?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
s3:rpc_server:netlogon: don't require NEG_AUTHENTICATED_RPC in netr_ServerAuthenticate*()
[Samba.git] / ctdb / ib / ibwrapper.c
blob5ca3b946f00f371bd8a6a0cd4eadcfe09fa14d9b
1 /*
2 * Unix SMB/CIFS implementation.
3 * Wrap Infiniband calls.
4 *
5 * Copyright (C) Sven Oehme <oehmes@de.ibm.com> 2006
6 *
7 * Major code contributions by Peter Somogyi <psomogyi@gamax.hu>
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 3 of the License, or
12 * (at your option) any later version.
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, see <http://www.gnu.org/licenses/>.
21 */
22
23 #include "replace.h"
24 #include "system/network.h"
25
26 #include <assert.h>
27 #include <talloc.h>
28 #include <tevent.h>
29
30 #include "lib/util/dlinklist.h"
31 #include "lib/util/debug.h"
32
33 #include "common/logging.h"
34
35 #include <infiniband/kern-abi.h>
36 #include <rdma/rdma_cma_abi.h>
37 #include <rdma/rdma_cma.h>
38
39 #include "ibwrapper.h"
40 #include "ibwrapper_internal.h"
41
42 #define IBW_LASTERR_BUFSIZE 512
43 static char ibw_lasterr[IBW_LASTERR_BUFSIZE];
44
45 #define IBW_MAX_SEND_WR 256
46 #define IBW_MAX_RECV_WR 1024
47 #define IBW_RECV_BUFSIZE 256
48 #define IBW_RECV_THRESHOLD (1 * 1024 * 1024)
49
50 static void ibw_event_handler_verbs(struct tevent_context *ev,
51 struct tevent_fd *fde, uint16_t flags, void *private_data);
52 static int ibw_fill_cq(struct ibw_conn *conn);
53 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc);
54 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc);
55 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len);
56
57 static void *ibw_alloc_mr(struct ibw_ctx_priv *pctx, struct ibw_conn_priv *pconn,
58 uint32_t n, struct ibv_mr **ppmr)
59 {
60 void *buf;
61
62 DEBUG(DEBUG_DEBUG, ("ibw_alloc_mr(cmid=%p, n=%u)\n", pconn->cm_id, n));
63 buf = memalign(pctx->pagesize, n);
64 if (!buf) {
65 sprintf(ibw_lasterr, "couldn't allocate memory\n");
66 return NULL;
67 }
68
69 *ppmr = ibv_reg_mr(pconn->pd, buf, n, IBV_ACCESS_LOCAL_WRITE);
70 if (!*ppmr) {
71 sprintf(ibw_lasterr, "couldn't allocate mr\n");
72 free(buf);
73 return NULL;
74 }
75
76 return buf;
77 }
78
79 static void ibw_free_mr(char **ppbuf, struct ibv_mr **ppmr)
80 {
81 DEBUG(DEBUG_DEBUG, ("ibw_free_mr(%p %p)\n", *ppbuf, *ppmr));
82 if (*ppmr!=NULL) {
83 ibv_dereg_mr(*ppmr);
84 *ppmr = NULL;
85 }
86 if (*ppbuf) {
87 free(*ppbuf);
88 *ppbuf = NULL;
89 }
90 }
91
92 static int ibw_init_memory(struct ibw_conn *conn)
93 {
94 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
95 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
96 struct ibw_opts *opts = &pctx->opts;
97 int i;
98 struct ibw_wr *p;
99
100 DEBUG(DEBUG_DEBUG, ("ibw_init_memory(cmid: %p)\n", pconn->cm_id));
101 pconn->buf_send = ibw_alloc_mr(pctx, pconn,
102 opts->max_send_wr * opts->recv_bufsize, &pconn->mr_send);
103 if (!pconn->buf_send) {
104 sprintf(ibw_lasterr, "couldn't allocate work send buf\n");
105 return -1;
106 }
107
108 pconn->buf_recv = ibw_alloc_mr(pctx, pconn,
109 opts->max_recv_wr * opts->recv_bufsize, &pconn->mr_recv);
110 if (!pconn->buf_recv) {
111 sprintf(ibw_lasterr, "couldn't allocate work recv buf\n");
112 return -1;
113 }
114
115 pconn->wr_index = talloc_size(pconn, opts->max_send_wr * sizeof(struct ibw_wr *));
116 assert(pconn->wr_index!=NULL);
117
118 for(i=0; i<opts->max_send_wr; i++) {
119 p = pconn->wr_index[i] = talloc_zero(pconn, struct ibw_wr);
120 p->buf = pconn->buf_send + (i * opts->recv_bufsize);
121 p->wr_id = i;
122
123 DLIST_ADD(pconn->wr_list_avail, p);
124 }
125
126 return 0;
127 }
128
129 static int ibw_ctx_priv_destruct(struct ibw_ctx_priv *pctx)
130 {
131 DEBUG(DEBUG_DEBUG, ("ibw_ctx_priv_destruct(%p)\n", pctx));
132
133 /*
134 * tevent_fd must be removed before the fd is closed
135 */
136 TALLOC_FREE(pctx->cm_channel_event);
137
138 /* destroy cm */
139 if (pctx->cm_channel) {
140 rdma_destroy_event_channel(pctx->cm_channel);
141 pctx->cm_channel = NULL;
142 }
143 if (pctx->cm_id) {
144 rdma_destroy_id(pctx->cm_id);
145 pctx->cm_id = NULL;
146 }
147
148 return 0;
149 }
150
151 static int ibw_ctx_destruct(struct ibw_ctx *ctx)
152 {
153 DEBUG(DEBUG_DEBUG, ("ibw_ctx_destruct(%p)\n", ctx));
154 return 0;
155 }
156
157 static int ibw_conn_priv_destruct(struct ibw_conn_priv *pconn)
158 {
159 DEBUG(DEBUG_DEBUG, ("ibw_conn_priv_destruct(%p, cmid: %p)\n",
160 pconn, pconn->cm_id));
161
162 /* pconn->wr_index is freed by talloc */
163 /* pconn->wr_index[i] are freed by talloc */
164
165 /*
166 * tevent_fd must be removed before the fd is closed
167 */
168 TALLOC_FREE(pconn->verbs_channel_event);
169
170 /* destroy verbs */
171 if (pconn->cm_id!=NULL && pconn->cm_id->qp!=NULL) {
172 rdma_destroy_qp(pconn->cm_id);
173 pconn->cm_id->qp = NULL;
174 }
175
176 if (pconn->cq!=NULL) {
177 ibv_destroy_cq(pconn->cq);
178 pconn->cq = NULL;
179 }
180
181 if (pconn->verbs_channel!=NULL) {
182 ibv_destroy_comp_channel(pconn->verbs_channel);
183 pconn->verbs_channel = NULL;
184 }
185
186 /* free memory regions */
187 ibw_free_mr(&pconn->buf_send, &pconn->mr_send);
