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 int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
79 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
80 struct rpcrdma_sendctx *sc);
81 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
82 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
83 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
84 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
85 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
86 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
87 static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
88 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
89 static struct rpcrdma_regbuf *
90 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
92 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
93 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
95 /* Wait for outstanding transport work to finish. ib_drain_qp
96 * handles the drains in the wrong order for us, so open code
99 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
101 struct rpcrdma_ep *ep = r_xprt->rx_ep;
102 struct rdma_cm_id *id = ep->re_id;
104 /* Flush Receives, then wait for deferred Reply work
109 /* Deferred Reply processing might have scheduled
110 * local invalidations.
118 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
119 * @event: details of the event
120 * @context: ep that owns QP where event occurred
122 * Called from the RDMA provider (device driver) possibly in an interrupt
123 * context. The QP is always destroyed before the ID, so the ID will be
124 * reliably available when this handler is invoked.
126 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
128 struct rpcrdma_ep *ep = context;
130 trace_xprtrdma_qp_event(ep, event);
133 /* Ensure xprt_force_disconnect() is invoked exactly once when a
134 * connection is closed or lost. (The important thing is it needs
135 * to be invoked "at least" once).
137 static void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
139 if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
140 xprt_force_disconnect(ep->re_xprt);
144 * rpcrdma_flush_disconnect - Disconnect on flushed completion
145 * @r_xprt: transport to disconnect
146 * @wc: work completion entry
148 * Must be called in process context.
150 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
152 if (wc->status != IB_WC_SUCCESS)
153 rpcrdma_force_disconnect(r_xprt->rx_ep);
157 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
158 * @cq: completion queue
159 * @wc: WCE for a completed Send WR
162 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
164 struct ib_cqe *cqe = wc->wr_cqe;
165 struct rpcrdma_sendctx *sc =
166 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
167 struct rpcrdma_xprt *r_xprt = cq->cq_context;
169 /* WARNING: Only wr_cqe and status are reliable at this point */
170 trace_xprtrdma_wc_send(wc, &sc->sc_cid);
171 rpcrdma_sendctx_put_locked(r_xprt, sc);
172 rpcrdma_flush_disconnect(r_xprt, wc);
176 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
177 * @cq: completion queue
178 * @wc: WCE for a completed Receive WR
181 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
183 struct ib_cqe *cqe = wc->wr_cqe;
184 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
186 struct rpcrdma_xprt *r_xprt = cq->cq_context;
188 /* WARNING: Only wr_cqe and status are reliable at this point */
189 trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
190 --r_xprt->rx_ep->re_receive_count;
191 if (wc->status != IB_WC_SUCCESS)
194 /* status == SUCCESS means all fields in wc are trustworthy */
195 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
196 rep->rr_wc_flags = wc->wc_flags;
197 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
199 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
200 rdmab_addr(rep->rr_rdmabuf),
201 wc->byte_len, DMA_FROM_DEVICE);
203 rpcrdma_reply_handler(rep);
207 rpcrdma_flush_disconnect(r_xprt, wc);
208 rpcrdma_rep_destroy(rep);
211 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
212 struct rdma_conn_param *param)
214 const struct rpcrdma_connect_private *pmsg = param->private_data;
215 unsigned int rsize, wsize;
217 /* Default settings for RPC-over-RDMA Version One */
218 ep->re_implicit_roundup = xprt_rdma_pad_optimize;
219 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
220 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
223 pmsg->cp_magic == rpcrdma_cmp_magic &&
224 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
225 ep->re_implicit_roundup = true;
226 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
227 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
230 if (rsize < ep->re_inline_recv)
231 ep->re_inline_recv = rsize;
232 if (wsize < ep->re_inline_send)
233 ep->re_inline_send = wsize;
235 rpcrdma_set_max_header_sizes(ep);
239 * rpcrdma_cm_event_handler - Handle RDMA CM events
240 * @id: rdma_cm_id on which an event has occurred
241 * @event: details of the event
243 * Called with @id's mutex held. Returns 1 if caller should
244 * destroy @id, otherwise 0.
