2 * The Guest network driver.
4 * This is very simple a virtual network driver, and our last Guest driver.
5 * The only trick is that it can talk directly to multiple other recipients
6 * (ie. other Guests on the same network). It can also be used with only the
10 /* Copyright 2006 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
27 #include <linux/netdevice.h>
28 #include <linux/etherdevice.h>
29 #include <linux/module.h>
30 #include <linux/mm_types.h>
32 #include <linux/lguest_bus.h>
34 #define SHARED_SIZE PAGE_SIZE
38 /*D:530 The "struct lguestnet_info" contains all the information we need to
39 * know about the network device. */
42 /* The mapped device page(s) (an array of "struct lguest_net"). */
43 struct lguest_net *peer;
44 /* The physical address of the device page(s) */
45 unsigned long peer_phys;
46 /* The size of the device page(s). */
47 unsigned long mapsize;
49 /* The lguest_device I come from */
50 struct lguest_device *lgdev;
52 /* My peerid (ie. my slot in the array). */
55 /* Receive queue: the network packets waiting to be filled. */
56 struct sk_buff *skb[NUM_SKBS];
57 struct lguest_dma dma[NUM_SKBS];
61 /* How many bytes left in this page. */
62 static unsigned int rest_of_page(void *data)
64 return PAGE_SIZE - ((unsigned long)data % PAGE_SIZE);
67 /*D:570 Each peer (ie. Guest or Host) on the network binds their receive
68 * buffers to a different key: we simply use the physical address of the
69 * device's memory page plus the peer number. The Host insists that all keys
70 * be a multiple of 4, so we multiply the peer number by 4. */
71 static unsigned long peer_key(struct lguestnet_info *info, unsigned peernum)
73 return info->peer_phys + 4 * peernum;
76 /* This is the routine which sets up a "struct lguest_dma" to point to a
77 * network packet, similar to req_to_dma() in lguest_blk.c. The structure of a
78 * "struct sk_buff" has grown complex over the years: it consists of a "head"
79 * linear section pointed to by "skb->data", and possibly an array of
80 * "fragments" in the case of a non-linear packet.
82 * Our receive buffers don't use fragments at all but outgoing skbs might, so
84 static void skb_to_dma(const struct sk_buff *skb, unsigned int headlen,
85 struct lguest_dma *dma)
89 /* First, we put the linear region into the "struct lguest_dma". Each
90 * entry can't go over a page boundary, so even though all our packets
91 * are 1514 bytes or less, we might need to use two entries here: */
92 for (i = seg = 0; i < headlen; seg++, i += rest_of_page(skb->data+i)) {
93 dma->addr[seg] = virt_to_phys(skb->data + i);
94 dma->len[seg] = min((unsigned)(headlen - i),
95 rest_of_page(skb->data + i));
98 /* Now we handle the fragments: at least they're guaranteed not to go
99 * over a page. skb_shinfo(skb) returns a pointer to the structure
100 * which tells us about the number of fragments and the fragment
102 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, seg++) {
103 const skb_frag_t *f = &skb_shinfo(skb)->frags[i];
104 /* Should not happen with MTU less than 64k - 2 * PAGE_SIZE. */
105 if (seg == LGUEST_MAX_DMA_SECTIONS) {
106 /* We will end up sending a truncated packet should
107 * this ever happen. Plus, a cool log message! */
108 printk("Woah dude! Megapacket!\n");
111 dma->addr[seg] = page_to_phys(f->page) + f->page_offset;
112 dma->len[seg] = f->size;
115 /* If after all that we didn't use the entire "struct lguest_dma"
116 * array, we terminate it with a 0 length. */
117 if (seg < LGUEST_MAX_DMA_SECTIONS)
122 * Packet transmission.
124 * Our packet transmission is a little unusual. A real network card would just
125 * send out the packet and leave the receivers to decide if they're interested.
126 * Instead, we look through the network device memory page and see if any of
127 * the ethernet addresses match the packet destination, and if so we send it to
130 * This is made a little more complicated in two cases. The first case is
131 * broadcast packets: for that we send the packet to all Guests on the network,
132 * one at a time. The second case is "promiscuous" mode, where a Guest wants
133 * to see all the packets on the network. We need a way for the Guest to tell
134 * us it wants to see all packets, so it sets the "multicast" bit on its
135 * published MAC address, which is never valid in a real ethernet address.
