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>
55 #include <linux/sunrpc/svc_rdma.h>
56 #include <linux/export.h>
57 #include "xprt_rdma.h"
59 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
61 static int svc_rdma_post_recv(struct svcxprt_rdma *xprt);
62 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int);
63 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
65 struct sockaddr *sa, int salen,
67 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
68 static void svc_rdma_release_rqst(struct svc_rqst *);
69 static void svc_rdma_detach(struct svc_xprt *xprt);
70 static void svc_rdma_free(struct svc_xprt *xprt);
71 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
72 static int svc_rdma_secure_port(struct svc_rqst *);
73 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
75 static const struct svc_xprt_ops svc_rdma_ops = {
76 .xpo_create = svc_rdma_create,
77 .xpo_recvfrom = svc_rdma_recvfrom,
78 .xpo_sendto = svc_rdma_sendto,
79 .xpo_release_rqst = svc_rdma_release_rqst,
80 .xpo_detach = svc_rdma_detach,
81 .xpo_free = svc_rdma_free,
82 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
83 .xpo_has_wspace = svc_rdma_has_wspace,
84 .xpo_accept = svc_rdma_accept,
85 .xpo_secure_port = svc_rdma_secure_port,
86 .xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
89 struct svc_xprt_class svc_rdma_class = {
91 .xcl_owner = THIS_MODULE,
92 .xcl_ops = &svc_rdma_ops,
93 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
94 .xcl_ident = XPRT_TRANSPORT_RDMA,
97 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
98 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
99 struct sockaddr *, int, int);
100 static void svc_rdma_bc_detach(struct svc_xprt *);
101 static void svc_rdma_bc_free(struct svc_xprt *);
103 static const struct svc_xprt_ops svc_rdma_bc_ops = {
104 .xpo_create = svc_rdma_bc_create,
105 .xpo_detach = svc_rdma_bc_detach,
106 .xpo_free = svc_rdma_bc_free,
107 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
108 .xpo_secure_port = svc_rdma_secure_port,
111 struct svc_xprt_class svc_rdma_bc_class = {
112 .xcl_name = "rdma-bc",
113 .xcl_owner = THIS_MODULE,
114 .xcl_ops = &svc_rdma_bc_ops,
115 .xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
118 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
120 struct sockaddr *sa, int salen,
123 struct svcxprt_rdma *cma_xprt;
124 struct svc_xprt *xprt;
126 cma_xprt = rdma_create_xprt(serv, 0);
128 return ERR_PTR(-ENOMEM);
129 xprt = &cma_xprt->sc_xprt;
131 svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
132 set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
133 serv->sv_bc_xprt = xprt;
135 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
139 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
141 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
144 static void svc_rdma_bc_free(struct svc_xprt *xprt)
146 struct svcxprt_rdma *rdma =
147 container_of(xprt, struct svcxprt_rdma, sc_xprt);
149 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
153 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
155 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
158 struct svc_rdma_op_ctxt *ctxt;
160 ctxt = kmalloc(sizeof(*ctxt), flags);
163 INIT_LIST_HEAD(&ctxt->list);
168 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
172 /* Each RPC/RDMA credit can consume one Receive and
173 * one Send WQE at the same time.
175 i = xprt->sc_sq_depth + xprt->sc_rq_depth;
178 struct svc_rdma_op_ctxt *ctxt;
180 ctxt = alloc_ctxt(xprt, GFP_KERNEL);
182 dprintk("svcrdma: No memory for RDMA ctxt\n");
185 list_add(&ctxt->list, &xprt->sc_ctxts);
190 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
192 struct svc_rdma_op_ctxt *ctxt = NULL;
194 spin_lock(&xprt->sc_ctxt_lock);
195 xprt->sc_ctxt_used++;
196 if (list_empty(&xprt->sc_ctxts))
199 ctxt = list_first_entry(&xprt->sc_ctxts,
200 struct svc_rdma_op_ctxt, list);
201 list_del(&ctxt->list);
202 spin_unlock(&xprt->sc_ctxt_lock);
206 ctxt->mapped_sges = 0;
210 /* Either pre-allocation missed the mark, or send
211 * queue accounting is broken.
