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 = vmalloc_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 memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
386
387 ic->i_recvs = vmalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
388 ibdev_to_node(dev));
389 if (!ic->i_recvs) {
390 ret = -ENOMEM;
391 rdsdebug("recv allocation failed\n");
392 goto out;
393 }
394 memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
395
396 rds_ib_recv_init_ack(ic);
397
398 rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
399 ic->i_send_cq, ic->i_recv_cq);
400
401out:
402 rds_ib_dev_put(rds_ibdev);
403 return ret;
404}
405
406static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
407{
408 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
409 u16 common;
410 u32 version = 0;
411
412 /*
413 * rdma_cm private data is odd - when there is any private data in the
414 * request, we will be given a pretty large buffer without telling us the
415 * original size. The only way to tell the difference is by looking at
416 * the contents, which are initialized to zero.
417 * If the protocol version fields aren't set, this is a connection attempt
418 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
419 * We really should have changed this for OFED 1.3 :-(
420 */
421
422 /* Be paranoid. RDS always has privdata */
423 if (!event->param.conn.private_data_len) {
424 printk(KERN_NOTICE "RDS incoming connection has no private data, "
425 "rejecting\n");
426 return 0;
427 }
428
429 /* Even if len is crap *now* I still want to check it. -ASG */
430 if (event->param.conn.private_data_len < sizeof (*dp) ||
431 dp->dp_protocol_major == 0)
432 return RDS_PROTOCOL_3_0;
433
434 common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
435 if (dp->dp_protocol_major == 3 && common) {
436 version = RDS_PROTOCOL_3_0;
437 while ((common >>= 1) != 0)
438 version++;
439 }
440 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using "
441 "incompatible protocol version %u.%u\n",
442 &dp->dp_saddr,
443 dp->dp_protocol_major,
444 dp->dp_protocol_minor);
445 return version;
446}
447
448int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
449 struct rdma_cm_event *event)
450{
451 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
452 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
453 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
454 struct rds_ib_connect_private dp_rep;
455 struct rds_connection *conn = NULL;
456 struct rds_ib_connection *ic = NULL;
457 struct rdma_conn_param conn_param;
458 u32 version;
459 int err = 1, destroy = 1;
460
461 /* Check whether the remote protocol version matches ours. */
462 version = rds_ib_protocol_compatible(event);
463 if (!version)
464 goto out;
465
466 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
467 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
468 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
469 (unsigned long long)be64_to_cpu(lguid),
470 (unsigned long long)be64_to_cpu(fguid));
471
472 conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
473 GFP_KERNEL);
474 if (IS_ERR(conn)) {
475 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
476 conn = NULL;
477 goto out;
478 }
479
480 /*
481 * The connection request may occur while the
482 * previous connection exist, e.g. in case of failover.
483 * But as connections may be initiated simultaneously
484 * by both hosts, we have a random backoff mechanism -
485 * see the comment above rds_queue_reconnect()
486 */
487 mutex_lock(&conn->c_cm_lock);
488 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
489 if (rds_conn_state(conn) == RDS_CONN_UP) {
490 rdsdebug("incoming connect while connecting\n");
491 rds_conn_drop(conn);
492 rds_ib_stats_inc(s_ib_listen_closed_stale);
493 } else
494 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
495 /* Wait and see - our connect may still be succeeding */
496 rds_ib_stats_inc(s_ib_connect_raced);
497 }
498 goto out;
499 }
500
501 ic = conn->c_transport_data;
502
503 rds_ib_set_protocol(conn, version);
504 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
505
506 /* If the peer gave us the last packet it saw, process this as if
507 * we had received a regular ACK. */
508 if (dp->dp_ack_seq)
509 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
510
511 BUG_ON(cm_id->context);
512 BUG_ON(ic->i_cm_id);
513
514 ic->i_cm_id = cm_id;
515 cm_id->context = conn;
516
517 /* We got halfway through setting up the ib_connection, if we
518 * fail now, we have to take the long route out of this mess. */
519 destroy = 0;
520
521 err = rds_ib_setup_qp(conn);
522 if (err) {
523 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
524 goto out;
525 }
526
527 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
528 event->param.conn.responder_resources,
529 event->param.conn.initiator_depth);
530
531 /* rdma_accept() calls rdma_reject() internally if it fails */
532 err = rdma_accept(cm_id, &conn_param);
533 if (err)
534 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
535
536out:
537 if (conn)
538 mutex_unlock(&conn->c_cm_lock);
539 if (err)
540 rdma_reject(cm_id, NULL, 0);
541 return destroy;
542}
543
544
545int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
546{
547 struct rds_connection *conn = cm_id->context;
548 struct rds_ib_connection *ic = conn->c_transport_data;
549 struct rdma_conn_param conn_param;
550 struct rds_ib_connect_private dp;
551 int ret;
552
553 /* If the peer doesn't do protocol negotiation, we must
554 * default to RDSv3.0 */
555 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
556 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
557
558 ret = rds_ib_setup_qp(conn);
559 if (ret) {
560 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
561 goto out;
562 }
563
564 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
565 UINT_MAX, UINT_MAX);
566 ret = rdma_connect(cm_id, &conn_param);
567 if (ret)
568 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
569
570out:
571 /* Beware - returning non-zero tells the rdma_cm to destroy
572 * the cm_id. We should certainly not do it as long as we still
573 * "own" the cm_id. */
574 if (ret) {
575 if (ic->i_cm_id == cm_id)
576 ret = 0;
577 }
578 return ret;
579}
580
581int rds_ib_conn_connect(struct rds_connection *conn)
582{
583 struct rds_ib_connection *ic = conn->c_transport_data;
584 struct sockaddr_in src, dest;
585 int ret;
586
587 /* XXX I wonder what affect the port space has */
588 /* delegate cm event handler to rdma_transport */
589 ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
590 RDMA_PS_TCP, IB_QPT_RC);
591 if (IS_ERR(ic->i_cm_id)) {
592 ret = PTR_ERR(ic->i_cm_id);
593 ic->i_cm_id = NULL;
594 rdsdebug("rdma_create_id() failed: %d\n", ret);
595 goto out;
596 }
597
598 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
599
600 src.sin_family = AF_INET;
601 src.sin_addr.s_addr = (__force u32)conn->c_laddr;
602 src.sin_port = (__force u16)htons(0);
603
604 dest.sin_family = AF_INET;
605 dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
606 dest.sin_port = (__force u16)htons(RDS_PORT);
607
608 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
609 (struct sockaddr *)&dest,
610 RDS_RDMA_RESOLVE_TIMEOUT_MS);
611 if (ret) {
612 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
613 ret);
614 rdma_destroy_id(ic->i_cm_id);
615 ic->i_cm_id = NULL;
616 }
617
618out:
619 return ret;
620}
621
622/*
623 * This is so careful about only cleaning up resources that were built up
624 * so that it can be called at any point during startup. In fact it
625 * can be called multiple times for a given connection.
