/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (C) 2002 Cluster File Systems, Inc.
+ * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
*
- * This code is issued under the GNU General Public License.
- * See the file COPYING in this distribution
+ * This file is part of the Lustre file system, http://www.lustre.org
+ * Lustre is a trademark of Cluster File Systems, Inc.
*
- * by Cluster File Systems, Inc.
+ * 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.
+ *
+ * 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.
+ *
+ * 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.
*/
#define DEBUG_SUBSYSTEM S_LDLM
+#ifndef __KERNEL__
+#include <signal.h>
+#include <liblustre.h>
+#endif
+
+#include <lustre_dlm.h>
+#include <obd_class.h>
+#include <obd.h>
-#include <linux/lustre_dlm.h>
-#include <linux/obd_class.h>
-#include <linux/obd.h>
+#include "ldlm_internal.h"
-static int interrupted_completion_wait(void *data)
+static void interrupted_completion_wait(void *data)
{
- RETURN(1);
}
-static int expired_completion_wait(void *data)
+struct lock_wait_data {
+ struct ldlm_lock *lwd_lock;
+ __u32 lwd_conn_cnt;
+};
+
+struct ldlm_async_args {
+ struct lustre_handle lock_handle;
+};
+
+int ldlm_expired_completion_wait(void *data)
{
- struct ldlm_lock *lock = data;
- struct ptlrpc_connection *conn;
+ struct lock_wait_data *lwd = data;
+ struct ldlm_lock *lock = lwd->lwd_lock;
+ struct obd_import *imp;
struct obd_device *obd;
- if (!lock)
- CERROR("NULL lock\n");
- else if (!lock->l_connh)
- CERROR("lock %p has NULL connh\n", lock);
- else if (!(obd = class_conn2obd(lock->l_connh)))
- CERROR("lock %p has NULL obd\n", lock);
- else if (!(conn = obd->u.cli.cl_import.imp_connection))
- CERROR("lock %p has NULL connection\n", lock);
- else {
- class_signal_connection_failure(conn);
+ ENTRY;
+ if (lock->l_conn_export == NULL) {
+ static cfs_time_t next_dump = 0, last_dump = 0;
+
+ 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);
+ 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);
+ if (last_dump == 0)
+ libcfs_debug_dumplog();
+ }
+ RETURN(0);
}
+
+ 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);
+
RETURN(0);
}
-#if 0
-static int expired_completion_wait(void *data)
+static int is_granted_or_cancelled(struct ldlm_lock *lock)
+{
+ int ret = 0;
+
+ lock_res_and_lock(lock);
+ if (((lock->l_req_mode == lock->l_granted_mode) &&
+ !(lock->l_flags & LDLM_FL_CP_REQD)) ||
+ (lock->l_flags & LDLM_FL_FAILED))
+ ret = 1;
+ unlock_res_and_lock(lock);
+
+ return ret;
+}
+
+int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
{
- struct ldlm_lock *lock = data;
- struct ptlrpc_connection *conn =
- class_conn2cliimp(lock->l_connh)->imp_connection;
+ /* XXX ALLOCATE - 160 bytes */
+ struct lock_wait_data lwd;
+ struct obd_device *obd;
+ struct obd_import *imp = NULL;
+ struct l_wait_info lwi;
+ int rc = 0;
+ ENTRY;
+
+ if (flags == LDLM_FL_WAIT_NOREPROC) {
+ LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
+ goto noreproc;
+ }
+
+ if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
+ LDLM_FL_BLOCK_CONV))) {
+ cfs_waitq_signal(&lock->l_waitq);
+ RETURN(0);
+ }
+
+ 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)
+ imp = obd->u.cli.cl_import;
+
+ lwd.lwd_lock = lock;
+
+ if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
+ 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),
+ ldlm_expired_completion_wait,
+ 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);
+ }
+
+ /* Go to sleep until the lock is granted or cancelled. */
+ rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
- if (!conn) {
- CERROR("lock %p has NULL import connection\n", lock);
- RETURN(1);
+ if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
+ LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
+ RETURN(-EIO);
}
- class_signal_connection_failure(conn);
+ if (rc) {
+ LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
+ rc);
+ RETURN(rc);
+ }
+
+ LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
RETURN(0);
}
-#endif
-int ldlm_completion_ast(struct ldlm_lock *lock, int flags)
+/*
+ * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
+ */
+int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
+ void *data, int flag)
{
- struct l_wait_info lwi =
- LWI_TIMEOUT_INTR(obd_timeout * HZ, expired_completion_wait,
- interrupted_completion_wait, lock);
- int rc = 0;
+ int do_ast;
ENTRY;
- if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
- LDLM_FL_BLOCK_CONV)) {
- /* Go to sleep until the lock is granted. */
- /* FIXME: or cancelled. */
- LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock,"
- " sleeping");
- ldlm_lock_dump(lock);
- ldlm_reprocess_all(lock->l_resource);
- rc = l_wait_event(lock->l_waitq,
- (lock->l_req_mode == lock->l_granted_mode),
- &lwi);
- if (rc) {
- LDLM_DEBUG(lock,
- "client-side enqueue waking up: failed (%d)",
- rc);
- } else {
- LDLM_DEBUG(lock,
- "client-side enqueue waking up: granted");
- }
- } else if (flags == LDLM_FL_WAIT_NOREPROC) {
- rc = l_wait_event(lock->l_waitq,
- (lock->l_req_mode == lock->l_granted_mode),
- &lwi);
- } else if (flags == 0) {
- wake_up(&lock->l_waitq);
+ if (flag == LDLM_CB_CANCELING) {
+ /* Don't need to do anything here. */
+ RETURN(0);
}
- RETURN(rc);
+ 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);
+
+ if (do_ast) {
+ struct lustre_handle lockh;
+ int rc;
+
+ LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
+ ldlm_lock2handle(lock, &lockh);
+ rc = ldlm_cli_cancel(&lockh);
+ if (rc < 0)
+ CERROR("ldlm_cli_cancel: %d\n", rc);
+ } else {
+ LDLM_DEBUG(lock, "Lock still has references, will be "
+ "cancelled later");
+ }
+ RETURN(0);
+}
+
+/*
+ * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
+ * comment in filter_intent_policy() on why you may need this.
