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>
56 #include <linux/log2.h>
58 #include <asm-generic/barrier.h>
59 #include <asm/bitops.h>
61 #include <rdma/ib_cm.h>
63 #include "xprt_rdma.h"
64 #include <trace/events/rpcrdma.h>
70 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
71 # define RPCDBG_FACILITY RPCDBG_TRANS
77 static void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
78 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
79 static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
80 static struct rpcrdma_regbuf *
81 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
83 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
84 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
85 static void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
87 /* Wait for outstanding transport work to finish. ib_drain_qp
88 * handles the drains in the wrong order for us, so open code
91 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
93 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
95 /* Flush Receives, then wait for deferred Reply work
98 ib_drain_rq(ia->ri_id->qp);
100 /* Deferred Reply processing might have scheduled
101 * local invalidations.
103 ib_drain_sq(ia->ri_id->qp);
107 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
108 * @event: details of the event
109 * @context: ep that owns QP where event occurred
111 * Called from the RDMA provider (device driver) possibly in an interrupt
115 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
117 struct rpcrdma_ep *ep = context;
118 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
121 trace_xprtrdma_qp_event(r_xprt, event);
125 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
126 * @cq: completion queue (ignored)
131 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
133 struct ib_cqe *cqe = wc->wr_cqe;
134 struct rpcrdma_sendctx *sc =
135 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
137 /* WARNING: Only wr_cqe and status are reliable at this point */
138 trace_xprtrdma_wc_send(sc, wc);
139 rpcrdma_sendctx_put_locked(sc);
143 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
144 * @cq: completion queue (ignored)
149 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
151 struct ib_cqe *cqe = wc->wr_cqe;
152 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
154 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
156 /* WARNING: Only wr_cqe and status are reliable at this point */
157 trace_xprtrdma_wc_receive(wc);
158 --r_xprt->rx_ep.rep_receive_count;
159 if (wc->status != IB_WC_SUCCESS)
162 /* status == SUCCESS means all fields in wc are trustworthy */
163 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
164 rep->rr_wc_flags = wc->wc_flags;
165 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
167 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
168 rdmab_addr(rep->rr_rdmabuf),
169 wc->byte_len, DMA_FROM_DEVICE);
171 rpcrdma_post_recvs(r_xprt, false);
172 rpcrdma_reply_handler(rep);
176 rpcrdma_recv_buffer_put(rep);
180 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
181 struct rdma_conn_param *param)
183 const struct rpcrdma_connect_private *pmsg = param->private_data;
184 unsigned int rsize, wsize;
186 /* Default settings for RPC-over-RDMA Version One */
187 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
188 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
189 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
192 pmsg->cp_magic == rpcrdma_cmp_magic &&
193 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
194 r_xprt->rx_ia.ri_implicit_roundup = true;
195 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
196 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
199 if (rsize < r_xprt->rx_ep.rep_inline_recv)
200 r_xprt->rx_ep.rep_inline_recv = rsize;
201 if (wsize < r_xprt->rx_ep.rep_inline_send)
202 r_xprt->rx_ep.rep_inline_send = wsize;
203 dprintk("RPC: %s: max send %u, max recv %u\n", __func__,
204 r_xprt->rx_ep.rep_inline_send,
205 r_xprt->rx_ep.rep_inline_recv);
206 rpcrdma_set_max_header_sizes(r_xprt);
210 * rpcrdma_cm_event_handler - Handle RDMA CM events
211 * @id: rdma_cm_id on which an event has occurred
212 * @event: details of the event
214 * Called with @id's mutex held. Returns 1 if caller should
215 * destroy @id, otherwise 0.
