RDS/IB: get the xmit max_sge from the RDS IB device on the connection
[linux-2.6.git] / net / rds / iw_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/slab.h>
36 #include <linux/vmalloc.h>
37
38 #include "rds.h"
39 #include "iw.h"
40
41 /*
42  * Set the selected protocol version
43  */
44 static void rds_iw_set_protocol(struct rds_connection *conn, unsigned int version)
45 {
46         conn->c_version = version;
47 }
48
49 /*
50  * Set up flow control
51  */
52 static void rds_iw_set_flow_control(struct rds_connection *conn, u32 credits)
53 {
54         struct rds_iw_connection *ic = conn->c_transport_data;
55
56         if (rds_iw_sysctl_flow_control && credits != 0) {
57                 /* We're doing flow control */
58                 ic->i_flowctl = 1;
59                 rds_iw_send_add_credits(conn, credits);
60         } else {
61                 ic->i_flowctl = 0;
62         }
63 }
64
65 /*
66  * Connection established.
67  * We get here for both outgoing and incoming connection.
68  */
69 void rds_iw_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
70 {
71         const struct rds_iw_connect_private *dp = NULL;
72         struct rds_iw_connection *ic = conn->c_transport_data;
73         struct rds_iw_device *rds_iwdev;
74         int err;
75
76         if (event->param.conn.private_data_len) {
77                 dp = event->param.conn.private_data;
78
79                 rds_iw_set_protocol(conn,
80                                 RDS_PROTOCOL(dp->dp_protocol_major,
81                                         dp->dp_protocol_minor));
82                 rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
83         }
84
85         /* update ib_device with this local ipaddr & conn */
86         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
87         err = rds_iw_update_cm_id(rds_iwdev, ic->i_cm_id);
88         if (err)
89                 printk(KERN_ERR "rds_iw_update_ipaddr failed (%d)\n", err);
90         rds_iw_add_conn(rds_iwdev, conn);
91
92         /* If the peer gave us the last packet it saw, process this as if
93          * we had received a regular ACK. */
94         if (dp && dp->dp_ack_seq)
95                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
96
97         printk(KERN_NOTICE "RDS/IW: connected to %pI4<->%pI4 version %u.%u%s\n",
98                         &conn->c_laddr, &conn->c_faddr,
99                         RDS_PROTOCOL_MAJOR(conn->c_version),
100                         RDS_PROTOCOL_MINOR(conn->c_version),
101                         ic->i_flowctl ? ", flow control" : "");
102
103         rds_connect_complete(conn);
104 }
105
106 static void rds_iw_cm_fill_conn_param(struct rds_connection *conn,
107                         struct rdma_conn_param *conn_param,
108                         struct rds_iw_connect_private *dp,
109                         u32 protocol_version)
110 {
111         struct rds_iw_connection *ic = conn->c_transport_data;
112
113         memset(conn_param, 0, sizeof(struct rdma_conn_param));
114         /* XXX tune these? */
115         conn_param->responder_resources = 1;
116         conn_param->initiator_depth = 1;
117
118         if (dp) {
119                 memset(dp, 0, sizeof(*dp));
120                 dp->dp_saddr = conn->c_laddr;
121                 dp->dp_daddr = conn->c_faddr;
122                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
123                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
124                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IW_SUPPORTED_PROTOCOLS);
125                 dp->dp_ack_seq = rds_iw_piggyb_ack(ic);
126
127                 /* Advertise flow control */
128                 if (ic->i_flowctl) {
129                         unsigned int credits;
130
131                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
132                         dp->dp_credit = cpu_to_be32(credits);
133                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
134                 }
135
136                 conn_param->private_data = dp;
137                 conn_param->private_data_len = sizeof(*dp);
138         }
139 }
140
141 static void rds_iw_cq_event_handler(struct ib_event *event, void *data)
142 {
143         rdsdebug("event %u data %p\n", event->event, data);
144 }
145
146 static void rds_iw_qp_event_handler(struct ib_event *event, void *data)
147 {
