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);
81 static void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
83 /* Wait for outstanding transport work to finish.
85 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
87 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
88 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
90 /* Flush Receives, then wait for deferred Reply work
93 ib_drain_qp(ia->ri_id->qp);
94 drain_workqueue(buf->rb_completion_wq);
96 /* Deferred Reply processing might have scheduled
97 * local invalidations.
99 ib_drain_sq(ia->ri_id->qp);
103 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
104 * @event: details of the event
105 * @context: ep that owns QP where event occurred
107 * Called from the RDMA provider (device driver) possibly in an interrupt
111 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
113 struct rpcrdma_ep *ep = context;
114 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
117 trace_xprtrdma_qp_event(r_xprt, event);
121 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
122 * @cq: completion queue (ignored)
127 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
129 struct ib_cqe *cqe = wc->wr_cqe;
130 struct rpcrdma_sendctx *sc =
131 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
133 /* WARNING: Only wr_cqe and status are reliable at this point */
134 trace_xprtrdma_wc_send(sc, wc);
135 if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
136 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
137 ib_wc_status_msg(wc->status),
138 wc->status, wc->vendor_err);
140 rpcrdma_sendctx_put_locked(sc);
144 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
145 * @cq: completion queue (ignored)
150 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
152 struct ib_cqe *cqe = wc->wr_cqe;
153 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
155 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
157 /* WARNING: Only wr_cqe and status are reliable at this point */
158 trace_xprtrdma_wc_receive(wc);
159 --r_xprt->rx_ep.rep_receive_count;
160 if (wc->status != IB_WC_SUCCESS)
163 /* status == SUCCESS means all fields in wc are trustworthy */
164 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
165 rep->rr_wc_flags = wc->wc_flags;
166 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
168 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
169 rdmab_addr(rep->rr_rdmabuf),
170 wc->byte_len, DMA_FROM_DEVICE);
172 rpcrdma_post_recvs(r_xprt, false);
173 rpcrdma_reply_handler(rep);
177 if (wc->status != IB_WC_WR_FLUSH_ERR)
178 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
179 ib_wc_status_msg(wc->status),
180 wc->status, wc->vendor_err);
181 rpcrdma_recv_buffer_put(rep);
185 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
186 struct rdma_conn_param *param)
188 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
189 const struct rpcrdma_connect_private *pmsg = param->private_data;
190 unsigned int rsize, wsize;
192 /* Default settings for RPC-over-RDMA Version One */
193 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
194 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
195 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
198 pmsg->cp_magic == rpcrdma_cmp_magic &&
199 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
200 r_xprt->rx_ia.ri_implicit_roundup = true;
201 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
202 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
205 if (rsize < cdata->inline_rsize)
206 cdata->inline_rsize = rsize;
207 if (wsize < cdata->inline_wsize)
208 cdata->inline_wsize = wsize;
209 dprintk("RPC: %s: max send %u, max recv %u\n",
210 __func__, cdata->inline_wsize, cdata->inline_rsize);
211 rpcrdma_set_max_header_sizes(r_xprt);
215 * rpcrdma_cm_event_handler - Handle RDMA CM events
216 * @id: rdma_cm_id on which an event has occurred
217 * @event: details of the event
219 * Called with @id's mutex held. Returns 1 if caller should
220 * destroy @id, otherwise 0.
