Re: [PATCH 0/5] make *_gate_vma accept mm_struct instead of task_struct

[Stephen Wilson <wilsons@start.ca>]

On Thu, Mar 10, 2011 at 08:00:32AM -0800, Andi Kleen wrote:
> On Tue, Mar 08, 2011 at 07:31:56PM -0500, Stephen Wilson wrote:
> > 
> > Morally, the question of whether an address lies in a gate vma should be asked
> > with respect to an mm, not a particular task.
> > 
> > Practically, dropping the dependency on task_struct will help make current and
> > future operations on mm's more flexible and convenient.  In particular, it
> > allows some code paths to avoid the need to hold task_lock.
> > 
> > The only architecture this change impacts in any significant way is x86_64.
> > The principle change on that architecture is to mirror TIF_IA32 via
> > a new flag in mm_context_t. 
> 
> The problem is -- you're adding a likely cache miss on mm_struct for
> every 32bit compat syscall now, even if they don't need mm_struct
> currently (and a lot of them do not) Unless there's a very good
> justification to make up for this performance issue elsewhere
> (including numbers) this seems like a bad idea.

I do not think this will result in cache misses on the scale you
suggest.  I am simply mirroring the *state* of the TIF_IA32 flag in
mm_struct, not testing/accessing it in the same way.

The only place where this flag is accessed (outside the exec() syscall
path) is in x86/mm/init_64.c, get_gate_vma(),  which in turn is needed
by a few, relatively heavy weight, page locking/pinning routines on the
mm side (get_user_pages, for example).  Patches 3 and 4 in the series
show the extent of the change.

Or am I missing something?


> > /proc/pid/mem.  I will be posting the second series to lkml shortly.  These
> 
> Making every syscall slower for /proc/pid/mem doesn't seem like a good
> tradeoff to me. Please solve this in some other way.
> 
> -Andi

-- 
steve