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;
40 request->rq_req_md.user_ptr = request;
42 switch (request->rq_type) {
43 case PTL_RPC_MSG_REQUEST:
44 request->rq_reqmsg->type = HTON__u32(request->rq_type);
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;
49 case PTL_RPC_MSG_REPLY:
50 request->rq_repmsg->type = HTON__u32(request->rq_type);
51 request->rq_req_md.start = request->rq_repmsg;
52 request->rq_req_md.length = request->rq_replen;
53 request->rq_req_md.eventq = reply_out_eq;
57 return -1; /* notreached */
59 request->rq_req_md.threshold = 1;
60 request->rq_req_md.options = PTL_MD_OP_PUT;
61 request->rq_req_md.user_ptr = request;
63 rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
65 CERROR("PtlMDBind failed: %d\n", rc);
70 remote_id.nid = conn->c_peer.peer_nid;
73 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %Ld\n",
74 request->rq_req_md.length, portal, request->rq_xid);
76 rc = PtlPut(md_h, PTL_NOACK_REQ, remote_id, portal, 0, request->rq_xid,
79 CERROR("PtlPut(%Lu, %d, %Ld) failed: %d\n", remote_id.nid,
80 portal, request->rq_xid, rc);
87 static inline struct iovec *
88 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
92 if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
93 return (desc->bd_iov);
95 OBD_ALLOC (iov, desc->bd_page_count * sizeof (struct iovec));
103 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, struct iovec *iov)
105 if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
108 OBD_FREE (iov, desc->bd_page_count * sizeof (struct iovec));
111 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
114 struct list_head *tmp, *next;
115 ptl_process_id_t remote_id;
120 iov = ptlrpc_get_bulk_iov (desc);
124 desc->bd_md.start = iov;
125 desc->bd_md.niov = 0;
126 desc->bd_md.length = 0;
127 desc->bd_md.eventq = bulk_source_eq;
128 desc->bd_md.threshold = 2; /* SENT and ACK */
129 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
130 desc->bd_md.user_ptr = desc;
132 list_for_each_safe(tmp, next, &desc->bd_page_list) {
133 struct ptlrpc_bulk_page *bulk;
134 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
136 LASSERT (desc->bd_md.niov < desc->bd_page_count);
138 if (desc->bd_md.niov == 0)
140 LASSERT (xid == bulk->bp_xid); /* should all be the same */
142 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
143 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
145 desc->bd_md.length += bulk->bp_buflen;
148 LASSERT (desc->bd_md.niov == desc->bd_page_count);
149 LASSERT (desc->bd_md.niov != 0);
151 rc = PtlMDBind(desc->bd_connection->c_peer.peer_ni, desc->bd_md,
154 ptlrpc_put_bulk_iov (desc, iov); /* move down to reduce latency to send */
157 CERROR("PtlMDBind failed: %d\n", rc);
162 remote_id.nid = desc->bd_connection->c_peer.peer_nid;
165 CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d nid %Lx pid %d xid %d\n",
166 desc->bd_md.niov, desc->bd_md.length,
167 desc->bd_portal, remote_id.nid, remote_id.pid, xid);
169 rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
170 desc->bd_portal, 0, xid, 0, 0);
172 CERROR("PtlPut(%Lu, %d, %d) failed: %d\n",
173 remote_id.nid, desc->bd_portal, xid, rc);
174 PtlMDUnlink(desc->bd_md_h);
182 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
184 struct list_head *tmp, *next;
190 iov = ptlrpc_get_bulk_iov (desc);
194 desc->bd_md.start = iov;
195 desc->bd_md.niov = 0;
196 desc->bd_md.length = 0;
197 desc->bd_md.threshold = 1;
198 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
199 desc->bd_md.user_ptr = desc;
200 desc->bd_md.eventq = bulk_sink_eq;
202 list_for_each_safe(tmp, next, &desc->bd_page_list) {
203 struct ptlrpc_bulk_page *bulk;
204 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
206 LASSERT (desc->bd_md.niov < desc->bd_page_count);
208 if (desc->bd_md.niov == 0)
210 LASSERT (xid == bulk->bp_xid); /* should all be the same */
212 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
213 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
215 desc->bd_md.length += bulk->bp_buflen;
218 LASSERT (desc->bd_md.niov == desc->bd_page_count);
219 LASSERT (desc->bd_md.niov != 0);
221 rc = PtlMEAttach(desc->bd_connection->c_peer.peer_ni,
222 desc->bd_portal, local_id, xid, 0,
223 PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
225 ptlrpc_put_bulk_iov (desc, iov);
228 CERROR("PtlMEAttach failed: %d\n", rc);
233 rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
236 CERROR("PtlMDAttach failed: %d\n", rc);
241 CDEBUG(D_NET, "Setup bulk sink buffers: %u pages %u bytes, xid %u, "