223 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
225 struct rpcrdma_xprt *r_xprt = id->context;
226 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
227 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
228 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
232 trace_xprtrdma_cm_event(r_xprt, event);
233 switch (event->event) {
234 case RDMA_CM_EVENT_ADDR_RESOLVED:
235 case RDMA_CM_EVENT_ROUTE_RESOLVED:
237 complete(&ia->ri_done);
239 case RDMA_CM_EVENT_ADDR_ERROR:
240 ia->ri_async_rc = -EPROTO;
241 complete(&ia->ri_done);
243 case RDMA_CM_EVENT_ROUTE_ERROR:
244 ia->ri_async_rc = -ENETUNREACH;
245 complete(&ia->ri_done);
247 case RDMA_CM_EVENT_DEVICE_REMOVAL:
248 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
249 pr_info("rpcrdma: removing device %s for %s:%s\n",
251 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
253 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
254 ep->rep_connected = -ENODEV;
255 xprt_force_disconnect(xprt);
256 wait_for_completion(&ia->ri_remove_done);
259 ia->ri_device = NULL;
260 /* Return 1 to ensure the core destroys the id. */
262 case RDMA_CM_EVENT_ESTABLISHED:
263 ++xprt->connect_cookie;
264 ep->rep_connected = 1;
265 rpcrdma_update_connect_private(r_xprt, &event->param.conn);
266 wake_up_all(&ep->rep_connect_wait);
268 case RDMA_CM_EVENT_CONNECT_ERROR:
269 ep->rep_connected = -ENOTCONN;
271 case RDMA_CM_EVENT_UNREACHABLE:
272 ep->rep_connected = -ENETUNREACH;
274 case RDMA_CM_EVENT_REJECTED:
275 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
276 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
277 rdma_reject_msg(id, event->status));
278 ep->rep_connected = -ECONNREFUSED;
279 if (event->status == IB_CM_REJ_STALE_CONN)
280 ep->rep_connected = -EAGAIN;
282 case RDMA_CM_EVENT_DISCONNECTED:
283 ep->rep_connected = -ECONNABORTED;
285 xprt_force_disconnect(xprt);
286 wake_up_all(&ep->rep_connect_wait);
292 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__,
293 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
294 ia->ri_device->name, rdma_event_msg(event->event));
298 static struct rdma_cm_id *
299 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
301 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
302 struct rdma_cm_id *id;
305 trace_xprtrdma_conn_start(xprt);
307 init_completion(&ia->ri_done);
308 init_completion(&ia->ri_remove_done);
310 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
311 xprt, RDMA_PS_TCP, IB_QPT_RC);
315 ia->ri_async_rc = -ETIMEDOUT;
316 rc = rdma_resolve_addr(id, NULL,
317 (struct sockaddr *)&xprt->rx_xprt.addr,
318 RDMA_RESOLVE_TIMEOUT);
321 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
323 trace_xprtrdma_conn_tout(xprt);
327 rc = ia->ri_async_rc;
331 ia->ri_async_rc = -ETIMEDOUT;
332 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
335 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
337 trace_xprtrdma_conn_tout(xprt);
340 rc = ia->ri_async_rc;
352 * Exported functions.
356 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
357 * @xprt: transport with IA to (re)initialize
359 * Returns 0 on success, negative errno if an appropriate
360 * Interface Adapter could not be found and opened.
363 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
365 struct rpcrdma_ia *ia = &xprt->rx_ia;
368 ia->ri_id = rpcrdma_create_id(xprt, ia);
369 if (IS_ERR(ia->ri_id)) {
370 rc = PTR_ERR(ia->ri_id);
373 ia->ri_device = ia->ri_id->device;
375 ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
376 if (IS_ERR(ia->ri_pd)) {
377 rc = PTR_ERR(ia->ri_pd);
378 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
382 switch (xprt_rdma_memreg_strategy) {
384 if (frwr_is_supported(ia))
388 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
389 ia->ri_device->name, xprt_rdma_memreg_strategy);
397 rpcrdma_ia_close(ia);
402 * rpcrdma_ia_remove - Handle device driver unload
403 * @ia: interface adapter being removed
405 * Divest transport H/W resources associated with this adapter,
406 * but allow it to be restored later.
409 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
411 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
413 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
414 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
415 struct rpcrdma_req *req;
416 struct rpcrdma_rep *rep;
418 cancel_delayed_work_sync(&buf->rb_refresh_worker);
420 /* This is similar to rpcrdma_ep_destroy, but:
421 * - Don't cancel the connect worker.
422 * - Don't call rpcrdma_ep_disconnect, which waits
423 * for another conn upcall, which will deadlock.
