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 #include <net/ip6_checksum.h>
43 
44 #include <xen/xen.h>
45 #include <xen/events.h>
46 #include <xen/interface/memory.h>
47 
48 #include <asm/xen/hypercall.h>
49 #include <asm/xen/page.h>
50 
51 /* Provide an option to disable split event channels at load time as
52  * event channels are limited resource. Split event channels are
53  * enabled by default.
54  */
55 bool separate_tx_rx_irq = 1;
56 module_param(separate_tx_rx_irq, bool, 0644);
57 
58 /*
59  * This is the maximum slots a skb can have. If a guest sends a skb
60  * which exceeds this limit it is considered malicious.
61  */
62 #define FATAL_SKB_SLOTS_DEFAULT 20
63 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
64 module_param(fatal_skb_slots, uint, 0444);
65 
66 /*
67  * To avoid confusion, we define XEN_NETBK_LEGACY_SLOTS_MAX indicating
68  * the maximum slots a valid packet can use. Now this value is defined
69  * to be XEN_NETIF_NR_SLOTS_MIN, which is supposed to be supported by
70  * all backend.
71  */
72 #define XEN_NETBK_LEGACY_SLOTS_MAX XEN_NETIF_NR_SLOTS_MIN
73 
74 /*
75  * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
76  * one or more merged tx requests, otherwise it is the continuation of
77  * previous tx request.
78  */
79 static inline int pending_tx_is_head(struct xenvif *vif, RING_IDX idx)
80 {
81 	return vif->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
82 }
83 
84 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
85 			       u8 status);
86 
87 static void make_tx_response(struct xenvif *vif,
88 			     struct xen_netif_tx_request *txp,
89 			     s8       st);
90 
91 static inline int tx_work_todo(struct xenvif *vif);
92 static inline int rx_work_todo(struct xenvif *vif);
93 
94 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
95 					     u16      id,
96 					     s8       st,
97 					     u16      offset,
98 					     u16      size,
99 					     u16      flags);
100 
101 static inline unsigned long idx_to_pfn(struct xenvif *vif,
102 				       u16 idx)
103 {
104 	return page_to_pfn(vif->mmap_pages[idx]);
105 }
106 
107 static inline unsigned long idx_to_kaddr(struct xenvif *vif,
108 					 u16 idx)
109 {
110 	return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
111 }
112 
113 /* This is a miniumum size for the linear area to avoid lots of
114  * calls to __pskb_pull_tail() as we set up checksum offsets. The
115  * value 128 was chosen as it covers all IPv4 and most likely
116  * IPv6 headers.
117  */
118 #define PKT_PROT_LEN 128
119 
120 static u16 frag_get_pending_idx(skb_frag_t *frag)
121 {
122 	return (u16)frag->page_offset;
123 }
124 
125 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
126 {
127 	frag->page_offset = pending_idx;
128 }
129 
130 static inline pending_ring_idx_t pending_index(unsigned i)
131 {
132 	return i & (MAX_PENDING_REQS-1);
133 }
134 
135 static inline pending_ring_idx_t nr_pending_reqs(struct xenvif *vif)
136 {
137 	return MAX_PENDING_REQS -
138 		vif->pending_prod + vif->pending_cons;
139 }
140 
141 static int max_required_rx_slots(struct xenvif *vif)
142 {
143 	int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
144 
145 	/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
146 	if (vif->can_sg || vif->gso_mask || vif->gso_prefix_mask)
147 		max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
148 
149 	return max;
150 }
151 
152 int xenvif_rx_ring_full(struct xenvif *vif)
153 {
154 	RING_IDX peek   = vif->rx_req_cons_peek;
155 	RING_IDX needed = max_required_rx_slots(vif);
156 
157 	return ((vif->rx.sring->req_prod - peek) < needed) ||
158 	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
159 }
160 
161 int xenvif_must_stop_queue(struct xenvif *vif)
162 {
163 	if (!xenvif_rx_ring_full(vif))
164 		return 0;
165 
166 	vif->rx.sring->req_event = vif->rx_req_cons_peek +
167 		max_required_rx_slots(vif);
168 	mb(); /* request notification /then/ check the queue */
169 
170 	return xenvif_rx_ring_full(vif);
171 }
172 
173 /*
174  * Returns true if we should start a new receive buffer instead of
175  * adding 'size' bytes to a buffer which currently contains 'offset'
176  * bytes.
177  */
178 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
179 {
180 	/* simple case: we have completely filled the current buffer. */
181 	if (offset == MAX_BUFFER_OFFSET)
182 		return true;
183 
184 	/*
185 	 * complex case: start a fresh buffer if the current frag
186 	 * would overflow the current buffer but only if:
187 	 *     (i)   this frag would fit completely in the next buffer
188 	 * and (ii)  there is already some data in the current buffer
189 	 * and (iii) this is not the head buffer.
190 	 *
191 	 * Where:
192 	 * - (i) stops us splitting a frag into two copies
193 	 *   unless the frag is too large for a single buffer.
194 	 * - (ii) stops us from leaving a buffer pointlessly empty.
195 	 * - (iii) stops us leaving the first buffer
196 	 *   empty. Strictly speaking this is already covered
197 	 *   by (ii) but is explicitly checked because
198 	 *   netfront relies on the first buffer being
199 	 *   non-empty and can crash otherwise.
200 	 *
201 	 * This means we will effectively linearise small
202 	 * frags but do not needlessly split large buffers
203 	 * into multiple copies tend to give large frags their
204 	 * own buffers as before.
205 	 */
206 	if ((offset + size > MAX_BUFFER_OFFSET) &&
207 	    (size <= MAX_BUFFER_OFFSET) && offset && !head)
208 		return true;
209 
210 	return false;
211 }
212 
213 struct xenvif_count_slot_state {
214 	unsigned long copy_off;
215 	bool head;
216 };
217 
218 unsigned int xenvif_count_frag_slots(struct xenvif *vif,
219 				     unsigned long offset, unsigned long size,
220 				     struct xenvif_count_slot_state *state)
221 {
222 	unsigned count = 0;
223 
224 	offset &= ~PAGE_MASK;
225 
226 	while (size > 0) {
227 		unsigned long bytes;
228 
229 		bytes = PAGE_SIZE - offset;
230 
231 		if (bytes > size)
232 			bytes = size;
233 
234 		if (start_new_rx_buffer(state->copy_off, bytes, state->head)) {
235 			count++;
236 			state->copy_off = 0;
237 		}
238 
239 		if (state->copy_off + bytes > MAX_BUFFER_OFFSET)
240 			bytes = MAX_BUFFER_OFFSET - state->copy_off;
241 
242 		state->copy_off += bytes;
243 
244 		offset += bytes;
245 		size -= bytes;
246 
247 		if (offset == PAGE_SIZE)
248 			offset = 0;
249 
250 		state->head = false;
251 	}
252 
253 	return count;
254 }
255 
256 /*
257  * Figure out how many ring slots we're going to need to send @skb to
258  * the guest. This function is essentially a dry run of
259  * xenvif_gop_frag_copy.
