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_cm_event_handler - Handle RDMA CM events
224 * @id: rdma_cm_id on which an event has occurred
225 * @event: details of the event
227 * Called with @id's mutex held. Returns 1 if caller should
228 * destroy @id, otherwise 0.
231 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
233 struct rpcrdma_xprt *xprt = id->context;
234 struct rpcrdma_ia *ia = &xprt->rx_ia;
235 struct rpcrdma_ep *ep = &xprt->rx_ep;
240 trace_xprtrdma_cm_event(xprt, event);
241 switch (event->event) {
242 case RDMA_CM_EVENT_ADDR_RESOLVED:
243 case RDMA_CM_EVENT_ROUTE_RESOLVED:
245 complete(&ia->ri_done);
247 case RDMA_CM_EVENT_ADDR_ERROR:
248 ia->ri_async_rc = -EPROTO;
249 complete(&ia->ri_done);
251 case RDMA_CM_EVENT_ROUTE_ERROR:
252 ia->ri_async_rc = -ENETUNREACH;
253 complete(&ia->ri_done);
255 case RDMA_CM_EVENT_DEVICE_REMOVAL:
256 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
257 pr_info("rpcrdma: removing device %s for %s:%s\n",
259 rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
261 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
262 ep->rep_connected = -ENODEV;
263 xprt_force_disconnect(&xprt->rx_xprt);
264 wait_for_completion(&ia->ri_remove_done);
267 ia->ri_device = NULL;
268 /* Return 1 to ensure the core destroys the id. */
270 case RDMA_CM_EVENT_ESTABLISHED:
271 ++xprt->rx_xprt.connect_cookie;
273 rpcrdma_update_connect_private(xprt, &event->param.conn);
275 case RDMA_CM_EVENT_CONNECT_ERROR:
276 connstate = -ENOTCONN;
278 case RDMA_CM_EVENT_UNREACHABLE:
279 connstate = -ENETUNREACH;
281 case RDMA_CM_EVENT_REJECTED:
282 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
283 rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
284 rdma_reject_msg(id, event->status));
285 connstate = -ECONNREFUSED;
286 if (event->status == IB_CM_REJ_STALE_CONN)
289 case RDMA_CM_EVENT_DISCONNECTED:
290 ++xprt->rx_xprt.connect_cookie;
291 connstate = -ECONNABORTED;
293 ep->rep_connected = connstate;
294 rpcrdma_conn_func(ep);
295 wake_up_all(&ep->rep_connect_wait);
298 dprintk("RPC: %s: %s:%s on %s/%s (ep 0x%p): %s\n",
300 rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
301 ia->ri_device->name, ia->ri_ops->ro_displayname,
302 ep, rdma_event_msg(event->event));
309 static struct rdma_cm_id *
310 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
312 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
313 struct rdma_cm_id *id;
316 trace_xprtrdma_conn_start(xprt);
318 init_completion(&ia->ri_done);
319 init_completion(&ia->ri_remove_done);
321 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
322 xprt, RDMA_PS_TCP, IB_QPT_RC);
325 dprintk("RPC: %s: rdma_create_id() failed %i\n",
330 ia->ri_async_rc = -ETIMEDOUT;
331 rc = rdma_resolve_addr(id, NULL,
332 (struct sockaddr *)&xprt->rx_xprt.addr,
333 RDMA_RESOLVE_TIMEOUT);
335 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
339 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
341 trace_xprtrdma_conn_tout(xprt);
345 rc = ia->ri_async_rc;
349 ia->ri_async_rc = -ETIMEDOUT;
350 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
352 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
356 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
358 trace_xprtrdma_conn_tout(xprt);
361 rc = ia->ri_async_rc;
373 * Exported functions.
377 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
378 * @xprt: transport with IA to (re)initialize
380 * Returns 0 on success, negative errno if an appropriate
381 * Interface Adapter could not be found and opened.
