Loading...
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/slab.h>
36#include <linux/vmalloc.h>
37#include <linux/ratelimit.h>
38
39#include "rds.h"
40#include "ib.h"
41
42static char *rds_ib_event_type_strings[] = {
43#define RDS_IB_EVENT_STRING(foo) \
44 [IB_EVENT_##foo] = __stringify(IB_EVENT_##foo)
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
66static char *rds_ib_event_str(enum ib_event_type type)
67{
68 return rds_str_array(rds_ib_event_type_strings,
69 ARRAY_SIZE(rds_ib_event_type_strings), type);
70};
71
72/*
73 * Set the selected protocol version
74 */
75static 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 */
83static 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 */
110static void
111rds_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 */
125void 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;
129 struct ib_qp_attr qp_attr;
130 int err;
131
132 if (event->param.conn.private_data_len >= sizeof(*dp)) {
133 dp = event->param.conn.private_data;
134
135 /* make sure it isn't empty data */
136 if (dp->dp_protocol_major) {
137 rds_ib_set_protocol(conn,
138 RDS_PROTOCOL(dp->dp_protocol_major,
139 dp->dp_protocol_minor));
140 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
141 }
142 }
143
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 }
159
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. */
168 rds_ib_recv_refill(conn, 1);
169
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
178 /* update ib_device with this local ipaddr */
179 err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
180 if (err)
181 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
182 err);
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
192static 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,
195 u32 protocol_version,
196 u32 max_responder_resources,
197 u32 max_initiator_depth)
198{
199 struct rds_ib_connection *ic = conn->c_transport_data;
200 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
201
202 memset(conn_param, 0, sizeof(struct rdma_conn_param));
203
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);
208 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
209 conn_param->rnr_retry_count = 7;
210
211 if (dp) {
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
234static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
235{
236 rdsdebug("event %u (%s) data %p\n",
237 event->event, rds_ib_event_str(event->event), data);
238}
239
240static 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
245 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
246 rds_ib_event_str(event->event));
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:
253 rdsdebug("Fatal QP Event %u (%s) "
254 "- connection %pI4->%pI4, reconnecting\n",
255 event->event, rds_ib_event_str(event->event),
256 &conn->c_laddr, &conn->c_faddr);
257 rds_conn_drop(conn);
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 */
266static 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
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.
277 */
278 rds_ibdev = rds_ib_get_client_data(dev);
279 if (!rds_ibdev)
280 return -EOPNOTSUPP;
281
282 /* add the conn now so that connection establishment has the dev */
283 rds_ib_add_conn(rds_ibdev, conn);
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);
354 if (!ic->i_send_hdrs) {
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);
364 if (!ic->i_recv_hdrs) {
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);
372 if (!ic->i_ack) {
373 ret = -ENOMEM;
374 rdsdebug("ib_dma_alloc_coherent ack failed\n");
375 goto out;
376 }
377
378 ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
379 ibdev_to_node(dev));
380 if (!ic->i_sends) {
381 ret = -ENOMEM;
382 rdsdebug("send allocation failed\n");
383 goto out;
384 }
385
386 ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
387 ibdev_to_node(dev));
388 if (!ic->i_recvs) {
389 ret = -ENOMEM;
390 rdsdebug("recv allocation failed\n");
391 goto out;
392 }
393
394 rds_ib_recv_init_ack(ic);
395
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
399out:
400 rds_ib_dev_put(rds_ibdev);
401 return ret;
402}
403
404static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
405{
406 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
407 u16 common;
408 u32 version = 0;
409
410 /*
411 * rdma_cm private data is odd - when there is any private data in the
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.
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 */
428 if (event->param.conn.private_data_len < sizeof (*dp) ||
429 dp->dp_protocol_major == 0)
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++;
437 }
438 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using "
439 "incompatible protocol version %u.%u\n",
440 &dp->dp_saddr,
441 dp->dp_protocol_major,
442 dp->dp_protocol_minor);
443 return version;
444}
445
446int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
447 struct rdma_cm_event *event)
448{
449 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
450 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
451 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
452 struct rds_ib_connect_private dp_rep;
453 struct rds_connection *conn = NULL;
454 struct rds_ib_connection *ic = NULL;
455 struct rdma_conn_param conn_param;
456 u32 version;
457 int err = 1, destroy = 1;
458
459 /* Check whether the remote protocol version matches ours. */
460 version = rds_ib_protocol_compatible(event);
461 if (!version)
462 goto out;
463
464 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
465 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
466 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
467 (unsigned long long)be64_to_cpu(lguid),
468 (unsigned long long)be64_to_cpu(fguid));
469
470 conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
471 GFP_KERNEL);
472 if (IS_ERR(conn)) {
473 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
474 conn = NULL;
475 goto out;
476 }
477
478 /*
479 * The connection request may occur while the
480 * previous connection exist, e.g. in case of failover.
