2 * Network-device interface management.
4 * Copyright (c) 2004-2005, Keir Fraser
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:
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:
19 * The above copyright notice and this permission notice shall be included in
20 * all copies or substantial portions of the Software.
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
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>
39 #include <xen/events.h>
40 #include <asm/xen/hypercall.h>
41 #include <xen/balloon.h>
43 #define XENVIF_QUEUE_LENGTH 32
44 #define XENVIF_NAPI_WEIGHT 64
46 static inline void xenvif_stop_queue(struct xenvif_queue *queue)
48 struct net_device *dev = queue->vif->dev;
50 if (!queue->vif->can_queue)
53 netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
56 int xenvif_schedulable(struct xenvif *vif)
58 return netif_running(vif->dev) && netif_carrier_ok(vif->dev);
61 static irqreturn_t xenvif_tx_interrupt(int irq, void *dev_id)
63 struct xenvif_queue *queue = dev_id;
65 if (RING_HAS_UNCONSUMED_REQUESTS(&queue->tx))
66 napi_schedule(&queue->napi);
71 int xenvif_poll(struct napi_struct *napi, int budget)
73 struct xenvif_queue *queue =
74 container_of(napi, struct xenvif_queue, napi);
77 /* This vif is rogue, we pretend we've there is nothing to do
78 * for this vif to deschedule it from NAPI. But this interface
79 * will be turned off in thread context later.
81 if (unlikely(queue->vif->disabled)) {
86 work_done = xenvif_tx_action(queue, budget);
88 if (work_done < budget) {
90 xenvif_napi_schedule_or_enable_events(queue);
96 static irqreturn_t xenvif_rx_interrupt(int irq, void *dev_id)
98 struct xenvif_queue *queue = dev_id;
100 xenvif_kick_thread(queue);
105 static irqreturn_t xenvif_interrupt(int irq, void *dev_id)
107 xenvif_tx_interrupt(irq, dev_id);
108 xenvif_rx_interrupt(irq, dev_id);
113 int xenvif_queue_stopped(struct xenvif_queue *queue)
115 struct net_device *dev = queue->vif->dev;
116 unsigned int id = queue->id;
117 return netif_tx_queue_stopped(netdev_get_tx_queue(dev, id));
120 void xenvif_wake_queue(struct xenvif_queue *queue)
122 struct net_device *dev = queue->vif->dev;
123 unsigned int id = queue->id;
124 netif_tx_wake_queue(netdev_get_tx_queue(dev, id));
127 /* Callback to wake the queue and drain it on timeout */
128 static void xenvif_wake_queue_callback(unsigned long data)
130 struct xenvif_queue *queue = (struct xenvif_queue *)data;
132 if (xenvif_queue_stopped(queue)) {
133 netdev_err(queue->vif->dev, "draining TX queue\n");
134 queue->rx_queue_purge = true;
135 xenvif_kick_thread(queue);
136 xenvif_wake_queue(queue);
140 static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
142 struct xenvif *vif = netdev_priv(dev);
143 struct xenvif_queue *queue = NULL;
144 unsigned int num_queues = vif->num_queues;
146 int min_slots_needed;
148 BUG_ON(skb->dev != dev);
150 /* Drop the packet if queues are not set up */
154 /* Obtain the queue to be used to transmit this packet */
155 index = skb_get_queue_mapping(skb);
156 if (index >= num_queues) {
157 pr_warn_ratelimited("Invalid queue %hu for packet on interface %s\n.",
158 index, vif->dev->name);
161 queue = &vif->queues[index];
163 /* Drop the packet if queue is not ready */
164 if (queue->task == NULL ||
165 queue->dealloc_task == NULL ||
166 !xenvif_schedulable(vif))
