xref: /openbmc/linux/drivers/net/xen-netback/rx.c (revision c4f7ac64) 
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 	unsigned long flags;
42 
43 	spin_lock_irqsave(&queue->rx_queue.lock, flags);
44 
45 	skb = skb_peek(&queue->rx_queue);
46 	if (!skb) {
47 		spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
48 		return false;
49 	}
50 
51 	needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
52 	if (skb_is_gso(skb))
53 		needed++;
54 	if (skb->sw_hash)
55 		needed++;
56 
57 	spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
58 
59 	do {
60 		prod = queue->rx.sring->req_prod;
61 		cons = queue->rx.req_cons;
62 
63 		if (prod - cons >= needed)
64 			return true;
65 
66 		queue->rx.sring->req_event = prod + 1;
67 
68 		/* Make sure event is visible before we check prod
69 		 * again.
70 		 */
71 		mb();
72 	} while (queue->rx.sring->req_prod != prod);
73 
74 	return false;
75 }
76 
77 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
78 {
79 	unsigned long flags;
80 
81 	spin_lock_irqsave(&queue->rx_queue.lock, flags);
82 
83 	__skb_queue_tail(&queue->rx_queue, skb);
84 
85 	queue->rx_queue_len += skb->len;
86 	if (queue->rx_queue_len > queue->rx_queue_max) {
87 		struct net_device *dev = queue->vif->dev;
88 
89 		netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
90 	}
91 
92 	spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
93 }
94 
95 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
96 {
97 	struct sk_buff *skb;
98 
99 	spin_lock_irq(&queue->rx_queue.lock);
100 
101 	skb = __skb_dequeue(&queue->rx_queue);
102 	if (skb) {
103 		queue->rx_queue_len -= skb->len;
104 		if (queue->rx_queue_len < queue->rx_queue_max) {
105 			struct netdev_queue *txq;
106 
107 			txq = netdev_get_tx_queue(queue->vif->dev, queue->id);
108 			netif_tx_wake_queue(txq);
109 		}
110 	}
111 
112 	spin_unlock_irq(&queue->rx_queue.lock);
113 
114 	return skb;
115 }
116 
117 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
118 {
119 	struct sk_buff *skb;
120 
121 	while ((skb = xenvif_rx_dequeue(queue)) != NULL)
122 		kfree_skb(skb);
123 }
124 
125 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
126 {
127 	struct sk_buff *skb;
128 
129 	for (;;) {
130 		skb = skb_peek(&queue->rx_queue);
131 		if (!skb)
132 			break;
133 		if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
134 			break;
135 		xenvif_rx_dequeue(queue);
136 		kfree_skb(skb);
137 	}
138 }
139 
140 static void xenvif_rx_copy_flush(struct xenvif_queue *queue)
141 {
142 	unsigned int i;
143 	int notify;
144 
145 	gnttab_batch_copy(queue->rx_copy.op, queue->rx_copy.num);
146 
147 	for (i = 0; i < queue->rx_copy.num; i++) {
148 		struct gnttab_copy *op;
149 
150 		op = &queue->rx_copy.op[i];
151 
152 		/* If the copy failed, overwrite the status field in
153 		 * the corresponding response.
