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) &&
59 test_bit(VIF_STATUS_CONNECTED, &vif->status);
62 static irqreturn_t xenvif_tx_interrupt(int irq, void *dev_id)
64 struct xenvif_queue *queue = dev_id;
66 if (RING_HAS_UNCONSUMED_REQUESTS(&queue->tx))
67 napi_schedule(&queue->napi);
72 int xenvif_poll(struct napi_struct *napi, int budget)
74 struct xenvif_queue *queue =
75 container_of(napi, struct xenvif_queue, napi);
78 /* This vif is rogue, we pretend we've there is nothing to do
79 * for this vif to deschedule it from NAPI. But this interface
80 * will be turned off in thread context later.
81 * Also, if a guest doesn't post enough slots to receive data on one of
82 * its queues, the carrier goes down and NAPI is descheduled here so
83 * the guest can't send more packets until it's ready to receive.
85 if (unlikely(queue->vif->disabled ||
86 !netif_carrier_ok(queue->vif->dev))) {
91 work_done = xenvif_tx_action(queue, budget);
93 if (work_done < budget) {
95 xenvif_napi_schedule_or_enable_events(queue);
101 static irqreturn_t xenvif_rx_interrupt(int irq, void *dev_id)
103 struct xenvif_queue *queue = dev_id;
104 struct netdev_queue *net_queue =
105 netdev_get_tx_queue(queue->vif->dev, queue->id);
107 /* QUEUE_STATUS_RX_PURGE_EVENT is only set if either QDisc was off OR
108 * the carrier went down and this queue was previously blocked
110 if (unlikely(netif_tx_queue_stopped(net_queue) ||
111 (!netif_carrier_ok(queue->vif->dev) &&
112 test_bit(QUEUE_STATUS_RX_STALLED, &queue->status))))
113 set_bit(QUEUE_STATUS_RX_PURGE_EVENT, &queue->status);
114 xenvif_kick_thread(queue);
119 irqreturn_t xenvif_interrupt(int irq, void *dev_id)
121 xenvif_tx_interrupt(irq, dev_id);
122 xenvif_rx_interrupt(irq, dev_id);
127 int xenvif_queue_stopped(struct xenvif_queue *queue)
129 struct net_device *dev = queue->vif->dev;
130 unsigned int id = queue->id;
131 return netif_tx_queue_stopped(netdev_get_tx_queue(dev, id));
134 void xenvif_wake_queue(struct xenvif_queue *queue)
136 struct net_device *dev = queue->vif->dev;
137 unsigned int id = queue->id;
138 netif_tx_wake_queue(netdev_get_tx_queue(dev, id));
141 /* Callback to wake the queue's thread and turn the carrier off on timeout */
142 static void xenvif_rx_stalled(unsigned long data)
144 struct xenvif_queue *queue = (struct xenvif_queue *)data;
146 if (xenvif_queue_stopped(queue)) {
147 set_bit(QUEUE_STATUS_RX_PURGE_EVENT, &queue->status);
148 xenvif_kick_thread(queue);
152 static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
154 struct xenvif *vif = netdev_priv(dev);
155 struct xenvif_queue *queue = NULL;
156 unsigned int num_queues = vif->num_queues;
158 int min_slots_needed;
160 BUG_ON(skb->dev != dev);
162 /* Drop the packet if queues are not set up */
166 /* Obtain the queue to be used to transmit this packet */
167 index = skb_get_queue_mapping(skb);
168 if (index >= num_queues) {
169 pr_warn_ratelimited("Invalid queue %hu for packet on interface %s\n.",
170 index, vif->dev->name);
173 queue = &vif->queues[index];
175 /* Drop the packet if queue is not ready */
176 if (queue->task == NULL ||
177 queue->dealloc_task == NULL ||
178 !xenvif_schedulable(vif))
181 /* At best we'll need one slot for the header and one for each
184 min_slots_needed = 1 + skb_shinfo(skb)->nr_frags;
