Whamcloud - gitweb
* add a bit in ptlrpc_request to indicate ptlrpcs error, gss recovery
[fs/lustre-release.git] / lustre / ptlrpc / niobuf.c
index 3b1d32f..7cdb908 100644 (file)
 #define DEBUG_SUBSYSTEM S_RPC
 #ifndef __KERNEL__
 #include <liblustre.h>
-#include <portals/lib-types.h>
 #endif
+#include <linux/obd_class.h>
 #include <linux/obd_support.h>
 #include <linux/lustre_net.h>
 #include <linux/lustre_lib.h>
 #include <linux/obd.h>
+#include <linux/lustre_sec.h>
+#include "ptlrpc_internal.h"
 
-static int ptl_send_buf(struct ptlrpc_request *request,
-                        struct ptlrpc_connection *conn, int portal)
+static int ptl_send_buf (ptl_handle_md_t *mdh, void *base, int len, 
+                         ptl_ack_req_t ack, struct ptlrpc_cb_id *cbid,
+                         struct ptlrpc_connection *conn, int portal, __u64 xid)
 {
-        int rc;
-        ptl_process_id_t remote_id;
-        ptl_handle_md_t md_h;
-        ptl_ack_req_t ack_req;
+        int              rc;
+        ptl_md_t         md;
+        char str[PTL_NALFMT_SIZE];
+        ENTRY;
 
-        LASSERT(conn);
-        CDEBUG (D_INFO, "conn=%p ni %s nid "LPX64" on %s\n", 
+        LASSERT (portal != 0);
+        LASSERT (conn != NULL);
+        CDEBUG (D_INFO, "conn=%p ni %s id %s on %s\n",
                 conn, conn->c_peer.peer_ni->pni_name,
-                conn->c_peer.peer_nid, conn->c_peer.peer_ni->pni_name);
-
-        request->rq_req_md.user_ptr = request;
-
-        switch (request->rq_type) {
-        case PTL_RPC_MSG_REQUEST:
-                request->rq_reqmsg->type = HTON__u32(request->rq_type);
-                request->rq_req_md.start = request->rq_reqmsg;
-                request->rq_req_md.length = request->rq_reqlen;
-                request->rq_req_md.eventq = conn->c_peer.peer_ni->pni_request_out_eq_h;
-                break;
-        case PTL_RPC_MSG_ERR:
-        case PTL_RPC_MSG_REPLY:
-                request->rq_repmsg->type = HTON__u32(request->rq_type);
-                request->rq_req_md.start = request->rq_repmsg;
-                request->rq_req_md.length = request->rq_replen;
-                request->rq_req_md.eventq = conn->c_peer.peer_ni->pni_reply_out_eq_h;
-                break;
-        default:
-                LBUG();
-                return -1; /* notreached */
-        }
-        if (request->rq_flags & PTL_RPC_FL_WANT_ACK) {
-                request->rq_req_md.threshold = 2; /* SENT and ACK */
-                ack_req = PTL_ACK_REQ;
-        } else {
-                request->rq_req_md.threshold = 1;
-                ack_req = PTL_NOACK_REQ;
-        }
-        request->rq_req_md.options = PTL_MD_OP_PUT;
-        request->rq_req_md.user_ptr = request;
-
-        if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
-                request->rq_req_md.options |= PTL_MD_ACK_DISABLE;
+                ptlrpc_id2str(&conn->c_peer, str),
+                conn->c_peer.peer_ni->pni_name);
+        md.start     = base;
+        md.length    = len;
+        md.threshold = (ack == PTL_ACK_REQ) ? 2 : 1;
+        md.options   = PTLRPC_MD_OPTIONS;
+        md.user_ptr  = cbid;
+        md.eq_handle    = conn->c_peer.peer_ni->pni_eq_h;
+
+        if (ack == PTL_ACK_REQ &&
+            OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
+                /* don't ask for the ack to simulate failing client */
+                ack = PTL_NOACK_REQ;
                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
         }
 
-        rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h, request->rq_req_md, &md_h);
-        if (rc != 0) {
-                CERROR("PtlMDBind failed: %d\n", rc);
-                LBUG();
-                return rc;
+        rc = PtlMDBind (conn->c_peer.peer_ni->pni_ni_h, md, 
+                        PTL_UNLINK, mdh);
+        if (rc != PTL_OK) {
+                CERROR ("PtlMDBind failed: %d\n", rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                RETURN (-ENOMEM);
         }
 
-        remote_id.nid = conn->c_peer.peer_nid;
-        remote_id.pid = 0;
-
         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
-               request->rq_req_md.length, portal, request->rq_xid);
+               len, portal, xid);
 
-        if (!portal)
-                LBUG();
-        rc = PtlPut(md_h, ack_req, remote_id, portal, 0, request->rq_xid, 0, 0);
+        rc = PtlPut (*mdh, ack, conn->c_peer.peer_id, portal, 0, xid, 0, 0);
         if (rc != PTL_OK) {
-                CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n",
-                       remote_id.nid, portal, request->rq_xid, rc);
-                PtlMDUnlink(md_h);
+                int rc2;   
+                /* 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("PtlPut(%s, %d, "LPD64") failed: %d\n",
+                       ptlrpc_id2str(&conn->c_peer, str),
+                       portal, xid, rc);
+                rc2 = PtlMDUnlink(*mdh);
+                LASSERTF(rc2 == PTL_OK, "rc2 = %d\n", rc2);
         }
 
-        return rc;
-}
-
-static inline struct iovec *
-ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
-{
-        struct iovec *iov;
-
-        if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
-                return (desc->bd_iov);
-
-        OBD_ALLOC (iov, desc->bd_page_count * sizeof (struct iovec));
-        if (iov == NULL)
-                LBUG();
-
-        return (iov);
+        RETURN (0);
 }
 
