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 u16 xenvif_select_queue(struct net_device *dev, struct sk_buff *skb,
141 void *accel_priv, select_queue_fallback_t fallback)
143 unsigned int num_queues = dev->real_num_tx_queues;
147 /* First, check if there is only one queue to optimise the
148 * single-queue or old frontend scenario.
150 if (num_queues == 1) {
153 /* Use skb_get_hash to obtain an L4 hash if available */
154 hash = skb_get_hash(skb);
155 queue_index = hash % num_queues;
161 static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
163 struct xenvif *vif = netdev_priv(dev);
164 struct xenvif_queue *queue = NULL;
165 unsigned int num_queues = dev->real_num_tx_queues;
167 int min_slots_needed;
169 BUG_ON(skb->dev != dev);
171 /* Drop the packet if queues are not set up */
175 /* Obtain the queue to be used to transmit this packet */
176 index = skb_get_queue_mapping(skb);
177 if (index >= num_queues) {
178 pr_warn_ratelimited("Invalid queue %hu for packet on interface %s\n.",
179 index, vif->dev->name);
182 queue = &vif->queues[index];
184 /* Drop the packet if queue is not ready */
185 if (queue->task == NULL ||
186 queue->dealloc_task == NULL ||
187 !xenvif_schedulable(vif))
190 /* At best we'll need one slot for the header and one for each
193 min_slots_needed = 1 + skb_shinfo(skb)->nr_frags;
195 /* If the skb is GSO then we'll also need an extra slot for the
201 /* If the skb can't possibly fit in the remaining slots
202 * then turn off the queue to give the ring a chance to
205 if (!xenvif_rx_ring_slots_available(queue, min_slots_needed)) {
206 queue->wake_queue.function = xenvif_wake_queue_callback;
207 queue->wake_queue.data = (unsigned long)queue;
208 xenvif_stop_queue(queue);
209 mod_timer(&queue->wake_queue,
210 jiffies + rx_drain_timeout_jiffies);
213 skb_queue_tail(&queue->rx_queue, skb);
214 xenvif_kick_thread(queue);
219 vif->dev->stats.tx_dropped++;
224 static struct net_device_stats *xenvif_get_stats(struct net_device *dev)
226 struct xenvif *vif = netdev_priv(dev);
227 struct xenvif_queue *queue = NULL;
228 unsigned int num_queues = dev->real_num_tx_queues;
229 unsigned long rx_bytes = 0;
230 unsigned long rx_packets = 0;
231 unsigned long tx_bytes = 0;
232 unsigned long tx_packets = 0;
235 if (vif->queues == NULL)
238 /* Aggregate tx and rx stats from each queue */
239 for (index = 0; index < num_queues; ++index) {
240 queue = &vif->queues[index];
241 rx_bytes += queue->stats.rx_bytes;
242 rx_packets += queue->stats.rx_packets;
243 tx_bytes += queue->stats.tx_bytes;
244 tx_packets += queue->stats.tx_packets;
248 vif->dev->stats.rx_bytes = rx_bytes;
249 vif->dev->stats.rx_packets = rx_packets;
250 vif->dev->stats.tx_bytes = tx_bytes;
251 vif->dev->stats.tx_packets = tx_packets;
253 return &vif->dev->stats;
256 static void xenvif_up(struct xenvif *vif)
258 struct xenvif_queue *queue = NULL;
259 unsigned int num_queues = vif->dev->real_num_tx_queues;
260 unsigned int queue_index;
262 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
263 queue = &vif->queues[queue_index];
264 napi_enable(&queue->napi);
265 enable_irq(queue->tx_irq);
266 if (queue->tx_irq != queue->rx_irq)
267 enable_irq(queue->rx_irq);
268 xenvif_napi_schedule_or_enable_events(queue);
272 static void xenvif_down(struct xenvif *vif)
274 struct xenvif_queue *queue = NULL;
275 unsigned int num_queues = vif->dev->real_num_tx_queues;
276 unsigned int queue_index;
278 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
279 queue = &vif->queues[queue_index];
280 napi_disable(&queue->napi);
281 disable_irq(queue->tx_irq);
282 if (queue->tx_irq != queue->rx_irq)
283 disable_irq(queue->rx_irq);
284 del_timer_sync(&queue->credit_timeout);
288 static int xenvif_open(struct net_device *dev)