626 */
627void rds_ib_conn_shutdown(struct rds_connection *conn)
628{
629 struct rds_ib_connection *ic = conn->c_transport_data;
630 int err = 0;
631
632 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
633 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
634 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
635
636 if (ic->i_cm_id) {
637 struct ib_device *dev = ic->i_cm_id->device;
638
639 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
640 err = rdma_disconnect(ic->i_cm_id);
641 if (err) {
642 /* Actually this may happen quite frequently, when
643 * an outgoing connect raced with an incoming connect.
644 */
645 rdsdebug("failed to disconnect, cm: %p err %d\n",
646 ic->i_cm_id, err);
647 }
648
649 /*
650 * We want to wait for tx and rx completion to finish
651 * before we tear down the connection, but we have to be
652 * careful not to get stuck waiting on a send ring that
653 * only has unsignaled sends in it. We've shutdown new
654 * sends before getting here so by waiting for signaled
655 * sends to complete we're ensured that there will be no
656 * more tx processing.
657 */
658 wait_event(rds_ib_ring_empty_wait,
659 rds_ib_ring_empty(&ic->i_recv_ring) &&
660 (atomic_read(&ic->i_signaled_sends) == 0));
661 tasklet_kill(&ic->i_recv_tasklet);
662
663 if (ic->i_send_hdrs)
664 ib_dma_free_coherent(dev,
665 ic->i_send_ring.w_nr *
666 sizeof(struct rds_header),
667 ic->i_send_hdrs,
668 ic->i_send_hdrs_dma);
669
670 if (ic->i_recv_hdrs)
671 ib_dma_free_coherent(dev,
672 ic->i_recv_ring.w_nr *
673 sizeof(struct rds_header),
674 ic->i_recv_hdrs,
675 ic->i_recv_hdrs_dma);
676
677 if (ic->i_ack)
678 ib_dma_free_coherent(dev, sizeof(struct rds_header),
679 ic->i_ack, ic->i_ack_dma);
680
681 if (ic->i_sends)
682 rds_ib_send_clear_ring(ic);
683 if (ic->i_recvs)
684 rds_ib_recv_clear_ring(ic);
685
686 if (ic->i_cm_id->qp)
687 rdma_destroy_qp(ic->i_cm_id);
688 if (ic->i_send_cq)
689 ib_destroy_cq(ic->i_send_cq);
690 if (ic->i_recv_cq)
691 ib_destroy_cq(ic->i_recv_cq);
692 rdma_destroy_id(ic->i_cm_id);
693
694 /*
695 * Move connection back to the nodev list.
696 */
697 if (ic->rds_ibdev)
698 rds_ib_remove_conn(ic->rds_ibdev, conn);
699
700 ic->i_cm_id = NULL;
701 ic->i_pd = NULL;
702 ic->i_mr = NULL;
703 ic->i_send_cq = NULL;
704 ic->i_recv_cq = NULL;
705 ic->i_send_hdrs = NULL;
706 ic->i_recv_hdrs = NULL;
707 ic->i_ack = NULL;
708 }
709 BUG_ON(ic->rds_ibdev);
710
711 /* Clear pending transmit */
712 if (ic->i_data_op) {
713 struct rds_message *rm;
714
715 rm = container_of(ic->i_data_op, struct rds_message, data);
716 rds_message_put(rm);
717 ic->i_data_op = NULL;
718 }
719
720 /* Clear the ACK state */
721 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
722#ifdef KERNEL_HAS_ATOMIC64
723 atomic64_set(&ic->i_ack_next, 0);
724#else
725 ic->i_ack_next = 0;
726#endif
727 ic->i_ack_recv = 0;
728
729 /* Clear flow control state */
730 ic->i_flowctl = 0;
731 atomic_set(&ic->i_credits, 0);
732
733 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
734 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
735
736 if (ic->i_ibinc) {
737 rds_inc_put(&ic->i_ibinc->ii_inc);
738 ic->i_ibinc = NULL;
739 }
740
741 vfree(ic->i_sends);
742 ic->i_sends = NULL;
743 vfree(ic->i_recvs);
744 ic->i_recvs = NULL;
745}
746
747int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
748{
749 struct rds_ib_connection *ic;
750 unsigned long flags;
751 int ret;
752
753 /* XXX too lazy? */
754 ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
755 if (!ic)
756 return -ENOMEM;
757
758 ret = rds_ib_recv_alloc_caches(ic);
759 if (ret) {
760 kfree(ic);
761 return ret;
762 }
763
764 INIT_LIST_HEAD(&ic->ib_node);
765 tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
766 (unsigned long) ic);
767 mutex_init(&ic->i_recv_mutex);
768#ifndef KERNEL_HAS_ATOMIC64
769 spin_lock_init(&ic->i_ack_lock);
770#endif
771 atomic_set(&ic->i_signaled_sends, 0);
772
773 /*
774 * rds_ib_conn_shutdown() waits for these to be emptied so they
775 * must be initialized before it can be called.
776 */
777 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
778 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
779
780 ic->conn = conn;
781 conn->c_transport_data = ic;
782
783 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
784 list_add_tail(&ic->ib_node, &ib_nodev_conns);
785 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
786
787
788 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
789 return 0;
790}
791
792/*
793 * Free a connection. Connection must be shut down and not set for reconnect.
