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 struct rpcrdma_regbuf *
80 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
82 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
83 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
84 static void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
86 /* Wait for outstanding transport work to finish. ib_drain_qp
87 * handles the drains in the wrong order for us, so open code
90 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
92 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
94 /* Flush Receives, then wait for deferred Reply work
97 ib_drain_rq(ia->ri_id->qp);
99 /* Deferred Reply processing might have scheduled
100 * local invalidations.
102 ib_drain_sq(ia->ri_id->qp);
106 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
107 * @event: details of the event
108 * @context: ep that owns QP where event occurred
110 * Called from the RDMA provider (device driver) possibly in an interrupt
114 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
116 struct rpcrdma_ep *ep = context;
117 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
120 trace_xprtrdma_qp_event(r_xprt, event);
124 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
125 * @cq: completion queue (ignored)
130 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
132 struct ib_cqe *cqe = wc->wr_cqe;
133 struct rpcrdma_sendctx *sc =
134 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
136 /* WARNING: Only wr_cqe and status are reliable at this point */
137 trace_xprtrdma_wc_send(sc, wc);
138 rpcrdma_sendctx_put_locked(sc);
142 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
143 * @cq: completion queue (ignored)
148 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
150 struct ib_cqe *cqe = wc->wr_cqe;
151 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
153 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
155 /* WARNING: Only wr_cqe and status are reliable at this point */
156 trace_xprtrdma_wc_receive(wc);
157 --r_xprt->rx_ep.rep_receive_count;
158 if (wc->status != IB_WC_SUCCESS)
161 /* status == SUCCESS means all fields in wc are trustworthy */
162 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
163 rep->rr_wc_flags = wc->wc_flags;
164 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
166 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
167 rdmab_addr(rep->rr_rdmabuf),
168 wc->byte_len, DMA_FROM_DEVICE);
170 rpcrdma_post_recvs(r_xprt, false);
171 rpcrdma_reply_handler(rep);
175 rpcrdma_recv_buffer_put(rep);
179 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
180 struct rdma_conn_param *param)
182 const struct rpcrdma_connect_private *pmsg = param->private_data;
183 unsigned int rsize, wsize;
185 /* Default settings for RPC-over-RDMA Version One */
186 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
187 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
188 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
191 pmsg->cp_magic == rpcrdma_cmp_magic &&
192 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
193 r_xprt->rx_ia.ri_implicit_roundup = true;
194 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
195 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
198 if (rsize < r_xprt->rx_ep.rep_inline_recv)
199 r_xprt->rx_ep.rep_inline_recv = rsize;
200 if (wsize < r_xprt->rx_ep.rep_inline_send)
201 r_xprt->rx_ep.rep_inline_send = wsize;
202 dprintk("RPC: %s: max send %u, max recv %u\n", __func__,
203 r_xprt->rx_ep.rep_inline_send,
204 r_xprt->rx_ep.rep_inline_recv);
205 rpcrdma_set_max_header_sizes(r_xprt);
209 * rpcrdma_cm_event_handler - Handle RDMA CM events
210 * @id: rdma_cm_id on which an event has occurred
211 * @event: details of the event
213 * Called with @id's mutex held. Returns 1 if caller should
214 * destroy @id, otherwise 0.