188 ibw_free_mr(&pconn->buf_recv, &pconn->mr_recv);
189
190 if (pconn->pd) {
191 ibv_dealloc_pd(pconn->pd);
192 pconn->pd = NULL;
193 DEBUG(DEBUG_DEBUG, ("pconn=%p pd deallocated\n", pconn));
194 }
195
196 if (pconn->cm_id) {
197 rdma_destroy_id(pconn->cm_id);
198 pconn->cm_id = NULL;
199 DEBUG(DEBUG_DEBUG, ("pconn=%p cm_id destroyed\n", pconn));
200 }
201
202 return 0;
203 }
204
205 static int ibw_wr_destruct(struct ibw_wr *wr)
206 {
207 if (wr->buf_large!=NULL)
208 ibw_free_mr(&wr->buf_large, &wr->mr_large);
209 return 0;
210 }
211
212 static int ibw_conn_destruct(struct ibw_conn *conn)
213 {
214 DEBUG(DEBUG_DEBUG, ("ibw_conn_destruct(%p)\n", conn));
215
216 /* important here: ctx is a talloc _parent_ */
217 DLIST_REMOVE(conn->ctx->conn_list, conn);
218 return 0;
219 }
220
221 struct ibw_conn *ibw_conn_new(struct ibw_ctx *ctx, TALLOC_CTX *mem_ctx)
222 {
223 struct ibw_conn *conn;
224 struct ibw_conn_priv *pconn;
225
226 assert(ctx!=NULL);
227
228 conn = talloc_zero(mem_ctx, struct ibw_conn);
229 assert(conn!=NULL);
230 talloc_set_destructor(conn, ibw_conn_destruct);
231
232 pconn = talloc_zero(conn, struct ibw_conn_priv);
233 assert(pconn!=NULL);
234 talloc_set_destructor(pconn, ibw_conn_priv_destruct);
235
236 conn->ctx = ctx;
237 conn->internal = (void *)pconn;
238
239 DLIST_ADD(ctx->conn_list, conn);
240
241 return conn;
242 }
243
244 static int ibw_setup_cq_qp(struct ibw_conn *conn)
245 {
246 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
247 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
248 struct ibv_qp_init_attr init_attr;
249 struct ibv_qp_attr attr;
250 int rc;
251
252 DEBUG(DEBUG_DEBUG, ("ibw_setup_cq_qp(cmid: %p)\n", pconn->cm_id));
253
254 /* init verbs */
255 pconn->verbs_channel = ibv_create_comp_channel(pconn->cm_id->verbs);
256 if (!pconn->verbs_channel) {
257 sprintf(ibw_lasterr, "ibv_create_comp_channel failed %d\n", errno);
258 return -1;
259 }
260 DEBUG(DEBUG_DEBUG, ("created channel %p\n", pconn->verbs_channel));
261
262 pconn->verbs_channel_event = tevent_add_fd(pctx->ectx, NULL, /* not pconn or conn */
263 pconn->verbs_channel->fd, TEVENT_FD_READ, ibw_event_handler_verbs, conn);
264
265 pconn->pd = ibv_alloc_pd(pconn->cm_id->verbs);
266 if (!pconn->pd) {
267 sprintf(ibw_lasterr, "ibv_alloc_pd failed %d\n", errno);
268 return -1;
269 }
270 DEBUG(DEBUG_DEBUG, ("created pd %p\n", pconn->pd));
271
272 /* init mr */
273 if (ibw_init_memory(conn))
274 return -1;
275
276 /* init cq */
277 pconn->cq = ibv_create_cq(pconn->cm_id->verbs,
278 pctx->opts.max_recv_wr + pctx->opts.max_send_wr,
279 conn, pconn->verbs_channel, 0);
280 if (pconn->cq==NULL) {
281 sprintf(ibw_lasterr, "ibv_create_cq failed\n");
282 return -1;
283 }
284
285 rc = ibv_req_notify_cq(pconn->cq, 0);
286 if (rc) {
287 sprintf(ibw_lasterr, "ibv_req_notify_cq failed with %d\n", rc);
288 return rc;
289 }
290
291 /* init qp */
292 memset(&init_attr, 0, sizeof(init_attr));
293 init_attr.cap.max_send_wr = pctx->opts.max_send_wr;
294 init_attr.cap.max_recv_wr = pctx->opts.max_recv_wr;
295 init_attr.cap.max_recv_sge = 1;
296 init_attr.cap.max_send_sge = 1;
297 init_attr.qp_type = IBV_QPT_RC;
298 init_attr.send_cq = pconn->cq;
299 init_attr.recv_cq = pconn->cq;
300
301 rc = rdma_create_qp(pconn->cm_id, pconn->pd, &init_attr);
302 if (rc) {
303 sprintf(ibw_lasterr, "rdma_create_qp failed with %d\n", rc);
304 return rc;
305 }
306 /* elase result is in pconn->cm_id->qp */
307
308 rc = ibv_query_qp(pconn->cm_id->qp, &attr, IBV_QP_PATH_MTU, &init_attr);
309 if (rc) {
310 sprintf(ibw_lasterr, "ibv_query_qp failed with %d\n", rc);
311 return rc;
312 }
313
314 return ibw_fill_cq(conn);
315 }
316
317 static int ibw_refill_cq_recv(struct ibw_conn *conn)
318 {
319 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
320 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
321 int rc;
322 struct ibv_sge list = {
323 .addr = (uintptr_t) NULL, /* filled below */
324 .length = pctx->opts.recv_bufsize,
325 .lkey = pconn->mr_recv->lkey /* always the same */
326 };
327 struct ibv_recv_wr wr = {
328 .wr_id = 0, /* filled below */
329 .sg_list = &list,
330 .num_sge = 1,
331 };
332 struct ibv_recv_wr *bad_wr;
333
334 DEBUG(DEBUG_DEBUG, ("ibw_refill_cq_recv(cmid: %p)\n", pconn->cm_id));
335
336 list.addr = (uintptr_t) pconn->buf_recv + pctx->opts.recv_bufsize * pconn->recv_index;
337 wr.wr_id = pconn->recv_index;
338 pconn->recv_index = (pconn->recv_index + 1) % pctx->opts.max_recv_wr;
339
340 rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
341 if (rc) {
342 sprintf(ibw_lasterr, "refill/ibv_post_recv failed with %d\n", rc);
343 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
344 return -2;
345 }
346
347 return 0;
348 }
349
350 static int ibw_fill_cq(struct ibw_conn *conn)
351 {
352 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
353 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
354 int i, rc;
355 struct ibv_sge list = {
356 .addr = (uintptr_t) NULL, /* filled below */
357 .length = pctx->opts.recv_bufsize,
358 .lkey = pconn->mr_recv->lkey /* always the same */
359 };
360 struct ibv_recv_wr wr = {
