1 /*
2  * Network-device interface management.
3  *
4  * Copyright (c) 2004-2005, Keir Fraser
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License version 2
8  * as published by the Free Software Foundation; or, when distributed
9  * separately from the Linux kernel or incorporated into other
10  * software packages, subject to the following license:
11  *
12  * Permission is hereby granted, free of charge, to any person obtaining a copy
13  * of this source file (the "Software"), to deal in the Software without
14  * restriction, including without limitation the rights to use, copy, modify,
15  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
16  * and to permit persons to whom the Software is furnished to do so, subject to
17  * the following conditions:
18  *
19  * The above copyright notice and this permission notice shall be included in
20  * all copies or substantial portions of the Software.
21  *
22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
23  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
24  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
25  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
26  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
27  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28  * IN THE SOFTWARE.
29  */
30 
31 #include "common.h"
32 
33 #include <linux/kthread.h>
34 #include <linux/ethtool.h>
35 #include <linux/rtnetlink.h>
36 #include <linux/if_vlan.h>
37 #include <linux/vmalloc.h>
38 
39 #include <xen/events.h>
40 #include <asm/xen/hypercall.h>
41 #include <xen/balloon.h>
42 
43 #define XENVIF_QUEUE_LENGTH 32
44 #define XENVIF_NAPI_WEIGHT  64
45 
46 int xenvif_schedulable(struct xenvif *vif)
47 {
48 	return netif_running(vif->dev) && netif_carrier_ok(vif->dev);
49 }
50 
51 static irqreturn_t xenvif_tx_interrupt(int irq, void *dev_id)
52 {
53 	struct xenvif *vif = dev_id;
54 
55 	if (RING_HAS_UNCONSUMED_REQUESTS(&vif->tx))
56 		napi_schedule(&vif->napi);
57 
58 	return IRQ_HANDLED;
59 }
60 
61 static int xenvif_poll(struct napi_struct *napi, int budget)
62 {
63 	struct xenvif *vif = container_of(napi, struct xenvif, napi);
64 	int work_done;
65 
66 	/* This vif is rogue, we pretend we've there is nothing to do
67 	 * for this vif to deschedule it from NAPI. But this interface
68 	 * will be turned off in thread context later.
69 	 */
70 	if (unlikely(vif->disabled)) {
71 		napi_complete(napi);
72 		return 0;
73 	}
74 
75 	work_done = xenvif_tx_action(vif, budget);
76 
77 	if (work_done < budget) {
78 		int more_to_do = 0;
79 		unsigned long flags;
80 
81 		/* It is necessary to disable IRQ before calling
82 		 * RING_HAS_UNCONSUMED_REQUESTS. Otherwise we might
83 		 * lose event from the frontend.
84 		 *
85 		 * Consider:
86 		 *   RING_HAS_UNCONSUMED_REQUESTS
87 		 *   <frontend generates event to trigger napi_schedule>
88 		 *   __napi_complete
89 		 *
90 		 * This handler is still in scheduled state so the
91 		 * event has no effect at all. After __napi_complete
92 		 * this handler is descheduled and cannot get
93 		 * scheduled again. We lose event in this case and the ring
94 		 * will be completely stalled.
