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 /* Wait for rpcrdma_post_recvs() to leave its critical
107 if (atomic_inc_return(&ep->re_receiving) > 1)
108 wait_for_completion(&ep->re_done);
110 /* Flush Receives, then wait for deferred Reply work
115 /* Deferred Reply processing might have scheduled
116 * local invalidations.
124 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
125 * @event: details of the event
126 * @context: ep that owns QP where event occurred
128 * Called from the RDMA provider (device driver) possibly in an interrupt
129 * context. The QP is always destroyed before the ID, so the ID will be
130 * reliably available when this handler is invoked.
132 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
134 struct rpcrdma_ep *ep = context;
136 trace_xprtrdma_qp_event(ep, event);
139 /* Ensure xprt_force_disconnect() is invoked exactly once when a
140 * connection is closed or lost. (The important thing is it needs
141 * to be invoked "at least" once).
143 static void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
145 if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
146 xprt_force_disconnect(ep->re_xprt);
150 * rpcrdma_flush_disconnect - Disconnect on flushed completion
151 * @r_xprt: transport to disconnect
152 * @wc: work completion entry
154 * Must be called in process context.
156 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
158 if (wc->status != IB_WC_SUCCESS)
159 rpcrdma_force_disconnect(r_xprt->rx_ep);
163 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
164 * @cq: completion queue
165 * @wc: WCE for a completed Send WR
168 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
170 struct ib_cqe *cqe = wc->wr_cqe;
171 struct rpcrdma_sendctx *sc =
172 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
173 struct rpcrdma_xprt *r_xprt = cq->cq_context;
175 /* WARNING: Only wr_cqe and status are reliable at this point */
176 trace_xprtrdma_wc_send(wc, &sc->sc_cid);
177 rpcrdma_sendctx_put_locked(r_xprt, sc);
178 rpcrdma_flush_disconnect(r_xprt, wc);
182 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
183 * @cq: completion queue
184 * @wc: WCE for a completed Receive WR
187 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
189 struct ib_cqe *cqe = wc->wr_cqe;
190 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
192 struct rpcrdma_xprt *r_xprt = cq->cq_context;
194 /* WARNING: Only wr_cqe and status are reliable at this point */
195 trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
196 --r_xprt->rx_ep->re_receive_count;
197 if (wc->status != IB_WC_SUCCESS)
200 /* status == SUCCESS means all fields in wc are trustworthy */
201 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
202 rep->rr_wc_flags = wc->wc_flags;
203 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
205 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
206 rdmab_addr(rep->rr_rdmabuf),
207 wc->byte_len, DMA_FROM_DEVICE);
209 rpcrdma_reply_handler(rep);
213 rpcrdma_flush_disconnect(r_xprt, wc);
214 rpcrdma_rep_put(&r_xprt->rx_buf, rep);
217 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
218 struct rdma_conn_param *param)
220 const struct rpcrdma_connect_private *pmsg = param->private_data;
221 unsigned int rsize, wsize;
223 /* Default settings for RPC-over-RDMA Version One */
224 ep->re_implicit_roundup = xprt_rdma_pad_optimize;
225 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
226 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
229 pmsg->cp_magic == rpcrdma_cmp_magic &&
230 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
231 ep->re_implicit_roundup = true;
232 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
233 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
236 if (rsize < ep->re_inline_recv)
237 ep->re_inline_recv = rsize;
238 if (wsize < ep->re_inline_send)
239 ep->re_inline_send = wsize;
241 rpcrdma_set_max_header_sizes(ep);
245 * rpcrdma_cm_event_handler - Handle RDMA CM events
246 * @id: rdma_cm_id on which an event has occurred
247 * @event: details of the event
249 * Called with @id's mutex held. Returns 1 if caller should
250 * destroy @id, otherwise 0.