481 * But as connections may be initiated simultaneously
482 * by both hosts, we have a random backoff mechanism -
483 * see the comment above rds_queue_reconnect()
484 */
485 mutex_lock(&conn->c_cm_lock);
486 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
487 if (rds_conn_state(conn) == RDS_CONN_UP) {
488 rdsdebug("incoming connect while connecting\n");
489 rds_conn_drop(conn);
490 rds_ib_stats_inc(s_ib_listen_closed_stale);
491 } else
492 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
493 /* Wait and see - our connect may still be succeeding */
494 rds_ib_stats_inc(s_ib_connect_raced);
495 }
496 goto out;
497 }
498
499 ic = conn->c_transport_data;
500
501 rds_ib_set_protocol(conn, version);
502 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
503
504 /* If the peer gave us the last packet it saw, process this as if
505 * we had received a regular ACK. */
506 if (dp->dp_ack_seq)
507 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
508
509 BUG_ON(cm_id->context);
510 BUG_ON(ic->i_cm_id);
511
512 ic->i_cm_id = cm_id;
513 cm_id->context = conn;
514
515 /* We got halfway through setting up the ib_connection, if we
516 * fail now, we have to take the long route out of this mess. */
517 destroy = 0;
518
519 err = rds_ib_setup_qp(conn);
520 if (err) {
521 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
522 goto out;
523 }
524
525 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
526 event->param.conn.responder_resources,
527 event->param.conn.initiator_depth);
528
529 /* rdma_accept() calls rdma_reject() internally if it fails */
530 err = rdma_accept(cm_id, &conn_param);
531 if (err)
532 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
533
534out:
535 if (conn)
536 mutex_unlock(&conn->c_cm_lock);
537 if (err)
538 rdma_reject(cm_id, NULL, 0);
539 return destroy;
540}
541
542
543int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
544{
545 struct rds_connection *conn = cm_id->context;
546 struct rds_ib_connection *ic = conn->c_transport_data;
547 struct rdma_conn_param conn_param;
548 struct rds_ib_connect_private dp;
549 int ret;
550
551 /* If the peer doesn't do protocol negotiation, we must
552 * default to RDSv3.0 */
553 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
554 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
555
556 ret = rds_ib_setup_qp(conn);
557 if (ret) {
558 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
559 goto out;
560 }
561
562 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
563 UINT_MAX, UINT_MAX);
564 ret = rdma_connect(cm_id, &conn_param);
565 if (ret)
566 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
567
568out:
569 /* Beware - returning non-zero tells the rdma_cm to destroy
570 * the cm_id. We should certainly not do it as long as we still
571 * "own" the cm_id. */
572 if (ret) {
573 if (ic->i_cm_id == cm_id)
574 ret = 0;
575 }
576 return ret;
577}
578
579int rds_ib_conn_connect(struct rds_connection *conn)
580{
581 struct rds_ib_connection *ic = conn->c_transport_data;
582 struct sockaddr_in src, dest;
583 int ret;
584
585 /* XXX I wonder what affect the port space has */
586 /* delegate cm event handler to rdma_transport */
587 ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
588 RDMA_PS_TCP, IB_QPT_RC);
589 if (IS_ERR(ic->i_cm_id)) {
590 ret = PTR_ERR(ic->i_cm_id);
591 ic->i_cm_id = NULL;
592 rdsdebug("rdma_create_id() failed: %d\n", ret);
593 goto out;
594 }
595
596 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
597
598 src.sin_family = AF_INET;
599 src.sin_addr.s_addr = (__force u32)conn->c_laddr;
600 src.sin_port = (__force u16)htons(0);
601
602 dest.sin_family = AF_INET;
603 dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
604 dest.sin_port = (__force u16)htons(RDS_PORT);
605
606 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
607 (struct sockaddr *)&dest,
608 RDS_RDMA_RESOLVE_TIMEOUT_MS);
609 if (ret) {
610 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
611 ret);
612 rdma_destroy_id(ic->i_cm_id);
613 ic->i_cm_id = NULL;
614 }
615
616out:
617 return ret;
618}
619
620/*
621 * This is so careful about only cleaning up resources that were built up
622 * so that it can be called at any point during startup. In fact it
623 * can be called multiple times for a given connection.
624 */
625void rds_ib_conn_shutdown(struct rds_connection *conn)
626{
627 struct rds_ib_connection *ic = conn->c_transport_data;
628 int err = 0;
629
630 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
631 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
632 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
633
634 if (ic->i_cm_id) {
635 struct ib_device *dev = ic->i_cm_id->device;
636
637 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
638 err = rdma_disconnect(ic->i_cm_id);
639 if (err) {
640 /* Actually this may happen quite frequently, when
641 * an outgoing connect raced with an incoming connect.
642 */
643 rdsdebug("failed to disconnect, cm: %p err %d\n",
644 ic->i_cm_id, err);
645 }
646
647 /*
648 * We want to wait for tx and rx completion to finish
649 * before we tear down the connection, but we have to be
650 * careful not to get stuck waiting on a send ring that
651 * only has unsignaled sends in it. We've shutdown new
652 * sends before getting here so by waiting for signaled
653 * sends to complete we're ensured that there will be no
654 * more tx processing.
