+Jack
> On 20 Apr 2018, at 17:34, Jason Gunthorpe <[email protected]> wrote:
>
> On Thu, Apr 19, 2018 at 11:55:55PM -0400, Doug Ledford wrote:
>> On Wed, 2018-04-18 at 16:24 +0200, Håkon Bugge wrote:
>>> Two kernel threads may get the same value for agent.hi_tid, if the
>>> agents are registered for different ports. As of now, this works, as
>>> the agent list is per port.
>>>
>>> It is however confusing and not future robust. Hence, making it
>>> atomic.
>>>
>>
>> People sometimes underestimate the performance penalty of atomic ops.
>> Every atomic op is the equivalent of a spin_lock/spin_unlock pair.
Well, may be this holds true if the mutex and the variable is located in the
same cacheline.
>> This
>> is why two atomics are worse than taking a spin_lock, doing what you
>> have to do, and releasing the spin_lock. Is this really what you want
>> for a "confusing, let's make it robust" issue?
>
> But it is on the ib_register_mad_agent() path which is not a
> performance path..
>
> This actually looks like a genuine bug, why is it described only as
> 'confusing'? ib_register_mad_agent is callable from userspace, so at
> least two userspace agents can race and get the same TID’s.
My understanding is that every lookup is using the {port, TID} tuple. As such,
it is not a bug, but, very confusing.
> TIDs need to be globally unique on the entire machine.
If you are correct Jason, let me reword the commit message.
Thxs, Håkon
