1 /*
2 * Copyright (C) 2009 Michael Brown <mbrown@fensystems.co.uk>.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License as
6 * published by the Free Software Foundation; either version 2 of the
7 * License, or any later version.
8 *
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
17 * 02110-1301, USA.
18 *
19 * You can also choose to distribute this program under the terms of
20 * the Unmodified Binary Distribution Licence (as given in the file
21 * COPYING.UBDL), provided that you have satisfied its requirements.
22 */
23
24 FILE_LICENCE ( GPL2_OR_LATER_OR_UBDL );
25
26 #include <stdint.h>
27 #include <stdlib.h>
28 #include <string.h>
29 #include <byteswap.h>
30 #include <errno.h>
31 #include <assert.h>
32 #include <ipxe/infiniband.h>
33 #include <ipxe/ib_mi.h>
34 #include <ipxe/ib_pathrec.h>
35 #include <ipxe/ib_cm.h>
36
37 /**
38 * @file
39 *
40 * Infiniband communication management
41 *
42 */
43
44 /** List of connections */
45 static LIST_HEAD ( ib_cm_conns );
46
47 /**
48 * Find connection by local communication ID
49 *
50 * @v local_id Local communication ID
51 * @ret conn Connection, or NULL
52 */
ib_cm_find(uint32_t local_id)53 static struct ib_connection * ib_cm_find ( uint32_t local_id ) {
54 struct ib_connection *conn;
55
56 list_for_each_entry ( conn, &ib_cm_conns, list ) {
57 if ( conn->local_id == local_id )
58 return conn;
59 }
60 return NULL;
61 }
62
63 /**
64 * Send "ready to use" response
65 *
66 * @v ibdev Infiniband device
67 * @v mi Management interface
68 * @v tid Transaction identifier
69 * @v av Address vector
70 * @v local_id Local communication ID
71 * @v remote_id Remote communication ID
72 * @ret rc Return status code
73 */
ib_cm_send_rtu(struct ib_device * ibdev,struct ib_mad_interface * mi,struct ib_mad_tid * tid,struct ib_address_vector * av,uint32_t local_id,uint32_t remote_id)74 static int ib_cm_send_rtu ( struct ib_device *ibdev,
75 struct ib_mad_interface *mi,
76 struct ib_mad_tid *tid,
77 struct ib_address_vector *av,
78 uint32_t local_id, uint32_t remote_id ) {
79 union ib_mad mad;
80 struct ib_cm_ready_to_use *rtu = &mad.cm.cm_data.ready_to_use;
81 int rc;
82
83 /* Construct "ready to use" response */
84 memset ( &mad, 0, sizeof ( mad ) );
85 mad.hdr.mgmt_class = IB_MGMT_CLASS_CM;
86 mad.hdr.class_version = IB_CM_CLASS_VERSION;
87 mad.hdr.method = IB_MGMT_METHOD_SEND;
88 memcpy ( &mad.hdr.tid, tid, sizeof ( mad.hdr.tid ) );
89 mad.hdr.attr_id = htons ( IB_CM_ATTR_READY_TO_USE );
90 rtu->local_id = htonl ( local_id );
91 rtu->remote_id = htonl ( remote_id );
92 if ( ( rc = ib_mi_send ( ibdev, mi, &mad, av ) ) != 0 ) {
93 DBGC ( local_id, "CM %08x could not send RTU: %s\n",
94 local_id, strerror ( rc ) );
95 return rc;
96 }
97
98 return 0;
99 }
100
101 /**
102 * Handle duplicate connection replies
103 *
104 * @v ibdev Infiniband device
105 * @v mi Management interface
106 * @v mad Received MAD
107 * @v av Source address vector
108 * @ret rc Return status code
109 *
110 * If a "ready to use" MAD is lost, the peer may resend the connection
111 * reply. We have to respond to these with duplicate "ready to use"
112 * MADs, otherwise the peer may time out and drop the connection.
