Merge tag 'nfs-for-4.19-1' of git://git.linux-nfs.org/projects/anna/linux-nfs
[sfrench/cifs-2.6.git] / net / sunrpc / xprtrdma / verbs.c
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  *
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
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
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.
23  *
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
27  *      permission.
28  *
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.
40  */
41
42 /*
43  * verbs.c
44  *
45  * Encapsulates the major functions managing:
46  *  o adapters
47  *  o endpoints
48  *  o connections
49  *  o buffer memory
50  */
51
52 #include <linux/interrupt.h>
53 #include <linux/slab.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/sunrpc/svc_rdma.h>
56
57 #include <asm-generic/barrier.h>
58 #include <asm/bitops.h>
59
60 #include <rdma/ib_cm.h>
61
62 #include "xprt_rdma.h"
63 #include <trace/events/rpcrdma.h>
64
65 /*
66  * Globals/Macros
67  */
68
69 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
70 # define RPCDBG_FACILITY        RPCDBG_TRANS
71 #endif
72
73 /*
74  * internal functions
75  */
76 static void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
77 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
79 static int rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp);
80 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
81
82 struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
83
84 int
85 rpcrdma_alloc_wq(void)
86 {
87         struct workqueue_struct *recv_wq;
88
89         recv_wq = alloc_workqueue("xprtrdma_receive",
90                                   WQ_MEM_RECLAIM | WQ_HIGHPRI,
91                                   0);
92         if (!recv_wq)
93                 return -ENOMEM;
94
95         rpcrdma_receive_wq = recv_wq;
96         return 0;
97 }
98
99 void
100 rpcrdma_destroy_wq(void)
101 {
102         struct workqueue_struct *wq;
103
104         if (rpcrdma_receive_wq) {
105                 wq = rpcrdma_receive_wq;
106                 rpcrdma_receive_wq = NULL;
107                 destroy_workqueue(wq);
108         }
109 }
110
111 static void
112 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
113 {
114         struct rpcrdma_ep *ep = context;
115         struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
116                                                    rx_ep);
117
118         trace_xprtrdma_qp_error(r_xprt, event);
119         pr_err("rpcrdma: %s on device %s ep %p\n",
120                ib_event_msg(event->event), event->device->name, context);
121
122         if (ep->rep_connected == 1) {
123                 ep->rep_connected = -EIO;
124                 rpcrdma_conn_func(ep);
125                 wake_up_all(&ep->rep_connect_wait);
126         }
127 }
128
129 /**
130  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
131  * @cq: completion queue (ignored)
132  * @wc: completed WR
133  *
134  */
135 static void
136 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
137 {
138         struct ib_cqe *cqe = wc->wr_cqe;
139         struct rpcrdma_sendctx *sc =
140                 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
141
142         /* WARNING: Only wr_cqe and status are reliable at this point */
143         trace_xprtrdma_wc_send(sc, wc);
144         if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
145                 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
146                        ib_wc_status_msg(wc->status),
147                        wc->status, wc->vendor_err);
148
149         rpcrdma_sendctx_put_locked(sc);
150 }
151
152 /**
153  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
154  * @cq: completion queue (ignored)
155  * @wc: completed WR
156  *
157  */
158 static void
159 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
160 {
161         struct ib_cqe *cqe = wc->wr_cqe;
162         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
163                                                rr_cqe);
164
165         /* WARNING: Only wr_id and status are reliable at this point */
166         trace_xprtrdma_wc_receive(wc);
167         if (wc->status != IB_WC_SUCCESS)
168                 goto out_fail;
169
170         /* status == SUCCESS means all fields in wc are trustworthy */
171         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
172         rep->rr_wc_flags = wc->wc_flags;
173         rep->rr_inv_rkey = wc->ex.invalidate_rkey;
174
175         ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
176                                    rdmab_addr(rep->rr_rdmabuf),
177                                    wc->byte_len, DMA_FROM_DEVICE);
178
179 out_schedule:
180         rpcrdma_reply_handler(rep);
181         return;
182
183 out_fail:
184         if (wc->status != IB_WC_WR_FLUSH_ERR)
185                 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
186                        ib_wc_status_msg(wc->status),
187                        wc->status, wc->vendor_err);
188         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
189         goto out_schedule;
190 }
191
192 static void
193 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
194                                struct rdma_conn_param *param)
195 {
196         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
197         const struct rpcrdma_connect_private *pmsg = param->private_data;
198         unsigned int rsize, wsize;
199
200         /* Default settings for RPC-over-RDMA Version One */
201         r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
202         rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
203         wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
204
205         if (pmsg &&
206             pmsg->cp_magic == rpcrdma_cmp_magic &&
207             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
208                 r_xprt->rx_ia.ri_implicit_roundup = true;
209                 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
210                 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
211         }
212
213         if (rsize < cdata->inline_rsize)
214                 cdata->inline_rsize = rsize;
215         if (wsize < cdata->inline_wsize)
216                 cdata->inline_wsize = wsize;
217         dprintk("RPC:       %s: max send %u, max recv %u\n",
218                 __func__, cdata->inline_wsize, cdata->inline_rsize);
219         rpcrdma_set_max_header_sizes(r_xprt);
220 }
221
222 static int
223 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
224 {
225         struct rpcrdma_xprt *xprt = id->context;
226         struct rpcrdma_ia *ia = &xprt->rx_ia;
227         struct rpcrdma_ep *ep = &xprt->rx_ep;
228         int connstate = 0;
229
230         trace_xprtrdma_conn_upcall(xprt, event);
231         switch (event->event) {
232         case RDMA_CM_EVENT_ADDR_RESOLVED:
233         case RDMA_CM_EVENT_ROUTE_RESOLVED:
234                 ia->ri_async_rc = 0;
