Re: [PATCH] smp_call_function WARN_ON

[Manfred Spraul <manfred@colorfullife.com>]

Zwane wrote:

>The WARN_ON() is really only true if you're waiting for the other
>processors.
>
>  
>
Wrong: At least on i386 the code always waits for delivery of the IPI, 
it just doesn't wait for completion of the callback.
Do you need IPIs from irq or bh context? It would be tricky to change 
the current code: an IPI just delivers an interrupt without any payload. 
The global 'call_data' variable contains the description of the IPI and 
accesses to it must be synchronized.

IPIs from BH context do no work either: they can deadlock against the 
tasklist_lock: the first cpu calls read_lock(&tasklist_lock) from 
process context, another cpu calls write_lock_irq(&tasklist_lock), and 
then the first cpu tries smp_call_function() from bh context.

--
    Manfred