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, 2019 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/dmapool.h>
34#include <linux/kernel.h>
35#include <linux/in.h>
36#include <linux/slab.h>
37#include <linux/vmalloc.h>
38#include <linux/ratelimit.h>
39#include <net/addrconf.h>
40#include <rdma/ib_cm.h>
41
42#include "rds_single_path.h"
43#include "rds.h"
44#include "ib.h"
45#include "ib_mr.h"
46
47/*
48 * Set the selected protocol version
49 */
50static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
51{
52 conn->c_version = version;
53}
54
55/*
56 * Set up flow control
57 */
58static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
59{
60 struct rds_ib_connection *ic = conn->c_transport_data;
61
62 if (rds_ib_sysctl_flow_control && credits != 0) {
63 /* We're doing flow control */
64 ic->i_flowctl = 1;
65 rds_ib_send_add_credits(conn, credits);
66 } else {
67 ic->i_flowctl = 0;
68 }
69}
70
71/*
72 * Tune RNR behavior. Without flow control, we use a rather
73 * low timeout, but not the absolute minimum - this should
74 * be tunable.
75 *
76 * We already set the RNR retry count to 7 (which is the
77 * smallest infinite number :-) above.
78 * If flow control is off, we want to change this back to 0
79 * so that we learn quickly when our credit accounting is
80 * buggy.
81 *
82 * Caller passes in a qp_attr pointer - don't waste stack spacv
83 * by allocation this twice.
84 */
85static void
86rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
87{
88 int ret;
89
90 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
91 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
92 if (ret)
93 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
94}
95
96/*
97 * Connection established.
98 * We get here for both outgoing and incoming connection.
99 */
100void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
101{
102 struct rds_ib_connection *ic = conn->c_transport_data;
103 const union rds_ib_conn_priv *dp = NULL;
104 struct ib_qp_attr qp_attr;
105 __be64 ack_seq = 0;
106 __be32 credit = 0;
107 u8 major = 0;
108 u8 minor = 0;
109 int err;
110
111 dp = event->param.conn.private_data;
112 if (conn->c_isv6) {
113 if (event->param.conn.private_data_len >=
114 sizeof(struct rds6_ib_connect_private)) {
115 major = dp->ricp_v6.dp_protocol_major;
116 minor = dp->ricp_v6.dp_protocol_minor;
117 credit = dp->ricp_v6.dp_credit;
118 /* dp structure start is not guaranteed to be 8 bytes
119 * aligned. Since dp_ack_seq is 64-bit extended load
120 * operations can be used so go through get_unaligned
121 * to avoid unaligned errors.
122 */
123 ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
124 }
125 } else if (event->param.conn.private_data_len >=
126 sizeof(struct rds_ib_connect_private)) {
127 major = dp->ricp_v4.dp_protocol_major;
128 minor = dp->ricp_v4.dp_protocol_minor;
129 credit = dp->ricp_v4.dp_credit;
130 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
131 }
132
133 /* make sure it isn't empty data */
134 if (major) {
135 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
136 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
137 }
138
139 if (conn->c_version < RDS_PROTOCOL_VERSION) {
140 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
141 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
142 &conn->c_laddr, &conn->c_faddr,
143 RDS_PROTOCOL_MAJOR(conn->c_version),
144 RDS_PROTOCOL_MINOR(conn->c_version));
145 rds_conn_destroy(conn);
146 return;
147 }
148 }
149
150 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
151 ic->i_active_side ? "Active" : "Passive",
152 &conn->c_laddr, &conn->c_faddr, conn->c_tos,
153 RDS_PROTOCOL_MAJOR(conn->c_version),
154 RDS_PROTOCOL_MINOR(conn->c_version),
155 ic->i_flowctl ? ", flow control" : "");
156
157 /* receive sl from the peer */
158 ic->i_sl = ic->i_cm_id->route.path_rec->sl;
159
160 atomic_set(&ic->i_cq_quiesce, 0);
161
162 /* Init rings and fill recv. this needs to wait until protocol
163 * negotiation is complete, since ring layout is different
164 * from 3.1 to 4.1.
