xref: /linux/drivers/net/xen-netback/netback.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
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 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
100 			       u8 status);
101 
102 static void make_tx_response(struct xenvif_queue *queue,
103 			     struct xen_netif_tx_request *txp,
104 			     unsigned int extra_count,
105 			     s8       st);
106 static void push_tx_responses(struct xenvif_queue *queue);
107 
108 static inline int tx_work_todo(struct xenvif_queue *queue);
109 
110 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
111 				       u16 idx)
112 {
113 	return page_to_pfn(queue->mmap_pages[idx]);
114 }
115 
116 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
117 					 u16 idx)
118 {
119 	return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
120 }
121 
122 #define callback_param(vif, pending_idx) \
123 	(vif->pending_tx_info[pending_idx].callback_struct)
124 
125 /* Find the containing VIF's structure from a pointer in pending_tx_info array
126  */
127 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
128 {
129 	u16 pending_idx = ubuf->desc;
130 	struct pending_tx_info *temp =
131 		container_of(ubuf, struct pending_tx_info, callback_struct);
132 	return container_of(temp - pending_idx,
133 			    struct xenvif_queue,
134 			    pending_tx_info[0]);
135 }
136 
137 static u16 frag_get_pending_idx(skb_frag_t *frag)
138 {
139 	return (u16)frag->page_offset;
140 }
141 
142 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
143 {
144 	frag->page_offset = pending_idx;
145 }
146 
147 static inline pending_ring_idx_t pending_index(unsigned i)
148 {
149 	return i & (MAX_PENDING_REQS-1);
150 }
151 
152 void xenvif_kick_thread(struct xenvif_queue *queue)
153 {
154 	wake_up(&queue->wq);
155 }
156 
157 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
158 {
159 	int more_to_do;
160 
161 	RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
162 
163 	if (more_to_do)
164 		napi_schedule(&queue->napi);
165 }
166 
167 static void tx_add_credit(struct xenvif_queue *queue)
168 {
169 	unsigned long max_burst, max_credit;
170 
171 	/*
172 	 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
173 	 * Otherwise the interface can seize up due to insufficient credit.
174 	 */
175 	max_burst = max(131072UL, queue->credit_bytes);
176 
177 	/* Take care that adding a new chunk of credit doesn't wrap to zero. */
178 	max_credit = queue->remaining_credit + queue->credit_bytes;
179 	if (max_credit < queue->remaining_credit)
180 		max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
181 
182 	queue->remaining_credit = min(max_credit, max_burst);
183 }
184 
185 void xenvif_tx_credit_callback(unsigned long data)
186 {
187 	struct xenvif_queue *queue = (struct xenvif_queue *)data;
188 	tx_add_credit(queue);
189 	xenvif_napi_schedule_or_enable_events(queue);
190 }
191 
192 static void xenvif_tx_err(struct xenvif_queue *queue,
193 			  struct xen_netif_tx_request *txp,
194 			  unsigned int extra_count, RING_IDX end)
195 {
196 	RING_IDX cons = queue->tx.req_cons;
197 	unsigned long flags;
198 
199 	do {
200 		spin_lock_irqsave(&queue->response_lock, flags);
201 		make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
202 		push_tx_responses(queue);
203 		spin_unlock_irqrestore(&queue->response_lock, flags);
204 		if (cons == end)
205 			break;
206 		RING_COPY_REQUEST(&queue->tx, cons++, txp);
207 		extra_count = 0; /* only the first frag can have extras */
208 	} while (1);
209 	queue->tx.req_cons = cons;
210 }
211 
212 static void xenvif_fatal_tx_err(struct xenvif *vif)
213 {
214 	netdev_err(vif->dev, "fatal error; disabling device\n");
215 	vif->disabled = true;
216 	/* Disable the vif from queue 0's kthread */
217 	if (vif->num_queues)
218 		xenvif_kick_thread(&vif->queues[0]);
219 }
220 
221 static int xenvif_count_requests(struct xenvif_queue *queue,
222 				 struct xen_netif_tx_request *first,
223 				 unsigned int extra_count,
224 				 struct xen_netif_tx_request *txp,
225 				 int work_to_do)
226 {
227 	RING_IDX cons = queue->tx.req_cons;
228 	int slots = 0;
229 	int drop_err = 0;
230 	int more_data;
231 
232 	if (!(first->flags & XEN_NETTXF_more_data))
233 		return 0;
234 
235 	do {
236 		struct xen_netif_tx_request dropped_tx = { 0 };
237 
238 		if (slots >= work_to_do) {
239 			netdev_err(queue->vif->dev,
240 				   "Asked for %d slots but exceeds this limit\n",
241 				   work_to_do);
242 			xenvif_fatal_tx_err(queue->vif);
243 			return -ENODATA;
244 		}
245 
246 		/* This guest is really using too many slots and
247 		 * considered malicious.
248 		 */
249 		if (unlikely(slots >= fatal_skb_slots)) {
250 			netdev_err(queue->vif->dev,
251 				   "Malicious frontend using %d slots, threshold %u\n",
252 				   slots, fatal_skb_slots);
253 			xenvif_fatal_tx_err(queue->vif);
254 			return -E2BIG;
255 		}
256 
257 		/* Xen network protocol had implicit dependency on
258 		 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
259 		 * the historical MAX_SKB_FRAGS value 18 to honor the
260 		 * same behavior as before. Any packet using more than
261 		 * 18 slots but less than fatal_skb_slots slots is
262 		 * dropped
263 		 */
264 		if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
265 			if (net_ratelimit())
266 				netdev_dbg(queue->vif->dev,
267 					   "Too many slots (%d) exceeding limit (%d), dropping packet\n",
268 					   slots, XEN_NETBK_LEGACY_SLOTS_MAX);
269 			drop_err = -E2BIG;
270 		}
271 
272 		if (drop_err)
273 			txp = &dropped_tx;
274 
275 		RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
276 
277 		/* If the guest submitted a frame >= 64 KiB then
278 		 * first->size overflowed and following slots will
279 		 * appear to be larger than the frame.
280 		 *
281 		 * This cannot be fatal error as there are buggy
282 		 * frontends that do this.
283 		 *
284 		 * Consume all slots and drop the packet.
