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 void rpcrdma_mr_free(struct rpcrdma_mr *mr);
81 static struct rpcrdma_regbuf *
82 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
84 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
85 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
86 static void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
88 /* Wait for outstanding transport work to finish. ib_drain_qp
89 * handles the drains in the wrong order for us, so open code
92 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
94 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
96 /* Flush Receives, then wait for deferred Reply work
99 ib_drain_rq(ia->ri_id->qp);
101 /* Deferred Reply processing might have scheduled
102 * local invalidations.
104 ib_drain_sq(ia->ri_id->qp);
108 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
109 * @event: details of the event
110 * @context: ep that owns QP where event occurred
112 * Called from the RDMA provider (device driver) possibly in an interrupt
116 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
118 struct rpcrdma_ep *ep = context;
119 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
122 trace_xprtrdma_qp_event(r_xprt, event);
126 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
127 * @cq: completion queue (ignored)
132 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
134 struct ib_cqe *cqe = wc->wr_cqe;
135 struct rpcrdma_sendctx *sc =
136 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
138 /* WARNING: Only wr_cqe and status are reliable at this point */
139 trace_xprtrdma_wc_send(sc, wc);
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 rpcrdma_recv_buffer_put(rep);
181 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
182 struct rdma_conn_param *param)
184 const struct rpcrdma_connect_private *pmsg = param->private_data;
185 unsigned int rsize, wsize;
187 /* Default settings for RPC-over-RDMA Version One */
188 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
189 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
190 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
193 pmsg->cp_magic == rpcrdma_cmp_magic &&
194 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
195 r_xprt->rx_ia.ri_implicit_roundup = true;
196 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
197 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
200 if (rsize < r_xprt->rx_ep.rep_inline_recv)
201 r_xprt->rx_ep.rep_inline_recv = rsize;
202 if (wsize < r_xprt->rx_ep.rep_inline_send)
203 r_xprt->rx_ep.rep_inline_send = wsize;
204 dprintk("RPC: %s: max send %u, max recv %u\n", __func__,
205 r_xprt->rx_ep.rep_inline_send,
206 r_xprt->rx_ep.rep_inline_recv);
207 rpcrdma_set_max_header_sizes(r_xprt);
211 * rpcrdma_cm_event_handler - Handle RDMA CM events
212 * @id: rdma_cm_id on which an event has occurred
213 * @event: details of the event
215 * Called with @id's mutex held. Returns 1 if caller should
216 * destroy @id, otherwise 0.
219 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
221 struct rpcrdma_xprt *r_xprt = id->context;
222 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
223 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
224 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
228 trace_xprtrdma_cm_event(r_xprt, event);
229 switch (event->event) {
230 case RDMA_CM_EVENT_ADDR_RESOLVED:
231 case RDMA_CM_EVENT_ROUTE_RESOLVED:
233 complete(&ia->ri_done);
235 case RDMA_CM_EVENT_ADDR_ERROR:
236 ia->ri_async_rc = -EPROTO;
237 complete(&ia->ri_done);
239 case RDMA_CM_EVENT_ROUTE_ERROR:
240 ia->ri_async_rc = -ENETUNREACH;
241 complete(&ia->ri_done);
243 case RDMA_CM_EVENT_DEVICE_REMOVAL:
244 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
245 pr_info("rpcrdma: removing device %s for %s:%s\n",
246 ia->ri_id->device->name,
247 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
249 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
250 ep->rep_connected = -ENODEV;
251 xprt_force_disconnect(xprt);
252 wait_for_completion(&ia->ri_remove_done);
255 /* Return 1 to ensure the core destroys the id. */
257 case RDMA_CM_EVENT_ESTABLISHED:
258 ++xprt->connect_cookie;
259 ep->rep_connected = 1;
260 rpcrdma_update_connect_private(r_xprt, &event->param.conn);
261 wake_up_all(&ep->rep_connect_wait);
263 case RDMA_CM_EVENT_CONNECT_ERROR:
264 ep->rep_connected = -ENOTCONN;
266 case RDMA_CM_EVENT_UNREACHABLE:
267 ep->rep_connected = -ENETUNREACH;
269 case RDMA_CM_EVENT_REJECTED:
