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>
66 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
67 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
68 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
69 struct rpcrdma_sendctx *sc);
70 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
71 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
72 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
73 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
74 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
75 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
76 static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
77 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
78 static struct rpcrdma_regbuf *
79 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
81 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
82 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
84 /* Wait for outstanding transport work to finish. ib_drain_qp
85 * handles the drains in the wrong order for us, so open code
88 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
90 struct rpcrdma_ep *ep = r_xprt->rx_ep;
91 struct rdma_cm_id *id = ep->re_id;
93 /* Wait for rpcrdma_post_recvs() to leave its critical
96 if (atomic_inc_return(&ep->re_receiving) > 1)
97 wait_for_completion(&ep->re_done);
99 /* Flush Receives, then wait for deferred Reply work
104 /* Deferred Reply processing might have scheduled
105 * local invalidations.
112 /* Ensure xprt_force_disconnect() is invoked exactly once when a
113 * connection is closed or lost. (The important thing is it needs
114 * to be invoked "at least" once).
116 void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
118 if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
119 xprt_force_disconnect(ep->re_xprt);
123 * rpcrdma_flush_disconnect - Disconnect on flushed completion
124 * @r_xprt: transport to disconnect
125 * @wc: work completion entry
127 * Must be called in process context.
129 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
131 if (wc->status != IB_WC_SUCCESS)
132 rpcrdma_force_disconnect(r_xprt->rx_ep);
136 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
137 * @cq: completion queue
138 * @wc: WCE for a completed Send WR
141 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
143 struct ib_cqe *cqe = wc->wr_cqe;
144 struct rpcrdma_sendctx *sc =
145 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
146 struct rpcrdma_xprt *r_xprt = cq->cq_context;
148 /* WARNING: Only wr_cqe and status are reliable at this point */
149 trace_xprtrdma_wc_send(wc, &sc->sc_cid);
150 rpcrdma_sendctx_put_locked(r_xprt, sc);
151 rpcrdma_flush_disconnect(r_xprt, wc);
155 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
156 * @cq: completion queue
157 * @wc: WCE for a completed Receive WR
160 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
162 struct ib_cqe *cqe = wc->wr_cqe;
163 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
165 struct rpcrdma_xprt *r_xprt = cq->cq_context;
167 /* WARNING: Only wr_cqe and status are reliable at this point */
168 trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
169 --r_xprt->rx_ep->re_receive_count;
170 if (wc->status != IB_WC_SUCCESS)
173 /* status == SUCCESS means all fields in wc are trustworthy */
174 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
175 rep->rr_wc_flags = wc->wc_flags;
176 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
178 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
179 rdmab_addr(rep->rr_rdmabuf),
180 wc->byte_len, DMA_FROM_DEVICE);
182 rpcrdma_reply_handler(rep);
186 rpcrdma_flush_disconnect(r_xprt, wc);
187 rpcrdma_rep_put(&r_xprt->rx_buf, rep);
190 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
191 struct rdma_conn_param *param)
193 const struct rpcrdma_connect_private *pmsg = param->private_data;
194 unsigned int rsize, wsize;
196 /* Default settings for RPC-over-RDMA Version One */
197 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
198 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
201 pmsg->cp_magic == rpcrdma_cmp_magic &&
202 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
203 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
204 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
207 if (rsize < ep->re_inline_recv)
208 ep->re_inline_recv = rsize;
209 if (wsize < ep->re_inline_send)
210 ep->re_inline_send = wsize;
212 rpcrdma_set_max_header_sizes(ep);
216 * rpcrdma_cm_event_handler - Handle RDMA CM events
217 * @id: rdma_cm_id on which an event has occurred
218 * @event: details of the event
220 * Called with @id's mutex held. Returns 1 if caller should
221 * destroy @id, otherwise 0.
