Re: [PATCH net 3/8] idpf: fix possible race in idpf_vport_stop()

[Jacob Keller <jacob.e.keller@intel.com>]

[-- Attachment #1.1: Type: text/plain, Size: 3672 bytes --]



On 10/6/2025 5:07 PM, Tantilov, Emil S wrote:
> 
> 
> On 10/6/2025 10:26 AM, Jakub Kicinski wrote:
>> On Mon, 6 Oct 2025 07:49:32 -0700 Tantilov, Emil S wrote:
>>>> Argh, please stop using the flag based state machines. They CANNOT
>>>> replace locking. If there was proper locking in place it wouldn't
>>>> have mattered when we clear the flag.
>>>
>>> This patch is resolving a bug in the current logic of how the flag is
>>> used (not being atomic and not being cleared properly). I don't think
>>> there is an existing lock in place to address this issue, though we are
>>> looking to refactor the code over time to remove and/or limit how these
>>> flags are used.
>>
>> Can you share more details about the race? If there is no lock in place
>> there's always the risk that:
>>
>>    CPU 0                         CPU 1
>>   idpf_vport_stop()             whatever()
>>                                   if (test_bit(UP))
>>                                    # sees true
>>                                  # !< long IRQ arrives
>>   test_and_clear(UP)
>>     ...
>>     all the rest
>>     ...
>>                                  # > long IRQ ends
>>                                  proceed but UP isn't really set any more
>>
> 
> The specific case I was targeting with this patch is for when both
> idpf_vport_stop() and idpf_addr_unsync(), called via set_rx_mode attempt
> to delete the MAC filters. At least in my testing I have not seen a case
> where the set_rx_mode callback will happen before idpf_vport_stop(). I
> am assuming due to userspace reacting to the removal of the netdevs.
> 
>             rmmod-6089    [021] .....  3521.291596: idpf_remove 
> <-pci_device_remove
>             rmmod-6089    [021] .....  3521.292686: idpf_vport_stop 
> <-idpf_vport_dealloc
>   systemd-resolve-1633    [022] b..1.  3521.295320: idpf_set_rx_mode 
> <-dev_mc_del
>   systemd-resolve-1633    [022] b..1.  3521.295338: idpf_addr_unsync 
> <-__hw_addr_sync_dev
>   systemd-resolve-1633    [022] b..1.  3521.295339: idpf_del_mac_filter 
> <-idpf_addr_unsync
>   systemd-resolve-1633    [022] b..1.  3521.295450: idpf_set_rx_mode 
> <-dev_mc_del
>   systemd-resolve-1633    [022] b..1.  3521.295451: idpf_addr_unsync 
> <-__hw_addr_sync_dev
>   systemd-resolve-1633    [022] b..1.  3521.295451: idpf_del_mac_filter 
> <-idpf_addr_unsync
>             rmmod-6089    [002] .....  3521.934980: idpf_vport_stop 
> <-idpf_vport_dealloc
>   systemd-resolve-1633    [022] b..1.  3522.297299: idpf_set_rx_mode 
> <-dev_mc_del
>   systemd-resolve-1633    [022] b..1.  3522.297316: idpf_addr_unsync 
> <-__hw_addr_sync_dev
>   systemd-resolve-1633    [022] b..1.  3522.297317: idpf_del_mac_filter 
> <-idpf_addr_unsync
>    kworker/u261:2-3157    [037] ...1.  3522.297931: 
> idpf_mac_filter_async_handler: Received invalid MAC filter payload (op 
> 536) (len 0)
>             rmmod-6089    [020] .....  3522.573251: idpf_vport_stop 
> <-idpf_vport_dealloc
>             rmmod-6089    [002] .....  3523.229936: idpf_vport_stop 
> <-idpf_vport_dealloc
>   systemd-resolve-1633    [022] b..1.  3523.311435: idpf_set_rx_mode 
> <-dev_mc_del
>   systemd-resolve-1633    [022] b..1.  3523.311452: idpf_addr_unsync 
> <-__hw_addr_sync_dev
>   systemd-resolve-1633    [022] b..1.  3523.311453: idpf_del_mac_filter 
> <-idpf_addr_unsync
> 
>

Jakub, from this thread it still seems like you won't accept this patch
as-is?

I'm working on a v3 of this series resolving the other issues Simon
pointed out, and want to know if I should continue excluding Emil's
patches for now.

Thanks,
Jake



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 236 bytes --]