Messages in this thread | | | Date | Thu, 4 Apr 2013 12:01:52 -0700 (PDT) | From | Hugh Dickins <> | Subject | Re: [PATCH] mm: prevent mmap_cache race in find_vma() |
| |
On Thu, 4 Apr 2013, Linus Torvalds wrote: > On Thu, Apr 4, 2013 at 11:35 AM, Hugh Dickins <hughd@google.com> wrote: > > > > find_vma() can be called by multiple threads with read lock > > held on mm->mmap_sem and any of them can update mm->mmap_cache. > > Prevent compiler from re-fetching mm->mmap_cache, because other > > readers could update it in the meantime: > > Ack. I do wonder if we should mark the unlocked update too some way > (also in find_vma()), although it's probably not a problem in practice > since there's no way the compiler can reasonably really do anything > odd with it. We *could* make that an ACCESS_ONCE() write too just to > highlight the fact that it's an unlocked write to this optimistic data > structure.
Hah, you beat me to it.
I wanted to get Jan's patch in first, seeing as it actually fixes his observed issue; and it is very nice to have such a good description of one of those, when ACCESS_ONCE() is usually just an insurance policy.
But then I was researching the much rarer "ACCESS_ONCE(x) = y" usage (popular in drivers/net/wireless/ath/ath9k and kernel/rcutree* and sound/firewire, but few places else).
When Paul reminded us of it yesterday, I came to wonder if actually every use of ACCESS_ONCE in the read form should strictly be matched by ACCESS_ONCE whenever modifying the location.
My uneducated guess is that strictly it ought to, in the sense of insurance policy; but that (apart from that strange split writing issue which came up a couple of months ago) in practice our compilers have not "advanced" to the point of making this an issue yet.
> > Anyway, applied.
Thanks, Hugh
| |