Re: [PATCH] can: vcan: use per-CPU device stats

[Oliver Hartkopp <socketcan@hartkopp.net>]

On 08.10.26 06:13, Yogesh Gaur wrote:
> On Wed, Oct 7, 2026 at 10:41 PM Oliver Hartkopp <socketcan@hartkopp.net> wrote:
>>
>>
>>
>> On 07.10.26 13:36, Yogesh Gaur wrote:
>>> vcan_tx() and vcan_rx() update dev->stats with plain increments and
>>> rely on the tx queue lock to serialise them.  That only holds for a
>>> single tx queue.  vcan does not implement get_num_tx_queues(), so
>>> rtnl_create_link() honours IFLA_NUM_TX_QUEUES and a vcan device can be
>>> created with several queues
>>
>> dev_dstats_tx_add() has been introduced in Linux 6.14 so it is probably
>> some can-next material for Linux 7.4+.
> 
> Agreed.
> 
>>
>> Btw. why don't we implement get_num_tx_queues() and return 1 as a
>> potential stable fix? This would add a correct locking via the existing
>> tx queue lock for all vcan statistics right?
> 
> The locking argument holds: vcan is not LLTX, and every counter update
> (tx, the echo-mode rx in vcan_rx(), and tx_dropped from
> can_dropped_invalid_skb()) happens in vcan_tx() under the xmit locak.
> 
> But get_num_tx_queues() alone does not stop the multi-queue case, as
> rtnl_create_link() only consults it when userspace did not ask.

Right, but there is no reason why userspace should set this value - 
except if you want to race in the stats now ;-)

>>
>> Best regards,
>> Oliver
>>
>> ps. vxcan should have a similar problem.
> 
> Yes - same pattern and the same single-queue argument works there.
> 
> Plan for v2: limiting vcan and vxcan to one TX queue, then the dstats
> conversion for can-next.

Great! This might then also cover the dstats for "real" CAN devices as 
Marc already pointed out.

Many thanks!
Oliver

> 
> Yogesh
>>
>> , each with its own xmit lock.  Two CPUs
>>> transmitting on different queues then update the same counters
>>> concurrently and lose increments:
>>>
>>>     BUG: KCSAN: data-race in vcan_tx / vcan_tx
>>>     read-write to 0xffff88811aad8228 of 8 bytes by interrupt on cpu 1:
>>>      vcan_tx+0x325/0x5d0 drivers/net/can/vcan.c:110
>>>      dev_hard_start_xmit+0x10c/0x380 net/core/dev.c:3969
>>>      __dev_queue_xmit+0xbfd/0x1ec0 net/core/dev.c:4958
>>>      can_send+0x584/0x720 net/can/af_can.c:279
>>>      bcm_can_tx+0x3ba/0x5b0 net/can/bcm.c:367
>>>     read-write to 0xffff88811aad8228 of 8 bytes by interrupt on cpu 0:
>>>      vcan_tx+0x325/0x5d0 drivers/net/can/vcan.c:110
>>>     value changed: 0x00000000000025ef -> 0x00000000000025f0
>>>
>>> Switch the packet and byte counters to the core's per-CPU dstats.
>>> The core allocates them for NETDEV_PCPU_STAT_DSTATS and folds them
>>> together with dev->stats in dev_get_stats(), so tx_dropped from
>>> can_dropped_invalid_skb() is still reported.
>>>
>>> Reported-by: syzbot+937a3a1fbfbc99d6fcaf@syzkaller.appspotmail.com
>>> Closes: https://syzkaller.appspot.com/bug?extid=937a3a1fbfbc99d6fcaf
>>> Assisted-by: LLM
>>> Signed-off-by: Yogesh Gaur <yogeshgaur.83@gmail.com>
>>> ---
>>>    drivers/net/can/vcan.c | 13 ++++---------
>>>    1 file changed, 4 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/drivers/net/can/vcan.c b/drivers/net/can/vcan.c
>>> index 76e6b7b5c6a1..2ff90cb7eee9 100644
>>> --- a/drivers/net/can/vcan.c
>>> +++ b/drivers/net/can/vcan.c
>>> @@ -71,10 +71,7 @@ MODULE_PARM_DESC(echo, "Echo sent frames (for testing). Default: 0 (Off)");
>>>
>>>    static void vcan_rx(struct sk_buff *skb, struct net_device *dev)
>>>    {
>>> -     struct net_device_stats *stats = &dev->stats;
>>> -
>>> -     stats->rx_packets++;
>>> -     stats->rx_bytes += can_skb_get_data_len(skb);
>>> +     dev_dstats_rx_add(dev, can_skb_get_data_len(skb));
>>>
>>>        skb->pkt_type  = PACKET_BROADCAST;
>>>        skb->dev       = dev;
>>> @@ -85,7 +82,6 @@ static void vcan_rx(struct sk_buff *skb, struct net_device *dev)
>>>
>>>    static netdev_tx_t vcan_tx(struct sk_buff *skb, struct net_device *dev)
>>>    {
>>> -     struct net_device_stats *stats = &dev->stats;
>>>        unsigned int len;
>>>        int loop;
>>>
>>> @@ -93,8 +89,7 @@ static netdev_tx_t vcan_tx(struct sk_buff *skb, struct net_device *dev)
>>>                return NETDEV_TX_OK;
>>>
>>>        len = can_skb_get_data_len(skb);
>>> -     stats->tx_packets++;
>>> -     stats->tx_bytes += len;
>>> +     dev_dstats_tx_add(dev, len);
>>>
>>>        /* set flag whether this packet has to be looped back */
>>>        loop = skb->pkt_type == PACKET_LOOPBACK;
>>> @@ -107,8 +102,7 @@ static netdev_tx_t vcan_tx(struct sk_buff *skb, struct net_device *dev)
>>>                        /* only count the packets here, because the
>>>                         * CAN core already did the echo for us
>>>                         */
>>> -                     stats->rx_packets++;
>>> -                     stats->rx_bytes += len;
>>> +                     dev_dstats_rx_add(dev, len);
>>>                }
>>>                consume_skb(skb);
>>>                return NETDEV_TX_OK;
>>> @@ -185,6 +179,7 @@ static void vcan_setup(struct net_device *dev)
>>>        dev->netdev_ops         = &vcan_netdev_ops;
>>>        dev->ethtool_ops        = &vcan_ethtool_ops;
>>>        dev->needs_free_netdev  = true;
>>> +     dev->pcpu_stat_type     = NETDEV_PCPU_STAT_DSTATS;
>>>    }
>>>
>>>    static struct rtnl_link_ops vcan_link_ops __read_mostly = {
>>