186 /* If the skb is GSO then we'll also need an extra slot for the
192 /* If the skb can't possibly fit in the remaining slots
193 * then turn off the queue to give the ring a chance to
196 if (!xenvif_rx_ring_slots_available(queue, min_slots_needed)) {
197 queue->rx_stalled.function = xenvif_rx_stalled;
198 queue->rx_stalled.data = (unsigned long)queue;
199 xenvif_stop_queue(queue);
200 mod_timer(&queue->rx_stalled,
201 jiffies + rx_drain_timeout_jiffies);
204 skb_queue_tail(&queue->rx_queue, skb);
205 xenvif_kick_thread(queue);
210 vif->dev->stats.tx_dropped++;
215 static struct net_device_stats *xenvif_get_stats(struct net_device *dev)
217 struct xenvif *vif = netdev_priv(dev);
218 struct xenvif_queue *queue = NULL;
219 unsigned int num_queues = vif->num_queues;
220 unsigned long rx_bytes = 0;
221 unsigned long rx_packets = 0;
222 unsigned long tx_bytes = 0;
223 unsigned long tx_packets = 0;
226 if (vif->queues == NULL)
229 /* Aggregate tx and rx stats from each queue */
230 for (index = 0; index < num_queues; ++index) {
231 queue = &vif->queues[index];
232 rx_bytes += queue->stats.rx_bytes;
233 rx_packets += queue->stats.rx_packets;
234 tx_bytes += queue->stats.tx_bytes;
235 tx_packets += queue->stats.tx_packets;
239 vif->dev->stats.rx_bytes = rx_bytes;
240 vif->dev->stats.rx_packets = rx_packets;
241 vif->dev->stats.tx_bytes = tx_bytes;
242 vif->dev->stats.tx_packets = tx_packets;
244 return &vif->dev->stats;
247 static void xenvif_up(struct xenvif *vif)
249 struct xenvif_queue *queue = NULL;
250 unsigned int num_queues = vif->num_queues;
251 unsigned int queue_index;
253 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
254 queue = &vif->queues[queue_index];
255 napi_enable(&queue->napi);
256 enable_irq(queue->tx_irq);
257 if (queue->tx_irq != queue->rx_irq)
258 enable_irq(queue->rx_irq);
259 xenvif_napi_schedule_or_enable_events(queue);
263 static void xenvif_down(struct xenvif *vif)
265 struct xenvif_queue *queue = NULL;
266 unsigned int num_queues = vif->num_queues;
267 unsigned int queue_index;
269 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
270 queue = &vif->queues[queue_index];
271 napi_disable(&queue->napi);
272 disable_irq(queue->tx_irq);
273 if (queue->tx_irq != queue->rx_irq)
274 disable_irq(queue->rx_irq);
275 del_timer_sync(&queue->credit_timeout);
279 static int xenvif_open(struct net_device *dev)
281 struct xenvif *vif = netdev_priv(dev);
282 if (test_bit(VIF_STATUS_CONNECTED, &vif->status))
284 netif_tx_start_all_queues(dev);
288 static int xenvif_close(struct net_device *dev)
290 struct xenvif *vif = netdev_priv(dev);
291 if (test_bit(VIF_STATUS_CONNECTED, &vif->status))
293 netif_tx_stop_all_queues(dev);
297 static int xenvif_change_mtu(struct net_device *dev, int mtu)
299 struct xenvif *vif = netdev_priv(dev);
300 int max = vif->can_sg ? 65535 - VLAN_ETH_HLEN : ETH_DATA_LEN;
308 static netdev_features_t xenvif_fix_features(struct net_device *dev,
309 netdev_features_t features)
311 struct xenvif *vif = netdev_priv(dev);
314 features &= ~NETIF_F_SG;
315 if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV4))
316 features &= ~NETIF_F_TSO;
317 if (~(vif->gso_mask | vif->gso_prefix_mask) & GSO_BIT(TCPV6))
318 features &= ~NETIF_F_TSO6;
320 features &= ~NETIF_F_IP_CSUM;
322 features &= ~NETIF_F_IPV6_CSUM;
