Re: [PATCH v3 x86 1/2] fix page faults by nmiaction in nmi ifkmemcheck is enabled

From: Li Zhong
Date: Fri Mar 09 2012 - 04:53:18 EST


On Tue, 2012-03-06 at 11:27 +0100, Vegard Nossum wrote:
> On 6 March 2012 11:09, Li Zhong <zhong@xxxxxxxxxxxxxxxxxx> wrote:
> > This patch tries to fix the problem of page fault exception caused by
> > accessing nmiaction structure in nmi if kmemcheck is enabled.
> >
> > If kmemcheck is enabled, the memory allocated through slab are in pages
> > that are marked non-present, so that some checks could be done in the
> > page fault handling code ( e.g. whether the memory is read before
> > written to ).
> > As nmiaction is allocated in this way, so it resides in a non-present
> > page. Then there is a page fault while the nmi code accessing the
> > nmiaction structure, which would then cause a warning by
> > WARN_ON_ONCE(in_nmi()) in kmemcheck_fault(), called by do_page_fault().
> >
> > v2: as Peter suggested, changed the nmiaction to use static storage.
> >
> > v3: as Peter suggested, use macro to shorten the codes. Also keep the
> > original usage of register_nmi_handler, so users of this call doesn't
> > need change.
> >
> > Signed-off-by: Li Zhong <zhong@xxxxxxxxxxxxxxxxxx>
>
> Looks like you've solved this now. Thanks.

There is still one place [2/2] not solved ... and I guess it might need
the way you suggested below.

>
> For the record, another way to prevent the page fault from happening
> in the first place is to set up a new slab cache with the flag
> SLAB_NOTRACK. This is different from the GFP_NOTRACK flag which, as
> you noted, doesn't prevent page faults, just inhibits
> checking/warnings for those memory areas.
>
> It's a bit of a hassle, I admit. Maybe we could create an additional,
> separate set of slab caches (using SLAB_NOTRACK) and a new GFP flag
> which selects this set of caches instead. This would allow anything
> that takes a gfp_t to allocate memory that is guaranteed not to page
> fault when using kmemcheck. Pekka, any thoughts?
>

I'm not sure whether I understand it correctly?
If CONFIG_KMEMCHECK is enabled, create another two sets of
malloc_sizes caches, one for cs_cachep, one for cs_dmacachep, both with
SLAB_NOTRACK flag.

Create a new GFP flag, like __GFP_NO_PF for those places where page
fault is not allowed, and return memory from the caches created above.
This new GFP flag is set to 0 if CONFIG_KMEMCHECK is not enabled.

I think there shouldn't be many such cases, so most of these caches
wouldn't actually be used ...

Thanks,
Zhong

>
> Vegard
>


--
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/