1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
19 * Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials provided
22 * with the distribution.
24 * Neither the name of the Network Appliance, Inc. nor the names of
25 * its contributors may be used to endorse or promote products
26 * derived from this software without specific prior written
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 * Encapsulates the major functions managing:
52 #include <linux/interrupt.h>
53 #include <linux/slab.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/sunrpc/svc_rdma.h>
56 #include <linux/log2.h>
58 #include <asm-generic/barrier.h>
59 #include <asm/bitops.h>
61 #include <rdma/ib_cm.h>
63 #include "xprt_rdma.h"
64 #include <trace/events/rpcrdma.h>
70 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
71 # define RPCDBG_FACILITY RPCDBG_TRANS
77 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
79 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
80 struct rpcrdma_sendctx *sc);
81 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
82 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
83 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
84 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
85 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
86 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
87 static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
88 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
89 static struct rpcrdma_regbuf *
90 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
92 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
93 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
95 /* Wait for outstanding transport work to finish. ib_drain_qp
96 * handles the drains in the wrong order for us, so open code
99 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
101 struct rpcrdma_ep *ep = r_xprt->rx_ep;
102 struct rdma_cm_id *id = ep->re_id;
104 /* Flush Receives, then wait for deferred Reply work
109 /* Deferred Reply processing might have scheduled
110 * local invalidations.
118 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
119 * @event: details of the event
120 * @context: ep that owns QP where event occurred
122 * Called from the RDMA provider (device driver) possibly in an interrupt
123 * context. The QP is always destroyed before the ID, so the ID will be
124 * reliably available when this handler is invoked.
126 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
128 struct rpcrdma_ep *ep = context;
130 trace_xprtrdma_qp_event(ep, event);
134 * rpcrdma_flush_disconnect - Disconnect on flushed completion
135 * @r_xprt: transport to disconnect
136 * @wc: work completion entry
138 * Must be called in process context.
140 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
142 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
144 if (wc->status != IB_WC_SUCCESS &&
145 r_xprt->rx_ep->re_connect_status == 1) {
146 r_xprt->rx_ep->re_connect_status = -ECONNABORTED;
147 xprt_force_disconnect(xprt);
152 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
153 * @cq: completion queue
154 * @wc: WCE for a completed Send WR
157 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
159 struct ib_cqe *cqe = wc->wr_cqe;
160 struct rpcrdma_sendctx *sc =
161 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
162 struct rpcrdma_xprt *r_xprt = cq->cq_context;
164 /* WARNING: Only wr_cqe and status are reliable at this point */
165 trace_xprtrdma_wc_send(sc, wc);
166 rpcrdma_sendctx_put_locked(r_xprt, sc);
167 rpcrdma_flush_disconnect(r_xprt, wc);
171 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
172 * @cq: completion queue
173 * @wc: WCE for a completed Receive WR
176 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
178 struct ib_cqe *cqe = wc->wr_cqe;
179 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
181 struct rpcrdma_xprt *r_xprt = cq->cq_context;
183 /* WARNING: Only wr_cqe and status are reliable at this point */
184 trace_xprtrdma_wc_receive(wc);
185 --r_xprt->rx_ep->re_receive_count;
186 if (wc->status != IB_WC_SUCCESS)
189 /* status == SUCCESS means all fields in wc are trustworthy */
190 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
191 rep->rr_wc_flags = wc->wc_flags;
192 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
194 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
195 rdmab_addr(rep->rr_rdmabuf),
196 wc->byte_len, DMA_FROM_DEVICE);
198 rpcrdma_reply_handler(rep);
202 rpcrdma_flush_disconnect(r_xprt, wc);
203 rpcrdma_rep_destroy(rep);
206 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
207 struct rdma_conn_param *param)
209 const struct rpcrdma_connect_private *pmsg = param->private_data;
210 unsigned int rsize, wsize;
212 /* Default settings for RPC-over-RDMA Version One */
213 ep->re_implicit_roundup = xprt_rdma_pad_optimize;
214 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
215 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
218 pmsg->cp_magic == rpcrdma_cmp_magic &&
219 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
220 ep->re_implicit_roundup = true;
221 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
222 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
225 if (rsize < ep->re_inline_recv)
226 ep->re_inline_recv = rsize;
227 if (wsize < ep->re_inline_send)
228 ep->re_inline_send = wsize;
230 rpcrdma_set_max_header_sizes(ep);
234 * rpcrdma_cm_event_handler - Handle RDMA CM events
235 * @id: rdma_cm_id on which an event has occurred
236 * @event: details of the event
238 * Called with @id's mutex held. Returns 1 if caller should
239 * destroy @id, otherwise 0.