154 		 */
155 		if (unlikely(op->status != GNTST_okay)) {
156 			struct xen_netif_rx_response *rsp;
157 
158 			rsp = RING_GET_RESPONSE(&queue->rx,
159 						queue->rx_copy.idx[i]);
160 			rsp->status = op->status;
161 		}
162 	}
163 
164 	queue->rx_copy.num = 0;
165 
166 	/* Push responses for all completed packets. */
167 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, notify);
168 	if (notify)
169 		notify_remote_via_irq(queue->rx_irq);
170 
171 	__skb_queue_purge(queue->rx_copy.completed);
172 }
173 
174 static void xenvif_rx_copy_add(struct xenvif_queue *queue,
175 			       struct xen_netif_rx_request *req,
176 			       unsigned int offset, void *data, size_t len)
177 {
178 	struct gnttab_copy *op;
179 	struct page *page;
180 	struct xen_page_foreign *foreign;
181 
182 	if (queue->rx_copy.num == COPY_BATCH_SIZE)
183 		xenvif_rx_copy_flush(queue);
184 
185 	op = &queue->rx_copy.op[queue->rx_copy.num];
186 
187 	page = virt_to_page(data);
188 
189 	op->flags = GNTCOPY_dest_gref;
190 
191 	foreign = xen_page_foreign(page);
192 	if (foreign) {
193 		op->source.domid = foreign->domid;
194 		op->source.u.ref = foreign->gref;
195 		op->flags |= GNTCOPY_source_gref;
196 	} else {
197 		op->source.u.gmfn = virt_to_gfn(data);
198 		op->source.domid  = DOMID_SELF;
199 	}
200 
201 	op->source.offset = xen_offset_in_page(data);
202 	op->dest.u.ref    = req->gref;
203 	op->dest.domid    = queue->vif->domid;
204 	op->dest.offset   = offset;
205 	op->len           = len;
206 
207 	queue->rx_copy.idx[queue->rx_copy.num] = queue->rx.req_cons;
208 	queue->rx_copy.num++;
209 }
210 
211 static unsigned int xenvif_gso_type(struct sk_buff *skb)
212 {
213 	if (skb_is_gso(skb)) {
214 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
215 			return XEN_NETIF_GSO_TYPE_TCPV4;
216 		else
217 			return XEN_NETIF_GSO_TYPE_TCPV6;
218 	}
219 	return XEN_NETIF_GSO_TYPE_NONE;
220 }
221 
222 struct xenvif_pkt_state {
223 	struct sk_buff *skb;
224 	size_t remaining_len;
225 	struct sk_buff *frag_iter;
226 	int frag; /* frag == -1 => frag_iter->head */
227 	unsigned int frag_offset;
228 	struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
229 	unsigned int extra_count;
230 	unsigned int slot;
231 };
232 
233 static void xenvif_rx_next_skb(struct xenvif_queue *queue,
234 			       struct xenvif_pkt_state *pkt)
235 {
236 	struct sk_buff *skb;
237 	unsigned int gso_type;
238 
239 	skb = xenvif_rx_dequeue(queue);
240 
241 	queue->stats.tx_bytes += skb->len;
242 	queue->stats.tx_packets++;
243 
244 	/* Reset packet state. */
245 	memset(pkt, 0, sizeof(struct xenvif_pkt_state));
246 
247 	pkt->skb = skb;
248 	pkt->frag_iter = skb;
249 	pkt->remaining_len = skb->len;
250 	pkt->frag = -1;
251 
252 	gso_type = xenvif_gso_type(skb);
253 	if ((1 << gso_type) & queue->vif->gso_mask) {
254 		struct xen_netif_extra_info *extra;
255 
256 		extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
257 
258 		extra->u.gso.type = gso_type;
259 		extra->u.gso.size = skb_shinfo(skb)->gso_size;
260 		extra->u.gso.pad = 0;
261 		extra->u.gso.features = 0;
262 		extra->type = XEN_NETIF_EXTRA_TYPE_GSO;
263 		extra->flags = 0;
264 
265 		pkt->extra_count++;
266 	}
267 
268 	if (queue->vif->xdp_headroom) {
269 		struct xen_netif_extra_info *extra;
270 
271 		extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_XDP - 1];
272 
273 		memset(extra, 0, sizeof(struct xen_netif_extra_info));
274 		extra->u.xdp.headroom = queue->vif->xdp_headroom;
275 		extra->type = XEN_NETIF_EXTRA_TYPE_XDP;
276 		extra->flags = 0;
277 
278 		pkt->extra_count++;
279 	}
280 
281 	if (skb->sw_hash) {
282 		struct xen_netif_extra_info *extra;
283 
284 		extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
285 
286 		extra->u.hash.algorithm =
287 			XEN_NETIF_CTRL_HASH_ALGORITHM_TOEPLITZ;
288 
289 		if (skb->l4_hash)
290 			extra->u.hash.type =
291 				skb->protocol == htons(ETH_P_IP) ?
292 				_XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP :
293 				_XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP;
294 		else
295 			extra->u.hash.type =
296 				skb->protocol == htons(ETH_P_IP) ?