218 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
220 struct rpcrdma_xprt *r_xprt = id->context;
221 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
222 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
223 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
227 trace_xprtrdma_cm_event(r_xprt, event);
228 switch (event->event) {
229 case RDMA_CM_EVENT_ADDR_RESOLVED:
230 case RDMA_CM_EVENT_ROUTE_RESOLVED:
232 complete(&ia->ri_done);
234 case RDMA_CM_EVENT_ADDR_ERROR:
235 ia->ri_async_rc = -EPROTO;
236 complete(&ia->ri_done);
238 case RDMA_CM_EVENT_ROUTE_ERROR:
239 ia->ri_async_rc = -ENETUNREACH;
240 complete(&ia->ri_done);
242 case RDMA_CM_EVENT_DEVICE_REMOVAL:
243 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
244 pr_info("rpcrdma: removing device %s for %s:%s\n",
245 ia->ri_id->device->name,
246 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
248 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
249 ep->rep_connected = -ENODEV;
250 xprt_force_disconnect(xprt);
251 wait_for_completion(&ia->ri_remove_done);
254 /* Return 1 to ensure the core destroys the id. */
256 case RDMA_CM_EVENT_ESTABLISHED:
257 ++xprt->connect_cookie;
258 ep->rep_connected = 1;
259 rpcrdma_update_connect_private(r_xprt, &event->param.conn);
260 wake_up_all(&ep->rep_connect_wait);
262 case RDMA_CM_EVENT_CONNECT_ERROR:
263 ep->rep_connected = -ENOTCONN;
265 case RDMA_CM_EVENT_UNREACHABLE:
266 ep->rep_connected = -ENETUNREACH;
268 case RDMA_CM_EVENT_REJECTED:
269 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
270 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
271 rdma_reject_msg(id, event->status));
272 ep->rep_connected = -ECONNREFUSED;
273 if (event->status == IB_CM_REJ_STALE_CONN)
274 ep->rep_connected = -EAGAIN;
276 case RDMA_CM_EVENT_DISCONNECTED:
277 ep->rep_connected = -ECONNABORTED;
279 xprt_force_disconnect(xprt);
280 wake_up_all(&ep->rep_connect_wait);
286 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__,
287 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
288 ia->ri_id->device->name, rdma_event_msg(event->event));
292 static struct rdma_cm_id *
293 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
295 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
296 struct rdma_cm_id *id;
299 trace_xprtrdma_conn_start(xprt);
301 init_completion(&ia->ri_done);
302 init_completion(&ia->ri_remove_done);
304 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
305 xprt, RDMA_PS_TCP, IB_QPT_RC);
309 ia->ri_async_rc = -ETIMEDOUT;
310 rc = rdma_resolve_addr(id, NULL,
311 (struct sockaddr *)&xprt->rx_xprt.addr,
312 RDMA_RESOLVE_TIMEOUT);
315 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
317 trace_xprtrdma_conn_tout(xprt);
321 rc = ia->ri_async_rc;
325 ia->ri_async_rc = -ETIMEDOUT;
326 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
329 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
331 trace_xprtrdma_conn_tout(xprt);
334 rc = ia->ri_async_rc;
346 * Exported functions.
350 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
351 * @xprt: transport with IA to (re)initialize
353 * Returns 0 on success, negative errno if an appropriate
354 * Interface Adapter could not be found and opened.
357 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
359 struct rpcrdma_ia *ia = &xprt->rx_ia;
362 ia->ri_id = rpcrdma_create_id(xprt, ia);
363 if (IS_ERR(ia->ri_id)) {
364 rc = PTR_ERR(ia->ri_id);
368 ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0);
369 if (IS_ERR(ia->ri_pd)) {
370 rc = PTR_ERR(ia->ri_pd);
371 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
375 switch (xprt_rdma_memreg_strategy) {
377 if (frwr_is_supported(ia->ri_id->device))
381 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
382 ia->ri_id->device->name, xprt_rdma_memreg_strategy);
390 rpcrdma_ia_close(ia);
395 * rpcrdma_ia_remove - Handle device driver unload
396 * @ia: interface adapter being removed
398 * Divest transport H/W resources associated with this adapter,
399 * but allow it to be restored later.