290 struct xenvif *vif = netdev_priv(dev);
291 if (netif_carrier_ok(dev))
293 netif_tx_start_all_queues(dev);
297 static int xenvif_close(struct net_device *dev)
299 struct xenvif *vif = netdev_priv(dev);
300 if (netif_carrier_ok(dev))
302 netif_tx_stop_all_queues(dev);
306 static int xenvif_change_mtu(struct net_device *dev, int mtu)
308 struct xenvif *vif = netdev_priv(dev);
309 int max = vif->can_sg ? 65535 - VLAN_ETH_HLEN : ETH_DATA_LEN;
317 static netdev_features_t xenvif_fix_features(struct net_device *dev,
318 netdev_features_t features)
320 struct xenvif *vif = netdev_priv(dev);
323 features &= ~NETIF_F_SG;
324 if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV4))
325 features &= ~NETIF_F_TSO;
326 if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV6))
327 features &= ~NETIF_F_TSO6;
329 features &= ~NETIF_F_IP_CSUM;
331 features &= ~NETIF_F_IPV6_CSUM;
336 static const struct xenvif_stat {
337 char name[ETH_GSTRING_LEN];
341 "rx_gso_checksum_fixup",
342 offsetof(struct xenvif_stats, rx_gso_checksum_fixup)
344 /* If (sent != success + fail), there are probably packets never
349 offsetof(struct xenvif_stats, tx_zerocopy_sent),
352 "tx_zerocopy_success",
353 offsetof(struct xenvif_stats, tx_zerocopy_success),
357 offsetof(struct xenvif_stats, tx_zerocopy_fail)
359 /* Number of packets exceeding MAX_SKB_FRAG slots. You should use
360 * a guest with the same MAX_SKB_FRAG
364 offsetof(struct xenvif_stats, tx_frag_overflow)
368 static int xenvif_get_sset_count(struct net_device *dev, int string_set)
370 switch (string_set) {
372 return ARRAY_SIZE(xenvif_stats);
378 static void xenvif_get_ethtool_stats(struct net_device *dev,
379 struct ethtool_stats *stats, u64 * data)
381 struct xenvif *vif = netdev_priv(dev);
382 unsigned int num_queues = dev->real_num_tx_queues;
384 unsigned int queue_index;
385 struct xenvif_stats *vif_stats;
387 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++) {
388 unsigned long accum = 0;
389 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
390 vif_stats = &vif->queues[queue_index].stats;
391 accum += *(unsigned long *)(vif_stats + xenvif_stats[i].offset);
397 static void xenvif_get_strings(struct net_device *dev, u32 stringset, u8 * data)
403 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++)
404 memcpy(data + i * ETH_GSTRING_LEN,
405 xenvif_stats[i].name, ETH_GSTRING_LEN);
410 static const struct ethtool_ops xenvif_ethtool_ops = {
411 .get_link = ethtool_op_get_link,
413 .get_sset_count = xenvif_get_sset_count,
414 .get_ethtool_stats = xenvif_get_ethtool_stats,
415 .get_strings = xenvif_get_strings,
418 static const struct net_device_ops xenvif_netdev_ops = {
419 .ndo_start_xmit = xenvif_start_xmit,
420 .ndo_get_stats = xenvif_get_stats,
421 .ndo_open = xenvif_open,
422 .ndo_stop = xenvif_close,
423 .ndo_change_mtu = xenvif_change_mtu,
424 .ndo_fix_features = xenvif_fix_features,
425 .ndo_set_mac_address = eth_mac_addr,
426 .ndo_validate_addr = eth_validate_addr,
427 .ndo_select_queue = xenvif_select_queue,
430 struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
434 struct net_device *dev;
436 char name[IFNAMSIZ] = {};
438 snprintf(name, IFNAMSIZ - 1, "vif%u.%u", domid, handle);
439 /* Allocate a netdev with the max. supported number of queues.
440 * When the guest selects the desired number, it will be updated
441 * via netif_set_real_num_tx_queues().
443 dev = alloc_netdev_mq(sizeof(struct xenvif), name, ether_setup,
446 pr_warn("Could not allocate netdev for %s\n", name);
447 return ERR_PTR(-ENOMEM);
450 SET_NETDEV_DEV(dev, parent);
452 vif = netdev_priv(dev);
455 vif->handle = handle;
459 vif->disabled = false;
461 /* Start out with no queues. The call below does not require
462 * rtnl_lock() as it happens before register_netdev().
