1 /*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19 *
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26 /*
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2011, 2012, Intel Corporation.
31 */
32 /*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 */
36
37 #define DEBUG_SUBSYSTEM S_RPC
38 #include "../include/obd_support.h"
39 #include "../include/lustre_net.h"
40 #include "../include/lustre_lib.h"
41 #include "../include/obd.h"
42 #include "../include/obd_class.h"
43 #include "ptlrpc_internal.h"
44
45 /**
46 * Helper function. Sends \a len bytes from \a base at offset \a offset
47 * over \a conn connection to portal \a portal.
48 * Returns 0 on success or error code.
49 */
ptl_send_buf(lnet_handle_md_t * mdh,void * base,int len,lnet_ack_req_t ack,struct ptlrpc_cb_id * cbid,struct ptlrpc_connection * conn,int portal,__u64 xid,unsigned int offset)50 static int ptl_send_buf(lnet_handle_md_t *mdh, void *base, int len,
51 lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
52 struct ptlrpc_connection *conn, int portal, __u64 xid,
53 unsigned int offset)
54 {
55 int rc;
56 lnet_md_t md;
57
58 LASSERT(portal != 0);
59 LASSERT(conn != NULL);
60 CDEBUG(D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
61 md.start = base;
62 md.length = len;
63 md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
64 md.options = PTLRPC_MD_OPTIONS;
65 md.user_ptr = cbid;
66 md.eq_handle = ptlrpc_eq_h;
67
68 if (unlikely(ack == LNET_ACK_REQ &&
69 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK,
70 OBD_FAIL_ONCE))) {
71 /* don't ask for the ack to simulate failing client */
72 ack = LNET_NOACK_REQ;
73 }
74
75 rc = LNetMDBind(md, LNET_UNLINK, mdh);
76 if (unlikely(rc != 0)) {
77 CERROR("LNetMDBind failed: %d\n", rc);
78 LASSERT(rc == -ENOMEM);
79 return -ENOMEM;
80 }
81
82 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
83 len, portal, xid, offset);
84
85 rc = LNetPut(conn->c_self, *mdh, ack,
86 conn->c_peer, portal, xid, offset, 0);
87 if (unlikely(rc != 0)) {
88 int rc2;
89 /* We're going to get an UNLINK event when I unlink below,
90 * which will complete just like any other failed send, so
91 * I fall through and return success here! */
92 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
93 libcfs_id2str(conn->c_peer), portal, xid, rc);
94 rc2 = LNetMDUnlink(*mdh);
95 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
96 }
97
98 return 0;
99 }
100
mdunlink_iterate_helper(lnet_handle_md_t * bd_mds,int count)101 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
102 {
103 int i;
104
105 for (i = 0; i < count; i++)
106 LNetMDUnlink(bd_mds[i]);
107 }
108
109
110 /**
111 * Register bulk at the sender for later transfer.
112 * Returns 0 on success or error code.
113 */
ptlrpc_register_bulk(struct ptlrpc_request * req)114 int ptlrpc_register_bulk(struct ptlrpc_request *req)
115 {
116 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
117 lnet_process_id_t peer;
118 int rc = 0;
119 int rc2;
120 int posted_md;
121 int total_md;
122 __u64 xid;
123 lnet_handle_me_t me_h;
124 lnet_md_t md;
125
126 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
127 return 0;
128
129 /* NB no locking required until desc is on the network */
130 LASSERT(desc->bd_nob > 0);
131 LASSERT(desc->bd_md_count == 0);
132 LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
133 LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
134 LASSERT(desc->bd_req != NULL);
135 LASSERT(desc->bd_type == BULK_PUT_SINK ||
136 desc->bd_type == BULK_GET_SOURCE);
137
138 /* cleanup the state of the bulk for it will be reused */
139 if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
140 desc->bd_nob_transferred = 0;
141 else
142 LASSERT(desc->bd_nob_transferred == 0);
143
144 desc->bd_failure = 0;
145
146 peer = desc->bd_import->imp_connection->c_peer;
147
148 LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
149 LASSERT(desc->bd_cbid.cbid_arg == desc);
150
151 /* An XID is only used for a single request from the client.
