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 | ||
42 | /* | |
43 | * Set the selected protocol version | |
44 | */ | |
45 | static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) | |
46 | { | |
47 | conn->c_version = version; | |
48 | } | |
49 | ||
50 | /* | |
51 | * Set up flow control | |
52 | */ | |
53 | static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) | |
54 | { | |
55 | struct rds_ib_connection *ic = conn->c_transport_data; | |
56 | ||
57 | if (rds_ib_sysctl_flow_control && credits != 0) { | |
58 | /* We're doing flow control */ | |
59 | ic->i_flowctl = 1; | |
60 | rds_ib_send_add_credits(conn, credits); | |
61 | } else { | |
62 | ic->i_flowctl = 0; | |
63 | } | |
64 | } | |
65 | ||
66 | /* | |
67 | * Tune RNR behavior. Without flow control, we use a rather | |
68 | * low timeout, but not the absolute minimum - this should | |
69 | * be tunable. | |
70 | * | |
71 | * We already set the RNR retry count to 7 (which is the | |
72 | * smallest infinite number :-) above. | |
73 | * If flow control is off, we want to change this back to 0 | |
74 | * so that we learn quickly when our credit accounting is | |
75 | * buggy. | |
76 | * | |
77 | * Caller passes in a qp_attr pointer - don't waste stack spacv | |
78 | * by allocation this twice. | |
79 | */ | |
80 | static void | |
81 | rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) | |
82 | { | |
83 | int ret; | |
84 | ||
85 | attr->min_rnr_timer = IB_RNR_TIMER_000_32; | |
86 | ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); | |
87 | if (ret) | |
88 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); | |
89 | } | |
90 | ||
91 | /* | |
92 | * Connection established. | |
93 | * We get here for both outgoing and incoming connection. | |
94 | */ | |
95 | void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) | |
96 | { | |
97 | const struct rds_ib_connect_private *dp = NULL; | |
98 | struct rds_ib_connection *ic = conn->c_transport_data; | |
ec16227e AG |
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. */ | |
73ce4317 | 138 | rds_ib_recv_refill(conn, 1, GFP_KERNEL); |
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 | ||
3e0249f9 ZB |
148 | /* update ib_device with this local ipaddr */ |
149 | err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr); | |
ec16227e | 150 | if (err) |
3e0249f9 ZB |
151 | printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", |
152 | err); | |
ec16227e AG |
153 | |
154 | /* If the peer gave us the last packet it saw, process this as if | |
155 | * we had received a regular ACK. */ | |
c0adf54a | 156 | if (dp) { |
157 | /* dp structure start is not guaranteed to be 8 bytes aligned. | |
158 | * Since dp_ack_seq is 64-bit extended load operations can be | |
159 | * used so go through get_unaligned to avoid unaligned errors. | |
160 | */ | |
e2783717 | 161 | __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq); |
c0adf54a | 162 | |
163 | if (dp_ack_seq) | |
164 | rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq), | |
165 | NULL); | |
166 | } | |
ec16227e AG |
167 | |
168 | rds_connect_complete(conn); | |
169 | } | |
170 | ||
171 | static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, | |
172 | struct rdma_conn_param *conn_param, | |
173 | struct rds_ib_connect_private *dp, | |
40589e74 AG |
174 | u32 protocol_version, |
175 | u32 max_responder_resources, | |
176 | u32 max_initiator_depth) | |
ec16227e | 177 | { |
40589e74 | 178 | struct rds_ib_connection *ic = conn->c_transport_data; |
3e0249f9 | 179 | struct rds_ib_device *rds_ibdev = ic->rds_ibdev; |
40589e74 | 180 | |
ec16227e | 181 | memset(conn_param, 0, sizeof(struct rdma_conn_param)); |
40589e74 | 182 | |
40589e74 AG |
183 | conn_param->responder_resources = |
184 | min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources); | |
185 | conn_param->initiator_depth = | |
186 | min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth); | |
