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 static 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 ep->re_implicit_roundup = xprt_rdma_pad_optimize;
209 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
210 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
213 pmsg->cp_magic == rpcrdma_cmp_magic &&
214 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
215 ep->re_implicit_roundup = true;
216 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
217 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
220 if (rsize < ep->re_inline_recv)
221 ep->re_inline_recv = rsize;
222 if (wsize < ep->re_inline_send)
223 ep->re_inline_send = wsize;
225 rpcrdma_set_max_header_sizes(ep);
229 * rpcrdma_cm_event_handler - Handle RDMA CM events
230 * @id: rdma_cm_id on which an event has occurred
231 * @event: details of the event
233 * Called with @id's mutex held. Returns 1 if caller should
234 * destroy @id, otherwise 0.
237 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
239 struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
240 struct rpcrdma_ep *ep = id->context;
244 switch (event->event) {
245 case RDMA_CM_EVENT_ADDR_RESOLVED:
246 case RDMA_CM_EVENT_ROUTE_RESOLVED:
248 complete(&ep->re_done);
250 case RDMA_CM_EVENT_ADDR_ERROR:
251 ep->re_async_rc = -EPROTO;
252 complete(&ep->re_done);
254 case RDMA_CM_EVENT_ROUTE_ERROR:
255 ep->re_async_rc = -ENETUNREACH;
256 complete(&ep->re_done);
258 case RDMA_CM_EVENT_DEVICE_REMOVAL:
259 pr_info("rpcrdma: removing device %s for %pISpc\n",
260 ep->re_id->device->name, sap);
262 case RDMA_CM_EVENT_ADDR_CHANGE:
263 ep->re_connect_status = -ENODEV;
265 case RDMA_CM_EVENT_ESTABLISHED:
267 ep->re_connect_status = 1;
268 rpcrdma_update_cm_private(ep, &event->param.conn);
269 trace_xprtrdma_inline_thresh(ep);
270 wake_up_all(&ep->re_connect_wait);
272 case RDMA_CM_EVENT_CONNECT_ERROR:
273 ep->re_connect_status = -ENOTCONN;
274 goto wake_connect_worker;
275 case RDMA_CM_EVENT_UNREACHABLE:
276 ep->re_connect_status = -ENETUNREACH;
277 goto wake_connect_worker;
278 case RDMA_CM_EVENT_REJECTED:
279 dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
280 sap, rdma_reject_msg(id, event->status));
281 ep->re_connect_status = -ECONNREFUSED;
282 if (event->status == IB_CM_REJ_STALE_CONN)
283 ep->re_connect_status = -ENOTCONN;
285 wake_up_all(&ep->re_connect_wait);
287 case RDMA_CM_EVENT_DISCONNECTED:
288 ep->re_connect_status = -ECONNABORTED;
290 rpcrdma_force_disconnect(ep);
291 return rpcrdma_ep_put(ep);
296 dprintk("RPC: %s: %pISpc on %s/frwr: %s\n", __func__, sap,
297 ep->re_id->device->name, rdma_event_msg(event->event));
301 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
302 struct rpcrdma_ep *ep)
304 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
305 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
306 struct rdma_cm_id *id;
309 init_completion(&ep->re_done);
311 id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
312 RDMA_PS_TCP, IB_QPT_RC);
316 ep->re_async_rc = -ETIMEDOUT;
317 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
318 RDMA_RESOLVE_TIMEOUT);
321 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
325 rc = ep->re_async_rc;
329 ep->re_async_rc = -ETIMEDOUT;
330 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
333 rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
336 rc = ep->re_async_rc;
347 static void rpcrdma_ep_destroy(struct kref *kref)
349 struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
352 rdma_destroy_qp(ep->re_id);
353 ep->re_id->qp = NULL;
356 if (ep->re_attr.recv_cq)
357 ib_free_cq(ep->re_attr.recv_cq);
358 ep->re_attr.recv_cq = NULL;
359 if (ep->re_attr.send_cq)
360 ib_free_cq(ep->re_attr.send_cq);
361 ep->re_attr.send_cq = NULL;
364 ib_dealloc_pd(ep->re_pd);
368 module_put(THIS_MODULE);
371 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
373 kref_get(&ep->re_kref);
377 * %0 if @ep still has a positive kref count, or
378 * %1 if @ep was destroyed successfully.