213 spin_unlock(&xprt->sc_ctxt_lock);
215 ctxt = alloc_ctxt(xprt, GFP_NOIO);
219 spin_lock(&xprt->sc_ctxt_lock);
220 xprt->sc_ctxt_used--;
221 spin_unlock(&xprt->sc_ctxt_lock);
222 WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
226 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
228 struct svcxprt_rdma *xprt = ctxt->xprt;
229 struct ib_device *device = xprt->sc_cm_id->device;
232 for (i = 0; i < ctxt->mapped_sges; i++)
233 ib_dma_unmap_page(device,
237 ctxt->mapped_sges = 0;
240 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
242 struct svcxprt_rdma *xprt = ctxt->xprt;
246 for (i = 0; i < ctxt->count; i++)
247 put_page(ctxt->pages[i]);
249 spin_lock(&xprt->sc_ctxt_lock);
250 xprt->sc_ctxt_used--;
251 list_add(&ctxt->list, &xprt->sc_ctxts);
252 spin_unlock(&xprt->sc_ctxt_lock);
255 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
257 while (!list_empty(&xprt->sc_ctxts)) {
258 struct svc_rdma_op_ctxt *ctxt;
260 ctxt = list_first_entry(&xprt->sc_ctxts,
261 struct svc_rdma_op_ctxt, list);
262 list_del(&ctxt->list);
267 /* QP event handler */
268 static void qp_event_handler(struct ib_event *event, void *context)
270 struct svc_xprt *xprt = context;
272 switch (event->event) {
273 /* These are considered benign events */
274 case IB_EVENT_PATH_MIG:
275 case IB_EVENT_COMM_EST:
276 case IB_EVENT_SQ_DRAINED:
277 case IB_EVENT_QP_LAST_WQE_REACHED:
278 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
279 ib_event_msg(event->event), event->event,
282 /* These are considered fatal events */
283 case IB_EVENT_PATH_MIG_ERR:
284 case IB_EVENT_QP_FATAL:
285 case IB_EVENT_QP_REQ_ERR:
286 case IB_EVENT_QP_ACCESS_ERR:
287 case IB_EVENT_DEVICE_FATAL:
289 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
290 "closing transport\n",
291 ib_event_msg(event->event), event->event,
293 set_bit(XPT_CLOSE, &xprt->xpt_flags);
294 svc_xprt_enqueue(xprt);
300 * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
301 * @cq: completion queue
305 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
307 struct svcxprt_rdma *xprt = cq->cq_context;
308 struct ib_cqe *cqe = wc->wr_cqe;
309 struct svc_rdma_op_ctxt *ctxt;
311 /* WARNING: Only wc->wr_cqe and wc->status are reliable */
312 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
313 svc_rdma_unmap_dma(ctxt);
315 if (wc->status != IB_WC_SUCCESS)
318 /* All wc fields are now known to be valid */
319 ctxt->byte_len = wc->byte_len;
320 spin_lock(&xprt->sc_rq_dto_lock);
321 list_add_tail(&ctxt->list, &xprt->sc_rq_dto_q);
322 spin_unlock(&xprt->sc_rq_dto_lock);
324 svc_rdma_post_recv(xprt);
326 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
327 if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
332 if (wc->status != IB_WC_WR_FLUSH_ERR)
333 pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
334 ib_wc_status_msg(wc->status),
335 wc->status, wc->vendor_err);
336 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
337 svc_rdma_put_context(ctxt, 1);
340 svc_xprt_enqueue(&xprt->sc_xprt);
342 svc_xprt_put(&xprt->sc_xprt);
346 * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
347 * @cq: completion queue
351 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
353 struct svcxprt_rdma *xprt = cq->cq_context;
354 struct ib_cqe *cqe = wc->wr_cqe;
355 struct svc_rdma_op_ctxt *ctxt;
357 atomic_inc(&xprt->sc_sq_avail);
358 wake_up(&xprt->sc_send_wait);
360 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
361 svc_rdma_unmap_dma(ctxt);
362 svc_rdma_put_context(ctxt, 1);
364 if (unlikely(wc->status != IB_WC_SUCCESS)) {
365 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
366 svc_xprt_enqueue(&xprt->sc_xprt);
367 if (wc->status != IB_WC_WR_FLUSH_ERR)
368 pr_err("svcrdma: Send: %s (%u/0x%x)\n",
369 ib_wc_status_msg(wc->status),
370 wc->status, wc->vendor_err);
373 svc_xprt_put(&xprt->sc_xprt);
376 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
379 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
383 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
384 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
385 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
386 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
387 INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
388 INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
389 init_waitqueue_head(&cma_xprt->sc_send_wait);
391 spin_lock_init(&cma_xprt->sc_lock);
392 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
393 spin_lock_init(&cma_xprt->sc_ctxt_lock);
394 spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
397 * Note that this implies that the underlying transport support
398 * has some form of congestion control (see RFC 7530 section 3.1
399 * paragraph 2). For now, we assume that all supported RDMA
400 * transports are suitable here.