-static inline void
-ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, struct iovec *iov)
+int ptlrpc_start_bulk_transfer (struct ptlrpc_bulk_desc *desc)
 {
-        if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
-                return;
-
-        OBD_FREE (iov, desc->bd_page_count * sizeof (struct iovec));
-}
-
-int ptlrpc_bulk_put(struct ptlrpc_bulk_desc *desc)
-{
-        int rc;
+        int                 rc;
+        int                 rc2;
         struct ptlrpc_peer *peer;
-        struct list_head *tmp, *next;
-        ptl_process_id_t remote_id;
-        __u32 xid = 0;
-        struct iovec *iov;
+        ptl_md_t            md;
+        __u64               xid;
+        char                str[PTL_NALFMT_SIZE];
         ENTRY;
 
-        iov = ptlrpc_get_bulk_iov (desc);
-        if (iov == NULL)
-                RETURN (-ENOMEM);
+        if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_PUT_NET)) 
+                RETURN(0);
 
-        peer = &desc->bd_connection->c_peer;
-
-        desc->bd_md.start = iov;
-        desc->bd_md.niov = 0;
-        desc->bd_md.length = 0;
-        desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_source_eq_h;
-        desc->bd_md.threshold = 2; /* SENT and ACK */
-        desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
-        desc->bd_md.user_ptr = desc;
-
-        atomic_set(&desc->bd_source_callback_count, 2);
-
-        list_for_each_safe(tmp, next, &desc->bd_page_list) {
-                struct ptlrpc_bulk_page *bulk;
-                bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
-
-                LASSERT(desc->bd_md.niov < desc->bd_page_count);
-
-                if (desc->bd_md.niov == 0)
-                        xid = bulk->bp_xid;
-                LASSERT(xid == bulk->bp_xid);   /* should all be the same */
-
-                iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
-                iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
-                if (iov[desc->bd_md.niov].iov_len <= 0) {
-                        CERROR("bad bp_buflen[%d] @ %p: %d\n", desc->bd_md.niov,
-                               bulk->bp_buf, bulk->bp_buflen);
-                        CERROR("desc: xid %u, pages %d, ptl %d, ref %d\n",
-                               xid, desc->bd_page_count, desc->bd_portal,
-                               atomic_read(&desc->bd_refcount));
-                        LBUG();
-                }
-                desc->bd_md.niov++;
-                desc->bd_md.length += bulk->bp_buflen;
-        }
+        /* NB no locking required until desc is on the network */
+        LASSERT (!desc->bd_network_rw);
+        LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
+                 desc->bd_type == BULK_GET_SINK);
+        desc->bd_success = 0;
+        peer = &desc->bd_export->exp_connection->c_peer;
 
-        LASSERT(desc->bd_md.niov == desc->bd_page_count);
-        LASSERT(desc->bd_md.niov != 0);
+        md.user_ptr = &desc->bd_cbid;
+        md.eq_handle = peer->peer_ni->pni_eq_h;
+        md.threshold = 2; /* SENT and ACK/REPLY */
+        md.options = PTLRPC_MD_OPTIONS;
+        ptlrpc_fill_bulk_md(&md, desc);
 
-        rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md,
-                       &desc->bd_md_h);
+        LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
+        LASSERT (desc->bd_cbid.cbid_arg == desc);
 
-        ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/
+        /* NB total length may be 0 for a read past EOF, so we send a 0
+         * length bulk, since the client expects a bulk event. */
 
+        rc = PtlMDBind(peer->peer_ni->pni_ni_h, md,
+                       PTL_UNLINK, &desc->bd_md_h);
         if (rc != PTL_OK) {
                 CERROR("PtlMDBind failed: %d\n", rc);
-                LBUG();
-                RETURN(rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                RETURN(-ENOMEM);
         }
 
-        remote_id.nid = peer->peer_nid;
-        remote_id.pid = 0;
+        /* Client's bulk and reply matchbits are the same */
+        xid = desc->bd_req->rq_xid;
+        CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d on %s "
+               "nid %s pid %d xid "LPX64"\n", desc->bd_iov_count,
+               desc->bd_nob, desc->bd_portal, peer->peer_ni->pni_name,
+               ptlrpc_id2str(peer, str), peer->peer_id.pid, xid);
+
+        /* Network is about to get at the memory */
+        desc->bd_network_rw = 1;
 
-        CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d on %s "
-               "nid "LPX64" pid %d xid %d\n", 
-               desc->bd_md.niov, desc->bd_md.length,
-               desc->bd_portal, peer->peer_ni->pni_name,
-               remote_id.nid, remote_id.pid, xid);
+        if (desc->bd_type == BULK_PUT_SOURCE)
+                rc = PtlPut (desc->bd_md_h, PTL_ACK_REQ, peer->peer_id,
+                             desc->bd_portal, 0, xid, 0, 0);
+        else
+                rc = PtlGet (desc->bd_md_h, peer->peer_id,
+                             desc->bd_portal, 0, xid, 0);
 
-        rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
-                    desc->bd_portal, 0, xid, 0, 0);
         if (rc != PTL_OK) {
-                CERROR("PtlPut("LPU64", %d, %d) failed: %d\n",
-                       remote_id.nid, desc->bd_portal, xid, rc);
-                PtlMDUnlink(desc->bd_md_h);
-                LBUG();
-                RETURN(rc);
+                /* Can't send, so we unlink the MD bound above.  The UNLINK
+                 * event this creates will signal completion with failure,
+                 * so we return SUCCESS here! */
+                CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
+                       ptlrpc_id2str(peer, str),
+                       desc->bd_portal, xid, rc);
+                rc2 = PtlMDUnlink(desc->bd_md_h);
+                LASSERT (rc2 == PTL_OK);
         }
 