217 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
219 struct rpcrdma_xprt *r_xprt = id->context;
220 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
221 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
222 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
226 trace_xprtrdma_cm_event(r_xprt, event);
227 switch (event->event) {
228 case RDMA_CM_EVENT_ADDR_RESOLVED:
229 case RDMA_CM_EVENT_ROUTE_RESOLVED:
231 complete(&ia->ri_done);
233 case RDMA_CM_EVENT_ADDR_ERROR:
234 ia->ri_async_rc = -EPROTO;
235 complete(&ia->ri_done);
237 case RDMA_CM_EVENT_ROUTE_ERROR:
238 ia->ri_async_rc = -ENETUNREACH;
239 complete(&ia->ri_done);
241 case RDMA_CM_EVENT_DEVICE_REMOVAL:
242 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
243 pr_info("rpcrdma: removing device %s for %s:%s\n",
244 ia->ri_id->device->name,
245 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
247 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
248 ep->rep_connected = -ENODEV;
249 xprt_force_disconnect(xprt);
250 wait_for_completion(&ia->ri_remove_done);
253 /* Return 1 to ensure the core destroys the id. */
255 case RDMA_CM_EVENT_ESTABLISHED:
256 ++xprt->connect_cookie;
257 ep->rep_connected = 1;
258 rpcrdma_update_connect_private(r_xprt, &event->param.conn);
259 wake_up_all(&ep->rep_connect_wait);
261 case RDMA_CM_EVENT_CONNECT_ERROR:
262 ep->rep_connected = -ENOTCONN;
264 case RDMA_CM_EVENT_UNREACHABLE:
265 ep->rep_connected = -ENETUNREACH;
267 case RDMA_CM_EVENT_REJECTED:
268 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
269 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
270 rdma_reject_msg(id, event->status));
271 ep->rep_connected = -ECONNREFUSED;
272 if (event->status == IB_CM_REJ_STALE_CONN)
273 ep->rep_connected = -EAGAIN;
275 case RDMA_CM_EVENT_DISCONNECTED:
276 ep->rep_connected = -ECONNABORTED;
278 xprt_force_disconnect(xprt);
279 wake_up_all(&ep->rep_connect_wait);
285 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__,
286 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
287 ia->ri_id->device->name, rdma_event_msg(event->event));
291 static struct rdma_cm_id *
292 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
294 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
295 struct rdma_cm_id *id;
298 trace_xprtrdma_conn_start(xprt);
300 init_completion(&ia->ri_done);
301 init_completion(&ia->ri_remove_done);
303 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
304 xprt, RDMA_PS_TCP, IB_QPT_RC);
308 ia->ri_async_rc = -ETIMEDOUT;
309 rc = rdma_resolve_addr(id, NULL,
310 (struct sockaddr *)&xprt->rx_xprt.addr,
311 RDMA_RESOLVE_TIMEOUT);
314 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
316 trace_xprtrdma_conn_tout(xprt);
320 rc = ia->ri_async_rc;
324 ia->ri_async_rc = -ETIMEDOUT;
325 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
328 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
330 trace_xprtrdma_conn_tout(xprt);
333 rc = ia->ri_async_rc;
345 * Exported functions.
349 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
350 * @xprt: transport with IA to (re)initialize
352 * Returns 0 on success, negative errno if an appropriate
353 * Interface Adapter could not be found and opened.
356 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
358 struct rpcrdma_ia *ia = &xprt->rx_ia;
361 ia->ri_id = rpcrdma_create_id(xprt, ia);
362 if (IS_ERR(ia->ri_id)) {
363 rc = PTR_ERR(ia->ri_id);
367 ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0);
368 if (IS_ERR(ia->ri_pd)) {
369 rc = PTR_ERR(ia->ri_pd);
370 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
374 switch (xprt_rdma_memreg_strategy) {
376 if (frwr_is_supported(ia->ri_id->device))
380 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
381 ia->ri_id->device->name, xprt_rdma_memreg_strategy);
389 rpcrdma_ia_close(ia);
394 * rpcrdma_ia_remove - Handle device driver unload
395 * @ia: interface adapter being removed
397 * Divest transport H/W resources associated with this adapter,
398 * but allow it to be restored later.
401 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
403 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
405 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
406 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
407 struct rpcrdma_req *req;
408 struct rpcrdma_rep *rep;
410 cancel_delayed_work_sync(&buf->rb_refresh_worker);
412 /* This is similar to rpcrdma_ep_destroy, but:
413 * - Don't cancel the connect worker.
414 * - Don't call rpcrdma_ep_disconnect, which waits
415 * for another conn upcall, which will deadlock.
416 * - rdma_disconnect is unneeded, the underlying
417 * connection is already gone.