242 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
244 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
245 struct rpcrdma_ep *ep = id->context;
246 struct rpc_xprt *xprt = ep->re_xprt;
250 switch (event->event) {
251 case RDMA_CM_EVENT_ADDR_RESOLVED:
252 case RDMA_CM_EVENT_ROUTE_RESOLVED:
254 complete(&ep->re_done);
256 case RDMA_CM_EVENT_ADDR_ERROR:
257 ep->re_async_rc = -EPROTO;
258 complete(&ep->re_done);
260 case RDMA_CM_EVENT_ROUTE_ERROR:
261 ep->re_async_rc = -ENETUNREACH;
262 complete(&ep->re_done);
264 case RDMA_CM_EVENT_DEVICE_REMOVAL:
265 pr_info("rpcrdma: removing device %s for %pISpc\n",
266 ep->re_id->device->name, sap);
268 case RDMA_CM_EVENT_ADDR_CHANGE:
269 ep->re_connect_status = -ENODEV;
270 xprt_force_disconnect(xprt);
272 case RDMA_CM_EVENT_ESTABLISHED:
274 ep->re_connect_status = 1;
275 rpcrdma_update_cm_private(ep, &event->param.conn);
276 trace_xprtrdma_inline_thresh(ep);
277 wake_up_all(&ep->re_connect_wait);
279 case RDMA_CM_EVENT_CONNECT_ERROR:
280 ep->re_connect_status = -ENOTCONN;
282 case RDMA_CM_EVENT_UNREACHABLE:
283 ep->re_connect_status = -ENETUNREACH;
285 case RDMA_CM_EVENT_REJECTED:
286 dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
287 sap, rdma_reject_msg(id, event->status));
288 ep->re_connect_status = -ECONNREFUSED;
289 if (event->status == IB_CM_REJ_STALE_CONN)
290 ep->re_connect_status = -EAGAIN;
292 case RDMA_CM_EVENT_DISCONNECTED:
293 ep->re_connect_status = -ECONNABORTED;
295 xprt_force_disconnect(xprt);
296 return rpcrdma_ep_put(ep);
301 dprintk("RPC: %s: %pISpc on %s/frwr: %s\n", __func__, sap,
302 ep->re_id->device->name, rdma_event_msg(event->event));
306 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
307 struct rpcrdma_ep *ep)
309 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
310 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
311 struct rdma_cm_id *id;
314 init_completion(&ep->re_done);
316 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
317 RDMA_PS_TCP, IB_QPT_RC);
321 ep->re_async_rc = -ETIMEDOUT;
322 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
323 RDMA_RESOLVE_TIMEOUT);
326 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
330 rc = ep->re_async_rc;
334 ep->re_async_rc = -ETIMEDOUT;
335 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
338 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
341 rc = ep->re_async_rc;
352 static void rpcrdma_ep_destroy(struct kref *kref)
354 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
357 rdma_destroy_qp(ep->re_id);
358 ep->re_id->qp = NULL;
361 if (ep->re_attr.recv_cq)
362 ib_free_cq(ep->re_attr.recv_cq);
363 ep->re_attr.recv_cq = NULL;
364 if (ep->re_attr.send_cq)
365 ib_free_cq(ep->re_attr.send_cq);
366 ep->re_attr.send_cq = NULL;
369 ib_dealloc_pd(ep->re_pd);
373 module_put(THIS_MODULE);
376 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
378 kref_get(&ep->re_kref);
382 * %0 if @ep still has a positive kref count, or
383 * %1 if @ep was destroyed successfully.
