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;
209 /* Either pre-allocation missed the mark, or send
210 * queue accounting is broken.
212 spin_unlock(&xprt->sc_ctxt_lock);
214 ctxt = alloc_ctxt(xprt, GFP_NOIO);
218 spin_lock(&xprt->sc_ctxt_lock);
219 xprt->sc_ctxt_used--;
220 spin_unlock(&xprt->sc_ctxt_lock);
221 WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
225 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
227 struct svcxprt_rdma *xprt = ctxt->xprt;
228 struct ib_device *device = xprt->sc_cm_id->device;
229 u32 lkey = xprt->sc_pd->local_dma_lkey;
232 for (i = 0; i < ctxt->mapped_sges; i++) {
234 * Unmap the DMA addr in the SGE if the lkey matches
235 * the local_dma_lkey, otherwise, ignore it since it is
236 * an FRMR lkey and will be unmapped later when the
237 * last WR that uses it completes.
239 if (ctxt->sge[i].lkey == lkey)
240 ib_dma_unmap_page(device,
245 ctxt->mapped_sges = 0;
248 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
250 struct svcxprt_rdma *xprt = ctxt->xprt;
254 for (i = 0; i < ctxt->count; i++)
255 put_page(ctxt->pages[i]);
257 spin_lock(&xprt->sc_ctxt_lock);
258 xprt->sc_ctxt_used--;
259 list_add(&ctxt->list, &xprt->sc_ctxts);
260 spin_unlock(&xprt->sc_ctxt_lock);
263 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
265 while (!list_empty(&xprt->sc_ctxts)) {
266 struct svc_rdma_op_ctxt *ctxt;
268 ctxt = list_first_entry(&xprt->sc_ctxts,
269 struct svc_rdma_op_ctxt, list);
270 list_del(&ctxt->list);
275 static struct svc_rdma_req_map *alloc_req_map(gfp_t flags)
277 struct svc_rdma_req_map *map;
279 map = kmalloc(sizeof(*map), flags);
281 INIT_LIST_HEAD(&map->free);
285 static bool svc_rdma_prealloc_maps(struct svcxprt_rdma *xprt)
289 /* One for each receive buffer on this connection. */
290 i = xprt->sc_max_requests;
293 struct svc_rdma_req_map *map;
295 map = alloc_req_map(GFP_KERNEL);
297 dprintk("svcrdma: No memory for request map\n");
300 list_add(&map->free, &xprt->sc_maps);
305 struct svc_rdma_req_map *svc_rdma_get_req_map(struct svcxprt_rdma *xprt)
307 struct svc_rdma_req_map *map = NULL;
309 spin_lock(&xprt->sc_map_lock);
310 if (list_empty(&xprt->sc_maps))
313 map = list_first_entry(&xprt->sc_maps,
314 struct svc_rdma_req_map, free);
315 list_del_init(&map->free);
316 spin_unlock(&xprt->sc_map_lock);
323 spin_unlock(&xprt->sc_map_lock);
325 /* Pre-allocation amount was incorrect */
326 map = alloc_req_map(GFP_NOIO);
330 WARN_ONCE(1, "svcrdma: empty request map list?\n");
334 void svc_rdma_put_req_map(struct svcxprt_rdma *xprt,
335 struct svc_rdma_req_map *map)
337 spin_lock(&xprt->sc_map_lock);
338 list_add(&map->free, &xprt->sc_maps);
339 spin_unlock(&xprt->sc_map_lock);
342 static void svc_rdma_destroy_maps(struct svcxprt_rdma *xprt)
344 while (!list_empty(&xprt->sc_maps)) {
345 struct svc_rdma_req_map *map;
347 map = list_first_entry(&xprt->sc_maps,
348 struct svc_rdma_req_map, free);
349 list_del(&map->free);
354 /* QP event handler */
355 static void qp_event_handler(struct ib_event *event, void *context)
357 struct svc_xprt *xprt = context;
359 switch (event->event) {
360 /* These are considered benign events */
361 case IB_EVENT_PATH_MIG:
362 case IB_EVENT_COMM_EST:
363 case IB_EVENT_SQ_DRAINED:
364 case IB_EVENT_QP_LAST_WQE_REACHED:
365 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
366 ib_event_msg(event->event), event->event,
369 /* These are considered fatal events */
370 case IB_EVENT_PATH_MIG_ERR:
371 case IB_EVENT_QP_FATAL:
372 case IB_EVENT_QP_REQ_ERR:
373 case IB_EVENT_QP_ACCESS_ERR:
374 case IB_EVENT_DEVICE_FATAL:
376 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
377 "closing transport\n",
378 ib_event_msg(event->event), event->event,
380 set_bit(XPT_CLOSE, &xprt->xpt_flags);
386 * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
387 * @cq: completion queue
391 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
393 struct svcxprt_rdma *xprt = cq->cq_context;
394 struct ib_cqe *cqe = wc->wr_cqe;
395 struct svc_rdma_op_ctxt *ctxt;
397 /* WARNING: Only wc->wr_cqe and wc->status are reliable */
398 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
399 svc_rdma_unmap_dma(ctxt);
401 if (wc->status != IB_WC_SUCCESS)
404 /* All wc fields are now known to be valid */
405 ctxt->byte_len = wc->byte_len;
