xref: /linux/drivers/net/xen-netback/netback.c (revision 1b0975ee3bdd3eb19a47371c26fd7ef8f7f6b599)
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 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
116 
117 static inline int tx_work_todo(struct xenvif_queue *queue);
118 
119 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
120 				       u16 idx)
121 {
122 	return page_to_pfn(queue->mmap_pages[idx]);
123 }
124 
125 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
126 					 u16 idx)
127 {
128 	return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
129 }
130 
131 #define callback_param(vif, pending_idx) \
132 	(vif->pending_tx_info[pending_idx].callback_struct)
133 
134 /* Find the containing VIF's structure from a pointer in pending_tx_info array
135  */
136 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info_msgzc *ubuf)
137 {
138 	u16 pending_idx = ubuf->desc;
139 	struct pending_tx_info *temp =
140 		container_of(ubuf, struct pending_tx_info, callback_struct);
141 	return container_of(temp - pending_idx,
142 			    struct xenvif_queue,
143 			    pending_tx_info[0]);
144 }
145 
146 static u16 frag_get_pending_idx(skb_frag_t *frag)
147 {
148 	return (u16)skb_frag_off(frag);
149 }
150 
151 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
152 {
153 	skb_frag_off_set(frag, pending_idx);
154 }
155 
156 static inline pending_ring_idx_t pending_index(unsigned i)
157 {
158 	return i & (MAX_PENDING_REQS-1);
159 }
160 
161 void xenvif_kick_thread(struct xenvif_queue *queue)
162 {
163 	wake_up(&queue->wq);
164 }
165 
166 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
167 {
168 	int more_to_do;
169 
170 	RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
171 
172 	if (more_to_do)
173 		napi_schedule(&queue->napi);
174 	else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
175 				     &queue->eoi_pending) &
176 		 (NETBK_TX_EOI | NETBK_COMMON_EOI))
177 		xen_irq_lateeoi(queue->tx_irq, 0);
178 }
179 
180 static void tx_add_credit(struct xenvif_queue *queue)
181 {
182 	unsigned long max_burst, max_credit;
183 
184 	/*
185 	 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
186 	 * Otherwise the interface can seize up due to insufficient credit.
187 	 */
188 	max_burst = max(131072UL, queue->credit_bytes);
189 
190 	/* Take care that adding a new chunk of credit doesn't wrap to zero. */
191 	max_credit = queue->remaining_credit + queue->credit_bytes;
192 	if (max_credit < queue->remaining_credit)
193 		max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
194 
195 	queue->remaining_credit = min(max_credit, max_burst);
196 	queue->rate_limited = false;
197 }
198 
199 void xenvif_tx_credit_callback(struct timer_list *t)
200 {
201 	struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
202 	tx_add_credit(queue);
203 	xenvif_napi_schedule_or_enable_events(queue);
204 }
205 
206 static void xenvif_tx_err(struct xenvif_queue *queue,
207 			  struct xen_netif_tx_request *txp,
208 			  unsigned int extra_count, RING_IDX end)
209 {
210 	RING_IDX cons = queue->tx.req_cons;
211 	unsigned long flags;
212 
213 	do {
214 		spin_lock_irqsave(&queue->response_lock, flags);
215 		make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
216 		push_tx_responses(queue);
217 		spin_unlock_irqrestore(&queue->response_lock, flags);
218 		if (cons == end)
219 			break;
220 		RING_COPY_REQUEST(&queue->tx, cons++, txp);
221 		extra_count = 0; /* only the first frag can have extras */
222 	} while (1);
223 	queue->tx.req_cons = cons;
224 }
225 
226 static void xenvif_fatal_tx_err(struct xenvif *vif)
227 {
228 	netdev_err(vif->dev, "fatal error; disabling device\n");
229 	vif->disabled = true;
230 	/* Disable the vif from queue 0's kthread */
231 	if (vif->num_queues)
232 		xenvif_kick_thread(&vif->queues[0]);
233 }
234 
235 static int xenvif_count_requests(struct xenvif_queue *queue,
236 				 struct xen_netif_tx_request *first,
237 				 unsigned int extra_count,
238 				 struct xen_netif_tx_request *txp,
239 				 int work_to_do)
240 {
241 	RING_IDX cons = queue->tx.req_cons;
242 	int slots = 0;
243 	int drop_err = 0;
244 	int more_data;
245 
246 	if (!(first->flags & XEN_NETTXF_more_data))
247 		return 0;
248 
249 	do {
250 		struct xen_netif_tx_request dropped_tx = { 0 };
251 
252 		if (slots >= work_to_do) {
253 			netdev_err(queue->vif->dev,
254 				   "Asked for %d slots but exceeds this limit\n",
255 				   work_to_do);
256 			xenvif_fatal_tx_err(queue->vif);
257 			return -ENODATA;
258 		}
259 
260 		/* This guest is really using too many slots and
261 		 * considered malicious.
262 		 */
263 		if (unlikely(slots >= fatal_skb_slots)) {
264 			netdev_err(queue->vif->dev,
265 				   "Malicious frontend using %d slots, threshold %u\n",
266 				   slots, fatal_skb_slots);
267 			xenvif_fatal_tx_err(queue->vif);
268 			return -E2BIG;
269 		}
270 
271 		/* Xen network protocol had implicit dependency on
272 		 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
273 		 * the historical MAX_SKB_FRAGS value 18 to honor the
274 		 * same behavior as before. Any packet using more than
275 		 * 18 slots but less than fatal_skb_slots slots is
276 		 * dropped
277 		 */
278 		if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
279 			if (net_ratelimit())
280 				netdev_dbg(queue->vif->dev,
281 					   "Too many slots (%d) exceeding limit (%d), dropping packet\n",
282 					   slots, XEN_NETBK_LEGACY_SLOTS_MAX);
283 			drop_err = -E2BIG;
284 		}
285 
286 		if (drop_err)
287 			txp = &dropped_tx;
288 
289 		RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
290 
291 		/* If the guest submitted a frame >= 64 KiB then
292 		 * first->size overflowed and following slots will
293 		 * appear to be larger than the frame.
294 		 *
295 		 * This cannot be fatal error as there are buggy
296 		 * frontends that do this.
297 		 *
298 		 * Consume all slots and drop the packet.