247 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
249 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
250 struct rpcrdma_ep *ep = id->context;
254 switch (event->event) {
255 case RDMA_CM_EVENT_ADDR_RESOLVED:
256 case RDMA_CM_EVENT_ROUTE_RESOLVED:
258 complete(&ep->re_done);
260 case RDMA_CM_EVENT_ADDR_ERROR:
261 ep->re_async_rc = -EPROTO;
262 complete(&ep->re_done);
264 case RDMA_CM_EVENT_ROUTE_ERROR:
265 ep->re_async_rc = -ENETUNREACH;
266 complete(&ep->re_done);
268 case RDMA_CM_EVENT_DEVICE_REMOVAL:
269 pr_info("rpcrdma: removing device %s for %pISpc\n",
270 ep->re_id->device->name, sap);
272 case RDMA_CM_EVENT_ADDR_CHANGE:
273 ep->re_connect_status = -ENODEV;
275 case RDMA_CM_EVENT_ESTABLISHED:
277 ep->re_connect_status = 1;
278 rpcrdma_update_cm_private(ep, &event->param.conn);
279 trace_xprtrdma_inline_thresh(ep);
280 wake_up_all(&ep->re_connect_wait);
282 case RDMA_CM_EVENT_CONNECT_ERROR:
283 ep->re_connect_status = -ENOTCONN;
284 goto wake_connect_worker;
285 case RDMA_CM_EVENT_UNREACHABLE:
286 ep->re_connect_status = -ENETUNREACH;
287 goto wake_connect_worker;
288 case RDMA_CM_EVENT_REJECTED:
289 dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
290 sap, rdma_reject_msg(id, event->status));
291 ep->re_connect_status = -ECONNREFUSED;
292 if (event->status == IB_CM_REJ_STALE_CONN)
293 ep->re_connect_status = -ENOTCONN;
295 wake_up_all(&ep->re_connect_wait);
297 case RDMA_CM_EVENT_DISCONNECTED:
298 ep->re_connect_status = -ECONNABORTED;
300 rpcrdma_force_disconnect(ep);
301 return rpcrdma_ep_put(ep);
306 dprintk("RPC: %s: %pISpc on %s/frwr: %s\n", __func__, sap,
307 ep->re_id->device->name, rdma_event_msg(event->event));
311 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
312 struct rpcrdma_ep *ep)
314 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
315 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
316 struct rdma_cm_id *id;
319 init_completion(&ep->re_done);
321 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
322 RDMA_PS_TCP, IB_QPT_RC);
326 ep->re_async_rc = -ETIMEDOUT;
327 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
328 RDMA_RESOLVE_TIMEOUT);
331 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
335 rc = ep->re_async_rc;
339 ep->re_async_rc = -ETIMEDOUT;
340 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
343 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
346 rc = ep->re_async_rc;
357 static void rpcrdma_ep_destroy(struct kref *kref)
359 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
362 rdma_destroy_qp(ep->re_id);
363 ep->re_id->qp = NULL;
366 if (ep->re_attr.recv_cq)
367 ib_free_cq(ep->re_attr.recv_cq);
368 ep->re_attr.recv_cq = NULL;
369 if (ep->re_attr.send_cq)
370 ib_free_cq(ep->re_attr.send_cq);
371 ep->re_attr.send_cq = NULL;
374 ib_dealloc_pd(ep->re_pd);
378 module_put(THIS_MODULE);
381 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
383 kref_get(&ep->re_kref);
387 * %0 if @ep still has a positive kref count, or
388 * %1 if @ep was destroyed successfully.