465 netif_set_real_num_tx_queues(dev, 0);
467 dev->netdev_ops = &xenvif_netdev_ops;
468 dev->hw_features = NETIF_F_SG |
469 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
470 NETIF_F_TSO | NETIF_F_TSO6;
471 dev->features = dev->hw_features | NETIF_F_RXCSUM;
472 dev->ethtool_ops = &xenvif_ethtool_ops;
474 dev->tx_queue_len = XENVIF_QUEUE_LENGTH;
477 * Initialise a dummy MAC address. We choose the numerically
478 * largest non-broadcast address to prevent the address getting
479 * stolen by an Ethernet bridge for STP purposes.
480 * (FE:FF:FF:FF:FF:FF)
482 memset(dev->dev_addr, 0xFF, ETH_ALEN);
483 dev->dev_addr[0] &= ~0x01;
485 netif_carrier_off(dev);
487 err = register_netdev(dev);
489 netdev_warn(dev, "Could not register device: err=%d\n", err);
494 netdev_dbg(dev, "Successfully created xenvif\n");
496 __module_get(THIS_MODULE);
501 int xenvif_init_queue(struct xenvif_queue *queue)
505 queue->credit_bytes = queue->remaining_credit = ~0UL;
506 queue->credit_usec = 0UL;
507 init_timer(&queue->credit_timeout);
508 queue->credit_window_start = get_jiffies_64();
510 skb_queue_head_init(&queue->rx_queue);
511 skb_queue_head_init(&queue->tx_queue);
513 queue->pending_cons = 0;
514 queue->pending_prod = MAX_PENDING_REQS;
515 for (i = 0; i < MAX_PENDING_REQS; ++i)
516 queue->pending_ring[i] = i;
518 spin_lock_init(&queue->callback_lock);
519 spin_lock_init(&queue->response_lock);
521 /* If ballooning is disabled, this will consume real memory, so you
522 * better enable it. The long term solution would be to use just a
523 * bunch of valid page descriptors, without dependency on ballooning
525 err = alloc_xenballooned_pages(MAX_PENDING_REQS,
529 netdev_err(queue->vif->dev, "Could not reserve mmap_pages\n");
533 for (i = 0; i < MAX_PENDING_REQS; i++) {
534 queue->pending_tx_info[i].callback_struct = (struct ubuf_info)
535 { .callback = xenvif_zerocopy_callback,
538 queue->grant_tx_handle[i] = NETBACK_INVALID_HANDLE;
541 init_timer(&queue->wake_queue);
543 netif_napi_add(queue->vif->dev, &queue->napi, xenvif_poll,
549 void xenvif_carrier_on(struct xenvif *vif)
552 if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
553 dev_set_mtu(vif->dev, ETH_DATA_LEN);
554 netdev_update_features(vif->dev);
555 netif_carrier_on(vif->dev);
556 if (netif_running(vif->dev))
561 int xenvif_connect(struct xenvif_queue *queue, unsigned long tx_ring_ref,
562 unsigned long rx_ring_ref, unsigned int tx_evtchn,
563 unsigned int rx_evtchn)
565 struct task_struct *task;
568 BUG_ON(queue->tx_irq);
570 BUG_ON(queue->dealloc_task);
572 err = xenvif_map_frontend_rings(queue, tx_ring_ref, rx_ring_ref);
576 init_waitqueue_head(&queue->wq);
577 init_waitqueue_head(&queue->dealloc_wq);
579 if (tx_evtchn == rx_evtchn) {
580 /* feature-split-event-channels == 0 */
581 err = bind_interdomain_evtchn_to_irqhandler(
582 queue->vif->domid, tx_evtchn, xenvif_interrupt, 0,
586 queue->tx_irq = queue->rx_irq = err;
587 disable_irq(queue->tx_irq);
589 /* feature-split-event-channels == 1 */
590 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name),
591 "%s-tx", queue->name);
592 err = bind_interdomain_evtchn_to_irqhandler(
593 queue->vif->domid, tx_evtchn, xenvif_tx_interrupt, 0,
594 queue->tx_irq_name, queue);
598 disable_irq(queue->tx_irq);
600 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name),
601 "%s-rx", queue->name);
602 err = bind_interdomain_evtchn_to_irqhandler(
603 queue->vif->domid, rx_evtchn, xenvif_rx_interrupt, 0,
604 queue->rx_irq_name, queue);
608 disable_irq(queue->rx_irq);
611 task = kthread_create(xenvif_kthread_guest_rx,
612 (void *)queue, "%s-guest-rx", queue->name);