95 		 */
96 
97 		local_irq_save(flags);
98 
99 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
100 		if (!more_to_do)
101 			__napi_complete(napi);
102 
103 		local_irq_restore(flags);
104 	}
105 
106 	return work_done;
107 }
108 
109 static irqreturn_t xenvif_rx_interrupt(int irq, void *dev_id)
110 {
111 	struct xenvif *vif = dev_id;
112 
113 	xenvif_kick_thread(vif);
114 
115 	return IRQ_HANDLED;
116 }
117 
118 static irqreturn_t xenvif_interrupt(int irq, void *dev_id)
119 {
120 	xenvif_tx_interrupt(irq, dev_id);
121 	xenvif_rx_interrupt(irq, dev_id);
122 
123 	return IRQ_HANDLED;
124 }
125 
126 static void xenvif_wake_queue(unsigned long data)
127 {
128 	struct xenvif *vif = (struct xenvif *)data;
129 
130 	if (netif_queue_stopped(vif->dev)) {
131 		netdev_err(vif->dev, "draining TX queue\n");
132 		vif->rx_queue_purge = true;
133 		xenvif_kick_thread(vif);
134 		netif_wake_queue(vif->dev);
135 	}
136 }
137 
138 static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
139 {
140 	struct xenvif *vif = netdev_priv(dev);
141 	int min_slots_needed;
142 
143 	BUG_ON(skb->dev != dev);
144 
145 	/* Drop the packet if vif is not ready */
146 	if (vif->task == NULL ||
147 	    vif->dealloc_task == NULL ||
148 	    !xenvif_schedulable(vif))
149 		goto drop;
150 
151 	/* At best we'll need one slot for the header and one for each
152 	 * frag.
153 	 */
154 	min_slots_needed = 1 + skb_shinfo(skb)->nr_frags;
155 
156 	/* If the skb is GSO then we'll also need an extra slot for the
157 	 * metadata.
158 	 */
159 	if (skb_is_gso(skb))
160 		min_slots_needed++;
161 
162 	/* If the skb can't possibly fit in the remaining slots
163 	 * then turn off the queue to give the ring a chance to
164 	 * drain.
165 	 */
166 	if (!xenvif_rx_ring_slots_available(vif, min_slots_needed)) {
167 		vif->wake_queue.function = xenvif_wake_queue;
168 		vif->wake_queue.data = (unsigned long)vif;
169 		xenvif_stop_queue(vif);
170 		mod_timer(&vif->wake_queue,
171 			jiffies + rx_drain_timeout_jiffies);
172 	}
173 
174 	skb_queue_tail(&vif->rx_queue, skb);
175 	xenvif_kick_thread(vif);
176 
177 	return NETDEV_TX_OK;
178 
179  drop:
180 	vif->dev->stats.tx_dropped++;
181 	dev_kfree_skb(skb);
182 	return NETDEV_TX_OK;
183 }
184 
185 static struct net_device_stats *xenvif_get_stats(struct net_device *dev)
186 {
187 	struct xenvif *vif = netdev_priv(dev);
188 	return &vif->dev->stats;
189 }
190 
191 static void xenvif_up(struct xenvif *vif)
192 {
193 	napi_enable(&vif->napi);
194 	enable_irq(vif->tx_irq);
195 	if (vif->tx_irq != vif->rx_irq)
196 		enable_irq(vif->rx_irq);
197 	xenvif_check_rx_xenvif(vif);
198 }
199 
200 static void xenvif_down(struct xenvif *vif)
201 {
202 	napi_disable(&vif->napi);
203 	disable_irq(vif->tx_irq);
204 	if (vif->tx_irq != vif->rx_irq)
205 		disable_irq(vif->rx_irq);
206 	del_timer_sync(&vif->credit_timeout);
207 }
208 
209 static int xenvif_open(struct net_device *dev)
210 {
211 	struct xenvif *vif = netdev_priv(dev);
212 	if (netif_carrier_ok(dev))
213 		xenvif_up(vif);
214 	netif_start_queue(dev);
215 	return 0;
216 }
217 
218 static int xenvif_close(struct net_device *dev)
219 {
220 	struct xenvif *vif = netdev_priv(dev);
221 	if (netif_carrier_ok(dev))
222 		xenvif_down(vif);
223 	netif_stop_queue(dev);
224 	return 0;
225 }
226 
227 static int xenvif_change_mtu(struct net_device *dev, int mtu)
228 {
229 	struct xenvif *vif = netdev_priv(dev);
230 	int max = vif->can_sg ? 65535 - VLAN_ETH_HLEN : ETH_DATA_LEN;
231 
232 	if (mtu > max)
233 		return -EINVAL;
234 	dev->mtu = mtu;
235 	return 0;
236 }
237 
238 static netdev_features_t xenvif_fix_features(struct net_device *dev,
239 	netdev_features_t features)
240 {
241 	struct xenvif *vif = netdev_priv(dev);
242 
243 	if (!vif->can_sg)
244 		features &= ~NETIF_F_SG;
245 	if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV4))
246 		features &= ~NETIF_F_TSO;
247 	if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV6))
248 		features &= ~NETIF_F_TSO6;
249 	if (!vif->ip_csum)
250 		features &= ~NETIF_F_IP_CSUM;
251 	if (!vif->ipv6_csum)
252 		features &= ~NETIF_F_IPV6_CSUM;
253 
254 	return features;
255 }
256 
257 static const struct xenvif_stat {
258 	char name[ETH_GSTRING_LEN];
259 	u16 offset;
260 } xenvif_stats[] = {
261 	{
262 		"rx_gso_checksum_fixup",
263 		offsetof(struct xenvif, rx_gso_checksum_fixup)
264 	},
265 	/* If (sent != success + fail), there are probably packets never
266 	 * freed up properly!