270 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
271 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
272 rdma_reject_msg(id, event->status));
273 ep->rep_connected = -ECONNREFUSED;
274 if (event->status == IB_CM_REJ_STALE_CONN)
275 ep->rep_connected = -EAGAIN;
277 case RDMA_CM_EVENT_DISCONNECTED:
278 ep->rep_connected = -ECONNABORTED;
280 xprt_force_disconnect(xprt);
281 wake_up_all(&ep->rep_connect_wait);
287 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__,
288 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
289 ia->ri_id->device->name, rdma_event_msg(event->event));
293 static struct rdma_cm_id *
294 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
296 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
297 struct rdma_cm_id *id;
300 trace_xprtrdma_conn_start(xprt);
302 init_completion(&ia->ri_done);
303 init_completion(&ia->ri_remove_done);
305 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
306 xprt, RDMA_PS_TCP, IB_QPT_RC);
310 ia->ri_async_rc = -ETIMEDOUT;
311 rc = rdma_resolve_addr(id, NULL,
312 (struct sockaddr *)&xprt->rx_xprt.addr,
313 RDMA_RESOLVE_TIMEOUT);
316 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
318 trace_xprtrdma_conn_tout(xprt);
322 rc = ia->ri_async_rc;
326 ia->ri_async_rc = -ETIMEDOUT;
327 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
330 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
332 trace_xprtrdma_conn_tout(xprt);
335 rc = ia->ri_async_rc;
347 * Exported functions.
351 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
352 * @xprt: transport with IA to (re)initialize
354 * Returns 0 on success, negative errno if an appropriate
355 * Interface Adapter could not be found and opened.
358 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
360 struct rpcrdma_ia *ia = &xprt->rx_ia;
363 ia->ri_id = rpcrdma_create_id(xprt, ia);
364 if (IS_ERR(ia->ri_id)) {
365 rc = PTR_ERR(ia->ri_id);
369 ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0);
370 if (IS_ERR(ia->ri_pd)) {
371 rc = PTR_ERR(ia->ri_pd);
372 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
376 switch (xprt_rdma_memreg_strategy) {
378 if (frwr_is_supported(ia->ri_id->device))
382 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
383 ia->ri_id->device->name, xprt_rdma_memreg_strategy);
391 rpcrdma_ia_close(ia);
396 * rpcrdma_ia_remove - Handle device driver unload
397 * @ia: interface adapter being removed
399 * Divest transport H/W resources associated with this adapter,
400 * but allow it to be restored later.
403 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
405 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
407 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
408 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
409 struct rpcrdma_req *req;
410 struct rpcrdma_rep *rep;
412 cancel_work_sync(&buf->rb_refresh_worker);
414 /* This is similar to rpcrdma_ep_destroy, but:
415 * - Don't cancel the connect worker.
416 * - Don't call rpcrdma_ep_disconnect, which waits
417 * for another conn upcall, which will deadlock.
418 * - rdma_disconnect is unneeded, the underlying
419 * connection is already gone.
422 rpcrdma_xprt_drain(r_xprt);
423 rdma_destroy_qp(ia->ri_id);
424 ia->ri_id->qp = NULL;
426 ib_free_cq(ep->rep_attr.recv_cq);
427 ep->rep_attr.recv_cq = NULL;
428 ib_free_cq(ep->rep_attr.send_cq);
429 ep->rep_attr.send_cq = NULL;
431 /* The ULP is responsible for ensuring all DMA
432 * mappings and MRs are gone.
434 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
435 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
436 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
437 rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf);
438 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
439 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
441 rpcrdma_mrs_destroy(buf);
442 ib_dealloc_pd(ia->ri_pd);
445 /* Allow waiters to continue */
446 complete(&ia->ri_remove_done);
448 trace_xprtrdma_remove(r_xprt);
452 * rpcrdma_ia_close - Clean up/close an IA.
453 * @ia: interface adapter to close
457 rpcrdma_ia_close(struct rpcrdma_ia *ia)
459 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
461 rdma_destroy_qp(ia->ri_id);
462 rdma_destroy_id(ia->ri_id);
466 /* If the pd is still busy, xprtrdma missed freeing a resource */
467 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
468 ib_dealloc_pd(ia->ri_pd);
473 * rpcrdma_ep_create - Create unconnected endpoint
474 * @r_xprt: transport to instantiate
476 * Returns zero on success, or a negative errno.
