x86: fix global_flush_tlb() bug
authorIngo Molnar <mingo@elte.hu>
Fri, 19 Oct 2007 10:19:26 +0000 (12:19 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Fri, 19 Oct 2007 10:19:26 +0000 (12:19 +0200)
While we were reviewing pageattr_32/64.c for unification,
Thomas Gleixner noticed the following serious SMP bug in
global_flush_tlb():

down_read(&init_mm.mmap_sem);
list_replace_init(&deferred_pages, &l);
up_read(&init_mm.mmap_sem);

this is SMP-unsafe because list_replace_init() done on two CPUs in
parallel can corrupt the list.

This bug has been introduced about a year ago in the 64-bit tree:

       commit ea7322decb974a4a3e804f96a0201e893ff88ce3
       Author: Andi Kleen <ak@suse.de>
       Date:   Thu Dec 7 02:14:05 2006 +0100

       [PATCH] x86-64: Speed and clean up cache flushing in change_page_attr

                down_read(&init_mm.mmap_sem);
        -       dpage = xchg(&deferred_pages, NULL);
        +       list_replace_init(&deferred_pages, &l);
                up_read(&init_mm.mmap_sem);

the xchg() based version was SMP-safe, but list_replace_init() is not.
So this "cleanup" introduced a nasty bug.

why this bug never become prominent is a mystery - it can probably be
explained with the (still) relative obscurity of the x86_64 architecture.

the safe fix for now is to write-lock init_mm.mmap_sem.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
arch/x86/mm/pageattr_64.c

index 8a4f65bf956ea85ef45ecb6157eccceca82f92a5..c7b7dfe1d405153717fa8f91fa15ecbde809b2b7 100644 (file)
@@ -230,9 +230,14 @@ void global_flush_tlb(void)
        struct page *pg, *next;
        struct list_head l;
 
-       down_read(&init_mm.mmap_sem);
+       /*
+        * Write-protect the semaphore, to exclude two contexts
+        * doing a list_replace_init() call in parallel and to
+        * exclude new additions to the deferred_pages list:
+        */
+       down_write(&init_mm.mmap_sem);
        list_replace_init(&deferred_pages, &l);
-       up_read(&init_mm.mmap_sem);
+       up_write(&init_mm.mmap_sem);
 
        flush_map(&l);