]>
Commit | Line | Data |
---|---|---|
ec16227e AG |
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> | |
5a0e3ad6 | 35 | #include <linux/slab.h> |
ec16227e AG |
36 | #include <linux/vmalloc.h> |
37 | ||
38 | #include "rds.h" | |
39 | #include "ib.h" | |
40 | ||
41 | /* | |
42 | * Set the selected protocol version | |
43 | */ | |
44 | static void rds_ib_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_ib_set_flow_control(struct rds_connection *conn, u32 credits) | |
53 | { | |
54 | struct rds_ib_connection *ic = conn->c_transport_data; | |
55 | ||
56 | if (rds_ib_sysctl_flow_control && credits != 0) { | |
57 | /* We're doing flow control */ | |
58 | ic->i_flowctl = 1; | |
59 | rds_ib_send_add_credits(conn, credits); | |
60 | } else { | |
61 | ic->i_flowctl = 0; | |
62 | } | |
63 | } | |
64 | ||
65 | /* | |
66 | * Tune RNR behavior. Without flow control, we use a rather | |
67 | * low timeout, but not the absolute minimum - this should | |
68 | * be tunable. | |
69 | * | |
70 | * We already set the RNR retry count to 7 (which is the | |
71 | * smallest infinite number :-) above. | |
72 | * If flow control is off, we want to change this back to 0 | |
73 | * so that we learn quickly when our credit accounting is | |
74 | * buggy. | |
75 | * | |
76 | * Caller passes in a qp_attr pointer - don't waste stack spacv | |
77 | * by allocation this twice. | |
78 | */ | |
79 | static void | |
80 | rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) | |
81 | { | |
82 | int ret; | |
83 | ||
84 | attr->min_rnr_timer = IB_RNR_TIMER_000_32; | |
85 | ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); | |
86 | if (ret) | |
87 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); | |
88 | } | |
89 | ||
90 | /* | |
91 | * Connection established. | |
92 | * We get here for both outgoing and incoming connection. | |
93 | */ | |
94 | void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) | |
95 | { | |
96 | const struct rds_ib_connect_private *dp = NULL; | |
97 | struct rds_ib_connection *ic = conn->c_transport_data; | |
98 | struct rds_ib_device *rds_ibdev; | |
99 | struct ib_qp_attr qp_attr; | |
100 | int err; | |
101 | ||
9ddbcfa0 | 102 | if (event->param.conn.private_data_len >= sizeof(*dp)) { |
ec16227e AG |
103 | dp = event->param.conn.private_data; |
104 | ||
02a6a259 AG |
105 | /* make sure it isn't empty data */ |
106 | if (dp->dp_protocol_major) { | |
107 | rds_ib_set_protocol(conn, | |
ec16227e | 108 | RDS_PROTOCOL(dp->dp_protocol_major, |
02a6a259 AG |
109 | dp->dp_protocol_minor)); |
110 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); | |
111 | } | |
ec16227e AG |
112 | } |
113 | ||
f147dd9e AG |
114 | if (conn->c_version < RDS_PROTOCOL(3,1)) { |
115 | printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed," | |
116 | " no longer supported\n", | |
117 | &conn->c_faddr, | |
118 | RDS_PROTOCOL_MAJOR(conn->c_version), | |
119 | RDS_PROTOCOL_MINOR(conn->c_version)); | |
120 | rds_conn_destroy(conn); | |
121 | return; | |
122 | } else { | |
123 | printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n", | |
124 | &conn->c_faddr, | |
125 | RDS_PROTOCOL_MAJOR(conn->c_version), | |
126 | RDS_PROTOCOL_MINOR(conn->c_version), | |
127 | ic->i_flowctl ? ", flow control" : ""); | |
128 | } | |
ec16227e | 129 | |
e11d912a AG |
130 | /* |
131 | * Init rings and fill recv. this needs to wait until protocol negotiation | |
132 | * is complete, since ring layout is different from 3.0 to 3.1. | |
