2 * Unix SMB/CIFS implementation.
3 * Wrap Infiniband calls.
5 * Copyright (C) Sven Oehme <oehmes@de.ibm.com> 2006
7 * Major code contributions by Peter Somogyi <psomogyi@gamax.hu>
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.
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.
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/>.
24 #include "system/network.h"
30 #include "lib/util/dlinklist.h"
31 #include "lib/util/debug.h"
33 #include "common/logging.h"
35 #include <infiniband/kern-abi.h>
36 #include <rdma/rdma_cma_abi.h>
37 #include <rdma/rdma_cma.h>
39 #include "ibwrapper.h"
40 #include "ibwrapper_internal.h"
42 #define IBW_LASTERR_BUFSIZE 512
43 static char ibw_lasterr[IBW_LASTERR_BUFSIZE];
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)
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);
57 static void *ibw_alloc_mr(struct ibw_ctx_priv *pctx, struct ibw_conn_priv *pconn,
58 uint32_t n, struct ibv_mr **ppmr)
62 DEBUG(DEBUG_DEBUG, ("ibw_alloc_mr(cmid=%p, n=%u)\n", pconn->cm_id, n));
63 buf = memalign(pctx->pagesize, n);
65 sprintf(ibw_lasterr, "couldn't allocate memory\n");
69 *ppmr = ibv_reg_mr(pconn->pd, buf, n, IBV_ACCESS_LOCAL_WRITE);
71 sprintf(ibw_lasterr, "couldn't allocate mr\n");
79 static void ibw_free_mr(char **ppbuf, struct ibv_mr **ppmr)
81 DEBUG(DEBUG_DEBUG, ("ibw_free_mr(%p %p)\n", *ppbuf, *ppmr));
92 static int ibw_init_memory(struct ibw_conn *conn)
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;
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");
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");
115 pconn->wr_index = talloc_size(pconn, opts->max_send_wr * sizeof(struct ibw_wr *));
116 assert(pconn->wr_index!=NULL);
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);
123 DLIST_ADD(pconn->wr_list_avail, p);
129 static int ibw_ctx_priv_destruct(struct ibw_ctx_priv *pctx)
131 DEBUG(DEBUG_DEBUG, ("ibw_ctx_priv_destruct(%p)\n", pctx));
134 * tevent_fd must be removed before the fd is closed
136 TALLOC_FREE(pctx->cm_channel_event);
139 if (pctx->cm_channel) {
140 rdma_destroy_event_channel(pctx->cm_channel);
141 pctx->cm_channel = NULL;
144 rdma_destroy_id(pctx->cm_id);
151 static int ibw_ctx_destruct(struct ibw_ctx *ctx)
153 DEBUG(DEBUG_DEBUG, ("ibw_ctx_destruct(%p)\n", ctx));
157 static int ibw_conn_priv_destruct(struct ibw_conn_priv *pconn)
159 DEBUG(DEBUG_DEBUG, ("ibw_conn_priv_destruct(%p, cmid: %p)\n",
160 pconn, pconn->cm_id));
162 /* pconn->wr_index is freed by talloc */
163 /* pconn->wr_index[i] are freed by talloc */
166 * tevent_fd must be removed before the fd is closed
168 TALLOC_FREE(pconn->verbs_channel_event);
171 if (pconn->cm_id!=NULL && pconn->cm_id->qp!=NULL) {
172 rdma_destroy_qp(pconn->cm_id);
173 pconn->cm_id->qp = NULL;
176 if (pconn->cq!=NULL) {
177 ibv_destroy_cq(pconn->cq);
181 if (pconn->verbs_channel!=NULL) {
182 ibv_destroy_comp_channel(pconn->verbs_channel);
183 pconn->verbs_channel = NULL;
186 /* free memory regions */
187 ibw_free_mr(&pconn->buf_send, &pconn->mr_send);
188 ibw_free_mr(&pconn->buf_recv, &pconn->mr_recv);
191 ibv_dealloc_pd(pconn->pd);
193 DEBUG(DEBUG_DEBUG, ("pconn=%p pd deallocated\n", pconn));
197 rdma_destroy_id(pconn->cm_id);
199 DEBUG(DEBUG_DEBUG, ("pconn=%p cm_id destroyed\n", pconn));
205 static int ibw_wr_destruct(struct ibw_wr *wr)
207 if (wr->buf_large!=NULL)
208 ibw_free_mr(&wr->buf_large, &wr->mr_large);
212 static int ibw_conn_destruct(struct ibw_conn *conn)
214 DEBUG(DEBUG_DEBUG, ("ibw_conn_destruct(%p)\n", conn));
216 /* important here: ctx is a talloc _parent_ */
217 DLIST_REMOVE(conn->ctx->conn_list, conn);
221 struct ibw_conn *ibw_conn_new(struct ibw_ctx *ctx, TALLOC_CTX *mem_ctx)
223 struct ibw_conn *conn;