152 * For retried bulk transfers, a new XID will be allocated in
153 * in ptlrpc_check_set() if it needs to be resent, so it is not
154 * using the same RDMA match bits after an error.
155 *
156 * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
157 * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
158 xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
159 LASSERTF(!(desc->bd_registered &&
160 req->rq_send_state != LUSTRE_IMP_REPLAY) ||
161 xid != desc->bd_last_xid,
162 "registered: %d rq_xid: %llu bd_last_xid: %llu\n",
163 desc->bd_registered, xid, desc->bd_last_xid);
164
165 total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
166 desc->bd_registered = 1;
167 desc->bd_last_xid = xid;
168 desc->bd_md_count = total_md;
169 md.user_ptr = &desc->bd_cbid;
170 md.eq_handle = ptlrpc_eq_h;
171 md.threshold = 1; /* PUT or GET */
172
173 for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
174 md.options = PTLRPC_MD_OPTIONS |
175 ((desc->bd_type == BULK_GET_SOURCE) ?
176 LNET_MD_OP_GET : LNET_MD_OP_PUT);
177 ptlrpc_fill_bulk_md(&md, desc, posted_md);
178
179 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
180 LNET_UNLINK, LNET_INS_AFTER, &me_h);
181 if (rc != 0) {
182 CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
183 desc->bd_import->imp_obd->obd_name, xid,
184 posted_md, rc);
185 break;
186 }
187
188 /* About to let the network at it... */
189 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
190 &desc->bd_mds[posted_md]);
191 if (rc != 0) {
192 CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
193 desc->bd_import->imp_obd->obd_name, xid,
194 posted_md, rc);
195 rc2 = LNetMEUnlink(me_h);
196 LASSERT(rc2 == 0);
197 break;
198 }
199 }
200
201 if (rc != 0) {
202 LASSERT(rc == -ENOMEM);
203 spin_lock(&desc->bd_lock);
204 desc->bd_md_count -= total_md - posted_md;
205 spin_unlock(&desc->bd_lock);
206 LASSERT(desc->bd_md_count >= 0);
207 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
208 req->rq_status = -ENOMEM;
209 return -ENOMEM;
210 }
211
212 /* Set rq_xid to matchbits of the final bulk so that server can
213 * infer the number of bulks that were prepared */
214 req->rq_xid = --xid;
215 LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
216 "bd_last_xid = x%llu, rq_xid = x%llu\n",
217 desc->bd_last_xid, req->rq_xid);
218
219 spin_lock(&desc->bd_lock);
220 /* Holler if peer manages to touch buffers before he knows the xid */
221 if (desc->bd_md_count != total_md)
222 CWARN("%s: Peer %s touched %d buffers while I registered\n",
223 desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
224 total_md - desc->bd_md_count);
225 spin_unlock(&desc->bd_lock);
226
227 CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, xid x%#llx-%#llx, portal %u\n",
228 desc->bd_md_count,
229 desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
230 desc->bd_iov_count, desc->bd_nob,
231 desc->bd_last_xid, req->rq_xid, desc->bd_portal);
232
233 return 0;
234 }
235 EXPORT_SYMBOL(ptlrpc_register_bulk);
236
237 /**
238 * Disconnect a bulk desc from the network. Idempotent. Not
239 * thread-safe (i.e. only interlocks with completion callback).
240 * Returns 1 on success or 0 if network unregistration failed for whatever
241 * reason.