384 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
386 struct rpcrdma_ia *ia = &xprt->rx_ia;
389 ia->ri_id = rpcrdma_create_id(xprt, ia);
390 if (IS_ERR(ia->ri_id)) {
391 rc = PTR_ERR(ia->ri_id);
394 ia->ri_device = ia->ri_id->device;
396 ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
397 if (IS_ERR(ia->ri_pd)) {
398 rc = PTR_ERR(ia->ri_pd);
399 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
403 switch (xprt_rdma_memreg_strategy) {
405 if (frwr_is_supported(ia)) {
406 ia->ri_ops = &rpcrdma_frwr_memreg_ops;
410 case RPCRDMA_MTHCAFMR:
411 if (fmr_is_supported(ia)) {
412 ia->ri_ops = &rpcrdma_fmr_memreg_ops;
417 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
418 ia->ri_device->name, xprt_rdma_memreg_strategy);
426 rpcrdma_ia_close(ia);
431 * rpcrdma_ia_remove - Handle device driver unload
432 * @ia: interface adapter being removed
434 * Divest transport H/W resources associated with this adapter,
435 * but allow it to be restored later.
438 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
440 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
442 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
443 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
444 struct rpcrdma_req *req;
445 struct rpcrdma_rep *rep;
447 cancel_delayed_work_sync(&buf->rb_refresh_worker);
449 /* This is similar to rpcrdma_ep_destroy, but:
450 * - Don't cancel the connect worker.
451 * - Don't call rpcrdma_ep_disconnect, which waits
452 * for another conn upcall, which will deadlock.
453 * - rdma_disconnect is unneeded, the underlying
454 * connection is already gone.
457 ib_drain_qp(ia->ri_id->qp);
458 rdma_destroy_qp(ia->ri_id);
459 ia->ri_id->qp = NULL;
461 ib_free_cq(ep->rep_attr.recv_cq);
462 ep->rep_attr.recv_cq = NULL;
463 ib_free_cq(ep->rep_attr.send_cq);
464 ep->rep_attr.send_cq = NULL;
466 /* The ULP is responsible for ensuring all DMA
467 * mappings and MRs are gone.
469 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
470 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
471 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
472 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
473 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
474 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
476 rpcrdma_mrs_destroy(buf);
477 ib_dealloc_pd(ia->ri_pd);
480 /* Allow waiters to continue */
481 complete(&ia->ri_remove_done);
483 trace_xprtrdma_remove(r_xprt);
487 * rpcrdma_ia_close - Clean up/close an IA.
488 * @ia: interface adapter to close
492 rpcrdma_ia_close(struct rpcrdma_ia *ia)
494 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
496 rdma_destroy_qp(ia->ri_id);
497 rdma_destroy_id(ia->ri_id);
500 ia->ri_device = NULL;
502 /* If the pd is still busy, xprtrdma missed freeing a resource */
503 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
504 ib_dealloc_pd(ia->ri_pd);
509 * Create unconnected endpoint.
512 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
513 struct rpcrdma_create_data_internal *cdata)
515 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
516 struct ib_cq *sendcq, *recvcq;
517 unsigned int max_sge;
520 max_sge = min_t(unsigned int, ia->ri_device->attrs.max_send_sge,
521 RPCRDMA_MAX_SEND_SGES);
522 if (max_sge < RPCRDMA_MIN_SEND_SGES) {
523 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
526 ia->ri_max_send_sges = max_sge;
528 rc = ia->ri_ops->ro_open(ia, ep, cdata);
532 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
533 ep->rep_attr.qp_context = ep;
534 ep->rep_attr.srq = NULL;
535 ep->rep_attr.cap.max_send_sge = max_sge;
536 ep->rep_attr.cap.max_recv_sge = 1;
537 ep->rep_attr.cap.max_inline_data = 0;
538 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
539 ep->rep_attr.qp_type = IB_QPT_RC;
540 ep->rep_attr.port_num = ~0;
542 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
543 "iovs: send %d recv %d\n",
545 ep->rep_attr.cap.max_send_wr,
546 ep->rep_attr.cap.max_recv_wr,
547 ep->rep_attr.cap.max_send_sge,
548 ep->rep_attr.cap.max_recv_sge);
550 /* set trigger for requesting send completion */
551 ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
552 cdata->max_requests >> 2);
553 ep->rep_send_count = ep->rep_send_batch;
554 init_waitqueue_head(&ep->rep_connect_wait);
555 INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
557 sendcq = ib_alloc_cq(ia->ri_device, NULL,
558 ep->rep_attr.cap.max_send_wr + 1,
559 1, IB_POLL_WORKQUEUE);
560 if (IS_ERR(sendcq)) {
561 rc = PTR_ERR(sendcq);
562 dprintk("RPC: %s: failed to create send CQ: %i\n",
567 recvcq = ib_alloc_cq(ia->ri_device, NULL,
568 ep->rep_attr.cap.max_recv_wr + 1,
569 0, IB_POLL_WORKQUEUE);
570 if (IS_ERR(recvcq)) {
571 rc = PTR_ERR(recvcq);
572 dprintk("RPC: %s: failed to create recv CQ: %i\n",
577 ep->rep_attr.send_cq = sendcq;
578 ep->rep_attr.recv_cq = recvcq;
580 /* Initialize cma parameters */
581 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
583 /* Prepare RDMA-CM private message */
584 pmsg->cp_magic = rpcrdma_cmp_magic;
585 pmsg->cp_version = RPCRDMA_CMP_VERSION;
586 pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
587 pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
588 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
589 ep->rep_remote_cma.private_data = pmsg;
590 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
592 /* Client offers RDMA Read but does not initiate */
593 ep->rep_remote_cma.initiator_depth = 0;
594 ep->rep_remote_cma.responder_resources =
595 min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
597 /* Limit transport retries so client can detect server
598 * GID changes quickly. RPC layer handles re-establishing
599 * transport connection and retransmission.