420 rpcrdma_xprt_drain(r_xprt);
421 rdma_destroy_qp(ia->ri_id);
422 ia->ri_id->qp = NULL;
424 ib_free_cq(ep->rep_attr.recv_cq);
425 ep->rep_attr.recv_cq = NULL;
426 ib_free_cq(ep->rep_attr.send_cq);
427 ep->rep_attr.send_cq = NULL;
429 /* The ULP is responsible for ensuring all DMA
430 * mappings and MRs are gone.
432 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
433 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
434 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
435 rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf);
436 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
437 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
439 rpcrdma_mrs_destroy(buf);
440 ib_dealloc_pd(ia->ri_pd);
443 /* Allow waiters to continue */
444 complete(&ia->ri_remove_done);
446 trace_xprtrdma_remove(r_xprt);
450 * rpcrdma_ia_close - Clean up/close an IA.
451 * @ia: interface adapter to close
455 rpcrdma_ia_close(struct rpcrdma_ia *ia)
457 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
459 rdma_destroy_qp(ia->ri_id);
460 rdma_destroy_id(ia->ri_id);
464 /* If the pd is still busy, xprtrdma missed freeing a resource */
465 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
466 ib_dealloc_pd(ia->ri_pd);
471 * rpcrdma_ep_create - Create unconnected endpoint
472 * @r_xprt: transport to instantiate
474 * Returns zero on success, or a negative errno.
476 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
478 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
479 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
480 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
481 struct ib_cq *sendcq, *recvcq;
482 unsigned int max_sge;
485 ep->rep_max_requests = xprt_rdma_slot_table_entries;
486 ep->rep_inline_send = xprt_rdma_max_inline_write;
487 ep->rep_inline_recv = xprt_rdma_max_inline_read;
489 max_sge = min_t(unsigned int, ia->ri_id->device->attrs.max_send_sge,
490 RPCRDMA_MAX_SEND_SGES);
491 if (max_sge < RPCRDMA_MIN_SEND_SGES) {
492 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
495 ia->ri_max_send_sges = max_sge;
497 rc = frwr_open(ia, ep);
501 ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
502 ep->rep_attr.qp_context = ep;
503 ep->rep_attr.srq = NULL;
504 ep->rep_attr.cap.max_send_sge = max_sge;
505 ep->rep_attr.cap.max_recv_sge = 1;
506 ep->rep_attr.cap.max_inline_data = 0;
507 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
508 ep->rep_attr.qp_type = IB_QPT_RC;
509 ep->rep_attr.port_num = ~0;
511 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
512 "iovs: send %d recv %d\n",
514 ep->rep_attr.cap.max_send_wr,
515 ep->rep_attr.cap.max_recv_wr,
516 ep->rep_attr.cap.max_send_sge,
517 ep->rep_attr.cap.max_recv_sge);
519 ep->rep_send_batch = ep->rep_max_requests >> 3;
520 ep->rep_send_count = ep->rep_send_batch;
521 init_waitqueue_head(&ep->rep_connect_wait);
522 ep->rep_receive_count = 0;
524 sendcq = ib_alloc_cq_any(ia->ri_id->device, NULL,
525 ep->rep_attr.cap.max_send_wr + 1,
527 if (IS_ERR(sendcq)) {
528 rc = PTR_ERR(sendcq);
532 recvcq = ib_alloc_cq_any(ia->ri_id->device, NULL,
533 ep->rep_attr.cap.max_recv_wr + 1,
535 if (IS_ERR(recvcq)) {
536 rc = PTR_ERR(recvcq);
540 ep->rep_attr.send_cq = sendcq;
541 ep->rep_attr.recv_cq = recvcq;
543 /* Initialize cma parameters */
544 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
546 /* Prepare RDMA-CM private message */
547 pmsg->cp_magic = rpcrdma_cmp_magic;
548 pmsg->cp_version = RPCRDMA_CMP_VERSION;
549 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
550 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send);
551 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv);
552 ep->rep_remote_cma.private_data = pmsg;
553 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
555 /* Client offers RDMA Read but does not initiate */
556 ep->rep_remote_cma.initiator_depth = 0;
557 ep->rep_remote_cma.responder_resources =
558 min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom);
560 /* Limit transport retries so client can detect server
561 * GID changes quickly. RPC layer handles re-establishing
562 * transport connection and retransmission.