299 		 */
300 		if (!drop_err && txp->size > first->size) {
301 			if (net_ratelimit())
302 				netdev_dbg(queue->vif->dev,
303 					   "Invalid tx request, slot size %u > remaining size %u\n",
304 					   txp->size, first->size);
305 			drop_err = -EIO;
306 		}
307 
308 		first->size -= txp->size;
309 		slots++;
310 
311 		if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
312 			netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
313 				 txp->offset, txp->size);
314 			xenvif_fatal_tx_err(queue->vif);
315 			return -EINVAL;
316 		}
317 
318 		more_data = txp->flags & XEN_NETTXF_more_data;
319 
320 		if (!drop_err)
321 			txp++;
322 
323 	} while (more_data);
324 
325 	if (drop_err) {
326 		xenvif_tx_err(queue, first, extra_count, cons + slots);
327 		return drop_err;
328 	}
329 
330 	return slots;
331 }
332 
333 
334 struct xenvif_tx_cb {
335 	u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
336 	u8 copy_count;
337 	u32 split_mask;
338 };
339 
340 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
341 #define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
342 #define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
343 
344 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
345 					   u16 pending_idx,
346 					   struct xen_netif_tx_request *txp,
347 					   unsigned int extra_count,
348 					   struct gnttab_map_grant_ref *mop)
349 {
350 	queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
351 	gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
352 			  GNTMAP_host_map | GNTMAP_readonly,
353 			  txp->gref, queue->vif->domid);
354 
355 	memcpy(&queue->pending_tx_info[pending_idx].req, txp,
356 	       sizeof(*txp));
357 	queue->pending_tx_info[pending_idx].extra_count = extra_count;
358 }
359 
360 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
361 {
362 	struct sk_buff *skb =
363 		alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
364 			  GFP_ATOMIC | __GFP_NOWARN);
365 
366 	BUILD_BUG_ON(sizeof(*XENVIF_TX_CB(skb)) > sizeof(skb->cb));
367 	if (unlikely(skb == NULL))
368 		return NULL;
369 
370 	/* Packets passed to netif_rx() must have some headroom. */
371 	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
372 
373 	/* Initialize it here to avoid later surprises */
374 	skb_shinfo(skb)->destructor_arg = NULL;
375 
376 	return skb;
377 }
378 
379 static void xenvif_get_requests(struct xenvif_queue *queue,
380 				struct sk_buff *skb,
381 				struct xen_netif_tx_request *first,
382 				struct xen_netif_tx_request *txfrags,
383 			        unsigned *copy_ops,
384 			        unsigned *map_ops,
385 				unsigned int frag_overflow,
386 				struct sk_buff *nskb,
387 				unsigned int extra_count,
388 				unsigned int data_len)
389 {
390 	struct skb_shared_info *shinfo = skb_shinfo(skb);
391 	skb_frag_t *frags = shinfo->frags;
392 	u16 pending_idx;
393 	pending_ring_idx_t index;
394 	unsigned int nr_slots;
395 	struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
396 	struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
397 	struct xen_netif_tx_request *txp = first;
398 
399 	nr_slots = shinfo->nr_frags + 1;
400 
401 	copy_count(skb) = 0;
402 	XENVIF_TX_CB(skb)->split_mask = 0;
403 
404 	/* Create copy ops for exactly data_len bytes into the skb head. */
405 	__skb_put(skb, data_len);
406 	while (data_len > 0) {
407 		int amount = data_len > txp->size ? txp->size : data_len;
408 		bool split = false;
409 
410 		cop->source.u.ref = txp->gref;
411 		cop->source.domid = queue->vif->domid;
412 		cop->source.offset = txp->offset;
413 
414 		cop->dest.domid = DOMID_SELF;
415 		cop->dest.offset = (offset_in_page(skb->data +
416 						   skb_headlen(skb) -
417 						   data_len)) & ~XEN_PAGE_MASK;
418 		cop->dest.u.gmfn = virt_to_gfn(skb->data + skb_headlen(skb)
419 				               - data_len);
420 
421 		/* Don't cross local page boundary! */
422 		if (cop->dest.offset + amount > XEN_PAGE_SIZE) {
423 			amount = XEN_PAGE_SIZE - cop->dest.offset;
424 			XENVIF_TX_CB(skb)->split_mask |= 1U << copy_count(skb);
425 			split = true;
426 		}
427 
428 		cop->len = amount;
429 		cop->flags = GNTCOPY_source_gref;
430 
431 		index = pending_index(queue->pending_cons);
432 		pending_idx = queue->pending_ring[index];
433 		callback_param(queue, pending_idx).ctx = NULL;
434 		copy_pending_idx(skb, copy_count(skb)) = pending_idx;
435 		if (!split)
436 			copy_count(skb)++;
437 
438 		cop++;
439 		data_len -= amount;
440 
441 		if (amount == txp->size) {
442 			/* The copy op covered the full tx_request */
443 
444 			memcpy(&queue->pending_tx_info[pending_idx].req,
445 			       txp, sizeof(*txp));
446 			queue->pending_tx_info[pending_idx].extra_count =
447 				(txp == first) ? extra_count : 0;
448 
449 			if (txp == first)
450 				txp = txfrags;
451 			else
452 				txp++;
453 			queue->pending_cons++;
454 			nr_slots--;
455 		} else {
456 			/* The copy op partially covered the tx_request.
457 			 * The remainder will be mapped or copied in the next
458 			 * iteration.
