Re: [PATCH] adjust gfp mask passed on nested vmalloc() invocation

From: Hugh Dickins
Date: Wed Oct 07 2009 - 08:09:47 EST


On Wed, 7 Oct 2009, Jan Beulich wrote:
> >>> Hugh Dickins <hugh.dickins@xxxxxxxxxxxxx> 06.10.09 23:58 >>>
> >On Mon, 5 Oct 2009, Jan Beulich wrote:
> >
> >> - fix a latent bug resulting from blindly or-ing in __GFP_ZERO, since
> >> the combination of this and __GFP_HIGHMEM (possibly passed into the
> >> function) is forbidden in interrupt context
> >> - avoid wasting more precious resources (DMA or DMA32 pools), when
> >> being called through vmalloc_32{,_user}()
> >> - explicitly allow using high memory here even if the outer allocation
> >> request doesn't allow it, unless is collides with __GFP_ZERO
> >>
> >> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxxxx>
> >
> >I thought vmalloc.c was a BUG_ON(in_interrupt()) zone?
> >The locking is all spin_lock stuff, not spin_lock_irq stuff.
> >That's probably why your "bug" has remained "latent".
>
> Then you probably mean BUG_ON(irqs_disabled()), which would seem
> correct.

I'm relieved you came to see that remark as bogus.

> But if the gfp mask massaging was needed for calling kmalloc(),
> it would seem odd that the same shouldn't be needed for calling
> vmalloc() recursively...
>
> >Using HIGHMEM for internal arrays looks reasonable to me; but if
> >__GFP_ZERO were a problem, wouldn't it be much cleaner to skip the
> >"unless it collides" and #ifdef CONFIG_HIGHMEM !in_interrupt() stuff,
> >just memset the array returned from __vmalloc_node()?
>
> The main goal was to change the existing code as little as possible - I
> did consider this alternative, but wasn't sure that would be accepted.
> If you view this as the better alternative, I'll certainly modify the
> patch to do it that way.

Well, now we've accepted that this code cannot be used in_interrupt(),
there's no need for your #ifdef CONFIG_HIGHMEM nor for my memset: just
use __GFP_ZERO as it was before, and your patch would amount to or'ing
__GFP_HIGHMEM into gfp_mask for the __vmalloc_node case - wouldn't it?

Hugh
--
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/