478 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
480 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
481 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
482 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
483 struct ib_cq *sendcq, *recvcq;
484 unsigned int max_sge;
487 ep->rep_max_requests = xprt_rdma_slot_table_entries;
488 ep->rep_inline_send = xprt_rdma_max_inline_write;
489 ep->rep_inline_recv = xprt_rdma_max_inline_read;
491 max_sge = min_t(unsigned int, ia->ri_id->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);
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 ep->rep_send_batch = ep->rep_max_requests >> 3;
522 ep->rep_send_count = ep->rep_send_batch;
523 init_waitqueue_head(&ep->rep_connect_wait);
524 ep->rep_receive_count = 0;
526 sendcq = ib_alloc_cq(ia->ri_id->device, NULL,
527 ep->rep_attr.cap.max_send_wr + 1,
528 ia->ri_id->device->num_comp_vectors > 1 ? 1 : 0,
530 if (IS_ERR(sendcq)) {
531 rc = PTR_ERR(sendcq);
535 recvcq = ib_alloc_cq(ia->ri_id->device, NULL,
536 ep->rep_attr.cap.max_recv_wr + 1,
537 0, IB_POLL_WORKQUEUE);
538 if (IS_ERR(recvcq)) {
539 rc = PTR_ERR(recvcq);
543 ep->rep_attr.send_cq = sendcq;
544 ep->rep_attr.recv_cq = recvcq;
546 /* Initialize cma parameters */
547 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
549 /* Prepare RDMA-CM private message */
550 pmsg->cp_magic = rpcrdma_cmp_magic;
551 pmsg->cp_version = RPCRDMA_CMP_VERSION;
552 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
553 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send);
554 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv);
555 ep->rep_remote_cma.private_data = pmsg;
556 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
558 /* Client offers RDMA Read but does not initiate */
559 ep->rep_remote_cma.initiator_depth = 0;
560 ep->rep_remote_cma.responder_resources =
561 min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom);
563 /* Limit transport retries so client can detect server
564 * GID changes quickly. RPC layer handles re-establishing
565 * transport connection and retransmission.
567 ep->rep_remote_cma.retry_count = 6;
569 /* RPC-over-RDMA handles its own flow control. In addition,
570 * make all RNR NAKs visible so we know that RPC-over-RDMA
571 * flow control is working correctly (no NAKs should be seen).
573 ep->rep_remote_cma.flow_control = 0;
574 ep->rep_remote_cma.rnr_retry_count = 0;
585 * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
586 * @r_xprt: transport instance to shut down
589 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
591 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
592 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
594 if (ia->ri_id && ia->ri_id->qp) {
595 rpcrdma_ep_disconnect(ep, ia);
596 rdma_destroy_qp(ia->ri_id);
597 ia->ri_id->qp = NULL;
600 if (ep->rep_attr.recv_cq)
601 ib_free_cq(ep->rep_attr.recv_cq);
602 if (ep->rep_attr.send_cq)
603 ib_free_cq(ep->rep_attr.send_cq);
606 /* Re-establish a connection after a device removal event.
607 * Unlike a normal reconnection, a fresh PD and a new set
608 * of MRs and buffers is needed.
611 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
612 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
616 trace_xprtrdma_reinsert(r_xprt);
619 if (rpcrdma_ia_open(r_xprt))
623 err = rpcrdma_ep_create(r_xprt);
625 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
630 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
632 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
636 rpcrdma_mrs_create(r_xprt);
640 rpcrdma_ep_destroy(r_xprt);
642 rpcrdma_ia_close(ia);
648 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
649 struct rpcrdma_ia *ia)
651 struct rdma_cm_id *id, *old;
654 trace_xprtrdma_reconnect(r_xprt);
656 rpcrdma_ep_disconnect(ep, ia);
659 id = rpcrdma_create_id(r_xprt, ia);
663 /* As long as the new ID points to the same device as the
664 * old ID, we can reuse the transport's existing PD and all
665 * previously allocated MRs. Also, the same device means
666 * the transport's previous DMA mappings are still valid.
668 * This is a sanity check only. There should be no way these
669 * point to two different devices here.