3ba23ade | 187 | conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); |
ec16227e AG |
188 | conn_param->rnr_retry_count = 7; |
189 | ||
190 | if (dp) { | |
ec16227e AG |
191 | memset(dp, 0, sizeof(*dp)); |
192 | dp->dp_saddr = conn->c_laddr; | |
193 | dp->dp_daddr = conn->c_faddr; | |
194 | dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version); | |
195 | dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version); | |
196 | dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); | |
197 | dp->dp_ack_seq = rds_ib_piggyb_ack(ic); | |
198 | ||
199 | /* Advertise flow control */ | |
200 | if (ic->i_flowctl) { | |
201 | unsigned int credits; | |
202 | ||
203 | credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)); | |
204 | dp->dp_credit = cpu_to_be32(credits); | |
205 | atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits); | |
206 | } | |
207 | ||
208 | conn_param->private_data = dp; | |
209 | conn_param->private_data_len = sizeof(*dp); | |
210 | } | |
211 | } | |
212 | ||
213 | static void rds_ib_cq_event_handler(struct ib_event *event, void *data) | |
214 | { | |
1bde04a6 | 215 | rdsdebug("event %u (%s) data %p\n", |
3c88f3dc | 216 | event->event, ib_event_msg(event->event), data); |
ec16227e AG |
217 | } |
218 | ||
f4f943c9 SS |
219 | /* Plucking the oldest entry from the ring can be done concurrently with |
220 | * the thread refilling the ring. Each ring operation is protected by | |
221 | * spinlocks and the transient state of refilling doesn't change the | |
222 | * recording of which entry is oldest. | |
223 | * | |
224 | * This relies on IB only calling one cq comp_handler for each cq so that | |
225 | * there will only be one caller of rds_recv_incoming() per RDS connection. | |
226 | */ | |
227 | static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context) | |
228 | { | |
229 | struct rds_connection *conn = context; | |
230 | struct rds_ib_connection *ic = conn->c_transport_data; | |
231 | ||
232 | rdsdebug("conn %p cq %p\n", conn, cq); | |
233 | ||
234 | rds_ib_stats_inc(s_ib_evt_handler_call); | |
235 | ||
236 | tasklet_schedule(&ic->i_recv_tasklet); | |
237 | } | |
238 | ||
239 | static void poll_cq(struct rds_ib_connection *ic, struct ib_cq *cq, | |
240 | struct ib_wc *wcs, | |
241 | struct rds_ib_ack_state *ack_state) | |
242 | { | |
243 | int nr; | |
244 | int i; | |
245 | struct ib_wc *wc; | |
246 | ||
247 | while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { | |
248 | for (i = 0; i < nr; i++) { | |
249 | wc = wcs + i; | |
250 | rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", | |
251 | (unsigned long long)wc->wr_id, wc->status, | |
252 | wc->byte_len, be32_to_cpu(wc->ex.imm_data)); | |
0c28c045 SS |
253 | |
254 | if (wc->wr_id & RDS_IB_SEND_OP) | |
255 | rds_ib_send_cqe_handler(ic, wc); | |
256 | else | |
257 | rds_ib_recv_cqe_handler(ic, wc, ack_state); | |
f4f943c9 SS |
258 | } |
259 | } | |
260 | } | |
261 | ||
0c28c045 SS |
262 | static void rds_ib_tasklet_fn_send(unsigned long data) |
263 | { | |
264 | struct rds_ib_connection *ic = (struct rds_ib_connection *)data; | |
265 | struct rds_connection *conn = ic->conn; | |
266 | struct rds_ib_ack_state state; | |
267 | ||
268 | rds_ib_stats_inc(s_ib_tasklet_call); | |
269 | ||
270 | memset(&state, 0, sizeof(state)); | |
271 | poll_cq(ic, ic->i_send_cq, ic->i_send_wc, &state); | |
272 | ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); | |
273 | poll_cq(ic, ic->i_send_cq, ic->i_send_wc, &state); | |
274 | ||
275 | if (rds_conn_up(conn) && | |
276 | (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) || | |
277 | test_bit(0, &conn->c_map_queued))) | |
278 | rds_send_xmit(ic->conn); | |
279 | } | |
280 | ||
f4f943c9 SS |
281 | static void rds_ib_tasklet_fn_recv(unsigned long data) |
282 | { | |
283 | struct rds_ib_connection *ic = (struct rds_ib_connection *)data; | |
284 | struct rds_connection *conn = ic->conn; | |
285 | struct rds_ib_device *rds_ibdev = ic->rds_ibdev; | |
