xref: /openbmc/linux/drivers/net/xen-netback/rx.c (revision e953aeaa) 
1 /*
2  * Copyright (c) 2016 Citrix Systems Inc.
3  * Copyright (c) 2002-2005, K A Fraser
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License version 2
7  * as published by the Free Software Foundation; or, when distributed
8  * separately from the Linux kernel or incorporated into other
9  * software packages, subject to the following license:
10  *
11  * Permission is hereby granted, free of charge, to any person obtaining a copy
12  * of this source file (the "Software"), to deal in the Software without
13  * restriction, including without limitation the rights to use, copy, modify,
14  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
15  * and to permit persons to whom the Software is furnished to do so, subject to
16  * the following conditions:
17  *
18  * The above copyright notice and this permission notice shall be included in
19  * all copies or substantial portions of the Software.
20  *
21  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
22  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
23  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
24  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
25  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
26  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
27  * IN THE SOFTWARE.
28  */
29 #include "common.h"
30 
31 #include <linux/kthread.h>
32 
33 #include <xen/xen.h>
34 #include <xen/events.h>
35 
36 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
37 {
38 	RING_IDX prod, cons;
39 	struct sk_buff *skb;
40 	int needed;
41 
42 	skb = skb_peek(&queue->rx_queue);
43 	if (!skb)
44 		return false;
45 
46 	needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
47 	if (skb_is_gso(skb))
48 		needed++;
49 	if (skb->sw_hash)
50 		needed++;
51 
52 	do {
53 		prod = queue->rx.sring->req_prod;
54 		cons = queue->rx.req_cons;
55 
56 		if (prod - cons >= needed)
57 			return true;
58 
59 		queue->rx.sring->req_event = prod + 1;
60 
61 		/* Make sure event is visible before we check prod
62 		 * again.
63 		 */
64 		mb();
65 	} while (queue->rx.sring->req_prod != prod);
66 
67 	return false;
68 }
69 
70 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
71 {
72 	unsigned long flags;
73 
74 	spin_lock_irqsave(&queue->rx_queue.lock, flags);
75 
76 	__skb_queue_tail(&queue->rx_queue, skb);
77 
78 	queue->rx_queue_len += skb->len;
79 	if (queue->rx_queue_len > queue->rx_queue_max) {
80 		struct net_device *dev = queue->vif->dev;
81 
82 		netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
83 	}
84 
85 	spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
86 }
87 
88 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
89 {
90 	struct sk_buff *skb;
91 
92 	spin_lock_irq(&queue->rx_queue.lock);
93 
94 	skb = __skb_dequeue(&queue->rx_queue);
95 	if (skb) {
96 		queue->rx_queue_len -= skb->len;
97 		if (queue->rx_queue_len < queue->rx_queue_max) {
98 			struct netdev_queue *txq;
99 
100 			txq = netdev_get_tx_queue(queue->vif->dev, queue->id);
101 			netif_tx_wake_queue(txq);
102 		}
103 	}
104 
105 	spin_unlock_irq(&queue->rx_queue.lock);
106 
107 	return skb;
108 }
109 
110 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
111 {
112 	struct sk_buff *skb;
113 
114 	while ((skb = xenvif_rx_dequeue(queue)) != NULL)
115 		kfree_skb(skb);
116 }
117 
118 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
119 {
120 	struct sk_buff *skb;
121 
122 	for (;;) {
123 		skb = skb_peek(&queue->rx_queue);
124 		if (!skb)
125 			break;
126 		if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
127 			break;
128 		xenvif_rx_dequeue(queue);
129 		kfree_skb(skb);
130 	}
131 }
132 
133 static void xenvif_rx_copy_flush(struct xenvif_queue *queue)
134 {
135 	unsigned int i;
136 	int notify;
137 
138 	gnttab_batch_copy(queue->rx_copy.op, queue->rx_copy.num);
139 
140 	for (i = 0; i < queue->rx_copy.num; i++) {
141 		struct gnttab_copy *op;
142 
143 		op = &queue->rx_copy.op[i];
144 
145 		/* If the copy failed, overwrite the status field in
146 		 * the corresponding response.
