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/sched/task.h>
35 #include <linux/ethtool.h>
36 #include <linux/rtnetlink.h>
37 #include <linux/if_vlan.h>
38 #include <linux/vmalloc.h>
40 #include <xen/events.h>
41 #include <asm/xen/hypercall.h>
42 #include <xen/balloon.h>
44 #define XENVIF_QUEUE_LENGTH 32
45 #define XENVIF_NAPI_WEIGHT 64
47 /* Number of bytes allowed on the internal guest Rx queue. */
48 #define XENVIF_RX_QUEUE_BYTES (XEN_NETIF_RX_RING_SIZE/2 * PAGE_SIZE)
50 /* This function is used to set SKBTX_DEV_ZEROCOPY as well as
51 * increasing the inflight counter. We need to increase the inflight
52 * counter because core driver calls into xenvif_zerocopy_callback
53 * which calls xenvif_skb_zerocopy_complete.
55 void xenvif_skb_zerocopy_prepare(struct xenvif_queue *queue,
58 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
59 atomic_inc(&queue->inflight_packets);
62 void xenvif_skb_zerocopy_complete(struct xenvif_queue *queue)
64 atomic_dec(&queue->inflight_packets);
66 /* Wake the dealloc thread _after_ decrementing inflight_packets so
67 * that if kthread_stop() has already been called, the dealloc thread
68 * does not wait forever with nothing to wake it.
70 wake_up(&queue->dealloc_wq);
73 int xenvif_schedulable(struct xenvif *vif)
75 return netif_running(vif->dev) &&
76 test_bit(VIF_STATUS_CONNECTED, &vif->status) &&
80 static irqreturn_t xenvif_tx_interrupt(int irq, void *dev_id)
82 struct xenvif_queue *queue = dev_id;
84 if (RING_HAS_UNCONSUMED_REQUESTS(&queue->tx))
85 napi_schedule(&queue->napi);
90 static int xenvif_poll(struct napi_struct *napi, int budget)
92 struct xenvif_queue *queue =
93 container_of(napi, struct xenvif_queue, napi);
96 /* This vif is rogue, we pretend we've there is nothing to do
97 * for this vif to deschedule it from NAPI. But this interface
98 * will be turned off in thread context later.
100 if (unlikely(queue->vif->disabled)) {
105 work_done = xenvif_tx_action(queue, budget);
107 if (work_done < budget) {
108 napi_complete_done(napi, work_done);
109 /* If the queue is rate-limited, it shall be
110 * rescheduled in the timer callback.
112 if (likely(!queue->rate_limited))
113 xenvif_napi_schedule_or_enable_events(queue);
119 static irqreturn_t xenvif_rx_interrupt(int irq, void *dev_id)
121 struct xenvif_queue *queue = dev_id;
123 xenvif_kick_thread(queue);
128 irqreturn_t xenvif_interrupt(int irq, void *dev_id)
130 xenvif_tx_interrupt(irq, dev_id);
131 xenvif_rx_interrupt(irq, dev_id);
136 int xenvif_queue_stopped(struct xenvif_queue *queue)
138 struct net_device *dev = queue->vif->dev;
139 unsigned int id = queue->id;
140 return netif_tx_queue_stopped(netdev_get_tx_queue(dev, id));
143 void xenvif_wake_queue(struct xenvif_queue *queue)
145 struct net_device *dev = queue->vif->dev;
146 unsigned int id = queue->id;
147 netif_tx_wake_queue(netdev_get_tx_queue(dev, id));
150 static u16 xenvif_select_queue(struct net_device *dev, struct sk_buff *skb,
152 select_queue_fallback_t fallback)
154 struct xenvif *vif = netdev_priv(dev);
155 unsigned int size = vif->hash.size;
156 unsigned int num_queues;
158 /* If queues are not set up internally - always return 0
159 * as the packet going to be dropped anyway */
160 num_queues = READ_ONCE(vif->num_queues);
164 if (vif->hash.alg == XEN_NETIF_CTRL_HASH_ALGORITHM_NONE)
165 return fallback(dev, skb) % dev->real_num_tx_queues;
167 xenvif_set_skb_hash(vif, skb);
170 return skb_get_hash_raw(skb) % dev->real_num_tx_queues;
172 return vif->hash.mapping[skb_get_hash_raw(skb) % size];
175 static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
177 struct xenvif *vif = netdev_priv(dev);
178 struct xenvif_queue *queue = NULL;
179 unsigned int num_queues;
181 struct xenvif_rx_cb *cb;
183 BUG_ON(skb->dev != dev);
185 /* Drop the packet if queues are not set up.
