1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
19 * Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials provided
22 * with the distribution.
24 * Neither the name of the Network Appliance, Inc. nor the names of
25 * its contributors may be used to endorse or promote products
26 * derived from this software without specific prior written
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 * Encapsulates the major functions managing:
52 #include <linux/interrupt.h>
53 #include <linux/slab.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/sunrpc/svc_rdma.h>
57 #include <asm-generic/barrier.h>
58 #include <asm/bitops.h>
60 #include <rdma/ib_cm.h>
62 #include "xprt_rdma.h"
63 #include <trace/events/rpcrdma.h>
69 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
70 # define RPCDBG_FACILITY RPCDBG_TRANS
76 static void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
77 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
79 static int rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp);
80 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
82 struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
85 rpcrdma_alloc_wq(void)
87 struct workqueue_struct *recv_wq;
89 recv_wq = alloc_workqueue("xprtrdma_receive",
90 WQ_MEM_RECLAIM | WQ_HIGHPRI,
95 rpcrdma_receive_wq = recv_wq;
100 rpcrdma_destroy_wq(void)
102 struct workqueue_struct *wq;
104 if (rpcrdma_receive_wq) {
105 wq = rpcrdma_receive_wq;
106 rpcrdma_receive_wq = NULL;
107 destroy_workqueue(wq);
112 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
114 struct rpcrdma_ep *ep = context;
115 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
118 trace_xprtrdma_qp_error(r_xprt, event);
119 pr_err("rpcrdma: %s on device %s ep %p\n",
120 ib_event_msg(event->event), event->device->name, context);
122 if (ep->rep_connected == 1) {
123 ep->rep_connected = -EIO;
124 rpcrdma_conn_func(ep);
125 wake_up_all(&ep->rep_connect_wait);
130 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
131 * @cq: completion queue (ignored)
136 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
138 struct ib_cqe *cqe = wc->wr_cqe;
139 struct rpcrdma_sendctx *sc =
140 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
142 /* WARNING: Only wr_cqe and status are reliable at this point */
143 trace_xprtrdma_wc_send(sc, wc);
144 if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
145 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
146 ib_wc_status_msg(wc->status),
147 wc->status, wc->vendor_err);
149 rpcrdma_sendctx_put_locked(sc);
153 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
154 * @cq: completion queue (ignored)
159 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
161 struct ib_cqe *cqe = wc->wr_cqe;
162 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
165 /* WARNING: Only wr_id and status are reliable at this point */
166 trace_xprtrdma_wc_receive(wc);
167 if (wc->status != IB_WC_SUCCESS)
170 /* status == SUCCESS means all fields in wc are trustworthy */
171 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
172 rep->rr_wc_flags = wc->wc_flags;
173 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
175 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
176 rdmab_addr(rep->rr_rdmabuf),
177 wc->byte_len, DMA_FROM_DEVICE);
180 rpcrdma_reply_handler(rep);
184 if (wc->status != IB_WC_WR_FLUSH_ERR)
185 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
186 ib_wc_status_msg(wc->status),
187 wc->status, wc->vendor_err);
188 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
193 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
194 struct rdma_conn_param *param)
196 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
197 const struct rpcrdma_connect_private *pmsg = param->private_data;
198 unsigned int rsize, wsize;
200 /* Default settings for RPC-over-RDMA Version One */
201 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
202 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
203 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
206 pmsg->cp_magic == rpcrdma_cmp_magic &&
207 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
208 r_xprt->rx_ia.ri_implicit_roundup = true;
209 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
210 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
213 if (rsize < cdata->inline_rsize)
214 cdata->inline_rsize = rsize;
215 if (wsize < cdata->inline_wsize)