286 | struct rds_ib_ack_state state; | |
287 | ||
9441c973 SS |
288 | if (!rds_ibdev) |
289 | rds_conn_drop(conn); | |
f4f943c9 SS |
290 | |
291 | rds_ib_stats_inc(s_ib_tasklet_call); | |
292 | ||
293 | memset(&state, 0, sizeof(state)); | |
294 | poll_cq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); | |
295 | ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); | |
296 | poll_cq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); | |
297 | ||
298 | if (state.ack_next_valid) | |
299 | rds_ib_set_ack(ic, state.ack_next, state.ack_required); | |
300 | if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { | |
301 | rds_send_drop_acked(conn, state.ack_recv, NULL); | |
302 | ic->i_ack_recv = state.ack_recv; | |
303 | } | |
304 | ||
305 | if (rds_conn_up(conn)) | |
306 | rds_ib_attempt_ack(ic); | |
307 | } | |
308 | ||
ec16227e AG |
309 | static void rds_ib_qp_event_handler(struct ib_event *event, void *data) |
310 | { | |
311 | struct rds_connection *conn = data; | |
312 | struct rds_ib_connection *ic = conn->c_transport_data; | |
313 | ||
1bde04a6 | 314 | rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event, |
3c88f3dc | 315 | ib_event_msg(event->event)); |
ec16227e AG |
316 | |
317 | switch (event->event) { | |
318 | case IB_EVENT_COMM_EST: | |
319 | rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); | |
320 | break; | |
321 | default: | |
1bde04a6 | 322 | rdsdebug("Fatal QP Event %u (%s) " |
fdf6e6b4 | 323 | "- connection %pI4->%pI4, reconnecting\n", |
3c88f3dc | 324 | event->event, ib_event_msg(event->event), |
1bde04a6 | 325 | &conn->c_laddr, &conn->c_faddr); |
97069788 | 326 | rds_conn_drop(conn); |
ec16227e AG |
327 | break; |
328 | } | |
329 | } | |
330 | ||
0c28c045 SS |
331 | static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context) |
332 | { | |
333 | struct rds_connection *conn = context; | |
334 | struct rds_ib_connection *ic = conn->c_transport_data; | |
335 | ||
336 | rdsdebug("conn %p cq %p\n", conn, cq); | |
337 | ||
338 | rds_ib_stats_inc(s_ib_evt_handler_call); | |
339 | ||
340 | tasklet_schedule(&ic->i_send_tasklet); | |
341 | } | |
342 | ||
ec16227e AG |
343 | /* |
344 | * This needs to be very careful to not leave IS_ERR pointers around for | |
345 | * cleanup to trip over. | |
346 | */ | |
347 | static int rds_ib_setup_qp(struct rds_connection *conn) | |
348 | { | |
349 | struct rds_ib_connection *ic = conn->c_transport_data; | |
350 | struct ib_device *dev = ic->i_cm_id->device; | |
351 | struct ib_qp_init_attr attr; | |
8e37210b | 352 | struct ib_cq_init_attr cq_attr = {}; |
ec16227e AG |
353 | struct rds_ib_device *rds_ibdev; |
354 | int ret; | |
355 | ||
3e0249f9 ZB |
356 | /* |
357 | * It's normal to see a null device if an incoming connection races | |
358 | * with device removal, so we don't print a warning. | |
ec16227e | 359 | */ |
3e0249f9 ZB |
360 | rds_ibdev = rds_ib_get_client_data(dev); |
361 | if (!rds_ibdev) | |
ec16227e | 362 | return -EOPNOTSUPP; |
3e0249f9 ZB |
363 | |
364 | /* add the conn now so that connection establishment has the dev */ | |
365 | rds_ib_add_conn(rds_ibdev, conn); | |
ec16227e AG |
366 | |
367 | if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) | |
368 | rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); | |
369 | if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) | |
370 | rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); | |
371 | ||
372 | /* Protection domain and memory range */ | |
373 | ic->i_pd = rds_ibdev->pd; | |
ec16227e | 374 | |
8e37210b | 375 | cq_attr.cqe = ic->i_send_ring.w_nr + 1; |
0c28c045 SS |
376 | |
377 | ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send, | |
ec16227e | 378 | rds_ib_cq_event_handler, conn, |
8e37210b | 379 | &cq_attr); |
ec16227e AG |
380 | if (IS_ERR(ic->i_send_cq)) { |
381 | ret = PTR_ERR(ic->i_send_cq); | |
382 | ic->i_send_cq = NULL; | |
383 | rdsdebug("ib_create_cq send failed: %d\n", ret); | |
384 | goto out; | |
