/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
+ * GPL HEADER START
*
- * This file is part of the Lustre file system, http://www.lustre.org
- * Lustre is a trademark of Cluster File Systems, Inc.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
- * You may have signed or agreed to another license before downloading
- * this software. If so, you are bound by the terms and conditions
- * of that agreement, and the following does not apply to you. See the
- * LICENSE file included with this distribution for more information.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
*
- * If you did not agree to a different license, then this copy of Lustre
- * is open source software; you can redistribute it and/or modify it
- * under the terms of version 2 of the GNU General Public License as
- * published by the Free Software Foundation.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
*
- * In either case, Lustre is distributed in the hope that it will be
- * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
- * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * license text for more details.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
*
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
*/
#define DEBUG_SUBSYSTEM S_RPC
return NULL;
}
- c = ptlrpc_get_connection(peer, self, uuid);
+ c = ptlrpc_connection_get(peer, self, uuid);
if (c) {
memcpy(c->c_remote_uuid.uuid,
uuid->uuid, sizeof(c->c_remote_uuid.uuid));
return c;
}
-void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
- struct obd_uuid *uuid)
-{
- lnet_nid_t self;
- lnet_process_id_t peer;
- int err;
-
- err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
- if (err != 0) {
- CERROR("cannot find peer %s!\n", uuid->uuid);
- return;
- }
-
- conn->c_peer = peer;
- conn->c_self = self;
- return;
-}
-
static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
{
struct ptlrpc_bulk_desc *desc;
cfs_waitq_init(&desc->bd_waitq);
desc->bd_max_iov = npages;
desc->bd_iov_count = 0;
- desc->bd_md_h = LNET_INVALID_HANDLE;
+ LNetInvalidateHandle(&desc->bd_md_h);
desc->bd_portal = portal;
desc->bd_type = type;
EXIT;
}
+/* Set server timelimit for this req */
+void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
+{
+ __u32 serv_est;
+ int idx;
+ struct imp_at *at;
+
+ LASSERT(req->rq_import);
+
+ if (AT_OFF) {
+ /* non-AT settings */
+ req->rq_timeout = req->rq_import->imp_server_timeout ?
+ obd_timeout / 2 : obd_timeout;
+ lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
+ return;
+ }
+
+ at = &req->rq_import->imp_at;
+ idx = import_at_get_index(req->rq_import,
+ req->rq_request_portal);
+ serv_est = at_get(&at->iat_service_estimate[idx]);
+ /* add an arbitrary minimum: 125% +5 sec */
+ req->rq_timeout = serv_est + (serv_est >> 2) + 5;
+ /* We could get even fancier here, using history to predict increased
+ loading... */
+
+ /* Let the server know what this RPC timeout is by putting it in the
+ reqmsg*/
+ lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
+}
+
+/* Adjust max service estimate based on server value */
+static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
+ unsigned int serv_est)
+{
+ int idx;
+ unsigned int oldse;
+ struct imp_at *at;
+
+ LASSERT(req->rq_import);
+ at = &req->rq_import->imp_at;
+
+ idx = import_at_get_index(req->rq_import, req->rq_request_portal);
+ /* max service estimates are tracked on the server side,
+ so just keep minimal history here */
+ oldse = at_add(&at->iat_service_estimate[idx], serv_est);
+ if (oldse != 0)
+ CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
+ "has changed from %d to %d\n",
+ req->rq_import->imp_obd->obd_name,req->rq_request_portal,
+ oldse, at_get(&at->iat_service_estimate[idx]));
+}
+
+/* Expected network latency per remote node (secs) */
+int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
+{
+ return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
+}
+
+/* Adjust expected network latency */
+static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
+ unsigned int service_time)
+{
+ unsigned int nl, oldnl;
+ struct imp_at *at;
+ time_t now = cfs_time_current_sec();
+
+ LASSERT(req->rq_import);
+ at = &req->rq_import->imp_at;
+
+ /* Network latency is total time less server processing time */
+ nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
+ if (service_time > now - req->rq_sent + 3 /* bz16408 */)
+ CWARN("Reported service time %u > total measured time "
+ CFS_DURATION_T"\n", service_time,
+ cfs_time_sub(now, req->rq_sent));
+
+ oldnl = at_add(&at->iat_net_latency, nl);
+ if (oldnl != 0)
+ CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
+ "has changed from %d to %d\n",
+ req->rq_import->imp_obd->obd_name,
+ obd_uuid2str(
+ &req->rq_import->imp_connection->c_remote_uuid),
+ oldnl, at_get(&at->iat_net_latency));
+}
+
+static int unpack_reply(struct ptlrpc_request *req)
+{
+ int rc;
+
+ /* Clear reply swab mask; we may have already swabbed an early reply */
+ req->rq_rep_swab_mask = 0;
+
+ rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
+ if (rc) {
+ DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
+ return(-EPROTO);
+ }
+
+ rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
+ if (rc) {
+ DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
+ return(-EPROTO);
+ }
+ return 0;
+}
+
+/*
+ * Handle an early reply message, called with the rq_lock held.