224 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
226 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
227 struct rpcrdma_ep *ep = id->context;
231 switch (event->event) {
232 case RDMA_CM_EVENT_ADDR_RESOLVED:
233 case RDMA_CM_EVENT_ROUTE_RESOLVED:
235 complete(&ep->re_done);
237 case RDMA_CM_EVENT_ADDR_ERROR:
238 ep->re_async_rc = -EPROTO;
239 complete(&ep->re_done);
241 case RDMA_CM_EVENT_ROUTE_ERROR:
242 ep->re_async_rc = -ENETUNREACH;
243 complete(&ep->re_done);
245 case RDMA_CM_EVENT_DEVICE_REMOVAL:
246 pr_info("rpcrdma: removing device %s for %pISpc\n",
247 ep->re_id->device->name, sap);
249 case RDMA_CM_EVENT_ADDR_CHANGE:
250 ep->re_connect_status = -ENODEV;
252 case RDMA_CM_EVENT_ESTABLISHED:
254 ep->re_connect_status = 1;
255 rpcrdma_update_cm_private(ep, &event->param.conn);
256 trace_xprtrdma_inline_thresh(ep);
257 wake_up_all(&ep->re_connect_wait);
259 case RDMA_CM_EVENT_CONNECT_ERROR:
260 ep->re_connect_status = -ENOTCONN;
261 goto wake_connect_worker;
262 case RDMA_CM_EVENT_UNREACHABLE:
263 ep->re_connect_status = -ENETUNREACH;
264 goto wake_connect_worker;
265 case RDMA_CM_EVENT_REJECTED:
266 ep->re_connect_status = -ECONNREFUSED;
267 if (event->status == IB_CM_REJ_STALE_CONN)
268 ep->re_connect_status = -ENOTCONN;
270 wake_up_all(&ep->re_connect_wait);
272 case RDMA_CM_EVENT_DISCONNECTED:
273 ep->re_connect_status = -ECONNABORTED;
275 rpcrdma_force_disconnect(ep);
276 return rpcrdma_ep_put(ep);
284 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
285 struct rpcrdma_ep *ep)
287 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
288 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
289 struct rdma_cm_id *id;
292 init_completion(&ep->re_done);
294 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
295 RDMA_PS_TCP, IB_QPT_RC);
299 ep->re_async_rc = -ETIMEDOUT;
300 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
301 RDMA_RESOLVE_TIMEOUT);
304 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
308 rc = ep->re_async_rc;
312 ep->re_async_rc = -ETIMEDOUT;
313 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
316 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
319 rc = ep->re_async_rc;
330 static void rpcrdma_ep_destroy(struct kref *kref)
332 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
335 rdma_destroy_qp(ep->re_id);
336 ep->re_id->qp = NULL;
339 if (ep->re_attr.recv_cq)
340 ib_free_cq(ep->re_attr.recv_cq);
341 ep->re_attr.recv_cq = NULL;
342 if (ep->re_attr.send_cq)
343 ib_free_cq(ep->re_attr.send_cq);
344 ep->re_attr.send_cq = NULL;
347 ib_dealloc_pd(ep->re_pd);
351 module_put(THIS_MODULE);
354 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
356 kref_get(&ep->re_kref);
360 * %0 if @ep still has a positive kref count, or
361 * %1 if @ep was destroyed successfully.
363 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
365 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
368 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
370 struct rpcrdma_connect_private *pmsg;
371 struct ib_device *device;
372 struct rdma_cm_id *id;
373 struct rpcrdma_ep *ep;
376 ep = kzalloc(sizeof(*ep), GFP_NOFS);
379 ep->re_xprt = &r_xprt->rx_xprt;
380 kref_init(&ep->re_kref);
382 id = rpcrdma_create_id(r_xprt, ep);
387 __module_get(THIS_MODULE);
390 reinit_completion(&ep->re_done);
392 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
393 ep->re_inline_send = xprt_rdma_max_inline_write;
394 ep->re_inline_recv = xprt_rdma_max_inline_read;
395 rc = frwr_query_device(ep, device);
399 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
401 ep->re_attr.srq = NULL;
402 ep->re_attr.cap.max_inline_data = 0;
403 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
404 ep->re_attr.qp_type = IB_QPT_RC;
405 ep->re_attr.port_num = ~0;
407 ep->re_send_batch = ep->re_max_requests >> 3;
408 ep->re_send_count = ep->re_send_batch;
409 init_waitqueue_head(&ep->re_connect_wait);
411 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
412 ep->re_attr.cap.max_send_wr,
414 if (IS_ERR(ep->re_attr.send_cq)) {
415 rc = PTR_ERR(ep->re_attr.send_cq);
419 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
420 ep->re_attr.cap.max_recv_wr,
422 if (IS_ERR(ep->re_attr.recv_cq)) {
423 rc = PTR_ERR(ep->re_attr.recv_cq);
426 ep->re_receive_count = 0;
428 /* Initialize cma parameters */
429 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
431 /* Prepare RDMA-CM private message */
432 pmsg = &ep->re_cm_private;
433 pmsg->cp_magic = rpcrdma_cmp_magic;
434 pmsg->cp_version = RPCRDMA_CMP_VERSION;
435 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
436 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
437 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
438 ep->re_remote_cma.private_data = pmsg;
439 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
441 /* Client offers RDMA Read but does not initiate */
442 ep->re_remote_cma.initiator_depth = 0;
443 ep->re_remote_cma.responder_resources =
444 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
446 /* Limit transport retries so client can detect server
447 * GID changes quickly. RPC layer handles re-establishing
448 * transport connection and retransmission.
