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 int rpcrdma_ep_destroy(struct rpcrdma_ep *ep);
88 static struct rpcrdma_regbuf *
89 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
91 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
92 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
94 /* Wait for outstanding transport work to finish. ib_drain_qp
95 * handles the drains in the wrong order for us, so open code
98 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
100 struct rdma_cm_id *id = r_xprt->rx_ep->re_id;
102 /* Flush Receives, then wait for deferred Reply work
107 /* Deferred Reply processing might have scheduled
108 * local invalidations.
114 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
115 * @event: details of the event
116 * @context: ep that owns QP where event occurred
118 * Called from the RDMA provider (device driver) possibly in an interrupt
119 * context. The QP is always destroyed before the ID, so the ID will be
120 * reliably available when this handler is invoked.
122 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
124 struct rpcrdma_ep *ep = context;
126 trace_xprtrdma_qp_event(ep, event);
130 * rpcrdma_flush_disconnect - Disconnect on flushed completion
131 * @cq: completion queue
132 * @wc: work completion entry
134 * Must be called in process context.
136 void rpcrdma_flush_disconnect(struct ib_cq *cq, struct ib_wc *wc)
138 struct rpcrdma_xprt *r_xprt = cq->cq_context;
139 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
141 if (wc->status != IB_WC_SUCCESS &&
142 r_xprt->rx_ep->re_connect_status == 1) {
143 r_xprt->rx_ep->re_connect_status = -ECONNABORTED;
144 trace_xprtrdma_flush_dct(r_xprt, wc->status);
145 xprt_force_disconnect(xprt);
150 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
151 * @cq: completion queue
152 * @wc: WCE for a completed Send WR
155 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
157 struct ib_cqe *cqe = wc->wr_cqe;
158 struct rpcrdma_sendctx *sc =
159 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
161 /* WARNING: Only wr_cqe and status are reliable at this point */
162 trace_xprtrdma_wc_send(sc, wc);
163 rpcrdma_sendctx_put_locked((struct rpcrdma_xprt *)cq->cq_context, sc);
164 rpcrdma_flush_disconnect(cq, wc);
168 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
169 * @cq: completion queue
170 * @wc: WCE for a completed Receive WR
173 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
175 struct ib_cqe *cqe = wc->wr_cqe;
176 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
178 struct rpcrdma_xprt *r_xprt = cq->cq_context;
180 /* WARNING: Only wr_cqe and status are reliable at this point */
181 trace_xprtrdma_wc_receive(wc);
182 --r_xprt->rx_ep->re_receive_count;
183 if (wc->status != IB_WC_SUCCESS)
186 /* status == SUCCESS means all fields in wc are trustworthy */
187 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
188 rep->rr_wc_flags = wc->wc_flags;
189 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
191 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
192 rdmab_addr(rep->rr_rdmabuf),
193 wc->byte_len, DMA_FROM_DEVICE);
195 rpcrdma_reply_handler(rep);
199 rpcrdma_flush_disconnect(cq, wc);
200 rpcrdma_rep_destroy(rep);
203 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
204 struct rdma_conn_param *param)
206 const struct rpcrdma_connect_private *pmsg = param->private_data;
207 unsigned int rsize, wsize;
209 /* Default settings for RPC-over-RDMA Version One */
210 ep->re_implicit_roundup = xprt_rdma_pad_optimize;
211 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
212 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
215 pmsg->cp_magic == rpcrdma_cmp_magic &&
216 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
217 ep->re_implicit_roundup = true;
218 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
219 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
222 if (rsize < ep->re_inline_recv)
223 ep->re_inline_recv = rsize;
224 if (wsize < ep->re_inline_send)
225 ep->re_inline_send = wsize;
227 rpcrdma_set_max_header_sizes(ep);
231 * rpcrdma_cm_event_handler - Handle RDMA CM events
232 * @id: rdma_cm_id on which an event has occurred
233 * @event: details of the event
235 * Called with @id's mutex held. Returns 1 if caller should
236 * destroy @id, otherwise 0.
