/**
* Register request buffer descriptor for request receiving.
*/
int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
{
struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY};
int rc;
lnet_md_t md;
lnet_handle_me_t me_h;
CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
service->srv_req_portal);
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
return -ENOMEM;
/* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
* which means buffer can only be attached on local CPT, and LND
* threads can find it by grabbing a local lock
*/
rc = LNetMEAttach(service->srv_req_portal,
match_id, 0, ~0, LNET_UNLINK,
rqbd->rqbd_svcpt->scp_cpt >= 0 ?
LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
if (rc != 0) {
CERROR("LNetMEAttach failed: %d\n", rc);
return -ENOMEM;
}
LASSERT(rqbd->rqbd_refcount == 0);
rqbd->rqbd_refcount = 1;
md.start = rqbd->rqbd_buffer;
md.length = service->srv_buf_size;
md.max_size = service->srv_max_req_size;
md.threshold = LNET_MD_THRESH_INF;
md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
md.user_ptr = &rqbd->rqbd_cbid;
md.eq_handle = ptlrpc_eq_h;
rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
if (rc == 0)
return 0;
CERROR("LNetMDAttach failed: %d;\n", rc);
LASSERT(rc == -ENOMEM);
rc = LNetMEUnlink(me_h);
LASSERT(rc == 0);
rqbd->rqbd_refcount = 0;
return -ENOMEM;
}
/**
* Send request \a request.
* if \a noreply is set, don't expect any reply back and don't set up
* reply buffers.
* Returns 0 on success or error code.
*/
int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
{
int rc;
int rc2;
int mpflag = 0;
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;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
return 0;
LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
LASSERT(request->rq_wait_ctx == 0);
/* If this is a re-transmit, we're required to have disengaged
* 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)));
if (unlikely(obd != NULL && obd->obd_fail)) {
CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
obd->obd_name);
/* this prevents us from waiting in ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = -ENODEV;
return -ENODEV;
}
connection = request->rq_import->imp_connection;
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_resend)
lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
if (request->rq_memalloc)
mpflag = cfs_memory_pressure_get_and_set();
rc = sptlrpc_cli_wrap_request(request);
if (rc)
goto out;
/* bulk register should be done after wrap_request() */
if (request->rq_bulk != NULL) {
rc = ptlrpc_register_bulk(request);
if (rc != 0)
goto out;
}
if (!noreply) {
LASSERT(request->rq_replen != 0);
if (request->rq_repbuf == NULL) {
LASSERT(request->rq_repdata == NULL);
LASSERT(request->rq_repmsg == NULL);
rc = sptlrpc_cli_alloc_repbuf(request,
request->rq_replen);
if (rc) {
/* this prevents us from looping in
* ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = rc;
goto cleanup_bulk;
}
} else {
request->rq_repdata = NULL;
request->rq_repmsg = NULL;
}
rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
connection->c_peer, request->rq_xid, 0,
LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
if (rc != 0) {
CERROR("LNetMEAttach failed: %d\n", rc);
LASSERT(rc == -ENOMEM);
rc = -ENOMEM;
goto cleanup_bulk;
}
}
spin_lock(&request->rq_lock);
/* If the MD attach succeeds, there _will_ be a reply_in callback */
request->rq_receiving_reply = !noreply;
request->rq_req_unlink = 1;
/* We are responsible for unlinking the reply buffer */
request->rq_reply_unlink = !noreply;
/* Clear any flags that may be present from previous sends. */
request->rq_replied = 0;
request->rq_err = 0;
request->rq_timedout = 0;
request->rq_net_err = 0;
request->rq_resend = 0;
request->rq_restart = 0;
request->rq_reply_truncate = 0;
