*
* Copyright (c) 2002, 2003 Cluster File Systems, Inc.
*
- * This file is part of Lustre, http://www.lustre.org.
+ * This file is part of the Lustre file system, http://www.lustre.org
+ * Lustre is a trademark of Cluster File Systems, Inc.
*
- * Lustre is free 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.
+ * 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.
*
- * 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
- * GNU General Public License for more details.
+ * 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.
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * 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.
*
*/
#include <liblustre.h>
#endif
-#include <linux/obd_support.h>
-#include <linux/obd_class.h>
-#include <linux/lustre_lib.h>
-#include <linux/lustre_ha.h>
-#include <linux/lustre_import.h>
+#include <obd_support.h>
+#include <obd_class.h>
+#include <lustre_lib.h>
+#include <lustre_ha.h>
+#include <lustre_import.h>
#include "ptlrpc_internal.h"
struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
{
struct ptlrpc_connection *c;
- struct ptlrpc_peer peer;
- int err;
+ lnet_nid_t self;
+ lnet_process_id_t peer;
+ int err;
- err = ptlrpc_uuid_to_peer(uuid, &peer);
+ err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
if (err != 0) {
CERROR("cannot find peer %s!\n", uuid->uuid);
return NULL;
}
- c = ptlrpc_get_connection(&peer, uuid);
+ c = ptlrpc_get_connection(peer, self, uuid);
if (c) {
memcpy(c->c_remote_uuid.uuid,
uuid->uuid, sizeof(c->c_remote_uuid.uuid));
void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
struct obd_uuid *uuid)
{
- struct ptlrpc_peer peer;
- int err;
+ lnet_nid_t self;
+ lnet_process_id_t peer;
+ int err;
- err = ptlrpc_uuid_to_peer(uuid, &peer);
+ err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
if (err != 0) {
CERROR("cannot find peer %s!\n", uuid->uuid);
return;
}
- memcpy(&conn->c_peer, &peer, sizeof (peer));
+ conn->c_peer = peer;
+ conn->c_self = self;
return;
}
return NULL;
spin_lock_init(&desc->bd_lock);
- init_waitqueue_head(&desc->bd_waitq);
+ cfs_waitq_init(&desc->bd_waitq);
desc->bd_max_iov = npages;
desc->bd_iov_count = 0;
- desc->bd_md_h = PTL_INVALID_HANDLE;
+ desc->bd_md_h = LNET_INVALID_HANDLE;
desc->bd_portal = portal;
desc->bd_type = type;
-
+
return desc;
}
struct obd_import *imp = req->rq_import;
struct ptlrpc_bulk_desc *desc;
+ ENTRY;
LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
desc = new_bulk(npages, type, portal);
if (desc == NULL)
struct obd_export *exp = req->rq_export;
struct ptlrpc_bulk_desc *desc;
+ ENTRY;
LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
desc = new_bulk(npages, type, portal);
}
void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
- struct page *page, int pageoffset, int len)
+ cfs_page_t *page, int pageoffset, int len)
{
LASSERT(desc->bd_iov_count < desc->bd_max_iov);
LASSERT(page != NULL);
LASSERT(pageoffset >= 0);
LASSERT(len > 0);
- LASSERT(pageoffset + len <= PAGE_SIZE);
+ LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
desc->bd_nob += len;
- pers_bulk_add_page(desc, page, pageoffset, len);
+ ptlrpc_add_bulk_page(desc, page, pageoffset, len);
}
void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
ENTRY;
LASSERT(desc != NULL);
- LASSERT(desc->bd_iov_count != 0x5a5a5a5a); /* not freed already */
+ LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
LASSERT(!desc->bd_network_rw); /* network hands off or */
LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
+
+ sptlrpc_enc_pool_put_pages(desc);
+
if (desc->bd_export)
class_export_put(desc->bd_export);
else
class_import_put(desc->bd_import);
- OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
+ OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
bd_iov[desc->bd_max_iov]));
EXIT;
}
-struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
- int count, int *lengths, char **bufs)
+void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
+{
+ struct list_head *l, *tmp;
+ struct ptlrpc_request *req;
+
+ if (!pool)
+ return;
+
+ list_for_each_safe(l, tmp, &pool->prp_req_list) {
+ req = list_entry(l, struct ptlrpc_request, rq_list);
+ list_del(&req->rq_list);
+ LASSERT(req->rq_reqbuf);
+ LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
+ OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
+ OBD_FREE(req, sizeof(*req));
+ }
+ OBD_FREE(pool, sizeof(*pool));
+}
+
+void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
+{
+ int i;
+ int size = 1;
+
+ while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
+ size <<= 1;
+
+ LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
+ "Trying to change pool size with nonempty pool "
+ "from %d to %d bytes\n", pool->prp_rq_size, size);
+
+ spin_lock(&pool->prp_lock);
+ pool->prp_rq_size = size;
+ for (i = 0; i < num_rq; i++) {
+ struct ptlrpc_request *req;
+ struct lustre_msg *msg;
+
+ spin_unlock(&pool->prp_lock);
+ OBD_ALLOC(req, sizeof(struct ptlrpc_request));
+ if (!req)
+ return;
+ OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
+ if (!msg) {
+ OBD_FREE(req, sizeof(struct ptlrpc_request));
+ return;
+ }
+ req->rq_reqbuf = msg;
+ req->rq_reqbuf_len = size;
+ req->rq_pool = pool;
+ spin_lock(&pool->prp_lock);
+ list_add_tail(&req->rq_list, &pool->prp_req_list);
+ }
+ spin_unlock(&pool->prp_lock);
+ return;
+}
+
+struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
+ void (*populate_pool)(struct ptlrpc_request_pool *, int))
+{
+ struct ptlrpc_request_pool *pool;
+
+ OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
+ if (!pool)
+ return NULL;
+
+ /* Request next power of two for the allocation, because internally
+ kernel would do exactly this */
+
+ spin_lock_init(&pool->prp_lock);
+ CFS_INIT_LIST_HEAD(&pool->prp_req_list);
+ pool->prp_rq_size = msgsize;
+ pool->prp_populate = populate_pool;
+
+ populate_pool(pool, num_rq);
+
+ if (list_empty(&pool->prp_req_list)) {
+ /* have not allocated a single request for the pool */
+ OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
+ pool = NULL;
+ }
+ return pool;
+}
+
+static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
{
struct ptlrpc_request *request;
+ struct lustre_msg *reqbuf;
+
+ if (!pool)
+ return NULL;
+
+ spin_lock(&pool->prp_lock);
+
+ /* See if we have anything in a pool, and bail out if nothing,
+ * in writeout path, where this matters, this is safe to do, because
+ * nothing is lost in this case, and when some in-flight requests
+ * complete, this code will be called again. */
+ if (unlikely(list_empty(&pool->prp_req_list))) {
+ spin_unlock(&pool->prp_lock);
+ return NULL;
+ }
+
+ request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
+ rq_list);
+ list_del(&request->rq_list);
+ spin_unlock(&pool->prp_lock);
+
+ LASSERT(request->rq_reqbuf);
+ LASSERT(request->rq_pool);
+
+ reqbuf = request->rq_reqbuf;
+ memset(request, 0, sizeof(*request));
+ request->rq_reqbuf = reqbuf;
+ request->rq_reqbuf_len = pool->prp_rq_size;
+ request->rq_pool = pool;
+ return request;
+}
+
+static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
+{
+ struct ptlrpc_request_pool *pool = request->rq_pool;
+
+ spin_lock(&pool->prp_lock);
+ LASSERT(list_empty(&request->rq_list));
+ list_add_tail(&request->rq_list, &pool->prp_req_list);
+ spin_unlock(&pool->prp_lock);
+}
+
+struct ptlrpc_request *
+ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
+ int count, int *lengths, char **bufs,
+ struct ptlrpc_request_pool *pool,
+ struct ptlrpc_cli_ctx *ctx)
+{
+ struct ptlrpc_request *request = NULL;
int rc;
ENTRY;
- LASSERT((unsigned long)imp > 0x1000);
+ /* The obd disconnected */
+ if (imp == NULL)
+ return NULL;
+
+ LASSERT(imp != LP_POISON);
+ LASSERT((unsigned long)imp->imp_client > 0x1000);
+ LASSERT(imp->imp_client != LP_POISON);
+
+ if (pool)
+ request = ptlrpc_prep_req_from_pool(pool);
+
+ if (!request)
+ OBD_ALLOC(request, sizeof(*request));
- OBD_ALLOC(request, sizeof(*request));
if (!request) {
CERROR("request allocation out of memory\n");
RETURN(NULL);
}
- rc = lustre_pack_request(request, count, lengths, bufs);
+ request->rq_import = class_import_get(imp);
+
+ if (unlikely(ctx))
+ request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
+ else {
+ rc = sptlrpc_req_get_ctx(request);
+ if (rc)
+ GOTO(out_free, rc);
+ }
+
+ sptlrpc_req_set_flavor(request, opcode);
+
+ rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
+ bufs);
if (rc) {
- CERROR("cannot pack request %d\n", rc);
- OBD_FREE(request, sizeof(*request));
- RETURN(NULL);
+ LASSERT(!request->rq_pool);
+ GOTO(out_ctx, rc);
}
+ 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_import = class_import_get(imp);
+ request->rq_export = NULL;
request->rq_req_cbid.cbid_fn = request_out_callback;
request->rq_req_cbid.cbid_arg = request;
request->rq_reply_cbid.cbid_fn = reply_in_callback;
request->rq_reply_cbid.cbid_arg = request;
-
+
request->rq_phase = RQ_PHASE_NEW;
/* XXX FIXME bug 249 */
request->rq_reply_portal = imp->imp_client->cli_reply_portal;
spin_lock_init(&request->rq_lock);
- INIT_LIST_HEAD(&request->rq_list);
- INIT_LIST_HEAD(&request->rq_replay_list);
- INIT_LIST_HEAD(&request->rq_set_chain);
- init_waitqueue_head(&request->rq_reply_waitq);
+ CFS_INIT_LIST_HEAD(&request->rq_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_waitq_init(&request->rq_reply_waitq);
request->rq_xid = ptlrpc_next_xid();
atomic_set(&request->rq_refcount, 1);
- request->rq_reqmsg->opc = opcode;
- request->rq_reqmsg->flags = 0;
+ lustre_msg_set_opc(request->rq_reqmsg, opcode);
+ lustre_msg_set_flags(request->rq_reqmsg, 0);
RETURN(request);
+out_ctx:
+ sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
+out_free:
+ class_import_put(imp);
+ if (request->rq_pool)
+ __ptlrpc_free_req_to_pool(request);
+ else
+ OBD_FREE(request, sizeof(*request));
+ return NULL;
+}
+
+struct ptlrpc_request *
+ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
+ int *lengths, char **bufs)
+{
+ return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
+ NULL, NULL);
}
struct ptlrpc_request_set *ptlrpc_prep_set(void)
{
struct ptlrpc_request_set *set;
+ ENTRY;
OBD_ALLOC(set, sizeof *set);
if (!set)
RETURN(NULL);
- INIT_LIST_HEAD(&set->set_requests);
- init_waitqueue_head(&set->set_waitq);
+ CFS_INIT_LIST_HEAD(&set->set_requests);
+ cfs_waitq_init(&set->set_waitq);
set->set_remaining = 0;
spin_lock_init(&set->set_new_req_lock);
- INIT_LIST_HEAD(&set->set_new_requests);
+ CFS_INIT_LIST_HEAD(&set->set_new_requests);
RETURN(set);
}
list_add_tail(&req->rq_set_chain, &set->set_requests);
req->rq_set = set;
set->set_remaining++;
+
atomic_inc(&req->rq_import->imp_inflight);
}
void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
struct ptlrpc_request *req)
{
- unsigned long flags;
- spin_lock_irqsave(&set->set_new_req_lock, flags);
+ spin_lock(&set->set_new_req_lock);
/* 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_irqrestore(&set->set_new_req_lock, flags);
+ spin_unlock(&set->set_new_req_lock);
}
/*
*
* The imp->imp_lock must be held.