459 			 */
460 			txp->offset += amount;
461 			txp->size -= amount;
462 		}
463 	}
464 
465 	for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots;
466 	     shinfo->nr_frags++, gop++) {
467 		index = pending_index(queue->pending_cons++);
468 		pending_idx = queue->pending_ring[index];
469 		xenvif_tx_create_map_op(queue, pending_idx, txp,
470 				        txp == first ? extra_count : 0, gop);
471 		frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
472 
473 		if (txp == first)
474 			txp = txfrags;
475 		else
476 			txp++;
477 	}
478 
479 	if (frag_overflow) {
480 
481 		shinfo = skb_shinfo(nskb);
482 		frags = shinfo->frags;
483 
484 		for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
485 		     shinfo->nr_frags++, txp++, gop++) {
486 			index = pending_index(queue->pending_cons++);
487 			pending_idx = queue->pending_ring[index];
488 			xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
489 						gop);
490 			frag_set_pending_idx(&frags[shinfo->nr_frags],
491 					     pending_idx);
492 		}
493 
494 		skb_shinfo(skb)->frag_list = nskb;
495 	}
496 
497 	(*copy_ops) = cop - queue->tx_copy_ops;
498 	(*map_ops) = gop - queue->tx_map_ops;
499 }
500 
501 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
502 					   u16 pending_idx,
503 					   grant_handle_t handle)
504 {
505 	if (unlikely(queue->grant_tx_handle[pending_idx] !=
506 		     NETBACK_INVALID_HANDLE)) {
507 		netdev_err(queue->vif->dev,
508 			   "Trying to overwrite active handle! pending_idx: 0x%x\n",
509 			   pending_idx);
510 		BUG();
511 	}
512 	queue->grant_tx_handle[pending_idx] = handle;
513 }
514 
515 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
516 					     u16 pending_idx)
517 {
518 	if (unlikely(queue->grant_tx_handle[pending_idx] ==
519 		     NETBACK_INVALID_HANDLE)) {
520 		netdev_err(queue->vif->dev,
521 			   "Trying to unmap invalid handle! pending_idx: 0x%x\n",
522 			   pending_idx);
523 		BUG();
524 	}
525 	queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
526 }
527 
528 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
529 			       struct sk_buff *skb,
530 			       struct gnttab_map_grant_ref **gopp_map,
531 			       struct gnttab_copy **gopp_copy)
532 {
533 	struct gnttab_map_grant_ref *gop_map = *gopp_map;
534 	u16 pending_idx;
535 	/* This always points to the shinfo of the skb being checked, which
536 	 * could be either the first or the one on the frag_list
537 	 */
538 	struct skb_shared_info *shinfo = skb_shinfo(skb);
539 	/* If this is non-NULL, we are currently checking the frag_list skb, and
540 	 * this points to the shinfo of the first one
541 	 */
542 	struct skb_shared_info *first_shinfo = NULL;
543 	int nr_frags = shinfo->nr_frags;
544 	const bool sharedslot = nr_frags &&
545 				frag_get_pending_idx(&shinfo->frags[0]) ==
546 				    copy_pending_idx(skb, copy_count(skb) - 1);
547 	int i, err = 0;
548 
549 	for (i = 0; i < copy_count(skb); i++) {
550 		int newerr;
551 
552 		/* Check status of header. */
553 		pending_idx = copy_pending_idx(skb, i);
554 
555 		newerr = (*gopp_copy)->status;
556 
557 		/* Split copies need to be handled together. */
558 		if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
559 			(*gopp_copy)++;
560 			if (!newerr)
561 				newerr = (*gopp_copy)->status;
562 		}
563 		if (likely(!newerr)) {
564 			/* The first frag might still have this slot mapped */
565 			if (i < copy_count(skb) - 1 || !sharedslot)
566 				xenvif_idx_release(queue, pending_idx,
567 						   XEN_NETIF_RSP_OKAY);
568 		} else {
569 			err = newerr;
570 			if (net_ratelimit())
571 				netdev_dbg(queue->vif->dev,
572 					   "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
573 					   (*gopp_copy)->status,
574 					   pending_idx,
575 					   (*gopp_copy)->source.u.ref);
576 			/* The first frag might still have this slot mapped */
577 			if (i < copy_count(skb) - 1 || !sharedslot)
578 				xenvif_idx_release(queue, pending_idx,
579 						   XEN_NETIF_RSP_ERROR);
580 		}
581 		(*gopp_copy)++;
582 	}
583 
584 check_frags:
585 	for (i = 0; i < nr_frags; i++, gop_map++) {
586 		int j, newerr;
587 
588 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
589 
590 		/* Check error status: if okay then remember grant handle. */
591 		newerr = gop_map->status;
592 
593 		if (likely(!newerr)) {
594 			xenvif_grant_handle_set(queue,
595 						pending_idx,
596 						gop_map->handle);
597 			/* Had a previous error? Invalidate this fragment. */
598 			if (unlikely(err)) {
599 				xenvif_idx_unmap(queue, pending_idx);
600 				/* If the mapping of the first frag was OK, but
601 				 * the header's copy failed, and they are
602 				 * sharing a slot, send an error
603 				 */
604 				if (i == 0 && !first_shinfo && sharedslot)
605 					xenvif_idx_release(queue, pending_idx,
606 							   XEN_NETIF_RSP_ERROR);
607 				else
608 					xenvif_idx_release(queue, pending_idx,
609 							   XEN_NETIF_RSP_OKAY);
610 			}
611 			continue;
612 		}
613 
614 		/* Error on this fragment: respond to client with an error. */
615 		if (net_ratelimit())
616 			netdev_dbg(queue->vif->dev,
617 				   "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
618 				   i,
619 				   gop_map->status,
620 				   pending_idx,
621 				   gop_map->ref);
622 
623 		xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
624 
625 		/* Not the first error? Preceding frags already invalidated. */
626 		if (err)
627 			continue;
628 
629 		/* Invalidate preceding fragments of this skb. */
630 		for (j = 0; j < i; j++) {
631 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
632 			xenvif_idx_unmap(queue, pending_idx);
633 			xenvif_idx_release(queue, pending_idx,
634 					   XEN_NETIF_RSP_OKAY);
635 		}
636 
637 		/* And if we found the error while checking the frag_list, unmap
638 		 * the first skb's frags
639 		 */
640 		if (first_shinfo) {
641 			for (j = 0; j < first_shinfo->nr_frags; j++) {
642 				pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
643 				xenvif_idx_unmap(queue, pending_idx);
644 				xenvif_idx_release(queue, pending_idx,
645 						   XEN_NETIF_RSP_OKAY);
646 			}
647 		}
648 
649 		/* Remember the error: invalidate all subsequent fragments. */
650 		err = newerr;
651 	}
652 
653 	if (skb_has_frag_list(skb) && !first_shinfo) {
654 		first_shinfo = shinfo;
655 		shinfo = skb_shinfo(shinfo->frag_list);
656 		nr_frags = shinfo->nr_frags;
657 
658 		goto check_frags;