424 * - rdma_disconnect is unneeded, the underlying
425 * connection is already gone.
428 rpcrdma_xprt_drain(r_xprt);
429 rdma_destroy_qp(ia->ri_id);
430 ia->ri_id->qp = NULL;
432 ib_free_cq(ep->rep_attr.recv_cq);
433 ep->rep_attr.recv_cq = NULL;
434 ib_free_cq(ep->rep_attr.send_cq);
435 ep->rep_attr.send_cq = NULL;
437 /* The ULP is responsible for ensuring all DMA
438 * mappings and MRs are gone.
440 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
441 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
442 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
443 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
444 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
445 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
447 rpcrdma_mrs_destroy(buf);
448 ib_dealloc_pd(ia->ri_pd);
451 /* Allow waiters to continue */
452 complete(&ia->ri_remove_done);
454 trace_xprtrdma_remove(r_xprt);
458 * rpcrdma_ia_close - Clean up/close an IA.
459 * @ia: interface adapter to close
463 rpcrdma_ia_close(struct rpcrdma_ia *ia)
465 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
467 rdma_destroy_qp(ia->ri_id);
468 rdma_destroy_id(ia->ri_id);
471 ia->ri_device = NULL;
473 /* If the pd is still busy, xprtrdma missed freeing a resource */
474 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
475 ib_dealloc_pd(ia->ri_pd);
480 * Create unconnected endpoint.
483 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
484 struct rpcrdma_create_data_internal *cdata)
486 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
487 struct ib_cq *sendcq, *recvcq;
488 unsigned int max_sge;
491 max_sge = min_t(unsigned int, ia->ri_device->attrs.max_send_sge,
492 RPCRDMA_MAX_SEND_SGES);
493 if (max_sge < RPCRDMA_MIN_SEND_SGES) {
494 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
497 ia->ri_max_send_sges = max_sge;
499 rc = frwr_open(ia, ep, cdata);
503 ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
504 ep->rep_attr.qp_context = ep;
505 ep->rep_attr.srq = NULL;
506 ep->rep_attr.cap.max_send_sge = max_sge;
507 ep->rep_attr.cap.max_recv_sge = 1;
508 ep->rep_attr.cap.max_inline_data = 0;
509 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
510 ep->rep_attr.qp_type = IB_QPT_RC;
511 ep->rep_attr.port_num = ~0;
513 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
514 "iovs: send %d recv %d\n",
516 ep->rep_attr.cap.max_send_wr,
517 ep->rep_attr.cap.max_recv_wr,
518 ep->rep_attr.cap.max_send_sge,
519 ep->rep_attr.cap.max_recv_sge);
521 /* set trigger for requesting send completion */
522 ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
523 cdata->max_requests >> 2);
524 ep->rep_send_count = ep->rep_send_batch;
525 init_waitqueue_head(&ep->rep_connect_wait);
526 ep->rep_receive_count = 0;
528 sendcq = ib_alloc_cq(ia->ri_device, NULL,
529 ep->rep_attr.cap.max_send_wr + 1,
530 1, IB_POLL_WORKQUEUE);
531 if (IS_ERR(sendcq)) {
532 rc = PTR_ERR(sendcq);
536 recvcq = ib_alloc_cq(ia->ri_device, NULL,
537 ep->rep_attr.cap.max_recv_wr + 1,
538 0, IB_POLL_WORKQUEUE);
539 if (IS_ERR(recvcq)) {
540 rc = PTR_ERR(recvcq);
544 ep->rep_attr.send_cq = sendcq;
545 ep->rep_attr.recv_cq = recvcq;
547 /* Initialize cma parameters */
548 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
550 /* Prepare RDMA-CM private message */
551 pmsg->cp_magic = rpcrdma_cmp_magic;
552 pmsg->cp_version = RPCRDMA_CMP_VERSION;
553 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
554 pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
555 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
556 ep->rep_remote_cma.private_data = pmsg;
557 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
559 /* Client offers RDMA Read but does not initiate */
560 ep->rep_remote_cma.initiator_depth = 0;
561 ep->rep_remote_cma.responder_resources =
562 min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
564 /* Limit transport retries so client can detect server
565 * GID changes quickly. RPC layer handles re-establishing
566 * transport connection and retransmission.