216 cdata->inline_wsize = wsize;
217 dprintk("RPC: %s: max send %u, max recv %u\n",
218 __func__, cdata->inline_wsize, cdata->inline_rsize);
219 rpcrdma_set_max_header_sizes(r_xprt);
223 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
225 struct rpcrdma_xprt *xprt = id->context;
226 struct rpcrdma_ia *ia = &xprt->rx_ia;
227 struct rpcrdma_ep *ep = &xprt->rx_ep;
230 trace_xprtrdma_conn_upcall(xprt, event);
231 switch (event->event) {
232 case RDMA_CM_EVENT_ADDR_RESOLVED:
233 case RDMA_CM_EVENT_ROUTE_RESOLVED:
235 complete(&ia->ri_done);
237 case RDMA_CM_EVENT_ADDR_ERROR:
238 ia->ri_async_rc = -EPROTO;
239 complete(&ia->ri_done);
241 case RDMA_CM_EVENT_ROUTE_ERROR:
242 ia->ri_async_rc = -ENETUNREACH;
243 complete(&ia->ri_done);
245 case RDMA_CM_EVENT_DEVICE_REMOVAL:
246 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
247 pr_info("rpcrdma: removing device %s for %s:%s\n",
249 rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
251 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
252 ep->rep_connected = -ENODEV;
253 xprt_force_disconnect(&xprt->rx_xprt);
254 wait_for_completion(&ia->ri_remove_done);
257 ia->ri_device = NULL;
258 /* Return 1 to ensure the core destroys the id. */
260 case RDMA_CM_EVENT_ESTABLISHED:
261 ++xprt->rx_xprt.connect_cookie;
263 rpcrdma_update_connect_private(xprt, &event->param.conn);
265 case RDMA_CM_EVENT_CONNECT_ERROR:
266 connstate = -ENOTCONN;
268 case RDMA_CM_EVENT_UNREACHABLE:
269 connstate = -ENETUNREACH;
271 case RDMA_CM_EVENT_REJECTED:
272 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
273 rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
274 rdma_reject_msg(id, event->status));
275 connstate = -ECONNREFUSED;
276 if (event->status == IB_CM_REJ_STALE_CONN)
279 case RDMA_CM_EVENT_DISCONNECTED:
280 ++xprt->rx_xprt.connect_cookie;
281 connstate = -ECONNABORTED;
283 ep->rep_connected = connstate;
284 rpcrdma_conn_func(ep);
285 wake_up_all(&ep->rep_connect_wait);
288 dprintk("RPC: %s: %s:%s on %s/%s (ep 0x%p): %s\n",
290 rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
291 ia->ri_device->name, ia->ri_ops->ro_displayname,
292 ep, rdma_event_msg(event->event));
299 static struct rdma_cm_id *
300 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
302 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
303 struct rdma_cm_id *id;
306 trace_xprtrdma_conn_start(xprt);
308 init_completion(&ia->ri_done);
309 init_completion(&ia->ri_remove_done);
311 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_conn_upcall,
312 xprt, RDMA_PS_TCP, IB_QPT_RC);
315 dprintk("RPC: %s: rdma_create_id() failed %i\n",
320 ia->ri_async_rc = -ETIMEDOUT;
321 rc = rdma_resolve_addr(id, NULL,
322 (struct sockaddr *)&xprt->rx_xprt.addr,
323 RDMA_RESOLVE_TIMEOUT);
325 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
329 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
331 trace_xprtrdma_conn_tout(xprt);
335 rc = ia->ri_async_rc;
339 ia->ri_async_rc = -ETIMEDOUT;
340 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
342 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
346 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
348 trace_xprtrdma_conn_tout(xprt);
351 rc = ia->ri_async_rc;
363 * Exported functions.
367 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
368 * @xprt: transport with IA to (re)initialize
370 * Returns 0 on success, negative errno if an appropriate
371 * Interface Adapter could not be found and opened.
374 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
376 struct rpcrdma_ia *ia = &xprt->rx_ia;
379 ia->ri_id = rpcrdma_create_id(xprt, ia);
380 if (IS_ERR(ia->ri_id)) {
381 rc = PTR_ERR(ia->ri_id);
384 ia->ri_device = ia->ri_id->device;
386 ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
387 if (IS_ERR(ia->ri_pd)) {
388 rc = PTR_ERR(ia->ri_pd);
389 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
393 switch (xprt_rdma_memreg_strategy) {
395 if (frwr_is_supported(ia)) {
396 ia->ri_ops = &rpcrdma_frwr_memreg_ops;
400 case RPCRDMA_MTHCAFMR:
401 if (fmr_is_supported(ia)) {
402 ia->ri_ops = &rpcrdma_fmr_memreg_ops;
407 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
408 ia->ri_device->name, xprt_rdma_memreg_strategy);
416 rpcrdma_ia_close(ia);
421 * rpcrdma_ia_remove - Handle device driver unload
422 * @ia: interface adapter being removed
424 * Divest transport H/W resources associated with this adapter,
425 * but allow it to be restored later.
