xref: /openbmc/linux/drivers/net/xen-netfront.c (revision 4800cd83)
1 /*
2  * Virtual network driver for conversing with remote driver backends.
3  *
4  * Copyright (c) 2002-2005, K A Fraser
5  * Copyright (c) 2005, XenSource Ltd
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License version 2
9  * as published by the Free Software Foundation; or, when distributed
10  * separately from the Linux kernel or incorporated into other
11  * software packages, subject to the following license:
12  *
13  * Permission is hereby granted, free of charge, to any person obtaining a copy
14  * of this source file (the "Software"), to deal in the Software without
15  * restriction, including without limitation the rights to use, copy, modify,
16  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
17  * and to permit persons to whom the Software is furnished to do so, subject to
18  * the following conditions:
19  *
20  * The above copyright notice and this permission notice shall be included in
21  * all copies or substantial portions of the Software.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
26  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
28  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
29  * IN THE SOFTWARE.
30  */
31 
32 #include <linux/module.h>
33 #include <linux/kernel.h>
34 #include <linux/netdevice.h>
35 #include <linux/etherdevice.h>
36 #include <linux/skbuff.h>
37 #include <linux/ethtool.h>
38 #include <linux/if_ether.h>
39 #include <linux/tcp.h>
40 #include <linux/udp.h>
41 #include <linux/moduleparam.h>
42 #include <linux/mm.h>
43 #include <linux/slab.h>
44 #include <net/ip.h>
45 
46 #include <xen/xen.h>
47 #include <xen/xenbus.h>
48 #include <xen/events.h>
49 #include <xen/page.h>
50 #include <xen/grant_table.h>
51 
52 #include <xen/interface/io/netif.h>
53 #include <xen/interface/memory.h>
54 #include <xen/interface/grant_table.h>
55 
56 static const struct ethtool_ops xennet_ethtool_ops;
57 
58 struct netfront_cb {
59 	struct page *page;
60 	unsigned offset;
61 };
62 
63 #define NETFRONT_SKB_CB(skb)	((struct netfront_cb *)((skb)->cb))
64 
65 #define RX_COPY_THRESHOLD 256
66 
67 #define GRANT_INVALID_REF	0
68 
69 #define NET_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, PAGE_SIZE)
70 #define NET_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, PAGE_SIZE)
71 #define TX_MAX_TARGET min_t(int, NET_RX_RING_SIZE, 256)
72 
73 struct netfront_info {
74 	struct list_head list;
75 	struct net_device *netdev;
76 
77 	struct napi_struct napi;
78 
79 	unsigned int evtchn;
80 	struct xenbus_device *xbdev;
81 
82 	spinlock_t   tx_lock;
83 	struct xen_netif_tx_front_ring tx;
84 	int tx_ring_ref;
85 
86 	/*
87 	 * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries
88 	 * are linked from tx_skb_freelist through skb_entry.link.
89 	 *
90 	 *  NB. Freelist index entries are always going to be less than
91 	 *  PAGE_OFFSET, whereas pointers to skbs will always be equal or
92 	 *  greater than PAGE_OFFSET: we use this property to distinguish
93 	 *  them.
94 	 */
95 	union skb_entry {
96 		struct sk_buff *skb;
97 		unsigned long link;
98 	} tx_skbs[NET_TX_RING_SIZE];
99 	grant_ref_t gref_tx_head;
100 	grant_ref_t grant_tx_ref[NET_TX_RING_SIZE];
101 	unsigned tx_skb_freelist;
102 
103 	spinlock_t   rx_lock ____cacheline_aligned_in_smp;
104 	struct xen_netif_rx_front_ring rx;
105 	int rx_ring_ref;
106 
107 	/* Receive-ring batched refills. */
108 #define RX_MIN_TARGET 8
109 #define RX_DFL_MIN_TARGET 64
110 #define RX_MAX_TARGET min_t(int, NET_RX_RING_SIZE, 256)
111 	unsigned rx_min_target, rx_max_target, rx_target;
112 	struct sk_buff_head rx_batch;
113 
114 	struct timer_list rx_refill_timer;
115 
116 	struct sk_buff *rx_skbs[NET_RX_RING_SIZE];
117 	grant_ref_t gref_rx_head;
118 	grant_ref_t grant_rx_ref[NET_RX_RING_SIZE];
119 
120 	unsigned long rx_pfn_array[NET_RX_RING_SIZE];
121 	struct multicall_entry rx_mcl[NET_RX_RING_SIZE+1];
122 	struct mmu_update rx_mmu[NET_RX_RING_SIZE];
123 
124 	/* Statistics */
125 	int rx_gso_checksum_fixup;
126 };
127 
128 struct netfront_rx_info {
129 	struct xen_netif_rx_response rx;
130 	struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
131 };
132 
133 static void skb_entry_set_link(union skb_entry *list, unsigned short id)
134 {
135 	list->link = id;
136 }
137 
138 static int skb_entry_is_link(const union skb_entry *list)
139 {
140 	BUILD_BUG_ON(sizeof(list->skb) != sizeof(list->link));
141 	return (unsigned long)list->skb < PAGE_OFFSET;
142 }
143 
144 /*
145  * Access macros for acquiring freeing slots in tx_skbs[].
146  */
147 
148 static void add_id_to_freelist(unsigned *head, union skb_entry *list,
149 			       unsigned short id)
150 {
151 	skb_entry_set_link(&list[id], *head);
152 	*head = id;
153 }
154 
155 static unsigned short get_id_from_freelist(unsigned *head,
156 					   union skb_entry *list)
157 {
158 	unsigned int id = *head;
159 	*head = list[id].link;
160 	return id;
161 }
162 
163 static int xennet_rxidx(RING_IDX idx)
164 {
165 	return idx & (NET_RX_RING_SIZE - 1);
166 }
167 
168 static struct sk_buff *xennet_get_rx_skb(struct netfront_info *np,
169 					 RING_IDX ri)
170 {
171 	int i = xennet_rxidx(ri);
172 	struct sk_buff *skb = np->rx_skbs[i];
173 	np->rx_skbs[i] = NULL;
174 	return skb;
175 }
176 
177 static grant_ref_t xennet_get_rx_ref(struct netfront_info *np,
178 					    RING_IDX ri)
179 {
180 	int i = xennet_rxidx(ri);
181 	grant_ref_t ref = np->grant_rx_ref[i];
182 	np->grant_rx_ref[i] = GRANT_INVALID_REF;
183 	return ref;
184 }
185 
186 #ifdef CONFIG_SYSFS
187 static int xennet_sysfs_addif(struct net_device *netdev);
188 static void xennet_sysfs_delif(struct net_device *netdev);
189 #else /* !CONFIG_SYSFS */
190 #define xennet_sysfs_addif(dev) (0)
191 #define xennet_sysfs_delif(dev) do { } while (0)
192 #endif
193 
194 static int xennet_can_sg(struct net_device *dev)
195 {
196 	return dev->features & NETIF_F_SG;
197 }
198 
199 
200 static void rx_refill_timeout(unsigned long data)
201 {
202 	struct net_device *dev = (struct net_device *)data;
203 	struct netfront_info *np = netdev_priv(dev);
204 	napi_schedule(&np->napi);
205 }
206 
207 static int netfront_tx_slot_available(struct netfront_info *np)
208 {
209 	return (np->tx.req_prod_pvt - np->tx.rsp_cons) <
210 		(TX_MAX_TARGET - MAX_SKB_FRAGS - 2);
211 }
212 
213 static void xennet_maybe_wake_tx(struct net_device *dev)
214 {
215 	struct netfront_info *np = netdev_priv(dev);
216 
217 	if (unlikely(netif_queue_stopped(dev)) &&
218 	    netfront_tx_slot_available(np) &&
219 	    likely(netif_running(dev)))
220 		netif_wake_queue(dev);
221 }
222 
223 static void xennet_alloc_rx_buffers(struct net_device *dev)
224 {
225 	unsigned short id;
226 	struct netfront_info *np = netdev_priv(dev);
227 	struct sk_buff *skb;
228 	struct page *page;
229 	int i, batch_target, notify;
230 	RING_IDX req_prod = np->rx.req_prod_pvt;
231 	grant_ref_t ref;
232 	unsigned long pfn;
233 	void *vaddr;
234 	struct xen_netif_rx_request *req;
235 
236 	if (unlikely(!netif_carrier_ok(dev)))
237 		return;
238 
239 	/*
240 	 * Allocate skbuffs greedily, even though we batch updates to the
241 	 * receive ring. This creates a less bursty demand on the memory
242 	 * allocator, so should reduce the chance of failed allocation requests
243 	 * both for ourself and for other kernel subsystems.