235                 complete(&ia->ri_done);
236                 break;
237         case RDMA_CM_EVENT_ADDR_ERROR:
238                 ia->ri_async_rc = -EPROTO;
239                 complete(&ia->ri_done);
240                 break;
241         case RDMA_CM_EVENT_ROUTE_ERROR:
242                 ia->ri_async_rc = -ENETUNREACH;
243                 complete(&ia->ri_done);
244                 break;
245         case RDMA_CM_EVENT_DEVICE_REMOVAL:
246 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
247                 pr_info("rpcrdma: removing device %s for %s:%s\n",
248                         ia->ri_device->name,
249                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
250 #endif
251                 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
252                 ep->rep_connected = -ENODEV;
253                 xprt_force_disconnect(&xprt->rx_xprt);
254                 wait_for_completion(&ia->ri_remove_done);
255
256                 ia->ri_id = NULL;
257                 ia->ri_device = NULL;
258                 /* Return 1 to ensure the core destroys the id. */
259                 return 1;
260         case RDMA_CM_EVENT_ESTABLISHED:
261                 ++xprt->rx_xprt.connect_cookie;
262                 connstate = 1;
263                 rpcrdma_update_connect_private(xprt, &event->param.conn);
264                 goto connected;
265         case RDMA_CM_EVENT_CONNECT_ERROR:
266                 connstate = -ENOTCONN;
267                 goto connected;
268         case RDMA_CM_EVENT_UNREACHABLE:
269                 connstate = -ENETUNREACH;
270                 goto connected;
271         case RDMA_CM_EVENT_REJECTED:
272                 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
273                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
274                         rdma_reject_msg(id, event->status));
275                 connstate = -ECONNREFUSED;
276                 if (event->status == IB_CM_REJ_STALE_CONN)
277                         connstate = -EAGAIN;
278                 goto connected;
279         case RDMA_CM_EVENT_DISCONNECTED:
280                 ++xprt->rx_xprt.connect_cookie;
281                 connstate = -ECONNABORTED;
282 connected:
283                 ep->rep_connected = connstate;
284                 rpcrdma_conn_func(ep);
285                 wake_up_all(&ep->rep_connect_wait);
286                 /*FALLTHROUGH*/
287         default:
288                 dprintk("RPC:       %s: %s:%s on %s/%s (ep 0x%p): %s\n",
289                         __func__,
290                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
291                         ia->ri_device->name, ia->ri_ops->ro_displayname,
292                         ep, rdma_event_msg(event->event));
293                 break;
294         }
295
296         return 0;
297 }
298
299 static struct rdma_cm_id *
300 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
301 {
302         unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
303         struct rdma_cm_id *id;
304         int rc;
305
306         trace_xprtrdma_conn_start(xprt);
307
308         init_completion(&ia->ri_done);
309         init_completion(&ia->ri_remove_done);
310
311         id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_conn_upcall,
312                             xprt, RDMA_PS_TCP, IB_QPT_RC);
313         if (IS_ERR(id)) {
314                 rc = PTR_ERR(id);
315                 dprintk("RPC:       %s: rdma_create_id() failed %i\n",
316                         __func__, rc);
317                 return id;
318         }
319
320         ia->ri_async_rc = -ETIMEDOUT;
321         rc = rdma_resolve_addr(id, NULL,
322                                (struct sockaddr *)&xprt->rx_xprt.addr,
323                                RDMA_RESOLVE_TIMEOUT);
324         if (rc) {
325                 dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
326                         __func__, rc);
327                 goto out;
328         }
329         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
330         if (rc < 0) {
331                 trace_xprtrdma_conn_tout(xprt);
332                 goto out;
333         }
334
335         rc = ia->ri_async_rc;
336         if (rc)
337                 goto out;
338
339         ia->ri_async_rc = -ETIMEDOUT;
340         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
341         if (rc) {
342                 dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
343                         __func__, rc);
344                 goto out;
345         }
346         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
347         if (rc < 0) {
348                 trace_xprtrdma_conn_tout(xprt);
349                 goto out;
350         }
351         rc = ia->ri_async_rc;
352         if (rc)
353                 goto out;
354
355         return id;
356
357 out:
358         rdma_destroy_id(id);
359         return ERR_PTR(rc);
360 }
361
362 /*
363  * Exported functions.
364  */
365
366 /**
367  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
368  * @xprt: transport with IA to (re)initialize
369  *
370  * Returns 0 on success, negative errno if an appropriate
371  * Interface Adapter could not be found and opened.
372  */
373 int
374 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
375 {
376         struct rpcrdma_ia *ia = &xprt->rx_ia;
377         int rc;
378
379         ia->ri_id = rpcrdma_create_id(xprt, ia);
380         if (IS_ERR(ia->ri_id)) {
381                 rc = PTR_ERR(ia->ri_id);
382                 goto out_err;
383         }
384         ia->ri_device = ia->ri_id->device;
385
386         ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
387         if (IS_ERR(ia->ri_pd)) {
388                 rc = PTR_ERR(ia->ri_pd);
389                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
390                 goto out_err;
391         }
392
393         switch (xprt_rdma_memreg_strategy) {
394         case RPCRDMA_FRWR:
395                 if (frwr_is_supported(ia)) {
396                         ia->ri_ops = &rpcrdma_frwr_memreg_ops;
397                         break;
398                 }
399                 /*FALLTHROUGH*/
400         case RPCRDMA_MTHCAFMR:
401                 if (fmr_is_supported(ia)) {
402                         ia->ri_ops = &rpcrdma_fmr_memreg_ops;
403                         break;
404                 }
405                 /*FALLTHROUGH*/
406         default:
407                 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
408                        ia->ri_device->name, xprt_rdma_memreg_strategy);
409                 rc = -EINVAL;
410                 goto out_err;
411         }
412
413         return 0;
414
415 out_err:
416         rpcrdma_ia_close(ia);
417         return rc;
418 }
419
420 /**
421  * rpcrdma_ia_remove - Handle device driver unload
422  * @ia: interface adapter being removed
423  *
424  * Divest transport H/W resources associated with this adapter,
425  * but allow it to be restored later.