147 		 */
148 		if (unlikely(op->status != GNTST_okay)) {
149 			struct xen_netif_rx_response *rsp;
150 
151 			rsp = RING_GET_RESPONSE(&queue->rx,
152 						queue->rx_copy.idx[i]);
153 			rsp->status = op->status;
154 		}
155 	}
156 
157 	queue->rx_copy.num = 0;
158 
159 	/* Push responses for all completed packets. */
160 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, notify);
161 	if (notify)
162 		notify_remote_via_irq(queue->rx_irq);
163 
164 	__skb_queue_purge(queue->rx_copy.completed);
165 }
166 
167 static void xenvif_rx_copy_add(struct xenvif_queue *queue,
168 			       struct xen_netif_rx_request *req,
169 			       unsigned int offset, void *data, size_t len)
170 {
171 	struct gnttab_copy *op;
172 	struct page *page;
173 	struct xen_page_foreign *foreign;
174 
175 	if (queue->rx_copy.num == COPY_BATCH_SIZE)
176 		xenvif_rx_copy_flush(queue);
177 
178 	op = &queue->rx_copy.op[queue->rx_copy.num];
179 
180 	page = virt_to_page(data);
181 
182 	op->flags = GNTCOPY_dest_gref;
183 
184 	foreign = xen_page_foreign(page);
185 	if (foreign) {
186 		op->source.domid = foreign->domid;
187 		op->source.u.ref = foreign->gref;
188 		op->flags |= GNTCOPY_source_gref;
189 	} else {
190 		op->source.u.gmfn = virt_to_gfn(data);
191 		op->source.domid  = DOMID_SELF;
192 	}
193 
194 	op->source.offset = xen_offset_in_page(data);
195 	op->dest.u.ref    = req->gref;
196 	op->dest.domid    = queue->vif->domid;
197 	op->dest.offset   = offset;
198 	op->len           = len;
199 
200 	queue->rx_copy.idx[queue->rx_copy.num] = queue->rx.req_cons;
201 	queue->rx_copy.num++;
202 }
203 
204 static unsigned int xenvif_gso_type(struct sk_buff *skb)
205 {
206 	if (skb_is_gso(skb)) {
207 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
208 			return XEN_NETIF_GSO_TYPE_TCPV4;
209 		else
210 			return XEN_NETIF_GSO_TYPE_TCPV6;
211 	}
212 	return XEN_NETIF_GSO_TYPE_NONE;
213 }
214 
215 struct xenvif_pkt_state {
216 	struct sk_buff *skb;
217 	size_t remaining_len;
218 	struct sk_buff *frag_iter;
219 	int frag; /* frag == -1 => frag_iter->head */
220 	unsigned int frag_offset;
221 	struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
222 	unsigned int extra_count;
223 	unsigned int slot;
224 };
225 
226 static void xenvif_rx_next_skb(struct xenvif_queue *queue,
227 			       struct xenvif_pkt_state *pkt)
228 {
229 	struct sk_buff *skb;
230 	unsigned int gso_type;
231 
232 	skb = xenvif_rx_dequeue(queue);
233 
234 	queue->stats.tx_bytes += skb->len;
235 	queue->stats.tx_packets++;
236 
237 	/* Reset packet state. */
238 	memset(pkt, 0, sizeof(struct xenvif_pkt_state));
239 
240 	pkt->skb = skb;
241 	pkt->frag_iter = skb;
242 	pkt->remaining_len = skb->len;
243 	pkt->frag = -1;
244 
245 	gso_type = xenvif_gso_type(skb);
246 	if ((1 << gso_type) & queue->vif->gso_mask) {
247 		struct xen_netif_extra_info *extra;
248 
249 		extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
250 
251 		extra->u.gso.type = gso_type;
252 		extra->u.gso.size = skb_shinfo(skb)->gso_size;
253 		extra->u.gso.pad = 0;
254 		extra->u.gso.features = 0;
255 		extra->type = XEN_NETIF_EXTRA_TYPE_GSO;
256 		extra->flags = 0;
257 
258 		pkt->extra_count++;
259 	}
260 
261 	if (queue->vif->xdp_headroom) {
262 		struct xen_netif_extra_info *extra;
263 
264 		extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_XDP - 1];
265 
266 		memset(extra, 0, sizeof(struct xen_netif_extra_info));
267 		extra->u.xdp.headroom = queue->vif->xdp_headroom;
268 		extra->type = XEN_NETIF_EXTRA_TYPE_XDP;
269 		extra->flags = 0;
270 
271 		pkt->extra_count++;
272 	}
273 
274 	if (skb->sw_hash) {
275 		struct xen_netif_extra_info *extra;
276 
277 		extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
278 
279 		extra->u.hash.algorithm =
280 			XEN_NETIF_CTRL_HASH_ALGORITHM_TOEPLITZ;
281 
282 		if (skb->l4_hash)
283 			extra->u.hash.type =
284 				skb->protocol == htons(ETH_P_IP) ?