428 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
430 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
432 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
433 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
434 struct rpcrdma_req *req;
435 struct rpcrdma_rep *rep;
437 cancel_delayed_work_sync(&buf->rb_refresh_worker);
439 /* This is similar to rpcrdma_ep_destroy, but:
440 * - Don't cancel the connect worker.
441 * - Don't call rpcrdma_ep_disconnect, which waits
442 * for another conn upcall, which will deadlock.
443 * - rdma_disconnect is unneeded, the underlying
444 * connection is already gone.
447 ib_drain_qp(ia->ri_id->qp);
448 rdma_destroy_qp(ia->ri_id);
449 ia->ri_id->qp = NULL;
451 ib_free_cq(ep->rep_attr.recv_cq);
452 ep->rep_attr.recv_cq = NULL;
453 ib_free_cq(ep->rep_attr.send_cq);
454 ep->rep_attr.send_cq = NULL;
456 /* The ULP is responsible for ensuring all DMA
457 * mappings and MRs are gone.
459 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
460 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
461 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
462 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
463 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
464 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
466 rpcrdma_mrs_destroy(buf);
467 ib_dealloc_pd(ia->ri_pd);
470 /* Allow waiters to continue */
471 complete(&ia->ri_remove_done);
473 trace_xprtrdma_remove(r_xprt);
477 * rpcrdma_ia_close - Clean up/close an IA.
478 * @ia: interface adapter to close
482 rpcrdma_ia_close(struct rpcrdma_ia *ia)
484 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
486 rdma_destroy_qp(ia->ri_id);
487 rdma_destroy_id(ia->ri_id);
490 ia->ri_device = NULL;
492 /* If the pd is still busy, xprtrdma missed freeing a resource */
493 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
494 ib_dealloc_pd(ia->ri_pd);
499 * Create unconnected endpoint.
502 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
503 struct rpcrdma_create_data_internal *cdata)
505 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
506 struct ib_cq *sendcq, *recvcq;
507 unsigned int max_sge;
510 max_sge = min_t(unsigned int, ia->ri_device->attrs.max_send_sge,
511 RPCRDMA_MAX_SEND_SGES);
512 if (max_sge < RPCRDMA_MIN_SEND_SGES) {
513 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
516 ia->ri_max_send_sges = max_sge;
518 rc = ia->ri_ops->ro_open(ia, ep, cdata);
522 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
523 ep->rep_attr.qp_context = ep;
524 ep->rep_attr.srq = NULL;
525 ep->rep_attr.cap.max_send_sge = max_sge;
526 ep->rep_attr.cap.max_recv_sge = 1;
527 ep->rep_attr.cap.max_inline_data = 0;
528 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
529 ep->rep_attr.qp_type = IB_QPT_RC;
530 ep->rep_attr.port_num = ~0;
532 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
533 "iovs: send %d recv %d\n",
535 ep->rep_attr.cap.max_send_wr,
536 ep->rep_attr.cap.max_recv_wr,
537 ep->rep_attr.cap.max_send_sge,
538 ep->rep_attr.cap.max_recv_sge);
540 /* set trigger for requesting send completion */
541 ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
542 cdata->max_requests >> 2);
543 ep->rep_send_count = ep->rep_send_batch;
544 init_waitqueue_head(&ep->rep_connect_wait);
545 INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
547 sendcq = ib_alloc_cq(ia->ri_device, NULL,
548 ep->rep_attr.cap.max_send_wr + 1,
549 1, IB_POLL_WORKQUEUE);
550 if (IS_ERR(sendcq)) {
551 rc = PTR_ERR(sendcq);
552 dprintk("RPC: %s: failed to create send CQ: %i\n",
557 recvcq = ib_alloc_cq(ia->ri_device, NULL,
558 ep->rep_attr.cap.max_recv_wr + 1,
559 0, IB_POLL_WORKQUEUE);
560 if (IS_ERR(recvcq)) {
561 rc = PTR_ERR(recvcq);
562 dprintk("RPC: %s: failed to create recv CQ: %i\n",
567 ep->rep_attr.send_cq = sendcq;
568 ep->rep_attr.recv_cq = recvcq;
570 /* Initialize cma parameters */
571 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
573 /* Prepare RDMA-CM private message */
574 pmsg->cp_magic = rpcrdma_cmp_magic;
575 pmsg->cp_version = RPCRDMA_CMP_VERSION;
576 pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
577 pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
578 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
579 ep->rep_remote_cma.private_data = pmsg;
580 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
582 /* Client offers RDMA Read but does not initiate */
583 ep->rep_remote_cma.initiator_depth = 0;
584 ep->rep_remote_cma.responder_resources =
585 min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
587 /* Limit transport retries so client can detect server
588 * GID changes quickly. RPC layer handles re-establishing
589 * transport connection and retransmission.