385 | } | |
386 | ||
8e37210b | 387 | cq_attr.cqe = ic->i_recv_ring.w_nr; |
f4f943c9 | 388 | ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv, |
ec16227e | 389 | rds_ib_cq_event_handler, conn, |
8e37210b | 390 | &cq_attr); |
ec16227e AG |
391 | if (IS_ERR(ic->i_recv_cq)) { |
392 | ret = PTR_ERR(ic->i_recv_cq); | |
393 | ic->i_recv_cq = NULL; | |
394 | rdsdebug("ib_create_cq recv failed: %d\n", ret); | |
395 | goto out; | |
396 | } | |
397 | ||
398 | ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); | |
399 | if (ret) { | |
400 | rdsdebug("ib_req_notify_cq send failed: %d\n", ret); | |
401 | goto out; | |
402 | } | |
403 | ||
404 | ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); | |
405 | if (ret) { | |
406 | rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); | |
407 | goto out; | |
408 | } | |
409 | ||
410 | /* XXX negotiate max send/recv with remote? */ | |
411 | memset(&attr, 0, sizeof(attr)); | |
412 | attr.event_handler = rds_ib_qp_event_handler; | |
413 | attr.qp_context = conn; | |
414 | /* + 1 to allow for the single ack message */ | |
415 | attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1; | |
416 | attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; | |
417 | attr.cap.max_send_sge = rds_ibdev->max_sge; | |
418 | attr.cap.max_recv_sge = RDS_IB_RECV_SGE; | |
419 | attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
420 | attr.qp_type = IB_QPT_RC; | |
421 | attr.send_cq = ic->i_send_cq; | |
422 | attr.recv_cq = ic->i_recv_cq; | |
423 | ||
424 | /* | |
425 | * XXX this can fail if max_*_wr is too large? Are we supposed | |
426 | * to back off until we get a value that the hardware can support? | |
427 | */ | |
428 | ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); | |
429 | if (ret) { | |
430 | rdsdebug("rdma_create_qp failed: %d\n", ret); | |
431 | goto out; | |
432 | } | |
433 | ||
434 | ic->i_send_hdrs = ib_dma_alloc_coherent(dev, | |
435 | ic->i_send_ring.w_nr * | |
436 | sizeof(struct rds_header), | |
437 | &ic->i_send_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 438 | if (!ic->i_send_hdrs) { |
ec16227e AG |
439 | ret = -ENOMEM; |
440 | rdsdebug("ib_dma_alloc_coherent send failed\n"); | |
441 | goto out; | |
442 | } | |
443 | ||
444 | ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, | |
445 | ic->i_recv_ring.w_nr * | |
446 | sizeof(struct rds_header), | |
447 | &ic->i_recv_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 448 | if (!ic->i_recv_hdrs) { |
ec16227e AG |
449 | ret = -ENOMEM; |
450 | rdsdebug("ib_dma_alloc_coherent recv failed\n"); | |
451 | goto out; | |
452 | } | |
453 | ||
454 | ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), | |
455 | &ic->i_ack_dma, GFP_KERNEL); | |
8690bfa1 | 456 | if (!ic->i_ack) { |
ec16227e AG |
457 | ret = -ENOMEM; |
458 | rdsdebug("ib_dma_alloc_coherent ack failed\n"); | |
459 | goto out; | |
460 | } | |
461 | ||
3dbd4439 | 462 | ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work), |
e4c52c98 | 463 | ibdev_to_node(dev)); |
8690bfa1 | 464 | if (!ic->i_sends) { |
ec16227e AG |
465 | ret = -ENOMEM; |
466 | rdsdebug("send allocation failed\n"); | |
467 | goto out; | |
468 | } | |
ec16227e | 469 | |
3dbd4439 | 470 | ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work), |
e4c52c98 | 471 | ibdev_to_node(dev)); |
8690bfa1 | 472 | if (!ic->i_recvs) { |
ec16227e AG |
473 | ret = -ENOMEM; |
474 | rdsdebug("recv allocation failed\n"); | |
475 | goto out; | |
476 | } | |
477 | ||
ec16227e AG |
478 | rds_ib_recv_init_ack(ic); |
479 | ||
e5580242 | 480 | rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd, |
ec16227e AG |
481 | ic->i_send_cq, ic->i_recv_cq); |
482 | ||
483 | out: | |
3e0249f9 | 484 | rds_ib_dev_put(rds_ibdev); |
ec16227e AG |
485 | return ret; |
486 | } | |
487 | ||
9ddbcfa0 | 488 | static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event) |
ec16227e | 489 | { |
9ddbcfa0 | 490 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; |