165 */
166 rds_ib_send_init_ring(ic);
167 rds_ib_recv_init_ring(ic);
168 /* Post receive buffers - as a side effect, this will update
169 * the posted credit count. */
170 rds_ib_recv_refill(conn, 1, GFP_KERNEL);
171
172 /* Tune RNR behavior */
173 rds_ib_tune_rnr(ic, &qp_attr);
174
175 qp_attr.qp_state = IB_QPS_RTS;
176 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
177 if (err)
178 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
179
180 /* update ib_device with this local ipaddr */
181 err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
182 if (err)
183 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
184 err);
185
186 /* If the peer gave us the last packet it saw, process this as if
187 * we had received a regular ACK. */
188 if (dp) {
189 if (ack_seq)
190 rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
191 NULL);
192 }
193
194 conn->c_proposed_version = conn->c_version;
195 rds_connect_complete(conn);
196}
197
198static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
199 struct rdma_conn_param *conn_param,
200 union rds_ib_conn_priv *dp,
201 u32 protocol_version,
202 u32 max_responder_resources,
203 u32 max_initiator_depth,
204 bool isv6)
205{
206 struct rds_ib_connection *ic = conn->c_transport_data;
207 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
208
209 memset(conn_param, 0, sizeof(struct rdma_conn_param));
210
211 conn_param->responder_resources =
212 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
213 conn_param->initiator_depth =
214 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
215 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
216 conn_param->rnr_retry_count = 7;
217
218 if (dp) {
219 memset(dp, 0, sizeof(*dp));
220 if (isv6) {
221 dp->ricp_v6.dp_saddr = conn->c_laddr;
222 dp->ricp_v6.dp_daddr = conn->c_faddr;
223 dp->ricp_v6.dp_protocol_major =
224 RDS_PROTOCOL_MAJOR(protocol_version);
225 dp->ricp_v6.dp_protocol_minor =
226 RDS_PROTOCOL_MINOR(protocol_version);
227 dp->ricp_v6.dp_protocol_minor_mask =
228 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
229 dp->ricp_v6.dp_ack_seq =
230 cpu_to_be64(rds_ib_piggyb_ack(ic));
231 dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
232
233 conn_param->private_data = &dp->ricp_v6;
234 conn_param->private_data_len = sizeof(dp->ricp_v6);
235 } else {
236 dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
237 dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
238 dp->ricp_v4.dp_protocol_major =
239 RDS_PROTOCOL_MAJOR(protocol_version);
240 dp->ricp_v4.dp_protocol_minor =
241 RDS_PROTOCOL_MINOR(protocol_version);
242 dp->ricp_v4.dp_protocol_minor_mask =
243 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
244 dp->ricp_v4.dp_ack_seq =
245 cpu_to_be64(rds_ib_piggyb_ack(ic));
246 dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
247
248 conn_param->private_data = &dp->ricp_v4;
249 conn_param->private_data_len = sizeof(dp->ricp_v4);
250 }
251
252 /* Advertise flow control */
253 if (ic->i_flowctl) {
254 unsigned int credits;
255
256 credits = IB_GET_POST_CREDITS
257 (atomic_read(&ic->i_credits));
258 if (isv6)
259 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
260 else
261 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
262 atomic_sub(IB_SET_POST_CREDITS(credits),
263 &ic->i_credits);
264 }
265 }
266}
267
268static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
269{
270 rdsdebug("event %u (%s) data %p\n",
271 event->event, ib_event_msg(event->event), data);
272}
273
274/* Plucking the oldest entry from the ring can be done concurrently with
275 * the thread refilling the ring. Each ring operation is protected by
276 * spinlocks and the transient state of refilling doesn't change the
277 * recording of which entry is oldest.
278 *
279 * This relies on IB only calling one cq comp_handler for each cq so that
280 * there will only be one caller of rds_recv_incoming() per RDS connection.
281 */
282static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
283{
284 struct rds_connection *conn = context;
285 struct rds_ib_connection *ic = conn->c_transport_data;
286
287 rdsdebug("conn %p cq %p\n", conn, cq);
288
289 rds_ib_stats_inc(s_ib_evt_handler_call);
290
291 tasklet_schedule(&ic->i_recv_tasklet);
292}
293
294static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
295 struct ib_wc *wcs)
296{
297 int nr, i;
298 struct ib_wc *wc;
299
300 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
301 for (i = 0; i < nr; i++) {
302 wc = wcs + i;
303 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
304 (unsigned long long)wc->wr_id, wc->status,
305 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
306
307 if (wc->wr_id <= ic->i_send_ring.w_nr ||
308 wc->wr_id == RDS_IB_ACK_WR_ID)
309 rds_ib_send_cqe_handler(ic, wc);
310 else
311 rds_ib_mr_cqe_handler(ic, wc);
312
313 }
314 }
315}
316
317static void rds_ib_tasklet_fn_send(unsigned long data)
318{
319 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
320 struct rds_connection *conn = ic->conn;
321
322 rds_ib_stats_inc(s_ib_tasklet_call);
323
324 /* if cq has been already reaped, ignore incoming cq event */
325 if (atomic_read(&ic->i_cq_quiesce))
326 return;
327
328 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
329 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
330 poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
331
332 if (rds_conn_up(conn) &&
333 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
334 test_bit(0, &conn->c_map_queued)))