244 	 */
245 	batch_target = np->rx_target - (req_prod - np->rx.rsp_cons);
246 	for (i = skb_queue_len(&np->rx_batch); i < batch_target; i++) {
247 		skb = __netdev_alloc_skb(dev, RX_COPY_THRESHOLD + NET_IP_ALIGN,
248 					 GFP_ATOMIC | __GFP_NOWARN);
249 		if (unlikely(!skb))
250 			goto no_skb;
251 
252 		/* Align ip header to a 16 bytes boundary */
253 		skb_reserve(skb, NET_IP_ALIGN);
254 
255 		page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
256 		if (!page) {
257 			kfree_skb(skb);
258 no_skb:
259 			/* Any skbuffs queued for refill? Force them out. */
260 			if (i != 0)
261 				goto refill;
262 			/* Could not allocate any skbuffs. Try again later. */
263 			mod_timer(&np->rx_refill_timer,
264 				  jiffies + (HZ/10));
265 			break;
266 		}
267 
268 		skb_shinfo(skb)->frags[0].page = page;
269 		skb_shinfo(skb)->nr_frags = 1;
270 		__skb_queue_tail(&np->rx_batch, skb);
271 	}
272 
273 	/* Is the batch large enough to be worthwhile? */
274 	if (i < (np->rx_target/2)) {
275 		if (req_prod > np->rx.sring->req_prod)
276 			goto push;
277 		return;
278 	}
279 
280 	/* Adjust our fill target if we risked running out of buffers. */
281 	if (((req_prod - np->rx.sring->rsp_prod) < (np->rx_target / 4)) &&
282 	    ((np->rx_target *= 2) > np->rx_max_target))
283 		np->rx_target = np->rx_max_target;
284 
285  refill:
286 	for (i = 0; ; i++) {
287 		skb = __skb_dequeue(&np->rx_batch);
288 		if (skb == NULL)
289 			break;
290 
291 		skb->dev = dev;
292 
293 		id = xennet_rxidx(req_prod + i);
294 
295 		BUG_ON(np->rx_skbs[id]);
296 		np->rx_skbs[id] = skb;
297 
298 		ref = gnttab_claim_grant_reference(&np->gref_rx_head);
299 		BUG_ON((signed short)ref < 0);
300 		np->grant_rx_ref[id] = ref;
301 
302 		pfn = page_to_pfn(skb_shinfo(skb)->frags[0].page);
303 		vaddr = page_address(skb_shinfo(skb)->frags[0].page);
304 
305 		req = RING_GET_REQUEST(&np->rx, req_prod + i);
306 		gnttab_grant_foreign_access_ref(ref,
307 						np->xbdev->otherend_id,
308 						pfn_to_mfn(pfn),
309 						0);
310 
311 		req->id = id;
312 		req->gref = ref;
313 	}
314 
315 	wmb();		/* barrier so backend seens requests */
316 
317 	/* Above is a suitable barrier to ensure backend will see requests. */
318 	np->rx.req_prod_pvt = req_prod + i;
319  push:
320 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->rx, notify);
321 	if (notify)
322 		notify_remote_via_irq(np->netdev->irq);
323 }
324 
325 static int xennet_open(struct net_device *dev)
326 {
327 	struct netfront_info *np = netdev_priv(dev);
328 
329 	napi_enable(&np->napi);
330 
331 	spin_lock_bh(&np->rx_lock);
332 	if (netif_carrier_ok(dev)) {
333 		xennet_alloc_rx_buffers(dev);
334 		np->rx.sring->rsp_event = np->rx.rsp_cons + 1;
335 		if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx))
336 			napi_schedule(&np->napi);
337 	}
338 	spin_unlock_bh(&np->rx_lock);
339 
340 	netif_start_queue(dev);
341 
342 	return 0;
343 }
344 
345 static void xennet_tx_buf_gc(struct net_device *dev)
346 {
347 	RING_IDX cons, prod;
348 	unsigned short id;
349 	struct netfront_info *np = netdev_priv(dev);
350 	struct sk_buff *skb;
351 
352 	BUG_ON(!netif_carrier_ok(dev));
353 
354 	do {
355 		prod = np->tx.sring->rsp_prod;
356 		rmb(); /* Ensure we see responses up to 'rp'. */
357 
358 		for (cons = np->tx.rsp_cons; cons != prod; cons++) {
359 			struct xen_netif_tx_response *txrsp;
360 
361 			txrsp = RING_GET_RESPONSE(&np->tx, cons);
362 			if (txrsp->status == NETIF_RSP_NULL)
363 				continue;
364 
365 			id  = txrsp->id;
366 			skb = np->tx_skbs[id].skb;
367 			if (unlikely(gnttab_query_foreign_access(
368 				np->grant_tx_ref[id]) != 0)) {
369 				printk(KERN_ALERT "xennet_tx_buf_gc: warning "
370 				       "-- grant still in use by backend "
371 				       "domain.\n");
372 				BUG();
373 			}
374 			gnttab_end_foreign_access_ref(
375 				np->grant_tx_ref[id], GNTMAP_readonly);
376 			gnttab_release_grant_reference(
377 				&np->gref_tx_head, np->grant_tx_ref[id]);
378 			np->grant_tx_ref[id] = GRANT_INVALID_REF;
379 			add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, id);
380 			dev_kfree_skb_irq(skb);
381 		}
382 
383 		np->tx.rsp_cons = prod;
384 
385 		/*
386 		 * Set a new event, then check for race with update of tx_cons.
387 		 * Note that it is essential to schedule a callback, no matter
388 		 * how few buffers are pending. Even if there is space in the
389 		 * transmit ring, higher layers may be blocked because too much
390 		 * data is outstanding: in such cases notification from Xen is
391 		 * likely to be the only kick that we'll get.
392 		 */
393 		np->tx.sring->rsp_event =
394 			prod + ((np->tx.sring->req_prod - prod) >> 1) + 1;
395 		mb();		/* update shared area */
396 	} while ((cons == prod) && (prod != np->tx.sring->rsp_prod));
397 
398 	xennet_maybe_wake_tx(dev);
399 }
400 
401 static void xennet_make_frags(struct sk_buff *skb, struct net_device *dev,
402 			      struct xen_netif_tx_request *tx)
403 {
404 	struct netfront_info *np = netdev_priv(dev);
405 	char *data = skb->data;
406 	unsigned long mfn;
407 	RING_IDX prod = np->tx.req_prod_pvt;
408 	int frags = skb_shinfo(skb)->nr_frags;
409 	unsigned int offset = offset_in_page(data);
410 	unsigned int len = skb_headlen(skb);
411 	unsigned int id;
412 	grant_ref_t ref;
413 	int i;
414 
415 	/* While the header overlaps a page boundary (including being
416 	   larger than a page), split it it into page-sized chunks. */
417 	while (len > PAGE_SIZE - offset) {
418 		tx->size = PAGE_SIZE - offset;
419 		tx->flags |= NETTXF_more_data;
420 		len -= tx->size;
421 		data += tx->size;
422 		offset = 0;
423 
424 		id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
425 		np->tx_skbs[id].skb = skb_get(skb);
426 		tx = RING_GET_REQUEST(&np->tx, prod++);
427 		tx->id = id;
428 		ref = gnttab_claim_grant_reference(&np->gref_tx_head);
429 		BUG_ON((signed short)ref < 0);
430 
431 		mfn = virt_to_mfn(data);
432 		gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id,
433 						mfn, GNTMAP_readonly);
434 
435 		tx->gref = np->grant_tx_ref[id] = ref;
436 		tx->offset = offset;
437 		tx->size = len;
438 		tx->flags = 0;
439 	}
440 
441 	/* Grant backend access to each skb fragment page. */
442 	for (i = 0; i < frags; i++) {
443 		skb_frag_t *frag = skb_shinfo(skb)->frags + i;
444 
445 		tx->flags |= NETTXF_more_data;
446 
447 		id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
448 		np->tx_skbs[id].skb = skb_get(skb);
449 		tx = RING_GET_REQUEST(&np->tx, prod++);
450 		tx->id = id;
451 		ref = gnttab_claim_grant_reference(&np->gref_tx_head);
