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 #include <linux/highmem.h>
41 
42 #include <net/tcp.h>
43 
44 #include <xen/xen.h>
45 #include <xen/events.h>
46 #include <xen/interface/memory.h>
47 #include <xen/page.h>
48 
49 #include <asm/xen/hypercall.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 = true;
56 module_param(separate_tx_rx_irq, bool, 0644);
57 
58 /* The time that packets can stay on the guest Rx internal queue
59  * before they are dropped.
60  */
61 unsigned int rx_drain_timeout_msecs = 10000;
62 module_param(rx_drain_timeout_msecs, uint, 0444);
63 
64 /* The length of time before the frontend is considered unresponsive
65  * because it isn't providing Rx slots.
66  */
67 unsigned int rx_stall_timeout_msecs = 60000;
68 module_param(rx_stall_timeout_msecs, uint, 0444);
69 
70 #define MAX_QUEUES_DEFAULT 8
71 unsigned int xenvif_max_queues;
72 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
73 MODULE_PARM_DESC(max_queues,
74 		 "Maximum number of queues per virtual interface");
75 
76 /*
77  * This is the maximum slots a skb can have. If a guest sends a skb
78  * which exceeds this limit it is considered malicious.
79  */
80 #define FATAL_SKB_SLOTS_DEFAULT 20
81 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
82 module_param(fatal_skb_slots, uint, 0444);
83 
84 /* The amount to copy out of the first guest Tx slot into the skb's
85  * linear area.  If the first slot has more data, it will be mapped
86  * and put into the first frag.
87  *
88  * This is sized to avoid pulling headers from the frags for most
89  * TCP/IP packets.
90  */
91 #define XEN_NETBACK_TX_COPY_LEN 128
92 
93 /* This is the maximum number of flows in the hash cache. */
94 #define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
95 unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
96 module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
97 MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
98 
99 /* The module parameter tells that we have to put data
100  * for xen-netfront with the XDP_PACKET_HEADROOM offset
101  * needed for XDP processing
102  */
103 bool provides_xdp_headroom = true;
104 module_param(provides_xdp_headroom, bool, 0644);
105 
106 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
107 			       u8 status);
108 
109 static void make_tx_response(struct xenvif_queue *queue,
110 			     struct xen_netif_tx_request *txp,
111 			     unsigned int extra_count,
112 			     s8       st);
113 static void push_tx_responses(struct xenvif_queue *queue);
114 
115 static inline int tx_work_todo(struct xenvif_queue *queue);
116 
117 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
118 				       u16 idx)
119 {
120 	return page_to_pfn(queue->mmap_pages[idx]);
121 }
122 
123 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
124 					 u16 idx)
125 {
126 	return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
127 }
128 
129 #define callback_param(vif, pending_idx) \
130 	(vif->pending_tx_info[pending_idx].callback_struct)
131 
132 /* Find the containing VIF's structure from a pointer in pending_tx_info array
133  */
134 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
135 {
136 	u16 pending_idx = ubuf->desc;
137 	struct pending_tx_info *temp =
138 		container_of(ubuf, struct pending_tx_info, callback_struct);
139 	return container_of(temp - pending_idx,
140 			    struct xenvif_queue,
141 			    pending_tx_info[0]);
142 }
143 
144 static u16 frag_get_pending_idx(skb_frag_t *frag)
145 {
146 	return (u16)skb_frag_off(frag);
147 }
148 
149 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
150 {
151 	skb_frag_off_set(frag, pending_idx);
152 }
153 
154 static inline pending_ring_idx_t pending_index(unsigned i)
155 {
156 	return i & (MAX_PENDING_REQS-1);
157 }
158 
159 void xenvif_kick_thread(struct xenvif_queue *queue)
160 {
161 	wake_up(&queue->wq);
162 }
163 
164 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
165 {
166 	int more_to_do;
167 
168 	RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
169 
170 	if (more_to_do)
171 		napi_schedule(&queue->napi);
172 }
173 
174 static void tx_add_credit(struct xenvif_queue *queue)
175 {
176 	unsigned long max_burst, max_credit;
177 
178 	/*
179 	 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
180 	 * Otherwise the interface can seize up due to insufficient credit.
181 	 */
182 	max_burst = max(131072UL, queue->credit_bytes);
183 
184 	/* Take care that adding a new chunk of credit doesn't wrap to zero. */
185 	max_credit = queue->remaining_credit + queue->credit_bytes;
186 	if (max_credit < queue->remaining_credit)
187 		max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
188 
189 	queue->remaining_credit = min(max_credit, max_burst);
190 	queue->rate_limited = false;
191 }
192 
193 void xenvif_tx_credit_callback(struct timer_list *t)
194 {
195 	struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
196 	tx_add_credit(queue);
197 	xenvif_napi_schedule_or_enable_events(queue);
198 }
199 
200 static void xenvif_tx_err(struct xenvif_queue *queue,
201 			  struct xen_netif_tx_request *txp,
202 			  unsigned int extra_count, RING_IDX end)
203 {
204 	RING_IDX cons = queue->tx.req_cons;
205 	unsigned long flags;
206 
207 	do {
208 		spin_lock_irqsave(&queue->response_lock, flags);
209 		make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
210 		push_tx_responses(queue);
211 		spin_unlock_irqrestore(&queue->response_lock, flags);
212 		if (cons == end)
213 			break;
214 		RING_COPY_REQUEST(&queue->tx, cons++, txp);
215 		extra_count = 0; /* only the first frag can have extras */
216 	} while (1);
217 	queue->tx.req_cons = cons;
218 }
219 
220 static void xenvif_fatal_tx_err(struct xenvif *vif)
221 {
222 	netdev_err(vif->dev, "fatal error; disabling device\n");
223 	vif->disabled = true;
224 	/* Disable the vif from queue 0's kthread */
225 	if (vif->num_queues)
226 		xenvif_kick_thread(&vif->queues[0]);
227 }
228 
229 static int xenvif_count_requests(struct xenvif_queue *queue,
230 				 struct xen_netif_tx_request *first,
231 				 unsigned int extra_count,
232 				 struct xen_netif_tx_request *txp,
233 				 int work_to_do)
234 {
235 	RING_IDX cons = queue->tx.req_cons;
236 	int slots = 0;
237 	int drop_err = 0;
238 	int more_data;
239 
240 	if (!(first->flags & XEN_NETTXF_more_data))
241 		return 0;
242 
243 	do {
244 		struct xen_netif_tx_request dropped_tx = { 0 };
245 
246 		if (slots >= work_to_do) {
247 			netdev_err(queue->vif->dev,
248 				   "Asked for %d slots but exceeds this limit\n",
249 				   work_to_do);
250 			xenvif_fatal_tx_err(queue->vif);
251 			return -ENODATA;
252 		}
253 
254 		/* This guest is really using too many slots and
255 		 * considered malicious.