335 rds_send_xmit(&ic->conn->c_path[0]);
336}
337
338static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
339 struct ib_wc *wcs,
340 struct rds_ib_ack_state *ack_state)
341{
342 int nr, i;
343 struct ib_wc *wc;
344
345 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
346 for (i = 0; i < nr; i++) {
347 wc = wcs + i;
348 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
349 (unsigned long long)wc->wr_id, wc->status,
350 wc->byte_len, be32_to_cpu(wc->ex.imm_data));
351
352 rds_ib_recv_cqe_handler(ic, wc, ack_state);
353 }
354 }
355}
356
357static void rds_ib_tasklet_fn_recv(unsigned long data)
358{
359 struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
360 struct rds_connection *conn = ic->conn;
361 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
362 struct rds_ib_ack_state state;
363
364 if (!rds_ibdev)
365 rds_conn_drop(conn);
366
367 rds_ib_stats_inc(s_ib_tasklet_call);
368
369 /* if cq has been already reaped, ignore incoming cq event */
370 if (atomic_read(&ic->i_cq_quiesce))
371 return;
372
373 memset(&state, 0, sizeof(state));
374 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
375 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
376 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
377
378 if (state.ack_next_valid)
379 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
380 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
381 rds_send_drop_acked(conn, state.ack_recv, NULL);
382 ic->i_ack_recv = state.ack_recv;
383 }
384
385 if (rds_conn_up(conn))
386 rds_ib_attempt_ack(ic);
387}
388
389static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
390{
391 struct rds_connection *conn = data;
392 struct rds_ib_connection *ic = conn->c_transport_data;
393
394 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
395 ib_event_msg(event->event));
396
397 switch (event->event) {
398 case IB_EVENT_COMM_EST:
399 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
400 break;
401 default:
402 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
403 event->event, ib_event_msg(event->event),
404 &conn->c_laddr, &conn->c_faddr);
405 rds_conn_drop(conn);
406 break;
407 }
408}
409
410static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
411{
412 struct rds_connection *conn = context;
413 struct rds_ib_connection *ic = conn->c_transport_data;
414
415 rdsdebug("conn %p cq %p\n", conn, cq);
416
417 rds_ib_stats_inc(s_ib_evt_handler_call);
418
419 tasklet_schedule(&ic->i_send_tasklet);
420}
421
422static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
423{
424 int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
425 int index = rds_ibdev->dev->num_comp_vectors - 1;
426 int i;
427
428 for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
429 if (rds_ibdev->vector_load[i] < min) {
430 index = i;
431 min = rds_ibdev->vector_load[i];
432 }
433 }
434
435 rds_ibdev->vector_load[index]++;
436 return index;
437}
438
439static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
440{
441 rds_ibdev->vector_load[index]--;
442}
443
444/* Allocate DMA coherent memory to be used to store struct rds_header for
445 * sending/receiving packets. The pointers to the DMA memory and the
446 * associated DMA addresses are stored in two arrays.
447 *
448 * @ibdev: the IB device
449 * @pool: the DMA memory pool
450 * @dma_addrs: pointer to the array for storing DMA addresses
451 * @num_hdrs: number of headers to allocate
452 *
453 * It returns the pointer to the array storing the DMA memory pointers. On
454 * error, NULL pointer is returned.
455 */
456struct rds_header **rds_dma_hdrs_alloc(struct ib_device *ibdev,
457 struct dma_pool *pool,
458 dma_addr_t **dma_addrs, u32 num_hdrs)
459{
460 struct rds_header **hdrs;
461 dma_addr_t *hdr_daddrs;
462 u32 i;
463
464 hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
465 ibdev_to_node(ibdev));
466 if (!hdrs)
467 return NULL;
468
469 hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
470 ibdev_to_node(ibdev));
471 if (!hdr_daddrs) {
472 kvfree(hdrs);
473 return NULL;
474 }
475
476 for (i = 0; i < num_hdrs; i++) {
477 hdrs[i] = dma_pool_zalloc(pool, GFP_KERNEL, &hdr_daddrs[i]);
478 if (!hdrs[i]) {
479 rds_dma_hdrs_free(pool, hdrs, hdr_daddrs, i);
480 return NULL;
481 }
482 }
483
484 *dma_addrs = hdr_daddrs;
485 return hdrs;
486}
487
488/* Free the DMA memory used to store struct rds_header.
489 *
490 * @pool: the DMA memory pool
491 * @hdrs: pointer to the array storing DMA memory pointers
492 * @dma_addrs: pointer to the array storing DMA addresses
493 * @num_hdars: number of headers to free.
494 */
495void rds_dma_hdrs_free(struct dma_pool *pool, struct rds_header **hdrs,
496 dma_addr_t *dma_addrs, u32 num_hdrs)
497{
498 u32 i;
499
500 for (i = 0; i < num_hdrs; i++)
501 dma_pool_free(pool, hdrs[i], dma_addrs[i]);
502 kvfree(hdrs);
503 kvfree(dma_addrs);
504}
505
506/*
507 * This needs to be very careful to not leave IS_ERR pointers around for
508 * cleanup to trip over.
509 */
510static int rds_ib_setup_qp(struct rds_connection *conn)
511{
512 struct rds_ib_connection *ic = conn->c_transport_data;
513 struct ib_device *dev = ic->i_cm_id->device;
514 struct ib_qp_init_attr attr;
515 struct ib_cq_init_attr cq_attr = {};
516 struct rds_ib_device *rds_ibdev;
517 unsigned long max_wrs;
518 int ret, fr_queue_space;
519 struct dma_pool *pool;
520
521 /*
522 * It's normal to see a null device if an incoming connection races
523 * with device removal, so we don't print a warning.
