]>
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 | 36 | #include <linux/vmalloc.h> |
cb0a6056 | 37 | #include <linux/ratelimit.h> |
ec16227e AG |
38 | |
39 | #include "rds.h" | |
40 | #include "ib.h" | |
41 | ||
1bde04a6 | 42 | static char *rds_ib_event_type_strings[] = { |
59f740a6 ZB |
43 | #define RDS_IB_EVENT_STRING(foo) \ |
44 | [IB_EVENT_##foo] = __stringify(IB_EVENT_##foo) | |
1bde04a6 ZB |
45 | RDS_IB_EVENT_STRING(CQ_ERR), |
46 | RDS_IB_EVENT_STRING(QP_FATAL), | |
47 | RDS_IB_EVENT_STRING(QP_REQ_ERR), | |
48 | RDS_IB_EVENT_STRING(QP_ACCESS_ERR), | |
49 | RDS_IB_EVENT_STRING(COMM_EST), | |
50 | RDS_IB_EVENT_STRING(SQ_DRAINED), | |
51 | RDS_IB_EVENT_STRING(PATH_MIG), | |
52 | RDS_IB_EVENT_STRING(PATH_MIG_ERR), | |
53 | RDS_IB_EVENT_STRING(DEVICE_FATAL), | |
54 | RDS_IB_EVENT_STRING(PORT_ACTIVE), | |
55 | RDS_IB_EVENT_STRING(PORT_ERR), | |
56 | RDS_IB_EVENT_STRING(LID_CHANGE), | |
57 | RDS_IB_EVENT_STRING(PKEY_CHANGE), | |
58 | RDS_IB_EVENT_STRING(SM_CHANGE), | |
59 | RDS_IB_EVENT_STRING(SRQ_ERR), | |
60 | RDS_IB_EVENT_STRING(SRQ_LIMIT_REACHED), | |
61 | RDS_IB_EVENT_STRING(QP_LAST_WQE_REACHED), | |
62 | RDS_IB_EVENT_STRING(CLIENT_REREGISTER), | |
63 | #undef RDS_IB_EVENT_STRING | |
64 | }; | |
65 | ||
66 | static char *rds_ib_event_str(enum ib_event_type type) | |
67 | { | |
59f740a6 ZB |
68 | return rds_str_array(rds_ib_event_type_strings, |
69 | ARRAY_SIZE(rds_ib_event_type_strings), type); | |
1bde04a6 ZB |
70 | }; |
71 | ||
ec16227e AG |
72 | /* |
73 | * Set the selected protocol version | |
74 | */ | |
75 | static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) | |
76 | { | |
77 | conn->c_version = version; | |
78 | } | |
79 | ||
80 | /* | |
81 | * Set up flow control | |
82 | */ | |
83 | static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) | |
84 | { | |
85 | struct rds_ib_connection *ic = conn->c_transport_data; | |
86 | ||
87 | if (rds_ib_sysctl_flow_control && credits != 0) { | |
88 | /* We're doing flow control */ | |
89 | ic->i_flowctl = 1; | |
90 | rds_ib_send_add_credits(conn, credits); | |
91 | } else { | |
92 | ic->i_flowctl = 0; | |
93 | } | |
94 | } | |
95 | ||
96 | /* | |
97 | * Tune RNR behavior. Without flow control, we use a rather | |
98 | * low timeout, but not the absolute minimum - this should | |
99 | * be tunable. | |
100 | * | |
101 | * We already set the RNR retry count to 7 (which is the | |
102 | * smallest infinite number :-) above. | |
103 | * If flow control is off, we want to change this back to 0 | |
104 | * so that we learn quickly when our credit accounting is | |
105 | * buggy. | |
106 | * | |
107 | * Caller passes in a qp_attr pointer - don't waste stack spacv | |
108 | * by allocation this twice. | |
109 | */ | |
110 | static void | |
111 | rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) | |
112 | { | |
113 | int ret; | |
114 | ||
115 | attr->min_rnr_timer = IB_RNR_TIMER_000_32; | |
116 | ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); | |
117 | if (ret) | |
118 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); | |
119 | } | |
120 | ||
121 | /* | |
122 | * Connection established. | |
123 | * We get here for both outgoing and incoming connection. | |
124 | */ | |
125 | void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) | |
126 | { | |
127 | const struct rds_ib_connect_private *dp = NULL; | |
128 | struct rds_ib_connection *ic = conn->c_transport_data; | |
ec16227e AG |
129 | struct ib_qp_attr qp_attr; |
130 | int err; | |
131 | ||
9ddbcfa0 | 132 | if (event->param.conn.private_data_len >= sizeof(*dp)) { |
ec16227e AG |
133 | dp = event->param.conn.private_data; |
134 | ||
02a6a259 AG |
135 | /* make sure it isn't empty data */ |
136 | if (dp->dp_protocol_major) { | |
