X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_request.c;h=f8088439824808d3a260396293555c3154e8f2e7;hb=675dd06e429ee9551d0f874f3461ac3e5091c039;hp=771ed62446ae0b27f535068cfe1b55cf231b842a;hpb=f83a8d69c1dc128af5c40811d222bd92ee73e437;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_request.c b/lustre/ldlm/ldlm_request.c index 771ed62..f808843 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 @@ -34,6 +46,10 @@ #include "ldlm_internal.h" +int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; +CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644, + "lock enqueue timeout minimum"); + static void interrupted_completion_wait(void *data) { } @@ -61,10 +77,12 @@ int ldlm_expired_completion_wait(void *data) if (ptlrpc_check_suspend()) RETURN(0); - LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); " - "not entering recovery in server code, just going " - "back to sleep", lock->l_enqueued_time.tv_sec, - CURRENT_SECONDS - lock->l_enqueued_time.tv_sec); + LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", " + CFS_DURATION_T"s ago); not entering recovery in " + "server code, just going back to sleep", + 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); @@ -79,14 +97,104 @@ int ldlm_expired_completion_wait(void *data) obd = lock->l_conn_export->exp_obd; imp = obd->u.cli.cl_import; ptlrpc_fail_import(imp, lwd->lwd_conn_cnt); - LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering " - "recovery for %s@%s", lock->l_enqueued_time.tv_sec, - CURRENT_SECONDS - lock->l_enqueued_time.tv_sec, - obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid); + LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", " + CFS_DURATION_T"s ago), entering recovery for %s@%s", + 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); } +/* 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 timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate); + 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); + +/** + * 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(&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)); + } + + 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 */ @@ -94,6 +202,7 @@ 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; @@ -111,15 +220,20 @@ 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; @@ -127,25 +241,26 @@ 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, - ((lock->l_req_mode == lock->l_granted_mode) || - (lock->l_flags & LDLM_FL_FAILED)), &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) { @@ -154,35 +269,25 @@ 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); @@ -203,6 +308,42 @@ 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 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); + } + 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. @@ -237,24 +378,27 @@ 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)); - if (unlikely(ns->ns_client)) { + if (unlikely(ns_is_client(ns))) { CERROR("Trying to enqueue local lock in a shadow namespace\n"); 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); @@ -262,10 +406,11 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, 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; @@ -276,46 +421,67 @@ 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 lustre_handle old_hash_key; struct ldlm_lock *lock; struct ldlm_reply *reply; int cleanup_phase = 1; @@ -334,19 +500,18 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED"); if (rc == ELDLM_LOCK_ABORTED) { /* Before we return, swab the reply */ - reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, - sizeof(*reply), - lustre_swab_ldlm_reply); - if (reply == NULL) { - CERROR("Can't unpack ldlm_reply\n"); + reply = req_capsule_server_get(&req->rq_pill, + &RMF_DLM_REP); + if (reply == NULL) rc = -EPROTO; - } if (lvb_len) { - void *tmplvb; - tmplvb = lustre_swab_repbuf(req, - DLM_REPLY_REC_OFF, - lvb_len, - lvb_swabber); + struct ost_lvb *tmplvb; + + 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) GOTO(cleanup, rc = -EPROTO); if (lvb != NULL) @@ -356,18 +521,23 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, GOTO(cleanup, rc); } - reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply), - lustre_swab_ldlm_reply); - if (reply == NULL) { - CERROR("Can't unpack ldlm_reply\n"); + reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); + if (reply == NULL) GOTO(cleanup, rc = -EPROTO); - } /* lock enqueued on the server */ 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) + cfs_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() @@ -403,12 +573,10 