385 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
387 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
390 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
392 struct rpcrdma_connect_private *pmsg;
393 struct ib_device *device;
394 struct rdma_cm_id *id;
395 struct rpcrdma_ep *ep;
398 ep = kzalloc(sizeof(*ep), GFP_NOFS);
401 ep->re_xprt = &r_xprt->rx_xprt;
402 kref_init(&ep->re_kref);
404 id = rpcrdma_create_id(r_xprt, ep);
409 __module_get(THIS_MODULE);
413 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
414 ep->re_inline_send = xprt_rdma_max_inline_write;
415 ep->re_inline_recv = xprt_rdma_max_inline_read;
416 rc = frwr_query_device(ep, device);
420 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
422 ep->re_attr.event_handler = rpcrdma_qp_event_handler;
423 ep->re_attr.qp_context = ep;
424 ep->re_attr.srq = NULL;
425 ep->re_attr.cap.max_inline_data = 0;
426 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
427 ep->re_attr.qp_type = IB_QPT_RC;
428 ep->re_attr.port_num = ~0;
430 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
431 "iovs: send %d recv %d\n",
433 ep->re_attr.cap.max_send_wr,
434 ep->re_attr.cap.max_recv_wr,
435 ep->re_attr.cap.max_send_sge,
436 ep->re_attr.cap.max_recv_sge);
438 ep->re_send_batch = ep->re_max_requests >> 3;
439 ep->re_send_count = ep->re_send_batch;
440 init_waitqueue_head(&ep->re_connect_wait);
442 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
443 ep->re_attr.cap.max_send_wr,
445 if (IS_ERR(ep->re_attr.send_cq)) {
446 rc = PTR_ERR(ep->re_attr.send_cq);
450 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
451 ep->re_attr.cap.max_recv_wr,
453 if (IS_ERR(ep->re_attr.recv_cq)) {
454 rc = PTR_ERR(ep->re_attr.recv_cq);
457 ep->re_receive_count = 0;
459 /* Initialize cma parameters */
460 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
462 /* Prepare RDMA-CM private message */
463 pmsg = &ep->re_cm_private;
464 pmsg->cp_magic = rpcrdma_cmp_magic;
465 pmsg->cp_version = RPCRDMA_CMP_VERSION;
466 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
467 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
468 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
469 ep->re_remote_cma.private_data = pmsg;
470 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
472 /* Client offers RDMA Read but does not initiate */
473 ep->re_remote_cma.initiator_depth = 0;
474 ep->re_remote_cma.responder_resources =
475 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
477 /* Limit transport retries so client can detect server
478 * GID changes quickly. RPC layer handles re-establishing
479 * transport connection and retransmission.
481 ep->re_remote_cma.retry_count = 6;
483 /* RPC-over-RDMA handles its own flow control. In addition,
484 * make all RNR NAKs visible so we know that RPC-over-RDMA
485 * flow control is working correctly (no NAKs should be seen).
487 ep->re_remote_cma.flow_control = 0;
488 ep->re_remote_cma.rnr_retry_count = 0;
490 ep->re_pd = ib_alloc_pd(device, 0);
491 if (IS_ERR(ep->re_pd)) {
492 rc = PTR_ERR(ep->re_pd);
496 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
508 r_xprt->rx_ep = NULL;
513 * rpcrdma_xprt_connect - Connect an unconnected transport
514 * @r_xprt: controlling transport instance
516 * Returns 0 on success or a negative errno.
