1 /*
2  * Back-end of the driver for virtual network devices. This portion of the
3  * driver exports a 'unified' network-device interface that can be accessed
4  * by any operating system that implements a compatible front end. A
5  * reference front-end implementation can be found in:
6  *  drivers/net/xen-netfront.c
7  *
8  * Copyright (c) 2002-2005, K A Fraser
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public License version 2
12  * as published by the Free Software Foundation; or, when distributed
13  * separately from the Linux kernel or incorporated into other
14  * software packages, subject to the following license:
15  *
16  * Permission is hereby granted, free of charge, to any person obtaining a copy
17  * of this source file (the "Software"), to deal in the Software without
18  * restriction, including without limitation the rights to use, copy, modify,
19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
20  * and to permit persons to whom the Software is furnished to do so, subject to
21  * the following conditions:
22  *
23  * The above copyright notice and this permission notice shall be included in
24  * all copies or substantial portions of the Software.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
32  * IN THE SOFTWARE.
33  */
34 
35 #include "common.h"
36 
37 #include <linux/kthread.h>
38 #include <linux/if_vlan.h>
39 #include <linux/udp.h>
40 
41 #include <net/tcp.h>
42 
43 #include <xen/xen.h>
44 #include <xen/events.h>
45 #include <xen/interface/memory.h>
46 
47 #include <asm/xen/hypercall.h>
48 #include <asm/xen/page.h>
49 
50 /* Provide an option to disable split event channels at load time as
51  * event channels are limited resource. Split event channels are
52  * enabled by default.
53  */
54 bool separate_tx_rx_irq = 1;
55 module_param(separate_tx_rx_irq, bool, 0644);
56 
57 /*
58  * This is the maximum slots a skb can have. If a guest sends a skb
59  * which exceeds this limit it is considered malicious.
60  */
61 #define FATAL_SKB_SLOTS_DEFAULT 20
62 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
63 module_param(fatal_skb_slots, uint, 0444);
64 
65 /*
66  * To avoid confusion, we define XEN_NETBK_LEGACY_SLOTS_MAX indicating
67  * the maximum slots a valid packet can use. Now this value is defined
68  * to be XEN_NETIF_NR_SLOTS_MIN, which is supposed to be supported by
69  * all backend.
70  */
71 #define XEN_NETBK_LEGACY_SLOTS_MAX XEN_NETIF_NR_SLOTS_MIN
72 
73 /*
74  * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
75  * one or more merged tx requests, otherwise it is the continuation of
76  * previous tx request.
77  */
78 static inline int pending_tx_is_head(struct xenvif *vif, RING_IDX idx)
79 {
80 	return vif->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
81 }
82 
83 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
84 			       u8 status);
85 
86 static void make_tx_response(struct xenvif *vif,
87 			     struct xen_netif_tx_request *txp,
88 			     s8       st);
89 
90 static inline int tx_work_todo(struct xenvif *vif);
91 static inline int rx_work_todo(struct xenvif *vif);
92 
93 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
94 					     u16      id,
95 					     s8       st,
96 					     u16      offset,
97 					     u16      size,
98 					     u16      flags);
99 
100 static inline unsigned long idx_to_pfn(struct xenvif *vif,
101 				       u16 idx)
102 {
103 	return page_to_pfn(vif->mmap_pages[idx]);
104 }
105 
106 static inline unsigned long idx_to_kaddr(struct xenvif *vif,
107 					 u16 idx)
108 {
109 	return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
110 }
111 
112 /*
113  * This is the amount of packet we copy rather than map, so that the
114  * guest can't fiddle with the contents of the headers while we do
115  * packet processing on them (netfilter, routing, etc).
116  */
117 #define PKT_PROT_LEN    (ETH_HLEN + \
118 			 VLAN_HLEN + \
119 			 sizeof(struct iphdr) + MAX_IPOPTLEN + \
120 			 sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
121 
122 static u16 frag_get_pending_idx(skb_frag_t *frag)
123 {
124 	return (u16)frag->page_offset;
125 }
126 
127 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
128 {
129 	frag->page_offset = pending_idx;
130 }
131 
132 static inline pending_ring_idx_t pending_index(unsigned i)
133 {
134 	return i & (MAX_PENDING_REQS-1);
135 }
136 
137 static inline pending_ring_idx_t nr_pending_reqs(struct xenvif *vif)
138 {
139 	return MAX_PENDING_REQS -
140 		vif->pending_prod + vif->pending_cons;
141 }
142 
143 static int max_required_rx_slots(struct xenvif *vif)
144 {
145 	int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
146 
147 	/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
148 	if (vif->can_sg || vif->gso || vif->gso_prefix)
149 		max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
150 
151 	return max;
152 }
153 
154 int xenvif_rx_ring_full(struct xenvif *vif)
155 {
156 	RING_IDX peek   = vif->rx_req_cons_peek;
157 	RING_IDX needed = max_required_rx_slots(vif);
158 
159 	return ((vif->rx.sring->req_prod - peek) < needed) ||
160 	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
161 }
162 
163 int xenvif_must_stop_queue(struct xenvif *vif)
164 {
165 	if (!xenvif_rx_ring_full(vif))
166 		return 0;
167 
168 	vif->rx.sring->req_event = vif->rx_req_cons_peek +
169 		max_required_rx_slots(vif);
170 	mb(); /* request notification /then/ check the queue */
171 
172 	return xenvif_rx_ring_full(vif);
173 }
174 
175 /*
176  * Returns true if we should start a new receive buffer instead of
177  * adding 'size' bytes to a buffer which currently contains 'offset'
178  * bytes.
179  */
180 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
181 {
182 	/* simple case: we have completely filled the current buffer. */
183 	if (offset == MAX_BUFFER_OFFSET)
184 		return true;
185 
186 	/*
187 	 * complex case: start a fresh buffer if the current frag
188 	 * would overflow the current buffer but only if:
189 	 *     (i)   this frag would fit completely in the next buffer
190 	 * and (ii)  there is already some data in the current buffer
191 	 * and (iii) this is not the head buffer.
192 	 *
193 	 * Where:
194 	 * - (i) stops us splitting a frag into two copies
195 	 *   unless the frag is too large for a single buffer.
196 	 * - (ii) stops us from leaving a buffer pointlessly empty.
197 	 * - (iii) stops us leaving the first buffer
198 	 *   empty. Strictly speaking this is already covered
199 	 *   by (ii) but is explicitly checked because
200 	 *   netfront relies on the first buffer being
201 	 *   non-empty and can crash otherwise.
202 	 *
203 	 * This means we will effectively linearise small
204 	 * frags but do not needlessly split large buffers
205 	 * into multiple copies tend to give large frags their
206 	 * own buffers as before.