361 .wr_id = 0, /* filled below */
362 .sg_list = &list,
363 .num_sge = 1,
364 };
365 struct ibv_recv_wr *bad_wr;
366
367 DEBUG(DEBUG_DEBUG, ("ibw_fill_cq(cmid: %p)\n", pconn->cm_id));
368
369 for(i = pctx->opts.max_recv_wr; i!=0; i--) {
370 list.addr = (uintptr_t) pconn->buf_recv + pctx->opts.recv_bufsize * pconn->recv_index;
371 wr.wr_id = pconn->recv_index;
372 pconn->recv_index = (pconn->recv_index + 1) % pctx->opts.max_recv_wr;
373
374 rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
375 if (rc) {
376 sprintf(ibw_lasterr, "fill/ibv_post_recv failed with %d\n", rc);
377 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
378 return -2;
379 }
380 }
381
382 return 0;
383 }
384
385 static int ibw_manage_connect(struct ibw_conn *conn)
386 {
387 struct rdma_conn_param conn_param;
388 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
389 int rc;
390
391 DEBUG(DEBUG_DEBUG, ("ibw_manage_connect(cmid: %p)\n", pconn->cm_id));
392
393 if (ibw_setup_cq_qp(conn))
394 return -1;
395
396 /* cm connect */
397 memset(&conn_param, 0, sizeof conn_param);
398 conn_param.responder_resources = 1;
399 conn_param.initiator_depth = 1;
400 conn_param.retry_count = 10;
401
402 rc = rdma_connect(pconn->cm_id, &conn_param);
403 if (rc)
404 sprintf(ibw_lasterr, "rdma_connect error %d\n", rc);
405
406 return rc;
407 }
408
409 static void ibw_event_handler_cm(struct tevent_context *ev,
410 struct tevent_fd *fde, uint16_t flags, void *private_data)
411 {
412 int rc;
413 struct ibw_ctx *ctx = talloc_get_type(private_data, struct ibw_ctx);
414 struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
415 struct ibw_conn *conn = NULL;
416 struct ibw_conn_priv *pconn = NULL;
417 struct rdma_cm_id *cma_id = NULL;
418 struct rdma_cm_event *event = NULL;
419
420 assert(ctx!=NULL);
421
422 rc = rdma_get_cm_event(pctx->cm_channel, &event);
423 if (rc) {
424 ctx->state = IBWS_ERROR;
425 event = NULL;
426 sprintf(ibw_lasterr, "rdma_get_cm_event error %d\n", rc);
427 goto error;
428 }
429 cma_id = event->id;
430
431 DEBUG(DEBUG_DEBUG, ("cma_event type %d cma_id %p (%s)\n", event->event, cma_id,
432 (cma_id == pctx->cm_id) ? "parent" : "child"));
433
434 switch (event->event) {
435 case RDMA_CM_EVENT_ADDR_RESOLVED:
436 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_ADDR_RESOLVED\n"));
437 /* continuing from ibw_connect ... */
438 rc = rdma_resolve_route(cma_id, 2000);
439 if (rc) {
440 sprintf(ibw_lasterr, "rdma_resolve_route error %d\n", rc);
441 goto error;
442 }
443 /* continued at RDMA_CM_EVENT_ROUTE_RESOLVED */
444 break;
445
446 case RDMA_CM_EVENT_ROUTE_RESOLVED:
447 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_ROUTE_RESOLVED\n"));
448 /* after RDMA_CM_EVENT_ADDR_RESOLVED: */
449 assert(cma_id->context!=NULL);
450 conn = talloc_get_type(cma_id->context, struct ibw_conn);
451
452 rc = ibw_manage_connect(conn);
453 if (rc)
454 goto error;
455
456 break;
457
458 case RDMA_CM_EVENT_CONNECT_REQUEST:
459 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_CONNECT_REQUEST\n"));
460 ctx->state = IBWS_CONNECT_REQUEST;
461 conn = ibw_conn_new(ctx, ctx);
462 pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
463 pconn->cm_id = cma_id; /* !!! event will be freed but id not */
464 cma_id->context = (void *)conn;
465 DEBUG(DEBUG_DEBUG, ("pconn->cm_id %p\n", pconn->cm_id));
466
467 if (ibw_setup_cq_qp(conn))
468 goto error;
469
470 conn->state = IBWC_INIT;
471 pctx->connstate_func(ctx, conn);
472
473 /* continued at ibw_accept when invoked by the func above */
474 if (!pconn->is_accepted) {
475 rc = rdma_reject(cma_id, NULL, 0);
476 if (rc)
477 DEBUG(DEBUG_ERR, ("rdma_reject failed with rc=%d\n", rc));
478 talloc_free(conn);
479 DEBUG(DEBUG_DEBUG, ("pconn->cm_id %p wasn't accepted\n", pconn->cm_id));
480 }
481
482 /* TODO: clarify whether if it's needed by upper layer: */
483 ctx->state = IBWS_READY;
484 pctx->connstate_func(ctx, NULL);
485
486 /* NOTE: more requests can arrive until RDMA_CM_EVENT_ESTABLISHED ! */
487 break;
488
489 case RDMA_CM_EVENT_ESTABLISHED:
490 /* expected after ibw_accept and ibw_connect[not directly] */
491 DEBUG(DEBUG_INFO, ("ESTABLISHED (conn: %p)\n", cma_id->context));
492 conn = talloc_get_type(cma_id->context, struct ibw_conn);
493 assert(conn!=NULL); /* important assumption */
494
495 DEBUG(DEBUG_DEBUG, ("ibw_setup_cq_qp succeeded (cmid=%p)\n", cma_id));
496
497 /* client conn is up */
498 conn->state = IBWC_CONNECTED;
499
500 /* both ctx and conn have changed */
501 pctx->connstate_func(ctx, conn);
502 break;
503
504 case RDMA_CM_EVENT_ADDR_ERROR:
505 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ADDR_ERROR, error %d\n", event->status);
506 goto error;
507 case RDMA_CM_EVENT_ROUTE_ERROR:
508 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ROUTE_ERROR, error %d\n", event->status);
509 goto error;
510 case RDMA_CM_EVENT_CONNECT_ERROR:
511 sprintf(ibw_lasterr, "RDMA_CM_EVENT_CONNECT_ERROR, error %d\n", event->status);
512 goto error;
513 case RDMA_CM_EVENT_UNREACHABLE:
514 sprintf(ibw_lasterr, "RDMA_CM_EVENT_UNREACHABLE, error %d\n", event->status);