655 */
656 wait_event(rds_ib_ring_empty_wait,
657 rds_ib_ring_empty(&ic->i_recv_ring) &&
658 (atomic_read(&ic->i_signaled_sends) == 0));
659 tasklet_kill(&ic->i_recv_tasklet);
660
661 if (ic->i_send_hdrs)
662 ib_dma_free_coherent(dev,
663 ic->i_send_ring.w_nr *
664 sizeof(struct rds_header),
665 ic->i_send_hdrs,
666 ic->i_send_hdrs_dma);
667
668 if (ic->i_recv_hdrs)
669 ib_dma_free_coherent(dev,
670 ic->i_recv_ring.w_nr *
671 sizeof(struct rds_header),
672 ic->i_recv_hdrs,
673 ic->i_recv_hdrs_dma);
674
675 if (ic->i_ack)
676 ib_dma_free_coherent(dev, sizeof(struct rds_header),
677 ic->i_ack, ic->i_ack_dma);
678
679 if (ic->i_sends)
680 rds_ib_send_clear_ring(ic);
681 if (ic->i_recvs)
682 rds_ib_recv_clear_ring(ic);
683
684 if (ic->i_cm_id->qp)
685 rdma_destroy_qp(ic->i_cm_id);
686 if (ic->i_send_cq)
687 ib_destroy_cq(ic->i_send_cq);
688 if (ic->i_recv_cq)
689 ib_destroy_cq(ic->i_recv_cq);
690 rdma_destroy_id(ic->i_cm_id);
691
692 /*
693 * Move connection back to the nodev list.
694 */
695 if (ic->rds_ibdev)
696 rds_ib_remove_conn(ic->rds_ibdev, conn);
697
698 ic->i_cm_id = NULL;
699 ic->i_pd = NULL;
700 ic->i_mr = NULL;
701 ic->i_send_cq = NULL;
702 ic->i_recv_cq = NULL;
703 ic->i_send_hdrs = NULL;
704 ic->i_recv_hdrs = NULL;
705 ic->i_ack = NULL;
706 }
707 BUG_ON(ic->rds_ibdev);
708
709 /* Clear pending transmit */
710 if (ic->i_data_op) {
711 struct rds_message *rm;
712
713 rm = container_of(ic->i_data_op, struct rds_message, data);
714 rds_message_put(rm);
715 ic->i_data_op = NULL;
716 }
717
718 /* Clear the ACK state */
719 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
720#ifdef KERNEL_HAS_ATOMIC64
721 atomic64_set(&ic->i_ack_next, 0);
722#else
723 ic->i_ack_next = 0;
724#endif
725 ic->i_ack_recv = 0;
726
727 /* Clear flow control state */
728 ic->i_flowctl = 0;
729 atomic_set(&ic->i_credits, 0);
730
731 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
732 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
733
734 if (ic->i_ibinc) {
735 rds_inc_put(&ic->i_ibinc->ii_inc);
736 ic->i_ibinc = NULL;
737 }
738
739 vfree(ic->i_sends);
740 ic->i_sends = NULL;
741 vfree(ic->i_recvs);
742 ic->i_recvs = NULL;
743}
744
745int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
746{
747 struct rds_ib_connection *ic;
748 unsigned long flags;
749 int ret;
750
751 /* XXX too lazy? */
752 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
753 if (!ic)
754 return -ENOMEM;
755
756 ret = rds_ib_recv_alloc_caches(ic);
757 if (ret) {
758 kfree(ic);
759 return ret;
760 }
761
762 INIT_LIST_HEAD(&ic->ib_node);
763 tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
764 (unsigned long) ic);
765 mutex_init(&ic->i_recv_mutex);
766#ifndef KERNEL_HAS_ATOMIC64
767 spin_lock_init(&ic->i_ack_lock);
768#endif
769 atomic_set(&ic->i_signaled_sends, 0);
770
771 /*
772 * rds_ib_conn_shutdown() waits for these to be emptied so they
773 * must be initialized before it can be called.
774 */
775 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
776 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
777
778 ic->conn = conn;
779 conn->c_transport_data = ic;
780
781 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
782 list_add_tail(&ic->ib_node, &ib_nodev_conns);
783 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
784
785
786 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
787 return 0;
788}
789
790/*
791 * Free a connection. Connection must be shut down and not set for reconnect.
792 */
793void rds_ib_conn_free(void *arg)
794{
795 struct rds_ib_connection *ic = arg;
796 spinlock_t *lock_ptr;
797
798 rdsdebug("ic %p\n", ic);
799
800 /*
801 * Conn is either on a dev's list or on the nodev list.
802 * A race with shutdown() or connect() would cause problems
803 * (since rds_ibdev would change) but that should never happen.
804 */
805 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
806
807 spin_lock_irq(lock_ptr);
808 list_del(&ic->ib_node);
809 spin_unlock_irq(lock_ptr);
810
811 rds_ib_recv_free_caches(ic);
812
813 kfree(ic);
814}
815
816
817/*
818 * An error occurred on the connection
819 */
820void
821__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
822{
823 va_list ap;
824
825 rds_conn_drop(conn);
826
827 va_start(ap, fmt);
828 vprintk(fmt, ap);
829 va_end(ap);
830}
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/slab.h>
36#include <linux/vmalloc.h>
37#include <linux/ratelimit.h>
38
39#include "rds_single_path.h"
40#include "rds.h"
41#include "ib.h"
42
43/*
44 * Set the selected protocol version
45 */
46static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
47{
48 conn->c_version = version;
49}
50
51/*
52 * Set up flow control
53 */
54static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
55{
56 struct rds_ib_connection *ic = conn->c_transport_data;
57
58 if (rds_ib_sysctl_flow_control && credits != 0) {
59 /* We're doing flow control */
60 ic->i_flowctl = 1;
61 rds_ib_send_add_credits(conn, credits);
62 } else {
63 ic->i_flowctl = 0;
64 }
65}
66
67/*
68 * Tune RNR behavior. Without flow control, we use a rather
69 * low timeout, but not the absolute minimum - this should
70 * be tunable.
