2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 * Author: Tom Tucker <tom@opengridcomputing.com>
43 #include <linux/sunrpc/svc_xprt.h>
44 #include <linux/sunrpc/addr.h>
45 #include <linux/sunrpc/debug.h>
46 #include <linux/sunrpc/rpc_rdma.h>
47 #include <linux/interrupt.h>
48 #include <linux/sched.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/workqueue.h>
52 #include <rdma/ib_verbs.h>
53 #include <rdma/rdma_cm.h>
54 #include <linux/sunrpc/svc_rdma.h>
55 #include <linux/export.h>
56 #include "xprt_rdma.h"
58 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
60 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int);
61 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
63 struct sockaddr *sa, int salen,
65 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
66 static void svc_rdma_release_rqst(struct svc_rqst *);
67 static void svc_rdma_detach(struct svc_xprt *xprt);
68 static void svc_rdma_free(struct svc_xprt *xprt);
69 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
70 static int svc_rdma_secure_port(struct svc_rqst *);
71 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
73 static struct svc_xprt_ops svc_rdma_ops = {
74 .xpo_create = svc_rdma_create,
75 .xpo_recvfrom = svc_rdma_recvfrom,
76 .xpo_sendto = svc_rdma_sendto,
77 .xpo_release_rqst = svc_rdma_release_rqst,
78 .xpo_detach = svc_rdma_detach,
79 .xpo_free = svc_rdma_free,
80 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
81 .xpo_has_wspace = svc_rdma_has_wspace,
82 .xpo_accept = svc_rdma_accept,
83 .xpo_secure_port = svc_rdma_secure_port,
84 .xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
87 struct svc_xprt_class svc_rdma_class = {
89 .xcl_owner = THIS_MODULE,
90 .xcl_ops = &svc_rdma_ops,
91 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
92 .xcl_ident = XPRT_TRANSPORT_RDMA,
95 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
96 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
97 struct sockaddr *, int, int);
98 static void svc_rdma_bc_detach(struct svc_xprt *);
99 static void svc_rdma_bc_free(struct svc_xprt *);
101 static struct svc_xprt_ops svc_rdma_bc_ops = {
102 .xpo_create = svc_rdma_bc_create,
103 .xpo_detach = svc_rdma_bc_detach,
104 .xpo_free = svc_rdma_bc_free,
105 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
106 .xpo_secure_port = svc_rdma_secure_port,
109 struct svc_xprt_class svc_rdma_bc_class = {
110 .xcl_name = "rdma-bc",
111 .xcl_owner = THIS_MODULE,
112 .xcl_ops = &svc_rdma_bc_ops,
113 .xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
116 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
118 struct sockaddr *sa, int salen,
121 struct svcxprt_rdma *cma_xprt;
122 struct svc_xprt *xprt;
124 cma_xprt = rdma_create_xprt(serv, 0);
126 return ERR_PTR(-ENOMEM);
127 xprt = &cma_xprt->sc_xprt;
129 svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
130 set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
131 serv->sv_bc_xprt = xprt;
133 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
137 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
139 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
142 static void svc_rdma_bc_free(struct svc_xprt *xprt)
144 struct svcxprt_rdma *rdma =
145 container_of(xprt, struct svcxprt_rdma, sc_xprt);
147 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
151 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
153 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
156 struct svc_rdma_op_ctxt *ctxt;
158 ctxt = kmalloc(sizeof(*ctxt), flags);
161 INIT_LIST_HEAD(&ctxt->list);
166 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
170 /* Each RPC/RDMA credit can consume a number of send
171 * and receive WQEs. One ctxt is allocated for each.