327 static const struct xenvif_stat {
328 char name[ETH_GSTRING_LEN];
332 "rx_gso_checksum_fixup",
333 offsetof(struct xenvif_stats, rx_gso_checksum_fixup)
335 /* If (sent != success + fail), there are probably packets never
340 offsetof(struct xenvif_stats, tx_zerocopy_sent),
343 "tx_zerocopy_success",
344 offsetof(struct xenvif_stats, tx_zerocopy_success),
348 offsetof(struct xenvif_stats, tx_zerocopy_fail)
350 /* Number of packets exceeding MAX_SKB_FRAG slots. You should use
351 * a guest with the same MAX_SKB_FRAG
355 offsetof(struct xenvif_stats, tx_frag_overflow)
359 static int xenvif_get_sset_count(struct net_device *dev, int string_set)
361 switch (string_set) {
363 return ARRAY_SIZE(xenvif_stats);
369 static void xenvif_get_ethtool_stats(struct net_device *dev,
370 struct ethtool_stats *stats, u64 * data)
372 struct xenvif *vif = netdev_priv(dev);
373 unsigned int num_queues = vif->num_queues;
375 unsigned int queue_index;
376 struct xenvif_stats *vif_stats;
378 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++) {
379 unsigned long accum = 0;
380 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
381 vif_stats = &vif->queues[queue_index].stats;
382 accum += *(unsigned long *)(vif_stats + xenvif_stats[i].offset);
388 static void xenvif_get_strings(struct net_device *dev, u32 stringset, u8 * data)
394 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++)
395 memcpy(data + i * ETH_GSTRING_LEN,
396 xenvif_stats[i].name, ETH_GSTRING_LEN);
401 static const struct ethtool_ops xenvif_ethtool_ops = {
402 .get_link = ethtool_op_get_link,
404 .get_sset_count = xenvif_get_sset_count,
405 .get_ethtool_stats = xenvif_get_ethtool_stats,
406 .get_strings = xenvif_get_strings,
409 static const struct net_device_ops xenvif_netdev_ops = {
410 .ndo_start_xmit = xenvif_start_xmit,
411 .ndo_get_stats = xenvif_get_stats,
412 .ndo_open = xenvif_open,
413 .ndo_stop = xenvif_close,
414 .ndo_change_mtu = xenvif_change_mtu,
415 .ndo_fix_features = xenvif_fix_features,
416 .ndo_set_mac_address = eth_mac_addr,
417 .ndo_validate_addr = eth_validate_addr,
420 struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
424 struct net_device *dev;
426 char name[IFNAMSIZ] = {};
428 snprintf(name, IFNAMSIZ - 1, "vif%u.%u", domid, handle);
429 /* Allocate a netdev with the max. supported number of queues.
430 * When the guest selects the desired number, it will be updated
431 * via netif_set_real_num_*_queues().
433 dev = alloc_netdev_mq(sizeof(struct xenvif), name, NET_NAME_UNKNOWN,
434 ether_setup, xenvif_max_queues);
436 pr_warn("Could not allocate netdev for %s\n", name);
437 return ERR_PTR(-ENOMEM);
440 SET_NETDEV_DEV(dev, parent);
442 vif = netdev_priv(dev);
445 vif->handle = handle;
449 vif->disabled = false;
451 /* Start out with no queues. */
455 dev->netdev_ops = &xenvif_netdev_ops;
456 dev->hw_features = NETIF_F_SG |
457 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
458 NETIF_F_TSO | NETIF_F_TSO6;
459 dev->features = dev->hw_features | NETIF_F_RXCSUM;
460 dev->ethtool_ops = &xenvif_ethtool_ops;
462 dev->tx_queue_len = XENVIF_QUEUE_LENGTH;
465 * Initialise a dummy MAC address. We choose the numerically
466 * largest non-broadcast address to prevent the address getting
467 * stolen by an Ethernet bridge for STP purposes.