673 if (ia->ri_id->device != id->device) {
674 pr_err("rpcrdma: can't reconnect on different device!\n");
678 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
682 /* Atomically replace the transport's ID and QP. */
686 rdma_destroy_qp(old);
689 rdma_destroy_id(old);
695 * Connect unconnected endpoint.
698 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
700 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
702 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
706 switch (ep->rep_connected) {
708 dprintk("RPC: %s: connecting...\n", __func__);
709 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
716 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
721 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
726 ep->rep_connected = 0;
727 xprt_clear_connected(xprt);
729 rpcrdma_post_recvs(r_xprt, true);
731 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
735 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
736 if (ep->rep_connected <= 0) {
737 if (ep->rep_connected == -EAGAIN)
739 rc = ep->rep_connected;
743 dprintk("RPC: %s: connected\n", __func__);
747 ep->rep_connected = rc;
754 * rpcrdma_ep_disconnect - Disconnect underlying transport
755 * @ep: endpoint to disconnect
756 * @ia: associated interface adapter
758 * This is separate from destroy to facilitate the ability
759 * to reconnect without recreating the endpoint.
761 * This call is not reentrant, and must not be made in parallel
762 * on the same endpoint.
765 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
767 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
771 /* returns without wait if ID is not connected */
772 rc = rdma_disconnect(ia->ri_id);
774 wait_event_interruptible(ep->rep_connect_wait,
775 ep->rep_connected != 1);
777 ep->rep_connected = rc;
778 trace_xprtrdma_disconnect(r_xprt, rc);
780 rpcrdma_xprt_drain(r_xprt);
783 /* Fixed-size circular FIFO queue. This implementation is wait-free and
786 * Consumer is the code path that posts Sends. This path dequeues a
787 * sendctx for use by a Send operation. Multiple consumer threads
788 * are serialized by the RPC transport lock, which allows only one
789 * ->send_request call at a time.
791 * Producer is the code path that handles Send completions. This path
792 * enqueues a sendctx that has been completed. Multiple producer
793 * threads are serialized by the ib_poll_cq() function.
796 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
797 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
800 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
804 for (i = 0; i <= buf->rb_sc_last; i++)
805 kfree(buf->rb_sc_ctxs[i]);
806 kfree(buf->rb_sc_ctxs);
809 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
811 struct rpcrdma_sendctx *sc;
813 sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges),
818 sc->sc_wr.wr_cqe = &sc->sc_cqe;
819 sc->sc_wr.sg_list = sc->sc_sges;
820 sc->sc_wr.opcode = IB_WR_SEND;
821 sc->sc_cqe.done = rpcrdma_wc_send;
825 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
827 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
828 struct rpcrdma_sendctx *sc;
831 /* Maximum number of concurrent outstanding Send WRs. Capping
832 * the circular queue size stops Send Queue overflow by causing
833 * the ->send_request call to fail temporarily before too many
836 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
837 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
838 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
839 if (!buf->rb_sc_ctxs)
842 buf->rb_sc_last = i - 1;
843 for (i = 0; i <= buf->rb_sc_last; i++) {
844 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
848 sc->sc_xprt = r_xprt;
849 buf->rb_sc_ctxs[i] = sc;
855 /* The sendctx queue is not guaranteed to have a size that is a
856 * power of two, thus the helpers in circ_buf.h cannot be used.
857 * The other option is to use modulus (%), which can be expensive.
859 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
862 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
866 * rpcrdma_sendctx_get_locked - Acquire a send context
867 * @r_xprt: controlling transport instance
869 * Returns pointer to a free send completion context; or NULL if
870 * the queue is empty.
872 * Usage: Called to acquire an SGE array before preparing a Send WR.
874 * The caller serializes calls to this function (per transport), and
875 * provides an effective memory barrier that flushes the new value
878 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
880 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
881 struct rpcrdma_sendctx *sc;
882 unsigned long next_head;
884 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
886 if (next_head == READ_ONCE(buf->rb_sc_tail))
889 /* ORDER: item must be accessed _before_ head is updated */
890 sc = buf->rb_sc_ctxs[next_head];
892 /* Releasing the lock in the caller acts as a memory
893 * barrier that flushes rb_sc_head.
895 buf->rb_sc_head = next_head;
900 /* The queue is "empty" if there have not been enough Send
901 * completions recently. This is a sign the Send Queue is
902 * backing up. Cause the caller to pause and try again.