133 | */ | |
134 | rds_ib_send_init_ring(ic); | |
135 | rds_ib_recv_init_ring(ic); | |
136 | /* Post receive buffers - as a side effect, this will update | |
137 | * the posted credit count. */ | |
f17a1a55 | 138 | rds_ib_recv_refill(conn, 1); |
e11d912a | 139 | |
ec16227e AG |
140 | /* Tune RNR behavior */ |
141 | rds_ib_tune_rnr(ic, &qp_attr); | |
142 | ||
143 | qp_attr.qp_state = IB_QPS_RTS; | |
144 | err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); | |
145 | if (err) | |
146 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); | |
147 | ||
148 | /* update ib_device with this local ipaddr & conn */ | |
149 | rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); | |
150 | err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr); | |
151 | if (err) | |
152 | printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err); | |
745cbcca | 153 | rds_ib_add_conn(rds_ibdev, conn); |
ec16227e AG |
154 | |
155 | /* If the peer gave us the last packet it saw, process this as if | |
156 | * we had received a regular ACK. */ | |
157 | if (dp && dp->dp_ack_seq) | |
158 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); | |
159 | ||
160 | rds_connect_complete(conn); | |
161 | } | |
162 | ||
163 | static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, | |
164 | struct rdma_conn_param *conn_param, | |
165 | struct rds_ib_connect_private *dp, | |
40589e74 AG |
166 | u32 protocol_version, |
167 | u32 max_responder_resources, | |
168 | u32 max_initiator_depth) | |
ec16227e | 169 | { |
40589e74 AG |
170 | struct rds_ib_connection *ic = conn->c_transport_data; |
171 | struct rds_ib_device *rds_ibdev; | |
172 | ||
ec16227e | 173 | memset(conn_param, 0, sizeof(struct rdma_conn_param)); |
40589e74 AG |
174 | |
175 | rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); | |
176 | ||
177 | conn_param->responder_resources = | |
178 | min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources); | |
179 | conn_param->initiator_depth = | |
180 | min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth); | |
3ba23ade | 181 | conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); |
ec16227e AG |
182 | conn_param->rnr_retry_count = 7; |
183 | ||
184 | if (dp) { | |
ec16227e AG |
185 | memset(dp, 0, sizeof(*dp)); |
186 | dp->dp_saddr = conn->c_laddr; | |
187 | dp->dp_daddr = conn->c_faddr; | |
188 | dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version); | |
189 | dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version); | |
190 | dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); | |
191 | dp->dp_ack_seq = rds_ib_piggyb_ack(ic); | |
192 | ||
193 | /* Advertise flow control */ | |
194 | if (ic->i_flowctl) { | |
195 | unsigned int credits; | |
196 | ||
197 | credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)); | |
198 | dp->dp_credit = cpu_to_be32(credits); | |
199 | atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits); | |
200 | } | |
201 | ||
202 | conn_param->private_data = dp; | |
203 | conn_param->private_data_len = sizeof(*dp); | |
204 | } | |
205 | } | |
206 | ||
207 | static void rds_ib_cq_event_handler(struct ib_event *event, void *data) | |
208 | { | |
209 | rdsdebug("event %u data %p\n", event->event, data); | |
210 | } | |
211 | ||
212 | static void rds_ib_qp_event_handler(struct ib_event *event, void *data) | |
213 | { | |
214 | struct rds_connection *conn = data; | |
215 | struct rds_ib_connection *ic = conn->c_transport_data; | |
216 | ||
