• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2007 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., 675 Mass Ave, Cambridge, MA 02139, USA.
17  */
18 
19 FILE_LICENCE ( GPL2_OR_LATER );
20 
21 #include <stdint.h>
22 #include <stdio.h>
23 #include <unistd.h>
24 #include <string.h>
25 #include <byteswap.h>
26 #include <errno.h>
27 #include <gpxe/errortab.h>
28 #include <gpxe/if_arp.h>
29 #include <gpxe/iobuf.h>
30 #include <gpxe/netdevice.h>
31 #include <gpxe/infiniband.h>
32 #include <gpxe/ib_pathrec.h>
33 #include <gpxe/ib_mcast.h>
34 #include <gpxe/ipoib.h>
35 
36 /** @file
37  *
38  * IP over Infiniband
39  */
40 
41 /** Number of IPoIB send work queue entries */
42 #define IPOIB_NUM_SEND_WQES 2
43 
44 /** Number of IPoIB receive work queue entries */
45 #define IPOIB_NUM_RECV_WQES 4
46 
47 /** Number of IPoIB completion entries */
48 #define IPOIB_NUM_CQES 8
49 
50 /** An IPoIB device */
51 struct ipoib_device {
52 	/** Network device */
53 	struct net_device *netdev;
54 	/** Underlying Infiniband device */
55 	struct ib_device *ibdev;
56 	/** Completion queue */
57 	struct ib_completion_queue *cq;
58 	/** Queue pair */
59 	struct ib_queue_pair *qp;
60 	/** Broadcast MAC */
61 	struct ipoib_mac broadcast;
62 	/** Joined to IPv4 broadcast multicast group
63 	 *
64 	 * This flag indicates whether or not we have initiated the
65 	 * join to the IPv4 broadcast multicast group.
66 	 */
67 	int broadcast_joined;
68 	/** IPv4 broadcast multicast group membership */
69 	struct ib_mc_membership broadcast_membership;
70 };
71 
72 /** Broadcast IPoIB address */
73 static struct ipoib_mac ipoib_broadcast = {
74 	.flags__qpn = htonl ( IB_QPN_BROADCAST ),
75 	.gid.u.bytes = 	{ 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
76 			  0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff },
77 };
78 
79 /** Link status for "broadcast join in progress" */
80 #define EINPROGRESS_JOINING ( EINPROGRESS | EUNIQ_01 )
81 
82 /** Human-readable message for the link status */
83 struct errortab ipoib_errors[] __errortab = {
84 	{ EINPROGRESS_JOINING, "Joining" },
85 };
86 
87 /****************************************************************************
88  *
89  * IPoIB peer cache
90  *
91  ****************************************************************************
92  */
93 
94 /**
95  * IPoIB peer address
96  *
97  * The IPoIB link-layer header is only four bytes long and so does not
98  * have sufficient room to store IPoIB MAC address(es).  We therefore
99  * maintain a cache of MAC addresses identified by a single-byte key,
100  * and abuse the spare two bytes within the link-layer header to
101  * communicate these MAC addresses between the link-layer code and the
102  * netdevice driver.
103  */
104 struct ipoib_peer {
105 	/** Key */
106 	uint8_t key;
107 	/** MAC address */
108 	struct ipoib_mac mac;
109 };
110 
111 /** Number of IPoIB peer cache entries
112  *
113  * Must be a power of two.