659 	}
660 
661 	*gopp_map = gop_map;
662 	return err;
663 }
664 
665 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
666 {
667 	struct skb_shared_info *shinfo = skb_shinfo(skb);
668 	int nr_frags = shinfo->nr_frags;
669 	int i;
670 	u16 prev_pending_idx = INVALID_PENDING_IDX;
671 
672 	for (i = 0; i < nr_frags; i++) {
673 		skb_frag_t *frag = shinfo->frags + i;
674 		struct xen_netif_tx_request *txp;
675 		struct page *page;
676 		u16 pending_idx;
677 
678 		pending_idx = frag_get_pending_idx(frag);
679 
680 		/* If this is not the first frag, chain it to the previous*/
681 		if (prev_pending_idx == INVALID_PENDING_IDX)
682 			skb_shinfo(skb)->destructor_arg =
683 				&callback_param(queue, pending_idx);
684 		else
685 			callback_param(queue, prev_pending_idx).ctx =
686 				&callback_param(queue, pending_idx);
687 
688 		callback_param(queue, pending_idx).ctx = NULL;
689 		prev_pending_idx = pending_idx;
690 
691 		txp = &queue->pending_tx_info[pending_idx].req;
692 		page = virt_to_page((void *)idx_to_kaddr(queue, pending_idx));
693 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
694 		skb->len += txp->size;
695 		skb->data_len += txp->size;
696 		skb->truesize += txp->size;
697 
698 		/* Take an extra reference to offset network stack's put_page */
699 		get_page(queue->mmap_pages[pending_idx]);
700 	}
701 }
702 
703 static int xenvif_get_extras(struct xenvif_queue *queue,
704 			     struct xen_netif_extra_info *extras,
705 			     unsigned int *extra_count,
706 			     int work_to_do)
707 {
708 	struct xen_netif_extra_info extra;
709 	RING_IDX cons = queue->tx.req_cons;
710 
711 	do {
712 		if (unlikely(work_to_do-- <= 0)) {
713 			netdev_err(queue->vif->dev, "Missing extra info\n");
714 			xenvif_fatal_tx_err(queue->vif);
715 			return -EBADR;
716 		}
717 
718 		RING_COPY_REQUEST(&queue->tx, cons, &extra);
719 
720 		queue->tx.req_cons = ++cons;
721 		(*extra_count)++;
722 
723 		if (unlikely(!extra.type ||
724 			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
725 			netdev_err(queue->vif->dev,
726 				   "Invalid extra type: %d\n", extra.type);
727 			xenvif_fatal_tx_err(queue->vif);
728 			return -EINVAL;
729 		}
730 
731 		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
732 	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
733 
734 	return work_to_do;
735 }
736 
737 static int xenvif_set_skb_gso(struct xenvif *vif,
738 			      struct sk_buff *skb,
739 			      struct xen_netif_extra_info *gso)
740 {
741 	if (!gso->u.gso.size) {
742 		netdev_err(vif->dev, "GSO size must not be zero.\n");
743 		xenvif_fatal_tx_err(vif);
744 		return -EINVAL;
745 	}
746 
747 	switch (gso->u.gso.type) {
748 	case XEN_NETIF_GSO_TYPE_TCPV4:
749 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
750 		break;
751 	case XEN_NETIF_GSO_TYPE_TCPV6:
752 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
753 		break;
754 	default:
755 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
756 		xenvif_fatal_tx_err(vif);
757 		return -EINVAL;
758 	}
759 
760 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
761 	/* gso_segs will be calculated later */
762 
763 	return 0;
764 }
765 
766 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
767 {
768 	bool recalculate_partial_csum = false;
769 
770 	/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
771 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
772 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
773 	 * recalculate the partial checksum.
774 	 */
775 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
776 		queue->stats.rx_gso_checksum_fixup++;
777 		skb->ip_summed = CHECKSUM_PARTIAL;
778 		recalculate_partial_csum = true;
779 	}
780 
781 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
782 	if (skb->ip_summed != CHECKSUM_PARTIAL)
783 		return 0;
784 
785 	return skb_checksum_setup(skb, recalculate_partial_csum);
786 }
787 
788 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
789 {
790 	u64 now = get_jiffies_64();
791 	u64 next_credit = queue->credit_window_start +
792 		msecs_to_jiffies(queue->credit_usec / 1000);
793 
794 	/* Timer could already be pending in rare cases. */
795 	if (timer_pending(&queue->credit_timeout)) {
796 		queue->rate_limited = true;
797 		return true;
798 	}
799 
800 	/* Passed the point where we can replenish credit? */
801 	if (time_after_eq64(now, next_credit)) {
802 		queue->credit_window_start = now;
803 		tx_add_credit(queue);
804 	}
805 
806 	/* Still too big to send right now? Set a callback. */
807 	if (size > queue->remaining_credit) {
808 		mod_timer(&queue->credit_timeout,
809 			  next_credit);
810 		queue->credit_window_start = next_credit;
811 		queue->rate_limited = true;
812 
813 		return true;
814 	}
815 
816 	return false;
817 }
818 
819 /* No locking is required in xenvif_mcast_add/del() as they are
820  * only ever invoked from NAPI poll. An RCU list is used because
821  * xenvif_mcast_match() is called asynchronously, during start_xmit.
822  */
823 
824 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
825 {
826 	struct xenvif_mcast_addr *mcast;
827 
828 	if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
829 		if (net_ratelimit())
830 			netdev_err(vif->dev,
831 				   "Too many multicast addresses\n");
832 		return -ENOSPC;
833 	}
834 
835 	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
836 	if (!mcast)
837 		return -ENOMEM;
838 
839 	ether_addr_copy(mcast->addr, addr);
840 	list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
841 	vif->fe_mcast_count++;
842 
843 	return 0;
844 }
845 
846 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
847 {
848 	struct xenvif_mcast_addr *mcast;
849 
850 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
851 		if (ether_addr_equal(addr, mcast->addr)) {
852 			--vif->fe_mcast_count;
853 			list_del_rcu(&mcast->entry);
854 			kfree_rcu(mcast, rcu);
855 			break;
856 		}
857 	}
858 }
859 
860 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
861 {
862 	struct xenvif_mcast_addr *mcast;
863 
864 	rcu_read_lock();
865 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
866 		if (ether_addr_equal(addr, mcast->addr)) {
867 			rcu_read_unlock();
868 			return true;
869 		}
870 	}
871 	rcu_read_unlock();
872 
873 	return false;
874 }
875 
876 void xenvif_mcast_addr_list_free(struct xenvif *vif)
877 {
878 	/* No need for locking or RCU here. NAPI poll and TX queue
879 	 * are stopped.