217 | rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event); | |
218 | ||
219 | switch (event->event) { | |
220 | case IB_EVENT_COMM_EST: | |
221 | rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); | |
222 | break; | |
223 | default: | |
97069788 | 224 | rdsdebug("Fatal QP Event %u " |
fdf6e6b4 AG |
225 | "- connection %pI4->%pI4, reconnecting\n", |
226 | event->event, &conn->c_laddr, &conn->c_faddr); | |
97069788 | 227 | rds_conn_drop(conn); |
ec16227e AG |
228 | break; |
229 | } | |
230 | } | |
231 | ||
232 | /* | |
233 | * This needs to be very careful to not leave IS_ERR pointers around for | |
234 | * cleanup to trip over. | |
235 | */ | |
236 | static int rds_ib_setup_qp(struct rds_connection *conn) | |
237 | { | |
238 | struct rds_ib_connection *ic = conn->c_transport_data; | |
239 | struct ib_device *dev = ic->i_cm_id->device; | |
240 | struct ib_qp_init_attr attr; | |
241 | struct rds_ib_device *rds_ibdev; | |
242 | int ret; | |
243 | ||
244 | /* rds_ib_add_one creates a rds_ib_device object per IB device, | |
245 | * and allocates a protection domain, memory range and FMR pool | |
246 | * for each. If that fails for any reason, it will not register | |
247 | * the rds_ibdev at all. | |
248 | */ | |
249 | rds_ibdev = ib_get_client_data(dev, &rds_ib_client); | |
8690bfa1 | 250 | if (!rds_ibdev) { |
ec16227e AG |
251 | if (printk_ratelimit()) |
252 | printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n", | |
253 | dev->name); | |
254 | return -EOPNOTSUPP; | |
255 | } | |
256 | ||
257 | if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) | |
258 | rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); | |
259 | if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) | |
260 | rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); | |
261 | ||
262 | /* Protection domain and memory range */ | |
263 | ic->i_pd = rds_ibdev->pd; | |
264 | ic->i_mr = rds_ibdev->mr; | |
265 | ||
266 | ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler, | |
267 | rds_ib_cq_event_handler, conn, | |
268 | ic->i_send_ring.w_nr + 1, 0); | |
269 | if (IS_ERR(ic->i_send_cq)) { | |
270 | ret = PTR_ERR(ic->i_send_cq); | |
271 | ic->i_send_cq = NULL; | |
272 | rdsdebug("ib_create_cq send failed: %d\n", ret); | |
273 | goto out; | |
274 | } | |
275 | ||
276 | ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler, | |
277 | rds_ib_cq_event_handler, conn, | |
278 | ic->i_recv_ring.w_nr, 0); | |
279 | if (IS_ERR(ic->i_recv_cq)) { | |
280 | ret = PTR_ERR(ic->i_recv_cq); | |
281 | ic->i_recv_cq = NULL; | |
282 | rdsdebug("ib_create_cq recv failed: %d\n", ret); | |
283 | goto out; | |
284 | } | |
285 | ||
286 | ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); | |
287 | if (ret) { | |
288 | rdsdebug("ib_req_notify_cq send failed: %d\n", ret); | |
289 | goto out; | |
290 | } | |
291 | ||
292 | ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); | |
293 | if (ret) { | |
294 | rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); | |
295 | goto out; | |
296 | } | |
297 | ||
298 | /* XXX negotiate max send/recv with remote? */ | |
299 | memset(&attr, 0, sizeof(attr)); | |
300 | attr.event_handler = rds_ib_qp_event_handler; | |
301 | attr.qp_context = conn; | |
302 | /* + 1 to allow for the single ack message */ | |
303 | attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1; | |
304 | attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; | |
305 | attr.cap.max_send_sge = rds_ibdev->max_sge; | |
