Re: [PATCH] Replace completions with semaphores

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2008-04-14 at 08:58 -0700, Roland Dreier wrote:
> > which ones exactly are these places that demand the use of a counting 
>  > semaphore? I cannot think of a single place where it's the best choice, 
>  > let alone one where it's the only choice.
> 
> Two of the places that use semaphores are drivers/infiniband/hw/mthca
> and drivers/net/mlx4 -- in both cases, the device firmware allows up to
> "N" outstanding firmware commands to be in flight, and the driver uses a
> semaphore to handle issuing firmware commands.  That is, down() when we
> want to issue a command, and up() when the firmware responds that the
> command is complete.
> 
> What would you suggest as a better way to code this?  This is an honest
> question -- there probably is a more elegant way to handle this
> situation and I really would like to learn about it.
> 
> Also, the argument that removing semaphores makes the kernel as a whole
> better does make sense to me; I wouldn't be opposed to basically
> open-coding semaphores in terms of wait_event() in the driver or
> something like that, but I wouldn't say that such an implementation is
> locally more readable or maintainable if we look only at the driver
> code.

Yeah, I would open code it.  But this is indeed a sane usage of the
counting semaphore because there is no priority inversion.