285 		 */
286 		if (!drop_err && txp->size > first->size) {
287 			if (net_ratelimit())
288 				netdev_dbg(queue->vif->dev,
289 					   "Invalid tx request, slot size %u > remaining size %u\n",
290 					   txp->size, first->size);
291 			drop_err = -EIO;
292 		}
293 
294 		first->size -= txp->size;
295 		slots++;
296 
297 		if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
298 			netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
299 				 txp->offset, txp->size);
300 			xenvif_fatal_tx_err(queue->vif);
301 			return -EINVAL;
302 		}
303 
304 		more_data = txp->flags & XEN_NETTXF_more_data;
305 
306 		if (!drop_err)
307 			txp++;
308 
309 	} while (more_data);
310 
311 	if (drop_err) {
312 		xenvif_tx_err(queue, first, extra_count, cons + slots);
313 		return drop_err;
314 	}
315 
316 	return slots;
317 }
318 
319 
320 struct xenvif_tx_cb {
321 	u16 pending_idx;
322 };
323 
324 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
325 
326 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
327 					   u16 pending_idx,
328 					   struct xen_netif_tx_request *txp,
329 					   unsigned int extra_count,
330 					   struct gnttab_map_grant_ref *mop)
331 {
332 	queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
333 	gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
334 			  GNTMAP_host_map | GNTMAP_readonly,
335 			  txp->gref, queue->vif->domid);
336 
337 	memcpy(&queue->pending_tx_info[pending_idx].req, txp,
338 	       sizeof(*txp));
339 	queue->pending_tx_info[pending_idx].extra_count = extra_count;
340 }
341 
342 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
343 {
344 	struct sk_buff *skb =
345 		alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
346 			  GFP_ATOMIC | __GFP_NOWARN);
347 	if (unlikely(skb == NULL))
348 		return NULL;
349 
350 	/* Packets passed to netif_rx() must have some headroom. */
351 	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
352 
353 	/* Initialize it here to avoid later surprises */
354 	skb_shinfo(skb)->destructor_arg = NULL;
355 
356 	return skb;
357 }
358 
359 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
360 							struct sk_buff *skb,
361 							struct xen_netif_tx_request *txp,
362 							struct gnttab_map_grant_ref *gop,
363 							unsigned int frag_overflow,
364 							struct sk_buff *nskb)
365 {
366 	struct skb_shared_info *shinfo = skb_shinfo(skb);
367 	skb_frag_t *frags = shinfo->frags;
368 	u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
369 	int start;
370 	pending_ring_idx_t index;
371 	unsigned int nr_slots;
372 
373 	nr_slots = shinfo->nr_frags;
374 
375 	/* Skip first skb fragment if it is on same page as header fragment. */
376 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
377 
378 	for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
379 	     shinfo->nr_frags++, txp++, gop++) {
380 		index = pending_index(queue->pending_cons++);
381 		pending_idx = queue->pending_ring[index];
382 		xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop);
383 		frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
384 	}
385 
386 	if (frag_overflow) {
387 
388 		shinfo = skb_shinfo(nskb);
389 		frags = shinfo->frags;
390 
391 		for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
392 		     shinfo->nr_frags++, txp++, gop++) {
393 			index = pending_index(queue->pending_cons++);
394 			pending_idx = queue->pending_ring[index];
395 			xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
396 						gop);
397 			frag_set_pending_idx(&frags[shinfo->nr_frags],
398 					     pending_idx);
399 		}
400 
401 		skb_shinfo(skb)->frag_list = nskb;
402 	}
403 
404 	return gop;
405 }
406 
407 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
408 					   u16 pending_idx,
409 					   grant_handle_t handle)
410 {
411 	if (unlikely(queue->grant_tx_handle[pending_idx] !=
412 		     NETBACK_INVALID_HANDLE)) {
413 		netdev_err(queue->vif->dev,
414 			   "Trying to overwrite active handle! pending_idx: 0x%x\n",
415 			   pending_idx);
416 		BUG();
417 	}
418 	queue->grant_tx_handle[pending_idx] = handle;
419 }
420 
421 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
422 					     u16 pending_idx)
423 {
424 	if (unlikely(queue->grant_tx_handle[pending_idx] ==
425 		     NETBACK_INVALID_HANDLE)) {
426 		netdev_err(queue->vif->dev,
427 			   "Trying to unmap invalid handle! pending_idx: 0x%x\n",
428 			   pending_idx);
429 		BUG();
430 	}
431 	queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
432 }
433 
434 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
435 			       struct sk_buff *skb,
436 			       struct gnttab_map_grant_ref **gopp_map,
437 			       struct gnttab_copy **gopp_copy)
438 {
439 	struct gnttab_map_grant_ref *gop_map = *gopp_map;
440 	u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
441 	/* This always points to the shinfo of the skb being checked, which
442 	 * could be either the first or the one on the frag_list
443 	 */
444 	struct skb_shared_info *shinfo = skb_shinfo(skb);
445 	/* If this is non-NULL, we are currently checking the frag_list skb, and
446 	 * this points to the shinfo of the first one
447 	 */
448 	struct skb_shared_info *first_shinfo = NULL;
449 	int nr_frags = shinfo->nr_frags;
450 	const bool sharedslot = nr_frags &&
451 				frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
452 	int i, err;
453 
454 	/* Check status of header. */
455 	err = (*gopp_copy)->status;
456 	if (unlikely(err)) {
457 		if (net_ratelimit())
458 			netdev_dbg(queue->vif->dev,
459 				   "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
460 				   (*gopp_copy)->status,
461 				   pending_idx,
462 				   (*gopp_copy)->source.u.ref);
463 		/* The first frag might still have this slot mapped */
464 		if (!sharedslot)
465 			xenvif_idx_release(queue, pending_idx,
466 					   XEN_NETIF_RSP_ERROR);
467 	}
468 	(*gopp_copy)++;
469 
470 check_frags:
471 	for (i = 0; i < nr_frags; i++, gop_map++) {
472 		int j, newerr;
473 
474 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
475 
476 		/* Check error status: if okay then remember grant handle. */
477 		newerr = gop_map->status;
478 
479 		if (likely(!newerr)) {
480 			xenvif_grant_handle_set(queue,
481 						pending_idx,
482 						gop_map->handle);
483 			/* Had a previous error? Invalidate this fragment. */
484 			if (unlikely(err)) {
485 				xenvif_idx_unmap(queue, pending_idx);
486 				/* If the mapping of the first frag was OK, but
487 				 * the header's copy failed, and they are
488 				 * sharing a slot, send an error
489 				 */
490 				if (i == 0 && sharedslot)
491 					xenvif_idx_release(queue, pending_idx,
492 							   XEN_NETIF_RSP_ERROR);
493 				else
494 					xenvif_idx_release(queue, pending_idx,
495 							   XEN_NETIF_RSP_OKAY);
496 			}
497 			continue;
498 		}
499 
500 		/* Error on this fragment: respond to client with an error. */
501 		if (net_ratelimit())
502 			netdev_dbg(queue->vif->dev,
503 				   "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
504 				   i,
505 				   gop_map->status,
506 				   pending_idx,
507 				   gop_map->ref);
508 
509 		xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
510 
511 		/* Not the first error? Preceding frags already invalidated. */
512 		if (err)
513 			continue;
514 
515 		/* First error: if the header haven't shared a slot with the
516 		 * first frag, release it as well.
