2002-09-04 21:23:42

by Juan M. de la Torre

[permalink] [raw]
Subject: Questions on semaphores


Hi people, I have two question regarding the i386 semaphore implementation
in kernel 2.4.19.

Please dont blame me if they are too obvius; i'm a newbie in kernel hacking
:)

The functions __down, __down_interruptible and __down_trylock (defined
in arch/i386/kernel/semaphore.c) use the global spinlock
'semaphore_lock' to access some fields of the semaphore they are
working on:

1) Is there any reason to do this?
2) Wouldn't it be more scalable to use a per-semaphore lock instead a
global spinlock?

The function __down_trylock try to get the spinlock using
spin_lock_irqsave, instead of using spin_lock_irq:

1) why? :)

Thanks in advance,
Juanma

--
/jm


2002-09-04 22:17:10

by Bob Miller

[permalink] [raw]
Subject: Re: Questions on semaphores

On Wed, Sep 04, 2002 at 11:29:31PM +0200, Juan M. de la Torre wrote:
>
> Hi people, I have two question regarding the i386 semaphore implementation
> in kernel 2.4.19.
>
> Please dont blame me if they are too obvius; i'm a newbie in kernel hacking
> :)
>
> The functions __down, __down_interruptible and __down_trylock (defined
> in arch/i386/kernel/semaphore.c) use the global spinlock
> 'semaphore_lock' to access some fields of the semaphore they are
> working on:
>
> 1) Is there any reason to do this?

It was easy to do.

> 2) Wouldn't it be more scalable to use a per-semaphore lock instead a
> global spinlock?
>

Yes it would be more scalable, but not as much as you would think.
The __down, __down_interruptible and __down_trylock code only gets
invoked when the semaphore is contended for.

> The function __down_trylock try to get the spinlock using
> spin_lock_irqsave, instead of using spin_lock_irq:
>
> 1) why? :)
>

The __down_trylock() code can be called with another lock held. The
spin_lock_irqsave()/spin_lock_irqrestore() interface is used to restore
the irq value for the lock that may already be held.

> Thanks in advance,
> Juanma
>
The code in the 2.5 tree was changed a while back to use the spinlock in
the wait_queue_head_t to replace the global semaphore spin lock. So, this
has been "FIXED" in 2.5.

--
Bob Miller Email: [email protected]
Open Source Development Lab Phone: 503.626.2455 Ext. 17