260  */
261 unsigned int xenvif_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
262 {
263 	struct xenvif_count_slot_state state;
264 	unsigned int count;
265 	unsigned char *data;
266 	unsigned i;
267 
268 	state.head = true;
269 	state.copy_off = 0;
270 
271 	/* Slot for the first (partial) page of data. */
272 	count = 1;
273 
274 	/* Need a slot for the GSO prefix for GSO extra data? */
275 	if (skb_shinfo(skb)->gso_size)
276 		count++;
277 
278 	data = skb->data;
279 	while (data < skb_tail_pointer(skb)) {
280 		unsigned long offset = offset_in_page(data);
281 		unsigned long size = PAGE_SIZE - offset;
282 
283 		if (data + size > skb_tail_pointer(skb))
284 			size = skb_tail_pointer(skb) - data;
285 
286 		count += xenvif_count_frag_slots(vif, offset, size, &state);
287 
288 		data += size;
289 	}
290 
291 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
292 		unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
293 		unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
294 
295 		count += xenvif_count_frag_slots(vif, offset, size, &state);
296 	}
297 	return count;
298 }
299 
300 struct netrx_pending_operations {
301 	unsigned copy_prod, copy_cons;
302 	unsigned meta_prod, meta_cons;
303 	struct gnttab_copy *copy;
304 	struct xenvif_rx_meta *meta;
305 	int copy_off;
306 	grant_ref_t copy_gref;
307 };
308 
309 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
310 						 struct netrx_pending_operations *npo)
311 {
312 	struct xenvif_rx_meta *meta;
313 	struct xen_netif_rx_request *req;
314 
315 	req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
316 
317 	meta = npo->meta + npo->meta_prod++;
318 	meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
319 	meta->gso_size = 0;
320 	meta->size = 0;
321 	meta->id = req->id;
322 
323 	npo->copy_off = 0;
324 	npo->copy_gref = req->gref;
325 
326 	return meta;
327 }
328 
329 /*
330  * Set up the grant operations for this fragment. If it's a flipping
331  * interface, we also set up the unmap request from here.
332  */
333 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
334 				 struct netrx_pending_operations *npo,
335 				 struct page *page, unsigned long size,
336 				 unsigned long offset, int *head)
337 {
338 	struct gnttab_copy *copy_gop;
339 	struct xenvif_rx_meta *meta;
340 	unsigned long bytes;
341 	int gso_type;
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 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
401 			gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
402 		else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
403 			gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
404 		else
405 			gso_type = XEN_NETIF_GSO_TYPE_NONE;
406 
407 		if (*head && ((1 << gso_type) & vif->gso_mask))
408 			vif->rx.req_cons++;
409 
410 		*head = 0; /* There must be something in this buffer now. */
411 
412 	}
413 }
414 
415 /*
416  * Prepare an SKB to be transmitted to the frontend.
417  *
418  * This function is responsible for allocating grant operations, meta
419  * structures, etc.
420  *
421  * It returns the number of meta structures consumed. The number of
422  * ring slots used is always equal to the number of meta slots used
423  * plus the number of GSO descriptors used. Currently, we use either
424  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
425  * frontend-side LRO).
426  */
427 static int xenvif_gop_skb(struct sk_buff *skb,
428 			  struct netrx_pending_operations *npo)
429 {
430 	struct xenvif *vif = netdev_priv(skb->dev);
431 	int nr_frags = skb_shinfo(skb)->nr_frags;
432 	int i;
433 	struct xen_netif_rx_request *req;
434 	struct xenvif_rx_meta *meta;
435 	unsigned char *data;
436 	int head = 1;
437 	int old_meta_prod;
438 	int gso_type;
439 	int gso_size;
440 
441 	old_meta_prod = npo->meta_prod;
442 
443 	if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
444 		gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
445 		gso_size = skb_shinfo(skb)->gso_size;
446 	} else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
447 		gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
448 		gso_size = skb_shinfo(skb)->gso_size;
449 	} else {
450 		gso_type = XEN_NETIF_GSO_TYPE_NONE;
451 		gso_size = 0;
452 	}
453 
454 	/* Set up a GSO prefix descriptor, if necessary */
455 	if ((1 << gso_type) & vif->gso_prefix_mask) {
456 		req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
457 		meta = npo->meta + npo->meta_prod++;
458 		meta->gso_type = gso_type;
459 		meta->gso_size = gso_size;
460 		meta->size = 0;
461 		meta->id = req->id;
462 	}
463 
464 	req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
465 	meta = npo->meta + npo->meta_prod++;
466 
467 	if ((1 << gso_type) & vif->gso_mask) {
468 		meta->gso_type = gso_type;
469 		meta->gso_size = gso_size;
470 	} else {
471 		meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
472 		meta->gso_size = 0;
473 	}
474 
475 	meta->size = 0;
476 	meta->id = req->id;
477 	npo->copy_off = 0;
478 	npo->copy_gref = req->gref;
479 
480 	data = skb->data;
481 	while (data < skb_tail_pointer(skb)) {
482 		unsigned int offset = offset_in_page(data);
483 		unsigned int len = PAGE_SIZE - offset;
484 
485 		if (data + len > skb_tail_pointer(skb))
486 			len = skb_tail_pointer(skb) - data;
487 
488 		xenvif_gop_frag_copy(vif, skb, npo,
489 				     virt_to_page(data), len, offset, &head);
490 		data += len;
491 	}
492 
493 	for (i = 0; i < nr_frags; i++) {
494 		xenvif_gop_frag_copy(vif, skb, npo,
495 				     skb_frag_page(&skb_shinfo(skb)->frags[i]),
496 				     skb_frag_size(&skb_shinfo(skb)->frags[i]),
497 				     skb_shinfo(skb)->frags[i].page_offset,
498 				     &head);
499 	}
500 
501 	return npo->meta_prod - old_meta_prod;
502 }
503 
504 /*
505  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
506  * used to set up the operations on the top of
507  * netrx_pending_operations, which have since been done.  Check that
508  * they didn't give any errors and advance over them.