794 */
795void rds_ib_conn_free(void *arg)
796{
797 struct rds_ib_connection *ic = arg;
798 spinlock_t *lock_ptr;
799
800 rdsdebug("ic %p\n", ic);
801
802 /*
803 * Conn is either on a dev's list or on the nodev list.
804 * A race with shutdown() or connect() would cause problems
805 * (since rds_ibdev would change) but that should never happen.
806 */
807 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
808
809 spin_lock_irq(lock_ptr);
810 list_del(&ic->ib_node);
811 spin_unlock_irq(lock_ptr);
812
813 rds_ib_recv_free_caches(ic);
814
815 kfree(ic);
816}
817
818
819/*
820 * An error occurred on the connection
821 */
822void
823__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
824{
825 va_list ap;
826
827 rds_conn_drop(conn);
828
829 va_start(ap, fmt);
830 vprintk(fmt, ap);
831 va_end(ap);
832}
1/*
2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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#include <net/addrconf.h>
39
40#include "rds_single_path.h"
41#include "rds.h"
42#include "ib.h"
43#include "ib_mr.h"
44
45/*
46 * Set the selected protocol version
47 */
48static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
49{
50 conn->c_version = version;
51}
52
53/*
54 * Set up flow control
55 */
56static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
57{
58 struct rds_ib_connection *ic = conn->c_transport_data;
59
60 if (rds_ib_sysctl_flow_control && credits != 0) {
61 /* We're doing flow control */
62 ic->i_flowctl = 1;
63 rds_ib_send_add_credits(conn, credits);
64 } else {
65 ic->i_flowctl = 0;
66 }
67}
68
69/*
70 * Tune RNR behavior. Without flow control, we use a rather
71 * low timeout, but not the absolute minimum - this should
72 * be tunable.
73 *
74 * We already set the RNR retry count to 7 (which is the
75 * smallest infinite number :-) above.
76 * If flow control is off, we want to change this back to 0
77 * so that we learn quickly when our credit accounting is
78 * buggy.
79 *
80 * Caller passes in a qp_attr pointer - don't waste stack spacv
81 * by allocation this twice.
82 */
83static void
84rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
85{
86 int ret;
87
88 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
89 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
90 if (ret)
91 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
92}
93
94/*
95 * Connection established.
96 * We get here for both outgoing and incoming connection.
97 */
98void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
99{
100 struct rds_ib_connection *ic = conn->c_transport_data;
101 const union rds_ib_conn_priv *dp = NULL;
102 struct ib_qp_attr qp_attr;
103 __be64 ack_seq = 0;
104 __be32 credit = 0;
105 u8 major = 0;
106 u8 minor = 0;
107 int err;
108
109 dp = event->param.conn.private_data;
110 if (conn->c_isv6) {
111 if (event->param.conn.private_data_len >=
112 sizeof(struct rds6_ib_connect_private)) {
113 major = dp->ricp_v6.dp_protocol_major;
114 minor = dp->ricp_v6.dp_protocol_minor;
115 credit = dp->ricp_v6.dp_credit;
116 /* dp structure start is not guaranteed to be 8 bytes
117 * aligned. Since dp_ack_seq is 64-bit extended load
118 * operations can be used so go through get_unaligned
119 * to avoid unaligned errors.
120 */
121 ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
122 }
123 } else if (event->param.conn.private_data_len >=
124 sizeof(struct rds_ib_connect_private)) {
125 major = dp->ricp_v4.dp_protocol_major;
126 minor = dp->ricp_v4.dp_protocol_minor;
127 credit = dp->ricp_v4.dp_credit;
128 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
129 }
130
131 /* make sure it isn't empty data */
132 if (major) {
133 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
134 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
135 }
136
137 if (conn->c_version < RDS_PROTOCOL_VERSION) {
138 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
139 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
140 &conn->c_laddr, &conn->c_faddr,
141 RDS_PROTOCOL_MAJOR(conn->c_version),
142 RDS_PROTOCOL_MINOR(conn->c_version));
143 rds_conn_destroy(conn);
144 return;
145 }
146 }
147
148 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
149 ic->i_active_side ? "Active" : "Passive",
150 &conn->c_laddr, &conn->c_faddr, conn->c_tos,
151 RDS_PROTOCOL_MAJOR(conn->c_version),
152 RDS_PROTOCOL_MINOR(conn->c_version),
153 ic->i_flowctl ? ", flow control" : "");
154
155 /* receive sl from the peer */
156 ic->i_sl = ic->i_cm_id->route.path_rec->sl;
157
158 atomic_set(&ic->i_cq_quiesce, 0);
159
160 /* Init rings and fill recv. this needs to wait until protocol
161 * negotiation is complete, since ring layout is different
162 * from 3.1 to 4.1.
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, GFP_KERNEL);
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) {
187 if (ack_seq)
188 rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
189 NULL);
190 }
191
192 conn->c_proposed_version = conn->c_version;
193 rds_connect_complete(conn);
194}
195
196static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
197 struct rdma_conn_param *conn_param,
198 union rds_ib_conn_priv *dp,
199 u32 protocol_version,
200 u32 max_responder_resources,
201 u32 max_initiator_depth,
202 bool isv6)
203{
204 struct rds_ib_connection *ic = conn->c_transport_data;
205 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
206
207 memset(conn_param, 0, sizeof(struct rdma_conn_param));
208
209 conn_param->responder_resources =
210 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
211 conn_param->initiator_depth =
212 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
213 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
214 conn_param->rnr_retry_count = 7;
215
216 if (dp) {
217 memset(dp, 0, sizeof(*dp));
218 if (isv6) {
219 dp->ricp_v6.dp_saddr = conn->c_laddr;
220 dp->ricp_v6.dp_daddr = conn->c_faddr;
221 dp->ricp_v6.dp_protocol_major =
222 RDS_PROTOCOL_MAJOR(protocol_version);
223 dp->ricp_v6.dp_protocol_minor =
224 RDS_PROTOCOL_MINOR(protocol_version);
225 dp->ricp_v6.dp_protocol_minor_mask =
226 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
227 dp->ricp_v6.dp_ack_seq =
228 cpu_to_be64(rds_ib_piggyb_ack(ic));
229 dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
230
231 conn_param->private_data = &dp->ricp_v6;
232 conn_param->private_data_len = sizeof(dp->ricp_v6);
233 } else {
234 dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
235 dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
236 dp->ricp_v4.dp_protocol_major =
237 RDS_PROTOCOL_MAJOR(protocol_version);
238 dp->ricp_v4.dp_protocol_minor =
239 RDS_PROTOCOL_MINOR(protocol_version);
240 dp->ricp_v4.dp_protocol_minor_mask =
241 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
242 dp->ricp_v4.dp_ack_seq =
243 cpu_to_be64(rds_ib_piggyb_ack(ic));
244 dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
245
246 conn_param->private_data = &dp->ricp_v4;
247 conn_param->private_data_len = sizeof(dp->ricp_v4);
248 }
249
250 /* Advertise flow control */
251 if (ic->i_flowctl) {
252 unsigned int credits;
253
254 credits = IB_GET_POST_CREDITS
255 (atomic_read(&ic->i_credits));
256 if (isv6)
257 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
258 else
259 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
260 atomic_sub(IB_SET_POST_CREDITS(credits),
261 &ic->i_credits);
262 }
263 }
264}
265
266static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
267{
268 rdsdebug("event %u (%s) data %p\n",
269 event->event, ib_event_msg(event->event), data);
270}
271
272/* Plucking the oldest entry from the ring can be done concurrently with
273 * the thread refilling the ring. Each ring operation is protected by
274 * spinlocks and the transient state of refilling doesn't change the
275 * recording of which entry is oldest.