253 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
255 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
256 struct rpcrdma_ep *ep = id->context;
260 switch (event->event) {
261 case RDMA_CM_EVENT_ADDR_RESOLVED:
262 case RDMA_CM_EVENT_ROUTE_RESOLVED:
264 complete(&ep->re_done);
266 case RDMA_CM_EVENT_ADDR_ERROR:
267 ep->re_async_rc = -EPROTO;
268 complete(&ep->re_done);
270 case RDMA_CM_EVENT_ROUTE_ERROR:
271 ep->re_async_rc = -ENETUNREACH;
272 complete(&ep->re_done);
274 case RDMA_CM_EVENT_DEVICE_REMOVAL:
275 pr_info("rpcrdma: removing device %s for %pISpc\n",
276 ep->re_id->device->name, sap);
278 case RDMA_CM_EVENT_ADDR_CHANGE:
279 ep->re_connect_status = -ENODEV;
281 case RDMA_CM_EVENT_ESTABLISHED:
283 ep->re_connect_status = 1;
284 rpcrdma_update_cm_private(ep, &event->param.conn);
285 trace_xprtrdma_inline_thresh(ep);
286 wake_up_all(&ep->re_connect_wait);
288 case RDMA_CM_EVENT_CONNECT_ERROR:
289 ep->re_connect_status = -ENOTCONN;
290 goto wake_connect_worker;
291 case RDMA_CM_EVENT_UNREACHABLE:
292 ep->re_connect_status = -ENETUNREACH;
293 goto wake_connect_worker;
294 case RDMA_CM_EVENT_REJECTED:
295 dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
296 sap, rdma_reject_msg(id, event->status));
297 ep->re_connect_status = -ECONNREFUSED;
298 if (event->status == IB_CM_REJ_STALE_CONN)
299 ep->re_connect_status = -ENOTCONN;
301 wake_up_all(&ep->re_connect_wait);
303 case RDMA_CM_EVENT_DISCONNECTED:
304 ep->re_connect_status = -ECONNABORTED;
306 rpcrdma_force_disconnect(ep);
307 return rpcrdma_ep_put(ep);
312 dprintk("RPC: %s: %pISpc on %s/frwr: %s\n", __func__, sap,
313 ep->re_id->device->name, rdma_event_msg(event->event));
317 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
318 struct rpcrdma_ep *ep)
320 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
321 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
322 struct rdma_cm_id *id;
325 init_completion(&ep->re_done);
327 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
328 RDMA_PS_TCP, IB_QPT_RC);
332 ep->re_async_rc = -ETIMEDOUT;
333 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
334 RDMA_RESOLVE_TIMEOUT);
337 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
341 rc = ep->re_async_rc;
345 ep->re_async_rc = -ETIMEDOUT;
346 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
349 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
352 rc = ep->re_async_rc;
363 static void rpcrdma_ep_destroy(struct kref *kref)
365 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
368 rdma_destroy_qp(ep->re_id);
369 ep->re_id->qp = NULL;
372 if (ep->re_attr.recv_cq)
373 ib_free_cq(ep->re_attr.recv_cq);
374 ep->re_attr.recv_cq = NULL;
375 if (ep->re_attr.send_cq)
376 ib_free_cq(ep->re_attr.send_cq);
377 ep->re_attr.send_cq = NULL;
380 ib_dealloc_pd(ep->re_pd);
384 module_put(THIS_MODULE);
387 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
389 kref_get(&ep->re_kref);
393 * %0 if @ep still has a positive kref count, or
394 * %1 if @ep was destroyed successfully.