517 		 */
518 		if (!sharedslot)
519 			xenvif_idx_release(queue,
520 					   XENVIF_TX_CB(skb)->pending_idx,
521 					   XEN_NETIF_RSP_OKAY);
522 
523 		/* Invalidate preceding fragments of this skb. */
524 		for (j = 0; j < i; j++) {
525 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
526 			xenvif_idx_unmap(queue, pending_idx);
527 			xenvif_idx_release(queue, pending_idx,
528 					   XEN_NETIF_RSP_OKAY);
529 		}
530 
531 		/* And if we found the error while checking the frag_list, unmap
532 		 * the first skb's frags
533 		 */
534 		if (first_shinfo) {
535 			for (j = 0; j < first_shinfo->nr_frags; j++) {
536 				pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
537 				xenvif_idx_unmap(queue, pending_idx);
538 				xenvif_idx_release(queue, pending_idx,
539 						   XEN_NETIF_RSP_OKAY);
540 			}
541 		}
542 
543 		/* Remember the error: invalidate all subsequent fragments. */
544 		err = newerr;
545 	}
546 
547 	if (skb_has_frag_list(skb) && !first_shinfo) {
548 		first_shinfo = skb_shinfo(skb);
549 		shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
550 		nr_frags = shinfo->nr_frags;
551 
552 		goto check_frags;
553 	}
554 
555 	*gopp_map = gop_map;
556 	return err;
557 }
558 
559 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
560 {
561 	struct skb_shared_info *shinfo = skb_shinfo(skb);
562 	int nr_frags = shinfo->nr_frags;
563 	int i;
564 	u16 prev_pending_idx = INVALID_PENDING_IDX;
565 
566 	for (i = 0; i < nr_frags; i++) {
567 		skb_frag_t *frag = shinfo->frags + i;
568 		struct xen_netif_tx_request *txp;
569 		struct page *page;
570 		u16 pending_idx;
571 
572 		pending_idx = frag_get_pending_idx(frag);
573 
574 		/* If this is not the first frag, chain it to the previous*/
575 		if (prev_pending_idx == INVALID_PENDING_IDX)
576 			skb_shinfo(skb)->destructor_arg =
577 				&callback_param(queue, pending_idx);
578 		else
579 			callback_param(queue, prev_pending_idx).ctx =
580 				&callback_param(queue, pending_idx);
581 
582 		callback_param(queue, pending_idx).ctx = NULL;
583 		prev_pending_idx = pending_idx;
584 
585 		txp = &queue->pending_tx_info[pending_idx].req;
586 		page = virt_to_page(idx_to_kaddr(queue, pending_idx));
587 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
588 		skb->len += txp->size;
589 		skb->data_len += txp->size;
590 		skb->truesize += txp->size;
591 
592 		/* Take an extra reference to offset network stack's put_page */
593 		get_page(queue->mmap_pages[pending_idx]);
594 	}
595 }
596 
597 static int xenvif_get_extras(struct xenvif_queue *queue,
598 			     struct xen_netif_extra_info *extras,
599 			     unsigned int *extra_count,
600 			     int work_to_do)
601 {
602 	struct xen_netif_extra_info extra;
603 	RING_IDX cons = queue->tx.req_cons;
604 
605 	do {
606 		if (unlikely(work_to_do-- <= 0)) {
607 			netdev_err(queue->vif->dev, "Missing extra info\n");
608 			xenvif_fatal_tx_err(queue->vif);
609 			return -EBADR;
610 		}
611 
612 		RING_COPY_REQUEST(&queue->tx, cons, &extra);
613 
614 		queue->tx.req_cons = ++cons;
615 		(*extra_count)++;
616 
617 		if (unlikely(!extra.type ||
618 			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
619 			netdev_err(queue->vif->dev,
620 				   "Invalid extra type: %d\n", extra.type);
621 			xenvif_fatal_tx_err(queue->vif);
622 			return -EINVAL;
623 		}
624 
625 		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
626 	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
627 
628 	return work_to_do;
629 }
630 
631 static int xenvif_set_skb_gso(struct xenvif *vif,
632 			      struct sk_buff *skb,
633 			      struct xen_netif_extra_info *gso)
634 {
635 	if (!gso->u.gso.size) {
636 		netdev_err(vif->dev, "GSO size must not be zero.\n");
637 		xenvif_fatal_tx_err(vif);
638 		return -EINVAL;
639 	}
640 
641 	switch (gso->u.gso.type) {
642 	case XEN_NETIF_GSO_TYPE_TCPV4:
643 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
644 		break;
645 	case XEN_NETIF_GSO_TYPE_TCPV6:
646 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
647 		break;
648 	default:
649 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
650 		xenvif_fatal_tx_err(vif);
651 		return -EINVAL;
652 	}
653 
654 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
655 	/* gso_segs will be calculated later */
656 
657 	return 0;
658 }
659 
660 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
661 {
662 	bool recalculate_partial_csum = false;
663 
664 	/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
665 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
666 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
667 	 * recalculate the partial checksum.
