Re: Lockless/Get_User_Pages_Fast causes Xorg 1.4.99.* to lock

[Hugh Dickins]

On Mon, 7 Jul 2008, Carsten Otte wrote:
> Nick Piggin wrote:
> > I think we need a similar fix for s390 too. If so, then it really should
> > get into 2.6.26, but this late in the release, I hope an s390 maintainer
> > might be able to test and verify the fix?
> 
> I've done my best to combine mprotect, mmap and munmap a MAP_PRIVATE mapping
> on a xip file system. The system runs stable with and without this patch.
> Could someone please enlighten me on how to reproduce the problem so that I
> can verify the fix?

Though it would be more obvious to use a MAP_SHARED mapping, we had that
earlier thread in which it emerged that you're not using shared writable
xip mappings, IIRC.  So, sticking to MAP_PRIVATE, I'd expect the following
sequence

ptr = mmap(NULL, PAGE_SIZE, PROT_READ, MAP_PRIVATE, xip_fd, 0);
var = *ptr;
mprotect(ptr, PAGE_SIZE, PROT_NONE);
munmap(ptr, PAGE_SIZE);

to do a put_page on a non-existent struct page derived from the pfn:
perhaps corrupting other memory without being noticed?  Or if
you have CONFIG_DEBUG_VM=y (that would be a good move), to hit
vm_normal_page's VM_BUG_ON(!pfn_valid(pte_pfn(pte))) before that.

(I think you can just as well use PROT_READ|PROT_WRITE rather than
PROT_NONE there, but in principle mprotect could optimize away that
pte modification - though I think it goes ahead and does it anyway.)

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/