Re: making asm-generic/futex.h completely usable

[Mike Frysinger <vapier.adi@gmail.com>]

On Fri, Jul 3, 2009 at 11:51, Arnd Bergmann wrote:
> On Friday 03 July 2009, Mike Frysinger wrote:
>> ive been reading up on futexes of late and have been poking around the
>> futex kernel pieces to see what is needed to get Blackfin working with
>> it.
>>
>> Blackfin falls into the "no hardware atomic instructions" category as
>> can easily be seen in our atomic.h: disable interrupts, do
>> load/stores, restore interrupts.  my understanding of the
>> futex_atomic_op_inuser() function is that this runs in process
>> context, so this same interrupt trick should work fine.  which leads
>> me to wonder why doesnt the asm-generic/futex.h header already take
>> this approach ?
>
> The code does not run in user space, the name indicates that it
> operates on variables in user space while running in the kernel.

when i said "process context" i didnt mean it was running in usermode.
 i meant the context of the kernel execution is of a process, so it
can sleep.  since it needs to use get/put user, that means it can
sleep fine.

> The irq-disable trick fundamentally does not work on SMP systems,
> which need architecture specific atomic operations for this.

i'm guessing spin lock with irq semantics wouldnt work due to the
possibility of get/put user sleeping ...

>> seems to me that the SuperH implementation fits the bill nicely.
>> their arch/sh/include/asm/futex-irq.h looks like it could be literally
>> straight copied into asm-generic/futex.h thus making it fully
>> functional for everyone by default.
>
> That sounds reasonable for the arch/sh/include/asm/futex.h file, but
> the futex-irq.h file would still be reserved for non-SMP architectures.
> I believe that out of the architectures currently using asm-generic/futex.h,
> blackfin is the only one that supports SMP, so you still lose ;-)

if the functions were able to be overridden on a per-function basis,
that is easy to handle with CONFIG_SMP in the Blackfin header.
-mike