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.
123 /* Ensure xprt_force_disconnect() is invoked exactly once when a
124 * connection is closed or lost. (The important thing is it needs
125 * to be invoked "at least" once).
127 void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
129 if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
130 xprt_force_disconnect(ep->re_xprt);
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 if (wc->status != IB_WC_SUCCESS)
143 rpcrdma_force_disconnect(r_xprt->rx_ep);
147 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
148 * @cq: completion queue
149 * @wc: WCE for a completed Send WR
152 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
154 struct ib_cqe *cqe = wc->wr_cqe;
155 struct rpcrdma_sendctx *sc =
156 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
157 struct rpcrdma_xprt *r_xprt = cq->cq_context;
159 /* WARNING: Only wr_cqe and status are reliable at this point */
160 trace_xprtrdma_wc_send(wc, &sc->sc_cid);
161 rpcrdma_sendctx_put_locked(r_xprt, sc);
162 rpcrdma_flush_disconnect(r_xprt, wc);
166 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
167 * @cq: completion queue
168 * @wc: WCE for a completed Receive WR
171 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
173 struct ib_cqe *cqe = wc->wr_cqe;
174 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
176 struct rpcrdma_xprt *r_xprt = cq->cq_context;
178 /* WARNING: Only wr_cqe and status are reliable at this point */
179 trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
180 --r_xprt->rx_ep->re_receive_count;
181 if (wc->status != IB_WC_SUCCESS)
184 /* status == SUCCESS means all fields in wc are trustworthy */
185 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
186 rep->rr_wc_flags = wc->wc_flags;
187 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
189 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
190 rdmab_addr(rep->rr_rdmabuf),
191 wc->byte_len, DMA_FROM_DEVICE);
193 rpcrdma_reply_handler(rep);
197 rpcrdma_flush_disconnect(r_xprt, wc);
198 rpcrdma_rep_put(&r_xprt->rx_buf, rep);
201 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
202 struct rdma_conn_param *param)
204 const struct rpcrdma_connect_private *pmsg = param->private_data;
205 unsigned int rsize, wsize;
207 /* Default settings for RPC-over-RDMA Version One */
208 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
209 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
212 pmsg->cp_magic == rpcrdma_cmp_magic &&
213 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
214 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
215 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
218 if (rsize < ep->re_inline_recv)
219 ep->re_inline_recv = rsize;
220 if (wsize < ep->re_inline_send)
221 ep->re_inline_send = wsize;
223 rpcrdma_set_max_header_sizes(ep);
227 * rpcrdma_cm_event_handler - Handle RDMA CM events
228 * @id: rdma_cm_id on which an event has occurred
229 * @event: details of the event
231 * Called with @id's mutex held. Returns 1 if caller should
232 * destroy @id, otherwise 0.
235 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
237 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
238 struct rpcrdma_ep *ep = id->context;
242 switch (event->event) {
243 case RDMA_CM_EVENT_ADDR_RESOLVED:
244 case RDMA_CM_EVENT_ROUTE_RESOLVED:
246 complete(&ep->re_done);
248 case RDMA_CM_EVENT_ADDR_ERROR:
249 ep->re_async_rc = -EPROTO;
250 complete(&ep->re_done);
252 case RDMA_CM_EVENT_ROUTE_ERROR:
253 ep->re_async_rc = -ENETUNREACH;
254 complete(&ep->re_done);
256 case RDMA_CM_EVENT_DEVICE_REMOVAL:
257 pr_info("rpcrdma: removing device %s for %pISpc\n",
258 ep->re_id->device->name, sap);
260 case RDMA_CM_EVENT_ADDR_CHANGE:
261 ep->re_connect_status = -ENODEV;
263 case RDMA_CM_EVENT_ESTABLISHED:
265 ep->re_connect_status = 1;
266 rpcrdma_update_cm_private(ep, &event->param.conn);
267 trace_xprtrdma_inline_thresh(ep);
268 wake_up_all(&ep->re_connect_wait);
270 case RDMA_CM_EVENT_CONNECT_ERROR:
271 ep->re_connect_status = -ENOTCONN;
272 goto wake_connect_worker;