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, (long)lock->l_resource->lr_name.name[1], (long)lock->l_resource->lr_name.name[2]); - ldlm_lock_change_resource(ns, lock, - &reply->lock_desc.l_resource.lr_name); - if (lock->l_resource == NULL) { - LBUG(); + rc = ldlm_lock_change_resource(ns, lock, + &reply->lock_desc.l_resource.lr_name); + if (rc || lock->l_resource == NULL) GOTO(cleanup, rc = -ENOMEM); - } LDLM_DEBUG(lock, "client-side enqueue, new resource"); } if (with_policy) @@ -426,7 +594,7 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, * bug 7311). */ (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) { lock_res_and_lock(lock); - lock->l_flags |= LDLM_FL_CBPENDING; + lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; unlock_res_and_lock(lock); LDLM_DEBUG(lock, "enqueue reply includes blocking AST"); } @@ -435,8 +603,11 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, * clobber the LVB with an older one. */ if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) { void *tmplvb; - tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len, - lvb_swabber); + + 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) GOTO(cleanup, rc = -EPROTO); memcpy(lock->l_lvb_data, tmplvb, lvb_len); @@ -448,7 +619,7 @@ 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; } } @@ -463,84 +634,117 @@ 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; } /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into * a single page on the send/receive side. XXX: 512 should be changed * to more adequate value. */ -static inline int ldlm_req_handles_avail(struct obd_export *exp, - int *size, int bufcount, int off) +static inline int ldlm_req_handles_avail(int req_size, int off) { - int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512); - int old_size = size[DLM_LOCKREQ_OFF]; + int avail; - size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request); - avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic, - bufcount, 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; - size[DLM_LOCKREQ_OFF] = old_size; return avail; } -static inline int ldlm_cancel_handles_avail(struct obd_export *exp) +static inline int ldlm_capsule_handles_avail(struct req_capsule *pill, + enum req_location loc, + int off) +{ + int size = req_capsule_msg_size(pill, loc); + return ldlm_req_handles_avail(size, off); +} + +static inline int ldlm_format_handles_avail(struct obd_import *imp, + const struct req_format *fmt, + enum req_location loc, int off) { - int size[2] = { sizeof(struct ptlrpc_body), - sizeof(struct ldlm_request) }; - return ldlm_req_handles_avail(exp, size, 2, 0); + int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc); + return ldlm_req_handles_avail(size, off); } /* Cancel lru locks and pack them into the enqueue request. Pack there the given * @count locks in @cancels. */ -struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp, - int bufcount, int *size, - struct list_head *cancels, - int count) +int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, + int version, int opc, int canceloff, + cfs_list_t *cancels, int count) { - struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; - struct ldlm_request *dlm = NULL; - struct ptlrpc_request *req; + struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; + struct req_capsule *pill = &req->rq_pill; + struct ldlm_request *dlm = NULL; + int flags, avail, to_free, bufcount, pack = 0; CFS_LIST_HEAD(head); + int rc; ENTRY; - + if (cancels == NULL) cancels = &head; if (exp_connect_cancelset(exp)) { /* Estimate the amount of available space in the request. */ - int avail = ldlm_req_handles_avail(exp, size, bufcount, - LDLM_ENQUEUE_CANCEL_OFF); - LASSERT(avail >= count); - - /* Cancel lru locks here _only_ if the server supports + bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT); + avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff); + + 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 * EARLY_CANCEL. Otherwise we have to send extra CANCEL - * rpc right on enqueue, what will make it slower, vs. - * asynchronous rpc in blocking thread. */ - count += ldlm_cancel_lru_local(ns, cancels, - exp_connect_lru_resize(exp) ? 