137 | rds_ib_set_protocol(conn, | |
ec16227e | 138 | RDS_PROTOCOL(dp->dp_protocol_major, |
02a6a259 AG |
139 | dp->dp_protocol_minor)); |
140 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); | |
141 | } | |
ec16227e AG |
142 | } |
143 | ||
f147dd9e AG |
144 | if (conn->c_version < RDS_PROTOCOL(3,1)) { |
145 | printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed," | |
146 | " no longer supported\n", | |
147 | &conn->c_faddr, | |
148 | RDS_PROTOCOL_MAJOR(conn->c_version), | |
149 | RDS_PROTOCOL_MINOR(conn->c_version)); | |
150 | rds_conn_destroy(conn); | |
151 | return; | |
152 | } else { | |
153 | printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n", | |
154 | &conn->c_faddr, | |
155 | RDS_PROTOCOL_MAJOR(conn->c_version), | |
156 | RDS_PROTOCOL_MINOR(conn->c_version), | |
157 | ic->i_flowctl ? ", flow control" : ""); | |
158 | } | |
ec16227e | 159 | |
e11d912a AG |
160 | /* |
161 | * Init rings and fill recv. this needs to wait until protocol negotiation | |
162 | * is complete, since ring layout is different from 3.0 to 3.1. | |
163 | */ | |
164 | rds_ib_send_init_ring(ic); | |
165 | rds_ib_recv_init_ring(ic); | |
166 | /* Post receive buffers - as a side effect, this will update | |
167 | * the posted credit count. */ | |
f17a1a55 | 168 | rds_ib_recv_refill(conn, 1); |
e11d912a | 169 | |
ec16227e AG |
170 | /* Tune RNR behavior */ |
171 | rds_ib_tune_rnr(ic, &qp_attr); | |
172 | ||
173 | qp_attr.qp_state = IB_QPS_RTS; | |
174 | err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); | |
175 | if (err) | |
176 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); | |
177 | ||
3e0249f9 ZB |
178 | /* update ib_device with this local ipaddr */ |
179 | err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr); | |
ec16227e | 180 | if (err) |
3e0249f9 ZB |
181 | printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", |
182 | err); | |
ec16227e AG |
183 | |
184 | /* If the peer gave us the last packet it saw, process this as if | |
185 | * we had received a regular ACK. */ | |
186 | if (dp && dp->dp_ack_seq) | |
187 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); | |
188 | ||
189 | rds_connect_complete(conn); | |
190 | } | |
191 | ||
192 | static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, | |
193 | struct rdma_conn_param *conn_param, | |
194 | struct rds_ib_connect_private *dp, | |
40589e74 AG |
195 | u32 protocol_version, |
196 | u32 max_responder_resources, | |
197 | u32 max_initiator_depth) | |
ec16227e | 198 | { |
40589e74 | 199 | struct rds_ib_connection *ic = conn->c_transport_data; |
3e0249f9 | 200 | struct rds_ib_device *rds_ibdev = ic->rds_ibdev; |
40589e74 | 201 | |
ec16227e | 202 | memset(conn_param, 0, sizeof(struct rdma_conn_param)); |
40589e74 | 203 | |
40589e74 AG |
204 | conn_param->responder_resources = |
205 | min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources); | |
206 | conn_param->initiator_depth = | |
207 | min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth); | |
3ba23ade | 208 | conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); |
ec16227e AG |
209 | conn_param->rnr_retry_count = 7; |
210 | ||
211 | if (dp) { | |
ec16227e AG |
212 | memset(dp, 0, sizeof(*dp)); |
213 | dp->dp_saddr = conn->c_laddr; | |
214 | dp->dp_daddr = conn->c_faddr; | |
215 | dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version); | |
216 | dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version); | |
217 | dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); | |
218 | dp->dp_ack_seq = rds_ib_piggyb_ack(ic); | |
219 | ||
220 | /* Advertise flow control */ | |
221 | if (ic->i_flowctl) { | |
222 | unsigned int credits; | |
223 | ||