450 ep->re_remote_cma.retry_count = 6;
452 /* RPC-over-RDMA handles its own flow control. In addition,
453 * make all RNR NAKs visible so we know that RPC-over-RDMA
454 * flow control is working correctly (no NAKs should be seen).
456 ep->re_remote_cma.flow_control = 0;
457 ep->re_remote_cma.rnr_retry_count = 0;
459 ep->re_pd = ib_alloc_pd(device, 0);
460 if (IS_ERR(ep->re_pd)) {
461 rc = PTR_ERR(ep->re_pd);
465 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
479 * rpcrdma_xprt_connect - Connect an unconnected transport
480 * @r_xprt: controlling transport instance
482 * Returns 0 on success or a negative errno.
484 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
486 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
487 struct rpcrdma_ep *ep;
490 rc = rpcrdma_ep_create(r_xprt);
495 xprt_clear_connected(xprt);
496 rpcrdma_reset_cwnd(r_xprt);
498 /* Bump the ep's reference count while there are
499 * outstanding Receives.
502 rpcrdma_post_recvs(r_xprt, 1, true);
504 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
508 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
509 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
510 wait_event_interruptible(ep->re_connect_wait,
511 ep->re_connect_status != 0);
512 if (ep->re_connect_status <= 0) {
513 rc = ep->re_connect_status;
517 rc = rpcrdma_sendctxs_create(r_xprt);
523 rc = rpcrdma_reqs_setup(r_xprt);
528 rpcrdma_mrs_create(r_xprt);
529 frwr_wp_create(r_xprt);
532 trace_xprtrdma_connect(r_xprt, rc);
537 * rpcrdma_xprt_disconnect - Disconnect underlying transport
538 * @r_xprt: controlling transport instance
540 * Caller serializes. Either the transport send lock is held,
541 * or we're being called to destroy the transport.
543 * On return, @r_xprt is completely divested of all hardware
544 * resources and prepared for the next ->connect operation.
546 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
548 struct rpcrdma_ep *ep = r_xprt->rx_ep;
549 struct rdma_cm_id *id;
556 rc = rdma_disconnect(id);
557 trace_xprtrdma_disconnect(r_xprt, rc);
559 rpcrdma_xprt_drain(r_xprt);
560 rpcrdma_reps_unmap(r_xprt);
561 rpcrdma_reqs_reset(r_xprt);
562 rpcrdma_mrs_destroy(r_xprt);
563 rpcrdma_sendctxs_destroy(r_xprt);
565 if (rpcrdma_ep_put(ep))
568 r_xprt->rx_ep = NULL;
571 /* Fixed-size circular FIFO queue. This implementation is wait-free and
574 * Consumer is the code path that posts Sends. This path dequeues a
575 * sendctx for use by a Send operation. Multiple consumer threads
576 * are serialized by the RPC transport lock, which allows only one
577 * ->send_request call at a time.
579 * Producer is the code path that handles Send completions. This path
580 * enqueues a sendctx that has been completed. Multiple producer
581 * threads are serialized by the ib_poll_cq() function.