137 #define PROMISC_BIT 0x01
139 /* This is the callback which is summoned whenever the network device's
140 * multicast or promiscuous state changes. If the card is in promiscuous mode,
141 * we advertise that in our ethernet address in the device's memory. We do the
142 * same if Linux wants any or all multicast traffic. */
143 static void lguestnet_set_multicast(struct net_device *dev)
145 struct lguestnet_info *info = netdev_priv(dev);
147 if ((dev->flags & (IFF_PROMISC|IFF_ALLMULTI)) || dev->mc_count)
148 info->peer[info->me].mac[0] |= PROMISC_BIT;
150 info->peer[info->me].mac[0] &= ~PROMISC_BIT;
153 /* A simple test function to see if a peer wants to see all packets.*/
154 static int promisc(struct lguestnet_info *info, unsigned int peer)
156 return info->peer[peer].mac[0] & PROMISC_BIT;
159 /* Another simple function to see if a peer's advertised ethernet address
160 * matches a packet's destination ethernet address. */
161 static int mac_eq(const unsigned char mac[ETH_ALEN],
162 struct lguestnet_info *info, unsigned int peer)
164 /* Ignore multicast bit, which peer turns on to mean promisc. */
165 if ((info->peer[peer].mac[0] & (~PROMISC_BIT)) != mac[0])
167 return memcmp(mac+1, info->peer[peer].mac+1, ETH_ALEN-1) == 0;
170 /* This is the function which actually sends a packet once we've decided a
172 static void transfer_packet(struct net_device *dev,
174 unsigned int peernum)
176 struct lguestnet_info *info = netdev_priv(dev);
177 struct lguest_dma dma;
179 /* We use our handy "struct lguest_dma" packing function to prepare
180 * the skb for sending. */
181 skb_to_dma(skb, skb_headlen(skb), &dma);
182 pr_debug("xfer length %04x (%u)\n", htons(skb->len), skb->len);
184 /* This is the actual send call which copies the packet. */
185 lguest_send_dma(peer_key(info, peernum), &dma);
187 /* Check that the entire packet was transmitted. If not, it could mean
188 * that the other Guest registered a short receive buffer, but this
189 * driver should never do that. More likely, the peer is dead. */
190 if (dma.used_len != skb->len) {
191 dev->stats.tx_carrier_errors++;
192 pr_debug("Bad xfer to peer %i: %i of %i (dma %p/%i)\n",
193 peernum, dma.used_len, skb->len,
194 (void *)dma.addr[0], dma.len[0]);
196 /* On success we update the stats. */
197 dev->stats.tx_bytes += skb->len;
198 dev->stats.tx_packets++;
202 /* Another helper function to tell is if a slot in the device memory is unused.
203 * Since we always set the Local Assignment bit in the ethernet address, the
204 * first byte can never be 0. */
205 static int unused_peer(const struct lguest_net peer[], unsigned int num)
207 return peer[num].mac[0] == 0;
210 /* Finally, here is the routine which handles an outgoing packet. It's called
211 * "start_xmit" for traditional reasons. */
212 static int lguestnet_start_xmit(struct sk_buff *skb, struct net_device *dev)
216 struct lguestnet_info *info = netdev_priv(dev);
217 /* Extract the destination ethernet address from the packet. */
218 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
220 pr_debug("%s: xmit %02x:%02x:%02x:%02x:%02x:%02x\n",
221 dev->name, dest[0],dest[1],dest[2],dest[3],dest[4],dest[5]);
223 /* If it's a multicast packet, we broadcast to everyone. That's not
224 * very efficient, but there are very few applications which actually
225 * use multicast, which is a shame really.
227 * As etherdevice.h points out: "By definition the broadcast address is
228 * also a multicast address." So we don't have to test for broadcast
229 * packets separately. */
230 broadcast = is_multicast_ether_addr(dest);
232 /* Look through all the published ethernet addresses to see if we
233 * should send this packet. */
234 for (i = 0; i < info->mapsize/sizeof(struct lguest_net); i++) {
235 /* We don't send to ourselves (we actually can't SEND_DMA to
236 * ourselves anyway), and don't send to unused slots.*/
237 if (i == info->me || unused_peer(info->peer, i))
240 /* If it's broadcast we send it. If they want every packet we
241 * send it. If the destination matches their address we send
242 * it. Otherwise we go to the next peer. */
243 if (!broadcast && !promisc(info, i) && !mac_eq(dest, info, i))
246 pr_debug("lguestnet %s: sending from %i to %i\n",
247 dev->name, info->me, i);
248 /* Our routine which actually does the transfer. */
249 transfer_packet(dev, skb, i);
252 /* An xmit routine is expected to dispose of the packet, so we do. */
255 /* As per kernel convention, 0 means success. This is why I love