+ * If anything goes wrong just ignore it - same as if it never happened
+ */
+static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
+{
+ struct ptlrpc_request *early_req;
+ time_t olddl;
+ int rc;
+ ENTRY;
+
+ req->rq_early = 0;
+ spin_unlock(&req->rq_lock);
+
+ rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
+ if (rc) {
+ spin_lock(&req->rq_lock);
+ RETURN(rc);
+ }
+
+ rc = unpack_reply(early_req);
+ if (rc == 0) {
+ /* Expecting to increase the service time estimate here */
+ ptlrpc_at_adj_service(req,
+ lustre_msg_get_timeout(early_req->rq_repmsg));
+ ptlrpc_at_adj_net_latency(req,
+ lustre_msg_get_service_time(early_req->rq_repmsg));
+ }
+
+ sptlrpc_cli_finish_early_reply(early_req);
+
+ spin_lock(&req->rq_lock);
+
+ if (rc == 0) {
+ /* Adjust the local timeout for this req */
+ ptlrpc_at_set_req_timeout(req);
+
+ olddl = req->rq_deadline;
+ /* server assumes it now has rq_timeout from when it sent the
+ early reply, so client should give it at least that long. */
+ req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
+ ptlrpc_at_get_net_latency(req);
+
+ DEBUG_REQ(D_ADAPTTO, req,
+ "Early reply #%d, new deadline in "CFS_DURATION_T"s "
+ "("CFS_DURATION_T"s)", req->rq_early_count,
+ cfs_time_sub(req->rq_deadline,
+ cfs_time_current_sec()),
+ cfs_time_sub(req->rq_deadline, olddl));
+ }
+
+ RETURN(rc);
+}
+
void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
{
struct list_head *l, *tmp;
request->rq_reqbuf = reqbuf;
request->rq_reqbuf_len = pool->prp_rq_size;
request->rq_pool = pool;
+
return request;
}
static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
__u32 version, int opcode,
- int count, int *lengths, char **bufs,
+ int count, __u32 *lengths, char **bufs,
struct ptlrpc_cli_ctx *ctx)
{
struct obd_import *imp = request->rq_import;
}
lustre_msg_add_version(request->rq_reqmsg, version);
-
- if (imp->imp_server_timeout)
- request->rq_timeout = obd_timeout / 2;
- else
- request->rq_timeout = obd_timeout;
request->rq_send_state = LUSTRE_IMP_FULL;
request->rq_type = PTL_RPC_MSG_REQUEST;
request->rq_export = NULL;
request->rq_phase = RQ_PHASE_NEW;
- /* XXX FIXME bug 249 */
request->rq_request_portal = imp->imp_client->cli_request_portal;
request->rq_reply_portal = imp->imp_client->cli_reply_portal;
+ ptlrpc_at_set_req_timeout(request);
+
spin_lock_init(&request->rq_lock);
CFS_INIT_LIST_HEAD(&request->rq_list);
+ CFS_INIT_LIST_HEAD(&request->rq_timed_list);
CFS_INIT_LIST_HEAD(&request->rq_replay_list);
CFS_INIT_LIST_HEAD(&request->rq_mod_list);
CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
CFS_INIT_LIST_HEAD(&request->rq_set_chain);
+ CFS_INIT_LIST_HEAD(&request->rq_history_list);
cfs_waitq_init(&request->rq_reply_waitq);
request->rq_xid = ptlrpc_next_xid();
atomic_set(&request->rq_refcount, 1);
lustre_msg_set_opc(request->rq_reqmsg, opcode);
- lustre_msg_set_flags(request->rq_reqmsg, 0);
RETURN(0);
out_ctx:
EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
int ptlrpc_request_pack(struct ptlrpc_request *request,
- __u32 version, int opcode)
+ __u32 version, int opcode)
{
return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
}
OBD_ALLOC_PTR(request);
if (request) {
- LASSERT((unsigned long)imp > 0x1000);
+ LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
LASSERT(imp != LP_POISON);