224 struct ibw_conn_priv *pconn;
228 conn = talloc_zero(mem_ctx, struct ibw_conn);
230 talloc_set_destructor(conn, ibw_conn_destruct);
232 pconn = talloc_zero(conn, struct ibw_conn_priv);
234 talloc_set_destructor(pconn, ibw_conn_priv_destruct);
237 conn->internal = (void *)pconn;
239 DLIST_ADD(ctx->conn_list, conn);
244 static int ibw_setup_cq_qp(struct ibw_conn *conn)
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;
252 DEBUG(DEBUG_DEBUG, ("ibw_setup_cq_qp(cmid: %p)\n", pconn->cm_id));
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);
260 DEBUG(DEBUG_DEBUG, ("created channel %p\n", pconn->verbs_channel));
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);
265 pconn->pd = ibv_alloc_pd(pconn->cm_id->verbs);
267 sprintf(ibw_lasterr, "ibv_alloc_pd failed %d\n", errno);
270 DEBUG(DEBUG_DEBUG, ("created pd %p\n", pconn->pd));
273 if (ibw_init_memory(conn))
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");
285 rc = ibv_req_notify_cq(pconn->cq, 0);
287 sprintf(ibw_lasterr, "ibv_req_notify_cq failed with %d\n", rc);
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;
301 rc = rdma_create_qp(pconn->cm_id, pconn->pd, &init_attr);
303 sprintf(ibw_lasterr, "rdma_create_qp failed with %d\n", rc);
306 /* elase result is in pconn->cm_id->qp */
308 rc = ibv_query_qp(pconn->cm_id->qp, &attr, IBV_QP_PATH_MTU, &init_attr);
310 sprintf(ibw_lasterr, "ibv_query_qp failed with %d\n", rc);
314 return ibw_fill_cq(conn);
317 static int ibw_refill_cq_recv(struct ibw_conn *conn)
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);
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 */
327 struct ibv_recv_wr wr = {
328 .wr_id = 0, /* filled below */
332 struct ibv_recv_wr *bad_wr;
334 DEBUG(DEBUG_DEBUG, ("ibw_refill_cq_recv(cmid: %p)\n", pconn->cm_id));
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;
340 rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
342 sprintf(ibw_lasterr, "refill/ibv_post_recv failed with %d\n", rc);
343 DEBUG(DEBUG_ERR, (ibw_lasterr));
350 static int ibw_fill_cq(struct ibw_conn *conn)
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);
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 */
360 struct ibv_recv_wr wr = {
361 .wr_id = 0, /* filled below */
365 struct ibv_recv_wr *bad_wr;
367 DEBUG(DEBUG_DEBUG, ("ibw_fill_cq(cmid: %p)\n", pconn->cm_id));
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;
374 rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
376 sprintf(ibw_lasterr, "fill/ibv_post_recv failed with %d\n", rc);
377 DEBUG(DEBUG_ERR, (ibw_lasterr));
385 static int ibw_manage_connect(struct ibw_conn *conn)
387 struct rdma_conn_param conn_param;
388 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
391 DEBUG(DEBUG_DEBUG, ("ibw_manage_connect(cmid: %p)\n", pconn->cm_id));
393 if (ibw_setup_cq_qp(conn))
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;
402 rc = rdma_connect(pconn->cm_id, &conn_param);
404 sprintf(ibw_lasterr, "rdma_connect error %d\n", rc);
409 static void ibw_event_handler_cm(struct tevent_context *ev,
410 struct tevent_fd *fde, uint16_t flags, void *private_data)
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;
422 rc = rdma_get_cm_event(pctx->cm_channel, &event);
424 ctx->state = IBWS_ERROR;
426 sprintf(ibw_lasterr, "rdma_get_cm_event error %d\n", rc);
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"));
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);
440 sprintf(ibw_lasterr, "rdma_resolve_route error %d\n", rc);
443 /* continued at RDMA_CM_EVENT_ROUTE_RESOLVED */
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);
452 rc = ibw_manage_connect(conn);
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));
