1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
19 * Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials provided
22 * with the distribution.
24 * Neither the name of the Network Appliance, Inc. nor the names of
25 * its contributors may be used to endorse or promote products
26 * derived from this software without specific prior written
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 * Encapsulates the major functions managing:
52 #include <linux/interrupt.h>
53 #include <linux/slab.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/sunrpc/svc_rdma.h>
56 #include <linux/log2.h>
58 #include <asm-generic/barrier.h>
59 #include <asm/bitops.h>
61 #include <rdma/ib_cm.h>
63 #include "xprt_rdma.h"
64 #include <trace/events/rpcrdma.h>
66 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
67 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
68 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
69 struct rpcrdma_sendctx *sc);
70 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
71 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
72 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
73 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
74 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
75 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
76 static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
77 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
78 static struct rpcrdma_regbuf *
79 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction);
80 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
81 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
83 /* Wait for outstanding transport work to finish. ib_drain_qp
84 * handles the drains in the wrong order for us, so open code
87 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
89 struct rpcrdma_ep *ep = r_xprt->rx_ep;
90 struct rdma_cm_id *id = ep->re_id;
92 /* Wait for rpcrdma_post_recvs() to leave its critical
95 if (atomic_inc_return(&ep->re_receiving) > 1)
96 wait_for_completion(&ep->re_done);
98 /* Flush Receives, then wait for deferred Reply work
103 /* Deferred Reply processing might have scheduled
104 * local invalidations.
111 /* Ensure xprt_force_disconnect() is invoked exactly once when a
112 * connection is closed or lost. (The important thing is it needs
113 * to be invoked "at least" once).
115 void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
117 if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
118 xprt_force_disconnect(ep->re_xprt);
122 * rpcrdma_flush_disconnect - Disconnect on flushed completion
123 * @r_xprt: transport to disconnect
124 * @wc: work completion entry
126 * Must be called in process context.
128 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
130 if (wc->status != IB_WC_SUCCESS)
131 rpcrdma_force_disconnect(r_xprt->rx_ep);
135 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
136 * @cq: completion queue
137 * @wc: WCE for a completed Send WR
140 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
142 struct ib_cqe *cqe = wc->wr_cqe;
143 struct rpcrdma_sendctx *sc =
144 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
145 struct rpcrdma_xprt *r_xprt = cq->cq_context;
147 /* WARNING: Only wr_cqe and status are reliable at this point */
148 trace_xprtrdma_wc_send(wc, &sc->sc_cid);
149 rpcrdma_sendctx_put_locked(r_xprt, sc);
150 rpcrdma_flush_disconnect(r_xprt, wc);
154 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
155 * @cq: completion queue
156 * @wc: WCE for a completed Receive WR
159 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
161 struct ib_cqe *cqe = wc->wr_cqe;
162 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
164 struct rpcrdma_xprt *r_xprt = cq->cq_context;
166 /* WARNING: Only wr_cqe and status are reliable at this point */
167 trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
168 --r_xprt->rx_ep->re_receive_count;
169 if (wc->status != IB_WC_SUCCESS)
172 /* status == SUCCESS means all fields in wc are trustworthy */
173 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
174 rep->rr_wc_flags = wc->wc_flags;
175 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
177 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
178 rdmab_addr(rep->rr_rdmabuf),
179 wc->byte_len, DMA_FROM_DEVICE);
181 rpcrdma_reply_handler(rep);
185 rpcrdma_flush_disconnect(r_xprt, wc);
186 rpcrdma_rep_put(&r_xprt->rx_buf, rep);
189 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
190 struct rdma_conn_param *param)
192 const struct rpcrdma_connect_private *pmsg = param->private_data;
193 unsigned int rsize, wsize;
195 /* Default settings for RPC-over-RDMA Version One */
196 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
197 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
200 pmsg->cp_magic == rpcrdma_cmp_magic &&
201 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
202 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
203 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
206 if (rsize < ep->re_inline_recv)
207 ep->re_inline_recv = rsize;
208 if (wsize < ep->re_inline_send)
209 ep->re_inline_send = wsize;
211 rpcrdma_set_max_header_sizes(ep);
215 * rpcrdma_cm_event_handler - Handle RDMA CM events
216 * @id: rdma_cm_id on which an event has occurred
217 * @event: details of the event
219 * Called with @id's mutex held. Returns 1 if caller should
220 * destroy @id, otherwise 0.