- LASSERT((unsigned long)imp->imp_client > 0x1000);
+ LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
+ imp->imp_client);
LASSERT(imp->imp_client != LP_POISON);
request->rq_import = class_import_get(imp);
struct ptlrpc_request *
ptlrpc_prep_req_pool(struct obd_import *imp,
__u32 version, int opcode,
- int count, int *lengths, char **bufs,
+ int count, __u32 *lengths, char **bufs,
struct ptlrpc_request_pool *pool)
{
struct ptlrpc_request *request;
struct ptlrpc_request *
ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
- int *lengths, char **bufs)
+ __u32 *lengths, char **bufs)
{
return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
NULL);
spin_lock_init(&set->set_new_req_lock);
CFS_INIT_LIST_HEAD(&set->set_new_requests);
CFS_INIT_LIST_HEAD(&set->set_cblist);
-
+
RETURN(set);
}
if (req->rq_phase == RQ_PHASE_NEW) {
if (req->rq_interpret_reply != NULL) {
- int (*interpreter)(struct ptlrpc_request *,
- void *, int) =
+ ptlrpc_interpterer_t interpreter =
req->rq_interpret_reply;
/* higher level (i.e. LOV) failed;
* let the sub reqs clean up */
req->rq_status = -EBADR;
- interpreter(req, &req->rq_async_args,
+ interpreter(NULL, req, &req->rq_async_args,
req->rq_status);
}
set->set_remaining--;
atomic_inc(&req->rq_import->imp_inflight);
}
-/* lock so many callers can add things, the context that owns the set
- * is supposed to notice these and move them into the set proper. */
-void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
- struct ptlrpc_request *req)
+/**
+ * Lock so many callers can add things, the context that owns the set
+ * is supposed to notice these and move them into the set proper.
+ */
+int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
+ struct ptlrpc_request *req)
{
+ struct ptlrpc_request_set *set = pc->pc_set;
+
+ /*
+ * Let caller know that we stopped and will not handle this request.
+ * It needs to take care itself of request.
+ */
+ if (test_bit(LIOD_STOP, &pc->pc_flags))
+ return -EALREADY;
+
spin_lock(&set->set_new_req_lock);
- /* The set takes over the caller's request reference */
+ /*
+ * The set takes over the caller's request reference.
+ */
list_add_tail(&req->rq_set_chain, &set->set_new_requests);
req->rq_set = set;
spin_unlock(&set->set_new_req_lock);
+
+ /*
+ * Let thead know that we added something and better it to wake up
+ * and process.
+ */
+ cfs_waitq_signal(&set->set_waitq);
+ return 0;
}
/*
DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
*status = -EIO;
} else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
- imp->imp_state == LUSTRE_IMP_CONNECTING) {
+ imp->imp_state == LUSTRE_IMP_CONNECTING) {
/* allow CONNECT even if import is invalid */ ;
- } else if (imp->imp_invalid) {
- /* if it is mgc, wait for recovry. b=13464 */
- if (imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
- delay = 1;
+ if (atomic_read(&imp->imp_inval_count) != 0) {
+ DEBUG_REQ(D_ERROR, req, "invalidate in flight");
+ *status = -EIO;
+ }
+ } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
+ imp->imp_obd->obd_no_recov) {
/* If the import has been invalidated (such as by an OST
- * failure) the request must fail with -ESHUTDOWN. This
- * indicates the requests should be discarded; an -EIO
+ * failure), and if the import(MGC) tried all of its connection
+ * list (Bug 13464), the request must fail with -ESHUTDOWN.