306 | attr.cap.max_recv_sge = RDS_IB_RECV_SGE; | |
307 | attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
308 | attr.qp_type = IB_QPT_RC; | |
309 | attr.send_cq = ic->i_send_cq; | |
310 | attr.recv_cq = ic->i_recv_cq; | |
311 | ||
312 | /* | |
313 | * XXX this can fail if max_*_wr is too large? Are we supposed | |
314 | * to back off until we get a value that the hardware can support? | |
315 | */ | |
316 | ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); | |
317 | if (ret) { | |
318 | rdsdebug("rdma_create_qp failed: %d\n", ret); | |
319 | goto out; | |
320 | } | |
321 | ||
322 | ic->i_send_hdrs = ib_dma_alloc_coherent(dev, | |
323 | ic->i_send_ring.w_nr * | |
324 | sizeof(struct rds_header), | |
325 | &ic->i_send_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 326 | if (!ic->i_send_hdrs) { |
ec16227e AG |
327 | ret = -ENOMEM; |
328 | rdsdebug("ib_dma_alloc_coherent send failed\n"); | |
329 | goto out; | |
330 | } | |
331 | ||
332 | ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, | |
333 | ic->i_recv_ring.w_nr * | |
334 | sizeof(struct rds_header), | |
335 | &ic->i_recv_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 336 | if (!ic->i_recv_hdrs) { |
ec16227e AG |
337 | ret = -ENOMEM; |
338 | rdsdebug("ib_dma_alloc_coherent recv failed\n"); | |
339 | goto out; | |
340 | } | |
341 | ||
342 | ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), | |
343 | &ic->i_ack_dma, GFP_KERNEL); | |
8690bfa1 | 344 | if (!ic->i_ack) { |
ec16227e AG |
345 | ret = -ENOMEM; |
346 | rdsdebug("ib_dma_alloc_coherent ack failed\n"); | |
347 | goto out; | |
348 | } | |
349 | ||
350 | ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); | |
8690bfa1 | 351 | if (!ic->i_sends) { |
ec16227e AG |
352 | ret = -ENOMEM; |
353 | rdsdebug("send allocation failed\n"); | |
354 | goto out; | |
355 | } | |
e11d912a | 356 | memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); |
ec16227e AG |
357 | |
358 | ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); | |
8690bfa1 | 359 | if (!ic->i_recvs) { |
ec16227e AG |
360 | ret = -ENOMEM; |
361 | rdsdebug("recv allocation failed\n"); | |
362 | goto out; | |
363 | } | |
e11d912a | 364 | memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); |
ec16227e | 365 | |
ec16227e AG |
366 | rds_ib_recv_init_ack(ic); |
367 | ||
ec16227e AG |
368 | rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr, |
369 | ic->i_send_cq, ic->i_recv_cq); | |
370 | ||
371 | out: | |
372 | return ret; | |
373 | } | |
374 | ||
9ddbcfa0 | 375 | static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event) |
ec16227e | 376 | { |
9ddbcfa0 | 377 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; |
ec16227e AG |
378 | u16 common; |
379 | u32 version = 0; | |
380 | ||
9ddbcfa0 AG |
381 | /* |
382 | * rdma_cm private data is odd - when there is any private data in the | |
ec16227e AG |
383 | * request, we will be given a pretty large buffer without telling us the |
384 | * original size. The only way to tell the difference is by looking at | |
385 | * the contents, which are initialized to zero. | |
386 | * If the protocol version fields aren't set, this is a connection attempt | |
387 | * from an older version. This could could be 3.0 or 2.0 - we can't tell. | |
9ddbcfa0 AG |
388 | * We really should have changed this for OFED 1.3 :-( |
389 | */ | |
390 | ||
391 | /* Be paranoid. RDS always has privdata */ | |
392 | if (!event->param.conn.private_data_len) { | |
