2 * Virtual network driver for conversing with remote driver backends.
4 * Copyright (c) 2002-2005, K A Fraser
5 * Copyright (c) 2005, XenSource Ltd
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License version 2
9 * as published by the Free Software Foundation; or, when distributed
10 * separately from the Linux kernel or incorporated into other
11 * software packages, subject to the following license:
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this source file (the "Software"), to deal in the Software without
15 * restriction, including without limitation the rights to use, copy, modify,
16 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
17 * and to permit persons to whom the Software is furnished to do so, subject to
18 * the following conditions:
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
26 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
28 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
32 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
34 #include <linux/module.h>
35 #include <linux/kernel.h>
36 #include <linux/netdevice.h>
37 #include <linux/etherdevice.h>
38 #include <linux/skbuff.h>
39 #include <linux/ethtool.h>
40 #include <linux/if_ether.h>
42 #include <linux/udp.h>
43 #include <linux/moduleparam.h>
45 #include <linux/slab.h>
49 #include <xen/xenbus.h>
50 #include <xen/events.h>
52 #include <xen/platform_pci.h>
53 #include <xen/grant_table.h>
55 #include <xen/interface/io/netif.h>
56 #include <xen/interface/memory.h>
57 #include <xen/interface/grant_table.h>
59 /* Module parameters */
60 #define MAX_QUEUES_DEFAULT 8
61 static unsigned int xennet_max_queues;
62 module_param_named(max_queues, xennet_max_queues, uint, 0644);
63 MODULE_PARM_DESC(max_queues,
64 "Maximum number of queues per virtual interface");
66 #define XENNET_TIMEOUT (5 * HZ)
68 static const struct ethtool_ops xennet_ethtool_ops;
74 #define NETFRONT_SKB_CB(skb) ((struct netfront_cb *)((skb)->cb))
76 #define RX_COPY_THRESHOLD 256
78 #define GRANT_INVALID_REF 0
80 #define NET_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, XEN_PAGE_SIZE)
81 #define NET_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, XEN_PAGE_SIZE)
83 /* Minimum number of Rx slots (includes slot for GSO metadata). */
84 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
86 /* Queue name is interface name with "-qNNN" appended */
87 #define QUEUE_NAME_SIZE (IFNAMSIZ + 6)
89 /* IRQ name is queue name with "-tx" or "-rx" appended */
90 #define IRQ_NAME_SIZE (QUEUE_NAME_SIZE + 3)
92 static DECLARE_WAIT_QUEUE_HEAD(module_wq);
94 struct netfront_stats {
97 struct u64_stats_sync syncp;
100 struct netfront_info;
102 struct netfront_queue {
103 unsigned int id; /* Queue ID, 0-based */
104 char name[QUEUE_NAME_SIZE]; /* DEVNAME-qN */
105 struct netfront_info *info;
107 struct napi_struct napi;
109 /* Split event channels support, tx_* == rx_* when using
110 * single event channel.
112 unsigned int tx_evtchn, rx_evtchn;
113 unsigned int tx_irq, rx_irq;
114 /* Only used when split event channels support is enabled */
115 char tx_irq_name[IRQ_NAME_SIZE]; /* DEVNAME-qN-tx */
116 char rx_irq_name[IRQ_NAME_SIZE]; /* DEVNAME-qN-rx */
119 struct xen_netif_tx_front_ring tx;
123 * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries
124 * are linked from tx_skb_freelist through tx_link.
126 struct sk_buff *tx_skbs[NET_TX_RING_SIZE];
127 unsigned short tx_link[NET_TX_RING_SIZE];
128 #define TX_LINK_NONE 0xffff
129 #define TX_PENDING 0xfffe
130 grant_ref_t gref_tx_head;
131 grant_ref_t grant_tx_ref[NET_TX_RING_SIZE];
132 struct page *grant_tx_page[NET_TX_RING_SIZE];
133 unsigned tx_skb_freelist;
134 unsigned int tx_pend_queue;
136 spinlock_t rx_lock ____cacheline_aligned_in_smp;
137 struct xen_netif_rx_front_ring rx;
140 struct timer_list rx_refill_timer;
142 struct sk_buff *rx_skbs[NET_RX_RING_SIZE];
143 grant_ref_t gref_rx_head;
144 grant_ref_t grant_rx_ref[NET_RX_RING_SIZE];
146 unsigned int rx_rsp_unconsumed;
147 spinlock_t rx_cons_lock;
150 struct netfront_info {
151 struct list_head list;
152 struct net_device *netdev;
154 struct xenbus_device *xbdev;
156 /* Multi-queue support */
157 struct netfront_queue *queues;
160 struct netfront_stats __percpu *rx_stats;
161 struct netfront_stats __percpu *tx_stats;
163 /* Is device behaving sane? */
166 atomic_t rx_gso_checksum_fixup;
169 struct netfront_rx_info {
170 struct xen_netif_rx_response rx;
171 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
175 * Access macros for acquiring freeing slots in tx_skbs[].
178 static void add_id_to_list(unsigned *head, unsigned short *list,
185 static unsigned short get_id_from_list(unsigned *head, unsigned short *list)
187 unsigned int id = *head;
189 if (id != TX_LINK_NONE) {
191 list[id] = TX_LINK_NONE;
196 static int xennet_rxidx(RING_IDX idx)
198 return idx & (NET_RX_RING_SIZE - 1);
201 static struct sk_buff *xennet_get_rx_skb(struct netfront_queue *queue,
204 int i = xennet_rxidx(ri);
205 struct sk_buff *skb = queue->rx_skbs[i];
206 queue->rx_skbs[i] = NULL;
210 static grant_ref_t xennet_get_rx_ref(struct netfront_queue *queue,
213 int i = xennet_rxidx(ri);
214 grant_ref_t ref = queue->grant_rx_ref[i];
215 queue->grant_rx_ref[i] = GRANT_INVALID_REF;
220 static const struct attribute_group xennet_dev_group;
223 static bool xennet_can_sg(struct net_device *dev)
225 return dev->features & NETIF_F_SG;
229 static void rx_refill_timeout(struct timer_list *t)
231 struct netfront_queue *queue = from_timer(queue, t, rx_refill_timer);
232 napi_schedule(&queue->napi);
235 static int netfront_tx_slot_available(struct netfront_queue *queue)
237 return (queue->tx.req_prod_pvt - queue->tx.rsp_cons) <
238 (NET_TX_RING_SIZE - XEN_NETIF_NR_SLOTS_MIN - 1);
241 static void xennet_maybe_wake_tx(struct netfront_queue *queue)
243 struct net_device *dev = queue->info->netdev;
244 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, queue->id);
246 if (unlikely(netif_tx_queue_stopped(dev_queue)) &&
247 netfront_tx_slot_available(queue) &&
248 likely(netif_running(dev)))
249 netif_tx_wake_queue(netdev_get_tx_queue(dev, queue->id));
253 static struct sk_buff *xennet_alloc_one_rx_buffer(struct netfront_queue *queue)
258 skb = __netdev_alloc_skb(queue->info->netdev,
259 RX_COPY_THRESHOLD + NET_IP_ALIGN,
260 GFP_ATOMIC | __GFP_NOWARN);
264 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
269 skb_add_rx_frag(skb, 0, page, 0, 0, PAGE_SIZE);
271 /* Align ip header to a 16 bytes boundary */
272 skb_reserve(skb, NET_IP_ALIGN);
273 skb->dev = queue->info->netdev;
279 static void xennet_alloc_rx_buffers(struct netfront_queue *queue)
281 RING_IDX req_prod = queue->rx.req_prod_pvt;
285 if (unlikely(!netif_carrier_ok(queue->info->netdev)))
288 for (req_prod = queue->rx.req_prod_pvt;
289 req_prod - queue->rx.rsp_cons < NET_RX_RING_SIZE;
295 struct xen_netif_rx_request *req;
297 skb = xennet_alloc_one_rx_buffer(queue);
303 id = xennet_rxidx(req_prod);
305 BUG_ON(queue->rx_skbs[id]);
306 queue->rx_skbs[id] = skb;
308 ref = gnttab_claim_grant_reference(&queue->gref_rx_head);
309 WARN_ON_ONCE(IS_ERR_VALUE((unsigned long)(int)ref));
310 queue->grant_rx_ref[id] = ref;
312 page = skb_frag_page(&skb_shinfo(skb)->frags[0]);
314 req = RING_GET_REQUEST(&queue->rx, req_prod);
315 gnttab_page_grant_foreign_access_ref_one(ref,
316 queue->info->xbdev->otherend_id,
323 queue->rx.req_prod_pvt = req_prod;
325 /* Try again later if there are not enough requests or skb allocation
327 * Enough requests is quantified as the sum of newly created slots and
328 * the unconsumed slots at the backend.
