - lock = ldlm_lock_create(ns, res_id, type, mode, blocking, - completion, glimpse, data, lvb_len); + lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len); if (unlikely(!lock)) GOTO(out_nolock, err = -ENOMEM); - LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created"); - ldlm_lock_addref_internal(lock, mode); ldlm_lock2handle(lock, lockh); - lock_res_and_lock(lock); + + /* NB: we don't have any lock now (lock_res_and_lock) + * because it's a new lock */ + ldlm_lock_addref_internal_nolock(lock, mode); lock->l_flags |= LDLM_FL_LOCAL; if (*flags & LDLM_FL_ATOMIC_CB) lock->l_flags |= LDLM_FL_ATOMIC_CB; - lock->l_lvb_swabber = lvb_swabber; - unlock_res_and_lock(lock); + if (policy != NULL) lock->l_policy_data = *policy; + if (client_cookie != NULL) + lock->l_client_cookie = *client_cookie; if (type == LDLM_EXTENT) lock->l_req_extent = policy->l_extent; @@ -290,48 +426,69 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, if (policy != NULL) *policy = lock->l_policy_data; - LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)", - lock); - if (lock->l_completion_ast) lock->l_completion_ast(lock, *flags, NULL); - LDLM_DEBUG(lock, "client-side local enqueue END"); + LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created"); EXIT; out: - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); out_nolock: return err; } static void failed_lock_cleanup(struct ldlm_namespace *ns, - struct ldlm_lock *lock, - struct lustre_handle *lockh, int mode) + struct ldlm_lock *lock, int mode) { + int need_cancel = 0; + /* Set a flag to prevent us from sending a CANCEL (bug 407) */ lock_res_and_lock(lock); - lock->l_flags |= LDLM_FL_LOCAL_ONLY; + /* Check that lock is not granted or failed, we might race. */ + if ((lock->l_req_mode != lock->l_granted_mode) && + !(lock->l_flags & LDLM_FL_FAILED)) { + /* Make sure that this lock will not be found by raced + * bl_ast and -EINVAL reply is sent to server anyways. + * bug 17645 */ + lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED | + LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING; + need_cancel = 1; + } unlock_res_and_lock(lock); - LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY"); - ldlm_lock_decref_and_cancel(lockh, mode); + if (need_cancel) + LDLM_DEBUG(lock, + "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | " + "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING"); + else + LDLM_DEBUG(lock, "lock was granted or failed in race"); + + ldlm_lock_decref_internal(lock, mode); /* XXX - HACK because we shouldn't call ldlm_lock_destroy() * from llite/file.c/ll_file_flock(). */ + /* This code makes for the fact that we do not have blocking handler on + * a client for flock locks. As such this is the place where we must + * completely kill failed locks. (interrupted and those that + * were waiting to be granted when server evicted us. */ if (lock->l_resource->lr_type == LDLM_FLOCK) { - ldlm_lock_destroy(lock); + lock_res_and_lock(lock); + ldlm_resource_unlink_lock(lock); + ldlm_lock_destroy_nolock(lock); + unlock_res_and_lock(lock); } } int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode, int *flags, void *lvb, __u32 lvb_len, - void *lvb_swabber, struct lustre_handle *lockh,int rc) + struct lustre_handle *lockh,int rc) { struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; int is_replay = *flags & LDLM_FL_REPLAY; struct ldlm_lock *lock; struct ldlm_reply *reply; + struct ost_lvb *tmplvb; int cleanup_phase = 1; ENTRY; @@ -353,14 +510,12 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, if (reply == NULL) rc = -EPROTO; if (lvb_len) { - struct ost_lvb *tmplvb; req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len); - tmplvb = req_capsule_server_swab_get(&req->rq_pill, - &RMF_DLM_LVB, - lvb_swabber); + tmplvb = req_capsule_server_get(&req->rq_pill, + &RMF_DLM_LVB); if (tmplvb == NULL) GOTO(cleanup, rc = -EPROTO); if (lvb != NULL) @@ -378,7 +533,16 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, cleanup_phase = 0; lock_res_and_lock(lock); - lock->l_remote_handle = reply->lock_handle; + /* Key change rehash lock in per-export hash with new key */ + if (exp->exp_lock_hash) { + cfs_hash_rehash_key(exp->exp_lock_hash, + &lock->l_remote_handle, + &reply->lock_handle, + &lock->l_exp_hash); + } else { + lock->l_remote_handle = reply->lock_handle; + } + *flags = reply->lock_flags; lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS; /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match() @@ -442,17 +606,25 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, /* If the lock has already been granted by a completion AST, don't * clobber the LVB with an older one. */ - if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) { - void *tmplvb; + if (lvb_len) { + /* We must lock or a racing completion might update lvb + without letting us know and we'll clobber the correct value. + Cannot unlock after the check either, a that still leaves + a tiny window for completion to get in */ + lock_res_and_lock(lock); + if (lock->l_req_mode != lock->l_granted_mode) { - req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, - lvb_len); - tmplvb = req_capsule_server_swab_get(&req->rq_pill, - &RMF_DLM_LVB, - lvb_swabber); - if (tmplvb == NULL) - GOTO(cleanup, rc = -EPROTO); - memcpy(lock->l_lvb_data, tmplvb, lvb_len); + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, + RCL_SERVER, lvb_len); + tmplvb = req_capsule_server_get(&req->rq_pill, + &RMF_DLM_LVB); + if (tmplvb == NULL) { + unlock_res_and_lock(lock); + GOTO(cleanup, rc = -EPROTO); + } + memcpy(lock->l_lvb_data, tmplvb, lvb_len); + } + unlock_res_and_lock(lock); } if (!is_replay) { @@ -461,7 +633,7 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, int err = lock->l_completion_ast(lock, *flags, NULL); if (!rc) rc = err; - if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */ + if (rc) cleanup_phase = 1; } } @@ -476,10 +648,10 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, EXIT; cleanup: if (cleanup_phase == 1 && rc) - failed_lock_cleanup(ns, lock, lockh, mode); + failed_lock_cleanup(ns, lock, mode); /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */ LDLM_LOCK_PUT(lock); - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); return rc; } @@ -490,8 +662,11 @@ static inline int ldlm_req_handles_avail(int req_size, int off) { int avail; - avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size; - avail /= sizeof(struct lustre_handle); + avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size; + if (likely(avail >= 0)) + avail /= (int)sizeof(struct lustre_handle); + else + avail = 0; avail += LDLM_LOCKREQ_HANDLES - off; return avail; @@ -517,7 +692,7 @@ static inline int ldlm_format_handles_avail(struct obd_import *imp, * @count locks in @cancels. */ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, int version, int opc, int canceloff, - struct list_head *cancels, int count) + cfs_list_t *cancels, int count) { struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; struct req_capsule *pill = &req->rq_pill; @@ -534,12 +709,12 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT); avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff); - flags = ns_connect_lru_resize(ns) ? + flags = ns_connect_lru_resize(ns) ? LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; to_free = !ns_connect_lru_resize(ns) && opc == LDLM_ENQUEUE ? 1 : 0; - /* Cancel lru locks here _only_ if the server supports + /* Cancel lru locks here _only_ if the server supports * EARLY_CANCEL. Otherwise we have to send extra CANCEL * rpc, what will make us slower. */ if (avail > count) @@ -580,7 +755,7 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, } int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, - struct list_head *cancels, int count) + cfs_list_t *cancels, int count) { return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, LDLM_ENQUEUE_CANCEL_OFF, cancels, count); @@ -595,9 +770,9 @@ int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, struct ldlm_enqueue_info *einfo, const struct ldlm_res_id *res_id, - ldlm_policy_data_t *policy, int *flags, - void *lvb, __u32 lvb_len, void *lvb_swabber, - struct lustre_handle *lockh, int async) + ldlm_policy_data_t const *policy, int *flags, + void *lvb, __u32 lvb_len, struct lustre_handle *lockh, + int async) { struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; struct ldlm_lock *lock; @@ -613,21 +788,25 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, /* If we're replaying this lock, just check some invariants. * If we're creating a new lock, get everything all setup nice. */ if (is_replay) { - lock = ldlm_handle2lock(lockh); + lock = ldlm_handle2lock_long(lockh, 0); LASSERT(lock != NULL); LDLM_DEBUG(lock, "client-side enqueue START"); LASSERT(exp == lock->l_conn_export); } else { + const struct ldlm_callback_suite cbs = { + .lcs_completion = einfo->ei_cb_cp, + .lcs_blocking = einfo->ei_cb_bl, + .lcs_glimpse = einfo->ei_cb_gl, + .lcs_weigh = einfo->ei_cb_wg + }; lock = ldlm_lock_create(ns, res_id, einfo->ei_type, - einfo->ei_mode, einfo->ei_cb_bl, - einfo->ei_cb_cp, einfo->ei_cb_gl, - einfo->ei_cbdata, lvb_len); + einfo->ei_mode, &cbs, einfo->ei_cbdata, + lvb_len); if (lock == NULL) RETURN(-ENOMEM); /* for the local lock, add the reference */ ldlm_lock_addref_internal(lock, einfo->ei_mode); ldlm_lock2handle(lock, lockh); - lock->l_lvb_swabber = lvb_swabber; if (policy != NULL) { /* INODEBITS_INTEROP: If the server does not support * inodebits, we will request a plain lock in the @@ -656,8 +835,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE); if (req == NULL) { - failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode); - LDLM_LOCK_PUT(lock); + failed_lock_cleanup(ns, lock, einfo->ei_mode); + LDLM_LOCK_RELEASE(lock); RETURN(-ENOMEM); } req_passed_in = 0; @@ -670,7 +849,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT); LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n", - DLM_LOCKREQ_OFF, len, sizeof(*body)); + DLM_LOCKREQ_OFF, len, (int)sizeof(*body)); } lock->l_conn_export = exp; @@ -708,15 +887,17 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, } LDLM_DEBUG(lock, "sending request"); + rc = ptlrpc_queue_wait(req); + err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0, einfo->ei_mode, flags, lvb, lvb_len, - lvb_swabber, lockh, rc); + lockh, rc); /* If ldlm_cli_enqueue_fini did not find the lock, we need to free * one reference that we took */ if (err == -ENOLCK) - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); else rc = err; @@ -735,7 +916,7 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, struct ldlm_resource *res; int rc; ENTRY; - if (ns_is_client(lock->l_resource->lr_namespace)) { + if (ns_is_client(ldlm_lock_to_ns(lock))) { CERROR("Trying to cancel local lock\n"); LBUG(); } @@ -836,7 +1017,7 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock) { int rc = LDLM_FL_LOCAL_ONLY; ENTRY; - + if (lock->l_conn_export) { int local_only; @@ -858,14 +1039,13 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock) } ldlm_lock_cancel(lock); } else { - if (ns_is_client(lock->l_resource->lr_namespace)) { + if (ns_is_client(ldlm_lock_to_ns(lock))) { LDLM_ERROR(lock, "Trying to cancel local lock"); LBUG(); } LDLM_DEBUG(lock, "server-side local cancel"); ldlm_lock_cancel(lock); ldlm_reprocess_all(lock->l_resource); - LDLM_DEBUG(lock, "server-side local cancel handler END"); } RETURN(rc); @@ -874,7 +1054,7 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock) /* Pack @count locks in @head into ldlm_request buffer at the offset @off, of the request @req. */ static void ldlm_cancel_pack(struct ptlrpc_request *req, - struct list_head *head, int count) + cfs_list_t *head, int count) { struct ldlm_request *dlm; struct ldlm_lock *lock; @@ -885,7 +1065,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, LASSERT(dlm != NULL); /* Check the room in the request buffer. */ - max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) - + max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) - sizeof(struct ldlm_request); max /= sizeof(struct lustre_handle); max += LDLM_LOCKREQ_HANDLES; @@ -894,7 +1074,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, /* XXX: it would be better to pack lock handles grouped by resource. * so that the server cancel would call filter_lvbo_update() less * frequently. */ - list_for_each_entry(lock, head, l_bl_ast) { + cfs_list_for_each_entry(lock, head, l_bl_ast) { if (!count--) break; LASSERT(lock->l_conn_export); @@ -909,8 +1089,8 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, /* Prepare and send a batched cancel rpc, it will include count lock handles * of locks given in @head. */ -int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, - int count, int