390 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
392 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
395 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
397 struct rpcrdma_connect_private *pmsg;
398 struct ib_device *device;
399 struct rdma_cm_id *id;
400 struct rpcrdma_ep *ep;
403 ep = kzalloc(sizeof(*ep), GFP_NOFS);
406 ep->re_xprt = &r_xprt->rx_xprt;
407 kref_init(&ep->re_kref);
409 id = rpcrdma_create_id(r_xprt, ep);
414 __module_get(THIS_MODULE);
418 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
419 ep->re_inline_send = xprt_rdma_max_inline_write;
420 ep->re_inline_recv = xprt_rdma_max_inline_read;
421 rc = frwr_query_device(ep, device);
425 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
427 ep->re_attr.event_handler = rpcrdma_qp_event_handler;
428 ep->re_attr.qp_context = ep;
429 ep->re_attr.srq = NULL;
430 ep->re_attr.cap.max_inline_data = 0;
431 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
432 ep->re_attr.qp_type = IB_QPT_RC;
433 ep->re_attr.port_num = ~0;
435 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
436 "iovs: send %d recv %d\n",
438 ep->re_attr.cap.max_send_wr,
439 ep->re_attr.cap.max_recv_wr,
440 ep->re_attr.cap.max_send_sge,
441 ep->re_attr.cap.max_recv_sge);
443 ep->re_send_batch = ep->re_max_requests >> 3;
444 ep->re_send_count = ep->re_send_batch;
445 init_waitqueue_head(&ep->re_connect_wait);
447 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
448 ep->re_attr.cap.max_send_wr,
450 if (IS_ERR(ep->re_attr.send_cq)) {
451 rc = PTR_ERR(ep->re_attr.send_cq);
455 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
456 ep->re_attr.cap.max_recv_wr,
458 if (IS_ERR(ep->re_attr.recv_cq)) {
459 rc = PTR_ERR(ep->re_attr.recv_cq);
462 ep->re_receive_count = 0;
464 /* Initialize cma parameters */
465 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
467 /* Prepare RDMA-CM private message */
468 pmsg = &ep->re_cm_private;
469 pmsg->cp_magic = rpcrdma_cmp_magic;
470 pmsg->cp_version = RPCRDMA_CMP_VERSION;
471 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
472 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
473 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
474 ep->re_remote_cma.private_data = pmsg;
475 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
477 /* Client offers RDMA Read but does not initiate */
478 ep->re_remote_cma.initiator_depth = 0;
479 ep->re_remote_cma.responder_resources =
480 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
482 /* Limit transport retries so client can detect server
483 * GID changes quickly. RPC layer handles re-establishing
484 * transport connection and retransmission.
486 ep->re_remote_cma.retry_count = 6;
488 /* RPC-over-RDMA handles its own flow control. In addition,
489 * make all RNR NAKs visible so we know that RPC-over-RDMA
490 * flow control is working correctly (no NAKs should be seen).
492 ep->re_remote_cma.flow_control = 0;
493 ep->re_remote_cma.rnr_retry_count = 0;
495 ep->re_pd = ib_alloc_pd(device, 0);
496 if (IS_ERR(ep->re_pd)) {
497 rc = PTR_ERR(ep->re_pd);
501 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
515 * rpcrdma_xprt_connect - Connect an unconnected transport
516 * @r_xprt: controlling transport instance
518 * Returns 0 on success or a negative errno.
520 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
522 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
523 struct rpcrdma_ep *ep;
526 rc = rpcrdma_ep_create(r_xprt);
531 xprt_clear_connected(xprt);
532 rpcrdma_reset_cwnd(r_xprt);
534 /* Bump the ep's reference count while there are
535 * outstanding Receives.
538 rpcrdma_post_recvs(r_xprt, true);
540 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
544 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
545 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
546 wait_event_interruptible(ep->re_connect_wait,
547 ep->re_connect_status != 0);
548 if (ep->re_connect_status <= 0) {
549 rc = ep->re_connect_status;
553 rc = rpcrdma_sendctxs_create(r_xprt);
559 rc = rpcrdma_reqs_setup(r_xprt);
564 rpcrdma_mrs_create(r_xprt);
567 trace_xprtrdma_connect(r_xprt, rc);
572 * rpcrdma_xprt_disconnect - Disconnect underlying transport
573 * @r_xprt: controlling transport instance
575 * Caller serializes. Either the transport send lock is held,
576 * or we're being called to destroy the transport.
578 * On return, @r_xprt is completely divested of all hardware
579 * resources and prepared for the next ->connect operation.
581 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
583 struct rpcrdma_ep *ep = r_xprt->rx_ep;
584 struct rdma_cm_id *id;
591 rc = rdma_disconnect(id);
592 trace_xprtrdma_disconnect(r_xprt, rc);
594 rpcrdma_xprt_drain(r_xprt);
595 rpcrdma_reps_unmap(r_xprt);
596 rpcrdma_reqs_reset(r_xprt);
597 rpcrdma_mrs_destroy(r_xprt);
598 rpcrdma_sendctxs_destroy(r_xprt);
600 if (rpcrdma_ep_put(ep))
603 r_xprt->rx_ep = NULL;
606 /* Fixed-size circular FIFO queue. This implementation is wait-free and
609 * Consumer is the code path that posts Sends. This path dequeues a
610 * sendctx for use by a Send operation. Multiple consumer threads
611 * are serialized by the RPC transport lock, which allows only one
612 * ->send_request call at a time.