114  */
115 #define IPOIB_NUM_CACHED_PEERS 4
116 
117 /** IPoIB peer address cache */
118 static struct ipoib_peer ipoib_peer_cache[IPOIB_NUM_CACHED_PEERS];
119 
120 /** Oldest IPoIB peer cache entry index */
121 static unsigned int ipoib_peer_cache_idx = 1;
122 
123 /**
124  * Look up cached peer by key
125  *
126  * @v key		Peer cache key
127  * @ret peer		Peer cache entry, or NULL
128  */
ipoib_lookup_peer_by_key(unsigned int key)129 static struct ipoib_peer * ipoib_lookup_peer_by_key ( unsigned int key ) {
130 	struct ipoib_peer *peer;
131 	unsigned int i;
132 
133 	for ( i = 0 ; i < IPOIB_NUM_CACHED_PEERS ; i++ ) {
134 		peer = &ipoib_peer_cache[i];
135 		if ( peer->key == key )
136 			return peer;
137 	}
138 
139 	if ( key != 0 ) {
140 		DBG ( "IPoIB warning: peer cache lost track of key %x while "
141 		      "still in use\n", key );
142 	}
143 	return NULL;
144 }
145 
146 /**
147  * Store GID and QPN in peer cache
148  *
149  * @v mac		Peer MAC address
150  * @ret peer		Peer cache entry
151  */
ipoib_cache_peer(const struct ipoib_mac * mac)152 static struct ipoib_peer * ipoib_cache_peer ( const struct ipoib_mac *mac ) {
153 	struct ipoib_peer *peer;
154 	unsigned int key;
155 	unsigned int i;
156 
157 	/* Look for existing cache entry */
158 	for ( i = 0 ; i < IPOIB_NUM_CACHED_PEERS ; i++ ) {
159 		peer = &ipoib_peer_cache[i];
160 		if ( memcmp ( &peer->mac, mac, sizeof ( peer->mac ) ) == 0 )
161 			return peer;
162 	}
163 
164 	/* No entry found: create a new one */
165 	key = ipoib_peer_cache_idx++;
166 	peer = &ipoib_peer_cache[ key % IPOIB_NUM_CACHED_PEERS ];
167 	if ( peer->key )
168 		DBG ( "IPoIB peer %x evicted from cache\n", peer->key );
169 
170 	memset ( peer, 0, sizeof ( *peer ) );
171 	peer->key = key;
172 	memcpy ( &peer->mac, mac, sizeof ( peer->mac ) );
173 	DBG ( "IPoIB peer %x has MAC %s\n",
174 	      peer->key, ipoib_ntoa ( &peer->mac ) );
175 	return peer;
176 }
177 
178 /****************************************************************************
179  *
180  * IPoIB link layer
181  *
182  ****************************************************************************
183  */
184 
185 /**
186  * Add IPoIB link-layer header
187  *
188  * @v netdev		Network device
189  * @v iobuf		I/O buffer
190  * @v ll_dest		Link-layer destination address
191  * @v ll_source		Source link-layer address
192  * @v net_proto		Network-layer protocol, in network-byte order
193  * @ret rc		Return status code
194  */
ipoib_push(struct net_device * netdev __unused,struct io_buffer * iobuf,const void * ll_dest,const void * ll_source __unused,uint16_t net_proto)195 static int ipoib_push ( struct net_device *netdev __unused,
196 			struct io_buffer *iobuf, const void *ll_dest,
197 			const void *ll_source __unused, uint16_t net_proto ) {
198 	struct ipoib_hdr *ipoib_hdr =
199 		iob_push ( iobuf, sizeof ( *ipoib_hdr ) );
200 	const struct ipoib_mac *dest_mac = ll_dest;
201 	const struct ipoib_mac *src_mac = ll_source;
202 	struct ipoib_peer *dest;
203 	struct ipoib_peer *src;
204 
205 	/* Add link-layer addresses to cache */
206 	dest = ipoib_cache_peer ( dest_mac );
