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 + frag_overflow + 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; nr_slots > 0 && shinfo->nr_frags < MAX_SKB_FRAGS;
466 	     shinfo->nr_frags++, gop++, nr_slots--) {
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 (nr_slots > 0) {
480 
481 		shinfo = skb_shinfo(nskb);
482 		frags = shinfo->frags;
483 
484 		for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots;
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 	} else if (nskb) {
496 		/* A frag_list skb was allocated but it is no longer needed
497 		 * because enough slots were converted to copy ops above.
498 		 */
499 		kfree_skb(nskb);
500 	}
501 
502 	(*copy_ops) = cop - queue->tx_copy_ops;
503 	(*map_ops) = gop - queue->tx_map_ops;
504 }
505 
506 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
507 					   u16 pending_idx,
508 					   grant_handle_t handle)
509 {
510 	if (unlikely(queue->grant_tx_handle[pending_idx] !=
511 		     NETBACK_INVALID_HANDLE)) {
512 		netdev_err(queue->vif->dev,
513 			   "Trying to overwrite active handle! pending_idx: 0x%x\n",
514 			   pending_idx);
515 		BUG();
516 	}
517 	queue->grant_tx_handle[pending_idx] = handle;
518 }
519 
520 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
521 					     u16 pending_idx)
522 {
523 	if (unlikely(queue->grant_tx_handle[pending_idx] ==
524 		     NETBACK_INVALID_HANDLE)) {
525 		netdev_err(queue->vif->dev,
526 			   "Trying to unmap invalid handle! pending_idx: 0x%x\n",
527 			   pending_idx);
528 		BUG();
529 	}
530 	queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
531 }
532 
533 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
534 			       struct sk_buff *skb,
535 			       struct gnttab_map_grant_ref **gopp_map,
536 			       struct gnttab_copy **gopp_copy)
537 {
538 	struct gnttab_map_grant_ref *gop_map = *gopp_map;
539 	u16 pending_idx;
540 	/* This always points to the shinfo of the skb being checked, which
541 	 * could be either the first or the one on the frag_list
542 	 */
543 	struct skb_shared_info *shinfo = skb_shinfo(skb);
544 	/* If this is non-NULL, we are currently checking the frag_list skb, and
545 	 * this points to the shinfo of the first one
546 	 */
547 	struct skb_shared_info *first_shinfo = NULL;
548 	int nr_frags = shinfo->nr_frags;
549 	const bool sharedslot = nr_frags &&
550 				frag_get_pending_idx(&shinfo->frags[0]) ==
551 				    copy_pending_idx(skb, copy_count(skb) - 1);
552 	int i, err = 0;
553 
554 	for (i = 0; i < copy_count(skb); i++) {
555 		int newerr;
556 
557 		/* Check status of header. */
558 		pending_idx = copy_pending_idx(skb, i);
559 
560 		newerr = (*gopp_copy)->status;
561 
562 		/* Split copies need to be handled together. */
563 		if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
564 			(*gopp_copy)++;
565 			if (!newerr)
566 				newerr = (*gopp_copy)->status;
567 		}
568 		if (likely(!newerr)) {
569 			/* The first frag might still have this slot mapped */
570 			if (i < copy_count(skb) - 1 || !sharedslot)
571 				xenvif_idx_release(queue, pending_idx,
572 						   XEN_NETIF_RSP_OKAY);
573 		} else {
574 			err = newerr;
575 			if (net_ratelimit())
576 				netdev_dbg(queue->vif->dev,
577 					   "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
578 					   (*gopp_copy)->status,
579 					   pending_idx,
580 					   (*gopp_copy)->source.u.ref);
581 			/* The first frag might still have this slot mapped */
582 			if (i < copy_count(skb) - 1 || !sharedslot)
583 				xenvif_idx_release(queue, pending_idx,
584 						   XEN_NETIF_RSP_ERROR);
585 		}
586 		(*gopp_copy)++;
587 	}
588 
589 check_frags:
590 	for (i = 0; i < nr_frags; i++, gop_map++) {
591 		int j, newerr;
592 
593 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
594 
595 		/* Check error status: if okay then remember grant handle. */
596 		newerr = gop_map->status;
597 
598 		if (likely(!newerr)) {
599 			xenvif_grant_handle_set(queue,
600 						pending_idx,
601 						gop_map->handle);
602 			/* Had a previous error? Invalidate this fragment. */
603 			if (unlikely(err)) {
604 				xenvif_idx_unmap(queue, pending_idx);
605 				/* If the mapping of the first frag was OK, but
606 				 * the header's copy failed, and they are
607 				 * sharing a slot, send an error
608 				 */
609 				if (i == 0 && !first_shinfo && sharedslot)
610 					xenvif_idx_release(queue, pending_idx,
611 							   XEN_NETIF_RSP_ERROR);
612 				else
613 					xenvif_idx_release(queue, pending_idx,
614 							   XEN_NETIF_RSP_OKAY);
615 			}
616 			continue;
617 		}
618 
619 		/* Error on this fragment: respond to client with an error. */
620 		if (net_ratelimit())
621 			netdev_dbg(queue->vif->dev,
622 				   "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