276 *
277 * This relies on IB only calling one cq comp_handler for each cq so that
278 * there will only be one caller of rds_recv_incoming() per RDS connection.
279 */
280static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
281{
282 struct rds_connection *conn = context;
283 struct rds_ib_connection *ic = conn->c_transport_data;
284
285 rdsdebug("conn %p cq %p\n", conn, cq);
286
287 rds_ib_stats_inc(s_ib_evt_handler_call);
288
289 tasklet_schedule(&ic->i_recv_tasklet);
290}
291
292static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
293 struct ib_wc *wcs)
294{
295 int nr, i;
296 struct ib_wc *wc;
297
298 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
299 for (i = 0; i < nr; i++) {
300 wc = wcs + i;
301 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
302 (unsigned long long)wc->wr_id, wc->status,
303 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
304
305 if (wc->wr_id <= ic->i_send_ring.w_nr ||
306 wc->wr_id == RDS_IB_ACK_WR_ID)
307 rds_ib_send_cqe_handler(ic, wc);
308 else
309 rds_ib_mr_cqe_handler(ic, wc);
310
311 }
312 }
313}
314
315static void rds_ib_tasklet_fn_send(unsigned long data)
316{
317 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
318 struct rds_connection *conn = ic->conn;
319
320 rds_ib_stats_inc(s_ib_tasklet_call);
321
322 /* if cq has been already reaped, ignore incoming cq event */
323 if (atomic_read(&ic->i_cq_quiesce))
324 return;
325
326 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
327 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
328 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
329
330 if (rds_conn_up(conn) &&
331 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
332 test_bit(0, &conn->c_map_queued)))
333 rds_send_xmit(&ic->conn->c_path[0]);
334}
335
336static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
337 struct ib_wc *wcs,
338 struct rds_ib_ack_state *ack_state)
339{
340 int nr, i;
341 struct ib_wc *wc;
342
343 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
344 for (i = 0; i < nr; i++) {
345 wc = wcs + i;
346 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
347 (unsigned long long)wc->wr_id, wc->status,
348 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
349
350 rds_ib_recv_cqe_handler(ic, wc, ack_state);
351 }
352 }
353}
354
355static void rds_ib_tasklet_fn_recv(unsigned long data)
356{
357 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
358 struct rds_connection *conn = ic->conn;
359 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
360 struct rds_ib_ack_state state;
361
362 if (!rds_ibdev)
363 rds_conn_drop(conn);
364
365 rds_ib_stats_inc(s_ib_tasklet_call);
366
367 /* if cq has been already reaped, ignore incoming cq event */
368 if (atomic_read(&ic->i_cq_quiesce))
369 return;
370
371 memset(&state, 0, sizeof(state));
372 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
373 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
374 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
375
376 if (state.ack_next_valid)
377 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
378 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
379 rds_send_drop_acked(conn, state.ack_recv, NULL);
380 ic->i_ack_recv = state.ack_recv;
381 }
382
383 if (rds_conn_up(conn))
384 rds_ib_attempt_ack(ic);
385}
386
387static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
388{
389 struct rds_connection *conn = data;
390 struct rds_ib_connection *ic = conn->c_transport_data;
391
392 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
393 ib_event_msg(event->event));
394
395 switch (event->event) {
396 case IB_EVENT_COMM_EST:
397 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
398 break;
399 default:
400 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
401 event->event, ib_event_msg(event->event),
402 &conn->c_laddr, &conn->c_faddr);
403 rds_conn_drop(conn);
404 break;
405 }
406}
407
408static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
409{
410 struct rds_connection *conn = context;
411 struct rds_ib_connection *ic = conn->c_transport_data;
412
413 rdsdebug("conn %p cq %p\n", conn, cq);
414
415 rds_ib_stats_inc(s_ib_evt_handler_call);
416
417 tasklet_schedule(&ic->i_send_tasklet);
418}
419
420static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
421{
422 int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
423 int index = rds_ibdev->dev->num_comp_vectors - 1;
424 int i;
425
426 for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
427 if (rds_ibdev->vector_load[i] < min) {
428 index = i;
429 min = rds_ibdev->vector_load[i];
430 }
431 }
432
433 rds_ibdev->vector_load[index]++;
434 return index;
435}
436
437static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
438{
439 rds_ibdev->vector_load[index]--;
440}
441
442/*
443 * This needs to be very careful to not leave IS_ERR pointers around for
444 * cleanup to trip over.
445 */
446static int rds_ib_setup_qp(struct rds_connection *conn)
447{
448 struct rds_ib_connection *ic = conn->c_transport_data;
449 struct ib_device *dev = ic->i_cm_id->device;
450 struct ib_qp_init_attr attr;
451 struct ib_cq_init_attr cq_attr = {};
452 struct rds_ib_device *rds_ibdev;
453 unsigned long max_wrs;
454 int ret, fr_queue_space;
455
456 /*
457 * It's normal to see a null device if an incoming connection races
458 * with device removal, so we don't print a warning.