224 | credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)); | |
225 | dp->dp_credit = cpu_to_be32(credits); | |
226 | atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits); | |
227 | } | |
228 | ||
229 | conn_param->private_data = dp; | |
230 | conn_param->private_data_len = sizeof(*dp); | |
231 | } | |
232 | } | |
233 | ||
234 | static void rds_ib_cq_event_handler(struct ib_event *event, void *data) | |
235 | { | |
1bde04a6 ZB |
236 | rdsdebug("event %u (%s) data %p\n", |
237 | event->event, rds_ib_event_str(event->event), data); | |
ec16227e AG |
238 | } |
239 | ||
240 | static void rds_ib_qp_event_handler(struct ib_event *event, void *data) | |
241 | { | |
242 | struct rds_connection *conn = data; | |
243 | struct rds_ib_connection *ic = conn->c_transport_data; | |
244 | ||
1bde04a6 ZB |
245 | rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event, |
246 | rds_ib_event_str(event->event)); | |
ec16227e AG |
247 | |
248 | switch (event->event) { | |
249 | case IB_EVENT_COMM_EST: | |
250 | rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); | |
251 | break; | |
252 | default: | |
1bde04a6 | 253 | rdsdebug("Fatal QP Event %u (%s) " |
fdf6e6b4 | 254 | "- connection %pI4->%pI4, reconnecting\n", |
1bde04a6 ZB |
255 | event->event, rds_ib_event_str(event->event), |
256 | &conn->c_laddr, &conn->c_faddr); | |
97069788 | 257 | rds_conn_drop(conn); |
ec16227e AG |
258 | break; |
259 | } | |
260 | } | |
261 | ||
262 | /* | |
263 | * This needs to be very careful to not leave IS_ERR pointers around for | |
264 | * cleanup to trip over. | |
265 | */ | |
266 | static int rds_ib_setup_qp(struct rds_connection *conn) | |
267 | { | |
268 | struct rds_ib_connection *ic = conn->c_transport_data; | |
269 | struct ib_device *dev = ic->i_cm_id->device; | |
270 | struct ib_qp_init_attr attr; | |
271 | struct rds_ib_device *rds_ibdev; | |
272 | int ret; | |
273 | ||
3e0249f9 ZB |
274 | /* |
275 | * It's normal to see a null device if an incoming connection races | |
276 | * with device removal, so we don't print a warning. | |
ec16227e | 277 | */ |
3e0249f9 ZB |
278 | rds_ibdev = rds_ib_get_client_data(dev); |
279 | if (!rds_ibdev) | |
ec16227e | 280 | return -EOPNOTSUPP; |
3e0249f9 ZB |
281 | |
282 | /* add the conn now so that connection establishment has the dev */ | |
283 | rds_ib_add_conn(rds_ibdev, conn); | |
ec16227e AG |
284 | |
285 | if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) | |
286 | rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); | |
287 | if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) | |
288 | rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); | |
289 | ||
290 | /* Protection domain and memory range */ | |
291 | ic->i_pd = rds_ibdev->pd; | |
292 | ic->i_mr = rds_ibdev->mr; | |
293 | ||
294 | ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler, | |
295 | rds_ib_cq_event_handler, conn, | |
296 | ic->i_send_ring.w_nr + 1, 0); | |
297 | if (IS_ERR(ic->i_send_cq)) { | |
298 | ret = PTR_ERR(ic->i_send_cq); | |
299 | ic->i_send_cq = NULL; | |
300 | rdsdebug("ib_create_cq send failed: %d\n", ret); | |
301 | goto out; | |
302 | } | |
303 | ||
304 | ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler, | |
305 | rds_ib_cq_event_handler, conn, | |
306 | ic->i_recv_ring.w_nr, 0); | |
307 | if (IS_ERR(ic->i_recv_cq)) { | |
308 | ret = PTR_ERR(ic->i_recv_cq); | |
309 | ic->i_recv_cq = NULL; | |
310 | rdsdebug("ib_create_cq recv failed: %d\n", ret); | |
311 | goto out; | |
312 | } | |
313 | ||
314 | ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); | |
315 | if (ret) { | |
316 | rdsdebug("ib_req_notify_cq send failed: %d\n", ret); | |
317 | goto out; | |
318 | } | |
319 | ||
320 | ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); | |