393 | printk(KERN_NOTICE "RDS incoming connection has no private data, " | |
394 | "rejecting\n"); | |
395 | return 0; | |
396 | } | |
397 | ||
398 | /* Even if len is crap *now* I still want to check it. -ASG */ | |
f64f9e71 JP |
399 | if (event->param.conn.private_data_len < sizeof (*dp) || |
400 | dp->dp_protocol_major == 0) | |
ec16227e AG |
401 | return RDS_PROTOCOL_3_0; |
402 | ||
403 | common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS; | |
404 | if (dp->dp_protocol_major == 3 && common) { | |
405 | version = RDS_PROTOCOL_3_0; | |
406 | while ((common >>= 1) != 0) | |
407 | version++; | |
408 | } else if (printk_ratelimit()) { | |
409 | printk(KERN_NOTICE "RDS: Connection from %pI4 using " | |
410 | "incompatible protocol version %u.%u\n", | |
411 | &dp->dp_saddr, | |
412 | dp->dp_protocol_major, | |
413 | dp->dp_protocol_minor); | |
414 | } | |
415 | return version; | |
416 | } | |
417 | ||
418 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, | |
419 | struct rdma_cm_event *event) | |
420 | { | |
421 | __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; | |
422 | __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; | |
423 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; | |
424 | struct rds_ib_connect_private dp_rep; | |
425 | struct rds_connection *conn = NULL; | |
426 | struct rds_ib_connection *ic = NULL; | |
427 | struct rdma_conn_param conn_param; | |
428 | u32 version; | |
429 | int err, destroy = 1; | |
430 | ||
431 | /* Check whether the remote protocol version matches ours. */ | |
9ddbcfa0 | 432 | version = rds_ib_protocol_compatible(event); |
ec16227e AG |
433 | if (!version) |
434 | goto out; | |
435 | ||
436 | rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid " | |
437 | "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr, | |
438 | RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), | |
439 | (unsigned long long)be64_to_cpu(lguid), | |
440 | (unsigned long long)be64_to_cpu(fguid)); | |
441 | ||
442 | conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport, | |
443 | GFP_KERNEL); | |
444 | if (IS_ERR(conn)) { | |
445 | rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); | |
446 | conn = NULL; | |
447 | goto out; | |
448 | } | |
449 | ||
450 | /* | |
451 | * The connection request may occur while the | |
452 | * previous connection exist, e.g. in case of failover. | |
453 | * But as connections may be initiated simultaneously | |
454 | * by both hosts, we have a random backoff mechanism - | |
455 | * see the comment above rds_queue_reconnect() | |
456 | */ | |
457 | mutex_lock(&conn->c_cm_lock); | |
458 | if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { | |
459 | if (rds_conn_state(conn) == RDS_CONN_UP) { | |
460 | rdsdebug("incoming connect while connecting\n"); | |
461 | rds_conn_drop(conn); | |
462 | rds_ib_stats_inc(s_ib_listen_closed_stale); | |
463 | } else | |
464 | if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { | |
465 | /* Wait and see - our connect may still be succeeding */ | |
466 | rds_ib_stats_inc(s_ib_connect_raced); | |
467 | } | |
468 | mutex_unlock(&conn->c_cm_lock); | |
469 | goto out; | |
470 | } | |
471 | ||
472 | ic = conn->c_transport_data; | |
473 | ||
474 | rds_ib_set_protocol(conn, version); | |
475 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); | |
476 | ||
477 | /* If the peer gave us the last packet it saw, process this as if | |
478 | * we had received a regular ACK. */ | |