         RETURN(0);
 }
 
-int ptlrpc_bulk_get(struct ptlrpc_bulk_desc *desc)
+void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
 {
-        int rc;
-        struct ptlrpc_peer *peer;
-        struct list_head *tmp, *next;
-        ptl_process_id_t remote_id;
-        __u32 xid = 0;
-        struct iovec *iov;
-        ENTRY;
-
-        iov = ptlrpc_get_bulk_iov (desc);
-        if (iov == NULL)
-                RETURN (-ENOMEM);
-
-        peer = &desc->bd_connection->c_peer;
-
-        desc->bd_md.start = iov;
-        desc->bd_md.niov = 0;
-        desc->bd_md.length = 0;
-        desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_sink_eq_h;
-        desc->bd_md.threshold = 2; /* SENT and REPLY */
-        desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_IOV;
-        desc->bd_md.user_ptr = desc;
-
-        atomic_set(&desc->bd_source_callback_count, 2);
-
-        list_for_each_safe(tmp, next, &desc->bd_page_list) {
-                struct ptlrpc_bulk_page *bulk;
-                bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
-
-                LASSERT(desc->bd_md.niov < desc->bd_page_count);
-
-                if (desc->bd_md.niov == 0)
-                        xid = bulk->bp_xid;
-                LASSERT(xid == bulk->bp_xid);   /* should all be the same */
-
-                iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
-                iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
-                if (iov[desc->bd_md.niov].iov_len <= 0) {
-                        CERROR("bad bulk %p bp_buflen[%d] @ %p: %d\n", bulk,
-                               desc->bd_md.niov, bulk->bp_buf, bulk->bp_buflen);
-                        CERROR("desc %p: xid %u, pages %d, ptl %d, ref %d\n",
-                               desc, xid, desc->bd_page_count, desc->bd_portal,
-                               atomic_read(&desc->bd_refcount));
-                        LBUG();
-                }
-                desc->bd_md.niov++;
-                desc->bd_md.length += bulk->bp_buflen;
-        }
-
-        LASSERT(desc->bd_md.niov == desc->bd_page_count);
-        LASSERT(desc->bd_md.niov != 0);
-
-        rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md,
-                       &desc->bd_md_h);
+        /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
+         * serialises with completion callback) */
+        struct l_wait_info lwi;
+        int                rc;
 
-        ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/
+        LASSERT (!in_interrupt ());             /* might sleep */
 
-        if (rc != PTL_OK) {
-                CERROR("PtlMDBind failed: %d\n", rc);
-                LBUG();
-                RETURN(rc);
-        }
-
-        remote_id.nid = desc->bd_connection->c_peer.peer_nid;
-        remote_id.pid = 0;
-
-        CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d on %s "
-               "nid "LPX64" pid %d xid %d\n", 
-               desc->bd_md.niov, desc->bd_md.length,
-               desc->bd_portal, peer->peer_ni->pni_name,
-               remote_id.nid, remote_id.pid, xid);
-
-        rc = PtlGet(desc->bd_md_h, remote_id, desc->bd_portal, 0, xid, 0);
-        if (rc != PTL_OK) {
-                CERROR("PtlGet("LPU64", %d, %d) failed: %d\n",
-                       remote_id.nid, desc->bd_portal, xid, rc);
-                PtlMDUnlink(desc->bd_md_h);
-                LBUG();
-                RETURN(rc);
+        if (!ptlrpc_bulk_active(desc))          /* completed or */
+                return;                         /* never started */
+        
+        /* The unlink ensures the callback happens ASAP and is the last
+         * 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()*/
+
+        PtlMDUnlink (desc->bd_md_h);
+
+        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);
+                rc = l_wait_event(desc->bd_waitq, 
+                                  !ptlrpc_bulk_active(desc), &lwi);
+                if (rc == 0)
+                        return;
+
+                LASSERT(rc == -ETIMEDOUT);
+                CWARN("Unexpectedly long timeout: desc %p\n", desc);
         }
-
-        RETURN(0);
 }
 
-static int ptlrpc_register_bulk_shared(struct ptlrpc_bulk_desc *desc)
+int ptlrpc_register_bulk (struct ptlrpc_request *req)
 {
+        struct ptlrpc_bulk_desc *desc = req->rq_bulk;
         struct ptlrpc_peer *peer;
-        struct list_head *tmp, *next;
         int rc;
-        __u32 xid = 0;
-        struct iovec *iov;
-        ptl_process_id_t source_id;
+        int rc2;
+        ptl_handle_me_t  me_h;
+        ptl_md_t         md;
         ENTRY;
 