515 goto error;
516 case RDMA_CM_EVENT_REJECTED:
517 sprintf(ibw_lasterr, "RDMA_CM_EVENT_REJECTED, error %d\n", event->status);
518 DEBUG(DEBUG_INFO, ("cm event handler: %s", ibw_lasterr));
519 conn = talloc_get_type(cma_id->context, struct ibw_conn);
520 if (conn) {
521 /* must be done BEFORE connstate */
522 if ((rc=rdma_ack_cm_event(event)))
523 DEBUG(DEBUG_ERR, ("reject/rdma_ack_cm_event failed with %d\n", rc));
524 event = NULL; /* not to touch cma_id or conn */
525 conn->state = IBWC_ERROR;
526 /* it should free the conn */
527 pctx->connstate_func(NULL, conn);
528 }
529 break; /* this is not strictly an error */
530
531 case RDMA_CM_EVENT_DISCONNECTED:
532 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_DISCONNECTED\n"));
533 if ((rc=rdma_ack_cm_event(event)))
534 DEBUG(DEBUG_ERR, ("disc/rdma_ack_cm_event failed with %d\n", rc));
535 event = NULL; /* don't ack more */
536
537 if (cma_id!=pctx->cm_id) {
538 DEBUG(DEBUG_ERR, ("client DISCONNECT event cm_id=%p\n", cma_id));
539 conn = talloc_get_type(cma_id->context, struct ibw_conn);
540 conn->state = IBWC_DISCONNECTED;
541 pctx->connstate_func(NULL, conn);
542 }
543 break;
544
545 case RDMA_CM_EVENT_DEVICE_REMOVAL:
546 sprintf(ibw_lasterr, "cma detected device removal!\n");
547 goto error;
548
549 default:
550 sprintf(ibw_lasterr, "unknown event %d\n", event->event);
551 goto error;
552 }
553
554 if (event!=NULL && (rc=rdma_ack_cm_event(event))) {
555 sprintf(ibw_lasterr, "rdma_ack_cm_event failed with %d\n", rc);
556 goto error;
557 }
558
559 return;
560 error:
561 DEBUG(DEBUG_ERR, ("cm event handler: %s", ibw_lasterr));
562
563 if (event!=NULL) {
564 if (cma_id!=NULL && cma_id!=pctx->cm_id) {
565 conn = talloc_get_type(cma_id->context, struct ibw_conn);
566 if (conn) {
567 conn->state = IBWC_ERROR;
568 pctx->connstate_func(NULL, conn);
569 }
570 } else {
571 ctx->state = IBWS_ERROR;
572 pctx->connstate_func(ctx, NULL);
573 }
574
575 if ((rc=rdma_ack_cm_event(event))!=0) {
576 DEBUG(DEBUG_ERR, ("rdma_ack_cm_event failed with %d\n", rc));
577 }
578 }
579
580 return;
581 }
582
583 static void ibw_event_handler_verbs(struct tevent_context *ev,
584 struct tevent_fd *fde, uint16_t flags, void *private_data)
585 {
586 struct ibw_conn *conn = talloc_get_type(private_data, struct ibw_conn);
587 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
588 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
589
590 struct ibv_wc wc;
591 int rc;
592 struct ibv_cq *ev_cq;
593 void *ev_ctx;
594
595 DEBUG(DEBUG_DEBUG, ("ibw_event_handler_verbs(%u)\n", (uint32_t)flags));
596
597 /* TODO: check whether if it's good to have more channels here... */
598 rc = ibv_get_cq_event(pconn->verbs_channel, &ev_cq, &ev_ctx);
599 if (rc) {
600 sprintf(ibw_lasterr, "Failed to get cq_event with %d\n", rc);
601 goto error;
602 }
603 if (ev_cq != pconn->cq) {
604 sprintf(ibw_lasterr, "ev_cq(%p) != pconn->cq(%p)\n", ev_cq, pconn->cq);
605 goto error;
606 }
607 rc = ibv_req_notify_cq(pconn->cq, 0);
608 if (rc) {
609 sprintf(ibw_lasterr, "Couldn't request CQ notification (%d)\n", rc);
610 goto error;
611 }
612
613 while((rc=ibv_poll_cq(pconn->cq, 1, &wc))==1) {
614 if (wc.status) {
615 sprintf(ibw_lasterr, "cq completion failed status=%d, opcode=%d, rc=%d\n",
616 wc.status, wc.opcode, rc);
617 goto error;
618 }
619
620 switch(wc.opcode) {
621 case IBV_WC_SEND:
622 DEBUG(DEBUG_DEBUG, ("send completion\n"));
623 if (ibw_wc_send(conn, &wc))
624 goto error;
625 break;
626
627 case IBV_WC_RDMA_WRITE:
628 DEBUG(DEBUG_DEBUG, ("rdma write completion\n"));
629 break;
630
631 case IBV_WC_RDMA_READ:
632 DEBUG(DEBUG_DEBUG, ("rdma read completion\n"));
633 break;
634
635 case IBV_WC_RECV:
636 DEBUG(DEBUG_DEBUG, ("recv completion\n"));
637 if (ibw_wc_recv(conn, &wc))
638 goto error;
639 break;
640
641 default:
642 sprintf(ibw_lasterr, "unknown completion %d\n", wc.opcode);
643 goto error;
644 }
645 }
646 if (rc!=0) {
647 sprintf(ibw_lasterr, "ibv_poll_cq error %d\n", rc);
648 goto error;
649 }
650
651 ibv_ack_cq_events(pconn->cq, 1);
652
653 return;
654 error:
655 ibv_ack_cq_events(pconn->cq, 1);
656
657 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
658
659 if (conn->state!=IBWC_ERROR) {
660 conn->state = IBWC_ERROR;
661 pctx->connstate_func(NULL, conn);
662 }
663 }
664
665 static int ibw_process_queue(struct ibw_conn *conn)
666 {
667 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
668 struct ibw_ctx_priv *pctx;
669 struct ibw_wr *p;
670 int rc;
671 uint32_t msg_size;
672
673 if (pconn->queue==NULL)
674 return 0; /* NOP */
675
676 p = pconn->queue;
677
678 /* we must have at least 1 fragment to send */
679 assert(p->queued_ref_cnt>0);
680 p->queued_ref_cnt--;
681
682 pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
683 msg_size = (p->queued_ref_cnt) ? pctx->opts.recv_bufsize : p->queued_rlen;
684
685 assert(p->queued_msg!=NULL);
686 assert(msg_size!=0);
687
688 DEBUG(DEBUG_DEBUG, ("ibw_process_queue refcnt=%d msgsize=%u\n",
689 p->queued_ref_cnt, msg_size));
690
691 rc = ibw_send_packet(conn, p->queued_msg, p, msg_size);
692
693 /* was this the last fragment? */