459 */
460 rds_ibdev = rds_ib_get_client_data(dev);
461 if (!rds_ibdev)
462 return -EOPNOTSUPP;
463
464 /* The fr_queue_space is currently set to 512, to add extra space on
465 * completion queue and send queue. This extra space is used for FRMR
466 * registration and invalidation work requests
467 */
468 fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
469
470 /* add the conn now so that connection establishment has the dev */
471 rds_ib_add_conn(rds_ibdev, conn);
472
473 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
474 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
475 if (ic->i_send_ring.w_nr != max_wrs)
476 rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
477
478 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
479 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
480 if (ic->i_recv_ring.w_nr != max_wrs)
481 rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
482
483 /* Protection domain and memory range */
484 ic->i_pd = rds_ibdev->pd;
485
486 ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
487 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
488 cq_attr.comp_vector = ic->i_scq_vector;
489 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
490 rds_ib_cq_event_handler, conn,
491 &cq_attr);
492 if (IS_ERR(ic->i_send_cq)) {
493 ret = PTR_ERR(ic->i_send_cq);
494 ic->i_send_cq = NULL;
495 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
496 rdsdebug("ib_create_cq send failed: %d\n", ret);
497 goto rds_ibdev_out;
498 }
499
500 ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
501 cq_attr.cqe = ic->i_recv_ring.w_nr;
502 cq_attr.comp_vector = ic->i_rcq_vector;
503 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
504 rds_ib_cq_event_handler, conn,
505 &cq_attr);
506 if (IS_ERR(ic->i_recv_cq)) {
507 ret = PTR_ERR(ic->i_recv_cq);
508 ic->i_recv_cq = NULL;
509 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
510 rdsdebug("ib_create_cq recv failed: %d\n", ret);
511 goto send_cq_out;
512 }
513
514 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
515 if (ret) {
516 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
517 goto recv_cq_out;
518 }
519
520 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
521 if (ret) {
522 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
523 goto recv_cq_out;
524 }
525
526 /* XXX negotiate max send/recv with remote? */
527 memset(&attr, 0, sizeof(attr));
528 attr.event_handler = rds_ib_qp_event_handler;
529 attr.qp_context = conn;
530 /* + 1 to allow for the single ack message */
531 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
532 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
533 attr.cap.max_send_sge = rds_ibdev->max_sge;
534 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
535 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
536 attr.qp_type = IB_QPT_RC;
537 attr.send_cq = ic->i_send_cq;
538 attr.recv_cq = ic->i_recv_cq;
539
540 /*
541 * XXX this can fail if max_*_wr is too large? Are we supposed
542 * to back off until we get a value that the hardware can support?
543 */
544 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
545 if (ret) {
546 rdsdebug("rdma_create_qp failed: %d\n", ret);
547 goto recv_cq_out;
548 }
549
550 ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
551 ic->i_send_ring.w_nr *
552 sizeof(struct rds_header),
553 &ic->i_send_hdrs_dma, GFP_KERNEL);
554 if (!ic->i_send_hdrs) {
555 ret = -ENOMEM;
556 rdsdebug("ib_dma_alloc_coherent send failed\n");
557 goto qp_out;
558 }
559
560 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
561 ic->i_recv_ring.w_nr *
562 sizeof(struct rds_header),
563 &ic->i_recv_hdrs_dma, GFP_KERNEL);
564 if (!ic->i_recv_hdrs) {
565 ret = -ENOMEM;
566 rdsdebug("ib_dma_alloc_coherent recv failed\n");
567 goto send_hdrs_dma_out;
568 }
569
570 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
571 &ic->i_ack_dma, GFP_KERNEL);
572 if (!ic->i_ack) {
573 ret = -ENOMEM;
574 rdsdebug("ib_dma_alloc_coherent ack failed\n");
575 goto recv_hdrs_dma_out;
576 }
577
578 ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
579 ic->i_send_ring.w_nr),
580 ibdev_to_node(dev));
581 if (!ic->i_sends) {
582 ret = -ENOMEM;
583 rdsdebug("send allocation failed\n");
584 goto ack_dma_out;
585 }
586
587 ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
588 ic->i_recv_ring.w_nr),
589 ibdev_to_node(dev));
590 if (!ic->i_recvs) {
591 ret = -ENOMEM;
592 rdsdebug("recv allocation failed\n");
593 goto sends_out;
594 }
595
596 rds_ib_recv_init_ack(ic);
597
598 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
599 ic->i_send_cq, ic->i_recv_cq);
600
601 goto out;
602
603sends_out:
604 vfree(ic->i_sends);
605ack_dma_out:
606 ib_dma_free_coherent(dev, sizeof(struct rds_header),
607 ic->i_ack, ic->i_ack_dma);
608recv_hdrs_dma_out:
609 ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr *
610 sizeof(struct rds_header),
611 ic->i_recv_hdrs, ic->i_recv_hdrs_dma);
612send_hdrs_dma_out:
613 ib_dma_free_coherent(dev, ic->i_send_ring.w_nr *
614 sizeof(struct rds_header),
615 ic->i_send_hdrs, ic->i_send_hdrs_dma);
616qp_out:
617 rdma_destroy_qp(ic->i_cm_id);
618recv_cq_out:
619 ib_destroy_cq(ic->i_recv_cq);
620 ic->i_recv_cq = NULL;
621send_cq_out:
622 ib_destroy_cq(ic->i_send_cq);
623 ic->i_send_cq = NULL;
624rds_ibdev_out:
625 rds_ib_remove_conn(rds_ibdev, conn);
626out:
627 rds_ib_dev_put(rds_ibdev);
628
629 return ret;
630}
631
632static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
633{
634 const union rds_ib_conn_priv *dp = event->param.conn.private_data;
635 u8 data_len, major, minor;
636 u32 version = 0;
637 __be16 mask;
638 u16 common;
639
640 /*
641 * rdma_cm private data is odd - when there is any private data in the
642 * request, we will be given a pretty large buffer without telling us the
643 * original size. The only way to tell the difference is by looking at
644 * the contents, which are initialized to zero.
