X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_request.c;h=95b2f6368dc65b643039f670753fd8285743f7ad;hp=b6fb1d1d0343cf9299ad7172e70cd207bc022536;hb=a1052417b78bc18898161c4ed44d79de4a1a2f23;hpb=e6ba48a08c9fe831b3f60985a72bb31b9806bb34;ds=sidebyside diff --git a/lustre/ldlm/ldlm_request.c b/lustre/ldlm/ldlm_request.c index b6fb1d1..95b2f63 100644 --- a/lustre/ldlm/ldlm_request.c +++ b/lustre/ldlm/ldlm_request.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,13 +24,40 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2010, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. */ +/** + * This file contains Asynchronous System Trap (AST) handlers and related + * LDLM request-processing routines. + * + * An AST is a callback issued on a lock when its state is changed. There are + * several different types of ASTs (callbacks) registered for each lock: + * + * - completion AST: when a lock is enqueued by some process, but cannot be + * granted immediately due to other conflicting locks on the same resource, + * the completion AST is sent to notify the caller when the lock is + * eventually granted + * + * - blocking AST: when a lock is granted to some process, if another process + * enqueues a conflicting (blocking) lock on a resource, a blocking AST is + * sent to notify the holder(s) of the lock(s) of the conflicting lock + * request. The lock holder(s) must release their lock(s) on that resource in + * a timely manner or be evicted by the server. + * + * - glimpse AST: this is used when a process wants information about a lock + * (i.e. the lock value block (LVB)) but does not necessarily require holding + * the lock. If the resource is locked, the lock holder(s) are sent glimpse + * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL + * their lock(s) if they are idle. If the resource is not locked, the server + * may grant the lock. + */ #define DEBUG_SUBSYSTEM S_LDLM #ifndef __KERNEL__ @@ -50,6 +75,9 @@ int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644, "lock enqueue timeout minimum"); +/* in client side, whether the cached locks will be canceled before replay */ +unsigned int ldlm_cancel_unused_locks_before_replay = 1; + static void interrupted_completion_wait(void *data) { } @@ -77,17 +105,22 @@ int ldlm_expired_completion_wait(void *data) if (ptlrpc_check_suspend()) RETURN(0); - 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)); + LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", " + CFS_DURATION_T"s ago)\n", + lock->l_last_activity, + cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity)); + LDLM_DEBUG(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); ldlm_namespace_dump(D_DLMTRACE, - lock->l_resource->lr_namespace); + ldlm_lock_to_ns(lock)); if (last_dump == 0) libcfs_debug_dumplog(); } @@ -105,18 +138,19 @@ int ldlm_expired_completion_wait(void *data) RETURN(0); } +EXPORT_SYMBOL(ldlm_expired_completion_wait); /* 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); + int timeout = at_get(ldlm_lock_to_ns_at(lock)); if (AT_OFF) return obd_timeout / 2; /* Since these are non-updating timeouts, we should be conservative. It would be nice to have some kind of "early reply" mechanism for lock callbacks too... */ - timeout = 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); @@ -127,23 +161,24 @@ EXPORT_SYMBOL(ldlm_get_enq_timeout); */ 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_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay); - result = 0; - } - return result; + long delay; + int result; + + if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) { + LDLM_DEBUG(lock, "client-side enqueue: destroyed"); + result = -EIO; + } else { + delay = cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity); + LDLM_DEBUG(lock, "client-side enqueue: granted after " + CFS_DURATION_T"s", delay); + + /* Update our time estimate */ + at_measured(ldlm_lock_to_ns_at(lock), + delay); + result = 0; + } + return result; } /** @@ -151,7 +186,7 @@ static int ldlm_completion_tail(struct ldlm_lock *lock) * 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) +int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data) { ENTRY; @@ -168,19 +203,19 @@ int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data) 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); } +EXPORT_SYMBOL(ldlm_completion_ast_async); /** - * Client side LDLM "completion" AST. This is called in several cases: + * Generic LDLM "completion" AST. This is called in several cases: * - * - when a reply to an ENQUEUE rpc is received from the server + * - 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 + * - 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 @@ -194,7 +229,7 @@ int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data) * or penultimate cases happen in some other thread. * */ -int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data) +int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) { /* XXX ALLOCATE - 160 bytes */ struct lock_wait_data lwd; @@ -218,7 +253,6 @@ 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); noreproc: @@ -245,13 +279,13 @@ noreproc: interrupted_completion_wait, &lwd); } - if (imp != NULL) { - spin_lock(&imp->imp_lock); - lwd.lwd_conn_cnt = imp->imp_conn_cnt; - spin_unlock(&imp->imp_lock); - } + if (imp != NULL) { + spin_lock(&imp->imp_lock); + lwd.lwd_conn_cnt = imp->imp_conn_cnt; + spin_unlock(&imp->imp_lock); + } - if (ns_is_client(lock->l_resource->lr_namespace) && + if (ns_is_client(ldlm_lock_to_ns(lock)) && OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST, OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) { lock->l_flags |= LDLM_FL_FAIL_LOC; @@ -270,14 +304,15 @@ noreproc: RETURN(ldlm_completion_tail(lock)); } +EXPORT_SYMBOL(ldlm_completion_ast); /** - * A helper to build a blocking ast function + * A helper to build a blocking AST function * - * Perform a common operation for blocking asts: + * Perform a common operation for blocking ASTs: * defferred lock cancellation. * - * \param lock the lock blocking or canceling ast was called on + * \param lock the lock blocking or canceling AST was called on * \retval 0 * \see mdt_blocking_ast * \see ldlm_blocking_ast @@ -297,7 +332,7 @@ int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock) LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel"); ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); + rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); if (rc < 0) CERROR("ldlm_cli_cancel: %d\n", rc); } else { @@ -306,6 +341,7 @@ int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock) } RETURN(0); } +EXPORT_SYMBOL(ldlm_blocking_ast_nocheck); /** * Server blocking AST @@ -333,7 +369,7 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, 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 + * takes the lr_lock, then by the time we get the lock, we might not * be the correct blocking function anymore. So check, and return * early, if so. */ if (lock->l_blocking_ast != ldlm_blocking_ast) { @@ -342,8 +378,9 