113 */
ib_cm_recv_rep(struct ib_device * ibdev,struct ib_mad_interface * mi,union ib_mad * mad,struct ib_address_vector * av)114 static void ib_cm_recv_rep ( struct ib_device *ibdev,
115 struct ib_mad_interface *mi,
116 union ib_mad *mad,
117 struct ib_address_vector *av ) {
118 struct ib_cm_connect_reply *rep = &mad->cm.cm_data.connect_reply;
119 struct ib_connection *conn;
120 uint32_t local_id = ntohl ( rep->remote_id );
121 int rc;
122
123 /* Identify connection */
124 conn = ib_cm_find ( local_id );
125 if ( conn ) {
126 /* Try to send "ready to use" reply */
127 if ( ( rc = ib_cm_send_rtu ( ibdev, mi, &mad->hdr.tid, av,
128 conn->local_id,
129 conn->remote_id ) ) != 0 ) {
130 /* Ignore errors; the remote end will retry */
131 }
132 } else {
133 DBGC ( local_id, "CM %08x unexpected REP\n", local_id );
134 }
135 }
136
137 /**
138 * Send reply to disconnection request
139 *
140 * @v ibdev Infiniband device
141 * @v mi Management interface
142 * @v tid Transaction identifier
143 * @v av Address vector
144 * @v local_id Local communication ID
145 * @v remote_id Remote communication ID
146 * @ret rc Return status code
147 */
ib_cm_send_drep(struct ib_device * ibdev,struct ib_mad_interface * mi,struct ib_mad_tid * tid,struct ib_address_vector * av,uint32_t local_id,uint32_t remote_id)148 static int ib_cm_send_drep ( struct ib_device *ibdev,
149 struct ib_mad_interface *mi,
150 struct ib_mad_tid *tid,
151 struct ib_address_vector *av,
152 uint32_t local_id, uint32_t remote_id ) {
153 union ib_mad mad;
154 struct ib_cm_disconnect_reply *drep = &mad.cm.cm_data.disconnect_reply;
155 int rc;
156
157 /* Construct reply to disconnection request */
158 memset ( &mad, 0, sizeof ( mad ) );
159 mad.hdr.mgmt_class = IB_MGMT_CLASS_CM;
160 mad.hdr.class_version = IB_CM_CLASS_VERSION;
161 mad.hdr.method = IB_MGMT_METHOD_SEND;
162 memcpy ( &mad.hdr.tid, tid, sizeof ( mad.hdr.tid ) );
163 mad.hdr.attr_id = htons ( IB_CM_ATTR_DISCONNECT_REPLY );
164 drep->local_id = htonl ( local_id );
165 drep->remote_id = htonl ( remote_id );
166 if ( ( rc = ib_mi_send ( ibdev, mi, &mad, av ) ) != 0 ) {
167 DBGC ( local_id, "CM %08x could not send DREP: %s\n",
168 local_id, strerror ( rc ) );
169 return rc;
170 }
171
172 return 0;
173 }
174
175 /**
176 * Handle disconnection requests
177 *
178 * @v ibdev Infiniband device
179 * @v mi Management interface
180 * @v mad Received MAD
181 * @v av Source address vector
182 * @ret rc Return status code
183 */
ib_cm_recv_dreq(struct ib_device * ibdev,struct ib_mad_interface * mi,union ib_mad * mad,struct ib_address_vector * av)184 static void ib_cm_recv_dreq ( struct ib_device *ibdev,
185 struct ib_mad_interface *mi,
186 union ib_mad *mad,
187 struct ib_address_vector *av ) {
188 struct ib_cm_disconnect_request *dreq =
189 &mad->cm.cm_data.disconnect_request;
190 struct ib_connection *conn;
191 uint32_t local_id = ntohl ( dreq->remote_id );
192 uint32_t remote_id = ntohl ( dreq->local_id );
193 int rc;
194
195 /* Identify connection */
196 conn = ib_cm_find ( local_id );
197 if ( conn ) {
198 /* Notify upper layer */
199 conn->op->changed ( ibdev, conn->qp, conn, -ENOTCONN,
200 &dreq->private_data,
201 sizeof ( dreq->private_data ) );
202 } else {
203 DBGC ( local_id, "CM %08x unexpected DREQ\n", local_id );
204 }
205
206 /* Send reply */
207 if ( ( rc = ib_cm_send_drep ( ibdev, mi, &mad->hdr.tid, av, local_id,
208 remote_id ) ) != 0 ) {
209 /* Ignore errors; the remote end will retry */
210 }
211 };
212
213 /** Communication management agents */