645 * If the protocol version fields aren't set, this is a connection attempt
646 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
647 * We really should have changed this for OFED 1.3 :-(
648 */
649
650 /* Be paranoid. RDS always has privdata */
651 if (!event->param.conn.private_data_len) {
652 printk(KERN_NOTICE "RDS incoming connection has no private data, "
653 "rejecting\n");
654 return 0;
655 }
656
657 if (isv6) {
658 data_len = sizeof(struct rds6_ib_connect_private);
659 major = dp->ricp_v6.dp_protocol_major;
660 minor = dp->ricp_v6.dp_protocol_minor;
661 mask = dp->ricp_v6.dp_protocol_minor_mask;
662 } else {
663 data_len = sizeof(struct rds_ib_connect_private);
664 major = dp->ricp_v4.dp_protocol_major;
665 minor = dp->ricp_v4.dp_protocol_minor;
666 mask = dp->ricp_v4.dp_protocol_minor_mask;
667 }
668
669 /* Even if len is crap *now* I still want to check it. -ASG */
670 if (event->param.conn.private_data_len < data_len || major == 0)
671 return RDS_PROTOCOL_4_0;
672
673 common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
674 if (major == 4 && common) {
675 version = RDS_PROTOCOL_4_0;
676 while ((common >>= 1) != 0)
677 version++;
678 } else if (RDS_PROTOCOL_COMPAT_VERSION ==
679 RDS_PROTOCOL(major, minor)) {
680 version = RDS_PROTOCOL_COMPAT_VERSION;
681 } else {
682 if (isv6)
683 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
684 &dp->ricp_v6.dp_saddr, major, minor);
685 else
686 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
687 &dp->ricp_v4.dp_saddr, major, minor);
688 }
689 return version;
690}
691
692#if IS_ENABLED(CONFIG_IPV6)
693/* Given an IPv6 address, find the net_device which hosts that address and
694 * return its index. This is used by the rds_ib_cm_handle_connect() code to
695 * find the interface index of where an incoming request comes from when
696 * the request is using a link local address.
697 *
698 * Note one problem in this search. It is possible that two interfaces have
699 * the same link local address. Unfortunately, this cannot be solved unless
700 * the underlying layer gives us the interface which an incoming RDMA connect
701 * request comes from.
702 */
703static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
704{
705 struct net_device *dev;
706 int idx = 0;
707
708 rcu_read_lock();
709 for_each_netdev_rcu(net, dev) {
710 if (ipv6_chk_addr(net, addr, dev, 1)) {
711 idx = dev->ifindex;
712 break;
713 }
714 }
715 rcu_read_unlock();
716
717 return idx;
718}
719#endif
720
721int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
722 struct rdma_cm_event *event, bool isv6)
723{
724 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
725 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
726 const struct rds_ib_conn_priv_cmn *dp_cmn;
727 struct rds_connection *conn = NULL;
728 struct rds_ib_connection *ic = NULL;
729 struct rdma_conn_param conn_param;
730 const union rds_ib_conn_priv *dp;
731 union rds_ib_conn_priv dp_rep;
732 struct in6_addr s_mapped_addr;
733 struct in6_addr d_mapped_addr;
734 const struct in6_addr *saddr6;
735 const struct in6_addr *daddr6;
736 int destroy = 1;
737 u32 ifindex = 0;
738 u32 version;
739 int err = 1;
740
741 /* Check whether the remote protocol version matches ours. */
742 version = rds_ib_protocol_compatible(event, isv6);
743 if (!version) {
744 err = RDS_RDMA_REJ_INCOMPAT;
745 goto out;
746 }
747
748 dp = event->param.conn.private_data;
749 if (isv6) {
750#if IS_ENABLED(CONFIG_IPV6)
751 dp_cmn = &dp->ricp_v6.dp_cmn;
752 saddr6 = &dp->ricp_v6.dp_saddr;
753 daddr6 = &dp->ricp_v6.dp_daddr;
754 /* If either address is link local, need to find the
755 * interface index in order to create a proper RDS
756 * connection.
757 */
758 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
759 /* Using init_net for now .. */
760 ifindex = __rds_find_ifindex(&init_net, daddr6);
761 /* No index found... Need to bail out. */
762 if (ifindex == 0) {
763 err = -EOPNOTSUPP;
764 goto out;
765 }
766 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
767 /* Use our address to find the correct index. */
768 ifindex = __rds_find_ifindex(&init_net, daddr6);
769 /* No index found... Need to bail out. */
770 if (ifindex == 0) {
771 err = -EOPNOTSUPP;
772 goto out;
773 }
774 }
775#else
776 err = -EOPNOTSUPP;
777 goto out;
778#endif
779 } else {
780 dp_cmn = &dp->ricp_v4.dp_cmn;
781 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
782 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
783 saddr6 = &s_mapped_addr;
784 daddr6 = &d_mapped_addr;
785 }
786
787 rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
788 saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
789 RDS_PROTOCOL_MINOR(version),
790 (unsigned long long)be64_to_cpu(lguid),
791 (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
792
793 /* RDS/IB is not currently netns aware, thus init_net */
794 conn = rds_conn_create(&init_net, daddr6, saddr6,
795 &rds_ib_transport, dp_cmn->ricpc_dp_toss,
796 GFP_KERNEL, ifindex);
797 if (IS_ERR(conn)) {
798 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
799 conn = NULL;
800 goto out;
801 }
802
803 /*
804 * The connection request may occur while the
805 * previous connection exist, e.g. in case of failover.
