Re: [PATCH] Fix OOPS in mmap_region() when merging adjacent VM_LOCKED file segments

[Lee Schermerhorn <Lee.Schermerhorn@hp.com>]

On Fri, 2009-01-30 at 10:14 -0800, Linus Torvalds wrote:
> 
> On Fri, 30 Jan 2009, Hugh Dickins wrote:
> > 
> > ... what I think you have done is break the vma merging on
> > ordinary files: because of that irritating VM_CAN_NONLINEAR
> > flag which generic_file_mmap() and some others add in.
> 
> Ahh. Yes. VM_CAN_NONLINEAR is a "reverse flag", ie unlike the other flags 
> it's to some degree about being extra _normal_, not about being odd. Most 
> special flags tend to disable the VM from doing some clever thing, this 
> one enables it.
> 
> > To break the merging won't cause anyone much trouble,
> > but is a slight regression we should fix.
> 
> Yeah. Just masking it off when comparing is probably the simplest option. 
> Make it a separate #define just for readability. There might be other 
> flags like this in the future.
> 
> > I'd have been very upset not to find something ;)
> 
> Yay for you ;)
> 
> And yes, this is the kind of thing that probably does mean that we're 
> better off with the no-semantic-changes patches in -stable.
> 
> 		Linus
> 
> ---
>  mm/mmap.c |    5 ++++-
>  1 files changed, 4 insertions(+), 1 deletions(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index d3fa10a..c581df1 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -658,6 +658,9 @@ again:			remove_next = 1 + (end > next->vm_end);
>  	validate_mm(mm);
>  }
>  
> +/* Flags that can be inherited from an existing mapping when merging */
> +#define VM_MERGEABLE_FLAGS (VM_CAN_NONLINEAR)
> +
>  /*
>   * If the vma has a ->close operation then the driver probably needs to release
>   * per-vma resources, so we don't attempt to merge those.
> @@ -665,7 +668,7 @@ again:			remove_next = 1 + (end > next->vm_end);
>  static inline int is_mergeable_vma(struct vm_area_struct *vma,
>  			struct file *file, unsigned long vm_flags)
>  {
> -	if (vma->vm_flags != vm_flags)
> +	if ((vma->vm_flags ^ vm_flags) & ~VM_MERGEABLE_FLAGS)
>  		return 0;
>  	if (vma->vm_file != file)
>  		return 0;

I tried this patch atop 29-rc3 + your patch from yesterday with my
simple test program at http://free.linux/hp.com/~lts/Tests/mmap_lock.c. 

The test program shows the /proc/<pid>/maps before and after the mmap
and attempted merge.  It's not merging:

7fd20a668000-7fd20a66a000 rw-p 7fd20a668000 00:00 0 
7fd20a68a000-7fd20a68b000 r--s 00000000 68:23 6608852                    /tmp/tmpfVx1SFL (deleted)
7fd20a68b000-7fd20a68c000 r--s 00001000 68:23 6608852                    /tmp/tmpfVx1SFL (deleted)
7fd20a68c000-7fd20a690000 rw-p 7fd20a68c000 00:00 0 

Ad hoc instrumentation shows that it's the VM_ACCOUNT flag that is
different between the existing file segment and the one attempting the
merge:

is_mergeable_vma: !mergable: vma flags:  0x80020f9:0x1020f9
                                           |         |-VM_ACCOUNT
                                           +-----------VM_CAN_NONLINEAR

So happens, I'm mapping with MAP_SHARED, so the VM_ACCOUNT flag gets
cleared later in mmap_region().  Comments say that this is for checking
memory availability during shmem_file_setup().  Maybe we can move the
temporary setting of VM_ACCOUNT until just before the call to
shmem_zero_setup()?

Lee