406 spin_lock(&xprt->sc_rq_dto_lock);
407 list_add_tail(&ctxt->list, &xprt->sc_rq_dto_q);
408 spin_unlock(&xprt->sc_rq_dto_lock);
410 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
411 if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
413 svc_xprt_enqueue(&xprt->sc_xprt);
417 if (wc->status != IB_WC_WR_FLUSH_ERR)
418 pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
419 ib_wc_status_msg(wc->status),
420 wc->status, wc->vendor_err);
421 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
422 svc_rdma_put_context(ctxt, 1);
425 svc_xprt_put(&xprt->sc_xprt);
428 static void svc_rdma_send_wc_common(struct svcxprt_rdma *xprt,
432 if (wc->status != IB_WC_SUCCESS)
436 atomic_inc(&xprt->sc_sq_avail);
437 wake_up(&xprt->sc_send_wait);
441 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
442 if (wc->status != IB_WC_WR_FLUSH_ERR)
443 pr_err("svcrdma: %s: %s (%u/0x%x)\n",
444 opname, ib_wc_status_msg(wc->status),
445 wc->status, wc->vendor_err);
449 static void svc_rdma_send_wc_common_put(struct ib_cq *cq, struct ib_wc *wc,
452 struct svcxprt_rdma *xprt = cq->cq_context;
454 svc_rdma_send_wc_common(xprt, wc, opname);
455 svc_xprt_put(&xprt->sc_xprt);
459 * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
460 * @cq: completion queue
464 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
466 struct ib_cqe *cqe = wc->wr_cqe;
467 struct svc_rdma_op_ctxt *ctxt;
469 svc_rdma_send_wc_common_put(cq, wc, "send");
471 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
472 svc_rdma_unmap_dma(ctxt);
473 svc_rdma_put_context(ctxt, 1);
477 * svc_rdma_wc_write - Invoked by RDMA provider for each polled Write WC
478 * @cq: completion queue
482 void svc_rdma_wc_write(struct ib_cq *cq, struct ib_wc *wc)
484 struct ib_cqe *cqe = wc->wr_cqe;
485 struct svc_rdma_op_ctxt *ctxt;
487 svc_rdma_send_wc_common_put(cq, wc, "write");
489 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
490 svc_rdma_unmap_dma(ctxt);
491 svc_rdma_put_context(ctxt, 0);
495 * svc_rdma_wc_reg - Invoked by RDMA provider for each polled FASTREG WC
496 * @cq: completion queue
500 void svc_rdma_wc_reg(struct ib_cq *cq, struct ib_wc *wc)
502 svc_rdma_send_wc_common_put(cq, wc, "fastreg");
506 * svc_rdma_wc_read - Invoked by RDMA provider for each polled Read WC
507 * @cq: completion queue
511 void svc_rdma_wc_read(struct ib_cq *cq, struct ib_wc *wc)
513 struct svcxprt_rdma *xprt = cq->cq_context;
514 struct ib_cqe *cqe = wc->wr_cqe;
515 struct svc_rdma_op_ctxt *ctxt;
517 svc_rdma_send_wc_common(xprt, wc, "read");
519 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
520 svc_rdma_unmap_dma(ctxt);
521 svc_rdma_put_frmr(xprt, ctxt->frmr);
523 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
524 struct svc_rdma_op_ctxt *read_hdr;
526 read_hdr = ctxt->read_hdr;
527 spin_lock(&xprt->sc_rq_dto_lock);
528 list_add_tail(&read_hdr->list,
529 &xprt->sc_read_complete_q);
530 spin_unlock(&xprt->sc_rq_dto_lock);
532 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
533 svc_xprt_enqueue(&xprt->sc_xprt);
536 svc_rdma_put_context(ctxt, 0);
537 svc_xprt_put(&xprt->sc_xprt);
541 * svc_rdma_wc_inv - Invoked by RDMA provider for each polled LOCAL_INV WC
542 * @cq: completion queue
546 void svc_rdma_wc_inv(struct ib_cq *cq, struct ib_wc *wc)
548 svc_rdma_send_wc_common_put(cq, wc, "localInv");
551 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
554 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
558 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
559 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
560 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
561 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
562 INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
563 INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
564 INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
565 INIT_LIST_HEAD(&cma_xprt->sc_maps);
566 init_waitqueue_head(&cma_xprt->sc_send_wait);
568 spin_lock_init(&cma_xprt->sc_lock);
569 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
570 spin_lock_init(&cma_xprt->sc_frmr_q_lock);
571 spin_lock_init(&cma_xprt->sc_ctxt_lock);
572 spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
573 spin_lock_init(&cma_xprt->sc_map_lock);
576 * Note that this implies that the underlying transport support