806 * But as connections may be initiated simultaneously
807 * by both hosts, we have a random backoff mechanism -
808 * see the comment above rds_queue_reconnect()
809 */
810 mutex_lock(&conn->c_cm_lock);
811 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
812 if (rds_conn_state(conn) == RDS_CONN_UP) {
813 rdsdebug("incoming connect while connecting\n");
814 rds_conn_drop(conn);
815 rds_ib_stats_inc(s_ib_listen_closed_stale);
816 } else
817 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
818 /* Wait and see - our connect may still be succeeding */
819 rds_ib_stats_inc(s_ib_connect_raced);
820 }
821 goto out;
822 }
823
824 ic = conn->c_transport_data;
825
826 rds_ib_set_protocol(conn, version);
827 rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
828
829 /* If the peer gave us the last packet it saw, process this as if
830 * we had received a regular ACK. */
831 if (dp_cmn->ricpc_ack_seq)
832 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
833 NULL);
834
835 BUG_ON(cm_id->context);
836 BUG_ON(ic->i_cm_id);
837
838 ic->i_cm_id = cm_id;
839 cm_id->context = conn;
840
841 /* We got halfway through setting up the ib_connection, if we
842 * fail now, we have to take the long route out of this mess. */
843 destroy = 0;
844
845 err = rds_ib_setup_qp(conn);
846 if (err) {
847 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
848 goto out;
849 }
850
851 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
852 event->param.conn.responder_resources,
853 event->param.conn.initiator_depth, isv6);
854
855 /* rdma_accept() calls rdma_reject() internally if it fails */
856 if (rdma_accept(cm_id, &conn_param))
857 rds_ib_conn_error(conn, "rdma_accept failed\n");
858
859out:
860 if (conn)
861 mutex_unlock(&conn->c_cm_lock);
862 if (err)
863 rdma_reject(cm_id, &err, sizeof(int));
864 return destroy;
865}
866
867
868int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
869{
870 struct rds_connection *conn = cm_id->context;
871 struct rds_ib_connection *ic = conn->c_transport_data;
872 struct rdma_conn_param conn_param;
873 union rds_ib_conn_priv dp;
874 int ret;
875
876 /* If the peer doesn't do protocol negotiation, we must
877 * default to RDSv3.0 */
878 rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
879 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
880
881 ret = rds_ib_setup_qp(conn);
882 if (ret) {
883 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
884 goto out;
885 }
886
887 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
888 conn->c_proposed_version,
889 UINT_MAX, UINT_MAX, isv6);
890 ret = rdma_connect(cm_id, &conn_param);
891 if (ret)
892 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
893
894out:
895 /* Beware - returning non-zero tells the rdma_cm to destroy
896 * the cm_id. We should certainly not do it as long as we still
897 * "own" the cm_id. */
898 if (ret) {
899 if (ic->i_cm_id == cm_id)
900 ret = 0;
901 }
902 ic->i_active_side = true;
903 return ret;
904}
905
906int rds_ib_conn_path_connect(struct rds_conn_path *cp)
907{
908 struct rds_connection *conn = cp->cp_conn;
909 struct sockaddr_storage src, dest;
910 rdma_cm_event_handler handler;
911 struct rds_ib_connection *ic;
912 int ret;
913
914 ic = conn->c_transport_data;
915
916 /* XXX I wonder what affect the port space has */
917 /* delegate cm event handler to rdma_transport */
918#if IS_ENABLED(CONFIG_IPV6)
919 if (conn->c_isv6)
920 handler = rds6_rdma_cm_event_handler;
921 else
922#endif
923 handler = rds_rdma_cm_event_handler;
924 ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
925 RDMA_PS_TCP, IB_QPT_RC);
926 if (IS_ERR(ic->i_cm_id)) {
927 ret = PTR_ERR(ic->i_cm_id);
928 ic->i_cm_id = NULL;
929 rdsdebug("rdma_create_id() failed: %d\n", ret);
930 goto out;
931 }
932
933 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
934
935 if (ipv6_addr_v4mapped(&conn->c_faddr)) {
936 struct sockaddr_in *sin;
937
938 sin = (struct sockaddr_in *)&src;
939 sin->sin_family = AF_INET;
940 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
941 sin->sin_port = 0;
942
943 sin = (struct sockaddr_in *)&dest;
944 sin->sin_family = AF_INET;
945 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
946 sin->sin_port = htons(RDS_PORT);
947 } else {
948 struct sockaddr_in6 *sin6;
949
950 sin6 = (struct sockaddr_in6 *)&src;
951 sin6->sin6_family = AF_INET6;
952 sin6->sin6_addr = conn->c_laddr;
953 sin6->sin6_port = 0;
954 sin6->sin6_scope_id = conn->c_dev_if;
955
956 sin6 = (struct sockaddr_in6 *)&dest;
957 sin6->sin6_family = AF_INET6;
958 sin6->sin6_addr = conn->c_faddr;
959 sin6->sin6_port = htons(RDS_CM_PORT);
960 sin6->sin6_scope_id = conn->c_dev_if;
961 }
962
963 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
964 (struct sockaddr *)&dest,
965 RDS_RDMA_RESOLVE_TIMEOUT_MS);
966 if (ret) {
967 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
968 ret);
969 rdma_destroy_id(ic->i_cm_id);
970 ic->i_cm_id = NULL;
971 }
972
973out:
974 return ret;
975}
976
977/*
978 * This is so careful about only cleaning up resources that were built up
979 * so that it can be called at any point during startup. In fact it
980 * can be called multiple times for a given connection.
981 */
982void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
983{
984 struct rds_connection *conn = cp->cp_conn;
985 struct rds_ib_connection *ic = conn->c_transport_data;
986 int err = 0;
987
988 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
989 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
990 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
991
992 if (ic->i_cm_id) {
993 struct ib_device *dev = ic->i_cm_id->device;
994
995 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
996 err = rdma_disconnect(ic->i_cm_id);
997 if (err) {
998 /* Actually this may happen quite frequently, when
999 * an outgoing connect raced with an incoming connect.
1000 */
1001 rdsdebug("failed to disconnect, cm: %p err %d\n",
1002 ic->i_cm_id, err);
1003 }
1004
1005 /* kick off "flush_worker" for all pools in order to reap
1006 * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1007 */
1008 rds_ib_flush_mrs();
1009
1010 /*
1011 * We want to wait for tx and rx completion to finish
1012 * before we tear down the connection, but we have to be
1013 * careful not to get stuck waiting on a send ring that
1014 * only has unsignaled sends in it. We've shutdown new
1015 * sends before getting here so by waiting for signaled
1016 * sends to complete we're ensured that there will be no
1017 * more tx processing.