256 * networking: even if we never sent to anyone, that's still
264 * First, here's a helper routine which fills one of our array of receive
266 static int fill_slot(struct net_device *dev, unsigned int slot)
268 struct lguestnet_info *info = netdev_priv(dev);
270 /* We can receive ETH_DATA_LEN (1500) byte packets, plus a standard
271 * ethernet header of ETH_HLEN (14) bytes. */
272 info->skb[slot] = netdev_alloc_skb(dev, ETH_HLEN + ETH_DATA_LEN);
273 if (!info->skb[slot]) {
274 printk("%s: could not fill slot %i\n", dev->name, slot);
278 /* skb_to_dma() is a helper which sets up the "struct lguest_dma" to
279 * point to the data in the skb: we also use it for sending out a
281 skb_to_dma(info->skb[slot], ETH_HLEN + ETH_DATA_LEN, &info->dma[slot]);
283 /* This is a Write Memory Barrier: it ensures that the entry in the
284 * receive buffer array is written *before* we set the "used_len" entry
285 * to 0. If the Host were looking at the receive buffer array from a
286 * different CPU, it could potentially see "used_len = 0" and not see
287 * the updated receive buffer information. This would be a horribly
288 * nasty bug, so make sure the compiler and CPU know this has to happen
291 /* Writing 0 to "used_len" tells the Host it can use this receive
293 info->dma[slot].used_len = 0;
297 /* This is the actual receive routine. When we receive an interrupt from the
298 * Host to tell us a packet has been delivered, we arrive here: */
299 static irqreturn_t lguestnet_rcv(int irq, void *dev_id)
301 struct net_device *dev = dev_id;
302 struct lguestnet_info *info = netdev_priv(dev);
303 unsigned int i, done = 0;
305 /* Look through our entire receive array for an entry which has data
307 for (i = 0; i < ARRAY_SIZE(info->dma); i++) {
311 length = info->dma[i].used_len;
315 /* We've found one! Remember the skb (we grabbed the length
316 * above), and immediately refill the slot we've taken it
322 /* This shouldn't happen: micropackets could be sent by a
323 * badly-behaved Guest on the network, but the Host will never
324 * stuff more data in the buffer than the buffer length. */
325 if (length < ETH_HLEN || length > ETH_HLEN + ETH_DATA_LEN) {
326 pr_debug(KERN_WARNING "%s: unbelievable skb len: %i\n",
332 /* skb_put(), what a great function! I've ranted about this
333 * function before (http://lkml.org/lkml/1999/9/26/24). You
334 * call it after you've added data to the end of an skb (in
335 * this case, it was the Host which wrote the data). */
336 skb_put(skb, length);
338 /* The ethernet header contains a protocol field: we use the
339 * standard helper to extract it, and place the result in
340 * skb->protocol. The helper also sets up skb->pkt_type and
341 * eats up the ethernet header from the front of the packet. */
342 skb->protocol = eth_type_trans(skb, dev);
344 /* If this device doesn't need checksums for sending, we also
345 * don't need to check the packets when they come in. */
346 if (dev->features & NETIF_F_NO_CSUM)
347 skb->ip_summed = CHECKSUM_UNNECESSARY;
349 /* As a last resort for debugging the driver or the lguest I/O
350 * subsystem, you can uncomment the "#define DEBUG" at the top
351 * of this file, which turns all the pr_debug() into printk()
352 * and floods the logs. */
353 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
354 ntohs(skb->protocol), skb->len, skb->pkt_type);
356 /* Update the packet and byte counts (visible from ifconfig,
357 * and good for debugging). */
358 dev->stats.rx_bytes += skb->len;
359 dev->stats.rx_packets++;
361 /* Hand our fresh network packet into the stack's "network
362 * interface receive" routine. That will free the packet
363 * itself when it's finished. */
367 /* If we found any packets, we assume the interrupt was for us. */
368 return done ? IRQ_HANDLED : IRQ_NONE;
371 /*D:550 This is where we start: when the device is brought up by dhcpd or
372 * ifconfig. At this point we advertise our MAC address to the rest of the
373 * network, and register receive buffers ready for incoming packets. */
374 static int lguestnet_open(struct net_device *dev)
377 struct lguestnet_info *info = netdev_priv(dev);
379 /* Copy our MAC address into the device page, so others on the network
381 memcpy(info->peer[info->me].mac, dev->dev_addr, ETH_ALEN);
383 /* We might already be in promisc mode (dev->flags & IFF_PROMISC). Our
384 * set_multicast callback handles this already, so we call it now. */
385 lguestnet_set_multicast(dev);
387 /* Allocate packets and put them into our "struct lguest_dma" array.