568 ep->rep_remote_cma.retry_count = 6;
570 /* RPC-over-RDMA handles its own flow control. In addition,
571 * make all RNR NAKs visible so we know that RPC-over-RDMA
572 * flow control is working correctly (no NAKs should be seen).
574 ep->rep_remote_cma.flow_control = 0;
575 ep->rep_remote_cma.rnr_retry_count = 0;
588 * Disconnect and destroy endpoint. After this, the only
589 * valid operations on the ep are to free it (if dynamically
590 * allocated) or re-create it.
593 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
595 if (ia->ri_id && ia->ri_id->qp) {
596 rpcrdma_ep_disconnect(ep, ia);
597 rdma_destroy_qp(ia->ri_id);
598 ia->ri_id->qp = NULL;
601 if (ep->rep_attr.recv_cq)
602 ib_free_cq(ep->rep_attr.recv_cq);
603 if (ep->rep_attr.send_cq)
604 ib_free_cq(ep->rep_attr.send_cq);
607 /* Re-establish a connection after a device removal event.
608 * Unlike a normal reconnection, a fresh PD and a new set
609 * of MRs and buffers is needed.
612 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
613 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
617 trace_xprtrdma_reinsert(r_xprt);
620 if (rpcrdma_ia_open(r_xprt))
624 err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
626 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
631 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
633 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
637 rpcrdma_mrs_create(r_xprt);
641 rpcrdma_ep_destroy(ep, ia);
643 rpcrdma_ia_close(ia);
649 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
650 struct rpcrdma_ia *ia)
652 struct rdma_cm_id *id, *old;
655 trace_xprtrdma_reconnect(r_xprt);
657 rpcrdma_ep_disconnect(ep, ia);
660 id = rpcrdma_create_id(r_xprt, ia);
664 /* As long as the new ID points to the same device as the
665 * old ID, we can reuse the transport's existing PD and all
666 * previously allocated MRs. Also, the same device means
667 * the transport's previous DMA mappings are still valid.
669 * This is a sanity check only. There should be no way these
670 * point to two different devices here.
674 if (ia->ri_device != id->device) {
675 pr_err("rpcrdma: can't reconnect on different device!\n");
679 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
683 /* Atomically replace the transport's ID and QP. */
687 rdma_destroy_qp(old);
690 rdma_destroy_id(old);
696 * Connect unconnected endpoint.
699 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
701 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
703 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
707 switch (ep->rep_connected) {
709 dprintk("RPC: %s: connecting...\n", __func__);
710 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
717 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
722 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
727 ep->rep_connected = 0;
728 xprt_clear_connected(xprt);
730 rpcrdma_post_recvs(r_xprt, true);
732 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
736 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
737 if (ep->rep_connected <= 0) {
738 if (ep->rep_connected == -EAGAIN)
740 rc = ep->rep_connected;
744 dprintk("RPC: %s: connected\n", __func__);
748 ep->rep_connected = rc;
755 * rpcrdma_ep_disconnect - Disconnect underlying transport
756 * @ep: endpoint to disconnect
757 * @ia: associated interface adapter
759 * This is separate from destroy to facilitate the ability
760 * to reconnect without recreating the endpoint.
762 * This call is not reentrant, and must not be made in parallel
763 * on the same endpoint.
766 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
768 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
772 /* returns without wait if ID is not connected */
773 rc = rdma_disconnect(ia->ri_id);
775 wait_event_interruptible(ep->rep_connect_wait,
776 ep->rep_connected != 1);
778 ep->rep_connected = rc;
779 trace_xprtrdma_disconnect(r_xprt, rc);
781 rpcrdma_xprt_drain(r_xprt);
784 /* Fixed-size circular FIFO queue. This implementation is wait-free and
787 * Consumer is the code path that posts Sends. This path dequeues a
788 * sendctx for use by a Send operation. Multiple consumer threads
789 * are serialized by the RPC transport lock, which allows only one
790 * ->send_request call at a time.