273 case RDMA_CM_EVENT_UNREACHABLE:
274 ep->re_connect_status = -ENETUNREACH;
275 goto wake_connect_worker;
276 case RDMA_CM_EVENT_REJECTED:
277 dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
278 sap, rdma_reject_msg(id, event->status));
279 ep->re_connect_status = -ECONNREFUSED;
280 if (event->status == IB_CM_REJ_STALE_CONN)
281 ep->re_connect_status = -ENOTCONN;
283 wake_up_all(&ep->re_connect_wait);
285 case RDMA_CM_EVENT_DISCONNECTED:
286 ep->re_connect_status = -ECONNABORTED;
288 rpcrdma_force_disconnect(ep);
289 return rpcrdma_ep_put(ep);
294 dprintk("RPC: %s: %pISpc on %s/frwr: %s\n", __func__, sap,
295 ep->re_id->device->name, rdma_event_msg(event->event));
299 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
300 struct rpcrdma_ep *ep)
302 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
303 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
304 struct rdma_cm_id *id;
307 init_completion(&ep->re_done);
309 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
310 RDMA_PS_TCP, IB_QPT_RC);
314 ep->re_async_rc = -ETIMEDOUT;
315 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
316 RDMA_RESOLVE_TIMEOUT);
319 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
323 rc = ep->re_async_rc;
327 ep->re_async_rc = -ETIMEDOUT;
328 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
331 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
334 rc = ep->re_async_rc;
345 static void rpcrdma_ep_destroy(struct kref *kref)
347 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
350 rdma_destroy_qp(ep->re_id);
351 ep->re_id->qp = NULL;
354 if (ep->re_attr.recv_cq)
355 ib_free_cq(ep->re_attr.recv_cq);
356 ep->re_attr.recv_cq = NULL;
357 if (ep->re_attr.send_cq)
358 ib_free_cq(ep->re_attr.send_cq);
359 ep->re_attr.send_cq = NULL;
362 ib_dealloc_pd(ep->re_pd);
366 module_put(THIS_MODULE);
369 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
371 kref_get(&ep->re_kref);
375 * %0 if @ep still has a positive kref count, or
376 * %1 if @ep was destroyed successfully.
378 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
380 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
383 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
385 struct rpcrdma_connect_private *pmsg;
386 struct ib_device *device;
387 struct rdma_cm_id *id;
388 struct rpcrdma_ep *ep;
391 ep = kzalloc(sizeof(*ep), GFP_NOFS);
394 ep->re_xprt = &r_xprt->rx_xprt;
395 kref_init(&ep->re_kref);
397 id = rpcrdma_create_id(r_xprt, ep);
402 __module_get(THIS_MODULE);
405 reinit_completion(&ep->re_done);
407 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
408 ep->re_inline_send = xprt_rdma_max_inline_write;
409 ep->re_inline_recv = xprt_rdma_max_inline_read;
410 rc = frwr_query_device(ep, device);
414 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
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);
502 * rpcrdma_xprt_connect - Connect an unconnected transport
503 * @r_xprt: controlling transport instance
505 * Returns 0 on success or a negative errno.
507 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
509 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
510 struct rpcrdma_ep *ep;
513 rc = rpcrdma_ep_create(r_xprt);
518 xprt_clear_connected(xprt);
519 rpcrdma_reset_cwnd(r_xprt);
521 /* Bump the ep's reference count while there are
522 * outstanding Receives.
525 rpcrdma_post_recvs(r_xprt, 1, true);
527 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
531 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
532 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
533 wait_event_interruptible(ep->re_connect_wait,
534 ep->re_connect_status != 0);
535 if (ep->re_connect_status <= 0) {
536 rc = ep->re_connect_status;
540 rc = rpcrdma_sendctxs_create(r_xprt);
546 rc = rpcrdma_reqs_setup(r_xprt);
551 rpcrdma_mrs_create(r_xprt);
552 frwr_wp_create(r_xprt);
555 trace_xprtrdma_connect(r_xprt, rc);
560 * rpcrdma_xprt_disconnect - Disconnect underlying transport
561 * @r_xprt: controlling transport instance
563 * Caller serializes. Either the transport send lock is held,
564 * or we're being called to destroy the transport.