207 	 */
208 	if ((offset + size > MAX_BUFFER_OFFSET) &&
209 	    (size <= MAX_BUFFER_OFFSET) && offset && !head)
210 		return true;
211 
212 	return false;
213 }
214 
215 struct xenvif_count_slot_state {
216 	unsigned long copy_off;
217 	bool head;
218 };
219 
220 unsigned int xenvif_count_frag_slots(struct xenvif *vif,
221 				     unsigned long offset, unsigned long size,
222 				     struct xenvif_count_slot_state *state)
223 {
224 	unsigned count = 0;
225 
226 	offset &= ~PAGE_MASK;
227 
228 	while (size > 0) {
229 		unsigned long bytes;
230 
231 		bytes = PAGE_SIZE - offset;
232 
233 		if (bytes > size)
234 			bytes = size;
235 
236 		if (start_new_rx_buffer(state->copy_off, bytes, state->head)) {
237 			count++;
238 			state->copy_off = 0;
239 		}
240 
241 		if (state->copy_off + bytes > MAX_BUFFER_OFFSET)
242 			bytes = MAX_BUFFER_OFFSET - state->copy_off;
243 
244 		state->copy_off += bytes;
245 
246 		offset += bytes;
247 		size -= bytes;
248 
249 		if (offset == PAGE_SIZE)
250 			offset = 0;
251 
252 		state->head = false;
253 	}
254 
255 	return count;
256 }
257 
258 /*
259  * Figure out how many ring slots we're going to need to send @skb to
260  * the guest. This function is essentially a dry run of
261  * xenvif_gop_frag_copy.
262  */
263 unsigned int xenvif_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
264 {
265 	struct xenvif_count_slot_state state;
266 	unsigned int count;
267 	unsigned char *data;
268 	unsigned i;
269 
270 	state.head = true;
271 	state.copy_off = 0;
272 
273 	/* Slot for the first (partial) page of data. */
274 	count = 1;
275 
276 	/* Need a slot for the GSO prefix for GSO extra data? */
277 	if (skb_shinfo(skb)->gso_size)
278 		count++;
279 
280 	data = skb->data;
281 	while (data < skb_tail_pointer(skb)) {
282 		unsigned long offset = offset_in_page(data);
283 		unsigned long size = PAGE_SIZE - offset;
284 
285 		if (data + size > skb_tail_pointer(skb))
286 			size = skb_tail_pointer(skb) - data;
287 
288 		count += xenvif_count_frag_slots(vif, offset, size, &state);
289 
290 		data += size;
291 	}
292 
293 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
294 		unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
295 		unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
296 
297 		count += xenvif_count_frag_slots(vif, offset, size, &state);
298 	}
299 	return count;
300 }
301 
302 struct netrx_pending_operations {
303 	unsigned copy_prod, copy_cons;
304 	unsigned meta_prod, meta_cons;
305 	struct gnttab_copy *copy;
306 	struct xenvif_rx_meta *meta;
307 	int copy_off;
308 	grant_ref_t copy_gref;
309 };
310 
311 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
312 						 struct netrx_pending_operations *npo)
313 {
314 	struct xenvif_rx_meta *meta;
315 	struct xen_netif_rx_request *req;
316 
317 	req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
318 
319 	meta = npo->meta + npo->meta_prod++;
320 	meta->gso_size = 0;
321 	meta->size = 0;
322 	meta->id = req->id;
323 
324 	npo->copy_off = 0;
325 	npo->copy_gref = req->gref;
326 
327 	return meta;
328 }
329 
330 /*
331  * Set up the grant operations for this fragment. If it's a flipping
332  * interface, we also set up the unmap request from here.
333  */
334 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
335 				 struct netrx_pending_operations *npo,
336 				 struct page *page, unsigned long size,
337 				 unsigned long offset, int *head)
338 {
339 	struct gnttab_copy *copy_gop;
340 	struct xenvif_rx_meta *meta;
341 	unsigned long bytes;
342 
343 	/* Data must not cross a page boundary. */
344 	BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
345 
346 	meta = npo->meta + npo->meta_prod - 1;
347 
348 	/* Skip unused frames from start of page */
349 	page += offset >> PAGE_SHIFT;
350 	offset &= ~PAGE_MASK;
351 
352 	while (size > 0) {
353 		BUG_ON(offset >= PAGE_SIZE);
354 		BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
355 
356 		bytes = PAGE_SIZE - offset;
357 
358 		if (bytes > size)
359 			bytes = size;
360 
361 		if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
362 			/*
363 			 * Netfront requires there to be some data in the head
364 			 * buffer.
365 			 */
366 			BUG_ON(*head);
367 
368 			meta = get_next_rx_buffer(vif, npo);
369 		}
370 
371 		if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
372 			bytes = MAX_BUFFER_OFFSET - npo->copy_off;
373 
374 		copy_gop = npo->copy + npo->copy_prod++;
375 		copy_gop->flags = GNTCOPY_dest_gref;
376 		copy_gop->len = bytes;
377 
378 		copy_gop->source.domid = DOMID_SELF;
379 		copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
380 		copy_gop->source.offset = offset;
381 
382 		copy_gop->dest.domid = vif->domid;
383 		copy_gop->dest.offset = npo->copy_off;
384 		copy_gop->dest.u.ref = npo->copy_gref;
385 
386 		npo->copy_off += bytes;
387 		meta->size += bytes;
388 
389 		offset += bytes;
390 		size -= bytes;
391 
392 		/* Next frame */
393 		if (offset == PAGE_SIZE && size) {
394 			BUG_ON(!PageCompound(page));
395 			page++;
396 			offset = 0;
397 		}
398 
399 		/* Leave a gap for the GSO descriptor. */
400 		if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
401 			vif->rx.req_cons++;
402 
403 		*head = 0; /* There must be something in this buffer now. */
404 
405 	}
406 }
407 
408 /*
409  * Prepare an SKB to be transmitted to the frontend.
410  *
411  * This function is responsible for allocating grant operations, meta
412  * structures, etc.
413  *
414  * It returns the number of meta structures consumed. The number of
415  * ring slots used is always equal to the number of meta slots used
416  * plus the number of GSO descriptors used. Currently, we use either
417  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
418  * frontend-side LRO).