256 		 */
257 		if (unlikely(slots >= fatal_skb_slots)) {
258 			netdev_err(queue->vif->dev,
259 				   "Malicious frontend using %d slots, threshold %u\n",
260 				   slots, fatal_skb_slots);
261 			xenvif_fatal_tx_err(queue->vif);
262 			return -E2BIG;
263 		}
264 
265 		/* Xen network protocol had implicit dependency on
266 		 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
267 		 * the historical MAX_SKB_FRAGS value 18 to honor the
268 		 * same behavior as before. Any packet using more than
269 		 * 18 slots but less than fatal_skb_slots slots is
270 		 * dropped
271 		 */
272 		if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
273 			if (net_ratelimit())
274 				netdev_dbg(queue->vif->dev,
275 					   "Too many slots (%d) exceeding limit (%d), dropping packet\n",
276 					   slots, XEN_NETBK_LEGACY_SLOTS_MAX);
277 			drop_err = -E2BIG;
278 		}
279 
280 		if (drop_err)
281 			txp = &dropped_tx;
282 
283 		RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
284 
285 		/* If the guest submitted a frame >= 64 KiB then
286 		 * first->size overflowed and following slots will
287 		 * appear to be larger than the frame.
288 		 *
289 		 * This cannot be fatal error as there are buggy
290 		 * frontends that do this.
291 		 *
292 		 * Consume all slots and drop the packet.
293 		 */
294 		if (!drop_err && txp->size > first->size) {
295 			if (net_ratelimit())
296 				netdev_dbg(queue->vif->dev,
297 					   "Invalid tx request, slot size %u > remaining size %u\n",
298 					   txp->size, first->size);
299 			drop_err = -EIO;
300 		}
301 
302 		first->size -= txp->size;
303 		slots++;
304 
305 		if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
306 			netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
307 				 txp->offset, txp->size);
308 			xenvif_fatal_tx_err(queue->vif);
309 			return -EINVAL;
310 		}
311 
312 		more_data = txp->flags & XEN_NETTXF_more_data;
313 
314 		if (!drop_err)
315 			txp++;
316 
317 	} while (more_data);
318 
319 	if (drop_err) {
320 		xenvif_tx_err(queue, first, extra_count, cons + slots);
321 		return drop_err;
322 	}
323 
324 	return slots;
325 }
326 
327 
328 struct xenvif_tx_cb {
329 	u16 pending_idx;
330 };
331 
332 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
333 
334 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
335 					   u16 pending_idx,
336 					   struct xen_netif_tx_request *txp,
337 					   unsigned int extra_count,
338 					   struct gnttab_map_grant_ref *mop)
339 {
340 	queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
341 	gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
342 			  GNTMAP_host_map | GNTMAP_readonly,
343 			  txp->gref, queue->vif->domid);
344 
345 	memcpy(&queue->pending_tx_info[pending_idx].req, txp,
346 	       sizeof(*txp));
347 	queue->pending_tx_info[pending_idx].extra_count = extra_count;
348 }
349 
350 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
351 {
352 	struct sk_buff *skb =
353 		alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
354 			  GFP_ATOMIC | __GFP_NOWARN);
355 	if (unlikely(skb == NULL))
356 		return NULL;
357 
358 	/* Packets passed to netif_rx() must have some headroom. */
359 	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
360 
361 	/* Initialize it here to avoid later surprises */
362 	skb_shinfo(skb)->destructor_arg = NULL;
363 
364 	return skb;
365 }
366 
367 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
368 							struct sk_buff *skb,
369 							struct xen_netif_tx_request *txp,
370 							struct gnttab_map_grant_ref *gop,
371 							unsigned int frag_overflow,
372 							struct sk_buff *nskb)
373 {
374 	struct skb_shared_info *shinfo = skb_shinfo(skb);
375 	skb_frag_t *frags = shinfo->frags;
376 	u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
377 	int start;
378 	pending_ring_idx_t index;
379 	unsigned int nr_slots;
380 
381 	nr_slots = shinfo->nr_frags;
382 
383 	/* Skip first skb fragment if it is on same page as header fragment. */
384 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
385 
386 	for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
387 	     shinfo->nr_frags++, txp++, gop++) {
388 		index = pending_index(queue->pending_cons++);
389 		pending_idx = queue->pending_ring[index];
390 		xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop);
391 		frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
392 	}
393 
394 	if (frag_overflow) {
395 
396 		shinfo = skb_shinfo(nskb);
397 		frags = shinfo->frags;
398 
399 		for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
400 		     shinfo->nr_frags++, txp++, gop++) {
401 			index = pending_index(queue->pending_cons++);
402 			pending_idx = queue->pending_ring[index];
403 			xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
404 						gop);
405 			frag_set_pending_idx(&frags[shinfo->nr_frags],
406 					     pending_idx);
407 		}
408 
409 		skb_shinfo(skb)->frag_list = nskb;
410 	}
411 
412 	return gop;
413 }
414 
415 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
416 					   u16 pending_idx,
417 					   grant_handle_t handle)
418 {
419 	if (unlikely(queue->grant_tx_handle[pending_idx] !=
420 		     NETBACK_INVALID_HANDLE)) {
421 		netdev_err(queue->vif->dev,
422 			   "Trying to overwrite active handle! pending_idx: 0x%x\n",
423 			   pending_idx);
424 		BUG();
425 	}
426 	queue->grant_tx_handle[pending_idx] = handle;
427 }
428 
429 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
430 					     u16 pending_idx)
431 {
432 	if (unlikely(queue->grant_tx_handle[pending_idx] ==
433 		     NETBACK_INVALID_HANDLE)) {
434 		netdev_err(queue->vif->dev,
435 			   "Trying to unmap invalid handle! pending_idx: 0x%x\n",
436 			   pending_idx);
437 		BUG();
438 	}
439 	queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
440 }
441 
442 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
443 			       struct sk_buff *skb,
444 			       struct gnttab_map_grant_ref **gopp_map,
445 			       struct gnttab_copy **gopp_copy)
446 {
447 	struct gnttab_map_grant_ref *gop_map = *gopp_map;
448 	u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
449 	/* This always points to the shinfo of the skb being checked, which
450 	 * could be either the first or the one on the frag_list
451 	 */
452 	struct skb_shared_info *shinfo = skb_shinfo(skb);
453 	/* If this is non-NULL, we are currently checking the frag_list skb, and
454 	 * this points to the shinfo of the first one
455 	 */
456 	struct skb_shared_info *first_shinfo = NULL;
457 	int nr_frags = shinfo->nr_frags;
458 	const bool sharedslot = nr_frags &&
459 				frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
460 	int i, err;