479 | if (dp->dp_ack_seq) | |
480 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); | |
481 | ||
482 | BUG_ON(cm_id->context); | |
483 | BUG_ON(ic->i_cm_id); | |
484 | ||
485 | ic->i_cm_id = cm_id; | |
486 | cm_id->context = conn; | |
487 | ||
488 | /* We got halfway through setting up the ib_connection, if we | |
489 | * fail now, we have to take the long route out of this mess. */ | |
490 | destroy = 0; | |
491 | ||
492 | err = rds_ib_setup_qp(conn); | |
493 | if (err) { | |
494 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); | |
5daf47bb | 495 | mutex_unlock(&conn->c_cm_lock); |
ec16227e AG |
496 | goto out; |
497 | } | |
498 | ||
40589e74 AG |
499 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version, |
500 | event->param.conn.responder_resources, | |
501 | event->param.conn.initiator_depth); | |
ec16227e AG |
502 | |
503 | /* rdma_accept() calls rdma_reject() internally if it fails */ | |
504 | err = rdma_accept(cm_id, &conn_param); | |
505 | mutex_unlock(&conn->c_cm_lock); | |
506 | if (err) { | |
507 | rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); | |
508 | goto out; | |
509 | } | |
510 | ||
511 | return 0; | |
512 | ||
513 | out: | |
514 | rdma_reject(cm_id, NULL, 0); | |
515 | return destroy; | |
516 | } | |
517 | ||
518 | ||
519 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id) | |
520 | { | |
521 | struct rds_connection *conn = cm_id->context; | |
522 | struct rds_ib_connection *ic = conn->c_transport_data; | |
523 | struct rdma_conn_param conn_param; | |
524 | struct rds_ib_connect_private dp; | |
525 | int ret; | |
526 | ||
527 | /* If the peer doesn't do protocol negotiation, we must | |
528 | * default to RDSv3.0 */ | |
529 | rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); | |
530 | ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ | |
531 | ||
532 | ret = rds_ib_setup_qp(conn); | |
533 | if (ret) { | |
534 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); | |
535 | goto out; | |
536 | } | |
537 | ||
40589e74 AG |
538 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION, |
539 | UINT_MAX, UINT_MAX); | |
ec16227e AG |
540 | ret = rdma_connect(cm_id, &conn_param); |
541 | if (ret) | |
542 | rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); | |
543 | ||
544 | out: | |
545 | /* Beware - returning non-zero tells the rdma_cm to destroy | |
546 | * the cm_id. We should certainly not do it as long as we still | |
547 | * "own" the cm_id. */ | |
548 | if (ret) { | |
549 | if (ic->i_cm_id == cm_id) | |
550 | ret = 0; | |
551 | } | |
552 | return ret; | |
553 | } | |
554 | ||
555 | int rds_ib_conn_connect(struct rds_connection *conn) | |
556 | { | |
557 | struct rds_ib_connection *ic = conn->c_transport_data; | |
558 | struct sockaddr_in src, dest; | |
559 | int ret; | |
560 | ||
561 | /* XXX I wonder what affect the port space has */ | |
562 | /* delegate cm event handler to rdma_transport */ | |
563 | ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn, | |
564 | RDMA_PS_TCP); | |
565 | if (IS_ERR(ic->i_cm_id)) { | |
566 | ret = PTR_ERR(ic->i_cm_id); | |
567 | ic->i_cm_id = NULL; | |
568 | rdsdebug("rdma_create_id() failed: %d\n", ret); | |
569 | goto out; | |
570 | } | |
571 | ||
572 | rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); | |
573 | ||
574 | src.sin_family = AF_INET; | |
575 | src.sin_addr.s_addr = (__force u32)conn->c_laddr; | |
576 | src.sin_port = (__force u16)htons(0); | |
577 | ||
578 | dest.sin_family = AF_INET; | |