564 ep->rep_remote_cma.retry_count = 6;
566 /* RPC-over-RDMA handles its own flow control. In addition,
567 * make all RNR NAKs visible so we know that RPC-over-RDMA
568 * flow control is working correctly (no NAKs should be seen).
570 ep->rep_remote_cma.flow_control = 0;
571 ep->rep_remote_cma.rnr_retry_count = 0;
582 * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
583 * @r_xprt: transport instance to shut down
586 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
588 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
589 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
591 if (ia->ri_id && ia->ri_id->qp) {
592 rpcrdma_ep_disconnect(ep, ia);
593 rdma_destroy_qp(ia->ri_id);
594 ia->ri_id->qp = NULL;
597 if (ep->rep_attr.recv_cq)
598 ib_free_cq(ep->rep_attr.recv_cq);
599 if (ep->rep_attr.send_cq)
600 ib_free_cq(ep->rep_attr.send_cq);
603 /* Re-establish a connection after a device removal event.
604 * Unlike a normal reconnection, a fresh PD and a new set
605 * of MRs and buffers is needed.
608 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
609 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
613 trace_xprtrdma_reinsert(r_xprt);
616 if (rpcrdma_ia_open(r_xprt))
620 err = rpcrdma_ep_create(r_xprt);
622 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
627 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
629 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
633 rpcrdma_mrs_create(r_xprt);
637 rpcrdma_ep_destroy(r_xprt);
639 rpcrdma_ia_close(ia);
645 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
646 struct rpcrdma_ia *ia)
648 struct rdma_cm_id *id, *old;
651 trace_xprtrdma_reconnect(r_xprt);
653 rpcrdma_ep_disconnect(ep, ia);
656 id = rpcrdma_create_id(r_xprt, ia);
660 /* As long as the new ID points to the same device as the
661 * old ID, we can reuse the transport's existing PD and all
662 * previously allocated MRs. Also, the same device means
663 * the transport's previous DMA mappings are still valid.
665 * This is a sanity check only. There should be no way these
666 * point to two different devices here.
670 if (ia->ri_id->device != id->device) {
671 pr_err("rpcrdma: can't reconnect on different device!\n");
675 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
679 /* Atomically replace the transport's ID and QP. */
683 rdma_destroy_qp(old);
686 rdma_destroy_id(old);
692 * Connect unconnected endpoint.
695 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
697 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
699 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
703 switch (ep->rep_connected) {
705 dprintk("RPC: %s: connecting...\n", __func__);
706 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
713 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
718 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
723 ep->rep_connected = 0;
724 xprt_clear_connected(xprt);
726 rpcrdma_post_recvs(r_xprt, true);
728 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
732 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
733 if (ep->rep_connected <= 0) {
734 if (ep->rep_connected == -EAGAIN)
736 rc = ep->rep_connected;
740 dprintk("RPC: %s: connected\n", __func__);
744 ep->rep_connected = rc;
751 * rpcrdma_ep_disconnect - Disconnect underlying transport
752 * @ep: endpoint to disconnect
753 * @ia: associated interface adapter
755 * This is separate from destroy to facilitate the ability
756 * to reconnect without recreating the endpoint.
758 * This call is not reentrant, and must not be made in parallel
759 * on the same endpoint.
762 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
764 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
768 /* returns without wait if ID is not connected */
769 rc = rdma_disconnect(ia->ri_id);
771 wait_event_interruptible(ep->rep_connect_wait,
772 ep->rep_connected != 1);
774 ep->rep_connected = rc;
775 trace_xprtrdma_disconnect(r_xprt, rc);
777 rpcrdma_xprt_drain(r_xprt);
780 /* Fixed-size circular FIFO queue. This implementation is wait-free and
783 * Consumer is the code path that posts Sends. This path dequeues a
784 * sendctx for use by a Send operation. Multiple consumer threads
785 * are serialized by the RPC transport lock, which allows only one
786 * ->send_request call at a time.