402 set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
405 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
411 svc_rdma_post_recv(struct svcxprt_rdma *xprt)
413 struct ib_recv_wr recv_wr, *bad_recv_wr;
414 struct svc_rdma_op_ctxt *ctxt;
421 ctxt = svc_rdma_get_context(xprt);
423 ctxt->direction = DMA_FROM_DEVICE;
424 ctxt->cqe.done = svc_rdma_wc_receive;
425 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
426 if (sge_no >= xprt->sc_max_sge) {
427 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
430 page = alloc_page(GFP_KERNEL);
433 ctxt->pages[sge_no] = page;
434 pa = ib_dma_map_page(xprt->sc_cm_id->device,
437 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
439 svc_rdma_count_mappings(xprt, ctxt);
440 ctxt->sge[sge_no].addr = pa;
441 ctxt->sge[sge_no].length = PAGE_SIZE;
442 ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
443 ctxt->count = sge_no + 1;
447 recv_wr.sg_list = &ctxt->sge[0];
448 recv_wr.num_sge = ctxt->count;
449 recv_wr.wr_cqe = &ctxt->cqe;
451 svc_xprt_get(&xprt->sc_xprt);
452 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
454 svc_rdma_unmap_dma(ctxt);
455 svc_rdma_put_context(ctxt, 1);
456 svc_xprt_put(&xprt->sc_xprt);
461 svc_rdma_unmap_dma(ctxt);
462 svc_rdma_put_context(ctxt, 1);
467 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
468 struct rdma_conn_param *param)
470 const struct rpcrdma_connect_private *pmsg = param->private_data;
473 pmsg->cp_magic == rpcrdma_cmp_magic &&
474 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
475 newxprt->sc_snd_w_inv = pmsg->cp_flags &
476 RPCRDMA_CMP_F_SND_W_INV_OK;
478 dprintk("svcrdma: client send_size %u, recv_size %u "
479 "remote inv %ssupported\n",
480 rpcrdma_decode_buffer_size(pmsg->cp_send_size),
481 rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
482 newxprt->sc_snd_w_inv ? "" : "un");
487 * This function handles the CONNECT_REQUEST event on a listening
488 * endpoint. It is passed the cma_id for the _new_ connection. The context in
489 * this cma_id is inherited from the listening cma_id and is the svc_xprt
490 * structure for the listening endpoint.
492 * This function creates a new xprt for the new connection and enqueues it on
493 * the accept queue for the listent xprt. When the listen thread is kicked, it
494 * will call the recvfrom method on the listen xprt which will accept the new
497 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
498 struct rdma_conn_param *param)
500 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
501 struct svcxprt_rdma *newxprt;
504 /* Create a new transport */
505 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
507 dprintk("svcrdma: failed to create new transport\n");
510 newxprt->sc_cm_id = new_cma_id;
511 new_cma_id->context = newxprt;
512 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
513 newxprt, newxprt->sc_cm_id, listen_xprt);
514 svc_rdma_parse_connect_private(newxprt, param);
516 /* Save client advertised inbound read limit for use later in accept. */
517 newxprt->sc_ord = param->initiator_depth;
519 /* Set the local and remote addresses in the transport */
520 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
521 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
522 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
523 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
526 * Enqueue the new transport on the accept queue of the listening
529 spin_lock_bh(&listen_xprt->sc_lock);
530 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
531 spin_unlock_bh(&listen_xprt->sc_lock);
533 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
534 svc_xprt_enqueue(&listen_xprt->sc_xprt);
538 * Handles events generated on the listening endpoint. These events will be
539 * either be incoming connect requests or adapter removal events.