207 	src = ipoib_cache_peer ( src_mac );
208 
209 	/* Build IPoIB header */
210 	ipoib_hdr->proto = net_proto;
211 	ipoib_hdr->u.peer.dest = dest->key;
212 	ipoib_hdr->u.peer.src = src->key;
213 
214 	return 0;
215 }
216 
217 /**
218  * Remove IPoIB link-layer header
219  *
220  * @v netdev		Network device
221  * @v iobuf		I/O buffer
222  * @ret ll_dest		Link-layer destination address
223  * @ret ll_source	Source link-layer address
224  * @ret net_proto	Network-layer protocol, in network-byte order
225  * @ret rc		Return status code
226  */
ipoib_pull(struct net_device * netdev,struct io_buffer * iobuf,const void ** ll_dest,const void ** ll_source,uint16_t * net_proto)227 static int ipoib_pull ( struct net_device *netdev,
228 			struct io_buffer *iobuf, const void **ll_dest,
229 			const void **ll_source, uint16_t *net_proto ) {
230 	struct ipoib_device *ipoib = netdev->priv;
231 	struct ipoib_hdr *ipoib_hdr = iobuf->data;
232 	struct ipoib_peer *dest;
233 	struct ipoib_peer *source;
234 
235 	/* Sanity check */
236 	if ( iob_len ( iobuf ) < sizeof ( *ipoib_hdr ) ) {
237 		DBG ( "IPoIB packet too short for link-layer header\n" );
238 		DBG_HD ( iobuf->data, iob_len ( iobuf ) );
239 		return -EINVAL;
240 	}
241 
242 	/* Strip off IPoIB header */
243 	iob_pull ( iobuf, sizeof ( *ipoib_hdr ) );
244 
245 	/* Identify source and destination addresses, and clear
246 	 * reserved word in IPoIB header
247 	 */
248 	dest = ipoib_lookup_peer_by_key ( ipoib_hdr->u.peer.dest );
249 	source = ipoib_lookup_peer_by_key ( ipoib_hdr->u.peer.src );
250 	ipoib_hdr->u.reserved = 0;
251 
252 	/* Fill in required fields */
253 	*ll_dest = ( dest ? &dest->mac : &ipoib->broadcast );
254 	*ll_source = ( source ? &source->mac : &ipoib->broadcast );
255 	*net_proto = ipoib_hdr->proto;
256 
257 	return 0;
258 }
259 
260 /**
261  * Initialise IPoIB link-layer address
262  *
263  * @v hw_addr		Hardware address
264  * @v ll_addr		Link-layer address
265  */
ipoib_init_addr(const void * hw_addr,void * ll_addr)266 static void ipoib_init_addr ( const void *hw_addr, void *ll_addr ) {
267 	const struct ib_gid_half *guid = hw_addr;
268 	struct ipoib_mac *mac = ll_addr;
269 
270 	memset ( mac, 0, sizeof ( *mac ) );
271 	memcpy ( &mac->gid.u.half[1], guid, sizeof ( mac->gid.u.half[1] ) );
272 }
273 
274 /**
275  * Transcribe IPoIB link-layer address
276  *
277  * @v ll_addr	Link-layer address
278  * @ret string	Link-layer address in human-readable format
279  */
ipoib_ntoa(const void * ll_addr)280 const char * ipoib_ntoa ( const void *ll_addr ) {
281 	static char buf[45];
282 	const struct ipoib_mac *mac = ll_addr;
283 
284 	snprintf ( buf, sizeof ( buf ), "%08x:%08x:%08x:%08x:%08x",
285 		   htonl ( mac->flags__qpn ), htonl ( mac->gid.u.dwords[0] ),
286 		   htonl ( mac->gid.u.dwords[1] ),
287 		   htonl ( mac->gid.u.dwords[2] ),
288 		   htonl ( mac->gid.u.dwords[3] ) );
289 	return buf;
290 }
291 
292 /**
293  * Hash multicast address
294  *
295  * @v af		Address family
296  * @v net_addr		Network-layer address
297  * @v ll_addr		Link-layer address to fill in
298  * @ret rc		Return status code
299  */
