From: Konrad Rzeszutek <konradr@redhat.com>
Subject: [RHEL5 U1 PATCH] 263281: LTC38081-[ TPC-C ] madvise call to kernel loops forever
Date: Thu, 30 Aug 2007 15:07:55 -0400
Bugzilla: 263281
Message-Id: <20070830190755.GA25128@mars.boston.redhat.com>
Changelog: [mm] madvise call to kernel loops forever


RHBZ#:
------
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=263281

Description:
------------

This git commnit was posted on RHKL as
"[PATCH RHEL5] reduce MADV_DONTNEED contention"
(http://post-office.corp.redhat.com/archives/rhkernel-list/2007-May/msg00623.html)
but missed the patch for the madvise_remove function. I spoke
to Rik van Riel (the poster) and he agreed that this should
have been posted.

commit 0a27a14a62921b438bb6f33772690d345a089be6
Author: Nick Piggin <npiggin@suse.de>

    mm: madvise avoid exclusive mmap_sem

    Avoid down_write of the mmap_sem in madvise when we can help it.

    Acked-by: Hugh Dickins <hugh@veritas.com>
    Signed-off-by: Nick Piggin <npiggin@suse.de>
    Cc: Rik van Riel <riel@redhat.com>
    Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
    Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>


RHEL Version Found:
------------------
RHEL5.0

kABI Status:
------------
No symbols were harmed.

Upstream Status:
----------------
It is upsteam in 2.6.23

Test Status:
------------
I have tested the stock and modified kernel with a test-case that exercises
this code and found it to respectivly not work, and work. I am also
running LTP test-cases against the modified kernel to make sure that
there are no regressions introduced.

Proposed Patch:
---------------
This patch is based on 2.6.18-44.el5

diff -uNrp linux-2.6.18.i386.orig/mm/madvise.c linux-2.6.18.i386/mm/madvise.c
--- linux-2.6.18.i386.orig/mm/madvise.c	2007-08-30 10:20:56.000000000 -0400
+++ linux-2.6.18.i386/mm/madvise.c	2007-08-30 10:31:04.000000000 -0400
@@ -201,9 +201,9 @@ static long madvise_remove(struct vm_are
 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 
 	/* vmtruncate_range needs to take i_mutex and i_alloc_sem */
-	up_write(&current->mm->mmap_sem);
+	up_read(&current->mm->mmap_sem);
 	error = vmtruncate_range(mapping->host, offset, endoff);
-	down_write(&current->mm->mmap_sem);
+	down_read(&current->mm->mmap_sem);
 	return error;
 }

 
-- 
Konrad Rzeszutek 1-(978)-392-3903 or 1-(617)-693-1718
IBM on-site partner.