Re: 2.4.1-pre10 deadlock (Re: ps hang in 241-pre10)

[Jens Axboe]

On Sun, Jan 28 2001, Linus Torvalds wrote:
> On Sun, 28 Jan 2001, Jens Axboe wrote:
> > 
> > How about this instead?
> 
> I really don't like this one. It will basically re-introduce the old
> behaviour of waking people up in a trickle, as far as I can tell. The
> reason we want the batching is to make people have more requests to sort
> in the elevator, and as far as I can tell this will just hurt that.
> 
> Are there any downsides to just _always_ batching, regardless of whether
> the request freelist is empty or not? Sure, it will make the "effective"
> size of the freelist a bit smaller, but that's probably not actually
> noticeable under any load except for the one that empties the freelist (in
> which case the old code would have triggered the batching anyway).

The problem with removing the !list_empty test like you suggested
is that batching is no longer controlled anymore. If we start
batching once the lists are empty and start wakeups once batch_requests
has been reached, we know we'll give the elevator enough to work
with to be effective. With !list_empty removed, batch_requests is no
longer a measure of how many requests we want to batch. Always
batching is not a in problem in itself, the effective smaller freelist
effect should be neglible.

The sent patch will only trickle wakeups in case of batching already
in effect, but batch_request wakeups were not enough to deplete
the freelist again. At least that was the intended effect :-)

> Performance numbers?

Don't have any right now, will test a bit later.

-- 
* Jens Axboe <[e-mail]>
* SuSE Labs