566 * On return, @r_xprt is completely divested of all hardware
567 * resources and prepared for the next ->connect operation.
569 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
571 struct rpcrdma_ep *ep = r_xprt->rx_ep;
572 struct rdma_cm_id *id;
579 rc = rdma_disconnect(id);
580 trace_xprtrdma_disconnect(r_xprt, rc);
582 rpcrdma_xprt_drain(r_xprt);
583 rpcrdma_reps_unmap(r_xprt);
584 rpcrdma_reqs_reset(r_xprt);
585 rpcrdma_mrs_destroy(r_xprt);
586 rpcrdma_sendctxs_destroy(r_xprt);
588 if (rpcrdma_ep_put(ep))
591 r_xprt->rx_ep = NULL;
594 /* Fixed-size circular FIFO queue. This implementation is wait-free and
597 * Consumer is the code path that posts Sends. This path dequeues a
598 * sendctx for use by a Send operation. Multiple consumer threads
599 * are serialized by the RPC transport lock, which allows only one
600 * ->send_request call at a time.
602 * Producer is the code path that handles Send completions. This path
603 * enqueues a sendctx that has been completed. Multiple producer
604 * threads are serialized by the ib_poll_cq() function.
607 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
608 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
611 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
613 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
616 if (!buf->rb_sc_ctxs)
618 for (i = 0; i <= buf->rb_sc_last; i++)
619 kfree(buf->rb_sc_ctxs[i]);
620 kfree(buf->rb_sc_ctxs);
621 buf->rb_sc_ctxs = NULL;
624 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
626 struct rpcrdma_sendctx *sc;
628 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
633 sc->sc_cqe.done = rpcrdma_wc_send;
634 sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
635 sc->sc_cid.ci_completion_id =
636 atomic_inc_return(&ep->re_completion_ids);
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);
931 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
932 * and thus can be walked without holding rb_lock. Eg. the
933 * caller is holding the transport send lock to exclude
934 * device removal or disconnection.
936 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
938 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
939 struct rpcrdma_req *req;
941 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
942 rpcrdma_req_reset(req);
946 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
949 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
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 rep->rr_cid.ci_completion_id =
965 atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
967 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
968 rdmab_length(rep->rr_rdmabuf));
969 rep->rr_cqe.done = rpcrdma_wc_receive;
970 rep->rr_rxprt = r_xprt;
971 rep->rr_recv_wr.next = NULL;
972 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
973 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
974 rep->rr_recv_wr.num_sge = 1;
977 spin_lock(&buf->rb_lock);
978 list_add(&rep->rr_all, &buf->rb_all_reps);
979 spin_unlock(&buf->rb_lock);
983 rpcrdma_regbuf_free(rep->rr_rdmabuf);
990 static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
992 rpcrdma_regbuf_free(rep->rr_rdmabuf);
996 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
998 struct rpcrdma_buffer *buf = &rep->rr_rxprt->rx_buf;
1000 spin_lock(&buf->rb_lock);
1001 list_del(&rep->rr_all);
1002 spin_unlock(&buf->rb_lock);
1004 rpcrdma_rep_free(rep);
1007 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1009 struct llist_node *node;
1011 /* Calls to llist_del_first are required to be serialized */
1012 node = llist_del_first(&buf->rb_free_reps);
1015 return llist_entry(node, struct rpcrdma_rep, rr_node);
1019 * rpcrdma_rep_put - Release rpcrdma_rep back to free list
1021 * @rep: rep to release
1024 void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
1026 llist_add(&rep->rr_node, &buf->rb_free_reps);
1029 /* Caller must ensure the QP is quiescent (RQ is drained) before
1030 * invoking this function, to guarantee rb_all_reps is not
1033 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1035 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1036 struct rpcrdma_rep *rep;
1038 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1039 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1040 rep->rr_temp = true; /* Mark this rep for destruction */
1044 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1046 struct rpcrdma_rep *rep;
1048 spin_lock(&buf->rb_lock);
1049 while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1052 list_del(&rep->rr_all);
1053 spin_unlock(&buf->rb_lock);
1055 rpcrdma_rep_free(rep);
1057 spin_lock(&buf->rb_lock);
1059 spin_unlock(&buf->rb_lock);
1063 * rpcrdma_buffer_create - Create initial set of req/rep objects
1064 * @r_xprt: transport instance to (re)initialize
1066 * Returns zero on success, otherwise a negative errno.