467 if (ibw_setup_cq_qp(conn))
470 conn->state = IBWC_INIT;
471 pctx->connstate_func(ctx, conn);
473 /* continued at ibw_accept when invoked by the func above */
474 if (!pconn->is_accepted) {
475 rc = rdma_reject(cma_id, NULL, 0);
477 DEBUG(DEBUG_ERR, ("rdma_reject failed with rc=%d\n", rc));
479 DEBUG(DEBUG_DEBUG, ("pconn->cm_id %p wasn't accepted\n", pconn->cm_id));
482 /* TODO: clarify whether if it's needed by upper layer: */
483 ctx->state = IBWS_READY;
484 pctx->connstate_func(ctx, NULL);
486 /* NOTE: more requests can arrive until RDMA_CM_EVENT_ESTABLISHED ! */
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 */
495 DEBUG(DEBUG_DEBUG, ("ibw_setup_cq_qp succeeded (cmid=%p)\n", cma_id));
497 /* client conn is up */
498 conn->state = IBWC_CONNECTED;
500 /* both ctx and conn have changed */
501 pctx->connstate_func(ctx, conn);
504 case RDMA_CM_EVENT_ADDR_ERROR:
505 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ADDR_ERROR, error %d\n", event->status);
506 case RDMA_CM_EVENT_ROUTE_ERROR:
507 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ROUTE_ERROR, error %d\n", event->status);
508 case RDMA_CM_EVENT_CONNECT_ERROR:
509 sprintf(ibw_lasterr, "RDMA_CM_EVENT_CONNECT_ERROR, error %d\n", event->status);
510 case RDMA_CM_EVENT_UNREACHABLE:
511 sprintf(ibw_lasterr, "RDMA_CM_EVENT_UNREACHABLE, error %d\n", event->status);
513 case RDMA_CM_EVENT_REJECTED:
514 sprintf(ibw_lasterr, "RDMA_CM_EVENT_REJECTED, error %d\n", event->status);
515 DEBUG(DEBUG_INFO, ("cm event handler: %s", ibw_lasterr));
516 conn = talloc_get_type(cma_id->context, struct ibw_conn);
518 /* must be done BEFORE connstate */
519 if ((rc=rdma_ack_cm_event(event)))
520 DEBUG(DEBUG_ERR, ("reject/rdma_ack_cm_event failed with %d\n", rc));
521 event = NULL; /* not to touch cma_id or conn */
522 conn->state = IBWC_ERROR;
523 /* it should free the conn */
524 pctx->connstate_func(NULL, conn);
526 break; /* this is not strictly an error */
528 case RDMA_CM_EVENT_DISCONNECTED:
529 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_DISCONNECTED\n"));
530 if ((rc=rdma_ack_cm_event(event)))
531 DEBUG(DEBUG_ERR, ("disc/rdma_ack_cm_event failed with %d\n", rc));
532 event = NULL; /* don't ack more */
534 if (cma_id!=pctx->cm_id) {
535 DEBUG(DEBUG_ERR, ("client DISCONNECT event cm_id=%p\n", cma_id));
536 conn = talloc_get_type(cma_id->context, struct ibw_conn);
537 conn->state = IBWC_DISCONNECTED;
538 pctx->connstate_func(NULL, conn);
542 case RDMA_CM_EVENT_DEVICE_REMOVAL:
543 sprintf(ibw_lasterr, "cma detected device removal!\n");
547 sprintf(ibw_lasterr, "unknown event %d\n", event->event);
551 if (event!=NULL && (rc=rdma_ack_cm_event(event))) {
552 sprintf(ibw_lasterr, "rdma_ack_cm_event failed with %d\n", rc);
558 DEBUG(DEBUG_ERR, ("cm event handler: %s", ibw_lasterr));
561 if (cma_id!=NULL && cma_id!=pctx->cm_id) {
562 conn = talloc_get_type(cma_id->context, struct ibw_conn);
564 conn->state = IBWC_ERROR;
565 pctx->connstate_func(NULL, conn);
568 ctx->state = IBWS_ERROR;
569 pctx->connstate_func(ctx, NULL);
572 if ((rc=rdma_ack_cm_event(event))!=0) {
573 DEBUG(DEBUG_ERR, ("rdma_ack_cm_event failed with %d\n", rc));
580 static void ibw_event_handler_verbs(struct tevent_context *ev,
581 struct tevent_fd *fde, uint16_t flags, void *private_data)
583 struct ibw_conn *conn = talloc_get_type(private_data, struct ibw_conn);
584 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
585 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
589 struct ibv_cq *ev_cq;
592 DEBUG(DEBUG_DEBUG, ("ibw_event_handler_verbs(%u)\n", (uint32_t)flags));
594 /* TODO: check whether if it's good to have more channels here... */
595 rc = ibv_get_cq_event(pconn->verbs_channel, &ev_cq, &ev_ctx);