426  */
427 void
428 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
429 {
430         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
431                                                    rx_ia);
432         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
433         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
434         struct rpcrdma_req *req;
435         struct rpcrdma_rep *rep;
436
437         cancel_delayed_work_sync(&buf->rb_refresh_worker);
438
439         /* This is similar to rpcrdma_ep_destroy, but:
440          * - Don't cancel the connect worker.
441          * - Don't call rpcrdma_ep_disconnect, which waits
442          *   for another conn upcall, which will deadlock.
443          * - rdma_disconnect is unneeded, the underlying
444          *   connection is already gone.
445          */
446         if (ia->ri_id->qp) {
447                 ib_drain_qp(ia->ri_id->qp);
448                 rdma_destroy_qp(ia->ri_id);
449                 ia->ri_id->qp = NULL;
450         }
451         ib_free_cq(ep->rep_attr.recv_cq);
452         ep->rep_attr.recv_cq = NULL;
453         ib_free_cq(ep->rep_attr.send_cq);
454         ep->rep_attr.send_cq = NULL;
455
456         /* The ULP is responsible for ensuring all DMA
457          * mappings and MRs are gone.
458          */
459         list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
460                 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
461         list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
462                 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
463                 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
464                 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
465         }
466         rpcrdma_mrs_destroy(buf);
467         ib_dealloc_pd(ia->ri_pd);
468         ia->ri_pd = NULL;
469
470         /* Allow waiters to continue */
471         complete(&ia->ri_remove_done);
472
473         trace_xprtrdma_remove(r_xprt);
474 }
475
476 /**
477  * rpcrdma_ia_close - Clean up/close an IA.
478  * @ia: interface adapter to close
479  *
480  */
481 void
482 rpcrdma_ia_close(struct rpcrdma_ia *ia)
483 {
484         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
485                 if (ia->ri_id->qp)
486                         rdma_destroy_qp(ia->ri_id);
487                 rdma_destroy_id(ia->ri_id);
488         }
489         ia->ri_id = NULL;
490         ia->ri_device = NULL;
491
492         /* If the pd is still busy, xprtrdma missed freeing a resource */
493         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
494                 ib_dealloc_pd(ia->ri_pd);
495         ia->ri_pd = NULL;
496 }
497
498 /*
499  * Create unconnected endpoint.
500  */
501 int
502 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
503                   struct rpcrdma_create_data_internal *cdata)
504 {
505         struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
506         struct ib_cq *sendcq, *recvcq;
507         unsigned int max_sge;
508         int rc;
509
510         max_sge = min_t(unsigned int, ia->ri_device->attrs.max_send_sge,
511                         RPCRDMA_MAX_SEND_SGES);
512         if (max_sge < RPCRDMA_MIN_SEND_SGES) {
513                 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
514                 return -ENOMEM;
515         }
516         ia->ri_max_send_sges = max_sge;
517
518         rc = ia->ri_ops->ro_open(ia, ep, cdata);
519         if (rc)
520                 return rc;
521
522         ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
523         ep->rep_attr.qp_context = ep;
524         ep->rep_attr.srq = NULL;
525         ep->rep_attr.cap.max_send_sge = max_sge;
526         ep->rep_attr.cap.max_recv_sge = 1;
527         ep->rep_attr.cap.max_inline_data = 0;
528         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
529         ep->rep_attr.qp_type = IB_QPT_RC;
530         ep->rep_attr.port_num = ~0;
531
532         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
533                 "iovs: send %d recv %d\n",
534                 __func__,
535                 ep->rep_attr.cap.max_send_wr,
536                 ep->rep_attr.cap.max_recv_wr,
537                 ep->rep_attr.cap.max_send_sge,
538                 ep->rep_attr.cap.max_recv_sge);
539
540         /* set trigger for requesting send completion */
541         ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
542                                    cdata->max_requests >> 2);
543         ep->rep_send_count = ep->rep_send_batch;
544         init_waitqueue_head(&ep->rep_connect_wait);
545         INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
546
547         sendcq = ib_alloc_cq(ia->ri_device, NULL,
548                              ep->rep_attr.cap.max_send_wr + 1,
549                              1, IB_POLL_WORKQUEUE);
550         if (IS_ERR(sendcq)) {
551                 rc = PTR_ERR(sendcq);
552                 dprintk("RPC:       %s: failed to create send CQ: %i\n",
553                         __func__, rc);
554                 goto out1;
555         }
556
557         recvcq = ib_alloc_cq(ia->ri_device, NULL,
558                              ep->rep_attr.cap.max_recv_wr + 1,
559                              0, IB_POLL_WORKQUEUE);
560         if (IS_ERR(recvcq)) {
561                 rc = PTR_ERR(recvcq);
562                 dprintk("RPC:       %s: failed to create recv CQ: %i\n",
563                         __func__, rc);
564                 goto out2;
565         }
566
567         ep->rep_attr.send_cq = sendcq;
568         ep->rep_attr.recv_cq = recvcq;
569
570         /* Initialize cma parameters */
571         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
572
573         /* Prepare RDMA-CM private message */
574         pmsg->cp_magic = rpcrdma_cmp_magic;
575         pmsg->cp_version = RPCRDMA_CMP_VERSION;
576         pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
577         pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
578         pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
579         ep->rep_remote_cma.private_data = pmsg;
580         ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
581
582         /* Client offers RDMA Read but does not initiate */
583         ep->rep_remote_cma.initiator_depth = 0;
584         ep->rep_remote_cma.responder_resources =
585                 min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
586
587         /* Limit transport retries so client can detect server
588          * GID changes quickly. RPC layer handles re-establishing
589          * transport connection and retransmission.
590          */
591         ep->rep_remote_cma.retry_count = 6;
592
593         /* RPC-over-RDMA handles its own flow control. In addition,
594          * make all RNR NAKs visible so we know that RPC-over-RDMA
595          * flow control is working correctly (no NAKs should be seen).