541 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
542 struct rdma_cm_event *event)
544 struct svcxprt_rdma *xprt = cma_id->context;
547 switch (event->event) {
548 case RDMA_CM_EVENT_CONNECT_REQUEST:
549 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
550 "event = %s (%d)\n", cma_id, cma_id->context,
551 rdma_event_msg(event->event), event->event);
552 handle_connect_req(cma_id, &event->param.conn);
555 case RDMA_CM_EVENT_ESTABLISHED:
556 /* Accept complete */
557 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
558 "cm_id=%p\n", xprt, cma_id);
561 case RDMA_CM_EVENT_DEVICE_REMOVAL:
562 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
565 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
566 svc_xprt_enqueue(&xprt->sc_xprt);
571 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
572 "event = %s (%d)\n", cma_id,
573 rdma_event_msg(event->event), event->event);
580 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
581 struct rdma_cm_event *event)
583 struct svc_xprt *xprt = cma_id->context;
584 struct svcxprt_rdma *rdma =
585 container_of(xprt, struct svcxprt_rdma, sc_xprt);
586 switch (event->event) {
587 case RDMA_CM_EVENT_ESTABLISHED:
588 /* Accept complete */
590 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
591 "cm_id=%p\n", xprt, cma_id);
592 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
593 svc_xprt_enqueue(xprt);
595 case RDMA_CM_EVENT_DISCONNECTED:
596 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
599 set_bit(XPT_CLOSE, &xprt->xpt_flags);
600 svc_xprt_enqueue(xprt);
604 case RDMA_CM_EVENT_DEVICE_REMOVAL:
605 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
606 "event = %s (%d)\n", cma_id, xprt,
607 rdma_event_msg(event->event), event->event);
609 set_bit(XPT_CLOSE, &xprt->xpt_flags);
610 svc_xprt_enqueue(xprt);
615 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
616 "event = %s (%d)\n", cma_id,
617 rdma_event_msg(event->event), event->event);
624 * Create a listening RDMA service endpoint.
626 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
628 struct sockaddr *sa, int salen,
631 struct rdma_cm_id *listen_id;
632 struct svcxprt_rdma *cma_xprt;
635 dprintk("svcrdma: Creating RDMA socket\n");
636 if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
637 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
638 return ERR_PTR(-EAFNOSUPPORT);
640 cma_xprt = rdma_create_xprt(serv, 1);
642 return ERR_PTR(-ENOMEM);
644 listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
645 RDMA_PS_TCP, IB_QPT_RC);
646 if (IS_ERR(listen_id)) {
647 ret = PTR_ERR(listen_id);
648 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
652 /* Allow both IPv4 and IPv6 sockets to bind a single port
655 #if IS_ENABLED(CONFIG_IPV6)
656 ret = rdma_set_afonly(listen_id, 1);
658 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
662 ret = rdma_bind_addr(listen_id, sa);
664 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
667 cma_xprt->sc_cm_id = listen_id;
669 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
671 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
676 * We need to use the address from the cm_id in case the
677 * caller specified 0 for the port number.
679 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
680 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
682 return &cma_xprt->sc_xprt;
685 rdma_destroy_id(listen_id);
692 * This is the xpo_recvfrom function for listening endpoints. Its
693 * purpose is to accept incoming connections. The CMA callback handler
694 * has already created a new transport and attached it to the new CMA
697 * There is a queue of pending connections hung on the listening
698 * transport. This queue contains the new svc_xprt structure. This
699 * function takes svc_xprt structures off the accept_q and completes
702 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
704 struct svcxprt_rdma *listen_rdma;
705 struct svcxprt_rdma *newxprt = NULL;
706 struct rdma_conn_param conn_param;
707 struct rpcrdma_connect_private pmsg;
708 struct ib_qp_init_attr qp_attr;
709 struct ib_device *dev;
710 struct sockaddr *sap;
711 unsigned int i, ctxts;
714 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
715 clear_bit(XPT_CONN, &xprt->xpt_flags);
716 /* Get the next entry off the accept list */
717 spin_lock_bh(&listen_rdma->sc_lock);
718 if (!list_empty(&listen_rdma->sc_accept_q)) {
719 newxprt = list_entry(listen_rdma->sc_accept_q.next,
720 struct svcxprt_rdma, sc_accept_q);
721 list_del_init(&newxprt->sc_accept_q);
723 if (!list_empty(&listen_rdma->sc_accept_q))
724 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
725 spin_unlock_bh(&listen_rdma->sc_lock);
729 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
730 newxprt, newxprt->sc_cm_id);
732 dev = newxprt->sc_cm_id->device;
733 newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
735 /* Qualify the transport resource defaults with the
736 * capabilities of this particular device */
737 newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
738 (size_t)RPCSVC_MAXPAGES);
739 newxprt->sc_max_req_size = svcrdma_max_req_size;
740 newxprt->sc_max_requests = svcrdma_max_requests;
741 newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
742 newxprt->sc_rq_depth = newxprt->sc_max_requests +
743 newxprt->sc_max_bc_requests;
744 if (newxprt->sc_rq_depth > dev->attrs.max_qp_wr) {
745 pr_warn("svcrdma: reducing receive depth to %d\n",
746 dev->attrs.max_qp_wr);
747 newxprt->sc_rq_depth = dev->attrs.max_qp_wr;
748 newxprt->sc_max_requests = newxprt->sc_rq_depth - 2;
749 newxprt->sc_max_bc_requests = 2;
751 newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
752 ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES);
753 ctxts *= newxprt->sc_max_requests;
754 newxprt->sc_sq_depth = newxprt->sc_rq_depth + ctxts;
755 if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) {
756 pr_warn("svcrdma: reducing send depth to %d\n",
757 dev->attrs.max_qp_wr);
758 newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
760 atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
762 if (!svc_rdma_prealloc_ctxts(newxprt))
766 * Limit ORD based on client limit, local device limit, and
767 * configured svcrdma limit.