788 * Producer is the code path that handles Send completions. This path
789 * enqueues a sendctx that has been completed. Multiple producer
790 * threads are serialized by the ib_poll_cq() function.
793 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
794 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
797 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
801 for (i = 0; i <= buf->rb_sc_last; i++)
802 kfree(buf->rb_sc_ctxs[i]);
803 kfree(buf->rb_sc_ctxs);
806 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
808 struct rpcrdma_sendctx *sc;
810 sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges),
815 sc->sc_wr.wr_cqe = &sc->sc_cqe;
816 sc->sc_wr.sg_list = sc->sc_sges;
817 sc->sc_wr.opcode = IB_WR_SEND;
818 sc->sc_cqe.done = rpcrdma_wc_send;
822 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
824 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
825 struct rpcrdma_sendctx *sc;
828 /* Maximum number of concurrent outstanding Send WRs. Capping
829 * the circular queue size stops Send Queue overflow by causing
830 * the ->send_request call to fail temporarily before too many
833 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
834 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
835 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
836 if (!buf->rb_sc_ctxs)
839 buf->rb_sc_last = i - 1;
840 for (i = 0; i <= buf->rb_sc_last; i++) {
841 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
845 sc->sc_xprt = r_xprt;
846 buf->rb_sc_ctxs[i] = sc;
852 /* The sendctx queue is not guaranteed to have a size that is a
853 * power of two, thus the helpers in circ_buf.h cannot be used.
854 * The other option is to use modulus (%), which can be expensive.
856 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
859 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
863 * rpcrdma_sendctx_get_locked - Acquire a send context
864 * @r_xprt: controlling transport instance
866 * Returns pointer to a free send completion context; or NULL if
867 * the queue is empty.
869 * Usage: Called to acquire an SGE array before preparing a Send WR.
871 * The caller serializes calls to this function (per transport), and
872 * provides an effective memory barrier that flushes the new value
875 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
877 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
878 struct rpcrdma_sendctx *sc;
879 unsigned long next_head;
881 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
883 if (next_head == READ_ONCE(buf->rb_sc_tail))
886 /* ORDER: item must be accessed _before_ head is updated */
887 sc = buf->rb_sc_ctxs[next_head];
889 /* Releasing the lock in the caller acts as a memory
890 * barrier that flushes rb_sc_head.
892 buf->rb_sc_head = next_head;
897 /* The queue is "empty" if there have not been enough Send
898 * completions recently. This is a sign the Send Queue is
899 * backing up. Cause the caller to pause and try again.
901 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
902 r_xprt->rx_stats.empty_sendctx_q++;
907 * rpcrdma_sendctx_put_locked - Release a send context
908 * @sc: send context to release
910 * Usage: Called from Send completion to return a sendctxt
913 * The caller serializes calls to this function (per transport).
916 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
918 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
919 unsigned long next_tail;
921 /* Unmap SGEs of previously completed but unsignaled
922 * Sends by walking up the queue until @sc is found.