419  */
420 static int xenvif_gop_skb(struct sk_buff *skb,
421 			  struct netrx_pending_operations *npo)
422 {
423 	struct xenvif *vif = netdev_priv(skb->dev);
424 	int nr_frags = skb_shinfo(skb)->nr_frags;
425 	int i;
426 	struct xen_netif_rx_request *req;
427 	struct xenvif_rx_meta *meta;
428 	unsigned char *data;
429 	int head = 1;
430 	int old_meta_prod;
431 
432 	old_meta_prod = npo->meta_prod;
433 
434 	/* Set up a GSO prefix descriptor, if necessary */
435 	if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
436 		req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
437 		meta = npo->meta + npo->meta_prod++;
438 		meta->gso_size = skb_shinfo(skb)->gso_size;
439 		meta->size = 0;
440 		meta->id = req->id;
441 	}
442 
443 	req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
444 	meta = npo->meta + npo->meta_prod++;
445 
446 	if (!vif->gso_prefix)
447 		meta->gso_size = skb_shinfo(skb)->gso_size;
448 	else
449 		meta->gso_size = 0;
450 
451 	meta->size = 0;
452 	meta->id = req->id;
453 	npo->copy_off = 0;
454 	npo->copy_gref = req->gref;
455 
456 	data = skb->data;
457 	while (data < skb_tail_pointer(skb)) {
458 		unsigned int offset = offset_in_page(data);
459 		unsigned int len = PAGE_SIZE - offset;
460 
461 		if (data + len > skb_tail_pointer(skb))
462 			len = skb_tail_pointer(skb) - data;
463 
464 		xenvif_gop_frag_copy(vif, skb, npo,
465 				     virt_to_page(data), len, offset, &head);
466 		data += len;
467 	}
468 
469 	for (i = 0; i < nr_frags; i++) {
470 		xenvif_gop_frag_copy(vif, skb, npo,
471 				     skb_frag_page(&skb_shinfo(skb)->frags[i]),
472 				     skb_frag_size(&skb_shinfo(skb)->frags[i]),
473 				     skb_shinfo(skb)->frags[i].page_offset,
474 				     &head);
475 	}
476 
477 	return npo->meta_prod - old_meta_prod;
478 }
479 
480 /*
481  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
482  * used to set up the operations on the top of
483  * netrx_pending_operations, which have since been done.  Check that
484  * they didn't give any errors and advance over them.
485  */
486 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
487 			    struct netrx_pending_operations *npo)
488 {
489 	struct gnttab_copy     *copy_op;
490 	int status = XEN_NETIF_RSP_OKAY;
491 	int i;
492 
493 	for (i = 0; i < nr_meta_slots; i++) {
494 		copy_op = npo->copy + npo->copy_cons++;
495 		if (copy_op->status != GNTST_okay) {
496 			netdev_dbg(vif->dev,
497 				   "Bad status %d from copy to DOM%d.\n",
498 				   copy_op->status, vif->domid);
499 			status = XEN_NETIF_RSP_ERROR;
500 		}
501 	}
502 
503 	return status;
504 }
505 
506 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
507 				      struct xenvif_rx_meta *meta,
508 				      int nr_meta_slots)
509 {
510 	int i;
511 	unsigned long offset;
512 
513 	/* No fragments used */
514 	if (nr_meta_slots <= 1)
515 		return;
516 
517 	nr_meta_slots--;
518 
519 	for (i = 0; i < nr_meta_slots; i++) {
520 		int flags;
521 		if (i == nr_meta_slots - 1)
522 			flags = 0;
523 		else
524 			flags = XEN_NETRXF_more_data;
525 
526 		offset = 0;
527 		make_rx_response(vif, meta[i].id, status, offset,
528 				 meta[i].size, flags);
529 	}
530 }
531 
532 struct skb_cb_overlay {
533 	int meta_slots_used;
534 };
535 
536 static void xenvif_kick_thread(struct xenvif *vif)
537 {
538 	wake_up(&vif->wq);
539 }
540 
541 void xenvif_rx_action(struct xenvif *vif)
542 {
543 	s8 status;
544 	u16 flags;
545 	struct xen_netif_rx_response *resp;
546 	struct sk_buff_head rxq;
547 	struct sk_buff *skb;
548 	LIST_HEAD(notify);
549 	int ret;
550 	int nr_frags;
551 	int count;
552 	unsigned long offset;
553 	struct skb_cb_overlay *sco;
554 	int need_to_notify = 0;
555 
556 	struct netrx_pending_operations npo = {
557 		.copy  = vif->grant_copy_op,
558 		.meta  = vif->meta,
559 	};
560 
561 	skb_queue_head_init(&rxq);
562 
563 	count = 0;
564 
565 	while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
566 		vif = netdev_priv(skb->dev);
567 		nr_frags = skb_shinfo(skb)->nr_frags;
568 
569 		sco = (struct skb_cb_overlay *)skb->cb;
570 		sco->meta_slots_used = xenvif_gop_skb(skb, &npo);
571 
572 		count += nr_frags + 1;
573 
574 		__skb_queue_tail(&rxq, skb);
575 
576 		/* Filled the batch queue? */
577 		/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
578 		if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
579 			break;
580 	}
581 
582 	BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
583 
584 	if (!npo.copy_prod)
585 		return;
586 
587 	BUG_ON(npo.copy_prod > ARRAY_SIZE(vif->grant_copy_op));
588 	gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
589 
590 	while ((skb = __skb_dequeue(&rxq)) != NULL) {
591 		sco = (struct skb_cb_overlay *)skb->cb;
592 
593 		vif = netdev_priv(skb->dev);
594 
595 		if (vif->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
596 			resp = RING_GET_RESPONSE(&vif->rx,
597 						 vif->rx.rsp_prod_pvt++);
598 
599 			resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
600 
601 			resp->offset = vif->meta[npo.meta_cons].gso_size;
602 			resp->id = vif->meta[npo.meta_cons].id;
603 			resp->status = sco->meta_slots_used;
604 
605 			npo.meta_cons++;
606 			sco->meta_slots_used--;
607 		}
608 
609 
610 		vif->dev->stats.tx_bytes += skb->len;
611 		vif->dev->stats.tx_packets++;
612 
613 		status = xenvif_check_gop(vif, sco->meta_slots_used, &npo);
614 
615 		if (sco->meta_slots_used == 1)
616 			flags = 0;
617 		else
618 			flags = XEN_NETRXF_more_data;
619 
620 		if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
621 			flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
622 		else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
623 			/* remote but checksummed. */
624 			flags |= XEN_NETRXF_data_validated;
625 
626 		offset = 0;
627 		resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
628 					status, offset,
629 					vif->meta[npo.meta_cons].size,
630 					flags);
631 
632 		if (vif->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
633 			struct xen_netif_extra_info *gso =
634 				(struct xen_netif_extra_info *)
635 				RING_GET_RESPONSE(&vif->rx,
636 						  vif->rx.rsp_prod_pvt++);
637 
638 			resp->flags |= XEN_NETRXF_extra_info;
639 
640 			gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
641 			gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
642 			gso->u.gso.pad = 0;
643 			gso->u.gso.features = 0;
644 
645 			gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
646 			gso->flags = 0;
647 		}
648 
649 		xenvif_add_frag_responses(vif, status,
650 					  vif->meta + npo.meta_cons + 1,
651 					  sco->meta_slots_used);
652 
653 		RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
654 
655 		if (ret)
656 			need_to_notify = 1;
657 
658 		xenvif_notify_tx_completion(vif);
659 
660 		npo.meta_cons += sco->meta_slots_used;
661 		dev_kfree_skb(skb);
662 	}
663 
664 	if (need_to_notify)
665 		notify_remote_via_irq(vif->rx_irq);
666 
667 	/* More work to do? */
668 	if (!skb_queue_empty(&vif->rx_queue))
669 		xenvif_kick_thread(vif);
670 }
671 
672 void xenvif_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
673 {
674 	skb_queue_tail(&vif->rx_queue, skb);
675 
676 	xenvif_kick_thread(vif);
677 }
678 
679 void xenvif_check_rx_xenvif(struct xenvif *vif)
680 {
681 	int more_to_do;
682 
683 	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
684 
685 	if (more_to_do)
686 		napi_schedule(&vif->napi);
687 }
688 
689 static void tx_add_credit(struct xenvif *vif)
690 {
691 	unsigned long max_burst, max_credit;
692 
693 	/*
694 	 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
695 	 * Otherwise the interface can seize up due to insufficient credit.