flags) +int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels, + int count, ldlm_cancel_flags_t flags) { struct ptlrpc_request *req = NULL; struct obd_import *imp; @@ -921,6 +1101,8 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, LASSERT(exp != NULL); LASSERT(count > 0); + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val); + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE)) RETURN(count); @@ -955,15 +1137,15 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, req->rq_no_resend = 1; req->rq_no_delay = 1; - /* XXX FIXME bug 249 */ req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL; req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL; + ptlrpc_at_set_req_timeout(req); ldlm_cancel_pack(req, cancels, count); ptlrpc_request_set_replen(req); - if (flags & LDLM_FL_ASYNC) { - ptlrpcd_add_req(req); + if (flags & LCF_ASYNC) { + ptlrpcd_add_req(req, PSCOPE_OTHER); sent = count; GOTO(out, 0); } else { @@ -1000,56 +1182,57 @@ static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) return &imp->imp_obd->obd_namespace->ns_pool; } +/** + * Update client's obd pool related fields with new SLV and Limit from \a req. + */ int ldlm_cli_update_pool(struct ptlrpc_request *req) { + struct obd_device *obd; __u64 old_slv, new_slv; - struct ldlm_pool *pl; __u32 new_limit; ENTRY; - - if (!imp_connect_lru_resize(req->rq_import)) + if (unlikely(!req->rq_import || !req->rq_import->imp_obd || + !imp_connect_lru_resize(req->rq_import))) + { + /* + * Do nothing for corner cases. + */ RETURN(0); + } - /* In some cases RPC may contain slv and limit zeroed out. This is + /* + * In some cases RPC may contain slv and limit zeroed out. This is * the case when server does not support lru resize feature. This is * also possible in some recovery cases when server side reqs have no - * ref to obd export and thus access to server side namespace is no - * possible. */ - if (lustre_msg_get_slv(req->rq_repmsg) == 0 || + * ref to obd export and thus access to server side namespace is no + * possible. + */ + if (lustre_msg_get_slv(req->rq_repmsg) == 0 || lustre_msg_get_limit(req->rq_repmsg) == 0) { - DEBUG_REQ(D_HA, req, "zero SLV or Limit found " - "(SLV: "LPU64", Limit: %u)", - lustre_msg_get_slv(req->rq_repmsg), + DEBUG_REQ(D_HA, req, "Zero SLV or Limit found " + "(SLV: "LPU64", Limit: %u)", + lustre_msg_get_slv(req->rq_repmsg), lustre_msg_get_limit(req->rq_repmsg)); RETURN(0); } new_limit = lustre_msg_get_limit(req->rq_repmsg); new_slv = lustre_msg_get_slv(req->rq_repmsg); - pl = ldlm_imp2pl(req->rq_import); - - spin_lock(&pl->pl_lock); - old_slv = ldlm_pool_get_slv(pl); - ldlm_pool_set_slv(pl, new_slv); - ldlm_pool_set_limit(pl, new_limit); - - /* Check if we need to wakeup pools thread for fast SLV change. - * This is only done when threads period is noticably long like - * 10s or more. */ -#if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10) - { - __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE; - do_div(fast_change, 100); + obd = req->rq_import->imp_obd; + + /* + * Set new SLV and Limit to obd fields to make accessible for pool + * thread. We do not access obd_namespace and pool directly here + * as there is no reliable way to make sure that they are still + * alive in cleanup time. Evil races are possible which may cause + * oops in that time. + */ + cfs_write_lock(&obd->obd_pool_lock); + old_slv = obd->obd_pool_slv; + obd->obd_pool_slv = new_slv; + obd->obd_pool_limit = new_limit; + cfs_write_unlock(&obd->obd_pool_lock); - /* Wake up pools thread only if SLV has changed more than - * 50% since last update. In this case we want to react asap. - * Otherwise it is no sense to wake up pools as they are - * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */ - if (old_slv > new_slv && old_slv - new_slv > fast_change) - ldlm_pools_wakeup(); - } -#endif - spin_unlock(&pl->pl_lock); RETURN(0); } EXPORT_SYMBOL(ldlm_cli_update_pool); @@ -1064,7 +1247,7 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) ENTRY; /* concurrent cancels on the same handle can happen */ - lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING); + lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING); if (lock == NULL) { LDLM_DEBUG_NOLOCK("lock is already being destroyed\n"); RETURN(0); @@ -1072,14 +1255,14 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) rc = ldlm_cli_cancel_local(lock); if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) { - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); RETURN(rc < 0 ? rc : 0); } /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL * rpc which goes to canceld portal, so we can cancel other lru locks * here and send them all as one LDLM_CANCEL rpc. */ - LASSERT(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, &cancels); + LASSERT(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, &cancels); exp = lock->l_conn_export; if (exp_connect_cancelset(exp)) { @@ -1088,11 +1271,11 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) RCL_CLIENT, 0); LASSERT(avail > 0); - ns = lock->l_resource->lr_namespace; + ns = ldlm_lock_to_ns(lock); flags = ns_connect_lru_resize(ns) ? LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1, - LDLM_FL_BL_AST, flags); + LCF_BL_AST, flags); } ldlm_cli_cancel_list(&cancels, count, NULL, 0); RETURN(0); @@ -1100,34 +1283,35 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) /* XXX until we will have compound requests and can cut cancels from generic rpc * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */ -static int ldlm_cancel_list(struct list_head *cancels, int count, int flags) +int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count, + ldlm_cancel_flags_t flags) { CFS_LIST_HEAD(head); struct ldlm_lock *lock, *next; int left = 0, bl_ast = 0, rc; left = count; - list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { + cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { if (left-- == 0) break; - if (flags & LDLM_FL_LOCAL_ONLY) { + if (flags & LCF_LOCAL) { rc = LDLM_FL_LOCAL_ONLY; ldlm_lock_cancel(lock); } else { rc = ldlm_cli_cancel_local(lock); } - if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) { + if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) { LDLM_DEBUG(lock, "Cancel lock separately"); - list_del_init(&lock->l_bl_ast); - list_add(&lock->l_bl_ast, &head); + cfs_list_del_init(&lock->l_bl_ast); + cfs_list_add(&lock->l_bl_ast, &head); bl_ast ++; continue; } if (rc == LDLM_FL_LOCAL_ONLY) { /* CANCEL RPC should not be sent to server. */ - list_del_init(&lock->l_bl_ast); - LDLM_LOCK_PUT(lock); + cfs_list_del_init(&lock->l_bl_ast); + LDLM_LOCK_RELEASE(lock); count--; } @@ -1140,59 +1324,49 @@ static int ldlm_cancel_list(struct list_head *cancels, int count, int flags) RETURN(count); } -/* Return 1 to stop lru processing and keep current lock cached. Return zero - * otherwise. */ -static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, - int unused, int added, - int count) +/** + * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back + * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g. + * readahead requests, ...) + */ +static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, + struct ldlm_lock *lock, + int unused, int added, + int count) { - int lock_cost; - __u64 page_nr; - - /* Stop lru processing when we reached passed @count or checked all - * locks in lru. */ - if (count && added >= count) - return LDLM_POLICY_KEEP_LOCK; + ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK; + ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery; + lock_res_and_lock(lock); - if (lock->l_resource->lr_type == LDLM_EXTENT) { - struct ldlm_extent *l_extent; - - /* For all extent locks cost is 1 + number of pages in - * their extent. */ - l_extent = &lock->l_policy_data.l_extent; - page_nr = (l_extent->end - l_extent->start); - do_div(page_nr, CFS_PAGE_SIZE); - -#ifdef __KERNEL__ - /* XXX: In fact this is evil hack, we can't access inode - * here. For doing it right we need somehow to have number - * of covered by lock. This should be fixed later when 10718 - * is landed. */ - if (lock->l_ast_data != NULL) { - struct inode *inode = lock->l_ast_data; - if (page_nr > inode->i_mapping->nrpages) - page_nr = inode->i_mapping->nrpages; - } -#endif - lock_cost = 1 + page_nr; - } else { - /* For all locks which are not extent ones cost is 1 */ - lock_cost = 1; + /* don't check added & count since we want to process all locks + * from unused list */ + switch (lock->l_resource->lr_type) { + case LDLM_EXTENT: + case LDLM_IBITS: + if (cb && cb(lock)) + break; + default: + result = LDLM_POLICY_SKIP_LOCK; + lock->l_flags |= LDLM_FL_SKIPPED; + break; } - /* Keep all expensive locks in lru for the memory pressure time - * cancel policy. They anyways may be canceled by lru resize - * pplicy if they have not small enough CLV. */ - return lock_cost > ns->ns_shrink_thumb ? - LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; + unlock_res_and_lock(lock); + RETURN(result); } -/* Return 1 to stop lru processing and keep current lock cached. Return zero - * otherwise. */ +/** + * Callback function for lru-resize policy. Makes decision whether to keep + * \a lock in LRU for current \a LRU size \a unused, added in current scan + * \a added and number of locks to be preferably canceled \a count. + * + * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning + * + * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU + */ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, - int unused, int added, + struct ldlm_lock *lock, + int unused, int added, int count) { cfs_time_t cur = cfs_time_current(); @@ -1200,78 +1374,115 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, __u64 slv, lvf, lv; cfs_time_t la; - /* Stop lru processing when we reached passed @count or checked all - * locks in lru. */ + /* + * Stop lru processing when we reached passed @count or checked all + * locks in lru. + */ if (count && added >= count) return LDLM_POLICY_KEEP_LOCK; - spin_lock(&pl->pl_lock); slv = ldlm_pool_get_slv(pl); - lvf = atomic_read(&pl->pl_lock_volume_factor); - spin_unlock(&pl->pl_lock); - - la = cfs_duration_sec(cfs_time_sub(cur, + lvf = ldlm_pool_get_lvf(pl); + la = cfs_duration_sec(cfs_time_sub(cur, lock->l_last_used)); - /* Stop when slv is not yet come from server or - * lv is smaller than it is. */ + /* + * Stop when slv is not yet come from server or lv is smaller than + * it is. + */ lv = lvf * la * unused; - return (slv == 1 || lv < slv) ? + + /* + * Inform pool about current CLV to see it via proc. + */ + ldlm_pool_set_clv(pl, lv); + return (slv == 0 || lv < slv) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } -/* Return 1 to stop lru processing and keep current lock cached. Return zero - * otherwise. */ +/** + * Callback function for proc used policy. Makes decision whether to keep + * \a lock in LRU for current \a LRU size \a unused, added in current scan + * \a added and number of locks to be preferably canceled \a count. + * + * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning + * + * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU + */ static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, + struct ldlm_lock *lock, int unused, int added, int count) { - /* Stop lru processing when we reached passed @count or checked all - * locks in lru. */ - return (added >= count) ? + /* + * Stop lru processing when we reached passed @count or checked all + * locks in lru. + */ + return (added >= count) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } -/* Return 1 to stop lru processing and keep current lock cached. Return zero - * otherwise. */ +/** + * Callback function for aged policy. Makes decision whether to keep + * \a lock in LRU for current \a LRU size \a unused, added in current scan + * \a added and number of locks to be preferably canceled \a count. + * + * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning + * + * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU + */ static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, + struct ldlm_lock *lock, int unused, int added, int count) { - /* Stop lru processing if young lock is found and we reached passed - * @count. */ - return ((added >= count) && + /* + * Stop lru processing if young lock is found and we reached passed + * @count. + */ + return ((added >= count) && cfs_time_before(cfs_time_current(), cfs_time_add(lock->l_last_used, - ns->ns_max_age))) ? + ns->ns_max_age))) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } -/* Return 1 to stop lru processing and keep current lock cached. Return zero - * otherwise. */ +/** + * Callback function for default policy. Makes decision whether to keep + * \a lock in LRU for current \a LRU size \a unused, added in current scan + * \a added and number of locks to be preferably