186 * This handler should be called inside an RCU read section
187 * so we don't need to enter it here explicitly.
189 num_queues = READ_ONCE(vif->num_queues);
193 /* Obtain the queue to be used to transmit this packet */
194 index = skb_get_queue_mapping(skb);
195 if (index >= num_queues) {
196 pr_warn_ratelimited("Invalid queue %hu for packet on interface %s\n.",
197 index, vif->dev->name);
200 queue = &vif->queues[index];
202 /* Drop the packet if queue is not ready */
203 if (queue->task == NULL ||
204 queue->dealloc_task == NULL ||
205 !xenvif_schedulable(vif))
208 if (vif->multicast_control && skb->pkt_type == PACKET_MULTICAST) {
209 struct ethhdr *eth = (struct ethhdr *)skb->data;
211 if (!xenvif_mcast_match(vif, eth->h_dest))
215 cb = XENVIF_RX_CB(skb);
216 cb->expires = jiffies + vif->drain_timeout;
218 /* If there is no hash algorithm configured then make sure there
219 * is no hash information in the socket buffer otherwise it
220 * would be incorrectly forwarded to the frontend.
222 if (vif->hash.alg == XEN_NETIF_CTRL_HASH_ALGORITHM_NONE)
225 xenvif_rx_queue_tail(queue, skb);
226 xenvif_kick_thread(queue);
231 vif->dev->stats.tx_dropped++;
236 static struct net_device_stats *xenvif_get_stats(struct net_device *dev)
238 struct xenvif *vif = netdev_priv(dev);
239 struct xenvif_queue *queue = NULL;
240 unsigned int num_queues;
248 num_queues = READ_ONCE(vif->num_queues);
250 /* Aggregate tx and rx stats from each queue */
251 for (index = 0; index < num_queues; ++index) {
252 queue = &vif->queues[index];
253 rx_bytes += queue->stats.rx_bytes;
254 rx_packets += queue->stats.rx_packets;
255 tx_bytes += queue->stats.tx_bytes;
256 tx_packets += queue->stats.tx_packets;
261 vif->dev->stats.rx_bytes = rx_bytes;
262 vif->dev->stats.rx_packets = rx_packets;
263 vif->dev->stats.tx_bytes = tx_bytes;
264 vif->dev->stats.tx_packets = tx_packets;
266 return &vif->dev->stats;
269 static void xenvif_up(struct xenvif *vif)
271 struct xenvif_queue *queue = NULL;
272 unsigned int num_queues = vif->num_queues;
273 unsigned int queue_index;
275 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
276 queue = &vif->queues[queue_index];
277 napi_enable(&queue->napi);
278 enable_irq(queue->tx_irq);
279 if (queue->tx_irq != queue->rx_irq)
280 enable_irq(queue->rx_irq);
281 xenvif_napi_schedule_or_enable_events(queue);
285 static void xenvif_down(struct xenvif *vif)
287 struct xenvif_queue *queue = NULL;
288 unsigned int num_queues = vif->num_queues;
289 unsigned int queue_index;
291 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
292 queue = &vif->queues[queue_index];
293 disable_irq(queue->tx_irq);
294 if (queue->tx_irq != queue->rx_irq)
295 disable_irq(queue->rx_irq);
296 napi_disable(&queue->napi);
297 del_timer_sync(&queue->credit_timeout);
301 static int xenvif_open(struct net_device *dev)
303 struct xenvif *vif = netdev_priv(dev);
304 if (test_bit(VIF_STATUS_CONNECTED, &vif->status))
306 netif_tx_start_all_queues(dev);
310 static int xenvif_close(struct net_device *dev)
312 struct xenvif *vif = netdev_priv(dev);
313 if (test_bit(VIF_STATUS_CONNECTED, &vif->status))
315 netif_tx_stop_all_queues(dev);
319 static int xenvif_change_mtu(struct net_device *dev, int mtu)
321 struct xenvif *vif = netdev_priv(dev);
322 int max = vif->can_sg ? ETH_MAX_MTU - VLAN_ETH_HLEN : ETH_DATA_LEN;
330 static netdev_features_t xenvif_fix_features(struct net_device *dev,
331 netdev_features_t features)
333 struct xenvif *vif = netdev_priv(dev);
336 features &= ~NETIF_F_SG;
337 if (~(vif->gso_mask) & GSO_BIT(TCPV4))
338 features &= ~NETIF_F_TSO;
339 if (~(vif->gso_mask) & GSO_BIT(TCPV6))
340 features &= ~NETIF_F_TSO6;