696 	 */
697 	max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
698 	max_burst = min(max_burst, 131072UL);
699 	max_burst = max(max_burst, vif->credit_bytes);
700 
701 	/* Take care that adding a new chunk of credit doesn't wrap to zero. */
702 	max_credit = vif->remaining_credit + vif->credit_bytes;
703 	if (max_credit < vif->remaining_credit)
704 		max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
705 
706 	vif->remaining_credit = min(max_credit, max_burst);
707 }
708 
709 static void tx_credit_callback(unsigned long data)
710 {
711 	struct xenvif *vif = (struct xenvif *)data;
712 	tx_add_credit(vif);
713 	xenvif_check_rx_xenvif(vif);
714 }
715 
716 static void xenvif_tx_err(struct xenvif *vif,
717 			  struct xen_netif_tx_request *txp, RING_IDX end)
718 {
719 	RING_IDX cons = vif->tx.req_cons;
720 
721 	do {
722 		make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
723 		if (cons == end)
724 			break;
725 		txp = RING_GET_REQUEST(&vif->tx, cons++);
726 	} while (1);
727 	vif->tx.req_cons = cons;
728 }
729 
730 static void xenvif_fatal_tx_err(struct xenvif *vif)
731 {
732 	netdev_err(vif->dev, "fatal error; disabling device\n");
733 	xenvif_carrier_off(vif);
734 }
735 
736 static int xenvif_count_requests(struct xenvif *vif,
737 				 struct xen_netif_tx_request *first,
738 				 struct xen_netif_tx_request *txp,
739 				 int work_to_do)
740 {
741 	RING_IDX cons = vif->tx.req_cons;
742 	int slots = 0;
743 	int drop_err = 0;
744 	int more_data;
745 
746 	if (!(first->flags & XEN_NETTXF_more_data))
747 		return 0;
748 
749 	do {
750 		struct xen_netif_tx_request dropped_tx = { 0 };
751 
752 		if (slots >= work_to_do) {
753 			netdev_err(vif->dev,
754 				   "Asked for %d slots but exceeds this limit\n",
755 				   work_to_do);
756 			xenvif_fatal_tx_err(vif);
757 			return -ENODATA;
758 		}
759 
760 		/* This guest is really using too many slots and
761 		 * considered malicious.
762 		 */
763 		if (unlikely(slots >= fatal_skb_slots)) {
764 			netdev_err(vif->dev,
765 				   "Malicious frontend using %d slots, threshold %u\n",
766 				   slots, fatal_skb_slots);
767 			xenvif_fatal_tx_err(vif);
768 			return -E2BIG;
769 		}
770 
771 		/* Xen network protocol had implicit dependency on
772 		 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
773 		 * the historical MAX_SKB_FRAGS value 18 to honor the
774 		 * same behavior as before. Any packet using more than
775 		 * 18 slots but less than fatal_skb_slots slots is
776 		 * dropped
777 		 */
778 		if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
779 			if (net_ratelimit())
780 				netdev_dbg(vif->dev,
781 					   "Too many slots (%d) exceeding limit (%d), dropping packet\n",
782 					   slots, XEN_NETBK_LEGACY_SLOTS_MAX);
783 			drop_err = -E2BIG;
784 		}
785 
786 		if (drop_err)
787 			txp = &dropped_tx;
788 
789 		memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
790 		       sizeof(*txp));
791 
792 		/* If the guest submitted a frame >= 64 KiB then
793 		 * first->size overflowed and following slots will
794 		 * appear to be larger than the frame.
795 		 *
796 		 * This cannot be fatal error as there are buggy
797 		 * frontends that do this.
798 		 *
799 		 * Consume all slots and drop the packet.
800 		 */
801 		if (!drop_err && txp->size > first->size) {
802 			if (net_ratelimit())
803 				netdev_dbg(vif->dev,
804 					   "Invalid tx request, slot size %u > remaining size %u\n",
805 					   txp->size, first->size);
806 			drop_err = -EIO;
807 		}
808 
809 		first->size -= txp->size;
810 		slots++;
811 
812 		if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
813 			netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
814 				 txp->offset, txp->size);
815 			xenvif_fatal_tx_err(vif);
816 			return -EINVAL;
817 		}
818 
819 		more_data = txp->flags & XEN_NETTXF_more_data;
820 
821 		if (!drop_err)
822 			txp++;
823 
824 	} while (more_data);
825 
826 	if (drop_err) {
827 		xenvif_tx_err(vif, first, cons + slots);
828 		return drop_err;
829 	}
830 
831 	return slots;
832 }
833 
834 static struct page *xenvif_alloc_page(struct xenvif *vif,
835 				      u16 pending_idx)
836 {
837 	struct page *page;
838 
839 	page = alloc_page(GFP_ATOMIC|__GFP_COLD);
840 	if (!page)
841 		return NULL;
842 	vif->mmap_pages[pending_idx] = page;
843 
844 	return page;
845 }
846 
847 static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
848 					       struct sk_buff *skb,
849 					       struct xen_netif_tx_request *txp,
850 					       struct gnttab_copy *gop)
851 {
852 	struct skb_shared_info *shinfo = skb_shinfo(skb);
853 	skb_frag_t *frags = shinfo->frags;
854 	u16 pending_idx = *((u16 *)skb->data);
855 	u16 head_idx = 0;
856 	int slot, start;
857 	struct page *page;
858 	pending_ring_idx_t index, start_idx = 0;
859 	uint16_t dst_offset;
860 	unsigned int nr_slots;
861 	struct pending_tx_info *first = NULL;
862 
863 	/* At this point shinfo->nr_frags is in fact the number of
864 	 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
865 	 */
866 	nr_slots = shinfo->nr_frags;
867 
868 	/* Skip first skb fragment if it is on same page as header fragment. */
869 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
870 
871 	/* Coalesce tx requests, at this point the packet passed in
872 	 * should be <= 64K. Any packets larger than 64K have been
873 	 * handled in xenvif_count_requests().