623 				   i,
624 				   gop_map->status,
625 				   pending_idx,
626 				   gop_map->ref);
627 
628 		xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
629 
630 		/* Not the first error? Preceding frags already invalidated. */
631 		if (err)
632 			continue;
633 
634 		/* Invalidate preceding fragments of this skb. */
635 		for (j = 0; j < i; j++) {
636 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
637 			xenvif_idx_unmap(queue, pending_idx);
638 			xenvif_idx_release(queue, pending_idx,
639 					   XEN_NETIF_RSP_OKAY);
640 		}
641 
642 		/* And if we found the error while checking the frag_list, unmap
643 		 * the first skb's frags
644 		 */
645 		if (first_shinfo) {
646 			for (j = 0; j < first_shinfo->nr_frags; j++) {
647 				pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
648 				xenvif_idx_unmap(queue, pending_idx);
649 				xenvif_idx_release(queue, pending_idx,
650 						   XEN_NETIF_RSP_OKAY);
651 			}
652 		}
653 
654 		/* Remember the error: invalidate all subsequent fragments. */
655 		err = newerr;
656 	}
657 
658 	if (skb_has_frag_list(skb) && !first_shinfo) {
659 		first_shinfo = shinfo;
660 		shinfo = skb_shinfo(shinfo->frag_list);
661 		nr_frags = shinfo->nr_frags;
662 
663 		goto check_frags;
664 	}
665 
666 	*gopp_map = gop_map;
667 	return err;
668 }
669 
670 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
671 {
672 	struct skb_shared_info *shinfo = skb_shinfo(skb);
673 	int nr_frags = shinfo->nr_frags;
674 	int i;
675 	u16 prev_pending_idx = INVALID_PENDING_IDX;
676 
677 	for (i = 0; i < nr_frags; i++) {
678 		skb_frag_t *frag = shinfo->frags + i;
679 		struct xen_netif_tx_request *txp;
680 		struct page *page;
681 		u16 pending_idx;
682 
683 		pending_idx = frag_get_pending_idx(frag);
684 
685 		/* If this is not the first frag, chain it to the previous*/
686 		if (prev_pending_idx == INVALID_PENDING_IDX)
687 			skb_shinfo(skb)->destructor_arg =
688 				&callback_param(queue, pending_idx);
689 		else
690 			callback_param(queue, prev_pending_idx).ctx =
691 				&callback_param(queue, pending_idx);
692 
693 		callback_param(queue, pending_idx).ctx = NULL;
694 		prev_pending_idx = pending_idx;
695 
696 		txp = &queue->pending_tx_info[pending_idx].req;
697 		page = virt_to_page((void *)idx_to_kaddr(queue, pending_idx));
698 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
699 		skb->len += txp->size;
700 		skb->data_len += txp->size;
701 		skb->truesize += txp->size;
702 
703 		/* Take an extra reference to offset network stack's put_page */
704 		get_page(queue->mmap_pages[pending_idx]);
705 	}
706 }
707 
708 static int xenvif_get_extras(struct xenvif_queue *queue,
709 			     struct xen_netif_extra_info *extras,
710 			     unsigned int *extra_count,
711 			     int work_to_do)
712 {
713 	struct xen_netif_extra_info extra;
714 	RING_IDX cons = queue->tx.req_cons;
715 
716 	do {
717 		if (unlikely(work_to_do-- <= 0)) {
718 			netdev_err(queue->vif->dev, "Missing extra info\n");
719 			xenvif_fatal_tx_err(queue->vif);
720 			return -EBADR;
721 		}
722 
723 		RING_COPY_REQUEST(&queue->tx, cons, &extra);
724 
725 		queue->tx.req_cons = ++cons;
726 		(*extra_count)++;
727 
728 		if (unlikely(!extra.type ||
729 			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
730 			netdev_err(queue->vif->dev,
731 				   "Invalid extra type: %d\n", extra.type);
732 			xenvif_fatal_tx_err(queue->vif);
733 			return -EINVAL;
734 		}
735 
736 		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
737 	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
738 
739 	return work_to_do;
740 }
741 
742 static int xenvif_set_skb_gso(struct xenvif *vif,
743 			      struct sk_buff *skb,
744 			      struct xen_netif_extra_info *gso)
745 {
746 	if (!gso->u.gso.size) {
747 		netdev_err(vif->dev, "GSO size must not be zero.\n");
748 		xenvif_fatal_tx_err(vif);
749 		return -EINVAL;
750 	}
751 
752 	switch (gso->u.gso.type) {
753 	case XEN_NETIF_GSO_TYPE_TCPV4:
754 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
755 		break;
756 	case XEN_NETIF_GSO_TYPE_TCPV6:
757 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
758 		break;
759 	default:
760 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
761 		xenvif_fatal_tx_err(vif);
762 		return -EINVAL;
763 	}
764 
765 	skb_shinfo(skb)->gso_size = gso->u.gso.size;
766 	/* gso_segs will be calculated later */
767 
768 	return 0;
769 }
770 
771 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
772 {
773 	bool recalculate_partial_csum = false;
774 
775 	/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
776 	 * peers can fail to set NETRXF_csum_blank when sending a GSO
777 	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
778 	 * recalculate the partial checksum.