ec16227e AG |
491 | u16 common; |
492 | u32 version = 0; | |
493 | ||
9ddbcfa0 AG |
494 | /* |
495 | * rdma_cm private data is odd - when there is any private data in the | |
ec16227e AG |
496 | * request, we will be given a pretty large buffer without telling us the |
497 | * original size. The only way to tell the difference is by looking at | |
498 | * the contents, which are initialized to zero. | |
499 | * If the protocol version fields aren't set, this is a connection attempt | |
500 | * from an older version. This could could be 3.0 or 2.0 - we can't tell. | |
9ddbcfa0 AG |
501 | * We really should have changed this for OFED 1.3 :-( |
502 | */ | |
503 | ||
504 | /* Be paranoid. RDS always has privdata */ | |
505 | if (!event->param.conn.private_data_len) { | |
506 | printk(KERN_NOTICE "RDS incoming connection has no private data, " | |
507 | "rejecting\n"); | |
508 | return 0; | |
509 | } | |
510 | ||
511 | /* Even if len is crap *now* I still want to check it. -ASG */ | |
f64f9e71 JP |
512 | if (event->param.conn.private_data_len < sizeof (*dp) || |
513 | dp->dp_protocol_major == 0) | |
ec16227e AG |
514 | return RDS_PROTOCOL_3_0; |
515 | ||
516 | common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS; | |
517 | if (dp->dp_protocol_major == 3 && common) { | |
518 | version = RDS_PROTOCOL_3_0; | |
519 | while ((common >>= 1) != 0) | |
520 | version++; | |
a4967598 MM |
521 | } else |
522 | printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n", | |
523 | &dp->dp_saddr, | |
524 | dp->dp_protocol_major, | |
525 | dp->dp_protocol_minor); | |
ec16227e AG |
526 | return version; |
527 | } | |
528 | ||
529 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, | |
530 | struct rdma_cm_event *event) | |
531 | { | |
532 | __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; | |
533 | __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; | |
534 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; | |
535 | struct rds_ib_connect_private dp_rep; | |
536 | struct rds_connection *conn = NULL; | |
537 | struct rds_ib_connection *ic = NULL; | |
538 | struct rdma_conn_param conn_param; | |
539 | u32 version; | |
a46ca94e | 540 | int err = 1, destroy = 1; |
ec16227e AG |
541 | |
542 | /* Check whether the remote protocol version matches ours. */ | |
9ddbcfa0 | 543 | version = rds_ib_protocol_compatible(event); |
ec16227e AG |
544 | if (!version) |
545 | goto out; | |
546 | ||
547 | rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid " | |
548 | "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr, | |
549 | RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), | |
550 | (unsigned long long)be64_to_cpu(lguid), | |
551 | (unsigned long long)be64_to_cpu(fguid)); | |
552 | ||
d5a8ac28 SV |
553 | /* RDS/IB is not currently netns aware, thus init_net */ |
554 | conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr, | |
555 | &rds_ib_transport, GFP_KERNEL); | |
ec16227e AG |
556 | if (IS_ERR(conn)) { |
557 | rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); | |
558 | conn = NULL; | |
559 | goto out; | |
560 | } | |
561 | ||
562 | /* | |
563 | * The connection request may occur while the | |
564 | * previous connection exist, e.g. in case of failover. | |
565 | * But as connections may be initiated simultaneously | |
566 | * by both hosts, we have a random backoff mechanism - | |
567 | * see the comment above rds_queue_reconnect() | |
568 | */ | |
569 | mutex_lock(&conn->c_cm_lock); | |
570 | if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { | |
571 | if (rds_conn_state(conn) == RDS_CONN_UP) { | |
572 | rdsdebug("incoming connect while connecting\n"); | |
573 | rds_conn_drop(conn); | |
574 | rds_ib_stats_inc(s_ib_listen_closed_stale); | |
575 | } else | |
576 | if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { | |
577 | /* Wait and see - our connect may still be succeeding */ | |
578 | rds_ib_stats_inc(s_ib_connect_raced); | |