874 	 */
875 	for (shinfo->nr_frags = slot = start; slot < nr_slots;
876 	     shinfo->nr_frags++) {
877 		struct pending_tx_info *pending_tx_info =
878 			vif->pending_tx_info;
879 
880 		page = alloc_page(GFP_ATOMIC|__GFP_COLD);
881 		if (!page)
882 			goto err;
883 
884 		dst_offset = 0;
885 		first = NULL;
886 		while (dst_offset < PAGE_SIZE && slot < nr_slots) {
887 			gop->flags = GNTCOPY_source_gref;
888 
889 			gop->source.u.ref = txp->gref;
890 			gop->source.domid = vif->domid;
891 			gop->source.offset = txp->offset;
892 
893 			gop->dest.domid = DOMID_SELF;
894 
895 			gop->dest.offset = dst_offset;
896 			gop->dest.u.gmfn = virt_to_mfn(page_address(page));
897 
898 			if (dst_offset + txp->size > PAGE_SIZE) {
899 				/* This page can only merge a portion
900 				 * of tx request. Do not increment any
901 				 * pointer / counter here. The txp
902 				 * will be dealt with in future
903 				 * rounds, eventually hitting the
904 				 * `else` branch.
905 				 */
906 				gop->len = PAGE_SIZE - dst_offset;
907 				txp->offset += gop->len;
908 				txp->size -= gop->len;
909 				dst_offset += gop->len; /* quit loop */
910 			} else {
911 				/* This tx request can be merged in the page */
912 				gop->len = txp->size;
913 				dst_offset += gop->len;
914 
915 				index = pending_index(vif->pending_cons++);
916 
917 				pending_idx = vif->pending_ring[index];
918 
919 				memcpy(&pending_tx_info[pending_idx].req, txp,
920 				       sizeof(*txp));
921 
922 				/* Poison these fields, corresponding
923 				 * fields for head tx req will be set
924 				 * to correct values after the loop.
925 				 */
926 				vif->mmap_pages[pending_idx] = (void *)(~0UL);
927 				pending_tx_info[pending_idx].head =
928 					INVALID_PENDING_RING_IDX;
929 
930 				if (!first) {
931 					first = &pending_tx_info[pending_idx];
932 					start_idx = index;
933 					head_idx = pending_idx;
934 				}
935 
936 				txp++;
937 				slot++;
938 			}
939 
940 			gop++;
941 		}
942 
943 		first->req.offset = 0;
944 		first->req.size = dst_offset;
945 		first->head = start_idx;
946 		vif->mmap_pages[head_idx] = page;
947 		frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
948 	}
949 
950 	BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
951 
952 	return gop;
953 err:
954 	/* Unwind, freeing all pages and sending error responses. */
955 	while (shinfo->nr_frags-- > start) {
956 		xenvif_idx_release(vif,
957 				frag_get_pending_idx(&frags[shinfo->nr_frags]),
958 				XEN_NETIF_RSP_ERROR);
959 	}
960 	/* The head too, if necessary. */
961 	if (start)
962 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
963 
964 	return NULL;
965 }
966 
967 static int xenvif_tx_check_gop(struct xenvif *vif,
968 			       struct sk_buff *skb,
969 			       struct gnttab_copy **gopp)
970 {
971 	struct gnttab_copy *gop = *gopp;
972 	u16 pending_idx = *((u16 *)skb->data);
973 	struct skb_shared_info *shinfo = skb_shinfo(skb);
974 	struct pending_tx_info *tx_info;
975 	int nr_frags = shinfo->nr_frags;
976 	int i, err, start;
977 	u16 peek; /* peek into next tx request */
978 
979 	/* Check status of header. */
980 	err = gop->status;
981 	if (unlikely(err))
982 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
983 
984 	/* Skip first skb fragment if it is on same page as header fragment. */
985 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
986 
987 	for (i = start; i < nr_frags; i++) {
988 		int j, newerr;
989 		pending_ring_idx_t head;
990 
991 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
992 		tx_info = &vif->pending_tx_info[pending_idx];
993 		head = tx_info->head;
994 
995 		/* Check error status: if okay then remember grant handle. */
996 		do {
997 			newerr = (++gop)->status;
998 			if (newerr)
999 				break;
1000 			peek = vif->pending_ring[pending_index(++head)];
1001 		} while (!pending_tx_is_head(vif, peek));
1002 
1003 		if (likely(!newerr)) {
1004 			/* Had a previous error? Invalidate this fragment. */
1005 			if (unlikely(err))
1006 				xenvif_idx_release(vif, pending_idx,
1007 						   XEN_NETIF_RSP_OKAY);
1008 			continue;
1009 		}
1010 
1011 		/* Error on this fragment: respond to client with an error. */
1012 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
1013 
1014 		/* Not the first error? Preceding frags already invalidated. */
1015 		if (err)
1016 			continue;
1017 
1018 		/* First error: invalidate header and preceding fragments. */
1019 		pending_idx = *((u16 *)skb->data);
1020 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1021 		for (j = start; j < i; j++) {
1022 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1023 			xenvif_idx_release(vif, pending_idx,
1024 					   XEN_NETIF_RSP_OKAY);
1025 		}
1026 
1027 		/* Remember the error: invalidate all subsequent fragments. */
1028 		err = newerr;
1029 	}
1030 
1031 	*gopp = gop + 1;
1032 	return err;
1033 }
1034 
1035 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
1036 {
1037 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1038 	int nr_frags = shinfo->nr_frags;
1039 	int i;
1040 
1041 	for (i = 0; i < nr_frags; i++) {
1042 		skb_frag_t *frag = shinfo->frags + i;
1043 		struct xen_netif_tx_request *txp;
1044 		struct page *page;
1045 		u16 pending_idx;
1046 
1047 		pending_idx = frag_get_pending_idx(frag);
1048 
1049 		txp = &vif->pending_tx_info[pending_idx].req;
1050 		page = virt_to_page(idx_to_kaddr(vif, pending_idx));
1051 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1052 		skb->len += txp->size;
1053 		skb->data_len += txp->size;
1054 		skb->truesize += txp->size;
1055 
1056 		/* Take an extra reference to offset xenvif_idx_release */
1057 		get_page(vif->mmap_pages[pending_idx]);
1058 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1059 	}
