Re: [PATCH] ipv4: synchronize bind() with RTM_NEWADDR notifications

["Timo Teräs" <timo.teras@iki.fi>]

On 10/21/2010 02:03 PM, David Miller wrote:
> From: Timo Teräs <timo.teras@iki.fi>
> Date: Thu, 21 Oct 2010 13:58:08 +0300
> 
>> On 10/21/2010 01:50 PM, David Miller wrote:
>>> From: Timo Teräs <timo.teras@iki.fi>
>>> Date: Thu, 21 Oct 2010 13:41:37 +0300
>>>
>>>> Is inet_bind() called from non-userland context? If yes, then this is a
>>>> bad idea. Otherwise I don't think it's that hot path...
>>>
>>> It is.
>>
>> Yet, almost immediately after that there is lock_sock() which can also
>> sleep. How does that work then?
> 
> RTNL interlocks globally with a heavy primitive, a mutex, lock_sock()
> grabs a spinlcok which is local to the socket's context.
> 
> So if we have 100,000 sockets binding we'll suck with you're change
> whereas the lock_sock() does not cause that problem.
> 
> Is this so difficult for you to comprehend?

I was confused with Dave's original reply "It is." as referring to that
inet_bind() can get called from non-userland context. But apparently you
just meant that "It is (bad idea regardless)."

I thought the problem was possible sleeping, and not contention. Which
became very obvious from Eric's example. I didn't realize that many do
bind()/recv()/send() as general workload.

Sorry for not seeing the obvious.

This is the third time asking, what would be a good way to fix the
problem described in the original commit log?

Changing RTM_NEWADDR after FIB update would break Netlink event
ordering. And this breaks performance. I can't really use RTN_LOCAL
RTM_NEWROUTE events since (at least IPv6 side) has incorrect ifindex.

Should inet_addr_type() be rewritten to not use FIB lookups?