ipoib_mc_hash(unsigned int af __unused,const void * net_addr __unused,void * ll_addr __unused)300 static int ipoib_mc_hash ( unsigned int af __unused,
301 			   const void *net_addr __unused,
302 			   void *ll_addr __unused ) {
303 
304 	return -ENOTSUP;
305 }
306 
307 /**
308  * Generate Mellanox Ethernet-compatible compressed link-layer address
309  *
310  * @v ll_addr		Link-layer address
311  * @v eth_addr		Ethernet-compatible address to fill in
312  */
ipoib_mlx_eth_addr(const struct ib_gid_half * guid,uint8_t * eth_addr)313 static int ipoib_mlx_eth_addr ( const struct ib_gid_half *guid,
314 				uint8_t *eth_addr ) {
315 	eth_addr[0] = ( ( guid->u.bytes[3] == 2 ) ? 0x00 : 0x02 );
316 	eth_addr[1] = guid->u.bytes[1];
317 	eth_addr[2] = guid->u.bytes[2];
318 	eth_addr[3] = guid->u.bytes[5];
319 	eth_addr[4] = guid->u.bytes[6];
320 	eth_addr[5] = guid->u.bytes[7];
321 	return 0;
322 }
323 
324 /** An IPoIB Ethernet-compatible compressed link-layer address generator */
325 struct ipoib_eth_addr_handler {
326 	/** GUID byte 1 */
327 	uint8_t byte1;
328 	/** GUID byte 2 */
329 	uint8_t byte2;
330 	/** Handler */
331 	int ( * eth_addr ) ( const struct ib_gid_half *guid,
332 			     uint8_t *eth_addr );
333 };
334 
335 /** IPoIB Ethernet-compatible compressed link-layer address generators */
336 static struct ipoib_eth_addr_handler ipoib_eth_addr_handlers[] = {
337 	{ 0x02, 0xc9, ipoib_mlx_eth_addr },
338 };
339 
340 /**
341  * Generate Ethernet-compatible compressed link-layer address
342  *
343  * @v ll_addr		Link-layer address
344  * @v eth_addr		Ethernet-compatible address to fill in
345  */
ipoib_eth_addr(const void * ll_addr,void * eth_addr)346 static int ipoib_eth_addr ( const void *ll_addr, void *eth_addr ) {
347 	const struct ipoib_mac *ipoib_addr = ll_addr;
348 	const struct ib_gid_half *guid = &ipoib_addr->gid.u.half[1];
349 	struct ipoib_eth_addr_handler *handler;
350 	unsigned int i;
351 
352 	for ( i = 0 ; i < ( sizeof ( ipoib_eth_addr_handlers ) /
353 			    sizeof ( ipoib_eth_addr_handlers[0] ) ) ; i++ ) {
354 		handler = &ipoib_eth_addr_handlers[i];
355 		if ( ( handler->byte1 == guid->u.bytes[1] ) &&
356 		     ( handler->byte2 == guid->u.bytes[2] ) ) {
357 			return handler->eth_addr ( guid, eth_addr );
358 		}
359 	}
360 	return -ENOTSUP;
361 }
362 
363 /** IPoIB protocol */
364 struct ll_protocol ipoib_protocol __ll_protocol = {
365 	.name		= "IPoIB",
366 	.ll_proto	= htons ( ARPHRD_INFINIBAND ),
367 	.hw_addr_len	= sizeof ( struct ib_gid_half ),
368 	.ll_addr_len	= IPOIB_ALEN,
369 	.ll_header_len	= IPOIB_HLEN,
370 	.push		= ipoib_push,
371 	.pull		= ipoib_pull,
372 	.init_addr	= ipoib_init_addr,
373 	.ntoa		= ipoib_ntoa,
374 	.mc_hash	= ipoib_mc_hash,
375 	.eth_addr	= ipoib_eth_addr,
376 };
377 
378 /**
379  * Allocate IPoIB device
380  *
381  * @v priv_size		Size of driver private data
382  * @ret netdev		Network device, or NULL
383  */
alloc_ipoibdev(size_t priv_size)384 struct net_device * alloc_ipoibdev ( size_t priv_size ) {
385 	struct net_device *netdev;
386 
387 	netdev = alloc_netdev ( priv_size );
388 	if ( netdev ) {
389 		netdev->ll_protocol = &ipoib_protocol;