+ * This indicates the requests should be discarded; an -EIO
* may result in a resend of the request. */
if (!imp->imp_deactive)
- DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
+ DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
*status = -ESHUTDOWN; /* bz 12940 */
} else if (req->rq_import_generation != imp->imp_generation) {
DEBUG_REQ(D_ERROR, req, "req wrong generation:");
*status = -EIO;
} else if (req->rq_send_state != imp->imp_state) {
- if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
- req->rq_no_delay)
+ /* invalidate in progress - any requests should be drop */
+ if (atomic_read(&imp->imp_inval_count) != 0) {
+ DEBUG_REQ(D_ERROR, req, "invalidate in flight");
+ *status = -EIO;
+ } else if (imp->imp_dlm_fake || req->rq_no_delay) {
*status = -EWOULDBLOCK;
- else
+ } else {
delay = 1;
+ }
}
RETURN(delay);
if (req->rq_restart)
GOTO(out, rc = 1);
+
+ if (req->rq_early) {
+ ptlrpc_at_recv_early_reply(req);
+ GOTO(out, rc = 0); /* keep waiting */
+ }
+
EXIT;
out:
spin_unlock(&req->rq_lock);
RETURN(err);
}
+/**
+ * Callback function called when client receives RPC reply for \a req.
+ */
static int after_reply(struct ptlrpc_request *req)
{
struct obd_import *imp = req->rq_import;
LASSERT(obd);
LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
- /* NB Until this point, the whole of the incoming message,
- * including buflens, status etc is in the sender's byte order. */
-
- /* Clear reply swab mask; this is a new reply in sender's byte order */
- req->rq_rep_swab_mask = 0;
+ /*
+ * NB Until this point, the whole of the incoming message,
+ * including buflens, status etc is in the sender's byte order.
+ */
rc = sptlrpc_cli_unwrap_reply(req);
if (rc) {
RETURN(rc);
}
- /* security layer unwrap might ask resend this request */
+ /*
+ * Security layer unwrap might ask resend this request.
+ */
if (req->rq_resend)
RETURN(0);
- rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
- if (rc) {
- DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
- RETURN(-EPROTO);
- }
-
- rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
- if (rc) {
- DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
- RETURN(-EPROTO);
- }
+ rc = unpack_reply(req);
+ if (rc)
+ RETURN(rc);
do_gettimeofday(&work_start);
timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
RETURN(-EPROTO);
}
+ OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
+ ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
+ ptlrpc_at_adj_net_latency(req,
+ lustre_msg_get_service_time(req->rq_repmsg));
+
rc = ptlrpc_check_status(req);
imp->imp_connect_error = rc;
if (rc) {
- /* Either we've been evicted, or the server has failed for
+ /*
+ * Either we've been evicted, or the server has failed for
* some reason. Try to reconnect, and if that fails, punt to
- * the upcall. */
+ * the upcall.
+ */
if (ll_rpc_recoverable_error(rc)) {
if (req->rq_send_state != LUSTRE_IMP_FULL ||
imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
RETURN(rc);
}
} else {
- /* Let's look if server send slv. Do it only for RPC with
- * rc == 0. */
- if (imp->imp_obd->obd_namespace) {
- /* Disconnect rpc is sent when namespace is already
- * destroyed. Let's check this and will not try update
- * pool. */
- ldlm_cli_update_pool(req);
- }
+ /*
+ * Let's look if server sent slv. Do it only for RPC with
+ * rc == 0.
+ */
+ ldlm_cli_update_pool(req);
}
- /* Store transno in reqmsg for replay. */
+ /*
+ * Store transno in reqmsg for replay.