173 i = xprt->sc_sq_depth + xprt->sc_rq_depth;
176 struct svc_rdma_op_ctxt *ctxt;
178 ctxt = alloc_ctxt(xprt, GFP_KERNEL);
180 dprintk("svcrdma: No memory for RDMA ctxt\n");
183 list_add(&ctxt->list, &xprt->sc_ctxts);
188 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
190 struct svc_rdma_op_ctxt *ctxt = NULL;
192 spin_lock(&xprt->sc_ctxt_lock);
193 xprt->sc_ctxt_used++;
194 if (list_empty(&xprt->sc_ctxts))
197 ctxt = list_first_entry(&xprt->sc_ctxts,
198 struct svc_rdma_op_ctxt, list);
199 list_del(&ctxt->list);
200 spin_unlock(&xprt->sc_ctxt_lock);
204 ctxt->mapped_sges = 0;
208 /* Either pre-allocation missed the mark, or send
209 * queue accounting is broken.
211 spin_unlock(&xprt->sc_ctxt_lock);
213 ctxt = alloc_ctxt(xprt, GFP_NOIO);
217 spin_lock(&xprt->sc_ctxt_lock);
218 xprt->sc_ctxt_used--;
219 spin_unlock(&xprt->sc_ctxt_lock);
220 WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
224 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
226 struct svcxprt_rdma *xprt = ctxt->xprt;
227 struct ib_device *device = xprt->sc_cm_id->device;
228 u32 lkey = xprt->sc_pd->local_dma_lkey;
231 for (i = 0; i < ctxt->mapped_sges; i++) {
233 * Unmap the DMA addr in the SGE if the lkey matches
234 * the local_dma_lkey, otherwise, ignore it since it is
235 * an FRMR lkey and will be unmapped later when the
236 * last WR that uses it completes.
238 if (ctxt->sge[i].lkey == lkey)
239 ib_dma_unmap_page(device,
244 ctxt->mapped_sges = 0;
247 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
249 struct svcxprt_rdma *xprt = ctxt->xprt;
253 for (i = 0; i < ctxt->count; i++)
254 put_page(ctxt->pages[i]);
256 spin_lock(&xprt->sc_ctxt_lock);
257 xprt->sc_ctxt_used--;
258 list_add(&ctxt->list, &xprt->sc_ctxts);
259 spin_unlock(&xprt->sc_ctxt_lock);
262 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
264 while (!list_empty(&xprt->sc_ctxts)) {
265 struct svc_rdma_op_ctxt *ctxt;
267 ctxt = list_first_entry(&xprt->sc_ctxts,
268 struct svc_rdma_op_ctxt, list);
269 list_del(&ctxt->list);
274 /* QP event handler */
275 static void qp_event_handler(struct ib_event *event, void *context)
277 struct svc_xprt *xprt = context;
279 switch (event->event) {
280 /* These are considered benign events */
281 case IB_EVENT_PATH_MIG:
282 case IB_EVENT_COMM_EST:
283 case IB_EVENT_SQ_DRAINED:
284 case IB_EVENT_QP_LAST_WQE_REACHED:
285 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
286 ib_event_msg(event->event), event->event,
289 /* These are considered fatal events */
290 case IB_EVENT_PATH_MIG_ERR:
291 case IB_EVENT_QP_FATAL:
292 case IB_EVENT_QP_REQ_ERR:
293 case IB_EVENT_QP_ACCESS_ERR:
294 case IB_EVENT_DEVICE_FATAL:
296 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
297 "closing transport\n",
298 ib_event_msg(event->event), event->event,
300 set_bit(XPT_CLOSE, &xprt->xpt_flags);
306 * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
307 * @cq: completion queue
311 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
313 struct svcxprt_rdma *xprt = cq->cq_context;
314 struct ib_cqe *cqe = wc->wr_cqe;
315 struct svc_rdma_op_ctxt *ctxt;
317 /* WARNING: Only wc->wr_cqe and wc->status are reliable */
318 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
319 svc_rdma_unmap_dma(ctxt);
321 if (wc->status != IB_WC_SUCCESS)
324 /* All wc fields are now known to be valid */
325 ctxt->byte_len = wc->byte_len;
326 spin_lock(&xprt->sc_rq_dto_lock);
327 list_add_tail(&ctxt->list, &xprt->sc_rq_dto_q);
328 spin_unlock(&xprt->sc_rq_dto_lock);
330 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
331 if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
333 svc_xprt_enqueue(&xprt->sc_xprt);
337 if (wc->status != IB_WC_WR_FLUSH_ERR)
338 pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
339 ib_wc_status_msg(wc->status),
340 wc->status, wc->vendor_err);
341 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
342 svc_rdma_put_context(ctxt, 1);
345 svc_xprt_put(&xprt->sc_xprt);
349 * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
350 * @cq: completion queue