330 if (req_prod - queue->rx.rsp_cons < NET_RX_SLOTS_MIN ||
332 mod_timer(&queue->rx_refill_timer, jiffies + (HZ/10));
336 wmb(); /* barrier so backend seens requests */
338 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->rx, notify);
340 notify_remote_via_irq(queue->rx_irq);
343 static int xennet_open(struct net_device *dev)
345 struct netfront_info *np = netdev_priv(dev);
346 unsigned int num_queues = dev->real_num_tx_queues;
348 struct netfront_queue *queue = NULL;
350 if (!np->queues || np->broken)
353 for (i = 0; i < num_queues; ++i) {
354 queue = &np->queues[i];
355 napi_enable(&queue->napi);
357 spin_lock_bh(&queue->rx_lock);
358 if (netif_carrier_ok(dev)) {
359 xennet_alloc_rx_buffers(queue);
360 queue->rx.sring->rsp_event = queue->rx.rsp_cons + 1;
361 if (RING_HAS_UNCONSUMED_RESPONSES(&queue->rx))
362 napi_schedule(&queue->napi);
364 spin_unlock_bh(&queue->rx_lock);
367 netif_tx_start_all_queues(dev);
372 static bool xennet_tx_buf_gc(struct netfront_queue *queue)
378 bool work_done = false;
379 const struct device *dev = &queue->info->netdev->dev;
381 BUG_ON(!netif_carrier_ok(queue->info->netdev));
384 prod = queue->tx.sring->rsp_prod;
385 if (RING_RESPONSE_PROD_OVERFLOW(&queue->tx, prod)) {
386 dev_alert(dev, "Illegal number of responses %u\n",
387 prod - queue->tx.rsp_cons);
390 rmb(); /* Ensure we see responses up to 'rp'. */
392 for (cons = queue->tx.rsp_cons; cons != prod; cons++) {
393 struct xen_netif_tx_response txrsp;
397 RING_COPY_RESPONSE(&queue->tx, cons, &txrsp);
398 if (txrsp.status == XEN_NETIF_RSP_NULL)
402 if (id >= RING_SIZE(&queue->tx)) {
404 "Response has incorrect id (%u)\n",
408 if (queue->tx_link[id] != TX_PENDING) {
410 "Response for inactive request\n");
414 queue->tx_link[id] = TX_LINK_NONE;
415 skb = queue->tx_skbs[id];
416 queue->tx_skbs[id] = NULL;
417 if (unlikely(!gnttab_end_foreign_access_ref(
418 queue->grant_tx_ref[id], GNTMAP_readonly))) {
420 "Grant still in use by backend domain\n");
423 gnttab_release_grant_reference(
424 &queue->gref_tx_head, queue->grant_tx_ref[id]);
425 queue->grant_tx_ref[id] = GRANT_INVALID_REF;
426 queue->grant_tx_page[id] = NULL;
427 add_id_to_list(&queue->tx_skb_freelist, queue->tx_link, id);
428 dev_kfree_skb_irq(skb);
431 queue->tx.rsp_cons = prod;
433 RING_FINAL_CHECK_FOR_RESPONSES(&queue->tx, more_to_do);
434 } while (more_to_do);
436 xennet_maybe_wake_tx(queue);
441 queue->info->broken = true;
442 dev_alert(dev, "Disabled for further use\n");
447 struct xennet_gnttab_make_txreq {
448 struct netfront_queue *queue;
451 struct xen_netif_tx_request *tx; /* Last request on ring page */
452 struct xen_netif_tx_request tx_local; /* Last request local copy*/
456 static void xennet_tx_setup_grant(unsigned long gfn, unsigned int offset,
457 unsigned int len, void *data)
459 struct xennet_gnttab_make_txreq *info = data;
461 struct xen_netif_tx_request *tx;
463 /* convenient aliases */
464 struct page *page = info->page;
465 struct netfront_queue *queue = info->queue;
466 struct sk_buff *skb = info->skb;
468 id = get_id_from_list(&queue->tx_skb_freelist, queue->tx_link);
469 tx = RING_GET_REQUEST(&queue->tx, queue->tx.req_prod_pvt++);
470 ref = gnttab_claim_grant_reference(&queue->gref_tx_head);
471 WARN_ON_ONCE(IS_ERR_VALUE((unsigned long)(int)ref));
473 gnttab_grant_foreign_access_ref(ref, queue->info->xbdev->otherend_id,
474 gfn, GNTMAP_readonly);
476 queue->tx_skbs[id] = skb;
477 queue->grant_tx_page[id] = page;
478 queue->grant_tx_ref[id] = ref;
480 info->tx_local.id = id;
481 info->tx_local.gref = ref;
482 info->tx_local.offset = offset;
483 info->tx_local.size = len;
484 info->tx_local.flags = 0;
486 *tx = info->tx_local;
489 * Put the request in the pending queue, it will be set to be pending
490 * when the producer index is about to be raised.
492 add_id_to_list(&queue->tx_pend_queue, queue->tx_link, id);
495 info->size += info->tx_local.size;
498 static struct xen_netif_tx_request *xennet_make_first_txreq(
499 struct xennet_gnttab_make_txreq *info,
500 unsigned int offset, unsigned int len)
504 gnttab_for_one_grant(info->page, offset, len, xennet_tx_setup_grant, info);
509 static void xennet_make_one_txreq(unsigned long gfn, unsigned int offset,
510 unsigned int len, void *data)
512 struct xennet_gnttab_make_txreq *info = data;
514 info->tx->flags |= XEN_NETTXF_more_data;
516 xennet_tx_setup_grant(gfn, offset, len, data);
519 static void xennet_make_txreqs(
520 struct xennet_gnttab_make_txreq *info,
522 unsigned int offset, unsigned int len)
524 /* Skip unused frames from start of page */
525 page += offset >> PAGE_SHIFT;
526 offset &= ~PAGE_MASK;
532 gnttab_foreach_grant_in_range(page, offset, len,
533 xennet_make_one_txreq,
543 * Count how many ring slots are required to send this skb. Each frag
544 * might be a compound page.
546 static int xennet_count_skb_slots(struct sk_buff *skb)
548 int i, frags = skb_shinfo(skb)->nr_frags;
551 slots = gnttab_count_grant(offset_in_page(skb->data),
554 for (i = 0; i < frags; i++) {
555 skb_frag_t *frag = skb_shinfo(skb)->frags + i;
556 unsigned long size = skb_frag_size(frag);
557 unsigned long offset = frag->page_offset;
559 /* Skip unused frames from start of page */
560 offset &= ~PAGE_MASK;
562 slots += gnttab_count_grant(offset, size);
568 static u16 xennet_select_queue(struct net_device *dev, struct sk_buff *skb,
569 struct net_device *sb_dev,
570 select_queue_fallback_t fallback)
572 unsigned int num_queues = dev->real_num_tx_queues;
576 /* First, check if there is only one queue */
577 if (num_queues == 1) {
580 hash = skb_get_hash(skb);
581 queue_idx = hash % num_queues;
587 static void xennet_mark_tx_pending(struct netfront_queue *queue)
591 while ((i = get_id_from_list(&queue->tx_pend_queue, queue->tx_link)) !=
593 queue->tx_link[i] = TX_PENDING;
596 #define MAX_XEN_SKB_FRAGS (65536 / XEN_PAGE_SIZE + 1)
598 static netdev_tx_t xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
600 struct netfront_info *np = netdev_priv(dev);
601 struct netfront_stats *tx_stats = this_cpu_ptr(np->tx_stats);
602 struct xen_netif_tx_request *first_tx;
610 struct netfront_queue *queue = NULL;
611 struct xennet_gnttab_make_txreq info = { };
612 unsigned int num_queues = dev->real_num_tx_queues;
614 struct sk_buff *nskb;
616 /* Drop the packet if no queues are set up */
619 if (unlikely(np->broken))
621 /* Determine which queue to transmit this SKB on */
622 queue_index = skb_get_queue_mapping(skb);
623 queue = &np->queues[queue_index];
625 /* If skb->len is too big for wire format, drop skb and alert
626 * user about misconfiguration.
