~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/net/sunrpc/xprtrdma/verbs.c

Version: ~ [ linux-5.8-rc4 ] ~ [ linux-5.7.7 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.50 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.131 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.187 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.229 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.229 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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 struct rpcrdma_regbuf *
 80 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
 81                      gfp_t flags);
 82 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
 83 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
 84 static void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
 85 
 86 /* Wait for outstanding transport work to finish. ib_drain_qp
 87  * handles the drains in the wrong order for us, so open code
 88  * them here.
 89  */
 90 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
 91 {
 92         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 93 
 94         /* Flush Receives, then wait for deferred Reply work
 95          * to complete.
 96          */
 97         ib_drain_rq(ia->ri_id->qp);
 98 
 99         /* Deferred Reply processing might have scheduled
100          * local invalidations.
101          */
102         ib_drain_sq(ia->ri_id->qp);
103 }
104 
105 /**
106  * rpcrdma_qp_event_handler - Handle one QP event (error notification)
107  * @event: details of the event
108  * @context: ep that owns QP where event occurred
109  *
110  * Called from the RDMA provider (device driver) possibly in an interrupt
111  * context.
112  */
113 static void
114 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
115 {
116         struct rpcrdma_ep *ep = context;
117         struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
118                                                    rx_ep);
119 
120         trace_xprtrdma_qp_event(r_xprt, event);
121 }
122 
123 /**
124  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
125  * @cq: completion queue (ignored)
126  * @wc: completed WR
127  *
128  */
129 static void
130 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
131 {
132         struct ib_cqe *cqe = wc->wr_cqe;
133         struct rpcrdma_sendctx *sc =
134                 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
135 
136         /* WARNING: Only wr_cqe and status are reliable at this point */
137         trace_xprtrdma_wc_send(sc, wc);
138         rpcrdma_sendctx_put_locked(sc);
139 }
140 
141 /**
142  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
143  * @cq: completion queue (ignored)
144  * @wc: completed WR
145  *
146  */
147 static void
148 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
149 {
150         struct ib_cqe *cqe = wc->wr_cqe;
151         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
152                                                rr_cqe);
153         struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
154 
155         /* WARNING: Only wr_cqe and status are reliable at this point */
156         trace_xprtrdma_wc_receive(wc);
157         --r_xprt->rx_ep.rep_receive_count;
158         if (wc->status != IB_WC_SUCCESS)
159                 goto out_flushed;
160 
161         /* status == SUCCESS means all fields in wc are trustworthy */
162         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
163         rep->rr_wc_flags = wc->wc_flags;
164         rep->rr_inv_rkey = wc->ex.invalidate_rkey;
165 
166         ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
167                                    rdmab_addr(rep->rr_rdmabuf),
168                                    wc->byte_len, DMA_FROM_DEVICE);
169 
170         rpcrdma_post_recvs(r_xprt, false);
171         rpcrdma_reply_handler(rep);
172         return;
173 
174 out_flushed:
175         rpcrdma_recv_buffer_put(rep);
176 }
177 
178 static void
179 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
180                                struct rdma_conn_param *param)
181 {
182         const struct rpcrdma_connect_private *pmsg = param->private_data;
183         unsigned int rsize, wsize;
184 
185         /* Default settings for RPC-over-RDMA Version One */
186         r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
187         rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
188         wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
189 
190         if (pmsg &&
191             pmsg->cp_magic == rpcrdma_cmp_magic &&
192             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
193                 r_xprt->rx_ia.ri_implicit_roundup = true;
194                 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
195                 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
196         }
197 
198         if (rsize < r_xprt->rx_ep.rep_inline_recv)
199                 r_xprt->rx_ep.rep_inline_recv = rsize;
200         if (wsize < r_xprt->rx_ep.rep_inline_send)
201                 r_xprt->rx_ep.rep_inline_send = wsize;
202         dprintk("RPC:       %s: max send %u, max recv %u\n", __func__,
203                 r_xprt->rx_ep.rep_inline_send,
204                 r_xprt->rx_ep.rep_inline_recv);
205         rpcrdma_set_max_header_sizes(r_xprt);
206 }
207 
208 /**
209  * rpcrdma_cm_event_handler - Handle RDMA CM events
210  * @id: rdma_cm_id on which an event has occurred
211  * @event: details of the event
212  *
213  * Called with @id's mutex held. Returns 1 if caller should
214  * destroy @id, otherwise 0.
