Merge branch 'next' of master.kernel.org:/pub/scm/linux/kernel/git/galak/powerpc...
[linux-2.6] / net / rds / ib_cm.c
1 /*
2  * Copyright (c) 2006 Oracle.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  */
33 #include <linux/kernel.h>
34 #include <linux/in.h>
35 #include <linux/vmalloc.h>
36
37 #include "rds.h"
38 #include "ib.h"
39
40 /*
41  * Set the selected protocol version
42  */
43 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
44 {
45         conn->c_version = version;
46 }
47
48 /*
49  * Set up flow control
50  */
51 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
52 {
53         struct rds_ib_connection *ic = conn->c_transport_data;
54
55         if (rds_ib_sysctl_flow_control && credits != 0) {
56                 /* We're doing flow control */
57                 ic->i_flowctl = 1;
58                 rds_ib_send_add_credits(conn, credits);
59         } else {
60                 ic->i_flowctl = 0;
61         }
62 }
63
64 /*
65  * Tune RNR behavior. Without flow control, we use a rather
66  * low timeout, but not the absolute minimum - this should
67  * be tunable.
68  *
69  * We already set the RNR retry count to 7 (which is the
70  * smallest infinite number :-) above.
71  * If flow control is off, we want to change this back to 0
72  * so that we learn quickly when our credit accounting is
73  * buggy.
74  *
75  * Caller passes in a qp_attr pointer - don't waste stack spacv
76  * by allocation this twice.
77  */
78 static void
79 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
80 {
81         int ret;
82
83         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
84         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
85         if (ret)
86                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
87 }
88
89 /*
90  * Connection established.
91  * We get here for both outgoing and incoming connection.
92  */
93 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
94 {
95         const struct rds_ib_connect_private *dp = NULL;
96         struct rds_ib_connection *ic = conn->c_transport_data;
97         struct rds_ib_device *rds_ibdev;
98         struct ib_qp_attr qp_attr;
99         int err;
100
101         if (event->param.conn.private_data_len) {
102                 dp = event->param.conn.private_data;
103
104                 rds_ib_set_protocol(conn,
105                                 RDS_PROTOCOL(dp->dp_protocol_major,
106                                         dp->dp_protocol_minor));
107                 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
108         }
109
110         printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
111                         &conn->c_laddr,
112                         RDS_PROTOCOL_MAJOR(conn->c_version),
113                         RDS_PROTOCOL_MINOR(conn->c_version),
114                         ic->i_flowctl ? ", flow control" : "");
115
116         /* Tune RNR behavior */
117         rds_ib_tune_rnr(ic, &qp_attr);
118
119         qp_attr.qp_state = IB_QPS_RTS;
120         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
121         if (err)
122                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
123
124         /* update ib_device with this local ipaddr & conn */
125         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
126         err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
127         if (err)
128                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
129         rds_ib_add_conn(rds_ibdev, conn);
130
131         /* If the peer gave us the last packet it saw, process this as if
132          * we had received a regular ACK. */
133         if (dp && dp->dp_ack_seq)
134                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
135
136         rds_connect_complete(conn);
137 }
138
139 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
140                         struct rdma_conn_param *conn_param,
141                         struct rds_ib_connect_private *dp,
142                         u32 protocol_version)
143 {
144         memset(conn_param, 0, sizeof(struct rdma_conn_param));
145         /* XXX tune these? */
146         conn_param->responder_resources = 1;
147         conn_param->initiator_depth = 1;
148         conn_param->retry_count = 7;
149         conn_param->rnr_retry_count = 7;
150
151         if (dp) {
152                 struct rds_ib_connection *ic = conn->c_transport_data;
153
154                 memset(dp, 0, sizeof(*dp));
155                 dp->dp_saddr = conn->c_laddr;
156                 dp->dp_daddr = conn->c_faddr;
157                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
158                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
159                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
160                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
161
162                 /* Advertise flow control */
163                 if (ic->i_flowctl) {
164                         unsigned int credits;
165
166                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
167                         dp->dp_credit = cpu_to_be32(credits);
168                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
169                 }
170
171                 conn_param->private_data = dp;
172                 conn_param->private_data_len = sizeof(*dp);
173         }
174 }
175
176 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
177 {
178         rdsdebug("event %u data %p\n", event->event, data);
179 }
180
181 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
182 {
183         struct rds_connection *conn = data;
184         struct rds_ib_connection *ic = conn->c_transport_data;
185
186         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
187
188         switch (event->event) {
189         case IB_EVENT_COMM_EST:
190                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
191                 break;
192         default:
193                 printk(KERN_WARNING "RDS/ib: unhandled QP event %u "
194                        "on connection to %pI4\n", event->event,
195                        &conn->c_faddr);
196                 break;
197         }
198 }
199
200 /*
201  * This needs to be very careful to not leave IS_ERR pointers around for
202  * cleanup to trip over.