694 if (p->queued_ref_cnt) {
695 p->queued_msg += pctx->opts.recv_bufsize;
696 } else {
697 DLIST_REMOVE2(pconn->queue, p, qprev, qnext);
698 p->queued_msg = NULL;
699 }
700
701 return rc;
702 }
703
704 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc)
705 {
706 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
707 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
708 struct ibw_wr *p;
709 int send_index;
710
711 DEBUG(DEBUG_DEBUG, ("ibw_wc_send(cmid: %p, wr_id: %u, bl: %u)\n",
712 pconn->cm_id, (uint32_t)wc->wr_id, (uint32_t)wc->byte_len));
713
714 assert(pconn->cm_id->qp->qp_num==wc->qp_num);
715 assert(wc->wr_id >= pctx->opts.max_recv_wr);
716 send_index = wc->wr_id - pctx->opts.max_recv_wr;
717 pconn->wr_sent--;
718
719 if (send_index < pctx->opts.max_send_wr) {
720 DEBUG(DEBUG_DEBUG, ("ibw_wc_send#1 %u\n", (int)wc->wr_id));
721 p = pconn->wr_index[send_index];
722 if (p->buf_large!=NULL) {
723 if (p->ref_cnt) {
724 /* awaiting more of it... */
725 p->ref_cnt--;
726 } else {
727 ibw_free_mr(&p->buf_large, &p->mr_large);
728 DLIST_REMOVE(pconn->wr_list_used, p);
729 DLIST_ADD(pconn->wr_list_avail, p);
730 }
731 } else { /* nasty - but necessary */
732 DLIST_REMOVE(pconn->wr_list_used, p);
733 DLIST_ADD(pconn->wr_list_avail, p);
734 }
735 } else { /* "extra" request - not optimized */
736 DEBUG(DEBUG_DEBUG, ("ibw_wc_send#2 %u\n", (int)wc->wr_id));
737 for(p=pconn->extra_sent; p!=NULL; p=p->next)
738 if ((p->wr_id + pctx->opts.max_recv_wr)==(int)wc->wr_id)
739 break;
740 if (p==NULL) {
741 sprintf(ibw_lasterr, "failed to find wr_id %d\n", (int)wc->wr_id);
742 return -1;
743 }
744 if (p->ref_cnt) {
745 p->ref_cnt--;
746 } else {
747 ibw_free_mr(&p->buf_large, &p->mr_large);
748 DLIST_REMOVE(pconn->extra_sent, p);
749 DLIST_ADD(pconn->extra_avail, p);
750 }
751 }
752
753 return ibw_process_queue(conn);
754 }
755
756 static int ibw_append_to_part(struct ibw_conn_priv *pconn,
757 struct ibw_part *part, char **pp, uint32_t add_len, int info)
758 {
759 DEBUG(DEBUG_DEBUG, ("ibw_append_to_part: cmid=%p, (bs=%u, len=%u, tr=%u), al=%u, i=%u\n",
760 pconn->cm_id, part->bufsize, part->len, part->to_read, add_len, info));
761
762 /* allocate more if necessary - it's an "evergrowing" buffer... */
763 if (part->len + add_len > part->bufsize) {
764 if (part->buf==NULL) {
765 assert(part->len==0);
766 part->buf = talloc_size(pconn, add_len);
767 if (part->buf==NULL) {
768 sprintf(ibw_lasterr, "recv talloc_size error (%u) #%d\n",
769 add_len, info);
770 return -1;
771 }
772 part->bufsize = add_len;
773 } else {
774 part->buf = talloc_realloc_size(pconn,
775 part->buf, part->len + add_len);
776 if (part->buf==NULL) {
777 sprintf(ibw_lasterr, "recv realloc error (%u + %u) #%d\n",
778 part->len, add_len, info);
779 return -1;
780 }
781 }
782 part->bufsize = part->len + add_len;
783 }
784
785 /* consume pp */
786 memcpy(part->buf + part->len, *pp, add_len);
787 *pp += add_len;
788 part->len += add_len;
789 part->to_read -= add_len;
790
791 return 0;
792 }
793
794 static int ibw_wc_mem_threshold(struct ibw_conn_priv *pconn,
795 struct ibw_part *part, uint32_t threshold)
796 {
797 DEBUG(DEBUG_DEBUG, ("ibw_wc_mem_threshold: cmid=%p, (bs=%u, len=%u, tr=%u), thr=%u\n",
798 pconn->cm_id, part->bufsize, part->len, part->to_read, threshold));
799
800 if (part->bufsize > threshold) {
801 DEBUG(DEBUG_DEBUG, ("ibw_wc_mem_threshold: cmid=%p, %u > %u\n",
802 pconn->cm_id, part->bufsize, threshold));
803 talloc_free(part->buf);
804 part->buf = talloc_size(pconn, threshold);
805 if (part->buf==NULL) {
806 sprintf(ibw_lasterr, "talloc_size failed\n");
807 return -1;
808 }
809 part->bufsize = threshold;
810 }
811 return 0;
812 }
813
814 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc)
815 {
816 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
817 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
818 struct ibw_part *part = &pconn->part;
819 char *p;
820 uint32_t remain = wc->byte_len;
821
822 DEBUG(DEBUG_DEBUG, ("ibw_wc_recv: cmid=%p, wr_id: %u, bl: %u\n",
823 pconn->cm_id, (uint32_t)wc->wr_id, remain));
824
825 assert(pconn->cm_id->qp->qp_num==wc->qp_num);
826 assert((int)wc->wr_id < pctx->opts.max_recv_wr);
827 assert(wc->byte_len <= pctx->opts.recv_bufsize);
828
829 p = pconn->buf_recv + ((int)wc->wr_id * pctx->opts.recv_bufsize);
830
831 while(remain) {
832 /* here always true: (part->len!=0 && part->to_read!=0) ||
833 (part->len==0 && part->to_read==0) */
834 if (part->len) { /* is there a partial msg to be continued? */
835 int read_len = (part->to_read<=remain) ? part->to_read : remain;
836 if (ibw_append_to_part(pconn, part, &p, read_len, 421))
837 goto error;
838 remain -= read_len;
839
840 if (part->len<=sizeof(uint32_t) && part->to_read==0) {
841 assert(part->len==sizeof(uint32_t));
842 /* set it again now... */
843 part->to_read = *((uint32_t *)(part->buf)); /* TODO: ntohl */
844 if (part->to_read<sizeof(uint32_t)) {
845 sprintf(ibw_lasterr, "got msglen=%u #2\n", part->to_read);
846 goto error;
847 }
848 part->to_read -= sizeof(uint32_t); /* it's already read */
849 }
850