0 : 1, - avail - count, LDLM_CANCEL_AGED); - size[DLM_LOCKREQ_OFF] = - ldlm_request_bufsize(count, LDLM_ENQUEUE); - } - req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION, - LDLM_ENQUEUE, bufcount, size, NULL); - if (exp_connect_cancelset(exp) && req) { - dlm = lustre_msg_buf(req->rq_reqmsg, - DLM_LOCKREQ_OFF, sizeof(*dlm)); - /* Skip first lock handler in ldlm_request_pack(), this method - * will incrment @lock_count according to the lock handle amount - * actually written to the buffer. */ - dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF; - ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF, 0); + * rpc, what will make us slower. */ + if (avail > count) + count += ldlm_cancel_lru_local(ns, cancels, to_free, + avail - count, 0, flags); + if (avail > count) + pack = count; + else + pack = avail; + req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT, + ldlm_request_bufsize(pack, opc)); + } + + rc = ptlrpc_request_pack(req, version, opc); + if (rc) { + ldlm_lock_list_put(cancels, l_bl_ast, count); + RETURN(rc); + } + + if (exp_connect_cancelset(exp)) { + if (canceloff) { + dlm = req_capsule_client_get(pill, &RMF_DLM_REQ); + LASSERT(dlm); + /* Skip first lock handler in ldlm_request_pack(), + * this method will incrment @lock_count according + * to the lock handle amount actually written to + * the buffer. */ + dlm->lock_count = canceloff; + } + /* Pack into the request @pack lock handles. */ + ldlm_cli_cancel_list(cancels, pack, req, 0); + /* Prepare and send separate cancel rpc for others. */ + ldlm_cli_cancel_list(cancels, count - pack, NULL, 0); } else { ldlm_lock_list_put(cancels, l_bl_ast, count); } - RETURN(req); + RETURN(0); +} + +int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, + cfs_list_t *cancels, int count) +{ + return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, + LDLM_ENQUEUE_CANCEL_OFF, cancels, count); } /* If a request has some specific initialisation it is passed in @reqp, @@ -552,19 +756,16 @@ struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp, 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; - struct ldlm_request *body; - struct ldlm_reply *reply; - int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body), - [DLM_LOCKREQ_OFF] = sizeof(*body), - [DLM_REPLY_REC_OFF] = lvb_len }; - int is_replay = *flags & LDLM_FL_REPLAY; - int req_passed_in = 1, rc, err; + struct ldlm_lock *lock; + struct ldlm_request *body; + int is_replay = *flags & LDLM_FL_REPLAY; + int req_passed_in = 1; + int rc, err; struct ptlrpc_request *req; ENTRY; @@ -573,21 +774,25 @@ 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 @@ -611,22 +816,26 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, /* lock not sent to server yet */ if (reqp == NULL || *reqp == NULL) { - req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0); + req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), + &RQF_LDLM_ENQUEUE, + 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; if (reqp) *reqp = req; } else { + int len; + req = *reqp; - LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >= - sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n", - DLM_LOCKREQ_OFF, - lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF), - sizeof(*body)); + 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, (int)sizeof(*body)); } lock->l_conn_export = exp; @@ -634,15 +843,20 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, lock->l_blocking_ast = einfo->ei_cb_bl; /* Dump lock data into the request buffer */ - body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body)); + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); ldlm_lock2desc(lock, &body->lock_desc); body->lock_flags = *flags; body->lock_handle[0] = *lockh; /* Continue as normal. */ if (!req_passed_in) { - size[DLM_LOCKREPLY_OFF] = sizeof(*reply); - ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size); + if (lvb_len > 0) { + req_capsule_extend(&req->rq_pill, + &RQF_LDLM_ENQUEUE_LVB); + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, + RCL_SERVER, lvb_len); + } + ptlrpc_request_set_replen(req); } /* @@ -659,15 +873,17 @@ 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; @@ -681,12 +897,12 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, } static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, - int *flags) + __u32 *flags) { struct ldlm_resource *res; int rc; ENTRY; - if (lock->l_resource->lr_namespace->ns_client) { + if (ns_is_client(lock->l_resource->lr_namespace)) { CERROR("Trying to cancel local lock\n"); LBUG(); } @@ -708,16 +924,14 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, * conversion of locks which are on the waiting or converting queue */ /* Caller of this code is supposed to take care of lock readers/writers accounting */ -int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags) +int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) { - struct ldlm_request *body; - struct ldlm_reply *reply; - struct ldlm_lock *lock; - struct ldlm_resource *res; + struct ldlm_request *body; + struct ldlm_reply *reply; + struct ldlm_lock *lock; + struct ldlm_resource *res; struct ptlrpc_request *req; - int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body), - [DLM_LOCKREQ_OFF] = sizeof(*body) }; - int