1060 }
1061 
1062 static int xenvif_get_extras(struct xenvif *vif,
1063 				struct xen_netif_extra_info *extras,
1064 				int work_to_do)
1065 {
1066 	struct xen_netif_extra_info extra;
1067 	RING_IDX cons = vif->tx.req_cons;
1068 
1069 	do {
1070 		if (unlikely(work_to_do-- <= 0)) {
1071 			netdev_err(vif->dev, "Missing extra info\n");
1072 			xenvif_fatal_tx_err(vif);
1073 			return -EBADR;
1074 		}
1075 
1076 		memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1077 		       sizeof(extra));
1078 		if (unlikely(!extra.type ||
1079 			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1080 			vif->tx.req_cons = ++cons;
1081 			netdev_err(vif->dev,
1082 				   "Invalid extra type: %d\n", extra.type);
1083 			xenvif_fatal_tx_err(vif);
1084 			return -EINVAL;
1085 		}
1086 
1087 		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1088 		vif->tx.req_cons = ++cons;
1089 	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1090 
1091 	return work_to_do;
1092 }
1093 
1094 static int xenvif_set_skb_gso(struct xenvif *vif,
1095 			      struct sk_buff *skb,
1096 			      struct xen_netif_extra_info *gso)
1097 {
1098 	if (!gso->u.gso.size) {
1099 		netdev_err(vif->dev, "GSO size must not be zero.\n");
1100 		xenvif_fatal_tx_err(vif);
1101 		return -EINVAL;
1102 	}
1103 
1104 	/* Currently only TCPv4 S.O. is supported. */
1105 	if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1106 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1107 		xenvif_fatal_tx_err(vif);
1108 		return -EINVAL;
1109 	}
1110 
1111 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
1112 	skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1113 
1114 	/* Header must be checked, and gso_segs computed. */
1115 	skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1116 	skb_shinfo(skb)->gso_segs = 0;
1117 
1118 	return 0;
1119 }
1120 
1121 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1122 {
1123 	struct iphdr *iph;
1124 	int err = -EPROTO;
1125 	int recalculate_partial_csum = 0;
1126 
1127 	/*
1128 	 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1129 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
1130 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1131 	 * recalculate the partial checksum.
1132 	 */
1133 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1134 		vif->rx_gso_checksum_fixup++;
1135 		skb->ip_summed = CHECKSUM_PARTIAL;
1136 		recalculate_partial_csum = 1;
1137 	}
1138 
1139 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1140 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1141 		return 0;
1142 
1143 	if (skb->protocol != htons(ETH_P_IP))
1144 		goto out;
1145 
1146 	iph = (void *)skb->data;
1147 	switch (iph->protocol) {
1148 	case IPPROTO_TCP:
1149 		if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1150 					  offsetof(struct tcphdr, check)))
1151 			goto out;
1152 
1153 		if (recalculate_partial_csum) {
1154 			struct tcphdr *tcph = tcp_hdr(skb);
1155 			tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1156 							 skb->len - iph->ihl*4,
1157 							 IPPROTO_TCP, 0);
1158 		}
1159 		break;
1160 	case IPPROTO_UDP:
1161 		if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1162 					  offsetof(struct udphdr, check)))
1163 			goto out;
1164 
1165 		if (recalculate_partial_csum) {
1166 			struct udphdr *udph = udp_hdr(skb);
1167 			udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1168 							 skb->len - iph->ihl*4,
1169 							 IPPROTO_UDP, 0);
1170 		}
1171 		break;
1172 	default:
1173 		if (net_ratelimit())
1174 			netdev_err(vif->dev,
1175 				   "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1176 				   iph->protocol);
1177 		goto out;
1178 	}
1179 
1180 	err = 0;
1181 
1182 out:
1183 	return err;
1184 }
1185 
1186 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1187 {
1188 	unsigned long now = jiffies;
1189 	unsigned long next_credit =
1190 		vif->credit_timeout.expires +
1191 		msecs_to_jiffies(vif->credit_usec / 1000);
1192 
1193 	/* Timer could already be pending in rare cases. */
1194 	if (timer_pending(&vif->credit_timeout))
1195 		return true;
1196 
1197 	/* Passed the point where we can replenish credit? */
1198 	if (time_after_eq(now, next_credit)) {
1199 		vif->credit_timeout.expires = now;
1200 		tx_add_credit(vif);
1201 	}
1202 
1203 	/* Still too big to send right now? Set a callback. */
1204 	if (size > vif->remaining_credit) {
1205 		vif->credit_timeout.data     =
1206 			(unsigned long)vif;
1207 		vif->credit_timeout.function =
1208 			tx_credit_callback;
1209 		mod_timer(&vif->credit_timeout,
1210 			  next_credit);
1211 
1212 		return true;
1213 	}
1214 
1215 	return false;
1216 }
1217 
1218 static unsigned xenvif_tx_build_gops(struct xenvif *vif)
1219 {
1220 	struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
1221 	struct sk_buff *skb;
1222 	int ret;
1223 
1224 	while ((nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1225 		< MAX_PENDING_REQS)) {
1226 		struct xen_netif_tx_request txreq;
1227 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1228 		struct page *page;
1229 		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1230 		u16 pending_idx;
1231 		RING_IDX idx;
1232 		int work_to_do;
1233 		unsigned int data_len;
1234 		pending_ring_idx_t index;
1235 
1236 		if (vif->tx.sring->req_prod - vif->tx.req_cons >
1237 		    XEN_NETIF_TX_RING_SIZE) {
1238 			netdev_err(vif->dev,
1239 				   "Impossible number of requests. "
1240 				   "req_prod %d, req_cons %d, size %ld\n",
1241 				   vif->tx.sring->req_prod, vif->tx.req_cons,
1242 				   XEN_NETIF_TX_RING_SIZE);
1243 			xenvif_fatal_tx_err(vif);
1244 			continue;
1245 		}
1246 
1247 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1248 		if (!work_to_do)