577 * has some form of congestion control (see RFC 7530 section 3.1
578 * paragraph 2). For now, we assume that all supported RDMA
579 * transports are suitable here.
581 set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
584 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
589 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
591 struct ib_recv_wr recv_wr, *bad_recv_wr;
592 struct svc_rdma_op_ctxt *ctxt;
599 ctxt = svc_rdma_get_context(xprt);
601 ctxt->direction = DMA_FROM_DEVICE;
602 ctxt->cqe.done = svc_rdma_wc_receive;
603 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
604 if (sge_no >= xprt->sc_max_sge) {
605 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
608 page = alloc_page(flags);
611 ctxt->pages[sge_no] = page;
612 pa = ib_dma_map_page(xprt->sc_cm_id->device,
615 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
617 svc_rdma_count_mappings(xprt, ctxt);
618 ctxt->sge[sge_no].addr = pa;
619 ctxt->sge[sge_no].length = PAGE_SIZE;
620 ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
621 ctxt->count = sge_no + 1;
625 recv_wr.sg_list = &ctxt->sge[0];
626 recv_wr.num_sge = ctxt->count;
627 recv_wr.wr_cqe = &ctxt->cqe;
629 svc_xprt_get(&xprt->sc_xprt);
630 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
632 svc_rdma_unmap_dma(ctxt);
633 svc_rdma_put_context(ctxt, 1);
634 svc_xprt_put(&xprt->sc_xprt);
639 svc_rdma_unmap_dma(ctxt);
640 svc_rdma_put_context(ctxt, 1);
644 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags)
648 ret = svc_rdma_post_recv(xprt, flags);
650 pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
652 pr_err("svcrdma: closing transport %p.\n", xprt);
653 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
660 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
661 struct rdma_conn_param *param)
663 const struct rpcrdma_connect_private *pmsg = param->private_data;
666 pmsg->cp_magic == rpcrdma_cmp_magic &&
667 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
668 newxprt->sc_snd_w_inv = pmsg->cp_flags &
669 RPCRDMA_CMP_F_SND_W_INV_OK;
671 dprintk("svcrdma: client send_size %u, recv_size %u "
672 "remote inv %ssupported\n",
673 rpcrdma_decode_buffer_size(pmsg->cp_send_size),
674 rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
675 newxprt->sc_snd_w_inv ? "" : "un");
680 * This function handles the CONNECT_REQUEST event on a listening
681 * endpoint. It is passed the cma_id for the _new_ connection. The context in
682 * this cma_id is inherited from the listening cma_id and is the svc_xprt
683 * structure for the listening endpoint.
685 * This function creates a new xprt for the new connection and enqueues it on
686 * the accept queue for the listent xprt. When the listen thread is kicked, it
687 * will call the recvfrom method on the listen xprt which will accept the new
690 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
691 struct rdma_conn_param *param)
693 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
694 struct svcxprt_rdma *newxprt;
697 /* Create a new transport */
698 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
700 dprintk("svcrdma: failed to create new transport\n");
703 newxprt->sc_cm_id = new_cma_id;
704 new_cma_id->context = newxprt;
705 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
706 newxprt, newxprt->sc_cm_id, listen_xprt);
707 svc_rdma_parse_connect_private(newxprt, param);
709 /* Save client advertised inbound read limit for use later in accept. */
710 newxprt->sc_ord = param->initiator_depth;
712 /* Set the local and remote addresses in the transport */
713 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
714 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
715 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
716 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
719 * Enqueue the new transport on the accept queue of the listening
722 spin_lock_bh(&listen_xprt->sc_lock);
723 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
724 spin_unlock_bh(&listen_xprt->sc_lock);
726 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
727 svc_xprt_enqueue(&listen_xprt->sc_xprt);
731 * Handles events generated on the listening endpoint. These events will be
732 * either be incoming connect requests or adapter removal events.