rc; + int rc; ENTRY; lock = ldlm_handle2lock(lockh); @@ -732,30 +946,29 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags) LDLM_DEBUG(lock, "client-side convert"); - req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export), - LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL); - if (!req) - GOTO(out, rc = -ENOMEM); + req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export), + &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION, + LDLM_CONVERT); + if (req == NULL) { + LDLM_LOCK_PUT(lock); + RETURN(-ENOMEM); + } - body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body)); + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); body->lock_handle[0] = lock->l_remote_handle; body->lock_desc.l_req_mode = new_mode; body->lock_flags = *flags; - size[DLM_LOCKREPLY_OFF] = sizeof(*reply); - ptlrpc_req_set_repsize(req, 2, size); + ptlrpc_request_set_replen(req); rc = ptlrpc_queue_wait(req); if (rc != ELDLM_OK) GOTO(out, rc); - reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply), - lustre_swab_ldlm_reply); - if (reply == NULL) { - CERROR ("Can't unpack ldlm_reply\n"); - GOTO (out, rc = -EPROTO); - } + reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); + if (reply == NULL) + GOTO(out, rc = -EPROTO); if (req->rq_status) GOTO(out, rc = req->rq_status); @@ -790,7 +1003,7 @@ 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; @@ -812,14 +1025,13 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock) } ldlm_lock_cancel(lock); } else { - if (lock->l_resource->lr_namespace->ns_client) { + if (ns_is_client(lock->l_resource->lr_namespace)) { 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); @@ -827,19 +1039,19 @@ 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, int off, - struct list_head *head, int count) +static void ldlm_cancel_pack(struct ptlrpc_request *req, + cfs_list_t *head, int count) { struct ldlm_request *dlm; struct ldlm_lock *lock; int max, packed = 0; ENTRY; - dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm)); + dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); LASSERT(dlm != NULL); /* Check the room in the request buffer. */ - max = lustre_msg_buflen(req->rq_reqmsg, off) - + 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; @@ -848,7 +1060,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, int off, /* 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); @@ -863,13 +1075,10 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, int off, /* 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 ldlm_request *body; - int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body), - [DLM_LOCKREQ_OFF] = sizeof(*body) }; struct obd_import *imp; int free, sent = 0; int rc = 0; @@ -878,41 +1087,51 @@ 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); - free = ldlm_req_handles_avail(exp, size, 2, 0); + free = ldlm_format_handles_avail(class_exp2cliimp(exp), + &RQF_LDLM_CANCEL, RCL_CLIENT, 0); if (count > free) count = free; - size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL); while (1) { + int bufcount; + imp = class_exp2cliimp(exp); if (imp == NULL || imp->imp_invalid) { - CDEBUG(D_HA, "skipping cancel on invalid import %p\n", - imp); - break; + CDEBUG(D_DLMTRACE, + "skipping cancel on invalid import %p\n", imp); + RETURN(count); } - req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2, - size, NULL); - if (!req) + req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL); + if (req == NULL) GOTO(out, rc = -ENOMEM); + bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT); + req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT, + ldlm_request_bufsize(count, LDLM_CANCEL)); + + rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL); + if (rc) { + ptlrpc_request_free(req); + GOTO(out, rc); + } 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); - body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, - sizeof(*body)); - ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count); + ldlm_cancel_pack(req, cancels, count); - ptlrpc_req_set_repsize(req, 1, NULL); - if (flags & LDLM_FL_ASYNC) { - ptlrpcd_add_req(req); + ptlrpc_request_set_replen(req); + if (flags & LCF_ASYNC) { + ptlrpcd_add_req(req, PSCOPE_OTHER); sent = count; GOTO(out, 0); } else { @@ -949,39 +1168,56 @@ 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 ldlm_pool *pl; + struct obd_device *obd; + __u64 old_slv, new_slv; + __u32 new_limit; ENTRY; - - if (!imp_connect_lru_resize(req->rq_import)) - RETURN(0); - - pl = ldlm_imp2pl(req->rq_import); - - spin_lock(&pl->pl_lock); -#ifdef __KERNEL__ + if (unlikely(!req->rq_import || !req->rq_import->imp_obd || + !imp_connect_lru_resize(req->rq_import))) { - __u64 old_slv, fast_slv_change; + /* + * Do nothing for corner cases. + */ + RETURN(0); + } - old_slv = ldlm_pool_get_slv(pl); - fast_slv_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE; - do_div(fast_slv_change, 100); -#endif - pl->pl_update_time = cfs_time_current(); - ldlm_pool_set_slv(pl, lustre_msg_get_slv(req->rq_repmsg)); - ldlm_pool_set_limit(pl, lustre_msg_get_limit(req->rq_repmsg)); -#ifdef __KERNEL__ - /* Wake up pools thread only if SLV has changed more than - * 5% 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 1s anyways. */ - if (old_slv > ldlm_pool_get_slv(pl) && - old_slv - ldlm_pool_get_slv(pl) > fast_slv_change) - ldlm_pools_wakeup(); + /* + * 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 || + 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), + lustre_msg_get_limit(req->rq_repmsg)); + RETURN(0); } -#endif - spin_unlock(&pl->pl_lock); + + new_limit = lustre_msg_get_limit(req->rq_repmsg); + 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 + * 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); RETURN(0); } @@ -989,35 +1225,226 @@ EXPORT_SYMBOL(ldlm_cli_update_pool); int ldlm_cli_cancel(struct lustre_handle *lockh) { + struct obd_export *exp; + int avail, flags, count = 1, rc = 0; + struct ldlm_namespace *ns; struct ldlm_lock *lock; CFS_LIST_HEAD(cancels); - int rc = 0; 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); } - + rc = ldlm_cli_cancel_local(lock); - list_add(&lock->l_bl_ast, &cancels); + if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) { + 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(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, &cancels); - if (rc == LDLM_FL_BL_AST) { - rc = ldlm_cli_cancel_req(lock->l_conn_export, &cancels, 1, 0); - } else if (rc == LDLM_FL_CANCELING) { - int avail = ldlm_cancel_handles_avail(lock->l_conn_export); - int count = 1; + exp = lock->l_conn_export; + if (exp_connect_cancelset(exp)) { + avail = ldlm_format_handles_avail(class_exp2cliimp(exp), + &RQF_LDLM_CANCEL, + RCL_CLIENT, 0); LASSERT(avail > 0); - count += ldlm_cancel_lru_local(lock->l_resource->lr_namespace, - &cancels, 0, avail - 1, - LDLM_CANCEL_AGED); - ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0); + + ns = lock->l_resource->lr_namespace; + flags = ns_connect_lru_resize(ns) ? + LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; + count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1, + LCF_BL_AST, flags); } - if (rc != LDLM_FL_CANCELING) - LDLM_LOCK_PUT(lock); - RETURN(rc < 0 ? rc : 0); + ldlm_cli_cancel_list(&cancels, count, NULL, 0); + RETURN(0); +} + +/* 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 */ +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; + cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { + if (left-- == 0) + break; + + if (flags & LCF_LOCAL) { + rc = LDLM_FL_LOCAL_ONLY; + ldlm_lock_cancel(lock); + } else { + rc = ldlm_cli_cancel_local(lock); + } + if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) { + LDLM_DEBUG(lock, "Cancel lock separately"); + 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. */ + cfs_list_del_init(&lock->l_bl_ast); + LDLM_LOCK_RELEASE(lock); + count--; + } + + } + if (bl_ast > 0) { + count -= bl_ast; + ldlm_cli_cancel_list(&head, bl_ast, NULL, 0); + } + + RETURN(count); +} + +/** + * 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, + int count) +{ + cfs_time_t cur = cfs_time_current(); + struct ldlm_pool *pl = &ns->ns_pool; + __u64 slv, lvf, lv; + cfs_time_t la; + + /* + * Stop lru processing when we reached passed @count or checked all + * locks in lru. + */ + if (count && added >= count) + return LDLM_POLICY_KEEP_LOCK; + + slv = ldlm_pool_get_slv(pl); + 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. + */ + lv = lvf * la * unused; + + /* + * 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; +} + +/** + * 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, + int unused, int added, + int 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; +} + +/** + * 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, + int unused, int added, + int 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))) ? + LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; +} + +/** + * 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, + int unused, int added, + int 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, + int, int); + +static ldlm_cancel_lru_policy_t +ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) +{ + if (ns_connect_lru_resize(ns)) { + if (flags & LDLM_CANCEL_SHRINK) + /* 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) + return