148         struct rds_connection *conn = data;
149         struct rds_iw_connection *ic = conn->c_transport_data;
150
151         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
152
153         switch (event->event) {
154         case IB_EVENT_COMM_EST:
155                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
156                 break;
157         case IB_EVENT_QP_REQ_ERR:
158         case IB_EVENT_QP_FATAL:
159         default:
160                 rdsdebug("Fatal QP Event %u "
161                         "- connection %pI4->%pI4, reconnecting\n",
162                         event->event, &conn->c_laddr,
163                         &conn->c_faddr);
164                 rds_conn_drop(conn);
165                 break;
166         }
167 }
168
169 /*
170  * Create a QP
171  */
172 static int rds_iw_init_qp_attrs(struct ib_qp_init_attr *attr,
173                 struct rds_iw_device *rds_iwdev,
174                 struct rds_iw_work_ring *send_ring,
175                 void (*send_cq_handler)(struct ib_cq *, void *),
176                 struct rds_iw_work_ring *recv_ring,
177                 void (*recv_cq_handler)(struct ib_cq *, void *),
178                 void *context)
179 {
180         struct ib_device *dev = rds_iwdev->dev;
181         unsigned int send_size, recv_size;
182         int ret;
183
184         /* The offset of 1 is to accomodate the additional ACK WR. */
185         send_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_send_wr + 1);
186         recv_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_recv_wr + 1);
187         rds_iw_ring_resize(send_ring, send_size - 1);
188         rds_iw_ring_resize(recv_ring, recv_size - 1);
189
190         memset(attr, 0, sizeof(*attr));
191         attr->event_handler = rds_iw_qp_event_handler;
192         attr->qp_context = context;
193         attr->cap.max_send_wr = send_size;
194         attr->cap.max_recv_wr = recv_size;
195         attr->cap.max_send_sge = rds_iwdev->max_sge;
196         attr->cap.max_recv_sge = RDS_IW_RECV_SGE;
197         attr->sq_sig_type = IB_SIGNAL_REQ_WR;
198         attr->qp_type = IB_QPT_RC;
199
200         attr->send_cq = ib_create_cq(dev, send_cq_handler,
201                                      rds_iw_cq_event_handler,
202                                      context, send_size, 0);
203         if (IS_ERR(attr->send_cq)) {
204                 ret = PTR_ERR(attr->send_cq);
205                 attr->send_cq = NULL;
206                 rdsdebug("ib_create_cq send failed: %d\n", ret);
207                 goto out;
208         }
209
210         attr->recv_cq = ib_create_cq(dev, recv_cq_handler,
211                                      rds_iw_cq_event_handler,
212                                      context, recv_size, 0);
213         if (IS_ERR(attr->recv_cq)) {
214                 ret = PTR_ERR(attr->recv_cq);
215                 attr->recv_cq = NULL;
216                 rdsdebug("ib_create_cq send failed: %d\n", ret);
217                 goto out;
218         }
219
220         ret = ib_req_notify_cq(attr->send_cq, IB_CQ_NEXT_COMP);
221         if (ret) {
222                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
223                 goto out;
224         }
225
226         ret = ib_req_notify_cq(attr->recv_cq, IB_CQ_SOLICITED);
227         if (ret) {
228                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
229                 goto out;
230         }
231
232 out:
233         if (ret) {
234                 if (attr->send_cq)
235                         ib_destroy_cq(attr->send_cq);
236                 if (attr->recv_cq)
237                         ib_destroy_cq(attr->recv_cq);
238         }
239         return ret;
240 }
241
242 /*
243  * This needs to be very careful to not leave IS_ERR pointers around for
244  * cleanup to trip over.
245  */
246 static int rds_iw_setup_qp(struct rds_connection *conn)
247 {
248         struct rds_iw_connection *ic = conn->c_transport_data;
249         struct ib_device *dev = ic->i_cm_id->device;
250         struct ib_qp_init_attr attr;
251         struct rds_iw_device *rds_iwdev;
252         int ret;
253
254         /* rds_iw_add_one creates a rds_iw_device object per IB device,
255          * and allocates a protection domain, memory range and MR pool
256          * for each.  If that fails for any reason, it will not register
257          * the rds_iwdev at all.