596          */
597         ep->rep_remote_cma.flow_control = 0;
598         ep->rep_remote_cma.rnr_retry_count = 0;
599
600         return 0;
601
602 out2:
603         ib_free_cq(sendcq);
604 out1:
605         return rc;
606 }
607
608 /*
609  * rpcrdma_ep_destroy
610  *
611  * Disconnect and destroy endpoint. After this, the only
612  * valid operations on the ep are to free it (if dynamically
613  * allocated) or re-create it.
614  */
615 void
616 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
617 {
618         cancel_delayed_work_sync(&ep->rep_connect_worker);
619
620         if (ia->ri_id && ia->ri_id->qp) {
621                 rpcrdma_ep_disconnect(ep, ia);
622                 rdma_destroy_qp(ia->ri_id);
623                 ia->ri_id->qp = NULL;
624         }
625
626         if (ep->rep_attr.recv_cq)
627                 ib_free_cq(ep->rep_attr.recv_cq);
628         if (ep->rep_attr.send_cq)
629                 ib_free_cq(ep->rep_attr.send_cq);
630 }
631
632 /* Re-establish a connection after a device removal event.
633  * Unlike a normal reconnection, a fresh PD and a new set
634  * of MRs and buffers is needed.
635  */
636 static int
637 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
638                          struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
639 {
640         int rc, err;
641
642         trace_xprtrdma_reinsert(r_xprt);
643
644         rc = -EHOSTUNREACH;
645         if (rpcrdma_ia_open(r_xprt))
646                 goto out1;
647
648         rc = -ENOMEM;
649         err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
650         if (err) {
651                 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
652                 goto out2;
653         }
654
655         rc = -ENETUNREACH;
656         err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
657         if (err) {
658                 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
659                 goto out3;
660         }
661
662         rpcrdma_mrs_create(r_xprt);
663         return 0;
664
665 out3:
666         rpcrdma_ep_destroy(ep, ia);
667 out2:
668         rpcrdma_ia_close(ia);
669 out1:
670         return rc;
671 }
672
673 static int
674 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
675                      struct rpcrdma_ia *ia)
676 {
677         struct rdma_cm_id *id, *old;
678         int err, rc;
679
680         trace_xprtrdma_reconnect(r_xprt);
681
682         rpcrdma_ep_disconnect(ep, ia);
683
684         rc = -EHOSTUNREACH;
685         id = rpcrdma_create_id(r_xprt, ia);
686         if (IS_ERR(id))
687                 goto out;
688
689         /* As long as the new ID points to the same device as the
690          * old ID, we can reuse the transport's existing PD and all
691          * previously allocated MRs. Also, the same device means
692          * the transport's previous DMA mappings are still valid.
693          *
694          * This is a sanity check only. There should be no way these
695          * point to two different devices here.
696          */
697         old = id;
698         rc = -ENETUNREACH;
699         if (ia->ri_device != id->device) {
700                 pr_err("rpcrdma: can't reconnect on different device!\n");
701                 goto out_destroy;
702         }
703
704         err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
705         if (err) {
706                 dprintk("RPC:       %s: rdma_create_qp returned %d\n",
707                         __func__, err);
708                 goto out_destroy;
709         }
710
711         /* Atomically replace the transport's ID and QP. */
712         rc = 0;
713         old = ia->ri_id;
714         ia->ri_id = id;
715         rdma_destroy_qp(old);
716
717 out_destroy:
718         rdma_destroy_id(old);
719 out:
720         return rc;
721 }
722
723 /*
724  * Connect unconnected endpoint.
725  */
726 int
727 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
728 {
729         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
730                                                    rx_ia);
731         int rc;
732
733 retry:
734         switch (ep->rep_connected) {
735         case 0:
736                 dprintk("RPC:       %s: connecting...\n", __func__);
737                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
738                 if (rc) {
739                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
740                                 __func__, rc);
741                         rc = -ENETUNREACH;
742                         goto out_noupdate;
743                 }
744                 break;
745         case -ENODEV:
746                 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
747                 if (rc)
748                         goto out_noupdate;
749                 break;
750         default:
751                 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
752                 if (rc)
753                         goto out;
754         }
755
756         ep->rep_connected = 0;
757         rpcrdma_post_recvs(r_xprt, true);
758
759         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
760         if (rc) {
761                 dprintk("RPC:       %s: rdma_connect() failed with %i\n",
762                                 __func__, rc);
763                 goto out;
764         }
765
766         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
767         if (ep->rep_connected <= 0) {
768                 if (ep->rep_connected == -EAGAIN)
769                         goto retry;
770                 rc = ep->rep_connected;
771                 goto out;
772         }
773
774         dprintk("RPC:       %s: connected\n", __func__);
775
776 out:
777         if (rc)
778                 ep->rep_connected = rc;
779
780 out_noupdate:
781         return rc;
782 }
783
784 /*
785  * rpcrdma_ep_disconnect
786  *
787  * This is separate from destroy to facilitate the ability
788  * to reconnect without recreating the endpoint.
789  *
790  * This call is not reentrant, and must not be made in parallel
791  * on the same endpoint.
792  */
793 void
794 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
795 {
796         int rc;
797
798         rc = rdma_disconnect(ia->ri_id);
799         if (!rc)
800                 /* returns without wait if not connected */
801                 wait_event_interruptible(ep->rep_connect_wait,
802                                                         ep->rep_connected != 1);
803         else
804                 ep->rep_connected = rc;
805         trace_xprtrdma_disconnect(container_of(ep, struct rpcrdma_xprt,
806                                                rx_ep), rc);
807
808         ib_drain_qp(ia->ri_id->qp);
809 }
810
811 /* Fixed-size circular FIFO queue. This implementation is wait-free and
812  * lock-free.