628 if (unlikely(skb->len > XEN_NETIF_MAX_TX_SIZE)) {
629 net_alert_ratelimited(
630 "xennet: skb->len = %u, too big for wire format\n",
635 slots = xennet_count_skb_slots(skb);
636 if (unlikely(slots > MAX_XEN_SKB_FRAGS + 1)) {
637 net_dbg_ratelimited("xennet: skb rides the rocket: %d slots, %d bytes\n",
639 if (skb_linearize(skb))
643 page = virt_to_page(skb->data);
644 offset = offset_in_page(skb->data);
646 /* The first req should be at least ETH_HLEN size or the packet will be
647 * dropped by netback.
649 if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) {
650 nskb = skb_copy(skb, GFP_ATOMIC);
653 dev_consume_skb_any(skb);
655 page = virt_to_page(skb->data);
656 offset = offset_in_page(skb->data);
659 len = skb_headlen(skb);
661 spin_lock_irqsave(&queue->tx_lock, flags);
663 if (unlikely(!netif_carrier_ok(dev) ||
664 (slots > 1 && !xennet_can_sg(dev)) ||
665 netif_needs_gso(skb, netif_skb_features(skb)))) {
666 spin_unlock_irqrestore(&queue->tx_lock, flags);
670 /* First request for the linear area. */
674 first_tx = xennet_make_first_txreq(&info, offset, len);
675 offset += info.tx_local.size;
676 if (offset == PAGE_SIZE) {
680 len -= info.tx_local.size;
682 if (skb->ip_summed == CHECKSUM_PARTIAL)
684 first_tx->flags |= XEN_NETTXF_csum_blank |
685 XEN_NETTXF_data_validated;
686 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
687 /* remote but checksummed. */
688 first_tx->flags |= XEN_NETTXF_data_validated;
690 /* Optional extra info after the first request. */
691 if (skb_shinfo(skb)->gso_size) {
692 struct xen_netif_extra_info *gso;
694 gso = (struct xen_netif_extra_info *)
695 RING_GET_REQUEST(&queue->tx, queue->tx.req_prod_pvt++);
697 first_tx->flags |= XEN_NETTXF_extra_info;
699 gso->u.gso.size = skb_shinfo(skb)->gso_size;
700 gso->u.gso.type = (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) ?
701 XEN_NETIF_GSO_TYPE_TCPV6 :
702 XEN_NETIF_GSO_TYPE_TCPV4;
704 gso->u.gso.features = 0;
706 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
710 /* Requests for the rest of the linear area. */
711 xennet_make_txreqs(&info, page, offset, len);
713 /* Requests for all the frags. */
714 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
715 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
716 xennet_make_txreqs(&info, skb_frag_page(frag),
718 skb_frag_size(frag));
721 /* First request has the packet length. */
722 first_tx->size = skb->len;
724 xennet_mark_tx_pending(queue);
726 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->tx, notify);
728 notify_remote_via_irq(queue->tx_irq);
730 u64_stats_update_begin(&tx_stats->syncp);
731 tx_stats->bytes += skb->len;
733 u64_stats_update_end(&tx_stats->syncp);
735 /* Note: It is not safe to access skb after xennet_tx_buf_gc()! */
736 xennet_tx_buf_gc(queue);
738 if (!netfront_tx_slot_available(queue))
739 netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
741 spin_unlock_irqrestore(&queue->tx_lock, flags);
746 dev->stats.tx_dropped++;
747 dev_kfree_skb_any(skb);
751 static int xennet_close(struct net_device *dev)
753 struct netfront_info *np = netdev_priv(dev);
754 unsigned int num_queues = dev->real_num_tx_queues;
756 struct netfront_queue *queue;
757 netif_tx_stop_all_queues(np->netdev);
758 for (i = 0; i < num_queues; ++i) {
759 queue = &np->queues[i];
760 napi_disable(&queue->napi);
765 static void xennet_destroy_queues(struct netfront_info *info)
769 for (i = 0; i < info->netdev->real_num_tx_queues; i++) {
770 struct netfront_queue *queue = &info->queues[i];
772 if (netif_running(info->netdev))
773 napi_disable(&queue->napi);
774 netif_napi_del(&queue->napi);
781 static void xennet_uninit(struct net_device *dev)
783 struct netfront_info *np = netdev_priv(dev);
784 xennet_destroy_queues(np);
787 static void xennet_set_rx_rsp_cons(struct netfront_queue *queue, RING_IDX val)
791 spin_lock_irqsave(&queue->rx_cons_lock, flags);
792 queue->rx.rsp_cons = val;
793 queue->rx_rsp_unconsumed = RING_HAS_UNCONSUMED_RESPONSES(&queue->rx);
794 spin_unlock_irqrestore(&queue->rx_cons_lock, flags);
797 static void xennet_move_rx_slot(struct netfront_queue *queue, struct sk_buff *skb,
800 int new = xennet_rxidx(queue->rx.req_prod_pvt);
802 BUG_ON(queue->rx_skbs[new]);
803 queue->rx_skbs[new] = skb;
804 queue->grant_rx_ref[new] = ref;
805 RING_GET_REQUEST(&queue->rx, queue->rx.req_prod_pvt)->id = new;
806 RING_GET_REQUEST(&queue->rx, queue->rx.req_prod_pvt)->gref = ref;
807 queue->rx.req_prod_pvt++;
810 static int xennet_get_extras(struct netfront_queue *queue,
811 struct xen_netif_extra_info *extras,
815 struct xen_netif_extra_info extra;
816 struct device *dev = &queue->info->netdev->dev;
817 RING_IDX cons = queue->rx.rsp_cons;
824 if (unlikely(cons + 1 == rp)) {
826 dev_warn(dev, "Missing extra info\n");
831 RING_COPY_RESPONSE(&queue->rx, ++cons, &extra);
833 if (unlikely(!extra.type ||
834 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
836 dev_warn(dev, "Invalid extra type: %d\n",
840 extras[extra.type - 1] = extra;
843 skb = xennet_get_rx_skb(queue, cons);
844 ref = xennet_get_rx_ref(queue, cons);
845 xennet_move_rx_slot(queue, skb, ref);
846 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
848 xennet_set_rx_rsp_cons(queue, cons);
852 static int xennet_get_responses(struct netfront_queue *queue,
853 struct netfront_rx_info *rinfo, RING_IDX rp,
854 struct sk_buff_head *list)
856 struct xen_netif_rx_response *rx = &rinfo->rx, rx_local;
857 struct xen_netif_extra_info *extras = rinfo->extras;
858 struct device *dev = &queue->info->netdev->dev;
859 RING_IDX cons = queue->rx.rsp_cons;
860 struct sk_buff *skb = xennet_get_rx_skb(queue, cons);
861 grant_ref_t ref = xennet_get_rx_ref(queue, cons);
862 int max = XEN_NETIF_NR_SLOTS_MIN + (rx->status <= RX_COPY_THRESHOLD);
866 if (rx->flags & XEN_NETRXF_extra_info) {
867 err = xennet_get_extras(queue, extras, rp);
868 cons = queue->rx.rsp_cons;
872 if (unlikely(rx->status < 0 ||
873 rx->offset + rx->status > XEN_PAGE_SIZE)) {
875 dev_warn(dev, "rx->offset: %u, size: %d\n",
876 rx->offset, rx->status);
877 xennet_move_rx_slot(queue, skb, ref);
883 * This definitely indicates a bug, either in this driver or in
884 * the backend driver. In future this should flag the bad
885 * situation to the system controller to reboot the backend.