169 /* At best we'll need one slot for the header and one for each
172 min_slots_needed = 1 + skb_shinfo(skb)->nr_frags;
174 /* If the skb is GSO then we'll also need an extra slot for the
180 /* If the skb can't possibly fit in the remaining slots
181 * then turn off the queue to give the ring a chance to
184 if (!xenvif_rx_ring_slots_available(queue, min_slots_needed)) {
185 queue->wake_queue.function = xenvif_wake_queue_callback;
186 queue->wake_queue.data = (unsigned long)queue;
187 xenvif_stop_queue(queue);
188 mod_timer(&queue->wake_queue,
189 jiffies + rx_drain_timeout_jiffies);
192 skb_queue_tail(&queue->rx_queue, skb);
193 xenvif_kick_thread(queue);
198 vif->dev->stats.tx_dropped++;
203 static struct net_device_stats *xenvif_get_stats(struct net_device *dev)
205 struct xenvif *vif = netdev_priv(dev);
206 struct xenvif_queue *queue = NULL;
207 unsigned int num_queues = vif->num_queues;
208 unsigned long rx_bytes = 0;
209 unsigned long rx_packets = 0;
210 unsigned long tx_bytes = 0;
211 unsigned long tx_packets = 0;
214 if (vif->queues == NULL)
217 /* Aggregate tx and rx stats from each queue */
218 for (index = 0; index < num_queues; ++index) {
219 queue = &vif->queues[index];
220 rx_bytes += queue->stats.rx_bytes;
221 rx_packets += queue->stats.rx_packets;
222 tx_bytes += queue->stats.tx_bytes;
223 tx_packets += queue->stats.tx_packets;
227 vif->dev->stats.rx_bytes = rx_bytes;
228 vif->dev->stats.rx_packets = rx_packets;
229 vif->dev->stats.tx_bytes = tx_bytes;
230 vif->dev->stats.tx_packets = tx_packets;
232 return &vif->dev->stats;
235 static void xenvif_up(struct xenvif *vif)
237 struct xenvif_queue *queue = NULL;
238 unsigned int num_queues = vif->num_queues;
239 unsigned int queue_index;
241 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
242 queue = &vif->queues[queue_index];
243 napi_enable(&queue->napi);
244 enable_irq(queue->tx_irq);
245 if (queue->tx_irq != queue->rx_irq)
246 enable_irq(queue->rx_irq);
247 xenvif_napi_schedule_or_enable_events(queue);
251 static void xenvif_down(struct xenvif *vif)
253 struct xenvif_queue *queue = NULL;
254 unsigned int num_queues = vif->num_queues;
255 unsigned int queue_index;
257 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
258 queue = &vif->queues[queue_index];
259 napi_disable(&queue->napi);
260 disable_irq(queue->tx_irq);
261 if (queue->tx_irq != queue->rx_irq)
262 disable_irq(queue->rx_irq);
263 del_timer_sync(&queue->credit_timeout);
267 static int xenvif_open(struct net_device *dev)
269 struct xenvif *vif = netdev_priv(dev);
270 if (netif_carrier_ok(dev))
272 netif_tx_start_all_queues(dev);
276 static int xenvif_close(struct net_device *dev)
278 struct xenvif *vif = netdev_priv(dev);
279 if (netif_carrier_ok(dev))
281 netif_tx_stop_all_queues(dev);
285 static int xenvif_change_mtu(struct net_device *dev, int mtu)
287 struct xenvif *vif = netdev_priv(dev);
288 int max = vif->can_sg ? 65535 - VLAN_ETH_HLEN : ETH_DATA_LEN;
296 static netdev_features_t xenvif_fix_features(struct net_device *dev,
297 netdev_features_t features)
299 struct xenvif *vif = netdev_priv(dev);
302 features &= ~NETIF_F_SG;
303 if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV4))
304 features &= ~NETIF_F_TSO;
305 if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV6))
306 features &= ~NETIF_F_TSO6;