402 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
404 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
406 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
407 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
408 struct rpcrdma_req *req;
409 struct rpcrdma_rep *rep;
411 cancel_delayed_work_sync(&buf->rb_refresh_worker);
413 /* This is similar to rpcrdma_ep_destroy, but:
414 * - Don't cancel the connect worker.
415 * - Don't call rpcrdma_ep_disconnect, which waits
416 * for another conn upcall, which will deadlock.
417 * - rdma_disconnect is unneeded, the underlying
418 * connection is already gone.
421 rpcrdma_xprt_drain(r_xprt);
422 rdma_destroy_qp(ia->ri_id);
423 ia->ri_id->qp = NULL;
425 ib_free_cq(ep->rep_attr.recv_cq);
426 ep->rep_attr.recv_cq = NULL;
427 ib_free_cq(ep->rep_attr.send_cq);
428 ep->rep_attr.send_cq = NULL;
430 /* The ULP is responsible for ensuring all DMA
431 * mappings and MRs are gone.
433 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
434 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
435 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
436 rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf);
437 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
438 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
440 rpcrdma_mrs_destroy(buf);
441 ib_dealloc_pd(ia->ri_pd);
444 /* Allow waiters to continue */
445 complete(&ia->ri_remove_done);
447 trace_xprtrdma_remove(r_xprt);
451 * rpcrdma_ia_close - Clean up/close an IA.
452 * @ia: interface adapter to close
456 rpcrdma_ia_close(struct rpcrdma_ia *ia)
458 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
460 rdma_destroy_qp(ia->ri_id);
461 rdma_destroy_id(ia->ri_id);
465 /* If the pd is still busy, xprtrdma missed freeing a resource */
466 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
467 ib_dealloc_pd(ia->ri_pd);
472 * rpcrdma_ep_create - Create unconnected endpoint
473 * @r_xprt: transport to instantiate
475 * Returns zero on success, or a negative errno.
477 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
479 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
480 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
481 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
482 struct ib_cq *sendcq, *recvcq;
483 unsigned int max_sge;
486 ep->rep_max_requests = xprt_rdma_slot_table_entries;
487 ep->rep_inline_send = xprt_rdma_max_inline_write;
488 ep->rep_inline_recv = xprt_rdma_max_inline_read;
490 max_sge = min_t(unsigned int, ia->ri_id->device->attrs.max_send_sge,
491 RPCRDMA_MAX_SEND_SGES);
492 if (max_sge < RPCRDMA_MIN_SEND_SGES) {
493 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
496 ia->ri_max_send_sges = max_sge;
498 rc = frwr_open(ia, ep);
502 ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
503 ep->rep_attr.qp_context = ep;
504 ep->rep_attr.srq = NULL;
505 ep->rep_attr.cap.max_send_sge = max_sge;
506 ep->rep_attr.cap.max_recv_sge = 1;
507 ep->rep_attr.cap.max_inline_data = 0;
508 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
509 ep->rep_attr.qp_type = IB_QPT_RC;
510 ep->rep_attr.port_num = ~0;
512 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
513 "iovs: send %d recv %d\n",
515 ep->rep_attr.cap.max_send_wr,
516 ep->rep_attr.cap.max_recv_wr,
517 ep->rep_attr.cap.max_send_sge,
518 ep->rep_attr.cap.max_recv_sge);
520 ep->rep_send_batch = ep->rep_max_requests >> 3;
521 ep->rep_send_count = ep->rep_send_batch;
522 init_waitqueue_head(&ep->rep_connect_wait);
523 ep->rep_receive_count = 0;
525 sendcq = ib_alloc_cq(ia->ri_id->device, NULL,
526 ep->rep_attr.cap.max_send_wr + 1,
527 ia->ri_id->device->num_comp_vectors > 1 ? 1 : 0,
529 if (IS_ERR(sendcq)) {
530 rc = PTR_ERR(sendcq);
534 recvcq = ib_alloc_cq(ia->ri_id->device, NULL,
535 ep->rep_attr.cap.max_recv_wr + 1,
536 0, IB_POLL_WORKQUEUE);
537 if (IS_ERR(recvcq)) {
538 rc = PTR_ERR(recvcq);
542 ep->rep_attr.send_cq = sendcq;
543 ep->rep_attr.recv_cq = recvcq;
545 /* Initialize cma parameters */
546 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
548 /* Prepare RDMA-CM private message */
549 pmsg->cp_magic = rpcrdma_cmp_magic;
550 pmsg->cp_version = RPCRDMA_CMP_VERSION;
551 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
552 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send);
553 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv);
554 ep->rep_remote_cma.private_data = pmsg;
555 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
557 /* Client offers RDMA Read but does not initiate */
558 ep->rep_remote_cma.initiator_depth = 0;
559 ep->rep_remote_cma.responder_resources =
560 min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom);
562 /* Limit transport retries so client can detect server
563 * GID changes quickly. RPC layer handles re-establishing
564 * transport connection and retransmission.