1068 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1070 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1073 buf->rb_bc_srv_max_requests = 0;
1074 spin_lock_init(&buf->rb_lock);
1075 INIT_LIST_HEAD(&buf->rb_mrs);
1076 INIT_LIST_HEAD(&buf->rb_all_mrs);
1077 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1079 INIT_LIST_HEAD(&buf->rb_send_bufs);
1080 INIT_LIST_HEAD(&buf->rb_allreqs);
1081 INIT_LIST_HEAD(&buf->rb_all_reps);
1084 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1085 struct rpcrdma_req *req;
1087 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1091 list_add(&req->rl_list, &buf->rb_send_bufs);
1094 init_llist_head(&buf->rb_free_reps);
1098 rpcrdma_buffer_destroy(buf);
1103 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1104 * @req: unused object to be destroyed
1106 * Relies on caller holding the transport send lock to protect
1107 * removing req->rl_all from buf->rb_all_reqs safely.
1109 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1111 struct rpcrdma_mr *mr;
1113 list_del(&req->rl_all);
1115 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1116 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1118 spin_lock(&buf->rb_lock);
1119 list_del(&mr->mr_all);
1120 spin_unlock(&buf->rb_lock);
1122 frwr_mr_release(mr);
1125 rpcrdma_regbuf_free(req->rl_recvbuf);
1126 rpcrdma_regbuf_free(req->rl_sendbuf);
1127 rpcrdma_regbuf_free(req->rl_rdmabuf);
1132 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1133 * @r_xprt: controlling transport instance
1135 * Relies on caller holding the transport send lock to protect
1136 * removing mr->mr_list from req->rl_free_mrs safely.
1138 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1140 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1141 struct rpcrdma_mr *mr;
1143 cancel_work_sync(&buf->rb_refresh_worker);
1145 spin_lock(&buf->rb_lock);
1146 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1149 list_del(&mr->mr_list);
1150 list_del(&mr->mr_all);
1151 spin_unlock(&buf->rb_lock);
1153 frwr_mr_release(mr);
1155 spin_lock(&buf->rb_lock);
1157 spin_unlock(&buf->rb_lock);
1161 * rpcrdma_buffer_destroy - Release all hw resources
1162 * @buf: root control block for resources
1164 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1165 * - No more Send or Receive completions can occur
1166 * - All MRs, reps, and reqs are returned to their free lists
1169 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1171 rpcrdma_reps_destroy(buf);
1173 while (!list_empty(&buf->rb_send_bufs)) {
1174 struct rpcrdma_req *req;
1176 req = list_first_entry(&buf->rb_send_bufs,
1177 struct rpcrdma_req, rl_list);
1178 list_del(&req->rl_list);
1179 rpcrdma_req_destroy(req);
1184 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1185 * @r_xprt: controlling transport
1187 * Returns an initialized rpcrdma_mr or NULL if no free
1188 * rpcrdma_mr objects are available.
1191 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1193 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1194 struct rpcrdma_mr *mr;
1196 spin_lock(&buf->rb_lock);
1197 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1198 spin_unlock(&buf->rb_lock);
1203 * rpcrdma_reply_put - Put reply buffers back into pool
1204 * @buffers: buffer pool
1205 * @req: object to return
1208 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1210 if (req->rl_reply) {
1211 rpcrdma_rep_put(buffers, req->rl_reply);
1212 req->rl_reply = NULL;
1217 * rpcrdma_buffer_get - Get a request buffer
1218 * @buffers: Buffer pool from which to obtain a buffer
1220 * Returns a fresh rpcrdma_req, or NULL if none are available.