223 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
225 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
226 struct rpcrdma_ep *ep = id->context;
230 switch (event->event) {
231 case RDMA_CM_EVENT_ADDR_RESOLVED:
232 case RDMA_CM_EVENT_ROUTE_RESOLVED:
234 complete(&ep->re_done);
236 case RDMA_CM_EVENT_ADDR_ERROR:
237 ep->re_async_rc = -EPROTO;
238 complete(&ep->re_done);
240 case RDMA_CM_EVENT_ROUTE_ERROR:
241 ep->re_async_rc = -ENETUNREACH;
242 complete(&ep->re_done);
244 case RDMA_CM_EVENT_DEVICE_REMOVAL:
245 pr_info("rpcrdma: removing device %s for %pISpc\n",
246 ep->re_id->device->name, sap);
248 case RDMA_CM_EVENT_ADDR_CHANGE:
249 ep->re_connect_status = -ENODEV;
251 case RDMA_CM_EVENT_ESTABLISHED:
253 ep->re_connect_status = 1;
254 rpcrdma_update_cm_private(ep, &event->param.conn);
255 trace_xprtrdma_inline_thresh(ep);
256 wake_up_all(&ep->re_connect_wait);
258 case RDMA_CM_EVENT_CONNECT_ERROR:
259 ep->re_connect_status = -ENOTCONN;
260 goto wake_connect_worker;
261 case RDMA_CM_EVENT_UNREACHABLE:
262 ep->re_connect_status = -ENETUNREACH;
263 goto wake_connect_worker;
264 case RDMA_CM_EVENT_REJECTED:
265 ep->re_connect_status = -ECONNREFUSED;
266 if (event->status == IB_CM_REJ_STALE_CONN)
267 ep->re_connect_status = -ENOTCONN;
269 wake_up_all(&ep->re_connect_wait);
271 case RDMA_CM_EVENT_DISCONNECTED:
272 ep->re_connect_status = -ECONNABORTED;
274 rpcrdma_force_disconnect(ep);
275 return rpcrdma_ep_put(ep);
283 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
284 struct rpcrdma_ep *ep)
286 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
287 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
288 struct rdma_cm_id *id;
291 init_completion(&ep->re_done);
293 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
294 RDMA_PS_TCP, IB_QPT_RC);
298 ep->re_async_rc = -ETIMEDOUT;
299 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
300 RDMA_RESOLVE_TIMEOUT);
303 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
307 rc = ep->re_async_rc;
311 ep->re_async_rc = -ETIMEDOUT;
312 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
315 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
318 rc = ep->re_async_rc;
329 static void rpcrdma_ep_destroy(struct kref *kref)
331 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
334 rdma_destroy_qp(ep->re_id);
335 ep->re_id->qp = NULL;
338 if (ep->re_attr.recv_cq)
339 ib_free_cq(ep->re_attr.recv_cq);
340 ep->re_attr.recv_cq = NULL;
341 if (ep->re_attr.send_cq)
342 ib_free_cq(ep->re_attr.send_cq);
343 ep->re_attr.send_cq = NULL;
346 ib_dealloc_pd(ep->re_pd);
350 module_put(THIS_MODULE);
353 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
355 kref_get(&ep->re_kref);
359 * %0 if @ep still has a positive kref count, or
360 * %1 if @ep was destroyed successfully.