734 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
735 struct rdma_cm_event *event)
737 struct svcxprt_rdma *xprt = cma_id->context;
740 switch (event->event) {
741 case RDMA_CM_EVENT_CONNECT_REQUEST:
742 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
743 "event = %s (%d)\n", cma_id, cma_id->context,
744 rdma_event_msg(event->event), event->event);
745 handle_connect_req(cma_id, &event->param.conn);
748 case RDMA_CM_EVENT_ESTABLISHED:
749 /* Accept complete */
750 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
751 "cm_id=%p\n", xprt, cma_id);
754 case RDMA_CM_EVENT_DEVICE_REMOVAL:
755 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
758 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
762 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
763 "event = %s (%d)\n", cma_id,
764 rdma_event_msg(event->event), event->event);
771 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
772 struct rdma_cm_event *event)
774 struct svc_xprt *xprt = cma_id->context;
775 struct svcxprt_rdma *rdma =
776 container_of(xprt, struct svcxprt_rdma, sc_xprt);
777 switch (event->event) {
778 case RDMA_CM_EVENT_ESTABLISHED:
779 /* Accept complete */
781 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
782 "cm_id=%p\n", xprt, cma_id);
783 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
784 svc_xprt_enqueue(xprt);
786 case RDMA_CM_EVENT_DISCONNECTED:
787 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
790 set_bit(XPT_CLOSE, &xprt->xpt_flags);
791 svc_xprt_enqueue(xprt);
795 case RDMA_CM_EVENT_DEVICE_REMOVAL:
796 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
797 "event = %s (%d)\n", cma_id, xprt,
798 rdma_event_msg(event->event), event->event);
800 set_bit(XPT_CLOSE, &xprt->xpt_flags);
801 svc_xprt_enqueue(xprt);
806 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
807 "event = %s (%d)\n", cma_id,
808 rdma_event_msg(event->event), event->event);
815 * Create a listening RDMA service endpoint.
817 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
819 struct sockaddr *sa, int salen,
822 struct rdma_cm_id *listen_id;
823 struct svcxprt_rdma *cma_xprt;
826 dprintk("svcrdma: Creating RDMA socket\n");
827 if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
828 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
829 return ERR_PTR(-EAFNOSUPPORT);
831 cma_xprt = rdma_create_xprt(serv, 1);
833 return ERR_PTR(-ENOMEM);
835 listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
836 RDMA_PS_TCP, IB_QPT_RC);
837 if (IS_ERR(listen_id)) {
838 ret = PTR_ERR(listen_id);
839 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
843 /* Allow both IPv4 and IPv6 sockets to bind a single port
846 #if IS_ENABLED(CONFIG_IPV6)
847 ret = rdma_set_afonly(listen_id, 1);
849 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
853 ret = rdma_bind_addr(listen_id, sa);
855 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
858 cma_xprt->sc_cm_id = listen_id;
860 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
862 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
867 * We need to use the address from the cm_id in case the
868 * caller specified 0 for the port number.