601 ep->rep_remote_cma.retry_count = 6;
603 /* RPC-over-RDMA handles its own flow control. In addition,
604 * make all RNR NAKs visible so we know that RPC-over-RDMA
605 * flow control is working correctly (no NAKs should be seen).
607 ep->rep_remote_cma.flow_control = 0;
608 ep->rep_remote_cma.rnr_retry_count = 0;
621 * Disconnect and destroy endpoint. After this, the only
622 * valid operations on the ep are to free it (if dynamically
623 * allocated) or re-create it.
626 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
628 cancel_delayed_work_sync(&ep->rep_connect_worker);
630 if (ia->ri_id && ia->ri_id->qp) {
631 rpcrdma_ep_disconnect(ep, ia);
632 rdma_destroy_qp(ia->ri_id);
633 ia->ri_id->qp = NULL;
636 if (ep->rep_attr.recv_cq)
637 ib_free_cq(ep->rep_attr.recv_cq);
638 if (ep->rep_attr.send_cq)
639 ib_free_cq(ep->rep_attr.send_cq);
642 /* Re-establish a connection after a device removal event.
643 * Unlike a normal reconnection, a fresh PD and a new set
644 * of MRs and buffers is needed.
647 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
648 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
652 trace_xprtrdma_reinsert(r_xprt);
655 if (rpcrdma_ia_open(r_xprt))
659 err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
661 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
666 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
668 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
672 rpcrdma_mrs_create(r_xprt);
676 rpcrdma_ep_destroy(ep, ia);
678 rpcrdma_ia_close(ia);
684 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
685 struct rpcrdma_ia *ia)
687 struct rdma_cm_id *id, *old;
690 trace_xprtrdma_reconnect(r_xprt);
692 rpcrdma_ep_disconnect(ep, ia);
695 id = rpcrdma_create_id(r_xprt, ia);
699 /* As long as the new ID points to the same device as the
700 * old ID, we can reuse the transport's existing PD and all
701 * previously allocated MRs. Also, the same device means
702 * the transport's previous DMA mappings are still valid.
704 * This is a sanity check only. There should be no way these
705 * point to two different devices here.
709 if (ia->ri_device != id->device) {
710 pr_err("rpcrdma: can't reconnect on different device!\n");
714 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
716 dprintk("RPC: %s: rdma_create_qp returned %d\n",
721 /* Atomically replace the transport's ID and QP. */
725 rdma_destroy_qp(old);
728 rdma_destroy_id(old);
734 * Connect unconnected endpoint.
737 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
739 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
744 switch (ep->rep_connected) {
746 dprintk("RPC: %s: connecting...\n", __func__);
747 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
749 dprintk("RPC: %s: rdma_create_qp failed %i\n",
756 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
761 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
766 ep->rep_connected = 0;
767 rpcrdma_post_recvs(r_xprt, true);
769 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
771 dprintk("RPC: %s: rdma_connect() failed with %i\n",
776 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
777 if (ep->rep_connected <= 0) {
778 if (ep->rep_connected == -EAGAIN)
780 rc = ep->rep_connected;
784 dprintk("RPC: %s: connected\n", __func__);
788 ep->rep_connected = rc;
795 * rpcrdma_ep_disconnect
797 * This is separate from destroy to facilitate the ability
798 * to reconnect without recreating the endpoint.