354 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
356 struct svcxprt_rdma *xprt = cq->cq_context;
357 struct ib_cqe *cqe = wc->wr_cqe;
358 struct svc_rdma_op_ctxt *ctxt;
360 atomic_inc(&xprt->sc_sq_avail);
361 wake_up(&xprt->sc_send_wait);
363 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
364 svc_rdma_unmap_dma(ctxt);
365 svc_rdma_put_context(ctxt, 1);
367 if (unlikely(wc->status != IB_WC_SUCCESS)) {
368 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
369 if (wc->status != IB_WC_WR_FLUSH_ERR)
370 pr_err("svcrdma: Send: %s (%u/0x%x)\n",
371 ib_wc_status_msg(wc->status),
372 wc->status, wc->vendor_err);
375 svc_xprt_put(&xprt->sc_xprt);
378 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
381 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
385 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
386 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
387 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
388 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
389 INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
390 INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
391 init_waitqueue_head(&cma_xprt->sc_send_wait);
393 spin_lock_init(&cma_xprt->sc_lock);
394 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
395 spin_lock_init(&cma_xprt->sc_ctxt_lock);
396 spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
399 * Note that this implies that the underlying transport support
400 * has some form of congestion control (see RFC 7530 section 3.1
401 * paragraph 2). For now, we assume that all supported RDMA
402 * transports are suitable here.
404 set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
407 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
412 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
414 struct ib_recv_wr recv_wr, *bad_recv_wr;
415 struct svc_rdma_op_ctxt *ctxt;
422 ctxt = svc_rdma_get_context(xprt);
424 ctxt->direction = DMA_FROM_DEVICE;
425 ctxt->cqe.done = svc_rdma_wc_receive;
426 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
427 if (sge_no >= xprt->sc_max_sge) {
428 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
431 page = alloc_page(flags);
434 ctxt->pages[sge_no] = page;
435 pa = ib_dma_map_page(xprt->sc_cm_id->device,
438 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
440 svc_rdma_count_mappings(xprt, ctxt);
441 ctxt->sge[sge_no].addr = pa;
442 ctxt->sge[sge_no].length = PAGE_SIZE;
443 ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
444 ctxt->count = sge_no + 1;
448 recv_wr.sg_list = &ctxt->sge[0];
449 recv_wr.num_sge = ctxt->count;
450 recv_wr.wr_cqe = &ctxt->cqe;
452 svc_xprt_get(&xprt->sc_xprt);
453 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
455 svc_rdma_unmap_dma(ctxt);
456 svc_rdma_put_context(ctxt, 1);
457 svc_xprt_put(&xprt->sc_xprt);
462 svc_rdma_unmap_dma(ctxt);
463 svc_rdma_put_context(ctxt, 1);
467 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags)
471 ret = svc_rdma_post_recv(xprt, flags);
473 pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
475 pr_err("svcrdma: closing transport %p.\n", xprt);
476 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
483 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
484 struct rdma_conn_param *param)
486 const struct rpcrdma_connect_private *pmsg = param->private_data;
489 pmsg->cp_magic == rpcrdma_cmp_magic &&
490 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
491 newxprt->sc_snd_w_inv = pmsg->cp_flags &
492 RPCRDMA_CMP_F_SND_W_INV_OK;
494 dprintk("svcrdma: client send_size %u, recv_size %u "
495 "remote inv %ssupported\n",
496 rpcrdma_decode_buffer_size(pmsg->cp_send_size),
497 rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
498 newxprt->sc_snd_w_inv ? "" : "un");
503 * This function handles the CONNECT_REQUEST event on a listening
504 * endpoint. It is passed the cma_id for the _new_ connection. The context in
505 * this cma_id is inherited from the listening cma_id and is the svc_xprt
506 * structure for the listening endpoint.