461 
462 	/* Check status of header. */
463 	err = (*gopp_copy)->status;
464 	if (unlikely(err)) {
465 		if (net_ratelimit())
466 			netdev_dbg(queue->vif->dev,
467 				   "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
468 				   (*gopp_copy)->status,
469 				   pending_idx,
470 				   (*gopp_copy)->source.u.ref);
471 		/* The first frag might still have this slot mapped */
472 		if (!sharedslot)
473 			xenvif_idx_release(queue, pending_idx,
474 					   XEN_NETIF_RSP_ERROR);
475 	}
476 	(*gopp_copy)++;
477 
478 check_frags:
479 	for (i = 0; i < nr_frags; i++, gop_map++) {
480 		int j, newerr;
481 
482 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
483 
484 		/* Check error status: if okay then remember grant handle. */
485 		newerr = gop_map->status;
486 
487 		if (likely(!newerr)) {
488 			xenvif_grant_handle_set(queue,
489 						pending_idx,
490 						gop_map->handle);
491 			/* Had a previous error? Invalidate this fragment. */
492 			if (unlikely(err)) {
493 				xenvif_idx_unmap(queue, pending_idx);
494 				/* If the mapping of the first frag was OK, but
495 				 * the header's copy failed, and they are
496 				 * sharing a slot, send an error
497 				 */
498 				if (i == 0 && sharedslot)
499 					xenvif_idx_release(queue, pending_idx,
500 							   XEN_NETIF_RSP_ERROR);
501 				else
502 					xenvif_idx_release(queue, pending_idx,
503 							   XEN_NETIF_RSP_OKAY);
504 			}
505 			continue;
506 		}
507 
508 		/* Error on this fragment: respond to client with an error. */
509 		if (net_ratelimit())
510 			netdev_dbg(queue->vif->dev,
511 				   "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
512 				   i,
513 				   gop_map->status,
514 				   pending_idx,
515 				   gop_map->ref);
516 
517 		xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
518 
519 		/* Not the first error? Preceding frags already invalidated. */
520 		if (err)
521 			continue;
522 
523 		/* First error: if the header haven't shared a slot with the
524 		 * first frag, release it as well.
525 		 */
526 		if (!sharedslot)
527 			xenvif_idx_release(queue,
528 					   XENVIF_TX_CB(skb)->pending_idx,
529 					   XEN_NETIF_RSP_OKAY);
530 
531 		/* Invalidate preceding fragments of this skb. */
532 		for (j = 0; j < i; j++) {
533 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
534 			xenvif_idx_unmap(queue, pending_idx);
535 			xenvif_idx_release(queue, pending_idx,
536 					   XEN_NETIF_RSP_OKAY);
537 		}
538 
539 		/* And if we found the error while checking the frag_list, unmap
540 		 * the first skb's frags
541 		 */
542 		if (first_shinfo) {
543 			for (j = 0; j < first_shinfo->nr_frags; j++) {
544 				pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
545 				xenvif_idx_unmap(queue, pending_idx);
546 				xenvif_idx_release(queue, pending_idx,
547 						   XEN_NETIF_RSP_OKAY);
548 			}
549 		}
550 
551 		/* Remember the error: invalidate all subsequent fragments. */
552 		err = newerr;
553 	}
554 
555 	if (skb_has_frag_list(skb) && !first_shinfo) {
556 		first_shinfo = skb_shinfo(skb);
557 		shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
558 		nr_frags = shinfo->nr_frags;
559 
560 		goto check_frags;
561 	}
562 
563 	*gopp_map = gop_map;
564 	return err;
565 }
566 
567 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
568 {
569 	struct skb_shared_info *shinfo = skb_shinfo(skb);
570 	int nr_frags = shinfo->nr_frags;
571 	int i;
572 	u16 prev_pending_idx = INVALID_PENDING_IDX;
573 
574 	for (i = 0; i < nr_frags; i++) {
575 		skb_frag_t *frag = shinfo->frags + i;
576 		struct xen_netif_tx_request *txp;
577 		struct page *page;
578 		u16 pending_idx;
579 
580 		pending_idx = frag_get_pending_idx(frag);
581 
582 		/* If this is not the first frag, chain it to the previous*/
583 		if (prev_pending_idx == INVALID_PENDING_IDX)
584 			skb_shinfo(skb)->destructor_arg =
585 				&callback_param(queue, pending_idx);
586 		else
587 			callback_param(queue, prev_pending_idx).ctx =
588 				&callback_param(queue, pending_idx);
589 
590 		callback_param(queue, pending_idx).ctx = NULL;
591 		prev_pending_idx = pending_idx;
592 
593 		txp = &queue->pending_tx_info[pending_idx].req;
594 		page = virt_to_page(idx_to_kaddr(queue, pending_idx));
595 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
596 		skb->len += txp->size;
597 		skb->data_len += txp->size;
598 		skb->truesize += txp->size;
599 
600 		/* Take an extra reference to offset network stack's put_page */
601 		get_page(queue->mmap_pages[pending_idx]);
602 	}
603 }
604 
605 static int xenvif_get_extras(struct xenvif_queue *queue,
606 			     struct xen_netif_extra_info *extras,
607 			     unsigned int *extra_count,
608 			     int work_to_do)
609 {
610 	struct xen_netif_extra_info extra;
611 	RING_IDX cons = queue->tx.req_cons;
612 
613 	do {
614 		if (unlikely(work_to_do-- <= 0)) {
615 			netdev_err(queue->vif->dev, "Missing extra info\n");
616 			xenvif_fatal_tx_err(queue->vif);
617 			return -EBADR;
618 		}
619 
620 		RING_COPY_REQUEST(&queue->tx, cons, &extra);
621 
622 		queue->tx.req_cons = ++cons;
623 		(*extra_count)++;
624 
625 		if (unlikely(!extra.type ||
626 			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
627 			netdev_err(queue->vif->dev,
628 				   "Invalid extra type: %d\n", extra.type);
629 			xenvif_fatal_tx_err(queue->vif);
630 			return -EINVAL;
631 		}
632 
633 		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
634 	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
635 
636 	return work_to_do;
637 }
638 
639 static int xenvif_set_skb_gso(struct xenvif *vif,
640 			      struct sk_buff *skb,
641 			      struct xen_netif_extra_info *gso)
642 {
643 	if (!gso->u.gso.size) {
644 		netdev_err(vif->dev, "GSO size must not be zero.\n");
645 		xenvif_fatal_tx_err(vif);
646 		return -EINVAL;
647 	}
648 
649 	switch (gso->u.gso.type) {
650 	case XEN_NETIF_GSO_TYPE_TCPV4:
651 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
652 		break;
653 	case XEN_NETIF_GSO_TYPE_TCPV6:
654 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
655 		break;
656 	default:
657 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
658 		xenvif_fatal_tx_err(vif);
659 		return -EINVAL;
660 	}
661 
662 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
663 	/* gso_segs will be calculated later */
664 
665 	return 0;
666 }
667 
668 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
669 {
670 	bool recalculate_partial_csum = false;
671 
672 	/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
673 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
674 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
675 	 * recalculate the partial checksum.