215  */
216 static int
217 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
218 {
219         struct rpcrdma_xprt *r_xprt = id->context;
220         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
221         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
222         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
223 
224         might_sleep();
225 
226         trace_xprtrdma_cm_event(r_xprt, event);
227         switch (event->event) {
228         case RDMA_CM_EVENT_ADDR_RESOLVED:
229         case RDMA_CM_EVENT_ROUTE_RESOLVED:
230                 ia->ri_async_rc = 0;
231                 complete(&ia->ri_done);
232                 return 0;
233         case RDMA_CM_EVENT_ADDR_ERROR:
234                 ia->ri_async_rc = -EPROTO;
235                 complete(&ia->ri_done);
236                 return 0;
237         case RDMA_CM_EVENT_ROUTE_ERROR:
238                 ia->ri_async_rc = -ENETUNREACH;
239                 complete(&ia->ri_done);
240                 return 0;
241         case RDMA_CM_EVENT_DEVICE_REMOVAL:
242 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
243                 pr_info("rpcrdma: removing device %s for %s:%s\n",
244                         ia->ri_id->device->name,
245                         rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
246 #endif
247                 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
248                 ep->rep_connected = -ENODEV;
249                 xprt_force_disconnect(xprt);
250                 wait_for_completion(&ia->ri_remove_done);
251 
252                 ia->ri_id = NULL;
253                 /* Return 1 to ensure the core destroys the id. */
254                 return 1;
255         case RDMA_CM_EVENT_ESTABLISHED:
256                 ++xprt->connect_cookie;
257                 ep->rep_connected = 1;
258                 rpcrdma_update_connect_private(r_xprt, &event->param.conn);
259                 wake_up_all(&ep->rep_connect_wait);
260                 break;
261         case RDMA_CM_EVENT_CONNECT_ERROR:
262                 ep->rep_connected = -ENOTCONN;
263                 goto disconnected;
264         case RDMA_CM_EVENT_UNREACHABLE:
265                 ep->rep_connected = -ENETUNREACH;
266                 goto disconnected;
267         case RDMA_CM_EVENT_REJECTED:
268                 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
269                         rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
270                         rdma_reject_msg(id, event->status));
271                 ep->rep_connected = -ECONNREFUSED;
272                 if (event->status == IB_CM_REJ_STALE_CONN)
273                         ep->rep_connected = -EAGAIN;
274                 goto disconnected;
275         case RDMA_CM_EVENT_DISCONNECTED:
276                 ep->rep_connected = -ECONNABORTED;
277 disconnected:
278                 xprt_force_disconnect(xprt);
279                 wake_up_all(&ep->rep_connect_wait);
280                 break;
281         default:
282                 break;
283         }
284 
285         dprintk("RPC:       %s: %s:%s on %s/frwr: %s\n", __func__,
286                 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
287                 ia->ri_id->device->name, rdma_event_msg(event->event));
288         return 0;
289 }
290 
291 static struct rdma_cm_id *
292 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
293 {
294         unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
295         struct rdma_cm_id *id;
296         int rc;
297 
298         trace_xprtrdma_conn_start(xprt);
299 
300         init_completion(&ia->ri_done);
301         init_completion(&ia->ri_remove_done);
302 
303         id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
304                             xprt, RDMA_PS_TCP, IB_QPT_RC);
305         if (IS_ERR(id))
306                 return id;
307 
308         ia->ri_async_rc = -ETIMEDOUT;
309         rc = rdma_resolve_addr(id, NULL,
310                                (struct sockaddr *)&xprt->rx_xprt.addr,
311                                RDMA_RESOLVE_TIMEOUT);
312         if (rc)
313                 goto out;
314         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
315         if (rc < 0) {
316                 trace_xprtrdma_conn_tout(xprt);
317                 goto out;
318         }
319 
320         rc = ia->ri_async_rc;
321         if (rc)
322                 goto out;
323 
324         ia->ri_async_rc = -ETIMEDOUT;
325         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
326         if (rc)
327                 goto out;
328         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
329         if (rc < 0) {
330                 trace_xprtrdma_conn_tout(xprt);
331                 goto out;
332         }
333         rc = ia->ri_async_rc;
334         if (rc)
335                 goto out;
336 
337         return id;
338 
339 out:
340         rdma_destroy_id(id);
341         return ERR_PTR(rc);
342 }
343 
344 /*
345  * Exported functions.
346  */
347 
348 /**
349  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
350  * @xprt: transport with IA to (re)initialize
351  *
352  * Returns 0 on success, negative errno if an appropriate
353  * Interface Adapter could not be found and opened.
354  */
355 int
356 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
357 {
358         struct rpcrdma_ia *ia = &xprt->rx_ia;
359         int rc;
360 
361         ia->ri_id = rpcrdma_create_id(xprt, ia);
362         if (IS_ERR(ia->ri_id)) {
363                 rc = PTR_ERR(ia->ri_id);
364                 goto out_err;
365         }
366 
367         ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0);
368         if (IS_ERR(ia->ri_pd)) {
369                 rc = PTR_ERR(ia->ri_pd);
370                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
371                 goto out_err;
372         }
373 
374         switch (xprt_rdma_memreg_strategy) {
375         case RPCRDMA_FRWR:
376                 if (frwr_is_supported(ia->ri_id->device))
377                         break;
378                 /*FALLTHROUGH*/
379         default:
380                 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
381                        ia->ri_id->device->name, xprt_rdma_memreg_strategy);
382                 rc = -EINVAL;
383                 goto out_err;
384         }
385 
386         return 0;
387 
388 out_err:
389         rpcrdma_ia_close(ia);
390         return rc;
391 }
392 
393 /**
394  * rpcrdma_ia_remove - Handle device driver unload
395  * @ia: interface adapter being removed
396  *
397  * Divest transport H/W resources associated with this adapter,
398  * but allow it to be restored later.
399  */
400 void
401 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
402 {
403         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
404                                                    rx_ia);
405         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
406         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
407         struct rpcrdma_req *req;
408         struct rpcrdma_rep *rep;
409 
410         cancel_delayed_work_sync(&buf->rb_refresh_worker);
411 
412         /* This is similar to rpcrdma_ep_destroy, but:
413          * - Don't cancel the connect worker.
414          * - Don't call rpcrdma_ep_disconnect, which waits
415          *   for another conn upcall, which will deadlock.
416          * - rdma_disconnect is unneeded, the underlying
417          *   connection is already gone.
418          */
419         if (ia->ri_id->qp) {
420                 rpcrdma_xprt_drain(r_xprt);
421                 rdma_destroy_qp(ia->ri_id);
422                 ia->ri_id->qp = NULL;
423         }
424         ib_free_cq(ep->rep_attr.recv_cq);
425         ep->rep_attr.recv_cq = NULL;
426         ib_free_cq(ep->rep_attr.send_cq);
427         ep->rep_attr.send_cq = NULL;
428 
429         /* The ULP is responsible for ensuring all DMA
430          * mappings and MRs are gone.