203  */
204 static int rds_ib_setup_qp(struct rds_connection *conn)
205 {
206         struct rds_ib_connection *ic = conn->c_transport_data;
207         struct ib_device *dev = ic->i_cm_id->device;
208         struct ib_qp_init_attr attr;
209         struct rds_ib_device *rds_ibdev;
210         int ret;
211
212         /* rds_ib_add_one creates a rds_ib_device object per IB device,
213          * and allocates a protection domain, memory range and FMR pool
214          * for each.  If that fails for any reason, it will not register
215          * the rds_ibdev at all.
216          */
217         rds_ibdev = ib_get_client_data(dev, &rds_ib_client);
218         if (rds_ibdev == NULL) {
219                 if (printk_ratelimit())
220                         printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
221                                         dev->name);
222                 return -EOPNOTSUPP;
223         }
224
225         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
226                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
227         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
228                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
229
230         /* Protection domain and memory range */
231         ic->i_pd = rds_ibdev->pd;
232         ic->i_mr = rds_ibdev->mr;
233
234         ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
235                                      rds_ib_cq_event_handler, conn,
236                                      ic->i_send_ring.w_nr + 1, 0);
237         if (IS_ERR(ic->i_send_cq)) {
238                 ret = PTR_ERR(ic->i_send_cq);
239                 ic->i_send_cq = NULL;
240                 rdsdebug("ib_create_cq send failed: %d\n", ret);
241                 goto out;
242         }
243
244         ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
245                                      rds_ib_cq_event_handler, conn,
246                                      ic->i_recv_ring.w_nr, 0);
247         if (IS_ERR(ic->i_recv_cq)) {
248                 ret = PTR_ERR(ic->i_recv_cq);
249                 ic->i_recv_cq = NULL;
250                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
251                 goto out;
252         }
253
254         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
255         if (ret) {
256                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
257                 goto out;
258         }
259
260         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
261         if (ret) {
262                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
263                 goto out;
264         }
265
266         /* XXX negotiate max send/recv with remote? */
267         memset(&attr, 0, sizeof(attr));
268         attr.event_handler = rds_ib_qp_event_handler;
269         attr.qp_context = conn;
270         /* + 1 to allow for the single ack message */
271         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
272         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
273         attr.cap.max_send_sge = rds_ibdev->max_sge;
274         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
275         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
276         attr.qp_type = IB_QPT_RC;
277         attr.send_cq = ic->i_send_cq;
278         attr.recv_cq = ic->i_recv_cq;
279
280         /*
281          * XXX this can fail if max_*_wr is too large?  Are we supposed
282          * to back off until we get a value that the hardware can support?
