Re: [PATCH -mm] swsusp: userland interface (rev 2)

["Rafael J. Wysocki" <rjw@sisk.pl>]

Hi,

On Wednesday, 25 January 2006 13:18, Pavel Machek wrote:
> > > > +	case SNAPSHOT_ATOMIC_RESTORE:
> > > > +		if (data->mode != O_WRONLY || !data->frozen ||
> > > > +		    !snapshot_image_loaded(&data->handle)) {
> > > > +			error = -EPERM;
> > > > +			break;
> > > > +		}
> > > > +		down(&pm_sem);
> > > > +		pm_prepare_console();
> > > > +		error = device_suspend(PMSG_FREEZE);
> > > > +		if (!error) {
> > > > +			mb();
> > > > +			error = swsusp_resume();
> > > > +			device_resume();
> > > > +		}
> > > 
> > > whee, what does the mystery barrier do?  It'd be nice to comment this
> > > (please always comment open-coded barriers).
> > 
> > Pavel should know. ;-)
> 
> Pavel does not known. That memory barrier should be part of assembly
> parts, anyway, and AFAIK it is. Should be safe to kill.

OK

> > > > +	case SNAPSHOT_GET_SWAP_PAGE:
> > > > +		if (!access_ok(VERIFY_WRITE, (unsigned long __user *)arg, _IOC_SIZE(cmd))) {
> > > > +			error = -EINVAL;
> > > > +			break;
> > > > +		}
> > > 
> > > Why do we need an access_ok() here?
> > 
> > Because we use __put_user() down the road?
> > 
> > The problem is if the address is wrong we should not try to call
> > alloc_swap_page() at all.  If we did, we wouldn't be able to return the result
> > and we would leak a swap page.
> 
> I think you need to watch for failing put_user and free the page at
> that point. Anything else is racy as __put_user() may fail.

The page will be freed anyway when the device is closed (I was wrong
saying it would be "leaked"), so I think I'll just use put_user().

Greetings,
Rafael