396 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
398 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
401 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
403 struct rpcrdma_connect_private *pmsg;
404 struct ib_device *device;
405 struct rdma_cm_id *id;
406 struct rpcrdma_ep *ep;
409 ep = kzalloc(sizeof(*ep), GFP_NOFS);
412 ep->re_xprt = &r_xprt->rx_xprt;
413 kref_init(&ep->re_kref);
415 id = rpcrdma_create_id(r_xprt, ep);
420 __module_get(THIS_MODULE);
423 reinit_completion(&ep->re_done);
425 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
426 ep->re_inline_send = xprt_rdma_max_inline_write;
427 ep->re_inline_recv = xprt_rdma_max_inline_read;
428 rc = frwr_query_device(ep, device);
432 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
434 ep->re_attr.event_handler = rpcrdma_qp_event_handler;
435 ep->re_attr.qp_context = ep;
436 ep->re_attr.srq = NULL;
437 ep->re_attr.cap.max_inline_data = 0;
438 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
439 ep->re_attr.qp_type = IB_QPT_RC;
440 ep->re_attr.port_num = ~0;
442 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
443 "iovs: send %d recv %d\n",
445 ep->re_attr.cap.max_send_wr,
446 ep->re_attr.cap.max_recv_wr,
447 ep->re_attr.cap.max_send_sge,
448 ep->re_attr.cap.max_recv_sge);
450 ep->re_send_batch = ep->re_max_requests >> 3;
451 ep->re_send_count = ep->re_send_batch;
452 init_waitqueue_head(&ep->re_connect_wait);
454 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
455 ep->re_attr.cap.max_send_wr,
457 if (IS_ERR(ep->re_attr.send_cq)) {
458 rc = PTR_ERR(ep->re_attr.send_cq);
462 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
463 ep->re_attr.cap.max_recv_wr,
465 if (IS_ERR(ep->re_attr.recv_cq)) {
466 rc = PTR_ERR(ep->re_attr.recv_cq);
469 ep->re_receive_count = 0;
471 /* Initialize cma parameters */
472 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
474 /* Prepare RDMA-CM private message */
475 pmsg = &ep->re_cm_private;
476 pmsg->cp_magic = rpcrdma_cmp_magic;
477 pmsg->cp_version = RPCRDMA_CMP_VERSION;
478 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
479 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
480 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
481 ep->re_remote_cma.private_data = pmsg;
482 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
484 /* Client offers RDMA Read but does not initiate */
485 ep->re_remote_cma.initiator_depth = 0;
486 ep->re_remote_cma.responder_resources =
487 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
489 /* Limit transport retries so client can detect server
490 * GID changes quickly. RPC layer handles re-establishing
491 * transport connection and retransmission.
493 ep->re_remote_cma.retry_count = 6;
495 /* RPC-over-RDMA handles its own flow control. In addition,
496 * make all RNR NAKs visible so we know that RPC-over-RDMA
497 * flow control is working correctly (no NAKs should be seen).
499 ep->re_remote_cma.flow_control = 0;
500 ep->re_remote_cma.rnr_retry_count = 0;
502 ep->re_pd = ib_alloc_pd(device, 0);
503 if (IS_ERR(ep->re_pd)) {
504 rc = PTR_ERR(ep->re_pd);
508 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
522 * rpcrdma_xprt_connect - Connect an unconnected transport
523 * @r_xprt: controlling transport instance
525 * Returns 0 on success or a negative errno.
527 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
529 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
530 struct rpcrdma_ep *ep;
533 rc = rpcrdma_ep_create(r_xprt);
538 xprt_clear_connected(xprt);
539 rpcrdma_reset_cwnd(r_xprt);
541 /* Bump the ep's reference count while there are
542 * outstanding Receives.
545 rpcrdma_post_recvs(r_xprt, 1, true);
547 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
551 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
552 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
553 wait_event_interruptible(ep->re_connect_wait,
554 ep->re_connect_status != 0);
555 if (ep->re_connect_status <= 0) {
556 rc = ep->re_connect_status;
560 rc = rpcrdma_sendctxs_create(r_xprt);
566 rc = rpcrdma_reqs_setup(r_xprt);
571 rpcrdma_mrs_create(r_xprt);
574 trace_xprtrdma_connect(r_xprt, rc);
579 * rpcrdma_xprt_disconnect - Disconnect underlying transport
580 * @r_xprt: controlling transport instance
582 * Caller serializes. Either the transport send lock is held,
583 * or we're being called to destroy the transport.
585 * On return, @r_xprt is completely divested of all hardware
586 * resources and prepared for the next ->connect operation.
588 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
590 struct rpcrdma_ep *ep = r_xprt->rx_ep;
591 struct rdma_cm_id *id;
598 rc = rdma_disconnect(id);
599 trace_xprtrdma_disconnect(r_xprt, rc);
601 rpcrdma_xprt_drain(r_xprt);
602 rpcrdma_reps_unmap(r_xprt);
603 rpcrdma_reqs_reset(r_xprt);
604 rpcrdma_mrs_destroy(r_xprt);
605 rpcrdma_sendctxs_destroy(r_xprt);
607 if (rpcrdma_ep_put(ep))
610 r_xprt->rx_ep = NULL;
613 /* Fixed-size circular FIFO queue. This implementation is wait-free and
616 * Consumer is the code path that posts Sends. This path dequeues a
617 * sendctx for use by a Send operation. Multiple consumer threads
618 * are serialized by the RPC transport lock, which allows only one
619 * ->send_request call at a time.