380 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
382 return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
385 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
387 struct rpcrdma_connect_private *pmsg;
388 struct ib_device *device;
389 struct rdma_cm_id *id;
390 struct rpcrdma_ep *ep;
393 ep = kzalloc(sizeof(*ep), GFP_NOFS);
396 ep->re_xprt = &r_xprt->rx_xprt;
397 kref_init(&ep->re_kref);
399 id = rpcrdma_create_id(r_xprt, ep);
404 __module_get(THIS_MODULE);
407 reinit_completion(&ep->re_done);
409 ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
410 ep->re_inline_send = xprt_rdma_max_inline_write;
411 ep->re_inline_recv = xprt_rdma_max_inline_read;
412 rc = frwr_query_device(ep, device);
416 r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
418 ep->re_attr.srq = NULL;
419 ep->re_attr.cap.max_inline_data = 0;
420 ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
421 ep->re_attr.qp_type = IB_QPT_RC;
422 ep->re_attr.port_num = ~0;
424 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
425 "iovs: send %d recv %d\n",
427 ep->re_attr.cap.max_send_wr,
428 ep->re_attr.cap.max_recv_wr,
429 ep->re_attr.cap.max_send_sge,
430 ep->re_attr.cap.max_recv_sge);
432 ep->re_send_batch = ep->re_max_requests >> 3;
433 ep->re_send_count = ep->re_send_batch;
434 init_waitqueue_head(&ep->re_connect_wait);
436 ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
437 ep->re_attr.cap.max_send_wr,
439 if (IS_ERR(ep->re_attr.send_cq)) {
440 rc = PTR_ERR(ep->re_attr.send_cq);
444 ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
445 ep->re_attr.cap.max_recv_wr,
447 if (IS_ERR(ep->re_attr.recv_cq)) {
448 rc = PTR_ERR(ep->re_attr.recv_cq);
451 ep->re_receive_count = 0;
453 /* Initialize cma parameters */
454 memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
456 /* Prepare RDMA-CM private message */
457 pmsg = &ep->re_cm_private;
458 pmsg->cp_magic = rpcrdma_cmp_magic;
459 pmsg->cp_version = RPCRDMA_CMP_VERSION;
460 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
461 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
462 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
463 ep->re_remote_cma.private_data = pmsg;
464 ep->re_remote_cma.private_data_len = sizeof(*pmsg);
466 /* Client offers RDMA Read but does not initiate */
467 ep->re_remote_cma.initiator_depth = 0;
468 ep->re_remote_cma.responder_resources =
469 min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
471 /* Limit transport retries so client can detect server
472 * GID changes quickly. RPC layer handles re-establishing
473 * transport connection and retransmission.
475 ep->re_remote_cma.retry_count = 6;
477 /* RPC-over-RDMA handles its own flow control. In addition,
478 * make all RNR NAKs visible so we know that RPC-over-RDMA
479 * flow control is working correctly (no NAKs should be seen).
481 ep->re_remote_cma.flow_control = 0;
482 ep->re_remote_cma.rnr_retry_count = 0;
484 ep->re_pd = ib_alloc_pd(device, 0);
485 if (IS_ERR(ep->re_pd)) {
486 rc = PTR_ERR(ep->re_pd);
490 rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
504 * rpcrdma_xprt_connect - Connect an unconnected transport
505 * @r_xprt: controlling transport instance
507 * Returns 0 on success or a negative errno.
