On Sun, Dec 16, 2012 at 10:05 AM, Andy Lutomirski [off-list ref] wrote:

On Sun, Dec 16, 2012 at 4:39 AM, Michel Lespinasse [off-list ref] wrote:

quoted

My main concern is that just downgrading the mmap_sem only hides the
problem: as soon as a writer gets queued on that mmap_sem,
reader/writer fairness kicks in and blocks any new readers, which
makes the problem reappear. So in order to completely fix the issue,
we should look for a way that doesn't require holding the mmap_sem
(even in read mode) for the entire duration of the populate or mlock
operation.

Ugh.

At least with my patch, mmap in MCL_FUTURE mode is no worse than mmap
+ mlock.  I suspect I haven't hit this because all my mmaping is done
by one thread, so it never ends up waiting for itself, and the other
thread have very short mmap_sem hold times.

Yes, you won't hit the problems with long read-side mmap_sem hold
times if you don't have other threads blocking for the write side.

quoted

I think this could be done by extending the mlock work I did as part
of v2.6.38-rc1. The commit message for
c explains the idea; basically
mlock() was split into do_mlock() which just sets the VM_LOCKED flag
on vmas as needed, and do_mlock_pages() which goes through a range of
addresses and actually populates/mlocks each individual page that is
part of a VM_LOCKED vma.

Doesn't this have the same problem?  It holds mmap_sem for read for a
long time, and if another writer comes in then r/w starvation
prevention will kick in.

Well, my point is that do_mlock_pages() doesn't need to hold the
mmap_sem read side for a long time. It currently releases it when
faulting a page requires a disk read, and could conceptually release
it more often if needed.

We can't easily release mmap_sem from within mmap_region() since
mmap_region's callers don't expect it; however we can defer the page
mlocking and we don't have to hold mmap_sem continuously until then.
The only constraints are the new VM_LOCKED region's pages must be
mlocked before we return to userspace, and that if a concurrent thread
modifies the mappings while we don't hold mmap_sem, and creates a new
non-mlocked region, we shouldn't mlock those pages in
do_mlock_pages().

-- 
Michel "Walken" Lespinasse
A program is never fully debugged until the last user dies.

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>