676 	 */
677 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
678 		queue->stats.rx_gso_checksum_fixup++;
679 		skb->ip_summed = CHECKSUM_PARTIAL;
680 		recalculate_partial_csum = true;
681 	}
682 
683 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
684 	if (skb->ip_summed != CHECKSUM_PARTIAL)
685 		return 0;
686 
687 	return skb_checksum_setup(skb, recalculate_partial_csum);
688 }
689 
690 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
691 {
692 	u64 now = get_jiffies_64();
693 	u64 next_credit = queue->credit_window_start +
694 		msecs_to_jiffies(queue->credit_usec / 1000);
695 
696 	/* Timer could already be pending in rare cases. */
697 	if (timer_pending(&queue->credit_timeout)) {
698 		queue->rate_limited = true;
699 		return true;
700 	}
701 
702 	/* Passed the point where we can replenish credit? */
703 	if (time_after_eq64(now, next_credit)) {
704 		queue->credit_window_start = now;
705 		tx_add_credit(queue);
706 	}
707 
708 	/* Still too big to send right now? Set a callback. */
709 	if (size > queue->remaining_credit) {
710 		mod_timer(&queue->credit_timeout,
711 			  next_credit);
712 		queue->credit_window_start = next_credit;
713 		queue->rate_limited = true;
714 
715 		return true;
716 	}
717 
718 	return false;
719 }
720 
721 /* No locking is required in xenvif_mcast_add/del() as they are
722  * only ever invoked from NAPI poll. An RCU list is used because
723  * xenvif_mcast_match() is called asynchronously, during start_xmit.
724  */
725 
726 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
727 {
728 	struct xenvif_mcast_addr *mcast;
729 
730 	if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
731 		if (net_ratelimit())
732 			netdev_err(vif->dev,
733 				   "Too many multicast addresses\n");
734 		return -ENOSPC;
735 	}
736 
737 	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
738 	if (!mcast)
739 		return -ENOMEM;
740 
741 	ether_addr_copy(mcast->addr, addr);
742 	list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
743 	vif->fe_mcast_count++;
744 
745 	return 0;
746 }
747 
748 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
749 {
750 	struct xenvif_mcast_addr *mcast;
751 
752 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
753 		if (ether_addr_equal(addr, mcast->addr)) {
754 			--vif->fe_mcast_count;
755 			list_del_rcu(&mcast->entry);
756 			kfree_rcu(mcast, rcu);
757 			break;
758 		}
759 	}
760 }
761 
762 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
763 {
764 	struct xenvif_mcast_addr *mcast;
765 
766 	rcu_read_lock();
767 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
768 		if (ether_addr_equal(addr, mcast->addr)) {
769 			rcu_read_unlock();
770 			return true;
771 		}
772 	}
773 	rcu_read_unlock();
774 
775 	return false;
776 }
777 
778 void xenvif_mcast_addr_list_free(struct xenvif *vif)
779 {
780 	/* No need for locking or RCU here. NAPI poll and TX queue
781 	 * are stopped.
782 	 */
783 	while (!list_empty(&vif->fe_mcast_addr)) {
784 		struct xenvif_mcast_addr *mcast;
785 
786 		mcast = list_first_entry(&vif->fe_mcast_addr,
787 					 struct xenvif_mcast_addr,
788 					 entry);
789 		--vif->fe_mcast_count;
790 		list_del(&mcast->entry);
791 		kfree(mcast);
792 	}
793 }
794 
795 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
796 				     int budget,
797 				     unsigned *copy_ops,
798 				     unsigned *map_ops)
799 {
800 	struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
801 	struct sk_buff *skb, *nskb;
802 	int ret;
803 	unsigned int frag_overflow;
804 
805 	while (skb_queue_len(&queue->tx_queue) < budget) {
806 		struct xen_netif_tx_request txreq;
807 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
808 		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
809 		unsigned int extra_count;
810 		u16 pending_idx;
811 		RING_IDX idx;
812 		int work_to_do;
813 		unsigned int data_len;
814 		pending_ring_idx_t index;
815 
816 		if (queue->tx.sring->req_prod - queue->tx.req_cons >
817 		    XEN_NETIF_TX_RING_SIZE) {
818 			netdev_err(queue->vif->dev,
819 				   "Impossible number of requests. "
820 				   "req_prod %d, req_cons %d, size %ld\n",
821 				   queue->tx.sring->req_prod, queue->tx.req_cons,
822 				   XEN_NETIF_TX_RING_SIZE);
823 			xenvif_fatal_tx_err(queue->vif);
824 			break;
825 		}
826 
827 		work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
828 		if (!work_to_do)
829 			break;
830 
831 		idx = queue->tx.req_cons;
832 		rmb(); /* Ensure that we see the request before we copy it. */
833 		RING_COPY_REQUEST(&queue->tx, idx, &txreq);
834 
835 		/* Credit-based scheduling. */
836 		if (txreq.size > queue->remaining_credit &&
837 		    tx_credit_exceeded(queue, txreq.size))
838 			break;
839 
840 		queue->remaining_credit -= txreq.size;
841 
842 		work_to_do--;
843 		queue->tx.req_cons = ++idx;
844 
845 		memset(extras, 0, sizeof(extras));
846 		extra_count = 0;
847 		if (txreq.flags & XEN_NETTXF_extra_info) {
848 			work_to_do = xenvif_get_extras(queue, extras,
849 						       &extra_count,
850 						       work_to_do);
851 			idx = queue->tx.req_cons;
852 			if (unlikely(work_to_do < 0))
853 				break;
854 		}
855 
856 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
857 			struct xen_netif_extra_info *extra;
858 
859 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
860 			ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
861 
862 			make_tx_response(queue, &txreq, extra_count,
863 					 (ret == 0) ?