579 | dest.sin_addr.s_addr = (__force u32)conn->c_faddr; | |
580 | dest.sin_port = (__force u16)htons(RDS_PORT); | |
581 | ||
582 | ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, | |
583 | (struct sockaddr *)&dest, | |
584 | RDS_RDMA_RESOLVE_TIMEOUT_MS); | |
585 | if (ret) { | |
586 | rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, | |
587 | ret); | |
588 | rdma_destroy_id(ic->i_cm_id); | |
589 | ic->i_cm_id = NULL; | |
590 | } | |
591 | ||
592 | out: | |
593 | return ret; | |
594 | } | |
595 | ||
596 | /* | |
597 | * This is so careful about only cleaning up resources that were built up | |
598 | * so that it can be called at any point during startup. In fact it | |
599 | * can be called multiple times for a given connection. | |
600 | */ | |
601 | void rds_ib_conn_shutdown(struct rds_connection *conn) | |
602 | { | |
603 | struct rds_ib_connection *ic = conn->c_transport_data; | |
604 | int err = 0; | |
605 | ||
606 | rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, | |
607 | ic->i_pd, ic->i_send_cq, ic->i_recv_cq, | |
608 | ic->i_cm_id ? ic->i_cm_id->qp : NULL); | |
609 | ||
610 | if (ic->i_cm_id) { | |
611 | struct ib_device *dev = ic->i_cm_id->device; | |
612 | ||
613 | rdsdebug("disconnecting cm %p\n", ic->i_cm_id); | |
614 | err = rdma_disconnect(ic->i_cm_id); | |
615 | if (err) { | |
616 | /* Actually this may happen quite frequently, when | |
617 | * an outgoing connect raced with an incoming connect. | |
618 | */ | |
619 | rdsdebug("failed to disconnect, cm: %p err %d\n", | |
620 | ic->i_cm_id, err); | |
621 | } | |
622 | ||
e32b4a70 AG |
623 | /* |
624 | * Don't wait for the send ring to be empty -- there may be completed | |
625 | * non-signaled entries sitting on there. We unmap these below. | |
626 | */ | |
ec16227e | 627 | wait_event(rds_ib_ring_empty_wait, |
ec16227e AG |
628 | rds_ib_ring_empty(&ic->i_recv_ring)); |
629 | ||
630 | if (ic->i_send_hdrs) | |
631 | ib_dma_free_coherent(dev, | |
632 | ic->i_send_ring.w_nr * | |
633 | sizeof(struct rds_header), | |
634 | ic->i_send_hdrs, | |
635 | ic->i_send_hdrs_dma); | |
636 | ||
637 | if (ic->i_recv_hdrs) | |
638 | ib_dma_free_coherent(dev, | |
639 | ic->i_recv_ring.w_nr * | |
640 | sizeof(struct rds_header), | |
641 | ic->i_recv_hdrs, | |
642 | ic->i_recv_hdrs_dma); | |
643 | ||
644 | if (ic->i_ack) | |
645 | ib_dma_free_coherent(dev, sizeof(struct rds_header), | |
646 | ic->i_ack, ic->i_ack_dma); | |
647 | ||
648 | if (ic->i_sends) | |
649 | rds_ib_send_clear_ring(ic); | |
650 | if (ic->i_recvs) | |
651 | rds_ib_recv_clear_ring(ic); | |
652 | ||
653 | if (ic->i_cm_id->qp) | |
654 | rdma_destroy_qp(ic->i_cm_id); | |
655 | if (ic->i_send_cq) | |
656 | ib_destroy_cq(ic->i_send_cq); | |
657 | if (ic->i_recv_cq) | |
658 | ib_destroy_cq(ic->i_recv_cq); | |
659 | rdma_destroy_id(ic->i_cm_id); | |
660 | ||
661 | /* | |
662 | * Move connection back to the nodev list. | |
663 | */ | |
745cbcca AG |
664 | if (ic->rds_ibdev) |
665 | rds_ib_remove_conn(ic->rds_ibdev, conn); | |
ec16227e AG |
666 | |
667 | ic->i_cm_id = NULL; | |
668 | ic->i_pd = NULL; | |
669 | ic->i_mr = NULL; | |
670 | ic->i_send_cq = NULL; | |
671 | ic->i_recv_cq = NULL; | |
672 | ic->i_send_hdrs = NULL; | |
673 | ic->i_recv_hdrs = NULL; | |
674 | ic->i_ack = NULL; | |
675 | } | |
676 | BUG_ON(ic->rds_ibdev); | |
677 | ||
678 | /* Clear pending transmit */ | |
ff3d7d36 AG |
679 | if (ic->i_data_op) { |
680 | struct rds_message *rm; | |
681 | ||