1222 struct rpcrdma_req *
1223 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1225 struct rpcrdma_req *req;
1227 spin_lock(&buffers->rb_lock);
1228 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1229 struct rpcrdma_req, rl_list);
1231 list_del_init(&req->rl_list);
1232 spin_unlock(&buffers->rb_lock);
1237 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1238 * @buffers: buffer pool
1239 * @req: object to return
1242 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1244 rpcrdma_reply_put(buffers, req);
1246 spin_lock(&buffers->rb_lock);
1247 list_add(&req->rl_list, &buffers->rb_send_bufs);
1248 spin_unlock(&buffers->rb_lock);
1251 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1253 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1254 * receiving the payload of RDMA RECV operations. During Long Calls
1255 * or Replies they may be registered externally via frwr_map.
1257 static struct rpcrdma_regbuf *
1258 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1261 struct rpcrdma_regbuf *rb;
1263 rb = kmalloc(sizeof(*rb), flags);
1266 rb->rg_data = kmalloc(size, flags);
1272 rb->rg_device = NULL;
1273 rb->rg_direction = direction;
1274 rb->rg_iov.length = size;
1279 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1280 * @rb: regbuf to reallocate
1281 * @size: size of buffer to be allocated, in bytes
1284 * Returns true if reallocation was successful. If false is
1285 * returned, @rb is left untouched.
1287 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1291 buf = kmalloc(size, flags);
1295 rpcrdma_regbuf_dma_unmap(rb);
1299 rb->rg_iov.length = size;
1304 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1305 * @r_xprt: controlling transport instance
1306 * @rb: regbuf to be mapped
1308 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1310 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1311 struct rpcrdma_regbuf *rb)
1313 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1315 if (rb->rg_direction == DMA_NONE)
1318 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1319 rdmab_length(rb), rb->rg_direction);
1320 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1321 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1325 rb->rg_device = device;
1326 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1330 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1335 if (!rpcrdma_regbuf_is_mapped(rb))
1338 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1340 rb->rg_device = NULL;
1343 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1345 rpcrdma_regbuf_dma_unmap(rb);
1352 * rpcrdma_post_recvs - Refill the Receive Queue
1353 * @r_xprt: controlling transport instance
1354 * @needed: current credit grant
1355 * @temp: mark Receive buffers to be deleted after one use
1358 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1360 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1361 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1362 struct ib_recv_wr *wr, *bad_wr;
1363 struct rpcrdma_rep *rep;
1369 if (likely(ep->re_receive_count > needed))
1371 needed -= ep->re_receive_count;
1373 needed += RPCRDMA_MAX_RECV_BATCH;
1375 if (atomic_inc_return(&ep->re_receiving) > 1)
1378 /* fast path: all needed reps can be found on the free list */
1381 rep = rpcrdma_rep_get_locked(buf);
1382 if (rep && rep->rr_temp) {
1383 rpcrdma_rep_destroy(rep);
1387 rep = rpcrdma_rep_create(r_xprt, temp);
1391 rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1392 trace_xprtrdma_post_recv(rep);
1393 rep->rr_recv_wr.next = wr;
1394 wr = &rep->rr_recv_wr;
1401 rc = ib_post_recv(ep->re_id->qp, wr,
1402 (const struct ib_recv_wr **)&bad_wr);
1404 trace_xprtrdma_post_recvs_err(r_xprt, rc);
1405 for (wr = bad_wr; wr;) {
1406 struct rpcrdma_rep *rep;
1408 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1410 rpcrdma_rep_put(buf, rep);
1414 if (atomic_dec_return(&ep->re_receiving) > 0)
1415 complete(&ep->re_done);
1418 trace_xprtrdma_post_recvs(r_xprt, count);
1419 ep->re_receive_count += count;