800 * This call is not reentrant, and must not be made in parallel
801 * on the same endpoint.
804 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
808 rc = rdma_disconnect(ia->ri_id);
810 /* returns without wait if not connected */
811 wait_event_interruptible(ep->rep_connect_wait,
812 ep->rep_connected != 1);
814 ep->rep_connected = rc;
815 trace_xprtrdma_disconnect(container_of(ep, struct rpcrdma_xprt,
818 ib_drain_qp(ia->ri_id->qp);
821 /* Fixed-size circular FIFO queue. This implementation is wait-free and
824 * Consumer is the code path that posts Sends. This path dequeues a
825 * sendctx for use by a Send operation. Multiple consumer threads
826 * are serialized by the RPC transport lock, which allows only one
827 * ->send_request call at a time.
829 * Producer is the code path that handles Send completions. This path
830 * enqueues a sendctx that has been completed. Multiple producer
831 * threads are serialized by the ib_poll_cq() function.
834 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
835 * queue activity, and ib_drain_qp has flushed all remaining Send
838 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
842 for (i = 0; i <= buf->rb_sc_last; i++)
843 kfree(buf->rb_sc_ctxs[i]);
844 kfree(buf->rb_sc_ctxs);
847 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
849 struct rpcrdma_sendctx *sc;
851 sc = kzalloc(sizeof(*sc) +
852 ia->ri_max_send_sges * sizeof(struct ib_sge),
857 sc->sc_wr.wr_cqe = &sc->sc_cqe;
858 sc->sc_wr.sg_list = sc->sc_sges;
859 sc->sc_wr.opcode = IB_WR_SEND;
860 sc->sc_cqe.done = rpcrdma_wc_send;
864 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
866 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
867 struct rpcrdma_sendctx *sc;
870 /* Maximum number of concurrent outstanding Send WRs. Capping
871 * the circular queue size stops Send Queue overflow by causing
872 * the ->send_request call to fail temporarily before too many
875 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
876 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
877 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
878 if (!buf->rb_sc_ctxs)
881 buf->rb_sc_last = i - 1;
882 for (i = 0; i <= buf->rb_sc_last; i++) {
883 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
887 sc->sc_xprt = r_xprt;
888 buf->rb_sc_ctxs[i] = sc;
895 rpcrdma_sendctxs_destroy(buf);
899 /* The sendctx queue is not guaranteed to have a size that is a
900 * power of two, thus the helpers in circ_buf.h cannot be used.
901 * The other option is to use modulus (%), which can be expensive.
903 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
906 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
910 * rpcrdma_sendctx_get_locked - Acquire a send context
911 * @buf: transport buffers from which to acquire an unused context
913 * Returns pointer to a free send completion context; or NULL if
914 * the queue is empty.
916 * Usage: Called to acquire an SGE array before preparing a Send WR.
918 * The caller serializes calls to this function (per rpcrdma_buffer),
919 * and provides an effective memory barrier that flushes the new value
922 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
924 struct rpcrdma_xprt *r_xprt;
925 struct rpcrdma_sendctx *sc;
926 unsigned long next_head;
928 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
930 if (next_head == READ_ONCE(buf->rb_sc_tail))
933 /* ORDER: item must be accessed _before_ head is updated */
934 sc = buf->rb_sc_ctxs[next_head];
936 /* Releasing the lock in the caller acts as a memory
937 * barrier that flushes rb_sc_head.
939 buf->rb_sc_head = next_head;
944 /* The queue is "empty" if there have not been enough Send
945 * completions recently. This is a sign the Send Queue is
946 * backing up. Cause the caller to pause and try again.
948 set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
949 r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
950 r_xprt->rx_stats.empty_sendctx_q++;
955 * rpcrdma_sendctx_put_locked - Release a send context
956 * @sc: send context to release
958 * Usage: Called from Send completion to return a sendctxt
961 * The caller serializes calls to this function (per rpcrdma_buffer).
964 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
966 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
967 unsigned long next_tail;
969 /* Unmap SGEs of previously completed by unsignaled
970 * Sends by walking up the queue until @sc is found.