591 ep->rep_remote_cma.retry_count = 6;
593 /* RPC-over-RDMA handles its own flow control. In addition,
594 * make all RNR NAKs visible so we know that RPC-over-RDMA
595 * flow control is working correctly (no NAKs should be seen).
597 ep->rep_remote_cma.flow_control = 0;
598 ep->rep_remote_cma.rnr_retry_count = 0;
611 * Disconnect and destroy endpoint. After this, the only
612 * valid operations on the ep are to free it (if dynamically
613 * allocated) or re-create it.
616 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
618 cancel_delayed_work_sync(&ep->rep_connect_worker);
620 if (ia->ri_id && ia->ri_id->qp) {
621 rpcrdma_ep_disconnect(ep, ia);
622 rdma_destroy_qp(ia->ri_id);
623 ia->ri_id->qp = NULL;
626 if (ep->rep_attr.recv_cq)
627 ib_free_cq(ep->rep_attr.recv_cq);
628 if (ep->rep_attr.send_cq)
629 ib_free_cq(ep->rep_attr.send_cq);
632 /* Re-establish a connection after a device removal event.
633 * Unlike a normal reconnection, a fresh PD and a new set
634 * of MRs and buffers is needed.
637 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
638 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
642 trace_xprtrdma_reinsert(r_xprt);
645 if (rpcrdma_ia_open(r_xprt))
649 err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
651 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
656 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
658 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
662 rpcrdma_mrs_create(r_xprt);
666 rpcrdma_ep_destroy(ep, ia);
668 rpcrdma_ia_close(ia);
674 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
675 struct rpcrdma_ia *ia)
677 struct rdma_cm_id *id, *old;
680 trace_xprtrdma_reconnect(r_xprt);
682 rpcrdma_ep_disconnect(ep, ia);
685 id = rpcrdma_create_id(r_xprt, ia);
689 /* As long as the new ID points to the same device as the
690 * old ID, we can reuse the transport's existing PD and all
691 * previously allocated MRs. Also, the same device means
692 * the transport's previous DMA mappings are still valid.
694 * This is a sanity check only. There should be no way these
695 * point to two different devices here.
699 if (ia->ri_device != id->device) {
700 pr_err("rpcrdma: can't reconnect on different device!\n");
704 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
706 dprintk("RPC: %s: rdma_create_qp returned %d\n",
711 /* Atomically replace the transport's ID and QP. */
715 rdma_destroy_qp(old);
718 rdma_destroy_id(old);
724 * Connect unconnected endpoint.
727 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
729 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
734 switch (ep->rep_connected) {
736 dprintk("RPC: %s: connecting...\n", __func__);
737 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
739 dprintk("RPC: %s: rdma_create_qp failed %i\n",
746 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
751 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
756 ep->rep_connected = 0;
757 rpcrdma_post_recvs(r_xprt, true);
759 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
761 dprintk("RPC: %s: rdma_connect() failed with %i\n",
766 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
767 if (ep->rep_connected <= 0) {
768 if (ep->rep_connected == -EAGAIN)
770 rc = ep->rep_connected;
774 dprintk("RPC: %s: connected\n", __func__);
778 ep->rep_connected = rc;
785 * rpcrdma_ep_disconnect
787 * This is separate from destroy to facilitate the ability
788 * to reconnect without recreating the endpoint.
790 * This call is not reentrant, and must not be made in parallel
791 * on the same endpoint.