509  */
510 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
511 			    struct netrx_pending_operations *npo)
512 {
513 	struct gnttab_copy     *copy_op;
514 	int status = XEN_NETIF_RSP_OKAY;
515 	int i;
516 
517 	for (i = 0; i < nr_meta_slots; i++) {
518 		copy_op = npo->copy + npo->copy_cons++;
519 		if (copy_op->status != GNTST_okay) {
520 			netdev_dbg(vif->dev,
521 				   "Bad status %d from copy to DOM%d.\n",
522 				   copy_op->status, vif->domid);
523 			status = XEN_NETIF_RSP_ERROR;
524 		}
525 	}
526 
527 	return status;
528 }
529 
530 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
531 				      struct xenvif_rx_meta *meta,
532 				      int nr_meta_slots)
533 {
534 	int i;
535 	unsigned long offset;
536 
537 	/* No fragments used */
538 	if (nr_meta_slots <= 1)
539 		return;
540 
541 	nr_meta_slots--;
542 
543 	for (i = 0; i < nr_meta_slots; i++) {
544 		int flags;
545 		if (i == nr_meta_slots - 1)
546 			flags = 0;
547 		else
548 			flags = XEN_NETRXF_more_data;
549 
550 		offset = 0;
551 		make_rx_response(vif, meta[i].id, status, offset,
552 				 meta[i].size, flags);
553 	}
554 }
555 
556 struct skb_cb_overlay {
557 	int meta_slots_used;
558 };
559 
560 static void xenvif_kick_thread(struct xenvif *vif)
561 {
562 	wake_up(&vif->wq);
563 }
564 
565 void xenvif_rx_action(struct xenvif *vif)
566 {
567 	s8 status;
568 	u16 flags;
569 	struct xen_netif_rx_response *resp;
570 	struct sk_buff_head rxq;
571 	struct sk_buff *skb;
572 	LIST_HEAD(notify);
573 	int ret;
574 	int nr_frags;
575 	int count;
576 	unsigned long offset;
577 	struct skb_cb_overlay *sco;
578 	int need_to_notify = 0;
579 
580 	struct netrx_pending_operations npo = {
581 		.copy  = vif->grant_copy_op,
582 		.meta  = vif->meta,
583 	};
584 
585 	skb_queue_head_init(&rxq);
586 
587 	count = 0;
588 
589 	while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
590 		vif = netdev_priv(skb->dev);
591 		nr_frags = skb_shinfo(skb)->nr_frags;
592 
593 		sco = (struct skb_cb_overlay *)skb->cb;
594 		sco->meta_slots_used = xenvif_gop_skb(skb, &npo);
595 
596 		count += nr_frags + 1;
597 
598 		__skb_queue_tail(&rxq, skb);
599 
600 		/* Filled the batch queue? */
601 		/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
602 		if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
603 			break;
604 	}
605 
606 	BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
607 
608 	if (!npo.copy_prod)
609 		return;
610 
611 	BUG_ON(npo.copy_prod > ARRAY_SIZE(vif->grant_copy_op));
612 	gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
613 
614 	while ((skb = __skb_dequeue(&rxq)) != NULL) {
615 		sco = (struct skb_cb_overlay *)skb->cb;
616 
617 		vif = netdev_priv(skb->dev);
618 
619 		if ((1 << vif->meta[npo.meta_cons].gso_type) &
620 		    vif->gso_prefix_mask) {
621 			resp = RING_GET_RESPONSE(&vif->rx,
622 						 vif->rx.rsp_prod_pvt++);
623 
624 			resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
625 
626 			resp->offset = vif->meta[npo.meta_cons].gso_size;
627 			resp->id = vif->meta[npo.meta_cons].id;
628 			resp->status = sco->meta_slots_used;
629 
630 			npo.meta_cons++;
631 			sco->meta_slots_used--;
632 		}
633 
634 
635 		vif->dev->stats.tx_bytes += skb->len;
636 		vif->dev->stats.tx_packets++;
637 
638 		status = xenvif_check_gop(vif, sco->meta_slots_used, &npo);
639 
640 		if (sco->meta_slots_used == 1)
641 			flags = 0;
642 		else
643 			flags = XEN_NETRXF_more_data;
644 
645 		if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
646 			flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
647 		else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
648 			/* remote but checksummed. */
649 			flags |= XEN_NETRXF_data_validated;
650 
651 		offset = 0;
652 		resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
653 					status, offset,
654 					vif->meta[npo.meta_cons].size,
655 					flags);
656 
657 		if ((1 << vif->meta[npo.meta_cons].gso_type) &
658 		    vif->gso_mask) {
659 			struct xen_netif_extra_info *gso =
660 				(struct xen_netif_extra_info *)
661 				RING_GET_RESPONSE(&vif->rx,
662 						  vif->rx.rsp_prod_pvt++);
663 
664 			resp->flags |= XEN_NETRXF_extra_info;
665 
666 			gso->u.gso.type = vif->meta[npo.meta_cons].gso_type;
667 			gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
668 			gso->u.gso.pad = 0;
669 			gso->u.gso.features = 0;
670 
671 			gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
672 			gso->flags = 0;
673 		}
674 
675 		xenvif_add_frag_responses(vif, status,
676 					  vif->meta + npo.meta_cons + 1,
677 					  sco->meta_slots_used);
678 
679 		RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
680 
681 		if (ret)
682 			need_to_notify = 1;
683 
684 		xenvif_notify_tx_completion(vif);
685 
686 		npo.meta_cons += sco->meta_slots_used;
687 		dev_kfree_skb(skb);
688 	}
689 
690 	if (need_to_notify)
691 		notify_remote_via_irq(vif->rx_irq);
692 
693 	/* More work to do? */
694 	if (!skb_queue_empty(&vif->rx_queue))
695 		xenvif_kick_thread(vif);
696 }
697 
698 void xenvif_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
699 {
700 	skb_queue_tail(&vif->rx_queue, skb);
701 
702 	xenvif_kick_thread(vif);
703 }
704 
705 void xenvif_check_rx_xenvif(struct xenvif *vif)
706 {
707 	int more_to_do;
708 
709 	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
710 
711 	if (more_to_do)
712 		napi_schedule(&vif->napi);
713 }
714 
715 static void tx_add_credit(struct xenvif *vif)
716 {
717 	unsigned long max_burst, max_credit;
718 
719 	/*
720 	 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
721 	 * Otherwise the interface can seize up due to insufficient credit.
722 	 */
723 	max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
724 	max_burst = min(max_burst, 131072UL);
725 	max_burst = max(max_burst, vif->credit_bytes);
726 
727 	/* Take care that adding a new chunk of credit doesn't wrap to zero. */
728 	max_credit = vif->remaining_credit + vif->credit_bytes;
729 	if (max_credit < vif->remaining_credit)
730 		max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
731 
732 	vif->remaining_credit = min(max_credit, max_burst);
733 }
734 
735 static void tx_credit_callback(unsigned long data)
736 {
737 	struct xenvif *vif = (struct xenvif *)data;
738 	tx_add_credit(vif);
739 	xenvif_check_rx_xenvif(vif);
740 }
741 
742 static void xenvif_tx_err(struct xenvif *vif,
743 			  struct xen_netif_tx_request *txp, RING_IDX end)
744 {
745 	RING_IDX cons = vif->tx.req_cons;
746 
747 	do {
748 		make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
749 		if (cons == end)
750 			break;
751 		txp = RING_GET_REQUEST(&vif->tx, cons++);
752 	} while (1);
753 	vif->tx.req_cons = cons;
754 }
755 
756 static void xenvif_fatal_tx_err(struct xenvif *vif)
757 {
758 	netdev_err(vif->dev, "fatal error; disabling device\n");
759 	xenvif_carrier_off(vif);
760 }
761 
762 static int xenvif_count_requests(struct xenvif *vif,
763 				 struct xen_netif_tx_request *first,
764 				 struct xen_netif_tx_request *txp,
765 				 int work_to_do)
766 {
767 	RING_IDX cons = vif->tx.req_cons;
768 	int slots = 0;
769 	int drop_err = 0;
770 	int more_data;
771 
772 	if (!(first->flags & XEN_NETTXF_more_data))
773 		return 0;
774 
775 	do {
776 		struct xen_netif_tx_request dropped_tx = { 0 };
777 
778 		if (slots >= work_to_do) {
779 			netdev_err(vif->dev,
780 				   "Asked for %d slots but exceeds this limit\n",
781 				   work_to_do);
782 			xenvif_fatal_tx_err(vif);
783 			return -ENODATA;
784 		}
785 
786 		/* This guest is really using too many slots and
787 		 * considered malicious.