509 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
511 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
512 struct rpcrdma_ep *ep;
515 rc = rpcrdma_ep_create(r_xprt);
520 xprt_clear_connected(xprt);
521 rpcrdma_reset_cwnd(r_xprt);
523 /* Bump the ep's reference count while there are
524 * outstanding Receives.
527 rpcrdma_post_recvs(r_xprt, 1, true);
529 rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
533 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
534 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
535 wait_event_interruptible(ep->re_connect_wait,
536 ep->re_connect_status != 0);
537 if (ep->re_connect_status <= 0) {
538 rc = ep->re_connect_status;
542 rc = rpcrdma_sendctxs_create(r_xprt);
548 rc = rpcrdma_reqs_setup(r_xprt);
553 rpcrdma_mrs_create(r_xprt);
556 trace_xprtrdma_connect(r_xprt, rc);
561 * rpcrdma_xprt_disconnect - Disconnect underlying transport
562 * @r_xprt: controlling transport instance
564 * Caller serializes. Either the transport send lock is held,
565 * or we're being called to destroy the transport.
567 * On return, @r_xprt is completely divested of all hardware
568 * resources and prepared for the next ->connect operation.
570 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
572 struct rpcrdma_ep *ep = r_xprt->rx_ep;
573 struct rdma_cm_id *id;
580 rc = rdma_disconnect(id);
581 trace_xprtrdma_disconnect(r_xprt, rc);
583 rpcrdma_xprt_drain(r_xprt);
584 rpcrdma_reps_unmap(r_xprt);
585 rpcrdma_reqs_reset(r_xprt);
586 rpcrdma_mrs_destroy(r_xprt);
587 rpcrdma_sendctxs_destroy(r_xprt);
589 if (rpcrdma_ep_put(ep))
592 r_xprt->rx_ep = NULL;
595 /* Fixed-size circular FIFO queue. This implementation is wait-free and
598 * Consumer is the code path that posts Sends. This path dequeues a
599 * sendctx for use by a Send operation. Multiple consumer threads
600 * are serialized by the RPC transport lock, which allows only one
601 * ->send_request call at a time.
603 * Producer is the code path that handles Send completions. This path
604 * enqueues a sendctx that has been completed. Multiple producer
605 * threads are serialized by the ib_poll_cq() function.
608 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
609 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
612 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
614 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
617 if (!buf->rb_sc_ctxs)
619 for (i = 0; i <= buf->rb_sc_last; i++)
620 kfree(buf->rb_sc_ctxs[i]);
621 kfree(buf->rb_sc_ctxs);
622 buf->rb_sc_ctxs = NULL;
625 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
627 struct rpcrdma_sendctx *sc;
629 sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
634 sc->sc_cqe.done = rpcrdma_wc_send;
635 sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
636 sc->sc_cid.ci_completion_id =
637 atomic_inc_return(&ep->re_completion_ids);
641 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
643 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
644 struct rpcrdma_sendctx *sc;
647 /* Maximum number of concurrent outstanding Send WRs. Capping
648 * the circular queue size stops Send Queue overflow by causing
649 * the ->send_request call to fail temporarily before too many
652 i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
653 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
654 if (!buf->rb_sc_ctxs)
657 buf->rb_sc_last = i - 1;
658 for (i = 0; i <= buf->rb_sc_last; i++) {
659 sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
663 buf->rb_sc_ctxs[i] = sc;
671 /* The sendctx queue is not guaranteed to have a size that is a
672 * power of two, thus the helpers in circ_buf.h cannot be used.
673 * The other option is to use modulus (%), which can be expensive.
675 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
678 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
682 * rpcrdma_sendctx_get_locked - Acquire a send context
683 * @r_xprt: controlling transport instance
685 * Returns pointer to a free send completion context; or NULL if
686 * the queue is empty.
688 * Usage: Called to acquire an SGE array before preparing a Send WR.