668 	 */
669 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
670 		queue->stats.rx_gso_checksum_fixup++;
671 		skb->ip_summed = CHECKSUM_PARTIAL;
672 		recalculate_partial_csum = true;
673 	}
674 
675 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
676 	if (skb->ip_summed != CHECKSUM_PARTIAL)
677 		return 0;
678 
679 	return skb_checksum_setup(skb, recalculate_partial_csum);
680 }
681 
682 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
683 {
684 	u64 now = get_jiffies_64();
685 	u64 next_credit = queue->credit_window_start +
686 		msecs_to_jiffies(queue->credit_usec / 1000);
687 
688 	/* Timer could already be pending in rare cases. */
689 	if (timer_pending(&queue->credit_timeout))
690 		return true;
691 
692 	/* Passed the point where we can replenish credit? */
693 	if (time_after_eq64(now, next_credit)) {
694 		queue->credit_window_start = now;
695 		tx_add_credit(queue);
696 	}
697 
698 	/* Still too big to send right now? Set a callback. */
699 	if (size > queue->remaining_credit) {
700 		queue->credit_timeout.data     =
701 			(unsigned long)queue;
702 		mod_timer(&queue->credit_timeout,
703 			  next_credit);
704 		queue->credit_window_start = next_credit;
705 
706 		return true;
707 	}
708 
709 	return false;
710 }
711 
712 /* No locking is required in xenvif_mcast_add/del() as they are
713  * only ever invoked from NAPI poll. An RCU list is used because
714  * xenvif_mcast_match() is called asynchronously, during start_xmit.
715  */
716 
717 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
718 {
719 	struct xenvif_mcast_addr *mcast;
720 
721 	if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
722 		if (net_ratelimit())
723 			netdev_err(vif->dev,
724 				   "Too many multicast addresses\n");
725 		return -ENOSPC;
726 	}
727 
728 	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
729 	if (!mcast)
730 		return -ENOMEM;
731 
732 	ether_addr_copy(mcast->addr, addr);
733 	list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
734 	vif->fe_mcast_count++;
735 
736 	return 0;
737 }
738 
739 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
740 {
741 	struct xenvif_mcast_addr *mcast;
742 
743 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
744 		if (ether_addr_equal(addr, mcast->addr)) {
745 			--vif->fe_mcast_count;
746 			list_del_rcu(&mcast->entry);
747 			kfree_rcu(mcast, rcu);
748 			break;
749 		}
750 	}
751 }
752 
753 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
754 {
755 	struct xenvif_mcast_addr *mcast;
756 
757 	rcu_read_lock();
758 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
759 		if (ether_addr_equal(addr, mcast->addr)) {
760 			rcu_read_unlock();
761 			return true;
762 		}
763 	}
764 	rcu_read_unlock();
765 
766 	return false;
767 }
768 
769 void xenvif_mcast_addr_list_free(struct xenvif *vif)
770 {
771 	/* No need for locking or RCU here. NAPI poll and TX queue
772 	 * are stopped.
773 	 */
774 	while (!list_empty(&vif->fe_mcast_addr)) {
775 		struct xenvif_mcast_addr *mcast;
776 
777 		mcast = list_first_entry(&vif->fe_mcast_addr,
778 					 struct xenvif_mcast_addr,
779 					 entry);
780 		--vif->fe_mcast_count;
781 		list_del(&mcast->entry);
782 		kfree(mcast);
783 	}
784 }
785 
786 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
787 				     int budget,
788 				     unsigned *copy_ops,
789 				     unsigned *map_ops)
790 {
791 	struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
792 	struct sk_buff *skb, *nskb;
793 	int ret;
794 	unsigned int frag_overflow;
795 
796 	while (skb_queue_len(&queue->tx_queue) < budget) {
797 		struct xen_netif_tx_request txreq;
798 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
799 		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
800 		unsigned int extra_count;
801 		u16 pending_idx;
802 		RING_IDX idx;
803 		int work_to_do;
804 		unsigned int data_len;
805 		pending_ring_idx_t index;
806 
807 		if (queue->tx.sring->req_prod - queue->tx.req_cons >
808 		    XEN_NETIF_TX_RING_SIZE) {
809 			netdev_err(queue->vif->dev,
810 				   "Impossible number of requests. "
811 				   "req_prod %d, req_cons %d, size %ld\n",
812 				   queue->tx.sring->req_prod, queue->tx.req_cons,
813 				   XEN_NETIF_TX_RING_SIZE);
814 			xenvif_fatal_tx_err(queue->vif);
815 			break;
816 		}
817 
818 		work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
819 		if (!work_to_do)
820 			break;
821 
822 		idx = queue->tx.req_cons;
823 		rmb(); /* Ensure that we see the request before we copy it. */
824 		RING_COPY_REQUEST(&queue->tx, idx, &txreq);
825 
826 		/* Credit-based scheduling. */
827 		if (txreq.size > queue->remaining_credit &&
828 		    tx_credit_exceeded(queue, txreq.size))
829 			break;
830 
831 		queue->remaining_credit -= txreq.size;
832 
833 		work_to_do--;
834 		queue->tx.req_cons = ++idx;
835 
836 		memset(extras, 0, sizeof(extras));
837 		extra_count = 0;
838 		if (txreq.flags & XEN_NETTXF_extra_info) {
839 			work_to_do = xenvif_get_extras(queue, extras,
840 						       &extra_count,
841 						       work_to_do);
842 			idx = queue->tx.req_cons;
843 			if (unlikely(work_to_do < 0))
844 				break;
845 		}
846 
847 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
848 			struct xen_netif_extra_info *extra;
849 
850 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
851 			ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
852 
853 			make_tx_response(queue, &txreq, extra_count,
854 					 (ret == 0) ?