524 */
525 rds_ibdev = rds_ib_get_client_data(dev);
526 if (!rds_ibdev)
527 return -EOPNOTSUPP;
528
529 /* The fr_queue_space is currently set to 512, to add extra space on
530 * completion queue and send queue. This extra space is used for FRWR
531 * registration and invalidation work requests
532 */
533 fr_queue_space = RDS_IB_DEFAULT_FR_WR;
534
535 /* add the conn now so that connection establishment has the dev */
536 rds_ib_add_conn(rds_ibdev, conn);
537
538 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
539 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
540 if (ic->i_send_ring.w_nr != max_wrs)
541 rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
542
543 max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
544 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
545 if (ic->i_recv_ring.w_nr != max_wrs)
546 rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
547
548 /* Protection domain and memory range */
549 ic->i_pd = rds_ibdev->pd;
550
551 ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
552 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
553 cq_attr.comp_vector = ic->i_scq_vector;
554 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
555 rds_ib_cq_event_handler, conn,
556 &cq_attr);
557 if (IS_ERR(ic->i_send_cq)) {
558 ret = PTR_ERR(ic->i_send_cq);
559 ic->i_send_cq = NULL;
560 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
561 rdsdebug("ib_create_cq send failed: %d\n", ret);
562 goto rds_ibdev_out;
563 }
564
565 ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
566 cq_attr.cqe = ic->i_recv_ring.w_nr;
567 cq_attr.comp_vector = ic->i_rcq_vector;
568 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
569 rds_ib_cq_event_handler, conn,
570 &cq_attr);
571 if (IS_ERR(ic->i_recv_cq)) {
572 ret = PTR_ERR(ic->i_recv_cq);
573 ic->i_recv_cq = NULL;
574 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
575 rdsdebug("ib_create_cq recv failed: %d\n", ret);
576 goto send_cq_out;
577 }
578
579 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
580 if (ret) {
581 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
582 goto recv_cq_out;
583 }
584
585 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
586 if (ret) {
587 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
588 goto recv_cq_out;
589 }
590
591 /* XXX negotiate max send/recv with remote? */
592 memset(&attr, 0, sizeof(attr));
593 attr.event_handler = rds_ib_qp_event_handler;
594 attr.qp_context = conn;
595 /* + 1 to allow for the single ack message */
596 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
597 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
598 attr.cap.max_send_sge = rds_ibdev->max_sge;
599 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
600 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
601 attr.qp_type = IB_QPT_RC;
602 attr.send_cq = ic->i_send_cq;
603 attr.recv_cq = ic->i_recv_cq;
604
605 /*
606 * XXX this can fail if max_*_wr is too large? Are we supposed
607 * to back off until we get a value that the hardware can support?
608 */
609 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
610 if (ret) {
611 rdsdebug("rdma_create_qp failed: %d\n", ret);
612 goto recv_cq_out;
613 }
614
615 pool = rds_ibdev->rid_hdrs_pool;
616 ic->i_send_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_send_hdrs_dma,
617 ic->i_send_ring.w_nr);
618 if (!ic->i_send_hdrs) {
619 ret = -ENOMEM;
620 rdsdebug("DMA send hdrs alloc failed\n");
621 goto qp_out;
622 }
623
624 ic->i_recv_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_recv_hdrs_dma,
625 ic->i_recv_ring.w_nr);
626 if (!ic->i_recv_hdrs) {
627 ret = -ENOMEM;
628 rdsdebug("DMA recv hdrs alloc failed\n");
629 goto send_hdrs_dma_out;
630 }
631
632 ic->i_ack = dma_pool_zalloc(pool, GFP_KERNEL,
633 &ic->i_ack_dma);
634 if (!ic->i_ack) {
635 ret = -ENOMEM;
636 rdsdebug("DMA ack header alloc failed\n");
637 goto recv_hdrs_dma_out;
638 }
639
640 ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
641 ic->i_send_ring.w_nr),
642 ibdev_to_node(dev));
643 if (!ic->i_sends) {
644 ret = -ENOMEM;
645 rdsdebug("send allocation failed\n");
646 goto ack_dma_out;
647 }
648
649 ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
650 ic->i_recv_ring.w_nr),
651 ibdev_to_node(dev));
652 if (!ic->i_recvs) {
653 ret = -ENOMEM;
654 rdsdebug("recv allocation failed\n");
655 goto sends_out;
656 }
657
658 rds_ib_recv_init_ack(ic);
659
660 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
661 ic->i_send_cq, ic->i_recv_cq);
662
663 goto out;
664
665sends_out:
666 vfree(ic->i_sends);
667
668ack_dma_out:
669 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
670 ic->i_ack = NULL;
671
672recv_hdrs_dma_out:
673 rds_dma_hdrs_free(pool, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
674 ic->i_recv_ring.w_nr);
675 ic->i_recv_hdrs = NULL;
676 ic->i_recv_hdrs_dma = NULL;
677
678send_hdrs_dma_out:
679 rds_dma_hdrs_free(pool, ic->i_send_hdrs, ic->i_send_hdrs_dma,
680 ic->i_send_ring.w_nr);
681 ic->i_send_hdrs = NULL;
682 ic->i_send_hdrs_dma = NULL;
683
684qp_out:
685 rdma_destroy_qp(ic->i_cm_id);
686recv_cq_out:
687 ib_destroy_cq(ic->i_recv_cq);
688 ic->i_recv_cq = NULL;
689send_cq_out:
690 ib_destroy_cq(ic->i_send_cq);
691 ic->i_send_cq = NULL;
692rds_ibdev_out:
693 rds_ib_remove_conn(rds_ibdev, conn);
694out:
695 rds_ib_dev_put(rds_ibdev);
696
697 return ret;
698}
699
700static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
701{
702 const union rds_ib_conn_priv *dp = event->param.conn.private_data;
703 u8 data_len, major, minor;
704 u32 version = 0;
705 __be16 mask;
706 u16 common;
707
708 /*
709 * rdma_cm private data is odd - when there is any private data in the
710 * request, we will be given a pretty large buffer without telling us the
711 * original size. The only way to tell the difference is by looking at
712 * the contents, which are initialized to zero.