566 ep->rep_remote_cma.retry_count = 6;
568 /* RPC-over-RDMA handles its own flow control. In addition,
569 * make all RNR NAKs visible so we know that RPC-over-RDMA
570 * flow control is working correctly (no NAKs should be seen).
572 ep->rep_remote_cma.flow_control = 0;
573 ep->rep_remote_cma.rnr_retry_count = 0;
584 * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
585 * @r_xprt: transport instance to shut down
588 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
590 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
591 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
593 if (ia->ri_id && ia->ri_id->qp) {
594 rpcrdma_ep_disconnect(ep, ia);
595 rdma_destroy_qp(ia->ri_id);
596 ia->ri_id->qp = NULL;
599 if (ep->rep_attr.recv_cq)
600 ib_free_cq(ep->rep_attr.recv_cq);
601 if (ep->rep_attr.send_cq)
602 ib_free_cq(ep->rep_attr.send_cq);
605 /* Re-establish a connection after a device removal event.
606 * Unlike a normal reconnection, a fresh PD and a new set
607 * of MRs and buffers is needed.
610 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
611 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
615 trace_xprtrdma_reinsert(r_xprt);
618 if (rpcrdma_ia_open(r_xprt))
622 err = rpcrdma_ep_create(r_xprt);
624 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
629 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
631 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
635 rpcrdma_mrs_create(r_xprt);
639 rpcrdma_ep_destroy(r_xprt);
641 rpcrdma_ia_close(ia);
647 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
648 struct rpcrdma_ia *ia)
650 struct rdma_cm_id *id, *old;
653 trace_xprtrdma_reconnect(r_xprt);
655 rpcrdma_ep_disconnect(ep, ia);
658 id = rpcrdma_create_id(r_xprt, ia);
662 /* As long as the new ID points to the same device as the
663 * old ID, we can reuse the transport's existing PD and all
664 * previously allocated MRs. Also, the same device means
665 * the transport's previous DMA mappings are still valid.
667 * This is a sanity check only. There should be no way these
668 * point to two different devices here.
672 if (ia->ri_id->device != id->device) {
673 pr_err("rpcrdma: can't reconnect on different device!\n");
677 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
681 /* Atomically replace the transport's ID and QP. */
685 rdma_destroy_qp(old);
688 rdma_destroy_id(old);
694 * Connect unconnected endpoint.
697 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
699 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
701 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
705 switch (ep->rep_connected) {
707 dprintk("RPC: %s: connecting...\n", __func__);
708 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
715 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
720 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
725 ep->rep_connected = 0;
726 xprt_clear_connected(xprt);
728 rpcrdma_post_recvs(r_xprt, true);
730 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
734 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
735 if (ep->rep_connected <= 0) {
736 if (ep->rep_connected == -EAGAIN)
738 rc = ep->rep_connected;
742 dprintk("RPC: %s: connected\n", __func__);
746 ep->rep_connected = rc;
753 * rpcrdma_ep_disconnect - Disconnect underlying transport
754 * @ep: endpoint to disconnect
755 * @ia: associated interface adapter
757 * This is separate from destroy to facilitate the ability
758 * to reconnect without recreating the endpoint.