855 					 XEN_NETIF_RSP_OKAY :
856 					 XEN_NETIF_RSP_ERROR);
857 			push_tx_responses(queue);
858 			continue;
859 		}
860 
861 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
862 			struct xen_netif_extra_info *extra;
863 
864 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
865 			xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
866 
867 			make_tx_response(queue, &txreq, extra_count,
868 					 XEN_NETIF_RSP_OKAY);
869 			push_tx_responses(queue);
870 			continue;
871 		}
872 
873 		ret = xenvif_count_requests(queue, &txreq, extra_count,
874 					    txfrags, work_to_do);
875 		if (unlikely(ret < 0))
876 			break;
877 
878 		idx += ret;
879 
880 		if (unlikely(txreq.size < ETH_HLEN)) {
881 			netdev_dbg(queue->vif->dev,
882 				   "Bad packet size: %d\n", txreq.size);
883 			xenvif_tx_err(queue, &txreq, extra_count, idx);
884 			break;
885 		}
886 
887 		/* No crossing a page as the payload mustn't fragment. */
888 		if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
889 			netdev_err(queue->vif->dev,
890 				   "txreq.offset: %u, size: %u, end: %lu\n",
891 				   txreq.offset, txreq.size,
892 				   (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
893 			xenvif_fatal_tx_err(queue->vif);
894 			break;
895 		}
896 
897 		index = pending_index(queue->pending_cons);
898 		pending_idx = queue->pending_ring[index];
899 
900 		data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
901 			    ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
902 			XEN_NETBACK_TX_COPY_LEN : txreq.size;
903 
904 		skb = xenvif_alloc_skb(data_len);
905 		if (unlikely(skb == NULL)) {
906 			netdev_dbg(queue->vif->dev,
907 				   "Can't allocate a skb in start_xmit.\n");
908 			xenvif_tx_err(queue, &txreq, extra_count, idx);
909 			break;
910 		}
911 
912 		skb_shinfo(skb)->nr_frags = ret;
913 		if (data_len < txreq.size)
914 			skb_shinfo(skb)->nr_frags++;
915 		/* At this point shinfo->nr_frags is in fact the number of
916 		 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
917 		 */
918 		frag_overflow = 0;
919 		nskb = NULL;
920 		if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
921 			frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
922 			BUG_ON(frag_overflow > MAX_SKB_FRAGS);
923 			skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
924 			nskb = xenvif_alloc_skb(0);
925 			if (unlikely(nskb == NULL)) {
926 				kfree_skb(skb);
927 				xenvif_tx_err(queue, &txreq, extra_count, idx);
928 				if (net_ratelimit())
929 					netdev_err(queue->vif->dev,
930 						   "Can't allocate the frag_list skb.\n");
931 				break;
932 			}
933 		}
934 
935 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
936 			struct xen_netif_extra_info *gso;
937 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
938 
939 			if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
940 				/* Failure in xenvif_set_skb_gso is fatal. */
941 				kfree_skb(skb);
942 				kfree_skb(nskb);
943 				break;
944 			}
945 		}
946 
947 		if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
948 			struct xen_netif_extra_info *extra;
949 			enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
950 
951 			extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
952 
953 			switch (extra->u.hash.type) {
954 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
955 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
956 				type = PKT_HASH_TYPE_L3;
957 				break;
958 
959 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
960 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
961 				type = PKT_HASH_TYPE_L4;
962 				break;
963 
964 			default:
965 				break;
966 			}
967 
968 			if (type != PKT_HASH_TYPE_NONE)
969 				skb_set_hash(skb,
970 					     *(u32 *)extra->u.hash.value,
971 					     type);
972 		}
973 
974 		XENVIF_TX_CB(skb)->pending_idx = pending_idx;
975 
976 		__skb_put(skb, data_len);
977 		queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
978 		queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
979 		queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
980 
981 		queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
982 			virt_to_gfn(skb->data);
983 		queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
984 		queue->tx_copy_ops[*copy_ops].dest.offset =
985 			offset_in_page(skb->data) & ~XEN_PAGE_MASK;
986 
987 		queue->tx_copy_ops[*copy_ops].len = data_len;
988 		queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
989 
990 		(*copy_ops)++;
991 
992 		if (data_len < txreq.size) {
993 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
994 					     pending_idx);
995 			xenvif_tx_create_map_op(queue, pending_idx, &txreq,
996 						extra_count, gop);
997 			gop++;
998 		} else {
999 			frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1000 					     INVALID_PENDING_IDX);
1001 			memcpy(&queue->pending_tx_info[pending_idx].req,
1002 			       &txreq, sizeof(txreq));
1003 			queue->pending_tx_info[pending_idx].extra_count =
1004 				extra_count;
1005 		}
1006 
1007 		queue->pending_cons++;
1008 
1009 		gop = xenvif_get_requests(queue, skb, txfrags, gop,
1010 				          frag_overflow, nskb);
1011 
1012 		__skb_queue_tail(&queue->tx_queue, skb);
1013 
1014 		queue->tx.req_cons = idx;
1015 
1016 		if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1017 		    (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1018 			break;
1019 	}
1020 
1021 	(*map_ops) = gop - queue->tx_map_ops;
1022 	return;
1023 }
1024 
1025 /* Consolidate skb with a frag_list into a brand new one with local pages on
1026  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1027  */
1028 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1029 {
1030 	unsigned int offset = skb_headlen(skb);
1031 	skb_frag_t frags[MAX_SKB_FRAGS];
1032 	int i, f;
1033 	struct ubuf_info *uarg;
1034 	struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1035 
1036 	queue->stats.tx_zerocopy_sent += 2;
1037 	queue->stats.tx_frag_overflow++;
1038 
1039 	xenvif_fill_frags(queue, nskb);
1040 	/* Subtract frags size, we will correct it later */
1041 	skb->truesize -= skb->data_len;
1042 	skb->len += nskb->len;
1043 	skb->data_len += nskb->len;
1044 
1045 	/* create a brand new frags array and coalesce there */
1046 	for (i = 0; offset < skb->len; i++) {
1047 		struct page *page;
1048 		unsigned int len;
1049 
1050 		BUG_ON(i >= MAX_SKB_FRAGS);
1051 		page = alloc_page(GFP_ATOMIC);
1052 		if (!page) {
1053 			int j;
1054 			skb->truesize += skb->data_len;