792 * Producer is the code path that handles Send completions. This path
793 * enqueues a sendctx that has been completed. Multiple producer
794 * threads are serialized by the ib_poll_cq() function.
797 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
798 * queue activity, and ib_drain_qp has flushed all remaining Send
801 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
805 for (i = 0; i <= buf->rb_sc_last; i++)
806 kfree(buf->rb_sc_ctxs[i]);
807 kfree(buf->rb_sc_ctxs);
810 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
812 struct rpcrdma_sendctx *sc;
814 sc = kzalloc(sizeof(*sc) +
815 ia->ri_max_send_sges * sizeof(struct ib_sge),
820 sc->sc_wr.wr_cqe = &sc->sc_cqe;
821 sc->sc_wr.sg_list = sc->sc_sges;
822 sc->sc_wr.opcode = IB_WR_SEND;
823 sc->sc_cqe.done = rpcrdma_wc_send;
827 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
829 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
830 struct rpcrdma_sendctx *sc;
833 /* Maximum number of concurrent outstanding Send WRs. Capping
834 * the circular queue size stops Send Queue overflow by causing
835 * the ->send_request call to fail temporarily before too many
838 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
839 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
840 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
841 if (!buf->rb_sc_ctxs)
844 buf->rb_sc_last = i - 1;
845 for (i = 0; i <= buf->rb_sc_last; i++) {
846 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
850 sc->sc_xprt = r_xprt;
851 buf->rb_sc_ctxs[i] = sc;
857 rpcrdma_sendctxs_destroy(buf);
861 /* The sendctx queue is not guaranteed to have a size that is a
862 * power of two, thus the helpers in circ_buf.h cannot be used.
863 * The other option is to use modulus (%), which can be expensive.
865 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
868 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
872 * rpcrdma_sendctx_get_locked - Acquire a send context
873 * @buf: transport buffers from which to acquire an unused context
875 * Returns pointer to a free send completion context; or NULL if
876 * the queue is empty.
878 * Usage: Called to acquire an SGE array before preparing a Send WR.
880 * The caller serializes calls to this function (per rpcrdma_buffer),
881 * and provides an effective memory barrier that flushes the new value
884 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
886 struct rpcrdma_xprt *r_xprt;
887 struct rpcrdma_sendctx *sc;
888 unsigned long next_head;
890 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
892 if (next_head == READ_ONCE(buf->rb_sc_tail))
895 /* ORDER: item must be accessed _before_ head is updated */
896 sc = buf->rb_sc_ctxs[next_head];
898 /* Releasing the lock in the caller acts as a memory
899 * barrier that flushes rb_sc_head.
901 buf->rb_sc_head = next_head;
906 /* The queue is "empty" if there have not been enough Send
907 * completions recently. This is a sign the Send Queue is
908 * backing up. Cause the caller to pause and try again.
910 set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
911 r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
912 r_xprt->rx_stats.empty_sendctx_q++;
917 * rpcrdma_sendctx_put_locked - Release a send context
918 * @sc: send context to release
920 * Usage: Called from Send completion to return a sendctxt
923 * The caller serializes calls to this function (per rpcrdma_buffer).
926 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
928 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
929 unsigned long next_tail;
931 /* Unmap SGEs of previously completed by unsignaled
932 * Sends by walking up the queue until @sc is found.