71 *
72 * We already set the RNR retry count to 7 (which is the
73 * smallest infinite number :-) above.
74 * If flow control is off, we want to change this back to 0
75 * so that we learn quickly when our credit accounting is
76 * buggy.
77 *
78 * Caller passes in a qp_attr pointer - don't waste stack spacv
79 * by allocation this twice.
80 */
81static void
82rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
83{
84 int ret;
85
86 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
87 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
88 if (ret)
89 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
90}
91
92/*
93 * Connection established.
94 * We get here for both outgoing and incoming connection.
95 */
96void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
97{
98 const struct rds_ib_connect_private *dp = NULL;
99 struct rds_ib_connection *ic = conn->c_transport_data;
100 struct ib_qp_attr qp_attr;
101 int err;
102
103 if (event->param.conn.private_data_len >= sizeof(*dp)) {
104 dp = event->param.conn.private_data;
105
106 /* make sure it isn't empty data */
107 if (dp->dp_protocol_major) {
108 rds_ib_set_protocol(conn,
109 RDS_PROTOCOL(dp->dp_protocol_major,
110 dp->dp_protocol_minor));
111 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
112 }
113 }
114
115 if (conn->c_version < RDS_PROTOCOL(3, 1)) {
116 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
117 " no longer supported\n",
118 &conn->c_faddr,
119 RDS_PROTOCOL_MAJOR(conn->c_version),
120 RDS_PROTOCOL_MINOR(conn->c_version));
121 rds_conn_destroy(conn);
122 return;
123 } else {
124 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
125 &conn->c_faddr,
126 RDS_PROTOCOL_MAJOR(conn->c_version),
127 RDS_PROTOCOL_MINOR(conn->c_version),
128 ic->i_flowctl ? ", flow control" : "");
129 }
130
131 /*
132 * Init rings and fill recv. this needs to wait until protocol negotiation
133 * is complete, since ring layout is different from 3.0 to 3.1.
134 */
135 rds_ib_send_init_ring(ic);
136 rds_ib_recv_init_ring(ic);
137 /* Post receive buffers - as a side effect, this will update
138 * the posted credit count. */
139 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
140
141 /* Tune RNR behavior */
142 rds_ib_tune_rnr(ic, &qp_attr);
143
144 qp_attr.qp_state = IB_QPS_RTS;
145 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
146 if (err)
147 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
148
149 /* update ib_device with this local ipaddr */
150 err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
151 if (err)
152 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
153 err);
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) {
158 /* dp structure start is not guaranteed to be 8 bytes aligned.
159 * Since dp_ack_seq is 64-bit extended load operations can be
160 * used so go through get_unaligned to avoid unaligned errors.
161 */
162 __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
163
164 if (dp_ack_seq)
165 rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
166 NULL);
167 }
168
169 rds_connect_complete(conn);
170}
171
172static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
173 struct rdma_conn_param *conn_param,
174 struct rds_ib_connect_private *dp,
175 u32 protocol_version,
176 u32 max_responder_resources,
177 u32 max_initiator_depth)
178{
179 struct rds_ib_connection *ic = conn->c_transport_data;
180 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
181
182 memset(conn_param, 0, sizeof(struct rdma_conn_param));
183
184 conn_param->responder_resources =
185 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
186 conn_param->initiator_depth =
187 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
188 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
189 conn_param->rnr_retry_count = 7;
190
191 if (dp) {
192 memset(dp, 0, sizeof(*dp));
193 dp->dp_saddr = conn->c_laddr;
194 dp->dp_daddr = conn->c_faddr;
195 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
196 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
197 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
198 dp->dp_ack_seq = cpu_to_be64(rds_ib_piggyb_ack(ic));
199
200 /* Advertise flow control */
201 if (ic->i_flowctl) {
202 unsigned int credits;
203
204 credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
205 dp->dp_credit = cpu_to_be32(credits);
206 atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
207 }
208
209 conn_param->private_data = dp;
210 conn_param->private_data_len = sizeof(*dp);
211 }
212}
213
214static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
215{
216 rdsdebug("event %u (%s) data %p\n",
217 event->event, ib_event_msg(event->event), data);
218}
219
220/* Plucking the oldest entry from the ring can be done concurrently with
221 * the thread refilling the ring. Each ring operation is protected by
222 * spinlocks and the transient state of refilling doesn't change the
223 * recording of which entry is oldest.
224 *
225 * This relies on IB only calling one cq comp_handler for each cq so that
226 * there will only be one caller of rds_recv_incoming() per RDS connection.