779 	 */
780 	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
781 		queue->stats.rx_gso_checksum_fixup++;
782 		skb->ip_summed = CHECKSUM_PARTIAL;
783 		recalculate_partial_csum = true;
784 	}
785 
786 	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
787 	if (skb->ip_summed != CHECKSUM_PARTIAL)
788 		return 0;
789 
790 	return skb_checksum_setup(skb, recalculate_partial_csum);
791 }
792 
793 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
794 {
795 	u64 now = get_jiffies_64();
796 	u64 next_credit = queue->credit_window_start +
797 		msecs_to_jiffies(queue->credit_usec / 1000);
798 
799 	/* Timer could already be pending in rare cases. */
800 	if (timer_pending(&queue->credit_timeout)) {
801 		queue->rate_limited = true;
802 		return true;
803 	}
804 
805 	/* Passed the point where we can replenish credit? */
806 	if (time_after_eq64(now, next_credit)) {
807 		queue->credit_window_start = now;
808 		tx_add_credit(queue);
809 	}
810 
811 	/* Still too big to send right now? Set a callback. */
812 	if (size > queue->remaining_credit) {
813 		mod_timer(&queue->credit_timeout,
814 			  next_credit);
815 		queue->credit_window_start = next_credit;
816 		queue->rate_limited = true;
817 
818 		return true;
819 	}
820 
821 	return false;
822 }
823 
824 /* No locking is required in xenvif_mcast_add/del() as they are
825  * only ever invoked from NAPI poll. An RCU list is used because
826  * xenvif_mcast_match() is called asynchronously, during start_xmit.
827  */
828 
829 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
830 {
831 	struct xenvif_mcast_addr *mcast;
832 
833 	if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
834 		if (net_ratelimit())
835 			netdev_err(vif->dev,
836 				   "Too many multicast addresses\n");
837 		return -ENOSPC;
838 	}
839 
840 	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
841 	if (!mcast)
842 		return -ENOMEM;
843 
844 	ether_addr_copy(mcast->addr, addr);
845 	list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
846 	vif->fe_mcast_count++;
847 
848 	return 0;
849 }
850 
851 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
852 {
853 	struct xenvif_mcast_addr *mcast;
854 
855 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
856 		if (ether_addr_equal(addr, mcast->addr)) {
857 			--vif->fe_mcast_count;
858 			list_del_rcu(&mcast->entry);
859 			kfree_rcu(mcast, rcu);
860 			break;
861 		}
862 	}
863 }
864 
865 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
866 {
867 	struct xenvif_mcast_addr *mcast;
868 
869 	rcu_read_lock();
870 	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
871 		if (ether_addr_equal(addr, mcast->addr)) {
872 			rcu_read_unlock();
873 			return true;
874 		}
875 	}
876 	rcu_read_unlock();
877 
878 	return false;
879 }
880 
881 void xenvif_mcast_addr_list_free(struct xenvif *vif)
882 {
883 	/* No need for locking or RCU here. NAPI poll and TX queue
884 	 * are stopped.
885 	 */
886 	while (!list_empty(&vif->fe_mcast_addr)) {
887 		struct xenvif_mcast_addr *mcast;
888 
889 		mcast = list_first_entry(&vif->fe_mcast_addr,
890 					 struct xenvif_mcast_addr,
891 					 entry);
892 		--vif->fe_mcast_count;
893 		list_del(&mcast->entry);
894 		kfree(mcast);
895 	}
896 }
897 
898 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
899 				     int budget,
900 				     unsigned *copy_ops,
901 				     unsigned *map_ops)
902 {
903 	struct sk_buff *skb, *nskb;
904 	int ret;
905 	unsigned int frag_overflow;
906 
907 	while (skb_queue_len(&queue->tx_queue) < budget) {
908 		struct xen_netif_tx_request txreq;
909 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
910 		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
911 		unsigned int extra_count;
912 		RING_IDX idx;
913 		int work_to_do;
914 		unsigned int data_len;
915 
916 		if (queue->tx.sring->req_prod - queue->tx.req_cons >
917 		    XEN_NETIF_TX_RING_SIZE) {
918 			netdev_err(queue->vif->dev,
919 				   "Impossible number of requests. "
920 				   "req_prod %d, req_cons %d, size %ld\n",
921 				   queue->tx.sring->req_prod, queue->tx.req_cons,
922 				   XEN_NETIF_TX_RING_SIZE);
923 			xenvif_fatal_tx_err(queue->vif);
924 			break;
925 		}
926 
927 		work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
928 		if (!work_to_do)
929 			break;
930 
931 		idx = queue->tx.req_cons;
932 		rmb(); /* Ensure that we see the request before we copy it. */
933 		RING_COPY_REQUEST(&queue->tx, idx, &txreq);
934 
935 		/* Credit-based scheduling. */
936 		if (txreq.size > queue->remaining_credit &&
937 		    tx_credit_exceeded(queue, txreq.size))
938 			break;
939 
940 		queue->remaining_credit -= txreq.size;
941 
942 		work_to_do--;
943 		queue->tx.req_cons = ++idx;
944 
945 		memset(extras, 0, sizeof(extras));
946 		extra_count = 0;
947 		if (txreq.flags & XEN_NETTXF_extra_info) {
948 			work_to_do = xenvif_get_extras(queue, extras,
949 						       &extra_count,
950 						       work_to_do);
951 			idx = queue->tx.req_cons;
952 			if (unlikely(work_to_do < 0))
953 				break;
954 		}
955 
956 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
957 			struct xen_netif_extra_info *extra;
958 
959 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
960 			ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
961 
962 			make_tx_response(queue, &txreq, extra_count,
963 					 (ret == 0) ?
