X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_request.c;h=1eaa0e1e6389e009d8bca7e104278db5a499e812;hp=32cfc160f23ff1e4f9c6879c22aaa3b2e791d591;hb=cf45b9d3ec9aa0872d3654a72f1b5cc9c5fff373;hpb=e6581008d8aa99688561da8200743d4d1368b11d diff --git a/lustre/ldlm/ldlm_request.c b/lustre/ldlm/ldlm_request.c index 32cfc16..1eaa0e1 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 2008 Sun Microsystems, Inc. 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 @@ -66,11 +78,11 @@ 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); @@ -87,8 +99,8 @@ 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); @@ -104,24 +116,84 @@ int ldlm_get_enq_timeout(struct ldlm_lock *lock) /* 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 = timeout + (timeout >> 1); /* 150% */ + timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */ return max(timeout, ldlm_enqueue_min); } +EXPORT_SYMBOL(ldlm_get_enq_timeout); -static int is_granted_or_cancelled(struct ldlm_lock *lock) +/** + * Helper function for ldlm_completion_ast(), updating timings when lock is + * actually granted. + */ +static int ldlm_completion_tail(struct ldlm_lock *lock) { - int ret = 0; + long delay; + int result; - 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); + 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_add(&lock->l_resource->lr_namespace->ns_at_estimate, 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 */ @@ -147,7 +219,6 @@ 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: @@ -180,12 +251,15 @@ noreproc: 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(lock->l_resource->lr_namespace) && + 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) { @@ -194,23 +268,61 @@ noreproc: RETURN(rc); } - LDLM_DEBUG(lock, "client-side enqueue waking up: granted after %lds", - cfs_time_current_sec() - lock->l_enqueued_time.tv_sec); + RETURN(ldlm_completion_tail(lock)); +} + +/** + * 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_nocheck(struct ldlm_lock *lock) +{ + int do_ast; + ENTRY; + + lock->l_flags |= LDLM_FL_CBPENDING; + do_ast = (!lock->l_readers && !lock->l_writers); + unlock_res_and_lock(lock); - /* Update our time estimate */ - at_add(&lock->l_resource->lr_namespace->ns_at_estimate, - cfs_time_current_sec() - lock->l_enqueued_time.tv_sec); + if (do_ast) { + struct lustre_handle lockh; + int rc; + LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel"); + ldlm_lock2handle(lock, &lockh); + rc = ldlm_cli_cancel(&lockh); + if (rc < 0) + CERROR("ldlm_cli_cancel: %d\n", rc); + } else { + LDLM_DEBUG(lock, "Lock still has references, will be " + "cancelled later"); + } RETURN(0); } -/* - * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs. +/** + * 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) { - int do_ast; ENTRY; if (flag == LDLM_CB_CANCELING) { @@ -228,25 +340,7 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, 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); - - if (do_ast) { - struct lustre_handle lockh; - int rc; - - LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel"); - ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); - if (rc < 0) - CERROR("ldlm_cli_cancel: %d\n", rc); - } else { - LDLM_DEBUG(lock, "Lock still has references, will be " - "cancelled later"); - } - RETURN(0); + RETURN(ldlm_blocking_ast_nocheck(lock)); } /* @@ -284,10 +378,15 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, ldlm_completion_callback completion, ldlm_glimpse_callback glimpse, void *data, __u32 lvb_len, void *lvb_swabber, + 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)); @@ -296,8 +395,7 @@ 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"); @@ -312,6 +410,8 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, 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; @@ -331,7 +431,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, LDLM_DEBUG(lock, "client-side local enqueue END"); EXIT; out: - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); out_nolock: return err; } @@ -340,13 +440,31 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns, struct ldlm_lock *lock, struct lustre_handle *lockh, 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; + 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_lock_decref_and_cancel(lockh, mode); + } else { + LDLM_DEBUG(lock, "lock was granted or failed in race"); + ldlm_lock_decref(lockh, mode); + } /* XXX - HACK because we shouldn't call ldlm_lock_destroy() * from llite/file.c/ll_file_flock(). */ @@ -362,6 +480,7 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, { struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; int is_replay = *flags & LDLM_FL_REPLAY; + struct lustre_handle old_hash_key; struct ldlm_lock *lock; struct ldlm_reply *reply; int cleanup_phase = 1; @@ -410,7 +529,15 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, cleanup_phase = 0; lock_res_and_lock(lock); + old_hash_key = lock->l_remote_handle; lock->l_remote_handle = reply->lock_handle; + + /* Key change rehash lock in per-export hash with new key */ + if (exp->exp_lock_hash) + lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key, + &lock->l_remote_handle, + &lock->l_exp_hash); + *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() @@ -511,7 +638,7 @@ cleanup: failed_lock_cleanup(ns, lock, lockh, 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; } @@ -523,7 +650,10 @@ static inline int ldlm_req_handles_avail(int req_size, int off) int avail; avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size; - avail /= sizeof(struct lustre_handle); + if (likely(avail >= 0)) + avail /= (int)sizeof(struct lustre_handle); + else + avail = 0; avail += LDLM_LOCKREQ_HANDLES - off; return avail; @@ -566,12 +696,12 @@ 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) @@ -645,15 +775,20 @@ 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 */ @@ -689,7 +824,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, LDLM_ENQUEUE); if (req == NULL) { failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode); - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); RETURN(-ENOMEM); } req_passed_in = 0; @@ -740,7 +875,9 @@ 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); @@ -748,7 +885,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, /* 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; @@ -868,7 +1005,7 @@ 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; @@ -917,7 +1054,7 @@ 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; @@ -997,7 +1134,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, ptlrpc_request_set_replen(req); if (flags & LDLM_FL_ASYNC) { - ptlrpcd_add_req(req); + ptlrpcd_add_req(req, PSCOPE_OTHER); sent = count; GOTO(out, 0); } else { @@ -1043,28 +1180,27 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) __u64 old_slv, new_slv; __u32 new_limit; ENTRY; - - if (unlikely(!req->rq_import || !req->rq_import->imp_obd || + if (unlikely(!req->rq_import || !req->rq_import->imp_obd || !imp_connect_lru_resize(req->rq_import))) { - /* - * Do nothing for corner cases. + /* + * 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. + * ref to obd export and thus access to server side namespace is no + * possible. */ - if (lustre_msg_get_slv(req->rq_repmsg) == 0 || + 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), + "(SLV: "LPU64", Limit: %u)", + lustre_msg_get_slv(req->rq_repmsg), lustre_msg_get_limit(req->rq_repmsg)); RETURN(0); } @@ -1073,12 +1209,12 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) new_slv = lustre_msg_get_slv(req->rq_repmsg); obd = req->rq_import->imp_obd; - /* - * Set new SLV and Limit to obd fields to make accessible for pool + /* + * 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. + * oops in that time. */ write_lock(&obd->obd_pool_lock); old_slv = obd->obd_pool_slv; @@ -1086,26 +1222,6 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) obd->obd_pool_limit = new_limit; write_unlock(&obd->obd_pool_lock); - /* - * 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) - if (old_slv > 0) { - __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE; - do_div(fast_change, 100); - - /* - * 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 RETURN(0); } EXPORT_SYMBOL(ldlm_cli_update_pool); @@ -1120,7 +1236,7 @@ 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); @@ -1128,7 +1244,7 @@ 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 @@ -1183,7 +1299,7 @@ static int ldlm_cancel_list(struct list_head *cancels, int count, int flags) 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); + LDLM_LOCK_RELEASE(lock); count--; } @@ -1196,71 +1312,6 @@ static int ldlm_cancel_list(struct list_head *cancels, int count, int flags) RETURN(count); } -/** - * Callback function for shrink 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_shrink_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; - - 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; - } - - /* - * 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; -} - /** * 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 @@ -1271,8 +1322,8 @@ static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns, * \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(); @@ -1280,8 +1331,8 @@ 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 + /* + * Stop lru processing when we reached passed @count or checked all * locks in lru. */ if (count && added >= count) @@ -1289,20 +1340,20 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, slv = ldlm_pool_get_slv(pl); lvf = ldlm_pool_get_lvf(pl); - la = cfs_duration_sec(cfs_time_sub(cur, + 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 + /* + * Stop when slv is not yet come from server or lv is smaller than * it is. */ lv = lvf * la * unused; - - /* - * Inform pool about current CLV to see it via proc. + + /* + * Inform pool about current CLV to see it via proc. */ ldlm_pool_set_clv(pl, lv); - return (slv == 1 || lv < slv) ? + return (slv == 1 || lv < slv) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } @@ -1316,15 +1367,15 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, * \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. + /* + * Stop lru processing when we reached passed @count or checked all + * locks in lru. */ - return (added >= count) ? + return (added >= count) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } @@ -1338,18 +1389,18 @@ static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns, * \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. + /* + * Stop lru processing if young lock is found and we reached passed + * @count. */ - return ((added >= 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; } @@ -1363,20 +1414,20 @@ static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns, * \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. + /* + * Stop lru processing when we reached passed @count or checked all + * locks in lru. */ - return (added >= count) ? + 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 @@ -1384,7 +1435,8 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) { 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) @@ -1393,10 +1445,10 @@ 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; @@ -1439,7 +1491,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, pf = ldlm_cancel_lru_policy(ns, flags); LASSERT(pf != NULL); - + while (!list_empty(&ns->ns_unused_list)) { /* For any flags, stop scanning if @max is reached. */ if (max && added >= max) @@ -1459,6 +1511,10 @@ 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); + spin_unlock(&ns->ns_unused_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. * @@ -1466,17 +1522,20 @@ 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) + if (pf(ns, lock, unused, added, count) == + LDLM_POLICY_KEEP_LOCK) { + lu_ref_del(&lock->l_reference, + __FUNCTION__, cfs_current()); + LDLM_LOCK_RELEASE(lock); + spin_lock(&ns->ns_unused_lock); break; - - LDLM_LOCK_GET(lock); /* dropped by bl thread */ - spin_unlock(&ns->ns_unused_lock); + } lock_res_and_lock(lock); /* Check flags again under the lock. */ @@ -1488,7 +1547,9 @@ 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); + lu_ref_del(&lock->l_reference, + __FUNCTION__, cfs_current()); + LDLM_LOCK_RELEASE(lock); spin_lock(&ns->ns_unused_lock); continue; } @@ -1496,8 +1557,8 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, /* 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; @@ -1518,6 +1579,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, LASSERT(list_empty(&lock->l_bl_ast)); list_add(&lock->l_bl_ast, cancels); unlock_res_and_lock(lock); + lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current()); spin_lock(&ns->ns_unused_lock); added++; unused--; @@ -1526,50 +1588,11 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, RETURN(ldlm_cancel_list(cancels, added, cancel_flags)); } -/* 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) -{ - 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); -} - /* 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 sync, int flags) { CFS_LIST_HEAD(cancels); @@ -1586,8 +1609,8 @@ int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync, RETURN(count); } - /* If an error occured in ASYNC mode, or - * this is SYNC mode, cancel the list. */ + /* 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); } @@ -1624,7 +1647,7 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, /* 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; @@ -1652,10 +1675,10 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, RETURN(ldlm_cancel_list(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, @@ -1667,8 +1690,8 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count, if (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 @@ -1721,12 +1744,14 @@ 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); 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); } @@ -1775,6 +1800,7 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, ldlm_resource_getref(res); spin_unlock(&ns->ns_hash_lock); + LDLM_RESOURCE_ADDREF(res); rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name, NULL, LCK_MINMODE, flags, opaque); @@ -1783,6 +1809,7 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n", res->lr_name.name[0], rc); + LDLM_RESOURCE_DELREF(res); spin_lock(&ns->ns_hash_lock); tmp = tmp->next; ldlm_resource_putref_locked(res); @@ -1793,56 +1820,6 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, 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++; - } - } - spin_unlock(&ns->ns_unused_lock); -unlock: - unlock_res(res); - ldlm_resource_putref(res); - RETURN(count); -} - /* Lock iterators. */ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, @@ -1921,9 +1898,11 @@ int ldlm_namespace_foreach_res(struct ldlm_namespace *ns, res = list_entry(tmp, struct ldlm_resource, lr_hash); ldlm_resource_getref(res); spin_unlock(&ns->ns_hash_lock); + LDLM_RESOURCE_ADDREF(res); rc = iter(res, closure); + LDLM_RESOURCE_DELREF(res); spin_lock(&ns->ns_hash_lock); tmp = tmp->next; ldlm_resource_putref_locked(res); @@ -1955,7 +1934,9 @@ void ldlm_resource_iterate(struct ldlm_namespace *ns, return; } + LDLM_RESOURCE_ADDREF(res); ldlm_resource_foreach(res, iter, data); + LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); EXIT; } @@ -1969,17 +1950,26 @@ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *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", 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)) + /* 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))) { 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 lustre_handle old_hash_key; + struct ldlm_lock *lock; + struct ldlm_reply *reply; + struct obd_export *exp; ENTRY; atomic_dec(&req->rq_import->imp_replay_inflight); @@ -2001,7 +1991,16 @@ static int replay_lock_interpret(struct ptlrpc_request *req, GOTO(out, rc = -ESTALE); } + old_hash_key = lock->l_remote_handle; 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) + lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key, + &lock->l_remote_handle, + &lock->l_exp_hash); + LDLM_DEBUG(lock, "replayed lock:"); ptlrpc_import_recovery_state_machine(req->rq_import); LDLM_LOCK_PUT(lock); @@ -2009,7 +2008,6 @@ out: if (rc != ELDLM_OK) ptlrpc_connect_import(req->rq_import, NULL); - RETURN(rc); } @@ -2090,8 +2088,8 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) CLASSERT(sizeof(*aa) <= sizeof(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); } @@ -2114,9 +2112,12 @@ int ldlm_replay_locks(struct obd_import *imp) list_for_each_entry_safe(lock, next, &list, l_pending_chain) { list_del_init(&lock->l_pending_chain); - if (rc) + if (rc) { + LDLM_LOCK_PUT(lock); continue; /* or try to do the rest? */ + } rc = replay_one_lock(imp, lock); + LDLM_LOCK_PUT(lock); } atomic_dec(&imp->imp_replay_inflight);