242 */
ptlrpc_unregister_bulk(struct ptlrpc_request * req,int async)243 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
244 {
245 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
246 wait_queue_head_t *wq;
247 struct l_wait_info lwi;
248 int rc;
249
250 LASSERT(!in_interrupt()); /* might sleep */
251
252 /* Let's setup deadline for reply unlink. */
253 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
254 async && req->rq_bulk_deadline == 0)
255 req->rq_bulk_deadline = get_seconds() + LONG_UNLINK;
256
257 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
258 return 1; /* never registered */
259
260 LASSERT(desc->bd_req == req); /* bd_req NULL until registered */
261
262 /* the unlink ensures the callback happens ASAP and is the last
263 * one. If it fails, it must be because completion just happened,
264 * but we must still l_wait_event() in this case to give liblustre
265 * a chance to run client_bulk_callback() */
266 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
267
268 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
269 return 1; /* never registered */
270
271 /* Move to "Unregistering" phase as bulk was not unlinked yet. */
272 ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
273
274 /* Do not wait for unlink to finish. */
275 if (async)
276 return 0;
277
278 if (req->rq_set != NULL)
279 wq = &req->rq_set->set_waitq;
280 else
281 wq = &req->rq_reply_waitq;
282
283 for (;;) {
284 /* Network access will complete in finite time but the HUGE
285 * timeout lets us CWARN for visibility of sluggish NALs */
286 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
287 cfs_time_seconds(1), NULL, NULL);
288 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
289 if (rc == 0) {
290 ptlrpc_rqphase_move(req, req->rq_next_phase);
291 return 1;
292 }
293
294 LASSERT(rc == -ETIMEDOUT);
295 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
296 desc);
297 }
298 return 0;
299 }
300 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
301
ptlrpc_at_set_reply(struct ptlrpc_request * req,int flags)302 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
303 {
304 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
305 struct ptlrpc_service *svc = svcpt->scp_service;
306 int service_time = max_t(int, get_seconds() -
307 req->rq_arrival_time.tv_sec, 1);
308
309 if (!(flags & PTLRPC_REPLY_EARLY) &&
310 (req->rq_type != PTL_RPC_MSG_ERR) &&
311 (req->rq_reqmsg != NULL) &&
312 !(lustre_msg_get_flags(req->rq_reqmsg) &
313 (MSG_RESENT | MSG_REPLAY |
314 MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
315 /* early replies, errors and recovery requests don't count
316 * toward our service time estimate */
317 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
318
319 if (oldse != 0) {
320 DEBUG_REQ(D_ADAPTTO, req,
321 "svc %s changed estimate from %d to %d",
322 svc->srv_name, oldse,
323 at_get(&svcpt->scp_at_estimate));
324 }
325 }
326 /* Report actual service time for client latency calc */
327 lustre_msg_set_service_time(req->rq_repmsg, service_time);
328 /* Report service time estimate for future client reqs, but report 0
329 * (to be ignored by client) if it's a error reply during recovery.
330 * (bz15815) */
331 if (req->rq_type == PTL_RPC_MSG_ERR &&
332 (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
333 lustre_msg_set_timeout(req->rq_repmsg, 0);
334 else
335 lustre_msg_set_timeout(req->rq_repmsg,
336 at_get(&svcpt->scp_at_estimate));
337
338 if (req->rq_reqmsg &&
339 !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
340 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x req_flags=%#x magic=%d:%x/%x len=%d\n",
341 flags, lustre_msg_get_flags(req->rq_reqmsg),
342 lustre_msg_is_v1(req->rq_reqmsg),
343 lustre_msg_get_magic(req->rq_reqmsg),
344 lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
345 }
346 }
347
348 /**
349 * Send request reply from request \a req reply buffer.
350 * \a flags defines reply types
351 * Returns 0 on success or error code
352 */
ptlrpc_send_reply(struct ptlrpc_request * req,int flags)353 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
354 {
355 struct ptlrpc_reply_state *rs = req->rq_reply_state;
356 struct ptlrpc_connection *conn;
357 int rc;
358
359 /* We must already have a reply buffer (only ptlrpc_error() may be
360 * called without one). The reply generated by sptlrpc layer (e.g.
361 * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
362 * have a request buffer which is either the actual (swabbed) incoming
363 * request, or a saved copy if this is a req saved in
364 * target_queue_final_reply().
365 */
366 LASSERT(req->rq_no_reply == 0);
367 LASSERT(req->rq_reqbuf != NULL);
368 LASSERT(rs != NULL);
369 LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
370 LASSERT(req->rq_repmsg != NULL);
371 LASSERT(req->rq_repmsg == rs->rs_msg);
372 LASSERT(rs->rs_cb_id.cbid_fn == reply_out_callback);
373 LASSERT(rs->rs_cb_id.cbid_arg == rs);
374
375 /* There may be no rq_export during failover */
376
377 if (unlikely(req->rq_export && req->rq_export->exp_obd &&
378 req->rq_export->exp_obd->obd_fail)) {
379 /* Failed obd's only send ENODEV */
380 req->rq_type = PTL_RPC_MSG_ERR;
381 req->rq_status = -ENODEV;
382 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
383 req->rq_export->exp_obd->obd_minor);
384 }
385
386 /* In order to keep interoperability with the client (< 2.3) which
387 * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
388 * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
389 * reply buffer on client will be overflow.