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, } RETURN(ldlm_blocking_ast_nocheck(lock)); } +EXPORT_SYMBOL(ldlm_blocking_ast); -/* +/** * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See * comment in filter_intent_policy() on why you may need this. */ @@ -369,15 +406,19 @@ int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp) */ return -ELDLM_NO_LOCK_DATA; } +EXPORT_SYMBOL(ldlm_glimpse_ast); +/** + * Enqueue a local lock (typically on a server). + */ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, const struct ldlm_res_id *res_id, ldlm_type_t type, ldlm_policy_data_t *policy, - ldlm_mode_t mode, int *flags, + ldlm_mode_t mode, __u64 *flags, 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, enum lvb_type lvb_type, const __u64 *client_cookie, struct lustre_handle *lockh) { @@ -395,19 +436,20 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, LBUG(); } - lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len); + lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len, + lvb_type); if (unlikely(!lock)) GOTO(out_nolock, err = -ENOMEM); - LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created"); - ldlm_lock_addref_internal(lock, mode); ldlm_lock2handle(lock, lockh); - lock_res_and_lock(lock); + + /* NB: we don't have any lock now (lock_res_and_lock) + * because it's a new lock */ + ldlm_lock_addref_internal_nolock(lock, mode); lock->l_flags |= LDLM_FL_LOCAL; if (*flags & LDLM_FL_ATOMIC_CB) lock->l_flags |= LDLM_FL_ATOMIC_CB; - lock->l_lvb_swabber = lvb_swabber; - unlock_res_and_lock(lock); + if (policy != NULL) lock->l_policy_data = *policy; if (client_cookie != NULL) @@ -422,23 +464,20 @@ 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_RELEASE(lock); out_nolock: return err; } +EXPORT_SYMBOL(ldlm_cli_enqueue_local); 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; @@ -451,39 +490,50 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns, * 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_ATOMIC_CB | LDLM_FL_CBPENDING; need_cancel = 1; } unlock_res_and_lock(lock); - if (need_cancel) { + 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_FL_ATOMIC_CB | LDLM_FL_CBPENDING"); + else LDLM_DEBUG(lock, "lock was granted or failed in race"); - ldlm_lock_decref(lockh, mode); - } + + 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); } } +/** + * Finishing portion of client lock enqueue code. + * + * Called after receiving reply from server. + */ 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) + __u64 *flags, void *lvb, __u32 lvb_len, + 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; + int size = 0; ENTRY; lock = ldlm_handle2lock(lockh); @@ -493,59 +543,73 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, RETURN(-ENOLCK); } + LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len), + "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len); + if (rc != ELDLM_OK) { LASSERT(!is_replay); LDLM_DEBUG(lock, "client-side enqueue END (%s)", rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED"); - if (rc == ELDLM_LOCK_ABORTED) { - /* Before we return, swab the reply */ - reply = req_capsule_server_get(&req->rq_pill, - &RMF_DLM_REP); - if (reply == NULL) - rc = -EPROTO; - if (lvb_len) { - struct ost_lvb *tmplvb; - - req_capsule_set_size(&req->rq_pill, - &RMF_DLM_LVB, RCL_SERVER, - lvb_len); - tmplvb = req_capsule_server_swab_get(&req->rq_pill, - &RMF_DLM_LVB, - lvb_swabber); - if (tmplvb == NULL) - GOTO(cleanup, rc = -EPROTO); - if (lvb != NULL) - memcpy(lvb, tmplvb, lvb_len); - } - } - GOTO(cleanup, rc); - } - reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); - if (reply == NULL) - GOTO(cleanup, rc = -EPROTO); + if (rc != ELDLM_LOCK_ABORTED) + GOTO(cleanup, rc); + } + + /* Before we return, swab the reply */ + reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); + if (reply == NULL) + GOTO(cleanup, rc = -EPROTO); + + if (lvb_len != 0) { + LASSERT(lvb != NULL); + + size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, + RCL_SERVER); + if (size < 0) { + LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size); + GOTO(cleanup, rc = size); + } else if (unlikely(size > lvb_len)) { + LDLM_ERROR(lock, "Replied LVB is larger than " + "expectation, expected = %d, replied = %d", + lvb_len, size); + GOTO(cleanup, rc = -EINVAL); + } + } + + if (rc == ELDLM_LOCK_ABORTED) { + if (lvb_len != 0) + rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, + lvb, size); + GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED)); + } /* 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) - lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key, - &lock->l_remote_handle, - &lock->l_exp_hash); + if (exp->exp_lock_hash) { + /* In the function below, .hs_keycmp resolves to + * ldlm_export_lock_keycmp() */ + /* coverity[overrun-buffer-val] */ + cfs_hash_rehash_key(exp->exp_lock_hash, + &lock->l_remote_handle, + &reply->lock_handle, + &lock->l_exp_hash); + } else { + lock->l_remote_handle = reply->lock_handle; + } - *flags = reply->lock_flags; - lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS; + *flags = ldlm_flags_from_wire(reply->lock_flags); + lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & + LDLM_INHERIT_FLAGS); /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match() * to wait with no timeout as well */ - lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT; + lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & + LDLM_FL_NO_TIMEOUT); unlock_res_and_lock(lock); - CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n", + CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n", lock, reply->lock_handle.cookie, *flags); /* If enqueue returned a blocked lock but the completion handler has @@ -560,30 +624,27 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, lock->l_req_mode = newmode; } - if (memcmp(reply->lock_desc.l_resource.lr_name.name, - lock->l_resource->lr_name.name, - sizeof(struct ldlm_res_id))) { - CDEBUG(D_INFO, "remote intent success, locking " - "(%ld,%ld,%ld) instead of " - "(%ld,%ld,%ld)\n", - (long)reply->lock_desc.l_resource.lr_name.name[0], - (long)reply->lock_desc.l_resource.lr_name.name[1], - (long)reply->lock_desc.l_resource.lr_name.name[2], - (long)lock->l_resource->lr_name.name[0], - (long)lock->l_resource->lr_name.name[1], - (long)lock->l_resource->lr_name.name[2]); - - 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) - if (!(type == LDLM_IBITS && !(exp->exp_connect_flags & - OBD_CONNECT_IBITS))) - lock->l_policy_data = - reply->lock_desc.l_policy_data; + if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name, + &lock->l_resource->lr_name)) { + CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES + " instead of "DLDLMRES"\n", + PLDLMRES(&reply->lock_desc.l_resource), + PLDLMRES(lock->l_resource)); + + 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) + if (!(type == LDLM_IBITS && + !