864 					 XEN_NETIF_RSP_OKAY :
865 					 XEN_NETIF_RSP_ERROR);
866 			push_tx_responses(queue);
867 			continue;
868 		}
869 
870 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
871 			struct xen_netif_extra_info *extra;
872 
873 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
874 			xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
875 
876 			make_tx_response(queue, &txreq, extra_count,
877 					 XEN_NETIF_RSP_OKAY);
878 			push_tx_responses(queue);
879 			continue;
880 		}
881 
882 		ret = xenvif_count_requests(queue, &txreq, extra_count,
883 					    txfrags, work_to_do);
884 		if (unlikely(ret < 0))
885 			break;
886 
887 		idx += ret;
888 
889 		if (unlikely(txreq.size < ETH_HLEN)) {
890 			netdev_dbg(queue->vif->dev,
891 				   "Bad packet size: %d\n", txreq.size);
892 			xenvif_tx_err(queue, &txreq, extra_count, idx);
893 			break;
894 		}
895 
896 		/* No crossing a page as the payload mustn't fragment. */
897 		if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
898 			netdev_err(queue->vif->dev,
899 				   "txreq.offset: %u, size: %u, end: %lu\n",
900 				   txreq.offset, txreq.size,
901 				   (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
902 			xenvif_fatal_tx_err(queue->vif);
903 			break;
904 		}
905 
906 		index = pending_index(queue->pending_cons);
907 		pending_idx = queue->pending_ring[index];
908 
909 		data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
910 			    ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
911 			XEN_NETBACK_TX_COPY_LEN : txreq.size;
912 
913 		skb = xenvif_alloc_skb(data_len);
914 		if (unlikely(skb == NULL)) {
915 			netdev_dbg(queue->vif->dev,
916 				   "Can't allocate a skb in start_xmit.\n");
917 			xenvif_tx_err(queue, &txreq, extra_count, idx);
918 			break;
919 		}
920 
921 		skb_shinfo(skb)->nr_frags = ret;
922 		if (data_len < txreq.size)
923 			skb_shinfo(skb)->nr_frags++;
924 		/* At this point shinfo->nr_frags is in fact the number of
925 		 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
926 		 */
927 		frag_overflow = 0;
928 		nskb = NULL;
929 		if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
930 			frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
931 			BUG_ON(frag_overflow > MAX_SKB_FRAGS);
932 			skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
933 			nskb = xenvif_alloc_skb(0);
934 			if (unlikely(nskb == NULL)) {
935 				skb_shinfo(skb)->nr_frags = 0;
936 				kfree_skb(skb);
937 				xenvif_tx_err(queue, &txreq, extra_count, idx);
938 				if (net_ratelimit())
939 					netdev_err(queue->vif->dev,
940 						   "Can't allocate the frag_list skb.\n");
941 				break;
942 			}
943 		}
944 
945 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
946 			struct xen_netif_extra_info *gso;
947 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
948 
949 			if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
950 				/* Failure in xenvif_set_skb_gso is fatal. */
951 				skb_shinfo(skb)->nr_frags = 0;
952 				kfree_skb(skb);
953 				kfree_skb(nskb);
954 				break;
955 			}
956 		}
957 
958 		if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
959 			struct xen_netif_extra_info *extra;
960 			enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
961 
962 			extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
963 
964 			switch (extra->u.hash.type) {
965 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
966 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
967 				type = PKT_HASH_TYPE_L3;
968 				break;
969 
970 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
971 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
972 				type = PKT_HASH_TYPE_L4;
973 				break;
974 
975 			default:
976 				break;
977 			}
978 
979 			if (type != PKT_HASH_TYPE_NONE)
980 				skb_set_hash(skb,
981 					     *(u32 *)extra->u.hash.value,
982 					     type);
983 		}
984 
985 		XENVIF_TX_CB(skb)->pending_idx = pending_idx;
986 
987 		__skb_put(skb, data_len);
988 		queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
989 		queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
990 		queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
991 
992 		queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
993 			virt_to_gfn(skb->data);
994 		queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
995 		queue->tx_copy_ops[*copy_ops].dest.offset =
996 			offset_in_page(skb->data) & ~XEN_PAGE_MASK;
997 
998 		queue->tx_copy_ops[*copy_ops].len = data_len;
999 		queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1000 
1001 		(*copy_ops)++;
1002 
1003 		if (data_len < txreq.size) {
1004 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1005 					     pending_idx);
1006 			xenvif_tx_create_map_op(queue, pending_idx, &txreq,
1007 						extra_count, gop);
1008 			gop++;
1009 		} else {
1010 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1011 					     INVALID_PENDING_IDX);
1012 			memcpy(&queue->pending_tx_info[pending_idx].req,
1013 			       &txreq, sizeof(txreq));
1014 			queue->pending_tx_info[pending_idx].extra_count =
1015 				extra_count;
1016 		}
1017 
1018 		queue->pending_cons++;
1019 
1020 		gop = xenvif_get_requests(queue, skb, txfrags, gop,
1021 				          frag_overflow, nskb);
1022 
1023 		__skb_queue_tail(&queue->tx_queue, skb);
1024 
1025 		queue->tx.req_cons = idx;
1026 
1027 		if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1028 		    (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1029 			break;
1030 	}
1031 
1032 	(*map_ops) = gop - queue->tx_map_ops;
1033 	return;
1034 }
1035 
1036 /* Consolidate skb with a frag_list into a brand new one with local pages on
1037  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1038  */
1039 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1040 {
1041 	unsigned int offset = skb_headlen(skb);
1042 	skb_frag_t frags[MAX_SKB_FRAGS];
1043 	int i, f;
1044 	struct ubuf_info *uarg;
1045 	struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1046 
1047 	queue->stats.tx_zerocopy_sent += 2;
1048 	queue->stats.tx_frag_overflow++;
1049 
1050 	xenvif_fill_frags(queue, nskb);
1051 	/* Subtract frags size, we will correct it later */
1052 	skb->truesize -= skb->data_len;
1053 	skb->len += nskb->len;
1054 	skb->data_len += nskb->len;
1055 