760 * This call is not reentrant, and must not be made in parallel
761 * on the same endpoint.
764 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
766 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
770 /* returns without wait if ID is not connected */
771 rc = rdma_disconnect(ia->ri_id);
773 wait_event_interruptible(ep->rep_connect_wait,
774 ep->rep_connected != 1);
776 ep->rep_connected = rc;
777 trace_xprtrdma_disconnect(r_xprt, rc);
779 rpcrdma_xprt_drain(r_xprt);
782 /* Fixed-size circular FIFO queue. This implementation is wait-free and
785 * Consumer is the code path that posts Sends. This path dequeues a
786 * sendctx for use by a Send operation. Multiple consumer threads
787 * are serialized by the RPC transport lock, which allows only one
788 * ->send_request call at a time.
790 * Producer is the code path that handles Send completions. This path
791 * enqueues a sendctx that has been completed. Multiple producer
792 * threads are serialized by the ib_poll_cq() function.
795 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
796 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
799 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
803 for (i = 0; i <= buf->rb_sc_last; i++)
804 kfree(buf->rb_sc_ctxs[i]);
805 kfree(buf->rb_sc_ctxs);
808 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
810 struct rpcrdma_sendctx *sc;
812 sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges),
817 sc->sc_wr.wr_cqe = &sc->sc_cqe;
818 sc->sc_wr.sg_list = sc->sc_sges;
819 sc->sc_wr.opcode = IB_WR_SEND;
820 sc->sc_cqe.done = rpcrdma_wc_send;
824 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
826 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
827 struct rpcrdma_sendctx *sc;
830 /* Maximum number of concurrent outstanding Send WRs. Capping
831 * the circular queue size stops Send Queue overflow by causing
832 * the ->send_request call to fail temporarily before too many
835 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
836 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
837 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
838 if (!buf->rb_sc_ctxs)
841 buf->rb_sc_last = i - 1;
842 for (i = 0; i <= buf->rb_sc_last; i++) {
843 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
847 sc->sc_xprt = r_xprt;
848 buf->rb_sc_ctxs[i] = sc;
854 /* The sendctx queue is not guaranteed to have a size that is a
855 * power of two, thus the helpers in circ_buf.h cannot be used.
856 * The other option is to use modulus (%), which can be expensive.
858 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
861 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
865 * rpcrdma_sendctx_get_locked - Acquire a send context
866 * @r_xprt: controlling transport instance
868 * Returns pointer to a free send completion context; or NULL if
869 * the queue is empty.
871 * Usage: Called to acquire an SGE array before preparing a Send WR.
873 * The caller serializes calls to this function (per transport), and
874 * provides an effective memory barrier that flushes the new value
877 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
879 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
880 struct rpcrdma_sendctx *sc;
881 unsigned long next_head;
883 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
885 if (next_head == READ_ONCE(buf->rb_sc_tail))
888 /* ORDER: item must be accessed _before_ head is updated */
889 sc = buf->rb_sc_ctxs[next_head];
891 /* Releasing the lock in the caller acts as a memory
892 * barrier that flushes rb_sc_head.
894 buf->rb_sc_head = next_head;
899 /* The queue is "empty" if there have not been enough Send
900 * completions recently. This is a sign the Send Queue is
901 * backing up. Cause the caller to pause and try again.
903 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
904 r_xprt->rx_stats.empty_sendctx_q++;
909 * rpcrdma_sendctx_put_locked - Release a send context
910 * @sc: send context to release
912 * Usage: Called from Send completion to return a sendctxt
915 * The caller serializes calls to this function (per transport).
918 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
920 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
921 unsigned long next_tail;
923 /* Unmap SGEs of previously completed but unsignaled
924 * Sends by walking up the queue until @sc is found.