870 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
871 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
873 return &cma_xprt->sc_xprt;
876 rdma_destroy_id(listen_id);
882 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
885 struct scatterlist *sg;
886 struct svc_rdma_fastreg_mr *frmr;
889 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
893 num_sg = min_t(u32, RPCSVC_MAXPAGES, xprt->sc_frmr_pg_list_len);
894 mr = ib_alloc_mr(xprt->sc_pd, IB_MR_TYPE_MEM_REG, num_sg);
898 sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL);
902 sg_init_table(sg, RPCSVC_MAXPAGES);
906 INIT_LIST_HEAD(&frmr->frmr_list);
914 return ERR_PTR(-ENOMEM);
917 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
919 struct svc_rdma_fastreg_mr *frmr;
921 while (!list_empty(&xprt->sc_frmr_q)) {
922 frmr = list_entry(xprt->sc_frmr_q.next,
923 struct svc_rdma_fastreg_mr, frmr_list);
924 list_del_init(&frmr->frmr_list);
926 ib_dereg_mr(frmr->mr);
931 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
933 struct svc_rdma_fastreg_mr *frmr = NULL;
935 spin_lock(&rdma->sc_frmr_q_lock);
936 if (!list_empty(&rdma->sc_frmr_q)) {
937 frmr = list_entry(rdma->sc_frmr_q.next,
938 struct svc_rdma_fastreg_mr, frmr_list);
939 list_del_init(&frmr->frmr_list);
942 spin_unlock(&rdma->sc_frmr_q_lock);
946 return rdma_alloc_frmr(rdma);
949 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
950 struct svc_rdma_fastreg_mr *frmr)
953 ib_dma_unmap_sg(rdma->sc_cm_id->device,
954 frmr->sg, frmr->sg_nents, frmr->direction);
955 spin_lock(&rdma->sc_frmr_q_lock);
956 WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
957 list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
958 spin_unlock(&rdma->sc_frmr_q_lock);
963 * This is the xpo_recvfrom function for listening endpoints. Its
964 * purpose is to accept incoming connections. The CMA callback handler
965 * has already created a new transport and attached it to the new CMA
968 * There is a queue of pending connections hung on the listening
969 * transport. This queue contains the new svc_xprt structure. This
970 * function takes svc_xprt structures off the accept_q and completes
973 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
975 struct svcxprt_rdma *listen_rdma;
976 struct svcxprt_rdma *newxprt = NULL;
977 struct rdma_conn_param conn_param;
978 struct rpcrdma_connect_private pmsg;
979 struct ib_qp_init_attr qp_attr;
980 struct ib_device *dev;
981 struct sockaddr *sap;
985 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
986 clear_bit(XPT_CONN, &xprt->xpt_flags);
987 /* Get the next entry off the accept list */
988 spin_lock_bh(&listen_rdma->sc_lock);
989 if (!list_empty(&listen_rdma->sc_accept_q)) {
990 newxprt = list_entry(listen_rdma->sc_accept_q.next,
991 struct svcxprt_rdma, sc_accept_q);
992 list_del_init(&newxprt->sc_accept_q);
994 if (!list_empty(&listen_rdma->sc_accept_q))
995 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
996 spin_unlock_bh(&listen_rdma->sc_lock);
1000 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
1001 newxprt, newxprt->sc_cm_id);
1003 dev = newxprt->sc_cm_id->device;
1004 newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
1006 /* Qualify the transport resource defaults with the
1007 * capabilities of this particular device */
1008 newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
1009 (size_t)RPCSVC_MAXPAGES);
1010 newxprt->sc_max_sge_rd = min_t(size_t, dev->attrs.max_sge_rd,
1012 newxprt->sc_max_req_size = svcrdma_max_req_size;
1013 newxprt->sc_max_requests = min_t(u32, dev->attrs.max_qp_wr,
1014 svcrdma_max_requests);
1015 newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
1016 newxprt->sc_max_bc_requests = min_t(u32, dev->attrs.max_qp_wr,
1017 svcrdma_max_bc_requests);
1018 newxprt->sc_rq_depth = newxprt->sc_max_requests +
1019 newxprt->sc_max_bc_requests;
1020 newxprt->sc_sq_depth = newxprt->sc_rq_depth;
1021 atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
1023 if (!svc_rdma_prealloc_ctxts(newxprt))
1025 if (!svc_rdma_prealloc_maps(newxprt))
1029 * Limit ORD based on client limit, local device limit, and
1030 * configured svcrdma limit.