924 next_tail = buf->rb_sc_tail;
926 next_tail = rpcrdma_sendctx_next(buf, next_tail);
928 /* ORDER: item must be accessed _before_ tail is updated */
929 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
931 } while (buf->rb_sc_ctxs[next_tail] != sc);
933 /* Paired with READ_ONCE */
934 smp_store_release(&buf->rb_sc_tail, next_tail);
936 xprt_write_space(&sc->sc_xprt->rx_xprt);
940 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
942 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
943 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
948 for (count = 0; count < ia->ri_max_segs; count++) {
949 struct rpcrdma_mr *mr;
952 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
956 rc = frwr_init_mr(ia, mr);
962 mr->mr_xprt = r_xprt;
964 list_add(&mr->mr_list, &free);
965 list_add(&mr->mr_all, &all);
968 spin_lock(&buf->rb_mrlock);
969 list_splice(&free, &buf->rb_mrs);
970 list_splice(&all, &buf->rb_all);
971 r_xprt->rx_stats.mrs_allocated += count;
972 spin_unlock(&buf->rb_mrlock);
973 trace_xprtrdma_createmrs(r_xprt, count);
977 rpcrdma_mr_refresh_worker(struct work_struct *work)
979 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
980 rb_refresh_worker.work);
981 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
984 rpcrdma_mrs_create(r_xprt);
985 xprt_write_space(&r_xprt->rx_xprt);
989 * rpcrdma_req_create - Allocate an rpcrdma_req object
990 * @r_xprt: controlling r_xprt
991 * @size: initial size, in bytes, of send and receive buffers
992 * @flags: GFP flags passed to memory allocators
994 * Returns an allocated and fully initialized rpcrdma_req or NULL.
996 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
999 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1000 struct rpcrdma_regbuf *rb;
1001 struct rpcrdma_req *req;
1003 req = kzalloc(sizeof(*req), flags);
1007 rb = rpcrdma_regbuf_alloc(RPCRDMA_HDRBUF_SIZE, DMA_TO_DEVICE, flags);
1010 req->rl_rdmabuf = rb;
1011 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1013 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1014 if (!req->rl_sendbuf)
1017 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1018 if (!req->rl_recvbuf)
1021 INIT_LIST_HEAD(&req->rl_registered);
1022 spin_lock(&buffer->rb_lock);
1023 list_add(&req->rl_all, &buffer->rb_allreqs);
1024 spin_unlock(&buffer->rb_lock);
1028 kfree(req->rl_sendbuf);
1030 kfree(req->rl_rdmabuf);
1037 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1040 struct rpcrdma_rep *rep;
1042 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1046 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv,
1047 DMA_FROM_DEVICE, GFP_KERNEL);
1048 if (!rep->rr_rdmabuf)
1051 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
1052 rdmab_length(rep->rr_rdmabuf));
1053 rep->rr_cqe.done = rpcrdma_wc_receive;
1054 rep->rr_rxprt = r_xprt;
1055 rep->rr_recv_wr.next = NULL;
1056 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1057 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1058 rep->rr_recv_wr.num_sge = 1;
1059 rep->rr_temp = temp;
1069 * rpcrdma_buffer_create - Create initial set of req/rep objects
1070 * @r_xprt: transport instance to (re)initialize
1072 * Returns zero on success, otherwise a negative errno.
1074 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1076 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1079 buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
1080 buf->rb_bc_srv_max_requests = 0;
1081 spin_lock_init(&buf->rb_mrlock);
1082 spin_lock_init(&buf->rb_lock);
1083 INIT_LIST_HEAD(&buf->rb_mrs);
1084 INIT_LIST_HEAD(&buf->rb_all);
1085 INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1086 rpcrdma_mr_refresh_worker);
1088 rpcrdma_mrs_create(r_xprt);
1090 INIT_LIST_HEAD(&buf->rb_send_bufs);
1091 INIT_LIST_HEAD(&buf->rb_allreqs);
1094 for (i = 0; i < buf->rb_max_requests; i++) {
1095 struct rpcrdma_req *req;
1097 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE,
1101 list_add(&req->rl_list, &buf->rb_send_bufs);
1104 buf->rb_credits = 1;
1105 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1107 rc = rpcrdma_sendctxs_create(r_xprt);
1113 rpcrdma_buffer_destroy(buf);
1117 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1119 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1124 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1125 * @req: unused object to be destroyed
1127 * This function assumes that the caller prevents concurrent device
1128 * unload and transport tear-down.
