1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_RPC
25 #include <linux/obd_support.h>
26 #include <linux/lustre_net.h>
27 #include <linux/lustre_lib.h>
29 extern ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq,
30 bulk_source_eq, bulk_sink_eq;
31 static ptl_process_id_t local_id = {PTL_NID_ANY, PTL_PID_ANY};
33 static int ptl_send_buf(struct ptlrpc_request *request,
34 struct ptlrpc_connection *conn, int portal)
37 ptl_process_id_t remote_id;
41 request->rq_req_md.user_ptr = request;
43 switch (request->rq_type) {
44 case PTL_RPC_TYPE_REQUEST:
45 request->rq_req_md.start = request->rq_reqmsg;
46 request->rq_req_md.length = request->rq_reqlen;
47 request->rq_req_md.eventq = request_out_eq;
48 request->rq_req_md.threshold = 1;
51 case PTL_RPC_TYPE_REPLY:
52 request->rq_req_md.start = request->rq_repmsg;
53 request->rq_req_md.length = request->rq_replen;
54 request->rq_req_md.eventq = reply_out_eq;
55 request->rq_req_md.threshold = 1;
60 return -1; /* notreached */
62 request->rq_req_md.options = PTL_MD_OP_PUT;
63 request->rq_req_md.user_ptr = request;
65 rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
66 //CERROR("MDBind (outgoing req/rep/bulk): %Lu\n", (__u64)md_h);
68 CERROR("PtlMDBind failed: %d\n", rc);
73 remote_id.nid = conn->c_peer.peer_nid;
76 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %Ld\n",
77 request->rq_req_md.length, portal, request->rq_xid);
79 rc = PtlPut(md_h, ack, remote_id, portal, 0, request->rq_xid,
82 CERROR("PtlPut(%Lu, %d, %Ld) failed: %d\n", remote_id.nid,
83 portal, request->rq_xid, rc);
90 static inline struct iovec *
91 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
95 if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
96 return (desc->bd_iov);
98 OBD_ALLOC (iov, desc->bd_page_count * sizeof (struct iovec));
106 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, struct iovec *iov)
108 if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
111 OBD_FREE (iov, desc->bd_page_count * sizeof (struct iovec));
114 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
117 struct list_head *tmp, *next;
118 ptl_process_id_t remote_id;
123 iov = ptlrpc_get_bulk_iov (desc);
127 desc->bd_md.start = iov;
128 desc->bd_md.niov = 0;
129 desc->bd_md.length = 0;
130 desc->bd_md.eventq = bulk_source_eq;
131 desc->bd_md.threshold = 2; /* SENT and ACK */
132 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
133 desc->bd_md.user_ptr = desc;
135 list_for_each_safe(tmp, next, &desc->bd_page_list) {
136 struct ptlrpc_bulk_page *bulk;
137 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
139 LASSERT (desc->bd_md.niov < desc->bd_page_count);
141 if (desc->bd_md.niov == 0)
143 LASSERT (xid == bulk->bp_xid); /* should all be the same */
145 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
146 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
148 desc->bd_md.length += bulk->bp_buflen;
151 LASSERT (desc->bd_md.niov == desc->bd_page_count);
152 LASSERT (desc->bd_md.niov != 0);
154 rc = PtlMDBind(desc->bd_connection->c_peer.peer_ni, desc->bd_md,
157 ptlrpc_put_bulk_iov (desc, iov); /* move down to reduce latency to send */
160 CERROR("PtlMDBind failed: %d\n", rc);
165 remote_id.nid = desc->bd_connection->c_peer.peer_nid;
168 CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d nid %Lx pid %d xid %d\n",
169 desc->bd_md.niov, desc->bd_md.length,
170 desc->bd_portal, remote_id.nid, remote_id.pid, xid);
172 rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
173 desc->bd_portal, 0, xid, 0, 0);
175 CERROR("PtlPut(%Lu, %d, %d) failed: %d\n",
176 remote_id.nid, desc->bd_portal, xid, rc);
177 PtlMDUnlink(desc->bd_md_h);
185 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
187 struct list_head *tmp, *next;
193 iov = ptlrpc_get_bulk_iov (desc);
197 desc->bd_md.start = iov;
198 desc->bd_md.niov = 0;
199 desc->bd_md.length = 0;
200 desc->bd_md.threshold = 1;
201 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
202 desc->bd_md.user_ptr = desc;
203 desc->bd_md.eventq = bulk_sink_eq;
205 list_for_each_safe(tmp, next, &desc->bd_page_list) {
206 struct ptlrpc_bulk_page *bulk;
207 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
209 LASSERT (desc->bd_md.niov < desc->bd_page_count);
211 if (desc->bd_md.niov == 0)
213 LASSERT (xid == bulk->bp_xid); /* should all be the same */
215 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
216 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
218 desc->bd_md.length += bulk->bp_buflen;
221 LASSERT (desc->bd_md.niov == desc->bd_page_count);
222 LASSERT (desc->bd_md.niov != 0);
224 rc = PtlMEAttach(desc->bd_connection->c_peer.peer_ni,
225 desc->bd_portal, local_id, xid, 0,
226 PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
228 ptlrpc_put_bulk_iov (desc, iov);