285 				_XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP :
286 				_XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP;
287 		else
288 			extra->u.hash.type =
289 				skb->protocol == htons(ETH_P_IP) ?
290 				_XEN_NETIF_CTRL_HASH_TYPE_IPV4 :
291 				_XEN_NETIF_CTRL_HASH_TYPE_IPV6;
292 
293 		*(uint32_t *)extra->u.hash.value = skb_get_hash_raw(skb);
294 
295 		extra->type = XEN_NETIF_EXTRA_TYPE_HASH;
296 		extra->flags = 0;
297 
298 		pkt->extra_count++;
299 	}
300 }
301 
302 static void xenvif_rx_complete(struct xenvif_queue *queue,
303 			       struct xenvif_pkt_state *pkt)
304 {
305 	/* All responses are ready to be pushed. */
306 	queue->rx.rsp_prod_pvt = queue->rx.req_cons;
307 
308 	__skb_queue_tail(queue->rx_copy.completed, pkt->skb);
309 }
310 
311 static void xenvif_rx_next_frag(struct xenvif_pkt_state *pkt)
312 {
313 	struct sk_buff *frag_iter = pkt->frag_iter;
314 	unsigned int nr_frags = skb_shinfo(frag_iter)->nr_frags;
315 
316 	pkt->frag++;
317 	pkt->frag_offset = 0;
318 
319 	if (pkt->frag >= nr_frags) {
320 		if (frag_iter == pkt->skb)
321 			pkt->frag_iter = skb_shinfo(frag_iter)->frag_list;
322 		else
323 			pkt->frag_iter = frag_iter->next;
324 
325 		pkt->frag = -1;
326 	}
327 }
328 
329 static void xenvif_rx_next_chunk(struct xenvif_queue *queue,
330 				 struct xenvif_pkt_state *pkt,
331 				 unsigned int offset, void **data,
332 				 size_t *len)
333 {
334 	struct sk_buff *frag_iter = pkt->frag_iter;
335 	void *frag_data;
336 	size_t frag_len, chunk_len;
337 
338 	BUG_ON(!frag_iter);
339 
340 	if (pkt->frag == -1) {
341 		frag_data = frag_iter->data;
342 		frag_len = skb_headlen(frag_iter);
343 	} else {
344 		skb_frag_t *frag = &skb_shinfo(frag_iter)->frags[pkt->frag];
345 
346 		frag_data = skb_frag_address(frag);
347 		frag_len = skb_frag_size(frag);
348 	}
349 
350 	frag_data += pkt->frag_offset;
351 	frag_len -= pkt->frag_offset;
352 
353 	chunk_len = min_t(size_t, frag_len, XEN_PAGE_SIZE - offset);
354 	chunk_len = min_t(size_t, chunk_len, XEN_PAGE_SIZE -
355 					     xen_offset_in_page(frag_data));
356 
357 	pkt->frag_offset += chunk_len;
358 
359 	/* Advance to next frag? */
360 	if (frag_len == chunk_len)
361 		xenvif_rx_next_frag(pkt);
362 
363 	*data = frag_data;
364 	*len = chunk_len;
365 }
366 
367 static void xenvif_rx_data_slot(struct xenvif_queue *queue,
368 				struct xenvif_pkt_state *pkt,
369 				struct xen_netif_rx_request *req,
370 				struct xen_netif_rx_response *rsp)
371 {
372 	unsigned int offset = queue->vif->xdp_headroom;
373 	unsigned int flags;
374 
375 	do {
376 		size_t len;
377 		void *data;
378 
379 		xenvif_rx_next_chunk(queue, pkt, offset, &data, &len);
380 		xenvif_rx_copy_add(queue, req, offset, data, len);
381 
382 		offset += len;
383 		pkt->remaining_len -= len;
384 
385 	} while (offset < XEN_PAGE_SIZE && pkt->remaining_len > 0);
386 
387 	if (pkt->remaining_len > 0)
388 		flags = XEN_NETRXF_more_data;
389 	else
390 		flags = 0;
391 
392 	if (pkt->slot == 0) {
393 		struct sk_buff *skb = pkt->skb;
394 
395 		if (skb->ip_summed == CHECKSUM_PARTIAL)
396 			flags |= XEN_NETRXF_csum_blank |
397 				 XEN_NETRXF_data_validated;
398 		else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
399 			flags |= XEN_NETRXF_data_validated;
400 
401 		if (pkt->extra_count != 0)