214 struct ib_mad_agent ib_cm_agent[] __ib_mad_agent = {
215 {
216 .mgmt_class = IB_MGMT_CLASS_CM,
217 .class_version = IB_CM_CLASS_VERSION,
218 .attr_id = htons ( IB_CM_ATTR_CONNECT_REPLY ),
219 .handle = ib_cm_recv_rep,
220 },
221 {
222 .mgmt_class = IB_MGMT_CLASS_CM,
223 .class_version = IB_CM_CLASS_VERSION,
224 .attr_id = htons ( IB_CM_ATTR_DISCONNECT_REQUEST ),
225 .handle = ib_cm_recv_dreq,
226 },
227 };
228
229 /**
230 * Convert connection rejection reason to return status code
231 *
232 * @v reason Rejection reason (in network byte order)
233 * @ret rc Return status code
234 */
ib_cm_rejection_reason_to_rc(uint16_t reason)235 static int ib_cm_rejection_reason_to_rc ( uint16_t reason ) {
236 switch ( reason ) {
237 case htons ( IB_CM_REJECT_BAD_SERVICE_ID ) :
238 return -ENODEV;
239 case htons ( IB_CM_REJECT_STALE_CONN ) :
240 return -EALREADY;
241 case htons ( IB_CM_REJECT_CONSUMER ) :
242 return -ENOTTY;
243 default:
244 return -EPERM;
245 }
246 }
247
248 /**
249 * Handle connection request transaction completion
250 *
251 * @v ibdev Infiniband device
252 * @v mi Management interface
253 * @v madx Management transaction
254 * @v rc Status code
255 * @v mad Received MAD (or NULL on error)
256 * @v av Source address vector (or NULL on error)
257 */
ib_cm_req_complete(struct ib_device * ibdev,struct ib_mad_interface * mi,struct ib_mad_transaction * madx,int rc,union ib_mad * mad,struct ib_address_vector * av)258 static void ib_cm_req_complete ( struct ib_device *ibdev,
259 struct ib_mad_interface *mi,
260 struct ib_mad_transaction *madx,
261 int rc, union ib_mad *mad,
262 struct ib_address_vector *av ) {
263 struct ib_connection *conn = ib_madx_get_ownerdata ( madx );
264 struct ib_queue_pair *qp = conn->qp;
265 struct ib_cm_common *common = &mad->cm.cm_data.common;
266 struct ib_cm_connect_reply *rep = &mad->cm.cm_data.connect_reply;
267 struct ib_cm_connect_reject *rej = &mad->cm.cm_data.connect_reject;
268 uint32_t local_id = conn->local_id;
269 void *private_data = NULL;
270 size_t private_data_len = 0;
271
272 /* Report failures */
273 if ( ( rc == 0 ) && ( mad->hdr.status != htons ( IB_MGMT_STATUS_OK ) ))
274 rc = -EIO;
275 if ( rc != 0 ) {
276 DBGC ( local_id, "CM %08x connection request failed: %s\n",
277 local_id, strerror ( rc ) );
278 goto out;
279 }
280
281 /* Record remote communication ID */
282 conn->remote_id = ntohl ( common->local_id );
283
284 /* Handle response */
285 switch ( mad->hdr.attr_id ) {
286
287 case htons ( IB_CM_ATTR_CONNECT_REPLY ) :
288 /* Extract fields */
289 qp->av.qpn = ( ntohl ( rep->local_qpn ) >> 8 );
290 qp->send.psn = ( ntohl ( rep->starting_psn ) >> 8 );
291 private_data = &rep->private_data;
292 private_data_len = sizeof ( rep->private_data );
293 DBGC ( local_id, "CM %08x connected to QPN %#lx PSN %#x\n",
294 local_id, qp->av.qpn, qp->send.psn );
295
296 /* Modify queue pair */
297 if ( ( rc = ib_modify_qp ( ibdev, qp ) ) != 0 ) {
298 DBGC ( local_id, "CM %08x could not modify queue "
299 "pair: %s\n", local_id, strerror ( rc ) );
300 goto out;
301 }
302
303 /* Send "ready to use" reply */
304 if ( ( rc = ib_cm_send_rtu ( ibdev, mi, &mad->hdr.tid, av,
305 conn->local_id,
306 conn->remote_id ) ) != 0 ) {
307 /* Treat as non-fatal */
308 rc = 0;
309 }
310 break;
311
312 case htons ( IB_CM_ATTR_CONNECT_REJECT ) :
313 /* Extract fields */
314 DBGC ( local_id, "CM %08x connection rejected (reason %d)\n",
315 local_id, ntohs ( rej->reason ) );
316 /* Private data is valid only for a Consumer Reject */
317 if ( rej->reason == htons ( IB_CM_REJECT_CONSUMER ) ) {