518 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
520 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
521 struct rpcrdma_ep *ep;
525 rpcrdma_xprt_disconnect(r_xprt);
526 rc = rpcrdma_ep_create(r_xprt);
531 xprt_clear_connected(xprt);
532 rpcrdma_reset_cwnd(r_xprt);
534 /* Bump the ep's reference count while there are
535 * outstanding Receives.
538 rpcrdma_post_recvs(r_xprt, true);
540 rc = rpcrdma_sendctxs_create(r_xprt);
544 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
548 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
549 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
550 wait_event_interruptible(ep->re_connect_wait,
551 ep->re_connect_status != 0);
552 if (ep->re_connect_status <= 0) {
553 if (ep->re_connect_status == -EAGAIN)
555 rc = ep->re_connect_status;
559 rc = rpcrdma_reqs_setup(r_xprt);
561 rpcrdma_xprt_disconnect(r_xprt);
564 rpcrdma_mrs_create(r_xprt);
567 trace_xprtrdma_connect(r_xprt, rc);
572 * rpcrdma_xprt_disconnect - Disconnect underlying transport
573 * @r_xprt: controlling transport instance
575 * Caller serializes. Either the transport send lock is held,
576 * or we're being called to destroy the transport.
578 * On return, @r_xprt is completely divested of all hardware
579 * resources and prepared for the next ->connect operation.
581 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
583 struct rpcrdma_ep *ep = r_xprt->rx_ep;
584 struct rdma_cm_id *id;
591 rc = rdma_disconnect(id);
592 trace_xprtrdma_disconnect(r_xprt, rc);
594 rpcrdma_xprt_drain(r_xprt);
595 rpcrdma_reps_unmap(r_xprt);
596 rpcrdma_reqs_reset(r_xprt);
597 rpcrdma_mrs_destroy(r_xprt);
598 rpcrdma_sendctxs_destroy(r_xprt);
600 if (rpcrdma_ep_put(ep))
603 r_xprt->rx_ep = NULL;
606 /* Fixed-size circular FIFO queue. This implementation is wait-free and
609 * Consumer is the code path that posts Sends. This path dequeues a
610 * sendctx for use by a Send operation. Multiple consumer threads
611 * are serialized by the RPC transport lock, which allows only one
612 * ->send_request call at a time.
614 * Producer is the code path that handles Send completions. This path
615 * enqueues a sendctx that has been completed. Multiple producer
616 * threads are serialized by the ib_poll_cq() function.
619 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
620 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
623 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
625 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
628 if (!buf->rb_sc_ctxs)
630 for (i = 0; i <= buf->rb_sc_last; i++)
631 kfree(buf->rb_sc_ctxs[i]);
632 kfree(buf->rb_sc_ctxs);
633 buf->rb_sc_ctxs = NULL;
636 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
638 struct rpcrdma_sendctx *sc;
640 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
645 sc->sc_cqe.done = rpcrdma_wc_send;
649 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
651 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
652 struct rpcrdma_sendctx *sc;
655 /* Maximum number of concurrent outstanding Send WRs. Capping
656 * the circular queue size stops Send Queue overflow by causing
657 * the ->send_request call to fail temporarily before too many
660 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
661 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
662 if (!buf->rb_sc_ctxs)
665 buf->rb_sc_last = i - 1;
666 for (i = 0; i <= buf->rb_sc_last; i++) {
667 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
671 buf->rb_sc_ctxs[i] = sc;
679 /* The sendctx queue is not guaranteed to have a size that is a
680 * power of two, thus the helpers in circ_buf.h cannot be used.
681 * The other option is to use modulus (%), which can be expensive.
683 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
686 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
690 * rpcrdma_sendctx_get_locked - Acquire a send context
691 * @r_xprt: controlling transport instance
693 * Returns pointer to a free send completion context; or NULL if
694 * the queue is empty.
696 * Usage: Called to acquire an SGE array before preparing a Send WR.