813  *
814  * Consumer is the code path that posts Sends. This path dequeues a
815  * sendctx for use by a Send operation. Multiple consumer threads
816  * are serialized by the RPC transport lock, which allows only one
817  * ->send_request call at a time.
818  *
819  * Producer is the code path that handles Send completions. This path
820  * enqueues a sendctx that has been completed. Multiple producer
821  * threads are serialized by the ib_poll_cq() function.
822  */
823
824 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
825  * queue activity, and ib_drain_qp has flushed all remaining Send
826  * requests.
827  */
828 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
829 {
830         unsigned long i;
831
832         for (i = 0; i <= buf->rb_sc_last; i++)
833                 kfree(buf->rb_sc_ctxs[i]);
834         kfree(buf->rb_sc_ctxs);
835 }
836
837 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
838 {
839         struct rpcrdma_sendctx *sc;
840
841         sc = kzalloc(sizeof(*sc) +
842                      ia->ri_max_send_sges * sizeof(struct ib_sge),
843                      GFP_KERNEL);
844         if (!sc)
845                 return NULL;
846
847         sc->sc_wr.wr_cqe = &sc->sc_cqe;
848         sc->sc_wr.sg_list = sc->sc_sges;
849         sc->sc_wr.opcode = IB_WR_SEND;
850         sc->sc_cqe.done = rpcrdma_wc_send;
851         return sc;
852 }
853
854 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
855 {
856         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
857         struct rpcrdma_sendctx *sc;
858         unsigned long i;
859
860         /* Maximum number of concurrent outstanding Send WRs. Capping
861          * the circular queue size stops Send Queue overflow by causing
862          * the ->send_request call to fail temporarily before too many
863          * Sends are posted.
864          */
865         i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
866         dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
867         buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
868         if (!buf->rb_sc_ctxs)
869                 return -ENOMEM;
870
871         buf->rb_sc_last = i - 1;
872         for (i = 0; i <= buf->rb_sc_last; i++) {
873                 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
874                 if (!sc)
875                         goto out_destroy;
876
877                 sc->sc_xprt = r_xprt;
878                 buf->rb_sc_ctxs[i] = sc;
879         }
880         buf->rb_flags = 0;
881
882         return 0;
883
884 out_destroy:
885         rpcrdma_sendctxs_destroy(buf);
886         return -ENOMEM;
887 }
888
889 /* The sendctx queue is not guaranteed to have a size that is a
890  * power of two, thus the helpers in circ_buf.h cannot be used.
891  * The other option is to use modulus (%), which can be expensive.
892  */
893 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
894                                           unsigned long item)
895 {
896         return likely(item < buf->rb_sc_last) ? item + 1 : 0;
897 }
898
899 /**
900  * rpcrdma_sendctx_get_locked - Acquire a send context
901  * @buf: transport buffers from which to acquire an unused context
902  *
903  * Returns pointer to a free send completion context; or NULL if
904  * the queue is empty.
905  *
906  * Usage: Called to acquire an SGE array before preparing a Send WR.
907  *
908  * The caller serializes calls to this function (per rpcrdma_buffer),
909  * and provides an effective memory barrier that flushes the new value
910  * of rb_sc_head.
911  */
912 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
913 {
914         struct rpcrdma_xprt *r_xprt;
915         struct rpcrdma_sendctx *sc;
916         unsigned long next_head;
917
918         next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
919
920         if (next_head == READ_ONCE(buf->rb_sc_tail))
921                 goto out_emptyq;
922
923         /* ORDER: item must be accessed _before_ head is updated */
924         sc = buf->rb_sc_ctxs[next_head];
925
926         /* Releasing the lock in the caller acts as a memory
927          * barrier that flushes rb_sc_head.
928          */
929         buf->rb_sc_head = next_head;
930
931         return sc;
932
933 out_emptyq:
934         /* The queue is "empty" if there have not been enough Send
935          * completions recently. This is a sign the Send Queue is
936          * backing up. Cause the caller to pause and try again.
937          */
938         set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
939         r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
940         r_xprt->rx_stats.empty_sendctx_q++;
941         return NULL;
942 }
943
944 /**
945  * rpcrdma_sendctx_put_locked - Release a send context
946  * @sc: send context to release
947  *
948  * Usage: Called from Send completion to return a sendctxt
949  * to the queue.
950  *
951  * The caller serializes calls to this function (per rpcrdma_buffer).
952  */
953 static void
954 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
955 {
956         struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
957         unsigned long next_tail;
958
959         /* Unmap SGEs of previously completed by unsignaled
960          * Sends by walking up the queue until @sc is found.