904 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
905 r_xprt->rx_stats.empty_sendctx_q++;
910 * rpcrdma_sendctx_put_locked - Release a send context
911 * @sc: send context to release
913 * Usage: Called from Send completion to return a sendctxt
916 * The caller serializes calls to this function (per transport).
919 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
921 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
922 unsigned long next_tail;
924 /* Unmap SGEs of previously completed but unsignaled
925 * Sends by walking up the queue until @sc is found.
927 next_tail = buf->rb_sc_tail;
929 next_tail = rpcrdma_sendctx_next(buf, next_tail);
931 /* ORDER: item must be accessed _before_ tail is updated */
932 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
934 } while (buf->rb_sc_ctxs[next_tail] != sc);
936 /* Paired with READ_ONCE */
937 smp_store_release(&buf->rb_sc_tail, next_tail);
939 xprt_write_space(&sc->sc_xprt->rx_xprt);
943 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
945 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
946 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
949 for (count = 0; count < ia->ri_max_segs; count++) {
950 struct rpcrdma_mr *mr;
953 mr = kzalloc(sizeof(*mr), GFP_NOFS);
957 rc = frwr_init_mr(ia, mr);
963 mr->mr_xprt = r_xprt;
965 spin_lock(&buf->rb_mrlock);
966 list_add(&mr->mr_list, &buf->rb_mrs);
967 list_add(&mr->mr_all, &buf->rb_all_mrs);
968 spin_unlock(&buf->rb_mrlock);
971 r_xprt->rx_stats.mrs_allocated += count;
972 trace_xprtrdma_createmrs(r_xprt, count);
976 rpcrdma_mr_refresh_worker(struct work_struct *work)
978 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
980 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
983 rpcrdma_mrs_create(r_xprt);
984 xprt_write_space(&r_xprt->rx_xprt);
988 * rpcrdma_req_create - Allocate an rpcrdma_req object
989 * @r_xprt: controlling r_xprt
990 * @size: initial size, in bytes, of send and receive buffers
991 * @flags: GFP flags passed to memory allocators
993 * Returns an allocated and fully initialized rpcrdma_req or NULL.
995 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
998 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
999 struct rpcrdma_regbuf *rb;
1000 struct rpcrdma_req *req;
1003 req = kzalloc(sizeof(*req), flags);
1007 /* Compute maximum header buffer size in bytes */
1008 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
1009 r_xprt->rx_ia.ri_max_segs * rpcrdma_readchunk_maxsz;
1010 maxhdrsize *= sizeof(__be32);
1011 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
1012 DMA_TO_DEVICE, flags);
1015 req->rl_rdmabuf = rb;
1016 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1018 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1019 if (!req->rl_sendbuf)
1022 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1023 if (!req->rl_recvbuf)
1026 INIT_LIST_HEAD(&req->rl_free_mrs);
1027 INIT_LIST_HEAD(&req->rl_registered);
1028 spin_lock(&buffer->rb_lock);
1029 list_add(&req->rl_all, &buffer->rb_allreqs);
1030 spin_unlock(&buffer->rb_lock);
1034 kfree(req->rl_sendbuf);
1036 kfree(req->rl_rdmabuf);
1043 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1046 struct rpcrdma_rep *rep;
1048 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1052 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv,
1053 DMA_FROM_DEVICE, GFP_KERNEL);
1054 if (!rep->rr_rdmabuf)
1057 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
1058 rdmab_length(rep->rr_rdmabuf));
1059 rep->rr_cqe.done = rpcrdma_wc_receive;
1060 rep->rr_rxprt = r_xprt;
1061 rep->rr_recv_wr.next = NULL;
1062 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1063 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1064 rep->rr_recv_wr.num_sge = 1;
1065 rep->rr_temp = temp;
1075 * rpcrdma_buffer_create - Create initial set of req/rep objects
1076 * @r_xprt: transport instance to (re)initialize
1078 * Returns zero on success, otherwise a negative errno.