934 next_tail = buf->rb_sc_tail;
936 next_tail = rpcrdma_sendctx_next(buf, next_tail);
938 /* ORDER: item must be accessed _before_ tail is updated */
939 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
941 } while (buf->rb_sc_ctxs[next_tail] != sc);
943 /* Paired with READ_ONCE */
944 smp_store_release(&buf->rb_sc_tail, next_tail);
946 if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
947 smp_mb__after_atomic();
948 xprt_write_space(&sc->sc_xprt->rx_xprt);
953 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
955 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
956 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
961 for (count = 0; count < ia->ri_max_segs; count++) {
962 struct rpcrdma_mr *mr;
965 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
969 rc = frwr_init_mr(ia, mr);
975 mr->mr_xprt = r_xprt;
977 list_add(&mr->mr_list, &free);
978 list_add(&mr->mr_all, &all);
981 spin_lock(&buf->rb_mrlock);
982 list_splice(&free, &buf->rb_mrs);
983 list_splice(&all, &buf->rb_all);
984 r_xprt->rx_stats.mrs_allocated += count;
985 spin_unlock(&buf->rb_mrlock);
986 trace_xprtrdma_createmrs(r_xprt, count);
988 xprt_write_space(&r_xprt->rx_xprt);
992 rpcrdma_mr_refresh_worker(struct work_struct *work)
994 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
995 rb_refresh_worker.work);
996 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
999 rpcrdma_mrs_create(r_xprt);
1002 struct rpcrdma_req *
1003 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1005 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1006 struct rpcrdma_regbuf *rb;
1007 struct rpcrdma_req *req;
1009 req = kzalloc(sizeof(*req), GFP_KERNEL);
1011 return ERR_PTR(-ENOMEM);
1013 rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1014 DMA_TO_DEVICE, GFP_KERNEL);
1017 return ERR_PTR(-ENOMEM);
1019 req->rl_rdmabuf = rb;
1020 xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1021 req->rl_buffer = buffer;
1022 INIT_LIST_HEAD(&req->rl_registered);
1024 spin_lock(&buffer->rb_lock);
1025 list_add(&req->rl_all, &buffer->rb_allreqs);
1026 spin_unlock(&buffer->rb_lock);
1031 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp)
1033 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1034 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1035 struct rpcrdma_rep *rep;
1039 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1043 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1044 DMA_FROM_DEVICE, GFP_KERNEL);
1045 if (IS_ERR(rep->rr_rdmabuf)) {
1046 rc = PTR_ERR(rep->rr_rdmabuf);
1049 xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1050 rdmab_length(rep->rr_rdmabuf));
1052 rep->rr_cqe.done = rpcrdma_wc_receive;
1053 rep->rr_rxprt = r_xprt;
1054 INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
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;
1061 spin_lock(&buf->rb_lock);
1062 list_add(&rep->rr_list, &buf->rb_recv_bufs);
1063 spin_unlock(&buf->rb_lock);
1073 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1075 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1079 buf->rb_max_requests = r_xprt->rx_data.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);
1092 for (i = 0; i < buf->rb_max_requests; i++) {
1093 struct rpcrdma_req *req;
1095 req = rpcrdma_create_req(r_xprt);
1097 dprintk("RPC: %s: request buffer %d alloc"
1098 " failed\n", __func__, i);
1102 list_add(&req->rl_list, &buf->rb_send_bufs);
1105 buf->rb_credits = 1;
1106 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1108 rc = rpcrdma_sendctxs_create(r_xprt);
1112 buf->rb_completion_wq = alloc_workqueue("rpcrdma-%s",
1113 WQ_MEM_RECLAIM | WQ_HIGHPRI,
1115 r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR]);
1116 if (!buf->rb_completion_wq)
1121 rpcrdma_buffer_destroy(buf);
1126 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1128 rpcrdma_free_regbuf(rep->rr_rdmabuf);
1133 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1134 * @req: unused object to be destroyed
1136 * This function assumes that the caller prevents concurrent device
1137 * unload and transport tear-down.