851 if (part->to_read==0) {
852 if (pctx->receive_func(conn, part->buf, part->len) != 0) {
853 goto error;
854 }
855 part->len = 0; /* tells not having partial data (any more) */
856 if (ibw_wc_mem_threshold(pconn, part, pctx->opts.recv_threshold))
857 goto error;
858 }
859 } else {
860 if (remain>=sizeof(uint32_t)) {
861 uint32_t msglen = *(uint32_t *)p; /* TODO: ntohl */
862 if (msglen<sizeof(uint32_t)) {
863 sprintf(ibw_lasterr, "got msglen=%u\n", msglen);
864 goto error;
865 }
866
867 /* mostly awaited case: */
868 if (msglen<=remain) {
869 if (pctx->receive_func(conn, p, msglen) != 0) {
870 goto error;
871 }
872 p += msglen;
873 remain -= msglen;
874 } else {
875 part->to_read = msglen;
876 /* part->len is already 0 */
877 if (ibw_append_to_part(pconn, part, &p, remain, 422))
878 goto error;
879 remain = 0; /* to be continued ... */
880 /* part->to_read > 0 here */
881 }
882 } else { /* edge case: */
883 part->to_read = sizeof(uint32_t);
884 /* part->len is already 0 */
885 if (ibw_append_to_part(pconn, part, &p, remain, 423))
886 goto error;
887 remain = 0;
888 /* part->to_read > 0 here */
889 }
890 }
891 } /* <remain> is always decreased at least by 1 */
892
893 if (ibw_refill_cq_recv(conn))
894 goto error;
895
896 return 0;
897
898 error:
899 DEBUG(DEBUG_ERR, ("ibw_wc_recv error: %s", ibw_lasterr));
900 return -1;
901 }
902
903 static int ibw_process_init_attrs(struct ibw_initattr *attr, int nattr, struct ibw_opts *opts)
904 {
905 int i;
906 const char *name, *value;
907
908 DEBUG(DEBUG_DEBUG, ("ibw_process_init_attrs: nattr: %d\n", nattr));
909
910 opts->max_send_wr = IBW_MAX_SEND_WR;
911 opts->max_recv_wr = IBW_MAX_RECV_WR;
912 opts->recv_bufsize = IBW_RECV_BUFSIZE;
913 opts->recv_threshold = IBW_RECV_THRESHOLD;
914
915 for(i=0; i<nattr; i++) {
916 name = attr[i].name;
917 value = attr[i].value;
918
919 assert(name!=NULL && value!=NULL);
920 if (strcmp(name, "max_send_wr")==0)
921 opts->max_send_wr = atoi(value);
922 else if (strcmp(name, "max_recv_wr")==0)
923 opts->max_recv_wr = atoi(value);
924 else if (strcmp(name, "recv_bufsize")==0)
925 opts->recv_bufsize = atoi(value);
926 else if (strcmp(name, "recv_threshold")==0)
927 opts->recv_threshold = atoi(value);
928 else {
929 sprintf(ibw_lasterr, "ibw_init: unknown name %s\n", name);
930 return -1;
931 }
932 }
933 return 0;
934 }
935
936 struct ibw_ctx *ibw_init(struct ibw_initattr *attr, int nattr,
937 void *ctx_userdata,
938 ibw_connstate_fn_t ibw_connstate,
939 ibw_receive_fn_t ibw_receive,
940 struct tevent_context *ectx)
941 {
942 struct ibw_ctx *ctx = talloc_zero(NULL, struct ibw_ctx);
943 struct ibw_ctx_priv *pctx;
944 int rc;
945
946 DEBUG(DEBUG_DEBUG, ("ibw_init(ctx_userdata: %p, ectx: %p)\n", ctx_userdata, ectx));
947
948 /* initialize basic data structures */
949 memset(ibw_lasterr, 0, IBW_LASTERR_BUFSIZE);
950
951 assert(ctx!=NULL);
952 ibw_lasterr[0] = '\0';
953 talloc_set_destructor(ctx, ibw_ctx_destruct);
954 ctx->ctx_userdata = ctx_userdata;
955
956 pctx = talloc_zero(ctx, struct ibw_ctx_priv);
957 talloc_set_destructor(pctx, ibw_ctx_priv_destruct);
958 ctx->internal = (void *)pctx;
959 assert(pctx!=NULL);
960
961 pctx->connstate_func = ibw_connstate;
962 pctx->receive_func = ibw_receive;
963
964 pctx->ectx = ectx;
965
966 /* process attributes */
967 if (ibw_process_init_attrs(attr, nattr, &pctx->opts))
968 goto cleanup;
969
970 /* init cm */
971 pctx->cm_channel = rdma_create_event_channel();
972 if (!pctx->cm_channel) {
973 sprintf(ibw_lasterr, "rdma_create_event_channel error %d\n", errno);
974 goto cleanup;
975 }
976
977 pctx->cm_channel_event = tevent_add_fd(pctx->ectx, pctx,
978 pctx->cm_channel->fd, TEVENT_FD_READ, ibw_event_handler_cm, ctx);
979
980 #if RDMA_USER_CM_MAX_ABI_VERSION >= 2
981 rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx, RDMA_PS_TCP);
982 #else
983 rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx);
984 #endif
985 if (rc) {
986 rc = errno;
987 sprintf(ibw_lasterr, "rdma_create_id error %d\n", rc);
988 goto cleanup;
989 }
990 DEBUG(DEBUG_DEBUG, ("created cm_id %p\n", pctx->cm_id));
991
992 pctx->pagesize = sysconf(_SC_PAGESIZE);
993
994 return ctx;
995 /* don't put code here */
996 cleanup:
997 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
998
999 if (ctx)
1000 talloc_free(ctx);
1001
1002 return NULL;
1003 }
1004
1005 int ibw_stop(struct ibw_ctx *ctx)
1006 {
1007 struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1008 struct ibw_conn *p;
1009
1010 DEBUG(DEBUG_DEBUG, ("ibw_stop\n"));
1011
1012 for(p=ctx->conn_list; p!=NULL; p=p->next) {
1013 if (p->state==IBWC_ERROR || p->state==IBWC_CONNECTED) {
1014 if (ibw_disconnect(p))
1015 return -1;
1016 }
1017 }
1018
1019 ctx->state = IBWS_STOPPED;
1020 pctx->connstate_func(ctx, NULL);
1021
1022 return 0;
1023 }
1024
1025 int ibw_bind(struct ibw_ctx *ctx, struct sockaddr_in *my_addr)
1026 {
1027 struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1028 int rc;
1029
1030 DEBUG(DEBUG_DEBUG, ("ibw_bind: addr=%s, port=%u\n",
1031 inet_ntoa(my_addr->sin_addr), ntohs(my_addr->sin_port)));
1032 rc = rdma_bind_addr(pctx->cm_id, (struct sockaddr *) my_addr);