880 	 */
881 	while (!list_empty(&vif->fe_mcast_addr)) {
882 		struct xenvif_mcast_addr *mcast;
883 
884 		mcast = list_first_entry(&vif->fe_mcast_addr,
885 					 struct xenvif_mcast_addr,
886 					 entry);
887 		--vif->fe_mcast_count;
888 		list_del(&mcast->entry);
889 		kfree(mcast);
890 	}
891 }
892 
893 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
894 				     int budget,
895 				     unsigned *copy_ops,
896 				     unsigned *map_ops)
897 {
898 	struct sk_buff *skb, *nskb;
899 	int ret;
900 	unsigned int frag_overflow;
901 
902 	while (skb_queue_len(&queue->tx_queue) < budget) {
903 		struct xen_netif_tx_request txreq;
904 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
905 		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
906 		unsigned int extra_count;
907 		RING_IDX idx;
908 		int work_to_do;
909 		unsigned int data_len;
910 
911 		if (queue->tx.sring->req_prod - queue->tx.req_cons >
912 		    XEN_NETIF_TX_RING_SIZE) {
913 			netdev_err(queue->vif->dev,
914 				   "Impossible number of requests. "
915 				   "req_prod %d, req_cons %d, size %ld\n",
916 				   queue->tx.sring->req_prod, queue->tx.req_cons,
917 				   XEN_NETIF_TX_RING_SIZE);
918 			xenvif_fatal_tx_err(queue->vif);
919 			break;
920 		}
921 
922 		work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
923 		if (!work_to_do)
924 			break;
925 
926 		idx = queue->tx.req_cons;
927 		rmb(); /* Ensure that we see the request before we copy it. */
928 		RING_COPY_REQUEST(&queue->tx, idx, &txreq);
929 
930 		/* Credit-based scheduling. */
931 		if (txreq.size > queue->remaining_credit &&
932 		    tx_credit_exceeded(queue, txreq.size))
933 			break;
934 
935 		queue->remaining_credit -= txreq.size;
936 
937 		work_to_do--;
938 		queue->tx.req_cons = ++idx;
939 
940 		memset(extras, 0, sizeof(extras));
941 		extra_count = 0;
942 		if (txreq.flags & XEN_NETTXF_extra_info) {
943 			work_to_do = xenvif_get_extras(queue, extras,
944 						       &extra_count,
945 						       work_to_do);
946 			idx = queue->tx.req_cons;
947 			if (unlikely(work_to_do < 0))
948 				break;
949 		}
950 
951 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
952 			struct xen_netif_extra_info *extra;
953 
954 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
955 			ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
956 
957 			make_tx_response(queue, &txreq, extra_count,
958 					 (ret == 0) ?
959 					 XEN_NETIF_RSP_OKAY :
960 					 XEN_NETIF_RSP_ERROR);
961 			push_tx_responses(queue);
962 			continue;
963 		}
964 
965 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
966 			struct xen_netif_extra_info *extra;
967 
968 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
969 			xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
970 
971 			make_tx_response(queue, &txreq, extra_count,
972 					 XEN_NETIF_RSP_OKAY);
973 			push_tx_responses(queue);
974 			continue;
975 		}
976 
977 		data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
978 			XEN_NETBACK_TX_COPY_LEN : txreq.size;
979 
980 		ret = xenvif_count_requests(queue, &txreq, extra_count,
981 					    txfrags, work_to_do);
982 
983 		if (unlikely(ret < 0))
984 			break;
985 
986 		idx += ret;
987 
988 		if (unlikely(txreq.size < ETH_HLEN)) {
989 			netdev_dbg(queue->vif->dev,
990 				   "Bad packet size: %d\n", txreq.size);
991 			xenvif_tx_err(queue, &txreq, extra_count, idx);
992 			break;
993 		}
994 
995 		/* No crossing a page as the payload mustn't fragment. */
996 		if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
997 			netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
998 				   txreq.offset, txreq.size);
999 			xenvif_fatal_tx_err(queue->vif);
1000 			break;
1001 		}
1002 
1003 		if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1004 			data_len = txreq.size;
1005 
1006 		skb = xenvif_alloc_skb(data_len);
1007 		if (unlikely(skb == NULL)) {
1008 			netdev_dbg(queue->vif->dev,
1009 				   "Can't allocate a skb in start_xmit.\n");
1010 			xenvif_tx_err(queue, &txreq, extra_count, idx);
1011 			break;
1012 		}
1013 
1014 		skb_shinfo(skb)->nr_frags = ret;
1015 		/* At this point shinfo->nr_frags is in fact the number of
1016 		 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1017 		 */
1018 		frag_overflow = 0;
1019 		nskb = NULL;
1020 		if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1021 			frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1022 			BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1023 			skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1024 			nskb = xenvif_alloc_skb(0);
1025 			if (unlikely(nskb == NULL)) {
1026 				skb_shinfo(skb)->nr_frags = 0;
1027 				kfree_skb(skb);
1028 				xenvif_tx_err(queue, &txreq, extra_count, idx);
1029 				if (net_ratelimit())
1030 					netdev_err(queue->vif->dev,
1031 						   "Can't allocate the frag_list skb.\n");
1032 				break;
1033 			}
1034 		}
1035 
1036 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1037 			struct xen_netif_extra_info *gso;
1038 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1039 
1040 			if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1041 				/* Failure in xenvif_set_skb_gso is fatal. */
1042 				skb_shinfo(skb)->nr_frags = 0;
1043 				kfree_skb(skb);
1044 				kfree_skb(nskb);
1045 				break;
1046 			}
1047 		}
1048 
1049 		if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1050 			struct xen_netif_extra_info *extra;
1051 			enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1052 
1053 			extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1054 
1055 			switch (extra->u.hash.type) {
1056 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1057 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1058 				type = PKT_HASH_TYPE_L3;
1059 				break;
1060 
1061 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1062 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1063 				type = PKT_HASH_TYPE_L4;
1064 				break;
1065 
1066 			default:
1067 				break;
1068 			}
1069 
1070 			if (type != PKT_HASH_TYPE_NONE)
1071 				skb_set_hash(skb,
1072 					     *(u32 *)extra->u.hash.value,
1073 					     type);
1074 		}
1075 
1076 		xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1077 				    map_ops, frag_overflow, nskb, extra_count,
1078 				    data_len);
1079 
1080 		__skb_queue_tail(&queue->tx_queue, skb);
1081 
1082 		queue->tx.req_cons = idx;
1083 	}
1084 
1085 	return;
1086 }
1087 
1088 /* Consolidate skb with a frag_list into a brand new one with local pages on
1089  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1090  */
1091 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1092 {
1093 	unsigned int offset = skb_headlen(skb);
1094 	skb_frag_t frags[MAX_SKB_FRAGS];
1095 	int i, f;
1096 	struct ubuf_info *uarg;
1097 	struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1098 
1099 	queue->stats.tx_zerocopy_sent += 2;
1100 	queue->stats.tx_frag_overflow++;
1101 