362 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
364 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
367 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
369 struct rpcrdma_connect_private *pmsg;
370 struct ib_device *device;
371 struct rdma_cm_id *id;
372 struct rpcrdma_ep *ep;
375 ep = kzalloc(sizeof(*ep), XPRTRDMA_GFP_FLAGS);
378 ep->re_xprt = &r_xprt->rx_xprt;
379 kref_init(&ep->re_kref);
381 id = rpcrdma_create_id(r_xprt, ep);
386 __module_get(THIS_MODULE);
389 reinit_completion(&ep->re_done);
391 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
392 ep->re_inline_send = xprt_rdma_max_inline_write;
393 ep->re_inline_recv = xprt_rdma_max_inline_read;
394 rc = frwr_query_device(ep, device);
398 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
400 ep->re_attr.srq = NULL;
401 ep->re_attr.cap.max_inline_data = 0;
402 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
403 ep->re_attr.qp_type = IB_QPT_RC;
404 ep->re_attr.port_num = ~0;
406 ep->re_send_batch = ep->re_max_requests >> 3;
407 ep->re_send_count = ep->re_send_batch;
408 init_waitqueue_head(&ep->re_connect_wait);
410 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
411 ep->re_attr.cap.max_send_wr,
413 if (IS_ERR(ep->re_attr.send_cq)) {
414 rc = PTR_ERR(ep->re_attr.send_cq);
415 ep->re_attr.send_cq = NULL;
419 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
420 ep->re_attr.cap.max_recv_wr,
422 if (IS_ERR(ep->re_attr.recv_cq)) {
423 rc = PTR_ERR(ep->re_attr.recv_cq);
424 ep->re_attr.recv_cq = NULL;
427 ep->re_receive_count = 0;
429 /* Initialize cma parameters */
430 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
432 /* Prepare RDMA-CM private message */
433 pmsg = &ep->re_cm_private;
434 pmsg->cp_magic = rpcrdma_cmp_magic;
435 pmsg->cp_version = RPCRDMA_CMP_VERSION;
436 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
437 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
438 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
439 ep->re_remote_cma.private_data = pmsg;
440 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
442 /* Client offers RDMA Read but does not initiate */
443 ep->re_remote_cma.initiator_depth = 0;
444 ep->re_remote_cma.responder_resources =
445 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
447 /* Limit transport retries so client can detect server
448 * GID changes quickly. RPC layer handles re-establishing
449 * transport connection and retransmission.
451 ep->re_remote_cma.retry_count = 6;
453 /* RPC-over-RDMA handles its own flow control. In addition,
454 * make all RNR NAKs visible so we know that RPC-over-RDMA
455 * flow control is working correctly (no NAKs should be seen).
457 ep->re_remote_cma.flow_control = 0;
458 ep->re_remote_cma.rnr_retry_count = 0;
460 ep->re_pd = ib_alloc_pd(device, 0);
461 if (IS_ERR(ep->re_pd)) {
462 rc = PTR_ERR(ep->re_pd);
467 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
481 * rpcrdma_xprt_connect - Connect an unconnected transport
482 * @r_xprt: controlling transport instance
484 * Returns 0 on success or a negative errno.
486 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
488 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
489 struct rpcrdma_ep *ep;
492 rc = rpcrdma_ep_create(r_xprt);
497 xprt_clear_connected(xprt);
498 rpcrdma_reset_cwnd(r_xprt);
500 /* Bump the ep's reference count while there are
501 * outstanding Receives.
504 rpcrdma_post_recvs(r_xprt, 1, true);
506 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
510 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
511 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
512 wait_event_interruptible(ep->re_connect_wait,
513 ep->re_connect_status != 0);
514 if (ep->re_connect_status <= 0) {
515 rc = ep->re_connect_status;
519 rc = rpcrdma_sendctxs_create(r_xprt);
525 rc = rpcrdma_reqs_setup(r_xprt);
530 rpcrdma_mrs_create(r_xprt);
531 frwr_wp_create(r_xprt);
534 trace_xprtrdma_connect(r_xprt, rc);
539 * rpcrdma_xprt_disconnect - Disconnect underlying transport
540 * @r_xprt: controlling transport instance
542 * Caller serializes. Either the transport send lock is held,
543 * or we're being called to destroy the transport.
545 * On return, @r_xprt is completely divested of all hardware
546 * resources and prepared for the next ->connect operation.
548 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
550 struct rpcrdma_ep *ep = r_xprt->rx_ep;
551 struct rdma_cm_id *id;
558 rc = rdma_disconnect(id);
559 trace_xprtrdma_disconnect(r_xprt, rc);
561 rpcrdma_xprt_drain(r_xprt);
562 rpcrdma_reps_unmap(r_xprt);
563 rpcrdma_reqs_reset(r_xprt);
564 rpcrdma_mrs_destroy(r_xprt);
565 rpcrdma_sendctxs_destroy(r_xprt);
567 if (rpcrdma_ep_put(ep))
570 r_xprt->rx_ep = NULL;
573 /* Fixed-size circular FIFO queue. This implementation is wait-free and
576 * Consumer is the code path that posts Sends. This path dequeues a
577 * sendctx for use by a Send operation. Multiple consumer threads
578 * are serialized by the RPC transport lock, which allows only one
579 * ->send_request call at a time.