1032 newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
1033 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
1035 newxprt->sc_pd = ib_alloc_pd(dev, 0);
1036 if (IS_ERR(newxprt->sc_pd)) {
1037 dprintk("svcrdma: error creating PD for connect request\n");
1040 newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
1041 0, IB_POLL_WORKQUEUE);
1042 if (IS_ERR(newxprt->sc_sq_cq)) {
1043 dprintk("svcrdma: error creating SQ CQ for connect request\n");
1046 newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
1047 0, IB_POLL_WORKQUEUE);
1048 if (IS_ERR(newxprt->sc_rq_cq)) {
1049 dprintk("svcrdma: error creating RQ CQ for connect request\n");
1053 memset(&qp_attr, 0, sizeof qp_attr);
1054 qp_attr.event_handler = qp_event_handler;
1055 qp_attr.qp_context = &newxprt->sc_xprt;
1056 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
1057 qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
1058 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
1059 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
1060 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1061 qp_attr.qp_type = IB_QPT_RC;
1062 qp_attr.send_cq = newxprt->sc_sq_cq;
1063 qp_attr.recv_cq = newxprt->sc_rq_cq;
1064 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
1065 newxprt->sc_cm_id, newxprt->sc_pd);
1066 dprintk(" cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
1067 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
1068 dprintk(" cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
1069 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
1071 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
1073 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
1076 newxprt->sc_qp = newxprt->sc_cm_id->qp;
1079 * Use the most secure set of MR resources based on the
1080 * transport type and available memory management features in
1081 * the device. Here's the table implemented below:
1083 * Fast Global DMA Remote WR
1084 * Reg LKEY MR Access
1085 * Sup'd Sup'd Needed Needed
1097 * NB: iWARP requires remote write access for the data sink
1098 * of an RDMA_READ. IB does not.
1100 newxprt->sc_reader = rdma_read_chunk_lcl;
1101 if (dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1102 newxprt->sc_frmr_pg_list_len =
1103 dev->attrs.max_fast_reg_page_list_len;
1104 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
1105 newxprt->sc_reader = rdma_read_chunk_frmr;
1107 newxprt->sc_snd_w_inv = false;
1110 * Determine if a DMA MR is required and if so, what privs are required
1112 if (!rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num) &&
1113 !rdma_ib_or_roce(dev, newxprt->sc_cm_id->port_num))
1116 if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num))
1117 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
1119 /* Post receive buffers */
1120 for (i = 0; i < newxprt->sc_max_requests; i++) {
1121 ret = svc_rdma_post_recv(newxprt, GFP_KERNEL);
1123 dprintk("svcrdma: failure posting receive buffers\n");
1128 /* Swap out the handler */
1129 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
1131 /* Construct RDMA-CM private message */
1132 pmsg.cp_magic = rpcrdma_cmp_magic;
1133 pmsg.cp_version = RPCRDMA_CMP_VERSION;
1135 pmsg.cp_send_size = pmsg.cp_recv_size =
1136 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
1138 /* Accept Connection */
1139 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
1140 memset(&conn_param, 0, sizeof conn_param);
1141 conn_param.responder_resources = 0;
1142 conn_param.initiator_depth = newxprt->sc_ord;
1143 conn_param.private_data = &pmsg;
1144 conn_param.private_data_len = sizeof(pmsg);
1145 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
1147 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1152 dprintk("svcrdma: new connection %p accepted:\n", newxprt);
1153 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
1154 dprintk(" local address : %pIS:%u\n", sap, rpc_get_port(sap));
1155 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
1156 dprintk(" remote address : %pIS:%u\n", sap, rpc_get_port(sap));
1157 dprintk(" max_sge : %d\n", newxprt->sc_max_sge);
1158 dprintk(" max_sge_rd : %d\n", newxprt->sc_max_sge_rd);
1159 dprintk(" sq_depth : %d\n", newxprt->sc_sq_depth);
1160 dprintk(" max_requests : %d\n", newxprt->sc_max_requests);
1161 dprintk(" ord : %d\n", newxprt->sc_ord);
1163 return &newxprt->sc_xprt;
1166 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
1167 /* Take a reference in case the DTO handler runs */
1168 svc_xprt_get(&newxprt->sc_xprt);
1169 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
1170 ib_destroy_qp(newxprt->sc_qp);
1171 rdma_destroy_id(newxprt->sc_cm_id);
1172 /* This call to put will destroy the transport */
1173 svc_xprt_put(&newxprt->sc_xprt);
1177 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
1182 * When connected, an svc_xprt has at least two references:
1184 * - A reference held by the cm_id between the ESTABLISHED and
1185 * DISCONNECTED events. If the remote peer disconnected first, this
1186 * reference could be gone.
1188 * - A reference held by the svc_recv code that called this function
1189 * as part of close processing.
1191 * At a minimum one references should still be held.
