1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 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 <liblustre.h>
26 #include <portals/lib-types.h>
28 #include <linux/obd_support.h>
29 #include <linux/lustre_net.h>
30 #include <linux/lustre_lib.h>
31 #include <linux/obd.h>
32 #include "ptlrpc_internal.h"
34 static int ptl_send_buf (ptl_handle_md_t *mdh, void *base, int len,
35 ptl_ack_req_t ack, struct ptlrpc_cb_id *cbid,
36 struct ptlrpc_connection *conn, int portal, __u64 xid)
38 ptl_process_id_t remote_id;
42 char str[PTL_NALFMT_SIZE];
45 LASSERT (portal != 0);
46 LASSERT (conn != NULL);
47 CDEBUG (D_INFO, "conn=%p ni %s nid "LPX64" (%s) on %s\n",
48 conn, conn->c_peer.peer_ni->pni_name,
49 conn->c_peer.peer_nid,
50 portals_nid2str(conn->c_peer.peer_ni->pni_number,
51 conn->c_peer.peer_nid, str),
52 conn->c_peer.peer_ni->pni_name);
54 remote_id.nid = conn->c_peer.peer_nid,
59 md.threshold = (ack == PTL_ACK_REQ) ? 2 : 1;
60 md.options = PTLRPC_MD_OPTIONS;
62 md.eventq = conn->c_peer.peer_ni->pni_eq_h;
64 if (ack == PTL_ACK_REQ &&
65 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
66 /* don't ask for the ack to simulate failing client */
68 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
71 rc = PtlMDBind (conn->c_peer.peer_ni->pni_ni_h, md,
74 CERROR ("PtlMDBind failed: %d\n", rc);
75 LASSERT (rc == PTL_NO_SPACE);
79 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
82 rc2 = PtlPut (*mdh, ack, remote_id, portal, 0, xid, 0, 0);
84 /* We're going to get an UNLINK event when I unlink below,
85 * which will complete just like any other failed send, so
86 * I fall through and return success here! */
87 CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n",
88 remote_id.nid, portal, xid, rc);
89 rc2 = PtlMDUnlink(*mdh);
90 LASSERT (rc2 == PTL_OK);
96 static void ptlrpc_fill_md(ptl_md_t *md, struct ptlrpc_bulk_desc *desc)
98 LASSERT(ptl_md_max_iovs() == 0 ||
99 (desc->bd_iov_count <= ptl_md_max_iovs()));
101 if (ptl_requires_iov() || desc->bd_iov_count > 0) {
102 md->options |= PTLRPC_PTL_MD_IOV;
103 md->start = &desc->bd_iov[0];
104 md->niov = desc->bd_iov_count;
106 md->start = ptl_iov_base(&desc->bd_iov[0]);
110 int ptlrpc_start_bulk_transfer (struct ptlrpc_bulk_desc *desc)
114 struct ptlrpc_peer *peer;
115 ptl_process_id_t remote_id;
120 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_PUT_NET))
123 /* NB no locking required until desc is on the network */
124 LASSERT (!desc->bd_network_rw);
125 LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
126 desc->bd_type == BULK_GET_SINK);
127 desc->bd_success = 0;
128 peer = &desc->bd_export->exp_connection->c_peer;
130 md.length = desc->bd_nob;
131 md.eventq = peer->peer_ni->pni_eq_h;
132 md.threshold = 2; /* SENT and ACK/REPLY */
133 md.options = PTLRPC_MD_OPTIONS;
135 ptlrpc_fill_md(&md, desc);
136 md.user_ptr = &desc->bd_cbid;
137 LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
138 LASSERT (desc->bd_cbid.cbid_arg == desc);
140 /* NB total length may be 0 for a read past EOF, so we send a 0
141 * length bulk, since the client expects a bulk event. */
143 rc = PtlMDBind(peer->peer_ni->pni_ni_h, md,
144 PTL_UNLINK, &desc->bd_md_h);
146 CERROR("PtlMDBind failed: %d\n", rc);
147 LASSERT (rc == PTL_NO_SPACE);
151 /* Client's bulk and reply matchbits are the same */
152 xid = desc->bd_req->rq_xid;
153 remote_id.nid = peer->peer_nid;
156 CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d on %s "
157 "nid "LPX64" pid %d xid "LPX64"\n",
158 md.niov, md.length, desc->bd_portal, peer->peer_ni->pni_name,
159 remote_id.nid, remote_id.pid, xid);
161 /* Network is about to get at the memory */
162 desc->bd_network_rw = 1;
164 if (desc->bd_type == BULK_PUT_SOURCE)
165 rc = PtlPut (desc->bd_md_h, PTL_ACK_REQ, remote_id,