690 * The caller serializes calls to this function (per transport), and
691 * provides an effective memory barrier that flushes the new value
694 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
696 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
697 struct rpcrdma_sendctx *sc;
698 unsigned long next_head;
700 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
702 if (next_head == READ_ONCE(buf->rb_sc_tail))
705 /* ORDER: item must be accessed _before_ head is updated */
706 sc = buf->rb_sc_ctxs[next_head];
708 /* Releasing the lock in the caller acts as a memory
709 * barrier that flushes rb_sc_head.
711 buf->rb_sc_head = next_head;
716 /* The queue is "empty" if there have not been enough Send
717 * completions recently. This is a sign the Send Queue is
718 * backing up. Cause the caller to pause and try again.
720 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
721 r_xprt->rx_stats.empty_sendctx_q++;
726 * rpcrdma_sendctx_put_locked - Release a send context
727 * @r_xprt: controlling transport instance
728 * @sc: send context to release
730 * Usage: Called from Send completion to return a sendctxt
733 * The caller serializes calls to this function (per transport).
735 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
736 struct rpcrdma_sendctx *sc)
738 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
739 unsigned long next_tail;
741 /* Unmap SGEs of previously completed but unsignaled
742 * Sends by walking up the queue until @sc is found.
744 next_tail = buf->rb_sc_tail;
746 next_tail = rpcrdma_sendctx_next(buf, next_tail);
748 /* ORDER: item must be accessed _before_ tail is updated */
749 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
751 } while (buf->rb_sc_ctxs[next_tail] != sc);
753 /* Paired with READ_ONCE */
754 smp_store_release(&buf->rb_sc_tail, next_tail);
756 xprt_write_space(&r_xprt->rx_xprt);
760 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
762 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
763 struct rpcrdma_ep *ep = r_xprt->rx_ep;
766 for (count = 0; count < ep->re_max_rdma_segs; count++) {
767 struct rpcrdma_mr *mr;
770 mr = kzalloc(sizeof(*mr), GFP_NOFS);
774 rc = frwr_mr_init(r_xprt, mr);
780 spin_lock(&buf->rb_lock);
781 rpcrdma_mr_push(mr, &buf->rb_mrs);
782 list_add(&mr->mr_all, &buf->rb_all_mrs);
783 spin_unlock(&buf->rb_lock);
786 r_xprt->rx_stats.mrs_allocated += count;
787 trace_xprtrdma_createmrs(r_xprt, count);
791 rpcrdma_mr_refresh_worker(struct work_struct *work)
793 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
795 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
798 rpcrdma_mrs_create(r_xprt);
799 xprt_write_space(&r_xprt->rx_xprt);
803 * rpcrdma_mrs_refresh - Wake the MR refresh worker
804 * @r_xprt: controlling transport instance
807 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
809 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
810 struct rpcrdma_ep *ep = r_xprt->rx_ep;
812 /* If there is no underlying connection, it's no use
813 * to wake the refresh worker.
815 if (ep->re_connect_status == 1) {
816 /* The work is scheduled on a WQ_MEM_RECLAIM
817 * workqueue in order to prevent MR allocation
818 * from recursing into NFS during direct reclaim.
820 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
825 * rpcrdma_req_create - Allocate an rpcrdma_req object
826 * @r_xprt: controlling r_xprt
827 * @size: initial size, in bytes, of send and receive buffers
828 * @flags: GFP flags passed to memory allocators
830 * Returns an allocated and fully initialized rpcrdma_req or NULL.
832 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
835 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
836 struct rpcrdma_req *req;
838 req = kzalloc(sizeof(*req), flags);
842 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
843 if (!req->rl_sendbuf)
846 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
847 if (!req->rl_recvbuf)
850 INIT_LIST_HEAD(&req->rl_free_mrs);
851 INIT_LIST_HEAD(&req->rl_registered);
852 spin_lock(&buffer->rb_lock);
853 list_add(&req->rl_all, &buffer->rb_allreqs);
854 spin_unlock(&buffer->rb_lock);
858 kfree(req->rl_sendbuf);
866 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
867 * @r_xprt: controlling transport instance
868 * @req: rpcrdma_req object to set up
870 * Returns zero on success, and a negative errno on failure.