+ */
+int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
+{
+ /*
+ * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
+ * that is rather subtle: with OST-side locking, it may so happen that
+ * _all_ extent locks are held by the OST. If client wants to obtain
+ * current file size it calls ll{,u}_glimpse_size(), and (as locks are
+ * on the server), dummy glimpse callback fires and does
+ * nothing. Client still receives correct file size due to the
+ * following fragment in filter_intent_policy():
+ *
+ * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
+ * if (rc != 0 && res->lr_namespace->ns_lvbo &&
+ * res->lr_namespace->ns_lvbo->lvbo_update) {
+ * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
+ * }
+ *
+ * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
+ * returns correct file size to the client.
+ */
+ return -ELDLM_NO_LOCK_DATA;
}
-static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
- struct lustre_handle *parent_lockh,
- __u64 *res_id,
- __u32 type,
- void *cookie, int cookielen,
- ldlm_mode_t mode,
- int *flags,
- ldlm_completion_callback completion,
- ldlm_blocking_callback blocking,
- void *data,
- __u32 data_len,
- struct lustre_handle *lockh)
+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_blocking_callback blocking,
+ ldlm_completion_callback completion,
+ ldlm_glimpse_callback glimpse,
+ void *data, __u32 lvb_len, void *lvb_swabber,
+ struct lustre_handle *lockh)
{
struct ldlm_lock *lock;
int err;
+ ENTRY;
- if (ns->ns_client) {
- CERROR("Trying to cancel local lock\n");
+ LASSERT(!(*flags & LDLM_FL_REPLAY));
+ if (unlikely(ns_is_client(ns))) {
+ CERROR("Trying to enqueue local lock in a shadow namespace\n");
LBUG();
}
- lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode, data, data_len);
- if (!lock)
+ lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
+ completion, glimpse, data, lvb_len);
+ if (unlikely(!lock))
GOTO(out_nolock, err = -ENOMEM);
LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
ldlm_lock_addref_internal(lock, mode);
ldlm_lock2handle(lock, lockh);
- lock->l_connh = NULL;
+ lock_res_and_lock(lock);
+ 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 (type == LDLM_EXTENT)
+ lock->l_req_extent = policy->l_extent;
- err = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
- blocking);
- if (err != ELDLM_OK)
+ err = ldlm_lock_enqueue(ns, &lock, policy, flags);
+ if (unlikely(err != ELDLM_OK))
GOTO(out, err);
- if (type == LDLM_EXTENT)
- memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
- if ((*flags) & LDLM_FL_LOCK_CHANGED)
- memcpy(res_id, lock->l_resource->lr_name, sizeof(*res_id));
+ 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);
+ lock->l_completion_ast(lock, *flags, NULL);
LDLM_DEBUG(lock, "client-side local enqueue END");
EXIT;
return err;
}
-int ldlm_cli_enqueue(struct lustre_handle *connh,
- struct ptlrpc_request *req,
- struct ldlm_namespace *ns,
- struct lustre_handle *parent_lock_handle,
- __u64 *res_id,
- __u32 type,
- void *cookie, int cookielen,
- ldlm_mode_t mode,
- int *flags,
- ldlm_completion_callback completion,
- ldlm_blocking_callback blocking,
- void *data,
- __u32 data_len,
- struct lustre_handle *lockh)
+static void failed_lock_cleanup(struct ldlm_namespace *ns,
+ struct ldlm_lock *lock,
+ struct lustre_handle *lockh, int mode)
{
- struct ldlm_lock *lock;
- struct ldlm_request *body;
- struct ldlm_reply *reply;
- int rc, size = sizeof(*body), req_passed_in = 1;
- ENTRY;
-
- if (connh == NULL)
- return ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
- type, cookie, cookielen, mode,
- flags, completion, blocking, data,
- data_len, lockh);
+ /* Set a flag to prevent us from sending a CANCEL (bug 407) */
+ lock_res_and_lock(lock);
+ lock->l_flags |= LDLM_FL_LOCAL_ONLY;
+ unlock_res_and_lock(lock);
+ LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
- *flags = 0;
- lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type, mode,
- data, data_len);
- if (lock == NULL)
- GOTO(out_nolock, rc = -ENOMEM);
- LDLM_DEBUG(lock, "client-side enqueue START");
- /* for the local lock, add the reference */
- ldlm_lock_addref_internal(lock, mode);
- ldlm_lock2handle(lock, lockh);
-
- if (req == NULL) {
- req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
- &size, NULL);
- if (!req)
- GOTO(out, rc = -ENOMEM);
- req_passed_in = 0;
- } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
- LBUG();
+ ldlm_lock_decref_and_cancel(lockh, mode);
- /* Dump all of this data into the request buffer */
- body = lustre_msg_buf(req->rq_reqmsg, 0);
- ldlm_lock2desc(lock, &body->lock_desc);
- /* Phil: make this part of ldlm_lock2desc */
- if (type == LDLM_EXTENT)
- memcpy(&body->lock_desc.l_extent, cookie,
- sizeof(body->lock_desc.l_extent));
- body->lock_flags = *flags;
+ /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
+ * from llite/file.c/ll_file_flock(). */
+ if (lock->l_resource->lr_type == LDLM_FLOCK) {
+ ldlm_lock_destroy(lock);
+ }
+}
- memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
- if (parent_lock_handle)
- memcpy(&body->lock_handle2, parent_lock_handle,
- sizeof(body->lock_handle2));
+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 ldlm_namespace *ns = exp->exp_obd->obd_namespace;
+ int is_replay = *flags & LDLM_FL_REPLAY;
+ struct ldlm_lock *lock;
+ struct ldlm_reply *reply;
+ int cleanup_phase = 1;
+ ENTRY;
- /* Continue as normal. */
- if (!req_passed_in) {
- size = sizeof(*reply);
- req->rq_replen = lustre_msg_size(1, &size);
+ lock = ldlm_handle2lock(lockh);
+ /* ldlm_cli_enqueue is holding a reference on this lock. */
+ if (!lock) {
+ LASSERT(type == LDLM_FLOCK);
+ RETURN(-ENOLCK);
}
- lock->l_connh = connh;
- lock->l_export = NULL;
-
- rc = ptlrpc_queue_wait(req);
- /* FIXME: status check here? */
- rc = ptlrpc_check_status(req, rc);
if (rc != ELDLM_OK) {
+ LASSERT(!is_replay);
LDLM_DEBUG(lock, "client-side enqueue END (%s)",
rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
- ldlm_lock_decref(lockh, mode);
- /* FIXME: if we've already received a completion AST, this will
- * LBUG! */
- ldlm_lock_destroy(lock);