468 * (FE:FF:FF:FF:FF:FF)
470 memset(dev->dev_addr, 0xFF, ETH_ALEN);
471 dev->dev_addr[0] &= ~0x01;
473 netif_carrier_off(dev);
475 err = register_netdev(dev);
477 netdev_warn(dev, "Could not register device: err=%d\n", err);
482 netdev_dbg(dev, "Successfully created xenvif\n");
484 __module_get(THIS_MODULE);
489 int xenvif_init_queue(struct xenvif_queue *queue)
493 queue->credit_bytes = queue->remaining_credit = ~0UL;
494 queue->credit_usec = 0UL;
495 init_timer(&queue->credit_timeout);
496 queue->credit_window_start = get_jiffies_64();
498 skb_queue_head_init(&queue->rx_queue);
499 skb_queue_head_init(&queue->tx_queue);
501 queue->pending_cons = 0;
502 queue->pending_prod = MAX_PENDING_REQS;
503 for (i = 0; i < MAX_PENDING_REQS; ++i)
504 queue->pending_ring[i] = i;
506 spin_lock_init(&queue->callback_lock);
507 spin_lock_init(&queue->response_lock);
509 /* If ballooning is disabled, this will consume real memory, so you
510 * better enable it. The long term solution would be to use just a
511 * bunch of valid page descriptors, without dependency on ballooning
513 err = alloc_xenballooned_pages(MAX_PENDING_REQS,
517 netdev_err(queue->vif->dev, "Could not reserve mmap_pages\n");
521 for (i = 0; i < MAX_PENDING_REQS; i++) {
522 queue->pending_tx_info[i].callback_struct = (struct ubuf_info)
523 { .callback = xenvif_zerocopy_callback,
526 queue->grant_tx_handle[i] = NETBACK_INVALID_HANDLE;
529 init_timer(&queue->rx_stalled);
531 netif_napi_add(queue->vif->dev, &queue->napi, xenvif_poll,
537 void xenvif_carrier_on(struct xenvif *vif)
540 if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
541 dev_set_mtu(vif->dev, ETH_DATA_LEN);
542 netdev_update_features(vif->dev);
543 set_bit(VIF_STATUS_CONNECTED, &vif->status);
544 netif_carrier_on(vif->dev);
545 if (netif_running(vif->dev))
550 int xenvif_connect(struct xenvif_queue *queue, unsigned long tx_ring_ref,
551 unsigned long rx_ring_ref, unsigned int tx_evtchn,
552 unsigned int rx_evtchn)
554 struct task_struct *task;
557 BUG_ON(queue->tx_irq);
559 BUG_ON(queue->dealloc_task);
561 err = xenvif_map_frontend_rings(queue, tx_ring_ref, rx_ring_ref);
565 init_waitqueue_head(&queue->wq);
566 init_waitqueue_head(&queue->dealloc_wq);
568 if (tx_evtchn == rx_evtchn) {
569 /* feature-split-event-channels == 0 */
570 err = bind_interdomain_evtchn_to_irqhandler(
571 queue->vif->domid, tx_evtchn, xenvif_interrupt, 0,
575 queue->tx_irq = queue->rx_irq = err;
576 disable_irq(queue->tx_irq);
578 /* feature-split-event-channels == 1 */
579 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name),
580 "%s-tx", queue->name);
581 err = bind_interdomain_evtchn_to_irqhandler(
582 queue->vif->domid, tx_evtchn, xenvif_tx_interrupt, 0,
583 queue->tx_irq_name, queue);
587 disable_irq(queue->tx_irq);
589 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name),
590 "%s-rx", queue->name);
591 err = bind_interdomain_evtchn_to_irqhandler(
592 queue->vif->domid, rx_evtchn, xenvif_rx_interrupt, 0,
593 queue->rx_irq_name, queue);
597 disable_irq(queue->rx_irq);
600 task = kthread_create(xenvif_kthread_guest_rx,
601 (void *)queue, "%s-guest-rx", queue->name);