*/
-static int ptlrpc_import_delay_req(struct obd_import *imp,
+static int ptlrpc_import_delay_req(struct obd_import *imp,
struct ptlrpc_request *req, int *status)
{
int delay = 0;
LASSERT (status != NULL);
*status = 0;
- if (imp->imp_state == LUSTRE_IMP_NEW) {
+ if (req->rq_ctx_init || req->rq_ctx_fini) {
+ /* always allow ctx init/fini rpc go through */
+ } else if (imp->imp_state == LUSTRE_IMP_NEW) {
DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
*status = -EIO;
LBUG();
- }
- else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
+ } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
*status = -EIO;
- }
- /* allow CONNECT even if import is invalid */
- else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
+ } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
imp->imp_state == LUSTRE_IMP_CONNECTING) {
- ;
- }
- /*
- * If the import has been invalidated (such as by an OST failure), the
- * request must fail with -EIO.
- */
- else if (imp->imp_invalid) {
- DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
- *status = -EIO;
- }
- else if (req->rq_import_generation != imp->imp_generation) {
+ /* allow CONNECT even if import is invalid */ ;
+ } else if (imp->imp_invalid) {
+ /* 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
+ * may result in a resend of the request. */
+ if (!imp->imp_deactive)
+ 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)
+ } else if (req->rq_send_state != imp->imp_state) {
+ if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
+ req->rq_no_delay)
*status = -EWOULDBLOCK;
else
delay = 1;
static int ptlrpc_check_reply(struct ptlrpc_request *req)
{
- unsigned long flags;
int rc = 0;
ENTRY;
/* serialise with network callback */
- spin_lock_irqsave (&req->rq_lock, flags);
+ spin_lock(&req->rq_lock);
- if (req->rq_replied) {
- DEBUG_REQ(D_NET, req, "REPLIED:");
+ if (req->rq_replied)
GOTO(out, rc = 1);
+
+ if (req->rq_net_err && !req->rq_timedout) {
+ spin_unlock(&req->rq_lock);
+ rc = ptlrpc_expire_one_request(req);
+ spin_lock(&req->rq_lock);
+ GOTO(out, rc);
}
- if (req->rq_err) {
- DEBUG_REQ(D_ERROR, req, "ABORTED:");
+ if (req->rq_err)
GOTO(out, rc = 1);
- }
- if (req->rq_resend) {
- DEBUG_REQ(D_ERROR, req, "RESEND:");
+ if (req->rq_resend)
GOTO(out, rc = 1);
- }
- if (req->rq_restart) {
- DEBUG_REQ(D_ERROR, req, "RESTART:");
+ if (req->rq_restart)
GOTO(out, rc = 1);
- }
EXIT;
out:
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ spin_unlock(&req->rq_lock);
DEBUG_REQ(D_NET, req, "rc = %d for", rc);
return rc;
}
int err;
ENTRY;
- err = req->rq_repmsg->status;
- if (req->rq_repmsg->type == PTL_RPC_MSG_ERR) {
- DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
- err);
+ err = lustre_msg_get_status(req->rq_repmsg);
+ if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
+ struct obd_import *imp = req->rq_import;
+ __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
+ LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
+ " with %s. The %s operation failed with %d\n",
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ ll_opcode2str(opc), err);
RETURN(err < 0 ? err : -EINVAL);
}
static int after_reply(struct ptlrpc_request *req)
{
- unsigned long flags;
struct obd_import *imp = req->rq_import;
int rc;
ENTRY;
LASSERT(!req->rq_receiving_reply);
+ 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. */
-#if SWAB_PARANOIA
/* Clear reply swab mask; this is a new reply in sender's byte order */
req->rq_rep_swab_mask = 0;
-#endif
- LASSERT (req->rq_nob_received <= req->rq_replen);
- rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
+
+ rc = sptlrpc_cli_unwrap_reply(req);
if (rc) {
- CERROR("unpack_rep failed: %d\n", rc);
+ DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
+ RETURN(rc);
+ }
+
+ /* 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);
}
- if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
- req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
- CERROR("invalid packet type received (type=%u)\n",
- req->rq_repmsg->type);
+ 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);
}
- /* Store transno in reqmsg for replay. */
- req->rq_reqmsg->transno = req->rq_transno = req->rq_repmsg->transno;
+ if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
+ lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
+ DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
+ lustre_msg_get_type(req->rq_repmsg));
+ RETURN(-EPROTO);
+ }
rc = ptlrpc_check_status(req);
+ imp->imp_connect_error = rc;
- /* Either we've been evicted, or the server has failed for
- * some reason. Try to reconnect, and if that fails, punt to the
- * upcall. */
- if (rc == -ENOTCONN) {
- if (req->rq_send_state != LUSTRE_IMP_FULL ||
- imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
- RETURN(-ENOTCONN);
+ if (rc) {
+ /* Either we've been evicted, or the server has failed for
+ * some reason. Try to reconnect, and if that fails, punt to
+ * 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);
+ }
+ ptlrpc_request_handle_notconn(req);
+ 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);
}
-
- ptlrpc_request_handle_notconn(req);
-
- RETURN(rc);
}
+ /* 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_irqsave(&imp->imp_lock, flags);
- if (req->rq_replay || req->rq_transno != 0)
+ 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 >
+ lustre_msg_get_last_committed(req->rq_repmsg) ||
+ req->rq_replay))
ptlrpc_retain_replayable_request(req, imp);
- else if (req->rq_commit_cb != NULL)
+ else if (req->rq_commit_cb != NULL) {
+ spin_unlock(&imp->imp_lock);
req->rq_commit_cb(req);
-