621 * Producer is the code path that handles Send completions. This path
622 * enqueues a sendctx that has been completed. Multiple producer
623 * threads are serialized by the ib_poll_cq() function.
626 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
627 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
630 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
632 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
635 if (!buf->rb_sc_ctxs)
637 for (i = 0; i <= buf->rb_sc_last; i++)
638 kfree(buf->rb_sc_ctxs[i]);
639 kfree(buf->rb_sc_ctxs);
640 buf->rb_sc_ctxs = NULL;
643 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
645 struct rpcrdma_sendctx *sc;
647 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
652 sc->sc_cqe.done = rpcrdma_wc_send;
653 sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
654 sc->sc_cid.ci_completion_id =
655 atomic_inc_return(&ep->re_completion_ids);
659 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
661 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
662 struct rpcrdma_sendctx *sc;
665 /* Maximum number of concurrent outstanding Send WRs. Capping
666 * the circular queue size stops Send Queue overflow by causing
667 * the ->send_request call to fail temporarily before too many
670 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
671 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
672 if (!buf->rb_sc_ctxs)
675 buf->rb_sc_last = i - 1;
676 for (i = 0; i <= buf->rb_sc_last; i++) {
677 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
681 buf->rb_sc_ctxs[i] = sc;
689 /* The sendctx queue is not guaranteed to have a size that is a
690 * power of two, thus the helpers in circ_buf.h cannot be used.
691 * The other option is to use modulus (%), which can be expensive.
693 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
696 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
700 * rpcrdma_sendctx_get_locked - Acquire a send context
701 * @r_xprt: controlling transport instance
703 * Returns pointer to a free send completion context; or NULL if
704 * the queue is empty.
706 * Usage: Called to acquire an SGE array before preparing a Send WR.
708 * The caller serializes calls to this function (per transport), and
709 * provides an effective memory barrier that flushes the new value
712 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
714 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
715 struct rpcrdma_sendctx *sc;
716 unsigned long next_head;
718 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
720 if (next_head == READ_ONCE(buf->rb_sc_tail))
723 /* ORDER: item must be accessed _before_ head is updated */
724 sc = buf->rb_sc_ctxs[next_head];
726 /* Releasing the lock in the caller acts as a memory
727 * barrier that flushes rb_sc_head.
729 buf->rb_sc_head = next_head;
734 /* The queue is "empty" if there have not been enough Send
735 * completions recently. This is a sign the Send Queue is
736 * backing up. Cause the caller to pause and try again.
738 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
739 r_xprt->rx_stats.empty_sendctx_q++;
744 * rpcrdma_sendctx_put_locked - Release a send context
745 * @r_xprt: controlling transport instance
746 * @sc: send context to release
748 * Usage: Called from Send completion to return a sendctxt
751 * The caller serializes calls to this function (per transport).
753 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
754 struct rpcrdma_sendctx *sc)
756 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
757 unsigned long next_tail;
759 /* Unmap SGEs of previously completed but unsignaled
760 * Sends by walking up the queue until @sc is found.
762 next_tail = buf->rb_sc_tail;
764 next_tail = rpcrdma_sendctx_next(buf, next_tail);
766 /* ORDER: item must be accessed _before_ tail is updated */
767 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
769 } while (buf->rb_sc_ctxs[next_tail] != sc);
771 /* Paired with READ_ONCE */
772 smp_store_release(&buf->rb_sc_tail, next_tail);
774 xprt_write_space(&r_xprt->rx_xprt);
778 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
780 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
781 struct rpcrdma_ep *ep = r_xprt->rx_ep;
784 for (count = 0; count < ep->re_max_rdma_segs; count++) {
785 struct rpcrdma_mr *mr;
788 mr = kzalloc(sizeof(*mr), GFP_NOFS);
792 rc = frwr_mr_init(r_xprt, mr);
798 spin_lock(&buf->rb_lock);
799 rpcrdma_mr_push(mr, &buf->rb_mrs);
800 list_add(&mr->mr_all, &buf->rb_all_mrs);
801 spin_unlock(&buf->rb_lock);
804 r_xprt->rx_stats.mrs_allocated += count;
805 trace_xprtrdma_createmrs(r_xprt, count);
809 rpcrdma_mr_refresh_worker(struct work_struct *work)
811 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
813 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
816 rpcrdma_mrs_create(r_xprt);
817 xprt_write_space(&r_xprt->rx_xprt);
821 * rpcrdma_mrs_refresh - Wake the MR refresh worker
822 * @r_xprt: controlling transport instance
825 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
827 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
828 struct rpcrdma_ep *ep = r_xprt->rx_ep;
830 /* If there is no underlying connection, it's no use
831 * to wake the refresh worker.