267 	 */
268 	{
269 		"tx_zerocopy_sent",
270 		offsetof(struct xenvif, tx_zerocopy_sent),
271 	},
272 	{
273 		"tx_zerocopy_success",
274 		offsetof(struct xenvif, tx_zerocopy_success),
275 	},
276 	{
277 		"tx_zerocopy_fail",
278 		offsetof(struct xenvif, tx_zerocopy_fail)
279 	},
280 	/* Number of packets exceeding MAX_SKB_FRAG slots. You should use
281 	 * a guest with the same MAX_SKB_FRAG
282 	 */
283 	{
284 		"tx_frag_overflow",
285 		offsetof(struct xenvif, tx_frag_overflow)
286 	},
287 };
288 
289 static int xenvif_get_sset_count(struct net_device *dev, int string_set)
290 {
291 	switch (string_set) {
292 	case ETH_SS_STATS:
293 		return ARRAY_SIZE(xenvif_stats);
294 	default:
295 		return -EINVAL;
296 	}
297 }
298 
299 static void xenvif_get_ethtool_stats(struct net_device *dev,
300 				     struct ethtool_stats *stats, u64 * data)
301 {
302 	void *vif = netdev_priv(dev);
303 	int i;
304 
305 	for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++)
306 		data[i] = *(unsigned long *)(vif + xenvif_stats[i].offset);
307 }
308 
309 static void xenvif_get_strings(struct net_device *dev, u32 stringset, u8 * data)
310 {
311 	int i;
312 
313 	switch (stringset) {
314 	case ETH_SS_STATS:
315 		for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++)
316 			memcpy(data + i * ETH_GSTRING_LEN,
317 			       xenvif_stats[i].name, ETH_GSTRING_LEN);
318 		break;
319 	}
320 }
321 
322 static const struct ethtool_ops xenvif_ethtool_ops = {
323 	.get_link	= ethtool_op_get_link,
324 
325 	.get_sset_count = xenvif_get_sset_count,
326 	.get_ethtool_stats = xenvif_get_ethtool_stats,
327 	.get_strings = xenvif_get_strings,
328 };
329 
330 static const struct net_device_ops xenvif_netdev_ops = {
331 	.ndo_start_xmit	= xenvif_start_xmit,
332 	.ndo_get_stats	= xenvif_get_stats,
333 	.ndo_open	= xenvif_open,
334 	.ndo_stop	= xenvif_close,
335 	.ndo_change_mtu	= xenvif_change_mtu,
336 	.ndo_fix_features = xenvif_fix_features,
337 	.ndo_set_mac_address = eth_mac_addr,
338 	.ndo_validate_addr   = eth_validate_addr,
339 };
340 
341 struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
342 			    unsigned int handle)
343 {
344 	int err;
345 	struct net_device *dev;
346 	struct xenvif *vif;
347 	char name[IFNAMSIZ] = {};
348 	int i;
349 
350 	snprintf(name, IFNAMSIZ - 1, "vif%u.%u", domid, handle);
351 	dev = alloc_netdev(sizeof(struct xenvif), name, ether_setup);
352 	if (dev == NULL) {