1055 			for (j = 0; j < i; j++)
1056 				put_page(frags[j].page.p);
1057 			return -ENOMEM;
1058 		}
1059 
1060 		if (offset + PAGE_SIZE < skb->len)
1061 			len = PAGE_SIZE;
1062 		else
1063 			len = skb->len - offset;
1064 		if (skb_copy_bits(skb, offset, page_address(page), len))
1065 			BUG();
1066 
1067 		offset += len;
1068 		frags[i].page.p = page;
1069 		frags[i].page_offset = 0;
1070 		skb_frag_size_set(&frags[i], len);
1071 	}
1072 
1073 	/* Copied all the bits from the frag list -- free it. */
1074 	skb_frag_list_init(skb);
1075 	xenvif_skb_zerocopy_prepare(queue, nskb);
1076 	kfree_skb(nskb);
1077 
1078 	/* Release all the original (foreign) frags. */
1079 	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1080 		skb_frag_unref(skb, f);
1081 	uarg = skb_shinfo(skb)->destructor_arg;
1082 	/* increase inflight counter to offset decrement in callback */
1083 	atomic_inc(&queue->inflight_packets);
1084 	uarg->callback(uarg, true);
1085 	skb_shinfo(skb)->destructor_arg = NULL;
1086 
1087 	/* Fill the skb with the new (local) frags. */
1088 	memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1089 	skb_shinfo(skb)->nr_frags = i;
1090 	skb->truesize += i * PAGE_SIZE;
1091 
1092 	return 0;
1093 }
1094 
1095 static int xenvif_tx_submit(struct xenvif_queue *queue)
1096 {
1097 	struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1098 	struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1099 	struct sk_buff *skb;
1100 	int work_done = 0;
1101 
1102 	while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1103 		struct xen_netif_tx_request *txp;
1104 		u16 pending_idx;
1105 		unsigned data_len;
1106 
1107 		pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1108 		txp = &queue->pending_tx_info[pending_idx].req;
1109 
1110 		/* Check the remap error code. */
1111 		if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1112 			/* If there was an error, xenvif_tx_check_gop is
1113 			 * expected to release all the frags which were mapped,
1114 			 * so kfree_skb shouldn't do it again
1115 			 */
1116 			skb_shinfo(skb)->nr_frags = 0;
1117 			if (skb_has_frag_list(skb)) {
1118 				struct sk_buff *nskb =
1119 						skb_shinfo(skb)->frag_list;
1120 				skb_shinfo(nskb)->nr_frags = 0;
1121 			}
1122 			kfree_skb(skb);
1123 			continue;
1124 		}
1125 
1126 		data_len = skb->len;
1127 		callback_param(queue, pending_idx).ctx = NULL;
1128 		if (data_len < txp->size) {
1129 			/* Append the packet payload as a fragment. */
1130 			txp->offset += data_len;
1131 			txp->size -= data_len;
1132 		} else {
1133 			/* Schedule a response immediately. */
1134 			xenvif_idx_release(queue, pending_idx,
1135 					   XEN_NETIF_RSP_OKAY);
1136 		}
1137 
1138 		if (txp->flags & XEN_NETTXF_csum_blank)
1139 			skb->ip_summed = CHECKSUM_PARTIAL;
1140 		else if (txp->flags & XEN_NETTXF_data_validated)
1141 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1142 
1143 		xenvif_fill_frags(queue, skb);
1144 
1145 		if (unlikely(skb_has_frag_list(skb))) {
1146 			if (xenvif_handle_frag_list(queue, skb)) {
1147 				if (net_ratelimit())
1148 					netdev_err(queue->vif->dev,
1149 						   "Not enough memory to consolidate frag_list!\n");
1150 				xenvif_skb_zerocopy_prepare(queue, skb);
1151 				kfree_skb(skb);
1152 				continue;
1153 			}
1154 		}
1155 
1156 		skb->dev      = queue->vif->dev;
1157 		skb->protocol = eth_type_trans(skb, skb->dev);
1158 		skb_reset_network_header(skb);
1159 
1160 		if (checksum_setup(queue, skb)) {
1161 			netdev_dbg(queue->vif->dev,
1162 				   "Can't setup checksum in net_tx_action\n");
1163 			/* We have to set this flag to trigger the callback */
1164 			if (skb_shinfo(skb)->destructor_arg)
1165 				xenvif_skb_zerocopy_prepare(queue, skb);
1166 			kfree_skb(skb);
1167 			continue;
1168 		}
1169 
1170 		skb_probe_transport_header(skb, 0);
1171 
1172 		/* If the packet is GSO then we will have just set up the
1173 		 * transport header offset in checksum_setup so it's now
1174 		 * straightforward to calculate gso_segs.
1175 		 */
1176 		if (skb_is_gso(skb)) {
1177 			int mss = skb_shinfo(skb)->gso_size;
1178 			int hdrlen = skb_transport_header(skb) -
1179 				skb_mac_header(skb) +
1180 				tcp_hdrlen(skb);
1181 
1182 			skb_shinfo(skb)->gso_segs =
1183 				DIV_ROUND_UP(skb->len - hdrlen, mss);
1184 		}
1185 
1186 		queue->stats.rx_bytes += skb->len;
1187 		queue->stats.rx_packets++;
1188 
1189 		work_done++;
1190 
1191 		/* Set this flag right before netif_receive_skb, otherwise
1192 		 * someone might think this packet already left netback, and
1193 		 * do a skb_copy_ubufs while we are still in control of the
1194 		 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1195 		 */
1196 		if (skb_shinfo(skb)->destructor_arg) {
1197 			xenvif_skb_zerocopy_prepare(queue, skb);
1198 			queue->stats.tx_zerocopy_sent++;
1199 		}
1200 
1201 		netif_receive_skb(skb);
1202 	}
1203 
1204 	return work_done;
1205 }
1206 
1207 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1208 {
1209 	unsigned long flags;
1210 	pending_ring_idx_t index;
1211 	struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1212 
1213 	/* This is the only place where we grab this lock, to protect callbacks
1214 	 * from each other.
1215 	 */
1216 	spin_lock_irqsave(&queue->callback_lock, flags);
1217 	do {
1218 		u16 pending_idx = ubuf->desc;
1219 		ubuf = (struct ubuf_info *) ubuf->ctx;
1220 		BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1221 			MAX_PENDING_REQS);
1222 		index = pending_index(queue->dealloc_prod);
1223 		queue->dealloc_ring[index] = pending_idx;
1224 		/* Sync with xenvif_tx_dealloc_action:
1225 		 * insert idx then incr producer.
1226 		 */
1227 		smp_wmb();
1228 		queue->dealloc_prod++;
1229 	} while (ubuf);
1230 	spin_unlock_irqrestore(&queue->callback_lock, flags);
1231 
1232 	if (likely(zerocopy_success))
1233 		queue->stats.tx_zerocopy_success++;
1234 	else
1235 		queue->stats.tx_zerocopy_fail++;
1236 	xenvif_skb_zerocopy_complete(queue);
1237 }
1238 
1239 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1240 {
1241 	struct gnttab_unmap_grant_ref *gop;
1242 	pending_ring_idx_t dc, dp;
1243 	u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1244 	unsigned int i = 0;
1245 
1246 	dc = queue->dealloc_cons;
1247 	gop = queue->tx_unmap_ops;
1248 
1249 	/* Free up any grants we have finished using */
1250 	do {
1251 		dp = queue->dealloc_prod;
1252 
1253 		/* Ensure we see all indices enqueued by all
1254 		 * xenvif_zerocopy_callback().