972 next_tail = buf->rb_sc_tail;
974 next_tail = rpcrdma_sendctx_next(buf, next_tail);
976 /* ORDER: item must be accessed _before_ tail is updated */
977 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
979 } while (buf->rb_sc_ctxs[next_tail] != sc);
981 /* Paired with READ_ONCE */
982 smp_store_release(&buf->rb_sc_tail, next_tail);
984 if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
985 smp_mb__after_atomic();
986 xprt_write_space(&sc->sc_xprt->rx_xprt);
991 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
993 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
994 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
999 for (count = 0; count < ia->ri_max_segs; count++) {
1000 struct rpcrdma_mr *mr;
1003 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1007 rc = ia->ri_ops->ro_init_mr(ia, mr);
1013 mr->mr_xprt = r_xprt;
1015 list_add(&mr->mr_list, &free);
1016 list_add(&mr->mr_all, &all);
1019 spin_lock(&buf->rb_mrlock);
1020 list_splice(&free, &buf->rb_mrs);
1021 list_splice(&all, &buf->rb_all);
1022 r_xprt->rx_stats.mrs_allocated += count;
1023 spin_unlock(&buf->rb_mrlock);
1024 trace_xprtrdma_createmrs(r_xprt, count);
1026 xprt_write_space(&r_xprt->rx_xprt);
1030 rpcrdma_mr_refresh_worker(struct work_struct *work)
1032 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1033 rb_refresh_worker.work);
1034 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1037 rpcrdma_mrs_create(r_xprt);
1040 struct rpcrdma_req *
1041 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1043 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1044 struct rpcrdma_regbuf *rb;
1045 struct rpcrdma_req *req;
1047 req = kzalloc(sizeof(*req), GFP_KERNEL);
1049 return ERR_PTR(-ENOMEM);
1051 rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1052 DMA_TO_DEVICE, GFP_KERNEL);
1055 return ERR_PTR(-ENOMEM);
1057 req->rl_rdmabuf = rb;
1058 xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1059 req->rl_buffer = buffer;
1060 INIT_LIST_HEAD(&req->rl_registered);
1062 spin_lock(&buffer->rb_reqslock);
1063 list_add(&req->rl_all, &buffer->rb_allreqs);
1064 spin_unlock(&buffer->rb_reqslock);
1069 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp)
1071 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1072 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1073 struct rpcrdma_rep *rep;
1077 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1081 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1082 DMA_FROM_DEVICE, GFP_KERNEL);
1083 if (IS_ERR(rep->rr_rdmabuf)) {
1084 rc = PTR_ERR(rep->rr_rdmabuf);
1087 xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1088 rdmab_length(rep->rr_rdmabuf));
1090 rep->rr_cqe.done = rpcrdma_wc_receive;
1091 rep->rr_rxprt = r_xprt;
1092 INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1093 rep->rr_recv_wr.next = NULL;
1094 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1095 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1096 rep->rr_recv_wr.num_sge = 1;
1097 rep->rr_temp = temp;
1099 spin_lock(&buf->rb_lock);
1100 list_add(&rep->rr_list, &buf->rb_recv_bufs);
1101 spin_unlock(&buf->rb_lock);
1107 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1113 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1115 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1118 buf->rb_max_requests = r_xprt->rx_data.max_requests;
1119 buf->rb_bc_srv_max_requests = 0;
1120 spin_lock_init(&buf->rb_mrlock);
1121 spin_lock_init(&buf->rb_lock);
1122 INIT_LIST_HEAD(&buf->rb_mrs);
1123 INIT_LIST_HEAD(&buf->rb_all);
1124 INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1125 rpcrdma_mr_refresh_worker);
1127 rpcrdma_mrs_create(r_xprt);
1129 INIT_LIST_HEAD(&buf->rb_send_bufs);
1130 INIT_LIST_HEAD(&buf->rb_allreqs);
1131 spin_lock_init(&buf->rb_reqslock);
1132 for (i = 0; i < buf->rb_max_requests; i++) {
1133 struct rpcrdma_req *req;
1135 req = rpcrdma_create_req(r_xprt);
1137 dprintk("RPC: %s: request buffer %d alloc"
1138 " failed\n", __func__, i);
1142 list_add(&req->rl_list, &buf->rb_send_bufs);
1145 buf->rb_credits = 1;
1146 buf->rb_posted_receives = 0;
1147 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1149 rc = rpcrdma_sendctxs_create(r_xprt);
1155 rpcrdma_buffer_destroy(buf);
1160 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1162 rpcrdma_free_regbuf(rep->rr_rdmabuf);
