cmpxchg is not available to generic code

cmpxchg is not available to generic code

[Andrew Morton]

arm:

drivers/char/drm/drm_lock.c: In function `drm_lock_take':
drivers/char/drm/drm_lock.c:221: error: implicit declaration of function `cmpxchg'

You might be able to use atomic_cmpxchg, which _is_ present
on all architectures.  Or use a spinlock.

What's that code doing anyway?  driver-private locking primitives?

Re: cmpxchg is not available to generic code

[David Miller]

From: Andrew Morton <akpm@linux-foundation.org>
Date: Thu, 19 Jul 2007 00:05:49 -0700

> What's that code doing anyway?  driver-private locking primitives?

It's an atomic lock shared with userspace.  Whatever implementation is
used to do the lock on that object must be identical in the userspace
DRM bits.

Unlike futex, the lock operation on the user side isn't optional.
So if the platform can't do a true cmpxchg it generally cannot
support DRM.

Re: cmpxchg is not available to generic code

[Dave Airlie]

> arm:
>
> drivers/char/drm/drm_lock.c: In function `drm_lock_take':
> drivers/char/drm/drm_lock.c:221: error: implicit declaration of function `cmpxchg'
>
> You might be able to use atomic_cmpxchg, which _is_ present
> on all architectures.  Or use a spinlock.
>
> What's that code doing anyway?  driver-private locking primitives?

When did arm suddenly start wanting DRM? they need to grow a userpsace 
cmpxchg as davem mentioned to go along with this, changing the drm now 
isn't possible due to backwards compat..

Dave.

-- 
David Airlie, Software Engineer
http://www.skynet.ie/~airlied / airlied at skynet.ie
Linux kernel - DRI, VAX / pam_smb / ILUG

Re: cmpxchg is not available to generic code

[Andrew Morton]

On Thu, 19 Jul 2007 09:02:03 +0100 (IST) Dave Airlie <airlied@linux.ie> wrote:

> 
> 
> > arm:
> >
> > drivers/char/drm/drm_lock.c: In function `drm_lock_take':
> > drivers/char/drm/drm_lock.c:221: error: implicit declaration of function `cmpxchg'
> >
> > You might be able to use atomic_cmpxchg, which _is_ present
> > on all architectures.  Or use a spinlock.
> >
> > What's that code doing anyway?  driver-private locking primitives?
> 
> When did arm suddenly start wanting DRM?

It's selectable in config.  allmodconfig broke.

> they need to grow a userpsace 
> cmpxchg as davem mentioned to go along with this, changing the drm now 
> isn't possible due to backwards compat..

For reference purposes, that position is not acceptable.  We _never_ accept the
"oh I can't change my proposed kernel interface because I already have
userspace relying on it" argument.

Hopefully that won't be an issue here.  I guess DRM now needs a
`depends on !ARM'.

Re: cmpxchg is not available to generic code

[Dave Airlie]

On 7/19/07, David Miller <davem@davemloft.net> wrote:
> From: Andrew Morton <akpm@linux-foundation.org>
> Date: Thu, 19 Jul 2007 00:05:49 -0700
>
> > What's that code doing anyway?  driver-private locking primitives?
>
> It's an atomic lock shared with userspace.  Whatever implementation is
> used to do the lock on that object must be identical in the userspace
> DRM bits.
>
> Unlike futex, the lock operation on the user side isn't optional.
> So if the platform can't do a true cmpxchg it generally cannot
> support DRM.

Actually in  theory the userspace side is optional, it should fallback
to always entering the kernel and being slow, but Ive no idea how
well that codepath is tested... but it's an area I'd hate to play with
now ..

Maybe we could add CONFIG_HAVE_CMPXCHG and let DRM depend on it..

Dave.

Re: cmpxchg is not available to generic code

[Dave Airlie]

>
>> they need to grow a userpsace
>> cmpxchg as davem mentioned to go along with this, changing the drm now
>> isn't possible due to backwards compat..
>
> For reference purposes, that position is not acceptable.  We _never_ accept the
> "oh I can't change my proposed kernel interface because I already have
> userspace relying on it" argument.

Yes it is, I thought breaking userspace API was the worst crime in kernel 
history, (unless you are sysfs...) the userspace DRM has been around since 
2.2 days at least, so there are lots of legacy userspaces to break..