390 *
391 * XXX Remove this whenever we drop the interoperability with
392 * such client.
393 */
394 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
395 sizeof(struct ptlrpc_body_v2), 1);
396
397 if (req->rq_type != PTL_RPC_MSG_ERR)
398 req->rq_type = PTL_RPC_MSG_REPLY;
399
400 lustre_msg_set_type(req->rq_repmsg, req->rq_type);
401 lustre_msg_set_status(req->rq_repmsg,
402 ptlrpc_status_hton(req->rq_status));
403 lustre_msg_set_opc(req->rq_repmsg,
404 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
405
406 target_pack_pool_reply(req);
407
408 ptlrpc_at_set_reply(req, flags);
409
410 if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
411 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
412 else
413 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
414
415 if (unlikely(conn == NULL)) {
416 CERROR("not replying on NULL connection\n"); /* bug 9635 */
417 return -ENOTCONN;
418 }
419 ptlrpc_rs_addref(rs); /* +1 ref for the network */
420
421 rc = sptlrpc_svc_wrap_reply(req);
422 if (unlikely(rc))
423 goto out;
424
425 req->rq_sent = get_seconds();
426
427 rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
428 (rs->rs_difficult && !rs->rs_no_ack) ?
429 LNET_ACK_REQ : LNET_NOACK_REQ,
430 &rs->rs_cb_id, conn,
431 ptlrpc_req2svc(req)->srv_rep_portal,
432 req->rq_xid, req->rq_reply_off);
433 out:
434 if (unlikely(rc != 0))
435 ptlrpc_req_drop_rs(req);
436 ptlrpc_connection_put(conn);
437 return rc;
438 }
439 EXPORT_SYMBOL(ptlrpc_send_reply);
440
ptlrpc_reply(struct ptlrpc_request * req)441 int ptlrpc_reply(struct ptlrpc_request *req)
442 {
443 if (req->rq_no_reply)
444 return 0;
445 return ptlrpc_send_reply(req, 0);
446 }
447 EXPORT_SYMBOL(ptlrpc_reply);
448
449 /**
450 * For request \a req send an error reply back. Create empty
451 * reply buffers if necessary.
452 */
ptlrpc_send_error(struct ptlrpc_request * req,int may_be_difficult)453 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
454 {
455 int rc;
456
457 if (req->rq_no_reply)
458 return 0;
459
460 if (!req->rq_repmsg) {
461 rc = lustre_pack_reply(req, 1, NULL, NULL);
462 if (rc)
463 return rc;
464 }
465
466 if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
467 req->rq_status != -EPERM && req->rq_status != -ENOENT &&
468 req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
469 req->rq_type = PTL_RPC_MSG_ERR;
470
471 rc = ptlrpc_send_reply(req, may_be_difficult);
472 return rc;
473 }
474 EXPORT_SYMBOL(ptlrpc_send_error);
475
ptlrpc_error(struct ptlrpc_request * req)476 int ptlrpc_error(struct ptlrpc_request *req)
477 {
478 return ptlrpc_send_error(req, 0);
479 }
480 EXPORT_SYMBOL(ptlrpc_error);
481
482 /**
483 * Send request \a request.
484 * if \a noreply is set, don't expect any reply back and don't set up
485 * reply buffers.
486 * Returns 0 on success or error code.