+ */
req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
if (req->rq_import->imp_replayable) {
spin_lock(&imp->imp_lock);
- /* no point in adding already-committed requests to the replay
- * list, we will just remove them immediately. b=9829 */
- if (req->rq_transno != 0 &&
- (req->rq_transno >
+ /*
+ * No point in adding already-committed requests to the replay
+ * list, we will just remove them immediately. b=9829
+ */
+ if (req->rq_transno != 0 &&
+ (req->rq_transno >
lustre_msg_get_last_committed(req->rq_repmsg) ||
req->rq_replay))
ptlrpc_retain_replayable_request(req, imp);
spin_lock(&imp->imp_lock);
}
- /* Replay-enabled imports return commit-status information. */
+ /*
+ * Replay-enabled imports return commit-status information.
+ */
if (lustre_msg_get_last_committed(req->rq_repmsg)) {
imp->imp_peer_committed_transno =
lustre_msg_get_last_committed(req->rq_repmsg);
ENTRY;
LASSERT(req->rq_phase == RQ_PHASE_NEW);
- if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
+ if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
RETURN (0);
-
+
req->rq_phase = RQ_PHASE_RPC;
imp = req->rq_import;
req->rq_status = rc;
RETURN(1);
} else {
- /* here begins timeout counting */
- req->rq_sent = CURRENT_SECONDS;
req->rq_wait_ctx = 1;
RETURN(0);
}
}
/* this sends any unsent RPCs in @set and returns TRUE if all are sent */
-int ptlrpc_check_set(struct ptlrpc_request_set *set)
+int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
{
struct list_head *tmp;
int force_timer_recalc = 0;
req->rq_waiting || req->rq_wait_ctx) {
int status;
- /* rq_wait_ctx is only touched in ptlrpcd,
- * no lock needed here.
- */
- if (req->rq_wait_ctx)
- goto check_ctx;
-
ptlrpc_unregister_reply(req);
spin_lock(&imp->imp_lock);
spin_unlock(&imp->imp_lock);
req->rq_waiting = 0;
- if (req->rq_resend) {
+
+ if (req->rq_timedout||req->rq_resend) {
+ /* This is re-sending anyways,
+ * let's mark req as resend. */
+ req->rq_resend = 1;
lustre_msg_add_flags(req->rq_reqmsg,
MSG_RESENT);
if (req->rq_bulk) {
old_xid, req->rq_xid);
}
}
-check_ctx:
+ /*
+ * rq_wait_ctx is only touched by ptlrpcd,
+ * so no lock is needed here.
+ */
status = sptlrpc_req_refresh_ctx(req, -1);
if (status) {
if (req->rq_err) {
req->rq_status = status;
force_timer_recalc = 1;
- }
- if (!req->rq_wait_ctx) {
- /* begins timeout counting */
- req->rq_sent = CURRENT_SECONDS;
+ } else {
req->rq_wait_ctx = 1;
}
+
continue;
} else {
- req->rq_sent = 0;
req->rq_wait_ctx = 0;
}
force_timer_recalc = 1;
}
+ spin_lock(&req->rq_lock);
+
+ if (req->rq_early) {
+ ptlrpc_at_recv_early_reply(req);
+ spin_unlock(&req->rq_lock);
+ continue;
+ }
+
/* Still waiting for a reply? */
- if (ptlrpc_client_receiving_reply(req))
+ if (req->rq_receiving_reply) {
+ spin_unlock(&req->rq_lock);
continue;
+ }
/* Did we actually receive a reply? */
- if (!ptlrpc_client_replied(req))
+ if (!req->rq_replied) {
+ spin_unlock(&req->rq_lock);
continue;
+ }
+
+ spin_unlock(&req->rq_lock);
spin_lock(&imp->imp_lock);
list_del_init(&req->rq_list);
ptlrpc_unregister_bulk (req);
if (req->rq_interpret_reply != NULL) {
- int (*interpreter)(struct ptlrpc_request *,void *,int) =
+ ptlrpc_interpterer_t interpreter =
req->rq_interpret_reply;
- req->rq_status = interpreter(req, &req->rq_async_args,
+ req->rq_status = interpreter(NULL, req,
+ &req->rq_async_args,
req->rq_status);
}
req->rq_phase = RQ_PHASE_COMPLETE;
libcfs_nid2str(imp->imp_connection->c_peer.nid),
lustre_msg_get_opc(req->rq_reqmsg));
- set->set_remaining--;
-
atomic_dec(&imp->imp_inflight);
+ set->set_remaining--;
cfs_waitq_signal(&imp->imp_recovery_waitq);
}
RETURN(set->set_remaining == 0 || force_timer_recalc);
}
