Re: Kernel Panic with latest patch from git repo

On Tue, Feb 21, 2006 at 09:07:48PM +0100, Matthieu Pérotin wrote:
> When migrating a process from a node A to a node B, the node B crashes
> on a kernel panic.

I think I've finally figured out what really happens. When I wrote
earlier (to the list) that Node B receives the process at interrupt
time, I was wrong. It only looks that way (and the kernel actually
thinks that way) because the preempt counter overflows. (Note that
that counter is kept even on a kernel with CONFIG_PREEMPT off.)

The counter overflows because there is an unbalanced call to
pte_alloc_map() in mig_do_receive_page(). Now, if CONFIG_HIGHPTE is
set, user space PTEs are allocated in high memory. pte_alloc_map()
will atomically map the PTE so that it can be used; however, that map
is never released.
If CONFIG_HIGHPTE is off, the mapping is #defined away, so the problem
does not surface.

I've created a patch that enables migration even with CONFIG_HIGHPTE
on. I've tested it here, and it seems to work OK, so I'd like to
submit it for inclusion with openMosix.

A.

=======================================================================

diff -ruN linux-2.6-om/hpc/migrecv.c linux-2.6-om-work/hpc/migrecv.c
--- linux-2.6-om/hpc/migrecv.c	2006-02-07 16:51:58.000000000 +0100
+++ linux-2.6-om-work/hpc/migrecv.c	2006-02-24 21:13:35.000000000 +0100
@@ -183,6 +183,7 @@
 		OMBUG("double page at addr %p\n", (void *) addr);
 
 	set_pte(pte, pte_mkdirty(mk_pte(recv_page, vma->vm_page_prot)));
+	pte_unmap(pte);
 	page_dup_rmap(recv_page);
 	inc_mm_counter(mm, file_rss);
 	
-- 
Ansgar Esztermann
Researcher & Sysadmin
http://www2.thphy.uni-duesseldorf.de/~ansgar

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.1 (GNU/Linux)

iD8DBQFD/2/jwAkt8h1/xRgRAhOmAJ48HiKQEfyQ++OcPob49ivpgqs0+gCgggoj
7Rz4MQVYzEJdpLy6lSYnky4=
=Y9Gs
-----END PGP SIGNATURE-----