614 pr_warn("Could not allocate kthread for %s\n", queue->name);
620 task = kthread_create(xenvif_dealloc_kthread,
621 (void *)queue, "%s-dealloc", queue->name);
623 pr_warn("Could not allocate kthread for %s\n", queue->name);
627 queue->dealloc_task = task;
629 wake_up_process(queue->task);
630 wake_up_process(queue->dealloc_task);
635 unbind_from_irqhandler(queue->rx_irq, queue);
638 unbind_from_irqhandler(queue->tx_irq, queue);
641 xenvif_unmap_frontend_rings(queue);
643 module_put(THIS_MODULE);
647 void xenvif_carrier_off(struct xenvif *vif)
649 struct net_device *dev = vif->dev;
652 netif_carrier_off(dev); /* discard queued packets */
653 if (netif_running(dev))
658 static void xenvif_wait_unmap_timeout(struct xenvif_queue *queue,
659 unsigned int worst_case_skb_lifetime)
661 int i, unmap_timeout = 0;
663 for (i = 0; i < MAX_PENDING_REQS; ++i) {
664 if (queue->grant_tx_handle[i] != NETBACK_INVALID_HANDLE) {
666 schedule_timeout(msecs_to_jiffies(1000));
667 if (unmap_timeout > worst_case_skb_lifetime &&
669 netdev_err(queue->vif->dev,
670 "Page still granted! Index: %x\n",
677 void xenvif_disconnect(struct xenvif *vif)
679 struct xenvif_queue *queue = NULL;
680 unsigned int num_queues = vif->dev->real_num_tx_queues;
681 unsigned int queue_index;
683 if (netif_carrier_ok(vif->dev))
684 xenvif_carrier_off(vif);
686 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
687 queue = &vif->queues[queue_index];
690 del_timer_sync(&queue->wake_queue);
691 kthread_stop(queue->task);
695 if (queue->dealloc_task) {
696 kthread_stop(queue->dealloc_task);
697 queue->dealloc_task = NULL;
701 if (queue->tx_irq == queue->rx_irq)
702 unbind_from_irqhandler(queue->tx_irq, queue);
704 unbind_from_irqhandler(queue->tx_irq, queue);
705 unbind_from_irqhandler(queue->rx_irq, queue);
710 xenvif_unmap_frontend_rings(queue);
714 /* Reverse the relevant parts of xenvif_init_queue().
715 * Used for queue teardown from xenvif_free(), and on the
716 * error handling paths in xenbus.c:connect().
718 void xenvif_deinit_queue(struct xenvif_queue *queue)
720 free_xenballooned_pages(MAX_PENDING_REQS, queue->mmap_pages);
721 netif_napi_del(&queue->napi);
724 void xenvif_free(struct xenvif *vif)
726 struct xenvif_queue *queue = NULL;
727 unsigned int num_queues = vif->dev->real_num_tx_queues;
728 unsigned int queue_index;
729 /* Here we want to avoid timeout messages if an skb can be legitimately
730 * stuck somewhere else. Realistically this could be an another vif's
731 * internal or QDisc queue. That another vif also has this
732 * rx_drain_timeout_msecs timeout, but the timer only ditches the
733 * internal queue. After that, the QDisc queue can put in worst case
734 * XEN_NETIF_RX_RING_SIZE / MAX_SKB_FRAGS skbs into that another vif's
735 * internal queue, so we need several rounds of such timeouts until we
736 * can be sure that no another vif should have skb's from us. We are
737 * not sending more skb's, so newly stuck packets are not interesting
740 unsigned int worst_case_skb_lifetime = (rx_drain_timeout_msecs/1000) *
741 DIV_ROUND_UP(XENVIF_QUEUE_LENGTH, (XEN_NETIF_RX_RING_SIZE / MAX_SKB_FRAGS));
743 unregister_netdev(vif->dev);
745 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
746 queue = &vif->queues[queue_index];
747 xenvif_wait_unmap_timeout(queue, worst_case_skb_lifetime);
748 xenvif_deinit_queue(queue);
751 /* Free the array of queues. The call below does not require
752 * rtnl_lock() because it happens after unregister_netdev().
754 netif_set_real_num_tx_queues(vif->dev, 0);
758 free_netdev(vif->dev);
760 module_put(THIS_MODULE);