508 * This function creates a new xprt for the new connection and enqueues it on
509 * the accept queue for the listent xprt. When the listen thread is kicked, it
510 * will call the recvfrom method on the listen xprt which will accept the new
513 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
514 struct rdma_conn_param *param)
516 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
517 struct svcxprt_rdma *newxprt;
520 /* Create a new transport */
521 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
523 dprintk("svcrdma: failed to create new transport\n");
526 newxprt->sc_cm_id = new_cma_id;
527 new_cma_id->context = newxprt;
528 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
529 newxprt, newxprt->sc_cm_id, listen_xprt);
530 svc_rdma_parse_connect_private(newxprt, param);
532 /* Save client advertised inbound read limit for use later in accept. */
533 newxprt->sc_ord = param->initiator_depth;
535 /* Set the local and remote addresses in the transport */
536 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
537 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
538 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
539 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
542 * Enqueue the new transport on the accept queue of the listening
545 spin_lock_bh(&listen_xprt->sc_lock);
546 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
547 spin_unlock_bh(&listen_xprt->sc_lock);
549 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
550 svc_xprt_enqueue(&listen_xprt->sc_xprt);
554 * Handles events generated on the listening endpoint. These events will be
555 * either be incoming connect requests or adapter removal events.
557 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
558 struct rdma_cm_event *event)
560 struct svcxprt_rdma *xprt = cma_id->context;
563 switch (event->event) {
564 case RDMA_CM_EVENT_CONNECT_REQUEST:
565 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
566 "event = %s (%d)\n", cma_id, cma_id->context,
567 rdma_event_msg(event->event), event->event);
568 handle_connect_req(cma_id, &event->param.conn);
571 case RDMA_CM_EVENT_ESTABLISHED:
572 /* Accept complete */
573 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
574 "cm_id=%p\n", xprt, cma_id);
577 case RDMA_CM_EVENT_DEVICE_REMOVAL:
578 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
581 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
585 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
586 "event = %s (%d)\n", cma_id,
587 rdma_event_msg(event->event), event->event);
594 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
595 struct rdma_cm_event *event)
597 struct svc_xprt *xprt = cma_id->context;
598 struct svcxprt_rdma *rdma =
599 container_of(xprt, struct svcxprt_rdma, sc_xprt);
600 switch (event->event) {
601 case RDMA_CM_EVENT_ESTABLISHED:
602 /* Accept complete */
604 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
605 "cm_id=%p\n", xprt, cma_id);
606 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
607 svc_xprt_enqueue(xprt);
609 case RDMA_CM_EVENT_DISCONNECTED:
610 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
613 set_bit(XPT_CLOSE, &xprt->xpt_flags);
614 svc_xprt_enqueue(xprt);
618 case RDMA_CM_EVENT_DEVICE_REMOVAL:
619 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
620 "event = %s (%d)\n", cma_id, xprt,
621 rdma_event_msg(event->event), event->event);
623 set_bit(XPT_CLOSE, &xprt->xpt_flags);
624 svc_xprt_enqueue(xprt);
629 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
630 "event = %s (%d)\n", cma_id,
631 rdma_event_msg(event->event), event->event);
638 * Create a listening RDMA service endpoint.