1018 */
1019 wait_event(rds_ib_ring_empty_wait,
1020 rds_ib_ring_empty(&ic->i_recv_ring) &&
1021 (atomic_read(&ic->i_signaled_sends) == 0) &&
1022 (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1023 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1024 tasklet_kill(&ic->i_send_tasklet);
1025 tasklet_kill(&ic->i_recv_tasklet);
1026
1027 atomic_set(&ic->i_cq_quiesce, 1);
1028
1029 /* first destroy the ib state that generates callbacks */
1030 if (ic->i_cm_id->qp)
1031 rdma_destroy_qp(ic->i_cm_id);
1032 if (ic->i_send_cq) {
1033 if (ic->rds_ibdev)
1034 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1035 ib_destroy_cq(ic->i_send_cq);
1036 }
1037
1038 if (ic->i_recv_cq) {
1039 if (ic->rds_ibdev)
1040 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1041 ib_destroy_cq(ic->i_recv_cq);
1042 }
1043
1044 /* then free the resources that ib callbacks use */
1045 if (ic->i_send_hdrs)
1046 ib_dma_free_coherent(dev,
1047 ic->i_send_ring.w_nr *
1048 sizeof(struct rds_header),
1049 ic->i_send_hdrs,
1050 ic->i_send_hdrs_dma);
1051
1052 if (ic->i_recv_hdrs)
1053 ib_dma_free_coherent(dev,
1054 ic->i_recv_ring.w_nr *
1055 sizeof(struct rds_header),
1056 ic->i_recv_hdrs,
1057 ic->i_recv_hdrs_dma);
1058
1059 if (ic->i_ack)
1060 ib_dma_free_coherent(dev, sizeof(struct rds_header),
1061 ic->i_ack, ic->i_ack_dma);
1062
1063 if (ic->i_sends)
1064 rds_ib_send_clear_ring(ic);
1065 if (ic->i_recvs)
1066 rds_ib_recv_clear_ring(ic);
1067
1068 rdma_destroy_id(ic->i_cm_id);
1069
1070 /*
1071 * Move connection back to the nodev list.
1072 */
1073 if (ic->rds_ibdev)
1074 rds_ib_remove_conn(ic->rds_ibdev, conn);
1075
1076 ic->i_cm_id = NULL;
1077 ic->i_pd = NULL;
1078 ic->i_send_cq = NULL;
1079 ic->i_recv_cq = NULL;
1080 ic->i_send_hdrs = NULL;
1081 ic->i_recv_hdrs = NULL;
1082 ic->i_ack = NULL;
1083 }
1084 BUG_ON(ic->rds_ibdev);
1085
1086 /* Clear pending transmit */
1087 if (ic->i_data_op) {
1088 struct rds_message *rm;
1089
1090 rm = container_of(ic->i_data_op, struct rds_message, data);
1091 rds_message_put(rm);
1092 ic->i_data_op = NULL;
1093 }
1094
1095 /* Clear the ACK state */
1096 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1097#ifdef KERNEL_HAS_ATOMIC64
1098 atomic64_set(&ic->i_ack_next, 0);
1099#else
1100 ic->i_ack_next = 0;
1101#endif
1102 ic->i_ack_recv = 0;
1103
1104 /* Clear flow control state */
1105 ic->i_flowctl = 0;
1106 atomic_set(&ic->i_credits, 0);
1107
1108 /* Re-init rings, but retain sizes. */
1109 rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1110 rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1111
1112 if (ic->i_ibinc) {
1113 rds_inc_put(&ic->i_ibinc->ii_inc);
1114 ic->i_ibinc = NULL;
1115 }
1116
1117 vfree(ic->i_sends);
1118 ic->i_sends = NULL;
1119 vfree(ic->i_recvs);
1120 ic->i_recvs = NULL;
1121 ic->i_active_side = false;
1122}
1123
1124int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1125{
1126 struct rds_ib_connection *ic;
1127 unsigned long flags;
1128 int ret;
1129
1130 /* XXX too lazy? */
1131 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1132 if (!ic)
1133 return -ENOMEM;
1134
1135 ret = rds_ib_recv_alloc_caches(ic, gfp);
1136 if (ret) {
1137 kfree(ic);
1138 return ret;
1139 }
1140
1141 INIT_LIST_HEAD(&ic->ib_node);
1142 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1143 (unsigned long)ic);
1144 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1145 (unsigned long)ic);
1146 mutex_init(&ic->i_recv_mutex);
1147#ifndef KERNEL_HAS_ATOMIC64
1148 spin_lock_init(&ic->i_ack_lock);
1149#endif
1150 atomic_set(&ic->i_signaled_sends, 0);
1151 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1152
1153 /*
1154 * rds_ib_conn_shutdown() waits for these to be emptied so they
1155 * must be initialized before it can be called.
1156 */
1157 rds_ib_ring_init(&ic->i_send_ring, 0);
1158 rds_ib_ring_init(&ic->i_recv_ring, 0);
1159
1160 ic->conn = conn;
1161 conn->c_transport_data = ic;
1162
1163 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1164 list_add_tail(&ic->ib_node, &ib_nodev_conns);
1165 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1166
1167
1168 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1169 return 0;
1170}
1171
1172/*
1173 * Free a connection. Connection must be shut down and not set for reconnect.
1174 */
1175void rds_ib_conn_free(void *arg)
1176{
1177 struct rds_ib_connection *ic = arg;
1178 spinlock_t *lock_ptr;
1179
1180 rdsdebug("ic %p\n", ic);
1181
1182 /*
1183 * Conn is either on a dev's list or on the nodev list.
1184 * A race with shutdown() or connect() would cause problems
1185 * (since rds_ibdev would change) but that should never happen.
1186 */
1187 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1188
1189 spin_lock_irq(lock_ptr);
1190 list_del(&ic->ib_node);
1191 spin_unlock_irq(lock_ptr);
1192
1193 rds_ib_recv_free_caches(ic);
1194
1195 kfree(ic);
1196}
1197
1198
1199/*
1200 * An error occurred on the connection
1201 */
1202void
1203__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1204{
1205 va_list ap;
1206
1207 rds_conn_drop(conn);
1208
1209 va_start(ap, fmt);
1210 vprintk(fmt, ap);
1211 va_end(ap);
1212}