Re: [BUG 3.0.0-rc1] ksm: NULL pointer dereference in ksm_do_scan()

[Hugh Dickins <hughd@google.com>]

On Thu, 2 Jun 2011, Andrea Arcangeli wrote:
> On Thu, Jun 02, 2011 at 10:29:39AM -0700, Hugh Dickins wrote:
> > AndreaA, I didn't study the patch you posted half an hour ago,
> > since by that time I'd worked it out and was preparing patch below.
> > I think your patch would be for a different bug, hopefully one we
> > don't have, it looks more complicated than we should need for this.
> 
> I didn't expect two different bugs leading to double free.

There wasn't a double free there, just failure to cope with race
emptying the list, so accessing head when expecting a full entry.

> 
> If you've time please review my other patch too because mmput runs
> with no mmap_sem hold and I think the ksm scan code runs under the
> assumption that __ksm_exit is waiting in down_write() when
> ksm_mmlist_lock is released (before freeing the mm_slot), and that
> assumption is wrong. ksm_test_exit may very well be true despite
> __ksm_exit didn't run yet, and ksm scan will proceed freeing after
> changing the mm_slot and ksm_exit will be free to run and free again
> immediately after the ksm scan releases the ksm_mmlist_lock and before
> it clears the MMF_VM_MERGEABLE (because the mm_slot has been changed
> before releasing the ksm_mmlist_lock).
> 
> The rmap_list being null will kind of hide it, the fact there's so
> little time between the unlock of the ksm_mmlist_lock and the clearing
> of MMF_VM_MERGEABLE (that will stop ksm_exit from calling __ksm_exit
> at all) will also hide it. At least in
> unmerge_and_remove_all_rmap_items remove_trailing_rmap_items will nuke
> the rmap_list just before this race runs so making it more likely
> possible.

You'll see from the "beware" comment in scan_get_next_rmap_item()
that this case is expected, that it sometimes reaches freeing the
slots before the exiting task reaches __ksm_exit().

That race should already be handled.  I believe your patch is unnecessary,
because get_mm_slot() is a hashlist lookup, and will return NULL once
either end has done the hlist_del(&mm_slot->link).

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>