579 | } | |
ec16227e AG |
580 | goto out; |
581 | } | |
582 | ||
583 | ic = conn->c_transport_data; | |
584 | ||
585 | rds_ib_set_protocol(conn, version); | |
586 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); | |
587 | ||
588 | /* If the peer gave us the last packet it saw, process this as if | |
589 | * we had received a regular ACK. */ | |
590 | if (dp->dp_ack_seq) | |
591 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); | |
592 | ||
593 | BUG_ON(cm_id->context); | |
594 | BUG_ON(ic->i_cm_id); | |
595 | ||
596 | ic->i_cm_id = cm_id; | |
597 | cm_id->context = conn; | |
598 | ||
599 | /* We got halfway through setting up the ib_connection, if we | |
600 | * fail now, we have to take the long route out of this mess. */ | |
601 | destroy = 0; | |
602 | ||
603 | err = rds_ib_setup_qp(conn); | |
604 | if (err) { | |
605 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); | |
606 | goto out; | |
607 | } | |
608 | ||
40589e74 AG |
609 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version, |
610 | event->param.conn.responder_resources, | |
611 | event->param.conn.initiator_depth); | |
ec16227e AG |
612 | |
613 | /* rdma_accept() calls rdma_reject() internally if it fails */ | |
614 | err = rdma_accept(cm_id, &conn_param); | |
a46ca94e | 615 | if (err) |
ec16227e | 616 | rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); |
ec16227e AG |
617 | |
618 | out: | |
a46ca94e ZB |
619 | if (conn) |
620 | mutex_unlock(&conn->c_cm_lock); | |
621 | if (err) | |
622 | rdma_reject(cm_id, NULL, 0); | |
ec16227e AG |
623 | return destroy; |
624 | } | |
625 | ||
626 | ||
627 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id) | |
628 | { | |
629 | struct rds_connection *conn = cm_id->context; | |
630 | struct rds_ib_connection *ic = conn->c_transport_data; | |
631 | struct rdma_conn_param conn_param; | |
632 | struct rds_ib_connect_private dp; | |
633 | int ret; | |
634 | ||
635 | /* If the peer doesn't do protocol negotiation, we must | |
636 | * default to RDSv3.0 */ | |
637 | rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); | |
638 | ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ | |
639 | ||
640 | ret = rds_ib_setup_qp(conn); | |
641 | if (ret) { | |
642 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); | |
643 | goto out; | |
644 | } | |
645 | ||
40589e74 AG |
646 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION, |
647 | UINT_MAX, UINT_MAX); | |
ec16227e AG |
648 | ret = rdma_connect(cm_id, &conn_param); |
649 | if (ret) | |
650 | rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); | |
651 | ||
652 | out: | |
653 | /* Beware - returning non-zero tells the rdma_cm to destroy | |
654 | * the cm_id. We should certainly not do it as long as we still | |
655 | * "own" the cm_id. */ | |
656 | if (ret) { | |
657 | if (ic->i_cm_id == cm_id) | |
658 | ret = 0; | |
659 | } | |
660 | return ret; | |
661 | } | |
662 | ||
663 | int rds_ib_conn_connect(struct rds_connection *conn) | |
664 | { | |
665 | struct rds_ib_connection *ic = conn->c_transport_data; | |
666 | struct sockaddr_in src, dest; | |
667 | int ret; | |
668 | ||
669 | /* XXX I wonder what affect the port space has */ | |
670 | /* delegate cm event handler to rdma_transport */ | |
fa20105e | 671 | ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn, |
b26f9b99 | 672 | RDMA_PS_TCP, IB_QPT_RC); |
ec16227e AG |
673 | if (IS_ERR(ic->i_cm_id)) { |
674 | ret = PTR_ERR(ic->i_cm_id); | |
675 | ic->i_cm_id = NULL; | |
676 | rdsdebug("rdma_create_id() failed: %d\n", ret); | |
677 | goto out; | |
678 | } | |
679 | ||
680 | rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); | |
681 | ||
682 | src.sin_family = AF_INET; | |
683 | src.sin_addr.s_addr = (__force u32)conn->c_laddr; | |
684 | src.sin_port = (__force u16)htons(0); | |
685 | ||
686 | dest.sin_family = AF_INET; | |
687 | dest.sin_addr.s_addr = (__force u32)conn->c_faddr; | |
688 | dest.sin_port = (__force u16)htons(RDS_PORT); | |