318 private_data = &rej->private_data;
319 private_data_len = sizeof ( rej->private_data );
320 }
321 rc = ib_cm_rejection_reason_to_rc ( rej->reason );
322 break;
323
324 default:
325 DBGC ( local_id, "CM %08x unexpected response (attribute "
326 "%04x)\n", local_id, ntohs ( mad->hdr.attr_id ) );
327 rc = -ENOTSUP;
328 break;
329 }
330
331 out:
332 /* Destroy the completed transaction */
333 ib_destroy_madx ( ibdev, ibdev->gsi, madx );
334 conn->madx = NULL;
335
336 /* Hand off to the upper completion handler */
337 conn->op->changed ( ibdev, qp, conn, rc, private_data,
338 private_data_len );
339 }
340
341 /** Connection request operations */
342 static struct ib_mad_transaction_operations ib_cm_req_op = {
343 .complete = ib_cm_req_complete,
344 };
345
346 /**
347 * Handle connection path transaction completion
348 *
349 * @v ibdev Infiniband device
350 * @v path Path
351 * @v rc Status code
352 * @v av Address vector, or NULL on error
353 */
ib_cm_path_complete(struct ib_device * ibdev,struct ib_path * path,int rc,struct ib_address_vector * av)354 static void ib_cm_path_complete ( struct ib_device *ibdev,
355 struct ib_path *path, int rc,
356 struct ib_address_vector *av ) {
357 struct ib_connection *conn = ib_path_get_ownerdata ( path );
358 struct ib_queue_pair *qp = conn->qp;
359 union ib_mad mad;
360 struct ib_cm_connect_request *req = &mad.cm.cm_data.connect_request;
361 uint32_t local_id = conn->local_id;
362 size_t private_data_len;
363
364 /* Report failures */
365 if ( rc != 0 ) {
366 DBGC ( local_id, "CM %08x path lookup failed: %s\n",
367 local_id, strerror ( rc ) );
368 conn->op->changed ( ibdev, qp, conn, rc, NULL, 0 );
369 goto out;
370 }
371
372 /* Update queue pair peer path */
373 memcpy ( &qp->av, av, sizeof ( qp->av ) );
374
375 /* Construct connection request */
376 memset ( &mad, 0, sizeof ( mad ) );
377 mad.hdr.mgmt_class = IB_MGMT_CLASS_CM;
378 mad.hdr.class_version = IB_CM_CLASS_VERSION;
379 mad.hdr.method = IB_MGMT_METHOD_SEND;
380 mad.hdr.attr_id = htons ( IB_CM_ATTR_CONNECT_REQUEST );
381 req->local_id = htonl ( conn->local_id );
382 memcpy ( &req->service_id, &conn->service_id,
383 sizeof ( req->service_id ) );
384 memcpy ( &req->local_ca, &ibdev->node_guid, sizeof ( req->local_ca ) );
385 req->local_qpn__responder_resources = htonl ( ( qp->qpn << 8 ) | 1 );
386 req->local_eecn__initiator_depth = htonl ( ( 0 << 8 ) | 1 );
387 req->remote_eecn__remote_timeout__service_type__ee_flow_ctrl =
388 htonl ( ( 0x14 << 3 ) | ( IB_CM_TRANSPORT_RC << 1 ) |
389 ( 0 << 0 ) );
390 req->starting_psn__local_timeout__retry_count =
391 htonl ( ( qp->recv.psn << 8 ) | ( 0x14 << 3 ) |
392 ( 0x07 << 0 ) );
393 req->pkey = htons ( ibdev->pkey );
394 req->payload_mtu__rdc_exists__rnr_retry =
395 ( ( IB_MTU_2048 << 4 ) | ( 1 << 3 ) | ( 0x07 << 0 ) );
396 req->max_cm_retries__srq = ( ( 0x0f << 4 ) | ( 0 << 3 ) );
397 req->primary.local_lid = htons ( ibdev->lid );
398 req->primary.remote_lid = htons ( conn->qp->av.lid );
399 memcpy ( &req->primary.local_gid, &ibdev->gid,
400 sizeof ( req->primary.local_gid ) );
401 memcpy ( &req->primary.remote_gid, &conn->qp->av.gid,
402 sizeof ( req->primary.remote_gid ) );
403 req->primary.flow_label__rate =
404 htonl ( ( 0 << 12 ) | ( conn->qp->av.rate << 0 ) );
405 req->primary.hop_limit = 0;
406 req->primary.sl__subnet_local =
407 ( ( conn->qp->av.sl << 4 ) | ( 1 << 3 ) );
408 req->primary.local_ack_timeout = ( 0x13 << 3 );
409 private_data_len = conn->private_data_len;
