Re: 2.6.5-rc1-aa3

[Andrea Arcangeli <andrea@suse.de>]

On Sat, Mar 20, 2004 at 10:17:16PM -0800, Martin J. Bligh wrote:
> > Fixed the sigbus in nopage and improved the page_t layout per Hugh's
> > suggestion. BUG() with discontigmem disabled if somebody returns non-ram
> > via do_no_page, that cannot work right on numa anyways.
> 
> OK, well it doesn't oops any more. But sshd still dies as soon as it starts,
> so accessing the box is tricky ;-) And now I have no obvious diagnostics
> either ...

Jens sent me the perfect strace log, after his help it has not been
difficult to spot the bug. this incremental should fix it
MAP_SHARED|MAP_ANONYMOUS isn't very common and my userspace never
triggered it. I placed the pgoff anon setting in the path of the shared
memory too, that generated the sigbus. Leaving the setting only in the
MAP_PRIVATE should fix it, the anonymous memory is only MAP_PRIVATE.

patch is untested at the moment, as soon as I get confirmation I'll
upload an update.

thanks!

--- x/mm/mmap.c.~1~	2004-03-20 22:12:43.000000000 +0100
+++ x/mm/mmap.c	2004-03-21 14:15:17.269882800 +0100
@@ -622,11 +622,11 @@ unsigned long do_mmap_pgoff(struct file 
 			return -EINVAL;
 		case MAP_PRIVATE:
 			vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
-			/* fall through */
+			pgoff = addr >> PAGE_SHIFT;
+			break;
 		case MAP_SHARED:
 			break;
 		}
-		pgoff = addr >> PAGE_SHIFT;
 	}
 
 	error = security_file_mmap(file, prot, flags);