887 if (ref == GRANT_INVALID_REF) {
889 dev_warn(dev, "Bad rx response id %d.\n",
895 if (!gnttab_end_foreign_access_ref(ref, 0)) {
897 "Grant still in use by backend domain\n");
898 queue->info->broken = true;
899 dev_alert(dev, "Disabled for further use\n");
903 gnttab_release_grant_reference(&queue->gref_rx_head, ref);
905 __skb_queue_tail(list, skb);
908 if (!(rx->flags & XEN_NETRXF_more_data))
911 if (cons + slots == rp) {
913 dev_warn(dev, "Need more slots\n");
918 RING_COPY_RESPONSE(&queue->rx, cons + slots, &rx_local);
920 skb = xennet_get_rx_skb(queue, cons + slots);
921 ref = xennet_get_rx_ref(queue, cons + slots);
925 if (unlikely(slots > max)) {
927 dev_warn(dev, "Too many slots\n");
932 xennet_set_rx_rsp_cons(queue, cons + slots);
937 static int xennet_set_skb_gso(struct sk_buff *skb,
938 struct xen_netif_extra_info *gso)
940 if (!gso->u.gso.size) {
942 pr_warn("GSO size must not be zero\n");
946 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4 &&
947 gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV6) {
949 pr_warn("Bad GSO type %d\n", gso->u.gso.type);
953 skb_shinfo(skb)->gso_size = gso->u.gso.size;
954 skb_shinfo(skb)->gso_type =
955 (gso->u.gso.type == XEN_NETIF_GSO_TYPE_TCPV4) ?
959 /* Header must be checked, and gso_segs computed. */
960 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
961 skb_shinfo(skb)->gso_segs = 0;
966 static int xennet_fill_frags(struct netfront_queue *queue,
968 struct sk_buff_head *list)
970 RING_IDX cons = queue->rx.rsp_cons;
971 struct sk_buff *nskb;
973 while ((nskb = __skb_dequeue(list))) {
974 struct xen_netif_rx_response rx;
975 skb_frag_t *nfrag = &skb_shinfo(nskb)->frags[0];
977 RING_COPY_RESPONSE(&queue->rx, ++cons, &rx);
979 if (skb_shinfo(skb)->nr_frags == MAX_SKB_FRAGS) {
980 unsigned int pull_to = NETFRONT_SKB_CB(skb)->pull_to;
982 BUG_ON(pull_to < skb_headlen(skb));
983 __pskb_pull_tail(skb, pull_to - skb_headlen(skb));
985 if (unlikely(skb_shinfo(skb)->nr_frags >= MAX_SKB_FRAGS)) {
986 xennet_set_rx_rsp_cons(queue,
987 ++cons + skb_queue_len(list));
992 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
993 skb_frag_page(nfrag),
994 rx.offset, rx.status, PAGE_SIZE);
996 skb_shinfo(nskb)->nr_frags = 0;
1000 xennet_set_rx_rsp_cons(queue, cons);
1005 static int checksum_setup(struct net_device *dev, struct sk_buff *skb)
1007 bool recalculate_partial_csum = false;
1010 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1011 * peers can fail to set NETRXF_csum_blank when sending a GSO
1012 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1013 * recalculate the partial checksum.
1015 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1016 struct netfront_info *np = netdev_priv(dev);
1017 atomic_inc(&np->rx_gso_checksum_fixup);
1018 skb->ip_summed = CHECKSUM_PARTIAL;
1019 recalculate_partial_csum = true;
1022 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1023 if (skb->ip_summed != CHECKSUM_PARTIAL)
1026 return skb_checksum_setup(skb, recalculate_partial_csum);
1029 static int handle_incoming_queue(struct netfront_queue *queue,
1030 struct sk_buff_head *rxq)
1032 struct netfront_stats *rx_stats = this_cpu_ptr(queue->info->rx_stats);
1033 int packets_dropped = 0;
1034 struct sk_buff *skb;
1036 while ((skb = __skb_dequeue(rxq)) != NULL) {
1037 int pull_to = NETFRONT_SKB_CB(skb)->pull_to;
1039 if (pull_to > skb_headlen(skb))
1040 __pskb_pull_tail(skb, pull_to - skb_headlen(skb));
1042 /* Ethernet work: Delayed to here as it peeks the header. */
1043 skb->protocol = eth_type_trans(skb, queue->info->netdev);
1044 skb_reset_network_header(skb);
1046 if (checksum_setup(queue->info->netdev, skb)) {
1049 queue->info->netdev->stats.rx_errors++;
1053 u64_stats_update_begin(&rx_stats->syncp);
1054 rx_stats->packets++;
1055 rx_stats->bytes += skb->len;
1056 u64_stats_update_end(&rx_stats->syncp);
1059 napi_gro_receive(&queue->napi, skb);
1062 return packets_dropped;
1065 static int xennet_poll(struct napi_struct *napi, int budget)
1067 struct netfront_queue *queue = container_of(napi, struct netfront_queue, napi);
1068 struct net_device *dev = queue->info->netdev;
1069 struct sk_buff *skb;
1070 struct netfront_rx_info rinfo;
1071 struct xen_netif_rx_response *rx = &rinfo.rx;
1072 struct xen_netif_extra_info *extras = rinfo.extras;
1075 struct sk_buff_head rxq;
1076 struct sk_buff_head errq;
1077 struct sk_buff_head tmpq;
1080 spin_lock(&queue->rx_lock);
1082 skb_queue_head_init(&rxq);
1083 skb_queue_head_init(&errq);
1084 skb_queue_head_init(&tmpq);
1086 rp = queue->rx.sring->rsp_prod;
1087 if (RING_RESPONSE_PROD_OVERFLOW(&queue->rx, rp)) {
1088 dev_alert(&dev->dev, "Illegal number of responses %u\n",
1089 rp - queue->rx.rsp_cons);
1090 queue->info->broken = true;
1091 spin_unlock(&queue->rx_lock);
1094 rmb(); /* Ensure we see queued responses up to 'rp'. */
1096 i = queue->rx.rsp_cons;
1098 while ((i != rp) && (work_done < budget)) {
1099 RING_COPY_RESPONSE(&queue->rx, i, rx);
1100 memset(extras, 0, sizeof(rinfo.extras));
1102 err = xennet_get_responses(queue, &rinfo, rp, &tmpq);
1104 if (unlikely(err)) {
1105 if (queue->info->broken) {
1106 spin_unlock(&queue->rx_lock);
1110 while ((skb = __skb_dequeue(&tmpq)))
1111 __skb_queue_tail(&errq, skb);
1112 dev->stats.rx_errors++;
1113 i = queue->rx.rsp_cons;
1117 skb = __skb_dequeue(&tmpq);
1119 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1120 struct xen_netif_extra_info *gso;
1121 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1123 if (unlikely(xennet_set_skb_gso(skb, gso))) {
1124 __skb_queue_head(&tmpq, skb);
1125 xennet_set_rx_rsp_cons(queue,
1126 queue->rx.rsp_cons +
1127 skb_queue_len(&tmpq));
1132 NETFRONT_SKB_CB(skb)->pull_to = rx->status;
1133 if (NETFRONT_SKB_CB(skb)->pull_to > RX_COPY_THRESHOLD)
1134 NETFRONT_SKB_CB(skb)->pull_to = RX_COPY_THRESHOLD;
1136 skb_shinfo(skb)->frags[0].page_offset = rx->offset;
1137 skb_frag_size_set(&skb_shinfo(skb)->frags[0], rx->status);
1138 skb->data_len = rx->status;
1139 skb->len += rx->status;
1141 if (unlikely(xennet_fill_frags(queue, skb, &tmpq)))
1144 if (rx->flags & XEN_NETRXF_csum_blank)
1145 skb->ip_summed = CHECKSUM_PARTIAL;
1146 else if (rx->flags & XEN_NETRXF_data_validated)
1147 skb->ip_summed = CHECKSUM_UNNECESSARY;
1149 __skb_queue_tail(&rxq, skb);
1151 i = queue->rx.rsp_cons + 1;
1152 xennet_set_rx_rsp_cons(queue, i);
1156 __skb_queue_purge(&errq);
1158 work_done -= handle_incoming_queue(queue, &rxq);
1160 xennet_alloc_rx_buffers(queue);
1162 if (work_done < budget) {
1165 napi_complete_done(napi, work_done);
1167 RING_FINAL_CHECK_FOR_RESPONSES(&queue->rx, more_to_do);
1169 napi_schedule(napi);
1172 spin_unlock(&queue->rx_lock);
1177 static int xennet_change_mtu(struct net_device *dev, int mtu)
1179 int max = xennet_can_sg(dev) ? XEN_NETIF_MAX_TX_SIZE : ETH_DATA_LEN;
1187 static void xennet_get_stats64(struct net_device *dev,
1188 struct rtnl_link_stats64 *tot)
1190 struct netfront_info *np = netdev_priv(dev);
1193 for_each_possible_cpu(cpu) {
1194 struct netfront_stats *rx_stats = per_cpu_ptr(np->rx_stats, cpu);
1195 struct netfront_stats *tx_stats = per_cpu_ptr(np->tx_stats, cpu);
1196 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
1200 start = u64_stats_fetch_begin_irq(&tx_stats->syncp);
1201 tx_packets = tx_stats->packets;
1202 tx_bytes = tx_stats->bytes;
1203 } while (u64_stats_fetch_retry_irq(&tx_stats->syncp, start));
1206 start = u64_stats_fetch_begin_irq(&rx_stats->syncp);
1207 rx_packets = rx_stats->packets;
1208 rx_bytes = rx_stats->bytes;
1209 } while (u64_stats_fetch_retry_irq(&rx_stats->syncp, start));
1211 tot->rx_packets += rx_packets;
1212 tot->tx_packets += tx_packets;
1213 tot->rx_bytes += rx_bytes;
1214 tot->tx_bytes += tx_bytes;
1217 tot->rx_errors = dev->stats.rx_errors;
1218 tot->tx_dropped = dev->stats.tx_dropped;
1221 static void xennet_release_tx_bufs(struct netfront_queue *queue)
1223 struct sk_buff *skb;
1226 for (i = 0; i < NET_TX_RING_SIZE; i++) {
1227 /* Skip over entries which are actually freelist references */
1228 if (!queue->tx_skbs[i])
1231 skb = queue->tx_skbs[i];
1232 queue->tx_skbs[i] = NULL;
1233 get_page(queue->grant_tx_page[i]);
1234 gnttab_end_foreign_access(queue->grant_tx_ref[i],
1236 (unsigned long)page_address(queue->grant_tx_page[i]));
1237 queue->grant_tx_page[i] = NULL;
1238 queue->grant_tx_ref[i] = GRANT_INVALID_REF;
1239 add_id_to_list(&queue->tx_skb_freelist, queue->tx_link, i);
1240 dev_kfree_skb_irq(skb);
1244 static void xennet_release_rx_bufs(struct netfront_queue *queue)
1248 spin_lock_bh(&queue->rx_lock);
1250 for (id = 0; id < NET_RX_RING_SIZE; id++) {
1251 struct sk_buff *skb;
1254 skb = queue->rx_skbs[id];
1258 ref = queue->grant_rx_ref[id];
1259 if (ref == GRANT_INVALID_REF)
1262 page = skb_frag_page(&skb_shinfo(skb)->frags[0]);
1264 /* gnttab_end_foreign_access() needs a page ref until
1265 * foreign access is ended (which may be deferred).