342 features &= ~NETIF_F_IP_CSUM;
344 features &= ~NETIF_F_IPV6_CSUM;
349 static const struct xenvif_stat {
350 char name[ETH_GSTRING_LEN];
354 "rx_gso_checksum_fixup",
355 offsetof(struct xenvif_stats, rx_gso_checksum_fixup)
357 /* If (sent != success + fail), there are probably packets never
362 offsetof(struct xenvif_stats, tx_zerocopy_sent),
365 "tx_zerocopy_success",
366 offsetof(struct xenvif_stats, tx_zerocopy_success),
370 offsetof(struct xenvif_stats, tx_zerocopy_fail)
372 /* Number of packets exceeding MAX_SKB_FRAG slots. You should use
373 * a guest with the same MAX_SKB_FRAG
377 offsetof(struct xenvif_stats, tx_frag_overflow)
381 static int xenvif_get_sset_count(struct net_device *dev, int string_set)
383 switch (string_set) {
385 return ARRAY_SIZE(xenvif_stats);
391 static void xenvif_get_ethtool_stats(struct net_device *dev,
392 struct ethtool_stats *stats, u64 * data)
394 struct xenvif *vif = netdev_priv(dev);
395 unsigned int num_queues;
397 unsigned int queue_index;
400 num_queues = READ_ONCE(vif->num_queues);
402 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++) {
403 unsigned long accum = 0;
404 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
405 void *vif_stats = &vif->queues[queue_index].stats;
406 accum += *(unsigned long *)(vif_stats + xenvif_stats[i].offset);
414 static void xenvif_get_strings(struct net_device *dev, u32 stringset, u8 * data)
420 for (i = 0; i < ARRAY_SIZE(xenvif_stats); i++)
421 memcpy(data + i * ETH_GSTRING_LEN,
422 xenvif_stats[i].name, ETH_GSTRING_LEN);
427 static const struct ethtool_ops xenvif_ethtool_ops = {
428 .get_link = ethtool_op_get_link,
430 .get_sset_count = xenvif_get_sset_count,
431 .get_ethtool_stats = xenvif_get_ethtool_stats,
432 .get_strings = xenvif_get_strings,
435 static const struct net_device_ops xenvif_netdev_ops = {
436 .ndo_select_queue = xenvif_select_queue,
437 .ndo_start_xmit = xenvif_start_xmit,
438 .ndo_get_stats = xenvif_get_stats,
439 .ndo_open = xenvif_open,
440 .ndo_stop = xenvif_close,
441 .ndo_change_mtu = xenvif_change_mtu,
442 .ndo_fix_features = xenvif_fix_features,
443 .ndo_set_mac_address = eth_mac_addr,
444 .ndo_validate_addr = eth_validate_addr,
447 struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
451 struct net_device *dev;
453 char name[IFNAMSIZ] = {};
455 snprintf(name, IFNAMSIZ - 1, "vif%u.%u", domid, handle);
456 /* Allocate a netdev with the max. supported number of queues.
457 * When the guest selects the desired number, it will be updated
458 * via netif_set_real_num_*_queues().
460 dev = alloc_netdev_mq(sizeof(struct xenvif), name, NET_NAME_UNKNOWN,
461 ether_setup, xenvif_max_queues);
463 pr_warn("Could not allocate netdev for %s\n", name);
464 return ERR_PTR(-ENOMEM);
467 SET_NETDEV_DEV(dev, parent);
469 vif = netdev_priv(dev);
472 vif->handle = handle;
476 vif->disabled = false;
477 vif->drain_timeout = msecs_to_jiffies(rx_drain_timeout_msecs);
478 vif->stall_timeout = msecs_to_jiffies(rx_stall_timeout_msecs);
480 /* Start out with no queues. */
484 spin_lock_init(&vif->lock);
485 INIT_LIST_HEAD(&vif->fe_mcast_addr);
487 dev->netdev_ops = &xenvif_netdev_ops;
488 dev->hw_features = NETIF_F_SG |
489 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
490 NETIF_F_TSO | NETIF_F_TSO6 | NETIF_F_FRAGLIST;
491 dev->features = dev->hw_features | NETIF_F_RXCSUM;
492 dev->ethtool_ops = &xenvif_ethtool_ops;
494 dev->tx_queue_len = XENVIF_QUEUE_LENGTH;
496 dev->min_mtu = ETH_MIN_MTU;
497 dev->max_mtu = ETH_MAX_MTU - VLAN_ETH_HLEN;
500 * Initialise a dummy MAC address. We choose the numerically
501 * largest non-broadcast address to prevent the address getting
502 * stolen by an Ethernet bridge for STP purposes.