597 sprintf(ibw_lasterr, "Failed to get cq_event with %d\n", rc);
600 if (ev_cq != pconn->cq) {
601 sprintf(ibw_lasterr, "ev_cq(%p) != pconn->cq(%p)\n", ev_cq, pconn->cq);
604 rc = ibv_req_notify_cq(pconn->cq, 0);
606 sprintf(ibw_lasterr, "Couldn't request CQ notification (%d)\n", rc);
610 while((rc=ibv_poll_cq(pconn->cq, 1, &wc))==1) {
612 sprintf(ibw_lasterr, "cq completion failed status=%d, opcode=%d, rc=%d\n",
613 wc.status, wc.opcode, rc);
619 DEBUG(DEBUG_DEBUG, ("send completion\n"));
620 if (ibw_wc_send(conn, &wc))
624 case IBV_WC_RDMA_WRITE:
625 DEBUG(DEBUG_DEBUG, ("rdma write completion\n"));
628 case IBV_WC_RDMA_READ:
629 DEBUG(DEBUG_DEBUG, ("rdma read completion\n"));
633 DEBUG(DEBUG_DEBUG, ("recv completion\n"));
634 if (ibw_wc_recv(conn, &wc))
639 sprintf(ibw_lasterr, "unknown completion %d\n", wc.opcode);
644 sprintf(ibw_lasterr, "ibv_poll_cq error %d\n", rc);
648 ibv_ack_cq_events(pconn->cq, 1);
652 ibv_ack_cq_events(pconn->cq, 1);
654 DEBUG(DEBUG_ERR, (ibw_lasterr));
656 if (conn->state!=IBWC_ERROR) {
657 conn->state = IBWC_ERROR;
658 pctx->connstate_func(NULL, conn);
662 static int ibw_process_queue(struct ibw_conn *conn)
664 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
665 struct ibw_ctx_priv *pctx;
670 if (pconn->queue==NULL)
675 /* we must have at least 1 fragment to send */
676 assert(p->queued_ref_cnt>0);
679 pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
680 msg_size = (p->queued_ref_cnt) ? pctx->opts.recv_bufsize : p->queued_rlen;
682 assert(p->queued_msg!=NULL);
685 DEBUG(DEBUG_DEBUG, ("ibw_process_queue refcnt=%d msgsize=%u\n",
686 p->queued_ref_cnt, msg_size));
688 rc = ibw_send_packet(conn, p->queued_msg, p, msg_size);
690 /* was this the last fragment? */
691 if (p->queued_ref_cnt) {
692 p->queued_msg += pctx->opts.recv_bufsize;
694 DLIST_REMOVE2(pconn->queue, p, qprev, qnext);
695 p->queued_msg = NULL;
701 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc)
703 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
704 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
708 DEBUG(DEBUG_DEBUG, ("ibw_wc_send(cmid: %p, wr_id: %u, bl: %u)\n",
709 pconn->cm_id, (uint32_t)wc->wr_id, (uint32_t)wc->byte_len));
711 assert(pconn->cm_id->qp->qp_num==wc->qp_num);
712 assert(wc->wr_id >= pctx->opts.max_recv_wr);
713 send_index = wc->wr_id - pctx->opts.max_recv_wr;
716 if (send_index < pctx->opts.max_send_wr) {
717 DEBUG(DEBUG_DEBUG, ("ibw_wc_send#1 %u\n", (int)wc->wr_id));
718 p = pconn->wr_index[send_index];
719 if (p->buf_large!=NULL) {
721 /* awaiting more of it... */
724 ibw_free_mr(&p->buf_large, &p->mr_large);
725 DLIST_REMOVE(pconn->wr_list_used, p);
726 DLIST_ADD(pconn->wr_list_avail, p);
728 } else { /* nasty - but necessary */
729 DLIST_REMOVE(pconn->wr_list_used, p);
730 DLIST_ADD(pconn->wr_list_avail, p);
732 } else { /* "extra" request - not optimized */
733 DEBUG(DEBUG_DEBUG, ("ibw_wc_send#2 %u\n", (int)wc->wr_id));
734 for(p=pconn->extra_sent; p!=NULL; p=p->next)
735 if ((p->wr_id + pctx->opts.max_recv_wr)==(int)wc->wr_id)
738 sprintf(ibw_lasterr, "failed to find wr_id %d\n", (int)wc->wr_id);
744 ibw_free_mr(&p->buf_large, &p->mr_large);
745 DLIST_REMOVE(pconn->extra_sent, p);
746 DLIST_ADD(pconn->extra_avail, p);
750 return ibw_process_queue(conn);
753 static int ibw_append_to_part(struct ibw_conn_priv *pconn,
754 struct ibw_part *part, char **pp, uint32_t add_len, int info)
756 DEBUG(DEBUG_DEBUG, ("ibw_append_to_part: cmid=%p, (bs=%u, len=%u, tr=%u), al=%u, i=%u\n",
757 pconn->cm_id, part->bufsize, part->len, part->to_read, add_len, info));
759 /* allocate more if necessary - it's an "evergrowing" buffer... */
760 if (part->len + add_len > part->bufsize) {