(exp_connect_flags(exp) & OBD_CONNECT_IBITS))) + /* We assume lock type cannot change on server*/ + ldlm_convert_policy_to_local(exp, + lock->l_resource->lr_type, + &reply->lock_desc.l_policy_data, + &lock->l_policy_data); if (type != LDLM_PLAIN) LDLM_DEBUG(lock,"client-side enqueue, new policy data"); } @@ -601,17 +662,20 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, /* If the lock has already been granted by a completion AST, don't * clobber the LVB with an older one. */ - if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) { - void *tmplvb; - - req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, - lvb_len); - tmplvb = req_capsule_server_swab_get(&req->rq_pill, - &RMF_DLM_LVB, - lvb_swabber); - if (tmplvb == NULL) - GOTO(cleanup, rc = -EPROTO); - memcpy(lock->l_lvb_data, tmplvb, lvb_len); + if (lvb_len != 0) { + /* We must lock or a racing completion might update lvb without + * letting us know and we'll clobber the correct value. + * Cannot unlock after the check either, a that still leaves + * a tiny window for completion to get in */ + lock_res_and_lock(lock); + if (lock->l_req_mode != lock->l_granted_mode) + rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, + lock->l_lvb_data, size); + unlock_res_and_lock(lock); + if (rc < 0) { + cleanup_phase = 1; + GOTO(cleanup, rc); + } } if (!is_replay) { @@ -620,7 +684,7 @@ 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; } } @@ -635,21 +699,25 @@ 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_RELEASE(lock); return rc; } +EXPORT_SYMBOL(ldlm_cli_enqueue_fini); -/* 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. */ +/** + * Estimate number of lock handles that would fit into request of given + * size. 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(int req_size, int off) { int avail; - avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size; + avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size; if (likely(avail >= 0)) avail /= (int)sizeof(struct lustre_handle); else @@ -675,25 +743,31 @@ static inline int ldlm_format_handles_avail(struct obd_import *imp, 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. */ +/** + * Cancel LRU locks and pack them into the enqueue request. Pack there the given + * \a count locks in \a cancels. + * + * This is to be called by functions preparing their own requests that + * might contain lists of locks to cancel in addition to actual operation + * that needs to be performed. + */ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, int version, int opc, int canceloff, - struct list_head *cancels, int count) + cfs_list_t *cancels, int count) { struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; struct req_capsule *pill = &req->rq_pill; struct ldlm_request *dlm = NULL; - int flags, avail, to_free, bufcount, pack = 0; + int flags, avail, to_free, pack = 0; CFS_LIST_HEAD(head); int rc; ENTRY; if (cancels == NULL) cancels = &head; - if (exp_connect_cancelset(exp)) { + if (ns_connect_cancelset(ns)) { /* Estimate the amount of available space in the request. */ - bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT); + req_capsule_filled_sizes(pill, RCL_CLIENT); avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff); flags = ns_connect_lru_resize(ns) ? @@ -701,9 +775,9 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, 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, what will make us slower. */ + /* Cancel LRU locks here _only_ if the server supports + * EARLY_CANCEL. Otherwise we have to send extra CANCEL + * RPC, which will make us slower. */ if (avail > count) count += ldlm_cancel_lru_local(ns, cancels, to_free, avail - count, 0, flags); @@ -721,7 +795,7 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, RETURN(rc); } - if (exp_connect_cancelset(exp)) { + if (ns_connect_cancelset(ns)) { if (canceloff) { dlm = req_capsule_client_get(pill, &RMF_DLM_REQ); LASSERT(dlm); @@ -733,35 +807,63 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, } /* Pack into the request @pack lock handles. */ ldlm_cli_cancel_list(cancels, pack, req, 0); - /* Prepare and send separate cancel rpc for others. */ + /* 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(0); } +EXPORT_SYMBOL(ldlm_prep_elc_req); int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, - struct list_head *cancels, int count) + cfs_list_t *cancels, int count) { return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, LDLM_ENQUEUE_CANCEL_OFF, cancels, count); } +EXPORT_SYMBOL(ldlm_prep_enqueue_req); + +struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len) +{ + struct ptlrpc_request *req; + int rc; + ENTRY; + + req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE); + if (req == NULL) + RETURN(ERR_PTR(-ENOMEM)); + + rc = ldlm_prep_enqueue_req(exp, req, NULL, 0); + if (rc) { + ptlrpc_request_free(req); + RETURN(ERR_PTR(rc)); + } + + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len); + ptlrpc_request_set_replen(req); + RETURN(req); +} +EXPORT_SYMBOL(ldlm_enqueue_pack); -/* If a request has some specific initialisation it is passed in @reqp, +/** + * Client-side lock enqueue. + * + * If a request has some specific initialisation it is passed in \a reqp, * otherwise it is created in ldlm_cli_enqueue. * - * Supports sync and async requests, pass @async flag accordingly. If a + * Supports sync and async requests, pass \a async flag accordingly. If a * request was created in ldlm_cli_enqueue and it is the async request, - * pass it to the caller in @reqp. */ + * pass it to the caller in \a reqp. + */ 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, __u64 *flags, + void *lvb, __u32 lvb_len, enum lvb_type lvb_type, + struct lustre_handle *lockh, int async) { - struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; + struct ldlm_namespace *ns; struct ldlm_lock *lock; struct ldlm_request *body; int is_replay = *flags & LDLM_FL_REPLAY; @@ -772,6 +874,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, LASSERT(exp != NULL); + ns = exp->exp_obd->obd_namespace; + /* 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) { @@ -780,21 +884,19 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, 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, &cbs, einfo->ei_cbdata, - lvb_len); - if (lock == NULL) - RETURN(-ENOMEM); + const struct ldlm_callback_suite cbs = { + .lcs_completion = einfo->ei_cb_cp, + .lcs_blocking = einfo->ei_cb_bl, + .lcs_glimpse = einfo->ei_cb_gl + }; + lock = ldlm_lock_create(ns, res_id, einfo->ei_type, + einfo->ei_mode, &cbs, einfo->ei_cbdata, + lvb_len, lvb_type); + 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 @@ -802,7 +904,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, * inodebits lock internally with both bits set. */ if (einfo->ei_type == LDLM_IBITS && - !(exp->exp_connect_flags & OBD_CONNECT_IBITS)) + !