1167 rpcrdma_destroy_req(struct rpcrdma_req *req)
1169 rpcrdma_free_regbuf(req->rl_recvbuf);
1170 rpcrdma_free_regbuf(req->rl_sendbuf);
1171 rpcrdma_free_regbuf(req->rl_rdmabuf);
1176 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1178 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1180 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1181 struct rpcrdma_mr *mr;
1185 spin_lock(&buf->rb_mrlock);
1186 while (!list_empty(&buf->rb_all)) {
1187 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1188 list_del(&mr->mr_all);
1190 spin_unlock(&buf->rb_mrlock);
1192 /* Ensure MW is not on any rl_registered list */
1193 if (!list_empty(&mr->mr_list))
1194 list_del(&mr->mr_list);
1196 ia->ri_ops->ro_release_mr(mr);
1198 spin_lock(&buf->rb_mrlock);
1200 spin_unlock(&buf->rb_mrlock);
1201 r_xprt->rx_stats.mrs_allocated = 0;
1203 dprintk("RPC: %s: released %u MRs\n", __func__, count);
1207 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1209 cancel_delayed_work_sync(&buf->rb_refresh_worker);
1211 rpcrdma_sendctxs_destroy(buf);
1213 while (!list_empty(&buf->rb_recv_bufs)) {
1214 struct rpcrdma_rep *rep;
1216 rep = list_first_entry(&buf->rb_recv_bufs,
1217 struct rpcrdma_rep, rr_list);
1218 list_del(&rep->rr_list);
1219 rpcrdma_destroy_rep(rep);
1222 spin_lock(&buf->rb_reqslock);
1223 while (!list_empty(&buf->rb_allreqs)) {
1224 struct rpcrdma_req *req;
1226 req = list_first_entry(&buf->rb_allreqs,
1227 struct rpcrdma_req, rl_all);
1228 list_del(&req->rl_all);
1230 spin_unlock(&buf->rb_reqslock);
1231 rpcrdma_destroy_req(req);
1232 spin_lock(&buf->rb_reqslock);
1234 spin_unlock(&buf->rb_reqslock);
1236 rpcrdma_mrs_destroy(buf);
1240 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1241 * @r_xprt: controlling transport
1243 * Returns an initialized rpcrdma_mr or NULL if no free
1244 * rpcrdma_mr objects are available.
1247 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1249 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1250 struct rpcrdma_mr *mr = NULL;
1252 spin_lock(&buf->rb_mrlock);
1253 if (!list_empty(&buf->rb_mrs))
1254 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1255 spin_unlock(&buf->rb_mrlock);
1262 trace_xprtrdma_nomrs(r_xprt);
1263 if (r_xprt->rx_ep.rep_connected != -ENODEV)
1264 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1266 /* Allow the reply handler and refresh worker to run */
1273 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1275 spin_lock(&buf->rb_mrlock);
1276 rpcrdma_mr_push(mr, &buf->rb_mrs);
1277 spin_unlock(&buf->rb_mrlock);
1281 * rpcrdma_mr_put - Release an rpcrdma_mr object
1282 * @mr: object to release
1286 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1288 __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1292 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1293 * @mr: object to release
1297 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1299 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1301 trace_xprtrdma_mr_unmap(mr);
1302 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1303 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1304 __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1308 * rpcrdma_buffer_get - Get a request buffer
1309 * @buffers: Buffer pool from which to obtain a buffer
1311 * Returns a fresh rpcrdma_req, or NULL if none are available.
1313 struct rpcrdma_req *
1314 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1316 struct rpcrdma_req *req;
1318 spin_lock(&buffers->rb_lock);
1319 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1320 struct rpcrdma_req, rl_list);
1322 list_del_init(&req->rl_list);
1323 spin_unlock(&buffers->rb_lock);
1328 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1329 * @req: object to return
1333 rpcrdma_buffer_put(struct rpcrdma_req *req)
1335 struct rpcrdma_buffer *buffers = req->rl_buffer;
1336 struct rpcrdma_rep *rep = req->rl_reply;
1338 req->rl_reply = NULL;
1340 spin_lock(&buffers->rb_lock);
1341 list_add(&req->rl_list, &buffers->rb_send_bufs);
1343 if (!rep->rr_temp) {
1344 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1348 spin_unlock(&buffers->rb_lock);
1350 rpcrdma_destroy_rep(rep);
1354 * Put reply buffers back into pool when not attached to
1355 * request. This happens in error conditions.