- GOTO(out, rc);
+ 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 = lustre_msg_buf(req->rq_repmsg, 0);
- memcpy(&lock->l_remote_handle, &reply->lock_handle,
- sizeof(lock->l_remote_handle));
- if (type == LDLM_EXTENT)
- memcpy(cookie, &reply->lock_extent, sizeof(reply->lock_extent));
+ 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);
+ lock->l_remote_handle = reply->lock_handle;
*flags = reply->lock_flags;
+ lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
+ /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
+ * to wait with no timeout as well */
+ lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
+ unlock_res_and_lock(lock);
- CDEBUG(D_INFO, "remote handle: %p, flags: %d\n",
- (void *)(unsigned long)reply->lock_handle.addr, *flags);
- CDEBUG(D_INFO, "extent: %Lu -> %Lu\n",
- (unsigned long long)reply->lock_extent.start,
- (unsigned long long)reply->lock_extent.end);
+ CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
+ lock, reply->lock_handle.cookie, *flags);
/* If enqueue returned a blocked lock but the completion handler has
* already run, then it fixed up the resource and we don't need to do it
* again. */
if ((*flags) & LDLM_FL_LOCK_CHANGED) {
- int newmode = reply->lock_mode;
+ int newmode = reply->lock_desc.l_req_mode;
+ LASSERT(!is_replay);
if (newmode && newmode != lock->l_req_mode) {
LDLM_DEBUG(lock, "server returned different mode %s",
ldlm_lockname[newmode]);
lock->l_req_mode = newmode;
}
- if (reply->lock_resource_name[0] !=
- lock->l_resource->lr_name[0]) {
- CDEBUG(D_INFO, "remote intent success, locking %ld "
- "instead of %ld\n",
- (long)reply->lock_resource_name[0],
- (long)lock->l_resource->lr_name[0]);
-
- ldlm_lock_change_resource(lock,
- reply->lock_resource_name);
+ 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]);
+
+ ldlm_lock_change_resource(ns, lock,
+ &reply->lock_desc.l_resource.lr_name);
if (lock->l_resource == NULL) {
LBUG();
- RETURN(-ENOMEM);
+ 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 (type != LDLM_PLAIN)
+ LDLM_DEBUG(lock,"client-side enqueue, new policy data");
}
- if (!req_passed_in)
- ptlrpc_free_req(req);
+ if ((*flags) & LDLM_FL_AST_SENT ||
+ /* Cancel extent locks as soon as possible on a liblustre client,
+ * because it cannot handle asynchronous ASTs robustly (see
+ * bug 7311). */
+ (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
+ lock_res_and_lock(lock);
+ lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
+ unlock_res_and_lock(lock);
+ LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
+ }
- rc = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
- blocking);
- if (lock->l_completion_ast)
- lock->l_completion_ast(lock, *flags);
+ /* 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 (!is_replay) {
+ rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
+ if (lock->l_completion_ast != NULL) {
+ int err = lock->l_completion_ast(lock, *flags, NULL);
+ if (!rc)
+ rc = err;
+ if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
+ cleanup_phase = 1;
+ }
+ }
+
+ if (lvb_len && lvb != NULL) {
+ /* Copy the LVB here, and not earlier, because the completion
+ * AST (if any) can override what we got in the reply */
+ memcpy(lvb, lock->l_lvb_data, lvb_len);
+ }
LDLM_DEBUG(lock, "client-side enqueue END");
EXIT;
- out:
+cleanup:
+ if (cleanup_phase == 1 && rc)
+ failed_lock_cleanup(ns, lock, lockh, mode);
+ /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
+ LDLM_LOCK_PUT(lock);
LDLM_LOCK_PUT(lock);
- out_nolock:
return rc;
}
-int ldlm_match_or_enqueue(struct lustre_handle *connh,
- struct ptlrpc_request *req,
- struct ldlm_namespace *ns,
- struct lustre_handle *parent_lock_handle,
- __u64 *res_id,
- __u32 type,
- void *cookie, int cookielen,
- ldlm_mode_t mode,
- int *flags,
- ldlm_completion_callback completion,
- ldlm_blocking_callback blocking,
- void *data,
- __u32 data_len,
- struct lustre_handle *lockh)
+/* 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 bufoff, int off)
+{
+ int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);
+ int old_size = size[bufoff];
+
+ size[bufoff] = sizeof(struct ldlm_request);
+ avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic,
+ bufcount, size);
+ avail /= sizeof(struct lustre_handle);
+ avail += LDLM_LOCKREQ_HANDLES - off;
+ size[bufoff] = old_size;
+
+ return avail;
+}
+
+static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
+{
+ int size[2] = { sizeof(struct ptlrpc_body),
+ sizeof(struct ldlm_request) };
+ return ldlm_req_handles_avail(exp, size, 2, DLM_LOCKREQ_OFF, 0);
+}
+
+/* Cancel lru locks and pack them into the enqueue request. Pack there the given
+ * @count locks in @cancels. */
+int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
+ int version, int opc, int canceloff,
+ struct list_head *cancels, int count)
{
+ 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 rc;
ENTRY;
- rc = ldlm_lock_match(ns, res_id, type, cookie, cookielen, mode, lockh);
- if (rc == 0) {
- rc = ldlm_cli_enqueue(connh, req, ns,
- parent_lock_handle, res_id, type, cookie,
- cookielen, mode, flags, completion,
- blocking, data, data_len, lockh);
- if (rc != ELDLM_OK)
- CERROR("ldlm_cli_enqueue: err: %d\n", rc);
+
+
+ LASSERT(cancels != NULL);
+
+ if (exp_connect_cancelset(exp)) {
+ /* Estimate the amount of available space in the request. */
+ bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
+ avail = ldlm_req_handles_avail(exp, pill->rc_area[RCL_CLIENT],
+ bufcount, bufcount - 1, 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, 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(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
+ ldlm_request_bufsize(count, opc));
+ }
+
+ rc = ptlrpc_request_pack(req, version, opc);
+ if (rc) {
+ ldlm_lock_list_put(cancels, l_bl_ast, count);
RETURN(rc);
- } else
+ }
+
+ 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(0);
+}
+
+int ldlm_prep_enqueue_req(struct obd_export *exp,
+ struct ptlrpc_request *req,
+ struct list_head *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,
+ * otherwise it is created in ldlm_cli_enqueue.
