Re: [PATCH -mm 1/5][AIO] - Rework compat_sys_io_submit

["Sébastien Dugué" <sebastien.dugue@bull.net>]

On Wed, 29 Nov 2006 16:47:47 -0800, Zach Brown <zach.brown@oracle.com> wrote:

> On Nov 29, 2006, at 2:32 AM, Sébastien Dugué wrote:
> 
> >                  compat_sys_io_submit() cleanup
> >
> >
> >   Cleanup compat_sys_io_submit by duplicating some of the native  
> > syscall
> > logic in the compat layer and directly calling io_submit_one() instead
> > of fooling the syscall into thinking it is called from a native 64-bit
> > caller.
> >
> >   This is needed for the completion notification patch to avoid having
> > to rewrite each iocb on the caller stack for sys_io_submit() to  
> > find the
> > sigevents.
> 
> You could explicitly mention that this eliminates:
> 
>   - the overhead of copying nr pointers on the userspace caller's stack
> 
>   - the arbitrary PAGE_SIZE/(sizeof(void *)) limit on the number of  
> iocbs that can be submitted
> 
> Those alone make this worth merging.

  Right, will add.

> 
> > +	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
> > +		return -EFAULT;
> 
> I'm glad you got that right :)  I no doubt would have initially  
> hoisted these little checks into a shared helper function and missed  
> that detail of getting the size of the access_ok() right in the  
> compat case.

  Thanks.

> 
> > +	put_ioctx(ctx);
> > +
> > +	return i? i: ret;
> 
> sys_io_getevents() reads:

 uh!     ^^^^^^^^^    you must be meaning sys_io_submit()?

> 
>          put_ioctx(ctx);
>          return i ? i : ret;
> 
> So while this compat_sys_io_submit() logic seems fine and I would be  
> comfortable with it landing as-is, I'd also appreciate it if we  
> didn't introduce differences between the two functions when it seems  
> just as easy to make them the same.  (That chunk is just one  
> example.  There's whitespace, missing unlikely()s, etc).
> 

  OK, will fix.

  Thanks,

  Sébastien.