1255 		 */
1256 		smp_rmb();
1257 
1258 		while (dc != dp) {
1259 			BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1260 			pending_idx =
1261 				queue->dealloc_ring[pending_index(dc++)];
1262 
1263 			pending_idx_release[gop - queue->tx_unmap_ops] =
1264 				pending_idx;
1265 			queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1266 				queue->mmap_pages[pending_idx];
1267 			gnttab_set_unmap_op(gop,
1268 					    idx_to_kaddr(queue, pending_idx),
1269 					    GNTMAP_host_map,
1270 					    queue->grant_tx_handle[pending_idx]);
1271 			xenvif_grant_handle_reset(queue, pending_idx);
1272 			++gop;
1273 		}
1274 
1275 	} while (dp != queue->dealloc_prod);
1276 
1277 	queue->dealloc_cons = dc;
1278 
1279 	if (gop - queue->tx_unmap_ops > 0) {
1280 		int ret;
1281 		ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1282 					NULL,
1283 					queue->pages_to_unmap,
1284 					gop - queue->tx_unmap_ops);
1285 		if (ret) {
1286 			netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1287 				   gop - queue->tx_unmap_ops, ret);
1288 			for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1289 				if (gop[i].status != GNTST_okay)
1290 					netdev_err(queue->vif->dev,
1291 						   " host_addr: 0x%llx handle: 0x%x status: %d\n",
1292 						   gop[i].host_addr,
1293 						   gop[i].handle,
1294 						   gop[i].status);
1295 			}
1296 			BUG();
1297 		}
1298 	}
1299 
1300 	for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1301 		xenvif_idx_release(queue, pending_idx_release[i],
1302 				   XEN_NETIF_RSP_OKAY);
1303 }
1304 
1305 
1306 /* Called after netfront has transmitted */
1307 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1308 {
1309 	unsigned nr_mops, nr_cops = 0;
1310 	int work_done, ret;
1311 
1312 	if (unlikely(!tx_work_todo(queue)))
1313 		return 0;
1314 
1315 	xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1316 
1317 	if (nr_cops == 0)
1318 		return 0;
1319 
1320 	gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1321 	if (nr_mops != 0) {
1322 		ret = gnttab_map_refs(queue->tx_map_ops,
1323 				      NULL,
1324 				      queue->pages_to_map,
1325 				      nr_mops);
1326 		BUG_ON(ret);
1327 	}
1328 
1329 	work_done = xenvif_tx_submit(queue);
1330 
1331 	return work_done;
1332 }
1333 
1334 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1335 			       u8 status)
1336 {
1337 	struct pending_tx_info *pending_tx_info;
1338 	pending_ring_idx_t index;
1339 	unsigned long flags;
1340 
1341 	pending_tx_info = &queue->pending_tx_info[pending_idx];
1342 
1343 	spin_lock_irqsave(&queue->response_lock, flags);
1344 
1345 	make_tx_response(queue, &pending_tx_info->req,
1346 			 pending_tx_info->extra_count, status);
1347 
1348 	/* Release the pending index before pusing the Tx response so
1349 	 * its available before a new Tx request is pushed by the
1350 	 * frontend.
1351 	 */
1352 	index = pending_index(queue->pending_prod++);
1353 	queue->pending_ring[index] = pending_idx;
1354 
1355 	push_tx_responses(queue);
1356 
1357 	spin_unlock_irqrestore(&queue->response_lock, flags);
1358 }
1359 
1360 
1361 static void make_tx_response(struct xenvif_queue *queue,
1362 			     struct xen_netif_tx_request *txp,
1363 			     unsigned int extra_count,
1364 			     s8       st)
1365 {
1366 	RING_IDX i = queue->tx.rsp_prod_pvt;
1367 	struct xen_netif_tx_response *resp;
1368 
1369 	resp = RING_GET_RESPONSE(&queue->tx, i);
1370 	resp->id     = txp->id;
1371 	resp->status = st;
1372 
1373 	while (extra_count-- != 0)
1374 		RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1375 
1376 	queue->tx.rsp_prod_pvt = ++i;
1377 }
1378 
1379 static void push_tx_responses(struct xenvif_queue *queue)
1380 {
1381 	int notify;
1382 
1383 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1384 	if (notify)
1385 		notify_remote_via_irq(queue->tx_irq);
1386 }
1387 
1388 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1389 {
1390 	int ret;
1391 	struct gnttab_unmap_grant_ref tx_unmap_op;
1392 
1393 	gnttab_set_unmap_op(&tx_unmap_op,
1394 			    idx_to_kaddr(queue, pending_idx),
1395 			    GNTMAP_host_map,
1396 			    queue->grant_tx_handle[pending_idx]);
1397 	xenvif_grant_handle_reset(queue, pending_idx);
1398 
1399 	ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1400 				&queue->mmap_pages[pending_idx], 1);
1401 	if (ret) {
1402 		netdev_err(queue->vif->dev,
1403 			   "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1404 			   ret,
1405 			   pending_idx,
1406 			   tx_unmap_op.host_addr,
1407 			   tx_unmap_op.handle,
1408 			   tx_unmap_op.status);
1409 		BUG();
1410 	}
1411 }
1412 
1413 static inline int tx_work_todo(struct xenvif_queue *queue)
1414 {
1415 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1416 		return 1;
1417 
1418 	return 0;
1419 }
1420 
1421 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1422 {
1423 	return queue->dealloc_cons != queue->dealloc_prod;
1424 }
1425 
1426 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1427 {
1428 	if (queue->tx.sring)
1429 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1430 					queue->tx.sring);
1431 	if (queue->rx.sring)
1432 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1433 					queue->rx.sring);
1434 }
1435 
1436 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1437 				   grant_ref_t tx_ring_ref,
1438 				   grant_ref_t rx_ring_ref)
1439 {
1440 	void *addr;
1441 	struct xen_netif_tx_sring *txs;
1442 	struct xen_netif_rx_sring *rxs;
1443 
1444 	int err = -ENOMEM;
1445 
1446 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1447 				     &tx_ring_ref, 1, &addr);
1448 	if (err)
1449 		goto err;
1450 
1451 	txs = (struct xen_netif_tx_sring *)addr;
1452 	BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1453 
1454 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1455 				     &rx_ring_ref, 1, &addr);
1456 	if (err)
1457 		goto err;
1458 
1459 	rxs = (struct xen_netif_rx_sring *)addr;
1460 	BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1461 
1462 	return 0;
1463 
1464 err:
1465 	xenvif_unmap_frontend_data_rings(queue);
1466 	return err;
1467 }
1468 
1469 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1470 {
1471 	/* Dealloc thread must remain running until all inflight
1472 	 * packets complete.
