[RFC PATCH 00/12] Changes to code that reads iovec from userspace

[RFC PATCH 00/12] Changes to code that reads iovec from userspace

[David Laight]

This is RFC because we seem to be in a merge window.

The canonical code to read iov[] is currently:
	struct iovec iovstack[UIO_FASTIOV];
	struct iovec *iov;
	...
	iov = iovstack;
	rc = import_iovec(..., UIO_FASTIOV, &iov, &iter);
	if (rc < 0)
		return rc;
	...
	kfree(iov);

Note that the 'iov' parameter is used for two different things.
On input it is an iov[] can can be used.
On output it is an iov[] array that must be freed.

If 'iovstack' is passed, the count is actually always UIO_FASTIOV (8)
although in some places the array definition is in a different file
(never mind function) from the constant used.

import_iovec() itself is just a wrapper to rw_copy_check_uvector().
So everything is passed through to a second function.
Several items are 'passed by reference' - adding to the code paths.

On success import_iovec() returned the transfer count.
Only one caller looks at it, the count is also in iter.count.

The new canonical code is:
	struct iov_cache cache;
	struct iovec *iov;
	...
	iov = iovec_import(..., &cache, &iter);
	if (IS_ERR(iov))
		return PTR_ERR(iov);
	...
	kfree(iov);

Since 'struct iov_cache' is a fixed size there is no need to pass in
a length (correct or not!). It can still be NULL (used by the scsi code).

iovec_import() contains the code that used to be in rw_copy_check_uvector()
and then sets up the iov_iter.

rw_copy_check_uvector() is no more.
The only other caller was in mm/process_vm_access.c when reading the
iov[] for the target process addresses when copying from a differ process.
This can extract the iov[] from an extra 'struct iov_iter'.

In passing I noticed an access_ok() call on each fragment.
I hope this is just there to bail out early!
It is also skipped in process_vm_rw(). I did a quick look but couldn't
see an obvious equivalent check.

Patches 1 and 2 tidy up existing code.
Patches 3 and 4 add the new interface.
Patches 5 through 10 change all the callers.
Patch 11 removes a 'hack' that allowed fs/io_uring be updated before the socket code.
Patch 12 removes the old interface.

I suspect the changes need to trickle through a merge window.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

FW: [RFC PATCH 00/12] Changes to code that reads iovec from userspace

[David Laight]

> -----Original Message-----
> From: David Laight
> Sent: 31 March 2020 14:52
> To: Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
> Cc: netdev@vger.kernel.org
> Subject: [RFC PATCH 00/12] Changes to code that reads iovec from userspace
> 
> This is RFC because we seem to be in a merge window.
> 
> The canonical code to read iov[] is currently:
> 	struct iovec iovstack[UIO_FASTIOV];
> 	struct iovec *iov;
> 	...
> 	iov = iovstack;
> 	rc = import_iovec(..., UIO_FASTIOV, &iov, &iter);
> 	if (rc < 0)
> 		return rc;
> 	...
> 	kfree(iov);
> 
> Note that the 'iov' parameter is used for two different things.
> On input it is an iov[] can can be used.
> On output it is an iov[] array that must be freed.
> 
> If 'iovstack' is passed, the count is actually always UIO_FASTIOV (8)
> although in some places the array definition is in a different file
> (never mind function) from the constant used.
> 
> import_iovec() itself is just a wrapper to rw_copy_check_uvector().
> So everything is passed through to a second function.
> Several items are 'passed by reference' - adding to the code paths.
> 
> On success import_iovec() returned the transfer count.
> Only one caller looks at it, the count is also in iter.count.
> 
> The new canonical code is:
> 	struct iov_cache cache;
> 	struct iovec *iov;
> 	...
> 	iov = iovec_import(..., &cache, &iter);
> 	if (IS_ERR(iov))
> 		return PTR_ERR(iov);
> 	...
> 	kfree(iov);
> 
> Since 'struct iov_cache' is a fixed size there is no need to pass in
> a length (correct or not!). It can still be NULL (used by the scsi code).
> 
> iovec_import() contains the code that used to be in rw_copy_check_uvector()
> and then sets up the iov_iter.
> 
> rw_copy_check_uvector() is no more.
> The only other caller was in mm/process_vm_access.c when reading the
> iov[] for the target process addresses when copying from a differ process.
> This can extract the iov[] from an extra 'struct iov_iter'.
> 
> In passing I noticed an access_ok() call on each fragment.
> I hope this is just there to bail out early!
> It is also skipped in process_vm_rw(). I did a quick look but couldn't
> see an obvious equivalent check.
> 
> Patches 1 and 2 tidy up existing code.
> Patches 3 and 4 add the new interface.
> Patches 5 through 10 change all the callers.
> Patch 11 removes a 'hack' that allowed fs/io_uring be updated before the socket code.
> Patch 12 removes the old interface.
> 
> I suspect the changes need to trickle through a merge window.
> 
> 	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)