390 		netdev->ll_broadcast = ( uint8_t * ) &ipoib_broadcast;
391 		netdev->max_pkt_len = IB_MAX_PAYLOAD_SIZE;
392 	}
393 	return netdev;
394 }
395 
396 /****************************************************************************
397  *
398  * IPoIB network device
399  *
400  ****************************************************************************
401  */
402 
403 /**
404  * Transmit packet via IPoIB network device
405  *
406  * @v netdev		Network device
407  * @v iobuf		I/O buffer
408  * @ret rc		Return status code
409  */
ipoib_transmit(struct net_device * netdev,struct io_buffer * iobuf)410 static int ipoib_transmit ( struct net_device *netdev,
411 			    struct io_buffer *iobuf ) {
412 	struct ipoib_device *ipoib = netdev->priv;
413 	struct ib_device *ibdev = ipoib->ibdev;
414 	struct ipoib_hdr *ipoib_hdr;
415 	struct ipoib_peer *dest;
416 	struct ib_address_vector av;
417 	int rc;
418 
419 	/* Sanity check */
420 	if ( iob_len ( iobuf ) < sizeof ( *ipoib_hdr ) ) {
421 		DBGC ( ipoib, "IPoIB %p buffer too short\n", ipoib );
422 		return -EINVAL;
423 	}
424 	ipoib_hdr = iobuf->data;
425 
426 	/* Attempting transmission while link is down will put the
427 	 * queue pair into an error state, so don't try it.
428 	 */
429 	if ( ! ib_link_ok ( ibdev ) )
430 		return -ENETUNREACH;
431 
432 	/* Identify destination address */
433 	dest = ipoib_lookup_peer_by_key ( ipoib_hdr->u.peer.dest );
434 	if ( ! dest )
435 		return -ENXIO;
436 	ipoib_hdr->u.reserved = 0;
437 
438 	/* Construct address vector */
439 	memset ( &av, 0, sizeof ( av ) );
440 	av.qpn = ( ntohl ( dest->mac.flags__qpn ) & IB_QPN_MASK );
441 	av.gid_present = 1;
442 	memcpy ( &av.gid, &dest->mac.gid, sizeof ( av.gid ) );
443 	if ( ( rc = ib_resolve_path ( ibdev, &av ) ) != 0 ) {
444 		/* Path not resolved yet */
445 		return rc;
446 	}
447 
448 	return ib_post_send ( ibdev, ipoib->qp, &av, iobuf );
449 }
450 
451 /**
452  * Handle IPoIB send completion
453  *
454  * @v ibdev		Infiniband device
455  * @v qp		Queue pair
456  * @v iobuf		I/O buffer
457  * @v rc		Completion status code
458  */
ipoib_complete_send(struct ib_device * ibdev __unused,struct ib_queue_pair * qp,struct io_buffer * iobuf,int rc)459 static void ipoib_complete_send ( struct ib_device *ibdev __unused,
460 				  struct ib_queue_pair *qp,
461 				  struct io_buffer *iobuf, int rc ) {
462 	struct ipoib_device *ipoib = ib_qp_get_ownerdata ( qp );
463 
464 	netdev_tx_complete_err ( ipoib->netdev, iobuf, rc );
465 }
466 
467 /**
468  * Handle IPoIB receive completion
469  *
470  * @v ibdev		Infiniband device
471  * @v qp		Queue pair
472  * @v av		Address vector, or NULL
473  * @v iobuf		I/O buffer
474  * @v rc		Completion status code
475  */
ipoib_complete_recv(struct ib_device * ibdev __unused,struct ib_queue_pair * qp,struct ib_address_vector * av,struct io_buffer * iobuf,int rc)476 static void ipoib_complete_recv ( struct ib_device *ibdev __unused,
477 				  struct ib_queue_pair *qp,
478 				  struct ib_address_vector *av,
479 				  struct io_buffer *iobuf, int rc ) {
480 	struct ipoib_device *ipoib = ib_qp_get_ownerdata ( qp );