166 desc->bd_portal, 0, xid, 0, 0);
168 rc = PtlGet (desc->bd_md_h, remote_id,
169 desc->bd_portal, 0, xid, 0);
172 /* Can't send, so we unlink the MD bound above. The UNLINK
173 * event this creates will signal completion with failure,
174 * so we return SUCCESS here! */
175 CERROR("Transfer("LPU64", %d, "LPX64") failed: %d\n",
176 remote_id.nid, desc->bd_portal, xid, rc);
177 rc2 = PtlMDUnlink(desc->bd_md_h);
178 LASSERT (rc2 == PTL_OK);
184 void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
186 /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
187 * serialises with completion callback) */
188 struct l_wait_info lwi;
191 LASSERT (!in_interrupt ()); /* might sleep */
193 if (!ptlrpc_bulk_active(desc)) /* completed or */
194 return; /* never started */
196 /* The unlink ensures the callback happens ASAP and is the last
197 * one. If it fails, it must be because completion just
200 rc = PtlMDUnlink (desc->bd_md_h);
201 if (rc == PTL_MD_INVALID) {
202 LASSERT(!ptlrpc_bulk_active(desc));
206 LASSERT (rc == PTL_OK);
209 /* Network access will complete in finite time but the HUGE
210 * timeout lets us CWARN for visibility of sluggish NALs */
211 lwi = LWI_TIMEOUT (300 * HZ, NULL, NULL);
212 rc = l_wait_event(desc->bd_waitq,
213 !ptlrpc_bulk_active(desc), &lwi);
217 LASSERT(rc == -ETIMEDOUT);
218 CWARN("Unexpectedly long timeout: desc %p\n", desc);
222 int ptlrpc_register_bulk (struct ptlrpc_request *req)
224 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
225 struct ptlrpc_peer *peer;
228 ptl_process_id_t source_id;
229 ptl_handle_me_t me_h;
233 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_GET_NET))
236 /* NB no locking required until desc is on the network */
237 LASSERT (desc->bd_nob > 0);
238 LASSERT (!desc->bd_network_rw);
239 LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
240 LASSERT (desc->bd_req != NULL);
241 LASSERT (desc->bd_type == BULK_PUT_SINK ||
242 desc->bd_type == BULK_GET_SOURCE);
244 desc->bd_success = 0;
246 peer = &desc->bd_import->imp_connection->c_peer;
248 md.length = desc->bd_nob;
249 md.eventq = peer->peer_ni->pni_eq_h;
250 md.threshold = 1; /* PUT or GET */
251 md.options = PTLRPC_MD_OPTIONS |
252 ((desc->bd_type == BULK_GET_SOURCE) ?
253 PTL_MD_OP_GET : PTL_MD_OP_PUT);
254 ptlrpc_fill_md(&md, desc);
255 md.user_ptr = &desc->bd_cbid;
256 LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
257 LASSERT (desc->bd_cbid.cbid_arg == desc);
259 /* XXX Registering the same xid on retried bulk makes my head
260 * explode trying to understand how the original request's bulk
261 * might interfere with the retried request -eeb */
262 LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
263 desc->bd_registered = 1;
264 desc->bd_last_xid = req->rq_xid;
266 source_id.nid = desc->bd_import->imp_connection->c_peer.peer_nid;
267 source_id.pid = PTL_PID_ANY;
269 rc = PtlMEAttach(peer->peer_ni->pni_ni_h,
270 desc->bd_portal, source_id, req->rq_xid, 0,
271 PTL_UNLINK, PTL_INS_AFTER, &me_h);
273 CERROR("PtlMEAttach failed: %d\n", rc);
274 LASSERT (rc == PTL_NO_SPACE);
278 /* About to let the network at it... */
279 desc->bd_network_rw = 1;
280 rc = PtlMDAttach(me_h, md, PTL_UNLINK, &desc->bd_md_h);
282 CERROR("PtlMDAttach failed: %d\n", rc);
283 LASSERT (rc == PTL_NO_SPACE);
284 desc->bd_network_rw = 0;
285 rc2 = PtlMEUnlink (me_h);
286 LASSERT (rc2 == PTL_OK);
290 CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
292 desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
294 req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
298 void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
300 /* Disconnect a bulk desc from the network. Idempotent. Not
301 * thread-safe (i.e. only interlocks with completion callback). */
302 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