+/* Return 1 if we should give up, else 0 */
int ptlrpc_expire_one_request(struct ptlrpc_request *req)
{
struct obd_import *imp = req->rq_import;
int rc = 0;
ENTRY;
- DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
+ DEBUG_REQ(D_ERROR|D_NETERROR, req,
+ "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
req->rq_net_err ? "network error" : "timeout",
- (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
+ req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
+ req->rq_sent));
+
+ if (imp) {
+ LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
+ CFS_DURATION_T"s ago has timed out "
+ "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
+ req->rq_import->imp_obd->obd_name,
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
+ cfs_time_sub(req->rq_deadline, req->rq_sent));
+ }
if (imp != NULL && obd_debug_peer_on_timeout)
LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
spin_lock(&req->rq_lock);
req->rq_timedout = 1;
- req->rq_wait_ctx = 0;
spin_unlock(&req->rq_lock);
ptlrpc_unregister_reply (req);
if (req->rq_ctx_init || req->rq_ctx_fini ||
req->rq_send_state != LUSTRE_IMP_FULL ||
imp->imp_obd->obd_no_recov) {
+ DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
+ ptlrpc_import_state_name(req->rq_send_state),
+ ptlrpc_import_state_name(imp->imp_state));
spin_lock(&req->rq_lock);
req->rq_status = -ETIMEDOUT;
req->rq_err = 1;
spin_unlock(&req->rq_lock);
RETURN(1);
}
-
- /* if request can't be resend we can't wait answer after timeout */
+
+ /* if a request can't be resent we can't wait for an answer after
+ the timeout */
if (req->rq_no_resend) {
DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
rc = 1;
{
struct ptlrpc_request_set *set = data;
struct list_head *tmp;
- time_t now = CURRENT_SECONDS;
+ time_t now = cfs_time_current_sec();
ENTRY;
LASSERT(set != NULL);
list_entry(tmp, struct ptlrpc_request, rq_set_chain);
/* request in-flight? */
- if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
+ if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting &&
!req->rq_resend) ||
(req->rq_phase == RQ_PHASE_BULK)))
continue;
if (req->rq_timedout || /* already dealt with */
- req->rq_sent + req->rq_timeout > now) /* not expired */
+ req->rq_deadline > now) /* not expired */
continue;
/* deal with this guy */
}
}
+/* get the smallest timeout in the set; this does NOT set a timeout. */
int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
{
struct list_head *tmp;
- time_t now = CURRENT_SECONDS;
- time_t deadline;
+ time_t now = cfs_time_current_sec();
int timeout = 0;
struct ptlrpc_request *req;
+ int deadline;
ENTRY;
SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
if (req->rq_timedout) /* already timed out */
continue;
+ if (req->rq_wait_ctx) /* waiting for ctx */
+ continue;
+
if (req->rq_phase == RQ_PHASE_NEW)
deadline = req->rq_sent;
else
lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
ptlrpc_expired_set,
ptlrpc_interrupted_set, set);
- rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
+ rc = l_wait_event(set->set_waitq,
+ ptlrpc_check_set(NULL, set), &lwi);
LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
+ LASSERTF(!request->rq_replay, "req %p\n", request);
LASSERT(request->rq_cli_ctx);
req_capsule_fini(&request->rq_pill);
EXIT;
}
-void ptlrpc_free_req(struct ptlrpc_request *request)
-{
- __ptlrpc_free_req(request, 0);
-}
-
static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
{
struct l_wait_info lwi;
LASSERT(!in_interrupt ()); /* might sleep */
-
- if (!ptlrpc_client_receiving_reply(request))
+ if (!ptlrpc_client_recv_or_unlink(request))
+ /* Nothing left to do */
return;
LNetMDUnlink (request->rq_reply_md_h);