-        if (desc->bd_page_count > PTL_MD_MAX_IOV) {
-                CERROR("iov longer than %d pages not supported (count=%d)\n",
-                       PTL_MD_MAX_IOV, desc->bd_page_count);
-                RETURN(-EINVAL);
-        }
-
-        iov = ptlrpc_get_bulk_iov (desc);
-        if (iov == NULL)
-                return (-ENOMEM);
-
-        peer = &desc->bd_connection->c_peer;
-        
-        desc->bd_md.start = iov;
-        desc->bd_md.niov = 0;
-        desc->bd_md.length = 0;
-        desc->bd_md.threshold = 1;
-        desc->bd_md.user_ptr = desc;
-
-        list_for_each_safe(tmp, next, &desc->bd_page_list) {
-                struct ptlrpc_bulk_page *bulk;
-                bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
-
-                LASSERT(desc->bd_md.niov < desc->bd_page_count);
-
-                if (desc->bd_md.niov == 0)
-                        xid = bulk->bp_xid;
-                LASSERT(xid == bulk->bp_xid);   /* should all be the same */
-
-                iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
-                iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
-                desc->bd_md.niov++;
-                desc->bd_md.length += bulk->bp_buflen;
-        }
-
-        LASSERT(desc->bd_md.niov == desc->bd_page_count);
-        LASSERT(desc->bd_md.niov != 0);
-
-        source_id.nid = desc->bd_connection->c_peer.peer_nid;
-        source_id.pid = PTL_PID_ANY;
-
-        rc = PtlMEAttach(peer->peer_ni->pni_ni_h,
-                         desc->bd_portal, source_id, xid, 0,
-                         PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
-
+        if (OBD_FAIL_CHECK_ONCE(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_network_rw);
+        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);
+
+        desc->bd_success = 0;
+
+        peer = &desc->bd_import->imp_connection->c_peer;
+
+        md.user_ptr = &desc->bd_cbid;
+        md.eq_handle = peer->peer_ni->pni_eq_h;
+        md.threshold = 1;                       /* PUT or GET */
+        md.options = PTLRPC_MD_OPTIONS | 
+                     ((desc->bd_type == BULK_GET_SOURCE) ? 
+                      PTL_MD_OP_GET : PTL_MD_OP_PUT);
+        ptlrpc_fill_bulk_md(&md, desc);
+
+        LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
+        LASSERT (desc->bd_cbid.cbid_arg == desc);
+
+        /* XXX Registering the same xid on retried bulk makes my head
+         * explode trying to understand how the original request's bulk
+         * might interfere with the retried request -eeb */
+        LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
+        desc->bd_registered = 1;
+        desc->bd_last_xid = req->rq_xid;
+
+        rc = PtlMEAttach(peer->peer_ni->pni_ni_h, desc->bd_portal,
+                         desc->bd_import->imp_connection->c_peer.peer_id, 
+                         req->rq_xid, 0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
         if (rc != PTL_OK) {
                 CERROR("PtlMEAttach failed: %d\n", rc);
-                LBUG();
-                GOTO(cleanup, rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                RETURN (-ENOMEM);
         }
 
-        rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
-                         &desc->bd_md_h);
+        /* About to let the network at it... */
+        desc->bd_network_rw = 1;
+        rc = PtlMDAttach(me_h, md, PTL_UNLINK, &desc->bd_md_h);
         if (rc != PTL_OK) {
                 CERROR("PtlMDAttach failed: %d\n", rc);
-                LBUG();
-                GOTO(cleanup, rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                desc->bd_network_rw = 0;
+                rc2 = PtlMEUnlink (me_h);
+                LASSERT (rc2 == PTL_OK);
+                RETURN (-ENOMEM);
         }
 
-        ptlrpc_put_bulk_iov (desc, iov);
-
-        CDEBUG(D_NET, "Setup bulk sink buffers: %u pages %u bytes, xid %u, "
-               "portal %u on %s\n", desc->bd_md.niov, desc->bd_md.length,
-               xid, desc->bd_portal, peer->peer_ni->pni_name);
-
+        CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
+               "portal %u on %s\n",
+               desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
+               desc->bd_iov_count, desc->bd_nob,
+               req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
         RETURN(0);
-
- cleanup:
-        ptlrpc_put_bulk_iov (desc, iov);
-        ptlrpc_abort_bulk(desc);
-
-        return rc;
 }
 
-int ptlrpc_register_bulk_get(struct ptlrpc_bulk_desc *desc)
+void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
 {
-        desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_IOV;
-        desc->bd_md.eventq = 
-                desc->bd_connection->c_peer.peer_ni->pni_bulk_get_source_eq_h;
+        /* Disconnect a bulk desc from the network. Idempotent. Not
+         * thread-safe (i.e. only interlocks with completion callback). */
+        struct ptlrpc_bulk_desc *desc = req->rq_bulk;
+        wait_queue_head_t       *wq;
+        struct l_wait_info       lwi;
+        int                      rc;
+
+        LASSERT (!in_interrupt ());             /* might sleep */
 
-        return ptlrpc_register_bulk_shared(desc);
+        if (!ptlrpc_bulk_active(desc))          /* completed or */
+                return;                         /* never registered */
+        
+        LASSERT (desc->bd_req == req);          /* bd_req NULL until registered */
+
+        /* the unlink ensures the callback happens ASAP and is the last
+         * 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 client_bulk_callback() */
+
+        PtlMDUnlink (desc->bd_md_h);
+       
+        if (req->rq_set != NULL) 
+                wq = &req->rq_set->set_waitq;
+        else
+                wq = &req->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);
+                rc = l_wait_event(*wq, !ptlrpc_bulk_active(desc), &lwi);
+                if (rc == 0)
+                        return;
+                
+                LASSERT (rc == -ETIMEDOUT);
+                CWARN("Unexpectedly long timeout: desc %p\n", desc);
+        }
 }
 
-int ptlrpc_register_bulk_put(struct ptlrpc_bulk_desc *desc)
+int ptlrpc_send_reply (struct ptlrpc_request *req, int may_be_difficult)
 {
-        desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
-        desc->bd_md.eventq = 
-                desc->bd_connection->c_peer.peer_ni->pni_bulk_put_sink_eq_h;
+        struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_srv_ni->sni_service;
+        struct ptlrpc_reply_state *rs = req->rq_reply_state;
+        struct ptlrpc_connection  *conn;
+        int                        rc;
+
+        /* We must already have a reply buffer (only ptlrpc_error() may be
+         * called without one).  We usually also have a request buffer which
+         * is either the actual (swabbed) incoming request, or a saved copy
+         * if this is a req saved in target_queue_final_reply(). but this
+         * will not be true since some security handling may skip the reqmsg
+         * setting and prepare reply under normal ptlrpc layer */
+        LASSERT (rs != NULL);
+        LASSERT (req->rq_repmsg != NULL);
+        LASSERT (may_be_difficult || !rs->rs_difficult);
+        LASSERT (req->rq_repmsg == rs->rs_msg);
+        LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
+        LASSERT (rs->rs_cb_id.cbid_arg == rs);
+
+        LASSERT (req->rq_repmsg != NULL);
+        if (req->rq_type != PTL_RPC_MSG_ERR)
+                req->rq_type = PTL_RPC_MSG_REPLY;
 
