2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Copyright (C) 1999-2008 Silicon Graphics, Inc. All rights reserved.
10 * Cross Partition Network Interface (XPNET) support
12 * XPNET provides a virtual network layered on top of the Cross
13 * Partition communication layer.
15 * XPNET provides direct point-to-point and broadcast-like support
16 * for an ethernet-like device. The ethernet broadcast medium is
17 * replaced with a point-to-point message structure which passes
18 * pointers to a DMA-capable block that a remote partition should
19 * retrieve and pass to the upper level networking layer.
23 #include <linux/module.h>
24 #include <linux/netdevice.h>
25 #include <linux/etherdevice.h>
29 * The message payload transferred by XPC.
31 * buf_pa is the physical address where the DMA should pull from.
33 * NOTE: for performance reasons, buf_pa should _ALWAYS_ begin on a
34 * cacheline boundary. To accomplish this, we record the number of
35 * bytes from the beginning of the first cacheline to the first useful
36 * byte of the skb (leadin_ignore) and the number of bytes from the
37 * last useful byte of the skb to the end of the last cacheline
40 * size is the number of bytes to transfer which includes the skb->len
41 * (useful bytes of the senders skb) plus the leadin and tailout
43 struct xpnet_message {
44 u16 version; /* Version for this message */
45 u16 embedded_bytes; /* #of bytes embedded in XPC message */
46 u32 magic; /* Special number indicating this is xpnet */
47 u64 buf_pa; /* phys address of buffer to retrieve */
48 u32 size; /* #of bytes in buffer */
49 u8 leadin_ignore; /* #of bytes to ignore at the beginning */
50 u8 tailout_ignore; /* #of bytes to ignore at the end */
51 unsigned char data; /* body of small packets */
55 * Determine the size of our message, the cacheline aligned size,
56 * and then the number of message will request from XPC.
58 * XPC expects each message to exist in an individual cacheline.
60 #define XPNET_MSG_SIZE (L1_CACHE_BYTES - XPC_MSG_PAYLOAD_OFFSET)
61 #define XPNET_MSG_DATA_MAX \
62 (XPNET_MSG_SIZE - (u64)(&((struct xpnet_message *)0)->data))
63 #define XPNET_MSG_ALIGNED_SIZE (L1_CACHE_ALIGN(XPNET_MSG_SIZE))
64 #define XPNET_MSG_NENTRIES (PAGE_SIZE / XPNET_MSG_ALIGNED_SIZE)
66 #define XPNET_MAX_KTHREADS (XPNET_MSG_NENTRIES + 1)
67 #define XPNET_MAX_IDLE_KTHREADS (XPNET_MSG_NENTRIES + 1)
70 * Version number of XPNET implementation. XPNET can always talk to versions
71 * with same major #, and never talk to versions with a different version.
73 #define _XPNET_VERSION(_major, _minor) (((_major) << 4) | (_minor))
74 #define XPNET_VERSION_MAJOR(_v) ((_v) >> 4)
75 #define XPNET_VERSION_MINOR(_v) ((_v) & 0xf)
77 #define XPNET_VERSION _XPNET_VERSION(1, 0) /* version 1.0 */
78 #define XPNET_VERSION_EMBED _XPNET_VERSION(1, 1) /* version 1.1 */
79 #define XPNET_MAGIC 0x88786984 /* "XNET" */
81 #define XPNET_VALID_MSG(_m) \
82 ((XPNET_VERSION_MAJOR(_m->version) == XPNET_VERSION_MAJOR(XPNET_VERSION)) \
83 && (msg->magic == XPNET_MAGIC))
85 #define XPNET_DEVICE_NAME "xp0"
88 * When messages are queued with xpc_send_notify, a kmalloc'd buffer
89 * of the following type is passed as a notification cookie. When the
90 * notification function is called, we use the cookie to decide
91 * whether all outstanding message sends have completed. The skb can
94 struct xpnet_pending_msg {
99 /* driver specific structure pointed to by the device structure */
100 struct xpnet_dev_private {
101 struct net_device_stats stats;
104 struct net_device *xpnet_device;
107 * When we are notified of other partitions activating, we add them to
108 * our bitmask of partitions to which we broadcast.
110 static unsigned long *xpnet_broadcast_partitions;
112 static DEFINE_SPINLOCK(xpnet_broadcast_lock);
115 * Since the Block Transfer Engine (BTE) is being used for the transfer
116 * and it relies upon cache-line size transfers, we need to reserve at
117 * least one cache-line for head and tail alignment. The BTE is
118 * limited to 8MB transfers.
120 * Testing has shown that changing MTU to greater than 64KB has no effect
121 * on TCP as the two sides negotiate a Max Segment Size that is limited
122 * to 64K. Other protocols May use packets greater than this, but for
123 * now, the default is 64KB.
125 #define XPNET_MAX_MTU (0x800000UL - L1_CACHE_BYTES)
126 /* 32KB has been determined to be the ideal */
127 #define XPNET_DEF_MTU (0x8000UL)
130 * The partid is encapsulated in the MAC address beginning in the following
131 * octet and it consists of two octets.
133 #define XPNET_PARTID_OCTET 2
135 /* Define the XPNET debug device structures to be used with dev_dbg() et al */
137 struct device_driver xpnet_dbg_name = {
141 struct device xpnet_dbg_subname = {
142 .bus_id = {0}, /* set to "" */
143 .driver = &xpnet_dbg_name
146 struct device *xpnet = &xpnet_dbg_subname;
149 * Packet was recevied by XPC and forwarded to us.
152 xpnet_receive(short partid, int channel, struct xpnet_message *msg)
156 struct xpnet_dev_private *priv =
157 (struct xpnet_dev_private *)xpnet_device->priv;
159 if (!XPNET_VALID_MSG(msg)) {
161 * Packet with a different XPC version. Ignore.
163 xpc_received(partid, channel, (void *)msg);
165 priv->stats.rx_errors++;
169 dev_dbg(xpnet, "received 0x%lx, %d, %d, %d\n",
170 (unsigned long)msg->buf_pa, msg->size, msg->leadin_ignore,
171 msg->tailout_ignore);
173 /* reserve an extra cache line */
174 skb = dev_alloc_skb(msg->size + L1_CACHE_BYTES);
176 dev_err(xpnet, "failed on dev_alloc_skb(%d)\n",
177 msg->size + L1_CACHE_BYTES);
179 xpc_received(partid, channel, (void *)msg);
181 priv->stats.rx_errors++;
187 * The allocated skb has some reserved space.
188 * In order to use xp_remote_memcpy(), we need to get the
189 * skb->data pointer moved forward.
191 skb_reserve(skb, (L1_CACHE_BYTES - ((u64)skb->data &
192 (L1_CACHE_BYTES - 1)) +
193 msg->leadin_ignore));
196 * Update the tail pointer to indicate data actually
199 skb_put(skb, (msg->size - msg->leadin_ignore - msg->tailout_ignore));
202 * Move the data over from the other side.
204 if ((XPNET_VERSION_MINOR(msg->version) == 1) &&
205 (msg->embedded_bytes != 0)) {
206 dev_dbg(xpnet, "copying embedded message. memcpy(0x%p, 0x%p, "
207 "%lu)\n", skb->data, &msg->data,
208 (size_t)msg->embedded_bytes);
210 skb_copy_to_linear_data(skb, &msg->data,
211 (size_t)msg->embedded_bytes);
213 dev_dbg(xpnet, "transferring buffer to the skb->data area;\n\t"
214 "xp_remote_memcpy(0x%p, 0x%p, %hu)\n", (void *)
215 ((u64)skb->data & ~(L1_CACHE_BYTES - 1)),
216 (void *)msg->buf_pa, msg->size);
218 ret = xp_remote_memcpy((void *)((u64)skb->data &
219 ~(L1_CACHE_BYTES - 1)),
220 (void *)msg->buf_pa, msg->size);
222 if (ret != xpSuccess) {
224 * !!! Need better way of cleaning skb. Currently skb
225 * !!! appears in_use and we can't just call
228 dev_err(xpnet, "xp_remote_memcpy(0x%p, 0x%p, 0x%hx) "
229 "returned error=0x%x\n", (void *)
230 ((u64)skb->data & ~(L1_CACHE_BYTES - 1)),
231 (void *)msg->buf_pa, msg->size, ret);
233 xpc_received(partid, channel, (void *)msg);
235 priv->stats.rx_errors++;
241 dev_dbg(xpnet, "<skb->head=0x%p skb->data=0x%p skb->tail=0x%p "
242 "skb->end=0x%p skb->len=%d\n", (void *)skb->head,
243 (void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb),
246 skb->protocol = eth_type_trans(skb, xpnet_device);
247 skb->ip_summed = CHECKSUM_UNNECESSARY;
249 dev_dbg(xpnet, "passing skb to network layer\n"
250 KERN_DEBUG "\tskb->head=0x%p skb->data=0x%p skb->tail=0x%p "
251 "skb->end=0x%p skb->len=%d\n",
252 (void *)skb->head, (void *)skb->data, skb_tail_pointer(skb),
253 skb_end_pointer(skb), skb->len);
255 xpnet_device->last_rx = jiffies;
256 priv->stats.rx_packets++;
257 priv->stats.rx_bytes += skb->len + ETH_HLEN;
260 xpc_received(partid, channel, (void *)msg);
264 * This is the handler which XPC calls during any sort of change in
265 * state or message reception on a connection.
268 xpnet_connection_activity(enum xp_retval reason, short partid, int channel,
269 void *data, void *key)
271 DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
272 DBUG_ON(channel != XPC_NET_CHANNEL);
275 case xpMsgReceived: /* message received */
276 DBUG_ON(data == NULL);
278 xpnet_receive(partid, channel, (struct xpnet_message *)data);
281 case xpConnected: /* connection completed to a partition */
282 spin_lock_bh(&xpnet_broadcast_lock);
283 __set_bit(partid, xpnet_broadcast_partitions);
284 spin_unlock_bh(&xpnet_broadcast_lock);
286 netif_carrier_on(xpnet_device);
288 dev_dbg(xpnet, "%s connected to partition %d\n",
289 xpnet_device->name, partid);
293 spin_lock_bh(&xpnet_broadcast_lock);
294 __clear_bit(partid, xpnet_broadcast_partitions);
295 spin_unlock_bh(&xpnet_broadcast_lock);
297 if (bitmap_empty((unsigned long *)xpnet_broadcast_partitions,
298 xp_max_npartitions)) {
299 netif_carrier_off(xpnet_device);
302 dev_dbg(xpnet, "%s disconnected from partition %d\n",
303 xpnet_device->name, partid);
309 xpnet_dev_open(struct net_device *dev)
313 dev_dbg(xpnet, "calling xpc_connect(%d, 0x%p, NULL, %ld, %ld, %ld, "
314 "%ld)\n", XPC_NET_CHANNEL, xpnet_connection_activity,
315 (unsigned long)XPNET_MSG_SIZE,
316 (unsigned long)XPNET_MSG_NENTRIES,
317 (unsigned long)XPNET_MAX_KTHREADS,
318 (unsigned long)XPNET_MAX_IDLE_KTHREADS);
320 ret = xpc_connect(XPC_NET_CHANNEL, xpnet_connection_activity, NULL,
321 XPNET_MSG_SIZE, XPNET_MSG_NENTRIES,
322 XPNET_MAX_KTHREADS, XPNET_MAX_IDLE_KTHREADS);
323 if (ret != xpSuccess) {
324 dev_err(xpnet, "ifconfig up of %s failed on XPC connect, "
325 "ret=%d\n", dev->name, ret);
330 dev_dbg(xpnet, "ifconfig up of %s; XPC connected\n", dev->name);
336 xpnet_dev_stop(struct net_device *dev)
338 xpc_disconnect(XPC_NET_CHANNEL);
340 dev_dbg(xpnet, "ifconfig down of %s; XPC disconnected\n", dev->name);
346 xpnet_dev_change_mtu(struct net_device *dev, int new_mtu)
348 /* 68 comes from min TCP+IP+MAC header */
349 if ((new_mtu < 68) || (new_mtu > XPNET_MAX_MTU)) {
350 dev_err(xpnet, "ifconfig %s mtu %d failed; value must be "
351 "between 68 and %ld\n", dev->name, new_mtu,
357 dev_dbg(xpnet, "ifconfig %s mtu set to %d\n", dev->name, new_mtu);
362 * Required for the net_device structure.
365 xpnet_dev_set_config(struct net_device *dev, struct ifmap *new_map)
371 * Return statistics to the caller.
373 static struct net_device_stats *
374 xpnet_dev_get_stats(struct net_device *dev)
376 struct xpnet_dev_private *priv;
378 priv = (struct xpnet_dev_private *)dev->priv;
384 * Notification that the other end has received the message and
385 * DMA'd the skb information. At this point, they are done with
386 * our side. When all recipients are done processing, we
387 * release the skb and then release our pending message structure.
390 xpnet_send_completed(enum xp_retval reason, short partid, int channel,
393 struct xpnet_pending_msg *queued_msg = (struct xpnet_pending_msg *)__qm;
395 DBUG_ON(queued_msg == NULL);
397 dev_dbg(xpnet, "message to %d notified with reason %d\n",
400 if (atomic_dec_return(&queued_msg->use_count) == 0) {
401 dev_dbg(xpnet, "all acks for skb->head=-x%p\n",
402 (void *)queued_msg->skb->head);
404 dev_kfree_skb_any(queued_msg->skb);
410 xpnet_send(struct sk_buff *skb, struct xpnet_pending_msg *queued_msg,
411 u64 start_addr, u64 end_addr, u16 embedded_bytes, int dest_partid)
413 u8 msg_buffer[XPNET_MSG_SIZE];
414 struct xpnet_message *msg = (struct xpnet_message *)&msg_buffer;
417 msg->embedded_bytes = embedded_bytes;
418 if (unlikely(embedded_bytes != 0)) {
419 msg->version = XPNET_VERSION_EMBED;
420 dev_dbg(xpnet, "calling memcpy(0x%p, 0x%p, 0x%lx)\n",
421 &msg->data, skb->data, (size_t)embedded_bytes);
422 skb_copy_from_linear_data(skb, &msg->data,
423 (size_t)embedded_bytes);
425 msg->version = XPNET_VERSION;
427 msg->magic = XPNET_MAGIC;
428 msg->size = end_addr - start_addr;
429 msg->leadin_ignore = (u64)skb->data - start_addr;
430 msg->tailout_ignore = end_addr - (u64)skb_tail_pointer(skb);
431 msg->buf_pa = __pa(start_addr);
433 dev_dbg(xpnet, "sending XPC message to %d:%d\n"
434 KERN_DEBUG "msg->buf_pa=0x%lx, msg->size=%u, "
435 "msg->leadin_ignore=%u, msg->tailout_ignore=%u\n",
436 dest_partid, XPC_NET_CHANNEL, (unsigned long)msg->buf_pa,
437 msg->size, msg->leadin_ignore, msg->tailout_ignore);
439 atomic_inc(&queued_msg->use_count);
441 ret = xpc_send_notify(dest_partid, XPC_NET_CHANNEL, XPC_NOWAIT, msg,
442 XPNET_MSG_SIZE, xpnet_send_completed, queued_msg);
443 if (unlikely(ret != xpSuccess))
444 atomic_dec(&queued_msg->use_count);
448 * Network layer has formatted a packet (skb) and is ready to place it
449 * "on the wire". Prepare and send an xpnet_message to all partitions
450 * which have connected with us and are targets of this packet.
452 * MAC-NOTE: For the XPNET driver, the MAC address contains the
453 * destination partid. If the destination partid octets are 0xffff,
454 * this packet is to be broadcast to all connected partitions.
457 xpnet_dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev)
459 struct xpnet_pending_msg *queued_msg;
460 u64 start_addr, end_addr;
462 struct xpnet_dev_private *priv = (struct xpnet_dev_private *)dev->priv;
463 u16 embedded_bytes = 0;
465 dev_dbg(xpnet, ">skb->head=0x%p skb->data=0x%p skb->tail=0x%p "
466 "skb->end=0x%p skb->len=%d\n", (void *)skb->head,
467 (void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb),
470 if (skb->data[0] == 0x33) {
472 return 0; /* nothing needed to be done */
476 * The xpnet_pending_msg tracks how many outstanding
477 * xpc_send_notifies are relying on this skb. When none
478 * remain, release the skb.
480 queued_msg = kmalloc(sizeof(struct xpnet_pending_msg), GFP_ATOMIC);
481 if (queued_msg == NULL) {
482 dev_warn(xpnet, "failed to kmalloc %ld bytes; dropping "
483 "packet\n", sizeof(struct xpnet_pending_msg));
485 priv->stats.tx_errors++;
489 /* get the beginning of the first cacheline and end of last */
490 start_addr = ((u64)skb->data & ~(L1_CACHE_BYTES - 1));
491 end_addr = L1_CACHE_ALIGN((u64)skb_tail_pointer(skb));
493 /* calculate how many bytes to embed in the XPC message */
494 if (unlikely(skb->len <= XPNET_MSG_DATA_MAX)) {
495 /* skb->data does fit so embed */
496 embedded_bytes = skb->len;
500 * Since the send occurs asynchronously, we set the count to one
501 * and begin sending. Any sends that happen to complete before
502 * we are done sending will not free the skb. We will be left
503 * with that task during exit. This also handles the case of
504 * a packet destined for a partition which is no longer up.
506 atomic_set(&queued_msg->use_count, 1);
507 queued_msg->skb = skb;
509 if (skb->data[0] == 0xff) {
510 /* we are being asked to broadcast to all partitions */
511 for_each_bit(dest_partid, xpnet_broadcast_partitions,
512 xp_max_npartitions) {
514 xpnet_send(skb, queued_msg, start_addr, end_addr,
515 embedded_bytes, dest_partid);
518 dest_partid = (short)skb->data[XPNET_PARTID_OCTET + 1];
519 dest_partid |= (short)skb->data[XPNET_PARTID_OCTET + 0] << 8;
521 if (dest_partid >= 0 &&
522 dest_partid < xp_max_npartitions &&
523 test_bit(dest_partid, xpnet_broadcast_partitions) != 0) {
525 xpnet_send(skb, queued_msg, start_addr, end_addr,
526 embedded_bytes, dest_partid);
530 if (atomic_dec_return(&queued_msg->use_count) == 0) {
535 priv->stats.tx_packets++;
536 priv->stats.tx_bytes += skb->len;
542 * Deal with transmit timeouts coming from the network layer.
545 xpnet_dev_tx_timeout(struct net_device *dev)
547 struct xpnet_dev_private *priv;
549 priv = (struct xpnet_dev_private *)dev->priv;
551 priv->stats.tx_errors++;
560 if (!is_shub() && !is_uv())
563 dev_info(xpnet, "registering network device %s\n", XPNET_DEVICE_NAME);
565 xpnet_broadcast_partitions = kzalloc(BITS_TO_LONGS(xp_max_npartitions) *
566 sizeof(long), GFP_KERNEL);
567 if (xpnet_broadcast_partitions == NULL)
571 * use ether_setup() to init the majority of our device
572 * structure and then override the necessary pieces.
574 xpnet_device = alloc_netdev(sizeof(struct xpnet_dev_private),
575 XPNET_DEVICE_NAME, ether_setup);
576 if (xpnet_device == NULL) {
577 kfree(xpnet_broadcast_partitions);
581 netif_carrier_off(xpnet_device);
583 xpnet_device->mtu = XPNET_DEF_MTU;
584 xpnet_device->change_mtu = xpnet_dev_change_mtu;
585 xpnet_device->open = xpnet_dev_open;
586 xpnet_device->get_stats = xpnet_dev_get_stats;
587 xpnet_device->stop = xpnet_dev_stop;
588 xpnet_device->hard_start_xmit = xpnet_dev_hard_start_xmit;
589 xpnet_device->tx_timeout = xpnet_dev_tx_timeout;
590 xpnet_device->set_config = xpnet_dev_set_config;
593 * Multicast assumes the LSB of the first octet is set for multicast
594 * MAC addresses. We chose the first octet of the MAC to be unlikely
595 * to collide with any vendor's officially issued MAC.
597 xpnet_device->dev_addr[0] = 0x02; /* locally administered, no OUI */
599 xpnet_device->dev_addr[XPNET_PARTID_OCTET + 1] = xp_partition_id;
600 xpnet_device->dev_addr[XPNET_PARTID_OCTET + 0] = (xp_partition_id >> 8);
603 * ether_setup() sets this to a multicast device. We are
604 * really not supporting multicast at this time.
606 xpnet_device->flags &= ~IFF_MULTICAST;
609 * No need to checksum as it is a DMA transfer. The BTE will
610 * report an error if the data is not retrievable and the
611 * packet will be dropped.
613 xpnet_device->features = NETIF_F_NO_CSUM;
615 result = register_netdev(xpnet_device);
617 free_netdev(xpnet_device);
618 kfree(xpnet_broadcast_partitions);
624 module_init(xpnet_init);
629 dev_info(xpnet, "unregistering network device %s\n",
630 xpnet_device[0].name);
632 unregister_netdev(xpnet_device);
633 free_netdev(xpnet_device);
634 kfree(xpnet_broadcast_partitions);
637 module_exit(xpnet_exit);
639 MODULE_AUTHOR("Silicon Graphics, Inc.");
640 MODULE_DESCRIPTION("Cross Partition Network adapter (XPNET)");
641 MODULE_LICENSE("GPL");