761 if (part->buf==NULL) {
762 assert(part->len==0);
763 part->buf = talloc_size(pconn, add_len);
764 if (part->buf==NULL) {
765 sprintf(ibw_lasterr, "recv talloc_size error (%u) #%d\n",
769 part->bufsize = add_len;
771 part->buf = talloc_realloc_size(pconn,
772 part->buf, part->len + add_len);
773 if (part->buf==NULL) {
774 sprintf(ibw_lasterr, "recv realloc error (%u + %u) #%d\n",
775 part->len, add_len, info);
779 part->bufsize = part->len + add_len;
783 memcpy(part->buf + part->len, *pp, add_len);
785 part->len += add_len;
786 part->to_read -= add_len;
791 static int ibw_wc_mem_threshold(struct ibw_conn_priv *pconn,
792 struct ibw_part *part, uint32_t threshold)
794 DEBUG(DEBUG_DEBUG, ("ibw_wc_mem_threshold: cmid=%p, (bs=%u, len=%u, tr=%u), thr=%u\n",
795 pconn->cm_id, part->bufsize, part->len, part->to_read, threshold));
797 if (part->bufsize > threshold) {
798 DEBUG(DEBUG_DEBUG, ("ibw_wc_mem_threshold: cmid=%p, %u > %u\n",
799 pconn->cm_id, part->bufsize, threshold));
800 talloc_free(part->buf);
801 part->buf = talloc_size(pconn, threshold);
802 if (part->buf==NULL) {
803 sprintf(ibw_lasterr, "talloc_size failed\n");
806 part->bufsize = threshold;
811 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc)
813 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
814 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
815 struct ibw_part *part = &pconn->part;
817 uint32_t remain = wc->byte_len;
819 DEBUG(DEBUG_DEBUG, ("ibw_wc_recv: cmid=%p, wr_id: %u, bl: %u\n",
820 pconn->cm_id, (uint32_t)wc->wr_id, remain));
822 assert(pconn->cm_id->qp->qp_num==wc->qp_num);
823 assert((int)wc->wr_id < pctx->opts.max_recv_wr);
824 assert(wc->byte_len <= pctx->opts.recv_bufsize);
826 p = pconn->buf_recv + ((int)wc->wr_id * pctx->opts.recv_bufsize);
829 /* here always true: (part->len!=0 && part->to_read!=0) ||
830 (part->len==0 && part->to_read==0) */
831 if (part->len) { /* is there a partial msg to be continued? */
832 int read_len = (part->to_read<=remain) ? part->to_read : remain;
833 if (ibw_append_to_part(pconn, part, &p, read_len, 421))
837 if (part->len<=sizeof(uint32_t) && part->to_read==0) {
838 assert(part->len==sizeof(uint32_t));
839 /* set it again now... */
840 part->to_read = *((uint32_t *)(part->buf)); /* TODO: ntohl */
841 if (part->to_read<sizeof(uint32_t)) {
842 sprintf(ibw_lasterr, "got msglen=%u #2\n", part->to_read);
845 part->to_read -= sizeof(uint32_t); /* it's already read */
848 if (part->to_read==0) {
849 if (pctx->receive_func(conn, part->buf, part->len) != 0) {
852 part->len = 0; /* tells not having partial data (any more) */
853 if (ibw_wc_mem_threshold(pconn, part, pctx->opts.recv_threshold))
857 if (remain>=sizeof(uint32_t)) {
858 uint32_t msglen = *(uint32_t *)p; /* TODO: ntohl */
859 if (msglen<sizeof(uint32_t)) {
860 sprintf(ibw_lasterr, "got msglen=%u\n", msglen);
864 /* mostly awaited case: */
865 if (msglen<=remain) {
866 if (pctx->receive_func(conn, p, msglen) != 0) {
872 part->to_read = msglen;
873 /* part->len is already 0 */
874 if (ibw_append_to_part(pconn, part, &p, remain, 422))
876 remain = 0; /* to be continued ... */
877 /* part->to_read > 0 here */
879 } else { /* edge case: */
880 part->to_read = sizeof(uint32_t);
881 /* part->len is already 0 */
882 if (ibw_append_to_part(pconn, part, &p, remain, 423))
885 /* part->to_read > 0 here */
888 } /* <remain> is always decreased at least by 1 */
890 if (ibw_refill_cq_recv(conn))
896 DEBUG(DEBUG_ERR, ("ibw_wc_recv error: %s", ibw_lasterr));
900 static int ibw_process_init_attrs(struct ibw_initattr *attr, int nattr, struct ibw_opts *opts)
903 const char *name, *value;
905 DEBUG(DEBUG_DEBUG, ("ibw_process_init_attrs: nattr: %d\n", nattr));