+ *
+ * Supports sync and async requests, pass @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. */
+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)
+{
+ struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
+ 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;
+
+ LASSERT(exp != NULL);
+
+ /* 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);
+ LASSERT(lock != NULL);
+ LDLM_DEBUG(lock, "client-side enqueue START");
+ LASSERT(exp == lock->l_conn_export);
+ } else {
+ 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);
+ 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
+ * descriptor (ldlm_lock2desc() below) but use an
+ * inodebits lock internally with both bits set.
+ */
+ if (einfo->ei_type == LDLM_IBITS &&
+ !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
+ lock->l_policy_data.l_inodebits.bits =
+ MDS_INODELOCK_LOOKUP |
+ MDS_INODELOCK_UPDATE;
+ else
+ lock->l_policy_data = *policy;
+ }
+
+ if (einfo->ei_type == LDLM_EXTENT)
+ lock->l_req_extent = policy->l_extent;
+ LDLM_DEBUG(lock, "client-side enqueue START");
+ }
+
+ /* lock not sent to server yet */
+
+ if (reqp == NULL || *reqp == NULL) {
+ 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);
+ RETURN(-ENOMEM);
+ }
+ req_passed_in = 0;
+ if (reqp)
+ *reqp = req;
+ } else {
+ int len;
+
+ req = *reqp;
+ len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
+ RCL_CLIENT);
+ LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
+ DLM_LOCKREQ_OFF, len, sizeof(*body));
+ }
+
+ 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_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);
+ }
+
+ /*
+ * Liblustre client doesn't get extent locks, except for O_APPEND case
+ * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
+ * [i_size, OBD_OBJECT_EOF] lock is taken.
+ */
+ LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
+ policy->l_extent.end == OBD_OBJECT_EOF));
+
+ if (async) {
+ LASSERT(reqp != NULL);
RETURN(0);
+ }
+
+ 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);
+
+ /* 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);
+ else
+ rc = err;
+
+ if (!req_passed_in && req != NULL) {
+ ptlrpc_req_finished(req);
+ if (reqp)
+ *reqp = NULL;
+ }
+
+ RETURN(rc);
}
static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
- int *flags)
+ __u32 *flags)
{
-
- if (lock->l_resource->lr_namespace->ns_client) {
+ struct ldlm_resource *res;
+ int rc;
+ ENTRY;
+ if (ns_is_client(lock->l_resource->lr_namespace)) {
CERROR("Trying to cancel local lock\n");
LBUG();
}
LDLM_DEBUG(lock, "client-side local convert");
- ldlm_lock_convert(lock, new_mode, flags);
- ldlm_reprocess_all(lock->l_resource);
-
+ res = ldlm_lock_convert(lock, new_mode, flags);
+ if (res) {
+ ldlm_reprocess_all(res);
+ rc = 0;
+ } else {
+ rc = EDEADLOCK;
+ }
LDLM_DEBUG(lock, "client-side local convert handler END");
LDLM_LOCK_PUT(lock);
- RETURN(0);
+ RETURN(rc);
}
-int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
+/* FIXME: one of ldlm_cli_convert or the server side should reject attempted
+ * 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, __u32 *flags)
{
- struct ldlm_request *body;
- struct lustre_handle *connh;
- 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 rc, size = sizeof(*body);
+ int rc;
ENTRY;
lock = ldlm_handle2lock(lockh);
RETURN(-EINVAL);
}
*flags = 0;
- connh = lock->l_connh;
- if (!connh)
- return ldlm_cli_convert_local(lock, new_mode, flags);
+ if (lock->l_conn_export == NULL)
+ RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
LDLM_DEBUG(lock, "client-side convert");
- req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &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, 0);
- memcpy(&body->lock_handle1, &lock->l_remote_handle,
- sizeof(body->lock_handle1));
+ 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 = sizeof(*reply);
- req->rq_replen = lustre_msg_size(1, &size);
+ ptlrpc_request_set_replen(req);
rc = ptlrpc_queue_wait(req);
- rc = ptlrpc_check_status(req, rc);
if (rc != ELDLM_OK)
GOTO(out, rc);
- reply = lustre_msg_buf(req->rq_repmsg, 0);
+ 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);
+
res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
- if (res != NULL)
+ if (res != NULL) {
ldlm_reprocess_all(res);
- /* Go to sleep until the lock is granted. */
- /* FIXME: or cancelled. */
- if (lock->l_completion_ast)
- lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
+ /* Go to sleep until the lock is granted. */
+ /* FIXME: or cancelled. */
+ if (lock->l_completion_ast) {
+ rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
+ NULL);
+ if (rc)
+ GOTO(out, rc);
+ }
+ } else {
+ rc = EDEADLOCK;
+ }
EXIT;
out:
LDLM_LOCK_PUT(lock);
- ptlrpc_free_req(req);
+ ptlrpc_req_finished(req);
return rc;
}
-int ldlm_cli_cancel(struct lustre_handle *lockh)
+/* 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)
{
- struct ptlrpc_request *req;
- struct ldlm_lock *lock;
- struct ldlm_request *body;
- int rc = 0, size = sizeof(*body);
+ int rc = LDLM_FL_LOCAL_ONLY;
ENTRY;
+
+ if (lock->l_conn_export) {
+ int local_only;
- lock = ldlm_handle2lock(lockh);
- if (!lock) {
- /* It's possible that the decref that we did just before this
- * cancel was the last reader/writer, and caused a cancel before
- * we could call this function. If we want to make this
- * impossible (by adding a dec_and_cancel() or similar), then
- * we can put the LBUG back. */
- //LBUG();
- RETURN(-EINVAL);
- }
-
- if (lock->l_connh) {
LDLM_DEBUG(lock, "client-side cancel");
/* Set this flag to prevent others from getting new references*/
- l_lock(&lock->l_resource->lr_namespace->ns_lock);
+ 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));
ldlm_cancel_callback(lock);
- l_unlock(&lock->l_resource->lr_namespace->ns_lock);
-
- req = ptlrpc_prep_req(class_conn2cliimp(lock->l_connh),
- LDLM_CANCEL, 1, &size, NULL);
- if (!req)
- GOTO(out, rc = -ENOMEM);
-
- body = lustre_msg_buf(req->rq_reqmsg, 0);
- memcpy(&body->lock_handle1, &lock->l_remote_handle,
- sizeof(body->lock_handle1));
-
- req->rq_replen = lustre_msg_size(0, NULL);
-
- rc = ptlrpc_queue_wait(req);
- rc = ptlrpc_check_status(req, rc);
- ptlrpc_free_req(req);
- if (rc != ELDLM_OK)
- GOTO(out, rc);
-
+ rc = (lock->l_flags & LDLM_FL_BL_AST) ?