689 | ||
690 | ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, | |
691 | (struct sockaddr *)&dest, | |
692 | RDS_RDMA_RESOLVE_TIMEOUT_MS); | |
693 | if (ret) { | |
694 | rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, | |
695 | ret); | |
696 | rdma_destroy_id(ic->i_cm_id); | |
697 | ic->i_cm_id = NULL; | |
698 | } | |
699 | ||
700 | out: | |
701 | return ret; | |
702 | } | |
703 | ||
704 | /* | |
705 | * This is so careful about only cleaning up resources that were built up | |
706 | * so that it can be called at any point during startup. In fact it | |
707 | * can be called multiple times for a given connection. | |
708 | */ | |
709 | void rds_ib_conn_shutdown(struct rds_connection *conn) | |
710 | { | |
711 | struct rds_ib_connection *ic = conn->c_transport_data; | |
712 | int err = 0; | |
713 | ||
714 | rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, | |
715 | ic->i_pd, ic->i_send_cq, ic->i_recv_cq, | |
716 | ic->i_cm_id ? ic->i_cm_id->qp : NULL); | |
717 | ||
718 | if (ic->i_cm_id) { | |
719 | struct ib_device *dev = ic->i_cm_id->device; | |
720 | ||
721 | rdsdebug("disconnecting cm %p\n", ic->i_cm_id); | |
722 | err = rdma_disconnect(ic->i_cm_id); | |
723 | if (err) { | |
724 | /* Actually this may happen quite frequently, when | |
725 | * an outgoing connect raced with an incoming connect. | |
726 | */ | |
727 | rdsdebug("failed to disconnect, cm: %p err %d\n", | |
728 | ic->i_cm_id, err); | |
729 | } | |
730 | ||
e32b4a70 | 731 | /* |
f046011c ZB |
732 | * We want to wait for tx and rx completion to finish |
733 | * before we tear down the connection, but we have to be | |
734 | * careful not to get stuck waiting on a send ring that | |
735 | * only has unsignaled sends in it. We've shutdown new | |
736 | * sends before getting here so by waiting for signaled | |
737 | * sends to complete we're ensured that there will be no | |
738 | * more tx processing. | |
e32b4a70 | 739 | */ |
ec16227e | 740 | wait_event(rds_ib_ring_empty_wait, |
f046011c ZB |
741 | rds_ib_ring_empty(&ic->i_recv_ring) && |
742 | (atomic_read(&ic->i_signaled_sends) == 0)); | |
0c28c045 | 743 | tasklet_kill(&ic->i_send_tasklet); |
f046011c | 744 | tasklet_kill(&ic->i_recv_tasklet); |
ec16227e | 745 | |
1bc7b863 | 746 | /* first destroy the ib state that generates callbacks */ |
747 | if (ic->i_cm_id->qp) | |
748 | rdma_destroy_qp(ic->i_cm_id); | |
749 | if (ic->i_send_cq) | |
750 | ib_destroy_cq(ic->i_send_cq); | |
751 | if (ic->i_recv_cq) | |
752 | ib_destroy_cq(ic->i_recv_cq); | |
1bc7b863 | 753 | |
754 | /* then free the resources that ib callbacks use */ | |
ec16227e AG |
755 | if (ic->i_send_hdrs) |
756 | ib_dma_free_coherent(dev, | |
757 | ic->i_send_ring.w_nr * | |
758 | sizeof(struct rds_header), | |
759 | ic->i_send_hdrs, | |
760 | ic->i_send_hdrs_dma); | |
761 | ||
762 | if (ic->i_recv_hdrs) | |
763 | ib_dma_free_coherent(dev, | |
764 | ic->i_recv_ring.w_nr * | |
765 | sizeof(struct rds_header), | |
766 | ic->i_recv_hdrs, | |
767 | ic->i_recv_hdrs_dma); | |
768 | ||
769 | if (ic->i_ack) | |
770 | ib_dma_free_coherent(dev, sizeof(struct rds_header), | |
771 | ic->i_ack, ic->i_ack_dma); | |
772 | ||
773 | if (ic->i_sends) | |
774 | rds_ib_send_clear_ring(ic); | |
775 | if (ic->i_recvs) | |
776 | rds_ib_recv_clear_ring(ic); | |
777 | ||
1c3be624 SS |
778 | rdma_destroy_id(ic->i_cm_id); |
779 | ||
ec16227e AG |
780 | /* |
781 | * Move connection back to the nodev list. | |
782 | */ | |
745cbcca AG |
783 | if (ic->rds_ibdev) |
784 | rds_ib_remove_conn(ic->rds_ibdev, conn); | |
ec16227e AG |
785 | |
786 | ic->i_cm_id = NULL; | |
787 | ic->i_pd = NULL; | |
ec16227e AG |
788 | ic->i_send_cq = NULL; |
789 | ic->i_recv_cq = NULL; | |
790 | ic->i_send_hdrs = NULL; | |
791 | ic->i_recv_hdrs = NULL; | |
792 | ic->i_ack = NULL; | |
793 | } | |
794 | BUG_ON(ic->rds_ibdev); | |
795 | ||
796 | /* Clear pending transmit */ | |
ff3d7d36 AG |
797 | if (ic->i_data_op) { |