321 | if (ret) { | |
322 | rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); | |
323 | goto out; | |
324 | } | |
325 | ||
326 | /* XXX negotiate max send/recv with remote? */ | |
327 | memset(&attr, 0, sizeof(attr)); | |
328 | attr.event_handler = rds_ib_qp_event_handler; | |
329 | attr.qp_context = conn; | |
330 | /* + 1 to allow for the single ack message */ | |
331 | attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1; | |
332 | attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; | |
333 | attr.cap.max_send_sge = rds_ibdev->max_sge; | |
334 | attr.cap.max_recv_sge = RDS_IB_RECV_SGE; | |
335 | attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
336 | attr.qp_type = IB_QPT_RC; | |
337 | attr.send_cq = ic->i_send_cq; | |
338 | attr.recv_cq = ic->i_recv_cq; | |
339 | ||
340 | /* | |
341 | * XXX this can fail if max_*_wr is too large? Are we supposed | |
342 | * to back off until we get a value that the hardware can support? | |
343 | */ | |
344 | ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); | |
345 | if (ret) { | |
346 | rdsdebug("rdma_create_qp failed: %d\n", ret); | |
347 | goto out; | |
348 | } | |
349 | ||
350 | ic->i_send_hdrs = ib_dma_alloc_coherent(dev, | |
351 | ic->i_send_ring.w_nr * | |
352 | sizeof(struct rds_header), | |
353 | &ic->i_send_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 354 | if (!ic->i_send_hdrs) { |
ec16227e AG |
355 | ret = -ENOMEM; |
356 | rdsdebug("ib_dma_alloc_coherent send failed\n"); | |
357 | goto out; | |
358 | } | |
359 | ||
360 | ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, | |
361 | ic->i_recv_ring.w_nr * | |
362 | sizeof(struct rds_header), | |
363 | &ic->i_recv_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 364 | if (!ic->i_recv_hdrs) { |
ec16227e AG |
365 | ret = -ENOMEM; |
366 | rdsdebug("ib_dma_alloc_coherent recv failed\n"); | |
367 | goto out; | |
368 | } | |
369 | ||
370 | ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), | |
371 | &ic->i_ack_dma, GFP_KERNEL); | |
8690bfa1 | 372 | if (!ic->i_ack) { |
ec16227e AG |
373 | ret = -ENOMEM; |
374 | rdsdebug("ib_dma_alloc_coherent ack failed\n"); | |
375 | goto out; | |
376 | } | |
377 | ||
3dbd4439 | 378 | ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work), |
e4c52c98 | 379 | ibdev_to_node(dev)); |
8690bfa1 | 380 | if (!ic->i_sends) { |
ec16227e AG |
381 | ret = -ENOMEM; |
382 | rdsdebug("send allocation failed\n"); | |
383 | goto out; | |
384 | } | |
ec16227e | 385 | |
3dbd4439 | 386 | ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work), |
e4c52c98 | 387 | ibdev_to_node(dev)); |
8690bfa1 | 388 | if (!ic->i_recvs) { |
ec16227e AG |
389 | ret = -ENOMEM; |
390 | rdsdebug("recv allocation failed\n"); | |
391 | goto out; | |
392 | } | |
393 | ||
ec16227e AG |
394 | rds_ib_recv_init_ack(ic); |
395 | ||
ec16227e AG |
396 | rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr, |
397 | ic->i_send_cq, ic->i_recv_cq); | |
398 | ||
399 | out: | |
3e0249f9 | 400 | rds_ib_dev_put(rds_ibdev); |
ec16227e AG |
401 | return ret; |
402 | } | |
403 | ||
9ddbcfa0 | 404 | static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event) |
ec16227e | 405 | { |
9ddbcfa0 | 406 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; |
ec16227e AG |
407 | u16 common; |
408 | u32 version = 0; | |
409 | ||
9ddbcfa0 AG |
410 | /* |
411 | * rdma_cm private data is odd - when there is any private data in the | |
ec16227e AG |
412 | * request, we will be given a pretty large buffer without telling us the |
413 | * original size. The only way to tell the difference is by looking at | |
414 | * the contents, which are initialized to zero. | |
415 | * If the protocol version fields aren't set, this is a connection attempt | |