303 wait_queue_head_t *wq;
304 struct l_wait_info lwi;
307 LASSERT (!in_interrupt ()); /* might sleep */
309 if (!ptlrpc_bulk_active(desc)) /* completed or */
310 return; /* never registered */
312 LASSERT (desc->bd_req == req); /* bd_req NULL until registered */
314 /* the unlink ensures the callback happens ASAP and is the last
315 * one. If it fails, it must be because completion just
318 rc = PtlMDUnlink (desc->bd_md_h);
319 if (rc == PTL_MD_INVALID) {
320 LASSERT(!ptlrpc_bulk_active(desc));
324 LASSERT (rc == PTL_OK);
326 if (desc->bd_req->rq_set != NULL)
327 wq = &req->rq_set->set_waitq;
329 wq = &req->rq_reply_waitq;
332 /* Network access will complete in finite time but the HUGE
333 * timeout lets us CWARN for visibility of sluggish NALs */
334 lwi = LWI_TIMEOUT (300 * HZ, NULL, NULL);
335 rc = l_wait_event(*wq, !ptlrpc_bulk_active(desc), &lwi);
339 LASSERT (rc == -ETIMEDOUT);
340 CWARN("Unexpectedly long timeout: desc %p\n", desc);
344 int ptlrpc_send_reply (struct ptlrpc_request *req, int may_be_difficult)
346 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_srv_ni->sni_service;
347 struct ptlrpc_reply_state *rs = req->rq_reply_state;
348 struct ptlrpc_connection *conn;
351 /* We must already have a reply buffer (only ptlrpc_error() may be
352 * called without one). We must also have a request buffer which
353 * is either the actual (swabbed) incoming request, or a saved copy
354 * if this is a req saved in target_queue_final_reply(). */
355 LASSERT (req->rq_reqmsg != NULL);
356 LASSERT (rs != NULL);
357 LASSERT (req->rq_repmsg != NULL);
358 LASSERT (may_be_difficult || !rs->rs_difficult);
359 LASSERT (req->rq_repmsg == &rs->rs_msg);
360 LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
361 LASSERT (rs->rs_cb_id.cbid_arg == rs);
363 LASSERT (req->rq_repmsg != NULL);
364 if (req->rq_type != PTL_RPC_MSG_ERR)
365 req->rq_type = PTL_RPC_MSG_REPLY;
367 req->rq_repmsg->type = req->rq_type;
368 req->rq_repmsg->status = req->rq_status;
369 req->rq_repmsg->opc = req->rq_reqmsg->opc;
371 if (req->rq_export == NULL)
372 conn = ptlrpc_get_connection(&req->rq_peer, NULL);
374 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
376 atomic_inc (&svc->srv_outstanding_replies);
378 rc = ptl_send_buf (&rs->rs_md_h, req->rq_repmsg, req->rq_replen,
379 rs->rs_difficult ? PTL_ACK_REQ : PTL_NOACK_REQ,
381 svc->srv_rep_portal, req->rq_xid);
383 atomic_dec (&svc->srv_outstanding_replies);
385 if (!rs->rs_difficult) {
386 /* Callers other than target_send_reply() expect me
387 * to clean up on a comms error */
388 lustre_free_reply_state (rs);
389 req->rq_reply_state = NULL;
390 req->rq_repmsg = NULL;
393 ptlrpc_put_connection(conn);
397 int ptlrpc_reply (struct ptlrpc_request *req)
399 return (ptlrpc_send_reply (req, 0));
402 int ptlrpc_error(struct ptlrpc_request *req)
407 if (!req->rq_repmsg) {
408 rc = lustre_pack_reply(req, 0, NULL, NULL);
413 req->rq_type = PTL_RPC_MSG_ERR;
415 rc = ptlrpc_send_reply (req, 0);
419 int ptl_send_rpc(struct ptlrpc_request *request)
423 struct ptlrpc_connection *connection;
425 ptl_process_id_t source_id;
426 ptl_handle_me_t reply_me_h;
430 LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
432 /* If this is a re-transmit, we're required to have disengaged
433 * cleanly from the previous attempt */
434 LASSERT (!request->rq_receiving_reply);
436 connection = request->rq_import->imp_connection;
438 if (request->rq_bulk != NULL) {
439 rc = ptlrpc_register_bulk (request);
444 request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
445 request->rq_reqmsg->type = PTL_RPC_MSG_REQUEST;
446 request->rq_reqmsg->conn_cnt = request->rq_import->imp_conn_cnt;
448 source_id.nid = connection->c_peer.peer_nid;
449 source_id.pid = PTL_PID_ANY;