581 * Producer is the code path that handles Send completions. This path
582 * enqueues a sendctx that has been completed. Multiple producer
583 * threads are serialized by the ib_poll_cq() function.
586 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
587 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
590 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
592 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
595 if (!buf->rb_sc_ctxs)
597 for (i = 0; i <= buf->rb_sc_last; i++)
598 kfree(buf->rb_sc_ctxs[i]);
599 kfree(buf->rb_sc_ctxs);
600 buf->rb_sc_ctxs = NULL;
603 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
605 struct rpcrdma_sendctx *sc;
607 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
612 sc->sc_cqe.done = rpcrdma_wc_send;
613 sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
614 sc->sc_cid.ci_completion_id =
615 atomic_inc_return(&ep->re_completion_ids);
619 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
621 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
622 struct rpcrdma_sendctx *sc;
625 /* Maximum number of concurrent outstanding Send WRs. Capping
626 * the circular queue size stops Send Queue overflow by causing
627 * the ->send_request call to fail temporarily before too many
630 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
631 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), XPRTRDMA_GFP_FLAGS);
632 if (!buf->rb_sc_ctxs)
635 buf->rb_sc_last = i - 1;
636 for (i = 0; i <= buf->rb_sc_last; i++) {
637 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
641 buf->rb_sc_ctxs[i] = sc;
649 /* The sendctx queue is not guaranteed to have a size that is a
650 * power of two, thus the helpers in circ_buf.h cannot be used.
651 * The other option is to use modulus (%), which can be expensive.
653 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
656 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
660 * rpcrdma_sendctx_get_locked - Acquire a send context
661 * @r_xprt: controlling transport instance
663 * Returns pointer to a free send completion context; or NULL if
664 * the queue is empty.
666 * Usage: Called to acquire an SGE array before preparing a Send WR.
668 * The caller serializes calls to this function (per transport), and
669 * provides an effective memory barrier that flushes the new value
672 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
674 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
675 struct rpcrdma_sendctx *sc;
676 unsigned long next_head;
678 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
680 if (next_head == READ_ONCE(buf->rb_sc_tail))
683 /* ORDER: item must be accessed _before_ head is updated */
684 sc = buf->rb_sc_ctxs[next_head];
686 /* Releasing the lock in the caller acts as a memory
687 * barrier that flushes rb_sc_head.
689 buf->rb_sc_head = next_head;
694 /* The queue is "empty" if there have not been enough Send
695 * completions recently. This is a sign the Send Queue is
696 * backing up. Cause the caller to pause and try again.
698 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
699 r_xprt->rx_stats.empty_sendctx_q++;
704 * rpcrdma_sendctx_put_locked - Release a send context
705 * @r_xprt: controlling transport instance
706 * @sc: send context to release
708 * Usage: Called from Send completion to return a sendctxt
711 * The caller serializes calls to this function (per transport).
713 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
714 struct rpcrdma_sendctx *sc)
716 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
717 unsigned long next_tail;
719 /* Unmap SGEs of previously completed but unsignaled
720 * Sends by walking up the queue until @sc is found.
722 next_tail = buf->rb_sc_tail;
724 next_tail = rpcrdma_sendctx_next(buf, next_tail);
726 /* ORDER: item must be accessed _before_ tail is updated */
727 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
729 } while (buf->rb_sc_ctxs[next_tail] != sc);
731 /* Paired with READ_ONCE */
732 smp_store_release(&buf->rb_sc_tail, next_tail);
734 xprt_write_space(&r_xprt->rx_xprt);
738 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
740 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
741 struct rpcrdma_ep *ep = r_xprt->rx_ep;
742 struct ib_device *device = ep->re_id->device;
745 /* Try to allocate enough to perform one full-sized I/O */
746 for (count = 0; count < ep->re_max_rdma_segs; count++) {
747 struct rpcrdma_mr *mr;
750 mr = kzalloc_node(sizeof(*mr), XPRTRDMA_GFP_FLAGS,
751 ibdev_to_node(device));
755 rc = frwr_mr_init(r_xprt, mr);
761 spin_lock(&buf->rb_lock);
762 rpcrdma_mr_push(mr, &buf->rb_mrs);
763 list_add(&mr->mr_all, &buf->rb_all_mrs);
764 spin_unlock(&buf->rb_lock);
767 r_xprt->rx_stats.mrs_allocated += count;
768 trace_xprtrdma_createmrs(r_xprt, count);
772 rpcrdma_mr_refresh_worker(struct work_struct *work)
774 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
776 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
779 rpcrdma_mrs_create(r_xprt);
780 xprt_write_space(&r_xprt->rx_xprt);
784 * rpcrdma_mrs_refresh - Wake the MR refresh worker
785 * @r_xprt: controlling transport instance
788 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
790 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
791 struct rpcrdma_ep *ep = r_xprt->rx_ep;
793 /* If there is no underlying connection, it's no use
794 * to wake the refresh worker.