388 * If we fail to allocate all the packets we could still limp along,
389 * but it's a sign of real stress so we should probably give up now. */
390 for (i = 0; i < ARRAY_SIZE(info->dma); i++) {
391 if (fill_slot(dev, i) != 0)
395 /* Finally we tell the Host where our array of "struct lguest_dma"
396 * receive buffers is, binding it to the key corresponding to the
397 * device's physical memory plus our peerid. */
398 if (lguest_bind_dma(peer_key(info,info->me), info->dma,
399 NUM_SKBS, lgdev_irq(info->lgdev)) != 0)
405 dev_kfree_skb(info->skb[i]);
410 /* The close routine is called when the device is no longer in use: we clean up
412 static int lguestnet_close(struct net_device *dev)
415 struct lguestnet_info *info = netdev_priv(dev);
417 /* Clear all trace of our existence out of the device memory by setting
418 * the slot which held our MAC address to 0 (unused). */
419 memset(&info->peer[info->me], 0, sizeof(info->peer[info->me]));
421 /* Unregister our array of receive buffers */
422 lguest_unbind_dma(peer_key(info, info->me), info->dma);
423 for (i = 0; i < ARRAY_SIZE(info->dma); i++)
424 dev_kfree_skb(info->skb[i]);
428 /*D:510 The network device probe function is basically a standard ethernet
429 * device setup. It reads the "struct lguest_device_desc" and sets the "struct
430 * net_device". Oh, the line-by-line excitement! Let's skip over it. :*/
431 static int lguestnet_probe(struct lguest_device *lgdev)
433 int err, irqf = IRQF_SHARED;
434 struct net_device *dev;
435 struct lguestnet_info *info;
436 struct lguest_device_desc *desc = &lguest_devices[lgdev->index];
438 pr_debug("lguest_net: probing for device %i\n", lgdev->index);
440 dev = alloc_etherdev(sizeof(struct lguestnet_info));
444 SET_MODULE_OWNER(dev);
446 /* Ethernet defaults with some changes */
448 dev->set_mac_address = NULL;
450 dev->dev_addr[0] = 0x02; /* set local assignment bit (IEEE802) */
451 dev->dev_addr[1] = 0x00;
452 memcpy(&dev->dev_addr[2], &lguest_data.guestid, 2);
453 dev->dev_addr[4] = 0x00;
454 dev->dev_addr[5] = 0x00;
456 dev->open = lguestnet_open;
457 dev->stop = lguestnet_close;
458 dev->hard_start_xmit = lguestnet_start_xmit;
460 /* We don't actually support multicast yet, but turning on/off
461 * promisc also calls dev->set_multicast_list. */
462 dev->set_multicast_list = lguestnet_set_multicast;
463 SET_NETDEV_DEV(dev, &lgdev->dev);
465 /* The network code complains if you have "scatter-gather" capability
466 * if you don't also handle checksums (it seem that would be
467 * "illogical"). So we use a lie of omission and don't tell it that we
468 * can handle scattered packets unless we also don't want checksums,
469 * even though to us they're completely independent. */
470 if (desc->features & LGUEST_NET_F_NOCSUM)
471 dev->features = NETIF_F_SG|NETIF_F_NO_CSUM;
473 info = netdev_priv(dev);
474 info->mapsize = PAGE_SIZE * desc->num_pages;
475 info->peer_phys = ((unsigned long)desc->pfn << PAGE_SHIFT);
477 info->peer = lguest_map(info->peer_phys, desc->num_pages);
483 /* This stores our peerid (upper bits reserved for future). */
484 info->me = (desc->features & (info->mapsize-1));
486 err = register_netdev(dev);
488 pr_debug("lguestnet: registering device failed\n");
492 if (lguest_devices[lgdev->index].features & LGUEST_DEVICE_F_RANDOMNESS)
493 irqf |= IRQF_SAMPLE_RANDOM;
494 if (request_irq(lgdev_irq(lgdev), lguestnet_rcv, irqf, "lguestnet",
496 pr_debug("lguestnet: cannot get irq %i\n", lgdev_irq(lgdev));
500 pr_debug("lguestnet: registered device %s\n", dev->name);
501 /* Finally, we put the "struct net_device" in the generic "struct
502 * lguest_device"s private pointer. Again, it's not necessary, but
503 * makes sure the cool kernel kids don't tease us. */
504 lgdev->private = dev;
508 unregister_netdev(dev);
510 lguest_unmap(info->peer);
516 static struct lguest_driver lguestnet_drv = {
518 .owner = THIS_MODULE,
519 .device_type = LGUEST_DEVICE_T_NET,
520 .probe = lguestnet_probe,
523 static __init int lguestnet_init(void)
525 return register_lguest_driver(&lguestnet_drv);
527 module_init(lguestnet_init);
529 MODULE_DESCRIPTION("Lguest network driver");
530 MODULE_LICENSE("GPL");
533 * This is the last of the Drivers, and with this we have covered the many and
534 * wonderous and fine (and boring) details of the Guest.
536 * "make Launcher" beckons, where we answer questions like "Where do Guests
537 * come from?", and "What do you do when someone asks for optimization?"