spin_unlock(&request->rq_lock);
if (!noreply) {
reply_md.start = request->rq_repbuf;
reply_md.length = request->rq_repbuf_len;
/* Allow multiple early replies */
reply_md.threshold = LNET_MD_THRESH_INF;
/* Manage remote for early replies */
reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
LNET_MD_MANAGE_REMOTE |
LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */
reply_md.user_ptr = &request->rq_reply_cbid;
reply_md.eq_handle = ptlrpc_eq_h;
/* We must see the unlink callback to unset rq_reply_unlink,
so we can't auto-unlink */
rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
&request->rq_reply_md_h);
if (rc != 0) {
CERROR("LNetMDAttach failed: %d\n", rc);
LASSERT(rc == -ENOMEM);
spin_lock(&request->rq_lock);
/* ...but the MD attach didn't succeed... */
request->rq_receiving_reply = 0;
spin_unlock(&request->rq_lock);
rc = -ENOMEM;
goto cleanup_me;
}
CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
request->rq_repbuf_len, request->rq_xid,
request->rq_reply_portal);
}
/* add references on request for request_out_callback */
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));
OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
ktime_get_real_ts64(&request->rq_arrival_time);
request->rq_sent = ktime_get_real_seconds();
/* We give the server rq_timeout secs to process the req, and
add the network latency for our local timeout. */
request->rq_deadline = request->rq_sent + request->rq_timeout +
ptlrpc_at_get_net_latency(request);
ptlrpc_pinger_sending_on_import(request->rq_import);
DEBUG_REQ(D_INFO, request, "send flg=%x",
lustre_msg_get_flags(request->rq_reqmsg));
rc = ptl_send_buf(&request->rq_req_md_h,
request->rq_reqbuf, request->rq_reqdata_len,
LNET_NOACK_REQ, &request->rq_req_cbid,
connection,
request->rq_request_portal,
request->rq_xid, 0);
if (rc == 0)
goto out;
ptlrpc_req_finished(request);
if (noreply)
goto out;
cleanup_me:
/* MEUnlink is safe; the PUT didn't even get off the ground, and
* nobody apart from the PUT's target has the right nid+XID to
* access the reply buffer. */
rc2 = LNetMEUnlink(reply_me_h);
LASSERT(rc2 == 0);
/* UNLINKED callback called synchronously */
LASSERT(!request->rq_receiving_reply);
cleanup_bulk:
/* We do sync unlink here as there was no real transfer here so
* the chance to have long unlink to sluggish net is smaller here. */
ptlrpc_unregister_bulk(request, 0);
out:
if (request->rq_memalloc)
cfs_memory_pressure_restore(mpflag);
return rc;
}
/**
* Register bulk at the sender for later transfer.
* Returns 0 on success or error code.
*/
static int ptlrpc_register_bulk(struct ptlrpc_request *req)
{
struct ptlrpc_bulk_desc *desc = req->rq_bulk;
lnet_process_id_t peer;
int rc = 0;
int rc2;
int posted_md;
int total_md;
__u64 xid;
lnet_handle_me_t me_h;
lnet_md_t md;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
return 0;
/* NB no locking required until desc is on the network */
LASSERT(desc->bd_nob > 0);
LASSERT(desc->bd_md_count == 0);
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);
/* cleanup the state of the bulk for it will be reused */
if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
desc->bd_nob_transferred = 0;
else
LASSERT(desc->bd_nob_transferred == 0);
desc->bd_failure = 0;
peer = desc->bd_import->imp_connection->c_peer;
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);
LASSERTF(!(desc->bd_registered &&
req->rq_send_state != LUSTRE_IMP_REPLAY) ||
xid != desc->bd_last_xid,
"registered: %d rq_xid: %llu bd_last_xid: %llu\n",
desc->bd_registered, xid, desc->bd_last_xid);
total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
desc->bd_registered = 1;
desc->bd_last_xid = xid;
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++) {
md.options = PTLRPC_MD_OPTIONS |
((desc->bd_type == BULK_GET_SOURCE) ?