452 		BUG_ON((signed short)ref < 0);
453 
454 		mfn = pfn_to_mfn(page_to_pfn(frag->page));
455 		gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id,
456 						mfn, GNTMAP_readonly);
457 
458 		tx->gref = np->grant_tx_ref[id] = ref;
459 		tx->offset = frag->page_offset;
460 		tx->size = frag->size;
461 		tx->flags = 0;
462 	}
463 
464 	np->tx.req_prod_pvt = prod;
465 }
466 
467 static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
468 {
469 	unsigned short id;
470 	struct netfront_info *np = netdev_priv(dev);
471 	struct xen_netif_tx_request *tx;
472 	struct xen_netif_extra_info *extra;
473 	char *data = skb->data;
474 	RING_IDX i;
475 	grant_ref_t ref;
476 	unsigned long mfn;
477 	int notify;
478 	int frags = skb_shinfo(skb)->nr_frags;
479 	unsigned int offset = offset_in_page(data);
480 	unsigned int len = skb_headlen(skb);
481 
482 	frags += DIV_ROUND_UP(offset + len, PAGE_SIZE);
483 	if (unlikely(frags > MAX_SKB_FRAGS + 1)) {
484 		printk(KERN_ALERT "xennet: skb rides the rocket: %d frags\n",
485 		       frags);
486 		dump_stack();
487 		goto drop;
488 	}
489 
490 	spin_lock_irq(&np->tx_lock);
491 
492 	if (unlikely(!netif_carrier_ok(dev) ||
493 		     (frags > 1 && !xennet_can_sg(dev)) ||
494 		     netif_needs_gso(skb, netif_skb_features(skb)))) {
495 		spin_unlock_irq(&np->tx_lock);
496 		goto drop;
497 	}
498 
499 	i = np->tx.req_prod_pvt;
500 
501 	id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
502 	np->tx_skbs[id].skb = skb;
503 
504 	tx = RING_GET_REQUEST(&np->tx, i);
505 
506 	tx->id   = id;
507 	ref = gnttab_claim_grant_reference(&np->gref_tx_head);
508 	BUG_ON((signed short)ref < 0);
509 	mfn = virt_to_mfn(data);
510 	gnttab_grant_foreign_access_ref(
511 		ref, np->xbdev->otherend_id, mfn, GNTMAP_readonly);
512 	tx->gref = np->grant_tx_ref[id] = ref;
513 	tx->offset = offset;
514 	tx->size = len;
515 	extra = NULL;
516 
517 	tx->flags = 0;
518 	if (skb->ip_summed == CHECKSUM_PARTIAL)
519 		/* local packet? */
520 		tx->flags |= NETTXF_csum_blank | NETTXF_data_validated;
521 	else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
522 		/* remote but checksummed. */
523 		tx->flags |= NETTXF_data_validated;
524 
525 	if (skb_shinfo(skb)->gso_size) {
526 		struct xen_netif_extra_info *gso;
527 
528 		gso = (struct xen_netif_extra_info *)
529 			RING_GET_REQUEST(&np->tx, ++i);
530 
531 		if (extra)
532 			extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
533 		else
534 			tx->flags |= NETTXF_extra_info;
535 
536 		gso->u.gso.size = skb_shinfo(skb)->gso_size;
537 		gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
538 		gso->u.gso.pad = 0;
539 		gso->u.gso.features = 0;
540 
541 		gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
542 		gso->flags = 0;
543 		extra = gso;
544 	}
545 
546 	np->tx.req_prod_pvt = i + 1;
547 
548 	xennet_make_frags(skb, dev, tx);
549 	tx->size = skb->len;
550 
551 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->tx, notify);
552 	if (notify)
553 		notify_remote_via_irq(np->netdev->irq);
554 
555 	dev->stats.tx_bytes += skb->len;
556 	dev->stats.tx_packets++;
557 
558 	/* Note: It is not safe to access skb after xennet_tx_buf_gc()! */
559 	xennet_tx_buf_gc(dev);
560 
561 	if (!netfront_tx_slot_available(np))
562 		netif_stop_queue(dev);
563 
564 	spin_unlock_irq(&np->tx_lock);
565 
566 	return NETDEV_TX_OK;
567 
568  drop:
569 	dev->stats.tx_dropped++;
570 	dev_kfree_skb(skb);
571 	return NETDEV_TX_OK;
572 }
573 
574 static int xennet_close(struct net_device *dev)
575 {
576 	struct netfront_info *np = netdev_priv(dev);
577 	netif_stop_queue(np->netdev);
578 	napi_disable(&np->napi);
579 	return 0;
580 }
581 
582 static void xennet_move_rx_slot(struct netfront_info *np, struct sk_buff *skb,
583 				grant_ref_t ref)
584 {
585 	int new = xennet_rxidx(np->rx.req_prod_pvt);
586 
587 	BUG_ON(np->rx_skbs[new]);
588 	np->rx_skbs[new] = skb;
589 	np->grant_rx_ref[new] = ref;
590 	RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->id = new;
591 	RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->gref = ref;
592 	np->rx.req_prod_pvt++;
593 }
594 
595 static int xennet_get_extras(struct netfront_info *np,
596 			     struct xen_netif_extra_info *extras,
597 			     RING_IDX rp)
598 
599 {
600 	struct xen_netif_extra_info *extra;
601 	struct device *dev = &np->netdev->dev;
602 	RING_IDX cons = np->rx.rsp_cons;
603 	int err = 0;
604 
605 	do {
606 		struct sk_buff *skb;
607 		grant_ref_t ref;
608 
609 		if (unlikely(cons + 1 == rp)) {
610 			if (net_ratelimit())
611 				dev_warn(dev, "Missing extra info\n");
612 			err = -EBADR;
613 			break;
614 		}
615 
616 		extra = (struct xen_netif_extra_info *)
617 			RING_GET_RESPONSE(&np->rx, ++cons);
618 
619 		if (unlikely(!extra->type ||
620 			     extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
621 			if (net_ratelimit())
622 				dev_warn(dev, "Invalid extra type: %d\n",
623 					extra->type);
624 			err = -EINVAL;
625 		} else {
626 			memcpy(&extras[extra->type - 1], extra,
627 			       sizeof(*extra));
628 		}
629 
630 		skb = xennet_get_rx_skb(np, cons);
631 		ref = xennet_get_rx_ref(np, cons);
632 		xennet_move_rx_slot(np, skb, ref);
633 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
634 
635 	np->rx.rsp_cons = cons;
636 	return err;
637 }
638 
639 static int xennet_get_responses(struct netfront_info *np,
640 				struct netfront_rx_info *rinfo, RING_IDX rp,
641 				struct sk_buff_head *list)
642 {
643 	struct xen_netif_rx_response *rx = &rinfo->rx;
644 	struct xen_netif_extra_info *extras = rinfo->extras;
645 	struct device *dev = &np->netdev->dev;
646 	RING_IDX cons = np->rx.rsp_cons;
647 	struct sk_buff *skb = xennet_get_rx_skb(np, cons);
648 	grant_ref_t ref = xennet_get_rx_ref(np, cons);
649 	int max = MAX_SKB_FRAGS + (rx->status <= RX_COPY_THRESHOLD);
650 	int frags = 1;
651 	int err = 0;
652 	unsigned long ret;
653 
654 	if (rx->flags & NETRXF_extra_info) {
655 		err = xennet_get_extras(np, extras, rp);
656 		cons = np->rx.rsp_cons;
657 	}
658 
659 	for (;;) {
660 		if (unlikely(rx->status < 0 ||
661 			     rx->offset + rx->status > PAGE_SIZE)) {
662 			if (net_ratelimit())
663 				dev_warn(dev, "rx->offset: %x, size: %u\n",
664 					 rx->offset, rx->status);
665 			xennet_move_rx_slot(np, skb, ref);
666 			err = -EINVAL;
667 			goto next;
668 		}
669 
670 		/*
671 		 * This definitely indicates a bug, either in this driver or in
672 		 * the backend driver. In future this should flag the bad
673 		 * situation to the system controller to reboot the backed.