+ LDLM_FL_BL_AST : LDLM_FL_CANCELING;
+ unlock_res_and_lock(lock);
+
+ if (local_only) {
+ CDEBUG(D_DLMTRACE, "not sending request (at caller's "
+ "instruction)\n");
+ rc = LDLM_FL_LOCAL_ONLY;
+ }
ldlm_lock_cancel(lock);
} else {
- LDLM_DEBUG(lock, "client-side local cancel");
- if (lock->l_resource->lr_namespace->ns_client) {
- CERROR("Trying to cancel local lock\n");
+ 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, "client-side local cancel handler END");
+ 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. */
+static void ldlm_cancel_pack(struct ptlrpc_request *req,
+ struct list_head *head, int count)
+{
+ struct ldlm_request *dlm;
+ struct ldlm_lock *lock;
+ int max, packed = 0;
+ ENTRY;
+
+ dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
+ LASSERT(dlm != NULL);
+
+ /* Check the room in the request buffer. */
+ 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;
+ LASSERT(max >= dlm->lock_count + count);
+
+ /* 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) {
+ if (!count--)
+ break;
+ LASSERT(lock->l_conn_export);
+ /* Pack the lock handle to the given request buffer. */
+ LDLM_DEBUG(lock, "packing");
+ dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
+ packed++;
+ }
+ CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
EXIT;
- out:
- LDLM_LOCK_PUT(lock);
- return rc;
}
-/* Cancel all locks on a given resource that have 0 readers/writers.
- *
- * If 'local_only' is true, throw the locks away without trying to notify the
- * server. */
-int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, __u64 *res_id,
- int local_only)
+/* 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)
{
- struct ldlm_resource *res;
- struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
- struct ldlm_ast_work *w;
+ struct ptlrpc_request *req = NULL;
+ struct obd_import *imp;
+ int free, sent = 0;
+ int rc = 0;
ENTRY;
- res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
- if (res == NULL)
- RETURN(-EINVAL);
+ LASSERT(exp != NULL);
+ LASSERT(count > 0);
- l_lock(&ns->ns_lock);
- list_for_each(tmp, &res->lr_granted) {
- struct ldlm_lock *lock;
- lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+ if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
+ RETURN(count);
- if (lock->l_readers || lock->l_writers)
- continue;
+ while (1) {
+ int bufcount;
+ struct req_capsule *pill;
+ imp = class_exp2cliimp(exp);
+ if (imp == NULL || imp->imp_invalid) {
+ CDEBUG(D_DLMTRACE,
+ "skipping cancel on invalid import %p\n", imp);
+ RETURN(count);
+ }
- /* 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;
+ req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
+ if (req == NULL)
+ GOTO(out, rc = -ENOMEM);
- OBD_ALLOC(w, sizeof(*w));
- LASSERT(w);
+ pill = &req->rq_pill;
+ bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
- w->w_lock = LDLM_LOCK_GET(lock);
- list_add(&w->w_list, &list);
- }
- l_unlock(&ns->ns_lock);
+ free = ldlm_req_handles_avail(exp, pill->rc_area[RCL_CLIENT],
+ bufcount, bufcount, 0);
+ if (count > free)
+ count = free;
- list_for_each_safe(tmp, next, &list) {
- struct lustre_handle lockh;
- int rc;
- w = list_entry(tmp, struct ldlm_ast_work, w_list);
-
- if (local_only)
- ldlm_lock_cancel(w->w_lock);
- else {
- ldlm_lock2handle(w->w_lock, &lockh);
- rc = ldlm_cli_cancel(&lockh);
- if (rc != ELDLM_OK)
- CERROR("ldlm_cli_cancel: %d\n", rc);
+ 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;
+
+ ldlm_cancel_pack(req, cancels, count);
+
+ ptlrpc_request_set_replen(req);
+ if (flags & LDLM_FL_ASYNC) {
+ ptlrpcd_add_req(req);
+ sent = count;
+ GOTO(out, 0);
+ } else {
+ rc = ptlrpc_queue_wait(req);
}
- LDLM_LOCK_PUT(w->w_lock);
- list_del(&w->w_list);
- OBD_FREE(w, sizeof(*w));
+ if (rc == ESTALE) {
+ CDEBUG(D_DLMTRACE, "client/server (nid %s) "
+ "out of sync -- not fatal\n",
+ libcfs_nid2str(req->rq_import->
+ imp_connection->c_peer.nid));
+ rc = 0;
+ } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
+ req->rq_import_generation == imp->imp_generation) {
+ ptlrpc_req_finished(req);
+ continue;
+ } else if (rc != ELDLM_OK) {
+ CERROR("Got rc %d from cancel RPC: canceling "
+ "anyway\n", rc);
+ break;
+ }
+ sent = count;
+ break;
}
- ldlm_resource_put(res);
+ ptlrpc_req_finished(req);
+ EXIT;
+out:
+ return sent ? sent : rc;
+}
+
+static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
+{
+ LASSERT(imp != NULL);
+ return &imp->imp_obd->obd_namespace->ns_pool;
+}
- RETURN(0);
+int ldlm_cli_update_pool(struct ptlrpc_request *req)
+{
+ __u64 old_slv, new_slv;
+ struct ldlm_pool *pl;
+ __u32 new_limit;
+ ENTRY;
+
+ if (!imp_connect_lru_resize(req->rq_import))
+ 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. */
+ 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);
+ }
+
+ new_limit = lustre_msg_get_limit(req->rq_repmsg);
+ new_slv = lustre_msg_get_slv(req->rq_repmsg);
+ pl = ldlm_imp2pl(req->rq_import);
+
+ spin_lock(&pl->pl_lock);
+ old_slv = ldlm_pool_get_slv(pl);
+ ldlm_pool_set_slv(pl, new_slv);
+ ldlm_pool_set_limit(pl, new_limit);
+
+ /* Check if we need to wakeup pools thread for fast SLV change.