1268 gnttab_end_foreign_access(ref, 0,
1269 (unsigned long)page_address(page));
1270 queue->grant_rx_ref[id] = GRANT_INVALID_REF;
1275 spin_unlock_bh(&queue->rx_lock);
1278 static netdev_features_t xennet_fix_features(struct net_device *dev,
1279 netdev_features_t features)
1281 struct netfront_info *np = netdev_priv(dev);
1283 if (features & NETIF_F_SG &&
1284 !xenbus_read_unsigned(np->xbdev->otherend, "feature-sg", 0))
1285 features &= ~NETIF_F_SG;
1287 if (features & NETIF_F_IPV6_CSUM &&
1288 !xenbus_read_unsigned(np->xbdev->otherend,
1289 "feature-ipv6-csum-offload", 0))
1290 features &= ~NETIF_F_IPV6_CSUM;
1292 if (features & NETIF_F_TSO &&
1293 !xenbus_read_unsigned(np->xbdev->otherend, "feature-gso-tcpv4", 0))
1294 features &= ~NETIF_F_TSO;
1296 if (features & NETIF_F_TSO6 &&
1297 !xenbus_read_unsigned(np->xbdev->otherend, "feature-gso-tcpv6", 0))
1298 features &= ~NETIF_F_TSO6;
1303 static int xennet_set_features(struct net_device *dev,
1304 netdev_features_t features)
1306 if (!(features & NETIF_F_SG) && dev->mtu > ETH_DATA_LEN) {
1307 netdev_info(dev, "Reducing MTU because no SG offload");
1308 dev->mtu = ETH_DATA_LEN;
1314 static bool xennet_handle_tx(struct netfront_queue *queue, unsigned int *eoi)
1316 unsigned long flags;
1318 if (unlikely(queue->info->broken))
1321 spin_lock_irqsave(&queue->tx_lock, flags);
1322 if (xennet_tx_buf_gc(queue))
1324 spin_unlock_irqrestore(&queue->tx_lock, flags);
1329 static irqreturn_t xennet_tx_interrupt(int irq, void *dev_id)
1331 unsigned int eoiflag = XEN_EOI_FLAG_SPURIOUS;
1333 if (likely(xennet_handle_tx(dev_id, &eoiflag)))
1334 xen_irq_lateeoi(irq, eoiflag);
1339 static bool xennet_handle_rx(struct netfront_queue *queue, unsigned int *eoi)
1341 unsigned int work_queued;
1342 unsigned long flags;
1344 if (unlikely(queue->info->broken))
1347 spin_lock_irqsave(&queue->rx_cons_lock, flags);
1348 work_queued = RING_HAS_UNCONSUMED_RESPONSES(&queue->rx);
1349 if (work_queued > queue->rx_rsp_unconsumed) {
1350 queue->rx_rsp_unconsumed = work_queued;
1352 } else if (unlikely(work_queued < queue->rx_rsp_unconsumed)) {
1353 const struct device *dev = &queue->info->netdev->dev;
1355 spin_unlock_irqrestore(&queue->rx_cons_lock, flags);
1356 dev_alert(dev, "RX producer index going backwards\n");
1357 dev_alert(dev, "Disabled for further use\n");
1358 queue->info->broken = true;
1361 spin_unlock_irqrestore(&queue->rx_cons_lock, flags);
1363 if (likely(netif_carrier_ok(queue->info->netdev) && work_queued))
1364 napi_schedule(&queue->napi);
1369 static irqreturn_t xennet_rx_interrupt(int irq, void *dev_id)
1371 unsigned int eoiflag = XEN_EOI_FLAG_SPURIOUS;
1373 if (likely(xennet_handle_rx(dev_id, &eoiflag)))
1374 xen_irq_lateeoi(irq, eoiflag);
1379 static irqreturn_t xennet_interrupt(int irq, void *dev_id)
1381 unsigned int eoiflag = XEN_EOI_FLAG_SPURIOUS;
1383 if (xennet_handle_tx(dev_id, &eoiflag) &&
1384 xennet_handle_rx(dev_id, &eoiflag))
1385 xen_irq_lateeoi(irq, eoiflag);
1390 #ifdef CONFIG_NET_POLL_CONTROLLER
1391 static void xennet_poll_controller(struct net_device *dev)
1393 /* Poll each queue */
1394 struct netfront_info *info = netdev_priv(dev);
1395 unsigned int num_queues = dev->real_num_tx_queues;
1401 for (i = 0; i < num_queues; ++i)
1402 xennet_interrupt(0, &info->queues[i]);
1406 static const struct net_device_ops xennet_netdev_ops = {
1407 .ndo_uninit = xennet_uninit,
1408 .ndo_open = xennet_open,
1409 .ndo_stop = xennet_close,
1410 .ndo_start_xmit = xennet_start_xmit,
1411 .ndo_change_mtu = xennet_change_mtu,
1412 .ndo_get_stats64 = xennet_get_stats64,
1413 .ndo_set_mac_address = eth_mac_addr,
1414 .ndo_validate_addr = eth_validate_addr,
1415 .ndo_fix_features = xennet_fix_features,
1416 .ndo_set_features = xennet_set_features,
1417 .ndo_select_queue = xennet_select_queue,
1418 #ifdef CONFIG_NET_POLL_CONTROLLER
1419 .ndo_poll_controller = xennet_poll_controller,
1423 static void xennet_free_netdev(struct net_device *netdev)
1425 struct netfront_info *np = netdev_priv(netdev);
1427 free_percpu(np->rx_stats);
1428 free_percpu(np->tx_stats);
1429 free_netdev(netdev);
1432 static struct net_device *xennet_create_dev(struct xenbus_device *dev)
1435 struct net_device *netdev;
1436 struct netfront_info *np;
1438 netdev = alloc_etherdev_mq(sizeof(struct netfront_info), xennet_max_queues);
1440 return ERR_PTR(-ENOMEM);
1442 np = netdev_priv(netdev);
1448 np->rx_stats = netdev_alloc_pcpu_stats(struct netfront_stats);
1449 if (np->rx_stats == NULL)
1451 np->tx_stats = netdev_alloc_pcpu_stats(struct netfront_stats);
1452 if (np->tx_stats == NULL)
1455 netdev->netdev_ops = &xennet_netdev_ops;
1457 netdev->features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM |
1459 netdev->hw_features = NETIF_F_SG |
1461 NETIF_F_TSO | NETIF_F_TSO6;
1464 * Assume that all hw features are available for now. This set
1465 * will be adjusted by the call to netdev_update_features() in
1466 * xennet_connect() which is the earliest point where we can
1467 * negotiate with the backend regarding supported features.