308 features &= ~NETIF_F_IP_CSUM;
310 features &= ~NETIF_F_IPV6_CSUM;
315 static const struct xenvif_stat {
316 char name[ETH_GSTRING_LEN];
320 "rx_gso_checksum_fixup",
321 offsetof(struct xenvif_stats, rx_gso_checksum_fixup)
323 /* If (sent != success + fail), there are probably packets never
328 offsetof(struct xenvif_stats, tx_zerocopy_sent),
331 "tx_zerocopy_success",
332 offsetof(struct xenvif_stats, tx_zerocopy_success),
336 offsetof(struct xenvif_stats, tx_zerocopy_fail)
338 /* Number of packets exceeding MAX_SKB_FRAG slots. You should use
339 * a guest with the same MAX_SKB_FRAG
343 offsetof(struct xenvif_stats, tx_frag_overflow)
347 static int xenvif_get_sset_count(struct net_device *dev, int string_set)
349 switch (string_set) {
351 return ARRAY_SIZE(xenvif_stats);
357 static void xenvif_get_ethtool_stats(struct net_device *dev,
358 struct ethtool_stats *stats, u64 * data)
360 struct xenvif *vif = netdev_priv(dev);
361 unsigned int num_queues = vif->num_queues;
363 unsigned int queue_index;
364 struct xenvif_stats *vif_stats;
366 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++) {
367 unsigned long accum = 0;
368 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
369 vif_stats = &vif->queues[queue_index].stats;
370 accum += *(unsigned long *)(vif_stats + xenvif_stats[i].offset);
376 static void xenvif_get_strings(struct net_device *dev, u32 stringset, u8 * data)
382 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++)
383 memcpy(data + i * ETH_GSTRING_LEN,
384 xenvif_stats[i].name, ETH_GSTRING_LEN);
389 static const struct ethtool_ops xenvif_ethtool_ops = {
390 .get_link = ethtool_op_get_link,
392 .get_sset_count = xenvif_get_sset_count,
393 .get_ethtool_stats = xenvif_get_ethtool_stats,
394 .get_strings = xenvif_get_strings,
397 static const struct net_device_ops xenvif_netdev_ops = {
398 .ndo_start_xmit = xenvif_start_xmit,
399 .ndo_get_stats = xenvif_get_stats,
400 .ndo_open = xenvif_open,
401 .ndo_stop = xenvif_close,
402 .ndo_change_mtu = xenvif_change_mtu,
403 .ndo_fix_features = xenvif_fix_features,
404 .ndo_set_mac_address = eth_mac_addr,
405 .ndo_validate_addr = eth_validate_addr,
408 struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
412 struct net_device *dev;
414 char name[IFNAMSIZ] = {};
416 snprintf(name, IFNAMSIZ - 1, "vif%u.%u", domid, handle);
417 /* Allocate a netdev with the max. supported number of queues.
418 * When the guest selects the desired number, it will be updated
419 * via netif_set_real_num_*_queues().
421 dev = alloc_netdev_mq(sizeof(struct xenvif), name, ether_setup,
424 pr_warn("Could not allocate netdev for %s\n", name);
425 return ERR_PTR(-ENOMEM);
428 SET_NETDEV_DEV(dev, parent);
430 vif = netdev_priv(dev);
433 vif->handle = handle;
437 vif->disabled = false;
439 /* Start out with no queues. */
443 dev->netdev_ops = &xenvif_netdev_ops;
444 dev->hw_features = NETIF_F_SG |
445 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
446 NETIF_F_TSO | NETIF_F_TSO6;
447 dev->features = dev->hw_features | NETIF_F_RXCSUM;
448 dev->ethtool_ops = &xenvif_ethtool_ops;
450 dev->tx_queue_len = XENVIF_QUEUE_LENGTH;
453 * Initialise a dummy MAC address. We choose the numerically
454 * largest non-broadcast address to prevent the address getting
455 * stolen by an Ethernet bridge for STP purposes.