713 * If the protocol version fields aren't set, this is a connection attempt
714 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
715 * We really should have changed this for OFED 1.3 :-(
716 */
717
718 /* Be paranoid. RDS always has privdata */
719 if (!event->param.conn.private_data_len) {
720 printk(KERN_NOTICE "RDS incoming connection has no private data, "
721 "rejecting\n");
722 return 0;
723 }
724
725 if (isv6) {
726 data_len = sizeof(struct rds6_ib_connect_private);
727 major = dp->ricp_v6.dp_protocol_major;
728 minor = dp->ricp_v6.dp_protocol_minor;
729 mask = dp->ricp_v6.dp_protocol_minor_mask;
730 } else {
731 data_len = sizeof(struct rds_ib_connect_private);
732 major = dp->ricp_v4.dp_protocol_major;
733 minor = dp->ricp_v4.dp_protocol_minor;
734 mask = dp->ricp_v4.dp_protocol_minor_mask;
735 }
736
737 /* Even if len is crap *now* I still want to check it. -ASG */
738 if (event->param.conn.private_data_len < data_len || major == 0)
739 return RDS_PROTOCOL_4_0;
740
741 common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
742 if (major == 4 && common) {
743 version = RDS_PROTOCOL_4_0;
744 while ((common >>= 1) != 0)
745 version++;
746 } else if (RDS_PROTOCOL_COMPAT_VERSION ==
747 RDS_PROTOCOL(major, minor)) {
748 version = RDS_PROTOCOL_COMPAT_VERSION;
749 } else {
750 if (isv6)
751 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
752 &dp->ricp_v6.dp_saddr, major, minor);
753 else
754 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
755 &dp->ricp_v4.dp_saddr, major, minor);
756 }
757 return version;
758}
759
760#if IS_ENABLED(CONFIG_IPV6)
761/* Given an IPv6 address, find the net_device which hosts that address and
762 * return its index. This is used by the rds_ib_cm_handle_connect() code to
763 * find the interface index of where an incoming request comes from when
764 * the request is using a link local address.
765 *
766 * Note one problem in this search. It is possible that two interfaces have
767 * the same link local address. Unfortunately, this cannot be solved unless
768 * the underlying layer gives us the interface which an incoming RDMA connect
769 * request comes from.
770 */
771static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
772{
773 struct net_device *dev;
774 int idx = 0;
775
776 rcu_read_lock();
777 for_each_netdev_rcu(net, dev) {
778 if (ipv6_chk_addr(net, addr, dev, 1)) {
779 idx = dev->ifindex;
780 break;
781 }
782 }
783 rcu_read_unlock();
784
785 return idx;
786}
787#endif
788
789int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
790 struct rdma_cm_event *event, bool isv6)
791{
792 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
793 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
794 const struct rds_ib_conn_priv_cmn *dp_cmn;
795 struct rds_connection *conn = NULL;
796 struct rds_ib_connection *ic = NULL;
797 struct rdma_conn_param conn_param;
798 const union rds_ib_conn_priv *dp;
799 union rds_ib_conn_priv dp_rep;
800 struct in6_addr s_mapped_addr;
801 struct in6_addr d_mapped_addr;
802 const struct in6_addr *saddr6;
803 const struct in6_addr *daddr6;
804 int destroy = 1;
805 u32 ifindex = 0;
806 u32 version;
807 int err = 1;
808
809 /* Check whether the remote protocol version matches ours. */
810 version = rds_ib_protocol_compatible(event, isv6);
811 if (!version) {
812 err = RDS_RDMA_REJ_INCOMPAT;
813 goto out;
814 }
815
816 dp = event->param.conn.private_data;
817 if (isv6) {
818#if IS_ENABLED(CONFIG_IPV6)
819 dp_cmn = &dp->ricp_v6.dp_cmn;
820 saddr6 = &dp->ricp_v6.dp_saddr;
821 daddr6 = &dp->ricp_v6.dp_daddr;
822 /* If either address is link local, need to find the
823 * interface index in order to create a proper RDS
824 * connection.