1469 netdev->features |= netdev->hw_features;
1471 netdev->ethtool_ops = &xennet_ethtool_ops;
1472 netdev->min_mtu = ETH_MIN_MTU;
1473 netdev->max_mtu = XEN_NETIF_MAX_TX_SIZE;
1474 SET_NETDEV_DEV(netdev, &dev->dev);
1476 np->netdev = netdev;
1478 netif_carrier_off(netdev);
1481 xenbus_switch_state(dev, XenbusStateInitialising);
1482 err = wait_event_timeout(module_wq,
1483 xenbus_read_driver_state(dev->otherend) !=
1484 XenbusStateClosed &&
1485 xenbus_read_driver_state(dev->otherend) !=
1486 XenbusStateUnknown, XENNET_TIMEOUT);
1492 xennet_free_netdev(netdev);
1493 return ERR_PTR(err);
1497 * Entry point to this code when a new device is created. Allocate the basic
1498 * structures and the ring buffers for communication with the backend, and
1499 * inform the backend of the appropriate details for those.
1501 static int netfront_probe(struct xenbus_device *dev,
1502 const struct xenbus_device_id *id)
1505 struct net_device *netdev;
1506 struct netfront_info *info;
1508 netdev = xennet_create_dev(dev);
1509 if (IS_ERR(netdev)) {
1510 err = PTR_ERR(netdev);
1511 xenbus_dev_fatal(dev, err, "creating netdev");
1515 info = netdev_priv(netdev);
1516 dev_set_drvdata(&dev->dev, info);
1518 info->netdev->sysfs_groups[0] = &xennet_dev_group;
1524 static void xennet_end_access(int ref, void *page)
1526 /* This frees the page as a side-effect */
1527 if (ref != GRANT_INVALID_REF)
1528 gnttab_end_foreign_access(ref, 0, (unsigned long)page);
1531 static void xennet_disconnect_backend(struct netfront_info *info)
1534 unsigned int num_queues = info->netdev->real_num_tx_queues;
1536 netif_carrier_off(info->netdev);
1538 for (i = 0; i < num_queues && info->queues; ++i) {
1539 struct netfront_queue *queue = &info->queues[i];
1541 del_timer_sync(&queue->rx_refill_timer);
1543 if (queue->tx_irq && (queue->tx_irq == queue->rx_irq))
1544 unbind_from_irqhandler(queue->tx_irq, queue);
1545 if (queue->tx_irq && (queue->tx_irq != queue->rx_irq)) {
1546 unbind_from_irqhandler(queue->tx_irq, queue);
1547 unbind_from_irqhandler(queue->rx_irq, queue);
1549 queue->tx_evtchn = queue->rx_evtchn = 0;
1550 queue->tx_irq = queue->rx_irq = 0;
1552 if (netif_running(info->netdev))
1553 napi_synchronize(&queue->napi);
1555 xennet_release_tx_bufs(queue);
1556 xennet_release_rx_bufs(queue);
1557 gnttab_free_grant_references(queue->gref_tx_head);
1558 gnttab_free_grant_references(queue->gref_rx_head);
1560 /* End access and free the pages */
1561 xennet_end_access(queue->tx_ring_ref, queue->tx.sring);
1562 xennet_end_access(queue->rx_ring_ref, queue->rx.sring);
1564 queue->tx_ring_ref = GRANT_INVALID_REF;
1565 queue->rx_ring_ref = GRANT_INVALID_REF;
1566 queue->tx.sring = NULL;
1567 queue->rx.sring = NULL;
1572 * We are reconnecting to the backend, due to a suspend/resume, or a backend
1573 * driver restart. We tear down our netif structure and recreate it, but
1574 * leave the device-layer structures intact so that this is transparent to the
1575 * rest of the kernel.
1577 static int netfront_resume(struct xenbus_device *dev)
1579 struct netfront_info *info = dev_get_drvdata(&dev->dev);
1581 dev_dbg(&dev->dev, "%s\n", dev->nodename);
1583 netif_tx_lock_bh(info->netdev);
1584 netif_device_detach(info->netdev);
1585 netif_tx_unlock_bh(info->netdev);
1587 xennet_disconnect_backend(info);
1591 static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[])
1593 char *s, *e, *macstr;
1596 macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL);
1598 return PTR_ERR(macstr);
1600 for (i = 0; i < ETH_ALEN; i++) {
1601 mac[i] = simple_strtoul(s, &e, 16);
1602 if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) {
1613 static int setup_netfront_single(struct netfront_queue *queue)
1617 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->tx_evtchn);
1621 err = bind_evtchn_to_irqhandler_lateeoi(queue->tx_evtchn,
1622 xennet_interrupt, 0,
1623 queue->info->netdev->name,
1627 queue->rx_evtchn = queue->tx_evtchn;
1628 queue->rx_irq = queue->tx_irq = err;
1633 xenbus_free_evtchn(queue->info->xbdev, queue->tx_evtchn);
1634 queue->tx_evtchn = 0;
1639 static int setup_netfront_split(struct netfront_queue *queue)
1643 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->tx_evtchn);
1646 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->rx_evtchn);
1648 goto alloc_rx_evtchn_fail;
1650 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name),
1651 "%s-tx", queue->name);
1652 err = bind_evtchn_to_irqhandler_lateeoi(queue->tx_evtchn,
1653 xennet_tx_interrupt, 0,
1654 queue->tx_irq_name, queue);
1657 queue->tx_irq = err;
1659 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name),
1660 "%s-rx", queue->name);
1661 err = bind_evtchn_to_irqhandler_lateeoi(queue->rx_evtchn,
1662 xennet_rx_interrupt, 0,
1663 queue->rx_irq_name, queue);
1666 queue->rx_irq = err;
1671 unbind_from_irqhandler(queue->tx_irq, queue);
1674 xenbus_free_evtchn(queue->info->xbdev, queue->rx_evtchn);
1675 queue->rx_evtchn = 0;
1676 alloc_rx_evtchn_fail:
1677 xenbus_free_evtchn(queue->info->xbdev, queue->tx_evtchn);
1678 queue->tx_evtchn = 0;
1683 static int setup_netfront(struct xenbus_device *dev,
1684 struct netfront_queue *queue, unsigned int feature_split_evtchn)
1686 struct xen_netif_tx_sring *txs;
1687 struct xen_netif_rx_sring *rxs = NULL;
1691 queue->tx_ring_ref = GRANT_INVALID_REF;
1692 queue->rx_ring_ref = GRANT_INVALID_REF;
1693 queue->rx.sring = NULL;
1694 queue->tx.sring = NULL;
1696 txs = (struct xen_netif_tx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH);
1699 xenbus_dev_fatal(dev, err, "allocating tx ring page");
1702 SHARED_RING_INIT(txs);
1703 FRONT_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1705 err = xenbus_grant_ring(dev, txs, 1, &gref);
1708 queue->tx_ring_ref = gref;
1710 rxs = (struct xen_netif_rx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH);
1713 xenbus_dev_fatal(dev, err, "allocating rx ring page");
1716 SHARED_RING_INIT(rxs);
1717 FRONT_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1719 err = xenbus_grant_ring(dev, rxs, 1, &gref);
1722 queue->rx_ring_ref = gref;
1724 if (feature_split_evtchn)
1725 err = setup_netfront_split(queue);
1726 /* setup single event channel if
1727 * a) feature-split-event-channels == 0
1728 * b) feature-split-event-channels == 1 but failed to setup
1730 if (!feature_split_evtchn || (feature_split_evtchn && err))
1731 err = setup_netfront_single(queue);
1738 /* If we fail to setup netfront, it is safe to just revoke access to
1739 * granted pages because backend is not accessing it at this point.