1131 rpcrdma_req_destroy(struct rpcrdma_req *req)
1133 list_del(&req->rl_all);
1135 rpcrdma_regbuf_free(req->rl_recvbuf);
1136 rpcrdma_regbuf_free(req->rl_sendbuf);
1137 rpcrdma_regbuf_free(req->rl_rdmabuf);
1142 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1144 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1146 struct rpcrdma_mr *mr;
1150 spin_lock(&buf->rb_mrlock);
1151 while (!list_empty(&buf->rb_all)) {
1152 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1153 list_del(&mr->mr_all);
1155 spin_unlock(&buf->rb_mrlock);
1157 /* Ensure MW is not on any rl_registered list */
1158 if (!list_empty(&mr->mr_list))
1159 list_del(&mr->mr_list);
1161 frwr_release_mr(mr);
1163 spin_lock(&buf->rb_mrlock);
1165 spin_unlock(&buf->rb_mrlock);
1166 r_xprt->rx_stats.mrs_allocated = 0;
1168 dprintk("RPC: %s: released %u MRs\n", __func__, count);
1172 * rpcrdma_buffer_destroy - Release all hw resources
1173 * @buf: root control block for resources
1175 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1176 * - No more Send or Receive completions can occur
1177 * - All MRs, reps, and reqs are returned to their free lists
1180 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1182 cancel_delayed_work_sync(&buf->rb_refresh_worker);
1184 rpcrdma_sendctxs_destroy(buf);
1186 while (!list_empty(&buf->rb_recv_bufs)) {
1187 struct rpcrdma_rep *rep;
1189 rep = list_first_entry(&buf->rb_recv_bufs,
1190 struct rpcrdma_rep, rr_list);
1191 list_del(&rep->rr_list);
1192 rpcrdma_rep_destroy(rep);
1195 while (!list_empty(&buf->rb_send_bufs)) {
1196 struct rpcrdma_req *req;
1198 req = list_first_entry(&buf->rb_send_bufs,
1199 struct rpcrdma_req, rl_list);
1200 list_del(&req->rl_list);
1201 rpcrdma_req_destroy(req);
1204 rpcrdma_mrs_destroy(buf);
1208 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1209 * @r_xprt: controlling transport
1211 * Returns an initialized rpcrdma_mr or NULL if no free
1212 * rpcrdma_mr objects are available.
1215 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1217 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1218 struct rpcrdma_mr *mr = NULL;
1220 spin_lock(&buf->rb_mrlock);
1221 if (!list_empty(&buf->rb_mrs))
1222 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1223 spin_unlock(&buf->rb_mrlock);
1230 trace_xprtrdma_nomrs(r_xprt);
1231 if (r_xprt->rx_ep.rep_connected != -ENODEV)
1232 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1234 /* Allow the reply handler and refresh worker to run */
1241 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1243 spin_lock(&buf->rb_mrlock);
1244 rpcrdma_mr_push(mr, &buf->rb_mrs);
1245 spin_unlock(&buf->rb_mrlock);
1249 * rpcrdma_mr_put - Release an rpcrdma_mr object
1250 * @mr: object to release
1254 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1256 __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1260 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1261 * @mr: object to release
1265 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1267 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1269 if (mr->mr_dir != DMA_NONE) {
1270 trace_xprtrdma_mr_unmap(mr);
1271 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1272 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1273 mr->mr_dir = DMA_NONE;
1275 __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1279 * rpcrdma_buffer_get - Get a request buffer
1280 * @buffers: Buffer pool from which to obtain a buffer
1282 * Returns a fresh rpcrdma_req, or NULL if none are available.
1284 struct rpcrdma_req *
1285 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1287 struct rpcrdma_req *req;
1289 spin_lock(&buffers->rb_lock);
1290 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1291 struct rpcrdma_req, rl_list);
1293 list_del_init(&req->rl_list);
1294 spin_unlock(&buffers->rb_lock);
1299 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1300 * @buffers: buffer pool
1301 * @req: object to return
1304 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1306 struct rpcrdma_rep *rep = req->rl_reply;
1308 req->rl_reply = NULL;
1310 spin_lock(&buffers->rb_lock);
1311 list_add(&req->rl_list, &buffers->rb_send_bufs);
1313 if (!rep->rr_temp) {
1314 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1318 spin_unlock(&buffers->rb_lock);
1320 rpcrdma_rep_destroy(rep);
1324 * Put reply buffers back into pool when not attached to
1325 * request. This happens in error conditions.
