Re: get_user_pages() with write=1 and force=1 gets read-only pages.

[Nick Piggin <nickpiggin@yahoo.com.au>]

[-- Attachment #1: Type: text/plain, Size: 430 bytes --]

Robin Holt wrote:

> So far, I think the case for setting VM_FAULT_RACE only when there
> is a conflict for writable seems strong.
> 

But that's 1162 collisions out of how many faults? And only
on the get_user_pages faults does the return value really
make a difference.

How about the following? This should catch some cases. You
still miss the case where you're racing with anotyher writer
though.

-- 
SUSE Labs, Novell Inc.


[-- Attachment #2: mm-gup-write-fix.patch --]
[-- Type: text/plain, Size: 469 bytes --]

Index: linux-2.6/mm/memory.c
===================================================================
--- linux-2.6.orig/mm/memory.c	2005-07-31 11:49:35.000000000 +1000
+++ linux-2.6/mm/memory.c	2005-08-01 08:21:21.000000000 +1000
@@ -971,7 +971,9 @@ int get_user_pages(struct task_struct *t
 					 * that we have actually performed
 					 * the write fault (below).
 					 */
-					continue;
+					if (write)
+						continue;
+					break;
 				default:
 					BUG();
 				}