Re: [PATCH -mm 3/4][AIO] - AIO completion signal notification

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

Zach Brown <zach.brown@oracle.com> wrote:

> Sébastien Dugué wrote:
> >                       AIO completion signal notification
> 
> This is looking a lot better, thanks for keeping at it.
> 
> > +static long aio_setup_sigevent(struct aio_notify *notify,
> > +			       struct sigevent __user *user_event)
> > +{
> > +	sigevent_t event;
> > +	struct task_struct *target;
> > +
> > +	if (copy_from_user(&event, user_event, sizeof (event)))
> > +		return -EFAULT;
> 
> Last time we talked about this needing to call get_compat_sigevent().  I
> think it still needs to.

  Yes, I could not find an elegant way to tackle this, so I left it
float for a while.

> 
> I think we should avoid the examples set by the current
> compat_sys_io_submit() and get_compat_sigevent() callers.  They copy
> translated data on to the userspace stack and pass it to the syscalls.
> That will get crazy for compat_sys_io_submit() because it would have to
> rewrite the iocb and the pointer to the iocb to get sys_io_submit() to
> find a copied sigevent on the stack.

  Right, that's what I really wanted to avoid at all costs.

> 
> I think the model is compat_do_readv_writev().  Hoist some of the
> syscall logic up into the compat layer so that one copying and
> translating pass is made instead of trying to fool the syscall logic
> into thinking that it's being called from a native word size caller.

  Yes, but you end up duplicating a lot of code, but if it's the price to
pay, I'm all for it.

> 
> So io_submit_one() should be given the kernel copies of the userspace
> structures it needs.  sys_io_submit() will pass it the copies it made
> for native word size callers.  compat_sys_io_submit() will pass in the
> copies it made after translating from 32bit arguments.  io_submit_one()
> and lookup_kioctx() will have to be made available to kernel/compat.c
> (via linux/aio.h, surely.).  aio_setup_sigevent() will be called from
> *_io_submit() and given the kernel sigevent, not a userspace pointer.
> 
> Reworking things this way should have the added benefit of making 32/64
> sys_io_submit() more efficient than it is today.

  I completely aggree. I will look into this.

  Thanks,

  Sébastien.