798 | struct rds_message *rm; | |
799 | ||
800 | rm = container_of(ic->i_data_op, struct rds_message, data); | |
801 | rds_message_put(rm); | |
802 | ic->i_data_op = NULL; | |
ec16227e AG |
803 | } |
804 | ||
805 | /* Clear the ACK state */ | |
806 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
8cbd9606 AG |
807 | #ifdef KERNEL_HAS_ATOMIC64 |
808 | atomic64_set(&ic->i_ack_next, 0); | |
809 | #else | |
810 | ic->i_ack_next = 0; | |
811 | #endif | |
ec16227e AG |
812 | ic->i_ack_recv = 0; |
813 | ||
814 | /* Clear flow control state */ | |
815 | ic->i_flowctl = 0; | |
816 | atomic_set(&ic->i_credits, 0); | |
817 | ||
818 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
819 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
820 | ||
821 | if (ic->i_ibinc) { | |
822 | rds_inc_put(&ic->i_ibinc->ii_inc); | |
823 | ic->i_ibinc = NULL; | |
824 | } | |
825 | ||
826 | vfree(ic->i_sends); | |
827 | ic->i_sends = NULL; | |
828 | vfree(ic->i_recvs); | |
829 | ic->i_recvs = NULL; | |
830 | } | |
831 | ||
832 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) | |
833 | { | |
834 | struct rds_ib_connection *ic; | |
835 | unsigned long flags; | |
33244125 | 836 | int ret; |
ec16227e AG |
837 | |
838 | /* XXX too lazy? */ | |
f0229eaa | 839 | ic = kzalloc(sizeof(struct rds_ib_connection), gfp); |
8690bfa1 | 840 | if (!ic) |
ec16227e AG |
841 | return -ENOMEM; |
842 | ||
33244125 CM |
843 | ret = rds_ib_recv_alloc_caches(ic); |
844 | if (ret) { | |
845 | kfree(ic); | |
846 | return ret; | |
847 | } | |
848 | ||
ec16227e | 849 | INIT_LIST_HEAD(&ic->ib_node); |
0c28c045 SS |
850 | tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send, |
851 | (unsigned long)ic); | |
f4f943c9 | 852 | tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv, |
0c28c045 | 853 | (unsigned long)ic); |
ec16227e | 854 | mutex_init(&ic->i_recv_mutex); |
8cbd9606 AG |
855 | #ifndef KERNEL_HAS_ATOMIC64 |
856 | spin_lock_init(&ic->i_ack_lock); | |
857 | #endif | |
f046011c | 858 | atomic_set(&ic->i_signaled_sends, 0); |
ec16227e AG |
859 | |
860 | /* | |
861 | * rds_ib_conn_shutdown() waits for these to be emptied so they | |
862 | * must be initialized before it can be called. | |
863 | */ | |
864 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
865 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
866 | ||
867 | ic->conn = conn; | |
868 | conn->c_transport_data = ic; | |
869 | ||
870 | spin_lock_irqsave(&ib_nodev_conns_lock, flags); | |
871 | list_add_tail(&ic->ib_node, &ib_nodev_conns); | |
872 | spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); | |
873 | ||
874 | ||
875 | rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); | |
876 | return 0; | |
877 | } | |
878 | ||
745cbcca AG |
879 | /* |
880 | * Free a connection. Connection must be shut down and not set for reconnect. | |
881 | */ | |
ec16227e AG |
882 | void rds_ib_conn_free(void *arg) |
883 | { | |
884 | struct rds_ib_connection *ic = arg; | |
745cbcca AG |
885 | spinlock_t *lock_ptr; |
886 | ||
ec16227e | 887 | rdsdebug("ic %p\n", ic); |
745cbcca AG |
888 | |
889 | /* | |
890 | * Conn is either on a dev's list or on the nodev list. | |
891 | * A race with shutdown() or connect() would cause problems | |
892 | * (since rds_ibdev would change) but that should never happen. | |
893 | */ | |
894 | lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; | |
895 | ||
896 | spin_lock_irq(lock_ptr); | |
ec16227e | 897 | list_del(&ic->ib_node); |
745cbcca AG |
898 | spin_unlock_irq(lock_ptr); |
899 | ||
33244125 CM |
900 | rds_ib_recv_free_caches(ic); |
901 | ||
ec16227e AG |
902 | kfree(ic); |
903 | } | |
904 | ||
905 | ||
906 | /* | |
907 | * An error occurred on the connection | |
908 | */ | |
909 | void | |
910 | __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) | |
911 | { | |
912 | va_list ap; | |
913 | ||
914 | rds_conn_drop(conn); | |
915 | ||
916 | va_start(ap, fmt); | |
917 | vprintk(fmt, ap); | |
918 | va_end(ap); | |
919 | } |