794 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
798 rc = rdma_disconnect(ia->ri_id);
800 /* returns without wait if not connected */
801 wait_event_interruptible(ep->rep_connect_wait,
802 ep->rep_connected != 1);
804 ep->rep_connected = rc;
805 trace_xprtrdma_disconnect(container_of(ep, struct rpcrdma_xprt,
808 ib_drain_qp(ia->ri_id->qp);
811 /* Fixed-size circular FIFO queue. This implementation is wait-free and
814 * Consumer is the code path that posts Sends. This path dequeues a
815 * sendctx for use by a Send operation. Multiple consumer threads
816 * are serialized by the RPC transport lock, which allows only one
817 * ->send_request call at a time.
819 * Producer is the code path that handles Send completions. This path
820 * enqueues a sendctx that has been completed. Multiple producer
821 * threads are serialized by the ib_poll_cq() function.
824 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
825 * queue activity, and ib_drain_qp has flushed all remaining Send
828 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
832 for (i = 0; i <= buf->rb_sc_last; i++)
833 kfree(buf->rb_sc_ctxs[i]);
834 kfree(buf->rb_sc_ctxs);
837 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
839 struct rpcrdma_sendctx *sc;
841 sc = kzalloc(sizeof(*sc) +
842 ia->ri_max_send_sges * sizeof(struct ib_sge),
847 sc->sc_wr.wr_cqe = &sc->sc_cqe;
848 sc->sc_wr.sg_list = sc->sc_sges;
849 sc->sc_wr.opcode = IB_WR_SEND;
850 sc->sc_cqe.done = rpcrdma_wc_send;
854 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
856 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
857 struct rpcrdma_sendctx *sc;
860 /* Maximum number of concurrent outstanding Send WRs. Capping
861 * the circular queue size stops Send Queue overflow by causing
862 * the ->send_request call to fail temporarily before too many
865 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
866 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
867 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
868 if (!buf->rb_sc_ctxs)
871 buf->rb_sc_last = i - 1;
872 for (i = 0; i <= buf->rb_sc_last; i++) {
873 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
877 sc->sc_xprt = r_xprt;
878 buf->rb_sc_ctxs[i] = sc;
885 rpcrdma_sendctxs_destroy(buf);
889 /* The sendctx queue is not guaranteed to have a size that is a
890 * power of two, thus the helpers in circ_buf.h cannot be used.
891 * The other option is to use modulus (%), which can be expensive.
893 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
896 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
900 * rpcrdma_sendctx_get_locked - Acquire a send context
901 * @buf: transport buffers from which to acquire an unused context
903 * Returns pointer to a free send completion context; or NULL if
904 * the queue is empty.
906 * Usage: Called to acquire an SGE array before preparing a Send WR.
908 * The caller serializes calls to this function (per rpcrdma_buffer),
909 * and provides an effective memory barrier that flushes the new value
912 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
914 struct rpcrdma_xprt *r_xprt;
915 struct rpcrdma_sendctx *sc;
916 unsigned long next_head;
918 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
920 if (next_head == READ_ONCE(buf->rb_sc_tail))
923 /* ORDER: item must be accessed _before_ head is updated */
924 sc = buf->rb_sc_ctxs[next_head];
926 /* Releasing the lock in the caller acts as a memory
927 * barrier that flushes rb_sc_head.
929 buf->rb_sc_head = next_head;
934 /* The queue is "empty" if there have not been enough Send
935 * completions recently. This is a sign the Send Queue is
936 * backing up. Cause the caller to pause and try again.
938 set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
939 r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
940 r_xprt->rx_stats.empty_sendctx_q++;
945 * rpcrdma_sendctx_put_locked - Release a send context
946 * @sc: send context to release
948 * Usage: Called from Send completion to return a sendctxt
951 * The caller serializes calls to this function (per rpcrdma_buffer).
954 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
956 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
957 unsigned long next_tail;
959 /* Unmap SGEs of previously completed by unsignaled
960 * Sends by walking up the queue until @sc is found.