481 	struct net_device *netdev = ipoib->netdev;
482 	struct ipoib_hdr *ipoib_hdr;
483 	struct ipoib_mac ll_src;
484 	struct ipoib_peer *src;
485 
486 	if ( rc != 0 ) {
487 		netdev_rx_err ( netdev, iobuf, rc );
488 		return;
489 	}
490 
491 	/* Sanity check */
492 	if ( iob_len ( iobuf ) < sizeof ( struct ipoib_hdr ) ) {
493 		DBGC ( ipoib, "IPoIB %p received packet too short to "
494 		       "contain IPoIB header\n", ipoib );
495 		DBGC_HD ( ipoib, iobuf->data, iob_len ( iobuf ) );
496 		netdev_rx_err ( netdev, iobuf, -EIO );
497 		return;
498 	}
499 	ipoib_hdr = iobuf->data;
500 
501 	/* Parse source address */
502 	if ( av->gid_present ) {
503 		ll_src.flags__qpn = htonl ( av->qpn );
504 		memcpy ( &ll_src.gid, &av->gid, sizeof ( ll_src.gid ) );
505 		src = ipoib_cache_peer ( &ll_src );
506 		ipoib_hdr->u.peer.src = src->key;
507 	}
508 
509 	/* Hand off to network layer */
510 	netdev_rx ( netdev, iobuf );
511 }
512 
513 /** IPoIB completion operations */
514 static struct ib_completion_queue_operations ipoib_cq_op = {
515 	.complete_send = ipoib_complete_send,
516 	.complete_recv = ipoib_complete_recv,
517 };
518 
519 /**
520  * Poll IPoIB network device
521  *
522  * @v netdev		Network device
523  */
ipoib_poll(struct net_device * netdev)524 static void ipoib_poll ( struct net_device *netdev ) {
525 	struct ipoib_device *ipoib = netdev->priv;
526 	struct ib_device *ibdev = ipoib->ibdev;
527 
528 	ib_poll_eq ( ibdev );
529 }
530 
531 /**
532  * Enable/disable interrupts on IPoIB network device
533  *
534  * @v netdev		Network device
535  * @v enable		Interrupts should be enabled
536  */
ipoib_irq(struct net_device * netdev __unused,int enable __unused)537 static void ipoib_irq ( struct net_device *netdev __unused,
538 			int enable __unused ) {
539 	/* No implementation */
540 }
541 
542 /**
543  * Handle IPv4 broadcast multicast group join completion
544  *
545  * @v ibdev		Infiniband device
546  * @v qp		Queue pair
547  * @v membership	Multicast group membership
548  * @v rc		Status code
549  * @v mad		Response MAD (or NULL on error)
550  */
ipoib_join_complete(struct ib_device * ibdev __unused,struct ib_queue_pair * qp __unused,struct ib_mc_membership * membership,int rc,union ib_mad * mad __unused)551 void ipoib_join_complete ( struct ib_device *ibdev __unused,
552 			   struct ib_queue_pair *qp __unused,
553 			   struct ib_mc_membership *membership, int rc,
554 			   union ib_mad *mad __unused ) {
555 	struct ipoib_device *ipoib = container_of ( membership,
556 				   struct ipoib_device, broadcast_membership );
557 
558 	/* Record join status as link status */
559 	netdev_link_err ( ipoib->netdev, rc );
560 }
561 
562 /**
563  * Join IPv4 broadcast multicast group
564  *
565  * @v ipoib		IPoIB device
566  * @ret rc		Return status code
567  */
ipoib_join_broadcast_group(struct ipoib_device * ipoib)568 static int ipoib_join_broadcast_group ( struct ipoib_device *ipoib ) {
569 	int rc;
570 
571 	if ( ( rc = ib_mcast_join ( ipoib->ibdev, ipoib->qp,
572 				    &ipoib->broadcast_membership,
573 				    &ipoib->broadcast.gid,
574 				    ipoib_join_complete ) ) != 0 ) {