907 opts->max_send_wr = IBW_MAX_SEND_WR;
908 opts->max_recv_wr = IBW_MAX_RECV_WR;
909 opts->recv_bufsize = IBW_RECV_BUFSIZE;
910 opts->recv_threshold = IBW_RECV_THRESHOLD;
912 for(i=0; i<nattr; i++) {
914 value = attr[i].value;
916 assert(name!=NULL && value!=NULL);
917 if (strcmp(name, "max_send_wr")==0)
918 opts->max_send_wr = atoi(value);
919 else if (strcmp(name, "max_recv_wr")==0)
920 opts->max_recv_wr = atoi(value);
921 else if (strcmp(name, "recv_bufsize")==0)
922 opts->recv_bufsize = atoi(value);
923 else if (strcmp(name, "recv_threshold")==0)
924 opts->recv_threshold = atoi(value);
926 sprintf(ibw_lasterr, "ibw_init: unknown name %s\n", name);
933 struct ibw_ctx *ibw_init(struct ibw_initattr *attr, int nattr,
935 ibw_connstate_fn_t ibw_connstate,
936 ibw_receive_fn_t ibw_receive,
937 struct tevent_context *ectx)
939 struct ibw_ctx *ctx = talloc_zero(NULL, struct ibw_ctx);
940 struct ibw_ctx_priv *pctx;
943 DEBUG(DEBUG_DEBUG, ("ibw_init(ctx_userdata: %p, ectx: %p)\n", ctx_userdata, ectx));
945 /* initialize basic data structures */
946 memset(ibw_lasterr, 0, IBW_LASTERR_BUFSIZE);
949 ibw_lasterr[0] = '\0';
950 talloc_set_destructor(ctx, ibw_ctx_destruct);
951 ctx->ctx_userdata = ctx_userdata;
953 pctx = talloc_zero(ctx, struct ibw_ctx_priv);
954 talloc_set_destructor(pctx, ibw_ctx_priv_destruct);
955 ctx->internal = (void *)pctx;
958 pctx->connstate_func = ibw_connstate;
959 pctx->receive_func = ibw_receive;
963 /* process attributes */
964 if (ibw_process_init_attrs(attr, nattr, &pctx->opts))
968 pctx->cm_channel = rdma_create_event_channel();
969 if (!pctx->cm_channel) {
970 sprintf(ibw_lasterr, "rdma_create_event_channel error %d\n", errno);
974 pctx->cm_channel_event = tevent_add_fd(pctx->ectx, pctx,
975 pctx->cm_channel->fd, TEVENT_FD_READ, ibw_event_handler_cm, ctx);
977 #if RDMA_USER_CM_MAX_ABI_VERSION >= 2
978 rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx, RDMA_PS_TCP);
980 rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx);
984 sprintf(ibw_lasterr, "rdma_create_id error %d\n", rc);
987 DEBUG(DEBUG_DEBUG, ("created cm_id %p\n", pctx->cm_id));
989 pctx->pagesize = sysconf(_SC_PAGESIZE);
992 /* don't put code here */
994 DEBUG(DEBUG_ERR, (ibw_lasterr));
1002 int ibw_stop(struct ibw_ctx *ctx)
1004 struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1007 DEBUG(DEBUG_DEBUG, ("ibw_stop\n"));
1009 for(p=ctx->conn_list; p!=NULL; p=p->next) {
1010 if (p->state==IBWC_ERROR || p->state==IBWC_CONNECTED) {
1011 if (ibw_disconnect(p))
1016 ctx->state = IBWS_STOPPED;
1017 pctx->connstate_func(ctx, NULL);
1022 int ibw_bind(struct ibw_ctx *ctx, struct sockaddr_in *my_addr)
1024 struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1027 DEBUG(DEBUG_DEBUG, ("ibw_bind: addr=%s, port=%u\n",
1028 inet_ntoa(my_addr->sin_addr), ntohs(my_addr->sin_port)));
1029 rc = rdma_bind_addr(pctx->cm_id, (struct sockaddr *) my_addr);
1031 sprintf(ibw_lasterr, "rdma_bind_addr error %d\n", rc);
1032 DEBUG(DEBUG_ERR, (ibw_lasterr));
1035 DEBUG(DEBUG_DEBUG, ("rdma_bind_addr successful\n"));
1040 int ibw_listen(struct ibw_ctx *ctx, int backlog)
1042 struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
1045 DEBUG(DEBUG_DEBUG, ("ibw_listen\n"));
1046 rc = rdma_listen(pctx->cm_id, backlog);
1048 sprintf(ibw_lasterr, "rdma_listen failed: %d\n", rc);
1049 DEBUG(DEBUG_ERR, (ibw_lasterr));
1056 int ibw_accept(struct ibw_ctx *ctx, struct ibw_conn *conn, void *conn_userdata)
1058 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1059 struct rdma_conn_param conn_param;
1062 DEBUG(DEBUG_DEBUG, ("ibw_accept: cmid=%p\n", pconn->cm_id));
1063 conn->conn_userdata = conn_userdata;
1065 memset(&conn_param, 0, sizeof(struct rdma_conn_param));