1056 	/* create a brand new frags array and coalesce there */
1057 	for (i = 0; offset < skb->len; i++) {
1058 		struct page *page;
1059 		unsigned int len;
1060 
1061 		BUG_ON(i >= MAX_SKB_FRAGS);
1062 		page = alloc_page(GFP_ATOMIC);
1063 		if (!page) {
1064 			int j;
1065 			skb->truesize += skb->data_len;
1066 			for (j = 0; j < i; j++)
1067 				put_page(skb_frag_page(&frags[j]));
1068 			return -ENOMEM;
1069 		}
1070 
1071 		if (offset + PAGE_SIZE < skb->len)
1072 			len = PAGE_SIZE;
1073 		else
1074 			len = skb->len - offset;
1075 		if (skb_copy_bits(skb, offset, page_address(page), len))
1076 			BUG();
1077 
1078 		offset += len;
1079 		__skb_frag_set_page(&frags[i], page);
1080 		skb_frag_off_set(&frags[i], 0);
1081 		skb_frag_size_set(&frags[i], len);
1082 	}
1083 
1084 	/* Release all the original (foreign) frags. */
1085 	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1086 		skb_frag_unref(skb, f);
1087 	uarg = skb_shinfo(skb)->destructor_arg;
1088 	/* increase inflight counter to offset decrement in callback */
1089 	atomic_inc(&queue->inflight_packets);
1090 	uarg->callback(uarg, true);
1091 	skb_shinfo(skb)->destructor_arg = NULL;
1092 
1093 	/* Fill the skb with the new (local) frags. */
1094 	memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1095 	skb_shinfo(skb)->nr_frags = i;
1096 	skb->truesize += i * PAGE_SIZE;
1097 
1098 	return 0;
1099 }
1100 
1101 static int xenvif_tx_submit(struct xenvif_queue *queue)
1102 {
1103 	struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1104 	struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1105 	struct sk_buff *skb;
1106 	int work_done = 0;
1107 
1108 	while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1109 		struct xen_netif_tx_request *txp;
1110 		u16 pending_idx;
1111 		unsigned data_len;
1112 
1113 		pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1114 		txp = &queue->pending_tx_info[pending_idx].req;
1115 
1116 		/* Check the remap error code. */
1117 		if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1118 			/* If there was an error, xenvif_tx_check_gop is
1119 			 * expected to release all the frags which were mapped,
1120 			 * so kfree_skb shouldn't do it again
1121 			 */
1122 			skb_shinfo(skb)->nr_frags = 0;
1123 			if (skb_has_frag_list(skb)) {
1124 				struct sk_buff *nskb =
1125 						skb_shinfo(skb)->frag_list;
1126 				skb_shinfo(nskb)->nr_frags = 0;
1127 			}
1128 			kfree_skb(skb);
1129 			continue;
1130 		}
1131 
1132 		data_len = skb->len;
1133 		callback_param(queue, pending_idx).ctx = NULL;
1134 		if (data_len < txp->size) {
1135 			/* Append the packet payload as a fragment. */
1136 			txp->offset += data_len;
1137 			txp->size -= data_len;
1138 		} else {
1139 			/* Schedule a response immediately. */
1140 			xenvif_idx_release(queue, pending_idx,
1141 					   XEN_NETIF_RSP_OKAY);
1142 		}
1143 
1144 		if (txp->flags & XEN_NETTXF_csum_blank)
1145 			skb->ip_summed = CHECKSUM_PARTIAL;
1146 		else if (txp->flags & XEN_NETTXF_data_validated)
1147 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1148 
1149 		xenvif_fill_frags(queue, skb);
1150 
1151 		if (unlikely(skb_has_frag_list(skb))) {
1152 			struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1153 			xenvif_skb_zerocopy_prepare(queue, nskb);
1154 			if (xenvif_handle_frag_list(queue, skb)) {
1155 				if (net_ratelimit())
1156 					netdev_err(queue->vif->dev,
1157 						   "Not enough memory to consolidate frag_list!\n");
1158 				xenvif_skb_zerocopy_prepare(queue, skb);
1159 				kfree_skb(skb);
1160 				continue;
1161 			}
1162 			/* Copied all the bits from the frag list -- free it. */
1163 			skb_frag_list_init(skb);
1164 			kfree_skb(nskb);
1165 		}
1166 
1167 		skb->dev      = queue->vif->dev;
1168 		skb->protocol = eth_type_trans(skb, skb->dev);
1169 		skb_reset_network_header(skb);
1170 
1171 		if (checksum_setup(queue, skb)) {
1172 			netdev_dbg(queue->vif->dev,
1173 				   "Can't setup checksum in net_tx_action\n");
1174 			/* We have to set this flag to trigger the callback */
1175 			if (skb_shinfo(skb)->destructor_arg)
1176 				xenvif_skb_zerocopy_prepare(queue, skb);
1177 			kfree_skb(skb);
1178 			continue;
1179 		}
1180 
1181 		skb_probe_transport_header(skb);
1182 
1183 		/* If the packet is GSO then we will have just set up the
1184 		 * transport header offset in checksum_setup so it's now
1185 		 * straightforward to calculate gso_segs.
1186 		 */
1187 		if (skb_is_gso(skb)) {
1188 			int mss, hdrlen;
1189 
1190 			/* GSO implies having the L4 header. */
1191 			WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1192 			if (unlikely(!skb_transport_header_was_set(skb))) {
1193 				kfree_skb(skb);
1194 				continue;
1195 			}
1196 
1197 			mss = skb_shinfo(skb)->gso_size;
1198 			hdrlen = skb_transport_header(skb) -
1199 				skb_mac_header(skb) +
1200 				tcp_hdrlen(skb);
1201 
1202 			skb_shinfo(skb)->gso_segs =
1203 				DIV_ROUND_UP(skb->len - hdrlen, mss);
1204 		}
1205 
1206 		queue->stats.rx_bytes += skb->len;
1207 		queue->stats.rx_packets++;
1208 
1209 		work_done++;
1210 
1211 		/* Set this flag right before netif_receive_skb, otherwise
1212 		 * someone might think this packet already left netback, and
1213 		 * do a skb_copy_ubufs while we are still in control of the
1214 		 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1215 		 */
1216 		if (skb_shinfo(skb)->destructor_arg) {
1217 			xenvif_skb_zerocopy_prepare(queue, skb);
1218 			queue->stats.tx_zerocopy_sent++;
1219 		}
1220 
1221 		netif_receive_skb(skb);
1222 	}
1223 
1224 	return work_done;
1225 }
1226 
1227 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1228 {
1229 	unsigned long flags;
1230 	pending_ring_idx_t index;
1231 	struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1232 
1233 	/* This is the only place where we grab this lock, to protect callbacks
1234 	 * from each other.
1235 	 */
1236 	spin_lock_irqsave(&queue->callback_lock, flags);
1237 	do {
1238 		u16 pending_idx = ubuf->desc;
1239 		ubuf = (struct ubuf_info *) ubuf->ctx;
1240 		BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1241 			MAX_PENDING_REQS);
1242 		index = pending_index(queue->dealloc_prod);
1243 		queue->dealloc_ring[index] = pending_idx;
1244 		/* Sync with xenvif_tx_dealloc_action:
1245 		 * insert idx then incr producer.
