Messages in this thread | | | From | Linus Torvalds <> | Date | Fri, 21 Dec 2012 13:51:26 -0800 | Subject | Re: [patch] mm, mempolicy: Introduce spinlock to read shared policy tree |
| |
On Fri, Dec 21, 2012 at 10:21 AM, Hugh Dickins <hughd@google.com> wrote: > On Fri, 21 Dec 2012, Linus Torvalds wrote: >> >> compared to the diseased abortion you just posted. > > I'm picking up a vibe that you don't entirely like Mel's approach.
Good job. I was a bit nervous that I was being too subtle.
> I don't understand David's and Mel's remarks about the "shared pages" > check making Sasha's warning unlikely: page_mapcount has nothing to do > with whether a page belongs to shm/shmem/tmpfs, and it's easy enough > to reproduce Sasha's warning on the current git tree. "mount -o > remount,mpol=local /tmp" or something like that is useful in testing.
I think that Mel and David may talk about the mutex actually blocking (not just the debug message possibly triggering).
> I wish wish wish I had time to spend on this today, but I don't. > And I've not looked to see (let alone tested) whether it's easy > to revert Mel's mutex then add in Kosaki's patch (which I didn't > look at so have no opinion on).
I don't actually have Kosaki's patch either, just the description of it. We've done that kind of "preallocate before taking the lock" before, though.
> Shall we go for Peter/David's mutex+spinlock for rc1 - I assume > they both tested that - with a promise to do better in rc2?
Well, if the plan is to fix it for rc2, then there is no point in putting a workaround in now, since actually hitting the problem (as opposed to seeing the warning) is presumably much harder.
Linus
| |