- if (req->rq_transno > imp->imp_max_transno)
- imp->imp_max_transno = req->rq_transno;
+ spin_lock(&imp->imp_lock);
+ }
/* Replay-enabled imports return commit-status information. */
- if (req->rq_repmsg->last_committed)
+ if (lustre_msg_get_last_committed(req->rq_repmsg)) {
imp->imp_peer_committed_transno =
- req->rq_repmsg->last_committed;
+ lustre_msg_get_last_committed(req->rq_repmsg);
+ }
ptlrpc_free_committed(imp);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
}
RETURN(rc);
static int ptlrpc_send_new_req(struct ptlrpc_request *req)
{
struct obd_import *imp;
- unsigned long flags;
int rc;
ENTRY;
LASSERT(req->rq_phase == RQ_PHASE_NEW);
+ if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
+ RETURN (0);
+
req->rq_phase = RQ_PHASE_RPC;
imp = req->rq_import;
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
req->rq_import_generation = imp->imp_generation;
DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
"(%s != %s)",
- req->rq_reqmsg->status,
+ lustre_msg_get_status(req->rq_reqmsg) ,
ptlrpc_import_state_name(req->rq_send_state),
ptlrpc_import_state_name(imp->imp_state));
LASSERT(list_empty (&req->rq_list));
list_add_tail(&req->rq_list, &imp->imp_delayed_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
RETURN(0);
}
if (rc != 0) {
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
req->rq_status = rc;
req->rq_phase = RQ_PHASE_INTERPRET;
RETURN(rc);
/* XXX this is the same as ptlrpc_queue_wait */
LASSERT(list_empty(&req->rq_list));
list_add_tail(&req->rq_list, &imp->imp_sending_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
-
- req->rq_reqmsg->status = current->pid;
- CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
- " %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
- imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
- req->rq_xid,
- imp->imp_connection->c_peer.peer_ni->pni_name,
- imp->imp_connection->c_peer.peer_nid,
- req->rq_reqmsg->opc);
-
- rc = ptl_send_rpc(req);
+ spin_unlock(&imp->imp_lock);
+
+ lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
+
+ rc = sptlrpc_req_refresh_ctx(req, -1);
+ if (rc) {
+ if (req->rq_err) {
+ req->rq_status = rc;
+ RETURN(1);
+ } else {
+ /* here begins timeout counting */
+ req->rq_sent = CURRENT_SECONDS;
+ req->rq_wait_ctx = 1;
+ RETURN(0);
+ }
+ }
+
+ CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
+ " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
+ imp->imp_obd->obd_uuid.uuid,
+ lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ lustre_msg_get_opc(req->rq_reqmsg));
+
+ rc = ptl_send_rpc(req, 0);
if (rc) {
DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
- req->rq_timeout = 1;
+ req->rq_net_err = 1;
RETURN(rc);
}
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)
{
- unsigned long flags;
struct list_head *tmp;
int force_timer_recalc = 0;
ENTRY;
ptlrpc_send_new_req(req)) {
force_timer_recalc = 1;
}
+ /* delayed send - skip */
+ if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
+ continue;
if (!(req->rq_phase == RQ_PHASE_RPC ||
req->rq_phase == RQ_PHASE_BULK ||
if (req->rq_phase == RQ_PHASE_INTERPRET)
GOTO(interpret, req->rq_status);
+ if (req->rq_net_err && !req->rq_timedout)
+ ptlrpc_expire_one_request(req);
+
if (req->rq_err) {
ptlrpc_unregister_reply(req);
+ req->rq_replied = 0;
if (req->rq_status == 0)
req->rq_status = -EIO;
req->rq_phase = RQ_PHASE_INTERPRET;
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
list_del_init(&req->rq_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
GOTO(interpret, req->rq_status);
}
/* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
* will only be set after rq_timedout, but the oig waiting
* path sets rq_intr irrespective of whether ptlrpcd has
- * seen a timeout. our policy is to only interpret
+ * seen a timeout. our policy is to only interpret
* interrupted rpcs after they have timed out */
- if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
+ if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
+ req->rq_wait_ctx)) {
/* NB could be on delayed list */
ptlrpc_unregister_reply(req);
req->rq_status = -EINTR;
req->rq_phase = RQ_PHASE_INTERPRET;
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
list_del_init(&req->rq_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
GOTO(interpret, req->rq_status);
}
if (req->rq_phase == RQ_PHASE_RPC) {
- if (req->rq_waiting || req->rq_resend) {
+ if (req->rq_timedout || req->rq_resend ||
+ req->rq_waiting || req->rq_wait_ctx) {
int status;
- spin_lock_irqsave(&imp->imp_lock, flags);
+ /* 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);
if (ptlrpc_import_delay_req(imp, req, &status)){
- spin_unlock_irqrestore(&imp->imp_lock,
- flags);
+ spin_unlock(&imp->imp_lock);
continue;
- }
+ }
list_del_init(&req->rq_list);
if (status != 0) {
req->rq_status = status;
req->rq_phase = RQ_PHASE_INTERPRET;
- spin_unlock_irqrestore(&imp->imp_lock,
- flags);
+ spin_unlock(&imp->imp_lock);
GOTO(interpret, req->rq_status);
}
- if (req->rq_no_resend) {
+ if (req->rq_no_resend && !req->rq_wait_ctx) {
req->rq_status = -ENOTCONN;
req->rq_phase = RQ_PHASE_INTERPRET;
- spin_unlock_irqrestore(&imp->imp_lock,
- flags);
+ spin_unlock(&imp->imp_lock);
GOTO(interpret, req->rq_status);
}
list_add_tail(&req->rq_list,
&imp->imp_sending_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
req->rq_waiting = 0;
if (req->rq_resend) {
lustre_msg_add_flags(req->rq_reqmsg,
MSG_RESENT);
- ptlrpc_unregister_reply(req);
if (req->rq_bulk) {
__u64 old_xid = req->rq_xid;
old_xid, req->rq_xid);
}
}
+check_ctx:
+ 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;