674 		 */
675 		if (ref == GRANT_INVALID_REF) {
676 			if (net_ratelimit())
677 				dev_warn(dev, "Bad rx response id %d.\n",
678 					 rx->id);
679 			err = -EINVAL;
680 			goto next;
681 		}
682 
683 		ret = gnttab_end_foreign_access_ref(ref, 0);
684 		BUG_ON(!ret);
685 
686 		gnttab_release_grant_reference(&np->gref_rx_head, ref);
687 
688 		__skb_queue_tail(list, skb);
689 
690 next:
691 		if (!(rx->flags & NETRXF_more_data))
692 			break;
693 
694 		if (cons + frags == rp) {
695 			if (net_ratelimit())
696 				dev_warn(dev, "Need more frags\n");
697 			err = -ENOENT;
698 			break;
699 		}
700 
701 		rx = RING_GET_RESPONSE(&np->rx, cons + frags);
702 		skb = xennet_get_rx_skb(np, cons + frags);
703 		ref = xennet_get_rx_ref(np, cons + frags);
704 		frags++;
705 	}
706 
707 	if (unlikely(frags > max)) {
708 		if (net_ratelimit())
709 			dev_warn(dev, "Too many frags\n");
710 		err = -E2BIG;
711 	}
712 
713 	if (unlikely(err))
714 		np->rx.rsp_cons = cons + frags;
715 
716 	return err;
717 }
718 
719 static int xennet_set_skb_gso(struct sk_buff *skb,
720 			      struct xen_netif_extra_info *gso)
721 {
722 	if (!gso->u.gso.size) {
723 		if (net_ratelimit())
724 			printk(KERN_WARNING "GSO size must not be zero.\n");
725 		return -EINVAL;
726 	}
727 
728 	/* Currently only TCPv4 S.O. is supported. */
729 	if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
730 		if (net_ratelimit())
731 			printk(KERN_WARNING "Bad GSO type %d.\n", gso->u.gso.type);
732 		return -EINVAL;
733 	}
734 
735 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
736 	skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
737 
738 	/* Header must be checked, and gso_segs computed. */
739 	skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
740 	skb_shinfo(skb)->gso_segs = 0;
741 
742 	return 0;
743 }
744 
745 static RING_IDX xennet_fill_frags(struct netfront_info *np,
746 				  struct sk_buff *skb,
747 				  struct sk_buff_head *list)
748 {
749 	struct skb_shared_info *shinfo = skb_shinfo(skb);
750 	int nr_frags = shinfo->nr_frags;
751 	RING_IDX cons = np->rx.rsp_cons;
752 	skb_frag_t *frag = shinfo->frags + nr_frags;
753 	struct sk_buff *nskb;
754 
755 	while ((nskb = __skb_dequeue(list))) {
756 		struct xen_netif_rx_response *rx =
757 			RING_GET_RESPONSE(&np->rx, ++cons);
758 
759 		frag->page = skb_shinfo(nskb)->frags[0].page;
760 		frag->page_offset = rx->offset;
761 		frag->size = rx->status;
762 
763 		skb->data_len += rx->status;
764 
765 		skb_shinfo(nskb)->nr_frags = 0;
766 		kfree_skb(nskb);
767 
768 		frag++;
769 		nr_frags++;
770 	}
771 
772 	shinfo->nr_frags = nr_frags;
773 	return cons;
774 }
775 
776 static int checksum_setup(struct net_device *dev, struct sk_buff *skb)
777 {
778 	struct iphdr *iph;
779 	unsigned char *th;
780 	int err = -EPROTO;
781 	int recalculate_partial_csum = 0;
782 
783 	/*
784 	 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
785 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
786 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
787 	 * recalculate the partial checksum.
788 	 */
789 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
790 		struct netfront_info *np = netdev_priv(dev);
791 		np->rx_gso_checksum_fixup++;
792 		skb->ip_summed = CHECKSUM_PARTIAL;
793 		recalculate_partial_csum = 1;
794 	}
795 
796 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
797 	if (skb->ip_summed != CHECKSUM_PARTIAL)
798 		return 0;
799 
800 	if (skb->protocol != htons(ETH_P_IP))
801 		goto out;
802 
803 	iph = (void *)skb->data;
804 	th = skb->data + 4 * iph->ihl;
805 	if (th >= skb_tail_pointer(skb))
806 		goto out;
807 
808 	skb->csum_start = th - skb->head;
809 	switch (iph->protocol) {
810 	case IPPROTO_TCP:
811 		skb->csum_offset = offsetof(struct tcphdr, check);
812 
813 		if (recalculate_partial_csum) {
814 			struct tcphdr *tcph = (struct tcphdr *)th;
815 			tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
816 							 skb->len - iph->ihl*4,
817 							 IPPROTO_TCP, 0);
818 		}
819 		break;
820 	case IPPROTO_UDP:
821 		skb->csum_offset = offsetof(struct udphdr, check);
822 
823 		if (recalculate_partial_csum) {
824 			struct udphdr *udph = (struct udphdr *)th;
825 			udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
826 							 skb->len - iph->ihl*4,
827 							 IPPROTO_UDP, 0);
828 		}
829 		break;
830 	default:
831 		if (net_ratelimit())
832 			printk(KERN_ERR "Attempting to checksum a non-"
833 			       "TCP/UDP packet, dropping a protocol"
834 			       " %d packet", iph->protocol);
835 		goto out;
836 	}
837 
838 	if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb))
839 		goto out;
840 
841 	err = 0;
842 
843 out:
844 	return err;
845 }
846 
847 static int handle_incoming_queue(struct net_device *dev,
848 				 struct sk_buff_head *rxq)
849 {
850 	int packets_dropped = 0;
851 	struct sk_buff *skb;
852 
853 	while ((skb = __skb_dequeue(rxq)) != NULL) {
854 		struct page *page = NETFRONT_SKB_CB(skb)->page;
855 		void *vaddr = page_address(page);
856 		unsigned offset = NETFRONT_SKB_CB(skb)->offset;
857 
858 		memcpy(skb->data, vaddr + offset,
859 		       skb_headlen(skb));
860 
861 		if (page != skb_shinfo(skb)->frags[0].page)
862 			__free_page(page);
863 
864 		/* Ethernet work: Delayed to here as it peeks the header. */
865 		skb->protocol = eth_type_trans(skb, dev);
866 
867 		if (checksum_setup(dev, skb)) {
868 			kfree_skb(skb);
869 			packets_dropped++;
870 			dev->stats.rx_errors++;
871 			continue;
872 		}
873 
874 		dev->stats.rx_packets++;
875 		dev->stats.rx_bytes += skb->len;
876 
877 		/* Pass it up. */
878 		netif_receive_skb(skb);
879 	}
880 
881 	return packets_dropped;
882 }
883 
884 static int xennet_poll(struct napi_struct *napi, int budget)
885 {
886 	struct netfront_info *np = container_of(napi, struct netfront_info, napi);
887 	struct net_device *dev = np->netdev;
888 	struct sk_buff *skb;
889 	struct netfront_rx_info rinfo;
890 	struct xen_netif_rx_response *rx = &rinfo.rx;
891 	struct xen_netif_extra_info *extras = rinfo.extras;
892 	RING_IDX i, rp;
893 	int work_done;
894 	struct sk_buff_head rxq;
895 	struct sk_buff_head errq;
896 	struct sk_buff_head tmpq;
897 	unsigned long flags;
898 	unsigned int len;
899 	int err;
900 
901 	spin_lock(&np->rx_lock);
902 
903 	skb_queue_head_init(&rxq);
904 	skb_queue_head_init(&errq);
905 	skb_queue_head_init(&tmpq);
906 
907 	rp = np->rx.sring->rsp_prod;
908 	rmb(); /* Ensure we see queued responses up to 'rp'. */
909 
910 	i = np->rx.rsp_cons;
911 	work_done = 0;
912 	while ((i != rp) && (work_done < budget)) {
913 		memcpy(rx, RING_GET_RESPONSE(&np->rx, i), sizeof(*rx));
914 		memset(extras, 0, sizeof(rinfo.extras));
915 
916 		err = xennet_get_responses(np, &rinfo, rp, &tmpq);
917 
918 		if (unlikely(err)) {
919 err:
920 			while ((skb = __skb_dequeue(&tmpq)))
921 				__skb_queue_tail(&errq, skb);
922 			dev->stats.rx_errors++;
923 			i = np->rx.rsp_cons;
924 			continue;
925 		}
926 
927 		skb = __skb_dequeue(&tmpq);
928 
929 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
930 			struct xen_netif_extra_info *gso;
931 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
932 
933 			if (unlikely(xennet_set_skb_gso(skb, gso))) {
934 				__skb_queue_head(&tmpq, skb);
935 				np->rx.rsp_cons += skb_queue_len(&tmpq);
936 				goto err;
937 			}
938 		}
939 
940 		NETFRONT_SKB_CB(skb)->page = skb_shinfo(skb)->frags[0].page;
941 		NETFRONT_SKB_CB(skb)->offset = rx->offset;
942 
943 		len = rx->status;
944 		if (len > RX_COPY_THRESHOLD)
945 			len = RX_COPY_THRESHOLD;
946 		skb_put(skb, len);
947 
948 		if (rx->status > len) {
949 			skb_shinfo(skb)->frags[0].page_offset =
950 				rx->offset + len;
951 			skb_shinfo(skb)->frags[0].size = rx->status - len;
952 			skb->data_len = rx->status - len;
953 		} else {
954 			skb_shinfo(skb)->frags[0].page = NULL;
955 			skb_shinfo(skb)->nr_frags = 0;
956 		}
957 
958 		i = xennet_fill_frags(np, skb, &tmpq);
959 
960 		/*
961 		 * Truesize approximates the size of true data plus
962 		 * any supervisor overheads. Adding hypervisor
963 		 * overheads has been shown to significantly reduce
964 		 * achievable bandwidth with the default receive
965 		 * buffer size. It is therefore not wise to account
966 		 * for it here.
967 		 *
968 		 * After alloc_skb(RX_COPY_THRESHOLD), truesize is set
969 		 * to RX_COPY_THRESHOLD + the supervisor
970 		 * overheads. Here, we add the size of the data pulled
971 		 * in xennet_fill_frags().
