* Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2013, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
RETURN (-ENOMEM);
}
- CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
+ CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
len, portal, xid, offset);
rc = LNetPut (conn->c_self, *mdh, ack,
/* We're going to get an UNLINK event when I unlink below,
* which will complete just like any other failed send, so
* I fall through and return success here! */
- CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
+ CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
libcfs_id2str(conn->c_peer), portal, xid, rc);
rc2 = LNetMDUnlink(*mdh);
LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
#ifdef HAVE_SERVER_SUPPORT
/**
* Prepare bulk descriptor for specified incoming request \a req that
- * can fit \a npages * pages. \a type is bulk type. \a portal is where
+ * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
* the bulk to be sent. Used on server-side after request was already
* received.
* Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
* error.
*/
struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
- unsigned npages, unsigned max_brw,
- unsigned type, unsigned portal)
+ unsigned nfrags, unsigned max_brw,
+ unsigned int type,
+ unsigned portal,
+ const struct ptlrpc_bulk_frag_ops
+ *ops)
{
struct obd_export *exp = req->rq_export;
struct ptlrpc_bulk_desc *desc;
ENTRY;
- LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
+ LASSERT(ptlrpc_is_bulk_op_active(type));
- desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
+ desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
if (desc == NULL)
RETURN(NULL);
struct obd_export *exp = desc->bd_export;
struct ptlrpc_connection *conn = exp->exp_connection;
int rc = 0;
- __u64 xid;
+ __u64 mbits;
int posted_md;
int total_md;
lnet_md_t md;
/* NB no locking required until desc is on the network */
LASSERT(desc->bd_md_count == 0);
- LASSERT(desc->bd_type == BULK_PUT_SOURCE ||
- desc->bd_type == BULK_GET_SINK);
+ LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
LASSERT(desc->bd_cbid.cbid_arg == desc);
/* NB total length may be 0 for a read past EOF, so we send 0
* length bulks, since the client expects bulk events.
*
- * The client may not need all of the bulk XIDs for the RPC. The RPC
- * used the XID of the highest bulk XID needed, and the server masks
+ * The client may not need all of the bulk mbits for the RPC. The RPC
+ * used the mbits of the highest bulk mbits needed, and the server masks
* off high bits to get bulk count for this RPC. LU-1431 */
- xid = desc->bd_req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
- total_md = desc->bd_req->rq_xid - xid + 1;
+ mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
+ total_md = desc->bd_req->rq_mbits - mbits + 1;
desc->bd_md_count = total_md;
desc->bd_failure = 0;
md.eq_handle = ptlrpc_eq_h;
md.threshold = 2; /* SENT and ACK/REPLY */
- for (posted_md = 0; posted_md < total_md; xid++) {
+ for (posted_md = 0; posted_md < total_md; mbits++) {
md.options = PTLRPC_MD_OPTIONS;
/* NB it's assumed that source and sink buffer frags are
}
break;
}
+
+ /* LU-6441: last md is not sent and desc->bd_md_count == 1 */
+ if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
+ CFS_FAIL_ONCE) &&
+ total_md > 1 && posted_md == total_md - 1) {
+ posted_md++;
+ continue;
+ }
+
/* Network is about to get at the memory */
- if (desc->bd_type == BULK_PUT_SOURCE)
+ if (ptlrpc_is_bulk_put_source(desc->bd_type))
rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
LNET_ACK_REQ, conn->c_peer,
- desc->bd_portal, xid, 0, 0);
+ desc->bd_portal, mbits, 0, 0);
else
rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
- conn->c_peer, desc->bd_portal, xid, 0);
+ conn->c_peer, desc->bd_portal, mbits, 0);
posted_md++;
if (rc != 0) {
- CERROR("%s: failed bulk transfer with %s:%u x"LPU64": "
+ CERROR("%s: failed bulk transfer with %s:%u x%llu: "
"rc = %d\n", exp->exp_obd->obd_name,
libcfs_id2str(conn->c_peer), desc->bd_portal,
- xid, rc);
+ mbits, rc);
break;
}
}
}
CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
- "id %s xid "LPX64"-"LPX64"\n", desc->bd_iov_count,
+ "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
- xid - posted_md, xid - 1);
+ mbits - posted_md, mbits - 1);
RETURN(0);
}
-EXPORT_SYMBOL(ptlrpc_start_bulk_transfer);
/**
* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
* one. If it fails, it must be because completion just happened,
* but we must still l_wait_event() in this case, to give liblustre
* a chance to run server_bulk_callback()*/
- mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_count);
+ mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
for (;;) {
/* Network access will complete in finite time but the HUGE
CWARN("Unexpectedly long timeout: desc %p\n", desc);
}
}
-EXPORT_SYMBOL(ptlrpc_abort_bulk);
#endif /* HAVE_SERVER_SUPPORT */
/**
int rc2;
int posted_md;
int total_md;
- __u64 xid;
+ __u64 mbits;
lnet_handle_me_t me_h;
lnet_md_t md;
ENTRY;
LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
LASSERT(desc->bd_req != NULL);
- LASSERT(desc->bd_type == BULK_PUT_SINK ||
- desc->bd_type == BULK_GET_SOURCE);
+ LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
/* cleanup the state of the bulk for it will be reused */
if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
LASSERT(desc->bd_cbid.cbid_arg == desc);
- /* An XID is only used for a single request from the client.