431          */
432         list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
433                 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
434         list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
435                 rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf);
436                 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
437                 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
438         }
439         rpcrdma_mrs_destroy(buf);
440         ib_dealloc_pd(ia->ri_pd);
441         ia->ri_pd = NULL;
442 
443         /* Allow waiters to continue */
444         complete(&ia->ri_remove_done);
445 
446         trace_xprtrdma_remove(r_xprt);
447 }
448 
449 /**
450  * rpcrdma_ia_close - Clean up/close an IA.
451  * @ia: interface adapter to close
452  *
453  */
454 void
455 rpcrdma_ia_close(struct rpcrdma_ia *ia)
456 {
457         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
458                 if (ia->ri_id->qp)
459                         rdma_destroy_qp(ia->ri_id);
460                 rdma_destroy_id(ia->ri_id);
461         }
462         ia->ri_id = NULL;
463 
464         /* If the pd is still busy, xprtrdma missed freeing a resource */
465         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
466                 ib_dealloc_pd(ia->ri_pd);
467         ia->ri_pd = NULL;
468 }
469 
470 /**
471  * rpcrdma_ep_create - Create unconnected endpoint
472  * @r_xprt: transport to instantiate
473  *
474  * Returns zero on success, or a negative errno.
475  */
476 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
477 {
478         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
479         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
480         struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
481         struct ib_cq *sendcq, *recvcq;
482         unsigned int max_sge;
483         int rc;
484 
485         ep->rep_max_requests = xprt_rdma_slot_table_entries;
486         ep->rep_inline_send = xprt_rdma_max_inline_write;
487         ep->rep_inline_recv = xprt_rdma_max_inline_read;
488 
489         max_sge = min_t(unsigned int, ia->ri_id->device->attrs.max_send_sge,
490                         RPCRDMA_MAX_SEND_SGES);
491         if (max_sge < RPCRDMA_MIN_SEND_SGES) {
492                 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
493                 return -ENOMEM;
494         }
495         ia->ri_max_send_sges = max_sge;
496 
497         rc = frwr_open(ia, ep);
498         if (rc)
499                 return rc;
500 
501         ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
502         ep->rep_attr.qp_context = ep;
503         ep->rep_attr.srq = NULL;
504         ep->rep_attr.cap.max_send_sge = max_sge;
505         ep->rep_attr.cap.max_recv_sge = 1;
506         ep->rep_attr.cap.max_inline_data = 0;
507         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
508         ep->rep_attr.qp_type = IB_QPT_RC;
509         ep->rep_attr.port_num = ~0;
510 
511         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
512                 "iovs: send %d recv %d\n",
513                 __func__,
514                 ep->rep_attr.cap.max_send_wr,
515                 ep->rep_attr.cap.max_recv_wr,
516                 ep->rep_attr.cap.max_send_sge,
517                 ep->rep_attr.cap.max_recv_sge);
518 
519         ep->rep_send_batch = ep->rep_max_requests >> 3;
520         ep->rep_send_count = ep->rep_send_batch;
521         init_waitqueue_head(&ep->rep_connect_wait);
522         ep->rep_receive_count = 0;
523 
524         sendcq = ib_alloc_cq(ia->ri_id->device, NULL,
525                              ep->rep_attr.cap.max_send_wr + 1,
526                              ia->ri_id->device->num_comp_vectors > 1 ? 1 : 0,
527                              IB_POLL_WORKQUEUE);
528         if (IS_ERR(sendcq)) {
529                 rc = PTR_ERR(sendcq);
530                 goto out1;
531         }
532 
533         recvcq = ib_alloc_cq(ia->ri_id->device, NULL,
534                              ep->rep_attr.cap.max_recv_wr + 1,
535                              0, IB_POLL_WORKQUEUE);
536         if (IS_ERR(recvcq)) {
537                 rc = PTR_ERR(recvcq);
538                 goto out2;
539         }
540 
541         ep->rep_attr.send_cq = sendcq;
542         ep->rep_attr.recv_cq = recvcq;
543 
544         /* Initialize cma parameters */
545         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
546 
547         /* Prepare RDMA-CM private message */
548         pmsg->cp_magic = rpcrdma_cmp_magic;
549         pmsg->cp_version = RPCRDMA_CMP_VERSION;
550         pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
551         pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send);
552         pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv);
553         ep->rep_remote_cma.private_data = pmsg;
554         ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
555 
556         /* Client offers RDMA Read but does not initiate */
557         ep->rep_remote_cma.initiator_depth = 0;
558         ep->rep_remote_cma.responder_resources =
559                 min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom);
560 
561         /* Limit transport retries so client can detect server
562          * GID changes quickly. RPC layer handles re-establishing
563          * transport connection and retransmission.
564          */
565         ep->rep_remote_cma.retry_count = 6;
566 
567         /* RPC-over-RDMA handles its own flow control. In addition,
568          * make all RNR NAKs visible so we know that RPC-over-RDMA
569          * flow control is working correctly (no NAKs should be seen).
570          */
571         ep->rep_remote_cma.flow_control = 0;
572         ep->rep_remote_cma.rnr_retry_count = 0;
573 
574         return 0;
575 
576 out2:
577         ib_free_cq(sendcq);
578 out1:
579         return rc;
580 }
581 
582 /**
583  * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
584  * @r_xprt: transport instance to shut down
585  *
586  */
587 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
588 {
589         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
590         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
591 
592         if (ia->ri_id && ia->ri_id->qp) {
593                 rpcrdma_ep_disconnect(ep, ia);
594                 rdma_destroy_qp(ia->ri_id);
595                 ia->ri_id->qp = NULL;
596         }
597 
598         if (ep->rep_attr.recv_cq)
599                 ib_free_cq(ep->rep_attr.recv_cq);
600         if (ep->rep_attr.send_cq)
601                 ib_free_cq(ep->rep_attr.send_cq);
602 }
603 
604 /* Re-establish a connection after a device removal event.