LNET_MD_OP_GET : LNET_MD_OP_PUT);
ptlrpc_fill_bulk_md(&md, desc, posted_md);
rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
LNET_UNLINK, LNET_INS_AFTER, &me_h);
if (rc != 0) {
CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
desc->bd_import->imp_obd->obd_name, xid,
posted_md, rc);
break;
}
/* About to let the network at it... */
rc = LNetMDAttach(me_h, md, LNET_UNLINK,
&desc->bd_mds[posted_md]);
if (rc != 0) {
CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
desc->bd_import->imp_obd->obd_name, xid,
posted_md, rc);
rc2 = LNetMEUnlink(me_h);
LASSERT(rc2 == 0);
break;
}
}
if (rc != 0) {
LASSERT(rc == -ENOMEM);
spin_lock(&desc->bd_lock);
desc->bd_md_count -= total_md - posted_md;
spin_unlock(&desc->bd_lock);
LASSERT(desc->bd_md_count >= 0);
mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
req->rq_status = -ENOMEM;
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%llu, rq_xid = x%llu\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 */
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),
total_md - desc->bd_md_count);
spin_unlock(&desc->bd_lock);
CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, xid x%#llx-%#llx, portal %u\n",
desc->bd_md_count,
desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
desc->bd_iov_count, desc->bd_nob,
desc->bd_last_xid, req->rq_xid, desc->bd_portal);
return 0;
}
/**
* Send request \a request.
* if \a noreply is set, don't expect any reply back and don't set up
* reply buffers.
* Returns 0 on success or error code.
*/
int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
{
int rc;
int rc2;
int mpflag = 0;
struct ptlrpc_connection *connection;
lnet_handle_me_t reply_me_h;
lnet_md_t reply_md;
struct obd_import *imp = request->rq_import;
struct obd_device *obd = imp->imp_obd;
ENTRY;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
RETURN(0);
LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
LASSERT(request->rq_wait_ctx == 0);
/* If this is a re-transmit, we're required to have disengaged
* cleanly from the previous attempt */
LASSERT(!request->rq_receiving_reply);
LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
(imp->imp_state == LUSTRE_IMP_FULL)));
if (unlikely(obd != NULL && obd->obd_fail)) {
CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
obd->obd_name);
/* this prevents us from waiting in ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = -ENODEV;
RETURN(-ENODEV);
}
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);
request->rq_mbits = request->rq_xid;
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");
} else 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: "LPU64", replied: "LPU64", "
"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 == -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) {
rc = ptlrpc_register_bulk (request);
if (rc != 0)
GOTO(out, rc);
}
if (!noreply) {
LASSERT (request->rq_replen != 0);
if (request->rq_repbuf == NULL) {
LASSERT(request->rq_repdata == NULL);
LASSERT(request->rq_repmsg == NULL);
rc = sptlrpc_cli_alloc_repbuf(request,
request->rq_replen);
if (rc) {
/* this prevents us from looping in
* ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = rc;
GOTO(cleanup_bulk, rc);
}
} else {
request->rq_repdata = NULL;
request->rq_repmsg = NULL;
}
rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
connection->c_peer, request->rq_xid, 0,
LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
if (rc != 0) {
CERROR("LNetMEAttach failed: %d\n", rc);
LASSERT (rc == -ENOMEM);
GOTO(cleanup_bulk, rc = -ENOMEM);
}
}
spin_lock(&request->rq_lock);
/* We are responsible for unlinking the reply buffer */
request->rq_reply_unlinked = noreply;
request->rq_receiving_reply = !noreply;
/* Clear any flags that may be present from previous sends. */
request->rq_req_unlinked = 0;
request->rq_replied = 0;
request->rq_err = 0;
request->rq_timedout = 0;
request->rq_net_err = 0;
request->rq_resend = 0;
request->rq_restart = 0;
request->rq_reply_truncated = 0;
spin_unlock(&request->rq_lock);
if (!noreply) {
reply_md.start = request->rq_repbuf;
reply_md.length = request->rq_repbuf_len;
/* Allow multiple early replies */
reply_md.threshold = LNET_MD_THRESH_INF;
/* Manage remote for early replies */
reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
LNET_MD_MANAGE_REMOTE |
LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
reply_md.user_ptr = &request->rq_reply_cbid;
reply_md.eq_handle = ptlrpc_eq_h;