(exp_connect_flags(exp) & + OBD_CONNECT_IBITS)) lock->l_policy_data.l_inodebits.bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; @@ -812,9 +915,15 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, if (einfo->ei_type == LDLM_EXTENT) lock->l_req_extent = policy->l_extent; - LDLM_DEBUG(lock, "client-side enqueue START"); + LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n", + *flags); } + lock->l_conn_export = exp; + lock->l_export = NULL; + lock->l_blocking_ast = einfo->ei_cb_bl; + lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL)); + /* lock not sent to server yet */ if (reqp == NULL || *reqp == NULL) { @@ -823,7 +932,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE); if (req == NULL) { - failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode); + failed_lock_cleanup(ns, lock, einfo->ei_mode); LDLM_LOCK_RELEASE(lock); RETURN(-ENOMEM); } @@ -840,25 +949,20 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, DLM_LOCKREQ_OFF, len, (int)sizeof(*body)); } - lock->l_conn_export = exp; - lock->l_export = NULL; - lock->l_blocking_ast = einfo->ei_cb_bl; - /* Dump lock data into the request buffer */ body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); ldlm_lock2desc(lock, &body->lock_desc); - body->lock_flags = *flags; + body->lock_flags = ldlm_flags_to_wire(*flags); body->lock_handle[0] = *lockh; /* Continue as normal. */ if (!req_passed_in) { - 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); + 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); } /* @@ -880,7 +984,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, 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 */ @@ -897,6 +1001,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, RETURN(rc); } +EXPORT_SYMBOL(ldlm_cli_enqueue); static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, __u32 *flags) @@ -904,7 +1009,7 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, struct ldlm_resource *res; int rc; ENTRY; - if (ns_is_client(lock->l_resource->lr_namespace)) { + if (ns_is_client(ldlm_lock_to_ns(lock))) { CERROR("Trying to cancel local lock\n"); LBUG(); } @@ -915,7 +1020,7 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, ldlm_reprocess_all(res); rc = 0; } else { - rc = EDEADLOCK; + rc = LUSTRE_EDEADLK; } LDLM_DEBUG(lock, "client-side local convert handler END"); LDLM_LOCK_PUT(lock); @@ -960,7 +1065,7 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) body->lock_handle[0] = lock->l_remote_handle; body->lock_desc.l_req_mode = new_mode; - body->lock_flags = *flags; + body->lock_flags = ldlm_flags_to_wire(*flags); ptlrpc_request_set_replen(req); @@ -987,7 +1092,7 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) GOTO(out, rc); } } else { - rc = EDEADLOCK; + rc = LUSTRE_EDEADLK; } EXIT; out: @@ -995,26 +1100,29 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) ptlrpc_req_finished(req); return rc; } +EXPORT_SYMBOL(ldlm_cli_convert); -/* Cancel locks locally. +/** + * Cancel locks locally. * Returns: - * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server; - * LDLM_FL_CANCELING otherwise; - * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */ -static int ldlm_cli_cancel_local(struct ldlm_lock *lock) + * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server + * \retval LDLM_FL_CANCELING otherwise; + * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC. + */ +static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock) { - int rc = LDLM_FL_LOCAL_ONLY; + __u64 rc = LDLM_FL_LOCAL_ONLY; ENTRY; if (lock->l_conn_export) { - int local_only; + bool local_only; LDLM_DEBUG(lock, "client-side cancel"); /* Set this flag to prevent others from getting new references*/ lock_res_and_lock(lock); lock->l_flags |= LDLM_FL_CBPENDING; - local_only = (lock->l_flags & - (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK)); + local_only = !!(lock->l_flags & + (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK)); ldlm_cancel_callback(lock); rc = (lock->l_flags & LDLM_FL_BL_AST) ? LDLM_FL_BL_AST : LDLM_FL_CANCELING; @@ -1027,23 +1135,23 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock) } ldlm_lock_cancel(lock); } else { - if (ns_is_client(lock->l_resource->lr_namespace)) { + if (ns_is_client(ldlm_lock_to_ns(lock))) { LDLM_ERROR(lock, "Trying to cancel local lock"); LBUG(); } LDLM_DEBUG(lock, "server-side local cancel"); ldlm_lock_cancel(lock); ldlm_reprocess_all(lock->l_resource); - LDLM_DEBUG(lock, "server-side local cancel handler END"); } RETURN(rc); } -/* Pack @count locks in @head into ldlm_request buffer at the offset @off, - of the request @req. */ +/** + * Pack \a count locks in \a head into ldlm_request buffer of request \a req. + */ static void ldlm_cancel_pack(struct ptlrpc_request *req, - struct list_head *head, int count) + cfs_list_t *head, int count) { struct ldlm_request *dlm; struct ldlm_lock *lock; @@ -1063,7 +1171,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, /* XXX: it would be better to pack lock handles grouped by resource. * so that the server cancel would call filter_lvbo_update() less * frequently. */ - list_for_each_entry(lock, head, l_bl_ast) { + cfs_list_for_each_entry(lock, head, l_bl_ast) { if (!count--) break; LASSERT(lock->l_conn_export); @@ -1076,10 +1184,11 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, EXIT; } -/* 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) +/** + * Prepare and send a batched cancel RPC. It will include \a count lock + * handles of locks given in \a cancels list. */ +int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels, + int count, ldlm_cancel_flags_t flags) { struct ptlrpc_request *req = NULL; struct obd_import *imp; @@ -1090,9 +1199,9 @@ 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); + CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE)) + if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE)) RETURN(count); free = ldlm_format_handles_avail(class_exp2cliimp(exp), @@ -1101,8 +1210,6 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, count = free; while (1) { - int bufcount; - imp = class_exp2cliimp(exp); if (imp == NULL || imp->imp_invalid) { CDEBUG(D_DLMTRACE, @@ -1114,7 +1221,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, if (req == NULL) GOTO(out, rc = -ENOMEM); - bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT); + 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)); @@ -1123,8 +1230,6 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, ptlrpc_request_free(req); GOTO(out, rc); } - req->rq_no_resend = 1; - req->rq_no_delay = 1; req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL; req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL; @@ -1133,14 +1238,14 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, ldlm_cancel_pack(req, cancels, count); ptlrpc_request_set_replen(req); - if (flags & LDLM_FL_ASYNC) { - ptlrpcd_add_req(req, PSCOPE_OTHER); + if (flags & LCF_ASYNC) { + ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); sent = count; GOTO(out, 0); } else { rc = ptlrpc_queue_wait(req); } - if (rc == ESTALE) { + if (rc == LUSTRE_ESTALE) { CDEBUG(D_DLMTRACE, "client/server (nid %s) " "out of sync -- not fatal\n", libcfs_nid2str(req->rq_import-> @@ -1151,9 +1256,11 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, ptlrpc_req_finished(req); continue; } else if (rc != ELDLM_OK) { - CERROR("Got rc %d from cancel RPC: canceling " - "anyway\n", rc); - break; + /* -ESHUTDOWN is common on umount */ + CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, + "Got