1140 rpcrdma_req_destroy(struct rpcrdma_req *req)
1142 list_del(&req->rl_all);
1144 rpcrdma_free_regbuf(req->rl_recvbuf);
1145 rpcrdma_free_regbuf(req->rl_sendbuf);
1146 rpcrdma_free_regbuf(req->rl_rdmabuf);
1151 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1153 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1155 struct rpcrdma_mr *mr;
1159 spin_lock(&buf->rb_mrlock);
1160 while (!list_empty(&buf->rb_all)) {
1161 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1162 list_del(&mr->mr_all);
1164 spin_unlock(&buf->rb_mrlock);
1166 /* Ensure MW is not on any rl_registered list */
1167 if (!list_empty(&mr->mr_list))
1168 list_del(&mr->mr_list);
1170 frwr_release_mr(mr);
1172 spin_lock(&buf->rb_mrlock);
1174 spin_unlock(&buf->rb_mrlock);
1175 r_xprt->rx_stats.mrs_allocated = 0;
1177 dprintk("RPC: %s: released %u MRs\n", __func__, count);
1181 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1183 cancel_delayed_work_sync(&buf->rb_refresh_worker);
1185 if (buf->rb_completion_wq) {
1186 destroy_workqueue(buf->rb_completion_wq);
1187 buf->rb_completion_wq = NULL;
1190 rpcrdma_sendctxs_destroy(buf);
1192 while (!list_empty(&buf->rb_recv_bufs)) {
1193 struct rpcrdma_rep *rep;
1195 rep = list_first_entry(&buf->rb_recv_bufs,
1196 struct rpcrdma_rep, rr_list);
1197 list_del(&rep->rr_list);
1198 rpcrdma_destroy_rep(rep);
1201 while (!list_empty(&buf->rb_send_bufs)) {
1202 struct rpcrdma_req *req;
1204 req = list_first_entry(&buf->rb_send_bufs,
1205 struct rpcrdma_req, rl_list);
1206 list_del(&req->rl_list);
1207 rpcrdma_req_destroy(req);
1210 rpcrdma_mrs_destroy(buf);
1214 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1215 * @r_xprt: controlling transport
1217 * Returns an initialized rpcrdma_mr or NULL if no free
1218 * rpcrdma_mr objects are available.
1221 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1223 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1224 struct rpcrdma_mr *mr = NULL;
1226 spin_lock(&buf->rb_mrlock);
1227 if (!list_empty(&buf->rb_mrs))
1228 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1229 spin_unlock(&buf->rb_mrlock);
1236 trace_xprtrdma_nomrs(r_xprt);
1237 if (r_xprt->rx_ep.rep_connected != -ENODEV)
1238 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1240 /* Allow the reply handler and refresh worker to run */
1247 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1249 spin_lock(&buf->rb_mrlock);
1250 rpcrdma_mr_push(mr, &buf->rb_mrs);
1251 spin_unlock(&buf->rb_mrlock);
1255 * rpcrdma_mr_put - Release an rpcrdma_mr object
1256 * @mr: object to release
1260 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1262 __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1266 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1267 * @mr: object to release
1271 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1273 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1275 if (mr->mr_dir != DMA_NONE) {
1276 trace_xprtrdma_mr_unmap(mr);
1277 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1278 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1279 mr->mr_dir = DMA_NONE;
1281 __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1285 * rpcrdma_buffer_get - Get a request buffer
1286 * @buffers: Buffer pool from which to obtain a buffer
1288 * Returns a fresh rpcrdma_req, or NULL if none are available.
1290 struct rpcrdma_req *
1291 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1293 struct rpcrdma_req *req;
1295 spin_lock(&buffers->rb_lock);
1296 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1297 struct rpcrdma_req, rl_list);
1299 list_del_init(&req->rl_list);
1300 spin_unlock(&buffers->rb_lock);
1305 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1306 * @req: object to return
1310 rpcrdma_buffer_put(struct rpcrdma_req *req)
1312 struct rpcrdma_buffer *buffers = req->rl_buffer;
1313 struct rpcrdma_rep *rep = req->rl_reply;
1315 req->rl_reply = NULL;
1317 spin_lock(&buffers->rb_lock);
1318 list_add(&req->rl_list, &buffers->rb_send_bufs);
1320 if (!rep->rr_temp) {
1321 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1325 spin_unlock(&buffers->rb_lock);
1327 rpcrdma_destroy_rep(rep);
1331 * Put reply buffers back into pool when not attached to
1332 * request. This happens in error conditions.