/* We must see the unlink callback to set rq_reply_unlinked,
* so we can't auto-unlink */
rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
&request->rq_reply_md_h);
if (rc != 0) {
CERROR("LNetMDAttach failed: %d\n", rc);
LASSERT (rc == -ENOMEM);
spin_lock(&request->rq_lock);
/* ...but the MD attach didn't succeed... */
request->rq_receiving_reply = 0;
spin_unlock(&request->rq_lock);
GOTO(cleanup_me, rc = -ENOMEM);
}
CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
", portal %u\n",
request->rq_repbuf_len, request->rq_xid,
request->rq_reply_portal);
}
/* add references on request for request_out_callback */
ptlrpc_request_addref(request);
if (obd != NULL && obd->obd_svc_stats != NULL)
lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
atomic_read(&imp->imp_inflight));
OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
do_gettimeofday(&request->rq_sent_tv);
request->rq_sent = cfs_time_current_sec();
/* We give the server rq_timeout secs to process the req, and
add the network latency for our local timeout. */
request->rq_deadline = request->rq_sent + request->rq_timeout +
ptlrpc_at_get_net_latency(request);
ptlrpc_pinger_sending_on_import(imp);
DEBUG_REQ(D_INFO, request, "send flg=%x",
lustre_msg_get_flags(request->rq_reqmsg));
rc = ptl_send_buf(&request->rq_req_md_h,
request->rq_reqbuf, request->rq_reqdata_len,
LNET_NOACK_REQ, &request->rq_req_cbid,
connection,
request->rq_request_portal,
request->rq_xid, 0);
if (likely(rc == 0))
GOTO(out, rc);
request->rq_req_unlinked = 1;
ptlrpc_req_finished(request);
if (noreply)
GOTO(out, rc);
cleanup_me:
/* MEUnlink is safe; the PUT didn't even get off the ground, and
* nobody apart from the PUT's target has the right nid+XID to
* access the reply buffer. */
rc2 = LNetMEUnlink(reply_me_h);
LASSERT (rc2 == 0);
/* UNLINKED callback called synchronously */
LASSERT(!request->rq_receiving_reply);
cleanup_bulk:
/* We do sync unlink here as there was no real transfer here so
* the chance to have long unlink to sluggish net is smaller here. */
ptlrpc_unregister_bulk(request, 0);
out:
if (request->rq_memalloc)
cfs_memory_pressure_restore(mpflag);
return rc;
}
/**
* Register bulk at the sender for later transfer.
* Returns 0 on success or error code.
*/
int ptlrpc_register_bulk(struct ptlrpc_request *req)
{
struct ptlrpc_bulk_desc *desc = req->rq_bulk;
lnet_process_id_t peer;
int rc = 0;
int rc2;
int posted_md;
int total_md;
__u64 mbits;
lnet_handle_me_t me_h;
lnet_md_t md;
ENTRY;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
RETURN(0);
/* NB no locking required until desc is on the network */
LASSERT(desc->bd_nob > 0);
LASSERT(desc->bd_md_count == 0);
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(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)
desc->bd_nob_transferred = 0;
else
LASSERT(desc->bd_nob_transferred == 0);
desc->bd_failure = 0;
peer = desc->bd_import->imp_connection->c_peer;
LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
LASSERT(desc->bd_cbid.cbid_arg == desc);
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"LPU64", last mbits = x"LPU64"\n",
mbits, req->rq_mbits);
LASSERTF(!(desc->bd_registered &&
req->rq_send_state != LUSTRE_IMP_REPLAY) ||
mbits != desc->bd_last_mbits,
"registered: %d rq_mbits: "LPU64" bd_last_mbits: "LPU64"\n",
desc->bd_registered, mbits, desc->bd_last_mbits);
desc->bd_registered = 1;
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++, mbits++) {
md.options = PTLRPC_MD_OPTIONS |
(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, 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, mbits,
posted_md, rc);
break;
}
/* About to let the network at it... */
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, mbits,
posted_md, rc);
rc2 = LNetMEUnlink(me_h);
LASSERT(rc2 == 0);
break;
}
}
if (rc != 0) {
LASSERT(rc == -ENOMEM);
spin_lock(&desc->bd_lock);
desc->bd_md_count -= total_md - posted_md;
spin_unlock(&desc->bd_lock);
LASSERT(desc->bd_md_count >= 0);
mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
req->rq_status = -ENOMEM;
RETURN(-ENOMEM);
}
spin_lock(&desc->bd_lock);
/* 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),
total_md - desc->bd_md_count);
spin_unlock(&desc->bd_lock);
CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
"mbits x"LPX64"-"LPX64", 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_mbits, req->rq_mbits, desc->bd_portal);
RETURN(0);
}