1358 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1360 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1362 if (!rep->rr_temp) {
1363 spin_lock(&buffers->rb_lock);
1364 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1365 spin_unlock(&buffers->rb_lock);
1367 rpcrdma_destroy_rep(rep);
1372 * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1373 * @size: size of buffer to be allocated, in bytes
1374 * @direction: direction of data movement
1377 * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1378 * can be persistently DMA-mapped for I/O.
1380 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1381 * receiving the payload of RDMA RECV operations. During Long Calls
1382 * or Replies they may be registered externally via ro_map.
1384 struct rpcrdma_regbuf *
1385 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1388 struct rpcrdma_regbuf *rb;
1390 rb = kmalloc(sizeof(*rb) + size, flags);
1392 return ERR_PTR(-ENOMEM);
1394 rb->rg_device = NULL;
1395 rb->rg_direction = direction;
1396 rb->rg_iov.length = size;
1402 * __rpcrdma_map_regbuf - DMA-map a regbuf
1403 * @ia: controlling rpcrdma_ia
1404 * @rb: regbuf to be mapped
1407 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1409 struct ib_device *device = ia->ri_device;
1411 if (rb->rg_direction == DMA_NONE)
1414 rb->rg_iov.addr = ib_dma_map_single(device,
1415 (void *)rb->rg_base,
1418 if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1421 rb->rg_device = device;
1422 rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1427 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1432 if (!rpcrdma_regbuf_is_mapped(rb))
1435 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1436 rdmab_length(rb), rb->rg_direction);
1437 rb->rg_device = NULL;
1441 * rpcrdma_free_regbuf - deregister and free registered buffer
1442 * @rb: regbuf to be deregistered and freed
1445 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1447 rpcrdma_dma_unmap_regbuf(rb);
1452 * Prepost any receive buffer, then post send.
1454 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1457 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1458 struct rpcrdma_ep *ep,
1459 struct rpcrdma_req *req)
1461 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1464 if (!ep->rep_send_count ||
1465 test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1466 send_wr->send_flags |= IB_SEND_SIGNALED;
1467 ep->rep_send_count = ep->rep_send_batch;
1469 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1470 --ep->rep_send_count;
1473 rc = ia->ri_ops->ro_send(ia, req);
1474 trace_xprtrdma_post_send(req, rc);
1481 * rpcrdma_post_recvs - Maybe post some Receive buffers
1482 * @r_xprt: controlling transport
1483 * @temp: when true, allocate temp rpcrdma_rep objects
1487 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1489 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1490 struct ib_recv_wr *wr, *bad_wr;
1491 int needed, count, rc;
1493 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1494 if (buf->rb_posted_receives > needed)
1496 needed -= buf->rb_posted_receives;
1501 struct rpcrdma_regbuf *rb;
1502 struct rpcrdma_rep *rep;
1504 spin_lock(&buf->rb_lock);
1505 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1506 struct rpcrdma_rep, rr_list);
1508 list_del(&rep->rr_list);
1509 spin_unlock(&buf->rb_lock);
1511 if (rpcrdma_create_rep(r_xprt, temp))
1516 rb = rep->rr_rdmabuf;
1517 if (!rpcrdma_regbuf_is_mapped(rb)) {
1518 if (!__rpcrdma_dma_map_regbuf(&r_xprt->rx_ia, rb)) {
1519 rpcrdma_recv_buffer_put(rep);
1524 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1525 rep->rr_recv_wr.next = wr;
1526 wr = &rep->rr_recv_wr;
1533 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1534 (const struct ib_recv_wr **)&bad_wr);
1536 for (wr = bad_wr; wr; wr = wr->next) {
1537 struct rpcrdma_rep *rep;
1539 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1540 rpcrdma_recv_buffer_put(rep);
1544 buf->rb_posted_receives += count;
1545 trace_xprtrdma_post_recvs(r_xprt, count, rc);