769 newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
770 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
772 newxprt->sc_pd = ib_alloc_pd(dev, 0);
773 if (IS_ERR(newxprt->sc_pd)) {
774 dprintk("svcrdma: error creating PD for connect request\n");
777 newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
778 0, IB_POLL_WORKQUEUE);
779 if (IS_ERR(newxprt->sc_sq_cq)) {
780 dprintk("svcrdma: error creating SQ CQ for connect request\n");
783 newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
784 0, IB_POLL_WORKQUEUE);
785 if (IS_ERR(newxprt->sc_rq_cq)) {
786 dprintk("svcrdma: error creating RQ CQ for connect request\n");
790 memset(&qp_attr, 0, sizeof qp_attr);
791 qp_attr.event_handler = qp_event_handler;
792 qp_attr.qp_context = &newxprt->sc_xprt;
793 qp_attr.port_num = newxprt->sc_port_num;
794 qp_attr.cap.max_rdma_ctxs = ctxts;
795 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
796 qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
797 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
798 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
799 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
800 qp_attr.qp_type = IB_QPT_RC;
801 qp_attr.send_cq = newxprt->sc_sq_cq;
802 qp_attr.recv_cq = newxprt->sc_rq_cq;
803 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
804 newxprt->sc_cm_id, newxprt->sc_pd);
805 dprintk(" cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
806 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
807 dprintk(" cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
808 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
810 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
812 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
815 newxprt->sc_qp = newxprt->sc_cm_id->qp;
817 if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
818 newxprt->sc_snd_w_inv = false;
819 if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
820 !rdma_ib_or_roce(dev, newxprt->sc_port_num))
823 /* Post receive buffers */
824 for (i = 0; i < newxprt->sc_max_requests; i++) {
825 ret = svc_rdma_post_recv(newxprt);
827 dprintk("svcrdma: failure posting receive buffers\n");
832 /* Swap out the handler */
833 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
835 /* Construct RDMA-CM private message */
836 pmsg.cp_magic = rpcrdma_cmp_magic;
837 pmsg.cp_version = RPCRDMA_CMP_VERSION;
839 pmsg.cp_send_size = pmsg.cp_recv_size =
840 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
842 /* Accept Connection */
843 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
844 memset(&conn_param, 0, sizeof conn_param);
845 conn_param.responder_resources = 0;
846 conn_param.initiator_depth = newxprt->sc_ord;
847 conn_param.private_data = &pmsg;
848 conn_param.private_data_len = sizeof(pmsg);
849 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
851 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
856 dprintk("svcrdma: new connection %p accepted:\n", newxprt);
857 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
858 dprintk(" local address : %pIS:%u\n", sap, rpc_get_port(sap));
859 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
860 dprintk(" remote address : %pIS:%u\n", sap, rpc_get_port(sap));
861 dprintk(" max_sge : %d\n", newxprt->sc_max_sge);
862 dprintk(" sq_depth : %d\n", newxprt->sc_sq_depth);
863 dprintk(" rdma_rw_ctxs : %d\n", ctxts);
864 dprintk(" max_requests : %d\n", newxprt->sc_max_requests);
865 dprintk(" ord : %d\n", newxprt->sc_ord);
867 return &newxprt->sc_xprt;
870 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
871 /* Take a reference in case the DTO handler runs */
872 svc_xprt_get(&newxprt->sc_xprt);
873 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
874 ib_destroy_qp(newxprt->sc_qp);
875 rdma_destroy_id(newxprt->sc_cm_id);
876 /* This call to put will destroy the transport */
877 svc_xprt_put(&newxprt->sc_xprt);
881 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
886 * When connected, an svc_xprt has at least two references:
888 * - A reference held by the cm_id between the ESTABLISHED and
889 * DISCONNECTED events. If the remote peer disconnected first, this
890 * reference could be gone.