canceled \a count. + * + * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning + * + * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU + */ static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, + struct ldlm_lock *lock, int unused, int added, int count) { - /* Stop lru processing when we reached passed @count or checked all - * locks in lru. */ - return (added >= count) ? + /* + * Stop lru processing when we reached passed @count or checked all + * locks in lru. + */ + return (added >= count) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } -typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *, - struct ldlm_lock *, int, +typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *, + struct ldlm_lock *, int, int, int); static ldlm_cancel_lru_policy_t ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) { + if (flags & LDLM_CANCEL_NO_WAIT) + return ldlm_cancel_no_wait_policy; + if (ns_connect_lru_resize(ns)) { if (flags & LDLM_CANCEL_SHRINK) - return ldlm_cancel_shrink_policy; + /* We kill passed number of old locks. */ + return ldlm_cancel_passed_policy; else if (flags & LDLM_CANCEL_LRUR) return ldlm_cancel_lrur_policy; else if (flags & LDLM_CANCEL_PASSED) @@ -1280,10 +1491,10 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) if (flags & LDLM_CANCEL_AGED) return ldlm_cancel_aged_policy; } - + return ldlm_cancel_default_policy; } - + /* - Free space in lru for @count new locks, * redundant unused locks are canceled locally; * - also cancel locally unused aged locks; @@ -1309,33 +1520,51 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) * memory pressre policy function; * * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy". + * + * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible + * (typically before replaying locks) w/o + * sending any rpcs or waiting for any + * outstanding rpc to complete. */ -int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, - int count, int max, int cancel_flags, int flags) +static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, + int count, int max, int flags) { ldlm_cancel_lru_policy_t pf; struct ldlm_lock *lock, *next; - int added = 0, unused; + int added = 0, unused, remained; ENTRY; - spin_lock(&ns->ns_unused_lock); + cfs_spin_lock(&ns->ns_lock); unused = ns->ns_nr_unused; + remained = unused; if (!ns_connect_lru_resize(ns)) count += unused - ns->ns_max_unused; pf = ldlm_cancel_lru_policy(ns, flags); LASSERT(pf != NULL); - - while (!list_empty(&ns->ns_unused_list)) { + + while (!cfs_list_empty(&ns->ns_unused_list)) { + ldlm_policy_res_t result; + + /* all unused locks */ + if (remained-- <= 0) + break; + /* For any flags, stop scanning if @max is reached. */ if (max && added >= max) break; - list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){ + cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list, + l_lru){ /* No locks which got blocking requests. */ LASSERT(!(lock->l_flags & LDLM_FL_BL_AST)); + if (flags & LDLM_CANCEL_NO_WAIT && + lock->l_flags & LDLM_FL_SKIPPED) + /* already processed */ + continue; + /* Somebody is already doing CANCEL. No need in this * lock in lru, do not traverse it again. */ if (!(lock->l_flags & LDLM_FL_CANCELING)) @@ -1346,6 +1575,10 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, if (&lock->l_lru == &ns->ns_unused_list) break; + LDLM_LOCK_GET(lock); + cfs_spin_unlock(&ns->ns_lock); + lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current()); + /* Pass the lock through the policy filter and see if it * should stay in lru. * @@ -1353,17 +1586,27 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, * we find a lock that should stay in the cache. * We should take into account lock age anyway * as new lock even if it is small of weight is - * valuable resource. + * valuable resource. * * That is, for shrinker policy we drop only * old locks, but additionally chose them by - * their weight. Big extent locks will stay in + * their weight. Big extent locks will stay in * the cache. */ - if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK) + result = pf(ns, lock, unused, added, count); + if (result == LDLM_POLICY_KEEP_LOCK) { + lu_ref_del(&lock->l_reference, + __FUNCTION__, cfs_current()); + LDLM_LOCK_RELEASE(lock); + cfs_spin_lock(&ns->ns_lock); break; - - LDLM_LOCK_GET(lock); /* dropped by bl thread */ - spin_unlock(&ns->ns_unused_lock); + } + if (result == LDLM_POLICY_SKIP_LOCK) { + lu_ref_del(&lock->l_reference, + __FUNCTION__, cfs_current()); + LDLM_LOCK_RELEASE(lock); + cfs_spin_lock(&ns->ns_lock); + continue; + } lock_res_and_lock(lock); /* Check flags again under the lock. */ @@ -1375,16 +1618,18 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, * cancel by itseft or the lock is matched * is already not unused. */ unlock_res_and_lock(lock); - LDLM_LOCK_PUT(lock); - spin_lock(&ns->ns_unused_lock); + lu_ref_del(&lock->l_reference, + __FUNCTION__, cfs_current()); + LDLM_LOCK_RELEASE(lock); + cfs_spin_lock(&ns->ns_lock); continue; } LASSERT(!lock->l_readers && !lock->l_writers); /* If we have chosen to cancel this lock voluntarily, we * better send cancel notification to server, so that it - * frees appropriate state. This might lead to a race - * where while we are doing cancel here, server is also + * frees appropriate state. This might lead to a race + * where while we are doing cancel here, server is also * silently cancelling this lock. */ lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK; @@ -1398,65 +1643,38 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, /* We can't re-add to l_lru as it confuses the * refcounting in ldlm_lock_remove_from_lru() if an AST - * arrives after we drop ns_lock below. We use l_bl_ast + * arrives after we drop lr_lock below. We use l_bl_ast * and can't use l_pending_chain as it is used both on * server and client nevertheless bug 5666 says it is * used only on server */ - LASSERT(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, cancels); + LASSERT(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, cancels); unlock_res_and_lock(lock); - spin_lock(&ns->ns_unused_lock); + lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current()); + cfs_spin_lock(&ns->ns_lock); added++; unused--; } - spin_unlock(&ns->ns_unused_lock); - RETURN(ldlm_cancel_list(cancels, added, cancel_flags)); + cfs_spin_unlock(&ns->ns_lock); + RETURN(added); } -/* Returns number of locks