605  * Unlike a normal reconnection, a fresh PD and a new set
606  * of MRs and buffers is needed.
607  */
608 static int rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
609                                     struct ib_qp_init_attr *qp_init_attr)
610 {
611         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
612         int rc, err;
613 
614         trace_xprtrdma_reinsert(r_xprt);
615 
616         rc = -EHOSTUNREACH;
617         if (rpcrdma_ia_open(r_xprt))
618                 goto out1;
619 
620         rc = -ENOMEM;
621         err = rpcrdma_ep_create(r_xprt);
622         if (err) {
623                 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
624                 goto out2;
625         }
626 
627         rc = -ENETUNREACH;
628         err = rdma_create_qp(ia->ri_id, ia->ri_pd, qp_init_attr);
629         if (err) {
630                 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
631                 goto out3;
632         }
633 
634         rpcrdma_mrs_create(r_xprt);
635         return 0;
636 
637 out3:
638         rpcrdma_ep_destroy(r_xprt);
639 out2:
640         rpcrdma_ia_close(ia);
641 out1:
642         return rc;
643 }
644 
645 static int rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt,
646                                 struct ib_qp_init_attr *qp_init_attr)
647 {
648         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
649         struct rdma_cm_id *id, *old;
650         int err, rc;
651 
652         trace_xprtrdma_reconnect(r_xprt);
653 
654         rpcrdma_ep_disconnect(&r_xprt->rx_ep, ia);
655 
656         rc = -EHOSTUNREACH;
657         id = rpcrdma_create_id(r_xprt, ia);
658         if (IS_ERR(id))
659                 goto out;
660 
661         /* As long as the new ID points to the same device as the
662          * old ID, we can reuse the transport's existing PD and all
663          * previously allocated MRs. Also, the same device means
664          * the transport's previous DMA mappings are still valid.
665          *
666          * This is a sanity check only. There should be no way these
667          * point to two different devices here.
668          */
669         old = id;
670         rc = -ENETUNREACH;
671         if (ia->ri_id->device != id->device) {
672                 pr_err("rpcrdma: can't reconnect on different device!\n");
673                 goto out_destroy;
674         }
675 
676         err = rdma_create_qp(id, ia->ri_pd, qp_init_attr);
677         if (err)
678                 goto out_destroy;
679 
680         /* Atomically replace the transport's ID and QP. */
681         rc = 0;
682         old = ia->ri_id;
683         ia->ri_id = id;
684         rdma_destroy_qp(old);
685 
686 out_destroy:
687         rdma_destroy_id(old);
688 out:
689         return rc;
690 }
691 
692 /*
693  * Connect unconnected endpoint.
694  */
695 int
696 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
697 {
698         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
699                                                    rx_ia);
700         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
701         struct ib_qp_init_attr qp_init_attr;
702         int rc;
703 
704 retry:
705         memcpy(&qp_init_attr, &ep->rep_attr, sizeof(qp_init_attr));
706         switch (ep->rep_connected) {
707         case 0:
708                 dprintk("RPC:       %s: connecting...\n", __func__);
709                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &qp_init_attr);
710                 if (rc) {
711                         rc = -ENETUNREACH;
712                         goto out_noupdate;
713                 }
714                 break;
715         case -ENODEV:
716                 rc = rpcrdma_ep_recreate_xprt(r_xprt, &qp_init_attr);
717                 if (rc)
718                         goto out_noupdate;
719                 break;
720         default:
721                 rc = rpcrdma_ep_reconnect(r_xprt, &qp_init_attr);
722                 if (rc)
723                         goto out;
724         }
725 
726         ep->rep_connected = 0;
727         xprt_clear_connected(xprt);
728 
729         rpcrdma_post_recvs(r_xprt, true);
730 
731         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
732         if (rc)
733                 goto out;
734 
735         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
736         if (ep->rep_connected <= 0) {
737                 if (ep->rep_connected == -EAGAIN)
738                         goto retry;
739                 rc = ep->rep_connected;
740                 goto out;
741         }
742 
743         dprintk("RPC:       %s: connected\n", __func__);
744 
745 out:
746         if (rc)
747                 ep->rep_connected = rc;
748 
749 out_noupdate:
750         return rc;
751 }
752 
753 /**
754  * rpcrdma_ep_disconnect - Disconnect underlying transport
755  * @ep: endpoint to disconnect
756  * @ia: associated interface adapter
757  *
758  * This is separate from destroy to facilitate the ability
759  * to reconnect without recreating the endpoint.
760  *
761  * This call is not reentrant, and must not be made in parallel
762  * on the same endpoint.
763  */
764 void
765 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
766 {
767         struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
768                                                    rx_ep);
769         int rc;
770 
771         /* returns without wait if ID is not connected */
772         rc = rdma_disconnect(ia->ri_id);
773         if (!rc)
774                 wait_event_interruptible(ep->rep_connect_wait,
775                                                         ep->rep_connected != 1);
776         else
777                 ep->rep_connected = rc;
778         trace_xprtrdma_disconnect(r_xprt, rc);
779 
780         rpcrdma_xprt_drain(r_xprt);
781 }
782 
783 /* Fixed-size circular FIFO queue. This implementation is wait-free and
784  * lock-free.