1328 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1330 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1332 if (!rep->rr_temp) {
1333 spin_lock(&buffers->rb_lock);
1334 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1335 spin_unlock(&buffers->rb_lock);
1337 rpcrdma_rep_destroy(rep);
1341 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1343 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1344 * receiving the payload of RDMA RECV operations. During Long Calls
1345 * or Replies they may be registered externally via frwr_map.
1347 static struct rpcrdma_regbuf *
1348 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1351 struct rpcrdma_regbuf *rb;
1353 rb = kmalloc(sizeof(*rb), flags);
1356 rb->rg_data = kmalloc(size, flags);
1362 rb->rg_device = NULL;
1363 rb->rg_direction = direction;
1364 rb->rg_iov.length = size;
1369 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1370 * @rb: regbuf to reallocate
1371 * @size: size of buffer to be allocated, in bytes
1374 * Returns true if reallocation was successful. If false is
1375 * returned, @rb is left untouched.
1377 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1381 buf = kmalloc(size, flags);
1385 rpcrdma_regbuf_dma_unmap(rb);
1389 rb->rg_iov.length = size;
1394 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1395 * @r_xprt: controlling transport instance
1396 * @rb: regbuf to be mapped
1398 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1400 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1401 struct rpcrdma_regbuf *rb)
1403 struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1405 if (rb->rg_direction == DMA_NONE)
1408 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1409 rdmab_length(rb), rb->rg_direction);
1410 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1411 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1415 rb->rg_device = device;
1416 rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1420 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1425 if (!rpcrdma_regbuf_is_mapped(rb))
1428 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1430 rb->rg_device = NULL;
1433 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1435 rpcrdma_regbuf_dma_unmap(rb);
1442 * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1443 * @ia: transport's device information
1444 * @ep: transport's RDMA endpoint information
1445 * @req: rpcrdma_req containing the Send WR to post
1447 * Returns 0 if the post was successful, otherwise -ENOTCONN
1451 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1452 struct rpcrdma_ep *ep,
1453 struct rpcrdma_req *req)
1455 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1458 if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1459 send_wr->send_flags |= IB_SEND_SIGNALED;
1460 ep->rep_send_count = ep->rep_send_batch;
1462 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1463 --ep->rep_send_count;
1466 rc = frwr_send(ia, req);
1467 trace_xprtrdma_post_send(req, rc);
1474 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1476 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1477 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1478 struct ib_recv_wr *i, *wr, *bad_wr;
1479 struct rpcrdma_rep *rep;
1480 int needed, count, rc;
1485 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1486 if (ep->rep_receive_count > needed)
1488 needed -= ep->rep_receive_count;
1490 needed += RPCRDMA_MAX_RECV_BATCH;
1492 /* fast path: all needed reps can be found on the free list */
1494 spin_lock(&buf->rb_lock);
1496 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1497 struct rpcrdma_rep, rr_list);
1501 list_del(&rep->rr_list);
1502 rep->rr_recv_wr.next = wr;
1503 wr = &rep->rr_recv_wr;
1506 spin_unlock(&buf->rb_lock);
1509 rep = rpcrdma_rep_create(r_xprt, temp);
1513 rep->rr_recv_wr.next = wr;
1514 wr = &rep->rr_recv_wr;
1520 for (i = wr; i; i = i->next) {
1521 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1523 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1526 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1530 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1531 (const struct ib_recv_wr **)&bad_wr);
1533 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1535 for (wr = bad_wr; wr;) {
1536 struct rpcrdma_rep *rep;
1538 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1540 rpcrdma_recv_buffer_put(rep);
1544 ep->rep_receive_count += count;
1549 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1551 rpcrdma_recv_buffer_put(rep);