- * For retried bulk transfers, a new XID will be allocated in
- * in ptlrpc_check_set() if it needs to be resent, so it is not
- * using the same RDMA match bits after an error.
- *
- * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
- * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
- xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
+ total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
+ /* rq_mbits is matchbits of the final bulk */
+ mbits = req->rq_mbits - total_md + 1;
+
+ LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
+ "first mbits = x%llu, last mbits = x%llu\n",
+ mbits, req->rq_mbits);
LASSERTF(!(desc->bd_registered &&
req->rq_send_state != LUSTRE_IMP_REPLAY) ||
- xid != desc->bd_last_xid,
- "registered: %d rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
- desc->bd_registered, xid, desc->bd_last_xid);
+ mbits != desc->bd_last_mbits,
+ "registered: %d rq_mbits: %llu bd_last_mbits: %llu\n",
+ desc->bd_registered, mbits, desc->bd_last_mbits);
- total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
desc->bd_registered = 1;
- desc->bd_last_xid = xid;
+ desc->bd_last_mbits = mbits;
desc->bd_md_count = total_md;
md.user_ptr = &desc->bd_cbid;
md.eq_handle = ptlrpc_eq_h;
md.threshold = 1; /* PUT or GET */
- for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
+ for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
md.options = PTLRPC_MD_OPTIONS |
- ((desc->bd_type == BULK_GET_SOURCE) ?
+ (ptlrpc_is_bulk_op_get(desc->bd_type) ?
LNET_MD_OP_GET : LNET_MD_OP_PUT);
ptlrpc_fill_bulk_md(&md, desc, posted_md);
- rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
+ rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
LNET_UNLINK, LNET_INS_AFTER, &me_h);
if (rc != 0) {
- CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
- desc->bd_import->imp_obd->obd_name, xid,
+ CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
+ desc->bd_import->imp_obd->obd_name, mbits,
posted_md, rc);
break;
}
rc = LNetMDAttach(me_h, md, LNET_UNLINK,
&desc->bd_mds[posted_md]);
if (rc != 0) {
- CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
- desc->bd_import->imp_obd->obd_name, xid,
+ CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
+ desc->bd_import->imp_obd->obd_name, mbits,
posted_md, rc);
rc2 = LNetMEUnlink(me_h);
LASSERT(rc2 == 0);
RETURN(-ENOMEM);
}
- /* Set rq_xid to matchbits of the final bulk so that server can
- * infer the number of bulks that were prepared */
- req->rq_xid = --xid;
- LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
- "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n",
- desc->bd_last_xid, req->rq_xid);
-
spin_lock(&desc->bd_lock);
- /* Holler if peer manages to touch buffers before he knows the xid */
+ /* Holler if peer manages to touch buffers before he knows the mbits */
if (desc->bd_md_count != total_md)
CWARN("%s: Peer %s touched %d buffers while I registered\n",
desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
spin_unlock(&desc->bd_lock);
CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
- "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
- desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
+ "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
+ ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
desc->bd_iov_count, desc->bd_nob,
- desc->bd_last_xid, req->rq_xid, desc->bd_portal);
+ desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
RETURN(0);
}
-EXPORT_SYMBOL(ptlrpc_register_bulk);
/**
* Disconnect a bulk desc from the network. Idempotent. Not
/* Let's setup deadline for reply unlink. */
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
- async && req->rq_bulk_deadline == 0)
+ async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
RETURN(1); /* never registered */
- /* Move to "Unregistering" phase as bulk was not unlinked yet. */
- ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
+ /* Move to "Unregistering" phase as bulk was not unlinked yet. */
+ ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
- /* Do not wait for unlink to finish. */
- if (async)
- RETURN(0);
+ /* Do not wait for unlink to finish. */
+ if (async)
+ RETURN(0);
- for (;;) {
+ for (;;) {
/* The wq argument is ignored by user-space wait_event macros */
wait_queue_head_t *wq = (req->rq_set != NULL) ?