1066 conn_param.responder_resources = 1;
1067 conn_param.initiator_depth = 1;
1068 rc = rdma_accept(pconn->cm_id, &conn_param);
1070 sprintf(ibw_lasterr, "rdma_accept failed %d\n", rc);
1071 DEBUG(DEBUG_ERR, (ibw_lasterr));
1075 pconn->is_accepted = 1;
1077 /* continued at RDMA_CM_EVENT_ESTABLISHED */
1082 int ibw_connect(struct ibw_conn *conn, struct sockaddr_in *serv_addr, void *conn_userdata)
1084 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1085 struct ibw_conn_priv *pconn = NULL;
1090 conn->conn_userdata = conn_userdata;
1091 pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1092 DEBUG(DEBUG_DEBUG, ("ibw_connect: addr=%s, port=%u\n", inet_ntoa(serv_addr->sin_addr),
1093 ntohs(serv_addr->sin_port)));
1095 /* clean previous - probably half - initialization */
1096 if (ibw_conn_priv_destruct(pconn)) {
1097 DEBUG(DEBUG_ERR, ("ibw_connect/ibw_pconn_destruct failed for cm_id=%p\n", pconn->cm_id));
1102 #if RDMA_USER_CM_MAX_ABI_VERSION >= 2
1103 rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn, RDMA_PS_TCP);
1105 rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn);
1109 sprintf(ibw_lasterr, "ibw_connect/rdma_create_id error %d\n", rc);
1113 DEBUG(DEBUG_DEBUG, ("ibw_connect: rdma_create_id succeeded, cm_id=%p\n", pconn->cm_id));
1115 rc = rdma_resolve_addr(pconn->cm_id, NULL, (struct sockaddr *) serv_addr, 2000);
1117 sprintf(ibw_lasterr, "rdma_resolve_addr error %d\n", rc);
1118 DEBUG(DEBUG_ERR, (ibw_lasterr));
1123 /* continued at RDMA_CM_EVENT_ADDR_RESOLVED */
1128 int ibw_disconnect(struct ibw_conn *conn)
1131 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1133 DEBUG(DEBUG_DEBUG, ("ibw_disconnect: cmid=%p\n", pconn->cm_id));
1135 assert(pconn!=NULL);
1137 switch(conn->state) {
1139 ibw_conn_priv_destruct(pconn); /* do this here right now */
1141 case IBWC_CONNECTED:
1142 rc = rdma_disconnect(pconn->cm_id);
1144 sprintf(ibw_lasterr, "ibw_disconnect failed with %d\n", rc);
1145 DEBUG(DEBUG_ERR, (ibw_lasterr));
1150 DEBUG(DEBUG_DEBUG, ("invalid state for disconnect: %d\n", conn->state));
1157 int ibw_alloc_send_buf(struct ibw_conn *conn, void **buf, void **key, uint32_t len)
1159 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1160 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1161 struct ibw_wr *p = pconn->wr_list_avail;
1164 DEBUG(DEBUG_DEBUG, ("ibw_alloc_send_buf#1: cmid=%p, len=%d\n", pconn->cm_id, len));
1166 DLIST_REMOVE(pconn->wr_list_avail, p);
1167 DLIST_ADD(pconn->wr_list_used, p);
1169 if (len <= pctx->opts.recv_bufsize) {
1170 *buf = (void *)p->buf;
1172 p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1173 if (p->buf_large==NULL) {
1174 sprintf(ibw_lasterr, "ibw_alloc_mr#1 failed\n");
1177 *buf = (void *)p->buf_large;
1179 /* p->wr_id is already filled in ibw_init_memory */
1181 DEBUG(DEBUG_DEBUG, ("ibw_alloc_send_buf#2: cmid=%p, len=%d\n", pconn->cm_id, len));
1183 p = pconn->extra_avail;
1185 p = pconn->extra_avail = talloc_zero(pconn, struct ibw_wr);
1186 talloc_set_destructor(p, ibw_wr_destruct);
1188 sprintf(ibw_lasterr, "talloc_zero failed (emax: %u)\n", pconn->extra_max);
1191 p->wr_id = pctx->opts.max_send_wr + pconn->extra_max;
1193 switch(pconn->extra_max) {
1194 case 1: DEBUG(DEBUG_INFO, ("warning: queue performed\n")); break;
1195 case 10: DEBUG(DEBUG_INFO, ("warning: queue reached 10\n")); break;
1196 case 100: DEBUG(DEBUG_INFO, ("warning: queue reached 100\n")); break;
1197 case 1000: DEBUG(DEBUG_INFO, ("warning: queue reached 1000\n")); break;
1202 p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1203 if (p->buf_large==NULL) {
1204 sprintf(ibw_lasterr, "ibw_alloc_mr#2 failed\n");
1207 *buf = (void *)p->buf_large;
1209 DLIST_REMOVE(pconn->extra_avail, p);