1102 	xenvif_fill_frags(queue, nskb);
1103 	/* Subtract frags size, we will correct it later */
1104 	skb->truesize -= skb->data_len;
1105 	skb->len += nskb->len;
1106 	skb->data_len += nskb->len;
1107 
1108 	/* create a brand new frags array and coalesce there */
1109 	for (i = 0; offset < skb->len; i++) {
1110 		struct page *page;
1111 		unsigned int len;
1112 
1113 		BUG_ON(i >= MAX_SKB_FRAGS);
1114 		page = alloc_page(GFP_ATOMIC);
1115 		if (!page) {
1116 			int j;
1117 			skb->truesize += skb->data_len;
1118 			for (j = 0; j < i; j++)
1119 				put_page(skb_frag_page(&frags[j]));
1120 			return -ENOMEM;
1121 		}
1122 
1123 		if (offset + PAGE_SIZE < skb->len)
1124 			len = PAGE_SIZE;
1125 		else
1126 			len = skb->len - offset;
1127 		if (skb_copy_bits(skb, offset, page_address(page), len))
1128 			BUG();
1129 
1130 		offset += len;
1131 		skb_frag_fill_page_desc(&frags[i], page, 0, len);
1132 	}
1133 
1134 	/* Release all the original (foreign) frags. */
1135 	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1136 		skb_frag_unref(skb, f);
1137 	uarg = skb_shinfo(skb)->destructor_arg;
1138 	/* increase inflight counter to offset decrement in callback */
1139 	atomic_inc(&queue->inflight_packets);
1140 	uarg->callback(NULL, uarg, true);
1141 	skb_shinfo(skb)->destructor_arg = NULL;
1142 
1143 	/* Fill the skb with the new (local) frags. */
1144 	memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1145 	skb_shinfo(skb)->nr_frags = i;
1146 	skb->truesize += i * PAGE_SIZE;
1147 
1148 	return 0;
1149 }
1150 
1151 static int xenvif_tx_submit(struct xenvif_queue *queue)
1152 {
1153 	struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1154 	struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1155 	struct sk_buff *skb;
1156 	int work_done = 0;
1157 
1158 	while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1159 		struct xen_netif_tx_request *txp;
1160 		u16 pending_idx;
1161 
1162 		pending_idx = copy_pending_idx(skb, 0);
1163 		txp = &queue->pending_tx_info[pending_idx].req;
1164 
1165 		/* Check the remap error code. */
1166 		if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1167 			/* If there was an error, xenvif_tx_check_gop is
1168 			 * expected to release all the frags which were mapped,
1169 			 * so kfree_skb shouldn't do it again
1170 			 */
1171 			skb_shinfo(skb)->nr_frags = 0;
1172 			if (skb_has_frag_list(skb)) {
1173 				struct sk_buff *nskb =
1174 						skb_shinfo(skb)->frag_list;
1175 				skb_shinfo(nskb)->nr_frags = 0;
1176 			}
1177 			kfree_skb(skb);
1178 			continue;
1179 		}
1180 
1181 		if (txp->flags & XEN_NETTXF_csum_blank)
1182 			skb->ip_summed = CHECKSUM_PARTIAL;
1183 		else if (txp->flags & XEN_NETTXF_data_validated)
1184 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1185 
1186 		xenvif_fill_frags(queue, skb);
1187 
1188 		if (unlikely(skb_has_frag_list(skb))) {
1189 			struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1190 			xenvif_skb_zerocopy_prepare(queue, nskb);
1191 			if (xenvif_handle_frag_list(queue, skb)) {
1192 				if (net_ratelimit())
1193 					netdev_err(queue->vif->dev,
1194 						   "Not enough memory to consolidate frag_list!\n");
1195 				xenvif_skb_zerocopy_prepare(queue, skb);
1196 				kfree_skb(skb);
1197 				continue;
1198 			}
1199 			/* Copied all the bits from the frag list -- free it. */
1200 			skb_frag_list_init(skb);
1201 			kfree_skb(nskb);
1202 		}
1203 
1204 		skb->dev      = queue->vif->dev;
1205 		skb->protocol = eth_type_trans(skb, skb->dev);
1206 		skb_reset_network_header(skb);
1207 
1208 		if (checksum_setup(queue, skb)) {
1209 			netdev_dbg(queue->vif->dev,
1210 				   "Can't setup checksum in net_tx_action\n");
1211 			/* We have to set this flag to trigger the callback */
1212 			if (skb_shinfo(skb)->destructor_arg)
1213 				xenvif_skb_zerocopy_prepare(queue, skb);
1214 			kfree_skb(skb);
1215 			continue;
1216 		}
1217 
1218 		skb_probe_transport_header(skb);
1219 
1220 		/* If the packet is GSO then we will have just set up the
1221 		 * transport header offset in checksum_setup so it's now
1222 		 * straightforward to calculate gso_segs.
1223 		 */
1224 		if (skb_is_gso(skb)) {
1225 			int mss, hdrlen;
1226 
1227 			/* GSO implies having the L4 header. */
1228 			WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1229 			if (unlikely(!skb_transport_header_was_set(skb))) {
1230 				kfree_skb(skb);
1231 				continue;
1232 			}
1233 
1234 			mss = skb_shinfo(skb)->gso_size;
1235 			hdrlen = skb_tcp_all_headers(skb);
1236 
1237 			skb_shinfo(skb)->gso_segs =
1238 				DIV_ROUND_UP(skb->len - hdrlen, mss);
1239 		}
1240 
1241 		queue->stats.rx_bytes += skb->len;
1242 		queue->stats.rx_packets++;
1243 
1244 		work_done++;
1245 
1246 		/* Set this flag right before netif_receive_skb, otherwise
1247 		 * someone might think this packet already left netback, and
1248 		 * do a skb_copy_ubufs while we are still in control of the
1249 		 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1250 		 */
1251 		if (skb_shinfo(skb)->destructor_arg) {
1252 			xenvif_skb_zerocopy_prepare(queue, skb);
1253 			queue->stats.tx_zerocopy_sent++;
1254 		}
1255 
1256 		netif_receive_skb(skb);
1257 	}
1258 
1259 	return work_done;
1260 }
1261 
1262 void xenvif_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *ubuf_base,
1263 			      bool zerocopy_success)
1264 {
1265 	unsigned long flags;
1266 	pending_ring_idx_t index;
1267 	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1268 	struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1269 
1270 	/* This is the only place where we grab this lock, to protect callbacks
1271 	 * from each other.
1272 	 */
1273 	spin_lock_irqsave(&queue->callback_lock, flags);
1274 	do {
1275 		u16 pending_idx = ubuf->desc;
1276 		ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1277 		BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1278 			MAX_PENDING_REQS);
1279 		index = pending_index(queue->dealloc_prod);
1280 		queue->dealloc_ring[index] = pending_idx;
1281 		/* Sync with xenvif_tx_dealloc_action:
1282 		 * insert idx then incr producer.
1283 		 */
1284 		smp_wmb();
1285 		queue->dealloc_prod++;
1286 	} while (ubuf);
1287 	spin_unlock_irqrestore(&queue->callback_lock, flags);
1288 
1289 	if (likely(zerocopy_success))
1290 		queue->stats.tx_zerocopy_success++;
1291 	else
1292 		queue->stats.tx_zerocopy_fail++;
1293 	xenvif_skb_zerocopy_complete(queue);
1294 }
1295 
1296 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1297 {
1298 	struct gnttab_unmap_grant_ref *gop;
1299 	pending_ring_idx_t dc, dp;
1300 	u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1301 	unsigned int i = 0;
1302 
1303 	dc = queue->dealloc_cons;
1304 	gop = queue->tx_unmap_ops;
1305 
1306 	/* Free up any grants we have finished using */
1307 	do {
1308 		dp = queue->dealloc_prod;
1309 
1310 		/* Ensure we see all indices enqueued by all
1311 		 * xenvif_zerocopy_callback().
