Re: oops in copy_page_rep()

[Linus Torvalds <torvalds@linux-foundation.org>]

[-- Attachment #1: Type: text/plain, Size: 916 bytes --]

On Tue, Jan 8, 2013 at 9:37 AM, Andrea Arcangeli <aarcange@redhat.com> wrote:
>
> The reason it returned to userland and retried the fault is that this
> should be infrequent enough not to worry about it and this was
> marginally simpler but it could be changed.

Yeah, that was my suspicion. And as mentioned, returning to user land
might actually help with scheduling and/or signal handling latencies
etc, so it might be the right thing to do.  Especially if the
alternative is to just busy-loop.

> If we don't want to return to userland we should wait on the splitting
> bit and then take the pte walking routines like if the pmd wasn't
> huge. This is not related to the below though.

How does this patch sound to people? It does the splitting check
before the access bit set (even though I don't think it matters), and
at least talks about the alternatives and the issues a bit.

Hmm?

                 Linus

[-- Attachment #2: mm.patch --]
[-- Type: application/octet-stream, Size: 750 bytes --]

 mm/memory.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index 49fb1cf08611..f5ec3ae03f44 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3715,6 +3715,18 @@ retry:
 				return do_huge_pmd_numa_page(mm, vma, address,
 							     orig_pmd, pmd);
 
+			/*
+			 * If the pmd is splitting, return and retry the
+			 * the fault. We *could* set just the accessed flag,
+			 * but it's better to just avoid the races with
+			 * splitting entirely.
+			 *
+			 * Alternative: wait until the split is done, and
+			 * goto retry.
+			 */
+			if (pmd_trans_splitting(orig_pmd))
+				return 0;
+
 			if (dirty && !pmd_write(orig_pmd)) {
 				ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
 							  orig_pmd);