1080 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1082 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1085 buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
1086 buf->rb_bc_srv_max_requests = 0;
1087 spin_lock_init(&buf->rb_mrlock);
1088 spin_lock_init(&buf->rb_lock);
1089 INIT_LIST_HEAD(&buf->rb_mrs);
1090 INIT_LIST_HEAD(&buf->rb_all_mrs);
1091 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1093 rpcrdma_mrs_create(r_xprt);
1095 INIT_LIST_HEAD(&buf->rb_send_bufs);
1096 INIT_LIST_HEAD(&buf->rb_allreqs);
1099 for (i = 0; i < buf->rb_max_requests; i++) {
1100 struct rpcrdma_req *req;
1102 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE,
1106 list_add(&req->rl_list, &buf->rb_send_bufs);
1109 buf->rb_credits = 1;
1110 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1112 rc = rpcrdma_sendctxs_create(r_xprt);
1118 rpcrdma_buffer_destroy(buf);
1122 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1124 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1129 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1130 * @req: unused object to be destroyed
1132 * This function assumes that the caller prevents concurrent device
1133 * unload and transport tear-down.
1135 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1137 list_del(&req->rl_all);
1139 while (!list_empty(&req->rl_free_mrs))
1140 rpcrdma_mr_free(rpcrdma_mr_pop(&req->rl_free_mrs));
1142 rpcrdma_regbuf_free(req->rl_recvbuf);
1143 rpcrdma_regbuf_free(req->rl_sendbuf);
1144 rpcrdma_regbuf_free(req->rl_rdmabuf);
1149 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1151 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1153 struct rpcrdma_mr *mr;
1157 spin_lock(&buf->rb_mrlock);
1158 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1161 list_del(&mr->mr_all);
1162 spin_unlock(&buf->rb_mrlock);
1164 frwr_release_mr(mr);
1166 spin_lock(&buf->rb_mrlock);
1168 spin_unlock(&buf->rb_mrlock);
1169 r_xprt->rx_stats.mrs_allocated = 0;
1173 * rpcrdma_buffer_destroy - Release all hw resources
1174 * @buf: root control block for resources
1176 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1177 * - No more Send or Receive completions can occur
1178 * - All MRs, reps, and reqs are returned to their free lists
1181 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1183 cancel_work_sync(&buf->rb_refresh_worker);
1185 rpcrdma_sendctxs_destroy(buf);
1187 while (!list_empty(&buf->rb_recv_bufs)) {
1188 struct rpcrdma_rep *rep;
1190 rep = list_first_entry(&buf->rb_recv_bufs,
1191 struct rpcrdma_rep, rr_list);
1192 list_del(&rep->rr_list);
1193 rpcrdma_rep_destroy(rep);
1196 while (!list_empty(&buf->rb_send_bufs)) {
1197 struct rpcrdma_req *req;
1199 req = list_first_entry(&buf->rb_send_bufs,
1200 struct rpcrdma_req, rl_list);
1201 list_del(&req->rl_list);
1202 rpcrdma_req_destroy(req);
1205 rpcrdma_mrs_destroy(buf);
1209 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1210 * @r_xprt: controlling transport
1212 * Returns an initialized rpcrdma_mr or NULL if no free
1213 * rpcrdma_mr objects are available.
1216 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1218 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1219 struct rpcrdma_mr *mr;
1221 spin_lock(&buf->rb_mrlock);
1222 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1223 spin_unlock(&buf->rb_mrlock);
1228 * rpcrdma_mr_put - DMA unmap an MR and release it
1229 * @mr: MR to release
1232 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1234 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1236 if (mr->mr_dir != DMA_NONE) {
1237 trace_xprtrdma_mr_unmap(mr);
1238 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1239 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1240 mr->mr_dir = DMA_NONE;
1243 rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1246 static void rpcrdma_mr_free(struct rpcrdma_mr *mr)
1248 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1249 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1252 spin_lock(&buf->rb_mrlock);
1253 rpcrdma_mr_push(mr, &buf->rb_mrs);
1254 spin_unlock(&buf->rb_mrlock);
1258 * rpcrdma_buffer_get - Get a request buffer
1259 * @buffers: Buffer pool from which to obtain a buffer
1261 * Returns a fresh rpcrdma_req, or NULL if none are available.