785  *
786  * Consumer is the code path that posts Sends. This path dequeues a
787  * sendctx for use by a Send operation. Multiple consumer threads
788  * are serialized by the RPC transport lock, which allows only one
789  * ->send_request call at a time.
790  *
791  * Producer is the code path that handles Send completions. This path
792  * enqueues a sendctx that has been completed. Multiple producer
793  * threads are serialized by the ib_poll_cq() function.
794  */
795 
796 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
797  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
798  * Send requests.
799  */
800 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
801 {
802         unsigned long i;
803 
804         for (i = 0; i <= buf->rb_sc_last; i++)
805                 kfree(buf->rb_sc_ctxs[i]);
806         kfree(buf->rb_sc_ctxs);
807 }
808 
809 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
810 {
811         struct rpcrdma_sendctx *sc;
812 
813         sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges),
814                      GFP_KERNEL);
815         if (!sc)
816                 return NULL;
817 
818         sc->sc_wr.wr_cqe = &sc->sc_cqe;
819         sc->sc_wr.sg_list = sc->sc_sges;
820         sc->sc_wr.opcode = IB_WR_SEND;
821         sc->sc_cqe.done = rpcrdma_wc_send;
822         return sc;
823 }
824 
825 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
826 {
827         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
828         struct rpcrdma_sendctx *sc;
829         unsigned long i;
830 
831         /* Maximum number of concurrent outstanding Send WRs. Capping
832          * the circular queue size stops Send Queue overflow by causing
833          * the ->send_request call to fail temporarily before too many
834          * Sends are posted.
835          */
836         i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
837         dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
838         buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
839         if (!buf->rb_sc_ctxs)
840                 return -ENOMEM;
841 
842         buf->rb_sc_last = i - 1;
843         for (i = 0; i <= buf->rb_sc_last; i++) {
844                 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
845                 if (!sc)
846                         return -ENOMEM;
847 
848                 sc->sc_xprt = r_xprt;
849                 buf->rb_sc_ctxs[i] = sc;
850         }
851 
852         return 0;
853 }
854 
855 /* The sendctx queue is not guaranteed to have a size that is a
856  * power of two, thus the helpers in circ_buf.h cannot be used.
857  * The other option is to use modulus (%), which can be expensive.
858  */
859 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
860                                           unsigned long item)
861 {
862         return likely(item < buf->rb_sc_last) ? item + 1 : 0;
863 }
864 
865 /**
866  * rpcrdma_sendctx_get_locked - Acquire a send context
867  * @r_xprt: controlling transport instance
868  *
869  * Returns pointer to a free send completion context; or NULL if
870  * the queue is empty.
871  *
872  * Usage: Called to acquire an SGE array before preparing a Send WR.
873  *
874  * The caller serializes calls to this function (per transport), and
875  * provides an effective memory barrier that flushes the new value
876  * of rb_sc_head.
877  */
878 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
879 {
880         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
881         struct rpcrdma_sendctx *sc;
882         unsigned long next_head;
883 
884         next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
885 
886         if (next_head == READ_ONCE(buf->rb_sc_tail))
887                 goto out_emptyq;
888 
889         /* ORDER: item must be accessed _before_ head is updated */
890         sc = buf->rb_sc_ctxs[next_head];
891 
892         /* Releasing the lock in the caller acts as a memory
893          * barrier that flushes rb_sc_head.
894          */
895         buf->rb_sc_head = next_head;
896 
897         return sc;
898 
899 out_emptyq:
900         /* The queue is "empty" if there have not been enough Send
901          * completions recently. This is a sign the Send Queue is
902          * backing up. Cause the caller to pause and try again.
903          */
904         xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
905         r_xprt->rx_stats.empty_sendctx_q++;
906         return NULL;
907 }
908 
909 /**
910  * rpcrdma_sendctx_put_locked - Release a send context
911  * @sc: send context to release
912  *
913  * Usage: Called from Send completion to return a sendctxt
914  * to the queue.
915  *
916  * The caller serializes calls to this function (per transport).
917  */
918 static void
919 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
920 {
921         struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
922         unsigned long next_tail;
923 
924         /* Unmap SGEs of previously completed but unsignaled
925          * Sends by walking up the queue until @sc is found.
926          */
927         next_tail = buf->rb_sc_tail;
928         do {
929                 next_tail = rpcrdma_sendctx_next(buf, next_tail);
930 
931                 /* ORDER: item must be accessed _before_ tail is updated */
932                 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
933 
934         } while (buf->rb_sc_ctxs[next_tail] != sc);
935 
936         /* Paired with READ_ONCE */
937         smp_store_release(&buf->rb_sc_tail, next_tail);
938 
939         xprt_write_space(&sc->sc_xprt->rx_xprt);
940 }
941 
942 static void
943 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
944 {
945         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
946         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
947         unsigned int count;
948         LIST_HEAD(free);
949         LIST_HEAD(all);
950 
951         for (count = 0; count < ia->ri_max_segs; count++) {
952                 struct rpcrdma_mr *mr;
953                 int rc;
954 
955                 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
956                 if (!mr)
957                         break;
958 
959                 rc = frwr_init_mr(ia, mr);
960                 if (rc) {
961                         kfree(mr);
962                         break;
963                 }
964 
965                 mr->mr_xprt = r_xprt;
966 
967                 list_add(&mr->mr_list, &free);
968                 list_add(&mr->mr_all, &all);
969         }
970 
971         spin_lock(&buf->rb_mrlock);
972         list_splice(&free, &buf->rb_mrs);
973         list_splice(&all, &buf->rb_all);
974         r_xprt->rx_stats.mrs_allocated += count;
975         spin_unlock(&buf->rb_mrlock);
976         trace_xprtrdma_createmrs(r_xprt, count);
977 }
978 
979 static void
980 rpcrdma_mr_refresh_worker(struct work_struct *work)
981 {
982         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
983                                                   rb_refresh_worker.work);
984         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
985                                                    rx_buf);
986 
987         rpcrdma_mrs_create(r_xprt);
988         xprt_write_space(&r_xprt->rx_xprt);
989 }
990 
991 /**
992  * rpcrdma_req_create - Allocate an rpcrdma_req object
993  * @r_xprt: controlling r_xprt
994  * @size: initial size, in bytes, of send and receive buffers
995  * @flags: GFP flags passed to memory allocators
996  *
997  * Returns an allocated and fully initialized rpcrdma_req or NULL.