503 * (FE:FF:FF:FF:FF:FF)
505 eth_broadcast_addr(dev->dev_addr);
506 dev->dev_addr[0] &= ~0x01;
508 netif_carrier_off(dev);
510 err = register_netdev(dev);
512 netdev_warn(dev, "Could not register device: err=%d\n", err);
517 netdev_dbg(dev, "Successfully created xenvif\n");
519 __module_get(THIS_MODULE);
524 int xenvif_init_queue(struct xenvif_queue *queue)
528 queue->credit_bytes = queue->remaining_credit = ~0UL;
529 queue->credit_usec = 0UL;
530 init_timer(&queue->credit_timeout);
531 queue->credit_timeout.function = xenvif_tx_credit_callback;
532 queue->credit_window_start = get_jiffies_64();
534 queue->rx_queue_max = XENVIF_RX_QUEUE_BYTES;
536 skb_queue_head_init(&queue->rx_queue);
537 skb_queue_head_init(&queue->tx_queue);
539 queue->pending_cons = 0;
540 queue->pending_prod = MAX_PENDING_REQS;
541 for (i = 0; i < MAX_PENDING_REQS; ++i)
542 queue->pending_ring[i] = i;
544 spin_lock_init(&queue->callback_lock);
545 spin_lock_init(&queue->response_lock);
547 /* If ballooning is disabled, this will consume real memory, so you
548 * better enable it. The long term solution would be to use just a
549 * bunch of valid page descriptors, without dependency on ballooning
551 err = gnttab_alloc_pages(MAX_PENDING_REQS,
554 netdev_err(queue->vif->dev, "Could not reserve mmap_pages\n");
558 for (i = 0; i < MAX_PENDING_REQS; i++) {
559 queue->pending_tx_info[i].callback_struct = (struct ubuf_info)
560 { .callback = xenvif_zerocopy_callback,
563 queue->grant_tx_handle[i] = NETBACK_INVALID_HANDLE;
569 void xenvif_carrier_on(struct xenvif *vif)
572 if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
573 dev_set_mtu(vif->dev, ETH_DATA_LEN);
574 netdev_update_features(vif->dev);
575 set_bit(VIF_STATUS_CONNECTED, &vif->status);
576 if (netif_running(vif->dev))
581 int xenvif_connect_ctrl(struct xenvif *vif, grant_ref_t ring_ref,
584 struct net_device *dev = vif->dev;
586 struct xen_netif_ctrl_sring *shared;
589 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
590 &ring_ref, 1, &addr);
594 shared = (struct xen_netif_ctrl_sring *)addr;
595 BACK_RING_INIT(&vif->ctrl, shared, XEN_PAGE_SIZE);
597 err = bind_interdomain_evtchn_to_irq(vif->domid, evtchn);
603 xenvif_init_hash(vif);
605 err = request_threaded_irq(vif->ctrl_irq, NULL, xenvif_ctrl_irq_fn,
606 IRQF_ONESHOT, "xen-netback-ctrl", vif);
608 pr_warn("Could not setup irq handler for %s\n", dev->name);
615 xenvif_deinit_hash(vif);
616 unbind_from_irqhandler(vif->ctrl_irq, vif);
620 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
622 vif->ctrl.sring = NULL;
628 int xenvif_connect_data(struct xenvif_queue *queue,
629 unsigned long tx_ring_ref,
630 unsigned long rx_ring_ref,
631 unsigned int tx_evtchn,
632 unsigned int rx_evtchn)
634 struct task_struct *task;
637 BUG_ON(queue->tx_irq);
639 BUG_ON(queue->dealloc_task);
641 err = xenvif_map_frontend_data_rings(queue, tx_ring_ref,
646 init_waitqueue_head(&queue->wq);
647 init_waitqueue_head(&queue->dealloc_wq);
648 atomic_set(&queue->inflight_packets, 0);
650 netif_napi_add(queue->vif->dev, &queue->napi, xenvif_poll,
653 if (tx_evtchn == rx_evtchn) {
654 /* feature-split-event-channels == 0 */
655 err = bind_interdomain_evtchn_to_irqhandler(
656 queue->vif->domid, tx_evtchn, xenvif_interrupt, 0,