283          */
284         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
285         if (ret) {
286                 rdsdebug("rdma_create_qp failed: %d\n", ret);
287                 goto out;
288         }
289
290         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
291                                            ic->i_send_ring.w_nr *
292                                                 sizeof(struct rds_header),
293                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
294         if (ic->i_send_hdrs == NULL) {
295                 ret = -ENOMEM;
296                 rdsdebug("ib_dma_alloc_coherent send failed\n");
297                 goto out;
298         }
299
300         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
301                                            ic->i_recv_ring.w_nr *
302                                                 sizeof(struct rds_header),
303                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
304         if (ic->i_recv_hdrs == NULL) {
305                 ret = -ENOMEM;
306                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
307                 goto out;
308         }
309
310         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
311                                        &ic->i_ack_dma, GFP_KERNEL);
312         if (ic->i_ack == NULL) {
313                 ret = -ENOMEM;
314                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
315                 goto out;
316         }
317
318         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
319         if (ic->i_sends == NULL) {
320                 ret = -ENOMEM;
321                 rdsdebug("send allocation failed\n");
322                 goto out;
323         }
324         rds_ib_send_init_ring(ic);
325
326         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
327         if (ic->i_recvs == NULL) {
328                 ret = -ENOMEM;
329                 rdsdebug("recv allocation failed\n");
330                 goto out;
331         }
332
333         rds_ib_recv_init_ring(ic);
334         rds_ib_recv_init_ack(ic);
335
336         /* Post receive buffers - as a side effect, this will update
337          * the posted credit count. */
338         rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
339
340         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
341                  ic->i_send_cq, ic->i_recv_cq);
342
343 out:
344         return ret;
345 }
346
347 static u32 rds_ib_protocol_compatible(const struct rds_ib_connect_private *dp)
348 {
349         u16 common;
350         u32 version = 0;
351
352         /* rdma_cm private data is odd - when there is any private data in the
353          * request, we will be given a pretty large buffer without telling us the
354          * original size. The only way to tell the difference is by looking at
355          * the contents, which are initialized to zero.
356          * If the protocol version fields aren't set, this is a connection attempt
357          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
358          * We really should have changed this for OFED 1.3 :-( */
359         if (dp->dp_protocol_major == 0)
360                 return RDS_PROTOCOL_3_0;
361
362         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
363         if (dp->dp_protocol_major == 3 && common) {
364                 version = RDS_PROTOCOL_3_0;
365                 while ((common >>= 1) != 0)
366                         version++;
367         } else if (printk_ratelimit()) {
368                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
369                         "incompatible protocol version %u.%u\n",
370                         &dp->dp_saddr,
371                         dp->dp_protocol_major,
372                         dp->dp_protocol_minor);
373         }
374         return version;
375 }
376
377 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
378                                     struct rdma_cm_event *event)
379 {
380         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
381         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
382         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
383         struct rds_ib_connect_private dp_rep;
384         struct rds_connection *conn = NULL;
385         struct rds_ib_connection *ic = NULL;
386         struct rdma_conn_param conn_param;
387         u32 version;
388         int err, destroy = 1;
389
390         /* Check whether the remote protocol version matches ours. */
391         version = rds_ib_protocol_compatible(dp);
392         if (!version)
393                 goto out;
394
395         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
396                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
397                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
398                  (unsigned long long)be64_to_cpu(lguid),
399                  (unsigned long long)be64_to_cpu(fguid));
400
401         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
402                                GFP_KERNEL);
403         if (IS_ERR(conn)) {
404                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
405                 conn = NULL;
406                 goto out;
407         }
408
409         /*
410          * The connection request may occur while the
411          * previous connection exist, e.g. in case of failover.
412          * But as connections may be initiated simultaneously
413          * by both hosts, we have a random backoff mechanism -
414          * see the comment above rds_queue_reconnect()
415          */
416         mutex_lock(&conn->c_cm_lock);
417         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
418                 if (rds_conn_state(conn) == RDS_CONN_UP) {
419                         rdsdebug("incoming connect while connecting\n");
420                         rds_conn_drop(conn);
421                         rds_ib_stats_inc(s_ib_listen_closed_stale);
422                 } else
423                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
424                         /* Wait and see - our connect may still be succeeding */
425                         rds_ib_stats_inc(s_ib_connect_raced);
426                 }
427                 mutex_unlock(&conn->c_cm_lock);
428                 goto out;
429         }
430
431         ic = conn->c_transport_data;
432
433         rds_ib_set_protocol(conn, version);
434         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
435
436         /* If the peer gave us the last packet it saw, process this as if
437          * we had received a regular ACK. */
438         if (dp->dp_ack_seq)
439                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
440
441         BUG_ON(cm_id->context);