926 next_tail = buf->rb_sc_tail;
928 next_tail = rpcrdma_sendctx_next(buf, next_tail);
930 /* ORDER: item must be accessed _before_ tail is updated */
931 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
933 } while (buf->rb_sc_ctxs[next_tail] != sc);
935 /* Paired with READ_ONCE */
936 smp_store_release(&buf->rb_sc_tail, next_tail);
938 xprt_write_space(&sc->sc_xprt->rx_xprt);
942 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
944 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
945 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 spin_lock(&buf->rb_mrlock);
965 list_add(&mr->mr_list, &buf->rb_mrs);
966 list_add(&mr->mr_all, &buf->rb_all_mrs);
967 spin_unlock(&buf->rb_mrlock);
970 r_xprt->rx_stats.mrs_allocated += count;
971 trace_xprtrdma_createmrs(r_xprt, count);
975 rpcrdma_mr_refresh_worker(struct work_struct *work)
977 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
978 rb_refresh_worker.work);
979 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
982 rpcrdma_mrs_create(r_xprt);
983 xprt_write_space(&r_xprt->rx_xprt);
987 * rpcrdma_req_create - Allocate an rpcrdma_req object
988 * @r_xprt: controlling r_xprt
989 * @size: initial size, in bytes, of send and receive buffers
990 * @flags: GFP flags passed to memory allocators
992 * Returns an allocated and fully initialized rpcrdma_req or NULL.
994 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
997 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
998 struct rpcrdma_regbuf *rb;
999 struct rpcrdma_req *req;
1002 req = kzalloc(sizeof(*req), flags);
1006 /* Compute maximum header buffer size in bytes */
1007 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
1008 r_xprt->rx_ia.ri_max_segs * rpcrdma_readchunk_maxsz;
1009 maxhdrsize *= sizeof(__be32);
1010 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
1011 DMA_TO_DEVICE, flags);
1014 req->rl_rdmabuf = rb;
1015 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1017 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1018 if (!req->rl_sendbuf)
1021 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1022 if (!req->rl_recvbuf)
1025 INIT_LIST_HEAD(&req->rl_registered);
1026 spin_lock(&buffer->rb_lock);
1027 list_add(&req->rl_all, &buffer->rb_allreqs);
1028 spin_unlock(&buffer->rb_lock);
1032 kfree(req->rl_sendbuf);
1034 kfree(req->rl_rdmabuf);
1041 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1044 struct rpcrdma_rep *rep;
1046 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1050 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv,
1051 DMA_FROM_DEVICE, GFP_KERNEL);
1052 if (!rep->rr_rdmabuf)
1055 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
1056 rdmab_length(rep->rr_rdmabuf));
1057 rep->rr_cqe.done = rpcrdma_wc_receive;
1058 rep->rr_rxprt = r_xprt;
1059 rep->rr_recv_wr.next = NULL;
1060 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1061 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1062 rep->rr_recv_wr.num_sge = 1;
1063 rep->rr_temp = temp;
1073 * rpcrdma_buffer_create - Create initial set of req/rep objects
1074 * @r_xprt: transport instance to (re)initialize
1076 * Returns zero on success, otherwise a negative errno.
1078 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1080 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1083 buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
1084 buf->rb_bc_srv_max_requests = 0;
1085 spin_lock_init(&buf->rb_mrlock);
1086 spin_lock_init(&buf->rb_lock);
1087 INIT_LIST_HEAD(&buf->rb_mrs);
1088 INIT_LIST_HEAD(&buf->rb_all_mrs);
1089 INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1090 rpcrdma_mr_refresh_worker);
1092 rpcrdma_mrs_create(r_xprt);
1094 INIT_LIST_HEAD(&buf->rb_send_bufs);
1095 INIT_LIST_HEAD(&buf->rb_allreqs);
1098 for (i = 0; i < buf->rb_max_requests; i++) {
1099 struct rpcrdma_req *req;
1101 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE,
1105 list_add(&req->rl_list, &buf->rb_send_bufs);
1108 buf->rb_credits = 1;
1109 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1111 rc = rpcrdma_sendctxs_create(r_xprt);
1117 rpcrdma_buffer_destroy(buf);
1121 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1123 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1128 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1129 * @req: unused object to be destroyed
1131 * This function assumes that the caller prevents concurrent device
1132 * unload and transport tear-down.