231 CERROR("PtlMEAttach failed: %d\n", rc);
236 rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
239 CERROR("PtlMDAttach failed: %d\n", rc);
244 CDEBUG(D_NET, "Setup bulk sink buffers: %u pages %u bytes, xid %u, "
245 "portal %u\n", desc->bd_md.niov, desc->bd_md.length,
246 xid, desc->bd_portal);
251 ptlrpc_abort_bulk(desc);
256 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
258 /* This should be safe: these handles are initialized to be
259 * invalid in ptlrpc_prep_bulk() */
260 PtlMDUnlink(desc->bd_md_h);
261 PtlMEUnlink(desc->bd_me_h);
266 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
268 if (req->rq_repmsg == NULL) {
269 CERROR("bad: someone called ptlrpc_reply when they meant "
274 /* FIXME: we need to increment the count of handled events */
275 req->rq_type = PTL_RPC_TYPE_REPLY;
276 //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
277 //req->rq_repmsg->token = req->rq_connection->c_remote_token;
278 req->rq_repmsg->status = HTON__u32(req->rq_status);
279 req->rq_repmsg->type = HTON__u32(req->rq_type);
280 return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
283 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
288 if (req->rq_repmsg) {
289 CERROR("req already has repmsg\n");
293 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
297 req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
299 rc = ptlrpc_reply(svc, req);
304 int ptl_send_rpc(struct ptlrpc_request *request)
311 if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
312 CERROR("wrong packet type sent %d\n",
313 NTOH__u32(request->rq_reqmsg->type));
317 if (request->rq_replen == 0) {
318 CERROR("request->rq_replen is 0!\n");
322 /* request->rq_repmsg is set only when the reply comes in, in
323 * client_packet_callback() */
324 if (request->rq_reply_md.start)
325 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
327 OBD_ALLOC(repbuf, request->rq_replen);
333 // down(&request->rq_client->cli_rpc_sem);
335 rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
336 request->rq_import->imp_client->cli_reply_portal,
337 local_id, request->rq_xid, 0, PTL_UNLINK,
338 PTL_INS_AFTER, &request->rq_reply_me_h);
340 CERROR("PtlMEAttach failed: %d\n", rc);
345 request->rq_type = PTL_RPC_TYPE_REQUEST;
346 request->rq_reply_md.start = repbuf;
347 request->rq_reply_md.length = request->rq_replen;
348 request->rq_reply_md.threshold = 1;
349 request->rq_reply_md.options = PTL_MD_OP_PUT;
350 request->rq_reply_md.user_ptr = request;
351 request->rq_reply_md.eventq = reply_in_eq;
353 rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
354 PTL_UNLINK, &request->rq_reply_md_h);
356 CERROR("PtlMDAttach failed: %d\n", rc);
361 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
362 request->rq_replen, request->rq_xid,
363 request->rq_import->imp_client->cli_reply_portal);
365 rc = ptl_send_buf(request, request->rq_connection,
366 request->rq_import->imp_client->cli_request_portal);
370 PtlMEUnlink(request->rq_reply_me_h);
372 OBD_FREE(repbuf, request->rq_replen);
373 // up(&request->rq_client->cli_rpc_sem);
378 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
382 ptl_handle_md_t md_h;
384 /* Attach the leading ME on which we build the ring */
385 rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
386 local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
387 &(service->srv_me_h[i]));
389 CERROR("PtlMEAttach failed: %d\n", rc);
393 if (service->srv_ref_count[i])
396 dummy.start = service->srv_buf[i];
397 dummy.length = service->srv_buf_size;
398 dummy.max_offset = service->srv_buf_size;
399 dummy.threshold = PTL_MD_THRESH_INF;
400 dummy.options = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
401 dummy.user_ptr = service;
402 dummy.eventq = service->srv_eq_h;
403 dummy.max_offset = service->srv_buf_size;
405 rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
408 CERROR("PtlMDAttach failed: %d\n", rc);
413 /* ptl_handled_rpc() should be called by the sleeping process once
414 * it finishes processing an event. This ensures the ref count is
415 * decremented and that the rpc ring buffer cycles properly.
417 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
421 spin_lock(&service->srv_lock);
422 for (index = 0; index < service->srv_ring_length; index++)
423 if (service->srv_buf[index] == start)
426 if (index == service->srv_ring_length)
429 CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
430 service->srv_ref_count[index]);
431 service->srv_ref_count[index]--;
433 if (service->srv_ref_count[index] < 0)
436 if (service->srv_ref_count[index] == 0 &&
437 !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
438 CDEBUG(D_NET, "relinking %d\n", index);
439 ptlrpc_link_svc_me(service, index);
442 spin_unlock(&service->srv_lock);