ldlm_cancel_passed_policy; + } else { + if (flags & LDLM_CANCEL_AGED) + return ldlm_cancel_aged_policy; + } + + return ldlm_cancel_default_policy; } /* - Free space in lru for @count new locks, @@ -1031,186 +1458,185 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) * There are the following use cases: ldlm_cancel_resource_local(), * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this * flag properly. As any attempt to cancel a lock rely on this flag, - * l_bl_ast list is accessed later without any special locking. */ -int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, - int count, int max, int flags) + * l_bl_ast list is accessed later without any special locking. + * + * Calling policies for enabled lru resize: + * ---------------------------------------- + * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to + * cancel not more than @count locks; + * + * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at + * the beginning of lru list); + * + * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to + * memory pressre policy function; + * + * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy". + */ +static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, + int count, int max, int flags) { - cfs_time_t cur = cfs_time_current(); - int rc, added = 0, left, unused; + ldlm_cancel_lru_policy_t pf; struct ldlm_lock *lock, *next; - __u64 slv, lvf, lv; + int added = 0, unused; ENTRY; - spin_lock(&ns->ns_unused_lock); + cfs_spin_lock(&ns->ns_unused_lock); unused = ns->ns_nr_unused; - + if (!ns_connect_lru_resize(ns)) count += unused - ns->ns_max_unused; - while (!list_empty(&ns->ns_unused_list)) { - struct ldlm_pool *pl = &ns->ns_pool; + pf = ldlm_cancel_lru_policy(ns, flags); + LASSERT(pf != NULL); + while (!cfs_list_empty(&ns->ns_unused_list)) { + /* For any flags, stop scanning if @max is reached. */ if (max && added >= max) break; - list_for_each_entry(lock, &ns->ns_unused_list, l_lru) { - /* somebody is already doing CANCEL or there is a - * blocking request will send cancel. */ - if (!(lock->l_flags & LDLM_FL_CANCELING) && - !(lock->l_flags & LDLM_FL_BL_AST)) + 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)); + + /* Somebody is already doing CANCEL. No need in this + * lock in lru, do not traverse it again. */ + if (!(lock->l_flags & LDLM_FL_CANCELING)) break; + + ldlm_lock_remove_from_lru_nolock(lock); } if (&lock->l_lru == &ns->ns_unused_list) break; - if (ns_connect_lru_resize(ns)) { - cfs_time_t la; - - /* Take into account SLV only if cpount == 0. */ - if (count == 0) { - /* Calculate lv for every 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, - lock->l_last_used)); - if (la == 0) - la = 1; - - /* Stop when slv is not yet come from server - * or lv is smaller than it is. */ - lv = lvf * la * unused; - if (slv == 1 || lv < slv) - break; - } else { - if (added >= count) - break; - } - } else { - if ((added >= count) && - (!(flags & LDLM_CANCEL_AGED) || - cfs_time_before_64(cur, ns->ns_max_age + - lock->l_last_used))) - break; + LDLM_LOCK_GET(lock); + cfs_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. + * + * Even for shrinker policy we stop scanning if + * 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. + * + * That is, for shrinker policy we drop only + * old locks, but additionally chose them by + * their weight. Big extent locks will stay in + * the cache. */ + if (pf(ns, lock, unused, added, count) == + LDLM_POLICY_KEEP_LOCK) { + lu_ref_del(&lock->l_reference, + __FUNCTION__, cfs_current()); + LDLM_LOCK_RELEASE(lock); + cfs_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. */ if ((lock->l_flags & LDLM_FL_CANCELING) || - (lock->l_flags & LDLM_FL_BL_AST) || (ldlm_lock_remove_from_lru(lock) == 0)) { /* other thread is removing lock from lru or * somebody is already doing CANCEL or * there is a blocking request which will send - * cancel by itseft. */ + * 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_unused_lock); continue; } LASSERT(!lock->l_readers && !lock->l_writers); - /* If we have chosen to canecl 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 silently - cancelling this lock. */ + /* 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 + * silently cancelling this lock. */ lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK; - /* Setting the CBPENDING flag is a little misleading, but - * prevents an important race; namely, once CBPENDING is set, - * the lock can accumulate no more readers/writers. Since - * readers and writers are already zero here, ldlm_lock_decref - * won't see this flag and call l_blocking_ast */ + /* Setting the CBPENDING flag