258          */
259         rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
260         if (!rds_iwdev) {
261                 if (printk_ratelimit())
262                         printk(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
263                                         dev->name);
264                 return -EOPNOTSUPP;
265         }
266
267         /* Protection domain and memory range */
268         ic->i_pd = rds_iwdev->pd;
269         ic->i_mr = rds_iwdev->mr;
270
271         ret = rds_iw_init_qp_attrs(&attr, rds_iwdev,
272                         &ic->i_send_ring, rds_iw_send_cq_comp_handler,
273                         &ic->i_recv_ring, rds_iw_recv_cq_comp_handler,
274                         conn);
275         if (ret < 0)
276                 goto out;
277
278         ic->i_send_cq = attr.send_cq;
279         ic->i_recv_cq = attr.recv_cq;
280
281         /*
282          * XXX this can fail if max_*_wr is too large?  Are we supposed
283          * to back off until we get a value that the hardware can support?
284          */
285         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
286         if (ret) {
287                 rdsdebug("rdma_create_qp failed: %d\n", ret);
288                 goto out;
289         }
290
291         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
292                                            ic->i_send_ring.w_nr *
293                                                 sizeof(struct rds_header),
294                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
295         if (!ic->i_send_hdrs) {
296                 ret = -ENOMEM;
297                 rdsdebug("ib_dma_alloc_coherent send failed\n");
298                 goto out;
299         }
300
301         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
302                                            ic->i_recv_ring.w_nr *
303                                                 sizeof(struct rds_header),
304                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
305         if (!ic->i_recv_hdrs) {
306                 ret = -ENOMEM;
307                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
308                 goto out;
309         }
310
311         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
312                                        &ic->i_ack_dma, GFP_KERNEL);
313         if (!ic->i_ack) {
314                 ret = -ENOMEM;
315                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
316                 goto out;
317         }
318
319         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
320         if (!ic->i_sends) {
321                 ret = -ENOMEM;
322                 rdsdebug("send allocation failed\n");
323                 goto out;
324         }
325         rds_iw_send_init_ring(ic);
326
327         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
328         if (!ic->i_recvs) {
329                 ret = -ENOMEM;
330                 rdsdebug("recv allocation failed\n");
331                 goto out;
332         }
333
334         rds_iw_recv_init_ring(ic);
335         rds_iw_recv_init_ack(ic);
336
337         /* Post receive buffers - as a side effect, this will update
338          * the posted credit count. */
339         rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
340
341         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
342                  ic->i_send_cq, ic->i_recv_cq);
343
344 out:
345         return ret;
346 }
347
348 static u32 rds_iw_protocol_compatible(const struct rds_iw_connect_private *dp)
349 {
350         u16 common;
351         u32 version = 0;
352
353         /* rdma_cm private data is odd - when there is any private data in the
354          * request, we will be given a pretty large buffer without telling us the
355          * original size. The only way to tell the difference is by looking at
356          * the contents, which are initialized to zero.
357          * If the protocol version fields aren't set, this is a connection attempt
358          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
359          * We really should have changed this for OFED 1.3 :-( */
360         if (dp->dp_protocol_major == 0)
361                 return RDS_PROTOCOL_3_0;
362
363         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IW_SUPPORTED_PROTOCOLS;
364         if (dp->dp_protocol_major == 3 && common) {
365                 version = RDS_PROTOCOL_3_0;
366                 while ((common >>= 1) != 0)
367                         version++;
368         } else if (printk_ratelimit()) {
369                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
370                         "incompatible protocol version %u.%u\n",
371                         &dp->dp_saddr,
372                         dp->dp_protocol_major,
373                         dp->dp_protocol_minor);
374         }
375         return version;
376 }
377
378 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
379                                     struct rdma_cm_event *event)
380 {
381         const struct rds_iw_connect_private *dp = event->param.conn.private_data;
382         struct rds_iw_connect_private dp_rep;
383         struct rds_connection *conn = NULL;
384         struct rds_iw_connection *ic = NULL;
385         struct rdma_conn_param conn_param;
386         struct rds_iw_device *rds_iwdev;
387         u32 version;
388         int err, destroy = 1;
389
390         /* Check whether the remote protocol version matches ours. */
391         version = rds_iw_protocol_compatible(dp);
392         if (!version)
393                 goto out;
394
395         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u\n",
396                  &dp->dp_saddr, &dp->dp_daddr,
397                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version));
398
399         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_iw_transport,
400                                GFP_KERNEL);
401         if (IS_ERR(conn)) {
402                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
403                 conn = NULL;
404                 goto out;
405         }
406
407         /*
408          * The connection request may occur while the
409          * previous connection exist, e.g. in case of failover.