614 * Producer is the code path that handles Send completions. This path
615 * enqueues a sendctx that has been completed. Multiple producer
616 * threads are serialized by the ib_poll_cq() function.
619 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
620 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
623 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
625 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
628 if (!buf->rb_sc_ctxs)
630 for (i = 0; i <= buf->rb_sc_last; i++)
631 kfree(buf->rb_sc_ctxs[i]);
632 kfree(buf->rb_sc_ctxs);
633 buf->rb_sc_ctxs = NULL;
636 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
638 struct rpcrdma_sendctx *sc;
640 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
645 sc->sc_cqe.done = rpcrdma_wc_send;
646 sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
647 sc->sc_cid.ci_completion_id =
648 atomic_inc_return(&ep->re_completion_ids);
652 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
654 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
655 struct rpcrdma_sendctx *sc;
658 /* Maximum number of concurrent outstanding Send WRs. Capping
659 * the circular queue size stops Send Queue overflow by causing
660 * the ->send_request call to fail temporarily before too many
663 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
664 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
665 if (!buf->rb_sc_ctxs)
668 buf->rb_sc_last = i - 1;
669 for (i = 0; i <= buf->rb_sc_last; i++) {
670 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
674 buf->rb_sc_ctxs[i] = sc;
682 /* The sendctx queue is not guaranteed to have a size that is a
683 * power of two, thus the helpers in circ_buf.h cannot be used.
684 * The other option is to use modulus (%), which can be expensive.
686 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
689 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
693 * rpcrdma_sendctx_get_locked - Acquire a send context
694 * @r_xprt: controlling transport instance
696 * Returns pointer to a free send completion context; or NULL if
697 * the queue is empty.
699 * Usage: Called to acquire an SGE array before preparing a Send WR.
701 * The caller serializes calls to this function (per transport), and
702 * provides an effective memory barrier that flushes the new value
705 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
707 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
708 struct rpcrdma_sendctx *sc;
709 unsigned long next_head;
711 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
713 if (next_head == READ_ONCE(buf->rb_sc_tail))
716 /* ORDER: item must be accessed _before_ head is updated */
717 sc = buf->rb_sc_ctxs[next_head];
719 /* Releasing the lock in the caller acts as a memory
720 * barrier that flushes rb_sc_head.
722 buf->rb_sc_head = next_head;
727 /* The queue is "empty" if there have not been enough Send
728 * completions recently. This is a sign the Send Queue is
729 * backing up. Cause the caller to pause and try again.
731 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
732 r_xprt->rx_stats.empty_sendctx_q++;
737 * rpcrdma_sendctx_put_locked - Release a send context
738 * @r_xprt: controlling transport instance
739 * @sc: send context to release
741 * Usage: Called from Send completion to return a sendctxt
744 * The caller serializes calls to this function (per transport).
746 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
747 struct rpcrdma_sendctx *sc)
749 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
750 unsigned long next_tail;
752 /* Unmap SGEs of previously completed but unsignaled
753 * Sends by walking up the queue until @sc is found.