892 * - A reference held by the svc_recv code that called this function
893 * as part of close processing.
895 * At a minimum one references should still be held.
897 static void svc_rdma_detach(struct svc_xprt *xprt)
899 struct svcxprt_rdma *rdma =
900 container_of(xprt, struct svcxprt_rdma, sc_xprt);
901 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
903 /* Disconnect and flush posted WQE */
904 rdma_disconnect(rdma->sc_cm_id);
907 static void __svc_rdma_free(struct work_struct *work)
909 struct svcxprt_rdma *rdma =
910 container_of(work, struct svcxprt_rdma, sc_work);
911 struct svc_xprt *xprt = &rdma->sc_xprt;
913 dprintk("svcrdma: %s(%p)\n", __func__, rdma);
915 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
916 ib_drain_qp(rdma->sc_qp);
918 /* We should only be called from kref_put */
919 if (kref_read(&xprt->xpt_ref) != 0)
920 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
921 kref_read(&xprt->xpt_ref));
923 while (!list_empty(&rdma->sc_read_complete_q)) {
924 struct svc_rdma_op_ctxt *ctxt;
925 ctxt = list_first_entry(&rdma->sc_read_complete_q,
926 struct svc_rdma_op_ctxt, list);
927 list_del(&ctxt->list);
928 svc_rdma_put_context(ctxt, 1);
930 while (!list_empty(&rdma->sc_rq_dto_q)) {
931 struct svc_rdma_op_ctxt *ctxt;
932 ctxt = list_first_entry(&rdma->sc_rq_dto_q,
933 struct svc_rdma_op_ctxt, list);
934 list_del(&ctxt->list);
935 svc_rdma_put_context(ctxt, 1);
938 /* Warn if we leaked a resource or under-referenced */
939 if (rdma->sc_ctxt_used != 0)
940 pr_err("svcrdma: ctxt still in use? (%d)\n",
943 /* Final put of backchannel client transport */
944 if (xprt->xpt_bc_xprt) {
945 xprt_put(xprt->xpt_bc_xprt);
946 xprt->xpt_bc_xprt = NULL;
949 svc_rdma_destroy_rw_ctxts(rdma);
950 svc_rdma_destroy_ctxts(rdma);
952 /* Destroy the QP if present (not a listener) */
953 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
954 ib_destroy_qp(rdma->sc_qp);
956 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
957 ib_free_cq(rdma->sc_sq_cq);
959 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
960 ib_free_cq(rdma->sc_rq_cq);
962 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
963 ib_dealloc_pd(rdma->sc_pd);
965 /* Destroy the CM ID */
966 rdma_destroy_id(rdma->sc_cm_id);
971 static void svc_rdma_free(struct svc_xprt *xprt)
973 struct svcxprt_rdma *rdma =
974 container_of(xprt, struct svcxprt_rdma, sc_xprt);
975 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
976 queue_work(svc_rdma_wq, &rdma->sc_work);
979 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
981 struct svcxprt_rdma *rdma =
982 container_of(xprt, struct svcxprt_rdma, sc_xprt);
985 * If there are already waiters on the SQ,
988 if (waitqueue_active(&rdma->sc_send_wait))
991 /* Otherwise return true. */
995 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1000 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1004 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1006 struct ib_send_wr *bad_wr, *n_wr;
1011 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1015 for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1018 /* If the SQ is full, wait until an SQ entry is available */
1020 if ((atomic_sub_return(wr_count, &xprt->sc_sq_avail) < 0)) {
1021 atomic_inc(&rdma_stat_sq_starve);
1023 /* Wait until SQ WR available if SQ still full */
1024 atomic_add(wr_count, &xprt->sc_sq_avail);
1025 wait_event(xprt->sc_send_wait,
1026 atomic_read(&xprt->sc_sq_avail) > wr_count);
1027 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1031 /* Take a transport ref for each WR posted */
1032 for (i = 0; i < wr_count; i++)
1033 svc_xprt_get(&xprt->sc_xprt);
1035 /* Bump used SQ WR count and post */
1036 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1038 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1039 for (i = 0; i < wr_count; i ++)
1040 svc_xprt_put(&xprt->sc_xprt);
1041 dprintk("svcrdma: failed to post SQ WR rc=%d\n", ret);
1042 dprintk(" sc_sq_avail=%d, sc_sq_depth=%d\n",
1043 atomic_read(&xprt->sc_sq_avail),
1045 wake_up(&xprt->sc_send_wait);