is a little misleading, + * but prevents an important race; namely, once + * CBPENDING is set, the lock can accumulate no more + * readers/writers. Since readers and writers are + * already zero here, ldlm_lock_decref() won't see + * this flag and call l_blocking_ast */ lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING; - /* 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 and can't use l_pending_chain - * as it is used both on server and client nevertheles bug 5666 - * says it is used only on server. --umka */ - - LASSERT(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, 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 + * 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(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_unused_lock); added++; unused--; } - spin_unlock(&ns->ns_unused_lock); - - /* Handle only @added inserted locks. */ - left = added; - list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { - if (left-- == 0) - break; - - rc = ldlm_cli_cancel_local(lock); - if (rc == LDLM_FL_BL_AST) { - CFS_LIST_HEAD(head); - - LDLM_DEBUG(lock, "Cancel lock separately"); - list_del_init(&lock->l_bl_ast); - list_add(&lock->l_bl_ast, &head); - ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0); - rc = LDLM_FL_LOCAL_ONLY; - } - 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); - added--; - } - - } + cfs_spin_unlock(&ns->ns_unused_lock); RETURN(added); } +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) +{ + 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); int count, rc; 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); - 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, 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 flags, void *opaque) + ldlm_cancel_flags_t cancel_flags, void *opaque) { - struct ldlm_lock *lock, *next; - int count = 0, left; + 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); @@ -1218,17 +1644,12 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, continue; } - if (lock->l_readers || lock->l_writers) { - if (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; @@ -1244,79 +1665,53 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, /* See CBPENDING comment in ldlm_cancel_lru */ lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | - lock_flags; + 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); - /* Handle only @count inserted locks. */ - left = count; - list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { - int rc = LDLM_FL_LOCAL_ONLY; - - if (left-- == 0) - break; - if (flags & LDLM_FL_LOCAL_ONLY) - ldlm_lock_cancel(lock); - else - rc = ldlm_cli_cancel_local(lock); - - if (rc == LDLM_FL_BL_AST) { - CFS_LIST_HEAD(head); - - LDLM_DEBUG(lock, "Cancel lock separately"); - list_del_init(&lock->l_bl_ast); - list_add(&lock->l_bl_ast, &head); - ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0); - rc = LDLM_FL_LOCAL_ONLY; - } - 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); - count--; - } - } - RETURN(count); + 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 off, 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)) { res = count; if (req) - ldlm_cancel_pack(req, off, cancels, count); + ldlm_cancel_pack(req, cancels, count); else res = ldlm_cli_cancel_req(lock->l_conn_export, - cancels, count, flags); + cancels, count, + flags); } else { res = ldlm_cli_cancel_req(lock->l_conn_export, cancels, 1, flags); @@ -1330,7 +1725,6 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count, count -= res; ldlm_lock_list_put(cancels, l_bl_ast, res); } - LASSERT(list_empty(cancels)); LASSERT(count == 0); RETURN(0); } @@ -1338,7 +1732,9 @@ 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); @@ -1353,12 +1749,14 @@ 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, 0, flags); + 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); } @@ -1367,10 +1765,10 @@ static inline int have_no_nsresource(struct ldlm_namespace *ns) { int no_resource = 0; - spin_lock(&ns->ns_hash_lock); + cfs_spin_lock(&ns->ns_hash_lock); if (ns->ns_resources == 0) no_resource = 1; - spin_unlock(&ns->ns_hash_lock); + cfs_spin_unlock(&ns->ns_hash_lock); RETURN(no_resource); } @@ -1378,11 +1776,11 @@ static inline int have_no_nsresource(struct ldlm_namespace *ns) /* 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; ENTRY; @@ -1395,18 +1793,20 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, LCK_MINMODE, flags, opaque)); - spin_lock(&ns->ns_hash_lock); + cfs_spin_lock(&ns->ns_hash_lock); for (i = 0; i < RES_HASH_SIZE; i++) { - struct list_head *tmp; + cfs_list_t *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); + res = cfs_list_entry(tmp, struct