961          */
962         next_tail = buf->rb_sc_tail;
963         do {
964                 next_tail = rpcrdma_sendctx_next(buf, next_tail);
965
966                 /* ORDER: item must be accessed _before_ tail is updated */
967                 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
968
969         } while (buf->rb_sc_ctxs[next_tail] != sc);
970
971         /* Paired with READ_ONCE */
972         smp_store_release(&buf->rb_sc_tail, next_tail);
973
974         if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
975                 smp_mb__after_atomic();
976                 xprt_write_space(&sc->sc_xprt->rx_xprt);
977         }
978 }
979
980 static void
981 rpcrdma_mr_recovery_worker(struct work_struct *work)
982 {
983         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
984                                                   rb_recovery_worker.work);
985         struct rpcrdma_mr *mr;
986
987         spin_lock(&buf->rb_recovery_lock);
988         while (!list_empty(&buf->rb_stale_mrs)) {
989                 mr = rpcrdma_mr_pop(&buf->rb_stale_mrs);
990                 spin_unlock(&buf->rb_recovery_lock);
991
992                 trace_xprtrdma_recover_mr(mr);
993                 mr->mr_xprt->rx_ia.ri_ops->ro_recover_mr(mr);
994
995                 spin_lock(&buf->rb_recovery_lock);
996         }
997         spin_unlock(&buf->rb_recovery_lock);
998 }
999
1000 void
1001 rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr)
1002 {
1003         struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1004         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1005
1006         spin_lock(&buf->rb_recovery_lock);
1007         rpcrdma_mr_push(mr, &buf->rb_stale_mrs);
1008         spin_unlock(&buf->rb_recovery_lock);
1009
1010         schedule_delayed_work(&buf->rb_recovery_worker, 0);
1011 }
1012
1013 static void
1014 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
1015 {
1016         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1017         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1018         unsigned int count;
1019         LIST_HEAD(free);
1020         LIST_HEAD(all);
1021
1022         for (count = 0; count < 3; count++) {
1023                 struct rpcrdma_mr *mr;
1024                 int rc;
1025
1026                 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1027                 if (!mr)
1028                         break;
1029
1030                 rc = ia->ri_ops->ro_init_mr(ia, mr);
1031                 if (rc) {
1032                         kfree(mr);
1033                         break;
1034                 }
1035
1036                 mr->mr_xprt = r_xprt;
1037
1038                 list_add(&mr->mr_list, &free);
1039                 list_add(&mr->mr_all, &all);
1040         }
1041
1042         spin_lock(&buf->rb_mrlock);
1043         list_splice(&free, &buf->rb_mrs);
1044         list_splice(&all, &buf->rb_all);
1045         r_xprt->rx_stats.mrs_allocated += count;
1046         spin_unlock(&buf->rb_mrlock);
1047         trace_xprtrdma_createmrs(r_xprt, count);
1048
1049         xprt_write_space(&r_xprt->rx_xprt);
1050 }
1051
1052 static void
1053 rpcrdma_mr_refresh_worker(struct work_struct *work)
1054 {
1055         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1056                                                   rb_refresh_worker.work);
1057         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1058                                                    rx_buf);
1059
1060         rpcrdma_mrs_create(r_xprt);
1061 }
1062
1063 struct rpcrdma_req *
1064 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1065 {
1066         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1067         struct rpcrdma_regbuf *rb;
1068         struct rpcrdma_req *req;
1069
1070         req = kzalloc(sizeof(*req), GFP_KERNEL);
1071         if (req == NULL)
1072                 return ERR_PTR(-ENOMEM);
1073
1074         rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1075                                   DMA_TO_DEVICE, GFP_KERNEL);
1076         if (IS_ERR(rb)) {
1077                 kfree(req);
1078                 return ERR_PTR(-ENOMEM);
1079         }
1080         req->rl_rdmabuf = rb;
1081         xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1082         req->rl_buffer = buffer;
1083         INIT_LIST_HEAD(&req->rl_registered);
1084
1085         spin_lock(&buffer->rb_reqslock);
1086         list_add(&req->rl_all, &buffer->rb_allreqs);
1087         spin_unlock(&buffer->rb_reqslock);
1088         return req;
1089 }
1090
1091 static int
1092 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp)
1093 {
1094         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1095         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1096         struct rpcrdma_rep *rep;
1097         int rc;
1098
1099         rc = -ENOMEM;
1100         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1101         if (rep == NULL)
1102                 goto out;
1103
1104         rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1105                                                DMA_FROM_DEVICE, GFP_KERNEL);
1106         if (IS_ERR(rep->rr_rdmabuf)) {
1107                 rc = PTR_ERR(rep->rr_rdmabuf);
1108                 goto out_free;
1109         }
1110         xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1111                      rdmab_length(rep->rr_rdmabuf));
1112
1113         rep->rr_cqe.done = rpcrdma_wc_receive;
1114         rep->rr_rxprt = r_xprt;
1115         INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1116         rep->rr_recv_wr.next = NULL;
1117         rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1118         rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1119         rep->rr_recv_wr.num_sge = 1;
1120         rep->rr_temp = temp;
1121
1122         spin_lock(&buf->rb_lock);
1123         list_add(&rep->rr_list, &buf->rb_recv_bufs);
1124         spin_unlock(&buf->rb_lock);
1125         return 0;
1126
1127 out_free:
1128         kfree(rep);
1129 out:
1130         dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1131                 __func__, rc);
1132         return rc;
1133 }
1134
1135 int
1136 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1137 {
1138         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1139         int i, rc;
1140
1141         buf->rb_max_requests = r_xprt->rx_data.max_requests;
1142         buf->rb_bc_srv_max_requests = 0;
1143         spin_lock_init(&buf->rb_mrlock);
1144         spin_lock_init(&buf->rb_lock);
1145         spin_lock_init(&buf->rb_recovery_lock);
1146         INIT_LIST_HEAD(&buf->rb_mrs);
1147         INIT_LIST_HEAD(&buf->rb_all);
1148         INIT_LIST_HEAD(&buf->rb_stale_mrs);
1149         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1150                           rpcrdma_mr_refresh_worker);
1151         INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1152                           rpcrdma_mr_recovery_worker);
1153
1154         rpcrdma_mrs_create(r_xprt);
1155
1156         INIT_LIST_HEAD(&buf->rb_send_bufs);
1157         INIT_LIST_HEAD(&buf->rb_allreqs);
1158         spin_lock_init(&buf->rb_reqslock);
1159         for (i = 0; i < buf->rb_max_requests; i++) {