239 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
241 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
242 struct rpcrdma_ep *ep = id->context;
243 struct rpc_xprt *xprt = ep->re_xprt;
247 switch (event->event) {
248 case RDMA_CM_EVENT_ADDR_RESOLVED:
249 case RDMA_CM_EVENT_ROUTE_RESOLVED:
251 complete(&ep->re_done);
253 case RDMA_CM_EVENT_ADDR_ERROR:
254 ep->re_async_rc = -EPROTO;
255 complete(&ep->re_done);
257 case RDMA_CM_EVENT_ROUTE_ERROR:
258 ep->re_async_rc = -ENETUNREACH;
259 complete(&ep->re_done);
261 case RDMA_CM_EVENT_DEVICE_REMOVAL:
262 pr_info("rpcrdma: removing device %s for %pISpc\n",
263 ep->re_id->device->name, sap);
265 case RDMA_CM_EVENT_ADDR_CHANGE:
266 ep->re_connect_status = -ENODEV;
267 xprt_force_disconnect(xprt);
269 case RDMA_CM_EVENT_ESTABLISHED:
270 kref_get(&ep->re_kref);
271 ep->re_connect_status = 1;
272 rpcrdma_update_cm_private(ep, &event->param.conn);
273 trace_xprtrdma_inline_thresh(ep);
274 wake_up_all(&ep->re_connect_wait);
276 case RDMA_CM_EVENT_CONNECT_ERROR:
277 ep->re_connect_status = -ENOTCONN;
279 case RDMA_CM_EVENT_UNREACHABLE:
280 ep->re_connect_status = -ENETUNREACH;
282 case RDMA_CM_EVENT_REJECTED:
283 dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
284 sap, rdma_reject_msg(id, event->status));
285 ep->re_connect_status = -ECONNREFUSED;
286 if (event->status == IB_CM_REJ_STALE_CONN)
287 ep->re_connect_status = -EAGAIN;
289 case RDMA_CM_EVENT_DISCONNECTED:
290 ep->re_connect_status = -ECONNABORTED;
292 xprt_force_disconnect(xprt);
293 return rpcrdma_ep_destroy(ep);
298 dprintk("RPC: %s: %pISpc on %s/frwr: %s\n", __func__, sap,
299 ep->re_id->device->name, rdma_event_msg(event->event));
303 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
304 struct rpcrdma_ep *ep)
306 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
307 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
308 struct rdma_cm_id *id;
311 init_completion(&ep->re_done);
313 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
314 RDMA_PS_TCP, IB_QPT_RC);
318 ep->re_async_rc = -ETIMEDOUT;
319 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
320 RDMA_RESOLVE_TIMEOUT);
323 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
327 rc = ep->re_async_rc;
331 ep->re_async_rc = -ETIMEDOUT;
332 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
335 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
338 rc = ep->re_async_rc;
349 static void rpcrdma_ep_put(struct kref *kref)
351 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
354 rdma_destroy_qp(ep->re_id);
355 ep->re_id->qp = NULL;
358 if (ep->re_attr.recv_cq)
359 ib_free_cq(ep->re_attr.recv_cq);
360 ep->re_attr.recv_cq = NULL;
361 if (ep->re_attr.send_cq)
362 ib_free_cq(ep->re_attr.send_cq);
363 ep->re_attr.send_cq = NULL;
366 ib_dealloc_pd(ep->re_pd);
370 module_put(THIS_MODULE);
374 * %0 if @ep still has a positive kref count, or
375 * %1 if @ep was destroyed successfully.
