Linux Ethernet Bridging - Apr 2020 - [Bridge] BUG: soft lockup while deleting tap interface from vlan aware bridge

On 30/04/2020 13:55, Ido Schimmel wrote:
> On Wed, Apr 29, 2020 at 10:52:35PM +0200, Stefan Priebe - Profihost AG
wrote:
>> Hello,
>>
>> while running a stable vanilla kernel 4.19.115 i'm reproducably get
this
>> one:
>>
>> watchdog: BUG: soft lockup - CPU#38 stuck for 22s! [bridge:3570653]
>>
>> ...
>>
>> Call
>>
Trace:nbp_vlan_delete+0x59/0xa0br_vlan_info+0x66/0xd0br_afspec+0x18c/0x1d0br_dellink+0x74/0xd0rtnl_bridge_dellink+0x110/0x220rtnetlink_rcv_msg+0x283/0x360
> 
> Nik, Stefan,
> 
> My theory is that 4K VLANs are deleted in a batch and preemption is
> disabled (please confirm). For each VLAN the kernel needs to go over the
Right, that's what I was expecting. :-)
> entire FDB and delete affected entries. If the FDB is very large or the
> FDB lock is contended this can cause the kernel to loop for more than 20
> seconds without calling schedule().
Indeed, we already have that issue also with expire which goes over all entries.
I have rough patches that improve the situation from way back, will have to go
over and
polish them to submit when I got more time. Long ago I've tested it with
expiring 10 million
entries but on a rather powerful CPU.
> 
> To reproduce I added mdelay(100) in br_fdb_delete_by_port() and ran
> this:
> 
> ip link add name br10 up type bridge vlan_filtering 1
> ip link add name dummy10 up type dummy
> ip link set dev dummy10 master br10
> bridge vlan add vid 1-4094 dev dummy10 master
> bridge vlan del vid 1-4094 dev dummy10 master
> 
> Got a similar trace to Stefan's. Seems to be fixed by attached:
> 
> diff --git a/net/bridge/br_netlink.c b/net/bridge/br_netlink.c
> index a774e19c41bb..240e260e3461 100644
> --- a/net/bridge/br_netlink.c
> +++ b/net/bridge/br_netlink.c
> @@ -615,6 +615,7 @@ int br_process_vlan_info(struct net_bridge *br,
>                                                v - 1, rtm_cmd);
>                                 v_change_start = 0;
>                         }
> +                       cond_resched();
>                 }
>                 /* v_change_start is set only if the last/whole range
changed */
>                 if (v_change_start)
> 
> WDYT?
> 
Maybe we can batch the deletes at say 32 at a time?
Otherwise looks good to me, thanks!

On Thu, Apr 30, 2020 at 02:20:23PM +0300, Nikolay Aleksandrov
wrote:
> Maybe we can batch the deletes at say 32 at a time?
Hi Nik,

Thanks for looking into this!

I don't really feel comfortable hard coding an arbitrary number of
entries before calling cond_resched(). I didn't see a noticeable
difference in execution time with the previous patch versus an unpatched
kernel. Also, in the examples I saw in the networking code
cond_resched() is always called after each loop iteration.

Let me know how you prefer it and I will send a patch.