998  */
999 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
1000                                        gfp_t flags)
1001 {
1002         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1003         struct rpcrdma_regbuf *rb;
1004         struct rpcrdma_req *req;
1005 
1006         req = kzalloc(sizeof(*req), flags);
1007         if (req == NULL)
1008                 goto out1;
1009 
1010         rb = rpcrdma_regbuf_alloc(RPCRDMA_HDRBUF_SIZE, DMA_TO_DEVICE, flags);
1011         if (!rb)
1012                 goto out2;
1013         req->rl_rdmabuf = rb;
1014         xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1015 
1016         req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1017         if (!req->rl_sendbuf)
1018                 goto out3;
1019 
1020         req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1021         if (!req->rl_recvbuf)
1022                 goto out4;
1023 
1024         INIT_LIST_HEAD(&req->rl_registered);
1025         spin_lock(&buffer->rb_lock);
1026         list_add(&req->rl_all, &buffer->rb_allreqs);
1027         spin_unlock(&buffer->rb_lock);
1028         return req;
1029 
1030 out4:
1031         kfree(req->rl_sendbuf);
1032 out3:
1033         kfree(req->rl_rdmabuf);
1034 out2:
1035         kfree(req);
1036 out1:
1037         return NULL;
1038 }
1039 
1040 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1041                                               bool temp)
1042 {
1043         struct rpcrdma_rep *rep;
1044 
1045         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1046         if (rep == NULL)
1047                 goto out;
1048 
1049         rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv,
1050                                                DMA_FROM_DEVICE, GFP_KERNEL);
1051         if (!rep->rr_rdmabuf)
1052                 goto out_free;
1053 
1054         xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
1055                      rdmab_length(rep->rr_rdmabuf));
1056         rep->rr_cqe.done = rpcrdma_wc_receive;
1057         rep->rr_rxprt = r_xprt;
1058         rep->rr_recv_wr.next = NULL;
1059         rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1060         rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1061         rep->rr_recv_wr.num_sge = 1;
1062         rep->rr_temp = temp;
1063         return rep;
1064 
1065 out_free:
1066         kfree(rep);
1067 out:
1068         return NULL;
1069 }
1070 
1071 /**
1072  * rpcrdma_buffer_create - Create initial set of req/rep objects
1073  * @r_xprt: transport instance to (re)initialize
1074  *
1075  * Returns zero on success, otherwise a negative errno.
1076  */
1077 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1078 {
1079         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1080         int i, rc;
1081 
1082         buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
1083         buf->rb_bc_srv_max_requests = 0;
1084         spin_lock_init(&buf->rb_mrlock);
1085         spin_lock_init(&buf->rb_lock);
1086         INIT_LIST_HEAD(&buf->rb_mrs);
1087         INIT_LIST_HEAD(&buf->rb_all);
1088         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1089                           rpcrdma_mr_refresh_worker);
1090 
1091         rpcrdma_mrs_create(r_xprt);
1092 
1093         INIT_LIST_HEAD(&buf->rb_send_bufs);
1094         INIT_LIST_HEAD(&buf->rb_allreqs);
1095 
1096         rc = -ENOMEM;
1097         for (i = 0; i < buf->rb_max_requests; i++) {
1098                 struct rpcrdma_req *req;
1099 
1100                 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE,
1101                                          GFP_KERNEL);
1102                 if (!req)
1103                         goto out;
1104                 list_add(&req->rl_list, &buf->rb_send_bufs);
1105         }
1106 
1107         buf->rb_credits = 1;
1108         INIT_LIST_HEAD(&buf->rb_recv_bufs);
1109 
1110         rc = rpcrdma_sendctxs_create(r_xprt);
1111         if (rc)
1112                 goto out;
1113 
1114         return 0;
1115 out:
1116         rpcrdma_buffer_destroy(buf);
1117         return rc;
1118 }
1119 
1120 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1121 {
1122         rpcrdma_regbuf_free(rep->rr_rdmabuf);
1123         kfree(rep);
1124 }
1125 
1126 /**
1127  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1128  * @req: unused object to be destroyed
1129  *
1130  * This function assumes that the caller prevents concurrent device
1131  * unload and transport tear-down.