788 		 */
789 		if (unlikely(slots >= fatal_skb_slots)) {
790 			netdev_err(vif->dev,
791 				   "Malicious frontend using %d slots, threshold %u\n",
792 				   slots, fatal_skb_slots);
793 			xenvif_fatal_tx_err(vif);
794 			return -E2BIG;
795 		}
796 
797 		/* Xen network protocol had implicit dependency on
798 		 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
799 		 * the historical MAX_SKB_FRAGS value 18 to honor the
800 		 * same behavior as before. Any packet using more than
801 		 * 18 slots but less than fatal_skb_slots slots is
802 		 * dropped
803 		 */
804 		if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
805 			if (net_ratelimit())
806 				netdev_dbg(vif->dev,
807 					   "Too many slots (%d) exceeding limit (%d), dropping packet\n",
808 					   slots, XEN_NETBK_LEGACY_SLOTS_MAX);
809 			drop_err = -E2BIG;
810 		}
811 
812 		if (drop_err)
813 			txp = &dropped_tx;
814 
815 		memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
816 		       sizeof(*txp));
817 
818 		/* If the guest submitted a frame >= 64 KiB then
819 		 * first->size overflowed and following slots will
820 		 * appear to be larger than the frame.
821 		 *
822 		 * This cannot be fatal error as there are buggy
823 		 * frontends that do this.
824 		 *
825 		 * Consume all slots and drop the packet.
826 		 */
827 		if (!drop_err && txp->size > first->size) {
828 			if (net_ratelimit())
829 				netdev_dbg(vif->dev,
830 					   "Invalid tx request, slot size %u > remaining size %u\n",
831 					   txp->size, first->size);
832 			drop_err = -EIO;
833 		}
834 
835 		first->size -= txp->size;
836 		slots++;
837 
838 		if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
839 			netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
840 				 txp->offset, txp->size);
841 			xenvif_fatal_tx_err(vif);
842 			return -EINVAL;
843 		}
844 
845 		more_data = txp->flags & XEN_NETTXF_more_data;
846 
847 		if (!drop_err)
848 			txp++;
849 
850 	} while (more_data);
851 
852 	if (drop_err) {
853 		xenvif_tx_err(vif, first, cons + slots);
854 		return drop_err;
855 	}
856 
857 	return slots;
858 }
859 
860 static struct page *xenvif_alloc_page(struct xenvif *vif,
861 				      u16 pending_idx)
862 {
863 	struct page *page;
864 
865 	page = alloc_page(GFP_ATOMIC|__GFP_COLD);
866 	if (!page)
867 		return NULL;
868 	vif->mmap_pages[pending_idx] = page;
869 
870 	return page;
871 }
872 
873 static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
874 					       struct sk_buff *skb,
875 					       struct xen_netif_tx_request *txp,
876 					       struct gnttab_copy *gop)
877 {
878 	struct skb_shared_info *shinfo = skb_shinfo(skb);
879 	skb_frag_t *frags = shinfo->frags;
880 	u16 pending_idx = *((u16 *)skb->data);
881 	u16 head_idx = 0;
882 	int slot, start;
883 	struct page *page;
884 	pending_ring_idx_t index, start_idx = 0;
885 	uint16_t dst_offset;
886 	unsigned int nr_slots;
887 	struct pending_tx_info *first = NULL;
888 
889 	/* At this point shinfo->nr_frags is in fact the number of
890 	 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
891 	 */
892 	nr_slots = shinfo->nr_frags;
893 
894 	/* Skip first skb fragment if it is on same page as header fragment. */
895 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
896 
897 	/* Coalesce tx requests, at this point the packet passed in
898 	 * should be <= 64K. Any packets larger than 64K have been
899 	 * handled in xenvif_count_requests().
900 	 */
901 	for (shinfo->nr_frags = slot = start; slot < nr_slots;
902 	     shinfo->nr_frags++) {
903 		struct pending_tx_info *pending_tx_info =
904 			vif->pending_tx_info;
905 
906 		page = alloc_page(GFP_ATOMIC|__GFP_COLD);
907 		if (!page)
908 			goto err;
909 
910 		dst_offset = 0;
911 		first = NULL;
912 		while (dst_offset < PAGE_SIZE && slot < nr_slots) {
913 			gop->flags = GNTCOPY_source_gref;
914 
915 			gop->source.u.ref = txp->gref;
916 			gop->source.domid = vif->domid;
917 			gop->source.offset = txp->offset;
918 
919 			gop->dest.domid = DOMID_SELF;
920 
921 			gop->dest.offset = dst_offset;
922 			gop->dest.u.gmfn = virt_to_mfn(page_address(page));
923 
924 			if (dst_offset + txp->size > PAGE_SIZE) {
925 				/* This page can only merge a portion
926 				 * of tx request. Do not increment any
927 				 * pointer / counter here. The txp
928 				 * will be dealt with in future
929 				 * rounds, eventually hitting the
930 				 * `else` branch.
931 				 */
932 				gop->len = PAGE_SIZE - dst_offset;
933 				txp->offset += gop->len;
934 				txp->size -= gop->len;
935 				dst_offset += gop->len; /* quit loop */
936 			} else {
937 				/* This tx request can be merged in the page */
938 				gop->len = txp->size;
939 				dst_offset += gop->len;
940 
941 				index = pending_index(vif->pending_cons++);
942 
943 				pending_idx = vif->pending_ring[index];
944 
945 				memcpy(&pending_tx_info[pending_idx].req, txp,
946 				       sizeof(*txp));
947 
948 				/* Poison these fields, corresponding
949 				 * fields for head tx req will be set
950 				 * to correct values after the loop.