rc %d from cancel RPC: " + "canceling anyway\n", rc); + break; } sent = count; break; @@ -1164,6 +1271,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, out: return sent ? sent : rc; } +EXPORT_SYMBOL(ldlm_cli_cancel_req); static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) { @@ -1172,12 +1280,12 @@ static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) } /** - * Update client's obd pool related fields with new SLV and Limit from \a req. + * Update client's OBD pool related fields with new SLV and Limit from \a req. */ int ldlm_cli_update_pool(struct ptlrpc_request *req) { struct obd_device *obd; - __u64 old_slv, new_slv; + __u64 new_slv; __u32 new_limit; ENTRY; if (unlikely(!req->rq_import || !req->rq_import->imp_obd || @@ -1189,13 +1297,11 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) RETURN(0); } - /* - * 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. - */ + /* 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 reference to the OBD export and thus access to + * server-side namespace is not 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 " @@ -1209,27 +1315,31 @@ 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 - * 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. - */ - write_lock(&obd->obd_pool_lock); - old_slv = obd->obd_pool_slv; + /* Set new SLV and limit in OBD fields to make them accessible + * to the 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 at cleanup time. Evil races are possible + * which may cause Oops at that time. */ + write_lock(&obd->obd_pool_lock); obd->obd_pool_slv = new_slv; obd->obd_pool_limit = new_limit; - write_unlock(&obd->obd_pool_lock); + write_unlock(&obd->obd_pool_lock); RETURN(0); } EXPORT_SYMBOL(ldlm_cli_update_pool); -int ldlm_cli_cancel(struct lustre_handle *lockh) +/** + * Client side lock cancel. + * + * Lock must not have any readers or writers by this time. + */ +int ldlm_cli_cancel(struct lustre_handle *lockh, + ldlm_cancel_flags_t cancel_flags) { struct obd_export *exp; - int avail, flags, count = 1, rc = 0; + int avail, flags, count = 1; + __u64 rc = 0; struct ldlm_namespace *ns; struct ldlm_lock *lock; CFS_LIST_HEAD(cancels); @@ -1243,15 +1353,15 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) } rc = ldlm_cli_cancel_local(lock); - 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(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, &cancels); + if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) { + LDLM_LOCK_RELEASE(lock); + RETURN(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); exp = lock->l_conn_export; if (exp_connect_cancelset(exp)) { @@ -1260,49 +1370,57 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) RCL_CLIENT, 0); LASSERT(avail > 0); - ns = lock->l_resource->lr_namespace; + ns = ldlm_lock_to_ns(lock); flags = ns_connect_lru_resize(ns) ? LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1, - LDLM_FL_BL_AST, flags); + LCF_BL_AST, flags); } - ldlm_cli_cancel_list(&cancels, count, NULL, 0); + ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags); RETURN(0); } +EXPORT_SYMBOL(ldlm_cli_cancel); -/* XXX until we will have compound requests and can cut cancels from generic rpc - * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */ -static int ldlm_cancel_list(struct list_head *cancels, int count, int flags) +/** + * Locally cancel up to \a count locks in list \a cancels. + * Return the number of cancelled locks. + */ +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; + int left = 0, bl_ast = 0; + __u64 rc; left = count; - list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { + cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { if (left-- == 0) break; - if (flags & LDLM_FL_LOCAL_ONLY) { + if (flags & LCF_LOCAL) { rc = LDLM_FL_LOCAL_ONLY; ldlm_lock_cancel(lock); } else { rc = ldlm_cli_cancel_local(lock); } - if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) { - LDLM_DEBUG(lock, "Cancel lock separately"); - list_del_init(&lock->l_bl_ast); - list_add(&lock->l_bl_ast, &head); - bl_ast ++; + /* Until we have compound requests and can send LDLM_CANCEL + * requests batched with generic RPCs, we need to send cancels + * with the LDLM_FL_BL_AST flag in a separate RPC from + * the one being generated now. */ + 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. */ - list_del_init(&lock->l_bl_ast); + cfs_list_del_init(&lock->l_bl_ast); LDLM_LOCK_RELEASE(lock); count--; } - } if (bl_ast > 0) { count -= bl_ast; @@ -1311,11 +1429,43 @@ static int ldlm_cancel_list(struct list_head *cancels, int count, int flags) RETURN(count); } +EXPORT_SYMBOL(ldlm_cli_cancel_list_local); /** - * 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. + * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back + * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g. + * readahead requests, ...) + */ +static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, + struct ldlm_lock *lock, + int unused, int added, + int count) +{ + ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK; + ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery; + lock_res_and_lock(lock); + + /* don't check added & count since we want to process all locks + * from unused list */ + switch (lock->l_resource->lr_type) { + case LDLM_EXTENT: + case LDLM_IBITS: + if (cb && cb(lock)) + break; + default: + result = LDLM_POLICY_SKIP_LOCK; + lock->l_flags |= LDLM_FL_SKIPPED; + break; + } + + unlock_res_and_lock(lock); + RETURN(result); +} + +/** + * Callback function for LRU-resize policy. Decides 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 * @@ -1331,10 +1481,8 @@ 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 - * locks in lru. - */ + /* Stop LRU processing when we reach past @count or have checked all + * locks in LRU. */ if (count && added >= count) return LDLM_POLICY_KEEP_LOCK; @@ -1342,25 +1490,21 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, lvf = ldlm_pool_get_lvf(pl); la = cfs_duration_sec(cfs_time_sub(cur, lock->l_last_used)); + lv = lvf * la * unused; - /* - * 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); - /* - * Inform pool about current CLV to see it via proc. - */ - ldlm_pool_set_clv(pl, lv); - return (slv == 1 || lv < slv) ? - LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; + /* Stop when SLV is not yet come from server or lv is smaller than + * it is. */ + 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. + * \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 * @@ -1371,32 +1515,27 @@ static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns, 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; + /* Stop LRU processing when we reach past @count or have 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. + * Callback function for aged policy. Makes decision whether to keep \a lock in + * LRU for current 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) + 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 reach past count */ return ((added >= count) && cfs_time_before(cfs_time_current(), cfs_time_add(lock->l_last_used, @@ -1405,23 +1544,21 @@ static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns, } /** - * 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. + * Callback function for default policy. Makes decision whether to keep \a lock + * in LRU for current 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) + 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 reach past count or have checked all + * locks in LRU. */ return (added >= count) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } @@ -1433,6 +1570,9 @@ typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *, static ldlm_cancel_lru_policy_t ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) { + if (flags & LDLM_CANCEL_NO_WAIT) + return ldlm_cancel_no_wait_policy; + if (ns_connect_lru_resize(ns)) { if (flags & LDLM_CANCEL_SHRINK) /* We kill passed number of old locks. */ @@ -1449,42 +1589,50 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) return ldlm_cancel_default_policy; } -/* - Free space in lru for @count new locks, +/** + * - Free space in LRU for \a count new locks, * redundant unused locks are canceled locally; * - also cancel locally unused aged locks; - * - do not cancel more than @max locks; - * - GET the found locks and add them into the @cancels list. + * - do not cancel more than \a max locks; + * - GET the found locks and add them into the \a cancels list. * * A client lock can be added to the l_bl_ast list only when it is - * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL. - * 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. + * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing + * CANCEL. There are the following use cases: + * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and + * ldlm_cli_cancel(), which check and 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. * - * Calling policies for enabled lru resize: + * 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_LRUR - use LRU resize policy (SLV from server) to + * cancel not more than \a count locks; * - * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at - * the beginning of lru list); + * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at + * the beginning of LRU list); * - * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to + * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to * memory pressre policy function; * - * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy". + * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy". + * + * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible + * (typically before replaying locks) w/o + * sending any RPCs or waiting for any + * outstanding RPC to complete. */ -int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, - int count, int max, int cancel_flags, int flags) +static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, + int count, int max, int flags) { - ldlm_cancel_lru_policy_t pf; - struct ldlm_lock *lock, *next; - int added = 0, unused; - ENTRY; + ldlm_cancel_lru_policy_t pf; + struct ldlm_lock *lock, *next; + int added = 0, unused, remained; + ENTRY; - spin_lock(&ns->ns_unused_lock); + spin_lock(&ns->ns_lock); unused = ns->ns_nr_unused; + remained = unused; if (!ns_connect_lru_resize(ns)) count += unused - ns->ns_max_unused; @@ -1492,17 +1640,29 @@ 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)) { + while (!cfs_list_empty(&ns->ns_unused_list)) { + ldlm_policy_res_t result; + + /* all unused locks */ + if (remained-- <= 0) + break; + /* For any flags, stop scanning if @max is reached. */ if (max && added >= max) break; - list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){ + cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list, + l_lru) { /* No locks which got blocking requests. */ LASSERT(!(lock->l_flags & LDLM_FL_BL_AST)); - /* Somebody is already doing CANCEL. No need in this - * lock in lru, do not traverse it again. */ + if (flags & LDLM_CANCEL_NO_WAIT && + lock->l_flags & LDLM_FL_SKIPPED) + /* already processed */ + continue; + + /* Somebody is already doing CANCEL. No need for this + * lock in LRU, do not traverse it again. */ if (!(lock->l_flags & LDLM_FL_CANCELING)) break; @@ -1512,45 +1672,51 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, break; LDLM_LOCK_GET(lock); - spin_unlock(&ns->ns_unused_lock); + spin_unlock(&ns->ns_lock); lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current()); - /* Pass the lock through the policy filter and see if it - * should stay in lru. - * - * 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) { + /* 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 a new lock is a valuable resource even if + * it has a low weight. + * + * That is, for shrinker policy we drop only + * old locks, but additionally choose them by + * their weight. Big extent locks will stay in + * the cache. */ + result = pf(ns, lock, unused, added, count); + if (result == LDLM_POLICY_KEEP_LOCK) { lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current()); LDLM_LOCK_RELEASE(lock); - spin_lock(&ns->ns_unused_lock); - break; + spin_lock(&ns->ns_lock); + break; + } + if (result == LDLM_POLICY_SKIP_LOCK) { + lu_ref_del(&lock->l_reference, + __func__, cfs_current()); + LDLM_LOCK_RELEASE(lock); + spin_lock(&ns->ns_lock); + continue; } lock_res_and_lock(lock); /* Check flags again under the lock. */ if ((lock->l_flags & LDLM_FL_CANCELING) || (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 or the lock is matched - * is already not unused. */ + /* Another thread is removing lock from LRU, or + * somebody is already doing CANCEL, or there + * is a blocking request which will send cancel + * by itself, or the lock is no longer unused. */ unlock_res_and_lock(lock); lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current()); LDLM_LOCK_RELEASE(lock); - spin_lock(&ns->ns_unused_lock); + spin_lock(&ns->ns_lock); continue; } LASSERT(!lock->l_readers && !lock->l_writers); @@ -1572,64 +1738,79 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, /* We can't re-add to l_lru as it confuses the * refcounting in ldlm_lock_remove_from_lru() if an AST - * arrives after we drop ns_lock below. We use l_bl_ast + * arrives after we drop lr_lock below. We use l_bl_ast * and can't use l_pending_chain as it is used both on * server and client nevertheless bug 5666 says it is * used only on server */ - LASSERT(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, cancels); + LASSERT(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, cancels); unlock_res_and_lock(lock); lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current()); - spin_lock(&ns->ns_unused_lock); - added++; - unused--; - } - spin_unlock(&ns->ns_unused_lock); - RETURN(ldlm_cancel_list(cancels, added, cancel_flags)); + spin_lock(&ns->ns_lock); + added++; + unused--; + } + spin_unlock(&ns->ns_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 +/** + * Cancel at least \a nr locks from given namespace LRU. + * + * When called with LCF_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 flags) + * asked to call the