825 */
826 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
827 /* Using init_net for now .. */
828 ifindex = __rds_find_ifindex(&init_net, daddr6);
829 /* No index found... Need to bail out. */
830 if (ifindex == 0) {
831 err = -EOPNOTSUPP;
832 goto out;
833 }
834 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
835 /* Use our address to find the correct index. */
836 ifindex = __rds_find_ifindex(&init_net, daddr6);
837 /* No index found... Need to bail out. */
838 if (ifindex == 0) {
839 err = -EOPNOTSUPP;
840 goto out;
841 }
842 }
843#else
844 err = -EOPNOTSUPP;
845 goto out;
846#endif
847 } else {
848 dp_cmn = &dp->ricp_v4.dp_cmn;
849 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
850 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
851 saddr6 = &s_mapped_addr;
852 daddr6 = &d_mapped_addr;
853 }
854
855 rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
856 saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
857 RDS_PROTOCOL_MINOR(version),
858 (unsigned long long)be64_to_cpu(lguid),
859 (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
860
861 /* RDS/IB is not currently netns aware, thus init_net */
862 conn = rds_conn_create(&init_net, daddr6, saddr6,
863 &rds_ib_transport, dp_cmn->ricpc_dp_toss,
864 GFP_KERNEL, ifindex);
865 if (IS_ERR(conn)) {
866 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
867 conn = NULL;
868 goto out;
869 }
870
871 /*
872 * The connection request may occur while the
873 * previous connection exist, e.g. in case of failover.
874 * But as connections may be initiated simultaneously
875 * by both hosts, we have a random backoff mechanism -
876 * see the comment above rds_queue_reconnect()
877 */
878 mutex_lock(&conn->c_cm_lock);
879 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
880 if (rds_conn_state(conn) == RDS_CONN_UP) {
881 rdsdebug("incoming connect while connecting\n");
882 rds_conn_drop(conn);
883 rds_ib_stats_inc(s_ib_listen_closed_stale);
884 } else
885 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
886 /* Wait and see - our connect may still be succeeding */
887 rds_ib_stats_inc(s_ib_connect_raced);
888 }
889 goto out;
890 }
891
892 ic = conn->c_transport_data;
893
894 rds_ib_set_protocol(conn, version);
895 rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
896
897 /* If the peer gave us the last packet it saw, process this as if
898 * we had received a regular ACK. */
899 if (dp_cmn->ricpc_ack_seq)
900 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
901 NULL);
902
903 BUG_ON(cm_id->context);
904 BUG_ON(ic->i_cm_id);
905
906 ic->i_cm_id = cm_id;
907 cm_id->context = conn;
908
909 /* We got halfway through setting up the ib_connection, if we
910 * fail now, we have to take the long route out of this mess. */
911 destroy = 0;
912
913 err = rds_ib_setup_qp(conn);
914 if (err) {
915 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
916 goto out;
917 }
918
919 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
920 event->param.conn.responder_resources,
921 event->param.conn.initiator_depth, isv6);
922
923 /* rdma_accept() calls rdma_reject() internally if it fails */
924 if (rdma_accept(cm_id, &conn_param))
925 rds_ib_conn_error(conn, "rdma_accept failed\n");
926
927out:
928 if (conn)
929 mutex_unlock(&conn->c_cm_lock);
930 if (err)
931 rdma_reject(cm_id, &err, sizeof(int),
932 IB_CM_REJ_CONSUMER_DEFINED);
933 return destroy;
934}
935
936
937int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
938{
939 struct rds_connection *conn = cm_id->context;
940 struct rds_ib_connection *ic = conn->c_transport_data;
941 struct rdma_conn_param conn_param;
942 union rds_ib_conn_priv dp;
943 int ret;
944
945 /* If the peer doesn't do protocol negotiation, we must
946 * default to RDSv3.0 */
947 rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
948 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
949
950 ret = rds_ib_setup_qp(conn);
951 if (ret) {
952 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
953 goto out;
954 }
955
956 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
957 conn->c_proposed_version,
958 UINT_MAX, UINT_MAX, isv6);
959 ret = rdma_connect(cm_id, &conn_param);
960 if (ret)
961 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
962
963out:
964 /* Beware - returning non-zero tells the rdma_cm to destroy
965 * the cm_id. We should certainly not do it as long as we still
966 * "own" the cm_id. */
967 if (ret) {
968 if (ic->i_cm_id == cm_id)
969 ret = 0;
970 }
971 ic->i_active_side = true;
972 return ret;
973}
974
975int rds_ib_conn_path_connect(struct rds_conn_path *cp)
976{
977 struct rds_connection *conn = cp->cp_conn;
978 struct sockaddr_storage src, dest;
979 rdma_cm_event_handler handler;
980 struct rds_ib_connection *ic;
981 int ret;
982
983 ic = conn->c_transport_data;
984
985 /* XXX I wonder what affect the port space has */
986 /* delegate cm event handler to rdma_transport */
987#if IS_ENABLED(CONFIG_IPV6)
988 if (conn->c_isv6)
989 handler = rds6_rdma_cm_event_handler;