964 					 XEN_NETIF_RSP_OKAY :
965 					 XEN_NETIF_RSP_ERROR);
966 			push_tx_responses(queue);
967 			continue;
968 		}
969 
970 		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
971 			struct xen_netif_extra_info *extra;
972 
973 			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
974 			xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
975 
976 			make_tx_response(queue, &txreq, extra_count,
977 					 XEN_NETIF_RSP_OKAY);
978 			push_tx_responses(queue);
979 			continue;
980 		}
981 
982 		data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
983 			XEN_NETBACK_TX_COPY_LEN : txreq.size;
984 
985 		ret = xenvif_count_requests(queue, &txreq, extra_count,
986 					    txfrags, work_to_do);
987 
988 		if (unlikely(ret < 0))
989 			break;
990 
991 		idx += ret;
992 
993 		if (unlikely(txreq.size < ETH_HLEN)) {
994 			netdev_dbg(queue->vif->dev,
995 				   "Bad packet size: %d\n", txreq.size);
996 			xenvif_tx_err(queue, &txreq, extra_count, idx);
997 			break;
998 		}
999 
1000 		/* No crossing a page as the payload mustn't fragment. */
1001 		if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1002 			netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
1003 				   txreq.offset, txreq.size);
1004 			xenvif_fatal_tx_err(queue->vif);
1005 			break;
1006 		}
1007 
1008 		if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1009 			data_len = txreq.size;
1010 
1011 		skb = xenvif_alloc_skb(data_len);
1012 		if (unlikely(skb == NULL)) {
1013 			netdev_dbg(queue->vif->dev,
1014 				   "Can't allocate a skb in start_xmit.\n");
1015 			xenvif_tx_err(queue, &txreq, extra_count, idx);
1016 			break;
1017 		}
1018 
1019 		skb_shinfo(skb)->nr_frags = ret;
1020 		/* At this point shinfo->nr_frags is in fact the number of
1021 		 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1022 		 */
1023 		frag_overflow = 0;
1024 		nskb = NULL;
1025 		if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1026 			frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1027 			BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1028 			skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1029 			nskb = xenvif_alloc_skb(0);
1030 			if (unlikely(nskb == NULL)) {
1031 				skb_shinfo(skb)->nr_frags = 0;
1032 				kfree_skb(skb);
1033 				xenvif_tx_err(queue, &txreq, extra_count, idx);
1034 				if (net_ratelimit())
1035 					netdev_err(queue->vif->dev,
1036 						   "Can't allocate the frag_list skb.\n");
1037 				break;
1038 			}
1039 		}
1040 
1041 		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1042 			struct xen_netif_extra_info *gso;
1043 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1044 
1045 			if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1046 				/* Failure in xenvif_set_skb_gso is fatal. */
1047 				skb_shinfo(skb)->nr_frags = 0;
1048 				kfree_skb(skb);
1049 				kfree_skb(nskb);
1050 				break;
1051 			}
1052 		}
1053 
1054 		if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1055 			struct xen_netif_extra_info *extra;
1056 			enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1057 
1058 			extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1059 
1060 			switch (extra->u.hash.type) {
1061 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1062 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1063 				type = PKT_HASH_TYPE_L3;
1064 				break;
1065 
1066 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1067 			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1068 				type = PKT_HASH_TYPE_L4;
1069 				break;
1070 
1071 			default:
1072 				break;
1073 			}
1074 
1075 			if (type != PKT_HASH_TYPE_NONE)
1076 				skb_set_hash(skb,
1077 					     *(u32 *)extra->u.hash.value,
1078 					     type);
1079 		}
1080 
1081 		xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1082 				    map_ops, frag_overflow, nskb, extra_count,
1083 				    data_len);
1084 
1085 		__skb_queue_tail(&queue->tx_queue, skb);
1086 
1087 		queue->tx.req_cons = idx;
1088 	}
1089 
1090 	return;
1091 }
1092 
1093 /* Consolidate skb with a frag_list into a brand new one with local pages on
1094  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1095  */
1096 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1097 {
1098 	unsigned int offset = skb_headlen(skb);
1099 	skb_frag_t frags[MAX_SKB_FRAGS];