callback. When called with LCF_ASYNC the blocking + * callback will be performed in this function. + */ +int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, + ldlm_cancel_flags_t cancel_flags, + int flags) { - CFS_LIST_HEAD(cancels); - int count, rc; - ENTRY; + CFS_LIST_HEAD(cancels); + int count, rc; + ENTRY; #ifndef __KERNEL__ - sync = LDLM_SYNC; /* force to be sync in user space */ + cancel_flags &= ~LCF_ASYNC; /* force to be sync in user space */ #endif - count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags); - if (sync == LDLM_ASYNC) { - rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count); - if (rc == 0) - RETURN(count); - } - - /* 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); + /* 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, cancel_flags); + if (rc == 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. */ +/** + * Find and cancel locally unused locks found on resource, matched to the + * given policy, mode. GET the found locks and add them into the \a cancels + * list. + */ int ldlm_cancel_resource_local(struct ldlm_resource *res, - struct list_head *cancels, + cfs_list_t *cancels, ldlm_policy_data_t *policy, ldlm_mode_t mode, int lock_flags, - int cancel_flags, void *opaque) + ldlm_cancel_flags_t cancel_flags, void *opaque) { struct ldlm_lock *lock; int count = 0; ENTRY; lock_res(res); - list_for_each_entry(lock, &res->lr_granted, l_res_link) { + cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) { if (opaque != NULL && lock->l_ast_data != opaque) { LDLM_ERROR(lock, "data %p doesn't match opaque %p", lock->l_ast_data, opaque); @@ -1637,16 +1818,11 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, continue; } - if (lock->l_readers || lock->l_writers) { - if (cancel_flags & LDLM_FL_WARN) { - LDLM_ERROR(lock, "lock in use"); - //LBUG(); - } + if (lock->l_readers || lock->l_writers) continue; - } - /* If somebody is already doing CANCEL, or blocking ast came, - * skip this lock. */ + /* If somebody is already doing CANCEL, or blocking AST came, + * skip this lock. */ if (lock->l_flags & LDLM_FL_BL_AST || lock->l_flags & LDLM_FL_CANCELING) continue; @@ -1665,40 +1841,46 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | lock_flags; - LASSERT(list_empty(&lock->l_bl_ast)); - list_add(&lock->l_bl_ast, cancels); + LASSERT(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, cancels); LDLM_LOCK_GET(lock); count++; } unlock_res(res); - RETURN(ldlm_cancel_list(cancels, count, cancel_flags)); + RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags)); } +EXPORT_SYMBOL(ldlm_cancel_resource_local); -/* 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 +/** + * Cancel client-side locks from a list and send/prepare cancel RPCs to the + * server. + * If \a req is NULL, send CANCEL request to server with handles of locks + * in the \a 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 - * buffer at the offset @off. - * Destroy @cancels at the end. */ -int ldlm_cli_cancel_list(struct list_head *cancels, int count, - struct ptlrpc_request *req, int flags) + * If \a req is not NULL, put handles of locks in \a cancels into the request + * buffer at the offset \a off. + * Destroy \a cancels at the end. + */ +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. * 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. + * there are too 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)) { @@ -1714,10 +1896,11 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count, cancels, 1, flags); } - if (res < 0) { - CERROR("ldlm_cli_cancel_list: %d\n", res); - res = count; - } + if (res < 0) { + CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, + "ldlm_cli_cancel_list: %d\n", res); + res = count; + } count -= res; ldlm_lock_list_put(cancels, l_bl_ast, res); @@ -1725,11 +1908,19 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count, LASSERT(count == 0); RETURN(0); } +EXPORT_SYMBOL(ldlm_cli_cancel_list); +/** + * Cancel all locks on a resource that have 0 readers/writers. + * + * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying + * to notify the server. */ 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); @@ -1746,86 +1937,74 @@ int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns, LDLM_RESOURCE_ADDREF(res); count = ldlm_cancel_resource_local(res, &cancels, policy, mode, - 0, flags, opaque); + 0, flags | LCF_BL_AST, opaque); rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags); if (rc != ELDLM_OK) - CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc); + CERROR("canceling unused lock "DLDLMRES": rc = %d\n", + PLDLMRES(res), rc); LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); RETURN(0); } +EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource); -static inline int have_no_nsresource(struct ldlm_namespace *ns) -{ - int no_resource = 0; - - spin_lock(&ns->ns_hash_lock); - if (ns->ns_resources == 0) - no_resource = 1; - spin_unlock(&ns->ns_hash_lock); +struct ldlm_cli_cancel_arg { + int lc_flags; + void *lc_opaque; +}; - RETURN(no_resource); +static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) +{ + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + struct ldlm_cli_cancel_arg *lc = arg; + + ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name, + NULL, LCK_MINMODE, lc->lc_flags, + lc->lc_opaque); + /* must return 0 for hash iteration */ + return 0; } -/* Cancel all locks on a namespace (or a specific resource, if given) +/** + * Cancel all locks on a namespace (or a specific resource, if given) * that have 0 readers/writers. * - * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying + * If flags & LCF_LOCAL, throw the locks away without trying * to notify the server. */ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, const struct ldlm_res_id *res_id, - int flags, void *opaque) + ldlm_cancel_flags_t flags, void *opaque) { - int i; + struct ldlm_cli_cancel_arg arg = { + .lc_flags = flags, + .lc_opaque = opaque, + }; + ENTRY; if (ns == NULL) RETURN(ELDLM_OK); - if (res_id) + if (res_id != NULL) { RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL, LCK_MINMODE, flags, opaque)); - - spin_lock(&ns->ns_hash_lock); - for (i = 0; i < RES_HASH_SIZE; i++) { - struct list_head *tmp; - tmp = ns->ns_hash[i].next; - while (tmp != &(ns->ns_hash[i])) { - struct ldlm_resource *res; - int rc; - - res = list_entry(tmp, struct ldlm_resource, lr_hash); - ldlm_resource_getref(res); - spin_unlock(&ns->ns_hash_lock); - - LDLM_RESOURCE_ADDREF(res); - rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name, - NULL, LCK_MINMODE, - flags, opaque); - - if (rc) - CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n", - res->lr_name.name[0], rc); - - LDLM_RESOURCE_DELREF(res); - spin_lock(&ns->ns_hash_lock); - tmp = tmp->next; - ldlm_resource_putref_locked(res); - } + } else { + cfs_hash_for_each_nolock(ns->ns_rs_hash, + ldlm_cli_hash_cancel_unused, &arg); + RETURN(ELDLM_OK); } - spin_unlock(&ns->ns_hash_lock); - - RETURN(ELDLM_OK); } +EXPORT_SYMBOL(ldlm_cli_cancel_unused); /* 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; @@ -1835,22 +2014,22 @@ 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); @@ -1859,6 +2038,7 @@ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, unlock_res(res); RETURN(rc); } +EXPORT_SYMBOL(ldlm_resource_foreach); struct iter_helper_data { ldlm_iterator_t iter; @@ -1871,56 +2051,39 @@ static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure) return helper->iter(lock, helper->closure); } -static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure) +static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) + { - return ldlm_resource_foreach(res, ldlm_iter_helper, closure); + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + + return ldlm_resource_foreach(res, ldlm_iter_helper, arg) == + LDLM_ITER_STOP; } -int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter, - void *closure) +void ldlm_namespace_foreach(struct ldlm_namespace *ns, + ldlm_iterator_t iter, void *closure) + { struct iter_helper_data helper = { iter: iter, closure: closure }; - return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper); -} -int ldlm_namespace_foreach_res(struct ldlm_namespace *ns, - ldlm_res_iterator_t iter, void *closure) -{ - int i, rc = LDLM_ITER_CONTINUE; - struct ldlm_resource *res; - struct list_head *tmp; + cfs_hash_for_each_nolock(ns->ns_rs_hash, + ldlm_res_iter_helper, &helper); - ENTRY; - spin_lock(&ns->ns_hash_lock); - for (i = 0; i < RES_HASH_SIZE; i++) { - tmp = ns->ns_hash[i].next; - while (tmp != &(ns->ns_hash[i])) { - res = list_entry(tmp, struct ldlm_resource, lr_hash); - ldlm_resource_getref(res); - spin_unlock(&ns->ns_hash_lock); - 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); - if (rc == LDLM_ITER_STOP) - GOTO(out, rc); - } - } - out: - spin_unlock(&ns->ns_hash_lock); - RETURN(rc); } +EXPORT_SYMBOL(ldlm_namespace_foreach); -/* 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) { @@ -1929,30 +2092,36 @@ 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_ADDREF(res); - ldlm_resource_foreach(res, iter, data); + rc = ldlm_resource_foreach(res, iter, data); LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); - EXIT; + RETURN(rc); } +EXPORT_SYMBOL(ldlm_resource_iterate); /* 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; } @@ -1960,13 +2129,12 @@ static int replay_lock_interpret(const struct lu_env *env, struct ptlrpc_request *req, struct ldlm_async_args *aa, int rc) { - 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); @@ -1985,22 +2153,26 @@ static int replay_lock_interpret(const struct lu_env *env, 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); + if (exp && exp->exp_lock_hash) { + /* In the function below, .hs_keycmp resolves to + * ldlm_export_lock_keycmp() */ + /* coverity[overrun-buffer-val] */ + cfs_hash_rehash_key(exp->exp_lock_hash, + &lock->l_remote_handle, + &reply->lock_handle, + &lock->l_exp_hash); + } else { + lock->l_remote_handle = reply->lock_handle; + } LDLM_DEBUG(lock, "replayed lock:"); ptlrpc_import_recovery_state_machine(req->rq_import); LDLM_LOCK_PUT(lock); out: if (rc != ELDLM_OK) - ptlrpc_connect_import(req->rq_import, NULL); + ptlrpc_connect_import(req->rq_import); RETURN(rc); } @@ -2028,6 +2200,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) ldlm_lock_cancel(lock); RETURN(0); } + /* * If granted mode matches the requested mode, this lock is granted. * @@ -2046,7 +2219,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED; else if (lock->l_granted_mode) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV; - else if (!list_empty(&lock->l_res_link)) + else if (!cfs_list_empty(&lock->l_res_link)) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT; else flags = LDLM_FL_REPLAY; @@ -2061,14 +2234,13 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); ldlm_lock2desc(lock, &body->lock_desc); - body->lock_flags = flags; + body->lock_flags = ldlm_flags_to_wire(flags); ldlm_lock2handle(lock, &body->lock_handle[0]); - if (lock->l_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, - lock->l_lvb_len); - } + if (lock->l_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, + lock->l_lvb_len); ptlrpc_request_set_replen(req); /* notify the server we've replayed all requests. * also, we mark the request to be put on a dedicated @@ -2078,16 +2250,45 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) LDLM_DEBUG(lock, "replaying lock:"); - atomic_inc(&req->rq_import->imp_replay_inflight); + cfs_atomic_inc(&req->rq_import->imp_replay_inflight); CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args)); aa = ptlrpc_req_async_args(req); aa->lock_handle = body->lock_handle[0]; req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret; - ptlrpcd_add_req(req, PSCOPE_OTHER); + ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); RETURN(0); } +/** + * Cancel as many unused locks as possible before replay. since we are + * in recovery, we can't wait for any outstanding RPCs to send any RPC + * to the server. + * + * Called only in recovery before replaying locks. there is no need to + * replay locks that are unused. since the clients may hold thousands of + * cached unused locks, dropping the unused locks can greatly reduce the + * load on the servers at recovery time. + */ +static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns) +{ + int canceled; + CFS_LIST_HEAD(cancels); + + CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before" + "replay for namespace %s (%d)\n", + ldlm_ns_name(ns), ns->ns_nr_unused); + + /* We don't need to care whether or not LRU resize is enabled + * because the LDLM_CANCEL_NO_WAIT policy doesn't use the + * count parameter */ + canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0, + LCF_LOCAL, LDLM_CANCEL_NO_WAIT); + + CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n", + canceled, ldlm_ns_name(ns)); +} + int ldlm_replay_locks(struct obd_import *imp) { struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; @@ -2097,21 +2298,32 @@ int ldlm_replay_locks(struct obd_import *imp) ENTRY; - LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); + LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0); + + /* don't replay locks if import failed recovery */ + if (imp->imp_vbr_failed) + RETURN(0); /* ensure this doesn't fall to 0 before all have been queued */ - atomic_inc(&imp->imp_replay_inflight); + cfs_atomic_inc(&imp->imp_replay_inflight); + + if (ldlm_cancel_unused_locks_before_replay) + ldlm_cancel_unused_locks_for_replay(ns); - (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); + ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); - list_for_each_entry_safe(lock, next, &list, l_pending_chain) { - list_del_init(&lock->l_pending_chain); - if (rc) + cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) { + cfs_list_del_init(&lock->l_pending_chain); + if (rc) { + LDLM_LOCK_RELEASE(lock); continue; /* or try to do the rest? */ + } rc = replay_one_lock(imp, lock); + LDLM_LOCK_RELEASE(lock); } - atomic_dec(&imp->imp_replay_inflight); + cfs_atomic_dec(&imp->imp_replay_inflight); RETURN(rc); } +EXPORT_SYMBOL(ldlm_replay_locks);