1135 rpcrdma_req_destroy(struct rpcrdma_req *req)
1137 list_del(&req->rl_all);
1139 rpcrdma_regbuf_free(req->rl_recvbuf);
1140 rpcrdma_regbuf_free(req->rl_sendbuf);
1141 rpcrdma_regbuf_free(req->rl_rdmabuf);
1146 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1148 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1150 struct rpcrdma_mr *mr;
1154 spin_lock(&buf->rb_mrlock);
1155 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1158 list_del(&mr->mr_all);
1159 spin_unlock(&buf->rb_mrlock);
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;
1170 * rpcrdma_buffer_destroy - Release all hw resources
1171 * @buf: root control block for resources
1173 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1174 * - No more Send or Receive completions can occur
1175 * - All MRs, reps, and reqs are returned to their free lists
1178 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1180 cancel_delayed_work_sync(&buf->rb_refresh_worker);
1182 rpcrdma_sendctxs_destroy(buf);
1184 while (!list_empty(&buf->rb_recv_bufs)) {
1185 struct rpcrdma_rep *rep;
1187 rep = list_first_entry(&buf->rb_recv_bufs,
1188 struct rpcrdma_rep, rr_list);
1189 list_del(&rep->rr_list);
1190 rpcrdma_rep_destroy(rep);
1193 while (!list_empty(&buf->rb_send_bufs)) {
1194 struct rpcrdma_req *req;
1196 req = list_first_entry(&buf->rb_send_bufs,
1197 struct rpcrdma_req, rl_list);
1198 list_del(&req->rl_list);
1199 rpcrdma_req_destroy(req);
1202 rpcrdma_mrs_destroy(buf);
1206 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1207 * @r_xprt: controlling transport
1209 * Returns an initialized rpcrdma_mr or NULL if no free
1210 * rpcrdma_mr objects are available.
1213 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1215 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1216 struct rpcrdma_mr *mr;
1218 spin_lock(&buf->rb_mrlock);
1219 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1220 spin_unlock(&buf->rb_mrlock);
1226 trace_xprtrdma_nomrs(r_xprt);
1227 if (r_xprt->rx_ep.rep_connected != -ENODEV)
1228 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1230 /* Allow the reply handler and refresh worker to run */
1237 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1239 spin_lock(&buf->rb_mrlock);
1240 rpcrdma_mr_push(mr, &buf->rb_mrs);
1241 spin_unlock(&buf->rb_mrlock);
1245 * rpcrdma_mr_put - Release an rpcrdma_mr object
1246 * @mr: object to release
1250 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1252 __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1256 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1257 * @mr: object to release
1261 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1263 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1265 if (mr->mr_dir != DMA_NONE) {
1266 trace_xprtrdma_mr_unmap(mr);
1267 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1268 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1269 mr->mr_dir = DMA_NONE;
1271 __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1275 * rpcrdma_buffer_get - Get a request buffer
1276 * @buffers: Buffer pool from which to obtain a buffer
1278 * Returns a fresh rpcrdma_req, or NULL if none are available.
1280 struct rpcrdma_req *
1281 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1283 struct rpcrdma_req *req;
1285 spin_lock(&buffers->rb_lock);
1286 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1287 struct rpcrdma_req, rl_list);
1289 list_del_init(&req->rl_list);
1290 spin_unlock(&buffers->rb_lock);
1295 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1296 * @buffers: buffer pool
1297 * @req: object to return
1300 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1302 struct rpcrdma_rep *rep = req->rl_reply;
1304 req->rl_reply = NULL;
1306 spin_lock(&buffers->rb_lock);
1307 list_add(&req->rl_list, &buffers->rb_send_bufs);
1309 if (!rep->rr_temp) {
1310 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1314 spin_unlock(&buffers->rb_lock);
1316 rpcrdma_rep_destroy(rep);
1320 * Put reply buffers back into pool when not attached to
1321 * request. This happens in error conditions.