990 else
991#endif
992 handler = rds_rdma_cm_event_handler;
993 ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
994 RDMA_PS_TCP, IB_QPT_RC);
995 if (IS_ERR(ic->i_cm_id)) {
996 ret = PTR_ERR(ic->i_cm_id);
997 ic->i_cm_id = NULL;
998 rdsdebug("rdma_create_id() failed: %d\n", ret);
999 goto out;
1000 }
1001
1002 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
1003
1004 if (ipv6_addr_v4mapped(&conn->c_faddr)) {
1005 struct sockaddr_in *sin;
1006
1007 sin = (struct sockaddr_in *)&src;
1008 sin->sin_family = AF_INET;
1009 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1010 sin->sin_port = 0;
1011
1012 sin = (struct sockaddr_in *)&dest;
1013 sin->sin_family = AF_INET;
1014 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1015 sin->sin_port = htons(RDS_PORT);
1016 } else {
1017 struct sockaddr_in6 *sin6;
1018
1019 sin6 = (struct sockaddr_in6 *)&src;
1020 sin6->sin6_family = AF_INET6;
1021 sin6->sin6_addr = conn->c_laddr;
1022 sin6->sin6_port = 0;
1023 sin6->sin6_scope_id = conn->c_dev_if;
1024
1025 sin6 = (struct sockaddr_in6 *)&dest;
1026 sin6->sin6_family = AF_INET6;
1027 sin6->sin6_addr = conn->c_faddr;
1028 sin6->sin6_port = htons(RDS_CM_PORT);
1029 sin6->sin6_scope_id = conn->c_dev_if;
1030 }
1031
1032 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1033 (struct sockaddr *)&dest,
1034 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1035 if (ret) {
1036 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1037 ret);
1038 rdma_destroy_id(ic->i_cm_id);
1039 ic->i_cm_id = NULL;
1040 }
1041
1042out:
1043 return ret;
1044}
1045
1046/*
1047 * This is so careful about only cleaning up resources that were built up
1048 * so that it can be called at any point during startup. In fact it
1049 * can be called multiple times for a given connection.
1050 */
1051void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1052{
1053 struct rds_connection *conn = cp->cp_conn;
1054 struct rds_ib_connection *ic = conn->c_transport_data;
1055 int err = 0;
1056
1057 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1058 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1059 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1060
1061 if (ic->i_cm_id) {
1062 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1063 err = rdma_disconnect(ic->i_cm_id);
1064 if (err) {
1065 /* Actually this may happen quite frequently, when
1066 * an outgoing connect raced with an incoming connect.
1067 */
1068 rdsdebug("failed to disconnect, cm: %p err %d\n",
1069 ic->i_cm_id, err);
1070 }
1071
1072 /* kick off "flush_worker" for all pools in order to reap
1073 * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1074 */
1075 rds_ib_flush_mrs();
1076
1077 /*
1078 * We want to wait for tx and rx completion to finish
1079 * before we tear down the connection, but we have to be
1080 * careful not to get stuck waiting on a send ring that
1081 * only has unsignaled sends in it. We've shutdown new
1082 * sends before getting here so by waiting for signaled
1083 * sends to complete we're ensured that there will be no
1084 * more tx processing.
1085 */
1086 wait_event(rds_ib_ring_empty_wait,
1087 rds_ib_ring_empty(&ic->i_recv_ring) &&
1088 (atomic_read(&ic->i_signaled_sends) == 0) &&
1089 (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1090 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1091 tasklet_kill(&ic->i_send_tasklet);
1092 tasklet_kill(&ic->i_recv_tasklet);
1093
1094 atomic_set(&ic->i_cq_quiesce, 1);
1095
1096 /* first destroy the ib state that generates callbacks */
1097 if (ic->i_cm_id->qp)
1098 rdma_destroy_qp(ic->i_cm_id);
1099 if (ic->i_send_cq) {
1100 if (ic->rds_ibdev)
1101 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1102 ib_destroy_cq(ic->i_send_cq);
1103 }
1104
1105 if (ic->i_recv_cq) {
1106 if (ic->rds_ibdev)
1107 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1108 ib_destroy_cq(ic->i_recv_cq);
1109 }
1110
1111 if (ic->rds_ibdev) {
1112 struct dma_pool *pool;
1113
1114 pool = ic->rds_ibdev->rid_hdrs_pool;
1115
1116 /* then free the resources that ib callbacks use */
1117 if (ic->i_send_hdrs) {
1118 rds_dma_hdrs_free(pool, ic->i_send_hdrs,
1119 ic->i_send_hdrs_dma,
1120 ic->i_send_ring.w_nr);
1121 ic->i_send_hdrs = NULL;
1122 ic->i_send_hdrs_dma = NULL;
1123 }
1124
1125 if (ic->i_recv_hdrs) {
1126 rds_dma_hdrs_free(pool, ic->i_recv_hdrs,
1127 ic->i_recv_hdrs_dma,
1128 ic->i_recv_ring.w_nr);
1129 ic->i_recv_hdrs = NULL;
1130 ic->i_recv_hdrs_dma = NULL;
1131 }
1132
1133 if (ic->i_ack) {
1134 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
1135 ic->i_ack = NULL;
1136 }
1137 } else {
1138 WARN_ON(ic->i_send_hdrs);
1139 WARN_ON(ic->i_send_hdrs_dma);
1140 WARN_ON(ic->i_recv_hdrs);
1141 WARN_ON(ic->i_recv_hdrs_dma);
1142 WARN_ON(ic->i_ack);
1143 }
1144
1145 if (ic->i_sends)
1146 rds_ib_send_clear_ring(ic);
1147 if (ic->i_recvs)
1148 rds_ib_recv_clear_ring(ic);
1149
1150 rdma_destroy_id(ic->i_cm_id);
1151
1152 /*
1153 * Move connection back to the nodev list.