410          * But as connections may be initiated simultaneously
411          * by both hosts, we have a random backoff mechanism -
412          * see the comment above rds_queue_reconnect()
413          */
414         mutex_lock(&conn->c_cm_lock);
415         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
416                 if (rds_conn_state(conn) == RDS_CONN_UP) {
417                         rdsdebug("incoming connect while connecting\n");
418                         rds_conn_drop(conn);
419                         rds_iw_stats_inc(s_iw_listen_closed_stale);
420                 } else
421                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
422                         /* Wait and see - our connect may still be succeeding */
423                         rds_iw_stats_inc(s_iw_connect_raced);
424                 }
425                 mutex_unlock(&conn->c_cm_lock);
426                 goto out;
427         }
428
429         ic = conn->c_transport_data;
430
431         rds_iw_set_protocol(conn, version);
432         rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
433
434         /* If the peer gave us the last packet it saw, process this as if
435          * we had received a regular ACK. */
436         if (dp->dp_ack_seq)
437                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
438
439         BUG_ON(cm_id->context);
440         BUG_ON(ic->i_cm_id);
441
442         ic->i_cm_id = cm_id;
443         cm_id->context = conn;
444
445         rds_iwdev = ib_get_client_data(cm_id->device, &rds_iw_client);
446         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
447
448         /* We got halfway through setting up the ib_connection, if we
449          * fail now, we have to take the long route out of this mess. */
450         destroy = 0;
451
452         err = rds_iw_setup_qp(conn);
453         if (err) {
454                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", err);
455                 mutex_unlock(&conn->c_cm_lock);
456                 goto out;
457         }
458
459         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
460
461         /* rdma_accept() calls rdma_reject() internally if it fails */
462         err = rdma_accept(cm_id, &conn_param);
463         mutex_unlock(&conn->c_cm_lock);
464         if (err) {
465                 rds_iw_conn_error(conn, "rdma_accept failed (%d)\n", err);
466                 goto out;
467         }
468
469         return 0;
470
471 out:
472         rdma_reject(cm_id, NULL, 0);
473         return destroy;
474 }
475
476
477 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id)
478 {
479         struct rds_connection *conn = cm_id->context;
480         struct rds_iw_connection *ic = conn->c_transport_data;
481         struct rdma_conn_param conn_param;
482         struct rds_iw_connect_private dp;
483         int ret;
484
485         /* If the peer doesn't do protocol negotiation, we must
486          * default to RDSv3.0 */
487         rds_iw_set_protocol(conn, RDS_PROTOCOL_3_0);
488         ic->i_flowctl = rds_iw_sysctl_flow_control;     /* advertise flow control */
489
490         ret = rds_iw_setup_qp(conn);
491         if (ret) {
492                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", ret);
493                 goto out;
494         }
495
496         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
497
498         ret = rdma_connect(cm_id, &conn_param);
499         if (ret)
500                 rds_iw_conn_error(conn, "rdma_connect failed (%d)\n", ret);
501
502 out:
503         /* Beware - returning non-zero tells the rdma_cm to destroy
504          * the cm_id. We should certainly not do it as long as we still
505          * "own" the cm_id. */
506         if (ret) {
507                 struct rds_iw_connection *ic = conn->c_transport_data;
508
509                 if (ic->i_cm_id == cm_id)
510                         ret = 0;
511         }
512         return ret;
513 }
514
515 int rds_iw_conn_connect(struct rds_connection *conn)
516 {
517         struct rds_iw_connection *ic = conn->c_transport_data;
518         struct rds_iw_device *rds_iwdev;
519         struct sockaddr_in src, dest;
520         int ret;
521
522         /* XXX I wonder what affect the port space has */
523         /* delegate cm event handler to rdma_transport */
524         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
525                                      RDMA_PS_TCP);
526         if (IS_ERR(ic->i_cm_id)) {
527                 ret = PTR_ERR(ic->i_cm_id);
528                 ic->i_cm_id = NULL;
529                 rdsdebug("rdma_create_id() failed: %d\n", ret);
530                 goto out;
531         }
