Re: [PATCH 2/2] mm: page_alloc: avoid slowpath for more thanMAX_ORDER allocation.

[Johannes Weiner]

On Mon, Jul 22, 2013 at 07:01:18PM -0700, PINTU KUMAR wrote:
> >Lastly, order >= MAX_ORDER is not supported by the page allocator, and
> >we do not want to punish 99.999% of all legitimate page allocations in
> >the fast path in order to catch an unlikely situation like this.
[...]
> >Having the check only in the slowpath is a good thing.
> >
> Sorry, I could not understand, why adding this check in slowpath is only good.
> We could have returned failure much before that.
> Without this check, we are actually allowing failure of "first allocation attempt" and then returning the cause of failure in slowpath.
> I thought it will be better to track the unlikely failure in the system as early as possible, at least from the embedded system prospective.
> Let me know your opinion.

This is a trade-off between two cases: we expect (almost) all
allocations to be order < MAX_ORDER, so we want that path as
lightweight as possible.  On the other hand, we expect that only very
rarely an allocation will specify order >= MAX_ORDER.  By doing the
check late, we make the common case faster at the expense of the rare
case.  That's the whole point of having a fast path and a slow path.

What you are proposing would punish 99.999% of all cases in order to
speed up the 0.001% cases.  In addition, these 0.001% of all cases
will fail the allocation, so performance is the least of their
worries.  It's a bad trade-off.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/