/* We have to l_wait_event() whatever the result, to give liblustre
- * a chance to run reply_in_callback() */
+ * a chance to run reply_in_callback(), and to make sure we've
+ * unlinked before returning a req to the pool */
if (request->rq_set != NULL)
wq = &request->rq_set->set_waitq;
for (;;) {
/* Network access will complete in finite time but the HUGE
* timeout lets us CWARN for visibility of sluggish NALs */
- lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
- rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
+ lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
+ rc = l_wait_event (*wq, !ptlrpc_client_recv_or_unlink(request),
+ &lwi);
if (rc == 0)
return;
LASSERT (rc == -ETIMEDOUT);
- DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
+ DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
+ "rvcng=%d unlnk=%d", request->rq_receiving_reply,
+ request->rq_must_unlink);
}
}
imp->imp_generation == imp->imp_last_generation_checked) {
CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
+ EXIT;
return;
}
if (ptlrpc_check_suspend())
RETURN(1);
+ /* deadline may have changed with an early reply */
+ if (req->rq_deadline > cfs_time_current_sec())
+ RETURN(1);
+
RETURN(ptlrpc_expire_one_request(req));
}
int brc;
struct l_wait_info lwi;
struct obd_import *imp = req->rq_import;
- cfs_duration_t timeout = 0;
+ cfs_duration_t timeout = CFS_TICK;
+ long timeoutl;
ENTRY;
LASSERT(req->rq_set == NULL);
req->rq_phase = RQ_PHASE_RPC;
spin_lock(&imp->imp_lock);
-restart:
req->rq_import_generation = imp->imp_generation;
+restart:
if (ptlrpc_import_delay_req(imp, req, &rc)) {
list_del(&req->rq_list);
list_del_init(&req->rq_list);
if (req->rq_err) {
+ /* rq_status was set locally */
rc = -EIO;
}
else if (req->rq_intr) {
}
rc = ptl_send_rpc(req, 0);
- if (rc) {
+ if (rc)
DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
- timeout = CFS_TICK;
- } else {
- timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
- DEBUG_REQ(D_NET, req,
- "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
- }
+
repeat:
+ timeoutl = req->rq_deadline - cfs_time_current_sec();
+ timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
+ cfs_time_seconds(timeoutl);
+ DEBUG_REQ(D_NET, req,
+ "-- sleeping for "CFS_DURATION_T" ticks", timeout);
lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
req);
- rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
- if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
+ brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
+ if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
+ ptlrpc_check_and_wait_suspend(req)))
goto repeat;
after_send:
* req->rq_receiving_reply is clear and returns. */
ptlrpc_unregister_reply (req);
- if (req->rq_err)
- GOTO(out, rc = -EIO);
- /* Resend if we need to, unless we were interrupted. */
- if (req->rq_resend && !req->rq_intr) {
- /* ...unless we were specifically told otherwise. */
- if (req->rq_no_resend)
- GOTO(out, rc = -ETIMEDOUT);
- spin_lock(&imp->imp_lock);
- goto restart;
+ if (req->rq_err) {
+ DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
+ rc, req->rq_status);
+ GOTO(out, rc = rc ? rc : -EIO);
}
if (req->rq_intr) {
GOTO(out, rc = -EINTR);
}
+ /* Resend if we need to */
+ if (req->rq_resend) {
+ /* ...unless we were specifically told otherwise. */
+ if (req->rq_no_resend)
+ GOTO(out, rc = -ETIMEDOUT);
+ spin_lock(&imp->imp_lock);
+ goto restart;
+ }
+
if (req->rq_timedout) { /* non-recoverable timeout */
GOTO(out, rc = -ETIMEDOUT);
}
int praa_old_status;
};
-static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
+static int ptlrpc_replay_interpret(const struct lu_env *env,