487 */
ptl_send_rpc(struct ptlrpc_request * request,int noreply)488 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
489 {
490 int rc;
491 int rc2;
492 int mpflag = 0;
493 struct ptlrpc_connection *connection;
494 lnet_handle_me_t reply_me_h;
495 lnet_md_t reply_md;
496 struct obd_device *obd = request->rq_import->imp_obd;
497
498 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
499 return 0;
500
501 LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
502 LASSERT(request->rq_wait_ctx == 0);
503
504 /* If this is a re-transmit, we're required to have disengaged
505 * cleanly from the previous attempt */
506 LASSERT(!request->rq_receiving_reply);
507 LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
508 (request->rq_import->imp_state == LUSTRE_IMP_FULL)));
509
510 if (unlikely(obd != NULL && obd->obd_fail)) {
511 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
512 obd->obd_name);
513 /* this prevents us from waiting in ptlrpc_queue_wait */
514 spin_lock(&request->rq_lock);
515 request->rq_err = 1;
516 spin_unlock(&request->rq_lock);
517 request->rq_status = -ENODEV;
518 return -ENODEV;
519 }
520
521 connection = request->rq_import->imp_connection;
522
523 lustre_msg_set_handle(request->rq_reqmsg,
524 &request->rq_import->imp_remote_handle);
525 lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
526 lustre_msg_set_conn_cnt(request->rq_reqmsg,
527 request->rq_import->imp_conn_cnt);
528 lustre_msghdr_set_flags(request->rq_reqmsg,
529 request->rq_import->imp_msghdr_flags);
530
531 if (request->rq_resend)
532 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
533
534 if (request->rq_memalloc)
535 mpflag = cfs_memory_pressure_get_and_set();
536
537 rc = sptlrpc_cli_wrap_request(request);
538 if (rc)
539 goto out;
540
541 /* bulk register should be done after wrap_request() */
542 if (request->rq_bulk != NULL) {
543 rc = ptlrpc_register_bulk(request);
544 if (rc != 0)
545 goto out;
546 }
547
548 if (!noreply) {
549 LASSERT(request->rq_replen != 0);
550 if (request->rq_repbuf == NULL) {
551 LASSERT(request->rq_repdata == NULL);
552 LASSERT(request->rq_repmsg == NULL);
553 rc = sptlrpc_cli_alloc_repbuf(request,
554 request->rq_replen);
555 if (rc) {
556 /* this prevents us from looping in
557 * ptlrpc_queue_wait */
558 spin_lock(&request->rq_lock);
559 request->rq_err = 1;
560 spin_unlock(&request->rq_lock);
561 request->rq_status = rc;
562 goto cleanup_bulk;
563 }
564 } else {
565 request->rq_repdata = NULL;
566 request->rq_repmsg = NULL;
567 }
568
569 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
570 connection->c_peer, request->rq_xid, 0,
571 LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
572 if (rc != 0) {
573 CERROR("LNetMEAttach failed: %d\n", rc);
574 LASSERT(rc == -ENOMEM);
575 rc = -ENOMEM;
576 goto cleanup_bulk;
577 }
578 }
579
580 spin_lock(&request->rq_lock);
581 /* If the MD attach succeeds, there _will_ be a reply_in callback */
582 request->rq_receiving_reply = !noreply;
583 request->rq_req_unlink = 1;
584 /* We are responsible for unlinking the reply buffer */
585 request->rq_reply_unlink = !noreply;
586 /* Clear any flags that may be present from previous sends. */
587 request->rq_replied = 0;
588 request->rq_err = 0;
589 request->rq_timedout = 0;
590 request->rq_net_err = 0;
591 request->rq_resend = 0;
592 request->rq_restart = 0;
593 request->rq_reply_truncate = 0;
594 spin_unlock(&request->rq_lock);
595
596 if (!noreply) {
597 reply_md.start = request->rq_repbuf;
598 reply_md.length = request->rq_repbuf_len;
599 /* Allow multiple early replies */
600 reply_md.threshold = LNET_MD_THRESH_INF;
601 /* Manage remote for early replies */
602 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
603 LNET_MD_MANAGE_REMOTE |
604 LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
605 reply_md.user_ptr = &request->rq_reply_cbid;
606 reply_md.eq_handle = ptlrpc_eq_h;
607
608 /* We must see the unlink callback to unset rq_reply_unlink,
609 so we can't auto-unlink */