1100 	int i, f;
1101 	struct ubuf_info *uarg;
1102 	struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1103 
1104 	queue->stats.tx_zerocopy_sent += 2;
1105 	queue->stats.tx_frag_overflow++;
1106 
1107 	xenvif_fill_frags(queue, nskb);
1108 	/* Subtract frags size, we will correct it later */
1109 	skb->truesize -= skb->data_len;
1110 	skb->len += nskb->len;
1111 	skb->data_len += nskb->len;
1112 
1113 	/* create a brand new frags array and coalesce there */
1114 	for (i = 0; offset < skb->len; i++) {
1115 		struct page *page;
1116 		unsigned int len;
1117 
1118 		BUG_ON(i >= MAX_SKB_FRAGS);
1119 		page = alloc_page(GFP_ATOMIC);
1120 		if (!page) {
1121 			int j;
1122 			skb->truesize += skb->data_len;
1123 			for (j = 0; j < i; j++)
1124 				put_page(skb_frag_page(&frags[j]));
1125 			return -ENOMEM;
1126 		}
1127 
1128 		if (offset + PAGE_SIZE < skb->len)
1129 			len = PAGE_SIZE;
1130 		else
1131 			len = skb->len - offset;
1132 		if (skb_copy_bits(skb, offset, page_address(page), len))
1133 			BUG();
1134 
1135 		offset += len;
1136 		skb_frag_fill_page_desc(&frags[i], page, 0, len);
1137 	}
1138 
1139 	/* Release all the original (foreign) frags. */
1140 	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1141 		skb_frag_unref(skb, f);
1142 	uarg = skb_shinfo(skb)->destructor_arg;
1143 	/* increase inflight counter to offset decrement in callback */
1144 	atomic_inc(&queue->inflight_packets);
1145 	uarg->callback(NULL, uarg, true);
1146 	skb_shinfo(skb)->destructor_arg = NULL;
1147 
1148 	/* Fill the skb with the new (local) frags. */
1149 	memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1150 	skb_shinfo(skb)->nr_frags = i;
1151 	skb->truesize += i * PAGE_SIZE;
1152 
1153 	return 0;
1154 }
1155 
1156 static int xenvif_tx_submit(struct xenvif_queue *queue)
1157 {
1158 	struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1159 	struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1160 	struct sk_buff *skb;
1161 	int work_done = 0;
1162 
1163 	while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1164 		struct xen_netif_tx_request *txp;
1165 		u16 pending_idx;
1166 
1167 		pending_idx = copy_pending_idx(skb, 0);
1168 		txp = &queue->pending_tx_info[pending_idx].req;
1169 
1170 		/* Check the remap error code. */
1171 		if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1172 			/* If there was an error, xenvif_tx_check_gop is
1173 			 * expected to release all the frags which were mapped,
1174 			 * so kfree_skb shouldn't do it again
1175 			 */
1176 			skb_shinfo(skb)->nr_frags = 0;
1177 			if (skb_has_frag_list(skb)) {
1178 				struct sk_buff *nskb =
1179 						skb_shinfo(skb)->frag_list;
1180 				skb_shinfo(nskb)->nr_frags = 0;
1181 			}
1182 			kfree_skb(skb);
1183 			continue;
1184 		}
1185 
1186 		if (txp->flags & XEN_NETTXF_csum_blank)
1187 			skb->ip_summed = CHECKSUM_PARTIAL;
1188 		else if (txp->flags & XEN_NETTXF_data_validated)
1189 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1190 
1191 		xenvif_fill_frags(queue, skb);
1192 
1193 		if (unlikely(skb_has_frag_list(skb))) {
1194 			struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1195 			xenvif_skb_zerocopy_prepare(queue, nskb);
1196 			if (xenvif_handle_frag_list(queue, skb)) {
1197 				if (net_ratelimit())
1198 					netdev_err(queue->vif->dev,
1199 						   "Not enough memory to consolidate frag_list!\n");
1200 				xenvif_skb_zerocopy_prepare(queue, skb);
1201 				kfree_skb(skb);
1202 				continue;
1203 			}
1204 			/* Copied all the bits from the frag list -- free it. */
1205 			skb_frag_list_init(skb);
1206 			kfree_skb(nskb);
1207 		}
1208 
1209 		skb->dev      = queue->vif->dev;
1210 		skb->protocol = eth_type_trans(skb, skb->dev);
1211 		skb_reset_network_header(skb);
1212 
1213 		if (checksum_setup(queue, skb)) {
1214 			netdev_dbg(queue->vif->dev,
1215 				   "Can't setup checksum in net_tx_action\n");
1216 			/* We have to set this flag to trigger the callback */
1217 			if (skb_shinfo(skb)->destructor_arg)
1218 				xenvif_skb_zerocopy_prepare(queue, skb);
1219 			kfree_skb(skb);
1220 			continue;
1221 		}
1222 
1223 		skb_probe_transport_header(skb);
1224 
1225 		/* If the packet is GSO then we will have just set up the
1226 		 * transport header offset in checksum_setup so it's now
1227 		 * straightforward to calculate gso_segs.