660 queue->tx_irq = queue->rx_irq = err;
661 disable_irq(queue->tx_irq);
663 /* feature-split-event-channels == 1 */
664 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name),
665 "%s-tx", queue->name);
666 err = bind_interdomain_evtchn_to_irqhandler(
667 queue->vif->domid, tx_evtchn, xenvif_tx_interrupt, 0,
668 queue->tx_irq_name, queue);
672 disable_irq(queue->tx_irq);
674 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name),
675 "%s-rx", queue->name);
676 err = bind_interdomain_evtchn_to_irqhandler(
677 queue->vif->domid, rx_evtchn, xenvif_rx_interrupt, 0,
678 queue->rx_irq_name, queue);
682 disable_irq(queue->rx_irq);
685 queue->stalled = true;
687 task = kthread_create(xenvif_kthread_guest_rx,
688 (void *)queue, "%s-guest-rx", queue->name);
690 pr_warn("Could not allocate kthread for %s\n", queue->name);
695 get_task_struct(task);
697 task = kthread_create(xenvif_dealloc_kthread,
698 (void *)queue, "%s-dealloc", queue->name);
700 pr_warn("Could not allocate kthread for %s\n", queue->name);
704 queue->dealloc_task = task;
706 wake_up_process(queue->task);
707 wake_up_process(queue->dealloc_task);
712 unbind_from_irqhandler(queue->rx_irq, queue);
715 unbind_from_irqhandler(queue->tx_irq, queue);
718 xenvif_unmap_frontend_data_rings(queue);
719 netif_napi_del(&queue->napi);
721 module_put(THIS_MODULE);
725 void xenvif_carrier_off(struct xenvif *vif)
727 struct net_device *dev = vif->dev;
730 if (test_and_clear_bit(VIF_STATUS_CONNECTED, &vif->status)) {
731 netif_carrier_off(dev); /* discard queued packets */
732 if (netif_running(dev))
738 void xenvif_disconnect_data(struct xenvif *vif)
740 struct xenvif_queue *queue = NULL;
741 unsigned int num_queues = vif->num_queues;
742 unsigned int queue_index;
744 xenvif_carrier_off(vif);
746 for (queue_index = 0; queue_index < num_queues; ++queue_index) {
747 queue = &vif->queues[queue_index];
749 netif_napi_del(&queue->napi);
752 kthread_stop(queue->task);
753 put_task_struct(queue->task);
757 if (queue->dealloc_task) {
758 kthread_stop(queue->dealloc_task);
759 queue->dealloc_task = NULL;
763 if (queue->tx_irq == queue->rx_irq)
764 unbind_from_irqhandler(queue->tx_irq, queue);
766 unbind_from_irqhandler(queue->tx_irq, queue);
767 unbind_from_irqhandler(queue->rx_irq, queue);
772 xenvif_unmap_frontend_data_rings(queue);
775 xenvif_mcast_addr_list_free(vif);
778 void xenvif_disconnect_ctrl(struct xenvif *vif)
781 xenvif_deinit_hash(vif);
782 unbind_from_irqhandler(vif->ctrl_irq, vif);
786 if (vif->ctrl.sring) {
787 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
789 vif->ctrl.sring = NULL;
793 /* Reverse the relevant parts of xenvif_init_queue().
794 * Used for queue teardown from xenvif_free(), and on the
795 * error handling paths in xenbus.c:connect().
797 void xenvif_deinit_queue(struct xenvif_queue *queue)
799 gnttab_free_pages(MAX_PENDING_REQS, queue->mmap_pages);
802 void xenvif_free(struct xenvif *vif)
804 struct xenvif_queue *queues = vif->queues;
805 unsigned int num_queues = vif->num_queues;
806 unsigned int queue_index;
808 unregister_netdev(vif->dev);
809 free_netdev(vif->dev);
811 for (queue_index = 0; queue_index < num_queues; ++queue_index)
812 xenvif_deinit_queue(&queues[queue_index]);
815 module_put(THIS_MODULE);