227 */
228static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
229{
230 struct rds_connection *conn = context;
231 struct rds_ib_connection *ic = conn->c_transport_data;
232
233 rdsdebug("conn %p cq %p\n", conn, cq);
234
235 rds_ib_stats_inc(s_ib_evt_handler_call);
236
237 tasklet_schedule(&ic->i_recv_tasklet);
238}
239
240static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
241 struct ib_wc *wcs)
242{
243 int nr, i;
244 struct ib_wc *wc;
245
246 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
247 for (i = 0; i < nr; i++) {
248 wc = wcs + i;
249 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
250 (unsigned long long)wc->wr_id, wc->status,
251 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
252
253 if (wc->wr_id <= ic->i_send_ring.w_nr ||
254 wc->wr_id == RDS_IB_ACK_WR_ID)
255 rds_ib_send_cqe_handler(ic, wc);
256 else
257 rds_ib_mr_cqe_handler(ic, wc);
258
259 }
260 }
261}
262
263static void rds_ib_tasklet_fn_send(unsigned long data)
264{
265 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
266 struct rds_connection *conn = ic->conn;
267
268 rds_ib_stats_inc(s_ib_tasklet_call);
269
270 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
271 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
272 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
273
274 if (rds_conn_up(conn) &&
275 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
276 test_bit(0, &conn->c_map_queued)))
277 rds_send_xmit(&ic->conn->c_path[0]);
278}
279
280static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
281 struct ib_wc *wcs,
282 struct rds_ib_ack_state *ack_state)
283{
284 int nr, i;
285 struct ib_wc *wc;
286
287 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
288 for (i = 0; i < nr; i++) {
289 wc = wcs + i;
290 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
291 (unsigned long long)wc->wr_id, wc->status,
292 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
293
294 rds_ib_recv_cqe_handler(ic, wc, ack_state);
295 }
296 }
297}
298
299static void rds_ib_tasklet_fn_recv(unsigned long data)
300{
301 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
302 struct rds_connection *conn = ic->conn;
303 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
304 struct rds_ib_ack_state state;
305
306 if (!rds_ibdev)
307 rds_conn_drop(conn);
308
309 rds_ib_stats_inc(s_ib_tasklet_call);
310
311 memset(&state, 0, sizeof(state));
312 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
313 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
314 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
315
316 if (state.ack_next_valid)
317 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
318 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
319 rds_send_drop_acked(conn, state.ack_recv, NULL);
320 ic->i_ack_recv = state.ack_recv;
321 }
322
323 if (rds_conn_up(conn))
324 rds_ib_attempt_ack(ic);
325}
326
327static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
328{
329 struct rds_connection *conn = data;
330 struct rds_ib_connection *ic = conn->c_transport_data;
331
332 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
333 ib_event_msg(event->event));
334
335 switch (event->event) {
336 case IB_EVENT_COMM_EST:
337 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
338 break;
339 default:
340 rdsdebug("Fatal QP Event %u (%s) "
341 "- connection %pI4->%pI4, reconnecting\n",
342 event->event, ib_event_msg(event->event),
343 &conn->c_laddr, &conn->c_faddr);
344 rds_conn_drop(conn);
345 break;
346 }
347}
348
349static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
350{
351 struct rds_connection *conn = context;
352 struct rds_ib_connection *ic = conn->c_transport_data;
353
354 rdsdebug("conn %p cq %p\n", conn, cq);
355
356 rds_ib_stats_inc(s_ib_evt_handler_call);
357
358 tasklet_schedule(&ic->i_send_tasklet);
359}
360
361/*
362 * This needs to be very careful to not leave IS_ERR pointers around for
363 * cleanup to trip over.
364 */
365static int rds_ib_setup_qp(struct rds_connection *conn)
366{
367 struct rds_ib_connection *ic = conn->c_transport_data;
368 struct ib_device *dev = ic->i_cm_id->device;
369 struct ib_qp_init_attr attr;
370 struct ib_cq_init_attr cq_attr = {};
371 struct rds_ib_device *rds_ibdev;
372 int ret, fr_queue_space;
373
374 /*
375 * It's normal to see a null device if an incoming connection races
376 * with device removal, so we don't print a warning.
377 */
378 rds_ibdev = rds_ib_get_client_data(dev);
379 if (!rds_ibdev)
380 return -EOPNOTSUPP;
381
382 /* The fr_queue_space is currently set to 512, to add extra space on
383 * completion queue and send queue. This extra space is used for FRMR
384 * registration and invalidation work requests
385 */
386 fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
387
388 /* add the conn now so that connection establishment has the dev */
389 rds_ib_add_conn(rds_ibdev, conn);
390
391 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
392 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
393 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
394 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
395
396 /* Protection domain and memory range */
397 ic->i_pd = rds_ibdev->pd;
398
399 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
400
401 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
402 rds_ib_cq_event_handler, conn,
403 &cq_attr);
404 if (IS_ERR(ic->i_send_cq)) {
405 ret = PTR_ERR(ic->i_send_cq);
406 ic->i_send_cq = NULL;
407 rdsdebug("ib_create_cq send failed: %d\n", ret);
408 goto out;
409 }
410
411 cq_attr.cqe = ic->i_recv_ring.w_nr;
412 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
413 rds_ib_cq_event_handler, conn,
414 &cq_attr);
415 if (IS_ERR(ic->i_recv_cq)) {
416 ret = PTR_ERR(ic->i_recv_cq);
417 ic->i_recv_cq = NULL;
418 rdsdebug("ib_create_cq recv failed: %d\n", ret);
419 goto out;
420 }
421
422 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
423 if (ret) {
424 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
425 goto out;
426 }
427
428 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
429 if (ret) {
430 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
431 goto out;
432 }
433
434 /* XXX negotiate max send/recv with remote? */
435 memset(&attr, 0, sizeof(attr));
436 attr.event_handler = rds_ib_qp_event_handler;
437 attr.qp_context = conn;
438 /* + 1 to allow for the single ack message */
439 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
440 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
441 attr.cap.max_send_sge = rds_ibdev->max_sge;
442 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
443 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
444 attr.qp_type = IB_QPT_RC;
445 attr.send_cq = ic->i_send_cq;
446 attr.recv_cq = ic->i_recv_cq;
447 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
448
449 /*
450 * XXX this can fail if max_*_wr is too large? Are we supposed
451 * to back off until we get a value that the hardware can support?