+ req->rq_wait_ctx = 1;
+ }
+ continue;
+ } else {
+ req->rq_sent = 0;
+ req->rq_wait_ctx = 0;
+ }
- rc = ptl_send_rpc(req);
+ rc = ptl_send_rpc(req, 0);
if (rc) {
DEBUG_REQ(D_HA, req, "send failed (%d)",
rc);
force_timer_recalc = 1;
- req->rq_timeout = 0;
+ req->rq_net_err = 1;
}
/* need to reset the timeout */
force_timer_recalc = 1;
if (!ptlrpc_client_replied(req))
continue;
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
list_del_init(&req->rq_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
req->rq_status = after_reply(req);
if (req->rq_resend) {
/* Add this req to the delayed list so
it can be errored if the import is
evicted after recovery. */
- spin_lock_irqsave (&req->rq_lock, flags);
- list_add_tail(&req->rq_list,
+ spin_lock(&imp->imp_lock);
+ list_add_tail(&req->rq_list,
&imp->imp_delayed_list);
- spin_unlock_irqrestore(&req->rq_lock, flags);
+ spin_unlock(&imp->imp_lock);
continue;
}
if (req->rq_bulk != NULL)
ptlrpc_unregister_bulk (req);
- req->rq_phase = RQ_PHASE_COMPLETE;
-
if (req->rq_interpret_reply != NULL) {
int (*interpreter)(struct ptlrpc_request *,void *,int) =
req->rq_interpret_reply;
req->rq_status = interpreter(req, &req->rq_async_args,
req->rq_status);
}
+ req->rq_phase = RQ_PHASE_COMPLETE;
- CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
- "opc %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
- imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
- req->rq_xid,
- imp->imp_connection->c_peer.peer_ni->pni_name,
- imp->imp_connection->c_peer.peer_nid,
- req->rq_reqmsg->opc);
+ CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
+ "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
+ imp->imp_obd->obd_uuid.uuid,
+ lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ lustre_msg_get_opc(req->rq_reqmsg));
set->set_remaining--;
atomic_dec(&imp->imp_inflight);
- wake_up(&imp->imp_recovery_waitq);
+ cfs_waitq_signal(&imp->imp_recovery_waitq);
}
/* If we hit an error, we want to recover promptly. */
int ptlrpc_expire_one_request(struct ptlrpc_request *req)
{
- unsigned long flags;
struct obd_import *imp = req->rq_import;
+ int rc = 0;
ENTRY;
- DEBUG_REQ(D_ERROR, req, "timeout");
+ DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
+ req->rq_net_err ? "network error" : "timeout",
+ (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
- spin_lock_irqsave (&req->rq_lock, flags);
+ 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;
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ req->rq_wait_ctx = 0;
+ spin_unlock(&req->rq_lock);
ptlrpc_unregister_reply (req);
+ if (obd_dump_on_timeout)
+ libcfs_debug_dumplog();
+
if (req->rq_bulk != NULL)
ptlrpc_unregister_bulk (req);
/* If this request is for recovery or other primordial tasks,
* then error it out here. */
- if (req->rq_send_state != LUSTRE_IMP_FULL ||
+ if (req->rq_ctx_init || req->rq_ctx_fini ||
+ req->rq_send_state != LUSTRE_IMP_FULL ||
imp->imp_obd->obd_no_recov) {
- spin_lock_irqsave (&req->rq_lock, flags);
+ spin_lock(&req->rq_lock);
req->rq_status = -ETIMEDOUT;
req->rq_err = 1;
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ spin_unlock(&req->rq_lock);
RETURN(1);
}
+
+ /* if request can't be resend we can't wait answer after timeout */
+ if (req->rq_no_resend) {
+ DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
+ rc = 1;
+ }
- ptlrpc_fail_import(imp, req->rq_import_generation);
+ ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
- RETURN(0);
+ RETURN(rc);
}
int ptlrpc_expired_set(void *data)
{
struct ptlrpc_request_set *set = data;
struct list_head *tmp;
- time_t now = LTIME_S (CURRENT_TIME);
+ time_t now = CURRENT_SECONDS;
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
- && !req->rq_resend) ||
+ if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
+ !req->rq_resend) ||
(req->rq_phase == RQ_PHASE_BULK)))
continue;
void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
{
- unsigned long flags;
- spin_lock_irqsave(&req->rq_lock, flags);
+ spin_lock(&req->rq_lock);
req->rq_intr = 1;
- spin_unlock_irqrestore(&req->rq_lock, flags);
+ spin_unlock(&req->rq_lock);
}
void ptlrpc_interrupted_set(void *data)
int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
{
struct list_head *tmp;
- time_t now = LTIME_S(CURRENT_TIME);
+ time_t now = CURRENT_SECONDS;
time_t deadline;
int timeout = 0;
struct ptlrpc_request *req;
/* request in-flight? */
if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
- (req->rq_phase == RQ_PHASE_BULK)))
+ (req->rq_phase == RQ_PHASE_BULK) ||
+ (req->rq_phase == RQ_PHASE_NEW)))
continue;
if (req->rq_timedout) /* already timed out */
continue;
- deadline = req->rq_sent + req->rq_timeout;
+ if (req->rq_phase == RQ_PHASE_NEW)
+ deadline = req->rq_sent;
+ else
+ deadline = req->rq_sent + req->rq_timeout;
+
if (deadline <= now) /* actually expired already */
timeout = 1; /* ASAP */
else if (timeout == 0 || timeout > deadline - now)
}
RETURN(timeout);
}
-
int ptlrpc_set_wait(struct ptlrpc_request_set *set)
{
int rc, timeout;
ENTRY;
- LASSERT(!list_empty(&set->set_requests));
+ if (list_empty(&set->set_requests))
+ RETURN(0);
+
list_for_each(tmp, &set->set_requests) {
req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
if (req->rq_phase == RQ_PHASE_NEW)
/* wait until all complete, interrupted, or an in-flight
* req times out */
- CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
+ CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
set, timeout);
- lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
- ptlrpc_expired_set,
+ 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);
* EINTR.