297 				_XEN_NETIF_CTRL_HASH_TYPE_IPV4 :
298 				_XEN_NETIF_CTRL_HASH_TYPE_IPV6;
299 
300 		*(uint32_t *)extra->u.hash.value = skb_get_hash_raw(skb);
301 
302 		extra->type = XEN_NETIF_EXTRA_TYPE_HASH;
303 		extra->flags = 0;
304 
305 		pkt->extra_count++;
306 	}
307 }
308 
309 static void xenvif_rx_complete(struct xenvif_queue *queue,
310 			       struct xenvif_pkt_state *pkt)
311 {
312 	/* All responses are ready to be pushed. */
313 	queue->rx.rsp_prod_pvt = queue->rx.req_cons;
314 
315 	__skb_queue_tail(queue->rx_copy.completed, pkt->skb);
316 }
317 
318 static void xenvif_rx_next_frag(struct xenvif_pkt_state *pkt)
319 {
320 	struct sk_buff *frag_iter = pkt->frag_iter;
321 	unsigned int nr_frags = skb_shinfo(frag_iter)->nr_frags;
322 
323 	pkt->frag++;
324 	pkt->frag_offset = 0;
325 
326 	if (pkt->frag >= nr_frags) {
327 		if (frag_iter == pkt->skb)
328 			pkt->frag_iter = skb_shinfo(frag_iter)->frag_list;
329 		else
330 			pkt->frag_iter = frag_iter->next;
331 
332 		pkt->frag = -1;
333 	}
334 }
335 
336 static void xenvif_rx_next_chunk(struct xenvif_queue *queue,
337 				 struct xenvif_pkt_state *pkt,
338 				 unsigned int offset, void **data,
339 				 size_t *len)
340 {
341 	struct sk_buff *frag_iter = pkt->frag_iter;
342 	void *frag_data;
343 	size_t frag_len, chunk_len;
344 
345 	BUG_ON(!frag_iter);
346 
347 	if (pkt->frag == -1) {
348 		frag_data = frag_iter->data;
349 		frag_len = skb_headlen(frag_iter);
350 	} else {
351 		skb_frag_t *frag = &skb_shinfo(frag_iter)->frags[pkt->frag];
352 
353 		frag_data = skb_frag_address(frag);
354 		frag_len = skb_frag_size(frag);
355 	}
356 
357 	frag_data += pkt->frag_offset;
358 	frag_len -= pkt->frag_offset;
359 
360 	chunk_len = min_t(size_t, frag_len, XEN_PAGE_SIZE - offset);
361 	chunk_len = min_t(size_t, chunk_len, XEN_PAGE_SIZE -
362 					     xen_offset_in_page(frag_data));
363 
364 	pkt->frag_offset += chunk_len;
365 
366 	/* Advance to next frag? */
367 	if (frag_len == chunk_len)
368 		xenvif_rx_next_frag(pkt);
369 
370 	*data = frag_data;
371 	*len = chunk_len;
372 }
373 
374 static void xenvif_rx_data_slot(struct xenvif_queue *queue,
375 				struct xenvif_pkt_state *pkt,
376 				struct xen_netif_rx_request *req,
377 				struct xen_netif_rx_response *rsp)
378 {
379 	unsigned int offset = queue->vif->xdp_headroom;
380 	unsigned int flags;
381 
382 	do {
383 		size_t len;
384 		void *data;
385 
386 		xenvif_rx_next_chunk(queue, pkt, offset, &data, &len);
387 		xenvif_rx_copy_add(queue, req, offset, data, len);
388 
389 		offset += len;
390 		pkt->remaining_len -= len;
391 
392 	} while (offset < XEN_PAGE_SIZE && pkt->remaining_len > 0);
393 
394 	if (pkt->remaining_len > 0)
395 		flags = XEN_NETRXF_more_data;
396 	else
397 		flags = 0;
398 
399 	if (pkt->slot == 0) {
400 		struct sk_buff *skb = pkt->skb;
401 
402 		if (skb->ip_summed == CHECKSUM_PARTIAL)
403 			flags |= XEN_NETRXF_csum_blank |
404 				 XEN_NETRXF_data_validated;
405 		else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
406 			flags |= XEN_NETRXF_data_validated;
407 
408 		if (pkt->extra_count != 0)
409 			flags |= XEN_NETRXF_extra_info;
410 	}
411 
412 	rsp->offset = 0;
413 	rsp->flags = flags;
414 	rsp->id = req->id;
415 	rsp->status = (s16)offset;