> Hopefully that won't be an issue here.  I guess DRM now needs a
> `depends on !ARM'.

Possibly the answer..

Dave.

-- 
David Airlie, Software Engineer
http://www.skynet.ie/~airlied / airlied at skynet.ie
Linux kernel - DRI, VAX / pam_smb / ILUG

Re: cmpxchg is not available to generic code

[Andrew Morton]

On Thu, 19 Jul 2007 18:15:03 +1000 "Dave Airlie" <airlied@gmail.com> wrote:

> Maybe we could add CONFIG_HAVE_CMPXCHG and let DRM depend on it..

That would certainly be better than adding a sprinkle of architectures
in DRM Kconfig dependencies.

I don't know how important DRM is on ARM.  Zero?

Re: cmpxchg is not available to generic code

[Dave Airlie]

On 7/19/07, Andrew Morton <akpm@linux-foundation.org> wrote:
> On Thu, 19 Jul 2007 18:15:03 +1000 "Dave Airlie" <airlied@gmail.com> wrote:
>
> > Maybe we could add CONFIG_HAVE_CMPXCHG and let DRM depend on it..
>
> That would certainly be better than adding a sprinkle of architectures
> in DRM Kconfig dependencies.
>
> I don't know how important DRM is on ARM.  Zero?
>

I'd guess zero I suppose if you wanted you could hook up a PCI
graphics card on ARM, but if you do that I think you could implement
cmpxchg :-)

Dave.

Re: cmpxchg is not available to generic code

[Andrew Morton]

On Thu, 19 Jul 2007 09:19:10 +0100 (IST) Dave Airlie <airlied@linux.ie> wrote:

> >
> >> they need to grow a userpsace
> >> cmpxchg as davem mentioned to go along with this, changing the drm now
> >> isn't possible due to backwards compat..
> >
> > For reference purposes, that position is not acceptable.  We _never_ accept the
> > "oh I can't change my proposed kernel interface because I already have
> > userspace relying on it" argument.
> 
> Yes it is, I thought breaking userspace API was the worst crime in kernel 
> history, (unless you are sysfs...) the userspace DRM has been around since 
> 2.2 days at least, so there are lots of legacy userspaces to break..

oh, sorry, I thought this cmpxchg stuff was newly-added.

I guess something changed which has now made DRM available under arm
allmodconfig.

Re: cmpxchg is not available to generic code

[David Miller]

From: Andrew Morton <akpm@linux-foundation.org>
Date: Thu, 19 Jul 2007 01:21:41 -0700

> On Thu, 19 Jul 2007 18:15:03 +1000 "Dave Airlie" <airlied@gmail.com> wrote:
> 
> > Maybe we could add CONFIG_HAVE_CMPXCHG and let DRM depend on it..
> 
> That would certainly be better than adding a sprinkle of architectures
> in DRM Kconfig dependencies.
> 
> I don't know how important DRM is on ARM.  Zero?

We already have EMULATED_CMPXCHG guarding DRM.

Re: cmpxchg is not available to generic code

[David Miller]

From: "Dave Airlie" <airlied@gmail.com>
Date: Thu, 19 Jul 2007 18:15:03 +1000

> Maybe we could add CONFIG_HAVE_CMPXCHG and let DRM depend on it..

There is already a "!EMULATED_CMPXCHG" guarding DRM, perhaps
ARM's Kconfig can set that if appropriate.

Re: cmpxchg is not available to generic code

[Nick Piggin]

Dave Airlie wrote:
> On 7/19/07, Andrew Morton <akpm@linux-foundation.org> wrote:
> 
>> On Thu, 19 Jul 2007 18:15:03 +1000 "Dave Airlie" <airlied@gmail.com> 
>> wrote:
>>
>> > Maybe we could add CONFIG_HAVE_CMPXCHG and let DRM depend on it..
>>
>> That would certainly be better than adding a sprinkle of architectures
>> in DRM Kconfig dependencies.
>>
>> I don't know how important DRM is on ARM.  Zero?
>>
> 
> I'd guess zero I suppose if you wanted you could hook up a PCI
> graphics card on ARM, but if you do that I think you could implement
> cmpxchg :-)

ARM does the locked load / store conditional thing which is at least as
strong as cmpxchg, so I imagine it could implement this API in kernel
and userspace quite easily if needed.

-- 
SUSE Labs, Novell Inc.