416 | * from an older version. This could could be 3.0 or 2.0 - we can't tell. | |
9ddbcfa0 AG |
417 | * We really should have changed this for OFED 1.3 :-( |
418 | */ | |
419 | ||
420 | /* Be paranoid. RDS always has privdata */ | |
421 | if (!event->param.conn.private_data_len) { | |
422 | printk(KERN_NOTICE "RDS incoming connection has no private data, " | |
423 | "rejecting\n"); | |
424 | return 0; | |
425 | } | |
426 | ||
427 | /* Even if len is crap *now* I still want to check it. -ASG */ | |
f64f9e71 JP |
428 | if (event->param.conn.private_data_len < sizeof (*dp) || |
429 | dp->dp_protocol_major == 0) | |
ec16227e AG |
430 | return RDS_PROTOCOL_3_0; |
431 | ||
432 | common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS; | |
433 | if (dp->dp_protocol_major == 3 && common) { | |
434 | version = RDS_PROTOCOL_3_0; | |
435 | while ((common >>= 1) != 0) | |
436 | version++; | |
a4967598 MM |
437 | } else |
438 | printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n", | |
439 | &dp->dp_saddr, | |
440 | dp->dp_protocol_major, | |
441 | dp->dp_protocol_minor); | |
ec16227e AG |
442 | return version; |
443 | } | |
444 | ||
445 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, | |
446 | struct rdma_cm_event *event) | |
447 | { | |
448 | __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; | |
449 | __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; | |
450 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; | |
451 | struct rds_ib_connect_private dp_rep; | |
452 | struct rds_connection *conn = NULL; | |
453 | struct rds_ib_connection *ic = NULL; | |
454 | struct rdma_conn_param conn_param; | |
455 | u32 version; | |
a46ca94e | 456 | int err = 1, destroy = 1; |
ec16227e AG |
457 | |
458 | /* Check whether the remote protocol version matches ours. */ | |
9ddbcfa0 | 459 | version = rds_ib_protocol_compatible(event); |
ec16227e AG |
460 | if (!version) |
461 | goto out; | |
462 | ||
463 | rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid " | |
464 | "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr, | |
465 | RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), | |
466 | (unsigned long long)be64_to_cpu(lguid), | |
467 | (unsigned long long)be64_to_cpu(fguid)); | |
468 | ||
469 | conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport, | |
470 | GFP_KERNEL); | |
471 | if (IS_ERR(conn)) { | |
472 | rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); | |
473 | conn = NULL; | |
474 | goto out; | |
475 | } | |
476 | ||
477 | /* | |
478 | * The connection request may occur while the | |
479 | * previous connection exist, e.g. in case of failover. | |
480 | * But as connections may be initiated simultaneously | |
481 | * by both hosts, we have a random backoff mechanism - | |
482 | * see the comment above rds_queue_reconnect() | |
483 | */ | |
484 | mutex_lock(&conn->c_cm_lock); | |
485 | if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { | |
486 | if (rds_conn_state(conn) == RDS_CONN_UP) { | |
487 | rdsdebug("incoming connect while connecting\n"); | |
488 | rds_conn_drop(conn); | |
489 | rds_ib_stats_inc(s_ib_listen_closed_stale); | |
490 | } else | |
491 | if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { | |
492 | /* Wait and see - our connect may still be succeeding */ | |
493 | rds_ib_stats_inc(s_ib_connect_raced); | |
494 | } | |
ec16227e AG |
495 | goto out; |
496 | } | |
497 | ||
498 | ic = conn->c_transport_data; | |
499 | ||
500 | rds_ib_set_protocol(conn, version); | |
501 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); | |
502 | ||
503 | /* If the peer gave us the last packet it saw, process this as if | |
504 | * we had received a regular ACK. */ | |