698 * The caller serializes calls to this function (per transport), and
699 * provides an effective memory barrier that flushes the new value
702 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
704 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
705 struct rpcrdma_sendctx *sc;
706 unsigned long next_head;
708 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
710 if (next_head == READ_ONCE(buf->rb_sc_tail))
713 /* ORDER: item must be accessed _before_ head is updated */
714 sc = buf->rb_sc_ctxs[next_head];
716 /* Releasing the lock in the caller acts as a memory
717 * barrier that flushes rb_sc_head.
719 buf->rb_sc_head = next_head;
724 /* The queue is "empty" if there have not been enough Send
725 * completions recently. This is a sign the Send Queue is
726 * backing up. Cause the caller to pause and try again.
728 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
729 r_xprt->rx_stats.empty_sendctx_q++;
734 * rpcrdma_sendctx_put_locked - Release a send context
735 * @r_xprt: controlling transport instance
736 * @sc: send context to release
738 * Usage: Called from Send completion to return a sendctxt
741 * The caller serializes calls to this function (per transport).
743 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
744 struct rpcrdma_sendctx *sc)
746 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
747 unsigned long next_tail;
749 /* Unmap SGEs of previously completed but unsignaled
750 * Sends by walking up the queue until @sc is found.
752 next_tail = buf->rb_sc_tail;
754 next_tail = rpcrdma_sendctx_next(buf, next_tail);
756 /* ORDER: item must be accessed _before_ tail is updated */
757 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
759 } while (buf->rb_sc_ctxs[next_tail] != sc);
761 /* Paired with READ_ONCE */
762 smp_store_release(&buf->rb_sc_tail, next_tail);
764 xprt_write_space(&r_xprt->rx_xprt);
768 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
770 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
771 struct rpcrdma_ep *ep = r_xprt->rx_ep;
774 for (count = 0; count < ep->re_max_rdma_segs; count++) {
775 struct rpcrdma_mr *mr;
778 mr = kzalloc(sizeof(*mr), GFP_NOFS);
782 rc = frwr_mr_init(r_xprt, mr);
788 spin_lock(&buf->rb_lock);
789 rpcrdma_mr_push(mr, &buf->rb_mrs);
790 list_add(&mr->mr_all, &buf->rb_all_mrs);
791 spin_unlock(&buf->rb_lock);
794 r_xprt->rx_stats.mrs_allocated += count;
795 trace_xprtrdma_createmrs(r_xprt, count);
799 rpcrdma_mr_refresh_worker(struct work_struct *work)
801 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
803 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
806 rpcrdma_mrs_create(r_xprt);
807 xprt_write_space(&r_xprt->rx_xprt);
811 * rpcrdma_mrs_refresh - Wake the MR refresh worker
812 * @r_xprt: controlling transport instance
815 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
817 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
818 struct rpcrdma_ep *ep = r_xprt->rx_ep;
820 /* If there is no underlying connection, it's no use
821 * to wake the refresh worker.
823 if (ep->re_connect_status == 1) {
824 /* The work is scheduled on a WQ_MEM_RECLAIM
825 * workqueue in order to prevent MR allocation
826 * from recursing into NFS during direct reclaim.
828 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
833 * rpcrdma_req_create - Allocate an rpcrdma_req object
834 * @r_xprt: controlling r_xprt
835 * @size: initial size, in bytes, of send and receive buffers
836 * @flags: GFP flags passed to memory allocators
838 * Returns an allocated and fully initialized rpcrdma_req or NULL.
840 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
843 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
844 struct rpcrdma_req *req;
846 req = kzalloc(sizeof(*req), flags);
850 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
851 if (!req->rl_sendbuf)
854 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
855 if (!req->rl_recvbuf)
858 INIT_LIST_HEAD(&req->rl_free_mrs);
859 INIT_LIST_HEAD(&req->rl_registered);
860 spin_lock(&buffer->rb_lock);
861 list_add(&req->rl_all, &buffer->rb_allreqs);
862 spin_unlock(&buffer->rb_lock);
866 kfree(req->rl_sendbuf);
874 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
875 * @r_xprt: controlling transport instance
876 * @req: rpcrdma_req object to set up
878 * Returns zero on success, and a negative errno on failure.