which could be canceled next time when - * ldlm_cancel_lru() is called. Used from locks pool shrinker. */ -int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns, - int count, int max, int flags) +int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels, + int count, int max, ldlm_cancel_flags_t cancel_flags, + int flags) { - ldlm_cancel_lru_policy_t pf; - struct ldlm_lock *lock; - int added = 0, unused; - ENTRY; - - pf = ldlm_cancel_lru_policy(ns, flags); - LASSERT(pf != NULL); - spin_lock(&ns->ns_unused_lock); - unused = ns->ns_nr_unused; - - list_for_each_entry(lock, &ns->ns_unused_list, l_lru) { - /* For any flags, stop scanning if @max is reached. */ - if (max && added >= max) - break; - - /* Somebody is already doing CANCEL or there is a - * blocking request will send cancel. Let's not count - * this lock. */ - if ((lock->l_flags & LDLM_FL_CANCELING) || - (lock->l_flags & LDLM_FL_BL_AST)) - continue; - - /* Pass the lock through the policy filter and see if it - * should stay in lru. */ - if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK) - break; - - added++; - unused--; - } - spin_unlock(&ns->ns_unused_lock); - RETURN(added); + int added; + added = ldlm_prepare_lru_list(ns, cancels, count, max, flags); + if (added <= 0) + return added; + return ldlm_cli_cancel_list_local(cancels, added, cancel_flags); } /* when called with LDLM_ASYNC the blocking callback will be handled * in a thread and this function will return after the thread has been * asked to call the callback. when called with LDLM_SYNC the blocking * callback will be performed in this function. */ -int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync, +int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode, int flags) { CFS_LIST_HEAD(cancels); @@ -1464,36 +1682,33 @@ int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync, ENTRY; #ifndef __KERNEL__ - sync = LDLM_SYNC; /* force to be sync in user space */ + mode = LDLM_SYNC; /* force to be sync in user space */ #endif - count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags); - if (sync == LDLM_ASYNC) { - rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count); - if (rc == 0) - RETURN(count); - } + /* Just prepare the list of locks, do not actually cancel them yet. + * Locks are cancelled later in a separate thread. */ + count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags); + rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode); + if (rc == 0) + RETURN(count); - /* If an error occured in ASYNC mode, or - * this is SYNC mode, cancel the list. */ - ldlm_cli_cancel_list(&cancels, count, NULL, 0); - RETURN(count); + RETURN(0); } /* Find and cancel locally unused locks found on resource, matched to the * given policy, mode. GET the found locks and add them into the @cancels * list. */ int ldlm_cancel_resource_local(struct ldlm_resource *res, - struct list_head *cancels, + cfs_list_t *cancels, ldlm_policy_data_t *policy, ldlm_mode_t mode, int lock_flags, - int cancel_flags, void *opaque) + ldlm_cancel_flags_t cancel_flags, void *opaque) { struct ldlm_lock *lock; int count = 0; ENTRY; lock_res(res); - list_for_each_entry(lock, &res->lr_granted, l_res_link) { + cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) { if (opaque != NULL && lock->l_ast_data != opaque) { LDLM_ERROR(lock, "data %p doesn't match opaque %p", lock->l_ast_data, opaque); @@ -1501,17 +1716,12 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, continue; } - if (lock->l_readers || lock->l_writers) { - if (cancel_flags & LDLM_FL_WARN) { - LDLM_ERROR(lock, "lock in use"); - //LBUG(); - } + if (lock->l_readers || lock->l_writers) continue; - } /* If somebody is already doing CANCEL, or blocking ast came, * skip this lock. */ - if (lock->l_flags & LDLM_FL_BL_AST || + if (lock->l_flags & LDLM_FL_BL_AST || lock->l_flags & LDLM_FL_CANCELING) continue; @@ -1529,40 +1739,41 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | lock_flags; - LASSERT(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, cancels); + LASSERT(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, cancels); LDLM_LOCK_GET(lock); count++; } unlock_res(res); - RETURN(ldlm_cancel_list(cancels, count, cancel_flags)); + RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags)); } -/* If @req is NULL, send CANCEL request to server with handles of locks - * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests +/* If @req is NULL, send CANCEL request to server with handles of locks + * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests * separately per lock. - * If @req is not NULL, put handles of locks in @cancels into the request + * If @req is not NULL, put handles of locks in @cancels into the request * buffer at the offset @off. * Destroy @cancels at the end. */ -int ldlm_cli_cancel_list(struct list_head *cancels, int count, - struct ptlrpc_request *req, int flags) +int ldlm_cli_cancel_list(cfs_list_t *cancels, int count, + struct ptlrpc_request *req, ldlm_cancel_flags_t flags) { struct ldlm_lock *lock; int res = 0; ENTRY; - if (list_empty(cancels) || count == 0) + if (cfs_list_empty(cancels) || count == 0) RETURN(0); - - /* XXX: requests (both batched and not) could be sent in parallel. + + /* XXX: requests (both batched and not) could be sent in parallel. * Usually it is enough to have just 1 RPC, but it is possible that * there are to many locks to be cancelled in LRU or on a resource. * It would also speed up the case when the server does not support * the feature. */ while (count > 0) { - LASSERT(!list_empty(cancels)); - lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast); + LASSERT(!cfs_list_empty(cancels)); + lock = cfs_list_entry(cancels->next, struct ldlm_lock, + l_bl_ast); LASSERT(lock->l_conn_export); if (exp_connect_cancelset(lock->l_conn_export)) { @@ -1593,7 +1804,9 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count, int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns, const struct ldlm_res_id *res_id, ldlm_policy_data_t *policy, - ldlm_mode_t mode, int flags, void *opaque) + ldlm_mode_t mode, + ldlm_cancel_flags_t flags, + void *opaque) { struct ldlm_resource *res; CFS_LIST_HEAD(cancels); @@ -1608,126 +1821,69 @@ int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns, RETURN(0); } + LDLM_RESOURCE_ADDREF(res); count = ldlm_cancel_resource_local(res, &cancels, policy, mode, - 0, flags, opaque); + 