353 		pr_warn("Could not allocate netdev for %s\n", name);
354 		return ERR_PTR(-ENOMEM);
355 	}
356 
357 	SET_NETDEV_DEV(dev, parent);
358 
359 	vif = netdev_priv(dev);
360 
361 	vif->grant_copy_op = vmalloc(sizeof(struct gnttab_copy) *
362 				     MAX_GRANT_COPY_OPS);
363 	if (vif->grant_copy_op == NULL) {
364 		pr_warn("Could not allocate grant copy space for %s\n", name);
365 		free_netdev(dev);
366 		return ERR_PTR(-ENOMEM);
367 	}
368 
369 	vif->domid  = domid;
370 	vif->handle = handle;
371 	vif->can_sg = 1;
372 	vif->ip_csum = 1;
373 	vif->dev = dev;
374 
375 	vif->disabled = false;
376 
377 	vif->credit_bytes = vif->remaining_credit = ~0UL;
378 	vif->credit_usec  = 0UL;
379 	init_timer(&vif->credit_timeout);
380 	vif->credit_window_start = get_jiffies_64();
381 
382 	init_timer(&vif->wake_queue);
383 
384 	dev->netdev_ops	= &xenvif_netdev_ops;
385 	dev->hw_features = NETIF_F_SG |
386 		NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
387 		NETIF_F_TSO | NETIF_F_TSO6;
388 	dev->features = dev->hw_features | NETIF_F_RXCSUM;
389 	SET_ETHTOOL_OPS(dev, &xenvif_ethtool_ops);
390 
391 	dev->tx_queue_len = XENVIF_QUEUE_LENGTH;
392 
393 	skb_queue_head_init(&vif->rx_queue);
394 	skb_queue_head_init(&vif->tx_queue);
395 
396 	vif->pending_cons = 0;
397 	vif->pending_prod = MAX_PENDING_REQS;
398 	for (i = 0; i < MAX_PENDING_REQS; i++)
399 		vif->pending_ring[i] = i;
400 	spin_lock_init(&vif->callback_lock);
401 	spin_lock_init(&vif->response_lock);
402 	/* If ballooning is disabled, this will consume real memory, so you
403 	 * better enable it. The long term solution would be to use just a
404 	 * bunch of valid page descriptors, without dependency on ballooning
405 	 */
406 	err = alloc_xenballooned_pages(MAX_PENDING_REQS,
407 				       vif->mmap_pages,
408 				       false);
409 	if (err) {
410 		netdev_err(dev, "Could not reserve mmap_pages\n");
411 		return ERR_PTR(-ENOMEM);
412 	}
413 	for (i = 0; i < MAX_PENDING_REQS; i++) {
414 		vif->pending_tx_info[i].callback_struct = (struct ubuf_info)
415 			{ .callback = xenvif_zerocopy_callback,
416 			  .ctx = NULL,
417 			  .desc = i };
418 		vif->grant_tx_handle[i] = NETBACK_INVALID_HANDLE;
419 	}
420 
421 	/*
422 	 * Initialise a dummy MAC address. We choose the numerically
423 	 * largest non-broadcast address to prevent the address getting
424 	 * stolen by an Ethernet bridge for STP purposes.