+ * This is only done when threads period is noticably long like
+ * 10s or more. */
+#if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
+ {
+ __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
+ do_div(fast_change, 100);
+
+ /* Wake up pools thread only if SLV has changed more than
+ * 50% since last update. In this case we want to react asap.
+ * Otherwise it is no sense to wake up pools as they are
+ * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */
+ if (old_slv > new_slv && old_slv - new_slv > fast_change)
+ ldlm_pools_wakeup();
+ }
+#endif
+ spin_unlock(&pl->pl_lock);
+ RETURN(0);
+}
+EXPORT_SYMBOL(ldlm_cli_update_pool);
+
+int ldlm_cli_cancel(struct lustre_handle *lockh)
+{
+ int avail, flags, count = 1, rc = 0;
+ struct ldlm_namespace *ns;
+ struct ldlm_lock *lock;
+ CFS_LIST_HEAD(cancels);
+ ENTRY;
+
+ /* concurrent cancels on the same handle can happen */
+ lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
+ if (lock == NULL) {
+ LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
+ RETURN(0);
+ }
+
+ rc = ldlm_cli_cancel_local(lock);
+ if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
+ LDLM_LOCK_PUT(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 (exp_connect_cancelset(lock->l_conn_export)) {
+ avail = ldlm_cancel_handles_avail(lock->l_conn_export);
+ LASSERT(avail > 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,
+ LDLM_FL_BL_AST, flags);
+ }
+ 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 */
+static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
+{
+ CFS_LIST_HEAD(head);
+ struct ldlm_lock *lock, *next;
+ int left = 0, bl_ast = 0, rc;
+
+ left = count;
+ list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
+ if (left-- == 0)
+ break;
+
+ if (flags & LDLM_FL_LOCAL_ONLY) {
+ 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 ++;
+ continue;
+ }
+ if (rc == LDLM_FL_LOCAL_ONLY) {
+ /* CANCEL RPC should not be sent to server. */
+ list_del_init(&lock->l_bl_ast);
+ LDLM_LOCK_PUT(lock);
+ count--;
+ }
+
+ }
+ if (bl_ast > 0) {
+ count -= bl_ast;
+ ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
+ }
+
+ RETURN(count);
+}
+
+/* Return 1 to stop lru processing and keep current lock cached. Return zero
+ * otherwise. */
+static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
+ struct ldlm_lock *lock,
+ int unused, int added,
+ int count)
+{
+ int lock_cost;
+ __u64 page_nr;
+
+ /* Stop lru processing when we reached passed @count or checked all
+ * locks in lru. */
+ if (count && added >= count)
+ return LDLM_POLICY_KEEP_LOCK;
+
+ if (lock->l_resource->lr_type == LDLM_EXTENT) {
+ struct ldlm_extent *l_extent;
+
+ /* For all extent locks cost is 1 + number of pages in
+ * their extent. */
+ l_extent = &lock->l_policy_data.l_extent;
+ page_nr = (l_extent->end - l_extent->start);
+ do_div(page_nr, CFS_PAGE_SIZE);
+
+#ifdef __KERNEL__
+ /* XXX: In fact this is evil hack, we can't access inode
+ * here. For doing it right we need somehow to have number
+ * of covered by lock. This should be fixed later when 10718
+ * is landed. */
+ if (lock->l_ast_data != NULL) {
+ struct inode *inode = lock->l_ast_data;
+ if (page_nr > inode->i_mapping->nrpages)
+ page_nr = inode->i_mapping->nrpages;
+ }
+#endif
+ lock_cost = 1 + page_nr;
+ } else {
+ /* For all locks which are not extent ones cost is 1 */
+ lock_cost = 1;
+ }
+
+ /* Keep all expensive locks in lru for the memory pressure time
+ * cancel policy. They anyways may be canceled by lru resize
+ * pplicy if they have not small enough CLV. */
+ return lock_cost > ns->ns_shrink_thumb ?
+ LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
+}
+
+/* Return 1 to stop lru processing and keep current lock cached. Return zero
+ * otherwise. */
+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;
+
+ 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));
+
+ /* Stop when slv is not yet come from server or
+ * lv is smaller than it is. */
+ lv = lvf * la * unused;
+ return (slv == 1 || lv < slv) ?
+ LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
+}
+
+/* Return 1 to stop lru processing and keep current lock cached. Return zero
+ * otherwise. */
+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;
+}
+
+/* Return 1 to stop lru processing and keep current lock cached. Return zero
+ * otherwise. */
+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;
+}
+
+/* Return 1 to stop lru processing and keep current lock cached. Return zero
+ * otherwise. */
+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)
+ return ldlm_cancel_shrink_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,
+ * 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.
+ *
+ * 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.
+ *
+ * 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".