416 }
417 
418 static void xenvif_rx_extra_slot(struct xenvif_queue *queue,
419 				 struct xenvif_pkt_state *pkt,
420 				 struct xen_netif_rx_request *req,
421 				 struct xen_netif_rx_response *rsp)
422 {
423 	struct xen_netif_extra_info *extra = (void *)rsp;
424 	unsigned int i;
425 
426 	pkt->extra_count--;
427 
428 	for (i = 0; i < ARRAY_SIZE(pkt->extras); i++) {
429 		if (pkt->extras[i].type) {
430 			*extra = pkt->extras[i];
431 
432 			if (pkt->extra_count != 0)
433 				extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
434 
435 			pkt->extras[i].type = 0;
436 			return;
437 		}
438 	}
439 	BUG();
440 }
441 
442 static void xenvif_rx_skb(struct xenvif_queue *queue)
443 {
444 	struct xenvif_pkt_state pkt;
445 
446 	xenvif_rx_next_skb(queue, &pkt);
447 
448 	queue->last_rx_time = jiffies;
449 
450 	do {
451 		struct xen_netif_rx_request *req;
452 		struct xen_netif_rx_response *rsp;
453 
454 		req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons);
455 		rsp = RING_GET_RESPONSE(&queue->rx, queue->rx.req_cons);
456 
457 		/* Extras must go after the first data slot */
458 		if (pkt.slot != 0 && pkt.extra_count != 0)
459 			xenvif_rx_extra_slot(queue, &pkt, req, rsp);
460 		else
461 			xenvif_rx_data_slot(queue, &pkt, req, rsp);
462 
463 		queue->rx.req_cons++;
464 		pkt.slot++;
465 	} while (pkt.remaining_len > 0 || pkt.extra_count != 0);
466 
467 	xenvif_rx_complete(queue, &pkt);
468 }
469 
470 #define RX_BATCH_SIZE 64
471 
472 void xenvif_rx_action(struct xenvif_queue *queue)
473 {
474 	struct sk_buff_head completed_skbs;
475 	unsigned int work_done = 0;
476 
477 	__skb_queue_head_init(&completed_skbs);
478 	queue->rx_copy.completed = &completed_skbs;
479 
480 	while (xenvif_rx_ring_slots_available(queue) &&
481 	       work_done < RX_BATCH_SIZE) {
482 		xenvif_rx_skb(queue);
483 		work_done++;
484 	}
485 
486 	/* Flush any pending copies and complete all skbs. */
487 	xenvif_rx_copy_flush(queue);
488 }
489 
490 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
491 {
492 	RING_IDX prod, cons;
493 
494 	prod = queue->rx.sring->req_prod;
495 	cons = queue->rx.req_cons;
496 
497 	return !queue->stalled &&
498 		prod - cons < 1 &&
499 		time_after(jiffies,
500 			   queue->last_rx_time + queue->vif->stall_timeout);
501 }
502 
503 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
504 {
505 	RING_IDX prod, cons;
506 
507 	prod = queue->rx.sring->req_prod;
508 	cons = queue->rx.req_cons;
509 
510 	return queue->stalled && prod - cons >= 1;
511 }
512 
513 bool xenvif_have_rx_work(struct xenvif_queue *queue, bool test_kthread)
514 {
515 	return xenvif_rx_ring_slots_available(queue) ||
516 		(queue->vif->stall_timeout &&
517 		 (xenvif_rx_queue_stalled(queue) ||
518 		  xenvif_rx_queue_ready(queue))) ||
519 		(test_kthread && kthread_should_stop()) ||
520 		queue->vif->disabled;
521 }
522 
523 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
524 {
525 	struct sk_buff *skb;
526 	long timeout;
527 
528 	skb = skb_peek(&queue->rx_queue);
529 	if (!skb)
530 		return MAX_SCHEDULE_TIMEOUT;
531 
532 	timeout = XENVIF_RX_CB(skb)->expires - jiffies;
533 	return timeout < 0 ? 0 : timeout;
534 }
535 
536 /* Wait until the guest Rx thread has work.