640 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
642 struct sockaddr *sa, int salen,
645 struct rdma_cm_id *listen_id;
646 struct svcxprt_rdma *cma_xprt;
649 dprintk("svcrdma: Creating RDMA socket\n");
650 if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
651 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
652 return ERR_PTR(-EAFNOSUPPORT);
654 cma_xprt = rdma_create_xprt(serv, 1);
656 return ERR_PTR(-ENOMEM);
658 listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
659 RDMA_PS_TCP, IB_QPT_RC);
660 if (IS_ERR(listen_id)) {
661 ret = PTR_ERR(listen_id);
662 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
666 /* Allow both IPv4 and IPv6 sockets to bind a single port
669 #if IS_ENABLED(CONFIG_IPV6)
670 ret = rdma_set_afonly(listen_id, 1);
672 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
676 ret = rdma_bind_addr(listen_id, sa);
678 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
681 cma_xprt->sc_cm_id = listen_id;
683 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
685 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
690 * We need to use the address from the cm_id in case the
691 * caller specified 0 for the port number.
693 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
694 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
696 return &cma_xprt->sc_xprt;
699 rdma_destroy_id(listen_id);
706 * This is the xpo_recvfrom function for listening endpoints. Its
707 * purpose is to accept incoming connections. The CMA callback handler
708 * has already created a new transport and attached it to the new CMA
711 * There is a queue of pending connections hung on the listening
712 * transport. This queue contains the new svc_xprt structure. This
713 * function takes svc_xprt structures off the accept_q and completes
716 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
718 struct svcxprt_rdma *listen_rdma;
719 struct svcxprt_rdma *newxprt = NULL;
720 struct rdma_conn_param conn_param;
721 struct rpcrdma_connect_private pmsg;
722 struct ib_qp_init_attr qp_attr;
723 struct ib_device *dev;
724 struct sockaddr *sap;
728 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
729 clear_bit(XPT_CONN, &xprt->xpt_flags);
730 /* Get the next entry off the accept list */
731 spin_lock_bh(&listen_rdma->sc_lock);
732 if (!list_empty(&listen_rdma->sc_accept_q)) {
733 newxprt = list_entry(listen_rdma->sc_accept_q.next,
734 struct svcxprt_rdma, sc_accept_q);
735 list_del_init(&newxprt->sc_accept_q);
737 if (!list_empty(&listen_rdma->sc_accept_q))
738 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
739 spin_unlock_bh(&listen_rdma->sc_lock);
743 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
744 newxprt, newxprt->sc_cm_id);
746 dev = newxprt->sc_cm_id->device;
747 newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
749 /* Qualify the transport resource defaults with the
750 * capabilities of this particular device */
751 newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
752 (size_t)RPCSVC_MAXPAGES);
753 newxprt->sc_max_req_size = svcrdma_max_req_size;
754 newxprt->sc_max_requests = min_t(u32, dev->attrs.max_qp_wr,
755 svcrdma_max_requests);
756 newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
757 newxprt->sc_max_bc_requests = min_t(u32, dev->attrs.max_qp_wr,
758 svcrdma_max_bc_requests);
759 newxprt->sc_rq_depth = newxprt->sc_max_requests +
760 newxprt->sc_max_bc_requests;
761 newxprt->sc_sq_depth = newxprt->sc_rq_depth;
762 atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
764 if (!svc_rdma_prealloc_ctxts(newxprt))
768 * Limit ORD based on client limit, local device limit, and
769 * configured svcrdma limit.