584 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
585 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
588 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
590 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
593 if (!buf->rb_sc_ctxs)
595 for (i = 0; i <= buf->rb_sc_last; i++)
596 kfree(buf->rb_sc_ctxs[i]);
597 kfree(buf->rb_sc_ctxs);
598 buf->rb_sc_ctxs = NULL;
601 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
603 struct rpcrdma_sendctx *sc;
605 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
610 sc->sc_cqe.done = rpcrdma_wc_send;
611 sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
612 sc->sc_cid.ci_completion_id =
613 atomic_inc_return(&ep->re_completion_ids);
617 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
619 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
620 struct rpcrdma_sendctx *sc;
623 /* Maximum number of concurrent outstanding Send WRs. Capping
624 * the circular queue size stops Send Queue overflow by causing
625 * the ->send_request call to fail temporarily before too many
628 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
629 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
630 if (!buf->rb_sc_ctxs)
633 buf->rb_sc_last = i - 1;
634 for (i = 0; i <= buf->rb_sc_last; i++) {
635 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
639 buf->rb_sc_ctxs[i] = sc;
647 /* The sendctx queue is not guaranteed to have a size that is a
648 * power of two, thus the helpers in circ_buf.h cannot be used.
649 * The other option is to use modulus (%), which can be expensive.
651 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
654 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
658 * rpcrdma_sendctx_get_locked - Acquire a send context
659 * @r_xprt: controlling transport instance
661 * Returns pointer to a free send completion context; or NULL if
662 * the queue is empty.
664 * Usage: Called to acquire an SGE array before preparing a Send WR.
666 * The caller serializes calls to this function (per transport), and
667 * provides an effective memory barrier that flushes the new value
670 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
672 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
673 struct rpcrdma_sendctx *sc;
674 unsigned long next_head;
676 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
678 if (next_head == READ_ONCE(buf->rb_sc_tail))
681 /* ORDER: item must be accessed _before_ head is updated */
682 sc = buf->rb_sc_ctxs[next_head];
684 /* Releasing the lock in the caller acts as a memory
685 * barrier that flushes rb_sc_head.
687 buf->rb_sc_head = next_head;
692 /* The queue is "empty" if there have not been enough Send
693 * completions recently. This is a sign the Send Queue is
694 * backing up. Cause the caller to pause and try again.
696 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
697 r_xprt->rx_stats.empty_sendctx_q++;
702 * rpcrdma_sendctx_put_locked - Release a send context
703 * @r_xprt: controlling transport instance
704 * @sc: send context to release
706 * Usage: Called from Send completion to return a sendctxt
709 * The caller serializes calls to this function (per transport).
711 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
712 struct rpcrdma_sendctx *sc)
714 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
715 unsigned long next_tail;
717 /* Unmap SGEs of previously completed but unsignaled
718 * Sends by walking up the queue until @sc is found.
720 next_tail = buf->rb_sc_tail;
722 next_tail = rpcrdma_sendctx_next(buf, next_tail);
724 /* ORDER: item must be accessed _before_ tail is updated */
725 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
727 } while (buf->rb_sc_ctxs[next_tail] != sc);
729 /* Paired with READ_ONCE */
730 smp_store_release(&buf->rb_sc_tail, next_tail);
732 xprt_write_space(&r_xprt->rx_xprt);
736 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
738 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
739 struct rpcrdma_ep *ep = r_xprt->rx_ep;
742 for (count = 0; count < ep->re_max_rdma_segs; count++) {
743 struct rpcrdma_mr *mr;
746 mr = kzalloc(sizeof(*mr), GFP_NOFS);
750 rc = frwr_mr_init(r_xprt, mr);
756 spin_lock(&buf->rb_lock);
757 rpcrdma_mr_push(mr, &buf->rb_mrs);
758 list_add(&mr->mr_all, &buf->rb_all_mrs);
759 spin_unlock(&buf->rb_lock);
762 r_xprt->rx_stats.mrs_allocated += count;
763 trace_xprtrdma_createmrs(r_xprt, count);
767 rpcrdma_mr_refresh_worker(struct work_struct *work)
769 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
771 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
774 rpcrdma_mrs_create(r_xprt);
775 xprt_write_space(&r_xprt->rx_xprt);
779 * rpcrdma_mrs_refresh - Wake the MR refresh worker
780 * @r_xprt: controlling transport instance
783 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
785 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
786 struct rpcrdma_ep *ep = r_xprt->rx_ep;
788 /* If there is no underlying connection, it's no use
789 * to wake the refresh worker.