&req->rq_set->set_waitq :
}
RETURN(0);
}
-EXPORT_SYMBOL(ptlrpc_unregister_bulk);
static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
{
/* Report actual service time for client latency calc */
lustre_msg_set_service_time(req->rq_repmsg, service_time);
/* Report service time estimate for future client reqs, but report 0
- * (to be ignored by client) if it's a error reply during recovery.
+ * (to be ignored by client) if it's an error reply during recovery.
* (bz15815) */
if (req->rq_type == PTL_RPC_MSG_ERR &&
- (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
+ (req->rq_export == NULL ||
+ req->rq_export->exp_obd->obd_recovering)) {
lustre_msg_set_timeout(req->rq_repmsg, 0);
- else
- lustre_msg_set_timeout(req->rq_repmsg,
- at_get(&svcpt->scp_at_estimate));
-
- if (req->rq_reqmsg &&
- !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
- CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
- "req_flags=%#x magic=%d:%x/%x len=%d\n",
- flags, lustre_msg_get_flags(req->rq_reqmsg),
- lustre_msg_is_v1(req->rq_reqmsg),
- lustre_msg_get_magic(req->rq_reqmsg),
- lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
- }
+ } else {
+ __u32 timeout;
+
+ if (req->rq_export && req->rq_reqmsg != NULL &&
+ lustre_msg_get_flags(req->rq_reqmsg) &
+ (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))
+ timeout = cfs_time_current_sec() -
+ req->rq_arrival_time.tv_sec +
+ min(at_extra,
+ req->rq_export->exp_obd->
+ obd_recovery_timeout / 4);
+ else
+ timeout = at_get(&svcpt->scp_at_estimate);
+ lustre_msg_set_timeout(req->rq_repmsg, timeout);
+ }
+
+ if (req->rq_reqmsg &&
+ !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
+ CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
+ "req_flags=%#x magic=%x/%x len=%d\n",
+ flags, lustre_msg_get_flags(req->rq_reqmsg),
+ lustre_msg_get_magic(req->rq_reqmsg),
+ lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
+ }
}
/**
* Send request reply from request \a req reply buffer.
* \a flags defines reply types
- * Returns 0 on sucess or error code
+ * Returns 0 on success or error code
*/
int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
{
ptlrpc_connection_put(conn);
return rc;
}
-EXPORT_SYMBOL(ptlrpc_send_reply);
int ptlrpc_reply (struct ptlrpc_request *req)
{
else
return (ptlrpc_send_reply(req, 0));
}
-EXPORT_SYMBOL(ptlrpc_reply);
/**
* For request \a req send an error reply back. Create empty
rc = ptlrpc_send_reply(req, may_be_difficult);
RETURN(rc);
}
-EXPORT_SYMBOL(ptlrpc_send_error);
int ptlrpc_error(struct ptlrpc_request *req)
{
return ptlrpc_send_error(req, 0);
}
-EXPORT_SYMBOL(ptlrpc_error);
/**
* Send request \a request.