532
533         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
534
535         src.sin_family = AF_INET;
536         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
537         src.sin_port = (__force u16)htons(0);
538
539         /* First, bind to the local address and device. */
540         ret = rdma_bind_addr(ic->i_cm_id, (struct sockaddr *) &src);
541         if (ret) {
542                 rdsdebug("rdma_bind_addr(%pI4) failed: %d\n",
543                                 &conn->c_laddr, ret);
544                 rdma_destroy_id(ic->i_cm_id);
545                 ic->i_cm_id = NULL;
546                 goto out;
547         }
548
549         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
550         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
551
552         dest.sin_family = AF_INET;
553         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
554         dest.sin_port = (__force u16)htons(RDS_PORT);
555
556         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
557                                 (struct sockaddr *)&dest,
558                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
559         if (ret) {
560                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
561                          ret);
562                 rdma_destroy_id(ic->i_cm_id);
563                 ic->i_cm_id = NULL;
564         }
565
566 out:
567         return ret;
568 }
569
570 /*
571  * This is so careful about only cleaning up resources that were built up
572  * so that it can be called at any point during startup.  In fact it
573  * can be called multiple times for a given connection.
574  */
575 void rds_iw_conn_shutdown(struct rds_connection *conn)
576 {
577         struct rds_iw_connection *ic = conn->c_transport_data;
578         int err = 0;
579         struct ib_qp_attr qp_attr;
580
581         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
582                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
583                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
584
585         if (ic->i_cm_id) {
586                 struct ib_device *dev = ic->i_cm_id->device;
587
588                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
589                 err = rdma_disconnect(ic->i_cm_id);
590                 if (err) {
591                         /* Actually this may happen quite frequently, when
592                          * an outgoing connect raced with an incoming connect.
593                          */
594                         rdsdebug("rds_iw_conn_shutdown: failed to disconnect,"
595                                    " cm: %p err %d\n", ic->i_cm_id, err);
596                 }
597
598                 if (ic->i_cm_id->qp) {
599                         qp_attr.qp_state = IB_QPS_ERR;
600                         ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
601                 }
602
603                 wait_event(rds_iw_ring_empty_wait,
604                         rds_iw_ring_empty(&ic->i_send_ring) &&
605                         rds_iw_ring_empty(&ic->i_recv_ring));
606
607                 if (ic->i_send_hdrs)
608                         ib_dma_free_coherent(dev,
609                                            ic->i_send_ring.w_nr *
610                                                 sizeof(struct rds_header),
611                                            ic->i_send_hdrs,
612                                            ic->i_send_hdrs_dma);
613
614                 if (ic->i_recv_hdrs)
615                         ib_dma_free_coherent(dev,
616                                            ic->i_recv_ring.w_nr *
617                                                 sizeof(struct rds_header),
618                                            ic->i_recv_hdrs,
619                                            ic->i_recv_hdrs_dma);
620
621                 if (ic->i_ack)
622                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
623                                              ic->i_ack, ic->i_ack_dma);
624
625                 if (ic->i_sends)
626                         rds_iw_send_clear_ring(ic);
627                 if (ic->i_recvs)
628                         rds_iw_recv_clear_ring(ic);
629
630                 if (ic->i_cm_id->qp)
631                         rdma_destroy_qp(ic->i_cm_id);
632                 if (ic->i_send_cq)
633                         ib_destroy_cq(ic->i_send_cq);
634                 if (ic->i_recv_cq)
635                         ib_destroy_cq(ic->i_recv_cq);
636
637                 /*
638                  * If associated with an rds_iw_device:
639                  *      Move connection back to the nodev list.