972 		 *
973 		 * We also adjust for any unused space in the main
974 		 * data area by subtracting (RX_COPY_THRESHOLD -
975 		 * len). This is especially important with drivers
976 		 * which split incoming packets into header and data,
977 		 * using only 66 bytes of the main data area (see the
978 		 * e1000 driver for example.)  On such systems,
979 		 * without this last adjustement, our achievable
980 		 * receive throughout using the standard receive
981 		 * buffer size was cut by 25%(!!!).
982 		 */
983 		skb->truesize += skb->data_len - (RX_COPY_THRESHOLD - len);
984 		skb->len += skb->data_len;
985 
986 		if (rx->flags & NETRXF_csum_blank)
987 			skb->ip_summed = CHECKSUM_PARTIAL;
988 		else if (rx->flags & NETRXF_data_validated)
989 			skb->ip_summed = CHECKSUM_UNNECESSARY;
990 
991 		__skb_queue_tail(&rxq, skb);
992 
993 		np->rx.rsp_cons = ++i;
994 		work_done++;
995 	}
996 
997 	__skb_queue_purge(&errq);
998 
999 	work_done -= handle_incoming_queue(dev, &rxq);
1000 
1001 	/* If we get a callback with very few responses, reduce fill target. */
1002 	/* NB. Note exponential increase, linear decrease. */
1003 	if (((np->rx.req_prod_pvt - np->rx.sring->rsp_prod) >
1004 	     ((3*np->rx_target) / 4)) &&
1005 	    (--np->rx_target < np->rx_min_target))
1006 		np->rx_target = np->rx_min_target;
1007 
1008 	xennet_alloc_rx_buffers(dev);
1009 
1010 	if (work_done < budget) {
1011 		int more_to_do = 0;
1012 
1013 		local_irq_save(flags);
1014 
1015 		RING_FINAL_CHECK_FOR_RESPONSES(&np->rx, more_to_do);
1016 		if (!more_to_do)
1017 			__napi_complete(napi);
1018 
1019 		local_irq_restore(flags);
1020 	}
1021 
1022 	spin_unlock(&np->rx_lock);
1023 
1024 	return work_done;
1025 }
1026 
1027 static int xennet_change_mtu(struct net_device *dev, int mtu)
1028 {
1029 	int max = xennet_can_sg(dev) ? 65535 - ETH_HLEN : ETH_DATA_LEN;
1030 
1031 	if (mtu > max)
1032 		return -EINVAL;
1033 	dev->mtu = mtu;
1034 	return 0;
1035 }
1036 
1037 static void xennet_release_tx_bufs(struct netfront_info *np)
1038 {
1039 	struct sk_buff *skb;
1040 	int i;
1041 
1042 	for (i = 0; i < NET_TX_RING_SIZE; i++) {
1043 		/* Skip over entries which are actually freelist references */
1044 		if (skb_entry_is_link(&np->tx_skbs[i]))
1045 			continue;
1046 
1047 		skb = np->tx_skbs[i].skb;
1048 		gnttab_end_foreign_access_ref(np->grant_tx_ref[i],
1049 					      GNTMAP_readonly);
1050 		gnttab_release_grant_reference(&np->gref_tx_head,
1051 					       np->grant_tx_ref[i]);
1052 		np->grant_tx_ref[i] = GRANT_INVALID_REF;
1053 		add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, i);
1054 		dev_kfree_skb_irq(skb);
1055 	}
1056 }
1057 
1058 static void xennet_release_rx_bufs(struct netfront_info *np)
1059 {
1060 	struct mmu_update      *mmu = np->rx_mmu;
1061 	struct multicall_entry *mcl = np->rx_mcl;
1062 	struct sk_buff_head free_list;
1063 	struct sk_buff *skb;
1064 	unsigned long mfn;
1065 	int xfer = 0, noxfer = 0, unused = 0;
1066 	int id, ref;
1067 
1068 	dev_warn(&np->netdev->dev, "%s: fix me for copying receiver.\n",
1069 			 __func__);
1070 	return;
1071 
1072 	skb_queue_head_init(&free_list);
1073 
1074 	spin_lock_bh(&np->rx_lock);
1075 
1076 	for (id = 0; id < NET_RX_RING_SIZE; id++) {
1077 		ref = np->grant_rx_ref[id];
1078 		if (ref == GRANT_INVALID_REF) {
1079 			unused++;
1080 			continue;
1081 		}
1082 
1083 		skb = np->rx_skbs[id];
1084 		mfn = gnttab_end_foreign_transfer_ref(ref);
1085 		gnttab_release_grant_reference(&np->gref_rx_head, ref);
1086 		np->grant_rx_ref[id] = GRANT_INVALID_REF;
1087 
1088 		if (0 == mfn) {
1089 			skb_shinfo(skb)->nr_frags = 0;
1090 			dev_kfree_skb(skb);
1091 			noxfer++;
1092 			continue;
1093 		}
1094 
1095 		if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1096 			/* Remap the page. */
1097 			struct page *page = skb_shinfo(skb)->frags[0].page;
1098 			unsigned long pfn = page_to_pfn(page);
1099 			void *vaddr = page_address(page);
1100 
1101 			MULTI_update_va_mapping(mcl, (unsigned long)vaddr,
1102 						mfn_pte(mfn, PAGE_KERNEL),
1103 						0);
1104 			mcl++;
1105 			mmu->ptr = ((u64)mfn << PAGE_SHIFT)
1106 				| MMU_MACHPHYS_UPDATE;
1107 			mmu->val = pfn;
1108 			mmu++;
1109 
1110 			set_phys_to_machine(pfn, mfn);
1111 		}
1112 		__skb_queue_tail(&free_list, skb);
1113 		xfer++;
1114 	}
1115 
1116 	dev_info(&np->netdev->dev, "%s: %d xfer, %d noxfer, %d unused\n",
1117 		 __func__, xfer, noxfer, unused);
1118 
1119 	if (xfer) {
1120 		if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1121 			/* Do all the remapping work and M2P updates. */
1122 			MULTI_mmu_update(mcl, np->rx_mmu, mmu - np->rx_mmu,
1123 					 NULL, DOMID_SELF);
1124 			mcl++;
1125 			HYPERVISOR_multicall(np->rx_mcl, mcl - np->rx_mcl);
1126 		}
1127 	}
1128 
1129 	__skb_queue_purge(&free_list);
1130 
1131 	spin_unlock_bh(&np->rx_lock);
1132 }
1133 
1134 static void xennet_uninit(struct net_device *dev)
1135 {
1136 	struct netfront_info *np = netdev_priv(dev);
1137 	xennet_release_tx_bufs(np);
1138 	xennet_release_rx_bufs(np);
1139 	gnttab_free_grant_references(np->gref_tx_head);
1140 	gnttab_free_grant_references(np->gref_rx_head);
1141 }
1142 
1143 static const struct net_device_ops xennet_netdev_ops = {
1144 	.ndo_open            = xennet_open,
1145 	.ndo_uninit          = xennet_uninit,
1146 	.ndo_stop            = xennet_close,
1147 	.ndo_start_xmit      = xennet_start_xmit,
1148 	.ndo_change_mtu	     = xennet_change_mtu,
1149 	.ndo_set_mac_address = eth_mac_addr,
1150 	.ndo_validate_addr   = eth_validate_addr,
1151 };
1152 
1153 static struct net_device * __devinit xennet_create_dev(struct xenbus_device *dev)
1154 {
1155 	int i, err;
1156 	struct net_device *netdev;
1157 	struct netfront_info *np;
1158 
1159 	netdev = alloc_etherdev(sizeof(struct netfront_info));
1160 	if (!netdev) {
1161 		printk(KERN_WARNING "%s> alloc_etherdev failed.\n",
1162 		       __func__);
1163 		return ERR_PTR(-ENOMEM);
1164 	}
1165 
1166 	np                   = netdev_priv(netdev);
1167 	np->xbdev            = dev;
1168 
1169 	spin_lock_init(&np->tx_lock);
1170 	spin_lock_init(&np->rx_lock);
1171 
1172 	skb_queue_head_init(&np->rx_batch);
1173 	np->rx_target     = RX_DFL_MIN_TARGET;
1174 	np->rx_min_target = RX_DFL_MIN_TARGET;
1175 	np->rx_max_target = RX_MAX_TARGET;
1176 
1177 	init_timer(&np->rx_refill_timer);
1178 	np->rx_refill_timer.data = (unsigned long)netdev;
1179 	np->rx_refill_timer.function = rx_refill_timeout;
1180 
1181 	/* Initialise tx_skbs as a free chain containing every entry. */
1182 	np->tx_skb_freelist = 0;
1183 	for (i = 0; i < NET_TX_RING_SIZE; i++) {
1184 		skb_entry_set_link(&np->tx_skbs[i], i+1);
1185 		np->grant_tx_ref[i] = GRANT_INVALID_REF;
1186 	}
1187 
1188 	/* Clear out rx_skbs */
1189 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1190 		np->rx_skbs[i] = NULL;
1191 		np->grant_rx_ref[i] = GRANT_INVALID_REF;
1192 	}
1193 
1194 	/* A grant for every tx ring slot */
1195 	if (gnttab_alloc_grant_references(TX_MAX_TARGET,
1196 					  &np->gref_tx_head) < 0) {
1197 		printk(KERN_ALERT "#### netfront can't alloc tx grant refs\n");
1198 		err = -ENOMEM;
1199 		goto exit;
1200 	}
1201 	/* A grant for every rx ring slot */
1202 	if (gnttab_alloc_grant_references(RX_MAX_TARGET,
1203 					  &np->gref_rx_head) < 0) {
1204 		printk(KERN_ALERT "#### netfront can't alloc rx grant refs\n");
1205 		err = -ENOMEM;
1206 		goto exit_free_tx;
1207 	}
1208 
1209 	netdev->netdev_ops	= &xennet_netdev_ops;
1210 
1211 	netif_napi_add(netdev, &np->napi, xennet_poll, 64);
1212 	netdev->features        = NETIF_F_IP_CSUM;
1213 
1214 	SET_ETHTOOL_OPS(netdev, &xennet_ethtool_ops);
1215 	SET_NETDEV_DEV(netdev, &dev->dev);
1216 
1217 	np->netdev = netdev;
1218 
1219 	netif_carrier_off(netdev);
1220 
1221 	return netdev;
1222 
1223  exit_free_tx:
1224 	gnttab_free_grant_references(np->gref_tx_head);
1225  exit:
1226 	free_netdev(netdev);
1227 	return ERR_PTR(err);
1228 }
1229 
1230 /**
1231  * Entry point to this code when a new device is created.  Allocate the basic
1232  * structures and the ring buffers for communication with the backend, and
1233  * inform the backend of the appropriate details for those.