1228 		 */
1229 		if (skb_is_gso(skb)) {
1230 			int mss, hdrlen;
1231 
1232 			/* GSO implies having the L4 header. */
1233 			WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1234 			if (unlikely(!skb_transport_header_was_set(skb))) {
1235 				kfree_skb(skb);
1236 				continue;
1237 			}
1238 
1239 			mss = skb_shinfo(skb)->gso_size;
1240 			hdrlen = skb_tcp_all_headers(skb);
1241 
1242 			skb_shinfo(skb)->gso_segs =
1243 				DIV_ROUND_UP(skb->len - hdrlen, mss);
1244 		}
1245 
1246 		queue->stats.rx_bytes += skb->len;
1247 		queue->stats.rx_packets++;
1248 
1249 		work_done++;
1250 
1251 		/* Set this flag right before netif_receive_skb, otherwise
1252 		 * someone might think this packet already left netback, and
1253 		 * do a skb_copy_ubufs while we are still in control of the
1254 		 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1255 		 */
1256 		if (skb_shinfo(skb)->destructor_arg) {
1257 			xenvif_skb_zerocopy_prepare(queue, skb);
1258 			queue->stats.tx_zerocopy_sent++;
1259 		}
1260 
1261 		netif_receive_skb(skb);
1262 	}
1263 
1264 	return work_done;
1265 }
1266 
1267 void xenvif_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *ubuf_base,
1268 			      bool zerocopy_success)
1269 {
1270 	unsigned long flags;
1271 	pending_ring_idx_t index;
1272 	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1273 	struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1274 
1275 	/* This is the only place where we grab this lock, to protect callbacks
1276 	 * from each other.
1277 	 */
1278 	spin_lock_irqsave(&queue->callback_lock, flags);
1279 	do {
1280 		u16 pending_idx = ubuf->desc;
1281 		ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1282 		BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1283 			MAX_PENDING_REQS);
1284 		index = pending_index(queue->dealloc_prod);
1285 		queue->dealloc_ring[index] = pending_idx;
1286 		/* Sync with xenvif_tx_dealloc_action:
1287 		 * insert idx then incr producer.
1288 		 */
1289 		smp_wmb();
1290 		queue->dealloc_prod++;
1291 	} while (ubuf);
1292 	spin_unlock_irqrestore(&queue->callback_lock, flags);
1293 
1294 	if (likely(zerocopy_success))
1295 		queue->stats.tx_zerocopy_success++;
1296 	else
1297 		queue->stats.tx_zerocopy_fail++;
1298 	xenvif_skb_zerocopy_complete(queue);
1299 }
1300 
1301 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1302 {
1303 	struct gnttab_unmap_grant_ref *gop;
1304 	pending_ring_idx_t dc, dp;
1305 	u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1306 	unsigned int i = 0;
1307 
1308 	dc = queue->dealloc_cons;
1309 	gop = queue->tx_unmap_ops;
1310 
1311 	/* Free up any grants we have finished using */
1312 	do {
1313 		dp = queue->dealloc_prod;
1314 
1315 		/* Ensure we see all indices enqueued by all
1316 		 * xenvif_zerocopy_callback().
1317 		 */
1318 		smp_rmb();
1319 
1320 		while (dc != dp) {
1321 			BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1322 			pending_idx =
1323 				queue->dealloc_ring[pending_index(dc++)];
1324 
1325 			pending_idx_release[gop - queue->tx_unmap_ops] =
1326 				pending_idx;
1327 			queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1328 				queue->mmap_pages[pending_idx];
1329 			gnttab_set_unmap_op(gop,
1330 					    idx_to_kaddr(queue, pending_idx),
1331 					    GNTMAP_host_map,
1332 					    queue->grant_tx_handle[pending_idx]);
1333 			xenvif_grant_handle_reset(queue, pending_idx);
1334 			++gop;
1335 		}
1336 
1337 	} while (dp != queue->dealloc_prod);
1338 
1339 	queue->dealloc_cons = dc;
1340 
1341 	if (gop - queue->tx_unmap_ops > 0) {
1342 		int ret;
1343 		ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1344 					NULL,
1345 					queue->pages_to_unmap,
1346 					gop - queue->tx_unmap_ops);
1347 		if (ret) {
1348 			netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1349 				   gop - queue->tx_unmap_ops, ret);
1350 			for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1351 				if (gop[i].status != GNTST_okay)
1352 					netdev_err(queue->vif->dev,
1353 						   " host_addr: 0x%llx handle: 0x%x status: %d\n",
1354 						   gop[i].host_addr,
1355 						   gop[i].handle,
1356 						   gop[i].status);
1357 			}
1358 			BUG();
1359 		}
1360 	}
1361 
1362 	for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1363 		xenvif_idx_release(queue, pending_idx_release[i],
1364 				   XEN_NETIF_RSP_OKAY);
1365 }
1366 
1367 
1368 /* Called after netfront has transmitted */
1369 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1370 {
1371 	unsigned nr_mops = 0, nr_cops = 0;
1372 	int work_done, ret;
1373 
1374 	if (unlikely(!tx_work_todo(queue)))
1375 		return 0;
1376 
1377 	xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1378 
1379 	if (nr_cops == 0)
1380 		return 0;
1381 
1382 	gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1383 	if (nr_mops != 0) {
1384 		ret = gnttab_map_refs(queue->tx_map_ops,
1385 				      NULL,
1386 				      queue->pages_to_map,
1387 				      nr_mops);
1388 		if (ret) {
1389 			unsigned int i;
1390 
1391 			netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1392 				   nr_mops, ret);
1393 			for (i = 0; i < nr_mops; ++i)
1394 				WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1395 				             GNTST_okay);
1396 		}
1397 	}
1398 
1399 	work_done = xenvif_tx_submit(queue);
1400 
1401 	return work_done;
1402 }
1403 
1404 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1405 			       u8 status)
1406 {
1407 	struct pending_tx_info *pending_tx_info;
1408 	pending_ring_idx_t index;
1409 	unsigned long flags;
1410 
1411 	pending_tx_info = &queue->pending_tx_info[pending_idx];
1412 
1413 	spin_lock_irqsave(&queue->response_lock, flags);
1414 
1415 	make_tx_response(queue, &pending_tx_info->req,
1416 			 pending_tx_info->extra_count, status);
1417 
1418 	/* Release the pending index before pusing the Tx response so
1419 	 * its available before a new Tx request is pushed by the
1420 	 * frontend.