* I don't really care if we go once more round the loop in
* the error cases -eeb. */
- } while (rc != 0);
+ } while (rc != 0 || set->set_remaining != 0);
LASSERT(set->set_remaining == 0);
if (set->set_interpret != NULL) {
int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
set->set_interpret;
- rc = interpreter (set, &set->set_args, rc);
+ rc = interpreter (set, set->set_arg, rc);
}
RETURN(rc);
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);
+ LASSERT(request->rq_cli_ctx);
/* We must take it off the imp_replay_list first. Otherwise, we'll set
* request->rq_reqmsg to NULL while osc_close is dereferencing it. */
if (request->rq_import != NULL) {
- unsigned long flags = 0;
if (!locked)
- spin_lock_irqsave(&request->rq_import->imp_lock, flags);
+ spin_lock(&request->rq_import->imp_lock);
+ list_del_init(&request->rq_mod_list);
list_del_init(&request->rq_replay_list);
if (!locked)
- spin_unlock_irqrestore(&request->rq_import->imp_lock,
- flags);
+ spin_unlock(&request->rq_import->imp_lock);
}
LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
LBUG();
}
- if (request->rq_repmsg != NULL) {
- OBD_FREE(request->rq_repmsg, request->rq_replen);
- request->rq_repmsg = NULL;
- }
- if (request->rq_reqmsg != NULL) {
- OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
- request->rq_reqmsg = NULL;
- }
+ if (request->rq_repbuf != NULL)
+ sptlrpc_cli_free_repbuf(request);
if (request->rq_export != NULL) {
class_export_put(request->rq_export);
request->rq_export = NULL;
if (request->rq_bulk != NULL)
ptlrpc_free_bulk(request->rq_bulk);
- OBD_FREE(request, sizeof(*request));
+ if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
+ sptlrpc_cli_free_reqbuf(request);
+
+ sptlrpc_req_put_ctx(request, !locked);
+
+ if (request->rq_pool)
+ __ptlrpc_free_req_to_pool(request);
+ else
+ OBD_FREE(request, sizeof(*request));
EXIT;
}
if (request == NULL)
RETURN(1);
- if (request == (void *)(unsigned long)(0x5a5a5a5a5a5a5a5a) ||
- request->rq_reqmsg == (void *)(unsigned long)(0x5a5a5a5a5a5a5a5a)) {
+ if (request == LP_POISON ||
+ request->rq_reqmsg == LP_POISON) {
CERROR("dereferencing freed request (bug 575)\n");
LBUG();
RETURN(1);
__ptlrpc_req_finished(request, 0);
}
+__u64 ptlrpc_req_xid(struct ptlrpc_request *request)
+{
+ return request->rq_xid;
+}
+EXPORT_SYMBOL(ptlrpc_req_xid);
+
/* Disengage the client's reply buffer from the network
* NB does _NOT_ unregister any client-side bulk.
* IDEMPOTENT, but _not_ safe against concurrent callers.
void ptlrpc_unregister_reply (struct ptlrpc_request *request)
{
int rc;
- wait_queue_head_t *wq;
+ cfs_waitq_t *wq;
struct l_wait_info lwi;
LASSERT(!in_interrupt ()); /* might sleep */
if (!ptlrpc_client_receiving_reply(request))
return;
- rc = PtlMDUnlink (request->rq_reply_md_h);
- if (rc == PTL_MD_INVALID) {
- LASSERT (!ptlrpc_client_receiving_reply(request));
- return;
- }
-
- LASSERT (rc == PTL_OK);
+ 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() */
- if (request->rq_set == NULL)
+ if (request->rq_set != NULL)
wq = &request->rq_set->set_waitq;
else
wq = &request->rq_reply_waitq;
for (;;) {
/* Network access will complete in finite time but the HUGE
* timeout lets us CWARN for visibility of sluggish NALs */
- lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
+ lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
if (rc == 0)
return;
LASSERT_SPIN_LOCKED(&imp->imp_lock);
- CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
- imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
+
+ if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
+ 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);
+ return;
+ }
+
+ CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
+ imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
+ imp->imp_generation);
+ imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
+ imp->imp_last_generation_checked = imp->imp_generation;
list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
last_req = req;
if (req->rq_import_generation < imp->imp_generation) {
- DEBUG_REQ(D_HA, req, "freeing request with old gen");
+ DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
GOTO(free_req, 0);
}
if (req->rq_replay) {
- DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
+ DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
continue;
}
/* not yet committed */
if (req->rq_transno > imp->imp_peer_committed_transno) {
- DEBUG_REQ(D_HA, req, "stopping search");
+ DEBUG_REQ(D_RPCTRACE, req, "stopping search");
break;
}
- DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
+ DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
imp->imp_peer_committed_transno);
free_req:
+ spin_lock(&req->rq_lock);
+ req->rq_replay = 0;
+ spin_unlock(&req->rq_lock);
if (req->rq_commit_cb != NULL)
req->rq_commit_cb(req);
list_del_init(&req->rq_replay_list);
void ptlrpc_resend_req(struct ptlrpc_request *req)
{
- unsigned long flags;
-
DEBUG_REQ(D_HA, req, "going to resend");
- req->rq_reqmsg->handle.cookie = 0;
+ lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
req->rq_status = -EAGAIN;
- spin_lock_irqsave (&req->rq_lock, flags);
+ spin_lock(&req->rq_lock);
req->rq_resend = 1;
+ req->rq_net_err = 0;