402 			flags |= XEN_NETRXF_extra_info;
403 	}
404 
405 	rsp->offset = 0;
406 	rsp->flags = flags;
407 	rsp->id = req->id;
408 	rsp->status = (s16)offset;
409 }
410 
411 static void xenvif_rx_extra_slot(struct xenvif_queue *queue,
412 				 struct xenvif_pkt_state *pkt,
413 				 struct xen_netif_rx_request *req,
414 				 struct xen_netif_rx_response *rsp)
415 {
416 	struct xen_netif_extra_info *extra = (void *)rsp;
417 	unsigned int i;
418 
419 	pkt->extra_count--;
420 
421 	for (i = 0; i < ARRAY_SIZE(pkt->extras); i++) {
422 		if (pkt->extras[i].type) {
423 			*extra = pkt->extras[i];
424 
425 			if (pkt->extra_count != 0)
426 				extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
427 
428 			pkt->extras[i].type = 0;
429 			return;
430 		}
431 	}
432 	BUG();
433 }
434 
435 static void xenvif_rx_skb(struct xenvif_queue *queue)
436 {
437 	struct xenvif_pkt_state pkt;
438 
439 	xenvif_rx_next_skb(queue, &pkt);
440 
441 	queue->last_rx_time = jiffies;
442 
443 	do {
444 		struct xen_netif_rx_request *req;
445 		struct xen_netif_rx_response *rsp;
446 
447 		req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons);
448 		rsp = RING_GET_RESPONSE(&queue->rx, queue->rx.req_cons);
449 
450 		/* Extras must go after the first data slot */
451 		if (pkt.slot != 0 && pkt.extra_count != 0)
452 			xenvif_rx_extra_slot(queue, &pkt, req, rsp);
453 		else
454 			xenvif_rx_data_slot(queue, &pkt, req, rsp);
455 
456 		queue->rx.req_cons++;
457 		pkt.slot++;
458 	} while (pkt.remaining_len > 0 || pkt.extra_count != 0);
459 
460 	xenvif_rx_complete(queue, &pkt);
461 }
462 
463 #define RX_BATCH_SIZE 64
464 
465 void xenvif_rx_action(struct xenvif_queue *queue)
466 {
467 	struct sk_buff_head completed_skbs;
468 	unsigned int work_done = 0;
469 
470 	__skb_queue_head_init(&completed_skbs);
471 	queue->rx_copy.completed = &completed_skbs;
472 
473 	while (xenvif_rx_ring_slots_available(queue) &&
474 	       work_done < RX_BATCH_SIZE) {
475 		xenvif_rx_skb(queue);
476 		work_done++;
477 	}
478 
479 	/* Flush any pending copies and complete all skbs. */
480 	xenvif_rx_copy_flush(queue);
481 }
482 
483 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
484 {
485 	RING_IDX prod, cons;
486 
487 	prod = queue->rx.sring->req_prod;
488 	cons = queue->rx.req_cons;
489 
490 	return !queue->stalled &&
491 		prod - cons < 1 &&
492 		time_after(jiffies,
493 			   queue->last_rx_time + queue->vif->stall_timeout);
494 }
495 
496 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
497 {
498 	RING_IDX prod, cons;
499 
500 	prod = queue->rx.sring->req_prod;
501 	cons = queue->rx.req_cons;
502 
503 	return queue->stalled && prod - cons >= 1;
504 }
505 
506 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
507 {
508 	return xenvif_rx_ring_slots_available(queue) ||
509 		(queue->vif->stall_timeout &&
510 		 (xenvif_rx_queue_stalled(queue) ||
511 		  xenvif_rx_queue_ready(queue))) ||
512 		kthread_should_stop() ||
513 		queue->vif->disabled;
514 }
515 
516 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
517 {
518 	struct sk_buff *skb;
519 	long timeout;
520 
521 	skb = skb_peek(&queue->rx_queue);
522 	if (!skb)
523 		return MAX_SCHEDULE_TIMEOUT;
524 
525 	timeout = XENVIF_RX_CB(skb)->expires - jiffies;
526 	return timeout < 0 ? 0 : timeout;
527 }
528 
529 /* Wait until the guest Rx thread has work.