1234  */
1235 static int __devinit netfront_probe(struct xenbus_device *dev,
1236 				    const struct xenbus_device_id *id)
1237 {
1238 	int err;
1239 	struct net_device *netdev;
1240 	struct netfront_info *info;
1241 
1242 	netdev = xennet_create_dev(dev);
1243 	if (IS_ERR(netdev)) {
1244 		err = PTR_ERR(netdev);
1245 		xenbus_dev_fatal(dev, err, "creating netdev");
1246 		return err;
1247 	}
1248 
1249 	info = netdev_priv(netdev);
1250 	dev_set_drvdata(&dev->dev, info);
1251 
1252 	err = register_netdev(info->netdev);
1253 	if (err) {
1254 		printk(KERN_WARNING "%s: register_netdev err=%d\n",
1255 		       __func__, err);
1256 		goto fail;
1257 	}
1258 
1259 	err = xennet_sysfs_addif(info->netdev);
1260 	if (err) {
1261 		unregister_netdev(info->netdev);
1262 		printk(KERN_WARNING "%s: add sysfs failed err=%d\n",
1263 		       __func__, err);
1264 		goto fail;
1265 	}
1266 
1267 	return 0;
1268 
1269  fail:
1270 	free_netdev(netdev);
1271 	dev_set_drvdata(&dev->dev, NULL);
1272 	return err;
1273 }
1274 
1275 static void xennet_end_access(int ref, void *page)
1276 {
1277 	/* This frees the page as a side-effect */
1278 	if (ref != GRANT_INVALID_REF)
1279 		gnttab_end_foreign_access(ref, 0, (unsigned long)page);
1280 }
1281 
1282 static void xennet_disconnect_backend(struct netfront_info *info)
1283 {
1284 	/* Stop old i/f to prevent errors whilst we rebuild the state. */
1285 	spin_lock_bh(&info->rx_lock);
1286 	spin_lock_irq(&info->tx_lock);
1287 	netif_carrier_off(info->netdev);
1288 	spin_unlock_irq(&info->tx_lock);
1289 	spin_unlock_bh(&info->rx_lock);
1290 
1291 	if (info->netdev->irq)
1292 		unbind_from_irqhandler(info->netdev->irq, info->netdev);
1293 	info->evtchn = info->netdev->irq = 0;
1294 
1295 	/* End access and free the pages */
1296 	xennet_end_access(info->tx_ring_ref, info->tx.sring);
1297 	xennet_end_access(info->rx_ring_ref, info->rx.sring);
1298 
1299 	info->tx_ring_ref = GRANT_INVALID_REF;
1300 	info->rx_ring_ref = GRANT_INVALID_REF;
1301 	info->tx.sring = NULL;
1302 	info->rx.sring = NULL;
1303 }
1304 
1305 /**
1306  * We are reconnecting to the backend, due to a suspend/resume, or a backend
1307  * driver restart.  We tear down our netif structure and recreate it, but
1308  * leave the device-layer structures intact so that this is transparent to the
1309  * rest of the kernel.
1310  */
1311 static int netfront_resume(struct xenbus_device *dev)
1312 {
1313 	struct netfront_info *info = dev_get_drvdata(&dev->dev);
1314 
1315 	dev_dbg(&dev->dev, "%s\n", dev->nodename);
1316 
1317 	xennet_disconnect_backend(info);
1318 	return 0;
1319 }
1320 
1321 static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[])
1322 {
1323 	char *s, *e, *macstr;
1324 	int i;
1325 
1326 	macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL);
1327 	if (IS_ERR(macstr))
1328 		return PTR_ERR(macstr);
1329 
1330 	for (i = 0; i < ETH_ALEN; i++) {
1331 		mac[i] = simple_strtoul(s, &e, 16);
1332 		if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) {
1333 			kfree(macstr);
1334 			return -ENOENT;
1335 		}
1336 		s = e+1;
1337 	}
1338 
1339 	kfree(macstr);
1340 	return 0;
1341 }
1342 
1343 static irqreturn_t xennet_interrupt(int irq, void *dev_id)
1344 {
1345 	struct net_device *dev = dev_id;
1346 	struct netfront_info *np = netdev_priv(dev);
1347 	unsigned long flags;
1348 
1349 	spin_lock_irqsave(&np->tx_lock, flags);
1350 
1351 	if (likely(netif_carrier_ok(dev))) {
1352 		xennet_tx_buf_gc(dev);
1353 		/* Under tx_lock: protects access to rx shared-ring indexes. */
1354 		if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx))
1355 			napi_schedule(&np->napi);
1356 	}
1357 
1358 	spin_unlock_irqrestore(&np->tx_lock, flags);
1359 
1360 	return IRQ_HANDLED;
1361 }
1362 
1363 static int setup_netfront(struct xenbus_device *dev, struct netfront_info *info)
1364 {
1365 	struct xen_netif_tx_sring *txs;
1366 	struct xen_netif_rx_sring *rxs;
1367 	int err;
1368 	struct net_device *netdev = info->netdev;
1369 
1370 	info->tx_ring_ref = GRANT_INVALID_REF;
1371 	info->rx_ring_ref = GRANT_INVALID_REF;
1372 	info->rx.sring = NULL;
1373 	info->tx.sring = NULL;
1374 	netdev->irq = 0;
1375 
1376 	err = xen_net_read_mac(dev, netdev->dev_addr);
1377 	if (err) {
1378 		xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename);
1379 		goto fail;
1380 	}
1381 
1382 	txs = (struct xen_netif_tx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH);
1383 	if (!txs) {
1384 		err = -ENOMEM;
1385 		xenbus_dev_fatal(dev, err, "allocating tx ring page");
1386 		goto fail;
1387 	}
1388 	SHARED_RING_INIT(txs);
1389 	FRONT_RING_INIT(&info->tx, txs, PAGE_SIZE);
1390 
1391 	err = xenbus_grant_ring(dev, virt_to_mfn(txs));
1392 	if (err < 0) {
1393 		free_page((unsigned long)txs);
1394 		goto fail;
1395 	}
1396 
1397 	info->tx_ring_ref = err;
1398 	rxs = (struct xen_netif_rx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH);
1399 	if (!rxs) {
1400 		err = -ENOMEM;
1401 		xenbus_dev_fatal(dev, err, "allocating rx ring page");
1402 		goto fail;
1403 	}
1404 	SHARED_RING_INIT(rxs);
1405 	FRONT_RING_INIT(&info->rx, rxs, PAGE_SIZE);
1406 
1407 	err = xenbus_grant_ring(dev, virt_to_mfn(rxs));
1408 	if (err < 0) {
1409 		free_page((unsigned long)rxs);
1410 		goto fail;
1411 	}
1412 	info->rx_ring_ref = err;
1413 
1414 	err = xenbus_alloc_evtchn(dev, &info->evtchn);
1415 	if (err)
1416 		goto fail;
1417 
1418 	err = bind_evtchn_to_irqhandler(info->evtchn, xennet_interrupt,
1419 					IRQF_SAMPLE_RANDOM, netdev->name,
1420 					netdev);
1421 	if (err < 0)
1422 		goto fail;
1423 	netdev->irq = err;
1424 	return 0;
1425 
1426  fail:
1427 	return err;
1428 }
1429 
1430 /* Common code used when first setting up, and when resuming. */