ldlm_resource, + lr_hash); ldlm_resource_getref(res); - spin_unlock(&ns->ns_hash_lock); + cfs_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); @@ -1415,71 +1815,23 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n", res->lr_name.name[0], rc); - spin_lock(&ns->ns_hash_lock); + LDLM_RESOURCE_DELREF(res); + cfs_spin_lock(&ns->ns_hash_lock); tmp = tmp->next; ldlm_resource_putref_locked(res); } } - spin_unlock(&ns->ns_hash_lock); + cfs_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->ns_client == LDLM_NAMESPACE_CLIENT); - - 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)) { - 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, void *closure) { - struct list_head *tmp, *next; + cfs_list_t *tmp, *next; struct ldlm_lock *lock; int rc = LDLM_ITER_CONTINUE; @@ -1489,22 +1841,22 @@ 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); @@ -1542,20 +1894,23 @@ int ldlm_namespace_foreach_res(struct ldlm_namespace *ns, { int i, rc = LDLM_ITER_CONTINUE; struct ldlm_resource *res; - struct list_head *tmp; + cfs_list_t *tmp; ENTRY; - spin_lock(&ns->ns_hash_lock); + cfs_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); + res = cfs_list_entry(tmp, struct ldlm_resource, + lr_hash); ldlm_resource_getref(res); - spin_unlock(&ns->ns_hash_lock); + cfs_spin_unlock(&ns->ns_hash_lock); + LDLM_RESOURCE_ADDREF(res); rc = iter(res, closure); - spin_lock(&ns->ns_hash_lock); + LDLM_RESOURCE_DELREF(res); + cfs_spin_lock(&ns->ns_hash_lock); tmp = tmp->next; ldlm_resource_putref_locked(res); if (rc == LDLM_ITER_STOP) @@ -1563,16 +1918,21 @@ int ldlm_namespace_foreach_res(struct ldlm_namespace *ns, } } out: - spin_unlock(&ns->ns_hash_lock); + cfs_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) { @@ -1581,49 +1941,56 @@ 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 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); + cfs_atomic_dec(&req->rq_import->imp_replay_inflight); if (rc != ELDLM_OK) GOTO(out, rc); - reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply), - lustre_swab_ldlm_reply); - if (reply == NULL) { - CERROR("Can't unpack ldlm_reply\n"); - GOTO (out, rc = -EPROTO); - } + reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); + if (reply == NULL) + GOTO(out, rc = -EPROTO); lock = ldlm_handle2lock(&aa->lock_handle); if (!lock) { @@ -1635,7 +2002,16 @@ 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) + cfs_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); @@ -1643,18 +2019,14 @@ out: if (rc != ELDLM_OK) ptlrpc_connect_import(req->rq_import, NULL); - RETURN(rc); } static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) { struct ptlrpc_request *req; - struct ldlm_request *body; - struct ldlm_reply *reply; struct ldlm_async_args *aa; - int buffers = 2; - int size[3] = { sizeof(struct ptlrpc_body) }; + struct ldlm_request *body; int flags; ENTRY; @@ -1691,31 +2063,30 @@ 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; - size[DLM_LOCKREQ_OFF] = sizeof(*body); - req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size, - NULL); - if (!req) + req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE, + LUSTRE_DLM_VERSION, LDLM_ENQUEUE); + if (req == NULL) RETURN(-ENOMEM); /* We're part of recovery, so don't wait for it. */ req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS; - body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body)); + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); ldlm_lock2desc(lock, &body->lock_desc); body->lock_flags = flags; ldlm_lock2handle(lock, &body->lock_handle[0]); - size[DLM_LOCKREPLY_OFF] = sizeof(*reply); if (lock->l_lvb_len != 0) { - buffers = 3; - size[DLM_REPLY_REC_OFF] = lock->l_lvb_len; + req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB); + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, + lock->l_lvb_len); } - ptlrpc_req_set_repsize(req, buffers, size); + ptlrpc_request_set_replen(req); /* notify the server we've replayed all requests. * also, we mark the request to be put on a dedicated * queue to be processed after all request replayes. @@ -1724,12 +2095,12 @@ 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); } @@ -1737,28 +2108,34 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) int ldlm_replay_locks(struct obd_import *imp) { struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; - struct list_head list; + CFS_LIST_HEAD(list); struct ldlm_lock *lock, *next; int rc = 0; ENTRY; - CFS_INIT_LIST_HEAD(&list); - 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); (void)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_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); + cfs_atomic_dec(&imp->imp_replay_inflight); RETURN(rc); }