771 newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
772 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
774 newxprt->sc_pd = ib_alloc_pd(dev, 0);
775 if (IS_ERR(newxprt->sc_pd)) {
776 dprintk("svcrdma: error creating PD for connect request\n");
779 newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
780 0, IB_POLL_WORKQUEUE);
781 if (IS_ERR(newxprt->sc_sq_cq)) {
782 dprintk("svcrdma: error creating SQ CQ for connect request\n");
785 newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
786 0, IB_POLL_WORKQUEUE);
787 if (IS_ERR(newxprt->sc_rq_cq)) {
788 dprintk("svcrdma: error creating RQ CQ for connect request\n");
792 memset(&qp_attr, 0, sizeof qp_attr);
793 qp_attr.event_handler = qp_event_handler;
794 qp_attr.qp_context = &newxprt->sc_xprt;
795 qp_attr.port_num = newxprt->sc_cm_id->port_num;
796 qp_attr.cap.max_rdma_ctxs = newxprt->sc_max_requests;
797 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
798 qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
799 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
800 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
801 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
802 qp_attr.qp_type = IB_QPT_RC;
803 qp_attr.send_cq = newxprt->sc_sq_cq;
804 qp_attr.recv_cq = newxprt->sc_rq_cq;
805 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
806 newxprt->sc_cm_id, newxprt->sc_pd);
807 dprintk(" cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
808 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
809 dprintk(" cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
810 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
812 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
814 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
817 newxprt->sc_qp = newxprt->sc_cm_id->qp;
820 * Use the most secure set of MR resources based on the
821 * transport type and available memory management features in
822 * the device. Here's the table implemented below:
824 * Fast Global DMA Remote WR
826 * Sup'd Sup'd Needed Needed
838 * NB: iWARP requires remote write access for the data sink
839 * of an RDMA_READ. IB does not.
841 if (dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
842 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
844 newxprt->sc_snd_w_inv = false;
847 * Determine if a DMA MR is required and if so, what privs are required
849 if (!rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num) &&
850 !rdma_ib_or_roce(dev, newxprt->sc_cm_id->port_num))
853 if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num))
854 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
856 /* Post receive buffers */
857 for (i = 0; i < newxprt->sc_max_requests; i++) {
858 ret = svc_rdma_post_recv(newxprt, GFP_KERNEL);
860 dprintk("svcrdma: failure posting receive buffers\n");
865 /* Swap out the handler */
866 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
868 /* Construct RDMA-CM private message */
869 pmsg.cp_magic = rpcrdma_cmp_magic;
870 pmsg.cp_version = RPCRDMA_CMP_VERSION;
872 pmsg.cp_send_size = pmsg.cp_recv_size =
873 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
875 /* Accept Connection */
876 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
877 memset(&conn_param, 0, sizeof conn_param);
878 conn_param.responder_resources = 0;
879 conn_param.initiator_depth = newxprt->sc_ord;
880 conn_param.private_data = &pmsg;
881 conn_param.private_data_len = sizeof(pmsg);
882 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
884 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
889 dprintk("svcrdma: new connection %p accepted:\n", newxprt);
890 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
891 dprintk(" local address : %pIS:%u\n", sap, rpc_get_port(sap));
892 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
893 dprintk(" remote address : %pIS:%u\n", sap, rpc_get_port(sap));
894 dprintk(" max_sge : %d\n", newxprt->sc_max_sge);
895 dprintk(" sq_depth : %d\n", newxprt->sc_sq_depth);
896 dprintk(" max_requests : %d\n", newxprt->sc_max_requests);
897 dprintk(" ord : %d\n", newxprt->sc_ord);
899 return &newxprt->sc_xprt;
902 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
903 /* Take a reference in case the DTO handler runs */
904 svc_xprt_get(&newxprt->sc_xprt);
905 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
906 ib_destroy_qp(newxprt->sc_qp);
907 rdma_destroy_id(newxprt->sc_cm_id);
908 /* This call to put will destroy the transport */
909 svc_xprt_put(&newxprt->sc_xprt);
913 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
918 * When connected, an svc_xprt has at least two references:
920 * - A reference held by the cm_id between the ESTABLISHED and
921 * DISCONNECTED events. If the remote peer disconnected first, this
922 * reference could be gone.
924 * - A reference held by the svc_recv code that called this function
925 * as part of close processing.
927 * At a minimum one references should still be held.