1431 static int talk_to_netback(struct xenbus_device *dev,
1432 			   struct netfront_info *info)
1433 {
1434 	const char *message;
1435 	struct xenbus_transaction xbt;
1436 	int err;
1437 
1438 	/* Create shared ring, alloc event channel. */
1439 	err = setup_netfront(dev, info);
1440 	if (err)
1441 		goto out;
1442 
1443 again:
1444 	err = xenbus_transaction_start(&xbt);
1445 	if (err) {
1446 		xenbus_dev_fatal(dev, err, "starting transaction");
1447 		goto destroy_ring;
1448 	}
1449 
1450 	err = xenbus_printf(xbt, dev->nodename, "tx-ring-ref", "%u",
1451 			    info->tx_ring_ref);
1452 	if (err) {
1453 		message = "writing tx ring-ref";
1454 		goto abort_transaction;
1455 	}
1456 	err = xenbus_printf(xbt, dev->nodename, "rx-ring-ref", "%u",
1457 			    info->rx_ring_ref);
1458 	if (err) {
1459 		message = "writing rx ring-ref";
1460 		goto abort_transaction;
1461 	}
1462 	err = xenbus_printf(xbt, dev->nodename,
1463 			    "event-channel", "%u", info->evtchn);
1464 	if (err) {
1465 		message = "writing event-channel";
1466 		goto abort_transaction;
1467 	}
1468 
1469 	err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u",
1470 			    1);
1471 	if (err) {
1472 		message = "writing request-rx-copy";
1473 		goto abort_transaction;
1474 	}
1475 
1476 	err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1);
1477 	if (err) {
1478 		message = "writing feature-rx-notify";
1479 		goto abort_transaction;
1480 	}
1481 
1482 	err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1);
1483 	if (err) {
1484 		message = "writing feature-sg";
1485 		goto abort_transaction;
1486 	}
1487 
1488 	err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1);
1489 	if (err) {
1490 		message = "writing feature-gso-tcpv4";
1491 		goto abort_transaction;
1492 	}
1493 
1494 	err = xenbus_transaction_end(xbt, 0);
1495 	if (err) {
1496 		if (err == -EAGAIN)
1497 			goto again;
1498 		xenbus_dev_fatal(dev, err, "completing transaction");
1499 		goto destroy_ring;
1500 	}
1501 
1502 	return 0;
1503 
1504  abort_transaction:
1505 	xenbus_transaction_end(xbt, 1);
1506 	xenbus_dev_fatal(dev, err, "%s", message);
1507  destroy_ring:
1508 	xennet_disconnect_backend(info);
1509  out:
1510 	return err;
1511 }
1512 
1513 static int xennet_set_sg(struct net_device *dev, u32 data)
1514 {
1515 	if (data) {
1516 		struct netfront_info *np = netdev_priv(dev);
1517 		int val;
1518 
1519 		if (xenbus_scanf(XBT_NIL, np->xbdev->otherend, "feature-sg",
1520 				 "%d", &val) < 0)
1521 			val = 0;
1522 		if (!val)
1523 			return -ENOSYS;
1524 	} else if (dev->mtu > ETH_DATA_LEN)
1525 		dev->mtu = ETH_DATA_LEN;
1526 
1527 	return ethtool_op_set_sg(dev, data);
1528 }
1529 
1530 static int xennet_set_tso(struct net_device *dev, u32 data)
1531 {
1532 	if (data) {
1533 		struct netfront_info *np = netdev_priv(dev);
1534 		int val;
1535 
1536 		if (xenbus_scanf(XBT_NIL, np->xbdev->otherend,
1537 				 "feature-gso-tcpv4", "%d", &val) < 0)
1538 			val = 0;
1539 		if (!val)
1540 			return -ENOSYS;
1541 	}
1542 
1543 	return ethtool_op_set_tso(dev, data);
1544 }
1545 
1546 static void xennet_set_features(struct net_device *dev)
1547 {
1548 	/* Turn off all GSO bits except ROBUST. */
1549 	dev->features &= ~NETIF_F_GSO_MASK;
1550 	dev->features |= NETIF_F_GSO_ROBUST;
1551 	xennet_set_sg(dev, 0);
1552 
1553 	/* We need checksum offload to enable scatter/gather and TSO. */
1554 	if (!(dev->features & NETIF_F_IP_CSUM))
1555 		return;
1556 
1557 	if (!xennet_set_sg(dev, 1))
1558 		xennet_set_tso(dev, 1);
1559 }
1560 
1561 static int xennet_connect(struct net_device *dev)
1562 {
1563 	struct netfront_info *np = netdev_priv(dev);
1564 	int i, requeue_idx, err;
1565 	struct sk_buff *skb;
1566 	grant_ref_t ref;
1567 	struct xen_netif_rx_request *req;
1568 	unsigned int feature_rx_copy;
1569 
1570 	err = xenbus_scanf(XBT_NIL, np->xbdev->otherend,
1571 			   "feature-rx-copy", "%u", &feature_rx_copy);
1572 	if (err != 1)
1573 		feature_rx_copy = 0;
1574 
1575 	if (!feature_rx_copy) {
1576 		dev_info(&dev->dev,
1577 			 "backend does not support copying receive path\n");
1578 		return -ENODEV;
1579 	}
1580 
1581 	err = talk_to_netback(np->xbdev, np);
1582 	if (err)
1583 		return err;
1584 
1585 	xennet_set_features(dev);
1586 
1587 	spin_lock_bh(&np->rx_lock);
1588 	spin_lock_irq(&np->tx_lock);
1589 
1590 	/* Step 1: Discard all pending TX packet fragments. */
1591 	xennet_release_tx_bufs(np);
1592 
1593 	/* Step 2: Rebuild the RX buffer freelist and the RX ring itself. */
1594 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1595 		if (!np->rx_skbs[i])
1596 			continue;
1597 
1598 		skb = np->rx_skbs[requeue_idx] = xennet_get_rx_skb(np, i);
1599 		ref = np->grant_rx_ref[requeue_idx] = xennet_get_rx_ref(np, i);
1600 		req = RING_GET_REQUEST(&np->rx, requeue_idx);
1601 
1602 		gnttab_grant_foreign_access_ref(
1603 			ref, np->xbdev->otherend_id,
1604 			pfn_to_mfn(page_to_pfn(skb_shinfo(skb)->
1605 					       frags->page)),
1606 			0);
1607 		req->gref = ref;
1608 		req->id   = requeue_idx;
1609 
1610 		requeue_idx++;
1611 	}
1612 
1613 	np->rx.req_prod_pvt = requeue_idx;
1614 
1615 	/*
1616 	 * Step 3: All public and private state should now be sane.  Get
1617 	 * ready to start sending and receiving packets and give the driver
1618 	 * domain a kick because we've probably just requeued some
1619 	 * packets.
1620 	 */
1621 	netif_carrier_on(np->netdev);
1622 	notify_remote_via_irq(np->netdev->irq);
1623 	xennet_tx_buf_gc(dev);
1624 	xennet_alloc_rx_buffers(dev);
1625 
1626 	spin_unlock_irq(&np->tx_lock);
1627 	spin_unlock_bh(&np->rx_lock);
1628 
1629 	return 0;
1630 }
1631 
1632 /**
1633  * Callback received when the backend's state changes.