951 				 */
952 				vif->mmap_pages[pending_idx] = (void *)(~0UL);
953 				pending_tx_info[pending_idx].head =
954 					INVALID_PENDING_RING_IDX;
955 
956 				if (!first) {
957 					first = &pending_tx_info[pending_idx];
958 					start_idx = index;
959 					head_idx = pending_idx;
960 				}
961 
962 				txp++;
963 				slot++;
964 			}
965 
966 			gop++;
967 		}
968 
969 		first->req.offset = 0;
970 		first->req.size = dst_offset;
971 		first->head = start_idx;
972 		vif->mmap_pages[head_idx] = page;
973 		frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
974 	}
975 
976 	BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
977 
978 	return gop;
979 err:
980 	/* Unwind, freeing all pages and sending error responses. */
981 	while (shinfo->nr_frags-- > start) {
982 		xenvif_idx_release(vif,
983 				frag_get_pending_idx(&frags[shinfo->nr_frags]),
984 				XEN_NETIF_RSP_ERROR);
985 	}
986 	/* The head too, if necessary. */
987 	if (start)
988 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
989 
990 	return NULL;
991 }
992 
993 static int xenvif_tx_check_gop(struct xenvif *vif,
994 			       struct sk_buff *skb,
995 			       struct gnttab_copy **gopp)
996 {
997 	struct gnttab_copy *gop = *gopp;
998 	u16 pending_idx = *((u16 *)skb->data);
999 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1000 	struct pending_tx_info *tx_info;
1001 	int nr_frags = shinfo->nr_frags;
1002 	int i, err, start;
1003 	u16 peek; /* peek into next tx request */
1004 
1005 	/* Check status of header. */
1006 	err = gop->status;
1007 	if (unlikely(err))
1008 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
1009 
1010 	/* Skip first skb fragment if it is on same page as header fragment. */
1011 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1012 
1013 	for (i = start; i < nr_frags; i++) {
1014 		int j, newerr;
1015 		pending_ring_idx_t head;
1016 
1017 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1018 		tx_info = &vif->pending_tx_info[pending_idx];
1019 		head = tx_info->head;
1020 
1021 		/* Check error status: if okay then remember grant handle. */
1022 		do {
1023 			newerr = (++gop)->status;
1024 			if (newerr)
1025 				break;
1026 			peek = vif->pending_ring[pending_index(++head)];
1027 		} while (!pending_tx_is_head(vif, peek));
1028 
1029 		if (likely(!newerr)) {
1030 			/* Had a previous error? Invalidate this fragment. */
1031 			if (unlikely(err))
1032 				xenvif_idx_release(vif, pending_idx,
1033 						   XEN_NETIF_RSP_OKAY);
1034 			continue;
1035 		}
1036 
1037 		/* Error on this fragment: respond to client with an error. */
1038 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
1039 
1040 		/* Not the first error? Preceding frags already invalidated. */
1041 		if (err)
1042 			continue;
1043 
1044 		/* First error: invalidate header and preceding fragments. */
1045 		pending_idx = *((u16 *)skb->data);
1046 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1047 		for (j = start; j < i; j++) {
1048 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1049 			xenvif_idx_release(vif, pending_idx,
1050 					   XEN_NETIF_RSP_OKAY);
1051 		}
1052 
1053 		/* Remember the error: invalidate all subsequent fragments. */
1054 		err = newerr;
1055 	}
1056 
1057 	*gopp = gop + 1;
1058 	return err;
1059 }
1060 
1061 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
1062 {
1063 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1064 	int nr_frags = shinfo->nr_frags;
1065 	int i;
1066 
1067 	for (i = 0; i < nr_frags; i++) {
1068 		skb_frag_t *frag = shinfo->frags + i;
1069 		struct xen_netif_tx_request *txp;
1070 		struct page *page;
1071 		u16 pending_idx;
1072 
1073 		pending_idx = frag_get_pending_idx(frag);
1074 
1075 		txp = &vif->pending_tx_info[pending_idx].req;
1076 		page = virt_to_page(idx_to_kaddr(vif, pending_idx));
1077 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1078 		skb->len += txp->size;
1079 		skb->data_len += txp->size;
1080 		skb->truesize += txp->size;
1081 
1082 		/* Take an extra reference to offset xenvif_idx_release */
1083 		get_page(vif->mmap_pages[pending_idx]);
1084 		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1085 	}
1086 }
1087 
1088 static int xenvif_get_extras(struct xenvif *vif,
1089 				struct xen_netif_extra_info *extras,
1090 				int work_to_do)
1091 {
1092 	struct xen_netif_extra_info extra;
1093 	RING_IDX cons = vif->tx.req_cons;
1094 
1095 	do {
1096 		if (unlikely(work_to_do-- <= 0)) {
1097 			netdev_err(vif->dev, "Missing extra info\n");
1098 			xenvif_fatal_tx_err(vif);
1099 			return -EBADR;
1100 		}
1101 
1102 		memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1103 		       sizeof(extra));
1104 		if (unlikely(!extra.type ||
1105 			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1106 			vif->tx.req_cons = ++cons;
1107 			netdev_err(vif->dev,
1108 				   "Invalid extra type: %d\n", extra.type);
1109 			xenvif_fatal_tx_err(vif);
1110 			return -EINVAL;
1111 		}
1112 
1113 		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1114 		vif->tx.req_cons = ++cons;
1115 	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1116 
1117 	return work_to_do;
1118 }
1119 
1120 static int xenvif_set_skb_gso(struct xenvif *vif,
1121 			      struct sk_buff *skb,
1122 			      struct xen_netif_extra_info *gso)
1123 {
1124 	if (!gso->u.gso.size) {
1125 		netdev_err(vif->dev, "GSO size must not be zero.\n");
1126 		xenvif_fatal_tx_err(vif);
1127 		return -EINVAL;
1128 	}
1129 
1130 	switch (gso->u.gso.type) {
1131 	case XEN_NETIF_GSO_TYPE_TCPV4:
1132 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1133 		break;
1134 	case XEN_NETIF_GSO_TYPE_TCPV6:
1135 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1136 		break;
1137 	default:
1138 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1139 		xenvif_fatal_tx_err(vif);
1140 		return -EINVAL;
1141 	}
1142 
1143 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
1144 
1145 	/* Header must be checked, and gso_segs computed. */
1146 	skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1147 	skb_shinfo(skb)->gso_segs = 0;
1148 
1149 	return 0;
1150 }
1151 
1152 static inline int maybe_pull_tail(struct sk_buff *skb, unsigned int len,
1153 				  unsigned int max)
1154 {
1155 	if (skb_headlen(skb) >= len)
1156 		return 0;
1157 
1158 	/* If we need to pullup then pullup to the max, so we
1159 	 * won't need to do it again.
1160 	 */
1161 	if (max > skb->len)
1162 		max = skb->len;
1163 
1164 	if (__pskb_pull_tail(skb, max - skb_headlen(skb)) == NULL)
1165 		return -ENOMEM;
1166 
1167 	if (skb_headlen(skb) < len)
1168 		return -EPROTO;
1169 
1170 	return 0;
1171 }
1172 
1173 /* This value should be large enough to cover a tagged ethernet header plus
1174  * maximally sized IP and TCP or UDP headers.
1175  */
1176 #define MAX_IP_HDR_LEN 128
1177 
1178 static int checksum_setup_ip(struct xenvif *vif, struct sk_buff *skb,
1179 			     int recalculate_partial_csum)
1180 {
1181 	unsigned int off;
1182 	bool fragment;
1183 	int err;
1184 
1185 	fragment = false;
1186 
1187 	err = maybe_pull_tail(skb,
1188 			      sizeof(struct iphdr),
1189 			      MAX_IP_HDR_LEN);
1190 	if (err < 0)
1191 		goto out;
1192 
1193 	if (ip_hdr(skb)->frag_off & htons(IP_OFFSET | IP_MF))
1194 		fragment = true;
1195 
1196 	off = ip_hdrlen(skb);
1197 
1198 	err = -EPROTO;
1199 
1200 	switch (ip_hdr(skb)->protocol) {
1201 	case IPPROTO_TCP:
1202 		err = maybe_pull_tail(skb,
1203 				      off + sizeof(struct tcphdr),
1204 				      MAX_IP_HDR_LEN);
1205 		if (err < 0)
1206 			goto out;
1207 
1208 		if (!skb_partial_csum_set(skb, off,
1209 					  offsetof(struct tcphdr, check)))
1210 			goto out;
1211 
1212 		if (recalculate_partial_csum)
1213 			tcp_hdr(skb)->check =
1214 				~csum_tcpudp_magic(ip_hdr(skb)->saddr,
1215 						   ip_hdr(skb)->daddr,
1216 						   skb->len - off,
1217 						   IPPROTO_TCP, 0);
1218 		break;
1219 	case IPPROTO_UDP:
1220 		err = maybe_pull_tail(skb,
1221 				      off + sizeof(struct udphdr),
1222 				      MAX_IP_HDR_LEN);
1223 		if (err < 0)
1224 			goto out;
1225 
1226 		if (!skb_partial_csum_set(skb, off,
1227 					  offsetof(struct udphdr, check)))
1228 			goto out;
1229 
1230 		if (recalculate_partial_csum)
1231 			udp_hdr(skb)->check =
1232 				~csum_tcpudp_magic(ip_hdr(skb)->saddr,
1233 						   ip_hdr(skb)->daddr,
1234 						   skb->len - off,
1235 						   IPPROTO_UDP, 0);
1236 		break;
1237 	default:
1238 		goto out;
1239 	}
1240 
1241 	err = 0;
1242 
1243 out:
1244 	return err;
1245 }
1246 
1247 /* This value should be large enough to cover a tagged ethernet header plus
1248  * an IPv6 header, all options, and a maximal TCP or UDP header.