575 		DBGC ( ipoib, "IPoIB %p could not join broadcast group: %s\n",
576 		       ipoib, strerror ( rc ) );
577 		return rc;
578 	}
579 	ipoib->broadcast_joined = 1;
580 
581 	return 0;
582 }
583 
584 /**
585  * Leave IPv4 broadcast multicast group
586  *
587  * @v ipoib		IPoIB device
588  */
ipoib_leave_broadcast_group(struct ipoib_device * ipoib)589 static void ipoib_leave_broadcast_group ( struct ipoib_device *ipoib ) {
590 
591 	if ( ipoib->broadcast_joined ) {
592 		ib_mcast_leave ( ipoib->ibdev, ipoib->qp,
593 				 &ipoib->broadcast_membership );
594 		ipoib->broadcast_joined = 0;
595 	}
596 }
597 
598 /**
599  * Open IPoIB network device
600  *
601  * @v netdev		Network device
602  * @ret rc		Return status code
603  */
ipoib_open(struct net_device * netdev)604 static int ipoib_open ( struct net_device *netdev ) {
605 	struct ipoib_device *ipoib = netdev->priv;
606 	struct ib_device *ibdev = ipoib->ibdev;
607 	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
608 	int rc;
609 
610 	/* Open IB device */
611 	if ( ( rc = ib_open ( ibdev ) ) != 0 ) {
612 		DBGC ( ipoib, "IPoIB %p could not open device: %s\n",
613 		       ipoib, strerror ( rc ) );
614 		goto err_ib_open;
615 	}
616 
617 	/* Allocate completion queue */
618 	ipoib->cq = ib_create_cq ( ibdev, IPOIB_NUM_CQES, &ipoib_cq_op );
619 	if ( ! ipoib->cq ) {
620 		DBGC ( ipoib, "IPoIB %p could not allocate completion queue\n",
621 		       ipoib );
622 		rc = -ENOMEM;
623 		goto err_create_cq;
624 	}
625 
626 	/* Allocate queue pair */
627 	ipoib->qp = ib_create_qp ( ibdev, IB_QPT_UD,
628 				   IPOIB_NUM_SEND_WQES, ipoib->cq,
629 				   IPOIB_NUM_RECV_WQES, ipoib->cq );
630 	if ( ! ipoib->qp ) {
631 		DBGC ( ipoib, "IPoIB %p could not allocate queue pair\n",
632 		       ipoib );
633 		rc = -ENOMEM;
634 		goto err_create_qp;
635 	}
636 	ib_qp_set_ownerdata ( ipoib->qp, ipoib );
637 
638 	/* Update MAC address with QPN */
639 	mac->flags__qpn = htonl ( ipoib->qp->qpn );
640 
641 	/* Fill receive rings */
642 	ib_refill_recv ( ibdev, ipoib->qp );
643 
644 	/* Fake a link status change to join the broadcast group */
645 	ipoib_link_state_changed ( ibdev );
646 
647 	return 0;
648 
649 	ib_destroy_qp ( ibdev, ipoib->qp );
650  err_create_qp:
651 	ib_destroy_cq ( ibdev, ipoib->cq );
652  err_create_cq:
653 	ib_close ( ibdev );
654  err_ib_open:
655 	return rc;
656 }
657 
658 /**
659  * Close IPoIB network device
660  *
661  * @v netdev		Network device
662  */
ipoib_close(struct net_device * netdev)663 static void ipoib_close ( struct net_device *netdev ) {
664 	struct ipoib_device *ipoib = netdev->priv;
665 	struct ib_device *ibdev = ipoib->ibdev;
666 	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
667 
668 	/* Leave broadcast group */
669 	ipoib_leave_broadcast_group ( ipoib );
670 
671 	/* Remove QPN from MAC address */
672 	mac->flags__qpn = 0;
673 
674 	/* Tear down the queues */
675 	ib_destroy_qp ( ibdev, ipoib->qp );
676 	ib_destroy_cq ( ibdev, ipoib->cq );
677 
678 	/* Close IB device */
679 	ib_close ( ibdev );
680 }
681 
682 /** IPoIB network device operations */
683 static struct net_device_operations ipoib_operations = {
684 	.open		= ipoib_open,
685 	.close		= ipoib_close,
686 	.transmit	= ipoib_transmit,
687 	.poll		= ipoib_poll,
688 	.irq		= ipoib_irq,
689 };
690 
691 /**
692  * Handle link status change
693  *
694  * @v ibdev		Infiniband device
695  */
ipoib_link_state_changed(struct ib_device * ibdev)696 void ipoib_link_state_changed ( struct ib_device *ibdev ) {
697 	struct net_device *netdev = ib_get_ownerdata ( ibdev );
698 	struct ipoib_device *ipoib = netdev->priv;
699 	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
700 	int rc;
701 
702 	/* Leave existing broadcast group */
703 	ipoib_leave_broadcast_group ( ipoib );
704 
705 	/* Update MAC address based on potentially-new GID prefix */
706 	memcpy ( &mac->gid.u.half[0], &ibdev->gid.u.half[0],
707 		 sizeof ( mac->gid.u.half[0] ) );
708 
709 	/* Update broadcast GID based on potentially-new partition key */
710 	ipoib->broadcast.gid.u.words[2] =
711 		htons ( ibdev->pkey | IB_PKEY_FULL );
712 
713 	/* Set net device link state to reflect Infiniband link state */
714 	rc = ib_link_rc ( ibdev );
715 	netdev_link_err ( netdev, ( rc ? rc : -EINPROGRESS_JOINING ) );
716 
717 	/* Join new broadcast group */
718 	if ( ib_link_ok ( ibdev ) &&
719 	     ( ( rc = ipoib_join_broadcast_group ( ipoib ) ) != 0 ) ) {
720 		DBGC ( ipoib, "IPoIB %p could not rejoin broadcast group: "
721 		       "%s\n", ipoib, strerror ( rc ) );
722 		netdev_link_err ( netdev, rc );
723 		return;
724 	}
725 }
726 
727 /**
728  * Probe IPoIB device
729  *
730  * @v ibdev		Infiniband device
731  * @ret rc		Return status code
732  */
ipoib_probe(struct ib_device * ibdev)733 int ipoib_probe ( struct ib_device *ibdev ) {
734 	struct net_device *netdev;
735 	struct ipoib_device *ipoib;
736 	int rc;
737 
738 	/* Allocate network device */
739 	netdev = alloc_ipoibdev ( sizeof ( *ipoib ) );
740 	if ( ! netdev )
741 		return -ENOMEM;
742 	netdev_init ( netdev, &ipoib_operations );
743 	ipoib = netdev->priv;
744 	ib_set_ownerdata ( ibdev, netdev );
745 	netdev->dev = ibdev->dev;
746 	memset ( ipoib, 0, sizeof ( *ipoib ) );
747 	ipoib->netdev = netdev;
748 	ipoib->ibdev = ibdev;
749 
750 	/* Extract hardware address */
751 	memcpy ( netdev->hw_addr, &ibdev->gid.u.half[1],
752 		 sizeof ( ibdev->gid.u.half[1] ) );
753 
754 	/* Set default broadcast address */
755 	memcpy ( &ipoib->broadcast, &ipoib_broadcast,
756 		 sizeof ( ipoib->broadcast ) );
757 	netdev->ll_broadcast = ( ( uint8_t * ) &ipoib->broadcast );
758 
759 	/* Register network device */
760 	if ( ( rc = register_netdev ( netdev ) ) != 0 )
761 		goto err_register_netdev;
762 
763 	return 0;
764 
765  err_register_netdev:
766 	netdev_nullify ( netdev );
767 	netdev_put ( netdev );
768 	return rc;
769 }
770 
771 /**
772  * Remove IPoIB device
773  *
774  * @v ibdev		Infiniband device
775  */
ipoib_remove(struct ib_device * ibdev)776 void ipoib_remove ( struct ib_device *ibdev ) {
777 	struct net_device *netdev = ib_get_ownerdata ( ibdev );
778 
779 	unregister_netdev ( netdev );
780 	netdev_nullify ( netdev );
781 	netdev_put ( netdev );
782 }
783