962 next_tail = buf->rb_sc_tail;
964 next_tail = rpcrdma_sendctx_next(buf, next_tail);
966 /* ORDER: item must be accessed _before_ tail is updated */
967 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
969 } while (buf->rb_sc_ctxs[next_tail] != sc);
971 /* Paired with READ_ONCE */
972 smp_store_release(&buf->rb_sc_tail, next_tail);
974 if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
975 smp_mb__after_atomic();
976 xprt_write_space(&sc->sc_xprt->rx_xprt);
981 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
983 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
984 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
989 for (count = 0; count < ia->ri_max_segs; count++) {
990 struct rpcrdma_mr *mr;
993 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
997 rc = ia->ri_ops->ro_init_mr(ia, mr);
1003 mr->mr_xprt = r_xprt;
1005 list_add(&mr->mr_list, &free);
1006 list_add(&mr->mr_all, &all);
1009 spin_lock(&buf->rb_mrlock);
1010 list_splice(&free, &buf->rb_mrs);
1011 list_splice(&all, &buf->rb_all);
1012 r_xprt->rx_stats.mrs_allocated += count;
1013 spin_unlock(&buf->rb_mrlock);
1014 trace_xprtrdma_createmrs(r_xprt, count);
1016 xprt_write_space(&r_xprt->rx_xprt);
1020 rpcrdma_mr_refresh_worker(struct work_struct *work)
1022 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1023 rb_refresh_worker.work);
1024 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1027 rpcrdma_mrs_create(r_xprt);
1030 struct rpcrdma_req *
1031 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1033 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1034 struct rpcrdma_regbuf *rb;
1035 struct rpcrdma_req *req;
1037 req = kzalloc(sizeof(*req), GFP_KERNEL);
1039 return ERR_PTR(-ENOMEM);
1041 rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1042 DMA_TO_DEVICE, GFP_KERNEL);
1045 return ERR_PTR(-ENOMEM);
1047 req->rl_rdmabuf = rb;
1048 xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1049 req->rl_buffer = buffer;
1050 INIT_LIST_HEAD(&req->rl_registered);
1052 spin_lock(&buffer->rb_reqslock);
1053 list_add(&req->rl_all, &buffer->rb_allreqs);
1054 spin_unlock(&buffer->rb_reqslock);
1059 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp)
1061 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1062 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1063 struct rpcrdma_rep *rep;
1067 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1071 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1072 DMA_FROM_DEVICE, GFP_KERNEL);
1073 if (IS_ERR(rep->rr_rdmabuf)) {
1074 rc = PTR_ERR(rep->rr_rdmabuf);
1077 xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1078 rdmab_length(rep->rr_rdmabuf));
1080 rep->rr_cqe.done = rpcrdma_wc_receive;
1081 rep->rr_rxprt = r_xprt;
1082 INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1083 rep->rr_recv_wr.next = NULL;
1084 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1085 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1086 rep->rr_recv_wr.num_sge = 1;
1087 rep->rr_temp = temp;
1089 spin_lock(&buf->rb_lock);
1090 list_add(&rep->rr_list, &buf->rb_recv_bufs);
1091 spin_unlock(&buf->rb_lock);
1097 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1103 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1105 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1108 buf->rb_max_requests = r_xprt->rx_data.max_requests;
1109 buf->rb_bc_srv_max_requests = 0;
1110 spin_lock_init(&buf->rb_mrlock);
1111 spin_lock_init(&buf->rb_lock);
1112 INIT_LIST_HEAD(&buf->rb_mrs);
1113 INIT_LIST_HEAD(&buf->rb_all);
1114 INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1115 rpcrdma_mr_refresh_worker);
1117 rpcrdma_mrs_create(r_xprt);
1119 INIT_LIST_HEAD(&buf->rb_send_bufs);
1120 INIT_LIST_HEAD(&buf->rb_allreqs);
1121 spin_lock_init(&buf->rb_reqslock);
1122 for (i = 0; i < buf->rb_max_requests; i++) {
1123 struct rpcrdma_req *req;
1125 req = rpcrdma_create_req(r_xprt);
1127 dprintk("RPC: %s: request buffer %d alloc"
1128 " failed\n", __func__, i);
1132 list_add(&req->rl_list, &buf->rb_send_bufs);
1135 buf->rb_credits = 1;
1136 buf->rb_posted_receives = 0;
1137 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1139 rc = rpcrdma_sendctxs_create(r_xprt);
1145 rpcrdma_buffer_destroy(buf);
1150 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1152 rpcrdma_free_regbuf(rep->rr_rdmabuf);