451 LASSERT (request->rq_replen != 0);
452 if (request->rq_repmsg == NULL)
453 OBD_ALLOC(request->rq_repmsg, request->rq_replen);
454 if (request->rq_repmsg == NULL)
455 GOTO(cleanup_bulk, rc = -ENOMEM);
457 rc = PtlMEAttach(connection->c_peer.peer_ni->pni_ni_h,
458 request->rq_reply_portal, /* XXX FIXME bug 249 */
459 source_id, request->rq_xid, 0, PTL_UNLINK,
460 PTL_INS_AFTER, &reply_me_h);
462 CERROR("PtlMEAttach failed: %d\n", rc);
463 LASSERT (rc == PTL_NO_SPACE);
464 GOTO(cleanup_repmsg, rc = -ENOMEM);
467 spin_lock_irqsave (&request->rq_lock, flags);
468 /* If the MD attach succeeds, there _will_ be a reply_in callback */
469 request->rq_receiving_reply = 1;
470 /* Clear any flags that may be present from previous sends. */
471 request->rq_replied = 0;
473 request->rq_timedout = 0;
474 request->rq_resend = 0;
475 request->rq_restart = 0;
476 spin_unlock_irqrestore (&request->rq_lock, flags);
478 reply_md.start = request->rq_repmsg;
479 reply_md.length = request->rq_replen;
480 reply_md.threshold = 1;
481 reply_md.options = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
482 reply_md.user_ptr = &request->rq_reply_cbid;
483 reply_md.eventq = connection->c_peer.peer_ni->pni_eq_h;
485 rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK,
486 &request->rq_reply_md_h);
488 CERROR("PtlMDAttach failed: %d\n", rc);
489 LASSERT (rc == PTL_NO_SPACE);
490 GOTO(cleanup_me, rc -ENOMEM);
493 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
494 ", portal %u on %s\n",
495 request->rq_replen, request->rq_xid,
496 request->rq_reply_portal,
497 connection->c_peer.peer_ni->pni_name);
499 ptlrpc_request_addref(request); /* +1 ref for the SENT callback */
501 request->rq_sent = LTIME_S(CURRENT_TIME);
502 ptlrpc_pinger_sending_on_import(request->rq_import);
503 rc = ptl_send_buf(&request->rq_req_md_h,
504 request->rq_reqmsg, request->rq_reqlen,
505 PTL_NOACK_REQ, &request->rq_req_cbid,
507 request->rq_request_portal,
510 ptlrpc_lprocfs_rpc_sent(request);
514 ptlrpc_req_finished (request); /* drop callback ref */
517 /* MEUnlink is safe; the PUT didn't even get off the ground, and
518 * nobody apart from the PUT's target has the right nid+XID to
519 * access the reply buffer. */
520 rc2 = PtlMEUnlink(reply_me_h);
521 LASSERT (rc2 == PTL_OK);
522 /* UNLINKED callback called synchronously */
523 LASSERT (!request->rq_receiving_reply);
526 OBD_FREE(request->rq_repmsg, request->rq_replen);
527 request->rq_repmsg = NULL;
530 if (request->rq_bulk != NULL)
531 ptlrpc_unregister_bulk(request);
536 int ptlrpc_register_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
538 struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
539 struct ptlrpc_service *service = srv_ni->sni_service;
540 static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
543 ptl_handle_me_t me_h;
545 CDEBUG(D_NET, "PtlMEAttach: portal %d on %s\n",
546 service->srv_req_portal, srv_ni->sni_ni->pni_name);
548 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_RQBD))
551 rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
552 match_id, 0, ~0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
554 CERROR("PtlMEAttach failed: %d\n", rc);
558 LASSERT(rqbd->rqbd_refcount == 0);
559 rqbd->rqbd_refcount = 1;
561 md.start = rqbd->rqbd_buffer;
562 md.length = service->srv_buf_size;
563 md.max_size = service->srv_max_req_size;
564 md.threshold = PTL_MD_THRESH_INF;
565 md.options = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT | PTL_MD_MAX_SIZE;
566 md.user_ptr = &rqbd->rqbd_cbid;
567 md.eventq = srv_ni->sni_ni->pni_eq_h;
569 rc = PtlMDAttach(me_h, md, PTL_UNLINK, &rqbd->rqbd_md_h);
573 CERROR("PtlMDAttach failed: %d; \n", rc);
574 LASSERT (rc == PTL_NO_SPACE);
575 rc = PtlMEUnlink (me_h);
576 LASSERT (rc == PTL_OK);
577 rqbd->rqbd_refcount = 0;