1033 if (rc) {
1034 sprintf(ibw_lasterr, "rdma_bind_addr error %d\n", rc);
1035 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
1036 return rc;
1037 }
1038 DEBUG(DEBUG_DEBUG, ("rdma_bind_addr successful\n"));
1039
1040 return 0;
1041 }
1042
1043 int ibw_listen(struct ibw_ctx *ctx, int backlog)
1044 {
1045 struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
1046 int rc;
1047
1048 DEBUG(DEBUG_DEBUG, ("ibw_listen\n"));
1049 rc = rdma_listen(pctx->cm_id, backlog);
1050 if (rc) {
1051 sprintf(ibw_lasterr, "rdma_listen failed: %d\n", rc);
1052 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
1053 return rc;
1054 }
1055
1056 return 0;
1057 }
1058
1059 int ibw_accept(struct ibw_ctx *ctx, struct ibw_conn *conn, void *conn_userdata)
1060 {
1061 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1062 struct rdma_conn_param conn_param;
1063 int rc;
1064
1065 DEBUG(DEBUG_DEBUG, ("ibw_accept: cmid=%p\n", pconn->cm_id));
1066 conn->conn_userdata = conn_userdata;
1067
1068 memset(&conn_param, 0, sizeof(struct rdma_conn_param));
1069 conn_param.responder_resources = 1;
1070 conn_param.initiator_depth = 1;
1071 rc = rdma_accept(pconn->cm_id, &conn_param);
1072 if (rc) {
1073 sprintf(ibw_lasterr, "rdma_accept failed %d\n", rc);
1074 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
1075 return -1;;
1076 }
1077
1078 pconn->is_accepted = 1;
1079
1080 /* continued at RDMA_CM_EVENT_ESTABLISHED */
1081
1082 return 0;
1083 }
1084
1085 int ibw_connect(struct ibw_conn *conn, struct sockaddr_in *serv_addr, void *conn_userdata)
1086 {
1087 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1088 struct ibw_conn_priv *pconn = NULL;
1089 int rc;
1090
1091 assert(conn!=NULL);
1092
1093 conn->conn_userdata = conn_userdata;
1094 pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1095 DEBUG(DEBUG_DEBUG, ("ibw_connect: addr=%s, port=%u\n", inet_ntoa(serv_addr->sin_addr),
1096 ntohs(serv_addr->sin_port)));
1097
1098 /* clean previous - probably half - initialization */
1099 if (ibw_conn_priv_destruct(pconn)) {
1100 DEBUG(DEBUG_ERR, ("ibw_connect/ibw_pconn_destruct failed for cm_id=%p\n", pconn->cm_id));
1101 return -1;
1102 }
1103
1104 /* init cm */
1105 #if RDMA_USER_CM_MAX_ABI_VERSION >= 2
1106 rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn, RDMA_PS_TCP);
1107 #else
1108 rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn);
1109 #endif
1110 if (rc) {
1111 rc = errno;
1112 sprintf(ibw_lasterr, "ibw_connect/rdma_create_id error %d\n", rc);
1113 talloc_free(conn);
1114 return -1;
1115 }
1116 DEBUG(DEBUG_DEBUG, ("ibw_connect: rdma_create_id succeeded, cm_id=%p\n", pconn->cm_id));
1117
1118 rc = rdma_resolve_addr(pconn->cm_id, NULL, (struct sockaddr *) serv_addr, 2000);
1119 if (rc) {
1120 sprintf(ibw_lasterr, "rdma_resolve_addr error %d\n", rc);
1121 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
1122 talloc_free(conn);
1123 return -1;
1124 }
1125
1126 /* continued at RDMA_CM_EVENT_ADDR_RESOLVED */
1127
1128 return 0;
1129 }
1130
1131 int ibw_disconnect(struct ibw_conn *conn)
1132 {
1133 int rc;
1134 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1135
1136 DEBUG(DEBUG_DEBUG, ("ibw_disconnect: cmid=%p\n", pconn->cm_id));
1137
1138 assert(pconn!=NULL);
1139
1140 switch(conn->state) {
1141 case IBWC_ERROR:
1142 ibw_conn_priv_destruct(pconn); /* do this here right now */
1143 break;
1144 case IBWC_CONNECTED:
1145 rc = rdma_disconnect(pconn->cm_id);
1146 if (rc) {
1147 sprintf(ibw_lasterr, "ibw_disconnect failed with %d\n", rc);
1148 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
1149 return rc;
1150 }
1151 break;
1152 default:
1153 DEBUG(DEBUG_DEBUG, ("invalid state for disconnect: %d\n", conn->state));
1154 break;
1155 }
1156
1157 return 0;
1158 }
1159
1160 int ibw_alloc_send_buf(struct ibw_conn *conn, void **buf, void **key, uint32_t len)
1161 {
1162 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1163 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1164 struct ibw_wr *p = pconn->wr_list_avail;
1165
1166 if (p!=NULL) {
1167 DEBUG(DEBUG_DEBUG, ("ibw_alloc_send_buf#1: cmid=%p, len=%d\n", pconn->cm_id, len));
1168
1169 DLIST_REMOVE(pconn->wr_list_avail, p);
1170 DLIST_ADD(pconn->wr_list_used, p);
1171
1172 if (len <= pctx->opts.recv_bufsize) {
1173 *buf = (void *)p->buf;
1174 } else {
1175 p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1176 if (p->buf_large==NULL) {
1177 sprintf(ibw_lasterr, "ibw_alloc_mr#1 failed\n");
1178 goto error;
1179 }
1180 *buf = (void *)p->buf_large;
1181 }
1182 /* p->wr_id is already filled in ibw_init_memory */
1183 } else {
1184 DEBUG(DEBUG_DEBUG, ("ibw_alloc_send_buf#2: cmid=%p, len=%d\n", pconn->cm_id, len));
1185 /* not optimized */
1186 p = pconn->extra_avail;
1187 if (!p) {
1188 p = pconn->extra_avail = talloc_zero(pconn, struct ibw_wr);
1189 talloc_set_destructor(p, ibw_wr_destruct);
1190 if (p==NULL) {
1191 sprintf(ibw_lasterr, "talloc_zero failed (emax: %u)\n", pconn->extra_max);
1192 goto error;
1193 }
1194 p->wr_id = pctx->opts.max_send_wr + pconn->extra_max;
1195 pconn->extra_max++;