791 if (ep->re_connect_status == 1) {
792 /* The work is scheduled on a WQ_MEM_RECLAIM
793 * workqueue in order to prevent MR allocation
794 * from recursing into NFS during direct reclaim.
796 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
801 * rpcrdma_req_create - Allocate an rpcrdma_req object
802 * @r_xprt: controlling r_xprt
803 * @size: initial size, in bytes, of send and receive buffers
804 * @flags: GFP flags passed to memory allocators
806 * Returns an allocated and fully initialized rpcrdma_req or NULL.
808 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
811 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
812 struct rpcrdma_req *req;
814 req = kzalloc(sizeof(*req), flags);
818 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
819 if (!req->rl_sendbuf)
822 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
823 if (!req->rl_recvbuf)
826 INIT_LIST_HEAD(&req->rl_free_mrs);
827 INIT_LIST_HEAD(&req->rl_registered);
828 spin_lock(&buffer->rb_lock);
829 list_add(&req->rl_all, &buffer->rb_allreqs);
830 spin_unlock(&buffer->rb_lock);
834 kfree(req->rl_sendbuf);
842 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
843 * @r_xprt: controlling transport instance
844 * @req: rpcrdma_req object to set up
846 * Returns zero on success, and a negative errno on failure.
848 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
850 struct rpcrdma_regbuf *rb;
853 /* Compute maximum header buffer size in bytes */
854 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
855 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
856 maxhdrsize *= sizeof(__be32);
857 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
858 DMA_TO_DEVICE, GFP_KERNEL);
862 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
865 req->rl_rdmabuf = rb;
866 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
870 rpcrdma_regbuf_free(rb);
875 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
876 * and thus can be walked without holding rb_lock. Eg. the
877 * caller is holding the transport send lock to exclude
878 * device removal or disconnection.
880 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
882 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
883 struct rpcrdma_req *req;
886 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
887 rc = rpcrdma_req_setup(r_xprt, req);
894 static void rpcrdma_req_reset(struct rpcrdma_req *req)
896 /* Credits are valid for only one connection */
897 req->rl_slot.rq_cong = 0;
899 rpcrdma_regbuf_free(req->rl_rdmabuf);
900 req->rl_rdmabuf = NULL;
902 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
903 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
908 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
909 * and thus can be walked without holding rb_lock. Eg. the
910 * caller is holding the transport send lock to exclude
911 * device removal or disconnection.
913 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
915 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
916 struct rpcrdma_req *req;
918 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
919 rpcrdma_req_reset(req);
923 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
926 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
927 struct rpcrdma_rep *rep;
929 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
933 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
934 DMA_FROM_DEVICE, GFP_KERNEL);
935 if (!rep->rr_rdmabuf)
938 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
939 goto out_free_regbuf;
941 rep->rr_cid.ci_completion_id =
942 atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
944 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
945 rdmab_length(rep->rr_rdmabuf));
946 rep->rr_cqe.done = rpcrdma_wc_receive;
947 rep->rr_rxprt = r_xprt;
948 rep->rr_recv_wr.next = NULL;
949 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
950 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
951 rep->rr_recv_wr.num_sge = 1;
954 spin_lock(&buf->rb_lock);
955 list_add(&rep->rr_all, &buf->rb_all_reps);
956 spin_unlock(&buf->rb_lock);
960 rpcrdma_regbuf_free(rep->rr_rdmabuf);
967 static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
969 rpcrdma_regbuf_free(rep->rr_rdmabuf);
973 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
975 struct rpcrdma_buffer *buf = &rep->rr_rxprt->rx_buf;
977 spin_lock(&buf->rb_lock);
978 list_del(&rep->rr_all);
979 spin_unlock(&buf->rb_lock);
981 rpcrdma_rep_free(rep);
984 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
986 struct llist_node *node;
988 /* Calls to llist_del_first are required to be serialized */
989 node = llist_del_first(&buf->rb_free_reps);
992 return llist_entry(node, struct rpcrdma_rep, rr_node);
996 * rpcrdma_rep_put - Release rpcrdma_rep back to free list
998 * @rep: rep to release
1001 void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
1003 llist_add(&rep->rr_node, &buf->rb_free_reps);
1006 /* Caller must ensure the QP is quiescent (RQ is drained) before
1007 * invoking this function, to guarantee rb_all_reps is not
1010 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1012 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1013 struct rpcrdma_rep *rep;
1015 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1016 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1017 rep->rr_temp = true; /* Mark this rep for destruction */
1021 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1023 struct rpcrdma_rep *rep;
1025 spin_lock(&buf->rb_lock);
1026 while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1029 list_del(&rep->rr_all);
1030 spin_unlock(&buf->rb_lock);
1032 rpcrdma_rep_free(rep);
1034 spin_lock(&buf->rb_lock);
1036 spin_unlock(&buf->rb_lock);
1040 * rpcrdma_buffer_create - Create initial set of req/rep objects
1041 * @r_xprt: transport instance to (re)initialize
1043 * Returns zero on success, otherwise a negative errno.