505 | if (dp->dp_ack_seq) | |
506 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); | |
507 | ||
508 | BUG_ON(cm_id->context); | |
509 | BUG_ON(ic->i_cm_id); | |
510 | ||
511 | ic->i_cm_id = cm_id; | |
512 | cm_id->context = conn; | |
513 | ||
514 | /* We got halfway through setting up the ib_connection, if we | |
515 | * fail now, we have to take the long route out of this mess. */ | |
516 | destroy = 0; | |
517 | ||
518 | err = rds_ib_setup_qp(conn); | |
519 | if (err) { | |
520 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); | |
521 | goto out; | |
522 | } | |
523 | ||
40589e74 AG |
524 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version, |
525 | event->param.conn.responder_resources, | |
526 | event->param.conn.initiator_depth); | |
ec16227e AG |
527 | |
528 | /* rdma_accept() calls rdma_reject() internally if it fails */ | |
529 | err = rdma_accept(cm_id, &conn_param); | |
a46ca94e | 530 | if (err) |
ec16227e | 531 | rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); |
ec16227e AG |
532 | |
533 | out: | |
a46ca94e ZB |
534 | if (conn) |
535 | mutex_unlock(&conn->c_cm_lock); | |
536 | if (err) | |
537 | rdma_reject(cm_id, NULL, 0); | |
ec16227e AG |
538 | return destroy; |
539 | } | |
540 | ||
541 | ||
542 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id) | |
543 | { | |
544 | struct rds_connection *conn = cm_id->context; | |
545 | struct rds_ib_connection *ic = conn->c_transport_data; | |
546 | struct rdma_conn_param conn_param; | |
547 | struct rds_ib_connect_private dp; | |
548 | int ret; | |
549 | ||
550 | /* If the peer doesn't do protocol negotiation, we must | |
551 | * default to RDSv3.0 */ | |
552 | rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); | |
553 | ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ | |
554 | ||
555 | ret = rds_ib_setup_qp(conn); | |
556 | if (ret) { | |
557 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); | |
558 | goto out; | |
559 | } | |
560 | ||
40589e74 AG |
561 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION, |
562 | UINT_MAX, UINT_MAX); | |
ec16227e AG |
563 | ret = rdma_connect(cm_id, &conn_param); |
564 | if (ret) | |
565 | rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); | |
566 | ||
567 | out: | |
568 | /* Beware - returning non-zero tells the rdma_cm to destroy | |
569 | * the cm_id. We should certainly not do it as long as we still | |
570 | * "own" the cm_id. */ | |
571 | if (ret) { | |
572 | if (ic->i_cm_id == cm_id) | |
573 | ret = 0; | |
574 | } | |
575 | return ret; | |
576 | } | |
577 | ||
578 | int rds_ib_conn_connect(struct rds_connection *conn) | |
579 | { | |
580 | struct rds_ib_connection *ic = conn->c_transport_data; | |
581 | struct sockaddr_in src, dest; | |
582 | int ret; | |
583 | ||
584 | /* XXX I wonder what affect the port space has */ | |
585 | /* delegate cm event handler to rdma_transport */ | |
586 | ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn, | |
b26f9b99 | 587 | RDMA_PS_TCP, IB_QPT_RC); |
ec16227e AG |
588 | if (IS_ERR(ic->i_cm_id)) { |
589 | ret = PTR_ERR(ic->i_cm_id); | |
590 | ic->i_cm_id = NULL; | |
591 | rdsdebug("rdma_create_id() failed: %d\n", ret); | |
592 | goto out; | |
593 | } | |
594 | ||
595 | rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); | |
596 | ||
597 | src.sin_family = AF_INET; | |
598 | src.sin_addr.s_addr = (__force u32)conn->c_laddr; | |
599 | src.sin_port = (__force u16)htons(0); | |
600 | ||
601 | dest.sin_family = AF_INET; | |
602 | dest.sin_addr.s_addr = (__force u32)conn->c_faddr; | |
603 | dest.sin_port = (__force u16)htons(RDS_PORT); | |
604 | ||
605 | ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, | |
606 | (struct sockaddr *)&dest, | |