1157 rpcrdma_destroy_req(struct rpcrdma_req *req)
1159 rpcrdma_free_regbuf(req->rl_recvbuf);
1160 rpcrdma_free_regbuf(req->rl_sendbuf);
1161 rpcrdma_free_regbuf(req->rl_rdmabuf);
1166 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1168 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1170 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1171 struct rpcrdma_mr *mr;
1175 spin_lock(&buf->rb_mrlock);
1176 while (!list_empty(&buf->rb_all)) {
1177 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1178 list_del(&mr->mr_all);
1180 spin_unlock(&buf->rb_mrlock);
1182 /* Ensure MW is not on any rl_registered list */
1183 if (!list_empty(&mr->mr_list))
1184 list_del(&mr->mr_list);
1186 ia->ri_ops->ro_release_mr(mr);
1188 spin_lock(&buf->rb_mrlock);
1190 spin_unlock(&buf->rb_mrlock);
1191 r_xprt->rx_stats.mrs_allocated = 0;
1193 dprintk("RPC: %s: released %u MRs\n", __func__, count);
1197 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1199 cancel_delayed_work_sync(&buf->rb_refresh_worker);
1201 rpcrdma_sendctxs_destroy(buf);
1203 while (!list_empty(&buf->rb_recv_bufs)) {
1204 struct rpcrdma_rep *rep;
1206 rep = list_first_entry(&buf->rb_recv_bufs,
1207 struct rpcrdma_rep, rr_list);
1208 list_del(&rep->rr_list);
1209 rpcrdma_destroy_rep(rep);
1212 spin_lock(&buf->rb_reqslock);
1213 while (!list_empty(&buf->rb_allreqs)) {
1214 struct rpcrdma_req *req;
1216 req = list_first_entry(&buf->rb_allreqs,
1217 struct rpcrdma_req, rl_all);
1218 list_del(&req->rl_all);
1220 spin_unlock(&buf->rb_reqslock);
1221 rpcrdma_destroy_req(req);
1222 spin_lock(&buf->rb_reqslock);
1224 spin_unlock(&buf->rb_reqslock);
1226 rpcrdma_mrs_destroy(buf);
1230 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1231 * @r_xprt: controlling transport
1233 * Returns an initialized rpcrdma_mr or NULL if no free
1234 * rpcrdma_mr objects are available.
1237 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1239 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1240 struct rpcrdma_mr *mr = NULL;
1242 spin_lock(&buf->rb_mrlock);
1243 if (!list_empty(&buf->rb_mrs))
1244 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1245 spin_unlock(&buf->rb_mrlock);
1252 trace_xprtrdma_nomrs(r_xprt);
1253 if (r_xprt->rx_ep.rep_connected != -ENODEV)
1254 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1256 /* Allow the reply handler and refresh worker to run */
1263 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1265 spin_lock(&buf->rb_mrlock);
1266 rpcrdma_mr_push(mr, &buf->rb_mrs);
1267 spin_unlock(&buf->rb_mrlock);
1271 * rpcrdma_mr_put - Release an rpcrdma_mr object
1272 * @mr: object to release
1276 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1278 __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1282 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1283 * @mr: object to release
1287 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1289 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1291 trace_xprtrdma_dma_unmap(mr);
1292 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1293 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1294 __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1298 * rpcrdma_buffer_get - Get a request buffer
1299 * @buffers: Buffer pool from which to obtain a buffer
1301 * Returns a fresh rpcrdma_req, or NULL if none are available.
1303 struct rpcrdma_req *
1304 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1306 struct rpcrdma_req *req;
1308 spin_lock(&buffers->rb_lock);
1309 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1310 struct rpcrdma_req, rl_list);
1312 list_del_init(&req->rl_list);
1313 spin_unlock(&buffers->rb_lock);
1318 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1319 * @req: object to return
1323 rpcrdma_buffer_put(struct rpcrdma_req *req)
1325 struct rpcrdma_buffer *buffers = req->rl_buffer;
1326 struct rpcrdma_rep *rep = req->rl_reply;
1328 req->rl_reply = NULL;
1330 spin_lock(&buffers->rb_lock);
1331 list_add(&req->rl_list, &buffers->rb_send_bufs);
1333 if (!rep->rr_temp) {
1334 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1338 spin_unlock(&buffers->rb_lock);
1340 rpcrdma_destroy_rep(rep);
1344 * Put reply buffers back into pool when not attached to
1345 * request. This happens in error conditions.