456 * (FE:FF:FF:FF:FF:FF)
458 memset(dev->dev_addr, 0xFF, ETH_ALEN);
459 dev->dev_addr[0] &= ~0x01;
461 netif_carrier_off(dev);
463 err = register_netdev(dev);
465 netdev_warn(dev, "Could not register device: err=%d\n", err);
470 netdev_dbg(dev, "Successfully created xenvif\n");
472 __module_get(THIS_MODULE);
477 int xenvif_init_queue(struct xenvif_queue *queue)
481 queue->credit_bytes = queue->remaining_credit = ~0UL;
482 queue->credit_usec = 0UL;
483 init_timer(&queue->credit_timeout);
484 queue->credit_window_start = get_jiffies_64();
486 skb_queue_head_init(&queue->rx_queue);
487 skb_queue_head_init(&queue->tx_queue);
489 queue->pending_cons = 0;
490 queue->pending_prod = MAX_PENDING_REQS;
491 for (i = 0; i < MAX_PENDING_REQS; ++i)
492 queue->pending_ring[i] = i;
494 spin_lock_init(&queue->callback_lock);
495 spin_lock_init(&queue->response_lock);
497 /* If ballooning is disabled, this will consume real memory, so you
498 * better enable it. The long term solution would be to use just a
499 * bunch of valid page descriptors, without dependency on ballooning
501 err = alloc_xenballooned_pages(MAX_PENDING_REQS,
505 netdev_err(queue->vif->dev, "Could not reserve mmap_pages\n");
509 for (i = 0; i < MAX_PENDING_REQS; i++) {
510 queue->pending_tx_info[i].callback_struct = (struct ubuf_info)
511 { .callback = xenvif_zerocopy_callback,
514 queue->grant_tx_handle[i] = NETBACK_INVALID_HANDLE;
517 init_timer(&queue->wake_queue);
519 netif_napi_add(queue->vif->dev, &queue->napi, xenvif_poll,
525 void xenvif_carrier_on(struct xenvif *vif)
528 if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
529 dev_set_mtu(vif->dev, ETH_DATA_LEN);
530 netdev_update_features(vif->dev);
531 netif_carrier_on(vif->dev);
532 if (netif_running(vif->dev))
537 int xenvif_connect(struct xenvif_queue *queue, unsigned long tx_ring_ref,
538 unsigned long rx_ring_ref, unsigned int tx_evtchn,
539 unsigned int rx_evtchn)
541 struct task_struct *task;
544 BUG_ON(queue->tx_irq);
546 BUG_ON(queue->dealloc_task);
548 err = xenvif_map_frontend_rings(queue, tx_ring_ref, rx_ring_ref);
552 init_waitqueue_head(&queue->wq);
553 init_waitqueue_head(&queue->dealloc_wq);
555 if (tx_evtchn == rx_evtchn) {
556 /* feature-split-event-channels == 0 */
557 err = bind_interdomain_evtchn_to_irqhandler(
558 queue->vif->domid, tx_evtchn, xenvif_interrupt, 0,
562 queue->tx_irq = queue->rx_irq = err;
563 disable_irq(queue->tx_irq);
565 /* feature-split-event-channels == 1 */
566 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name),
567 "%s-tx", queue->name);
568 err = bind_interdomain_evtchn_to_irqhandler(
569 queue->vif->domid, tx_evtchn, xenvif_tx_interrupt, 0,
570 queue->tx_irq_name, queue);
574 disable_irq(queue->tx_irq);
576 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name),
577 "%s-rx", queue->name);
578 err = bind_interdomain_evtchn_to_irqhandler(
579 queue->vif->domid, rx_evtchn, xenvif_rx_interrupt, 0,
580 queue->rx_irq_name, queue);
584 disable_irq(queue->rx_irq);
587 task = kthread_create(xenvif_kthread_guest_rx,
588 (void *)queue, "%s-guest-rx", queue->name);
590 pr_warn("Could not allocate kthread for %s\n", queue->name);
596 task = kthread_create(xenvif_dealloc_kthread,
597 (void *)queue, "%s-dealloc", queue->name);