796 if (ep->re_connect_status != 1)
798 queue_work(system_highpri_wq, &buf->rb_refresh_worker);
802 * rpcrdma_req_create - Allocate an rpcrdma_req object
803 * @r_xprt: controlling r_xprt
804 * @size: initial size, in bytes, of send and receive buffers
806 * Returns an allocated and fully initialized rpcrdma_req or NULL.
808 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt,
811 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
812 struct rpcrdma_req *req;
814 req = kzalloc(sizeof(*req), XPRTRDMA_GFP_FLAGS);
818 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE);
819 if (!req->rl_sendbuf)
822 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE);
823 if (!req->rl_recvbuf)
826 INIT_LIST_HEAD(&req->rl_free_mrs);
827 INIT_LIST_HEAD(&req->rl_registered);
828 spin_lock(&buffer->rb_lock);
829 list_add(&req->rl_all, &buffer->rb_allreqs);
830 spin_unlock(&buffer->rb_lock);
834 rpcrdma_regbuf_free(req->rl_sendbuf);
842 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
843 * @r_xprt: controlling transport instance
844 * @req: rpcrdma_req object to set up
846 * Returns zero on success, and a negative errno on failure.
848 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
850 struct rpcrdma_regbuf *rb;
853 /* Compute maximum header buffer size in bytes */
854 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
855 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
856 maxhdrsize *= sizeof(__be32);
857 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
862 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
865 req->rl_rdmabuf = rb;
866 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
870 rpcrdma_regbuf_free(rb);
875 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
876 * and thus can be walked without holding rb_lock. Eg. the
877 * caller is holding the transport send lock to exclude
878 * device removal or disconnection.
880 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
882 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
883 struct rpcrdma_req *req;
886 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
887 rc = rpcrdma_req_setup(r_xprt, req);
894 static void rpcrdma_req_reset(struct rpcrdma_req *req)
896 /* Credits are valid for only one connection */
897 req->rl_slot.rq_cong = 0;
899 rpcrdma_regbuf_free(req->rl_rdmabuf);
900 req->rl_rdmabuf = NULL;
902 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
903 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
908 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
909 * and thus can be walked without holding rb_lock. Eg. the
910 * caller is holding the transport send lock to exclude
911 * device removal or disconnection.
913 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
915 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
916 struct rpcrdma_req *req;
918 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
919 rpcrdma_req_reset(req);
923 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
926 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
927 struct rpcrdma_rep *rep;
929 rep = kzalloc(sizeof(*rep), XPRTRDMA_GFP_FLAGS);
933 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
935 if (!rep->rr_rdmabuf)
938 rep->rr_cid.ci_completion_id =
939 atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
941 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
942 rdmab_length(rep->rr_rdmabuf));
943 rep->rr_cqe.done = rpcrdma_wc_receive;
944 rep->rr_rxprt = r_xprt;
945 rep->rr_recv_wr.next = NULL;
946 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
947 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
948 rep->rr_recv_wr.num_sge = 1;
951 spin_lock(&buf->rb_lock);
952 list_add(&rep->rr_all, &buf->rb_all_reps);
953 spin_unlock(&buf->rb_lock);
962 static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
964 rpcrdma_regbuf_free(rep->rr_rdmabuf);
968 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
970 struct rpcrdma_buffer *buf = &rep->rr_rxprt->rx_buf;
972 spin_lock(&buf->rb_lock);
973 list_del(&rep->rr_all);
974 spin_unlock(&buf->rb_lock);
976 rpcrdma_rep_free(rep);
979 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
981 struct llist_node *node;
983 /* Calls to llist_del_first are required to be serialized */
984 node = llist_del_first(&buf->rb_free_reps);
987 return llist_entry(node, struct rpcrdma_rep, rr_node);
991 * rpcrdma_rep_put - Release rpcrdma_rep back to free list
993 * @rep: rep to release
996 void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
998 llist_add(&rep->rr_node, &buf->rb_free_reps);
1001 /* Caller must ensure the QP is quiescent (RQ is drained) before
1002 * invoking this function, to guarantee rb_all_reps is not
1005 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1007 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1008 struct rpcrdma_rep *rep;
1010 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1011 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1012 rep->rr_temp = true; /* Mark this rep for destruction */
1016 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1018 struct rpcrdma_rep *rep;
1020 spin_lock(&buf->rb_lock);
1021 while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1024 list_del(&rep->rr_all);
1025 spin_unlock(&buf->rb_lock);
1027 rpcrdma_rep_free(rep);
1029 spin_lock(&buf->rb_lock);
1031 spin_unlock(&buf->rb_lock);
1035 * rpcrdma_buffer_create - Create initial set of req/rep objects
1036 * @r_xprt: transport instance to (re)initialize
1038 * Returns zero on success, otherwise a negative errno.