929 static void svc_rdma_detach(struct svc_xprt *xprt)
931 struct svcxprt_rdma *rdma =
932 container_of(xprt, struct svcxprt_rdma, sc_xprt);
933 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
935 /* Disconnect and flush posted WQE */
936 rdma_disconnect(rdma->sc_cm_id);
939 static void __svc_rdma_free(struct work_struct *work)
941 struct svcxprt_rdma *rdma =
942 container_of(work, struct svcxprt_rdma, sc_work);
943 struct svc_xprt *xprt = &rdma->sc_xprt;
945 dprintk("svcrdma: %s(%p)\n", __func__, rdma);
947 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
948 ib_drain_qp(rdma->sc_qp);
950 /* We should only be called from kref_put */
951 if (kref_read(&xprt->xpt_ref) != 0)
952 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
953 kref_read(&xprt->xpt_ref));
955 while (!list_empty(&rdma->sc_read_complete_q)) {
956 struct svc_rdma_op_ctxt *ctxt;
957 ctxt = list_first_entry(&rdma->sc_read_complete_q,
958 struct svc_rdma_op_ctxt, list);
959 list_del(&ctxt->list);
960 svc_rdma_put_context(ctxt, 1);
962 while (!list_empty(&rdma->sc_rq_dto_q)) {
963 struct svc_rdma_op_ctxt *ctxt;
964 ctxt = list_first_entry(&rdma->sc_rq_dto_q,
965 struct svc_rdma_op_ctxt, list);
966 list_del(&ctxt->list);
967 svc_rdma_put_context(ctxt, 1);
970 /* Warn if we leaked a resource or under-referenced */
971 if (rdma->sc_ctxt_used != 0)
972 pr_err("svcrdma: ctxt still in use? (%d)\n",
975 /* Final put of backchannel client transport */
976 if (xprt->xpt_bc_xprt) {
977 xprt_put(xprt->xpt_bc_xprt);
978 xprt->xpt_bc_xprt = NULL;
981 svc_rdma_destroy_rw_ctxts(rdma);
982 svc_rdma_destroy_ctxts(rdma);
984 /* Destroy the QP if present (not a listener) */
985 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
986 ib_destroy_qp(rdma->sc_qp);
988 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
989 ib_free_cq(rdma->sc_sq_cq);
991 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
992 ib_free_cq(rdma->sc_rq_cq);
994 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
995 ib_dealloc_pd(rdma->sc_pd);
997 /* Destroy the CM ID */
998 rdma_destroy_id(rdma->sc_cm_id);
1003 static void svc_rdma_free(struct svc_xprt *xprt)
1005 struct svcxprt_rdma *rdma =
1006 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1007 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1008 queue_work(svc_rdma_wq, &rdma->sc_work);
1011 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1013 struct svcxprt_rdma *rdma =
1014 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1017 * If there are already waiters on the SQ,
1020 if (waitqueue_active(&rdma->sc_send_wait))
1023 /* Otherwise return true. */
1027 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1032 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1036 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1038 struct ib_send_wr *bad_wr, *n_wr;
1043 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1047 for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1050 /* If the SQ is full, wait until an SQ entry is available */
1052 if ((atomic_sub_return(wr_count, &xprt->sc_sq_avail) < 0)) {
1053 atomic_inc(&rdma_stat_sq_starve);
1055 /* Wait until SQ WR available if SQ still full */
1056 atomic_add(wr_count, &xprt->sc_sq_avail);
1057 wait_event(xprt->sc_send_wait,
1058 atomic_read(&xprt->sc_sq_avail) > wr_count);
1059 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1063 /* Take a transport ref for each WR posted */
1064 for (i = 0; i < wr_count; i++)
1065 svc_xprt_get(&xprt->sc_xprt);
1067 /* Bump used SQ WR count and post */
1068 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1070 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1071 for (i = 0; i < wr_count; i ++)
1072 svc_xprt_put(&xprt->sc_xprt);
1073 dprintk("svcrdma: failed to post SQ WR rc=%d\n", ret);
1074 dprintk(" sc_sq_avail=%d, sc_sq_depth=%d\n",
1075 atomic_read(&xprt->sc_sq_avail),
1077 wake_up(&xprt->sc_send_wait);