1742 if (queue->rx_ring_ref != GRANT_INVALID_REF) {
1743 gnttab_end_foreign_access(queue->rx_ring_ref, 0,
1744 (unsigned long)rxs);
1745 queue->rx_ring_ref = GRANT_INVALID_REF;
1747 free_page((unsigned long)rxs);
1749 if (queue->tx_ring_ref != GRANT_INVALID_REF) {
1750 gnttab_end_foreign_access(queue->tx_ring_ref, 0,
1751 (unsigned long)txs);
1752 queue->tx_ring_ref = GRANT_INVALID_REF;
1754 free_page((unsigned long)txs);
1759 /* Queue-specific initialisation
1760 * This used to be done in xennet_create_dev() but must now
1763 static int xennet_init_queue(struct netfront_queue *queue)
1769 spin_lock_init(&queue->tx_lock);
1770 spin_lock_init(&queue->rx_lock);
1771 spin_lock_init(&queue->rx_cons_lock);
1773 timer_setup(&queue->rx_refill_timer, rx_refill_timeout, 0);
1775 devid = strrchr(queue->info->xbdev->nodename, '/') + 1;
1776 snprintf(queue->name, sizeof(queue->name), "vif%s-q%u",
1779 /* Initialise tx_skb_freelist as a free chain containing every entry. */
1780 queue->tx_skb_freelist = 0;
1781 queue->tx_pend_queue = TX_LINK_NONE;
1782 for (i = 0; i < NET_TX_RING_SIZE; i++) {
1783 queue->tx_link[i] = i + 1;
1784 queue->grant_tx_ref[i] = GRANT_INVALID_REF;
1785 queue->grant_tx_page[i] = NULL;
1787 queue->tx_link[NET_TX_RING_SIZE - 1] = TX_LINK_NONE;
1789 /* Clear out rx_skbs */
1790 for (i = 0; i < NET_RX_RING_SIZE; i++) {
1791 queue->rx_skbs[i] = NULL;
1792 queue->grant_rx_ref[i] = GRANT_INVALID_REF;
1795 /* A grant for every tx ring slot */
1796 if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
1797 &queue->gref_tx_head) < 0) {
1798 pr_alert("can't alloc tx grant refs\n");
1803 /* A grant for every rx ring slot */
1804 if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
1805 &queue->gref_rx_head) < 0) {
1806 pr_alert("can't alloc rx grant refs\n");
1814 gnttab_free_grant_references(queue->gref_tx_head);
1819 static int write_queue_xenstore_keys(struct netfront_queue *queue,
1820 struct xenbus_transaction *xbt, int write_hierarchical)
1822 /* Write the queue-specific keys into XenStore in the traditional
1823 * way for a single queue, or in a queue subkeys for multiple
1826 struct xenbus_device *dev = queue->info->xbdev;
1828 const char *message;
1832 /* Choose the correct place to write the keys */
1833 if (write_hierarchical) {
1834 pathsize = strlen(dev->nodename) + 10;
1835 path = kzalloc(pathsize, GFP_KERNEL);
1838 message = "out of memory while writing ring references";
1841 snprintf(path, pathsize, "%s/queue-%u",
1842 dev->nodename, queue->id);
1844 path = (char *)dev->nodename;
1847 /* Write ring references */
1848 err = xenbus_printf(*xbt, path, "tx-ring-ref", "%u",
1849 queue->tx_ring_ref);
1851 message = "writing tx-ring-ref";
1855 err = xenbus_printf(*xbt, path, "rx-ring-ref", "%u",
1856 queue->rx_ring_ref);
1858 message = "writing rx-ring-ref";
1862 /* Write event channels; taking into account both shared
1863 * and split event channel scenarios.
1865 if (queue->tx_evtchn == queue->rx_evtchn) {
1866 /* Shared event channel */
1867 err = xenbus_printf(*xbt, path,
1868 "event-channel", "%u", queue->tx_evtchn);
1870 message = "writing event-channel";
1874 /* Split event channels */
1875 err = xenbus_printf(*xbt, path,
1876 "event-channel-tx", "%u", queue->tx_evtchn);
1878 message = "writing event-channel-tx";
1882 err = xenbus_printf(*xbt, path,
1883 "event-channel-rx", "%u", queue->rx_evtchn);
1885 message = "writing event-channel-rx";
1890 if (write_hierarchical)
1895 if (write_hierarchical)
1897 xenbus_dev_fatal(dev, err, "%s", message);
1901 static int xennet_create_queues(struct netfront_info *info,
1902 unsigned int *num_queues)
1907 info->queues = kcalloc(*num_queues, sizeof(struct netfront_queue),
1912 for (i = 0; i < *num_queues; i++) {
1913 struct netfront_queue *queue = &info->queues[i];
1918 ret = xennet_init_queue(queue);
1920 dev_warn(&info->xbdev->dev,
1921 "only created %d queues\n", i);
1926 netif_napi_add(queue->info->netdev, &queue->napi,
1928 if (netif_running(info->netdev))
1929 napi_enable(&queue->napi);
1932 netif_set_real_num_tx_queues(info->netdev, *num_queues);
1934 if (*num_queues == 0) {
1935 dev_err(&info->xbdev->dev, "no queues\n");
1941 /* Common code used when first setting up, and when resuming. */
1942 static int talk_to_netback(struct xenbus_device *dev,
1943 struct netfront_info *info)
1945 const char *message;
1946 struct xenbus_transaction xbt;
1948 unsigned int feature_split_evtchn;
1950 unsigned int max_queues = 0;
1951 struct netfront_queue *queue = NULL;
1952 unsigned int num_queues = 1;
1954 info->netdev->irq = 0;
1956 /* Check if backend supports multiple queues */
1957 max_queues = xenbus_read_unsigned(info->xbdev->otherend,
1958 "multi-queue-max-queues", 1);
1959 num_queues = min(max_queues, xennet_max_queues);
1961 /* Check feature-split-event-channels */
1962 feature_split_evtchn = xenbus_read_unsigned(info->xbdev->otherend,
1963 "feature-split-event-channels", 0);
1965 /* Read mac addr. */
1966 err = xen_net_read_mac(dev, info->netdev->dev_addr);
1968 xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename);
1974 xennet_destroy_queues(info);
1976 /* For the case of a reconnect reset the "broken" indicator. */
1977 info->broken = false;
1979 err = xennet_create_queues(info, &num_queues);
1981 xenbus_dev_fatal(dev, err, "creating queues");
1982 kfree(info->queues);
1983 info->queues = NULL;
1988 /* Create shared ring, alloc event channel -- for each queue */
1989 for (i = 0; i < num_queues; ++i) {
1990 queue = &info->queues[i];
1991 err = setup_netfront(dev, queue, feature_split_evtchn);
1997 err = xenbus_transaction_start(&xbt);
1999 xenbus_dev_fatal(dev, err, "starting transaction");
2003 if (xenbus_exists(XBT_NIL,
2004 info->xbdev->otherend, "multi-queue-max-queues")) {
2005 /* Write the number of queues */
2006 err = xenbus_printf(xbt, dev->nodename,
2007 "multi-queue-num-queues", "%u", num_queues);
2009 message = "writing multi-queue-num-queues";
2010 goto abort_transaction_no_dev_fatal;
2014 if (num_queues == 1) {
2015 err = write_queue_xenstore_keys(&info->queues[0], &xbt, 0); /* flat */
2017 goto abort_transaction_no_dev_fatal;
2019 /* Write the keys for each queue */
2020 for (i = 0; i < num_queues; ++i) {
2021 queue = &info->queues[i];
2022 err = write_queue_xenstore_keys(queue, &xbt, 1); /* hierarchical */
2024 goto abort_transaction_no_dev_fatal;
2028 /* The remaining keys are not queue-specific */
2029 err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u",
2032 message = "writing request-rx-copy";
2033 goto abort_transaction;
2036 err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1);
2038 message = "writing feature-rx-notify";
2039 goto abort_transaction;
2042 err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1);
2044 message = "writing feature-sg";
2045 goto abort_transaction;
2048 err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1);
2050 message = "writing feature-gso-tcpv4";
2051 goto abort_transaction;
2054 err = xenbus_write(xbt, dev->nodename, "feature-gso-tcpv6", "1");
2056 message = "writing feature-gso-tcpv6";
2057 goto abort_transaction;
2060 err = xenbus_write(xbt, dev->nodename, "feature-ipv6-csum-offload",
2063 message = "writing feature-ipv6-csum-offload";
2064 goto abort_transaction;
2067 err = xenbus_transaction_end(xbt, 0);
2071 xenbus_dev_fatal(dev, err, "completing transaction");
2078 xenbus_dev_fatal(dev, err, "%s", message);
2079 abort_transaction_no_dev_fatal:
2080 xenbus_transaction_end(xbt, 1);
2082 xennet_disconnect_backend(info);
2084 xennet_destroy_queues(info);
2088 device_unregister(&dev->dev);
2092 static int xennet_connect(struct net_device *dev)
2094 struct netfront_info *np = netdev_priv(dev);
2095 unsigned int num_queues = 0;
2098 struct netfront_queue *queue = NULL;
2100 if (!xenbus_read_unsigned(np->xbdev->otherend, "feature-rx-copy", 0)) {
2102 "backend does not support copying receive path\n");
2106 err = talk_to_netback(np->xbdev, np);
2110 /* talk_to_netback() sets the correct number of queues */
2111 num_queues = dev->real_num_tx_queues;
2113 if (dev->reg_state == NETREG_UNINITIALIZED) {
2114 err = register_netdev(dev);
2116 pr_warn("%s: register_netdev err=%d\n", __func__, err);
2117 device_unregister(&np->xbdev->dev);
2123 netdev_update_features(dev);
2127 * All public and private state should now be sane. Get
2128 * ready to start sending and receiving packets and give the driver
2129 * domain a kick because we've probably just requeued some
2132 netif_tx_lock_bh(np->netdev);
2133 netif_device_attach(np->netdev);
2134 netif_tx_unlock_bh(np->netdev);
2136 netif_carrier_on(np->netdev);
2137 for (j = 0; j < num_queues; ++j) {
2138 queue = &np->queues[j];
2140 notify_remote_via_irq(queue->tx_irq);
2141 if (queue->tx_irq != queue->rx_irq)
2142 notify_remote_via_irq(queue->rx_irq);
2144 spin_lock_irq(&queue->tx_lock);
2145 xennet_tx_buf_gc(queue);
2146 spin_unlock_irq(&queue->tx_lock);
2148 spin_lock_bh(&queue->rx_lock);
2149 xennet_alloc_rx_buffers(queue);
2150 spin_unlock_bh(&queue->rx_lock);
2157 * Callback received when the backend's state changes.