537  *
538  * The timeout needs to be adjusted based on the current head of the
539  * queue (and not just the head at the beginning).  In particular, if
540  * the queue is initially empty an infinite timeout is used and this
541  * needs to be reduced when a skb is queued.
542  *
543  * This cannot be done with wait_event_timeout() because it only
544  * calculates the timeout once.
545  */
546 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
547 {
548 	DEFINE_WAIT(wait);
549 
550 	if (xenvif_have_rx_work(queue, true))
551 		return;
552 
553 	for (;;) {
554 		long ret;
555 
556 		prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
557 		if (xenvif_have_rx_work(queue, true))
558 			break;
559 		if (atomic_fetch_andnot(NETBK_RX_EOI | NETBK_COMMON_EOI,
560 					&queue->eoi_pending) &
561 		    (NETBK_RX_EOI | NETBK_COMMON_EOI))
562 			xen_irq_lateeoi(queue->rx_irq, 0);
563 
564 		ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
565 		if (!ret)
566 			break;
567 	}
568 	finish_wait(&queue->wq, &wait);
569 }
570 
571 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
572 {
573 	struct xenvif *vif = queue->vif;
574 
575 	queue->stalled = true;
576 
577 	/* At least one queue has stalled? Disable the carrier. */
578 	spin_lock(&vif->lock);
579 	if (vif->stalled_queues++ == 0) {
580 		netdev_info(vif->dev, "Guest Rx stalled");
581 		netif_carrier_off(vif->dev);
582 	}
583 	spin_unlock(&vif->lock);
584 }
585 
586 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
587 {
588 	struct xenvif *vif = queue->vif;
589 
590 	queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
591 	queue->stalled = false;
592 
593 	/* All queues are ready? Enable the carrier. */
594 	spin_lock(&vif->lock);
595 	if (--vif->stalled_queues == 0) {
596 		netdev_info(vif->dev, "Guest Rx ready");
597 		netif_carrier_on(vif->dev);
598 	}
599 	spin_unlock(&vif->lock);
600 }
601 
602 int xenvif_kthread_guest_rx(void *data)
603 {
604 	struct xenvif_queue *queue = data;
605 	struct xenvif *vif = queue->vif;
606 
607 	if (!vif->stall_timeout)
608 		xenvif_queue_carrier_on(queue);
609 
610 	for (;;) {
611 		xenvif_wait_for_rx_work(queue);
612 
613 		if (kthread_should_stop())
614 			break;
615 
616 		/* This frontend is found to be rogue, disable it in
617 		 * kthread context. Currently this is only set when
618 		 * netback finds out frontend sends malformed packet,
619 		 * but we cannot disable the interface in softirq
620 		 * context so we defer it here, if this thread is
621 		 * associated with queue 0.
622 		 */
623 		if (unlikely(vif->disabled && queue->id == 0)) {
624 			xenvif_carrier_off(vif);
625 			break;
626 		}
627 
628 		if (!skb_queue_empty(&queue->rx_queue))
629 			xenvif_rx_action(queue);
630 
631 		/* If the guest hasn't provided any Rx slots for a
632 		 * while it's probably not responsive, drop the
633 		 * carrier so packets are dropped earlier.
634 		 */
635 		if (vif->stall_timeout) {
636 			if (xenvif_rx_queue_stalled(queue))
637 				xenvif_queue_carrier_off(queue);
638 			else if (xenvif_rx_queue_ready(queue))
639 				xenvif_queue_carrier_on(queue);
640 		}
641 
642 		/* Queued packets may have foreign pages from other
643 		 * domains.  These cannot be queued indefinitely as
644 		 * this would starve guests of grant refs and transmit
645 		 * slots.
646 		 */
647 		xenvif_rx_queue_drop_expired(queue);
648 
649 		cond_resched();
650 	}
651 
652 	/* Bin any remaining skbs */
653 	xenvif_rx_queue_purge(queue);
654 
655 	return 0;
656 }
657