1246 		 */
1247 		smp_wmb();
1248 		queue->dealloc_prod++;
1249 	} while (ubuf);
1250 	spin_unlock_irqrestore(&queue->callback_lock, flags);
1251 
1252 	if (likely(zerocopy_success))
1253 		queue->stats.tx_zerocopy_success++;
1254 	else
1255 		queue->stats.tx_zerocopy_fail++;
1256 	xenvif_skb_zerocopy_complete(queue);
1257 }
1258 
1259 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1260 {
1261 	struct gnttab_unmap_grant_ref *gop;
1262 	pending_ring_idx_t dc, dp;
1263 	u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1264 	unsigned int i = 0;
1265 
1266 	dc = queue->dealloc_cons;
1267 	gop = queue->tx_unmap_ops;
1268 
1269 	/* Free up any grants we have finished using */
1270 	do {
1271 		dp = queue->dealloc_prod;
1272 
1273 		/* Ensure we see all indices enqueued by all
1274 		 * xenvif_zerocopy_callback().
1275 		 */
1276 		smp_rmb();
1277 
1278 		while (dc != dp) {
1279 			BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1280 			pending_idx =
1281 				queue->dealloc_ring[pending_index(dc++)];
1282 
1283 			pending_idx_release[gop - queue->tx_unmap_ops] =
1284 				pending_idx;
1285 			queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1286 				queue->mmap_pages[pending_idx];
1287 			gnttab_set_unmap_op(gop,
1288 					    idx_to_kaddr(queue, pending_idx),
1289 					    GNTMAP_host_map,
1290 					    queue->grant_tx_handle[pending_idx]);
1291 			xenvif_grant_handle_reset(queue, pending_idx);
1292 			++gop;
1293 		}
1294 
1295 	} while (dp != queue->dealloc_prod);
1296 
1297 	queue->dealloc_cons = dc;
1298 
1299 	if (gop - queue->tx_unmap_ops > 0) {
1300 		int ret;
1301 		ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1302 					NULL,
1303 					queue->pages_to_unmap,
1304 					gop - queue->tx_unmap_ops);
1305 		if (ret) {
1306 			netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1307 				   gop - queue->tx_unmap_ops, ret);
1308 			for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1309 				if (gop[i].status != GNTST_okay)
1310 					netdev_err(queue->vif->dev,
1311 						   " host_addr: 0x%llx handle: 0x%x status: %d\n",
1312 						   gop[i].host_addr,
1313 						   gop[i].handle,
1314 						   gop[i].status);
1315 			}
1316 			BUG();
1317 		}
1318 	}
1319 
1320 	for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1321 		xenvif_idx_release(queue, pending_idx_release[i],
1322 				   XEN_NETIF_RSP_OKAY);
1323 }
1324 
1325 
1326 /* Called after netfront has transmitted */
1327 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1328 {
1329 	unsigned nr_mops, nr_cops = 0;
1330 	int work_done, ret;
1331 
1332 	if (unlikely(!tx_work_todo(queue)))
1333 		return 0;
1334 
1335 	xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1336 
1337 	if (nr_cops == 0)
1338 		return 0;
1339 
1340 	gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1341 	if (nr_mops != 0) {
1342 		ret = gnttab_map_refs(queue->tx_map_ops,
1343 				      NULL,
1344 				      queue->pages_to_map,
1345 				      nr_mops);
1346 		BUG_ON(ret);
1347 	}
1348 
1349 	work_done = xenvif_tx_submit(queue);
1350 
1351 	return work_done;
1352 }
1353 
1354 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1355 			       u8 status)
1356 {
1357 	struct pending_tx_info *pending_tx_info;
1358 	pending_ring_idx_t index;
1359 	unsigned long flags;
1360 
1361 	pending_tx_info = &queue->pending_tx_info[pending_idx];
1362 
1363 	spin_lock_irqsave(&queue->response_lock, flags);
1364 
1365 	make_tx_response(queue, &pending_tx_info->req,
1366 			 pending_tx_info->extra_count, status);
1367 
1368 	/* Release the pending index before pusing the Tx response so
1369 	 * its available before a new Tx request is pushed by the
1370 	 * frontend.
1371 	 */
1372 	index = pending_index(queue->pending_prod++);
1373 	queue->pending_ring[index] = pending_idx;
1374 
1375 	push_tx_responses(queue);
1376 
1377 	spin_unlock_irqrestore(&queue->response_lock, flags);
1378 }
1379 
1380 
1381 static void make_tx_response(struct xenvif_queue *queue,
1382 			     struct xen_netif_tx_request *txp,
1383 			     unsigned int extra_count,
1384 			     s8       st)
1385 {
1386 	RING_IDX i = queue->tx.rsp_prod_pvt;
1387 	struct xen_netif_tx_response *resp;
1388 
1389 	resp = RING_GET_RESPONSE(&queue->tx, i);
1390 	resp->id     = txp->id;
1391 	resp->status = st;
1392 
1393 	while (extra_count-- != 0)
1394 		RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1395 
1396 	queue->tx.rsp_prod_pvt = ++i;
1397 }
1398 
1399 static void push_tx_responses(struct xenvif_queue *queue)
1400 {
1401 	int notify;
1402 
1403 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1404 	if (notify)
1405 		notify_remote_via_irq(queue->tx_irq);
1406 }
1407 
1408 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1409 {
1410 	int ret;
1411 	struct gnttab_unmap_grant_ref tx_unmap_op;
1412 
1413 	gnttab_set_unmap_op(&tx_unmap_op,
1414 			    idx_to_kaddr(queue, pending_idx),
1415 			    GNTMAP_host_map,
1416 			    queue->grant_tx_handle[pending_idx]);
1417 	xenvif_grant_handle_reset(queue, pending_idx);
1418 
1419 	ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1420 				&queue->mmap_pages[pending_idx], 1);
1421 	if (ret) {
1422 		netdev_err(queue->vif->dev,
1423 			   "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1424 			   ret,
1425 			   pending_idx,
1426 			   tx_unmap_op.host_addr,
1427 			   tx_unmap_op.handle,
1428 			   tx_unmap_op.status);
1429 		BUG();
1430 	}
1431 }
1432 
1433 static inline int tx_work_todo(struct xenvif_queue *queue)
1434 {
1435 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1436 		return 1;
1437 
1438 	return 0;
1439 }
1440 
1441 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1442 {
1443 	return queue->dealloc_cons != queue->dealloc_prod;
1444 }
1445 
1446 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1447 {
1448 	if (queue->tx.sring)
1449 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1450 					queue->tx.sring);
1451 	if (queue->rx.sring)
1452 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1453 					queue->rx.sring);
1454 }
1455 
1456 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1457 				   grant_ref_t tx_ring_ref,
1458 				   grant_ref_t rx_ring_ref)
1459 {
1460 	void *addr;
1461 	struct xen_netif_tx_sring *txs;
1462 	struct xen_netif_rx_sring *rxs;
1463 	RING_IDX rsp_prod, req_prod;
1464 	int err = -ENOMEM;
1465 
1466 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1467 				     &tx_ring_ref, 1, &addr);
1468 	if (err)
1469 		goto err;
1470 
1471 	txs = (struct xen_netif_tx_sring *)addr;
1472 	rsp_prod = READ_ONCE(txs->rsp_prod);
1473 	req_prod = READ_ONCE(txs->req_prod);
1474 
1475 	BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1476 
1477 	err = -EIO;
1478 	if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1479 		goto err;
1480 
1481 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1482 				     &rx_ring_ref, 1, &addr);
1483 	if (err)
1484 		goto err;
1485 
1486 	rxs = (struct xen_netif_rx_sring *)addr;
1487 	rsp_prod = READ_ONCE(rxs->rsp_prod);
1488 	req_prod = READ_ONCE(rxs->req_prod);
1489 
1490 	BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1491 
1492 	err = -EIO;
1493 	if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1494 		goto err;
1495 
1496 	return 0;
1497 
1498 err:
1499 	xenvif_unmap_frontend_data_rings(queue);
1500 	return err;
1501 }
1502 
1503 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1504 {
1505 	/* Dealloc thread must remain running until all inflight
1506 	 * packets complete.