425 	 * (FE:FF:FF:FF:FF:FF)
426 	 */
427 	memset(dev->dev_addr, 0xFF, ETH_ALEN);
428 	dev->dev_addr[0] &= ~0x01;
429 
430 	netif_napi_add(dev, &vif->napi, xenvif_poll, XENVIF_NAPI_WEIGHT);
431 
432 	netif_carrier_off(dev);
433 
434 	err = register_netdev(dev);
435 	if (err) {
436 		netdev_warn(dev, "Could not register device: err=%d\n", err);
437 		free_netdev(dev);
438 		return ERR_PTR(err);
439 	}
440 
441 	netdev_dbg(dev, "Successfully created xenvif\n");
442 
443 	__module_get(THIS_MODULE);
444 
445 	return vif;
446 }
447 
448 int xenvif_connect(struct xenvif *vif, unsigned long tx_ring_ref,
449 		   unsigned long rx_ring_ref, unsigned int tx_evtchn,
450 		   unsigned int rx_evtchn)
451 {
452 	struct task_struct *task;
453 	int err = -ENOMEM;
454 
455 	BUG_ON(vif->tx_irq);
456 	BUG_ON(vif->task);
457 	BUG_ON(vif->dealloc_task);
458 
459 	err = xenvif_map_frontend_rings(vif, tx_ring_ref, rx_ring_ref);
460 	if (err < 0)
461 		goto err;
462 
463 	init_waitqueue_head(&vif->wq);
464 	init_waitqueue_head(&vif->dealloc_wq);
465 
466 	if (tx_evtchn == rx_evtchn) {
467 		/* feature-split-event-channels == 0 */
468 		err = bind_interdomain_evtchn_to_irqhandler(
469 			vif->domid, tx_evtchn, xenvif_interrupt, 0,
470 			vif->dev->name, vif);
471 		if (err < 0)
472 			goto err_unmap;
473 		vif->tx_irq = vif->rx_irq = err;
474 		disable_irq(vif->tx_irq);
475 	} else {
476 		/* feature-split-event-channels == 1 */
477 		snprintf(vif->tx_irq_name, sizeof(vif->tx_irq_name),
478 			 "%s-tx", vif->dev->name);
479 		err = bind_interdomain_evtchn_to_irqhandler(
480 			vif->domid, tx_evtchn, xenvif_tx_interrupt, 0,
481 			vif->tx_irq_name, vif);
482 		if (err < 0)
483 			goto err_unmap;
484 		vif->tx_irq = err;
485 		disable_irq(vif->tx_irq);
486 
487 		snprintf(vif->rx_irq_name, sizeof(vif->rx_irq_name),
488 			 "%s-rx", vif->dev->name);
489 		err = bind_interdomain_evtchn_to_irqhandler(
490 			vif->domid, rx_evtchn, xenvif_rx_interrupt, 0,
491 			vif->rx_irq_name, vif);
492 		if (err < 0)
493 			goto err_tx_unbind;
494 		vif->rx_irq = err;
495 		disable_irq(vif->rx_irq);
496 	}
497 
498 	task = kthread_create(xenvif_kthread_guest_rx,
499 			      (void *)vif, "%s-guest-rx", vif->dev->name);
500 	if (IS_ERR(task)) {
501 		pr_warn("Could not allocate kthread for %s\n", vif->dev->name);
502 		err = PTR_ERR(task);
503 		goto err_rx_unbind;
504 	}
505 
506 	vif->task = task;
507 
508 	task = kthread_create(xenvif_dealloc_kthread,
509 			      (void *)vif, "%s-dealloc", vif->dev->name);
510 	if (IS_ERR(task)) {
511 		pr_warn("Could not allocate kthread for %s\n", vif->dev->name);
512 		err = PTR_ERR(task);
513 		goto err_rx_unbind;
514 	}
515 
516 	vif->dealloc_task = task;
517 
518 	rtnl_lock();
519 	if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
520 		dev_set_mtu(vif->dev, ETH_DATA_LEN);
521 	netdev_update_features(vif->dev);
522 	netif_carrier_on(vif->dev);
523 	if (netif_running(vif->dev))
524 		xenvif_up(vif);
525 	rtnl_unlock();
526 
527 	wake_up_process(vif->task);
528 	wake_up_process(vif->dealloc_task);
529 
530 	return 0;
531 
532 err_rx_unbind:
533 	unbind_from_irqhandler(vif->rx_irq, vif);
534 	vif->rx_irq = 0;
535 err_tx_unbind:
536 	unbind_from_irqhandler(vif->tx_irq, vif);
537 	vif->tx_irq = 0;
538 err_unmap:
539 	xenvif_unmap_frontend_rings(vif);
540 err:
541 	module_put(THIS_MODULE);
542 	return err;
543 }
544 
545 void xenvif_carrier_off(struct xenvif *vif)
546 {
547 	struct net_device *dev = vif->dev;
548 
549 	rtnl_lock();
550 	netif_carrier_off(dev); /* discard queued packets */
551 	if (netif_running(dev))
552 		xenvif_down(vif);
553 	rtnl_unlock();
554 }
555 
556 void xenvif_disconnect(struct xenvif *vif)
557 {
558 	if (netif_carrier_ok(vif->dev))
559 		xenvif_carrier_off(vif);
560 
561 	if (vif->task) {
562 		del_timer_sync(&vif->wake_queue);
563 		kthread_stop(vif->task);
564 		vif->task = NULL;
565 	}
566 
567 	if (vif->dealloc_task) {
568 		kthread_stop(vif->dealloc_task);
569 		vif->dealloc_task = NULL;
570 	}
571 
572 	if (vif->tx_irq) {
573 		if (vif->tx_irq == vif->rx_irq)
574 			unbind_from_irqhandler(vif->tx_irq, vif);
575 		else {
576 			unbind_from_irqhandler(vif->tx_irq, vif);
577 			unbind_from_irqhandler(vif->rx_irq, vif);
578 		}
579 		vif->tx_irq = 0;
580 	}
581 
582 	xenvif_unmap_frontend_rings(vif);
583 }
584 
585 void xenvif_free(struct xenvif *vif)
586 {
587 	int i, unmap_timeout = 0;
588 	/* Here we want to avoid timeout messages if an skb can be legitimately
589 	 * stuck somewhere else. Realistically this could be an another vif's
590 	 * internal or QDisc queue. That another vif also has this
591 	 * rx_drain_timeout_msecs timeout, but the timer only ditches the
592 	 * internal queue. After that, the QDisc queue can put in worst case
593 	 * XEN_NETIF_RX_RING_SIZE / MAX_SKB_FRAGS skbs into that another vif's
594 	 * internal queue, so we need several rounds of such timeouts until we
595 	 * can be sure that no another vif should have skb's from us. We are
596 	 * not sending more skb's, so newly stuck packets are not interesting
597 	 * for us here.
598 	 */
599 	unsigned int worst_case_skb_lifetime = (rx_drain_timeout_msecs/1000) *
600 		DIV_ROUND_UP(XENVIF_QUEUE_LENGTH, (XEN_NETIF_RX_RING_SIZE / MAX_SKB_FRAGS));
601 
602 	for (i = 0; i < MAX_PENDING_REQS; ++i) {
603 		if (vif->grant_tx_handle[i] != NETBACK_INVALID_HANDLE) {
604 			unmap_timeout++;
605 			schedule_timeout(msecs_to_jiffies(1000));
606 			if (unmap_timeout > worst_case_skb_lifetime &&
607 			    net_ratelimit())
608 				netdev_err(vif->dev,
609 					   "Page still granted! Index: %x\n",
610 					   i);
611 			/* If there are still unmapped pages, reset the loop to
612 			 * start checking again. We shouldn't exit here until
613 			 * dealloc thread and NAPI instance release all the
614 			 * pages. If a kernel bug causes the skbs to stall
615 			 * somewhere, the interface cannot be brought down
616 			 * properly.
617 			 */
618 			i = -1;
619 		}
620 	}
621 
622 	free_xenballooned_pages(MAX_PENDING_REQS, vif->mmap_pages);
623 
624 	netif_napi_del(&vif->napi);
625 
626 	unregister_netdev(vif->dev);
627 
628 	vfree(vif->grant_copy_op);
629 	free_netdev(vif->dev);
630 
631 	module_put(THIS_MODULE);
632 }
633