377 static int rpcrdma_ep_destroy(struct rpcrdma_ep *ep)
379 return kref_put(&ep->re_kref, rpcrdma_ep_put);
382 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
384 struct rpcrdma_connect_private *pmsg;
385 struct ib_device *device;
386 struct rdma_cm_id *id;
387 struct rpcrdma_ep *ep;
390 ep = kzalloc(sizeof(*ep), GFP_NOFS);
393 ep->re_xprt = &r_xprt->rx_xprt;
394 kref_init(&ep->re_kref);
396 id = rpcrdma_create_id(r_xprt, ep);
401 __module_get(THIS_MODULE);
405 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
406 ep->re_inline_send = xprt_rdma_max_inline_write;
407 ep->re_inline_recv = xprt_rdma_max_inline_read;
408 rc = frwr_query_device(ep, device);
412 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
414 ep->re_attr.event_handler = rpcrdma_qp_event_handler;
415 ep->re_attr.qp_context = ep;
416 ep->re_attr.srq = NULL;
417 ep->re_attr.cap.max_inline_data = 0;
418 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
419 ep->re_attr.qp_type = IB_QPT_RC;
420 ep->re_attr.port_num = ~0;
422 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
423 "iovs: send %d recv %d\n",
425 ep->re_attr.cap.max_send_wr,
426 ep->re_attr.cap.max_recv_wr,
427 ep->re_attr.cap.max_send_sge,
428 ep->re_attr.cap.max_recv_sge);
430 ep->re_send_batch = ep->re_max_requests >> 3;
431 ep->re_send_count = ep->re_send_batch;
432 init_waitqueue_head(&ep->re_connect_wait);
434 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
435 ep->re_attr.cap.max_send_wr,
437 if (IS_ERR(ep->re_attr.send_cq)) {
438 rc = PTR_ERR(ep->re_attr.send_cq);
442 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
443 ep->re_attr.cap.max_recv_wr,
445 if (IS_ERR(ep->re_attr.recv_cq)) {
446 rc = PTR_ERR(ep->re_attr.recv_cq);
449 ep->re_receive_count = 0;
451 /* Initialize cma parameters */
452 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
454 /* Prepare RDMA-CM private message */
455 pmsg = &ep->re_cm_private;
456 pmsg->cp_magic = rpcrdma_cmp_magic;
457 pmsg->cp_version = RPCRDMA_CMP_VERSION;
458 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
459 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
460 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
461 ep->re_remote_cma.private_data = pmsg;
462 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
464 /* Client offers RDMA Read but does not initiate */
465 ep->re_remote_cma.initiator_depth = 0;
466 ep->re_remote_cma.responder_resources =
467 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
469 /* Limit transport retries so client can detect server
470 * GID changes quickly. RPC layer handles re-establishing
471 * transport connection and retransmission.
473 ep->re_remote_cma.retry_count = 6;
475 /* RPC-over-RDMA handles its own flow control. In addition,
476 * make all RNR NAKs visible so we know that RPC-over-RDMA
477 * flow control is working correctly (no NAKs should be seen).
479 ep->re_remote_cma.flow_control = 0;
480 ep->re_remote_cma.rnr_retry_count = 0;
482 ep->re_pd = ib_alloc_pd(device, 0);
483 if (IS_ERR(ep->re_pd)) {
484 rc = PTR_ERR(ep->re_pd);
488 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
496 rpcrdma_ep_destroy(ep);
500 r_xprt->rx_ep = NULL;
505 * rpcrdma_xprt_connect - Connect an unconnected transport
506 * @r_xprt: controlling transport instance
508 * Returns 0 on success or a negative errno.
510 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
512 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
513 struct rpcrdma_ep *ep;
517 rpcrdma_xprt_disconnect(r_xprt);
518 rc = rpcrdma_ep_create(r_xprt);
523 ep->re_connect_status = 0;
524 xprt_clear_connected(xprt);
526 rpcrdma_reset_cwnd(r_xprt);
527 rpcrdma_post_recvs(r_xprt, true);
529 rc = rpcrdma_sendctxs_create(r_xprt);
533 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
537 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
538 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
539 wait_event_interruptible(ep->re_connect_wait,
540 ep->re_connect_status != 0);
541 if (ep->re_connect_status <= 0) {
542 if (ep->re_connect_status == -EAGAIN)
544 rc = ep->re_connect_status;
548 rc = rpcrdma_reqs_setup(r_xprt);
550 rpcrdma_xprt_disconnect(r_xprt);
553 rpcrdma_mrs_create(r_xprt);
557 ep->re_connect_status = rc;
558 trace_xprtrdma_connect(r_xprt, rc);
563 * rpcrdma_xprt_disconnect - Disconnect underlying transport
564 * @r_xprt: controlling transport instance
566 * Caller serializes. Either the transport send lock is held,
567 * or we're being called to destroy the transport.
569 * On return, @r_xprt is completely divested of all hardware
570 * resources and prepared for the next ->connect operation.