1634  */
1635 static void netback_changed(struct xenbus_device *dev,
1636 			    enum xenbus_state backend_state)
1637 {
1638 	struct netfront_info *np = dev_get_drvdata(&dev->dev);
1639 	struct net_device *netdev = np->netdev;
1640 
1641 	dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state));
1642 
1643 	switch (backend_state) {
1644 	case XenbusStateInitialising:
1645 	case XenbusStateInitialised:
1646 	case XenbusStateReconfiguring:
1647 	case XenbusStateReconfigured:
1648 	case XenbusStateConnected:
1649 	case XenbusStateUnknown:
1650 	case XenbusStateClosed:
1651 		break;
1652 
1653 	case XenbusStateInitWait:
1654 		if (dev->state != XenbusStateInitialising)
1655 			break;
1656 		if (xennet_connect(netdev) != 0)
1657 			break;
1658 		xenbus_switch_state(dev, XenbusStateConnected);
1659 		netif_notify_peers(netdev);
1660 		break;
1661 
1662 	case XenbusStateClosing:
1663 		xenbus_frontend_closed(dev);
1664 		break;
1665 	}
1666 }
1667 
1668 static const struct xennet_stat {
1669 	char name[ETH_GSTRING_LEN];
1670 	u16 offset;
1671 } xennet_stats[] = {
1672 	{
1673 		"rx_gso_checksum_fixup",
1674 		offsetof(struct netfront_info, rx_gso_checksum_fixup)
1675 	},
1676 };
1677 
1678 static int xennet_get_sset_count(struct net_device *dev, int string_set)
1679 {
1680 	switch (string_set) {
1681 	case ETH_SS_STATS:
1682 		return ARRAY_SIZE(xennet_stats);
1683 	default:
1684 		return -EINVAL;
1685 	}
1686 }
1687 
1688 static void xennet_get_ethtool_stats(struct net_device *dev,
1689 				     struct ethtool_stats *stats, u64 * data)
1690 {
1691 	void *np = netdev_priv(dev);
1692 	int i;
1693 
1694 	for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
1695 		data[i] = *(int *)(np + xennet_stats[i].offset);
1696 }
1697 
1698 static void xennet_get_strings(struct net_device *dev, u32 stringset, u8 * data)
1699 {
1700 	int i;
1701 
1702 	switch (stringset) {
1703 	case ETH_SS_STATS:
1704 		for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
1705 			memcpy(data + i * ETH_GSTRING_LEN,
1706 			       xennet_stats[i].name, ETH_GSTRING_LEN);
1707 		break;
1708 	}
1709 }
1710 
1711 static const struct ethtool_ops xennet_ethtool_ops =
1712 {
1713 	.set_tx_csum = ethtool_op_set_tx_csum,
1714 	.set_sg = xennet_set_sg,
1715 	.set_tso = xennet_set_tso,
1716 	.get_link = ethtool_op_get_link,
1717 
1718 	.get_sset_count = xennet_get_sset_count,
1719 	.get_ethtool_stats = xennet_get_ethtool_stats,
1720 	.get_strings = xennet_get_strings,
1721 };
1722 
1723 #ifdef CONFIG_SYSFS
1724 static ssize_t show_rxbuf_min(struct device *dev,
1725 			      struct device_attribute *attr, char *buf)
1726 {
1727 	struct net_device *netdev = to_net_dev(dev);
1728 	struct netfront_info *info = netdev_priv(netdev);
1729 
1730 	return sprintf(buf, "%u\n", info->rx_min_target);
1731 }
1732 
1733 static ssize_t store_rxbuf_min(struct device *dev,
1734 			       struct device_attribute *attr,
1735 			       const char *buf, size_t len)
1736 {
1737 	struct net_device *netdev = to_net_dev(dev);
1738 	struct netfront_info *np = netdev_priv(netdev);
1739 	char *endp;
1740 	unsigned long target;
1741 
1742 	if (!capable(CAP_NET_ADMIN))
1743 		return -EPERM;
1744 
1745 	target = simple_strtoul(buf, &endp, 0);
1746 	if (endp == buf)
1747 		return -EBADMSG;
1748 
1749 	if (target < RX_MIN_TARGET)
1750 		target = RX_MIN_TARGET;
1751 	if (target > RX_MAX_TARGET)
1752 		target = RX_MAX_TARGET;
1753 
1754 	spin_lock_bh(&np->rx_lock);
1755 	if (target > np->rx_max_target)
1756 		np->rx_max_target = target;
1757 	np->rx_min_target = target;
1758 	if (target > np->rx_target)
1759 		np->rx_target = target;
1760 
1761 	xennet_alloc_rx_buffers(netdev);
1762 
1763 	spin_unlock_bh(&np->rx_lock);
1764 	return len;
1765 }
1766 
1767 static ssize_t show_rxbuf_max(struct device *dev,
1768 			      struct device_attribute *attr, char *buf)
1769 {
1770 	struct net_device *netdev = to_net_dev(dev);
1771 	struct netfront_info *info = netdev_priv(netdev);
1772 
1773 	return sprintf(buf, "%u\n", info->rx_max_target);
1774 }
1775 
1776 static ssize_t store_rxbuf_max(struct device *dev,
1777 			       struct device_attribute *attr,
1778 			       const char *buf, size_t len)
1779 {
1780 	struct net_device *netdev = to_net_dev(dev);
1781 	struct netfront_info *np = netdev_priv(netdev);
1782 	char *endp;
1783 	unsigned long target;
1784 
1785 	if (!capable(CAP_NET_ADMIN))
1786 		return -EPERM;
1787 
1788 	target = simple_strtoul(buf, &endp, 0);
1789 	if (endp == buf)
1790 		return -EBADMSG;
1791 
1792 	if (target < RX_MIN_TARGET)
1793 		target = RX_MIN_TARGET;
1794 	if (target > RX_MAX_TARGET)
1795 		target = RX_MAX_TARGET;
1796 
1797 	spin_lock_bh(&np->rx_lock);
1798 	if (target < np->rx_min_target)
1799 		np->rx_min_target = target;
1800 	np->rx_max_target = target;
1801 	if (target < np->rx_target)
1802 		np->rx_target = target;
1803 
1804 	xennet_alloc_rx_buffers(netdev);
1805 
1806 	spin_unlock_bh(&np->rx_lock);
1807 	return len;
1808 }
1809 
1810 static ssize_t show_rxbuf_cur(struct device *dev,
1811 			      struct device_attribute *attr, char *buf)
1812 {
1813 	struct net_device *netdev = to_net_dev(dev);
1814 	struct netfront_info *info = netdev_priv(netdev);
1815 
1816 	return sprintf(buf, "%u\n", info->rx_target);
1817 }
1818 
1819 static struct device_attribute xennet_attrs[] = {
1820 	__ATTR(rxbuf_min, S_IRUGO|S_IWUSR, show_rxbuf_min, store_rxbuf_min),
1821 	__ATTR(rxbuf_max, S_IRUGO|S_IWUSR, show_rxbuf_max, store_rxbuf_max),
1822 	__ATTR(rxbuf_cur, S_IRUGO, show_rxbuf_cur, NULL),
1823 };
1824 
1825 static int xennet_sysfs_addif(struct net_device *netdev)
1826 {
1827 	int i;
1828 	int err;
1829 
1830 	for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++) {
1831 		err = device_create_file(&netdev->dev,
1832 					   &xennet_attrs[i]);
1833 		if (err)
1834 			goto fail;
1835 	}
1836 	return 0;
1837 
1838  fail:
1839 	while (--i >= 0)
1840 		device_remove_file(&netdev->dev, &xennet_attrs[i]);
1841 	return err;
1842 }
1843 
1844 static void xennet_sysfs_delif(struct net_device *netdev)
1845 {
1846 	int i;
1847 
1848 	for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++)
1849 		device_remove_file(&netdev->dev, &xennet_attrs[i]);
1850 }
1851 
1852 #endif /* CONFIG_SYSFS */
1853 
1854 static struct xenbus_device_id netfront_ids[] = {
1855 	{ "vif" },
1856 	{ "" }
1857 };
1858 
1859 
1860 static int __devexit xennet_remove(struct xenbus_device *dev)
1861 {
1862 	struct netfront_info *info = dev_get_drvdata(&dev->dev);
1863 
1864 	dev_dbg(&dev->dev, "%s\n", dev->nodename);
1865 
1866 	unregister_netdev(info->netdev);
1867 
1868 	xennet_disconnect_backend(info);
1869 
1870 	del_timer_sync(&info->rx_refill_timer);
1871 
1872 	xennet_sysfs_delif(info->netdev);
1873 
1874 	free_netdev(info->netdev);
1875 
1876 	return 0;
1877 }
1878 
1879 static struct xenbus_driver netfront_driver = {
1880 	.name = "vif",
1881 	.owner = THIS_MODULE,
1882 	.ids = netfront_ids,
1883 	.probe = netfront_probe,
1884 	.remove = __devexit_p(xennet_remove),
1885 	.resume = netfront_resume,
1886 	.otherend_changed = netback_changed,
1887 };
1888 
1889 static int __init netif_init(void)
1890 {
1891 	if (!xen_domain())
1892 		return -ENODEV;
1893 
1894 	if (xen_initial_domain())
1895 		return 0;
1896 
1897 	printk(KERN_INFO "Initialising Xen virtual ethernet driver.\n");
1898 
1899 	return xenbus_register_frontend(&netfront_driver);
1900 }
1901 module_init(netif_init);
1902 
1903 
1904 static void __exit netif_exit(void)
1905 {
1906 	if (xen_initial_domain())
1907 		return;
1908 
1909 	xenbus_unregister_driver(&netfront_driver);
1910 }
1911 module_exit(netif_exit);
1912 
1913 MODULE_DESCRIPTION("Xen virtual network device frontend");
1914 MODULE_LICENSE("GPL");
1915 MODULE_ALIAS("xen:vif");
1916 MODULE_ALIAS("xennet");
1917