req->rq_timedout = 0;
if (req->rq_bulk) {
__u64 old_xid = req->rq_xid;
-
+
/* ensure previous bulk fails */
req->rq_xid = ptlrpc_next_xid();
CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
old_xid, req->rq_xid);
}
ptlrpc_wake_client_req(req);
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ spin_unlock(&req->rq_lock);
}
/* XXX: this function and rq_status are currently unused */
void ptlrpc_restart_req(struct ptlrpc_request *req)
{
- unsigned long flags;
-
DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
req->rq_status = -ERESTARTSYS;
- spin_lock_irqsave (&req->rq_lock, flags);
+ spin_lock(&req->rq_lock);
req->rq_restart = 1;
req->rq_timedout = 0;
ptlrpc_wake_client_req(req);
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ spin_unlock(&req->rq_lock);
}
static int expired_request(void *data)
struct ptlrpc_request *req = data;
ENTRY;
+ /* some failure can suspend regular timeouts */
+ if (ptlrpc_check_suspend())
+ RETURN(1);
+
RETURN(ptlrpc_expire_one_request(req));
}
static void interrupted_request(void *data)
{
- unsigned long flags;
-
struct ptlrpc_request *req = data;
DEBUG_REQ(D_HA, req, "request interrupted");
- spin_lock_irqsave (&req->rq_lock, flags);
+ spin_lock(&req->rq_lock);
req->rq_intr = 1;
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ spin_unlock(&req->rq_lock);
}
struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
LASSERT_SPIN_LOCKED(&imp->imp_lock);
+ /* clear this for new requests that were resent as well
+ as resent replayed requests. */
+ lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
+
/* don't re-add requests that have been replayed */
if (!list_empty(&req->rq_replay_list))
return;
- lustre_msg_add_flags(req->rq_reqmsg,
- MSG_REPLAY);
+ lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
LASSERT(imp->imp_replayable);
/* Balanced in ptlrpc_free_committed, usually. */
int brc;
struct l_wait_info lwi;
struct obd_import *imp = req->rq_import;
- unsigned long flags;
- int timeout = 0;
+ cfs_duration_t timeout = 0;
ENTRY;
LASSERT(req->rq_set == NULL);
atomic_inc(&imp->imp_inflight);
/* for distributed debugging */
- req->rq_reqmsg->status = current->pid;
+ lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
LASSERT(imp->imp_obd != NULL);
- CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
- "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
+ CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
+ "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
imp->imp_obd->obd_uuid.uuid,
- req->rq_reqmsg->status, req->rq_xid,
- imp->imp_connection->c_peer.peer_ni->pni_name,
- imp->imp_connection->c_peer.peer_nid,
- req->rq_reqmsg->opc);
+ lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ lustre_msg_get_opc(req->rq_reqmsg));
/* Mark phase here for a little debug help */
req->rq_phase = RQ_PHASE_RPC;
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
req->rq_import_generation = imp->imp_generation;
restart:
if (ptlrpc_import_delay_req(imp, req, &rc)) {
list_del(&req->rq_list);
list_add_tail(&req->rq_list, &imp->imp_delayed_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
- current->comm,
- ptlrpc_import_state_name(req->rq_send_state),
+ cfs_curproc_comm(),
+ ptlrpc_import_state_name(req->rq_send_state),
ptlrpc_import_state_name(imp->imp_state));
lwi = LWI_INTR(interrupted_request, req);
rc = l_wait_event(req->rq_reply_waitq,
(req->rq_send_state == imp->imp_state ||
- req->rq_err),
+ req->rq_err || req->rq_intr),
&lwi);
- DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
- current->comm,
- ptlrpc_import_state_name(imp->imp_state),
+ DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
+ cfs_curproc_comm(),
+ ptlrpc_import_state_name(imp->imp_state),
ptlrpc_import_state_name(req->rq_send_state),
- req->rq_err);
+ req->rq_err, req->rq_intr);
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
list_del_init(&req->rq_list);
if (req->rq_err) {
rc = -EIO;
- }
+ }
else if (req->rq_intr) {
rc = -EINTR;
}
else if (req->rq_no_resend) {
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
GOTO(out, rc = -ETIMEDOUT);
}
else {
GOTO(restart, rc);
}
- }
+ }
if (rc != 0) {
list_del_init(&req->rq_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
req->rq_status = rc; // XXX this ok?
GOTO(out, rc);
}
if (req->rq_resend) {
lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
- if (req->rq_bulk != NULL)
+ if (req->rq_bulk != NULL) {
ptlrpc_unregister_bulk (req);
+ /* bulk requests are supposed to be
+ * idempotent, so we are free to bump the xid
+ * here, which we need to do before
+ * registering the bulk again (bug 6371).
+ * print the old xid first for sanity.
+ */
+ DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
+ req->rq_xid = ptlrpc_next_xid();
+ }
+
DEBUG_REQ(D_HA, req, "resending: ");
}
/* XXX this is the same as ptlrpc_set_wait */
LASSERT(list_empty(&req->rq_list));
list_add_tail(&req->rq_list, &imp->imp_sending_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
- rc = ptl_send_rpc(req);
+ rc = sptlrpc_req_refresh_ctx(req, 0);
+ if (rc) {
+ if (req->rq_err) {
+ /* we got fatal ctx refresh error, directly jump out
+ * thus we can pass back the actual error code.