603 pr_warn("Could not allocate kthread for %s\n", queue->name);
609 task = kthread_create(xenvif_dealloc_kthread,
610 (void *)queue, "%s-dealloc", queue->name);
612 pr_warn("Could not allocate kthread for %s\n", queue->name);
616 queue->dealloc_task = task;
618 wake_up_process(queue->task);
619 wake_up_process(queue->dealloc_task);
624 unbind_from_irqhandler(queue->rx_irq, queue);
627 unbind_from_irqhandler(queue->tx_irq, queue);
630 xenvif_unmap_frontend_rings(queue);
632 module_put(THIS_MODULE);
636 void xenvif_carrier_off(struct xenvif *vif)
638 struct net_device *dev = vif->dev;
641 if (test_and_clear_bit(VIF_STATUS_CONNECTED, &vif->status)) {
642 netif_carrier_off(dev); /* discard queued packets */
643 if (netif_running(dev))
649 static void xenvif_wait_unmap_timeout(struct xenvif_queue *queue,
650 unsigned int worst_case_skb_lifetime)
652 int i, unmap_timeout = 0;
654 for (i = 0; i < MAX_PENDING_REQS; ++i) {
655 if (queue->grant_tx_handle[i] != NETBACK_INVALID_HANDLE) {
657 schedule_timeout(msecs_to_jiffies(1000));
658 if (unmap_timeout > worst_case_skb_lifetime &&
660 netdev_err(queue->vif->dev,
661 "Page still granted! Index: %x\n",
668 void xenvif_disconnect(struct xenvif *vif)
670 struct xenvif_queue *queue = NULL;
671 unsigned int num_queues = vif->num_queues;
672 unsigned int queue_index;
674 xenvif_carrier_off(vif);
676 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
677 queue = &vif->queues[queue_index];
680 del_timer_sync(&queue->rx_stalled);
681 kthread_stop(queue->task);
685 if (queue->dealloc_task) {
686 kthread_stop(queue->dealloc_task);
687 queue->dealloc_task = NULL;
691 if (queue->tx_irq == queue->rx_irq)
692 unbind_from_irqhandler(queue->tx_irq, queue);
694 unbind_from_irqhandler(queue->tx_irq, queue);
695 unbind_from_irqhandler(queue->rx_irq, queue);
700 xenvif_unmap_frontend_rings(queue);
704 /* Reverse the relevant parts of xenvif_init_queue().
705 * Used for queue teardown from xenvif_free(), and on the
706 * error handling paths in xenbus.c:connect().
708 void xenvif_deinit_queue(struct xenvif_queue *queue)
710 free_xenballooned_pages(MAX_PENDING_REQS, queue->mmap_pages);
711 netif_napi_del(&queue->napi);
714 void xenvif_free(struct xenvif *vif)
716 struct xenvif_queue *queue = NULL;
717 unsigned int num_queues = vif->num_queues;
718 unsigned int queue_index;
719 /* Here we want to avoid timeout messages if an skb can be legitimately
720 * stuck somewhere else. Realistically this could be an another vif's
721 * internal or QDisc queue. That another vif also has this
722 * rx_drain_timeout_msecs timeout, so give it time to drain out.
723 * Although if that other guest wakes up just before its timeout happens
724 * and takes only one skb from QDisc, it can hold onto other skbs for a
727 unsigned int worst_case_skb_lifetime = (rx_drain_timeout_msecs/1000);
729 unregister_netdev(vif->dev);
731 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
732 queue = &vif->queues[queue_index];
733 xenvif_wait_unmap_timeout(queue, worst_case_skb_lifetime);
734 xenvif_deinit_queue(queue);
741 free_netdev(vif->dev);
743 module_put(THIS_MODULE);