607 | RDS_RDMA_RESOLVE_TIMEOUT_MS); | |
608 | if (ret) { | |
609 | rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, | |
610 | ret); | |
611 | rdma_destroy_id(ic->i_cm_id); | |
612 | ic->i_cm_id = NULL; | |
613 | } | |
614 | ||
615 | out: | |
616 | return ret; | |
617 | } | |
618 | ||
619 | /* | |
620 | * This is so careful about only cleaning up resources that were built up | |
621 | * so that it can be called at any point during startup. In fact it | |
622 | * can be called multiple times for a given connection. | |
623 | */ | |
624 | void rds_ib_conn_shutdown(struct rds_connection *conn) | |
625 | { | |
626 | struct rds_ib_connection *ic = conn->c_transport_data; | |
627 | int err = 0; | |
628 | ||
629 | rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, | |
630 | ic->i_pd, ic->i_send_cq, ic->i_recv_cq, | |
631 | ic->i_cm_id ? ic->i_cm_id->qp : NULL); | |
632 | ||
633 | if (ic->i_cm_id) { | |
634 | struct ib_device *dev = ic->i_cm_id->device; | |
635 | ||
636 | rdsdebug("disconnecting cm %p\n", ic->i_cm_id); | |
637 | err = rdma_disconnect(ic->i_cm_id); | |
638 | if (err) { | |
639 | /* Actually this may happen quite frequently, when | |
640 | * an outgoing connect raced with an incoming connect. | |
641 | */ | |
642 | rdsdebug("failed to disconnect, cm: %p err %d\n", | |
643 | ic->i_cm_id, err); | |
644 | } | |
645 | ||
e32b4a70 | 646 | /* |
f046011c ZB |
647 | * We want to wait for tx and rx completion to finish |
648 | * before we tear down the connection, but we have to be | |
649 | * careful not to get stuck waiting on a send ring that | |
650 | * only has unsignaled sends in it. We've shutdown new | |
651 | * sends before getting here so by waiting for signaled | |
652 | * sends to complete we're ensured that there will be no | |
653 | * more tx processing. | |
e32b4a70 | 654 | */ |
ec16227e | 655 | wait_event(rds_ib_ring_empty_wait, |
f046011c ZB |
656 | rds_ib_ring_empty(&ic->i_recv_ring) && |
657 | (atomic_read(&ic->i_signaled_sends) == 0)); | |
658 | tasklet_kill(&ic->i_recv_tasklet); | |
ec16227e AG |
659 | |
660 | if (ic->i_send_hdrs) | |
661 | ib_dma_free_coherent(dev, | |
662 | ic->i_send_ring.w_nr * | |
663 | sizeof(struct rds_header), | |
664 | ic->i_send_hdrs, | |
665 | ic->i_send_hdrs_dma); | |
666 | ||
667 | if (ic->i_recv_hdrs) | |
668 | ib_dma_free_coherent(dev, | |
669 | ic->i_recv_ring.w_nr * | |
670 | sizeof(struct rds_header), | |
671 | ic->i_recv_hdrs, | |
672 | ic->i_recv_hdrs_dma); | |
673 | ||
674 | if (ic->i_ack) | |
675 | ib_dma_free_coherent(dev, sizeof(struct rds_header), | |
676 | ic->i_ack, ic->i_ack_dma); | |
677 | ||
678 | if (ic->i_sends) | |
679 | rds_ib_send_clear_ring(ic); | |
680 | if (ic->i_recvs) | |
681 | rds_ib_recv_clear_ring(ic); | |
682 | ||
683 | if (ic->i_cm_id->qp) | |
684 | rdma_destroy_qp(ic->i_cm_id); | |
685 | if (ic->i_send_cq) | |
686 | ib_destroy_cq(ic->i_send_cq); | |
687 | if (ic->i_recv_cq) | |
688 | ib_destroy_cq(ic->i_recv_cq); | |
689 | rdma_destroy_id(ic->i_cm_id); | |
690 | ||
691 | /* | |
692 | * Move connection back to the nodev list. | |
693 | */ | |
745cbcca AG |
694 | if (ic->rds_ibdev) |
695 | rds_ib_remove_conn(ic->rds_ibdev, conn); | |
ec16227e AG |
696 | |
697 | ic->i_cm_id = NULL; | |
698 | ic->i_pd = NULL; | |
699 | ic->i_mr = NULL; | |
700 | ic->i_send_cq = NULL; | |
701 | ic->i_recv_cq = NULL; | |
702 | ic->i_send_hdrs = NULL; | |
703 | ic->i_recv_hdrs = NULL; | |
704 | ic->i_ack = NULL; | |
705 | } | |
706 | BUG_ON(ic->rds_ibdev); | |
707 | ||
708 | /* Clear pending transmit */ | |
ff3d7d36 AG |
709 | if (ic->i_data_op) { |
710 | struct rds_message *rm; | |
711 | ||
712 | rm = container_of(ic->i_data_op, struct rds_message, data); | |
713 | rds_message_put(rm); | |