1040 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1042 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1045 buf->rb_bc_srv_max_requests = 0;
1046 spin_lock_init(&buf->rb_lock);
1047 INIT_LIST_HEAD(&buf->rb_mrs);
1048 INIT_LIST_HEAD(&buf->rb_all_mrs);
1049 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1051 INIT_LIST_HEAD(&buf->rb_send_bufs);
1052 INIT_LIST_HEAD(&buf->rb_allreqs);
1053 INIT_LIST_HEAD(&buf->rb_all_reps);
1056 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1057 struct rpcrdma_req *req;
1059 req = rpcrdma_req_create(r_xprt,
1060 RPCRDMA_V1_DEF_INLINE_SIZE * 2);
1063 list_add(&req->rl_list, &buf->rb_send_bufs);
1066 init_llist_head(&buf->rb_free_reps);
1070 rpcrdma_buffer_destroy(buf);
1075 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1076 * @req: unused object to be destroyed
1078 * Relies on caller holding the transport send lock to protect
1079 * removing req->rl_all from buf->rb_all_reqs safely.
1081 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1083 struct rpcrdma_mr *mr;
1085 list_del(&req->rl_all);
1087 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1088 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1090 spin_lock(&buf->rb_lock);
1091 list_del(&mr->mr_all);
1092 spin_unlock(&buf->rb_lock);
1094 frwr_mr_release(mr);
1097 rpcrdma_regbuf_free(req->rl_recvbuf);
1098 rpcrdma_regbuf_free(req->rl_sendbuf);
1099 rpcrdma_regbuf_free(req->rl_rdmabuf);
1104 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1105 * @r_xprt: controlling transport instance
1107 * Relies on caller holding the transport send lock to protect
1108 * removing mr->mr_list from req->rl_free_mrs safely.
1110 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1112 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1113 struct rpcrdma_mr *mr;
1115 cancel_work_sync(&buf->rb_refresh_worker);
1117 spin_lock(&buf->rb_lock);
1118 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1121 list_del(&mr->mr_list);
1122 list_del(&mr->mr_all);
1123 spin_unlock(&buf->rb_lock);
1125 frwr_mr_release(mr);
1127 spin_lock(&buf->rb_lock);
1129 spin_unlock(&buf->rb_lock);
1133 * rpcrdma_buffer_destroy - Release all hw resources
1134 * @buf: root control block for resources
1136 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1137 * - No more Send or Receive completions can occur
1138 * - All MRs, reps, and reqs are returned to their free lists
1141 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1143 rpcrdma_reps_destroy(buf);
1145 while (!list_empty(&buf->rb_send_bufs)) {
1146 struct rpcrdma_req *req;
1148 req = list_first_entry(&buf->rb_send_bufs,
1149 struct rpcrdma_req, rl_list);
1150 list_del(&req->rl_list);
1151 rpcrdma_req_destroy(req);
1156 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1157 * @r_xprt: controlling transport
1159 * Returns an initialized rpcrdma_mr or NULL if no free
1160 * rpcrdma_mr objects are available.