610 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
611 &request->rq_reply_md_h);
612 if (rc != 0) {
613 CERROR("LNetMDAttach failed: %d\n", rc);
614 LASSERT(rc == -ENOMEM);
615 spin_lock(&request->rq_lock);
616 /* ...but the MD attach didn't succeed... */
617 request->rq_receiving_reply = 0;
618 spin_unlock(&request->rq_lock);
619 rc = -ENOMEM;
620 goto cleanup_me;
621 }
622
623 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
624 request->rq_repbuf_len, request->rq_xid,
625 request->rq_reply_portal);
626 }
627
628 /* add references on request for request_out_callback */
629 ptlrpc_request_addref(request);
630 if (obd != NULL && obd->obd_svc_stats != NULL)
631 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
632 atomic_read(&request->rq_import->imp_inflight));
633
634 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
635
636 do_gettimeofday(&request->rq_arrival_time);
637 request->rq_sent = get_seconds();
638 /* We give the server rq_timeout secs to process the req, and
639 add the network latency for our local timeout. */
640 request->rq_deadline = request->rq_sent + request->rq_timeout +
641 ptlrpc_at_get_net_latency(request);
642
643 ptlrpc_pinger_sending_on_import(request->rq_import);
644
645 DEBUG_REQ(D_INFO, request, "send flg=%x",
646 lustre_msg_get_flags(request->rq_reqmsg));
647 rc = ptl_send_buf(&request->rq_req_md_h,
648 request->rq_reqbuf, request->rq_reqdata_len,
649 LNET_NOACK_REQ, &request->rq_req_cbid,
650 connection,
651 request->rq_request_portal,
652 request->rq_xid, 0);
653 if (rc == 0)
654 goto out;
655
656 ptlrpc_req_finished(request);
657 if (noreply)
658 goto out;
659
660 cleanup_me:
661 /* MEUnlink is safe; the PUT didn't even get off the ground, and
662 * nobody apart from the PUT's target has the right nid+XID to
663 * access the reply buffer. */
664 rc2 = LNetMEUnlink(reply_me_h);
665 LASSERT(rc2 == 0);
666 /* UNLINKED callback called synchronously */
667 LASSERT(!request->rq_receiving_reply);
668
669 cleanup_bulk:
670 /* We do sync unlink here as there was no real transfer here so
671 * the chance to have long unlink to sluggish net is smaller here. */
672 ptlrpc_unregister_bulk(request, 0);
673 out:
674 if (request->rq_memalloc)
675 cfs_memory_pressure_restore(mpflag);
676 return rc;
677 }
678 EXPORT_SYMBOL(ptl_send_rpc);
679
680 /**
681 * Register request buffer descriptor for request receiving.
682 */
ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc * rqbd)683 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
684 {
685 struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
686 static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY};
687 int rc;
688 lnet_md_t md;
689 lnet_handle_me_t me_h;
690
691 CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
692 service->srv_req_portal);
693
694 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
695 return -ENOMEM;
696
697 /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
698 * which means buffer can only be attached on local CPT, and LND
699 * threads can find it by grabbing a local lock */
700 rc = LNetMEAttach(service->srv_req_portal,
701 match_id, 0, ~0, LNET_UNLINK,
702 rqbd->rqbd_svcpt->scp_cpt >= 0 ?
703 LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
704 if (rc != 0) {
705 CERROR("LNetMEAttach failed: %d\n", rc);
706 return -ENOMEM;
707 }
708
709 LASSERT(rqbd->rqbd_refcount == 0);
710 rqbd->rqbd_refcount = 1;
711
712 md.start = rqbd->rqbd_buffer;
713 md.length = service->srv_buf_size;
714 md.max_size = service->srv_max_req_size;
715 md.threshold = LNET_MD_THRESH_INF;
716 md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
717 md.user_ptr = &rqbd->rqbd_cbid;
718 md.eq_handle = ptlrpc_eq_h;
719
720 rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
721 if (rc == 0)
722 return 0;
723
724 CERROR("LNetMDAttach failed: %d;\n", rc);
725 LASSERT(rc == -ENOMEM);
726 rc = LNetMEUnlink(me_h);
727 LASSERT(rc == 0);
728 rqbd->rqbd_refcount = 0;
729
730 return -ENOMEM;
731 }
732