572 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
574 struct rpcrdma_ep *ep = r_xprt->rx_ep;
575 struct rdma_cm_id *id;
582 rc = rdma_disconnect(id);
583 trace_xprtrdma_disconnect(r_xprt, rc);
585 rpcrdma_xprt_drain(r_xprt);
586 rpcrdma_reps_unmap(r_xprt);
587 rpcrdma_reqs_reset(r_xprt);
588 rpcrdma_mrs_destroy(r_xprt);
589 rpcrdma_sendctxs_destroy(r_xprt);
591 if (rpcrdma_ep_destroy(ep))
594 r_xprt->rx_ep = NULL;
597 /* Fixed-size circular FIFO queue. This implementation is wait-free and
600 * Consumer is the code path that posts Sends. This path dequeues a
601 * sendctx for use by a Send operation. Multiple consumer threads
602 * are serialized by the RPC transport lock, which allows only one
603 * ->send_request call at a time.
605 * Producer is the code path that handles Send completions. This path
606 * enqueues a sendctx that has been completed. Multiple producer
607 * threads are serialized by the ib_poll_cq() function.
610 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
611 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
614 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
616 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
619 if (!buf->rb_sc_ctxs)
621 for (i = 0; i <= buf->rb_sc_last; i++)
622 kfree(buf->rb_sc_ctxs[i]);
623 kfree(buf->rb_sc_ctxs);
624 buf->rb_sc_ctxs = NULL;
627 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
629 struct rpcrdma_sendctx *sc;
631 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
636 sc->sc_cqe.done = rpcrdma_wc_send;
640 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
642 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
643 struct rpcrdma_sendctx *sc;
646 /* Maximum number of concurrent outstanding Send WRs. Capping
647 * the circular queue size stops Send Queue overflow by causing
648 * the ->send_request call to fail temporarily before too many
651 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
652 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
653 if (!buf->rb_sc_ctxs)
656 buf->rb_sc_last = i - 1;
657 for (i = 0; i <= buf->rb_sc_last; i++) {
658 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
662 buf->rb_sc_ctxs[i] = sc;
670 /* The sendctx queue is not guaranteed to have a size that is a
671 * power of two, thus the helpers in circ_buf.h cannot be used.
672 * The other option is to use modulus (%), which can be expensive.
674 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
677 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
681 * rpcrdma_sendctx_get_locked - Acquire a send context
682 * @r_xprt: controlling transport instance
684 * Returns pointer to a free send completion context; or NULL if
685 * the queue is empty.
687 * Usage: Called to acquire an SGE array before preparing a Send WR.
689 * The caller serializes calls to this function (per transport), and
690 * provides an effective memory barrier that flushes the new value
693 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
695 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
696 struct rpcrdma_sendctx *sc;
697 unsigned long next_head;
699 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
701 if (next_head == READ_ONCE(buf->rb_sc_tail))
704 /* ORDER: item must be accessed _before_ head is updated */
705 sc = buf->rb_sc_ctxs[next_head];
707 /* Releasing the lock in the caller acts as a memory
708 * barrier that flushes rb_sc_head.
710 buf->rb_sc_head = next_head;
715 /* The queue is "empty" if there have not been enough Send
716 * completions recently. This is a sign the Send Queue is
717 * backing up. Cause the caller to pause and try again.
719 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
720 r_xprt->rx_stats.empty_sendctx_q++;
725 * rpcrdma_sendctx_put_locked - Release a send context
726 * @r_xprt: controlling transport instance
727 * @sc: send context to release
729 * Usage: Called from Send completion to return a sendctxt
732 * The caller serializes calls to this function (per transport).
734 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
735 struct rpcrdma_sendctx *sc)
737 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
738 unsigned long next_tail;
740 /* Unmap SGEs of previously completed but unsignaled
741 * Sends by walking up the queue until @sc is found.
743 next_tail = buf->rb_sc_tail;
745 next_tail = rpcrdma_sendctx_next(buf, next_tail);
747 /* ORDER: item must be accessed _before_ tail is updated */
748 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
750 } while (buf->rb_sc_ctxs[next_tail] != sc);
752 /* Paired with READ_ONCE */
753 smp_store_release(&buf->rb_sc_tail, next_tail);
755 xprt_write_space(&r_xprt->rx_xprt);
759 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
761 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
762 struct rpcrdma_ep *ep = r_xprt->rx_ep;
765 for (count = 0; count < ep->re_max_rdma_segs; count++) {
766 struct rpcrdma_mr *mr;
769 mr = kzalloc(sizeof(*mr), GFP_NOFS);
773 rc = frwr_mr_init(r_xprt, mr);
779 spin_lock(&buf->rb_lock);
780 rpcrdma_mr_push(mr, &buf->rb_mrs);
781 list_add(&mr->mr_all, &buf->rb_all_mrs);
782 spin_unlock(&buf->rb_lock);
785 r_xprt->rx_stats.mrs_allocated += count;
786 trace_xprtrdma_createmrs(r_xprt, count);
790 rpcrdma_mr_refresh_worker(struct work_struct *work)
792 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
794 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
797 rpcrdma_mrs_create(r_xprt);
798 xprt_write_space(&r_xprt->rx_xprt);
802 * rpcrdma_mrs_refresh - Wake the MR refresh worker
803 * @r_xprt: controlling transport instance
806 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
808 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
809 struct rpcrdma_ep *ep = r_xprt->rx_ep;
811 /* If there is no underlying connection, it's no use
812 * to wake the refresh worker.