833 if (ep->re_connect_status == 1) {
834 /* The work is scheduled on a WQ_MEM_RECLAIM
835 * workqueue in order to prevent MR allocation
836 * from recursing into NFS during direct reclaim.
838 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
843 * rpcrdma_req_create - Allocate an rpcrdma_req object
844 * @r_xprt: controlling r_xprt
845 * @size: initial size, in bytes, of send and receive buffers
846 * @flags: GFP flags passed to memory allocators
848 * Returns an allocated and fully initialized rpcrdma_req or NULL.
850 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
853 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
854 struct rpcrdma_req *req;
856 req = kzalloc(sizeof(*req), flags);
860 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
861 if (!req->rl_sendbuf)
864 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
865 if (!req->rl_recvbuf)
868 INIT_LIST_HEAD(&req->rl_free_mrs);
869 INIT_LIST_HEAD(&req->rl_registered);
870 spin_lock(&buffer->rb_lock);
871 list_add(&req->rl_all, &buffer->rb_allreqs);
872 spin_unlock(&buffer->rb_lock);
876 kfree(req->rl_sendbuf);
884 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
885 * @r_xprt: controlling transport instance
886 * @req: rpcrdma_req object to set up
888 * Returns zero on success, and a negative errno on failure.
890 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
892 struct rpcrdma_regbuf *rb;
895 /* Compute maximum header buffer size in bytes */
896 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
897 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
898 maxhdrsize *= sizeof(__be32);
899 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
900 DMA_TO_DEVICE, GFP_KERNEL);
904 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
907 req->rl_rdmabuf = rb;
908 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
912 rpcrdma_regbuf_free(rb);
917 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
918 * and thus can be walked without holding rb_lock. Eg. the
919 * caller is holding the transport send lock to exclude
920 * device removal or disconnection.
922 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
924 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
925 struct rpcrdma_req *req;
928 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
929 rc = rpcrdma_req_setup(r_xprt, req);
936 static void rpcrdma_req_reset(struct rpcrdma_req *req)
938 /* Credits are valid for only one connection */
939 req->rl_slot.rq_cong = 0;
941 rpcrdma_regbuf_free(req->rl_rdmabuf);
942 req->rl_rdmabuf = NULL;
944 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
945 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
950 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
951 * and thus can be walked without holding rb_lock. Eg. the
952 * caller is holding the transport send lock to exclude
953 * device removal or disconnection.
