Re: [PATCH] macros to detect existance of unlocked_ioctl and ioctl_compat

[Andi Kleen <ak@suse.de>]

> > Agree that's a problem. We just need an alternative interface here
> > or I try to come up with an x86-64 specific hack (I think it's possible
> > to do the compatibility x86-64 specific, just not in compat code for
> > all architectures who have truly separated user/kernel address spaces) 
> > 
> > I hope the USB people will eventually add a better interface here.
> > Greg?
> 
> Yes, we've been talking about a usbfs2 for a long time now, but no one
> has stepped up and done the work.  It would look a lot like gadgetfs,
> which only has a limited number of ioctls needed for its operation.
> 
> But I don't see where the problem is unique to usbfs here, don't we
> properly use the 32->64bit ioctl conversion code to solve this kind of
> issue?  Are we using it incorrectly?

DaveM can probably give you more details since he tried unsucessfully
to make it work. I think the problem is that there is no enough
information for the compat layer to convert everything.

Here's his comment from fs/compat_ioctl.c.

On x86-64 it can be a bit easier because you can get away with
only partially converting a data structure because the 32bit
space is still visible from the kernel. But that doesn't work
on some other archs. 

-Andi


/* This needs more work before we can enable it.  Unfortunately
 * because of the fancy asynchronous way URB status/error is written
 * back to userspace, we'll need to fiddle with USB devio internals
 * and/or reimplement entirely the frontend of it ourselves. -DaveM
 *
 * The issue is:
 *
 *      When an URB is submitted via usbdevicefs it is put onto an
 *      asynchronous queue.  When the URB completes, it may be reaped
 *      via another ioctl.  During this reaping the status is written
 *      back to userspace along with the length of the transfer.
 *
 *      We must translate into 64-bit kernel types so we pass in a kernel
 *      space copy of the usbdevfs_urb structure.  This would mean that we
 *      must do something to deal with the async entry reaping.  First we
 *      have to deal somehow with this transitory memory we've allocated.
 *      This is problematic since there are many call sites from which the
 *      async entries can be destroyed (and thus when we'd need to free up
 *      this kernel memory).  One of which is the close() op of usbdevicefs.
 *      To handle that we'd need to make our own file_operations struct which
 *      overrides usbdevicefs's release op with our own which runs usbdevicefs's
 *      real release op then frees up the kernel memory.
 *
 *      But how to keep track of these kernel buffers?  We'd need to either
 *      keep track of them in some table _or_ know about usbdevicefs internals
 *      (ie. the exact layout of its file private, which is actually defined
 *      in linux/usbdevice_fs.h, the layout of the async queues are private to
 *      devio.c)
 *
 * There is one possible other solution I considered, also involving knowledge
 * of usbdevicefs internals:
 *
 *      After an URB is submitted, we "fix up" the address back to the user
 *      space one.  This would work if the status/length fields written back
 *      by the async URB completion lines up perfectly in the 32-bit type with
 *      the 64-bit kernel type.  Unfortunately, it does not because the iso
 *      frame descriptors, at the end of the struct, can be written back.
 *
 * I think we'll just need to simply duplicate the devio URB engine here.
 */