+ */
+int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
+ int count, int max, int cancel_flags, int flags)
+{
+ ldlm_cancel_lru_policy_t pf;
+ struct ldlm_lock *lock;
+ int added = 0, unused;
+ ENTRY;
+
+ spin_lock(&ns->ns_unused_lock);
+ unused = ns->ns_nr_unused;
+
+ if (!ns_connect_lru_resize(ns))
+ count += unused - ns->ns_max_unused;
+
+ pf = ldlm_cancel_lru_policy(ns, flags);
+ LASSERT(pf != NULL);
+
+ while (!list_empty(&ns->ns_unused_list)) {
+ /* 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))
+ break;
+ }
+ if (&lock->l_lru == &ns->ns_unused_list)
+ break;
+
+ /* 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)
+ 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. */
+ unlock_res_and_lock(lock);
+ LDLM_LOCK_PUT(lock);
+ spin_lock(&ns->ns_unused_lock);
+ continue;
+ }
+ LASSERT(!lock->l_readers && !lock->l_writers);
+
+ /* If we have chosen to cancel this lock voluntarily, we
+ * better send cancel notification to server, so that it
+ * frees appropriate state. This might lead to a race
+ * where while we are doing cancel here, server is also
+ * 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 */
+ 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 nevertheless bug 5666 says it is
+ * used only on server */
+ LASSERT(list_empty(&lock->l_bl_ast));
+ list_add(&lock->l_bl_ast, cancels);
+ unlock_res_and_lock(lock);
+ spin_lock(&ns->ns_unused_lock);
+ added++;
+ unused--;
+ }
+ spin_unlock(&ns->ns_unused_lock);
+ RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
+}
+
+/* Returns number of locks which could be canceled next time when
+ * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
+int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
+ int count, int max, int flags)
+{
+ ldlm_cancel_lru_policy_t pf;
+ struct ldlm_lock *lock;
+ int added = 0, unused;
+ ENTRY;
+
+ pf = ldlm_cancel_lru_policy(ns, flags);
+ LASSERT(pf != NULL);
+ spin_lock(&ns->ns_unused_lock);
+ unused = ns->ns_nr_unused;
+
+ list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
+ /* For any flags, stop scanning if @max is reached. */
+ if (max && added >= max)
+ break;
+
+ /* Somebody is already doing CANCEL or there is a
+ * blocking request will send cancel. Let's not count
+ * this lock. */
+ if ((lock->l_flags & LDLM_FL_CANCELING) ||
+ (lock->l_flags & LDLM_FL_BL_AST))
+ continue;
+
+ /* Pass the lock through the policy filter and see if it
+ * should stay in lru. */
+ if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
+ break;
+
+ added++;
+ unused--;
+ }
+ spin_unlock(&ns->ns_unused_lock);
+ RETURN(added);
+}
+
+/* when called with LDLM_ASYNC the blocking callback will be handled
+ * in a thread and this function will return after the thread has been
+ * asked to call the callback. when called with LDLM_SYNC the blocking
+ * callback will be performed in this function. */
+int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
+ int flags)
+{
+ CFS_LIST_HEAD(cancels);
+ int count, rc;
+ ENTRY;
+
+#ifndef __KERNEL__
+ sync = LDLM_SYNC; /* force to be sync in user space */
+#endif
+ count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
+ if (sync == LDLM_ASYNC) {
+ rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
+ if (rc == 0)
+ RETURN(count);
+ }
+
+ /* 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);
+}
+
+/* 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,
+ ldlm_policy_data_t *policy,
+ ldlm_mode_t mode, int lock_flags,
+ int 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) {
+ if (opaque != NULL && lock->l_ast_data != opaque) {
+ LDLM_ERROR(lock, "data %p doesn't match opaque %p",
+ lock->l_ast_data, opaque);
+ //LBUG();
+ continue;
+ }
+
+ if (lock->l_readers || lock->l_writers) {
+ if (cancel_flags & LDLM_FL_WARN) {
+ LDLM_ERROR(lock, "lock in use");
+ //LBUG();
+ }
+ continue;
+ }
+
+ /* 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;
+
+ if (lockmode_compat(lock->l_granted_mode, mode))
+ continue;
+
+ /* If policy is given and this is IBITS lock, add to list only
+ * those locks that match by policy. */
+ if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
+ !(lock->l_policy_data.l_inodebits.bits &
+ policy->l_inodebits.bits))
+ continue;
+
+ /* See CBPENDING comment in ldlm_cancel_lru */
+ 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);
+ LDLM_LOCK_GET(lock);
+ count++;
+ }
+ unlock_res(res);
+
+ RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
+}
+
+/* If @req is NULL, send CANCEL request to server with handles of locks
+ * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
+ * 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)
+{
+ struct ldlm_lock *lock;
+ int res = 0;
+ ENTRY;
+
+ if (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.
+ * 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(lock->l_conn_export);
+
+ if (exp_connect_cancelset(lock->l_conn_export)) {
+ res = count;
+ if (req)
+ ldlm_cancel_pack(req, cancels, count);
+ else
+ res = ldlm_cli_cancel_req(lock->l_conn_export,
+ cancels, count,
+ flags);
+ } else {
+ res = ldlm_cli_cancel_req(lock->l_conn_export,
+ cancels, 1, flags);
+ }
+
+ if (res < 0) {
+ CERROR("ldlm_cli_cancel_list: %d\n", res);
+ res = count;
+ }
+
+ count -= res;
+ ldlm_lock_list_put(cancels, l_bl_ast, res);
+ }
+ LASSERT(count == 0);
+ RETURN(0);
+}
+
+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)
+{
+ struct ldlm_resource *res;
+ CFS_LIST_HEAD(cancels);
+ int count;
+ int rc;
+ ENTRY;
+
+ res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
+ if (res == NULL) {
+ /* This is not a problem. */
+ CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
+ RETURN(0);
+ }
+
+ count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
+ 0, flags, opaque);
+ rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
+ if (rc != ELDLM_OK)
+ CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
+
+ ldlm_resource_putref(res);
+ RETURN(0);
+}
+
+static inline int have_no_nsresource(struct ldlm_namespace *ns)
+{
+ int no_resource = 0;
+
+ spin_lock(&ns->ns_hash_lock);
+ if (ns->ns_resources == 0)
+ no_resource = 1;
+ spin_unlock(&ns->ns_hash_lock);
+
+ RETURN(no_resource);
+}
+
+/* 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
+ * to notify the server. */
+int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
+ const struct ldlm_res_id *res_id,
+ int flags, void *opaque)
+{
+ int i;
+ ENTRY;
+
+ if (ns == NULL)
+ RETURN(ELDLM_OK);
+
+ if (res_id)
+ RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
+ LCK_MINMODE, flags,
+ opaque));
+
+ spin_lock(&ns->ns_hash_lock);