755 next_tail = buf->rb_sc_tail;
757 next_tail = rpcrdma_sendctx_next(buf, next_tail);
759 /* ORDER: item must be accessed _before_ tail is updated */
760 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
762 } while (buf->rb_sc_ctxs[next_tail] != sc);
764 /* Paired with READ_ONCE */
765 smp_store_release(&buf->rb_sc_tail, next_tail);
767 xprt_write_space(&r_xprt->rx_xprt);
771 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
773 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
774 struct rpcrdma_ep *ep = r_xprt->rx_ep;
777 for (count = 0; count < ep->re_max_rdma_segs; count++) {
778 struct rpcrdma_mr *mr;
781 mr = kzalloc(sizeof(*mr), GFP_NOFS);
785 rc = frwr_mr_init(r_xprt, mr);
791 spin_lock(&buf->rb_lock);
792 rpcrdma_mr_push(mr, &buf->rb_mrs);
793 list_add(&mr->mr_all, &buf->rb_all_mrs);
794 spin_unlock(&buf->rb_lock);
797 r_xprt->rx_stats.mrs_allocated += count;
798 trace_xprtrdma_createmrs(r_xprt, count);
802 rpcrdma_mr_refresh_worker(struct work_struct *work)
804 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
806 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
809 rpcrdma_mrs_create(r_xprt);
810 xprt_write_space(&r_xprt->rx_xprt);
814 * rpcrdma_mrs_refresh - Wake the MR refresh worker
815 * @r_xprt: controlling transport instance
818 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
820 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
821 struct rpcrdma_ep *ep = r_xprt->rx_ep;
823 /* If there is no underlying connection, it's no use
824 * to wake the refresh worker.
826 if (ep->re_connect_status == 1) {
827 /* The work is scheduled on a WQ_MEM_RECLAIM
828 * workqueue in order to prevent MR allocation
829 * from recursing into NFS during direct reclaim.
831 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
836 * rpcrdma_req_create - Allocate an rpcrdma_req object
837 * @r_xprt: controlling r_xprt
838 * @size: initial size, in bytes, of send and receive buffers
839 * @flags: GFP flags passed to memory allocators
841 * Returns an allocated and fully initialized rpcrdma_req or NULL.
843 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
846 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
847 struct rpcrdma_req *req;
849 req = kzalloc(sizeof(*req), flags);
853 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
854 if (!req->rl_sendbuf)
857 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
858 if (!req->rl_recvbuf)
861 INIT_LIST_HEAD(&req->rl_free_mrs);
862 INIT_LIST_HEAD(&req->rl_registered);
863 spin_lock(&buffer->rb_lock);
864 list_add(&req->rl_all, &buffer->rb_allreqs);
865 spin_unlock(&buffer->rb_lock);
869 kfree(req->rl_sendbuf);
877 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
878 * @r_xprt: controlling transport instance
879 * @req: rpcrdma_req object to set up
881 * Returns zero on success, and a negative errno on failure.
883 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
885 struct rpcrdma_regbuf *rb;
888 /* Compute maximum header buffer size in bytes */
889 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
890 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
891 maxhdrsize *= sizeof(__be32);
892 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
893 DMA_TO_DEVICE, GFP_KERNEL);
897 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
900 req->rl_rdmabuf = rb;
901 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
905 rpcrdma_regbuf_free(rb);
910 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
911 * and thus can be walked without holding rb_lock. Eg. the
912 * caller is holding the transport send lock to exclude
913 * device removal or disconnection.
915 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
917 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
918 struct rpcrdma_req *req;
921 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
922 rc = rpcrdma_req_setup(r_xprt, req);
929 static void rpcrdma_req_reset(struct rpcrdma_req *req)
931 /* Credits are valid for only one connection */
932 req->rl_slot.rq_cong = 0;
934 rpcrdma_regbuf_free(req->rl_rdmabuf);
935 req->rl_rdmabuf = NULL;
937 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
938 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
943 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
944 * and thus can be walked without holding rb_lock. Eg. the
945 * caller is holding the transport send lock to exclude
946 * device removal or disconnection.
948 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
950 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
951 struct rpcrdma_req *req;
953 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
954 rpcrdma_req_reset(req);
957 /* No locking needed here. This function is called only by the
958 * Receive completion handler.
961 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
964 struct rpcrdma_rep *rep;
966 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
970 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
971 DMA_FROM_DEVICE, GFP_KERNEL);
972 if (!rep->rr_rdmabuf)
975 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
976 goto out_free_regbuf;
978 rep->rr_cid.ci_completion_id =
979 atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
981 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
982 rdmab_length(rep->rr_rdmabuf));
983 rep->rr_cqe.done = rpcrdma_wc_receive;
984 rep->rr_rxprt = r_xprt;
985 rep->rr_recv_wr.next = NULL;
986 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
987 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
988 rep->rr_recv_wr.num_sge = 1;
990 list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
994 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1001 /* No locking needed here. This function is invoked only by the
1002 * Receive completion handler, or during transport shutdown.
