Le 07.06.2006 08:22, Jens Axboe a écrit :
> On Tue, Jun 06 2006, Andrew Morton wrote:
>> On Wed, 7 Jun 2006 00:05:07 +0200
>> Ingo Molnar <mingo@elte.hu> wrote:
>>
>>> * Laurent Riffard <laurent.riffard@free.fr> wrote:
>>>
>>>>>> Results:
>>>>>> - 2.6.17-rc4-mm3 with 4K stack works fine (this is the latest good 4K-kernel).
>>>>>> - 2.6.17-rc5-mm3 with 4K stack crashes, the stack seems to be corrupted.
>>>>> that's vanilla mm3, or mm3 patched with extra lockdep patches? If it's 
>>>>> patched then you should try vanilla mm3 too.
>>>> It was vanilla mm3.
>>> ok, i'll check the stack impact of the block_dev.c changes tomorrow.
>>>
>> Note that Laurent is also passing through ide_cdrom_packet(), which has a
>> `struct request' on the stack.  The kernel does this in a lot of places,
>> and at 168 bytes on x86, it'd really be best if we were to dynamically
>> allocate these things.
> 
> That's an old peeve of mine, on-stack requests... It's nasty from
> several angles, stack usage just being one of them. Perhaps I'll give it
> a go for 2.6.18 and add checks for request being thrown at the block
> layer which didn't originate from get_request().

Let me say I successfuly run a 2.6.17-rc6-mm1 kerenel with 4K-stack 
and CONFIG_DEBUG_LOCKDEP=y, so I guess my stack problems are fixed.

I got a "possible circular locking deadlock detected" while loading
pktcdvd service, I'll make a detailed report in another thread.
 
.config is attached for information.

-- 
laurent