]> asedeno.scripts.mit.edu Git - linux.git/commit
sparc64: disable fast-GUP due to unexplained oopses
authorLinus Torvalds <torvalds@linux-foundation.org>
Tue, 15 Oct 2019 16:56:36 +0000 (09:56 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 15 Oct 2019 16:56:36 +0000 (09:56 -0700)
commit8e0d0ad206f08506c893326ca7c9c3d9cc042cef
treef60a7b84591b0e6f861c5c8789f7c0bd3e86a6be
parent02755af0f322fd1d2d7995c1e9fd6f9668511560
sparc64: disable fast-GUP due to unexplained oopses

HAVE_FAST_GUP enables the lockless quick page table walker for simple
cases, and is a nice optimization for some random loads that can then
use get_user_pages_fast() rather than the more careful page walker.

However, for some unexplained reason, it seems to be subtly broken on
sparc64.  The breakage is only with some compiler versions and some
hardware, and nobody seems to have figured out what triggers it,
although there's a simple reprodicer for the problem when it does
trigger.

The problem was introduced with the conversion to the generic GUP code
in commit 7b9afb86b632 ("sparc64: use the generic get_user_pages_fast
code"), but nothing looks obviously wrong in that conversion.  It may be
a compiler bug that just hits us with the code reorganization.  Or it
may be something very specific to sparc64.

This disables HAVE_FAST_GUP entirely.  That makes things like futexes a
bit slower, but at least they work.  If we can figure out the trigger,
that would be lovely, but it's been three months already..

Link: https://lore.kernel.org/lkml/20190717215956.GA30369@altlinux.org/
Fixes: 7b9afb86b632 ("sparc64: use the generic get_user_pages_fast code")
Reported-by: Dmitry V Levin <ldv@altlinux.org>
Reported-by: Anatoly Pugachev <matorola@gmail.com>
Requested-by: Meelis Roos <mroos@linux.ee>
Suggested-by: Christoph Hellwig <hch@infradead.org>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
arch/sparc/Kconfig