1196 switch(pconn->extra_max) {
1197 case 1: DEBUG(DEBUG_INFO, ("warning: queue performed\n")); break;
1198 case 10: DEBUG(DEBUG_INFO, ("warning: queue reached 10\n")); break;
1199 case 100: DEBUG(DEBUG_INFO, ("warning: queue reached 100\n")); break;
1200 case 1000: DEBUG(DEBUG_INFO, ("warning: queue reached 1000\n")); break;
1201 default: break;
1202 }
1203 }
1204
1205 p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1206 if (p->buf_large==NULL) {
1207 sprintf(ibw_lasterr, "ibw_alloc_mr#2 failed\n");
1208 goto error;
1209 }
1210 *buf = (void *)p->buf_large;
1211
1212 DLIST_REMOVE(pconn->extra_avail, p);
1213 /* we don't have prepared index for this, so that
1214 * we will have to find this by wr_id later on */
1215 DLIST_ADD(pconn->extra_sent, p);
1216 }
1217
1218 *key = (void *)p;
1219
1220 return 0;
1221 error:
1222 DEBUG(DEBUG_ERR, ("ibw_alloc_send_buf error: %s", ibw_lasterr));
1223 return -1;
1224 }
1225
1226
1227 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len)
1228 {
1229 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1230 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1231 int rc;
1232
1233 /* can we send it right now? */
1234 if (pconn->wr_sent<pctx->opts.max_send_wr) {
1235 struct ibv_send_wr *bad_wr;
1236 struct ibv_sge list = {
1237 .addr = (uintptr_t)buf,
1238 .length = len,
1239 .lkey = pconn->mr_send->lkey
1240 };
1241 struct ibv_send_wr wr = {
1242 .wr_id = p->wr_id + pctx->opts.max_recv_wr,
1243 .sg_list = &list,
1244 .num_sge = 1,
1245 .opcode = IBV_WR_SEND,
1246 .send_flags = IBV_SEND_SIGNALED,
1247 };
1248
1249 if (p->buf_large==NULL) {
1250 DEBUG(DEBUG_DEBUG, ("ibw_send#normal(cmid: %p, wrid: %u, n: %d)\n",
1251 pconn->cm_id, (uint32_t)wr.wr_id, len));
1252 } else {
1253 DEBUG(DEBUG_DEBUG, ("ibw_send#large(cmid: %p, wrid: %u, n: %d)\n",
1254 pconn->cm_id, (uint32_t)wr.wr_id, len));
1255 list.lkey = p->mr_large->lkey;
1256 }
1257
1258 rc = ibv_post_send(pconn->cm_id->qp, &wr, &bad_wr);
1259 if (rc) {
1260 sprintf(ibw_lasterr, "ibv_post_send error %d (%d)\n",
1261 rc, pconn->wr_sent);
1262 goto error;
1263 }
1264
1265 pconn->wr_sent++;
1266
1267 return rc;
1268 } /* else put the request into our own queue: */
1269
1270 DEBUG(DEBUG_DEBUG, ("ibw_send#queued(cmid: %p, len: %u)\n", pconn->cm_id, len));
1271
1272 /* TODO: clarify how to continue when state==IBWC_STOPPED */
1273
1274 /* to be sent by ibw_wc_send */
1275 /* regardless "normal" or [a part of] "large" packet */
1276 if (!p->queued_ref_cnt) {
1277 DLIST_ADD_END2(pconn->queue, p, struct ibw_wr *,
1278 qprev, qnext); /* TODO: optimize */
1279 p->queued_msg = buf;
1280 }
1281 p->queued_ref_cnt++;
1282 p->queued_rlen = len; /* last wins; see ibw_wc_send */
1283
1284 return 0;
1285 error:
1286 DEBUG(DEBUG_ERR, ("%s", ibw_lasterr));
1287 return -1;
1288 }
1289
1290 int ibw_send(struct ibw_conn *conn, void *buf, void *key, uint32_t len)
1291 {
1292 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1293 struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1294 int rc;
1295
1296 assert(len>=sizeof(uint32_t));
1297 assert((*((uint32_t *)buf)==len)); /* TODO: htonl */
1298
1299 if (len > pctx->opts.recv_bufsize) {
1300 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1301 int rlen = len;
1302 char *packet = (char *)buf;
1303 uint32_t recv_bufsize = pctx->opts.recv_bufsize;
1304
1305 DEBUG(DEBUG_DEBUG, ("ibw_send#frag(cmid: %p, buf: %p, len: %u)\n",
1306 pconn->cm_id, buf, len));
1307
1308 /* single threaded => no race here: */
1309 assert(p->ref_cnt==0);
1310 while(rlen > recv_bufsize) {
1311 rc = ibw_send_packet(conn, packet, p, recv_bufsize);
1312 if (rc)
1313 return rc;
1314 packet += recv_bufsize;
1315 rlen -= recv_bufsize;
1316 p->ref_cnt++; /* not good to have it in ibw_send_packet */
1317 }
1318 if (rlen) {
1319 rc = ibw_send_packet(conn, packet, p, rlen);
1320 p->ref_cnt++; /* not good to have it in ibw_send_packet */
1321 }
1322 p->ref_cnt--; /* for the same handling */
1323 } else {
1324 assert(p->ref_cnt==0);
1325 assert(p->queued_ref_cnt==0);
1326
1327 rc = ibw_send_packet(conn, buf, p, len);
1328 }
1329 return rc;
1330 }
1331
1332 int ibw_cancel_send_buf(struct ibw_conn *conn, void *buf, void *key)
1333 {
1334 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1335 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1336 struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1337
1338 assert(p!=NULL);
1339 assert(buf!=NULL);
1340 assert(conn!=NULL);
1341
1342 if (p->buf_large!=NULL)
1343 ibw_free_mr(&p->buf_large, &p->mr_large);
1344
1345 /* parallel case */
1346 if (p->wr_id < pctx->opts.max_send_wr) {
1347 DEBUG(DEBUG_DEBUG, ("ibw_cancel_send_buf#1 %u", (int)p->wr_id));
1348 DLIST_REMOVE(pconn->wr_list_used, p);
1349 DLIST_ADD(pconn->wr_list_avail, p);
1350 } else { /* "extra" packet */
1351 DEBUG(DEBUG_DEBUG, ("ibw_cancel_send_buf#2 %u", (int)p->wr_id));
1352 DLIST_REMOVE(pconn->extra_sent, p);
1353 DLIST_ADD(pconn->extra_avail, p);
1354 }
1355
1356 return 0;
1357 }
1358
1359 const char *ibw_getLastError(void)
1360 {
1361 return ibw_lasterr;
1362 }
Samba GIT mirror