1324 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1326 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1328 if (!rep->rr_temp) {
1329 spin_lock(&buffers->rb_lock);
1330 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1331 spin_unlock(&buffers->rb_lock);
1333 rpcrdma_rep_destroy(rep);
1337 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1339 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1340 * receiving the payload of RDMA RECV operations. During Long Calls
1341 * or Replies they may be registered externally via frwr_map.
1343 static struct rpcrdma_regbuf *
1344 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1347 struct rpcrdma_regbuf *rb;
1349 rb = kmalloc(sizeof(*rb), flags);
1352 rb->rg_data = kmalloc(size, flags);
1358 rb->rg_device = NULL;
1359 rb->rg_direction = direction;
1360 rb->rg_iov.length = size;
1365 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1366 * @rb: regbuf to reallocate
1367 * @size: size of buffer to be allocated, in bytes
1370 * Returns true if reallocation was successful. If false is
1371 * returned, @rb is left untouched.
1373 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1377 buf = kmalloc(size, flags);
1381 rpcrdma_regbuf_dma_unmap(rb);
1385 rb->rg_iov.length = size;
1390 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1391 * @r_xprt: controlling transport instance
1392 * @rb: regbuf to be mapped
1394 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1396 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1397 struct rpcrdma_regbuf *rb)
1399 struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1401 if (rb->rg_direction == DMA_NONE)
1404 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1405 rdmab_length(rb), rb->rg_direction);
1406 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1407 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1411 rb->rg_device = device;
1412 rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1416 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1421 if (!rpcrdma_regbuf_is_mapped(rb))
1424 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1426 rb->rg_device = NULL;
1429 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1431 rpcrdma_regbuf_dma_unmap(rb);
1438 * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1439 * @ia: transport's device information
1440 * @ep: transport's RDMA endpoint information
1441 * @req: rpcrdma_req containing the Send WR to post
1443 * Returns 0 if the post was successful, otherwise -ENOTCONN
1447 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1448 struct rpcrdma_ep *ep,
1449 struct rpcrdma_req *req)
1451 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1454 if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1455 send_wr->send_flags |= IB_SEND_SIGNALED;
1456 ep->rep_send_count = ep->rep_send_batch;
1458 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1459 --ep->rep_send_count;
1462 rc = frwr_send(ia, req);
1463 trace_xprtrdma_post_send(req, rc);
1470 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1472 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1473 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1474 struct ib_recv_wr *i, *wr, *bad_wr;
1475 struct rpcrdma_rep *rep;
1476 int needed, count, rc;
1481 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1482 if (ep->rep_receive_count > needed)
1484 needed -= ep->rep_receive_count;
1486 needed += RPCRDMA_MAX_RECV_BATCH;
1488 /* fast path: all needed reps can be found on the free list */
1490 spin_lock(&buf->rb_lock);
1492 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1493 struct rpcrdma_rep, rr_list);
1497 list_del(&rep->rr_list);
1498 rep->rr_recv_wr.next = wr;
1499 wr = &rep->rr_recv_wr;
1502 spin_unlock(&buf->rb_lock);
1505 rep = rpcrdma_rep_create(r_xprt, temp);
1509 rep->rr_recv_wr.next = wr;
1510 wr = &rep->rr_recv_wr;
1516 for (i = wr; i; i = i->next) {
1517 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1519 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1522 trace_xprtrdma_post_recv(rep);
1526 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1527 (const struct ib_recv_wr **)&bad_wr);
1529 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1531 for (wr = bad_wr; wr;) {
1532 struct rpcrdma_rep *rep;
1534 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1536 rpcrdma_recv_buffer_put(rep);
1540 ep->rep_receive_count += count;
1545 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1547 rpcrdma_recv_buffer_put(rep);