1193 static void svc_rdma_detach(struct svc_xprt *xprt)
1195 struct svcxprt_rdma *rdma =
1196 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1197 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1199 /* Disconnect and flush posted WQE */
1200 rdma_disconnect(rdma->sc_cm_id);
1203 static void __svc_rdma_free(struct work_struct *work)
1205 struct svcxprt_rdma *rdma =
1206 container_of(work, struct svcxprt_rdma, sc_work);
1207 struct svc_xprt *xprt = &rdma->sc_xprt;
1209 dprintk("svcrdma: %s(%p)\n", __func__, rdma);
1211 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1212 ib_drain_qp(rdma->sc_qp);
1214 /* We should only be called from kref_put */
1215 if (kref_read(&xprt->xpt_ref) != 0)
1216 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1217 kref_read(&xprt->xpt_ref));
1220 * Destroy queued, but not processed read completions. Note
1221 * that this cleanup has to be done before destroying the
1222 * cm_id because the device ptr is needed to unmap the dma in
1223 * svc_rdma_put_context.
1225 while (!list_empty(&rdma->sc_read_complete_q)) {
1226 struct svc_rdma_op_ctxt *ctxt;
1227 ctxt = list_first_entry(&rdma->sc_read_complete_q,
1228 struct svc_rdma_op_ctxt, list);
1229 list_del(&ctxt->list);
1230 svc_rdma_put_context(ctxt, 1);
1233 /* Destroy queued, but not processed recv completions */
1234 while (!list_empty(&rdma->sc_rq_dto_q)) {
1235 struct svc_rdma_op_ctxt *ctxt;
1236 ctxt = list_first_entry(&rdma->sc_rq_dto_q,
1237 struct svc_rdma_op_ctxt, list);
1238 list_del(&ctxt->list);
1239 svc_rdma_put_context(ctxt, 1);
1242 /* Warn if we leaked a resource or under-referenced */
1243 if (rdma->sc_ctxt_used != 0)
1244 pr_err("svcrdma: ctxt still in use? (%d)\n",
1245 rdma->sc_ctxt_used);
1247 /* Final put of backchannel client transport */
1248 if (xprt->xpt_bc_xprt) {
1249 xprt_put(xprt->xpt_bc_xprt);
1250 xprt->xpt_bc_xprt = NULL;
1253 rdma_dealloc_frmr_q(rdma);
1254 svc_rdma_destroy_rw_ctxts(rdma);
1255 svc_rdma_destroy_ctxts(rdma);
1256 svc_rdma_destroy_maps(rdma);
1258 /* Destroy the QP if present (not a listener) */
1259 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1260 ib_destroy_qp(rdma->sc_qp);
1262 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1263 ib_free_cq(rdma->sc_sq_cq);
1265 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1266 ib_free_cq(rdma->sc_rq_cq);
1268 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1269 ib_dealloc_pd(rdma->sc_pd);
1271 /* Destroy the CM ID */
1272 rdma_destroy_id(rdma->sc_cm_id);
1277 static void svc_rdma_free(struct svc_xprt *xprt)
1279 struct svcxprt_rdma *rdma =
1280 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1281 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1282 queue_work(svc_rdma_wq, &rdma->sc_work);
1285 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1287 struct svcxprt_rdma *rdma =
1288 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1291 * If there are already waiters on the SQ,
1294 if (waitqueue_active(&rdma->sc_send_wait))
1297 /* Otherwise return true. */
1301 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1306 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1310 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1312 struct ib_send_wr *bad_wr, *n_wr;
1317 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1321 for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1324 /* If the SQ is full, wait until an SQ entry is available */
1326 if ((atomic_sub_return(wr_count, &xprt->sc_sq_avail) < 0)) {
1327 atomic_inc(&rdma_stat_sq_starve);
1329 /* Wait until SQ WR available if SQ still full */
1330 atomic_add(wr_count, &xprt->sc_sq_avail);
1331 wait_event(xprt->sc_send_wait,
1332 atomic_read(&xprt->sc_sq_avail) > wr_count);
1333 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1337 /* Take a transport ref for each WR posted */
1338 for (i = 0; i < wr_count; i++)
1339 svc_xprt_get(&xprt->sc_xprt);
1341 /* Bump used SQ WR count and post */
1342 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1344 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1345 for (i = 0; i < wr_count; i ++)
1346 svc_xprt_put(&xprt->sc_xprt);
1347 dprintk("svcrdma: failed to post SQ WR rc=%d\n", ret);
1348 dprintk(" sc_sq_avail=%d, sc_sq_depth=%d\n",
1349 atomic_read(&xprt->sc_sq_avail),
1351 wake_up(&xprt->sc_send_wait);