814 if (ep->re_connect_status == 1) {
815 /* The work is scheduled on a WQ_MEM_RECLAIM
816 * workqueue in order to prevent MR allocation
817 * from recursing into NFS during direct reclaim.
819 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
824 * rpcrdma_req_create - Allocate an rpcrdma_req object
825 * @r_xprt: controlling r_xprt
826 * @size: initial size, in bytes, of send and receive buffers
827 * @flags: GFP flags passed to memory allocators
829 * Returns an allocated and fully initialized rpcrdma_req or NULL.
831 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
834 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
835 struct rpcrdma_req *req;
837 req = kzalloc(sizeof(*req), flags);
841 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
842 if (!req->rl_sendbuf)
845 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
846 if (!req->rl_recvbuf)
849 INIT_LIST_HEAD(&req->rl_free_mrs);
850 INIT_LIST_HEAD(&req->rl_registered);
851 spin_lock(&buffer->rb_lock);
852 list_add(&req->rl_all, &buffer->rb_allreqs);
853 spin_unlock(&buffer->rb_lock);
857 kfree(req->rl_sendbuf);
865 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
866 * @r_xprt: controlling transport instance
867 * @req: rpcrdma_req object to set up
869 * Returns zero on success, and a negative errno on failure.
871 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
873 struct rpcrdma_regbuf *rb;
876 /* Compute maximum header buffer size in bytes */
877 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
878 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
879 maxhdrsize *= sizeof(__be32);
880 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
881 DMA_TO_DEVICE, GFP_KERNEL);
885 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
888 req->rl_rdmabuf = rb;
889 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
893 rpcrdma_regbuf_free(rb);
898 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
899 * and thus can be walked without holding rb_lock. Eg. the
900 * caller is holding the transport send lock to exclude
901 * device removal or disconnection.
903 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
905 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
906 struct rpcrdma_req *req;
909 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
910 rc = rpcrdma_req_setup(r_xprt, req);
917 static void rpcrdma_req_reset(struct rpcrdma_req *req)
919 /* Credits are valid for only one connection */
920 req->rl_slot.rq_cong = 0;
922 rpcrdma_regbuf_free(req->rl_rdmabuf);
923 req->rl_rdmabuf = NULL;
925 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
926 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
929 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
930 * and thus can be walked without holding rb_lock. Eg. the
931 * caller is holding the transport send lock to exclude
932 * device removal or disconnection.
934 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
936 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
937 struct rpcrdma_req *req;
939 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
940 rpcrdma_req_reset(req);
943 /* No locking needed here. This function is called only by the
944 * Receive completion handler.
947 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
950 struct rpcrdma_rep *rep;
952 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
956 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
957 DMA_FROM_DEVICE, GFP_KERNEL);
958 if (!rep->rr_rdmabuf)
961 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
962 goto out_free_regbuf;
964 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
965 rdmab_length(rep->rr_rdmabuf));
966 rep->rr_cqe.done = rpcrdma_wc_receive;
967 rep->rr_rxprt = r_xprt;
968 rep->rr_recv_wr.next = NULL;
969 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
970 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
971 rep->rr_recv_wr.num_sge = 1;
973 list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
977 rpcrdma_regbuf_free(rep->rr_rdmabuf);
984 /* No locking needed here. This function is invoked only by the
985 * Receive completion handler, or during transport shutdown.