1312 		 */
1313 		smp_rmb();
1314 
1315 		while (dc != dp) {
1316 			BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1317 			pending_idx =
1318 				queue->dealloc_ring[pending_index(dc++)];
1319 
1320 			pending_idx_release[gop - queue->tx_unmap_ops] =
1321 				pending_idx;
1322 			queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1323 				queue->mmap_pages[pending_idx];
1324 			gnttab_set_unmap_op(gop,
1325 					    idx_to_kaddr(queue, pending_idx),
1326 					    GNTMAP_host_map,
1327 					    queue->grant_tx_handle[pending_idx]);
1328 			xenvif_grant_handle_reset(queue, pending_idx);
1329 			++gop;
1330 		}
1331 
1332 	} while (dp != queue->dealloc_prod);
1333 
1334 	queue->dealloc_cons = dc;
1335 
1336 	if (gop - queue->tx_unmap_ops > 0) {
1337 		int ret;
1338 		ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1339 					NULL,
1340 					queue->pages_to_unmap,
1341 					gop - queue->tx_unmap_ops);
1342 		if (ret) {
1343 			netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1344 				   gop - queue->tx_unmap_ops, ret);
1345 			for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1346 				if (gop[i].status != GNTST_okay)
1347 					netdev_err(queue->vif->dev,
1348 						   " host_addr: 0x%llx handle: 0x%x status: %d\n",
1349 						   gop[i].host_addr,
1350 						   gop[i].handle,
1351 						   gop[i].status);
1352 			}
1353 			BUG();
1354 		}
1355 	}
1356 
1357 	for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1358 		xenvif_idx_release(queue, pending_idx_release[i],
1359 				   XEN_NETIF_RSP_OKAY);
1360 }
1361 
1362 
1363 /* Called after netfront has transmitted */
1364 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1365 {
1366 	unsigned nr_mops = 0, nr_cops = 0;
1367 	int work_done, ret;
1368 
1369 	if (unlikely(!tx_work_todo(queue)))
1370 		return 0;
1371 
1372 	xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1373 
1374 	if (nr_cops == 0)
1375 		return 0;
1376 
1377 	gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1378 	if (nr_mops != 0) {
1379 		ret = gnttab_map_refs(queue->tx_map_ops,
1380 				      NULL,
1381 				      queue->pages_to_map,
1382 				      nr_mops);
1383 		if (ret) {
1384 			unsigned int i;
1385 
1386 			netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1387 				   nr_mops, ret);
1388 			for (i = 0; i < nr_mops; ++i)
1389 				WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1390 				             GNTST_okay);
1391 		}
1392 	}
1393 
1394 	work_done = xenvif_tx_submit(queue);
1395 
1396 	return work_done;
1397 }
1398 
1399 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1400 			       u8 status)
1401 {
1402 	struct pending_tx_info *pending_tx_info;
1403 	pending_ring_idx_t index;
1404 	unsigned long flags;
1405 
1406 	pending_tx_info = &queue->pending_tx_info[pending_idx];
1407 
1408 	spin_lock_irqsave(&queue->response_lock, flags);
1409 
1410 	make_tx_response(queue, &pending_tx_info->req,
1411 			 pending_tx_info->extra_count, status);
1412 
1413 	/* Release the pending index before pusing the Tx response so
1414 	 * its available before a new Tx request is pushed by the
1415 	 * frontend.
1416 	 */
1417 	index = pending_index(queue->pending_prod++);
1418 	queue->pending_ring[index] = pending_idx;
1419 
1420 	push_tx_responses(queue);
1421 
1422 	spin_unlock_irqrestore(&queue->response_lock, flags);
1423 }
1424 
1425 
1426 static void make_tx_response(struct xenvif_queue *queue,
1427 			     struct xen_netif_tx_request *txp,
1428 			     unsigned int extra_count,
1429 			     s8       st)
1430 {
1431 	RING_IDX i = queue->tx.rsp_prod_pvt;
1432 	struct xen_netif_tx_response *resp;
1433 
1434 	resp = RING_GET_RESPONSE(&queue->tx, i);
1435 	resp->id     = txp->id;
1436 	resp->status = st;
1437 
1438 	while (extra_count-- != 0)
1439 		RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1440 
1441 	queue->tx.rsp_prod_pvt = ++i;
1442 }
1443 
1444 static void push_tx_responses(struct xenvif_queue *queue)
1445 {
1446 	int notify;
1447 
1448 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1449 	if (notify)
1450 		notify_remote_via_irq(queue->tx_irq);
1451 }
1452 
1453 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1454 {
1455 	int ret;
1456 	struct gnttab_unmap_grant_ref tx_unmap_op;
1457 
1458 	gnttab_set_unmap_op(&tx_unmap_op,
1459 			    idx_to_kaddr(queue, pending_idx),
1460 			    GNTMAP_host_map,
1461 			    queue->grant_tx_handle[pending_idx]);
1462 	xenvif_grant_handle_reset(queue, pending_idx);
1463 
1464 	ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1465 				&queue->mmap_pages[pending_idx], 1);
1466 	if (ret) {
1467 		netdev_err(queue->vif->dev,
1468 			   "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1469 			   ret,
1470 			   pending_idx,
1471 			   tx_unmap_op.host_addr,
1472 			   tx_unmap_op.handle,
1473 			   tx_unmap_op.status);
1474 		BUG();
1475 	}
1476 }
1477 
1478 static inline int tx_work_todo(struct xenvif_queue *queue)
1479 {
1480 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1481 		return 1;
1482 
1483 	return 0;
1484 }
1485 
1486 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1487 {
1488 	return queue->dealloc_cons != queue->dealloc_prod;
1489 }
1490 
1491 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1492 {
1493 	if (queue->tx.sring)
1494 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1495 					queue->tx.sring);
1496 	if (queue->rx.sring)
1497 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1498 					queue->rx.sring);
1499 }
1500 
1501 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1502 				   grant_ref_t tx_ring_ref,
1503 				   grant_ref_t rx_ring_ref)
1504 {
1505 	void *addr;
1506 	struct xen_netif_tx_sring *txs;
1507 	struct xen_netif_rx_sring *rxs;
1508 	RING_IDX rsp_prod, req_prod;
1509 	int err;
1510 
1511 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1512 				     &tx_ring_ref, 1, &addr);
1513 	if (err)
1514 		goto err;
1515 
1516 	txs = (struct xen_netif_tx_sring *)addr;
1517 	rsp_prod = READ_ONCE(txs->rsp_prod);
1518 	req_prod = READ_ONCE(txs->req_prod);
1519 
1520 	BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1521 
1522 	err = -EIO;
1523 	if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1524 		goto err;
1525 
1526 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1527 				     &rx_ring_ref, 1, &addr);
1528 	if (err)
1529 		goto err;
1530 
1531 	rxs = (struct xen_netif_rx_sring *)addr;
1532 	rsp_prod = READ_ONCE(rxs->rsp_prod);
1533 	req_prod = READ_ONCE(rxs->req_prod);
1534 
1535 	BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1536 
1537 	err = -EIO;
1538 	if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1539 		goto err;
1540 
1541 	return 0;
1542 
1543 err:
1544 	xenvif_unmap_frontend_data_rings(queue);
1545 	return err;
1546 }
1547 
1548 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1549 {
1550 	/* Dealloc thread must remain running until all inflight
1551 	 * packets complete.