0, flags | LCF_BL_AST, opaque); rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags); if (rc != ELDLM_OK) CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc); + LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); RETURN(0); } -static inline int have_no_nsresource(struct ldlm_namespace *ns) -{ - int no_resource = 0; - - spin_lock(&ns->ns_hash_lock); - if (ns->ns_resources == 0) - no_resource = 1; - spin_unlock(&ns->ns_hash_lock); +struct ldlm_cli_cancel_arg { + int lc_flags; + void *lc_opaque; +}; - RETURN(no_resource); +static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) +{ + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + struct ldlm_cli_cancel_arg *lc = arg; + int rc; + + rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name, + NULL, LCK_MINMODE, + lc->lc_flags, lc->lc_opaque); + if (rc != 0) { + CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n", + res->lr_name.name[0], rc); + } + /* must return 0 for hash iteration */ + return 0; } /* Cancel all locks on a namespace (or a specific resource, if given) * that have 0 readers/writers. * - * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying + * If flags & LCF_LOCAL, throw the locks away without trying * to notify the server. */ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, const struct ldlm_res_id *res_id, - int flags, void *opaque) + ldlm_cancel_flags_t flags, void *opaque) { - int i; + struct ldlm_cli_cancel_arg arg = { + .lc_flags = flags, + .lc_opaque = opaque, + }; + ENTRY; if (ns == NULL) RETURN(ELDLM_OK); - if (res_id) + if (res_id != NULL) { RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL, LCK_MINMODE, flags, opaque)); - - spin_lock(&ns->ns_hash_lock); - for (i = 0; i < RES_HASH_SIZE; i++) { - struct list_head *tmp; - tmp = ns->ns_hash[i].next; - while (tmp != &(ns->ns_hash[i])) { - struct ldlm_resource *res; - int rc; - - res = list_entry(tmp, struct ldlm_resource, lr_hash); - ldlm_resource_getref(res); - spin_unlock(&ns->ns_hash_lock); - - rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name, - NULL, LCK_MINMODE, - flags, opaque); - - if (rc) - CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n", - res->lr_name.name[0], rc); - - spin_lock(&ns->ns_hash_lock); - tmp = tmp->next; - ldlm_resource_putref_locked(res); - } - } - spin_unlock(&ns->ns_hash_lock); - - RETURN(ELDLM_OK); -} - -/* join/split resource locks to/from lru list */ -int ldlm_cli_join_lru(struct ldlm_namespace *ns, - const struct ldlm_res_id *res_id, int join) -{ - struct ldlm_resource *res; - struct ldlm_lock *lock, *n; - int count = 0; - ENTRY; - - LASSERT(ns_is_client(ns)); - - res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0); - if (res == NULL) - RETURN(count); - LASSERT(res->lr_type == LDLM_EXTENT); - - lock_res(res); - if (!join) - goto split; - - list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) { - if (list_empty(&lock->l_lru) && - !lock->l_readers && !lock->l_writers && - !(lock->l_flags & LDLM_FL_LOCAL) && - !(lock->l_flags & LDLM_FL_CBPENDING) && - !(lock->l_flags & LDLM_FL_BL_AST)) { - ldlm_lock_add_to_lru(lock); - lock->l_flags &= ~LDLM_FL_NO_LRU; - LDLM_DEBUG(lock, "join lock to lru"); - count++; - } - } - goto unlock; -split: - spin_lock(&ns->ns_unused_lock); - list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) { - if (lock->l_resource == res) { - ldlm_lock_remove_from_lru_nolock(lock); - lock->l_flags |= LDLM_FL_NO_LRU; - LDLM_DEBUG(lock, "split lock from lru"); - count++; - } + } else { + cfs_hash_for_each_nolock(ns->ns_rs_hash, + ldlm_cli_hash_cancel_unused, &arg); + RETURN(ELDLM_OK); } - spin_unlock(&ns->ns_unused_lock); -unlock: - unlock_res(res); - ldlm_resource_putref(res); - RETURN(count); } /* Lock iterators. */ @@ -1735,7 +1891,7 @@ unlock: int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, void *closure) { - struct list_head *tmp, *next; + cfs_list_t *tmp, *next; struct ldlm_lock *lock; int rc = LDLM_ITER_CONTINUE; @@ -1745,22 +1901,22 @@ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, RETURN(LDLM_ITER_CONTINUE); lock_res(res); - list_for_each_safe(tmp, next, &res->lr_granted) { - lock = list_entry(tmp, struct ldlm_lock, l_res_link); + cfs_list_for_each_safe(tmp, next, &res->lr_granted) { + lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); if (iter(lock, closure) == LDLM_ITER_STOP) GOTO(out, rc = LDLM_ITER_STOP); } - list_for_each_safe(tmp, next, &res->lr_converting) { - lock = list_entry(tmp, struct ldlm_lock, l_res_link); + cfs_list_for_each_safe(tmp, next, &res->lr_converting) { + lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); if (iter(lock, closure) == LDLM_ITER_STOP) GOTO(out, rc = LDLM_ITER_STOP); } - list_for_each_safe(tmp, next, &res->lr_waiting) { - lock = list_entry(tmp, struct ldlm_lock, l_res_link); + cfs_list_for_each_safe(tmp, next, &res->lr_waiting) { + lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); if (iter(lock, closure) == LDLM_ITER_STOP) GOTO(out, rc = LDLM_ITER_STOP); @@ -1781,54 +1937,38 @@ static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure) return helper->iter(lock, helper->closure); } -static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure) +static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) + { - return ldlm_resource_foreach(res, ldlm_iter_helper, closure); + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + + return ldlm_resource_foreach(res, ldlm_iter_helper, arg) == + LDLM_ITER_STOP; } -int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter, - void *closure) +void ldlm_namespace_foreach(struct ldlm_namespace *ns, + ldlm_iterator_t iter, void *closure) + { struct iter_helper_data helper = { iter: iter, closure: closure }; - return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper); -} -int ldlm_namespace_foreach_res(struct ldlm_namespace *ns, - ldlm_res_iterator_t iter, void *closure) -{ - int i, rc = LDLM_ITER_CONTINUE; - struct ldlm_resource *res; - struct list_head *tmp; + cfs_hash_for_each_nolock(ns->ns_rs_hash, + ldlm_res_iter_helper, &helper); - ENTRY; - spin_lock(&ns->ns_hash_lock); - for (i = 0; i < RES_HASH_SIZE; i++) { - tmp = ns->ns_hash[i].next; - while (tmp != &(ns->ns_hash[i])) { - res = list_entry(tmp, struct ldlm_resource, lr_hash); - ldlm_resource_getref(res); - spin_unlock(&ns->ns_hash_lock); - - rc = iter(res, closure); - - spin_lock(&ns->ns_hash_lock); - tmp = tmp->next; - ldlm_resource_putref_locked(res); - if (rc == LDLM_ITER_STOP) - GOTO(out, rc); - } - } - out: - spin_unlock(&ns->ns_hash_lock); - RETURN(rc); } -/* non-blocking function to manipulate a lock whose cb_data is being put away.