+ struct ptlrpc_request *req,
void * data, int rc)
{
struct ptlrpc_replay_async_args *aa = data;
ENTRY;
LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
-
/* Not handling automatic bulk replay yet (or ever?) */
LASSERT(req->rq_bulk == NULL);
- DEBUG_REQ(D_HA, req, "REPLAY");
-
LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
- aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
+ aa = ptlrpc_req_async_args(req);
memset(aa, 0, sizeof *aa);
/* Prepare request to be resent with ptlrpcd */
aa->praa_old_state = req->rq_send_state;
req->rq_send_state = LUSTRE_IMP_REPLAY;
req->rq_phase = RQ_PHASE_NEW;
- aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
+ if (req->rq_repmsg)
+ aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
req->rq_status = 0;
-
req->rq_interpret_reply = ptlrpc_replay_interpret;
+ /* Readjust the timeout for current conditions */
+ ptlrpc_at_set_req_timeout(req);
+
+ DEBUG_REQ(D_HA, req, "REPLAY");
+
atomic_inc(&req->rq_import->imp_replay_inflight);
ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
spin_lock (&req->rq_lock);
if (req->rq_import_generation < imp->imp_generation) {
req->rq_err = 1;
+ req->rq_status = -EINTR;
ptlrpc_wake_client_req(req);
}
spin_unlock (&req->rq_lock);
spin_lock (&req->rq_lock);
if (req->rq_import_generation < imp->imp_generation) {
req->rq_err = 1;
+ req->rq_status = -EINTR;
ptlrpc_wake_client_req(req);
}
spin_unlock (&req->rq_lock);
EXIT;
}
-static __u64 ptlrpc_last_xid = 0;
-spinlock_t ptlrpc_last_xid_lock;
+void ptlrpc_abort_set(struct ptlrpc_request_set *set)
+{
+ struct list_head *tmp, *n;
+
+ LASSERT(set != NULL);
+
+ list_for_each_safe(tmp, n, &set->set_requests) {
+ struct ptlrpc_request *req =
+ list_entry(tmp, struct ptlrpc_request, rq_set_chain);
+
+ spin_lock (&req->rq_lock);
+ if (req->rq_phase != RQ_PHASE_RPC) {
+ spin_unlock (&req->rq_lock);
+ continue;
+ }
+
+ req->rq_err = 1;
+ req->rq_status = -EINTR;
+ ptlrpc_wake_client_req(req);
+ spin_unlock (&req->rq_lock);
+ }
+}
+
+static __u64 ptlrpc_last_xid;
+static spinlock_t ptlrpc_last_xid_lock;
+
+/* Initialize the XID for the node. This is common among all requests on
+ * this node, and only requires the property that it is monotonically
+ * increasing. It does not need to be sequential. Since this is also used
+ * as the RDMA match bits, it is important that a single client NOT have
+ * the same match bits for two different in-flight requests, hence we do
+ * NOT want to have an XID per target or similar.
+ *
+ * To avoid an unlikely collision between match bits after a client reboot
+ * (which would cause old to be delivered into the wrong buffer) we initialize
+ * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
+ * If the time is clearly incorrect, we instead use a 62-bit random number.
+ * In the worst case the random number will overflow 1M RPCs per second in
+ * 9133 years, or permutations thereof.
+ */
+#define YEAR_2004 (1ULL << 30)
+void ptlrpc_init_xid(void)
+{
+ time_t now = cfs_time_current_sec();
+
+ spin_lock_init(&ptlrpc_last_xid_lock);
+ if (now < YEAR_2004) {
+ ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
+ ptlrpc_last_xid >>= 2;
+ ptlrpc_last_xid |= (1ULL << 61);
+ } else {
+ ptlrpc_last_xid = (now << 20);
+ }
+}
__u64 ptlrpc_next_xid(void)
{
__u64 ptlrpc_sample_next_xid(void)
{
+#if BITS_PER_LONG == 32
+ /* need to avoid possible word tearing on 32-bit systems */
__u64 tmp;
spin_lock(&ptlrpc_last_xid_lock);
tmp = ptlrpc_last_xid + 1;
spin_unlock(&ptlrpc_last_xid_lock);
return tmp;
+#else
+ /* No need to lock, since returned value is racy anyways */
+ return ptlrpc_last_xid + 1;
+#endif
}
EXPORT_SYMBOL(ptlrpc_sample_next_xid);