880 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
882 struct rpcrdma_regbuf *rb;
885 /* Compute maximum header buffer size in bytes */
886 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
887 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
888 maxhdrsize *= sizeof(__be32);
889 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
890 DMA_TO_DEVICE, GFP_KERNEL);
894 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
897 req->rl_rdmabuf = rb;
898 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
902 rpcrdma_regbuf_free(rb);
907 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
908 * and thus can be walked without holding rb_lock. Eg. the
909 * caller is holding the transport send lock to exclude
910 * device removal or disconnection.
912 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
914 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
915 struct rpcrdma_req *req;
918 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
919 rc = rpcrdma_req_setup(r_xprt, req);
926 static void rpcrdma_req_reset(struct rpcrdma_req *req)
928 /* Credits are valid for only one connection */
929 req->rl_slot.rq_cong = 0;
931 rpcrdma_regbuf_free(req->rl_rdmabuf);
932 req->rl_rdmabuf = NULL;
934 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
935 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
938 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
939 * and thus can be walked without holding rb_lock. Eg. the
940 * caller is holding the transport send lock to exclude
941 * device removal or disconnection.
943 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
945 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
946 struct rpcrdma_req *req;
948 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
949 rpcrdma_req_reset(req);
952 /* No locking needed here. This function is called only by the
953 * Receive completion handler.
956 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
959 struct rpcrdma_rep *rep;
961 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
965 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
966 DMA_FROM_DEVICE, GFP_KERNEL);
967 if (!rep->rr_rdmabuf)
970 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
971 goto out_free_regbuf;
973 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
974 rdmab_length(rep->rr_rdmabuf));
975 rep->rr_cqe.done = rpcrdma_wc_receive;
976 rep->rr_rxprt = r_xprt;
977 rep->rr_recv_wr.next = NULL;
978 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
979 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
980 rep->rr_recv_wr.num_sge = 1;
982 list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
986 rpcrdma_regbuf_free(rep->rr_rdmabuf);
993 /* No locking needed here. This function is invoked only by the
994 * Receive completion handler, or during transport shutdown.
996 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
998 list_del(&rep->rr_all);
999 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1003 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1005 struct llist_node *node;
1007 /* Calls to llist_del_first are required to be serialized */
1008 node = llist_del_first(&buf->rb_free_reps);
1011 return llist_entry(node, struct rpcrdma_rep, rr_node);
1014 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1015 struct rpcrdma_rep *rep)
1017 llist_add(&rep->rr_node, &buf->rb_free_reps);
1020 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1022 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1023 struct rpcrdma_rep *rep;
1025 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1026 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1027 rep->rr_temp = true;
1031 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1033 struct rpcrdma_rep *rep;
1035 while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1036 rpcrdma_rep_destroy(rep);
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_release_mr(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_release_mr(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_mr_put - DMA unmap an MR and release it
1181 * @mr: MR to release
1184 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1186 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1188 if (mr->mr_dir != DMA_NONE) {
1189 trace_xprtrdma_mr_unmap(mr);
1190 ib_dma_unmap_sg(r_xprt->rx_ep->re_id->device,
1191 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1192 mr->mr_dir = DMA_NONE;
1195 rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1199 * rpcrdma_buffer_get - Get a request buffer
1200 * @buffers: Buffer pool from which to obtain a buffer
1202 * Returns a fresh rpcrdma_req, or NULL if none are available.