442         BUG_ON(ic->i_cm_id);
443
444         ic->i_cm_id = cm_id;
445         cm_id->context = conn;
446
447         /* We got halfway through setting up the ib_connection, if we
448          * fail now, we have to take the long route out of this mess. */
449         destroy = 0;
450
451         err = rds_ib_setup_qp(conn);
452         if (err) {
453                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
454                 goto out;
455         }
456
457         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
458
459         /* rdma_accept() calls rdma_reject() internally if it fails */
460         err = rdma_accept(cm_id, &conn_param);
461         mutex_unlock(&conn->c_cm_lock);
462         if (err) {
463                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
464                 goto out;
465         }
466
467         return 0;
468
469 out:
470         rdma_reject(cm_id, NULL, 0);
471         return destroy;
472 }
473
474
475 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
476 {
477         struct rds_connection *conn = cm_id->context;
478         struct rds_ib_connection *ic = conn->c_transport_data;
479         struct rdma_conn_param conn_param;
480         struct rds_ib_connect_private dp;
481         int ret;
482
483         /* If the peer doesn't do protocol negotiation, we must
484          * default to RDSv3.0 */
485         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
486         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
487
488         ret = rds_ib_setup_qp(conn);
489         if (ret) {
490                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
491                 goto out;
492         }
493
494         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
495
496         ret = rdma_connect(cm_id, &conn_param);
497         if (ret)
498                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
499
500 out:
501         /* Beware - returning non-zero tells the rdma_cm to destroy
502          * the cm_id. We should certainly not do it as long as we still
503          * "own" the cm_id. */
504         if (ret) {
505                 if (ic->i_cm_id == cm_id)
506                         ret = 0;
507         }
508         return ret;
509 }
510
511 int rds_ib_conn_connect(struct rds_connection *conn)
512 {
513         struct rds_ib_connection *ic = conn->c_transport_data;
514         struct sockaddr_in src, dest;
515         int ret;
516
517         /* XXX I wonder what affect the port space has */
518         /* delegate cm event handler to rdma_transport */
519         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
520                                      RDMA_PS_TCP);
521         if (IS_ERR(ic->i_cm_id)) {
522                 ret = PTR_ERR(ic->i_cm_id);
523                 ic->i_cm_id = NULL;
524                 rdsdebug("rdma_create_id() failed: %d\n", ret);
525                 goto out;
526         }
527
528         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
529
530         src.sin_family = AF_INET;
531         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
532         src.sin_port = (__force u16)htons(0);
533
534         dest.sin_family = AF_INET;
535         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
536         dest.sin_port = (__force u16)htons(RDS_PORT);
537
538         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
539                                 (struct sockaddr *)&dest,
540                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
541         if (ret) {
542                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
543                          ret);
544                 rdma_destroy_id(ic->i_cm_id);
545                 ic->i_cm_id = NULL;
546         }
547
548 out:
549         return ret;
550 }
551
552 /*
553  * This is so careful about only cleaning up resources that were built up
554  * so that it can be called at any point during startup.  In fact it
555  * can be called multiple times for a given connection.
556  */
557 void rds_ib_conn_shutdown(struct rds_connection *conn)
558 {
559         struct rds_ib_connection *ic = conn->c_transport_data;
560         int err = 0;
561
562         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
563                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
564                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
565
566         if (ic->i_cm_id) {
567                 struct ib_device *dev = ic->i_cm_id->device;
568
569                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
570                 err = rdma_disconnect(ic->i_cm_id);
571                 if (err) {
572                         /* Actually this may happen quite frequently, when
573                          * an outgoing connect raced with an incoming connect.
574                          */
575                         rdsdebug("failed to disconnect, cm: %p err %d\n",
576                                 ic->i_cm_id, err);
577                 }
578
579                 wait_event(rds_ib_ring_empty_wait,
580                         rds_ib_ring_empty(&ic->i_send_ring) &&
581                         rds_ib_ring_empty(&ic->i_recv_ring));
582
583                 if (ic->i_send_hdrs)
584                         ib_dma_free_coherent(dev,
585                                            ic->i_send_ring.w_nr *
586                                                 sizeof(struct rds_header),
587                                            ic->i_send_hdrs,
588                                            ic->i_send_hdrs_dma);
589
590                 if (ic->i_recv_hdrs)
591                         ib_dma_free_coherent(dev,
592                                            ic->i_recv_ring.w_nr *
593                                                 sizeof(struct rds_header),
594                                            ic->i_recv_hdrs,
595                                            ic->i_recv_hdrs_dma);
596
597                 if (ic->i_ack)
598                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
599                                              ic->i_ack, ic->i_ack_dma);
600
601                 if (ic->i_sends)
602                         rds_ib_send_clear_ring(ic);
603                 if (ic->i_recvs)
604                         rds_ib_recv_clear_ring(ic);
605
606                 if (ic->i_cm_id->qp)
607                         rdma_destroy_qp(ic->i_cm_id);
608                 if (ic->i_send_cq)
609                         ib_destroy_cq(ic->i_send_cq);
610                 if (ic->i_recv_cq)
611                         ib_destroy_cq(ic->i_recv_cq);
612                 rdma_destroy_id(ic->i_cm_id);
613
614                 /*
615                  * Move connection back to the nodev list.