1263 struct rpcrdma_req *
1264 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1266 struct rpcrdma_req *req;
1268 spin_lock(&buffers->rb_lock);
1269 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1270 struct rpcrdma_req, rl_list);
1272 list_del_init(&req->rl_list);
1273 spin_unlock(&buffers->rb_lock);
1278 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1279 * @buffers: buffer pool
1280 * @req: object to return
1283 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1285 struct rpcrdma_rep *rep = req->rl_reply;
1287 req->rl_reply = NULL;
1289 spin_lock(&buffers->rb_lock);
1290 list_add(&req->rl_list, &buffers->rb_send_bufs);
1292 if (!rep->rr_temp) {
1293 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1297 spin_unlock(&buffers->rb_lock);
1299 rpcrdma_rep_destroy(rep);
1303 * Put reply buffers back into pool when not attached to
1304 * request. This happens in error conditions.
1307 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1309 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1311 if (!rep->rr_temp) {
1312 spin_lock(&buffers->rb_lock);
1313 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1314 spin_unlock(&buffers->rb_lock);
1316 rpcrdma_rep_destroy(rep);
1320 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1322 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1323 * receiving the payload of RDMA RECV operations. During Long Calls
1324 * or Replies they may be registered externally via frwr_map.
1326 static struct rpcrdma_regbuf *
1327 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1330 struct rpcrdma_regbuf *rb;
1332 rb = kmalloc(sizeof(*rb), flags);
1335 rb->rg_data = kmalloc(size, flags);
1341 rb->rg_device = NULL;
1342 rb->rg_direction = direction;
1343 rb->rg_iov.length = size;
1348 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1349 * @rb: regbuf to reallocate
1350 * @size: size of buffer to be allocated, in bytes
1353 * Returns true if reallocation was successful. If false is
1354 * returned, @rb is left untouched.
1356 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1360 buf = kmalloc(size, flags);
1364 rpcrdma_regbuf_dma_unmap(rb);
1368 rb->rg_iov.length = size;
1373 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1374 * @r_xprt: controlling transport instance
1375 * @rb: regbuf to be mapped
1377 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1379 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1380 struct rpcrdma_regbuf *rb)
1382 struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1384 if (rb->rg_direction == DMA_NONE)
1387 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1388 rdmab_length(rb), rb->rg_direction);
1389 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1390 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1394 rb->rg_device = device;
1395 rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1399 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1404 if (!rpcrdma_regbuf_is_mapped(rb))
1407 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1409 rb->rg_device = NULL;
1412 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1414 rpcrdma_regbuf_dma_unmap(rb);
1421 * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1422 * @ia: transport's device information
1423 * @ep: transport's RDMA endpoint information
1424 * @req: rpcrdma_req containing the Send WR to post
1426 * Returns 0 if the post was successful, otherwise -ENOTCONN
1430 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1431 struct rpcrdma_ep *ep,
1432 struct rpcrdma_req *req)
1434 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1437 if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1438 send_wr->send_flags |= IB_SEND_SIGNALED;
1439 ep->rep_send_count = ep->rep_send_batch;
1441 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1442 --ep->rep_send_count;
1445 rc = frwr_send(ia, req);
1446 trace_xprtrdma_post_send(req, rc);
1453 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1455 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1456 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1457 struct ib_recv_wr *i, *wr, *bad_wr;
1458 struct rpcrdma_rep *rep;
1459 int needed, count, rc;
1464 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1465 if (ep->rep_receive_count > needed)
1467 needed -= ep->rep_receive_count;
1469 needed += RPCRDMA_MAX_RECV_BATCH;
1471 /* fast path: all needed reps can be found on the free list */
1473 spin_lock(&buf->rb_lock);
1475 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1476 struct rpcrdma_rep, rr_list);
1480 list_del(&rep->rr_list);
1481 rep->rr_recv_wr.next = wr;
1482 wr = &rep->rr_recv_wr;
1485 spin_unlock(&buf->rb_lock);
1488 rep = rpcrdma_rep_create(r_xprt, temp);
1492 rep->rr_recv_wr.next = wr;
1493 wr = &rep->rr_recv_wr;
1499 for (i = wr; i; i = i->next) {
1500 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1502 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1505 trace_xprtrdma_post_recv(rep);
1509 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1510 (const struct ib_recv_wr **)&bad_wr);
1512 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1514 for (wr = bad_wr; wr;) {
1515 struct rpcrdma_rep *rep;
1517 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1519 rpcrdma_recv_buffer_put(rep);
1523 ep->rep_receive_count += count;
1528 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1530 rpcrdma_recv_buffer_put(rep);