452 */
453 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
454 if (ret) {
455 rdsdebug("rdma_create_qp failed: %d\n", ret);
456 goto out;
457 }
458
459 ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
460 ic->i_send_ring.w_nr *
461 sizeof(struct rds_header),
462 &ic->i_send_hdrs_dma, GFP_KERNEL);
463 if (!ic->i_send_hdrs) {
464 ret = -ENOMEM;
465 rdsdebug("ib_dma_alloc_coherent send failed\n");
466 goto out;
467 }
468
469 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
470 ic->i_recv_ring.w_nr *
471 sizeof(struct rds_header),
472 &ic->i_recv_hdrs_dma, GFP_KERNEL);
473 if (!ic->i_recv_hdrs) {
474 ret = -ENOMEM;
475 rdsdebug("ib_dma_alloc_coherent recv failed\n");
476 goto out;
477 }
478
479 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
480 &ic->i_ack_dma, GFP_KERNEL);
481 if (!ic->i_ack) {
482 ret = -ENOMEM;
483 rdsdebug("ib_dma_alloc_coherent ack failed\n");
484 goto out;
485 }
486
487 ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
488 ibdev_to_node(dev));
489 if (!ic->i_sends) {
490 ret = -ENOMEM;
491 rdsdebug("send allocation failed\n");
492 goto out;
493 }
494
495 ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
496 ibdev_to_node(dev));
497 if (!ic->i_recvs) {
498 ret = -ENOMEM;
499 rdsdebug("recv allocation failed\n");
500 goto out;
501 }
502
503 rds_ib_recv_init_ack(ic);
504
505 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
506 ic->i_send_cq, ic->i_recv_cq);
507
508out:
509 rds_ib_dev_put(rds_ibdev);
510 return ret;
511}
512
513static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
514{
515 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
516 u16 common;
517 u32 version = 0;
518
519 /*
520 * rdma_cm private data is odd - when there is any private data in the
521 * request, we will be given a pretty large buffer without telling us the
522 * original size. The only way to tell the difference is by looking at
523 * the contents, which are initialized to zero.
524 * If the protocol version fields aren't set, this is a connection attempt
525 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
526 * We really should have changed this for OFED 1.3 :-(
527 */
528
529 /* Be paranoid. RDS always has privdata */
530 if (!event->param.conn.private_data_len) {
531 printk(KERN_NOTICE "RDS incoming connection has no private data, "
532 "rejecting\n");
533 return 0;
534 }
535
536 /* Even if len is crap *now* I still want to check it. -ASG */
537 if (event->param.conn.private_data_len < sizeof (*dp) ||
538 dp->dp_protocol_major == 0)
539 return RDS_PROTOCOL_3_0;
540
541 common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
542 if (dp->dp_protocol_major == 3 && common) {
543 version = RDS_PROTOCOL_3_0;
544 while ((common >>= 1) != 0)
545 version++;
546 } else
547 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
548 &dp->dp_saddr,
549 dp->dp_protocol_major,
550 dp->dp_protocol_minor);
551 return version;
552}
553
554int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
555 struct rdma_cm_event *event)
556{
557 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
558 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
559 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
560 struct rds_ib_connect_private dp_rep;
561 struct rds_connection *conn = NULL;
562 struct rds_ib_connection *ic = NULL;
563 struct rdma_conn_param conn_param;
564 u32 version;
565 int err = 1, destroy = 1;
566
567 /* Check whether the remote protocol version matches ours. */
568 version = rds_ib_protocol_compatible(event);
569 if (!version)
570 goto out;
571
572 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
573 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
574 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
575 (unsigned long long)be64_to_cpu(lguid),
576 (unsigned long long)be64_to_cpu(fguid));
577
578 /* RDS/IB is not currently netns aware, thus init_net */
579 conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
580 &rds_ib_transport, GFP_KERNEL);
581 if (IS_ERR(conn)) {
582 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
583 conn = NULL;
584 goto out;
585 }
586
587 /*
588 * The connection request may occur while the
589 * previous connection exist, e.g. in case of failover.