955 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
957 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
958 struct rpcrdma_req *req;
960 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
961 rpcrdma_req_reset(req);
965 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
968 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
969 struct rpcrdma_rep *rep;
971 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
975 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
976 DMA_FROM_DEVICE, GFP_KERNEL);
977 if (!rep->rr_rdmabuf)
980 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
981 goto out_free_regbuf;
983 rep->rr_cid.ci_completion_id =
984 atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
986 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
987 rdmab_length(rep->rr_rdmabuf));
988 rep->rr_cqe.done = rpcrdma_wc_receive;
989 rep->rr_rxprt = r_xprt;
990 rep->rr_recv_wr.next = NULL;
991 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
992 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
993 rep->rr_recv_wr.num_sge = 1;
996 spin_lock(&buf->rb_lock);
997 list_add(&rep->rr_all, &buf->rb_all_reps);
998 spin_unlock(&buf->rb_lock);
1002 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1009 static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
1011 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1015 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1017 struct rpcrdma_buffer *buf = &rep->rr_rxprt->rx_buf;
1019 spin_lock(&buf->rb_lock);
1020 list_del(&rep->rr_all);
1021 spin_unlock(&buf->rb_lock);
1023 rpcrdma_rep_free(rep);
1026 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1028 struct llist_node *node;
1030 /* Calls to llist_del_first are required to be serialized */
1031 node = llist_del_first(&buf->rb_free_reps);
1034 return llist_entry(node, struct rpcrdma_rep, rr_node);
1038 * rpcrdma_rep_put - Release rpcrdma_rep back to free list
1040 * @rep: rep to release
1043 void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
1045 llist_add(&rep->rr_node, &buf->rb_free_reps);
1048 /* Caller must ensure the QP is quiescent (RQ is drained) before
1049 * invoking this function, to guarantee rb_all_reps is not
1052 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1054 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1055 struct rpcrdma_rep *rep;
1057 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1058 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1059 rep->rr_temp = true; /* Mark this rep for destruction */
1063 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1065 struct rpcrdma_rep *rep;
1067 spin_lock(&buf->rb_lock);
1068 while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1071 list_del(&rep->rr_all);
1072 spin_unlock(&buf->rb_lock);
1074 rpcrdma_rep_free(rep);
1076 spin_lock(&buf->rb_lock);
1078 spin_unlock(&buf->rb_lock);
1082 * rpcrdma_buffer_create - Create initial set of req/rep objects
1083 * @r_xprt: transport instance to (re)initialize
1085 * Returns zero on success, otherwise a negative errno.
1087 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1089 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1092 buf->rb_bc_srv_max_requests = 0;
1093 spin_lock_init(&buf->rb_lock);
1094 INIT_LIST_HEAD(&buf->rb_mrs);
1095 INIT_LIST_HEAD(&buf->rb_all_mrs);
1096 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1098 INIT_LIST_HEAD(&buf->rb_send_bufs);
1099 INIT_LIST_HEAD(&buf->rb_allreqs);
1100 INIT_LIST_HEAD(&buf->rb_all_reps);
1103 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1104 struct rpcrdma_req *req;
1106 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1110 list_add(&req->rl_list, &buf->rb_send_bufs);
1113 init_llist_head(&buf->rb_free_reps);
1117 rpcrdma_buffer_destroy(buf);
1122 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1123 * @req: unused object to be destroyed
1125 * Relies on caller holding the transport send lock to protect
1126 * removing req->rl_all from buf->rb_all_reqs safely.
1128 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1130 struct rpcrdma_mr *mr;
1132 list_del(&req->rl_all);
1134 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1135 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1137 spin_lock(&buf->rb_lock);
1138 list_del(&mr->mr_all);
1139 spin_unlock(&buf->rb_lock);
1141 frwr_mr_release(mr);
1144 rpcrdma_regbuf_free(req->rl_recvbuf);
1145 rpcrdma_regbuf_free(req->rl_sendbuf);
1146 rpcrdma_regbuf_free(req->rl_rdmabuf);
1151 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1152 * @r_xprt: controlling transport instance
1154 * Relies on caller holding the transport send lock to protect
1155 * removing mr->mr_list from req->rl_free_mrs safely.
1157 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1159 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1160 struct rpcrdma_mr *mr;
1162 cancel_work_sync(&buf->rb_refresh_worker);
1164 spin_lock(&buf->rb_lock);
1165 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1168 list_del(&mr->mr_list);
1169 list_del(&mr->mr_all);
1170 spin_unlock(&buf->rb_lock);
1172 frwr_mr_release(mr);
1174 spin_lock(&buf->rb_lock);
1176 spin_unlock(&buf->rb_lock);
1180 * rpcrdma_buffer_destroy - Release all hw resources
1181 * @buf: root control block for resources
1183 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1184 * - No more Send or Receive completions can occur
1185 * - All MRs, reps, and reqs are returned to their free lists
1188 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1190 rpcrdma_reps_destroy(buf);
1192 while (!list_empty(&buf->rb_send_bufs)) {
1193 struct rpcrdma_req *req;
1195 req = list_first_entry(&buf->rb_send_bufs,
1196 struct rpcrdma_req, rl_list);
1197 list_del(&req->rl_list);
1198 rpcrdma_req_destroy(req);
1203 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1204 * @r_xprt: controlling transport
1206 * Returns an initialized rpcrdma_mr or NULL if no free
1207 * rpcrdma_mr objects are available.