1132  */
1133 void
1134 rpcrdma_req_destroy(struct rpcrdma_req *req)
1135 {
1136         list_del(&req->rl_all);
1137 
1138         rpcrdma_regbuf_free(req->rl_recvbuf);
1139         rpcrdma_regbuf_free(req->rl_sendbuf);
1140         rpcrdma_regbuf_free(req->rl_rdmabuf);
1141         kfree(req);
1142 }
1143 
1144 static void
1145 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1146 {
1147         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1148                                                    rx_buf);
1149         struct rpcrdma_mr *mr;
1150         unsigned int count;
1151 
1152         count = 0;
1153         spin_lock(&buf->rb_mrlock);
1154         while (!list_empty(&buf->rb_all)) {
1155                 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1156                 list_del(&mr->mr_all);
1157 
1158                 spin_unlock(&buf->rb_mrlock);
1159 
1160                 /* Ensure MW is not on any rl_registered list */
1161                 if (!list_empty(&mr->mr_list))
1162                         list_del(&mr->mr_list);
1163 
1164                 frwr_release_mr(mr);
1165                 count++;
1166                 spin_lock(&buf->rb_mrlock);
1167         }
1168         spin_unlock(&buf->rb_mrlock);
1169         r_xprt->rx_stats.mrs_allocated = 0;
1170 
1171         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1172 }
1173 
1174 /**
1175  * rpcrdma_buffer_destroy - Release all hw resources
1176  * @buf: root control block for resources
1177  *
1178  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1179  * - No more Send or Receive completions can occur
1180  * - All MRs, reps, and reqs are returned to their free lists
1181  */
1182 void
1183 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1184 {
1185         cancel_delayed_work_sync(&buf->rb_refresh_worker);
1186 
1187         rpcrdma_sendctxs_destroy(buf);
1188 
1189         while (!list_empty(&buf->rb_recv_bufs)) {
1190                 struct rpcrdma_rep *rep;
1191 
1192                 rep = list_first_entry(&buf->rb_recv_bufs,
1193                                        struct rpcrdma_rep, rr_list);
1194                 list_del(&rep->rr_list);
1195                 rpcrdma_rep_destroy(rep);
1196         }
1197 
1198         while (!list_empty(&buf->rb_send_bufs)) {
1199                 struct rpcrdma_req *req;
1200 
1201                 req = list_first_entry(&buf->rb_send_bufs,
1202                                        struct rpcrdma_req, rl_list);
1203                 list_del(&req->rl_list);
1204                 rpcrdma_req_destroy(req);
1205         }
1206 
1207         rpcrdma_mrs_destroy(buf);
1208 }
1209 
1210 /**
1211  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1212  * @r_xprt: controlling transport
1213  *
1214  * Returns an initialized rpcrdma_mr or NULL if no free
1215  * rpcrdma_mr objects are available.
1216  */
1217 struct rpcrdma_mr *
1218 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1219 {
1220         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1221         struct rpcrdma_mr *mr = NULL;
1222 
1223         spin_lock(&buf->rb_mrlock);
1224         if (!list_empty(&buf->rb_mrs))
1225                 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1226         spin_unlock(&buf->rb_mrlock);
1227 
1228         if (!mr)
1229                 goto out_nomrs;
1230         return mr;
1231 
1232 out_nomrs:
1233         trace_xprtrdma_nomrs(r_xprt);
1234         if (r_xprt->rx_ep.rep_connected != -ENODEV)
1235                 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1236 
1237         /* Allow the reply handler and refresh worker to run */
1238         cond_resched();
1239 
1240         return NULL;
1241 }
1242 
1243 static void
1244 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1245 {
1246         spin_lock(&buf->rb_mrlock);
1247         rpcrdma_mr_push(mr, &buf->rb_mrs);
1248         spin_unlock(&buf->rb_mrlock);
1249 }
1250 
1251 /**
1252  * rpcrdma_mr_put - Release an rpcrdma_mr object
1253  * @mr: object to release
1254  *
1255  */
1256 void
1257 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1258 {
1259         __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1260 }
1261 
1262 /**
1263  * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1264  * @mr: object to release
1265  *
1266  */
1267 void
1268 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1269 {
1270         struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1271 
1272         if (mr->mr_dir != DMA_NONE) {
1273                 trace_xprtrdma_mr_unmap(mr);
1274                 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1275                                 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1276                 mr->mr_dir = DMA_NONE;
1277         }
1278         __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1279 }
1280 
1281 /**
1282  * rpcrdma_buffer_get - Get a request buffer
1283  * @buffers: Buffer pool from which to obtain a buffer
1284  *
1285  * Returns a fresh rpcrdma_req, or NULL if none are available.
1286  */
1287 struct rpcrdma_req *
1288 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1289 {
1290         struct rpcrdma_req *req;
1291 
1292         spin_lock(&buffers->rb_lock);
1293         req = list_first_entry_or_null(&buffers->rb_send_bufs,
1294                                        struct rpcrdma_req, rl_list);
1295         if (req)
1296                 list_del_init(&req->rl_list);
1297         spin_unlock(&buffers->rb_lock);
1298         return req;
1299 }
1300 
1301 /**
1302  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1303  * @buffers: buffer pool
1304  * @req: object to return
1305  *
1306  */
1307 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1308 {
1309         struct rpcrdma_rep *rep = req->rl_reply;
1310 
1311         req->rl_reply = NULL;
1312 
1313         spin_lock(&buffers->rb_lock);
1314         list_add(&req->rl_list, &buffers->rb_send_bufs);
1315         if (rep) {
1316                 if (!rep->rr_temp) {
1317                         list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1318                         rep = NULL;
1319                 }
1320         }
1321         spin_unlock(&buffers->rb_lock);
1322         if (rep)
1323                 rpcrdma_rep_destroy(rep);
1324 }
1325 
1326 /*
1327  * Put reply buffers back into pool when not attached to
1328  * request. This happens in error conditions.
1329  */
1330 void
1331 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1332 {
1333         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1334 
1335         if (!rep->rr_temp) {
1336                 spin_lock(&buffers->rb_lock);
1337                 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1338                 spin_unlock(&buffers->rb_lock);
1339         } else {
1340                 rpcrdma_rep_destroy(rep);
1341         }
1342 }
1343 
1344 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1345  *
1346  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1347  * receiving the payload of RDMA RECV operations. During Long Calls
1348  * or Replies they may be registered externally via frwr_map.