590 * But as connections may be initiated simultaneously
591 * by both hosts, we have a random backoff mechanism -
592 * see the comment above rds_queue_reconnect()
593 */
594 mutex_lock(&conn->c_cm_lock);
595 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
596 if (rds_conn_state(conn) == RDS_CONN_UP) {
597 rdsdebug("incoming connect while connecting\n");
598 rds_conn_drop(conn);
599 rds_ib_stats_inc(s_ib_listen_closed_stale);
600 } else
601 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
602 /* Wait and see - our connect may still be succeeding */
603 rds_ib_stats_inc(s_ib_connect_raced);
604 }
605 goto out;
606 }
607
608 ic = conn->c_transport_data;
609
610 rds_ib_set_protocol(conn, version);
611 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
612
613 /* If the peer gave us the last packet it saw, process this as if
614 * we had received a regular ACK. */
615 if (dp->dp_ack_seq)
616 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
617
618 BUG_ON(cm_id->context);
619 BUG_ON(ic->i_cm_id);
620
621 ic->i_cm_id = cm_id;
622 cm_id->context = conn;
623
624 /* We got halfway through setting up the ib_connection, if we
625 * fail now, we have to take the long route out of this mess. */
626 destroy = 0;
627
628 err = rds_ib_setup_qp(conn);
629 if (err) {
630 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
631 goto out;
632 }
633
634 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
635 event->param.conn.responder_resources,
636 event->param.conn.initiator_depth);
637
638 /* rdma_accept() calls rdma_reject() internally if it fails */
639 err = rdma_accept(cm_id, &conn_param);
640 if (err)
641 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
642
643out:
644 if (conn)
645 mutex_unlock(&conn->c_cm_lock);
646 if (err)
647 rdma_reject(cm_id, NULL, 0);
648 return destroy;
649}
650
651
652int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
653{
654 struct rds_connection *conn = cm_id->context;
655 struct rds_ib_connection *ic = conn->c_transport_data;
656 struct rdma_conn_param conn_param;
657 struct rds_ib_connect_private dp;
658 int ret;
659
660 /* If the peer doesn't do protocol negotiation, we must
661 * default to RDSv3.0 */
662 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
663 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
664
665 ret = rds_ib_setup_qp(conn);
666 if (ret) {
667 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
668 goto out;
669 }
670
671 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
672 UINT_MAX, UINT_MAX);
673 ret = rdma_connect(cm_id, &conn_param);
674 if (ret)
675 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
676
677out:
678 /* Beware - returning non-zero tells the rdma_cm to destroy
679 * the cm_id. We should certainly not do it as long as we still
680 * "own" the cm_id. */
681 if (ret) {
682 if (ic->i_cm_id == cm_id)
683 ret = 0;
684 }
685 return ret;
686}
687
688int rds_ib_conn_path_connect(struct rds_conn_path *cp)
689{
690 struct rds_connection *conn = cp->cp_conn;
691 struct rds_ib_connection *ic = conn->c_transport_data;
692 struct sockaddr_in src, dest;
693 int ret;
694
695 /* XXX I wonder what affect the port space has */
696 /* delegate cm event handler to rdma_transport */
697 ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn,
698 RDMA_PS_TCP, IB_QPT_RC);
699 if (IS_ERR(ic->i_cm_id)) {
700 ret = PTR_ERR(ic->i_cm_id);
701 ic->i_cm_id = NULL;
702 rdsdebug("rdma_create_id() failed: %d\n", ret);
703 goto out;
704 }
705
706 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
707
708 src.sin_family = AF_INET;
709 src.sin_addr.s_addr = (__force u32)conn->c_laddr;
710 src.sin_port = (__force u16)htons(0);
711
712 dest.sin_family = AF_INET;
713 dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
714 dest.sin_port = (__force u16)htons(RDS_PORT);
715
716 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
717 (struct sockaddr *)&dest,
718 RDS_RDMA_RESOLVE_TIMEOUT_MS);
719 if (ret) {
720 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
721 ret);
722 rdma_destroy_id(ic->i_cm_id);
723 ic->i_cm_id = NULL;
724 }
725
726out:
727 return ret;
728}
729
730/*
731 * This is so careful about only cleaning up resources that were built up
732 * so that it can be called at any point during startup. In fact it
733 * can be called multiple times for a given connection.
734 */
735void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
736{
737 struct rds_connection *conn = cp->cp_conn;
738 struct rds_ib_connection *ic = conn->c_transport_data;
739 int err = 0;
740
741 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
742 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
743 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
744
745 if (ic->i_cm_id) {
746 struct ib_device *dev = ic->i_cm_id->device;
747
748 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
749 err = rdma_disconnect(ic->i_cm_id);
750 if (err) {
751 /* Actually this may happen quite frequently, when
752 * an outgoing connect raced with an incoming connect.
753 */
754 rdsdebug("failed to disconnect, cm: %p err %d\n",
755 ic->i_cm_id, err);
756 }
757
758 /*
759 * We want to wait for tx and rx completion to finish
760 * before we tear down the connection, but we have to be
761 * careful not to get stuck waiting on a send ring that
762 * only has unsignaled sends in it. We've shutdown new
763 * sends before getting here so by waiting for signaled
764 * sends to complete we're ensured that there will be no
765 * more tx processing.