1004 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1006 list_del(&rep->rr_all);
1007 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1011 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1013 struct llist_node *node;
1015 /* Calls to llist_del_first are required to be serialized */
1016 node = llist_del_first(&buf->rb_free_reps);
1019 return llist_entry(node, struct rpcrdma_rep, rr_node);
1022 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1023 struct rpcrdma_rep *rep)
1025 llist_add(&rep->rr_node, &buf->rb_free_reps);
1028 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1030 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1031 struct rpcrdma_rep *rep;
1033 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1034 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1035 rep->rr_temp = true;
1039 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1041 struct rpcrdma_rep *rep;
1043 while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1044 rpcrdma_rep_destroy(rep);
1048 * rpcrdma_buffer_create - Create initial set of req/rep objects
1049 * @r_xprt: transport instance to (re)initialize
1051 * Returns zero on success, otherwise a negative errno.
1053 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1055 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1058 buf->rb_bc_srv_max_requests = 0;
1059 spin_lock_init(&buf->rb_lock);
1060 INIT_LIST_HEAD(&buf->rb_mrs);
1061 INIT_LIST_HEAD(&buf->rb_all_mrs);
1062 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1064 INIT_LIST_HEAD(&buf->rb_send_bufs);
1065 INIT_LIST_HEAD(&buf->rb_allreqs);
1066 INIT_LIST_HEAD(&buf->rb_all_reps);
1069 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1070 struct rpcrdma_req *req;
1072 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1076 list_add(&req->rl_list, &buf->rb_send_bufs);
1079 init_llist_head(&buf->rb_free_reps);
1083 rpcrdma_buffer_destroy(buf);
1088 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1089 * @req: unused object to be destroyed
1091 * Relies on caller holding the transport send lock to protect
1092 * removing req->rl_all from buf->rb_all_reqs safely.
1094 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1096 struct rpcrdma_mr *mr;
1098 list_del(&req->rl_all);
1100 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1101 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1103 spin_lock(&buf->rb_lock);
1104 list_del(&mr->mr_all);
1105 spin_unlock(&buf->rb_lock);
1107 frwr_release_mr(mr);
1110 rpcrdma_regbuf_free(req->rl_recvbuf);
1111 rpcrdma_regbuf_free(req->rl_sendbuf);
1112 rpcrdma_regbuf_free(req->rl_rdmabuf);
1117 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1118 * @r_xprt: controlling transport instance
1120 * Relies on caller holding the transport send lock to protect
1121 * removing mr->mr_list from req->rl_free_mrs safely.
1123 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1125 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1126 struct rpcrdma_mr *mr;
1128 cancel_work_sync(&buf->rb_refresh_worker);
1130 spin_lock(&buf->rb_lock);
1131 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1134 list_del(&mr->mr_list);
1135 list_del(&mr->mr_all);
1136 spin_unlock(&buf->rb_lock);
1138 frwr_release_mr(mr);
1140 spin_lock(&buf->rb_lock);
1142 spin_unlock(&buf->rb_lock);
1146 * rpcrdma_buffer_destroy - Release all hw resources
1147 * @buf: root control block for resources
1149 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1150 * - No more Send or Receive completions can occur
1151 * - All MRs, reps, and reqs are returned to their free lists
1154 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1156 rpcrdma_reps_destroy(buf);
1158 while (!list_empty(&buf->rb_send_bufs)) {
1159 struct rpcrdma_req *req;
1161 req = list_first_entry(&buf->rb_send_bufs,
1162 struct rpcrdma_req, rl_list);
1163 list_del(&req->rl_list);
1164 rpcrdma_req_destroy(req);
1169 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1170 * @r_xprt: controlling transport
1172 * Returns an initialized rpcrdma_mr or NULL if no free
1173 * rpcrdma_mr objects are available.
1176 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1178 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1179 struct rpcrdma_mr *mr;
1181 spin_lock(&buf->rb_lock);
1182 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1183 spin_unlock(&buf->rb_lock);
1188 * rpcrdma_buffer_get - Get a request buffer
1189 * @buffers: Buffer pool from which to obtain a buffer
1191 * Returns a fresh rpcrdma_req, or NULL if none are available.