714 | ic->i_data_op = NULL; | |
ec16227e AG |
715 | } |
716 | ||
717 | /* Clear the ACK state */ | |
718 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
8cbd9606 AG |
719 | #ifdef KERNEL_HAS_ATOMIC64 |
720 | atomic64_set(&ic->i_ack_next, 0); | |
721 | #else | |
722 | ic->i_ack_next = 0; | |
723 | #endif | |
ec16227e AG |
724 | ic->i_ack_recv = 0; |
725 | ||
726 | /* Clear flow control state */ | |
727 | ic->i_flowctl = 0; | |
728 | atomic_set(&ic->i_credits, 0); | |
729 | ||
730 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
731 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
732 | ||
733 | if (ic->i_ibinc) { | |
734 | rds_inc_put(&ic->i_ibinc->ii_inc); | |
735 | ic->i_ibinc = NULL; | |
736 | } | |
737 | ||
738 | vfree(ic->i_sends); | |
739 | ic->i_sends = NULL; | |
740 | vfree(ic->i_recvs); | |
741 | ic->i_recvs = NULL; | |
742 | } | |
743 | ||
744 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) | |
745 | { | |
746 | struct rds_ib_connection *ic; | |
747 | unsigned long flags; | |
33244125 | 748 | int ret; |
ec16227e AG |
749 | |
750 | /* XXX too lazy? */ | |
f0229eaa | 751 | ic = kzalloc(sizeof(struct rds_ib_connection), gfp); |
8690bfa1 | 752 | if (!ic) |
ec16227e AG |
753 | return -ENOMEM; |
754 | ||
33244125 CM |
755 | ret = rds_ib_recv_alloc_caches(ic); |
756 | if (ret) { | |
757 | kfree(ic); | |
758 | return ret; | |
759 | } | |
760 | ||
ec16227e | 761 | INIT_LIST_HEAD(&ic->ib_node); |
d521b63b AG |
762 | tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn, |
763 | (unsigned long) ic); | |
ec16227e | 764 | mutex_init(&ic->i_recv_mutex); |
8cbd9606 AG |
765 | #ifndef KERNEL_HAS_ATOMIC64 |
766 | spin_lock_init(&ic->i_ack_lock); | |
767 | #endif | |
f046011c | 768 | atomic_set(&ic->i_signaled_sends, 0); |
ec16227e AG |
769 | |
770 | /* | |
771 | * rds_ib_conn_shutdown() waits for these to be emptied so they | |
772 | * must be initialized before it can be called. | |
773 | */ | |
774 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
775 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
776 | ||
777 | ic->conn = conn; | |
778 | conn->c_transport_data = ic; | |
779 | ||
780 | spin_lock_irqsave(&ib_nodev_conns_lock, flags); | |
781 | list_add_tail(&ic->ib_node, &ib_nodev_conns); | |
782 | spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); | |
783 | ||
784 | ||
785 | rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); | |
786 | return 0; | |
787 | } | |
788 | ||
745cbcca AG |
789 | /* |
790 | * Free a connection. Connection must be shut down and not set for reconnect. | |
791 | */ | |
ec16227e AG |
792 | void rds_ib_conn_free(void *arg) |
793 | { | |
794 | struct rds_ib_connection *ic = arg; | |
745cbcca AG |
795 | spinlock_t *lock_ptr; |
796 | ||
ec16227e | 797 | rdsdebug("ic %p\n", ic); |
745cbcca AG |
798 | |
799 | /* | |
800 | * Conn is either on a dev's list or on the nodev list. | |
801 | * A race with shutdown() or connect() would cause problems | |
802 | * (since rds_ibdev would change) but that should never happen. | |
803 | */ | |
804 | lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; | |
805 | ||
806 | spin_lock_irq(lock_ptr); | |
ec16227e | 807 | list_del(&ic->ib_node); |
745cbcca AG |
808 | spin_unlock_irq(lock_ptr); |
809 | ||
33244125 CM |
810 | rds_ib_recv_free_caches(ic); |
811 | ||
ec16227e AG |
812 | kfree(ic); |
813 | } | |
814 | ||
815 | ||
816 | /* | |
817 | * An error occurred on the connection | |
818 | */ | |
819 | void | |
820 | __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) | |
821 | { | |
822 | va_list ap; | |
823 | ||
824 | rds_conn_drop(conn); | |
825 | ||
826 | va_start(ap, fmt); | |
827 | vprintk(fmt, ap); | |
828 | va_end(ap); | |
829 | } |