682 | rm = container_of(ic->i_data_op, struct rds_message, data); | |
683 | rds_message_put(rm); | |
684 | ic->i_data_op = NULL; | |
ec16227e AG |
685 | } |
686 | ||
687 | /* Clear the ACK state */ | |
688 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
8cbd9606 AG |
689 | #ifdef KERNEL_HAS_ATOMIC64 |
690 | atomic64_set(&ic->i_ack_next, 0); | |
691 | #else | |
692 | ic->i_ack_next = 0; | |
693 | #endif | |
ec16227e AG |
694 | ic->i_ack_recv = 0; |
695 | ||
696 | /* Clear flow control state */ | |
697 | ic->i_flowctl = 0; | |
698 | atomic_set(&ic->i_credits, 0); | |
699 | ||
700 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
701 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
702 | ||
703 | if (ic->i_ibinc) { | |
704 | rds_inc_put(&ic->i_ibinc->ii_inc); | |
705 | ic->i_ibinc = NULL; | |
706 | } | |
707 | ||
708 | vfree(ic->i_sends); | |
709 | ic->i_sends = NULL; | |
710 | vfree(ic->i_recvs); | |
711 | ic->i_recvs = NULL; | |
712 | } | |
713 | ||
714 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) | |
715 | { | |
716 | struct rds_ib_connection *ic; | |
717 | unsigned long flags; | |
718 | ||
719 | /* XXX too lazy? */ | |
720 | ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL); | |
8690bfa1 | 721 | if (!ic) |
ec16227e AG |
722 | return -ENOMEM; |
723 | ||
724 | INIT_LIST_HEAD(&ic->ib_node); | |
d521b63b AG |
725 | tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn, |
726 | (unsigned long) ic); | |
ec16227e | 727 | mutex_init(&ic->i_recv_mutex); |
8cbd9606 AG |
728 | #ifndef KERNEL_HAS_ATOMIC64 |
729 | spin_lock_init(&ic->i_ack_lock); | |
730 | #endif | |
ec16227e AG |
731 | |
732 | /* | |
733 | * rds_ib_conn_shutdown() waits for these to be emptied so they | |
734 | * must be initialized before it can be called. | |
735 | */ | |
736 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
737 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
738 | ||
739 | ic->conn = conn; | |
740 | conn->c_transport_data = ic; | |
741 | ||
742 | spin_lock_irqsave(&ib_nodev_conns_lock, flags); | |
743 | list_add_tail(&ic->ib_node, &ib_nodev_conns); | |
744 | spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); | |
745 | ||
746 | ||
747 | rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); | |
748 | return 0; | |
749 | } | |
750 | ||
745cbcca AG |
751 | /* |
752 | * Free a connection. Connection must be shut down and not set for reconnect. | |
753 | */ | |
ec16227e AG |
754 | void rds_ib_conn_free(void *arg) |
755 | { | |
756 | struct rds_ib_connection *ic = arg; | |
745cbcca AG |
757 | spinlock_t *lock_ptr; |
758 | ||
ec16227e | 759 | rdsdebug("ic %p\n", ic); |
745cbcca AG |
760 | |
761 | /* | |
762 | * Conn is either on a dev's list or on the nodev list. | |
763 | * A race with shutdown() or connect() would cause problems | |
764 | * (since rds_ibdev would change) but that should never happen. | |
765 | */ | |
766 | lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; | |
767 | ||
768 | spin_lock_irq(lock_ptr); | |
ec16227e | 769 | list_del(&ic->ib_node); |
745cbcca AG |
770 | spin_unlock_irq(lock_ptr); |
771 | ||
ec16227e AG |
772 | kfree(ic); |
773 | } | |
774 | ||
775 | ||
776 | /* | |
777 | * An error occurred on the connection | |
778 | */ | |
779 | void | |
780 | __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) | |
781 | { | |
782 | va_list ap; | |
783 | ||
784 | rds_conn_drop(conn); | |
785 | ||
786 | va_start(ap, fmt); | |
787 | vprintk(fmt, ap); | |
788 | va_end(ap); | |
789 | } |