1552 	 */
1553 	return kthread_should_stop() &&
1554 		!atomic_read(&queue->inflight_packets);
1555 }
1556 
1557 int xenvif_dealloc_kthread(void *data)
1558 {
1559 	struct xenvif_queue *queue = data;
1560 
1561 	for (;;) {
1562 		wait_event_interruptible(queue->dealloc_wq,
1563 					 tx_dealloc_work_todo(queue) ||
1564 					 xenvif_dealloc_kthread_should_stop(queue));
1565 		if (xenvif_dealloc_kthread_should_stop(queue))
1566 			break;
1567 
1568 		xenvif_tx_dealloc_action(queue);
1569 		cond_resched();
1570 	}
1571 
1572 	/* Unmap anything remaining*/
1573 	if (tx_dealloc_work_todo(queue))
1574 		xenvif_tx_dealloc_action(queue);
1575 
1576 	return 0;
1577 }
1578 
1579 static void make_ctrl_response(struct xenvif *vif,
1580 			       const struct xen_netif_ctrl_request *req,
1581 			       u32 status, u32 data)
1582 {
1583 	RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1584 	struct xen_netif_ctrl_response rsp = {
1585 		.id = req->id,
1586 		.type = req->type,
1587 		.status = status,
1588 		.data = data,
1589 	};
1590 
1591 	*RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1592 	vif->ctrl.rsp_prod_pvt = ++idx;
1593 }
1594 
1595 static void push_ctrl_response(struct xenvif *vif)
1596 {
1597 	int notify;
1598 
1599 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1600 	if (notify)
1601 		notify_remote_via_irq(vif->ctrl_irq);
1602 }
1603 
1604 static void process_ctrl_request(struct xenvif *vif,
1605 				 const struct xen_netif_ctrl_request *req)
1606 {
1607 	u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1608 	u32 data = 0;
1609 
1610 	switch (req->type) {
1611 	case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1612 		status = xenvif_set_hash_alg(vif, req->data[0]);
1613 		break;
1614 
1615 	case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1616 		status = xenvif_get_hash_flags(vif, &data);
1617 		break;
1618 
1619 	case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1620 		status = xenvif_set_hash_flags(vif, req->data[0]);
1621 		break;
1622 
1623 	case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1624 		status = xenvif_set_hash_key(vif, req->data[0],
1625 					     req->data[1]);
1626 		break;
1627 
1628 	case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1629 		status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1630 		data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1631 		break;
1632 
1633 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1634 		status = xenvif_set_hash_mapping_size(vif,
1635 						      req->data[0]);
1636 		break;
1637 
1638 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1639 		status = xenvif_set_hash_mapping(vif, req->data[0],
1640 						 req->data[1],
1641 						 req->data[2]);
1642 		break;
1643 
1644 	default:
1645 		break;
1646 	}
1647 
1648 	make_ctrl_response(vif, req, status, data);
1649 	push_ctrl_response(vif);
1650 }
1651 
1652 static void xenvif_ctrl_action(struct xenvif *vif)
1653 {
1654 	for (;;) {
1655 		RING_IDX req_prod, req_cons;
1656 
1657 		req_prod = vif->ctrl.sring->req_prod;
1658 		req_cons = vif->ctrl.req_cons;
1659 
1660 		/* Make sure we can see requests before we process them. */
1661 		rmb();
1662 
1663 		if (req_cons == req_prod)
1664 			break;
1665 
1666 		while (req_cons != req_prod) {
1667 			struct xen_netif_ctrl_request req;
1668 
1669 			RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1670 			req_cons++;
1671 
1672 			process_ctrl_request(vif, &req);
1673 		}
1674 
1675 		vif->ctrl.req_cons = req_cons;
1676 		vif->ctrl.sring->req_event = req_cons + 1;
1677 	}
1678 }
1679 
1680 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1681 {
1682 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1683 		return true;
1684 
1685 	return false;
1686 }
1687 
1688 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1689 {
1690 	struct xenvif *vif = data;
1691 	unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1692 
1693 	while (xenvif_ctrl_work_todo(vif)) {
1694 		xenvif_ctrl_action(vif);
1695 		eoi_flag = 0;
1696 	}
1697 
1698 	xen_irq_lateeoi(irq, eoi_flag);
1699 
1700 	return IRQ_HANDLED;
1701 }
1702 
1703 static int __init netback_init(void)
1704 {
1705 	int rc = 0;
1706 
1707 	if (!xen_domain())
1708 		return -ENODEV;
1709 
1710 	/* Allow as many queues as there are CPUs but max. 8 if user has not
1711 	 * specified a value.
1712 	 */
1713 	if (xenvif_max_queues == 0)
1714 		xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1715 					  num_online_cpus());
1716 
1717 	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1718 		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1719 			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1720 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1721 	}
1722 
1723 	rc = xenvif_xenbus_init();
1724 	if (rc)
1725 		goto failed_init;
1726 
1727 #ifdef CONFIG_DEBUG_FS
1728 	xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1729 #endif /* CONFIG_DEBUG_FS */
1730 
1731 	return 0;
1732 
1733 failed_init:
1734 	return rc;
1735 }
1736 
1737 module_init(netback_init);
1738 
1739 static void __exit netback_fini(void)
1740 {
1741 #ifdef CONFIG_DEBUG_FS
1742 	debugfs_remove_recursive(xen_netback_dbg_root);
1743 #endif /* CONFIG_DEBUG_FS */
1744 	xenvif_xenbus_fini();
1745 }
1746 module_exit(netback_fini);
1747 
1748 MODULE_LICENSE("Dual BSD/GPL");
1749 MODULE_ALIAS("xen-backend:vif");
1750