-        return ptlrpc_register_bulk_shared(desc);
-}
+        req->rq_repmsg->type   = req->rq_type;
+        req->rq_repmsg->status = req->rq_status;
+        req->rq_repmsg->opc    = req->rq_reqmsg ? req->rq_reqmsg->opc : 0;
 
-int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
-{
-        int rc1, rc2;
-        /* This should be safe: these handles are initialized to be
-         * invalid in ptlrpc_prep_bulk() */
-        rc1 = PtlMDUnlink(desc->bd_md_h);
-        if (rc1 != PTL_OK)
-                CERROR("PtlMDUnlink: %d\n", rc1);
-        rc2 = PtlMEUnlink(desc->bd_me_h);
-        if (rc2 != PTL_OK)
-                CERROR("PtlMEUnlink: %d\n", rc2);
-
-        return rc1 ? rc1 : rc2;
-}
+        if (req->rq_export == NULL) 
+                conn = ptlrpc_get_connection(&req->rq_peer, NULL);
+        else
+                conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 
-void obd_brw_set_addref(struct obd_brw_set *set)
-{
-        atomic_inc(&set->brw_refcount);
-}
+        atomic_inc (&svc->srv_outstanding_replies);
 
-void obd_brw_set_add(struct obd_brw_set *set, struct ptlrpc_bulk_desc *desc)
-{
-        LASSERT(list_empty(&desc->bd_set_chain));
+        rc = svcsec_authorize(req);
+        if (rc) {
+                CERROR("Error wrap reply message "LPX64"\n", req->rq_xid);
+                goto out;
+        }
 
-        ptlrpc_bulk_addref(desc);
-        atomic_inc(&set->brw_desc_count);
-        desc->bd_brw_set = set;
-        list_add(&desc->bd_set_chain, &set->brw_desc_head);
+        rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
+                           rs->rs_difficult ? PTL_ACK_REQ : PTL_NOACK_REQ,
+                           &rs->rs_cb_id, conn,
+                           svc->srv_rep_portal, req->rq_xid);
+out:
+        if (rc != 0) {
+                atomic_dec (&svc->srv_outstanding_replies);
+
+                if (!rs->rs_difficult) {
+                        /* Callers other than target_send_reply() expect me
+                         * to clean up on a comms error */
+                        lustre_free_reply_state (rs);
+                        req->rq_reply_state = NULL;
+                        req->rq_repmsg = NULL;
+                }
+        }
+        ptlrpc_put_connection(conn);
+        return rc;
 }
 
-void obd_brw_set_del(struct ptlrpc_bulk_desc *desc)
+int ptlrpc_reply (struct ptlrpc_request *req)
 {
-        atomic_dec(&desc->bd_brw_set->brw_desc_count);
-        list_del_init(&desc->bd_set_chain);
-        ptlrpc_bulk_decref(desc);
+        return (ptlrpc_send_reply (req, 0));
 }
 
-struct obd_brw_set *obd_brw_set_new(void)
+int ptlrpc_error(struct ptlrpc_request *req)
 {
-        struct obd_brw_set *set;
-
-        OBD_ALLOC(set, sizeof(*set));
+        int rc;
+        ENTRY;
 
-        if (set != NULL) {
-                init_waitqueue_head(&set->brw_waitq);
-                INIT_LIST_HEAD(&set->brw_desc_head);
-                atomic_set(&set->brw_refcount, 1);
-                atomic_set(&set->brw_desc_count, 0);
+        if (!req->rq_repmsg) {
+                rc = lustre_pack_reply(req, 0, NULL, NULL);
+                if (rc)
+                        RETURN(rc);
         }
 
-        return set;
+        req->rq_type = PTL_RPC_MSG_ERR;
+
+        rc = ptlrpc_send_reply (req, 0);
+        RETURN(rc);
 }
 
-static void obd_brw_set_free(struct obd_brw_set *set)
+int ptl_send_rpc(struct ptlrpc_request *request)
 {
-        struct list_head *tmp, *next;
+        int rc;
+        int rc2;
+        struct ptlrpc_connection *connection;
+        unsigned long flags;
+        ptl_handle_me_t  reply_me_h;
+        ptl_md_t         reply_md;
         ENTRY;
 
-        list_for_each_safe(tmp, next, &set->brw_desc_head) {
-                struct ptlrpc_bulk_desc *desc =
-                        list_entry(tmp, struct ptlrpc_bulk_desc, bd_set_chain);
+        LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 
-                CERROR("Unfinished bulk descriptor: %p\n", desc);
+        /* If this is a re-transmit, we're required to have disengaged
+         * cleanly from the previous attempt */
+        LASSERT (!request->rq_receiving_reply);
 
-                ptlrpc_abort_bulk(desc);
-        }
-        OBD_FREE(set, sizeof(*set));
-        EXIT;
-        return;
-}
+        connection = request->rq_import->imp_connection;
 
-void obd_brw_set_decref(struct obd_brw_set *set)
-{
-        ENTRY;
-        if (atomic_dec_and_test(&set->brw_refcount))
-                obd_brw_set_free(set);
-        EXIT;
-}
+        if (request->rq_bulk != NULL) {
+                rc = ptlrpc_register_bulk (request);
+                if (rc != 0)
+                        RETURN(rc);
+        }
 
-int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
-{
-        if (req->rq_repmsg == NULL) {
-                CERROR("bad: someone called ptlrpc_reply when they meant "
-                       "ptlrpc_error\n");
-                return -EINVAL;
+        request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
+        request->rq_reqmsg->type = PTL_RPC_MSG_REQUEST;
+        request->rq_reqmsg->conn_cnt = request->rq_import->imp_conn_cnt;
+
+        /* wrap_request might need to refresh gss cred, if this is called
+         * in ptlrpcd then the whole daemon thread will be waiting on
+         * gss negotiate rpc. FIXME
+         */
+        rc = ptlrpcs_cli_wrap_request(request);
+        if (rc)
+                GOTO(cleanup_bulk, rc);
+
+        LASSERT (request->rq_replen != 0);
+        if (request->rq_repbuf == NULL) {
+                rc = ptlrpcs_cli_alloc_repbuf(request, request->rq_replen);
+                if (rc)
+                        GOTO(cleanup_bulk, rc);
         }
 