872 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
874 struct rpcrdma_regbuf *rb;
877 /* Compute maximum header buffer size in bytes */
878 maxhdrsize = rpcrdma_fixed_maxsz + 3 +
879 r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
880 maxhdrsize *= sizeof(__be32);
881 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
882 DMA_TO_DEVICE, GFP_KERNEL);
886 if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
889 req->rl_rdmabuf = rb;
890 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
894 rpcrdma_regbuf_free(rb);
899 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
900 * and thus can be walked without holding rb_lock. Eg. the
901 * caller is holding the transport send lock to exclude
902 * device removal or disconnection.
904 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
906 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
907 struct rpcrdma_req *req;
910 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
911 rc = rpcrdma_req_setup(r_xprt, req);
918 static void rpcrdma_req_reset(struct rpcrdma_req *req)
920 /* Credits are valid for only one connection */
921 req->rl_slot.rq_cong = 0;
923 rpcrdma_regbuf_free(req->rl_rdmabuf);
924 req->rl_rdmabuf = NULL;
926 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
927 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
932 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
933 * and thus can be walked without holding rb_lock. Eg. the
934 * caller is holding the transport send lock to exclude
935 * device removal or disconnection.
937 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
939 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
940 struct rpcrdma_req *req;
942 list_for_each_entry(req, &buf->rb_allreqs, rl_all)
943 rpcrdma_req_reset(req);
947 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
950 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
951 struct rpcrdma_rep *rep;
953 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
957 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
958 DMA_FROM_DEVICE, GFP_KERNEL);
959 if (!rep->rr_rdmabuf)
962 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
963 goto out_free_regbuf;
965 rep->rr_cid.ci_completion_id =
966 atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
968 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
969 rdmab_length(rep->rr_rdmabuf));
970 rep->rr_cqe.done = rpcrdma_wc_receive;
971 rep->rr_rxprt = r_xprt;
972 rep->rr_recv_wr.next = NULL;
973 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
974 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
975 rep->rr_recv_wr.num_sge = 1;
978 spin_lock(&buf->rb_lock);
979 list_add(&rep->rr_all, &buf->rb_all_reps);
980 spin_unlock(&buf->rb_lock);
984 rpcrdma_regbuf_free(rep->rr_rdmabuf);
991 static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
993 rpcrdma_regbuf_free(rep->rr_rdmabuf);
997 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
999 struct rpcrdma_buffer *buf = &rep->rr_rxprt->rx_buf;
1001 spin_lock(&buf->rb_lock);
1002 list_del(&rep->rr_all);
1003 spin_unlock(&buf->rb_lock);
1005 rpcrdma_rep_free(rep);
1008 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1010 struct llist_node *node;
1012 /* Calls to llist_del_first are required to be serialized */
1013 node = llist_del_first(&buf->rb_free_reps);
1016 return llist_entry(node, struct rpcrdma_rep, rr_node);
1020 * rpcrdma_rep_put - Release rpcrdma_rep back to free list
1022 * @rep: rep to release
1025 void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
1027 llist_add(&rep->rr_node, &buf->rb_free_reps);
1030 /* Caller must ensure the QP is quiescent (RQ is drained) before
1031 * invoking this function, to guarantee rb_all_reps is not
1034 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1036 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1037 struct rpcrdma_rep *rep;
1039 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1040 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1041 rep->rr_temp = true; /* Mark this rep for destruction */
1045 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1047 struct rpcrdma_rep *rep;
1049 spin_lock(&buf->rb_lock);
1050 while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1053 list_del(&rep->rr_all);
1054 spin_unlock(&buf->rb_lock);
1056 rpcrdma_rep_free(rep);
1058 spin_lock(&buf->rb_lock);
1060 spin_unlock(&buf->rb_lock);
1064 * rpcrdma_buffer_create - Create initial set of req/rep objects
1065 * @r_xprt: transport instance to (re)initialize
1067 * Returns zero on success, otherwise a negative errno.