1249  */
1250 #define MAX_IPV6_HDR_LEN 256
1251 
1252 #define OPT_HDR(type, skb, off) \
1253 	(type *)(skb_network_header(skb) + (off))
1254 
1255 static int checksum_setup_ipv6(struct xenvif *vif, struct sk_buff *skb,
1256 			       int recalculate_partial_csum)
1257 {
1258 	int err;
1259 	u8 nexthdr;
1260 	unsigned int off;
1261 	unsigned int len;
1262 	bool fragment;
1263 	bool done;
1264 
1265 	fragment = false;
1266 	done = false;
1267 
1268 	off = sizeof(struct ipv6hdr);
1269 
1270 	err = maybe_pull_tail(skb, off, MAX_IPV6_HDR_LEN);
1271 	if (err < 0)
1272 		goto out;
1273 
1274 	nexthdr = ipv6_hdr(skb)->nexthdr;
1275 
1276 	len = sizeof(struct ipv6hdr) + ntohs(ipv6_hdr(skb)->payload_len);
1277 	while (off <= len && !done) {
1278 		switch (nexthdr) {
1279 		case IPPROTO_DSTOPTS:
1280 		case IPPROTO_HOPOPTS:
1281 		case IPPROTO_ROUTING: {
1282 			struct ipv6_opt_hdr *hp;
1283 
1284 			err = maybe_pull_tail(skb,
1285 					      off +
1286 					      sizeof(struct ipv6_opt_hdr),
1287 					      MAX_IPV6_HDR_LEN);
1288 			if (err < 0)
1289 				goto out;
1290 
1291 			hp = OPT_HDR(struct ipv6_opt_hdr, skb, off);
1292 			nexthdr = hp->nexthdr;
1293 			off += ipv6_optlen(hp);
1294 			break;
1295 		}
1296 		case IPPROTO_AH: {
1297 			struct ip_auth_hdr *hp;
1298 
1299 			err = maybe_pull_tail(skb,
1300 					      off +
1301 					      sizeof(struct ip_auth_hdr),
1302 					      MAX_IPV6_HDR_LEN);
1303 			if (err < 0)
1304 				goto out;
1305 
1306 			hp = OPT_HDR(struct ip_auth_hdr, skb, off);
1307 			nexthdr = hp->nexthdr;
1308 			off += ipv6_authlen(hp);
1309 			break;
1310 		}
1311 		case IPPROTO_FRAGMENT: {
1312 			struct frag_hdr *hp;
1313 
1314 			err = maybe_pull_tail(skb,
1315 					      off +
1316 					      sizeof(struct frag_hdr),
1317 					      MAX_IPV6_HDR_LEN);
1318 			if (err < 0)
1319 				goto out;
1320 
1321 			hp = OPT_HDR(struct frag_hdr, skb, off);
1322 
1323 			if (hp->frag_off & htons(IP6_OFFSET | IP6_MF))
1324 				fragment = true;
1325 
1326 			nexthdr = hp->nexthdr;
1327 			off += sizeof(struct frag_hdr);
1328 			break;
1329 		}
1330 		default:
1331 			done = true;
1332 			break;
1333 		}
1334 	}
1335 
1336 	err = -EPROTO;
1337 
1338 	if (!done || fragment)
1339 		goto out;
1340 
1341 	switch (nexthdr) {
1342 	case IPPROTO_TCP:
1343 		err = maybe_pull_tail(skb,
1344 				      off + sizeof(struct tcphdr),
1345 				      MAX_IPV6_HDR_LEN);
1346 		if (err < 0)
1347 			goto out;
1348 
1349 		if (!skb_partial_csum_set(skb, off,
1350 					  offsetof(struct tcphdr, check)))
1351 			goto out;
1352 
1353 		if (recalculate_partial_csum)
1354 			tcp_hdr(skb)->check =
1355 				~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
1356 						 &ipv6_hdr(skb)->daddr,
1357 						 skb->len - off,
1358 						 IPPROTO_TCP, 0);
1359 		break;
1360 	case IPPROTO_UDP:
1361 		err = maybe_pull_tail(skb,
1362 				      off + sizeof(struct udphdr),
1363 				      MAX_IPV6_HDR_LEN);
1364 		if (err < 0)
1365 			goto out;
1366 
1367 		if (!skb_partial_csum_set(skb, off,
1368 					  offsetof(struct udphdr, check)))
1369 			goto out;
1370 
1371 		if (recalculate_partial_csum)
1372 			udp_hdr(skb)->check =
1373 				~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
1374 						 &ipv6_hdr(skb)->daddr,
1375 						 skb->len - off,
1376 						 IPPROTO_UDP, 0);
1377 		break;
1378 	default:
1379 		goto out;
1380 	}
1381 
1382 	err = 0;
1383 
1384 out:
1385 	return err;
1386 }
1387 
1388 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1389 {
1390 	int err = -EPROTO;
1391 	int recalculate_partial_csum = 0;
1392 
1393 	/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1394 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
1395 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1396 	 * recalculate the partial checksum.