2159 static void netback_changed(struct xenbus_device *dev,
2160 enum xenbus_state backend_state)
2162 struct netfront_info *np = dev_get_drvdata(&dev->dev);
2163 struct net_device *netdev = np->netdev;
2165 dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state));
2167 wake_up_all(&module_wq);
2169 switch (backend_state) {
2170 case XenbusStateInitialising:
2171 case XenbusStateInitialised:
2172 case XenbusStateReconfiguring:
2173 case XenbusStateReconfigured:
2174 case XenbusStateUnknown:
2177 case XenbusStateInitWait:
2178 if (dev->state != XenbusStateInitialising)
2180 if (xennet_connect(netdev) != 0)
2182 xenbus_switch_state(dev, XenbusStateConnected);
2185 case XenbusStateConnected:
2186 netdev_notify_peers(netdev);
2189 case XenbusStateClosed:
2190 if (dev->state == XenbusStateClosed)
2192 /* Missed the backend's CLOSING state -- fallthrough */
2193 case XenbusStateClosing:
2194 xenbus_frontend_closed(dev);
2199 static const struct xennet_stat {
2200 char name[ETH_GSTRING_LEN];
2202 } xennet_stats[] = {
2204 "rx_gso_checksum_fixup",
2205 offsetof(struct netfront_info, rx_gso_checksum_fixup)
2209 static int xennet_get_sset_count(struct net_device *dev, int string_set)
2211 switch (string_set) {
2213 return ARRAY_SIZE(xennet_stats);
2219 static void xennet_get_ethtool_stats(struct net_device *dev,
2220 struct ethtool_stats *stats, u64 * data)
2222 void *np = netdev_priv(dev);
2225 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
2226 data[i] = atomic_read((atomic_t *)(np + xennet_stats[i].offset));
2229 static void xennet_get_strings(struct net_device *dev, u32 stringset, u8 * data)
2233 switch (stringset) {
2235 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
2236 memcpy(data + i * ETH_GSTRING_LEN,
2237 xennet_stats[i].name, ETH_GSTRING_LEN);
2242 static const struct ethtool_ops xennet_ethtool_ops =
2244 .get_link = ethtool_op_get_link,
2246 .get_sset_count = xennet_get_sset_count,
2247 .get_ethtool_stats = xennet_get_ethtool_stats,
2248 .get_strings = xennet_get_strings,
2252 static ssize_t show_rxbuf(struct device *dev,
2253 struct device_attribute *attr, char *buf)
2255 return sprintf(buf, "%lu\n", NET_RX_RING_SIZE);
2258 static ssize_t store_rxbuf(struct device *dev,
2259 struct device_attribute *attr,
2260 const char *buf, size_t len)
2263 unsigned long target;
2265 if (!capable(CAP_NET_ADMIN))
2268 target = simple_strtoul(buf, &endp, 0);
2272 /* rxbuf_min and rxbuf_max are no longer configurable. */
2277 static DEVICE_ATTR(rxbuf_min, 0644, show_rxbuf, store_rxbuf);
2278 static DEVICE_ATTR(rxbuf_max, 0644, show_rxbuf, store_rxbuf);
2279 static DEVICE_ATTR(rxbuf_cur, 0444, show_rxbuf, NULL);
2281 static struct attribute *xennet_dev_attrs[] = {
2282 &dev_attr_rxbuf_min.attr,
2283 &dev_attr_rxbuf_max.attr,
2284 &dev_attr_rxbuf_cur.attr,
2288 static const struct attribute_group xennet_dev_group = {
2289 .attrs = xennet_dev_attrs
2291 #endif /* CONFIG_SYSFS */
2293 static void xennet_bus_close(struct xenbus_device *dev)
2297 if (xenbus_read_driver_state(dev->otherend) == XenbusStateClosed)
2300 xenbus_switch_state(dev, XenbusStateClosing);
2301 ret = wait_event_timeout(module_wq,
2302 xenbus_read_driver_state(dev->otherend) ==
2303 XenbusStateClosing ||
2304 xenbus_read_driver_state(dev->otherend) ==
2305 XenbusStateClosed ||
2306 xenbus_read_driver_state(dev->otherend) ==
2311 if (xenbus_read_driver_state(dev->otherend) == XenbusStateClosed)
2315 xenbus_switch_state(dev, XenbusStateClosed);
2316 ret = wait_event_timeout(module_wq,
2317 xenbus_read_driver_state(dev->otherend) ==
2318 XenbusStateClosed ||
2319 xenbus_read_driver_state(dev->otherend) ==
2325 static int xennet_remove(struct xenbus_device *dev)
2327 struct netfront_info *info = dev_get_drvdata(&dev->dev);
2329 xennet_bus_close(dev);
2330 xennet_disconnect_backend(info);
2332 if (info->netdev->reg_state == NETREG_REGISTERED)
2333 unregister_netdev(info->netdev);
2337 xennet_destroy_queues(info);
2340 xennet_free_netdev(info->netdev);
2345 static const struct xenbus_device_id netfront_ids[] = {
2350 static struct xenbus_driver netfront_driver = {
2351 .ids = netfront_ids,
2352 .probe = netfront_probe,
2353 .remove = xennet_remove,
2354 .resume = netfront_resume,
2355 .otherend_changed = netback_changed,
2358 static int __init netif_init(void)
2363 if (!xen_has_pv_nic_devices())
2366 pr_info("Initialising Xen virtual ethernet driver\n");
2368 /* Allow as many queues as there are CPUs inut max. 8 if user has not
2369 * specified a value.
2371 if (xennet_max_queues == 0)
2372 xennet_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
2375 return xenbus_register_frontend(&netfront_driver);
2377 module_init(netif_init);
2380 static void __exit netif_exit(void)
2382 xenbus_unregister_driver(&netfront_driver);
2384 module_exit(netif_exit);
2386 MODULE_DESCRIPTION("Xen virtual network device frontend");
2387 MODULE_LICENSE("GPL");
2388 MODULE_ALIAS("xen:vif");
2389 MODULE_ALIAS("xennet");