530  *
531  * The timeout needs to be adjusted based on the current head of the
532  * queue (and not just the head at the beginning).  In particular, if
533  * the queue is initially empty an infinite timeout is used and this
534  * needs to be reduced when a skb is queued.
535  *
536  * This cannot be done with wait_event_timeout() because it only
537  * calculates the timeout once.
538  */
539 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
540 {
541 	DEFINE_WAIT(wait);
542 
543 	if (xenvif_have_rx_work(queue))
544 		return;
545 
546 	for (;;) {
547 		long ret;
548 
549 		prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
550 		if (xenvif_have_rx_work(queue))
551 			break;
552 		ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
553 		if (!ret)
554 			break;
555 	}
556 	finish_wait(&queue->wq, &wait);
557 }
558 
559 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
560 {
561 	struct xenvif *vif = queue->vif;
562 
563 	queue->stalled = true;
564 
565 	/* At least one queue has stalled? Disable the carrier. */
566 	spin_lock(&vif->lock);
567 	if (vif->stalled_queues++ == 0) {
568 		netdev_info(vif->dev, "Guest Rx stalled");
569 		netif_carrier_off(vif->dev);
570 	}
571 	spin_unlock(&vif->lock);
572 }
573 
574 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
575 {
576 	struct xenvif *vif = queue->vif;
577 
578 	queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
579 	queue->stalled = false;
580 
581 	/* All queues are ready? Enable the carrier. */
582 	spin_lock(&vif->lock);
583 	if (--vif->stalled_queues == 0) {
584 		netdev_info(vif->dev, "Guest Rx ready");
585 		netif_carrier_on(vif->dev);
586 	}
587 	spin_unlock(&vif->lock);
588 }
589 
590 int xenvif_kthread_guest_rx(void *data)
591 {
592 	struct xenvif_queue *queue = data;
593 	struct xenvif *vif = queue->vif;
594 
595 	if (!vif->stall_timeout)
596 		xenvif_queue_carrier_on(queue);
597 
598 	for (;;) {
599 		xenvif_wait_for_rx_work(queue);
600 
601 		if (kthread_should_stop())
602 			break;
603 
604 		/* This frontend is found to be rogue, disable it in
605 		 * kthread context. Currently this is only set when
606 		 * netback finds out frontend sends malformed packet,
607 		 * but we cannot disable the interface in softirq
608 		 * context so we defer it here, if this thread is
609 		 * associated with queue 0.
610 		 */
611 		if (unlikely(vif->disabled && queue->id == 0)) {
612 			xenvif_carrier_off(vif);
613 			break;
614 		}
615 
616 		if (!skb_queue_empty(&queue->rx_queue))
617 			xenvif_rx_action(queue);
618 
619 		/* If the guest hasn't provided any Rx slots for a
620 		 * while it's probably not responsive, drop the
621 		 * carrier so packets are dropped earlier.
622 		 */
623 		if (vif->stall_timeout) {
624 			if (xenvif_rx_queue_stalled(queue))
625 				xenvif_queue_carrier_off(queue);
626 			else if (xenvif_rx_queue_ready(queue))
627 				xenvif_queue_carrier_on(queue);
628 		}
629 
630 		/* Queued packets may have foreign pages from other
631 		 * domains.  These cannot be queued indefinitely as
632 		 * this would starve guests of grant refs and transmit
633 		 * slots.
634 		 */
635 		xenvif_rx_queue_drop_expired(queue);
636 
637 		cond_resched();
638 	}
639 
640 	/* Bin any remaining skbs */
641 	xenvif_rx_queue_purge(queue);
642 
643 	return 0;
644 }
645