1249 			break;
1250 
1251 		idx = vif->tx.req_cons;
1252 		rmb(); /* Ensure that we see the request before we copy it. */
1253 		memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1254 
1255 		/* Credit-based scheduling. */
1256 		if (txreq.size > vif->remaining_credit &&
1257 		    tx_credit_exceeded(vif, txreq.size))
1258 			break;
1259 
1260 		vif->remaining_credit -= txreq.size;
1261 
1262 		work_to_do--;
1263 		vif->tx.req_cons = ++idx;
1264 
1265 		memset(extras, 0, sizeof(extras));
1266 		if (txreq.flags & XEN_NETTXF_extra_info) {
1267 			work_to_do = xenvif_get_extras(vif, extras,
1268 						       work_to_do);
1269 			idx = vif->tx.req_cons;
1270 			if (unlikely(work_to_do < 0))
1271 				break;
1272 		}
1273 
1274 		ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1275 		if (unlikely(ret < 0))
1276 			break;
1277 
1278 		idx += ret;
1279 
1280 		if (unlikely(txreq.size < ETH_HLEN)) {
1281 			netdev_dbg(vif->dev,
1282 				   "Bad packet size: %d\n", txreq.size);
1283 			xenvif_tx_err(vif, &txreq, idx);
1284 			break;
1285 		}
1286 
1287 		/* No crossing a page as the payload mustn't fragment. */
1288 		if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1289 			netdev_err(vif->dev,
1290 				   "txreq.offset: %x, size: %u, end: %lu\n",
1291 				   txreq.offset, txreq.size,
1292 				   (txreq.offset&~PAGE_MASK) + txreq.size);
1293 			xenvif_fatal_tx_err(vif);
1294 			break;
1295 		}
1296 
1297 		index = pending_index(vif->pending_cons);
1298 		pending_idx = vif->pending_ring[index];
1299 
1300 		data_len = (txreq.size > PKT_PROT_LEN &&
1301 			    ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1302 			PKT_PROT_LEN : txreq.size;
1303 
1304 		skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1305 				GFP_ATOMIC | __GFP_NOWARN);
1306 		if (unlikely(skb == NULL)) {
1307 			netdev_dbg(vif->dev,
1308 				   "Can't allocate a skb in start_xmit.\n");
1309 			xenvif_tx_err(vif, &txreq, idx);
1310 			break;
1311 		}
1312 
1313 		/* Packets passed to netif_rx() must have some headroom. */
1314 		skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1315 
1316 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1317 			struct xen_netif_extra_info *gso;
1318 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1319 
1320 			if (xenvif_set_skb_gso(vif, skb, gso)) {
1321 				/* Failure in xenvif_set_skb_gso is fatal. */
1322 				kfree_skb(skb);
1323 				break;
1324 			}
1325 		}
1326 
1327 		/* XXX could copy straight to head */
1328 		page = xenvif_alloc_page(vif, pending_idx);
1329 		if (!page) {
1330 			kfree_skb(skb);
1331 			xenvif_tx_err(vif, &txreq, idx);
1332 			break;
1333 		}
1334 
1335 		gop->source.u.ref = txreq.gref;
1336 		gop->source.domid = vif->domid;
1337 		gop->source.offset = txreq.offset;
1338 
1339 		gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1340 		gop->dest.domid = DOMID_SELF;
1341 		gop->dest.offset = txreq.offset;
1342 
1343 		gop->len = txreq.size;
1344 		gop->flags = GNTCOPY_source_gref;
1345 
1346 		gop++;
1347 
1348 		memcpy(&vif->pending_tx_info[pending_idx].req,
1349 		       &txreq, sizeof(txreq));
1350 		vif->pending_tx_info[pending_idx].head = index;
1351 		*((u16 *)skb->data) = pending_idx;
1352 
1353 		__skb_put(skb, data_len);
1354 
1355 		skb_shinfo(skb)->nr_frags = ret;
1356 		if (data_len < txreq.size) {
1357 			skb_shinfo(skb)->nr_frags++;
1358 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1359 					     pending_idx);
1360 		} else {
1361 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1362 					     INVALID_PENDING_IDX);
1363 		}
1364 
1365 		vif->pending_cons++;
1366 
1367 		request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1368 		if (request_gop == NULL) {
1369 			kfree_skb(skb);
1370 			xenvif_tx_err(vif, &txreq, idx);
1371 			break;
1372 		}
1373 		gop = request_gop;
1374 
1375 		__skb_queue_tail(&vif->tx_queue, skb);
1376 
1377 		vif->tx.req_cons = idx;
1378 
1379 		if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
1380 			break;
1381 	}
1382 
1383 	return gop - vif->tx_copy_ops;
1384 }
1385 
1386 
1387 static int xenvif_tx_submit(struct xenvif *vif, int budget)
1388 {
1389 	struct gnttab_copy *gop = vif->tx_copy_ops;
1390 	struct sk_buff *skb;
1391 	int work_done = 0;
1392 
1393 	while (work_done < budget &&
1394 	       (skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1395 		struct xen_netif_tx_request *txp;
1396 		u16 pending_idx;
1397 		unsigned data_len;
1398 
1399 		pending_idx = *((u16 *)skb->data);
1400 		txp = &vif->pending_tx_info[pending_idx].req;
1401 
1402 		/* Check the remap error code. */
1403 		if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1404 			netdev_dbg(vif->dev, "netback grant failed.\n");
1405 			skb_shinfo(skb)->nr_frags = 0;
1406 			kfree_skb(skb);
1407 			continue;
1408 		}
1409 
1410 		data_len = skb->len;
1411 		memcpy(skb->data,
1412 		       (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1413 		       data_len);
1414 		if (data_len < txp->size) {
1415 			/* Append the packet payload as a fragment. */
1416 			txp->offset += data_len;
1417 			txp->size -= data_len;
1418 		} else {
1419 			/* Schedule a response immediately. */
1420 			xenvif_idx_release(vif, pending_idx,
1421 					   XEN_NETIF_RSP_OKAY);
1422 		}
1423 
1424 		if (txp->flags & XEN_NETTXF_csum_blank)
1425 			skb->ip_summed = CHECKSUM_PARTIAL;
1426 		else if (txp->flags & XEN_NETTXF_data_validated)
1427 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1428 
1429 		xenvif_fill_frags(vif, skb);
1430 
1431 		/*
1432 		 * If the initial fragment was < PKT_PROT_LEN then
1433 		 * pull through some bytes from the other fragments to
1434 		 * increase the linear region to PKT_PROT_LEN bytes.