599 pr_warn("Could not allocate kthread for %s\n", queue->name);
603 queue->dealloc_task = task;
605 wake_up_process(queue->task);
606 wake_up_process(queue->dealloc_task);
611 unbind_from_irqhandler(queue->rx_irq, queue);
614 unbind_from_irqhandler(queue->tx_irq, queue);
617 xenvif_unmap_frontend_rings(queue);
619 module_put(THIS_MODULE);
623 void xenvif_carrier_off(struct xenvif *vif)
625 struct net_device *dev = vif->dev;
628 netif_carrier_off(dev); /* discard queued packets */
629 if (netif_running(dev))
634 static void xenvif_wait_unmap_timeout(struct xenvif_queue *queue,
635 unsigned int worst_case_skb_lifetime)
637 int i, unmap_timeout = 0;
639 for (i = 0; i < MAX_PENDING_REQS; ++i) {
640 if (queue->grant_tx_handle[i] != NETBACK_INVALID_HANDLE) {
642 schedule_timeout(msecs_to_jiffies(1000));
643 if (unmap_timeout > worst_case_skb_lifetime &&
645 netdev_err(queue->vif->dev,
646 "Page still granted! Index: %x\n",
653 void xenvif_disconnect(struct xenvif *vif)
655 struct xenvif_queue *queue = NULL;
656 unsigned int num_queues = vif->num_queues;
657 unsigned int queue_index;
659 if (netif_carrier_ok(vif->dev))
660 xenvif_carrier_off(vif);
662 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
663 queue = &vif->queues[queue_index];
666 del_timer_sync(&queue->wake_queue);
667 kthread_stop(queue->task);
671 if (queue->dealloc_task) {
672 kthread_stop(queue->dealloc_task);
673 queue->dealloc_task = NULL;
677 if (queue->tx_irq == queue->rx_irq)
678 unbind_from_irqhandler(queue->tx_irq, queue);
680 unbind_from_irqhandler(queue->tx_irq, queue);
681 unbind_from_irqhandler(queue->rx_irq, queue);
686 xenvif_unmap_frontend_rings(queue);
690 /* Reverse the relevant parts of xenvif_init_queue().
691 * Used for queue teardown from xenvif_free(), and on the
692 * error handling paths in xenbus.c:connect().
694 void xenvif_deinit_queue(struct xenvif_queue *queue)
696 free_xenballooned_pages(MAX_PENDING_REQS, queue->mmap_pages);
697 netif_napi_del(&queue->napi);
700 void xenvif_free(struct xenvif *vif)
702 struct xenvif_queue *queue = NULL;
703 unsigned int num_queues = vif->num_queues;
704 unsigned int queue_index;
705 /* Here we want to avoid timeout messages if an skb can be legitimately
706 * stuck somewhere else. Realistically this could be an another vif's
707 * internal or QDisc queue. That another vif also has this
708 * rx_drain_timeout_msecs timeout, but the timer only ditches the
709 * internal queue. After that, the QDisc queue can put in worst case
710 * XEN_NETIF_RX_RING_SIZE / MAX_SKB_FRAGS skbs into that another vif's
711 * internal queue, so we need several rounds of such timeouts until we
712 * can be sure that no another vif should have skb's from us. We are
713 * not sending more skb's, so newly stuck packets are not interesting
716 unsigned int worst_case_skb_lifetime = (rx_drain_timeout_msecs/1000) *
717 DIV_ROUND_UP(XENVIF_QUEUE_LENGTH, (XEN_NETIF_RX_RING_SIZE / MAX_SKB_FRAGS));
719 unregister_netdev(vif->dev);
721 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
722 queue = &vif->queues[queue_index];
723 xenvif_wait_unmap_timeout(queue, worst_case_skb_lifetime);
724 xenvif_deinit_queue(queue);
731 free_netdev(vif->dev);
733 module_put(THIS_MODULE);