+ for (i = 0; i < RES_HASH_SIZE; i++) {
+ struct list_head *tmp;
+ tmp = ns->ns_hash[i].next;
+ while (tmp != &(ns->ns_hash[i])) {
+ struct ldlm_resource *res;
+ int rc;
+
+ res = list_entry(tmp, struct ldlm_resource, lr_hash);
+ ldlm_resource_getref(res);
+ spin_unlock(&ns->ns_hash_lock);
+
+ rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
+ NULL, LCK_MINMODE,
+ flags, opaque);
+
+ if (rc)
+ CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
+ res->lr_name.name[0], rc);
+
+ spin_lock(&ns->ns_hash_lock);
+ tmp = tmp->next;
+ ldlm_resource_putref_locked(res);
+ }
+ }
+ spin_unlock(&ns->ns_hash_lock);
+
+ RETURN(ELDLM_OK);
+}
+
+/* join/split resource locks to/from lru list */
+int ldlm_cli_join_lru(struct ldlm_namespace *ns,
+ const struct ldlm_res_id *res_id, int join)
+{
+ struct ldlm_resource *res;
+ struct ldlm_lock *lock, *n;
+ int count = 0;
+ ENTRY;
+
+ LASSERT(ns_is_client(ns));
+
+ res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
+ if (res == NULL)
+ RETURN(count);
+ LASSERT(res->lr_type == LDLM_EXTENT);
+
+ lock_res(res);
+ if (!join)
+ goto split;
+
+ list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
+ if (list_empty(&lock->l_lru) &&
+ !lock->l_readers && !lock->l_writers &&
+ !(lock->l_flags & LDLM_FL_LOCAL) &&
+ !(lock->l_flags & LDLM_FL_CBPENDING)) {
+ 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;
+ struct ldlm_lock *lock;
+ int rc = LDLM_ITER_CONTINUE;
+
+ ENTRY;
+
+ if (!res)
+ 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);
+
+ 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);
+
+ 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);
+
+ if (iter(lock, closure) == LDLM_ITER_STOP)
+ GOTO(out, rc = LDLM_ITER_STOP);
+ }
+ out:
+ unlock_res(res);
+ RETURN(rc);
+}
+
+struct iter_helper_data {
+ ldlm_iterator_t iter;
+ void *closure;
+};
+
+static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
+{
+ struct iter_helper_data *helper = closure;
+ return helper->iter(lock, helper->closure);
+}
+
+static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
+{
+ return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
+}
+
+int 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;
+
+ ENTRY;
+ spin_lock(&ns->ns_hash_lock);
+ for (i = 0; i < RES_HASH_SIZE; i++) {
+ tmp = ns->ns_hash[i].next;
+ while (tmp != &(ns->ns_hash[i])) {
+ res = list_entry(tmp, struct ldlm_resource, lr_hash);
+ ldlm_resource_getref(res);
+ spin_unlock(&ns->ns_hash_lock);
+
+ rc = iter(res, closure);
+
+ spin_lock(&ns->ns_hash_lock);
+ tmp = tmp->next;
+ ldlm_resource_putref_locked(res);
+ if (rc == LDLM_ITER_STOP)
+ GOTO(out, rc);
+ }
+ }
+ out:
+ spin_unlock(&ns->ns_hash_lock);
+ RETURN(rc);
+}
+
+/* non-blocking function to manipulate a lock whose cb_data is being put away.*/
+void ldlm_resource_iterate(struct ldlm_namespace *ns,
+ const struct ldlm_res_id *res_id,
+ ldlm_iterator_t iter, void *data)
+{
+ struct ldlm_resource *res;
+ ENTRY;
+
+ if (ns == NULL) {
+ CERROR("must pass in namespace\n");
+ LBUG();
+ }
+
+ res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
+ if (res == NULL) {
+ EXIT;
+ return;
+ }
+
+ ldlm_resource_foreach(res, iter, data);
+ ldlm_resource_putref(res);
+ EXIT;
+}
+
+/* Lock replay */
+
+static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
+{
+ struct list_head *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",
+ 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);
+ return LDLM_ITER_CONTINUE;
+}
+
+static int replay_lock_interpret(struct ptlrpc_request *req,
+ struct ldlm_async_args *aa, int rc)
+{
+ struct ldlm_lock *lock;
+ struct ldlm_reply *reply;
+
+ ENTRY;
+ atomic_dec(&req->rq_import->imp_replay_inflight);
+ if (rc != ELDLM_OK)
+ GOTO(out, rc);
+
+
+ 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) {
+ CERROR("received replay ack for unknown local cookie "LPX64
+ " remote cookie "LPX64 " from server %s id %s\n",
+ aa->lock_handle.cookie, reply->lock_handle.cookie,
+ req->rq_export->exp_client_uuid.uuid,
+ libcfs_id2str(req->rq_peer));
+ GOTO(out, rc = -ESTALE);
+ }
+
+ 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);
+
+
+ RETURN(rc);
+}
+
+static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
+{
+ struct ptlrpc_request *req;
+ struct ldlm_async_args *aa;
+ struct ldlm_request *body;
+ int flags;
+ ENTRY;
+
+
+ /* Bug 11974: Do not replay a lock which is actively being canceled */
+ if (lock->l_flags & LDLM_FL_CANCELING) {
+ LDLM_DEBUG(lock, "Not replaying canceled lock:");
+ RETURN(0);
+ }
+
+ /* If this is reply-less callback lock, we cannot replay it, since
+ * server might have long dropped it, but notification of that event was
+ * lost by network. (and server granted conflicting lock already) */
+ if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
+ LDLM_DEBUG(lock, "Not replaying reply-less lock:");
+ ldlm_lock_cancel(lock);
+ RETURN(0);
+ }
+ /*
+ * If granted mode matches the requested mode, this lock is granted.
+ *
+ * If they differ, but we have a granted mode, then we were granted
+ * one mode and now want another: ergo, converting.
+ *
+ * If we haven't been granted anything and are on a resource list,
+ * then we're blocked/waiting.
+ *
+ * If we haven't been granted anything and we're NOT on a resource list,
+ * then we haven't got a reply yet and don't have a known disposition.
+ * This happens whenever a lock enqueue is the request that triggers
+ * recovery.
+ */
+ if (lock->l_granted_mode == lock->l_req_mode)
+ 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))
+ flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
+ else
+ flags = LDLM_FL_REPLAY;
+
+ 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 = 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]);
+ 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
+ * queue to be processed after all request replayes.
+ * bug 6063 */
+ lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
+
+ LDLM_DEBUG(lock, "replaying lock:");
+
+ 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->lock_handle = body->lock_handle[0];
+ req->rq_interpret_reply = replay_lock_interpret;
+ ptlrpcd_add_req(req);
+
+ RETURN(0);
+}
+
+int ldlm_replay_locks(struct obd_import *imp)
+{
+ struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
+ struct 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);
+
+ /* ensure this doesn't fall to 0 before all have been queued */
+ 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)
+ continue; /* or try to do the rest? */
+ rc = replay_one_lock(imp, lock);
+ }
+
+ atomic_dec(&imp->imp_replay_inflight);
+
+ RETURN(rc);
}