766 */
767 wait_event(rds_ib_ring_empty_wait,
768 rds_ib_ring_empty(&ic->i_recv_ring) &&
769 (atomic_read(&ic->i_signaled_sends) == 0) &&
770 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
771 tasklet_kill(&ic->i_send_tasklet);
772 tasklet_kill(&ic->i_recv_tasklet);
773
774 /* first destroy the ib state that generates callbacks */
775 if (ic->i_cm_id->qp)
776 rdma_destroy_qp(ic->i_cm_id);
777 if (ic->i_send_cq)
778 ib_destroy_cq(ic->i_send_cq);
779 if (ic->i_recv_cq)
780 ib_destroy_cq(ic->i_recv_cq);
781
782 /* then free the resources that ib callbacks use */
783 if (ic->i_send_hdrs)
784 ib_dma_free_coherent(dev,
785 ic->i_send_ring.w_nr *
786 sizeof(struct rds_header),
787 ic->i_send_hdrs,
788 ic->i_send_hdrs_dma);
789
790 if (ic->i_recv_hdrs)
791 ib_dma_free_coherent(dev,
792 ic->i_recv_ring.w_nr *
793 sizeof(struct rds_header),
794 ic->i_recv_hdrs,
795 ic->i_recv_hdrs_dma);
796
797 if (ic->i_ack)
798 ib_dma_free_coherent(dev, sizeof(struct rds_header),
799 ic->i_ack, ic->i_ack_dma);
800
801 if (ic->i_sends)
802 rds_ib_send_clear_ring(ic);
803 if (ic->i_recvs)
804 rds_ib_recv_clear_ring(ic);
805
806 rdma_destroy_id(ic->i_cm_id);
807
808 /*
809 * Move connection back to the nodev list.
810 */
811 if (ic->rds_ibdev)
812 rds_ib_remove_conn(ic->rds_ibdev, conn);
813
814 ic->i_cm_id = NULL;
815 ic->i_pd = NULL;
816 ic->i_send_cq = NULL;
817 ic->i_recv_cq = NULL;
818 ic->i_send_hdrs = NULL;
819 ic->i_recv_hdrs = NULL;
820 ic->i_ack = NULL;
821 }
822 BUG_ON(ic->rds_ibdev);
823
824 /* Clear pending transmit */
825 if (ic->i_data_op) {
826 struct rds_message *rm;
827
828 rm = container_of(ic->i_data_op, struct rds_message, data);
829 rds_message_put(rm);
830 ic->i_data_op = NULL;
831 }
832
833 /* Clear the ACK state */
834 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
835#ifdef KERNEL_HAS_ATOMIC64
836 atomic64_set(&ic->i_ack_next, 0);
837#else
838 ic->i_ack_next = 0;
839#endif
840 ic->i_ack_recv = 0;
841
842 /* Clear flow control state */
843 ic->i_flowctl = 0;
844 atomic_set(&ic->i_credits, 0);
845
846 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
847 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
848
849 if (ic->i_ibinc) {
850 rds_inc_put(&ic->i_ibinc->ii_inc);
851 ic->i_ibinc = NULL;
852 }
853
854 vfree(ic->i_sends);
855 ic->i_sends = NULL;
856 vfree(ic->i_recvs);
857 ic->i_recvs = NULL;
858}
859
860int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
861{
862 struct rds_ib_connection *ic;
863 unsigned long flags;
864 int ret;
865
866 /* XXX too lazy? */
867 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
868 if (!ic)
869 return -ENOMEM;
870
871 ret = rds_ib_recv_alloc_caches(ic);
872 if (ret) {
873 kfree(ic);
874 return ret;
875 }
876
877 INIT_LIST_HEAD(&ic->ib_node);
878 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
879 (unsigned long)ic);
880 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
881 (unsigned long)ic);
882 mutex_init(&ic->i_recv_mutex);
883#ifndef KERNEL_HAS_ATOMIC64
884 spin_lock_init(&ic->i_ack_lock);
885#endif
886 atomic_set(&ic->i_signaled_sends, 0);
887
888 /*
889 * rds_ib_conn_shutdown() waits for these to be emptied so they
890 * must be initialized before it can be called.
891 */
892 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
893 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
894
895 ic->conn = conn;
896 conn->c_transport_data = ic;
897
898 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
899 list_add_tail(&ic->ib_node, &ib_nodev_conns);
900 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
901
902
903 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
904 return 0;
905}
906
907/*
908 * Free a connection. Connection must be shut down and not set for reconnect.
909 */
910void rds_ib_conn_free(void *arg)
911{
912 struct rds_ib_connection *ic = arg;
913 spinlock_t *lock_ptr;
914
915 rdsdebug("ic %p\n", ic);
916
917 /*
918 * Conn is either on a dev's list or on the nodev list.
919 * A race with shutdown() or connect() would cause problems
920 * (since rds_ibdev would change) but that should never happen.
921 */
922 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
923
924 spin_lock_irq(lock_ptr);
925 list_del(&ic->ib_node);
926 spin_unlock_irq(lock_ptr);
927
928 rds_ib_recv_free_caches(ic);
929
930 kfree(ic);
931}
932
933
934/*
935 * An error occurred on the connection
936 */
937void
938__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
939{
940 va_list ap;
941
942 rds_conn_drop(conn);
943
944 va_start(ap, fmt);
945 vprintk(fmt, ap);
946 va_end(ap);
947}