1507 	 */
1508 	return kthread_should_stop() &&
1509 		!atomic_read(&queue->inflight_packets);
1510 }
1511 
1512 int xenvif_dealloc_kthread(void *data)
1513 {
1514 	struct xenvif_queue *queue = data;
1515 
1516 	for (;;) {
1517 		wait_event_interruptible(queue->dealloc_wq,
1518 					 tx_dealloc_work_todo(queue) ||
1519 					 xenvif_dealloc_kthread_should_stop(queue));
1520 		if (xenvif_dealloc_kthread_should_stop(queue))
1521 			break;
1522 
1523 		xenvif_tx_dealloc_action(queue);
1524 		cond_resched();
1525 	}
1526 
1527 	/* Unmap anything remaining*/
1528 	if (tx_dealloc_work_todo(queue))
1529 		xenvif_tx_dealloc_action(queue);
1530 
1531 	return 0;
1532 }
1533 
1534 static void make_ctrl_response(struct xenvif *vif,
1535 			       const struct xen_netif_ctrl_request *req,
1536 			       u32 status, u32 data)
1537 {
1538 	RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1539 	struct xen_netif_ctrl_response rsp = {
1540 		.id = req->id,
1541 		.type = req->type,
1542 		.status = status,
1543 		.data = data,
1544 	};
1545 
1546 	*RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1547 	vif->ctrl.rsp_prod_pvt = ++idx;
1548 }
1549 
1550 static void push_ctrl_response(struct xenvif *vif)
1551 {
1552 	int notify;
1553 
1554 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1555 	if (notify)
1556 		notify_remote_via_irq(vif->ctrl_irq);
1557 }
1558 
1559 static void process_ctrl_request(struct xenvif *vif,
1560 				 const struct xen_netif_ctrl_request *req)
1561 {
1562 	u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1563 	u32 data = 0;
1564 
1565 	switch (req->type) {
1566 	case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1567 		status = xenvif_set_hash_alg(vif, req->data[0]);
1568 		break;
1569 
1570 	case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1571 		status = xenvif_get_hash_flags(vif, &data);
1572 		break;
1573 
1574 	case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1575 		status = xenvif_set_hash_flags(vif, req->data[0]);
1576 		break;
1577 
1578 	case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1579 		status = xenvif_set_hash_key(vif, req->data[0],
1580 					     req->data[1]);
1581 		break;
1582 
1583 	case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1584 		status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1585 		data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1586 		break;
1587 
1588 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1589 		status = xenvif_set_hash_mapping_size(vif,
1590 						      req->data[0]);
1591 		break;
1592 
1593 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1594 		status = xenvif_set_hash_mapping(vif, req->data[0],
1595 						 req->data[1],
1596 						 req->data[2]);
1597 		break;
1598 
1599 	default:
1600 		break;
1601 	}
1602 
1603 	make_ctrl_response(vif, req, status, data);
1604 	push_ctrl_response(vif);
1605 }
1606 
1607 static void xenvif_ctrl_action(struct xenvif *vif)
1608 {
1609 	for (;;) {
1610 		RING_IDX req_prod, req_cons;
1611 
1612 		req_prod = vif->ctrl.sring->req_prod;
1613 		req_cons = vif->ctrl.req_cons;
1614 
1615 		/* Make sure we can see requests before we process them. */
1616 		rmb();
1617 
1618 		if (req_cons == req_prod)
1619 			break;
1620 
1621 		while (req_cons != req_prod) {
1622 			struct xen_netif_ctrl_request req;
1623 
1624 			RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1625 			req_cons++;
1626 
1627 			process_ctrl_request(vif, &req);
1628 		}
1629 
1630 		vif->ctrl.req_cons = req_cons;
1631 		vif->ctrl.sring->req_event = req_cons + 1;
1632 	}
1633 }
1634 
1635 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1636 {
1637 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1638 		return true;
1639 
1640 	return false;
1641 }
1642 
1643 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1644 {
1645 	struct xenvif *vif = data;
1646 
1647 	while (xenvif_ctrl_work_todo(vif))
1648 		xenvif_ctrl_action(vif);
1649 
1650 	return IRQ_HANDLED;
1651 }
1652 
1653 static int __init netback_init(void)
1654 {
1655 	int rc = 0;
1656 
1657 	if (!xen_domain())
1658 		return -ENODEV;
1659 
1660 	/* Allow as many queues as there are CPUs but max. 8 if user has not
1661 	 * specified a value.
1662 	 */
1663 	if (xenvif_max_queues == 0)
1664 		xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1665 					  num_online_cpus());
1666 
1667 	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1668 		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1669 			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1670 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1671 	}
1672 
1673 	rc = xenvif_xenbus_init();
1674 	if (rc)
1675 		goto failed_init;
1676 
1677 #ifdef CONFIG_DEBUG_FS
1678 	xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1679 #endif /* CONFIG_DEBUG_FS */
1680 
1681 	return 0;
1682 
1683 failed_init:
1684 	return rc;
1685 }
1686 
1687 module_init(netback_init);
1688 
1689 static void __exit netback_fini(void)
1690 {
1691 #ifdef CONFIG_DEBUG_FS
1692 	debugfs_remove_recursive(xen_netback_dbg_root);
1693 #endif /* CONFIG_DEBUG_FS */
1694 	xenvif_xenbus_fini();
1695 }
1696 module_exit(netback_fini);
1697 
1698 MODULE_LICENSE("Dual BSD/GPL");
1699 MODULE_ALIAS("xen-backend:vif");
1700