242 "portal %u\n", desc->bd_md.niov, desc->bd_md.length,
243 xid, desc->bd_portal);
248 ptlrpc_abort_bulk(desc);
253 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
255 /* This should be safe: these handles are initialized to be
256 * invalid in ptlrpc_prep_bulk() */
257 PtlMDUnlink(desc->bd_md_h);
258 PtlMEUnlink(desc->bd_me_h);
263 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
265 if (req->rq_repmsg == NULL) {
266 CERROR("bad: someone called ptlrpc_reply when they meant "
271 /* FIXME: we need to increment the count of handled events */
272 if (req->rq_type != PTL_RPC_MSG_ERR)
273 req->rq_type = PTL_RPC_MSG_REPLY;
274 //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
275 //req->rq_repmsg->token = req->rq_connection->c_remote_token;
276 req->rq_repmsg->status = HTON__u32(req->rq_status);
277 return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
280 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
285 if (req->rq_repmsg) {
286 CERROR("req already has repmsg\n");
290 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
294 req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
296 rc = ptlrpc_reply(svc, req);
300 int ptl_send_rpc(struct ptlrpc_request *request)
307 if (request->rq_type != PTL_RPC_MSG_REQUEST) {
308 CERROR("wrong packet type sent %d\n",
309 NTOH__u32(request->rq_reqmsg->type));
313 if (request->rq_replen == 0) {
314 CERROR("request->rq_replen is 0!\n");
318 /* request->rq_repmsg is set only when the reply comes in, in
319 * client_packet_callback() */
320 if (request->rq_reply_md.start)
321 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
323 OBD_ALLOC(repbuf, request->rq_replen);
329 // down(&request->rq_client->cli_rpc_sem);
331 rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
332 request->rq_import->imp_client->cli_reply_portal,
333 local_id, request->rq_xid, 0, PTL_UNLINK,
334 PTL_INS_AFTER, &request->rq_reply_me_h);
336 CERROR("PtlMEAttach failed: %d\n", rc);
341 request->rq_reply_md.start = repbuf;
342 request->rq_reply_md.length = request->rq_replen;
343 request->rq_reply_md.threshold = 1;
344 request->rq_reply_md.options = PTL_MD_OP_PUT;
345 request->rq_reply_md.user_ptr = request;
346 request->rq_reply_md.eventq = reply_in_eq;
348 rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
349 PTL_UNLINK, &request->rq_reply_md_h);
351 CERROR("PtlMDAttach failed: %d\n", rc);
356 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
357 request->rq_replen, request->rq_xid,
358 request->rq_import->imp_client->cli_reply_portal);
360 rc = ptl_send_buf(request, request->rq_connection,
361 request->rq_import->imp_client->cli_request_portal);
365 PtlMEUnlink(request->rq_reply_me_h);
367 OBD_FREE(repbuf, request->rq_replen);
368 // up(&request->rq_client->cli_rpc_sem);
373 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
377 ptl_handle_md_t md_h;
379 /* Attach the leading ME on which we build the ring */
380 rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
381 local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
382 &(service->srv_me_h[i]));
384 CERROR("PtlMEAttach failed: %d\n", rc);
388 if (service->srv_ref_count[i])
391 dummy.start = service->srv_buf[i];
392 dummy.length = service->srv_buf_size;
393 dummy.max_offset = service->srv_buf_size;
394 dummy.threshold = PTL_MD_THRESH_INF;
395 dummy.options = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
396 dummy.user_ptr = service;
397 dummy.eventq = service->srv_eq_h;
398 dummy.max_offset = service->srv_buf_size;
400 rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
403 CERROR("PtlMDAttach failed: %d\n", rc);
408 /* ptl_handled_rpc() should be called by the sleeping process once
409 * it finishes processing an event. This ensures the ref count is
410 * decremented and that the rpc ring buffer cycles properly.
412 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
416 spin_lock(&service->srv_lock);
417 for (index = 0; index < service->srv_ring_length; index++)
418 if (service->srv_buf[index] == start)
421 if (index == service->srv_ring_length)
424 CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
425 service->srv_ref_count[index]);
426 service->srv_ref_count[index]--;
428 if (service->srv_ref_count[index] < 0)
431 if (service->srv_ref_count[index] == 0 &&
432 !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
433 CDEBUG(D_NET, "relinking %d\n", index);
434 ptlrpc_link_svc_me(service, index);
437 spin_unlock(&service->srv_lock);