1421 	 */
1422 	index = pending_index(queue->pending_prod++);
1423 	queue->pending_ring[index] = pending_idx;
1424 
1425 	push_tx_responses(queue);
1426 
1427 	spin_unlock_irqrestore(&queue->response_lock, flags);
1428 }
1429 
1430 
1431 static void make_tx_response(struct xenvif_queue *queue,
1432 			     struct xen_netif_tx_request *txp,
1433 			     unsigned int extra_count,
1434 			     s8       st)
1435 {
1436 	RING_IDX i = queue->tx.rsp_prod_pvt;
1437 	struct xen_netif_tx_response *resp;
1438 
1439 	resp = RING_GET_RESPONSE(&queue->tx, i);
1440 	resp->id     = txp->id;
1441 	resp->status = st;
1442 
1443 	while (extra_count-- != 0)
1444 		RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1445 
1446 	queue->tx.rsp_prod_pvt = ++i;
1447 }
1448 
1449 static void push_tx_responses(struct xenvif_queue *queue)
1450 {
1451 	int notify;
1452 
1453 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1454 	if (notify)
1455 		notify_remote_via_irq(queue->tx_irq);
1456 }
1457 
1458 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1459 {
1460 	int ret;
1461 	struct gnttab_unmap_grant_ref tx_unmap_op;
1462 
1463 	gnttab_set_unmap_op(&tx_unmap_op,
1464 			    idx_to_kaddr(queue, pending_idx),
1465 			    GNTMAP_host_map,
1466 			    queue->grant_tx_handle[pending_idx]);
1467 	xenvif_grant_handle_reset(queue, pending_idx);
1468 
1469 	ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1470 				&queue->mmap_pages[pending_idx], 1);
1471 	if (ret) {
1472 		netdev_err(queue->vif->dev,
1473 			   "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1474 			   ret,
1475 			   pending_idx,
1476 			   tx_unmap_op.host_addr,
1477 			   tx_unmap_op.handle,
1478 			   tx_unmap_op.status);
1479 		BUG();
1480 	}
1481 }
1482 
1483 static inline int tx_work_todo(struct xenvif_queue *queue)
1484 {
1485 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1486 		return 1;
1487 
1488 	return 0;
1489 }
1490 
1491 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1492 {
1493 	return queue->dealloc_cons != queue->dealloc_prod;
1494 }
1495 
1496 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1497 {
1498 	if (queue->tx.sring)
1499 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1500 					queue->tx.sring);
1501 	if (queue->rx.sring)
1502 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1503 					queue->rx.sring);
1504 }
1505 
1506 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1507 				   grant_ref_t tx_ring_ref,
1508 				   grant_ref_t rx_ring_ref)
1509 {
1510 	void *addr;
1511 	struct xen_netif_tx_sring *txs;
1512 	struct xen_netif_rx_sring *rxs;
1513 	RING_IDX rsp_prod, req_prod;
1514 	int err;
1515 
1516 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1517 				     &tx_ring_ref, 1, &addr);
1518 	if (err)
1519 		goto err;
1520 
1521 	txs = (struct xen_netif_tx_sring *)addr;
1522 	rsp_prod = READ_ONCE(txs->rsp_prod);
1523 	req_prod = READ_ONCE(txs->req_prod);
1524 
1525 	BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1526 
1527 	err = -EIO;
1528 	if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1529 		goto err;
1530 
1531 	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1532 				     &rx_ring_ref, 1, &addr);
1533 	if (err)
1534 		goto err;
1535 
1536 	rxs = (struct xen_netif_rx_sring *)addr;
1537 	rsp_prod = READ_ONCE(rxs->rsp_prod);
1538 	req_prod = READ_ONCE(rxs->req_prod);
1539 
1540 	BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1541 
1542 	err = -EIO;
1543 	if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1544 		goto err;
1545 
1546 	return 0;
1547 
1548 err:
1549 	xenvif_unmap_frontend_data_rings(queue);
1550 	return err;
1551 }
1552 
1553 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1554 {
1555 	/* Dealloc thread must remain running until all inflight
1556 	 * packets complete.