-        /* FIXME: we need to increment the count of handled events */
-        if (req->rq_type != PTL_RPC_MSG_ERR)
-                req->rq_type = PTL_RPC_MSG_REPLY;
-        //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
-        //req->rq_repmsg->token = req->rq_connection->c_remote_token;
-        req->rq_repmsg->status = HTON__u32(req->rq_status);
-        return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
-}
+        rc = PtlMEAttach(connection->c_peer.peer_ni->pni_ni_h,
+                         request->rq_reply_portal, /* XXX FIXME bug 249 */
+                         connection->c_peer.peer_id, request->rq_xid, 0,
+                         PTL_UNLINK, PTL_INS_AFTER, &reply_me_h);
+        if (rc != PTL_OK) {
+                CERROR("PtlMEAttach failed: %d\n", rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                GOTO(cleanup_repmsg, rc = -ENOMEM);
+        }
 
-int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
-{
-        int rc;
-        ENTRY;
+        spin_lock_irqsave (&request->rq_lock, flags);
+        /* If the MD attach succeeds, there _will_ be a reply_in callback */
+        request->rq_receiving_reply = 1;
+        /* 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_ptlrpcs_restart = 0;
+        request->rq_ptlrpcs_err = 0;
+        spin_unlock_irqrestore (&request->rq_lock, flags);
+
+        reply_md.start     = request->rq_repbuf;
+        reply_md.length    = request->rq_repbuf_len;
+        reply_md.threshold = 1;
+        reply_md.options   = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
+        reply_md.user_ptr  = &request->rq_reply_cbid;
+        reply_md.eq_handle    = connection->c_peer.peer_ni->pni_eq_h;
+
+        rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK, 
+                         &request->rq_reply_md_h);
+        if (rc != PTL_OK) {
+                CERROR("PtlMDAttach failed: %d\n", rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                GOTO(cleanup_me, rc -ENOMEM);
+        }
 
-        if (!req->rq_repmsg) {
-                rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen,
-                                     &req->rq_repmsg);
-                if (rc)
-                        RETURN(rc);
+        CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
+               ", portal %u on %s\n",
+               request->rq_replen, request->rq_xid,
+               request->rq_reply_portal,
+               connection->c_peer.peer_ni->pni_name);
+
+        ptlrpc_request_addref(request);        /* +1 ref for the SENT callback */
+
+        request->rq_sent = LTIME_S(CURRENT_TIME);
+        rc = ptl_send_buf(&request->rq_req_md_h, 
+                          request->rq_reqbuf, request->rq_reqdata_len,
+                          PTL_NOACK_REQ, &request->rq_req_cbid, 
+                          connection,
+                          request->rq_request_portal,
+                          request->rq_xid);
+        if (rc == 0) {
+                ptlrpc_lprocfs_rpc_sent(request);
+                RETURN(rc);
         }
 
+        ptlrpc_req_finished (request);          /* drop callback ref */
 
-        req->rq_type = PTL_RPC_MSG_ERR;
+ 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 = PtlMEUnlink(reply_me_h);
+        LASSERT (rc2 == PTL_OK);
+        /* UNLINKED callback called synchronously */
+        LASSERT (!request->rq_receiving_reply);
 
-        rc = ptlrpc_reply(svc, req);
-        RETURN(rc);
+ cleanup_repmsg:
+        ptlrpcs_cli_free_repbuf(request);
+
+ cleanup_bulk:
+        if (request->rq_bulk != NULL)
+                ptlrpc_unregister_bulk(request);
+
+        return rc;
 }
 
-int ptl_send_rpc(struct ptlrpc_request *request)
+int ptlrpc_register_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 {
-        int rc;
-        char *repbuf;
-        ptl_process_id_t source_id;
+        struct ptlrpc_srv_ni    *srv_ni = rqbd->rqbd_srv_ni;
+        struct ptlrpc_service   *service = srv_ni->sni_service;
+        static ptl_process_id_t  match_id = {PTL_NID_ANY, PTL_PID_ANY};
+        int                      rc;
+        ptl_md_t                 md;
+        ptl_handle_me_t          me_h;
 
-        ENTRY;
+        CDEBUG(D_NET, "PtlMEAttach: portal %d on %s\n",
+               service->srv_req_portal, srv_ni->sni_ni->pni_name);
+
+        if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_RQBD))
+                return (-ENOMEM);
 
-        if (request->rq_type != PTL_RPC_MSG_REQUEST) {
-                CERROR("wrong packet type sent %d\n",
-                       NTOH__u32(request->rq_reqmsg->type));
-                LBUG();
-                RETURN(EINVAL);
+        rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
+                         match_id, 0, ~0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
+        if (rc != PTL_OK) {
+                CERROR("PtlMEAttach failed: %d\n", rc);
+                return (-ENOMEM);
         }
 
-        source_id.nid = request->rq_connection->c_peer.peer_nid;
-        source_id.pid = PTL_PID_ANY;
-
-        /* add a ref, which will be balanced in request_out_callback */
-        ptlrpc_request_addref(request);
-        if (request->rq_replen != 0) {
-                if (request->rq_reply_md.start != NULL) {
-                        rc = PtlMEUnlink(request->rq_reply_me_h);
-                        if (rc != PTL_OK && rc != PTL_INV_ME) {
-                                CERROR("rc %d\n", rc);
-                                LBUG();
-                        }
-                        repbuf = (char *)request->rq_reply_md.start;
-                        request->rq_repmsg = NULL;
-                } else {
-                        OBD_ALLOC(repbuf, request->rq_replen);
-                        if (!repbuf) {
-                                LBUG();
-                                RETURN(ENOMEM);
-                        }
-                }
+        LASSERT(rqbd->rqbd_refcount == 0);
+        rqbd->rqbd_refcount = 1;
 
-                rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni->pni_ni_h,
-                             request->rq_reply_portal,/* XXX FIXME bug 625069 */
-                                 source_id, request->rq_xid, 0, PTL_UNLINK,
-                                 PTL_INS_AFTER, &request->rq_reply_me_h);
-                if (rc != PTL_OK) {
-                        CERROR("PtlMEAttach failed: %d\n", rc);
-                        LBUG();
-                        GOTO(cleanup, rc);
-                }
+        md.start     = rqbd->rqbd_buffer;
+        md.length    = service->srv_buf_size;
+        md.max_size  = service->srv_max_req_size;
+        md.threshold = PTL_MD_THRESH_INF;
+        md.options   = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT | PTL_MD_MAX_SIZE;
+        md.user_ptr  = &rqbd->rqbd_cbid;
+        md.eq_handle = srv_ni->sni_ni->pni_eq_h;
+        
+        rc = PtlMDAttach(me_h, md, PTL_UNLINK, &rqbd->rqbd_md_h);
+        if (rc == PTL_OK)
+                return (0);
+
+        CERROR("PtlMDAttach failed: %d; \n", rc);
+        LASSERT (rc == PTL_NO_SPACE);
+        rc = PtlMEUnlink (me_h);
+        LASSERT (rc == PTL_OK);
+        rqbd->rqbd_refcount = 0;
+        
+        return (-ENOMEM);
+}
 
-                request->rq_reply_md.start = repbuf;
-                request->rq_reply_md.length = request->rq_replen;
-                request->rq_reply_md.threshold = 1;
-                request->rq_reply_md.options = PTL_MD_OP_PUT;
-                request->rq_reply_md.user_ptr = request;
-                request->rq_reply_md.eventq =
-                        request->rq_connection->c_peer.peer_ni->pni_reply_in_eq_h;
-
-                rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
-                                 PTL_UNLINK, NULL);
-                if (rc != PTL_OK) {
-                        CERROR("PtlMDAttach failed: %d\n", rc);
-                        LBUG();
-                        GOTO(cleanup2, rc);
-                }
+static int rawrpc_timedout(void *data)
+{
+        struct ptlrpc_request *req = (struct ptlrpc_request *) data;
+        unsigned long flags;
 
-                CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
-                       ", portal %u on %s\n",
-                       request->rq_replen, request->rq_xid,
-                       request->rq_reply_portal,
-                       request->rq_connection->c_peer.peer_ni->pni_name);
-        }
+        spin_lock_irqsave(&req->rq_lock, flags);
+        if (!req->rq_replied)
+                req->rq_timedout = 1;
+        spin_unlock_irqrestore(&req->rq_lock, flags);
 
-        /* Clear any flags that may be present from previous sends,
-         * except for REPLAY, NO_RESEND and WANT_ACK. */
-        request->rq_flags &= (PTL_RPC_FL_REPLAY | PTL_RPC_FL_NO_RESEND |
-                              PTL_RPC_FL_WANT_ACK);
-        rc = ptl_send_buf(request, request->rq_connection,
-                          request->rq_request_portal);
-        RETURN(rc);
+        return 1;
+}
 
- cleanup2:
-        PtlMEUnlink(request->rq_reply_me_h);
- cleanup:
-        OBD_FREE(repbuf, request->rq_replen);
-        // up(&request->rq_client->cli_rpc_sem);
+/* to make things as simple as possible */
+static int rawrpc_check_reply(struct ptlrpc_request *req)
+{
+        unsigned long flags;
+        int rc;
 
+        spin_lock_irqsave (&req->rq_lock, flags);
+        rc = req->rq_replied || req->rq_net_err || req->rq_err ||
+             req->rq_resend || req->rq_restart;
+        spin_unlock_irqrestore(&req->rq_lock, flags);
         return rc;
 }
 
-void ptlrpc_link_svc_me(struct ptlrpc_request_buffer_desc *rqbd)
+/*
+ * Construct a fake ptlrpc_request to do the work, in order to
+ * user the existing callback/wakeup facilities
+ */
+int ptlrpc_do_rawrpc(struct obd_import *imp,
+                     char *reqbuf, int reqlen,
+                     char *repbuf, int *replenp,
+                     int timeout)
 {
-        struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
-        struct ptlrpc_service *service = srv_ni->sni_service;
-        static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
+        struct ptlrpc_connection *conn;
+        struct ptlrpc_request request; /* just a fake one */
+        ptl_handle_me_t reply_me_h;
+        ptl_md_t reply_md, req_md;
+        struct l_wait_info lwi;
+        unsigned long irq_flags;
         int rc;
-        ptl_md_t dummy;
-        ptl_handle_md_t md_h;
-
-        LASSERT(atomic_read(&rqbd->rqbd_refcount) == 0);
+        ENTRY;
 
-        CDEBUG(D_NET, "PtlMEAttach: portal %d on %s h %lx.%lx\n",
-               service->srv_req_portal, srv_ni->sni_ni->pni_name,
-               srv_ni->sni_ni->pni_ni_h.nal_idx,
-               srv_ni->sni_ni->pni_ni_h.handle_idx);
+        LASSERT(imp);
+        class_import_get(imp);
+        if (imp->imp_state == LUSTRE_IMP_CLOSED) {
+                CDEBUG(D_SEC, "raw rpc on closed imp(=>%s)? send anyway\n",
+                       imp->imp_target_uuid.uuid);
+        }
 
-        /* Attach the leading ME on which we build the ring */
-        rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
-                         match_id, 0, ~0,
-                         PTL_UNLINK, PTL_INS_AFTER, &rqbd->rqbd_me_h);
+        conn = imp->imp_connection;
+
+        /* initialize request */
+        memset(&request, 0, sizeof(request));
+        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_reqbuf = reqbuf;
+        request.rq_reqbuf_len = reqlen;
+        request.rq_repbuf = repbuf;
+        request.rq_repbuf_len = *replenp;
+        request.rq_set = NULL;
+        spin_lock_init(&request.rq_lock);
+        init_waitqueue_head(&request.rq_reply_waitq);
+        atomic_set(&request.rq_refcount, 1000000); /* never be droped */
+        request.rq_xid = ptlrpc_next_xid();
+
+        /* add into sending list */
+        spin_lock_irqsave(&imp->imp_lock, irq_flags);
+        list_add_tail(&request.rq_list, &imp->imp_rawrpc_list);
+        spin_unlock_irqrestore(&imp->imp_lock, irq_flags);
+
+        /* prepare reply buffer */
+        rc = PtlMEAttach(conn->c_peer.peer_ni->pni_ni_h,
+                         imp->imp_client->cli_reply_portal,
+                         conn->c_peer.peer_id, request.rq_xid, 0, PTL_UNLINK,
+                         PTL_INS_AFTER, &reply_me_h);
         if (rc != PTL_OK) {
                 CERROR("PtlMEAttach failed: %d\n", rc);
-                LBUG();
+                LASSERT (rc == PTL_NO_SPACE);
+                GOTO(cleanup_imp, rc = -ENOMEM);
         }
 
-        dummy.start      = rqbd->rqbd_buffer;
-        dummy.length     = service->srv_buf_size;
-        dummy.max_size   = service->srv_max_req_size;
-        dummy.threshold  = PTL_MD_THRESH_INF;
-        dummy.options    = PTL_MD_OP_PUT | PTL_MD_MAX_SIZE | PTL_MD_AUTO_UNLINK;
-        dummy.user_ptr   = rqbd;
-        dummy.eventq     = srv_ni->sni_eq_h;
+        spin_lock_irqsave(&request.rq_lock, irq_flags);
+        request.rq_receiving_reply = 1;
+        spin_unlock_irqrestore(&request.rq_lock, irq_flags);
 
-        atomic_inc(&srv_ni->sni_nrqbds_receiving);
-        atomic_set(&rqbd->rqbd_refcount, 1);   /* 1 ref for portals */
+        reply_md.start          = repbuf;
+        reply_md.length         = *replenp;
+        reply_md.threshold      = 1;
+        reply_md.options        = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
+        reply_md.user_ptr       = &request.rq_reply_cbid;
+        reply_md.eq_handle      = conn->c_peer.peer_ni->pni_eq_h;
 
-        rc = PtlMDAttach(rqbd->rqbd_me_h, dummy, PTL_UNLINK, &md_h);
+        rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK,
+                         &request.rq_reply_md_h);
         if (rc != PTL_OK) {
                 CERROR("PtlMDAttach failed: %d\n", rc);
-                LBUG();
-#warning proper cleanup required
-                PtlMEUnlink (rqbd->rqbd_me_h);
-                atomic_set(&rqbd->rqbd_refcount, 0);
-                atomic_dec(&srv_ni->sni_nrqbds_receiving);
+                LASSERT (rc == PTL_NO_SPACE);
+                GOTO(cleanup_me, rc = -ENOMEM);
+        }
+
+        /* prepare request buffer */
+        req_md.start            = reqbuf;
+        req_md.length           = reqlen;
+        req_md.threshold        = 1;
+        req_md.options          = PTLRPC_MD_OPTIONS;
+        req_md.user_ptr         = &request.rq_req_cbid;
+        req_md.eq_handle        = conn->c_peer.peer_ni->pni_eq_h;
+
+        rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h,
+                       req_md, PTL_UNLINK, &request.rq_req_md_h);
+        if (rc != PTL_OK) {
+                CERROR("PtlMDBind failed %d\n", rc);
+                LASSERT (rc == PTL_NO_SPACE);
+                GOTO(cleanup_me, rc = -ENOMEM);
+        }
+
+        rc = PtlPut(request.rq_req_md_h, PTL_NOACK_REQ, conn->c_peer.peer_id,
+                    imp->imp_client->cli_request_portal,
+                    0, request.rq_xid, 0, 0);
+        if (rc != PTL_OK) {
+                CERROR("PtlPut failed %d\n", rc);
+                GOTO(cleanup_md, rc);
         }
+
+        lwi = LWI_TIMEOUT(timeout * HZ, rawrpc_timedout, &request);
+        l_wait_event(request.rq_reply_waitq,
+                     rawrpc_check_reply(&request), &lwi);
+
+        ptlrpc_unregister_reply(&request);
+
+        if (request.rq_err || request.rq_resend || request.rq_intr ||
+            request.rq_timedout || !request.rq_replied) {
+                CERROR("secinit rpc error: err %d, resend %d, "
+                       "intr %d, timedout %d, replied %d\n",
+                        request.rq_err, request.rq_resend, request.rq_intr,
+                        request.rq_timedout, request.rq_replied);
+                if (request.rq_timedout)
+                        rc = -ETIMEDOUT;
+                else
+                        rc = -EINVAL;
+        } else {
+                *replenp = request.rq_nob_received;
+                rc = 0;
+        }
+        GOTO(cleanup_imp, rc);
+
+cleanup_md:
+        PtlMDUnlink(request.rq_req_md_h);
+cleanup_me:
+        PtlMEUnlink(reply_me_h);
+cleanup_imp:
+        spin_lock_irqsave(&imp->imp_lock, irq_flags);
+        list_del_init(&request.rq_list);
+        spin_unlock_irqrestore(&imp->imp_lock, irq_flags);
+
+        class_import_put(imp);
+        RETURN(rc);
 }