*/ -void ldlm_resource_iterate(struct ldlm_namespace *ns, - const struct ldlm_res_id *res_id, - ldlm_iterator_t iter, void *data) +/* non-blocking function to manipulate a lock whose cb_data is being put away. + * return 0: find no resource + * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE. + * < 0: errors + */ +int ldlm_resource_iterate(struct ldlm_namespace *ns, + const struct ldlm_res_id *res_id, + ldlm_iterator_t iter, void *data) { struct ldlm_resource *res; + int rc; ENTRY; if (ns == NULL) { @@ -1837,39 +1977,48 @@ void ldlm_resource_iterate(struct ldlm_namespace *ns, } res = ldlm_resource_get(ns, NULL, res_id, 0, 0); - if (res == NULL) { - EXIT; - return; - } + if (res == NULL) + RETURN(0); - ldlm_resource_foreach(res, iter, data); + LDLM_RESOURCE_ADDREF(res); + rc = ldlm_resource_foreach(res, iter, data); + LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); - EXIT; + RETURN(rc); } /* Lock replay */ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure) { - struct list_head *list = closure; + cfs_list_t *list = closure; /* we use l_pending_chain here, because it's unused on clients. */ - LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n", + LASSERTF(cfs_list_empty(&lock->l_pending_chain), + "lock %p next %p prev %p\n", lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev); - /* bug 9573: don't replay locks left after eviction */ - if (!(lock->l_flags & LDLM_FL_FAILED)) - list_add(&lock->l_pending_chain, list); + /* bug 9573: don't replay locks left after eviction, or + * bug 17614: locks being actively cancelled. Get a reference + * on a lock so that it does not disapear under us (e.g. due to cancel) + */ + if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) { + cfs_list_add(&lock->l_pending_chain, list); + LDLM_LOCK_GET(lock); + } + return LDLM_ITER_CONTINUE; } -static int replay_lock_interpret(struct ptlrpc_request *req, +static int replay_lock_interpret(const struct lu_env *env, + struct ptlrpc_request *req, struct ldlm_async_args *aa, int rc) { - struct ldlm_lock *lock; - struct ldlm_reply *reply; + struct ldlm_lock *lock; + struct ldlm_reply *reply; + struct obd_export *exp; ENTRY; - atomic_dec(&req->rq_import->imp_replay_inflight); + cfs_atomic_dec(&req->rq_import->imp_replay_inflight); if (rc != ELDLM_OK) GOTO(out, rc); @@ -1888,7 +2037,17 @@ static int replay_lock_interpret(struct ptlrpc_request *req, GOTO(out, rc = -ESTALE); } - lock->l_remote_handle = reply->lock_handle; + /* Key change rehash lock in per-export hash with new key */ + exp = req->rq_export; + if (exp && exp->exp_lock_hash) { + cfs_hash_rehash_key(exp->exp_lock_hash, + &lock->l_remote_handle, + &reply->lock_handle, + &lock->l_exp_hash); + } else { + lock->l_remote_handle = reply->lock_handle; + } + LDLM_DEBUG(lock, "replayed lock:"); ptlrpc_import_recovery_state_machine(req->rq_import); LDLM_LOCK_PUT(lock); @@ -1896,7 +2055,6 @@ out: if (rc != ELDLM_OK) ptlrpc_connect_import(req->rq_import, NULL); - RETURN(rc); } @@ -1941,7 +2099,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED; else if (lock->l_granted_mode) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV; - else if (!list_empty(&lock->l_res_link)) + else if (!cfs_list_empty(&lock->l_res_link)) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT; else flags = LDLM_FL_REPLAY; @@ -1973,16 +2131,45 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) LDLM_DEBUG(lock, "replaying lock:"); - atomic_inc(&req->rq_import->imp_replay_inflight); + cfs_atomic_inc(&req->rq_import->imp_replay_inflight); CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args)); - aa = (struct ldlm_async_args *)&req->rq_async_args; + aa = ptlrpc_req_async_args(req); aa->lock_handle = body->lock_handle[0]; - req->rq_interpret_reply = replay_lock_interpret; - ptlrpcd_add_req(req); + req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret; + ptlrpcd_add_req(req, PSCOPE_OTHER); RETURN(0); } +/** + * Cancel as many unused locks as possible before replay. since we are + * in recovery, we can't wait for any outstanding RPCs to send any RPC + * to the server. + * + * Called only in recovery before replaying locks. there is no need to + * replay locks that are unused. since the clients may hold thousands of + * cached unused locks, dropping the unused locks can greatly reduce the + * load on the servers at recovery time. + */ +static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns) +{ + int canceled; + CFS_LIST_HEAD(cancels); + + CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before" + "replay for namespace %s (%d)\n", + ldlm_ns_name(ns), ns->ns_nr_unused); + + /* We don't need to care whether or not LRU resize is enabled + * because the LDLM_CANCEL_NO_WAIT policy doesn't use the + * count parameter */ + canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0, + LCF_LOCAL, LDLM_CANCEL_NO_WAIT); + + CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n", + canceled, ldlm_ns_name(ns)); +} + int ldlm_replay_locks(struct obd_import *imp) { struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; @@ -1992,21 +2179,31 @@ int ldlm_replay_locks(struct obd_import *imp) ENTRY; - LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); + LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0); + + /* don't replay locks if import failed recovery */ + if (imp->imp_vbr_failed) + RETURN(0); /* ensure this doesn't fall to 0 before all have been queued */ - atomic_inc(&imp->imp_replay_inflight); + cfs_atomic_inc(&imp->imp_replay_inflight); + + if (ldlm_cancel_unused_locks_before_replay) + ldlm_cancel_unused_locks_for_replay(ns); - (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); + ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); - list_for_each_entry_safe(lock, next, &list, l_pending_chain) { - list_del_init(&lock->l_pending_chain); - if (rc) + cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) { + cfs_list_del_init(&lock->l_pending_chain); + if (rc) { + LDLM_LOCK_RELEASE(lock); continue; /* or try to do the rest? */ + } rc = replay_one_lock(imp, lock); + LDLM_LOCK_RELEASE(lock); } - atomic_dec(&imp->imp_replay_inflight); + cfs_atomic_dec(&imp->imp_replay_inflight); RETURN(rc); }