1335 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1337 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1339 if (!rep->rr_temp) {
1340 spin_lock(&buffers->rb_lock);
1341 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1342 spin_unlock(&buffers->rb_lock);
1344 rpcrdma_destroy_rep(rep);
1349 * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1350 * @size: size of buffer to be allocated, in bytes
1351 * @direction: direction of data movement
1354 * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1355 * can be persistently DMA-mapped for I/O.
1357 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1358 * receiving the payload of RDMA RECV operations. During Long Calls
1359 * or Replies they may be registered externally via frwr_map.
1361 struct rpcrdma_regbuf *
1362 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1365 struct rpcrdma_regbuf *rb;
1367 rb = kmalloc(sizeof(*rb) + size, flags);
1369 return ERR_PTR(-ENOMEM);
1371 rb->rg_device = NULL;
1372 rb->rg_direction = direction;
1373 rb->rg_iov.length = size;
1379 * __rpcrdma_map_regbuf - DMA-map a regbuf
1380 * @ia: controlling rpcrdma_ia
1381 * @rb: regbuf to be mapped
1384 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1386 struct ib_device *device = ia->ri_device;
1388 if (rb->rg_direction == DMA_NONE)
1391 rb->rg_iov.addr = ib_dma_map_single(device,
1392 (void *)rb->rg_base,
1395 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1396 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1400 rb->rg_device = device;
1401 rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1406 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1411 if (!rpcrdma_regbuf_is_mapped(rb))
1414 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1415 rdmab_length(rb), rb->rg_direction);
1416 rb->rg_device = NULL;
1420 * rpcrdma_free_regbuf - deregister and free registered buffer
1421 * @rb: regbuf to be deregistered and freed
1424 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1426 rpcrdma_dma_unmap_regbuf(rb);
1431 * Prepost any receive buffer, then post send.
1433 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1436 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1437 struct rpcrdma_ep *ep,
1438 struct rpcrdma_req *req)
1440 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1443 if (!ep->rep_send_count ||
1444 test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1445 send_wr->send_flags |= IB_SEND_SIGNALED;
1446 ep->rep_send_count = ep->rep_send_batch;
1448 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1449 --ep->rep_send_count;
1452 rc = frwr_send(ia, req);
1453 trace_xprtrdma_post_send(req, rc);
1460 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1462 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1463 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1464 struct ib_recv_wr *wr, *bad_wr;
1465 int needed, count, rc;
1469 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1470 if (ep->rep_receive_count > needed)
1472 needed -= ep->rep_receive_count;
1477 struct rpcrdma_regbuf *rb;
1478 struct rpcrdma_rep *rep;
1480 spin_lock(&buf->rb_lock);
1481 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1482 struct rpcrdma_rep, rr_list);
1484 list_del(&rep->rr_list);
1485 spin_unlock(&buf->rb_lock);
1487 if (rpcrdma_create_rep(r_xprt, temp))
1492 rb = rep->rr_rdmabuf;
1493 if (!rpcrdma_regbuf_is_mapped(rb)) {
1494 if (!__rpcrdma_dma_map_regbuf(&r_xprt->rx_ia, rb)) {
1495 rpcrdma_recv_buffer_put(rep);
1500 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1501 rep->rr_recv_wr.next = wr;
1502 wr = &rep->rr_recv_wr;
1509 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1510 (const struct ib_recv_wr **)&bad_wr);
1512 for (wr = bad_wr; wr; wr = wr->next) {
1513 struct rpcrdma_rep *rep;
1515 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1516 rpcrdma_recv_buffer_put(rep);
1520 ep->rep_receive_count += count;
1522 trace_xprtrdma_post_recvs(r_xprt, count, rc);