1193 struct rpcrdma_req *
1194 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1196 struct rpcrdma_req *req;
1198 spin_lock(&buffers->rb_lock);
1199 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1200 struct rpcrdma_req, rl_list);
1202 list_del_init(&req->rl_list);
1203 spin_unlock(&buffers->rb_lock);
1208 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1209 * @buffers: buffer pool
1210 * @req: object to return
1213 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1216 rpcrdma_rep_put(buffers, req->rl_reply);
1217 req->rl_reply = NULL;
1219 spin_lock(&buffers->rb_lock);
1220 list_add(&req->rl_list, &buffers->rb_send_bufs);
1221 spin_unlock(&buffers->rb_lock);
1225 * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1226 * @rep: rep to release
1228 * Used after error conditions.
1230 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1232 rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1235 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1237 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1238 * receiving the payload of RDMA RECV operations. During Long Calls
1239 * or Replies they may be registered externally via frwr_map.
1241 static struct rpcrdma_regbuf *
1242 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1245 struct rpcrdma_regbuf *rb;
1247 rb = kmalloc(sizeof(*rb), flags);
1250 rb->rg_data = kmalloc(size, flags);
1256 rb->rg_device = NULL;
1257 rb->rg_direction = direction;
1258 rb->rg_iov.length = size;
1263 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1264 * @rb: regbuf to reallocate
1265 * @size: size of buffer to be allocated, in bytes
1268 * Returns true if reallocation was successful. If false is
1269 * returned, @rb is left untouched.
1271 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1275 buf = kmalloc(size, flags);
1279 rpcrdma_regbuf_dma_unmap(rb);
1283 rb->rg_iov.length = size;
1288 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1289 * @r_xprt: controlling transport instance
1290 * @rb: regbuf to be mapped
1292 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1294 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1295 struct rpcrdma_regbuf *rb)
1297 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1299 if (rb->rg_direction == DMA_NONE)
1302 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1303 rdmab_length(rb), rb->rg_direction);
1304 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1305 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1309 rb->rg_device = device;
1310 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1314 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1319 if (!rpcrdma_regbuf_is_mapped(rb))
1322 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1324 rb->rg_device = NULL;
1327 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1329 rpcrdma_regbuf_dma_unmap(rb);
1336 * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1337 * @r_xprt: controlling transport instance
1338 * @req: rpcrdma_req containing the Send WR to post
1340 * Returns 0 if the post was successful, otherwise -ENOTCONN
1343 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1345 struct ib_send_wr *send_wr = &req->rl_wr;
1346 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1349 if (!ep->re_send_count || kref_read(&req->rl_kref) > 1) {
1350 send_wr->send_flags |= IB_SEND_SIGNALED;
1351 ep->re_send_count = ep->re_send_batch;
1353 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1354 --ep->re_send_count;
1357 trace_xprtrdma_post_send(req);
1358 rc = frwr_send(r_xprt, req);
1365 * rpcrdma_post_recvs - Refill the Receive Queue
1366 * @r_xprt: controlling transport instance
1367 * @temp: mark Receive buffers to be deleted after use
1370 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1372 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1373 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1374 struct ib_recv_wr *wr, *bad_wr;
1375 struct rpcrdma_rep *rep;
1376 int needed, count, rc;
1381 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1382 if (likely(ep->re_receive_count > needed))
1384 needed -= ep->re_receive_count;
1386 needed += RPCRDMA_MAX_RECV_BATCH;
1388 /* fast path: all needed reps can be found on the free list */
1391 rep = rpcrdma_rep_get_locked(buf);
1392 if (rep && rep->rr_temp) {
1393 rpcrdma_rep_destroy(rep);
1397 rep = rpcrdma_rep_create(r_xprt, temp);
1401 rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1402 trace_xprtrdma_post_recv(rep);
1403 rep->rr_recv_wr.next = wr;
1404 wr = &rep->rr_recv_wr;
1411 rc = ib_post_recv(ep->re_id->qp, wr,
1412 (const struct ib_recv_wr **)&bad_wr);
1414 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1416 for (wr = bad_wr; wr;) {
1417 struct rpcrdma_rep *rep;
1419 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1421 rpcrdma_recv_buffer_put(rep);
1425 ep->re_receive_count += count;