1435 		 */
1436 		if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1437 			int target = min_t(int, skb->len, PKT_PROT_LEN);
1438 			__pskb_pull_tail(skb, target - skb_headlen(skb));
1439 		}
1440 
1441 		skb->dev      = vif->dev;
1442 		skb->protocol = eth_type_trans(skb, skb->dev);
1443 		skb_reset_network_header(skb);
1444 
1445 		if (checksum_setup(vif, skb)) {
1446 			netdev_dbg(vif->dev,
1447 				   "Can't setup checksum in net_tx_action\n");
1448 			kfree_skb(skb);
1449 			continue;
1450 		}
1451 
1452 		skb_probe_transport_header(skb, 0);
1453 
1454 		vif->dev->stats.rx_bytes += skb->len;
1455 		vif->dev->stats.rx_packets++;
1456 
1457 		work_done++;
1458 
1459 		netif_receive_skb(skb);
1460 	}
1461 
1462 	return work_done;
1463 }
1464 
1465 /* Called after netfront has transmitted */
1466 int xenvif_tx_action(struct xenvif *vif, int budget)
1467 {
1468 	unsigned nr_gops;
1469 	int work_done;
1470 
1471 	if (unlikely(!tx_work_todo(vif)))
1472 		return 0;
1473 
1474 	nr_gops = xenvif_tx_build_gops(vif);
1475 
1476 	if (nr_gops == 0)
1477 		return 0;
1478 
1479 	gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
1480 
1481 	work_done = xenvif_tx_submit(vif, nr_gops);
1482 
1483 	return work_done;
1484 }
1485 
1486 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1487 			       u8 status)
1488 {
1489 	struct pending_tx_info *pending_tx_info;
1490 	pending_ring_idx_t head;
1491 	u16 peek; /* peek into next tx request */
1492 
1493 	BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
1494 
1495 	/* Already complete? */
1496 	if (vif->mmap_pages[pending_idx] == NULL)
1497 		return;
1498 
1499 	pending_tx_info = &vif->pending_tx_info[pending_idx];
1500 
1501 	head = pending_tx_info->head;
1502 
1503 	BUG_ON(!pending_tx_is_head(vif, head));
1504 	BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
1505 
1506 	do {
1507 		pending_ring_idx_t index;
1508 		pending_ring_idx_t idx = pending_index(head);
1509 		u16 info_idx = vif->pending_ring[idx];
1510 
1511 		pending_tx_info = &vif->pending_tx_info[info_idx];
1512 		make_tx_response(vif, &pending_tx_info->req, status);
1513 
1514 		/* Setting any number other than
1515 		 * INVALID_PENDING_RING_IDX indicates this slot is
1516 		 * starting a new packet / ending a previous packet.
1517 		 */
1518 		pending_tx_info->head = 0;
1519 
1520 		index = pending_index(vif->pending_prod++);
1521 		vif->pending_ring[index] = vif->pending_ring[info_idx];
1522 
1523 		peek = vif->pending_ring[pending_index(++head)];
1524 
1525 	} while (!pending_tx_is_head(vif, peek));
1526 
1527 	put_page(vif->mmap_pages[pending_idx]);
1528 	vif->mmap_pages[pending_idx] = NULL;
1529 }
1530 
1531 
1532 static void make_tx_response(struct xenvif *vif,
1533 			     struct xen_netif_tx_request *txp,
1534 			     s8       st)
1535 {
1536 	RING_IDX i = vif->tx.rsp_prod_pvt;
1537 	struct xen_netif_tx_response *resp;
1538 	int notify;
1539 
1540 	resp = RING_GET_RESPONSE(&vif->tx, i);
1541 	resp->id     = txp->id;
1542 	resp->status = st;
1543 
1544 	if (txp->flags & XEN_NETTXF_extra_info)
1545 		RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1546 
1547 	vif->tx.rsp_prod_pvt = ++i;
1548 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1549 	if (notify)
1550 		notify_remote_via_irq(vif->tx_irq);
1551 }
1552 
1553 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1554 					     u16      id,
1555 					     s8       st,
1556 					     u16      offset,
1557 					     u16      size,
1558 					     u16      flags)
1559 {
1560 	RING_IDX i = vif->rx.rsp_prod_pvt;
1561 	struct xen_netif_rx_response *resp;
1562 
1563 	resp = RING_GET_RESPONSE(&vif->rx, i);
1564 	resp->offset     = offset;
1565 	resp->flags      = flags;
1566 	resp->id         = id;
1567 	resp->status     = (s16)size;
1568 	if (st < 0)
1569 		resp->status = (s16)st;
1570 
1571 	vif->rx.rsp_prod_pvt = ++i;
1572 
1573 	return resp;
1574 }
1575 
1576 static inline int rx_work_todo(struct xenvif *vif)
1577 {
1578 	return !skb_queue_empty(&vif->rx_queue);
1579 }
1580 
1581 static inline int tx_work_todo(struct xenvif *vif)
1582 {
1583 
1584 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1585 	    (nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1586 	     < MAX_PENDING_REQS))
1587 		return 1;
1588 
1589 	return 0;
1590 }
1591 
1592 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1593 {
1594 	if (vif->tx.sring)
1595 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1596 					vif->tx.sring);
1597 	if (vif->rx.sring)
1598 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1599 					vif->rx.sring);
1600 }
1601 
1602 int xenvif_map_frontend_rings(struct xenvif *vif,
1603 			      grant_ref_t tx_ring_ref,
1604 			      grant_ref_t rx_ring_ref)
1605 {
1606 	void *addr;
1607 	struct xen_netif_tx_sring *txs;
1608 	struct xen_netif_rx_sring *rxs;
1609 
1610 	int err = -ENOMEM;
1611 
1612 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1613 				     tx_ring_ref, &addr);
1614 	if (err)
1615 		goto err;
1616 
1617 	txs = (struct xen_netif_tx_sring *)addr;
1618 	BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1619 
1620 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1621 				     rx_ring_ref, &addr);
1622 	if (err)
1623 		goto err;
1624 
1625 	rxs = (struct xen_netif_rx_sring *)addr;
1626 	BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1627 
1628 	vif->rx_req_cons_peek = 0;
1629 
1630 	return 0;
1631 
1632 err:
1633 	xenvif_unmap_frontend_rings(vif);
1634 	return err;
1635 }
1636 
1637 int xenvif_kthread(void *data)
1638 {
1639 	struct xenvif *vif = data;
1640 
1641 	while (!kthread_should_stop()) {
1642 		wait_event_interruptible(vif->wq,
1643 					 rx_work_todo(vif) ||
1644 					 kthread_should_stop());
1645 		if (kthread_should_stop())
1646 			break;
1647 
1648 		if (rx_work_todo(vif))
1649 			xenvif_rx_action(vif);
1650 
1651 		cond_resched();
1652 	}
1653 
1654 	return 0;
1655 }
1656 
1657 static int __init netback_init(void)
1658 {
1659 	int rc = 0;
1660 
1661 	if (!xen_domain())
1662 		return -ENODEV;
1663 
1664 	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1665 		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1666 			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1667 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1668 	}
1669 
1670 	rc = xenvif_xenbus_init();
1671 	if (rc)
1672 		goto failed_init;
1673 
1674 	return 0;
1675 
1676 failed_init:
1677 	return rc;
1678 }
1679 
1680 module_init(netback_init);
1681 
1682 static void __exit netback_fini(void)
1683 {
1684 	xenvif_xenbus_fini();
1685 }
1686 module_exit(netback_fini);
1687 
1688 MODULE_LICENSE("Dual BSD/GPL");
1689 MODULE_ALIAS("xen-backend:vif");
1690