987 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
989 list_del(&rep->rr_all);
990 rpcrdma_regbuf_free(rep->rr_rdmabuf);
994 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
996 struct llist_node *node;
998 /* Calls to llist_del_first are required to be serialized */
999 node = llist_del_first(&buf->rb_free_reps);
1002 return llist_entry(node, struct rpcrdma_rep, rr_node);
1005 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1006 struct rpcrdma_rep *rep)
1008 llist_add(&rep->rr_node, &buf->rb_free_reps);
1011 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1013 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1014 struct rpcrdma_rep *rep;
1016 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1017 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1018 rep->rr_temp = true;
1022 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1024 struct rpcrdma_rep *rep;
1026 while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1027 rpcrdma_rep_destroy(rep);
1031 * rpcrdma_buffer_create - Create initial set of req/rep objects
1032 * @r_xprt: transport instance to (re)initialize
1034 * Returns zero on success, otherwise a negative errno.
1036 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1038 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1041 buf->rb_bc_srv_max_requests = 0;
1042 spin_lock_init(&buf->rb_lock);
1043 INIT_LIST_HEAD(&buf->rb_mrs);
1044 INIT_LIST_HEAD(&buf->rb_all_mrs);
1045 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1047 INIT_LIST_HEAD(&buf->rb_send_bufs);
1048 INIT_LIST_HEAD(&buf->rb_allreqs);
1049 INIT_LIST_HEAD(&buf->rb_all_reps);
1052 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1053 struct rpcrdma_req *req;
1055 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1059 list_add(&req->rl_list, &buf->rb_send_bufs);
1062 init_llist_head(&buf->rb_free_reps);
1066 rpcrdma_buffer_destroy(buf);
1071 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1072 * @req: unused object to be destroyed
1074 * Relies on caller holding the transport send lock to protect
1075 * removing req->rl_all from buf->rb_all_reqs safely.
1077 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1079 struct rpcrdma_mr *mr;
1081 list_del(&req->rl_all);
1083 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1084 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1086 spin_lock(&buf->rb_lock);
1087 list_del(&mr->mr_all);
1088 spin_unlock(&buf->rb_lock);
1090 frwr_release_mr(mr);
1093 rpcrdma_regbuf_free(req->rl_recvbuf);
1094 rpcrdma_regbuf_free(req->rl_sendbuf);
1095 rpcrdma_regbuf_free(req->rl_rdmabuf);
1100 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1101 * @r_xprt: controlling transport instance
1103 * Relies on caller holding the transport send lock to protect
1104 * removing mr->mr_list from req->rl_free_mrs safely.
1106 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1108 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1109 struct rpcrdma_mr *mr;
1111 cancel_work_sync(&buf->rb_refresh_worker);
1113 spin_lock(&buf->rb_lock);
1114 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1117 list_del(&mr->mr_list);
1118 list_del(&mr->mr_all);
1119 spin_unlock(&buf->rb_lock);
1121 frwr_release_mr(mr);
1123 spin_lock(&buf->rb_lock);
1125 spin_unlock(&buf->rb_lock);
1129 * rpcrdma_buffer_destroy - Release all hw resources
1130 * @buf: root control block for resources
1132 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1133 * - No more Send or Receive completions can occur
1134 * - All MRs, reps, and reqs are returned to their free lists
1137 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1139 rpcrdma_reps_destroy(buf);
1141 while (!list_empty(&buf->rb_send_bufs)) {
1142 struct rpcrdma_req *req;
1144 req = list_first_entry(&buf->rb_send_bufs,
1145 struct rpcrdma_req, rl_list);
1146 list_del(&req->rl_list);
1147 rpcrdma_req_destroy(req);
1152 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1153 * @r_xprt: controlling transport
1155 * Returns an initialized rpcrdma_mr or NULL if no free
1156 * rpcrdma_mr objects are available.
1159 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1161 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1162 struct rpcrdma_mr *mr;
1164 spin_lock(&buf->rb_lock);
1165 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1166 spin_unlock(&buf->rb_lock);
1171 * rpcrdma_mr_put - DMA unmap an MR and release it
1172 * @mr: MR to release
1175 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1177 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1179 if (mr->mr_dir != DMA_NONE) {
1180 trace_xprtrdma_mr_unmap(mr);
1181 ib_dma_unmap_sg(r_xprt->rx_ep->re_id->device,
1182 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1183 mr->mr_dir = DMA_NONE;
1186 rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1190 * rpcrdma_buffer_get - Get a request buffer
1191 * @buffers: Buffer pool from which to obtain a buffer
1193 * Returns a fresh rpcrdma_req, or NULL if none are available.
1195 struct rpcrdma_req *
1196 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1198 struct rpcrdma_req *req;
1200 spin_lock(&buffers->rb_lock);
1201 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1202 struct rpcrdma_req, rl_list);
1204 list_del_init(&req->rl_list);
1205 spin_unlock(&buffers->rb_lock);
1210 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1211 * @buffers: buffer pool
1212 * @req: object to return
1215 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1218 rpcrdma_rep_put(buffers, req->rl_reply);
1219 req->rl_reply = NULL;
1221 spin_lock(&buffers->rb_lock);
1222 list_add(&req->rl_list, &buffers->rb_send_bufs);
1223 spin_unlock(&buffers->rb_lock);
1227 * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1228 * @rep: rep to release
1230 * Used after error conditions.
1232 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1234 rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1237 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1239 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1240 * receiving the payload of RDMA RECV operations. During Long Calls
1241 * or Replies they may be registered externally via frwr_map.
1243 static struct rpcrdma_regbuf *
1244 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1247 struct rpcrdma_regbuf *rb;
1249 rb = kmalloc(sizeof(*rb), flags);
1252 rb->rg_data = kmalloc(size, flags);
1258 rb->rg_device = NULL;
1259 rb->rg_direction = direction;
1260 rb->rg_iov.length = size;
1265 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1266 * @rb: regbuf to reallocate
1267 * @size: size of buffer to be allocated, in bytes
1270 * Returns true if reallocation was successful. If false is
1271 * returned, @rb is left untouched.
1273 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1277 buf = kmalloc(size, flags);
1281 rpcrdma_regbuf_dma_unmap(rb);
1285 rb->rg_iov.length = size;
1290 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1291 * @r_xprt: controlling transport instance
1292 * @rb: regbuf to be mapped
1294 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1296 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1297 struct rpcrdma_regbuf *rb)
1299 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1301 if (rb->rg_direction == DMA_NONE)
1304 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1305 rdmab_length(rb), rb->rg_direction);
1306 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1307 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1311 rb->rg_device = device;
1312 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1316 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1321 if (!rpcrdma_regbuf_is_mapped(rb))
1324 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1326 rb->rg_device = NULL;
1329 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1331 rpcrdma_regbuf_dma_unmap(rb);
1338 * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1339 * @r_xprt: controlling transport instance
1340 * @req: rpcrdma_req containing the Send WR to post
1342 * Returns 0 if the post was successful, otherwise -ENOTCONN
1345 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1347 struct ib_send_wr *send_wr = &req->rl_wr;
1348 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1351 if (!ep->re_send_count || kref_read(&req->rl_kref) > 1) {
1352 send_wr->send_flags |= IB_SEND_SIGNALED;
1353 ep->re_send_count = ep->re_send_batch;
1355 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1356 --ep->re_send_count;
1359 trace_xprtrdma_post_send(req);
1360 rc = frwr_send(r_xprt, req);
1367 * rpcrdma_post_recvs - Refill the Receive Queue
1368 * @r_xprt: controlling transport instance
1369 * @temp: mark Receive buffers to be deleted after use
1372 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1374 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1375 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1376 struct ib_recv_wr *wr, *bad_wr;
1377 struct rpcrdma_rep *rep;
1378 int needed, count, rc;
1383 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1384 if (likely(ep->re_receive_count > needed))
1386 needed -= ep->re_receive_count;
1388 needed += RPCRDMA_MAX_RECV_BATCH;
1390 /* fast path: all needed reps can be found on the free list */
1393 rep = rpcrdma_rep_get_locked(buf);
1394 if (rep && rep->rr_temp) {
1395 rpcrdma_rep_destroy(rep);
1399 rep = rpcrdma_rep_create(r_xprt, temp);
1403 trace_xprtrdma_post_recv(rep);
1404 rep->rr_recv_wr.next = wr;
1405 wr = &rep->rr_recv_wr;
1412 rc = ib_post_recv(ep->re_id->qp, wr,
1413 (const struct ib_recv_wr **)&bad_wr);
1415 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1417 for (wr = bad_wr; wr;) {
1418 struct rpcrdma_rep *rep;
1420 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1422 rpcrdma_recv_buffer_put(rep);
1426 ep->re_receive_count += count;