1045 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1047 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1050 buf->rb_bc_srv_max_requests = 0;
1051 spin_lock_init(&buf->rb_lock);
1052 INIT_LIST_HEAD(&buf->rb_mrs);
1053 INIT_LIST_HEAD(&buf->rb_all_mrs);
1054 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1056 INIT_LIST_HEAD(&buf->rb_send_bufs);
1057 INIT_LIST_HEAD(&buf->rb_allreqs);
1058 INIT_LIST_HEAD(&buf->rb_all_reps);
1061 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1062 struct rpcrdma_req *req;
1064 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1068 list_add(&req->rl_list, &buf->rb_send_bufs);
1071 init_llist_head(&buf->rb_free_reps);
1075 rpcrdma_buffer_destroy(buf);
1080 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1081 * @req: unused object to be destroyed
1083 * Relies on caller holding the transport send lock to protect
1084 * removing req->rl_all from buf->rb_all_reqs safely.
1086 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1088 struct rpcrdma_mr *mr;
1090 list_del(&req->rl_all);
1092 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1093 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1095 spin_lock(&buf->rb_lock);
1096 list_del(&mr->mr_all);
1097 spin_unlock(&buf->rb_lock);
1099 frwr_mr_release(mr);
1102 rpcrdma_regbuf_free(req->rl_recvbuf);
1103 rpcrdma_regbuf_free(req->rl_sendbuf);
1104 rpcrdma_regbuf_free(req->rl_rdmabuf);
1109 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1110 * @r_xprt: controlling transport instance
1112 * Relies on caller holding the transport send lock to protect
1113 * removing mr->mr_list from req->rl_free_mrs safely.
1115 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1117 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1118 struct rpcrdma_mr *mr;
1120 cancel_work_sync(&buf->rb_refresh_worker);
1122 spin_lock(&buf->rb_lock);
1123 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1126 list_del(&mr->mr_list);
1127 list_del(&mr->mr_all);
1128 spin_unlock(&buf->rb_lock);
1130 frwr_mr_release(mr);
1132 spin_lock(&buf->rb_lock);
1134 spin_unlock(&buf->rb_lock);
1138 * rpcrdma_buffer_destroy - Release all hw resources
1139 * @buf: root control block for resources
1141 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1142 * - No more Send or Receive completions can occur
1143 * - All MRs, reps, and reqs are returned to their free lists
1146 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1148 rpcrdma_reps_destroy(buf);
1150 while (!list_empty(&buf->rb_send_bufs)) {
1151 struct rpcrdma_req *req;
1153 req = list_first_entry(&buf->rb_send_bufs,
1154 struct rpcrdma_req, rl_list);
1155 list_del(&req->rl_list);
1156 rpcrdma_req_destroy(req);
1161 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1162 * @r_xprt: controlling transport
1164 * Returns an initialized rpcrdma_mr or NULL if no free
1165 * rpcrdma_mr objects are available.