1163 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1165 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1166 struct rpcrdma_mr *mr;
1168 spin_lock(&buf->rb_lock);
1169 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1170 spin_unlock(&buf->rb_lock);
1175 * rpcrdma_reply_put - Put reply buffers back into pool
1176 * @buffers: buffer pool
1177 * @req: object to return
1180 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1182 if (req->rl_reply) {
1183 rpcrdma_rep_put(buffers, req->rl_reply);
1184 req->rl_reply = NULL;
1189 * rpcrdma_buffer_get - Get a request buffer
1190 * @buffers: Buffer pool from which to obtain a buffer
1192 * Returns a fresh rpcrdma_req, or NULL if none are available.
1194 struct rpcrdma_req *
1195 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1197 struct rpcrdma_req *req;
1199 spin_lock(&buffers->rb_lock);
1200 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1201 struct rpcrdma_req, rl_list);
1203 list_del_init(&req->rl_list);
1204 spin_unlock(&buffers->rb_lock);
1209 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1210 * @buffers: buffer pool
1211 * @req: object to return
1214 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1216 rpcrdma_reply_put(buffers, req);
1218 spin_lock(&buffers->rb_lock);
1219 list_add(&req->rl_list, &buffers->rb_send_bufs);
1220 spin_unlock(&buffers->rb_lock);
1223 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1225 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1226 * receiving the payload of RDMA RECV operations. During Long Calls
1227 * or Replies they may be registered externally via frwr_map.
1229 static struct rpcrdma_regbuf *
1230 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction)
1232 struct rpcrdma_regbuf *rb;
1234 rb = kmalloc(sizeof(*rb), XPRTRDMA_GFP_FLAGS);
1237 rb->rg_data = kmalloc(size, XPRTRDMA_GFP_FLAGS);
1243 rb->rg_device = NULL;
1244 rb->rg_direction = direction;
1245 rb->rg_iov.length = size;
1250 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1251 * @rb: regbuf to reallocate
1252 * @size: size of buffer to be allocated, in bytes
1255 * Returns true if reallocation was successful. If false is
1256 * returned, @rb is left untouched.
1258 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1262 buf = kmalloc(size, flags);
1266 rpcrdma_regbuf_dma_unmap(rb);
1270 rb->rg_iov.length = size;
1275 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1276 * @r_xprt: controlling transport instance
1277 * @rb: regbuf to be mapped
1279 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1281 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1282 struct rpcrdma_regbuf *rb)
1284 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1286 if (rb->rg_direction == DMA_NONE)
1289 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1290 rdmab_length(rb), rb->rg_direction);
1291 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1292 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1296 rb->rg_device = device;
1297 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1301 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1306 if (!rpcrdma_regbuf_is_mapped(rb))
1309 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1311 rb->rg_device = NULL;
1314 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1316 rpcrdma_regbuf_dma_unmap(rb);
1323 * rpcrdma_post_recvs - Refill the Receive Queue
1324 * @r_xprt: controlling transport instance
1325 * @needed: current credit grant
1326 * @temp: mark Receive buffers to be deleted after one use
1329 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1331 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1332 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1333 struct ib_recv_wr *wr, *bad_wr;
1334 struct rpcrdma_rep *rep;
1340 if (likely(ep->re_receive_count > needed))
1342 needed -= ep->re_receive_count;
1344 needed += RPCRDMA_MAX_RECV_BATCH;
1346 if (atomic_inc_return(&ep->re_receiving) > 1)
1349 /* fast path: all needed reps can be found on the free list */
1352 rep = rpcrdma_rep_get_locked(buf);
1353 if (rep && rep->rr_temp) {
1354 rpcrdma_rep_destroy(rep);
1358 rep = rpcrdma_rep_create(r_xprt, temp);
1361 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf)) {
1362 rpcrdma_rep_put(buf, rep);
1366 rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1367 trace_xprtrdma_post_recv(&rep->rr_cid);
1368 rep->rr_recv_wr.next = wr;
1369 wr = &rep->rr_recv_wr;
1376 rc = ib_post_recv(ep->re_id->qp, wr,
1377 (const struct ib_recv_wr **)&bad_wr);
1379 trace_xprtrdma_post_recvs_err(r_xprt, rc);
1380 for (wr = bad_wr; wr;) {
1381 struct rpcrdma_rep *rep;
1383 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1385 rpcrdma_rep_put(buf, rep);
1389 if (atomic_dec_return(&ep->re_receiving) > 0)
1390 complete(&ep->re_done);
1393 trace_xprtrdma_post_recvs(r_xprt, count);
1394 ep->re_receive_count += count;