1397 	 */
1398 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1399 		vif->rx_gso_checksum_fixup++;
1400 		skb->ip_summed = CHECKSUM_PARTIAL;
1401 		recalculate_partial_csum = 1;
1402 	}
1403 
1404 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1405 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1406 		return 0;
1407 
1408 	if (skb->protocol == htons(ETH_P_IP))
1409 		err = checksum_setup_ip(vif, skb, recalculate_partial_csum);
1410 	else if (skb->protocol == htons(ETH_P_IPV6))
1411 		err = checksum_setup_ipv6(vif, skb, recalculate_partial_csum);
1412 
1413 	return err;
1414 }
1415 
1416 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1417 {
1418 	u64 now = get_jiffies_64();
1419 	u64 next_credit = vif->credit_window_start +
1420 		msecs_to_jiffies(vif->credit_usec / 1000);
1421 
1422 	/* Timer could already be pending in rare cases. */
1423 	if (timer_pending(&vif->credit_timeout))
1424 		return true;
1425 
1426 	/* Passed the point where we can replenish credit? */
1427 	if (time_after_eq64(now, next_credit)) {
1428 		vif->credit_window_start = now;
1429 		tx_add_credit(vif);
1430 	}
1431 
1432 	/* Still too big to send right now? Set a callback. */
1433 	if (size > vif->remaining_credit) {
1434 		vif->credit_timeout.data     =
1435 			(unsigned long)vif;
1436 		vif->credit_timeout.function =
1437 			tx_credit_callback;
1438 		mod_timer(&vif->credit_timeout,
1439 			  next_credit);
1440 		vif->credit_window_start = next_credit;
1441 
1442 		return true;
1443 	}
1444 
1445 	return false;
1446 }
1447 
1448 static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
1449 {
1450 	struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
1451 	struct sk_buff *skb;
1452 	int ret;
1453 
1454 	while ((nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1455 		< MAX_PENDING_REQS) &&
1456 	       (skb_queue_len(&vif->tx_queue) < budget)) {
1457 		struct xen_netif_tx_request txreq;
1458 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1459 		struct page *page;
1460 		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1461 		u16 pending_idx;
1462 		RING_IDX idx;
1463 		int work_to_do;
1464 		unsigned int data_len;
1465 		pending_ring_idx_t index;
1466 
1467 		if (vif->tx.sring->req_prod - vif->tx.req_cons >
1468 		    XEN_NETIF_TX_RING_SIZE) {
1469 			netdev_err(vif->dev,
1470 				   "Impossible number of requests. "
1471 				   "req_prod %d, req_cons %d, size %ld\n",
1472 				   vif->tx.sring->req_prod, vif->tx.req_cons,
1473 				   XEN_NETIF_TX_RING_SIZE);
1474 			xenvif_fatal_tx_err(vif);
1475 			continue;
1476 		}
1477 
1478 		work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
1479 		if (!work_to_do)
1480 			break;
1481 
1482 		idx = vif->tx.req_cons;
1483 		rmb(); /* Ensure that we see the request before we copy it. */
1484 		memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1485 
1486 		/* Credit-based scheduling. */
1487 		if (txreq.size > vif->remaining_credit &&
1488 		    tx_credit_exceeded(vif, txreq.size))
1489 			break;
1490 
1491 		vif->remaining_credit -= txreq.size;
1492 
1493 		work_to_do--;
1494 		vif->tx.req_cons = ++idx;
1495 
1496 		memset(extras, 0, sizeof(extras));
1497 		if (txreq.flags & XEN_NETTXF_extra_info) {
1498 			work_to_do = xenvif_get_extras(vif, extras,
1499 						       work_to_do);
1500 			idx = vif->tx.req_cons;
1501 			if (unlikely(work_to_do < 0))
1502 				break;
1503 		}
1504 
1505 		ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1506 		if (unlikely(ret < 0))
1507 			break;
1508 
1509 		idx += ret;
1510 
1511 		if (unlikely(txreq.size < ETH_HLEN)) {
1512 			netdev_dbg(vif->dev,
1513 				   "Bad packet size: %d\n", txreq.size);
1514 			xenvif_tx_err(vif, &txreq, idx);
1515 			break;
1516 		}
1517 
1518 		/* No crossing a page as the payload mustn't fragment. */
1519 		if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1520 			netdev_err(vif->dev,
1521 				   "txreq.offset: %x, size: %u, end: %lu\n",
1522 				   txreq.offset, txreq.size,
1523 				   (txreq.offset&~PAGE_MASK) + txreq.size);
1524 			xenvif_fatal_tx_err(vif);
1525 			break;
1526 		}
1527 
1528 		index = pending_index(vif->pending_cons);
1529 		pending_idx = vif->pending_ring[index];
1530 
1531 		data_len = (txreq.size > PKT_PROT_LEN &&
1532 			    ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1533 			PKT_PROT_LEN : txreq.size;
1534 
1535 		skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1536 				GFP_ATOMIC | __GFP_NOWARN);
1537 		if (unlikely(skb == NULL)) {
1538 			netdev_dbg(vif->dev,
1539 				   "Can't allocate a skb in start_xmit.\n");
1540 			xenvif_tx_err(vif, &txreq, idx);
1541 			break;
1542 		}
1543 
1544 		/* Packets passed to netif_rx() must have some headroom. */
1545 		skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1546 
1547 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1548 			struct xen_netif_extra_info *gso;
1549 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1550 
1551 			if (xenvif_set_skb_gso(vif, skb, gso)) {
1552 				/* Failure in xenvif_set_skb_gso is fatal. */
1553 				kfree_skb(skb);
1554 				break;
1555 			}
1556 		}
1557 
1558 		/* XXX could copy straight to head */
1559 		page = xenvif_alloc_page(vif, pending_idx);
1560 		if (!page) {
1561 			kfree_skb(skb);
1562 			xenvif_tx_err(vif, &txreq, idx);
1563 			break;
1564 		}
1565 
1566 		gop->source.u.ref = txreq.gref;
1567 		gop->source.domid = vif->domid;
1568 		gop->source.offset = txreq.offset;
1569 
1570 		gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1571 		gop->dest.domid = DOMID_SELF;
1572 		gop->dest.offset = txreq.offset;
1573 
1574 		gop->len = txreq.size;
1575 		gop->flags = GNTCOPY_source_gref;
1576 
1577 		gop++;
1578 
1579 		memcpy(&vif->pending_tx_info[pending_idx].req,
1580 		       &txreq, sizeof(txreq));
1581 		vif->pending_tx_info[pending_idx].head = index;
1582 		*((u16 *)skb->data) = pending_idx;
1583 
1584 		__skb_put(skb, data_len);
1585 
1586 		skb_shinfo(skb)->nr_frags = ret;
1587 		if (data_len < txreq.size) {
1588 			skb_shinfo(skb)->nr_frags++;
1589 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1590 					     pending_idx);
1591 		} else {
1592 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1593 					     INVALID_PENDING_IDX);
1594 		}
1595 
1596 		vif->pending_cons++;
1597 
1598 		request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1599 		if (request_gop == NULL) {
1600 			kfree_skb(skb);
1601 			xenvif_tx_err(vif, &txreq, idx);
1602 			break;
1603 		}
1604 		gop = request_gop;
1605 
1606 		__skb_queue_tail(&vif->tx_queue, skb);
1607 
1608 		vif->tx.req_cons = idx;
1609 
1610 		if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
1611 			break;
1612 	}
1613 
1614 	return gop - vif->tx_copy_ops;
1615 }
1616 
1617 
1618 static int xenvif_tx_submit(struct xenvif *vif)
1619 {
1620 	struct gnttab_copy *gop = vif->tx_copy_ops;
1621 	struct sk_buff *skb;
1622 	int work_done = 0;
1623 
1624 	while ((skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1625 		struct xen_netif_tx_request *txp;
1626 		u16 pending_idx;
1627 		unsigned data_len;
1628 
1629 		pending_idx = *((u16 *)skb->data);
1630 		txp = &vif->pending_tx_info[pending_idx].req;
1631 
1632 		/* Check the remap error code. */
1633 		if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1634 			netdev_dbg(vif->dev, "netback grant failed.\n");
1635 			skb_shinfo(skb)->nr_frags = 0;
1636 			kfree_skb(skb);