1154 */
1155 if (ic->rds_ibdev)
1156 rds_ib_remove_conn(ic->rds_ibdev, conn);
1157
1158 ic->i_cm_id = NULL;
1159 ic->i_pd = NULL;
1160 ic->i_send_cq = NULL;
1161 ic->i_recv_cq = NULL;
1162 }
1163 BUG_ON(ic->rds_ibdev);
1164
1165 /* Clear pending transmit */
1166 if (ic->i_data_op) {
1167 struct rds_message *rm;
1168
1169 rm = container_of(ic->i_data_op, struct rds_message, data);
1170 rds_message_put(rm);
1171 ic->i_data_op = NULL;
1172 }
1173
1174 /* Clear the ACK state */
1175 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1176#ifdef KERNEL_HAS_ATOMIC64
1177 atomic64_set(&ic->i_ack_next, 0);
1178#else
1179 ic->i_ack_next = 0;
1180#endif
1181 ic->i_ack_recv = 0;
1182
1183 /* Clear flow control state */
1184 ic->i_flowctl = 0;
1185 atomic_set(&ic->i_credits, 0);
1186
1187 /* Re-init rings, but retain sizes. */
1188 rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1189 rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1190
1191 if (ic->i_ibinc) {
1192 rds_inc_put(&ic->i_ibinc->ii_inc);
1193 ic->i_ibinc = NULL;
1194 }
1195
1196 vfree(ic->i_sends);
1197 ic->i_sends = NULL;
1198 vfree(ic->i_recvs);
1199 ic->i_recvs = NULL;
1200 ic->i_active_side = false;
1201}
1202
1203int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1204{
1205 struct rds_ib_connection *ic;
1206 unsigned long flags;
1207 int ret;
1208
1209 /* XXX too lazy? */
1210 ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1211 if (!ic)
1212 return -ENOMEM;
1213
1214 ret = rds_ib_recv_alloc_caches(ic, gfp);
1215 if (ret) {
1216 kfree(ic);
1217 return ret;
1218 }
1219
1220 INIT_LIST_HEAD(&ic->ib_node);
1221 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1222 (unsigned long)ic);
1223 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1224 (unsigned long)ic);
1225 mutex_init(&ic->i_recv_mutex);
1226#ifndef KERNEL_HAS_ATOMIC64
1227 spin_lock_init(&ic->i_ack_lock);
1228#endif
1229 atomic_set(&ic->i_signaled_sends, 0);
1230 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1231
1232 /*
1233 * rds_ib_conn_shutdown() waits for these to be emptied so they
1234 * must be initialized before it can be called.
1235 */
1236 rds_ib_ring_init(&ic->i_send_ring, 0);
1237 rds_ib_ring_init(&ic->i_recv_ring, 0);
1238
1239 ic->conn = conn;
1240 conn->c_transport_data = ic;
1241
1242 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1243 list_add_tail(&ic->ib_node, &ib_nodev_conns);
1244 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1245
1246
1247 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1248 return 0;
1249}
1250
1251/*
1252 * Free a connection. Connection must be shut down and not set for reconnect.
1253 */
1254void rds_ib_conn_free(void *arg)
1255{
1256 struct rds_ib_connection *ic = arg;
1257 spinlock_t *lock_ptr;
1258
1259 rdsdebug("ic %p\n", ic);
1260
1261 /*
1262 * Conn is either on a dev's list or on the nodev list.
1263 * A race with shutdown() or connect() would cause problems
1264 * (since rds_ibdev would change) but that should never happen.
1265 */
1266 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1267
1268 spin_lock_irq(lock_ptr);
1269 list_del(&ic->ib_node);
1270 spin_unlock_irq(lock_ptr);
1271
1272 rds_ib_recv_free_caches(ic);
1273
1274 kfree(ic);
1275}
1276
1277
1278/*
1279 * An error occurred on the connection
1280 */
1281void
1282__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1283{
1284 va_list ap;
1285
1286 rds_conn_drop(conn);
1287
1288 va_start(ap, fmt);
1289 vprintk(fmt, ap);
1290 va_end(ap);
1291}