1557 	 */
1558 	return kthread_should_stop() &&
1559 		!atomic_read(&queue->inflight_packets);
1560 }
1561 
1562 int xenvif_dealloc_kthread(void *data)
1563 {
1564 	struct xenvif_queue *queue = data;
1565 
1566 	for (;;) {
1567 		wait_event_interruptible(queue->dealloc_wq,
1568 					 tx_dealloc_work_todo(queue) ||
1569 					 xenvif_dealloc_kthread_should_stop(queue));
1570 		if (xenvif_dealloc_kthread_should_stop(queue))
1571 			break;
1572 
1573 		xenvif_tx_dealloc_action(queue);
1574 		cond_resched();
1575 	}
1576 
1577 	/* Unmap anything remaining*/
1578 	if (tx_dealloc_work_todo(queue))
1579 		xenvif_tx_dealloc_action(queue);
1580 
1581 	return 0;
1582 }
1583 
1584 static void make_ctrl_response(struct xenvif *vif,
1585 			       const struct xen_netif_ctrl_request *req,
1586 			       u32 status, u32 data)
1587 {
1588 	RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1589 	struct xen_netif_ctrl_response rsp = {
1590 		.id = req->id,
1591 		.type = req->type,
1592 		.status = status,
1593 		.data = data,
1594 	};
1595 
1596 	*RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1597 	vif->ctrl.rsp_prod_pvt = ++idx;
1598 }
1599 
1600 static void push_ctrl_response(struct xenvif *vif)
1601 {
1602 	int notify;
1603 
1604 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1605 	if (notify)
1606 		notify_remote_via_irq(vif->ctrl_irq);
1607 }
1608 
1609 static void process_ctrl_request(struct xenvif *vif,
1610 				 const struct xen_netif_ctrl_request *req)
1611 {
1612 	u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1613 	u32 data = 0;
1614 
1615 	switch (req->type) {
1616 	case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1617 		status = xenvif_set_hash_alg(vif, req->data[0]);
1618 		break;
1619 
1620 	case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1621 		status = xenvif_get_hash_flags(vif, &data);
1622 		break;
1623 
1624 	case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1625 		status = xenvif_set_hash_flags(vif, req->data[0]);
1626 		break;
1627 
1628 	case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1629 		status = xenvif_set_hash_key(vif, req->data[0],
1630 					     req->data[1]);
1631 		break;
1632 
1633 	case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1634 		status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1635 		data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1636 		break;
1637 
1638 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1639 		status = xenvif_set_hash_mapping_size(vif,
1640 						      req->data[0]);
1641 		break;
1642 
1643 	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1644 		status = xenvif_set_hash_mapping(vif, req->data[0],
1645 						 req->data[1],
1646 						 req->data[2]);
1647 		break;
1648 
1649 	default:
1650 		break;
1651 	}
1652 
1653 	make_ctrl_response(vif, req, status, data);
1654 	push_ctrl_response(vif);
1655 }
1656 
1657 static void xenvif_ctrl_action(struct xenvif *vif)
1658 {
1659 	for (;;) {
1660 		RING_IDX req_prod, req_cons;
1661 
1662 		req_prod = vif->ctrl.sring->req_prod;
1663 		req_cons = vif->ctrl.req_cons;
1664 
1665 		/* Make sure we can see requests before we process them. */
1666 		rmb();
1667 
1668 		if (req_cons == req_prod)
1669 			break;
1670 
1671 		while (req_cons != req_prod) {
1672 			struct xen_netif_ctrl_request req;
1673 
1674 			RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1675 			req_cons++;
1676 
1677 			process_ctrl_request(vif, &req);
1678 		}
1679 
1680 		vif->ctrl.req_cons = req_cons;
1681 		vif->ctrl.sring->req_event = req_cons + 1;
1682 	}
1683 }
1684 
1685 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1686 {
1687 	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1688 		return true;
1689 
1690 	return false;
1691 }
1692 
1693 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1694 {
1695 	struct xenvif *vif = data;
1696 	unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1697 
1698 	while (xenvif_ctrl_work_todo(vif)) {
1699 		xenvif_ctrl_action(vif);
1700 		eoi_flag = 0;
1701 	}
1702 
1703 	xen_irq_lateeoi(irq, eoi_flag);
1704 
1705 	return IRQ_HANDLED;
1706 }
1707 
1708 static int __init netback_init(void)
1709 {
1710 	int rc = 0;
1711 
1712 	if (!xen_domain())
1713 		return -ENODEV;
1714 
1715 	/* Allow as many queues as there are CPUs but max. 8 if user has not
1716 	 * specified a value.
1717 	 */
1718 	if (xenvif_max_queues == 0)
1719 		xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1720 					  num_online_cpus());
1721 
1722 	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1723 		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1724 			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1725 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1726 	}
1727 
1728 	rc = xenvif_xenbus_init();
1729 	if (rc)
1730 		goto failed_init;
1731 
1732 #ifdef CONFIG_DEBUG_FS
1733 	xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1734 #endif /* CONFIG_DEBUG_FS */
1735 
1736 	return 0;
1737 
1738 failed_init:
1739 	return rc;
1740 }
1741 
1742 module_init(netback_init);
1743 
1744 static void __exit netback_fini(void)
1745 {
1746 #ifdef CONFIG_DEBUG_FS
1747 	debugfs_remove_recursive(xen_netback_dbg_root);
1748 #endif /* CONFIG_DEBUG_FS */
1749 	xenvif_xenbus_fini();
1750 }
1751 module_exit(netback_fini);
1752 
1753 MODULE_LICENSE("Dual BSD/GPL");
1754 MODULE_ALIAS("xen-backend:vif");
1755