640                  *      Remove cm_id from the device cm_id list.
641                  */
642                 if (ic->rds_iwdev)
643                         rds_iw_remove_conn(ic->rds_iwdev, conn);
644
645                 rdma_destroy_id(ic->i_cm_id);
646
647                 ic->i_cm_id = NULL;
648                 ic->i_pd = NULL;
649                 ic->i_mr = NULL;
650                 ic->i_send_cq = NULL;
651                 ic->i_recv_cq = NULL;
652                 ic->i_send_hdrs = NULL;
653                 ic->i_recv_hdrs = NULL;
654                 ic->i_ack = NULL;
655         }
656         BUG_ON(ic->rds_iwdev);
657
658         /* Clear pending transmit */
659         if (ic->i_rm) {
660                 rds_message_put(ic->i_rm);
661                 ic->i_rm = NULL;
662         }
663
664         /* Clear the ACK state */
665         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
666 #ifdef KERNEL_HAS_ATOMIC64
667         atomic64_set(&ic->i_ack_next, 0);
668 #else
669         ic->i_ack_next = 0;
670 #endif
671         ic->i_ack_recv = 0;
672
673         /* Clear flow control state */
674         ic->i_flowctl = 0;
675         atomic_set(&ic->i_credits, 0);
676
677         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
678         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
679
680         if (ic->i_iwinc) {
681                 rds_inc_put(&ic->i_iwinc->ii_inc);
682                 ic->i_iwinc = NULL;
683         }
684
685         vfree(ic->i_sends);
686         ic->i_sends = NULL;
687         vfree(ic->i_recvs);
688         ic->i_recvs = NULL;
689         rdsdebug("shutdown complete\n");
690 }
691
692 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
693 {
694         struct rds_iw_connection *ic;
695         unsigned long flags;
696
697         /* XXX too lazy? */
698         ic = kzalloc(sizeof(struct rds_iw_connection), GFP_KERNEL);
699         if (!ic)
700                 return -ENOMEM;
701
702         INIT_LIST_HEAD(&ic->iw_node);
703         tasklet_init(&ic->i_recv_tasklet, rds_iw_recv_tasklet_fn,
704                      (unsigned long) ic);
705         mutex_init(&ic->i_recv_mutex);
706 #ifndef KERNEL_HAS_ATOMIC64
707         spin_lock_init(&ic->i_ack_lock);
708 #endif
709
710         /*
711          * rds_iw_conn_shutdown() waits for these to be emptied so they
712          * must be initialized before it can be called.
713          */
714         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
715         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
716
717         ic->conn = conn;
718         conn->c_transport_data = ic;
719
720         spin_lock_irqsave(&iw_nodev_conns_lock, flags);
721         list_add_tail(&ic->iw_node, &iw_nodev_conns);
722         spin_unlock_irqrestore(&iw_nodev_conns_lock, flags);
723
724
725         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
726         return 0;
727 }
728
729 /*
730  * Free a connection. Connection must be shut down and not set for reconnect.
731  */
732 void rds_iw_conn_free(void *arg)
733 {
734         struct rds_iw_connection *ic = arg;
735         spinlock_t      *lock_ptr;
736
737         rdsdebug("ic %p\n", ic);
738
739         /*
740          * Conn is either on a dev's list or on the nodev list.
741          * A race with shutdown() or connect() would cause problems
742          * (since rds_iwdev would change) but that should never happen.
743          */
744         lock_ptr = ic->rds_iwdev ? &ic->rds_iwdev->spinlock : &iw_nodev_conns_lock;
745
746         spin_lock_irq(lock_ptr);
747         list_del(&ic->iw_node);
748         spin_unlock_irq(lock_ptr);
749
750         kfree(ic);
751 }
752
753 /*
754  * An error occurred on the connection
755  */
756 void
757 __rds_iw_conn_error(struct rds_connection *conn, const char *fmt, ...)
758 {
759         va_list ap;
760
761         rds_conn_drop(conn);
762
763         va_start(ap, fmt);
764         vprintk(fmt, ap);
765         va_end(ap);
766 }