616                  */
617                 if (ic->rds_ibdev)
618                         rds_ib_remove_conn(ic->rds_ibdev, conn);
619
620                 ic->i_cm_id = NULL;
621                 ic->i_pd = NULL;
622                 ic->i_mr = NULL;
623                 ic->i_send_cq = NULL;
624                 ic->i_recv_cq = NULL;
625                 ic->i_send_hdrs = NULL;
626                 ic->i_recv_hdrs = NULL;
627                 ic->i_ack = NULL;
628         }
629         BUG_ON(ic->rds_ibdev);
630
631         /* Clear pending transmit */
632         if (ic->i_rm) {
633                 rds_message_put(ic->i_rm);
634                 ic->i_rm = NULL;
635         }
636
637         /* Clear the ACK state */
638         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
639 #ifdef KERNEL_HAS_ATOMIC64
640         atomic64_set(&ic->i_ack_next, 0);
641 #else
642         ic->i_ack_next = 0;
643 #endif
644         ic->i_ack_recv = 0;
645
646         /* Clear flow control state */
647         ic->i_flowctl = 0;
648         atomic_set(&ic->i_credits, 0);
649
650         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
651         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
652
653         if (ic->i_ibinc) {
654                 rds_inc_put(&ic->i_ibinc->ii_inc);
655                 ic->i_ibinc = NULL;
656         }
657
658         vfree(ic->i_sends);
659         ic->i_sends = NULL;
660         vfree(ic->i_recvs);
661         ic->i_recvs = NULL;
662 }
663
664 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
665 {
666         struct rds_ib_connection *ic;
667         unsigned long flags;
668
669         /* XXX too lazy? */
670         ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
671         if (ic == NULL)
672                 return -ENOMEM;
673
674         INIT_LIST_HEAD(&ic->ib_node);
675         mutex_init(&ic->i_recv_mutex);
676 #ifndef KERNEL_HAS_ATOMIC64
677         spin_lock_init(&ic->i_ack_lock);
678 #endif
679
680         /*
681          * rds_ib_conn_shutdown() waits for these to be emptied so they
682          * must be initialized before it can be called.
683          */
684         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
685         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
686
687         ic->conn = conn;
688         conn->c_transport_data = ic;
689
690         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
691         list_add_tail(&ic->ib_node, &ib_nodev_conns);
692         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
693
694
695         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
696         return 0;
697 }
698
699 /*
700  * Free a connection. Connection must be shut down and not set for reconnect.
701  */
702 void rds_ib_conn_free(void *arg)
703 {
704         struct rds_ib_connection *ic = arg;
705         spinlock_t      *lock_ptr;
706
707         rdsdebug("ic %p\n", ic);
708
709         /*
710          * Conn is either on a dev's list or on the nodev list.
711          * A race with shutdown() or connect() would cause problems
712          * (since rds_ibdev would change) but that should never happen.
713          */
714         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
715
716         spin_lock_irq(lock_ptr);
717         list_del(&ic->ib_node);
718         spin_unlock_irq(lock_ptr);
719
720         kfree(ic);
721 }
722
723
724 /*
725  * An error occurred on the connection
726  */
727 void
728 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
729 {
730         va_list ap;
731
732         rds_conn_drop(conn);
733
734         va_start(ap, fmt);
735         vprintk(fmt, ap);
736         va_end(ap);
737 }