1069 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1071 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1074 buf->rb_bc_srv_max_requests = 0;
1075 spin_lock_init(&buf->rb_lock);
1076 INIT_LIST_HEAD(&buf->rb_mrs);
1077 INIT_LIST_HEAD(&buf->rb_all_mrs);
1078 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1080 INIT_LIST_HEAD(&buf->rb_send_bufs);
1081 INIT_LIST_HEAD(&buf->rb_allreqs);
1082 INIT_LIST_HEAD(&buf->rb_all_reps);
1085 for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1086 struct rpcrdma_req *req;
1088 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1092 list_add(&req->rl_list, &buf->rb_send_bufs);
1095 init_llist_head(&buf->rb_free_reps);
1099 rpcrdma_buffer_destroy(buf);
1104 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1105 * @req: unused object to be destroyed
1107 * Relies on caller holding the transport send lock to protect
1108 * removing req->rl_all from buf->rb_all_reqs safely.
1110 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1112 struct rpcrdma_mr *mr;
1114 list_del(&req->rl_all);
1116 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1117 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1119 spin_lock(&buf->rb_lock);
1120 list_del(&mr->mr_all);
1121 spin_unlock(&buf->rb_lock);
1123 frwr_mr_release(mr);
1126 rpcrdma_regbuf_free(req->rl_recvbuf);
1127 rpcrdma_regbuf_free(req->rl_sendbuf);
1128 rpcrdma_regbuf_free(req->rl_rdmabuf);
1133 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1134 * @r_xprt: controlling transport instance
1136 * Relies on caller holding the transport send lock to protect
1137 * removing mr->mr_list from req->rl_free_mrs safely.
1139 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1141 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1142 struct rpcrdma_mr *mr;
1144 cancel_work_sync(&buf->rb_refresh_worker);
1146 spin_lock(&buf->rb_lock);
1147 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1150 list_del(&mr->mr_list);
1151 list_del(&mr->mr_all);
1152 spin_unlock(&buf->rb_lock);
1154 frwr_mr_release(mr);
1156 spin_lock(&buf->rb_lock);
1158 spin_unlock(&buf->rb_lock);
1162 * rpcrdma_buffer_destroy - Release all hw resources
1163 * @buf: root control block for resources
1165 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1166 * - No more Send or Receive completions can occur
1167 * - All MRs, reps, and reqs are returned to their free lists
1170 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1172 rpcrdma_reps_destroy(buf);
1174 while (!list_empty(&buf->rb_send_bufs)) {
1175 struct rpcrdma_req *req;
1177 req = list_first_entry(&buf->rb_send_bufs,
1178 struct rpcrdma_req, rl_list);
1179 list_del(&req->rl_list);
1180 rpcrdma_req_destroy(req);
1185 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1186 * @r_xprt: controlling transport
1188 * Returns an initialized rpcrdma_mr or NULL if no free
1189 * rpcrdma_mr objects are available.
1192 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1194 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1195 struct rpcrdma_mr *mr;
1197 spin_lock(&buf->rb_lock);
1198 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1199 spin_unlock(&buf->rb_lock);
1204 * rpcrdma_reply_put - Put reply buffers back into pool
1205 * @buffers: buffer pool
1206 * @req: object to return
1209 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1211 if (req->rl_reply) {
1212 rpcrdma_rep_put(buffers, req->rl_reply);
1213 req->rl_reply = NULL;
1218 * rpcrdma_buffer_get - Get a request buffer
1219 * @buffers: Buffer pool from which to obtain a buffer
1221 * Returns a fresh rpcrdma_req, or NULL if none are available.