1637 			continue;
1638 		}
1639 
1640 		data_len = skb->len;
1641 		memcpy(skb->data,
1642 		       (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1643 		       data_len);
1644 		if (data_len < txp->size) {
1645 			/* Append the packet payload as a fragment. */
1646 			txp->offset += data_len;
1647 			txp->size -= data_len;
1648 		} else {
1649 			/* Schedule a response immediately. */
1650 			xenvif_idx_release(vif, pending_idx,
1651 					   XEN_NETIF_RSP_OKAY);
1652 		}
1653 
1654 		if (txp->flags & XEN_NETTXF_csum_blank)
1655 			skb->ip_summed = CHECKSUM_PARTIAL;
1656 		else if (txp->flags & XEN_NETTXF_data_validated)
1657 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1658 
1659 		xenvif_fill_frags(vif, skb);
1660 
1661 		if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1662 			int target = min_t(int, skb->len, PKT_PROT_LEN);
1663 			__pskb_pull_tail(skb, target - skb_headlen(skb));
1664 		}
1665 
1666 		skb->dev      = vif->dev;
1667 		skb->protocol = eth_type_trans(skb, skb->dev);
1668 		skb_reset_network_header(skb);
1669 
1670 		if (checksum_setup(vif, skb)) {
1671 			netdev_dbg(vif->dev,
1672 				   "Can't setup checksum in net_tx_action\n");
1673 			kfree_skb(skb);
1674 			continue;
1675 		}
1676 
1677 		skb_probe_transport_header(skb, 0);
1678 
1679 		vif->dev->stats.rx_bytes += skb->len;
1680 		vif->dev->stats.rx_packets++;
1681 
1682 		work_done++;
1683 
1684 		netif_receive_skb(skb);
1685 	}
1686 
1687 	return work_done;
1688 }
1689 
1690 /* Called after netfront has transmitted */
1691 int xenvif_tx_action(struct xenvif *vif, int budget)
1692 {
1693 	unsigned nr_gops;
1694 	int work_done;
1695 
1696 	if (unlikely(!tx_work_todo(vif)))
1697 		return 0;
1698 
1699 	nr_gops = xenvif_tx_build_gops(vif, budget);
1700 
1701 	if (nr_gops == 0)
1702 		return 0;
1703 
1704 	gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
1705 
1706 	work_done = xenvif_tx_submit(vif);
1707 
1708 	return work_done;
1709 }
1710 
1711 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1712 			       u8 status)
1713 {
1714 	struct pending_tx_info *pending_tx_info;
1715 	pending_ring_idx_t head;
1716 	u16 peek; /* peek into next tx request */
1717 
1718 	BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
1719 
1720 	/* Already complete? */
1721 	if (vif->mmap_pages[pending_idx] == NULL)
1722 		return;
1723 
1724 	pending_tx_info = &vif->pending_tx_info[pending_idx];
1725 
1726 	head = pending_tx_info->head;
1727 
1728 	BUG_ON(!pending_tx_is_head(vif, head));
1729 	BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
1730 
1731 	do {
1732 		pending_ring_idx_t index;
1733 		pending_ring_idx_t idx = pending_index(head);
1734 		u16 info_idx = vif->pending_ring[idx];
1735 
1736 		pending_tx_info = &vif->pending_tx_info[info_idx];
1737 		make_tx_response(vif, &pending_tx_info->req, status);
1738 
1739 		/* Setting any number other than
1740 		 * INVALID_PENDING_RING_IDX indicates this slot is
1741 		 * starting a new packet / ending a previous packet.
1742 		 */
1743 		pending_tx_info->head = 0;
1744 
1745 		index = pending_index(vif->pending_prod++);
1746 		vif->pending_ring[index] = vif->pending_ring[info_idx];
1747 
1748 		peek = vif->pending_ring[pending_index(++head)];
1749 
1750 	} while (!pending_tx_is_head(vif, peek));
1751 
1752 	put_page(vif->mmap_pages[pending_idx]);
1753 	vif->mmap_pages[pending_idx] = NULL;
1754 }
1755 
1756 
1757 static void make_tx_response(struct xenvif *vif,
1758 			     struct xen_netif_tx_request *txp,
1759 			     s8       st)
1760 {
1761 	RING_IDX i = vif->tx.rsp_prod_pvt;
1762 	struct xen_netif_tx_response *resp;
1763 	int notify;
1764 
1765 	resp = RING_GET_RESPONSE(&vif->tx, i);
1766 	resp->id     = txp->id;
1767 	resp->status = st;
1768 
1769 	if (txp->flags & XEN_NETTXF_extra_info)
1770 		RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1771 
1772 	vif->tx.rsp_prod_pvt = ++i;
1773 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1774 	if (notify)
1775 		notify_remote_via_irq(vif->tx_irq);
1776 }
1777 
1778 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1779 					     u16      id,
1780 					     s8       st,
1781 					     u16      offset,
1782 					     u16      size,
1783 					     u16      flags)
1784 {
1785 	RING_IDX i = vif->rx.rsp_prod_pvt;
1786 	struct xen_netif_rx_response *resp;
1787 
1788 	resp = RING_GET_RESPONSE(&vif->rx, i);
1789 	resp->offset     = offset;
1790 	resp->flags      = flags;
1791 	resp->id         = id;
1792 	resp->status     = (s16)size;
1793 	if (st < 0)
1794 		resp->status = (s16)st;
1795 
1796 	vif->rx.rsp_prod_pvt = ++i;
1797 
1798 	return resp;
1799 }
1800 
1801 static inline int rx_work_todo(struct xenvif *vif)
1802 {
1803 	return !skb_queue_empty(&vif->rx_queue);
1804 }
1805 
1806 static inline int tx_work_todo(struct xenvif *vif)
1807 {
1808 
1809 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1810 	    (nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1811 	     < MAX_PENDING_REQS))
1812 		return 1;
1813 
1814 	return 0;
1815 }
1816 
1817 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1818 {
1819 	if (vif->tx.sring)
1820 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1821 					vif->tx.sring);
1822 	if (vif->rx.sring)
1823 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1824 					vif->rx.sring);
1825 }
1826 
1827 int xenvif_map_frontend_rings(struct xenvif *vif,
1828 			      grant_ref_t tx_ring_ref,
1829 			      grant_ref_t rx_ring_ref)
1830 {
1831 	void *addr;
1832 	struct xen_netif_tx_sring *txs;
1833 	struct xen_netif_rx_sring *rxs;
1834 
1835 	int err = -ENOMEM;
1836 
1837 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1838 				     tx_ring_ref, &addr);
1839 	if (err)
1840 		goto err;
1841 
1842 	txs = (struct xen_netif_tx_sring *)addr;
1843 	BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1844 
1845 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1846 				     rx_ring_ref, &addr);
1847 	if (err)
1848 		goto err;
1849 
1850 	rxs = (struct xen_netif_rx_sring *)addr;
1851 	BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1852 
1853 	vif->rx_req_cons_peek = 0;
1854 
1855 	return 0;
1856 
1857 err:
1858 	xenvif_unmap_frontend_rings(vif);
1859 	return err;
1860 }
1861 
1862 int xenvif_kthread(void *data)
1863 {
1864 	struct xenvif *vif = data;
1865 
1866 	while (!kthread_should_stop()) {
1867 		wait_event_interruptible(vif->wq,
1868 					 rx_work_todo(vif) ||
1869 					 kthread_should_stop());
1870 		if (kthread_should_stop())
1871 			break;
1872 
1873 		if (rx_work_todo(vif))
1874 			xenvif_rx_action(vif);
1875 
1876 		cond_resched();
1877 	}
1878 
1879 	return 0;
1880 }
1881 
1882 static int __init netback_init(void)
1883 {
1884 	int rc = 0;
1885 
1886 	if (!xen_domain())
1887 		return -ENODEV;
1888 
1889 	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1890 		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1891 			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1892 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1893 	}
1894 
1895 	rc = xenvif_xenbus_init();
1896 	if (rc)
1897 		goto failed_init;
1898 
1899 	return 0;
1900 
1901 failed_init:
1902 	return rc;
1903 }
1904 
1905 module_init(netback_init);
1906 
1907 static void __exit netback_fini(void)
1908 {
1909 	xenvif_xenbus_fini();
1910 }
1911 module_exit(netback_fini);
1912 
1913 MODULE_LICENSE("Dual BSD/GPL");
1914 MODULE_ALIAS("xen-backend:vif");
1915