1348 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1350 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1352 if (!rep->rr_temp) {
1353 spin_lock(&buffers->rb_lock);
1354 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1355 spin_unlock(&buffers->rb_lock);
1357 rpcrdma_destroy_rep(rep);
1362 * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1363 * @size: size of buffer to be allocated, in bytes
1364 * @direction: direction of data movement
1367 * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1368 * can be persistently DMA-mapped for I/O.
1370 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1371 * receiving the payload of RDMA RECV operations. During Long Calls
1372 * or Replies they may be registered externally via ro_map.
1374 struct rpcrdma_regbuf *
1375 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1378 struct rpcrdma_regbuf *rb;
1380 rb = kmalloc(sizeof(*rb) + size, flags);
1382 return ERR_PTR(-ENOMEM);
1384 rb->rg_device = NULL;
1385 rb->rg_direction = direction;
1386 rb->rg_iov.length = size;
1392 * __rpcrdma_map_regbuf - DMA-map a regbuf
1393 * @ia: controlling rpcrdma_ia
1394 * @rb: regbuf to be mapped
1397 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1399 struct ib_device *device = ia->ri_device;
1401 if (rb->rg_direction == DMA_NONE)
1404 rb->rg_iov.addr = ib_dma_map_single(device,
1405 (void *)rb->rg_base,
1408 if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1411 rb->rg_device = device;
1412 rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1417 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1422 if (!rpcrdma_regbuf_is_mapped(rb))
1425 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1426 rdmab_length(rb), rb->rg_direction);
1427 rb->rg_device = NULL;
1431 * rpcrdma_free_regbuf - deregister and free registered buffer
1432 * @rb: regbuf to be deregistered and freed
1435 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1437 rpcrdma_dma_unmap_regbuf(rb);
1442 * Prepost any receive buffer, then post send.
1444 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1447 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1448 struct rpcrdma_ep *ep,
1449 struct rpcrdma_req *req)
1451 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1454 if (!ep->rep_send_count ||
1455 test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1456 send_wr->send_flags |= IB_SEND_SIGNALED;
1457 ep->rep_send_count = ep->rep_send_batch;
1459 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1460 --ep->rep_send_count;
1463 rc = ia->ri_ops->ro_send(ia, req);
1464 trace_xprtrdma_post_send(req, rc);
1471 * rpcrdma_post_recvs - Maybe post some Receive buffers
1472 * @r_xprt: controlling transport
1473 * @temp: when true, allocate temp rpcrdma_rep objects
1477 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1479 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1480 struct ib_recv_wr *wr, *bad_wr;
1481 int needed, count, rc;
1483 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1484 if (buf->rb_posted_receives > needed)
1486 needed -= buf->rb_posted_receives;
1491 struct rpcrdma_regbuf *rb;
1492 struct rpcrdma_rep *rep;
1494 spin_lock(&buf->rb_lock);
1495 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1496 struct rpcrdma_rep, rr_list);
1498 list_del(&rep->rr_list);
1499 spin_unlock(&buf->rb_lock);
1501 if (rpcrdma_create_rep(r_xprt, temp))
1506 rb = rep->rr_rdmabuf;
1507 if (!rpcrdma_regbuf_is_mapped(rb)) {
1508 if (!__rpcrdma_dma_map_regbuf(&r_xprt->rx_ia, rb)) {
1509 rpcrdma_recv_buffer_put(rep);
1514 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1515 rep->rr_recv_wr.next = wr;
1516 wr = &rep->rr_recv_wr;
1523 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1524 (const struct ib_recv_wr **)&bad_wr);
1526 for (wr = bad_wr; wr; wr = wr->next) {
1527 struct rpcrdma_rep *rep;
1529 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1530 rpcrdma_recv_buffer_put(rep);
1534 buf->rb_posted_receives += count;
1535 trace_xprtrdma_post_recvs(r_xprt, count, rc);