1223 struct rpcrdma_req *
1224 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1226 struct rpcrdma_req *req;
1228 spin_lock(&buffers->rb_lock);
1229 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1230 struct rpcrdma_req, rl_list);
1232 list_del_init(&req->rl_list);
1233 spin_unlock(&buffers->rb_lock);
1238 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1239 * @buffers: buffer pool
1240 * @req: object to return
1243 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1245 rpcrdma_reply_put(buffers, req);
1247 spin_lock(&buffers->rb_lock);
1248 list_add(&req->rl_list, &buffers->rb_send_bufs);
1249 spin_unlock(&buffers->rb_lock);
1252 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1254 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1255 * receiving the payload of RDMA RECV operations. During Long Calls
1256 * or Replies they may be registered externally via frwr_map.
1258 static struct rpcrdma_regbuf *
1259 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1262 struct rpcrdma_regbuf *rb;
1264 rb = kmalloc(sizeof(*rb), flags);
1267 rb->rg_data = kmalloc(size, flags);
1273 rb->rg_device = NULL;
1274 rb->rg_direction = direction;
1275 rb->rg_iov.length = size;
1280 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1281 * @rb: regbuf to reallocate
1282 * @size: size of buffer to be allocated, in bytes
1285 * Returns true if reallocation was successful. If false is
1286 * returned, @rb is left untouched.
1288 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1292 buf = kmalloc(size, flags);
1296 rpcrdma_regbuf_dma_unmap(rb);
1300 rb->rg_iov.length = size;
1305 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1306 * @r_xprt: controlling transport instance
1307 * @rb: regbuf to be mapped
1309 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1311 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1312 struct rpcrdma_regbuf *rb)
1314 struct ib_device *device = r_xprt->rx_ep->re_id->device;
1316 if (rb->rg_direction == DMA_NONE)
1319 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1320 rdmab_length(rb), rb->rg_direction);
1321 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1322 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1326 rb->rg_device = device;
1327 rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1331 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1336 if (!rpcrdma_regbuf_is_mapped(rb))
1339 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1341 rb->rg_device = NULL;
1344 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1346 rpcrdma_regbuf_dma_unmap(rb);
1353 * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1354 * @r_xprt: controlling transport instance
1355 * @req: rpcrdma_req containing the Send WR to post
1357 * Returns 0 if the post was successful, otherwise -ENOTCONN
1360 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1362 if (frwr_send(r_xprt, req))
1368 * rpcrdma_post_recvs - Refill the Receive Queue
1369 * @r_xprt: controlling transport instance
1370 * @needed: current credit grant
1371 * @temp: mark Receive buffers to be deleted after one use
1374 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1376 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1377 struct rpcrdma_ep *ep = r_xprt->rx_ep;
1378 struct ib_recv_wr *wr, *bad_wr;
1379 struct rpcrdma_rep *rep;
1385 if (likely(ep->re_receive_count > needed))
1387 needed -= ep->re_receive_count;
1389 needed += RPCRDMA_MAX_RECV_BATCH;
1391 if (atomic_inc_return(&ep->re_receiving) > 1)
1394 /* fast path: all needed reps can be found on the free list */
1397 rep = rpcrdma_rep_get_locked(buf);
1398 if (rep && rep->rr_temp) {
1399 rpcrdma_rep_destroy(rep);
1403 rep = rpcrdma_rep_create(r_xprt, temp);
1407 rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1408 trace_xprtrdma_post_recv(rep);
1409 rep->rr_recv_wr.next = wr;
1410 wr = &rep->rr_recv_wr;
1417 rc = ib_post_recv(ep->re_id->qp, wr,
1418 (const struct ib_recv_wr **)&bad_wr);
1419 if (atomic_dec_return(&ep->re_receiving) > 0)
1420 complete(&ep->re_done);
1423 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1425 for (wr = bad_wr; wr;) {
1426 struct rpcrdma_rep *rep;
1428 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1430 rpcrdma_rep_put(buf, rep);
1434 ep->re_receive_count += count;