1210 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1212 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1213 struct rpcrdma_mr *mr;
1215 spin_lock(&buf->rb_lock);
1216 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1217 spin_unlock(&buf->rb_lock);
1222 * rpcrdma_buffer_get - Get a request buffer
1223 * @buffers: Buffer pool from which to obtain a buffer
1225 * Returns a fresh rpcrdma_req, or NULL if none are available.
1227 struct rpcrdma_req *
1228 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1230 struct rpcrdma_req *req;
1232 spin_lock(&buffers->rb_lock);
1233 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1234 struct rpcrdma_req, rl_list);
1236 list_del_init(&req->rl_list);
1237 spin_unlock(&buffers->rb_lock);
1242 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1243 * @buffers: buffer pool
1244 * @req: object to return
1247 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1250 rpcrdma_rep_put(buffers, req->rl_reply);
1251 req->rl_reply = NULL;
1253 spin_lock(&buffers->rb_lock);
1254 list_add(&req->rl_list, &buffers->rb_send_bufs);
1255 spin_unlock(&buffers->rb_lock);
1258 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1260 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1261 * receiving the payload of RDMA RECV operations. During Long Calls
1262 * or Replies they may be registered externally via frwr_map.
1264 static struct rpcrdma_regbuf *
1265 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1268 struct rpcrdma_regbuf *rb;
1270 rb = kmalloc(sizeof(*rb), flags);
1273 rb->rg_data = kmalloc(size, flags);
1279 rb->rg_device = NULL;
1280 rb->rg_direction = direction;
1281 rb->rg_iov.length = size;
1286 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1287 * @rb: regbuf to reallocate
1288 * @size: size of buffer to be allocated, in bytes
1291 * Returns true if reallocation was successful. If false is
1292 * returned, @rb is left untouched.
1294 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1298 buf = kmalloc(size, flags);
1302 rpcrdma_regbuf_dma_unmap(rb);
1306 rb->rg_iov.length = size;
1311 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1312 * @r_xprt: controlling transport instance
1313 * @rb: regbuf to be mapped
1315 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1317 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1318 struct rpcrdma_regbuf *rb)
1320 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1322 if (rb->rg_direction == DMA_NONE)
1325 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1326 rdmab_length(rb), rb->rg_direction);
1327 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1328 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1332 rb->rg_device = device;
1333 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1337 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1342 if (!rpcrdma_regbuf_is_mapped(rb))
1345 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1347 rb->rg_device = NULL;
1350 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1352 rpcrdma_regbuf_dma_unmap(rb);
1359 * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1360 * @r_xprt: controlling transport instance
1361 * @req: rpcrdma_req containing the Send WR to post
1363 * Returns 0 if the post was successful, otherwise -ENOTCONN
1366 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1368 if (frwr_send(r_xprt, req))
1374 * rpcrdma_post_recvs - Refill the Receive Queue
1375 * @r_xprt: controlling transport instance
1376 * @needed: current credit grant
1377 * @temp: mark Receive buffers to be deleted after one use
1380 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1382 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1383 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1384 struct ib_recv_wr *wr, *bad_wr;
1385 struct rpcrdma_rep *rep;
1391 if (likely(ep->re_receive_count > needed))
1393 needed -= ep->re_receive_count;
1395 needed += RPCRDMA_MAX_RECV_BATCH;
1397 if (atomic_inc_return(&ep->re_receiving) > 1)
1400 /* fast path: all needed reps can be found on the free list */
1403 rep = rpcrdma_rep_get_locked(buf);
1404 if (rep && rep->rr_temp) {
1405 rpcrdma_rep_destroy(rep);
1409 rep = rpcrdma_rep_create(r_xprt, temp);
1413 rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1414 trace_xprtrdma_post_recv(rep);
1415 rep->rr_recv_wr.next = wr;
1416 wr = &rep->rr_recv_wr;
1423 rc = ib_post_recv(ep->re_id->qp, wr,
1424 (const struct ib_recv_wr **)&bad_wr);
1425 if (atomic_dec_return(&ep->re_receiving) > 0)
1426 complete(&ep->re_done);
1429 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1431 for (wr = bad_wr; wr;) {
1432 struct rpcrdma_rep *rep;
1434 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1436 rpcrdma_rep_put(buf, rep);
1440 ep->re_receive_count += count;