410 if ( private_data_len > sizeof ( req->private_data ) )
411 private_data_len = sizeof ( req->private_data );
412 memcpy ( &req->private_data, &conn->private_data, private_data_len );
413
414 /* Create connection request */
415 av->qpn = IB_QPN_GSI;
416 av->qkey = IB_QKEY_GSI;
417 conn->madx = ib_create_madx ( ibdev, ibdev->gsi, &mad, av,
418 &ib_cm_req_op );
419 if ( ! conn->madx ) {
420 DBGC ( local_id, "CM %08x could not create connection "
421 "request\n", local_id );
422 conn->op->changed ( ibdev, qp, conn, rc, NULL, 0 );
423 goto out;
424 }
425 ib_madx_set_ownerdata ( conn->madx, conn );
426
427 out:
428 /* Destroy the completed transaction */
429 ib_destroy_path ( ibdev, path );
430 conn->path = NULL;
431 }
432
433 /** Connection path operations */
434 static struct ib_path_operations ib_cm_path_op = {
435 .complete = ib_cm_path_complete,
436 };
437
438 /**
439 * Create connection to remote QP
440 *
441 * @v ibdev Infiniband device
442 * @v qp Queue pair
443 * @v dgid Target GID
444 * @v service_id Target service ID
445 * @v private_data Connection request private data
446 * @v private_data_len Length of connection request private data
447 * @v op Connection operations
448 * @ret conn Connection
449 */
450 struct ib_connection *
ib_create_conn(struct ib_device * ibdev,struct ib_queue_pair * qp,union ib_gid * dgid,union ib_guid * service_id,void * private_data,size_t private_data_len,struct ib_connection_operations * op)451 ib_create_conn ( struct ib_device *ibdev, struct ib_queue_pair *qp,
452 union ib_gid *dgid, union ib_guid *service_id,
453 void *private_data, size_t private_data_len,
454 struct ib_connection_operations *op ) {
455 struct ib_connection *conn;
456 uint32_t local_id;
457
458 /* Allocate and initialise request */
459 conn = zalloc ( sizeof ( *conn ) + private_data_len );
460 if ( ! conn )
461 goto err_alloc_conn;
462 conn->ibdev = ibdev;
463 conn->qp = qp;
464 memset ( &qp->av, 0, sizeof ( qp->av ) );
465 qp->av.gid_present = 1;
466 memcpy ( &qp->av.gid, dgid, sizeof ( qp->av.gid ) );
467 conn->local_id = local_id = random();
468 memcpy ( &conn->service_id, service_id, sizeof ( conn->service_id ) );
469 conn->op = op;
470 conn->private_data_len = private_data_len;
471 memcpy ( &conn->private_data, private_data, private_data_len );
472
473 /* Create path */
474 conn->path = ib_create_path ( ibdev, &qp->av, &ib_cm_path_op );
475 if ( ! conn->path )
476 goto err_create_path;
477 ib_path_set_ownerdata ( conn->path, conn );
478
479 /* Add to list of connections */
480 list_add ( &conn->list, &ib_cm_conns );
481
482 DBGC ( local_id, "CM %08x created for IBDEV %s QPN %#lx\n",
483 local_id, ibdev->name, qp->qpn );
484 DBGC ( local_id, "CM %08x connecting to " IB_GID_FMT " "
485 IB_GUID_FMT "\n", local_id, IB_GID_ARGS ( dgid ),
486 IB_GUID_ARGS ( service_id ) );
487
488 return conn;
489
490 ib_destroy_path ( ibdev, conn->path );
491 err_create_path:
492 free ( conn );
493 err_alloc_conn:
494 return NULL;
495 }
496
497 /**
498 * Destroy connection to remote QP
499 *
500 * @v ibdev Infiniband device
501 * @v qp Queue pair
502 * @v conn Connection
503 */
ib_destroy_conn(struct ib_device * ibdev,struct ib_queue_pair * qp __unused,struct ib_connection * conn)504 void ib_destroy_conn ( struct ib_device *ibdev,
505 struct ib_queue_pair *qp __unused,
506 struct ib_connection *conn ) {
507
508 list_del ( &conn->list );
509 if ( conn->madx )
510 ib_destroy_madx ( ibdev, ibdev->gsi, conn->madx );
511 if ( conn->path )
512 ib_destroy_path ( ibdev, conn->path );
513 free ( conn );
514 }
515