1168 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1170 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1171 struct rpcrdma_mr *mr;
1173 spin_lock(&buf->rb_lock);
1174 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1175 spin_unlock(&buf->rb_lock);
1180 * rpcrdma_reply_put - Put reply buffers back into pool
1181 * @buffers: buffer pool
1182 * @req: object to return
1185 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1187 if (req->rl_reply) {
1188 rpcrdma_rep_put(buffers, req->rl_reply);
1189 req->rl_reply = NULL;
1194 * rpcrdma_buffer_get - Get a request buffer
1195 * @buffers: Buffer pool from which to obtain a buffer
1197 * Returns a fresh rpcrdma_req, or NULL if none are available.
1199 struct rpcrdma_req *
1200 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1202 struct rpcrdma_req *req;
1204 spin_lock(&buffers->rb_lock);
1205 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1206 struct rpcrdma_req, rl_list);
1208 list_del_init(&req->rl_list);
1209 spin_unlock(&buffers->rb_lock);
1214 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1215 * @buffers: buffer pool
1216 * @req: object to return
1219 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1221 rpcrdma_reply_put(buffers, req);
1223 spin_lock(&buffers->rb_lock);
1224 list_add(&req->rl_list, &buffers->rb_send_bufs);
1225 spin_unlock(&buffers->rb_lock);
1228 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1230 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1231 * receiving the payload of RDMA RECV operations. During Long Calls
1232 * or Replies they may be registered externally via frwr_map.
1234 static struct rpcrdma_regbuf *
1235 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1238 struct rpcrdma_regbuf *rb;
1240 rb = kmalloc(sizeof(*rb), flags);
1243 rb->rg_data = kmalloc(size, flags);
1249 rb->rg_device = NULL;
1250 rb->rg_direction = direction;
1251 rb->rg_iov.length = size;
1256 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1257 * @rb: regbuf to reallocate
1258 * @size: size of buffer to be allocated, in bytes
1261 * Returns true if reallocation was successful. If false is
1262 * returned, @rb is left untouched.
1264 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1268 buf = kmalloc(size, flags);
1272 rpcrdma_regbuf_dma_unmap(rb);
1276 rb->rg_iov.length = size;
1281 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1282 * @r_xprt: controlling transport instance
1283 * @rb: regbuf to be mapped
1285 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1287 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1288 struct rpcrdma_regbuf *rb)
1290 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1292 if (rb->rg_direction == DMA_NONE)
1295 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1296 rdmab_length(rb), rb->rg_direction);
1297 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1298 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1302 rb->rg_device = device;
1303 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1307 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1312 if (!rpcrdma_regbuf_is_mapped(rb))
1315 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1317 rb->rg_device = NULL;
1320 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1322 rpcrdma_regbuf_dma_unmap(rb);
1329 * rpcrdma_post_recvs - Refill the Receive Queue
1330 * @r_xprt: controlling transport instance
1331 * @needed: current credit grant
1332 * @temp: mark Receive buffers to be deleted after one use
1335 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1337 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1338 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1339 struct ib_recv_wr *wr, *bad_wr;
1340 struct rpcrdma_rep *rep;
1346 if (likely(ep->re_receive_count > needed))
1348 needed -= ep->re_receive_count;
1350 needed += RPCRDMA_MAX_RECV_BATCH;
1352 if (atomic_inc_return(&ep->re_receiving) > 1)
1355 /* fast path: all needed reps can be found on the free list */
1358 rep = rpcrdma_rep_get_locked(buf);
1359 if (rep && rep->rr_temp) {
1360 rpcrdma_rep_destroy(rep);
1364 rep = rpcrdma_rep_create(r_xprt, temp);
1368 rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1369 trace_xprtrdma_post_recv(rep);
1370 rep->rr_recv_wr.next = wr;
1371 wr = &rep->rr_recv_wr;
1378 rc = ib_post_recv(ep->re_id->qp, wr,
1379 (const struct ib_recv_wr **)&bad_wr);
1381 trace_xprtrdma_post_recvs_err(r_xprt, rc);
1382 for (wr = bad_wr; wr;) {
1383 struct rpcrdma_rep *rep;
1385 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1387 rpcrdma_rep_put(buf, rep);
1391 if (atomic_dec_return(&ep->re_receiving) > 0)
1392 complete(&ep->re_done);
1395 trace_xprtrdma_post_recvs(r_xprt, count);
1396 ep->re_receive_count += count;