1160                 struct rpcrdma_req *req;
1161
1162                 req = rpcrdma_create_req(r_xprt);
1163                 if (IS_ERR(req)) {
1164                         dprintk("RPC:       %s: request buffer %d alloc"
1165                                 " failed\n", __func__, i);
1166                         rc = PTR_ERR(req);
1167                         goto out;
1168                 }
1169                 list_add(&req->rl_list, &buf->rb_send_bufs);
1170         }
1171
1172         buf->rb_credits = 1;
1173         buf->rb_posted_receives = 0;
1174         INIT_LIST_HEAD(&buf->rb_recv_bufs);
1175
1176         rc = rpcrdma_sendctxs_create(r_xprt);
1177         if (rc)
1178                 goto out;
1179
1180         return 0;
1181 out:
1182         rpcrdma_buffer_destroy(buf);
1183         return rc;
1184 }
1185
1186 static void
1187 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1188 {
1189         rpcrdma_free_regbuf(rep->rr_rdmabuf);
1190         kfree(rep);
1191 }
1192
1193 void
1194 rpcrdma_destroy_req(struct rpcrdma_req *req)
1195 {
1196         rpcrdma_free_regbuf(req->rl_recvbuf);
1197         rpcrdma_free_regbuf(req->rl_sendbuf);
1198         rpcrdma_free_regbuf(req->rl_rdmabuf);
1199         kfree(req);
1200 }
1201
1202 static void
1203 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1204 {
1205         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1206                                                    rx_buf);
1207         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1208         struct rpcrdma_mr *mr;
1209         unsigned int count;
1210
1211         count = 0;
1212         spin_lock(&buf->rb_mrlock);
1213         while (!list_empty(&buf->rb_all)) {
1214                 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1215                 list_del(&mr->mr_all);
1216
1217                 spin_unlock(&buf->rb_mrlock);
1218
1219                 /* Ensure MW is not on any rl_registered list */
1220                 if (!list_empty(&mr->mr_list))
1221                         list_del(&mr->mr_list);
1222
1223                 ia->ri_ops->ro_release_mr(mr);
1224                 count++;
1225                 spin_lock(&buf->rb_mrlock);
1226         }
1227         spin_unlock(&buf->rb_mrlock);
1228         r_xprt->rx_stats.mrs_allocated = 0;
1229
1230         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1231 }
1232
1233 void
1234 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1235 {
1236         cancel_delayed_work_sync(&buf->rb_recovery_worker);
1237         cancel_delayed_work_sync(&buf->rb_refresh_worker);
1238
1239         rpcrdma_sendctxs_destroy(buf);
1240
1241         while (!list_empty(&buf->rb_recv_bufs)) {
1242                 struct rpcrdma_rep *rep;
1243
1244                 rep = list_first_entry(&buf->rb_recv_bufs,
1245                                        struct rpcrdma_rep, rr_list);
1246                 list_del(&rep->rr_list);
1247                 rpcrdma_destroy_rep(rep);
1248         }
1249
1250         spin_lock(&buf->rb_reqslock);
1251         while (!list_empty(&buf->rb_allreqs)) {
1252                 struct rpcrdma_req *req;
1253
1254                 req = list_first_entry(&buf->rb_allreqs,
1255                                        struct rpcrdma_req, rl_all);
1256                 list_del(&req->rl_all);
1257
1258                 spin_unlock(&buf->rb_reqslock);
1259                 rpcrdma_destroy_req(req);
1260                 spin_lock(&buf->rb_reqslock);
1261         }
1262         spin_unlock(&buf->rb_reqslock);
1263
1264         rpcrdma_mrs_destroy(buf);
1265 }
1266
1267 /**
1268  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1269  * @r_xprt: controlling transport
1270  *
1271  * Returns an initialized rpcrdma_mr or NULL if no free
1272  * rpcrdma_mr objects are available.
1273  */
1274 struct rpcrdma_mr *
1275 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1276 {
1277         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1278         struct rpcrdma_mr *mr = NULL;
1279
1280         spin_lock(&buf->rb_mrlock);
1281         if (!list_empty(&buf->rb_mrs))
1282                 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1283         spin_unlock(&buf->rb_mrlock);
1284
1285         if (!mr)
1286                 goto out_nomrs;
1287         return mr;
1288
1289 out_nomrs:
1290         trace_xprtrdma_nomrs(r_xprt);
1291         if (r_xprt->rx_ep.rep_connected != -ENODEV)
1292                 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1293
1294         /* Allow the reply handler and refresh worker to run */
1295         cond_resched();
1296
1297         return NULL;
1298 }
1299
1300 static void
1301 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1302 {
1303         spin_lock(&buf->rb_mrlock);
1304         rpcrdma_mr_push(mr, &buf->rb_mrs);
1305         spin_unlock(&buf->rb_mrlock);
1306 }
1307
1308 /**
1309  * rpcrdma_mr_put - Release an rpcrdma_mr object
1310  * @mr: object to release
1311  *
1312  */
1313 void
1314 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1315 {
1316         __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1317 }
1318
1319 /**
1320  * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1321  * @mr: object to release
1322  *
1323  */
1324 void
1325 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1326 {
1327         struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1328
1329         trace_xprtrdma_dma_unmap(mr);
1330         ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1331                         mr->mr_sg, mr->mr_nents, mr->mr_dir);
1332         __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1333 }
1334
1335 /**
1336  * rpcrdma_buffer_get - Get a request buffer
1337  * @buffers: Buffer pool from which to obtain a buffer
1338  *
1339  * Returns a fresh rpcrdma_req, or NULL if none are available.
1340  */
1341 struct rpcrdma_req *
1342 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1343 {
1344         struct rpcrdma_req *req;
1345
1346         spin_lock(&buffers->rb_lock);
1347         req = list_first_entry_or_null(&buffers->rb_send_bufs,
1348                                        struct rpcrdma_req, rl_list);
1349         if (req)
1350                 list_del_init(&req->rl_list);
1351         spin_unlock(&buffers->rb_lock);
1352         return req;
1353 }
1354
1355 /**
1356  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1357  * @req: object to return
1358  *
1359  */
1360 void
1361 rpcrdma_buffer_put(struct rpcrdma_req *req)
1362 {
1363         struct rpcrdma_buffer *buffers = req->rl_buffer;
1364         struct rpcrdma_rep *rep = req->rl_reply;
1365
1366         req->rl_reply = NULL;
1367
1368         spin_lock(&buffers->rb_lock);
1369         list_add(&req->rl_list, &buffers->rb_send_bufs);
1370         if (rep) {
1371                 if (!rep->rr_temp) {
1372                         list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1373                         rep = NULL;
1374                 }
1375         }
1376         spin_unlock(&buffers->rb_lock);
1377         if (rep)
1378                 rpcrdma_destroy_rep(rep);
1379 }
1380
1381 /*
1382  * Put reply buffers back into pool when not attached to
1383  * request. This happens in error conditions.
1384  */
1385 void
1386 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1387 {
1388         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1389
1390         if (!rep->rr_temp) {
1391                 spin_lock(&buffers->rb_lock);
1392                 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1393                 spin_unlock(&buffers->rb_lock);
1394         } else {
1395                 rpcrdma_destroy_rep(rep);
1396         }
1397 }
1398
1399 /**
1400  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1401  * @size: size of buffer to be allocated, in bytes
1402  * @direction: direction of data movement
1403  * @flags: GFP flags
1404  *
1405  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1406  * can be persistently DMA-mapped for I/O.
1407  *
1408  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1409  * receiving the payload of RDMA RECV operations. During Long Calls
1410  * or Replies they may be registered externally via ro_map.
1411  */
1412 struct rpcrdma_regbuf *
1413 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1414                      gfp_t flags)
1415 {
1416         struct rpcrdma_regbuf *rb;
1417
1418         rb = kmalloc(sizeof(*rb) + size, flags);
1419         if (rb == NULL)
1420                 return ERR_PTR(-ENOMEM);
1421
1422         rb->rg_device = NULL;
1423         rb->rg_direction = direction;
1424         rb->rg_iov.length = size;
1425
1426         return rb;
1427 }
1428
1429 /**
1430  * __rpcrdma_map_regbuf - DMA-map a regbuf
1431  * @ia: controlling rpcrdma_ia
1432  * @rb: regbuf to be mapped
1433  */
1434 bool
1435 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1436 {
1437         struct ib_device *device = ia->ri_device;
1438
1439         if (rb->rg_direction == DMA_NONE)
1440                 return false;
1441
1442         rb->rg_iov.addr = ib_dma_map_single(device,
1443                                             (void *)rb->rg_base,
1444                                             rdmab_length(rb),
1445                                             rb->rg_direction);
1446         if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1447                 return false;
1448
1449         rb->rg_device = device;
1450         rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1451         return true;
1452 }
1453
1454 static void
1455 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1456 {
1457         if (!rb)
1458                 return;
1459
1460         if (!rpcrdma_regbuf_is_mapped(rb))
1461                 return;
1462
1463         ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1464                             rdmab_length(rb), rb->rg_direction);
1465         rb->rg_device = NULL;
1466 }
1467
1468 /**
1469  * rpcrdma_free_regbuf - deregister and free registered buffer
1470  * @rb: regbuf to be deregistered and freed
1471  */
1472 void
1473 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1474 {
1475         rpcrdma_dma_unmap_regbuf(rb);
1476         kfree(rb);
1477 }
1478
1479 /*
1480  * Prepost any receive buffer, then post send.
1481  *
1482  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1483  */
1484 int
1485 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1486                 struct rpcrdma_ep *ep,
1487                 struct rpcrdma_req *req)
1488 {
1489         struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1490         int rc;
1491
1492         if (!ep->rep_send_count ||
1493             test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1494                 send_wr->send_flags |= IB_SEND_SIGNALED;
1495                 ep->rep_send_count = ep->rep_send_batch;
1496         } else {
1497                 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1498                 --ep->rep_send_count;
1499         }
1500
1501         rc = ia->ri_ops->ro_send(ia, req);
1502         trace_xprtrdma_post_send(req, rc);
1503         if (rc)
1504                 return -ENOTCONN;
1505         return 0;
1506 }
1507
1508 /**
1509  * rpcrdma_post_recvs - Maybe post some Receive buffers
1510  * @r_xprt: controlling transport
1511  * @temp: when true, allocate temp rpcrdma_rep objects
1512  *
1513  */
1514 void
1515 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1516 {
1517         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1518         struct ib_recv_wr *wr, *bad_wr;
1519         int needed, count, rc;
1520
1521         needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1522         if (buf->rb_posted_receives > needed)
1523                 return;
1524         needed -= buf->rb_posted_receives;
1525
1526         count = 0;
1527         wr = NULL;
1528         while (needed) {
1529                 struct rpcrdma_regbuf *rb;
1530                 struct rpcrdma_rep *rep;
1531
1532                 spin_lock(&buf->rb_lock);
1533                 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1534                                                struct rpcrdma_rep, rr_list);
1535                 if (likely(rep))
1536                         list_del(&rep->rr_list);
1537                 spin_unlock(&buf->rb_lock);
1538                 if (!rep) {
1539                         if (rpcrdma_create_rep(r_xprt, temp))
1540                                 break;
1541                         continue;
1542                 }
1543
1544                 rb = rep->rr_rdmabuf;
1545                 if (!rpcrdma_regbuf_is_mapped(rb)) {
1546                         if (!__rpcrdma_dma_map_regbuf(&r_xprt->rx_ia, rb)) {
1547                                 rpcrdma_recv_buffer_put(rep);
1548                                 break;
1549                         }
1550                 }
1551
1552                 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1553                 rep->rr_recv_wr.next = wr;
1554                 wr = &rep->rr_recv_wr;
1555                 ++count;
1556                 --needed;
1557         }
1558         if (!count)
1559                 return;
1560
1561         rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1562                           (const struct ib_recv_wr **)&bad_wr);
1563         if (rc) {
1564                 for (wr = bad_wr; wr; wr = wr->next) {
1565                         struct rpcrdma_rep *rep;
1566
1567                         rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1568                         rpcrdma_recv_buffer_put(rep);
1569                         --count;
1570                 }
1571         }
1572         buf->rb_posted_receives += count;
1573         trace_xprtrdma_post_recvs(r_xprt, count, rc);
1574 }