1349  */
1350 static struct rpcrdma_regbuf *
1351 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1352                      gfp_t flags)
1353 {
1354         struct rpcrdma_regbuf *rb;
1355 
1356         rb = kmalloc(sizeof(*rb), flags);
1357         if (!rb)
1358                 return NULL;
1359         rb->rg_data = kmalloc(size, flags);
1360         if (!rb->rg_data) {
1361                 kfree(rb);
1362                 return NULL;
1363         }
1364 
1365         rb->rg_device = NULL;
1366         rb->rg_direction = direction;
1367         rb->rg_iov.length = size;
1368         return rb;
1369 }
1370 
1371 /**
1372  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1373  * @rb: regbuf to reallocate
1374  * @size: size of buffer to be allocated, in bytes
1375  * @flags: GFP flags
1376  *
1377  * Returns true if reallocation was successful. If false is
1378  * returned, @rb is left untouched.
1379  */
1380 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1381 {
1382         void *buf;
1383 
1384         buf = kmalloc(size, flags);
1385         if (!buf)
1386                 return false;
1387 
1388         rpcrdma_regbuf_dma_unmap(rb);
1389         kfree(rb->rg_data);
1390 
1391         rb->rg_data = buf;
1392         rb->rg_iov.length = size;
1393         return true;
1394 }
1395 
1396 /**
1397  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1398  * @r_xprt: controlling transport instance
1399  * @rb: regbuf to be mapped
1400  *
1401  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1402  */
1403 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1404                               struct rpcrdma_regbuf *rb)
1405 {
1406         struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1407 
1408         if (rb->rg_direction == DMA_NONE)
1409                 return false;
1410 
1411         rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1412                                             rdmab_length(rb), rb->rg_direction);
1413         if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1414                 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1415                 return false;
1416         }
1417 
1418         rb->rg_device = device;
1419         rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1420         return true;
1421 }
1422 
1423 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1424 {
1425         if (!rb)
1426                 return;
1427 
1428         if (!rpcrdma_regbuf_is_mapped(rb))
1429                 return;
1430 
1431         ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1432                             rb->rg_direction);
1433         rb->rg_device = NULL;
1434 }
1435 
1436 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1437 {
1438         rpcrdma_regbuf_dma_unmap(rb);
1439         if (rb)
1440                 kfree(rb->rg_data);
1441         kfree(rb);
1442 }
1443 
1444 /**
1445  * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1446  * @ia: transport's device information
1447  * @ep: transport's RDMA endpoint information
1448  * @req: rpcrdma_req containing the Send WR to post
1449  *
1450  * Returns 0 if the post was successful, otherwise -ENOTCONN
1451  * is returned.
1452  */
1453 int
1454 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1455                 struct rpcrdma_ep *ep,
1456                 struct rpcrdma_req *req)
1457 {
1458         struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1459         int rc;
1460 
1461         if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1462                 send_wr->send_flags |= IB_SEND_SIGNALED;
1463                 ep->rep_send_count = ep->rep_send_batch;
1464         } else {
1465                 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1466                 --ep->rep_send_count;
1467         }
1468 
1469         rc = frwr_send(ia, req);
1470         trace_xprtrdma_post_send(req, rc);
1471         if (rc)
1472                 return -ENOTCONN;
1473         return 0;
1474 }
1475 
1476 static void
1477 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1478 {
1479         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1480         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1481         struct ib_recv_wr *i, *wr, *bad_wr;
1482         struct rpcrdma_rep *rep;
1483         int needed, count, rc;
1484 
1485         rc = 0;
1486         count = 0;
1487 
1488         needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1489         if (ep->rep_receive_count > needed)
1490                 goto out;
1491         needed -= ep->rep_receive_count;
1492         if (!temp)
1493                 needed += RPCRDMA_MAX_RECV_BATCH;
1494 
1495         /* fast path: all needed reps can be found on the free list */
1496         wr = NULL;
1497         spin_lock(&buf->rb_lock);
1498         while (needed) {
1499                 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1500                                                struct rpcrdma_rep, rr_list);
1501                 if (!rep)
1502                         break;
1503 
1504                 list_del(&rep->rr_list);
1505                 rep->rr_recv_wr.next = wr;
1506                 wr = &rep->rr_recv_wr;
1507                 --needed;
1508         }
1509         spin_unlock(&buf->rb_lock);
1510 
1511         while (needed) {
1512                 rep = rpcrdma_rep_create(r_xprt, temp);
1513                 if (!rep)
1514                         break;
1515 
1516                 rep->rr_recv_wr.next = wr;
1517                 wr = &rep->rr_recv_wr;
1518                 --needed;
1519         }
1520         if (!wr)
1521                 goto out;
1522 
1523         for (i = wr; i; i = i->next) {
1524                 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1525 
1526                 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1527                         goto release_wrs;
1528 
1529                 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1530                 ++count;
1531         }
1532 
1533         rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1534                           (const struct ib_recv_wr **)&bad_wr);
1535 out:
1536         trace_xprtrdma_post_recvs(r_xprt, count, rc);
1537         if (rc) {
1538                 for (wr = bad_wr; wr;) {
1539                         struct rpcrdma_rep *rep;
1540 
1541                         rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1542                         wr = wr->next;
1543                         rpcrdma_recv_buffer_put(rep);
1544                         --count;
1545                 }
1546         }
1547         ep->rep_receive_count += count;
1548         return;
1549 
1550 release_wrs:
1551         for (i = wr; i;) {
1552                 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1553                 i = i->next;
1554                 rpcrdma_recv_buffer_put(rep);
1555         }
1556 }
1557 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp