1/*
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
4 * for more details.
5 *
6 * Copyright (C) 1999-2009 Silicon Graphics, Inc. All rights reserved.
7 */
8
9/*
10 * Cross Partition Network Interface (XPNET) support
11 *
12 *	XPNET provides a virtual network layered on top of the Cross
13 *	Partition communication layer.
14 *
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.
20 *
21 */
22
23#include <linux/slab.h>
24#include <linux/module.h>
25#include <linux/netdevice.h>
26#include <linux/etherdevice.h>
27#include "xp.h"
28
29/*
30 * The message payload transferred by XPC.
31 *
32 * buf_pa is the physical address where the DMA should pull from.
33 *
34 * NOTE: for performance reasons, buf_pa should _ALWAYS_ begin on a
35 * cacheline boundary.  To accomplish this, we record the number of
36 * bytes from the beginning of the first cacheline to the first useful
37 * byte of the skb (leadin_ignore) and the number of bytes from the
38 * last useful byte of the skb to the end of the last cacheline
39 * (tailout_ignore).
40 *
41 * size is the number of bytes to transfer which includes the skb->len
42 * (useful bytes of the senders skb) plus the leadin and tailout
43 */
44struct xpnet_message {
45	u16 version;		/* Version for this message */
46	u16 embedded_bytes;	/* #of bytes embedded in XPC message */
47	u32 magic;		/* Special number indicating this is xpnet */
48	unsigned long buf_pa;	/* phys address of buffer to retrieve */
49	u32 size;		/* #of bytes in buffer */
50	u8 leadin_ignore;	/* #of bytes to ignore at the beginning */
51	u8 tailout_ignore;	/* #of bytes to ignore at the end */
52	unsigned char data;	/* body of small packets */
53};
54
55/*
56 * Determine the size of our message, the cacheline aligned size,
57 * and then the number of message will request from XPC.
58 *
59 * XPC expects each message to exist in an individual cacheline.
60 */
61#define XPNET_MSG_SIZE		XPC_MSG_PAYLOAD_MAX_SIZE
62#define XPNET_MSG_DATA_MAX	\
63		(XPNET_MSG_SIZE - offsetof(struct xpnet_message, data))
64#define XPNET_MSG_NENTRIES	(PAGE_SIZE / XPC_MSG_MAX_SIZE)
65
66#define XPNET_MAX_KTHREADS	(XPNET_MSG_NENTRIES + 1)
67#define XPNET_MAX_IDLE_KTHREADS	(XPNET_MSG_NENTRIES + 1)
68
69/*
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.
72 */
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)
76
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" */
80
81#define XPNET_VALID_MSG(_m)						     \
82   ((XPNET_VERSION_MAJOR(_m->version) == XPNET_VERSION_MAJOR(XPNET_VERSION)) \
83    && (msg->magic == XPNET_MAGIC))
84
85#define XPNET_DEVICE_NAME		"xp0"
86
87/*
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
92 * then be released.
93 */
94struct xpnet_pending_msg {
95	struct sk_buff *skb;
96	atomic_t use_count;
97};
98
99struct net_device *xpnet_device;
100
101/*
102 * When we are notified of other partitions activating, we add them to
103 * our bitmask of partitions to which we broadcast.
104 */
105static unsigned long *xpnet_broadcast_partitions;
106/* protect above */
107static DEFINE_SPINLOCK(xpnet_broadcast_lock);
108
109/*
110 * Since the Block Transfer Engine (BTE) is being used for the transfer
111 * and it relies upon cache-line size transfers, we need to reserve at
112 * least one cache-line for head and tail alignment.  The BTE is
113 * limited to 8MB transfers.
114 *
115 * Testing has shown that changing MTU to greater than 64KB has no effect
116 * on TCP as the two sides negotiate a Max Segment Size that is limited
117 * to 64K.  Other protocols May use packets greater than this, but for
118 * now, the default is 64KB.
119 */
120#define XPNET_MAX_MTU (0x800000UL - L1_CACHE_BYTES)
121/* 32KB has been determined to be the ideal */
122#define XPNET_DEF_MTU (0x8000UL)
123
124/*
125 * The partid is encapsulated in the MAC address beginning in the following
126 * octet and it consists of two octets.
127 */
128#define XPNET_PARTID_OCTET	2
129
130/* Define the XPNET debug device structures to be used with dev_dbg() et al */
131
132struct device_driver xpnet_dbg_name = {
133	.name = "xpnet"
134};
135
136struct device xpnet_dbg_subname = {
137	.init_name = "",	/* set to "" */
138	.driver = &xpnet_dbg_name
139};
140
141struct device *xpnet = &xpnet_dbg_subname;
142
143/*
144 * Packet was recevied by XPC and forwarded to us.
145 */
146static void
147xpnet_receive(short partid, int channel, struct xpnet_message *msg)
148{
149	struct sk_buff *skb;
150	void *dst;
151	enum xp_retval ret;
152
153	if (!XPNET_VALID_MSG(msg)) {
154		/*
155		 * Packet with a different XPC version.  Ignore.
156		 */
157		xpc_received(partid, channel, (void *)msg);
158
159		xpnet_device->stats.rx_errors++;
160
161		return;
162	}
163	dev_dbg(xpnet, "received 0x%lx, %d, %d, %d\n", msg->buf_pa, msg->size,
164		msg->leadin_ignore, msg->tailout_ignore);
165
166	/* reserve an extra cache line */
167	skb = dev_alloc_skb(msg->size + L1_CACHE_BYTES);
168	if (!skb) {
169		dev_err(xpnet, "failed on dev_alloc_skb(%d)\n",
170			msg->size + L1_CACHE_BYTES);
171
172		xpc_received(partid, channel, (void *)msg);
173
174		xpnet_device->stats.rx_errors++;
175
176		return;
177	}
178
179	/*
180	 * The allocated skb has some reserved space.
181	 * In order to use xp_remote_memcpy(), we need to get the
182	 * skb->data pointer moved forward.
183	 */
184	skb_reserve(skb, (L1_CACHE_BYTES - ((u64)skb->data &
185					    (L1_CACHE_BYTES - 1)) +
186			  msg->leadin_ignore));
187
188	/*
189	 * Update the tail pointer to indicate data actually
190	 * transferred.
191	 */
192	skb_put(skb, (msg->size - msg->leadin_ignore - msg->tailout_ignore));
193
194	/*
195	 * Move the data over from the other side.
196	 */
197	if ((XPNET_VERSION_MINOR(msg->version) == 1) &&
198	    (msg->embedded_bytes != 0)) {
199		dev_dbg(xpnet, "copying embedded message. memcpy(0x%p, 0x%p, "
200			"%lu)\n", skb->data, &msg->data,
201			(size_t)msg->embedded_bytes);
202
203		skb_copy_to_linear_data(skb, &msg->data,
204					(size_t)msg->embedded_bytes);
205	} else {
206		dst = (void *)((u64)skb->data & ~(L1_CACHE_BYTES - 1));
207		dev_dbg(xpnet, "transferring buffer to the skb->data area;\n\t"
208			"xp_remote_memcpy(0x%p, 0x%p, %hu)\n", dst,
209					  (void *)msg->buf_pa, msg->size);
210
211		ret = xp_remote_memcpy(xp_pa(dst), msg->buf_pa, msg->size);
212		if (ret != xpSuccess) {
213			/*
214			 * !!! Need better way of cleaning skb.  Currently skb
215			 * !!! appears in_use and we can't just call
216			 * !!! dev_kfree_skb.
217			 */
218			dev_err(xpnet, "xp_remote_memcpy(0x%p, 0x%p, 0x%hx) "
219				"returned error=0x%x\n", dst,
220				(void *)msg->buf_pa, msg->size, ret);
221
222			xpc_received(partid, channel, (void *)msg);
223
224			xpnet_device->stats.rx_errors++;
225
226			return;
227		}
228	}
229
230	dev_dbg(xpnet, "<skb->head=0x%p skb->data=0x%p skb->tail=0x%p "
231		"skb->end=0x%p skb->len=%d\n", (void *)skb->head,
232		(void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb),
233		skb->len);
234
235	skb->protocol = eth_type_trans(skb, xpnet_device);
236	skb->ip_summed = CHECKSUM_UNNECESSARY;
237
238	dev_dbg(xpnet, "passing skb to network layer\n"
239		"\tskb->head=0x%p skb->data=0x%p skb->tail=0x%p "
240		"skb->end=0x%p skb->len=%d\n",
241		(void *)skb->head, (void *)skb->data, skb_tail_pointer(skb),
242		skb_end_pointer(skb), skb->len);
243
244	xpnet_device->stats.rx_packets++;
245	xpnet_device->stats.rx_bytes += skb->len + ETH_HLEN;
246
247	netif_rx_ni(skb);
248	xpc_received(partid, channel, (void *)msg);
249}
250
251/*
252 * This is the handler which XPC calls during any sort of change in
253 * state or message reception on a connection.
254 */
255static void
256xpnet_connection_activity(enum xp_retval reason, short partid, int channel,
257			  void *data, void *key)
258{
259	DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
260	DBUG_ON(channel != XPC_NET_CHANNEL);
261
262	switch (reason) {
263	case xpMsgReceived:	/* message received */
264		DBUG_ON(data == NULL);
265
266		xpnet_receive(partid, channel, (struct xpnet_message *)data);
267		break;
268
269	case xpConnected:	/* connection completed to a partition */
270		spin_lock_bh(&xpnet_broadcast_lock);
271		__set_bit(partid, xpnet_broadcast_partitions);
272		spin_unlock_bh(&xpnet_broadcast_lock);
273
274		netif_carrier_on(xpnet_device);
275
276		dev_dbg(xpnet, "%s connected to partition %d\n",
277			xpnet_device->name, partid);
278		break;
279
280	default:
281		spin_lock_bh(&xpnet_broadcast_lock);
282		__clear_bit(partid, xpnet_broadcast_partitions);
283		spin_unlock_bh(&xpnet_broadcast_lock);
284
285		if (bitmap_empty((unsigned long *)xpnet_broadcast_partitions,
286				 xp_max_npartitions)) {
287			netif_carrier_off(xpnet_device);
288		}
289
290		dev_dbg(xpnet, "%s disconnected from partition %d\n",
291			xpnet_device->name, partid);
292		break;
293	}
294}
295
296static int
297xpnet_dev_open(struct net_device *dev)
298{
299	enum xp_retval ret;
300
301	dev_dbg(xpnet, "calling xpc_connect(%d, 0x%p, NULL, %ld, %ld, %ld, "
302		"%ld)\n", XPC_NET_CHANNEL, xpnet_connection_activity,
303		(unsigned long)XPNET_MSG_SIZE,
304		(unsigned long)XPNET_MSG_NENTRIES,
305		(unsigned long)XPNET_MAX_KTHREADS,
306		(unsigned long)XPNET_MAX_IDLE_KTHREADS);
307
308	ret = xpc_connect(XPC_NET_CHANNEL, xpnet_connection_activity, NULL,
309			  XPNET_MSG_SIZE, XPNET_MSG_NENTRIES,
310			  XPNET_MAX_KTHREADS, XPNET_MAX_IDLE_KTHREADS);
311	if (ret != xpSuccess) {
312		dev_err(xpnet, "ifconfig up of %s failed on XPC connect, "
313			"ret=%d\n", dev->name, ret);
314
315		return -ENOMEM;
316	}
317
318	dev_dbg(xpnet, "ifconfig up of %s; XPC connected\n", dev->name);
319
320	return 0;
321}
322
323static int
324xpnet_dev_stop(struct net_device *dev)
325{
326	xpc_disconnect(XPC_NET_CHANNEL);
327
328	dev_dbg(xpnet, "ifconfig down of %s; XPC disconnected\n", dev->name);
329
330	return 0;
331}
332
333static int
334xpnet_dev_change_mtu(struct net_device *dev, int new_mtu)
335{
336	/* 68 comes from min TCP+IP+MAC header */
337	if ((new_mtu < 68) || (new_mtu > XPNET_MAX_MTU)) {
338		dev_err(xpnet, "ifconfig %s mtu %d failed; value must be "
339			"between 68 and %ld\n", dev->name, new_mtu,
340			XPNET_MAX_MTU);
341		return -EINVAL;
342	}
343
344	dev->mtu = new_mtu;
345	dev_dbg(xpnet, "ifconfig %s mtu set to %d\n", dev->name, new_mtu);
346	return 0;
347}
348
349/*
350 * Notification that the other end has received the message and
351 * DMA'd the skb information.  At this point, they are done with
352 * our side.  When all recipients are done processing, we
353 * release the skb and then release our pending message structure.
354 */
355static void
356xpnet_send_completed(enum xp_retval reason, short partid, int channel,
357		     void *__qm)
358{
359	struct xpnet_pending_msg *queued_msg = (struct xpnet_pending_msg *)__qm;
360
361	DBUG_ON(queued_msg == NULL);
362
363	dev_dbg(xpnet, "message to %d notified with reason %d\n",
364		partid, reason);
365
366	if (atomic_dec_return(&queued_msg->use_count) == 0) {
367		dev_dbg(xpnet, "all acks for skb->head=-x%p\n",
368			(void *)queued_msg->skb->head);
369
370		dev_kfree_skb_any(queued_msg->skb);
371		kfree(queued_msg);
372	}
373}
374
375static void
376xpnet_send(struct sk_buff *skb, struct xpnet_pending_msg *queued_msg,
377	   u64 start_addr, u64 end_addr, u16 embedded_bytes, int dest_partid)
378{
379	u8 msg_buffer[XPNET_MSG_SIZE];
380	struct xpnet_message *msg = (struct xpnet_message *)&msg_buffer;
381	u16 msg_size = sizeof(struct xpnet_message);
382	enum xp_retval ret;
383
384	msg->embedded_bytes = embedded_bytes;
385	if (unlikely(embedded_bytes != 0)) {
386		msg->version = XPNET_VERSION_EMBED;
387		dev_dbg(xpnet, "calling memcpy(0x%p, 0x%p, 0x%lx)\n",
388			&msg->data, skb->data, (size_t)embedded_bytes);
389		skb_copy_from_linear_data(skb, &msg->data,
390					  (size_t)embedded_bytes);
391		msg_size += embedded_bytes - 1;
392	} else {
393		msg->version = XPNET_VERSION;
394	}
395	msg->magic = XPNET_MAGIC;
396	msg->size = end_addr - start_addr;
397	msg->leadin_ignore = (u64)skb->data - start_addr;
398	msg->tailout_ignore = end_addr - (u64)skb_tail_pointer(skb);
399	msg->buf_pa = xp_pa((void *)start_addr);
400
401	dev_dbg(xpnet, "sending XPC message to %d:%d\n"
402		"msg->buf_pa=0x%lx, msg->size=%u, "
403		"msg->leadin_ignore=%u, msg->tailout_ignore=%u\n",
404		dest_partid, XPC_NET_CHANNEL, msg->buf_pa, msg->size,
405		msg->leadin_ignore, msg->tailout_ignore);
406
407	atomic_inc(&queued_msg->use_count);
408
409	ret = xpc_send_notify(dest_partid, XPC_NET_CHANNEL, XPC_NOWAIT, msg,
410			      msg_size, xpnet_send_completed, queued_msg);
411	if (unlikely(ret != xpSuccess))
412		atomic_dec(&queued_msg->use_count);
413}
414
415/*
416 * Network layer has formatted a packet (skb) and is ready to place it
417 * "on the wire".  Prepare and send an xpnet_message to all partitions
418 * which have connected with us and are targets of this packet.
419 *
420 * MAC-NOTE:  For the XPNET driver, the MAC address contains the
421 * destination partid.  If the destination partid octets are 0xffff,
422 * this packet is to be broadcast to all connected partitions.
423 */
424static int
425xpnet_dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev)
426{
427	struct xpnet_pending_msg *queued_msg;
428	u64 start_addr, end_addr;
429	short dest_partid;
430	u16 embedded_bytes = 0;
431
432	dev_dbg(xpnet, ">skb->head=0x%p skb->data=0x%p skb->tail=0x%p "
433		"skb->end=0x%p skb->len=%d\n", (void *)skb->head,
434		(void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb),
435		skb->len);
436
437	if (skb->data[0] == 0x33) {
438		dev_kfree_skb(skb);
439		return NETDEV_TX_OK;	/* nothing needed to be done */
440	}
441
442	/*
443	 * The xpnet_pending_msg tracks how many outstanding
444	 * xpc_send_notifies are relying on this skb.  When none
445	 * remain, release the skb.
446	 */
447	queued_msg = kmalloc(sizeof(struct xpnet_pending_msg), GFP_ATOMIC);
448	if (queued_msg == NULL) {
449		dev_warn(xpnet, "failed to kmalloc %ld bytes; dropping "
450			 "packet\n", sizeof(struct xpnet_pending_msg));
451
452		dev->stats.tx_errors++;
453		dev_kfree_skb(skb);
454		return NETDEV_TX_OK;
455	}
456
457	/* get the beginning of the first cacheline and end of last */
458	start_addr = ((u64)skb->data & ~(L1_CACHE_BYTES - 1));
459	end_addr = L1_CACHE_ALIGN((u64)skb_tail_pointer(skb));
460
461	/* calculate how many bytes to embed in the XPC message */
462	if (unlikely(skb->len <= XPNET_MSG_DATA_MAX)) {
463		/* skb->data does fit so embed */
464		embedded_bytes = skb->len;
465	}
466
467	/*
468	 * Since the send occurs asynchronously, we set the count to one
469	 * and begin sending.  Any sends that happen to complete before
470	 * we are done sending will not free the skb.  We will be left
471	 * with that task during exit.  This also handles the case of
472	 * a packet destined for a partition which is no longer up.
473	 */
474	atomic_set(&queued_msg->use_count, 1);
475	queued_msg->skb = skb;
476
477	if (skb->data[0] == 0xff) {
478		/* we are being asked to broadcast to all partitions */
479		for_each_set_bit(dest_partid, xpnet_broadcast_partitions,
480			     xp_max_npartitions) {
481
482			xpnet_send(skb, queued_msg, start_addr, end_addr,
483				   embedded_bytes, dest_partid);
484		}
485	} else {
486		dest_partid = (short)skb->data[XPNET_PARTID_OCTET + 1];
487		dest_partid |= (short)skb->data[XPNET_PARTID_OCTET + 0] << 8;
488
489		if (dest_partid >= 0 &&
490		    dest_partid < xp_max_npartitions &&
491		    test_bit(dest_partid, xpnet_broadcast_partitions) != 0) {
492
493			xpnet_send(skb, queued_msg, start_addr, end_addr,
494				   embedded_bytes, dest_partid);
495		}
496	}
497
498	dev->stats.tx_packets++;
499	dev->stats.tx_bytes += skb->len;
500
501	if (atomic_dec_return(&queued_msg->use_count) == 0) {
502		dev_kfree_skb(skb);
503		kfree(queued_msg);
504	}
505
506	return NETDEV_TX_OK;
507}
508
509/*
510 * Deal with transmit timeouts coming from the network layer.
511 */
512static void
513xpnet_dev_tx_timeout(struct net_device *dev)
514{
515	dev->stats.tx_errors++;
516}
517
518static const struct net_device_ops xpnet_netdev_ops = {
519	.ndo_open		= xpnet_dev_open,
520	.ndo_stop		= xpnet_dev_stop,
521	.ndo_start_xmit		= xpnet_dev_hard_start_xmit,
522	.ndo_change_mtu		= xpnet_dev_change_mtu,
523	.ndo_tx_timeout		= xpnet_dev_tx_timeout,
524	.ndo_set_mac_address 	= eth_mac_addr,
525	.ndo_validate_addr	= eth_validate_addr,
526};
527
528static int __init
529xpnet_init(void)
530{
531	int result;
532
533	if (!is_shub() && !is_uv())
534		return -ENODEV;
535
536	dev_info(xpnet, "registering network device %s\n", XPNET_DEVICE_NAME);
537
538	xpnet_broadcast_partitions = kzalloc(BITS_TO_LONGS(xp_max_npartitions) *
539					     sizeof(long), GFP_KERNEL);
540	if (xpnet_broadcast_partitions == NULL)
541		return -ENOMEM;
542
543	/*
544	 * use ether_setup() to init the majority of our device
545	 * structure and then override the necessary pieces.
546	 */
547	xpnet_device = alloc_netdev(0, XPNET_DEVICE_NAME, NET_NAME_UNKNOWN,
548				    ether_setup);
549	if (xpnet_device == NULL) {
550		kfree(xpnet_broadcast_partitions);
551		return -ENOMEM;
552	}
553
554	netif_carrier_off(xpnet_device);
555
556	xpnet_device->netdev_ops = &xpnet_netdev_ops;
557	xpnet_device->mtu = XPNET_DEF_MTU;
558
559	/*
560	 * Multicast assumes the LSB of the first octet is set for multicast
561	 * MAC addresses.  We chose the first octet of the MAC to be unlikely
562	 * to collide with any vendor's officially issued MAC.
563	 */
564	xpnet_device->dev_addr[0] = 0x02;     /* locally administered, no OUI */
565
566	xpnet_device->dev_addr[XPNET_PARTID_OCTET + 1] = xp_partition_id;
567	xpnet_device->dev_addr[XPNET_PARTID_OCTET + 0] = (xp_partition_id >> 8);
568
569	/*
570	 * ether_setup() sets this to a multicast device.  We are
571	 * really not supporting multicast at this time.
572	 */
573	xpnet_device->flags &= ~IFF_MULTICAST;
574
575	/*
576	 * No need to checksum as it is a DMA transfer.  The BTE will
577	 * report an error if the data is not retrievable and the
578	 * packet will be dropped.
579	 */
580	xpnet_device->features = NETIF_F_HW_CSUM;
581
582	result = register_netdev(xpnet_device);
583	if (result != 0) {
584		free_netdev(xpnet_device);
585		kfree(xpnet_broadcast_partitions);
586	}
587
588	return result;
589}
590
591module_init(xpnet_init);
592
593static void __exit
594xpnet_exit(void)
595{
596	dev_info(xpnet, "unregistering network device %s\n",
597		 xpnet_device[0].name);
598
599	unregister_netdev(xpnet_device);
600	free_netdev(xpnet_device);
601	kfree(xpnet_broadcast_partitions);
602}
603
604module_exit(xpnet_exit);
605
606MODULE_AUTHOR("Silicon Graphics, Inc.");
607MODULE_DESCRIPTION("Cross Partition Network adapter (XPNET)");
608MODULE_LICENSE("GPL");
609