+ */
+ spin_lock(&imp->imp_lock);
+ list_del_init(&req->rq_list);
+ spin_unlock(&imp->imp_lock);
+
+ CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
+ GOTO(out, rc);
+ }
+ /* simulating we got error during send rpc */
+ goto after_send;
+ }
+
+ rc = ptl_send_rpc(req, 0);
if (rc) {
DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
- timeout = 1;
+ timeout = CFS_TICK;
} else {
- timeout = MAX(req->rq_timeout * HZ, 1);
- DEBUG_REQ(D_NET, req, "-- sleeping");
+ timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
+ DEBUG_REQ(D_NET, req,
+ "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
}
+repeat:
lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
req);
- l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
- DEBUG_REQ(D_NET, req, "-- done sleeping");
+ rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
+ if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
+ goto repeat;
- CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
- "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
+after_send:
+ CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
+ "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
imp->imp_obd->obd_uuid.uuid,
- req->rq_reqmsg->status, req->rq_xid,
- imp->imp_connection->c_peer.peer_ni->pni_name,
- imp->imp_connection->c_peer.peer_nid,
- req->rq_reqmsg->opc);
+ lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ lustre_msg_get_opc(req->rq_reqmsg));
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
list_del_init(&req->rq_list);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
/* If the reply was received normally, this just grabs the spinlock
* (ensuring the reply callback has returned), sees that
/* ...unless we were specifically told otherwise. */
if (req->rq_no_resend)
GOTO(out, rc = -ETIMEDOUT);
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
goto restart;
}
GOTO(out, rc = req->rq_status);
}
- rc = after_reply (req);
+ rc = after_reply(req);
/* NB may return +ve success rc */
if (req->rq_resend) {
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
goto restart;
}
out:
if (req->rq_bulk != NULL) {
- if (rc >= 0) {
+ if (rc >= 0) {
/* success so far. Note that anything going wrong
* with bulk now, is EXTREMELY strange, since the
* server must have believed that the bulk
req->rq_phase = RQ_PHASE_INTERPRET;
atomic_dec(&imp->imp_inflight);
- wake_up(&imp->imp_recovery_waitq);
+ cfs_waitq_signal(&imp->imp_recovery_waitq);
RETURN(rc);
}
{
struct ptlrpc_replay_async_args *aa = data;
struct obd_import *imp = req->rq_import;
- unsigned long flags;
+ ENTRY;
atomic_dec(&imp->imp_replay_inflight);
-
+
if (!req->rq_replied) {
CERROR("request replay timed out, restarting recovery\n");
GOTO(out, rc = -ETIMEDOUT);
}
-#if SWAB_PARANOIA
- /* Clear reply swab mask; this is a new reply in sender's byte order */
- req->rq_rep_swab_mask = 0;
-#endif
- LASSERT (req->rq_nob_received <= req->rq_replen);
- rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
- if (rc) {
- CERROR("unpack_rep failed: %d\n", rc);
- GOTO(out, rc = -EPROTO);
- }
-
- if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
- req->rq_repmsg->status == -ENOTCONN)
- GOTO(out, rc = req->rq_repmsg->status);
+ if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
+ (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
+ lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
+ GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
/* The transno had better not change over replay. */
- LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
+ LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
+ lustre_msg_get_transno(req->rq_repmsg));
DEBUG_REQ(D_HA, req, "got rep");
if (req->rq_replay_cb)
req->rq_replay_cb(req);
- if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
+ if (req->rq_replied &&
+ lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
- req->rq_repmsg->status, aa->praa_old_status);
+ lustre_msg_get_status(req->rq_repmsg),
+ aa->praa_old_status);
} else {
/* Put it back for re-replay. */
- req->rq_repmsg->status = aa->praa_old_status;
+ lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
}
- spin_lock_irqsave(&imp->imp_lock, flags);
- imp->imp_last_replay_transno = req->rq_transno;
- spin_unlock_irqrestore(&imp->imp_lock, flags);
-
+ /*
+ * Errors while replay can set transno to 0, but
+ * imp_last_replay_transno shouldn't be set to 0 anyway
+ */
+ if (req->rq_transno > 0) {
+ spin_lock(&imp->imp_lock);
+ LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
+ imp->imp_last_replay_transno = req->rq_transno;
+ spin_unlock(&imp->imp_lock);
+ } else
+ CERROR("Transno is 0 during replay!\n");
/* continue with recovery */
rc = ptlrpc_import_recovery_state_machine(imp);
out:
req->rq_send_state = aa->praa_old_state;
-
+
if (rc != 0)
/* this replay failed, so restart recovery */
ptlrpc_connect_import(imp, NULL);
RETURN(rc);
}
-
int ptlrpc_replay_req(struct ptlrpc_request *req)
{
struct ptlrpc_replay_async_args *aa;
aa->praa_old_state = req->rq_send_state;
req->rq_send_state = LUSTRE_IMP_REPLAY;
req->rq_phase = RQ_PHASE_NEW;
- /*
- * Q: "How can a req get on the replay list if it wasn't replied?"
- * A: "If we failed during the replay of this request, it will still
- * be on the list, but rq_replied will have been reset to 0."
- */
- if (req->rq_replied) {
- aa->praa_old_status = req->rq_repmsg->status;
- req->rq_status = 0;
- req->rq_replied = 0;
- }
+ aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
+ req->rq_status = 0;
req->rq_interpret_reply = ptlrpc_replay_interpret;
atomic_inc(&req->rq_import->imp_replay_inflight);
void ptlrpc_abort_inflight(struct obd_import *imp)
{
- unsigned long flags;
struct list_head *tmp, *n;
ENTRY;
* ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
* this flag and then putting requests on sending_list or delayed_list.
*/
- spin_lock_irqsave(&imp->imp_lock, flags);
+ spin_lock(&imp->imp_lock);
/* XXX locking? Maybe we should remove each request with the list
* locked? Also, how do we know if the requests on the list are
struct ptlrpc_request *req =
list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "inflight");
+ DEBUG_REQ(D_RPCTRACE, req, "inflight");
spin_lock (&req->rq_lock);
if (req->rq_import_generation < imp->imp_generation) {
struct ptlrpc_request *req =
list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "aborting waiting req");
+ DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
spin_lock (&req->rq_lock);
if (req->rq_import_generation < imp->imp_generation) {
}
/* Last chance to free reqs left on the replay list, but we
- * will still leak reqs that haven't comitted. */
+ * will still leak reqs that haven't committed. */
if (imp->imp_replayable)
ptlrpc_free_committed(imp);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock(&imp->imp_lock);
EXIT;
}
static __u64 ptlrpc_last_xid = 0;
-static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
+spinlock_t ptlrpc_last_xid_lock;
__u64 ptlrpc_next_xid(void)
{
return tmp;
}
-
+__u64 ptlrpc_sample_next_xid(void)
+{
+ __u64 tmp;
+ spin_lock(&ptlrpc_last_xid_lock);
+ tmp = ptlrpc_last_xid + 1;
+ spin_unlock(&ptlrpc_last_xid_lock);
+ return tmp;
+}
+EXPORT_SYMBOL(ptlrpc_sample_next_xid);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff