Beispiel #1
0
/**
 * Send request reply from request \a req reply buffer.
 * \a flags defines reply types
 * Returns 0 on success or error code
 */
int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
{
	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). The reply generated by sptlrpc layer (e.g.
	 * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
	 * 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().
	 */
	LASSERT(req->rq_no_reply == 0);
	LASSERT(req->rq_reqbuf != NULL);
	LASSERT(rs != NULL);
	LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
	LASSERT(req->rq_repmsg != NULL);
	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);

	/* There may be no rq_export during failover */

	if (unlikely(req->rq_export && req->rq_export->exp_obd &&
		     req->rq_export->exp_obd->obd_fail)) {
		/* Failed obd's only send ENODEV */
		req->rq_type = PTL_RPC_MSG_ERR;
		req->rq_status = -ENODEV;
		CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
		       req->rq_export->exp_obd->obd_minor);
	}

	/* In order to keep interoperability with the client (< 2.3) which
	 * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
	 * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
	 * reply buffer on client will be overflow.
	 *
	 * XXX Remove this whenever we drop the interoperability with
	 * such client.
	 */
	req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
					   sizeof(struct ptlrpc_body_v2), 1);

	if (req->rq_type != PTL_RPC_MSG_ERR)
		req->rq_type = PTL_RPC_MSG_REPLY;

	lustre_msg_set_type(req->rq_repmsg, req->rq_type);
	lustre_msg_set_status(req->rq_repmsg,
			      ptlrpc_status_hton(req->rq_status));
	lustre_msg_set_opc(req->rq_repmsg,
		req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);

	target_pack_pool_reply(req);

	ptlrpc_at_set_reply(req, flags);

	if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
		conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
	else
		conn = ptlrpc_connection_addref(req->rq_export->exp_connection);

	if (unlikely(conn == NULL)) {
		CERROR("not replying on NULL connection\n"); /* bug 9635 */
		return -ENOTCONN;
	}
	ptlrpc_rs_addref(rs);		   /* +1 ref for the network */

	rc = sptlrpc_svc_wrap_reply(req);
	if (unlikely(rc))
		goto out;

	req->rq_sent = ktime_get_real_seconds();

	rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
			  (rs->rs_difficult && !rs->rs_no_ack) ?
			  LNET_ACK_REQ : LNET_NOACK_REQ,
			  &rs->rs_cb_id, conn,
			  ptlrpc_req2svc(req)->srv_rep_portal,
			  req->rq_xid, req->rq_reply_off);
out:
	if (unlikely(rc != 0))
		ptlrpc_req_drop_rs(req);
	ptlrpc_connection_put(conn);
	return rc;
}
Beispiel #2
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_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;
}
Beispiel #3
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;
}