1473 	 */
1474 	return kthread_should_stop() &&
1475 		!atomic_read(&queue->inflight_packets);
1476 }
1477 
1478 int xenvif_dealloc_kthread(void *data)
1479 {
1480 	struct xenvif_queue *queue = data;
1481 
1482 	for (;;) {
1483 		wait_event_interruptible(queue->dealloc_wq,
1484 					 tx_dealloc_work_todo(queue) ||
1485 					 xenvif_dealloc_kthread_should_stop(queue));
1486 		if (xenvif_dealloc_kthread_should_stop(queue))
1487 			break;
1488 
1489 		xenvif_tx_dealloc_action(queue);
1490 		cond_resched();
1491 	}
1492 
1493 	/* Unmap anything remaining*/
1494 	if (tx_dealloc_work_todo(queue))
1495 		xenvif_tx_dealloc_action(queue);
1496 
1497 	return 0;
1498 }
1499 
1500 static void make_ctrl_response(struct xenvif *vif,
1501 			       const struct xen_netif_ctrl_request *req,
1502 			       u32 status, u32 data)
1503 {
1504 	RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1505 	struct xen_netif_ctrl_response rsp = {
1506 		.id = req->id,
1507 		.type = req->type,
1508 		.status = status,
1509 		.data = data,
1510 	};
1511 
1512 	*RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1513 	vif->ctrl.rsp_prod_pvt = ++idx;
1514 }
1515 
1516 static void push_ctrl_response(struct xenvif *vif)
1517 {
1518 	int notify;
1519 
1520 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1521 	if (notify)
1522 		notify_remote_via_irq(vif->ctrl_irq);
1523 }
1524 
1525 static void process_ctrl_request(struct xenvif *vif,
1526 				 const struct xen_netif_ctrl_request *req)
1527 {
1528 	u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1529 	u32 data = 0;
1530 
1531 	switch (req->type) {
1532 	case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1533 		status = xenvif_set_hash_alg(vif, req->data[0]);
1534 		break;
1535 
1536 	case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1537 		status = xenvif_get_hash_flags(vif, &data);
1538 		break;
1539 
1540 	case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1541 		status = xenvif_set_hash_flags(vif, req->data[0]);
1542 		break;
1543 
1544 	case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1545 		status = xenvif_set_hash_key(vif, req->data[0],
1546 					     req->data[1]);
1547 		break;
1548 
1549 	case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1550 		status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1551 		data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1552 		break;
1553 
1554 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1555 		status = xenvif_set_hash_mapping_size(vif,
1556 						      req->data[0]);
1557 		break;
1558 
1559 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1560 		status = xenvif_set_hash_mapping(vif, req->data[0],
1561 						 req->data[1],
1562 						 req->data[2]);
1563 		break;
1564 
1565 	default:
1566 		break;
1567 	}
1568 
1569 	make_ctrl_response(vif, req, status, data);
1570 	push_ctrl_response(vif);
1571 }
1572 
1573 static void xenvif_ctrl_action(struct xenvif *vif)
1574 {
1575 	for (;;) {
1576 		RING_IDX req_prod, req_cons;
1577 
1578 		req_prod = vif->ctrl.sring->req_prod;
1579 		req_cons = vif->ctrl.req_cons;
1580 
1581 		/* Make sure we can see requests before we process them. */
1582 		rmb();
1583 
1584 		if (req_cons == req_prod)
1585 			break;
1586 
1587 		while (req_cons != req_prod) {
1588 			struct xen_netif_ctrl_request req;
1589 
1590 			RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1591 			req_cons++;
1592 
1593 			process_ctrl_request(vif, &req);
1594 		}
1595 
1596 		vif->ctrl.req_cons = req_cons;
1597 		vif->ctrl.sring->req_event = req_cons + 1;
1598 	}
1599 }
1600 
1601 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1602 {
1603 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1604 		return 1;
1605 
1606 	return 0;
1607 }
1608 
1609 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1610 {
1611 	struct xenvif *vif = data;
1612 
1613 	while (xenvif_ctrl_work_todo(vif))
1614 		xenvif_ctrl_action(vif);
1615 
1616 	return IRQ_HANDLED;
1617 }
1618 
1619 static int __init netback_init(void)
1620 {
1621 	int rc = 0;
1622 
1623 	if (!xen_domain())
1624 		return -ENODEV;
1625 
1626 	/* Allow as many queues as there are CPUs but max. 8 if user has not
1627 	 * specified a value.
1628 	 */
1629 	if (xenvif_max_queues == 0)
1630 		xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1631 					  num_online_cpus());
1632 
1633 	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1634 		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1635 			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1636 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1637 	}
1638 
1639 	rc = xenvif_xenbus_init();
1640 	if (rc)
1641 		goto failed_init;
1642 
1643 #ifdef CONFIG_DEBUG_FS
1644 	xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1645 	if (IS_ERR_OR_NULL(xen_netback_dbg_root))
1646 		pr_warn("Init of debugfs returned %ld!\n",
1647 			PTR_ERR(xen_netback_dbg_root));
1648 #endif /* CONFIG_DEBUG_FS */
1649 
1650 	return 0;
1651 
1652 failed_init:
1653 	return rc;
1654 }
1655 
1656 module_init(netback_init);
1657 
1658 static void __exit netback_fini(void)
1659 {
1660 #ifdef CONFIG_DEBUG_FS
1661 	if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
1662 		debugfs_remove_recursive(xen_netback_dbg_root);
1663 #endif /* CONFIG_DEBUG_FS */
1664 	xenvif_xenbus_fini();
1665 }
1666 module_exit(netback_fini);
1667 
1668 MODULE_LICENSE("Dual BSD/GPL");
1669 MODULE_ALIAS("xen-backend:vif");
1670