struct ptlrpc_connection *connection;
lnet_handle_me_t reply_me_h;
lnet_md_t reply_md;
- struct obd_device *obd = request->rq_import->imp_obd;
+ struct obd_import *imp = request->rq_import;
+ struct obd_device *obd = imp->imp_obd;
ENTRY;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
* cleanly from the previous attempt */
LASSERT(!request->rq_receiving_reply);
LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
- (request->rq_import->imp_state == LUSTRE_IMP_FULL)));
+ (imp->imp_state == LUSTRE_IMP_FULL)));
if (unlikely(obd != NULL && obd->obd_fail)) {
CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
RETURN(-ENODEV);
}
- connection = request->rq_import->imp_connection;
+ connection = imp->imp_connection;
+
+ lustre_msg_set_handle(request->rq_reqmsg,
+ &imp->imp_remote_handle);
+ lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
+ lustre_msg_set_conn_cnt(request->rq_reqmsg,
+ imp->imp_conn_cnt);
+ lustre_msghdr_set_flags(request->rq_reqmsg,
+ imp->imp_msghdr_flags);
+
+ /* If it's the first time to resend the request for EINPROGRESS,
+ * we need to allocate a new XID (see after_reply()), it's different
+ * from the resend for reply timeout. */
+ if (request->rq_nr_resend != 0 &&
+ list_empty(&request->rq_unreplied_list)) {
+ __u64 min_xid = 0;
+ /* resend for EINPROGRESS, allocate new xid to avoid reply
+ * reconstruction */
+ spin_lock(&imp->imp_lock);
+ ptlrpc_assign_next_xid_nolock(request);
+ min_xid = ptlrpc_known_replied_xid(imp);
+ spin_unlock(&imp->imp_lock);
+
+ lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
+ DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
+ "resend on EINPROGRESS");
+ }
- lustre_msg_set_handle(request->rq_reqmsg,
- &request->rq_import->imp_remote_handle);
- lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
- lustre_msg_set_conn_cnt(request->rq_reqmsg,
- request->rq_import->imp_conn_cnt);
- lustre_msghdr_set_flags(request->rq_reqmsg,
- request->rq_import->imp_msghdr_flags);
+ if (request->rq_bulk != NULL) {
+ ptlrpc_set_bulk_mbits(request);
+ lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
+ }
+
+ if (list_empty(&request->rq_unreplied_list) ||
+ request->rq_xid <= imp->imp_known_replied_xid) {
+ DEBUG_REQ(D_ERROR, request, "xid: %llu, replied: %llu, "
+ "list_empty:%d\n", request->rq_xid,
+ imp->imp_known_replied_xid,
+ list_empty(&request->rq_unreplied_list));
+ LBUG();
+ }
+
+ /** For enabled AT all request should have AT_SUPPORT in the
+ * FULL import state when OBD_CONNECT_AT is set */
+ LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
+ (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
+ !(imp->imp_connect_data.ocd_connect_flags &
+ OBD_CONNECT_AT));
if (request->rq_resend) {
lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
-
if (request->rq_resend_cb != NULL)
request->rq_resend_cb(request, &request->rq_async_args);
}
if (request->rq_memalloc)
mpflag = cfs_memory_pressure_get_and_set();
- rc = sptlrpc_cli_wrap_request(request);
- if (rc)
- GOTO(out, rc);
+ rc = sptlrpc_cli_wrap_request(request);
+ if (rc == -ENOMEM)
+ /* set rq_sent so that this request is treated
+ * as a delayed send in the upper layers */
+ request->rq_sent = cfs_time_current_sec();
+ if (rc)
+ GOTO(out, rc);
/* bulk register should be done after wrap_request() */
if (request->rq_bulk != NULL) {
GOTO(cleanup_me, rc = -ENOMEM);
}
- CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
+ CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
", portal %u\n",
request->rq_repbuf_len, request->rq_xid,
request->rq_reply_portal);
ptlrpc_request_addref(request);
if (obd != NULL && obd->obd_svc_stats != NULL)
lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
- atomic_read(&request->rq_import->imp_inflight));
+ atomic_read(&imp->imp_inflight));
OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
request->rq_deadline = request->rq_sent + request->rq_timeout +
ptlrpc_at_get_net_latency(request);
- ptlrpc_pinger_sending_on_import(request->rq_import);
+ ptlrpc_pinger_sending_on_import(imp);
DEBUG_REQ(D_INFO, request, "send flg=%x",
lustre_msg_get_flags(request->rq_reqmsg));
git://git.whamcloud.com / fs / lustre-release.git / blobdiff