1204 struct rpcrdma_req *
1205 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1207 struct rpcrdma_req *req;
1209 spin_lock(&buffers->rb_lock);
1210 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1211 struct rpcrdma_req, rl_list);
1213 list_del_init(&req->rl_list);
1214 spin_unlock(&buffers->rb_lock);
1219 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1220 * @buffers: buffer pool
1221 * @req: object to return
1224 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1227 rpcrdma_rep_put(buffers, req->rl_reply);
1228 req->rl_reply = NULL;
1230 spin_lock(&buffers->rb_lock);
1231 list_add(&req->rl_list, &buffers->rb_send_bufs);
1232 spin_unlock(&buffers->rb_lock);
1236 * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1237 * @rep: rep to release
1239 * Used after error conditions.
1241 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1243 rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1246 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1248 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1249 * receiving the payload of RDMA RECV operations. During Long Calls
1250 * or Replies they may be registered externally via frwr_map.
1252 static struct rpcrdma_regbuf *
1253 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1256 struct rpcrdma_regbuf *rb;
1258 rb = kmalloc(sizeof(*rb), flags);
1261 rb->rg_data = kmalloc(size, flags);
1267 rb->rg_device = NULL;
1268 rb->rg_direction = direction;
1269 rb->rg_iov.length = size;
1274 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1275 * @rb: regbuf to reallocate
1276 * @size: size of buffer to be allocated, in bytes
1279 * Returns true if reallocation was successful. If false is
1280 * returned, @rb is left untouched.
1282 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1286 buf = kmalloc(size, flags);
1290 rpcrdma_regbuf_dma_unmap(rb);
1294 rb->rg_iov.length = size;
1299 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1300 * @r_xprt: controlling transport instance
1301 * @rb: regbuf to be mapped
1303 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1305 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1306 struct rpcrdma_regbuf *rb)
1308 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1310 if (rb->rg_direction == DMA_NONE)
1313 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1314 rdmab_length(rb), rb->rg_direction);
1315 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1316 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1320 rb->rg_device = device;
1321 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1325 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1330 if (!rpcrdma_regbuf_is_mapped(rb))
1333 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1335 rb->rg_device = NULL;
1338 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1340 rpcrdma_regbuf_dma_unmap(rb);
1347 * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1348 * @r_xprt: controlling transport instance
1349 * @req: rpcrdma_req containing the Send WR to post
1351 * Returns 0 if the post was successful, otherwise -ENOTCONN
1354 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1356 struct ib_send_wr *send_wr = &req->rl_wr;
1357 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1360 if (!ep->re_send_count || kref_read(&req->rl_kref) > 1) {
1361 send_wr->send_flags |= IB_SEND_SIGNALED;
1362 ep->re_send_count = ep->re_send_batch;
1364 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1365 --ep->re_send_count;
1368 trace_xprtrdma_post_send(req);
1369 rc = frwr_send(r_xprt, req);
1376 * rpcrdma_post_recvs - Refill the Receive Queue
1377 * @r_xprt: controlling transport instance
1378 * @temp: mark Receive buffers to be deleted after use
1381 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1383 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1384 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1385 struct ib_recv_wr *wr, *bad_wr;
1386 struct rpcrdma_rep *rep;
1387 int needed, count, rc;
1392 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1393 if (likely(ep->re_receive_count > needed))
1395 needed -= ep->re_receive_count;
1397 needed += RPCRDMA_MAX_RECV_BATCH;
1399 /* fast path: all needed reps can be found on the free list */
1402 rep = rpcrdma_rep_get_locked(buf);
1403 if (rep && rep->rr_temp) {
1404 rpcrdma_rep_destroy(rep);
1408 rep = rpcrdma_rep_create(r_xprt, temp);
1412 trace_xprtrdma_post_recv(rep);
1413 rep->rr_recv_wr.next = wr;
1414 wr = &rep->rr_recv_wr;
1421 rc = ib_post_recv(ep->re_id->qp, wr,
1422 (const struct ib_recv_wr **)&bad_wr);
1424 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1426 for (wr = bad_wr; wr;) {
1427 struct rpcrdma_rep *rep;
1429 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1431 rpcrdma_recv_buffer_put(rep);
1435 ep->re_receive_count += count;