1210 /* we don't have prepared index for this, so that
1211 * we will have to find this by wr_id later on */
1212 DLIST_ADD(pconn->extra_sent, p);
1219 DEBUG(DEBUG_ERR, ("ibw_alloc_send_buf error: %s", ibw_lasterr));
1224 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len)
1226 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1227 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1230 /* can we send it right now? */
1231 if (pconn->wr_sent<pctx->opts.max_send_wr) {
1232 struct ibv_send_wr *bad_wr;
1233 struct ibv_sge list = {
1234 .addr = (uintptr_t)buf,
1236 .lkey = pconn->mr_send->lkey
1238 struct ibv_send_wr wr = {
1239 .wr_id = p->wr_id + pctx->opts.max_recv_wr,
1242 .opcode = IBV_WR_SEND,
1243 .send_flags = IBV_SEND_SIGNALED,
1246 if (p->buf_large==NULL) {
1247 DEBUG(DEBUG_DEBUG, ("ibw_send#normal(cmid: %p, wrid: %u, n: %d)\n",
1248 pconn->cm_id, (uint32_t)wr.wr_id, len));
1250 DEBUG(DEBUG_DEBUG, ("ibw_send#large(cmid: %p, wrid: %u, n: %d)\n",
1251 pconn->cm_id, (uint32_t)wr.wr_id, len));
1252 list.lkey = p->mr_large->lkey;
1255 rc = ibv_post_send(pconn->cm_id->qp, &wr, &bad_wr);
1257 sprintf(ibw_lasterr, "ibv_post_send error %d (%d)\n",
1258 rc, pconn->wr_sent);
1265 } /* else put the request into our own queue: */
1267 DEBUG(DEBUG_DEBUG, ("ibw_send#queued(cmid: %p, len: %u)\n", pconn->cm_id, len));
1269 /* TODO: clarify how to continue when state==IBWC_STOPPED */
1271 /* to be sent by ibw_wc_send */
1272 /* regardless "normal" or [a part of] "large" packet */
1273 if (!p->queued_ref_cnt) {
1274 DLIST_ADD_END2(pconn->queue, p, struct ibw_wr *,
1275 qprev, qnext); /* TODO: optimize */
1276 p->queued_msg = buf;
1278 p->queued_ref_cnt++;
1279 p->queued_rlen = len; /* last wins; see ibw_wc_send */
1283 DEBUG(DEBUG_ERR, (ibw_lasterr));
1287 int ibw_send(struct ibw_conn *conn, void *buf, void *key, uint32_t len)
1289 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1290 struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1293 assert(len>=sizeof(uint32_t));
1294 assert((*((uint32_t *)buf)==len)); /* TODO: htonl */
1296 if (len > pctx->opts.recv_bufsize) {
1297 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1299 char *packet = (char *)buf;
1300 uint32_t recv_bufsize = pctx->opts.recv_bufsize;
1302 DEBUG(DEBUG_DEBUG, ("ibw_send#frag(cmid: %p, buf: %p, len: %u)\n",
1303 pconn->cm_id, buf, len));
1305 /* single threaded => no race here: */
1306 assert(p->ref_cnt==0);
1307 while(rlen > recv_bufsize) {
1308 rc = ibw_send_packet(conn, packet, p, recv_bufsize);
1311 packet += recv_bufsize;
1312 rlen -= recv_bufsize;
1313 p->ref_cnt++; /* not good to have it in ibw_send_packet */
1316 rc = ibw_send_packet(conn, packet, p, rlen);
1317 p->ref_cnt++; /* not good to have it in ibw_send_packet */
1319 p->ref_cnt--; /* for the same handling */
1321 assert(p->ref_cnt==0);
1322 assert(p->queued_ref_cnt==0);
1324 rc = ibw_send_packet(conn, buf, p, len);
1329 int ibw_cancel_send_buf(struct ibw_conn *conn, void *buf, void *key)
1331 struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1332 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1333 struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1339 if (p->buf_large!=NULL)
1340 ibw_free_mr(&p->buf_large, &p->mr_large);
1343 if (p->wr_id < pctx->opts.max_send_wr) {
1344 DEBUG(DEBUG_DEBUG, ("ibw_cancel_send_buf#1 %u", (int)p->wr_id));
1345 DLIST_REMOVE(pconn->wr_list_used, p);
1346 DLIST_ADD(pconn->wr_list_avail, p);
1347 } else { /* "extra" packet */
1348 DEBUG(DEBUG_DEBUG, ("ibw_cancel_send_buf#2 %u", (int)p->wr_id));
1349 DLIST_REMOVE(pconn->extra_sent, p);
1350 DLIST_ADD(pconn->extra_avail, p);
1356 const char *ibw_getLastError(void)