On 21/09/2020 00:13, David Laight wrote:

From: Arnd Bergmann

quoted

Sent: 20 September 2020 21:49

On Sun, Sep 20, 2020 at 9:28 PM Andy Lutomirski [off-list ref] wrote:

quoted

On Sun, Sep 20, 2020 at 12:23 PM Matthew Wilcox [off-list ref] wrote:

quoted

On Sun, Sep 20, 2020 at 08:10:31PM +0100, Al Viro wrote:

quoted

IMO it's much saner to mark those and refuse to touch them from io_uring...

Simpler solution is to remove io_uring from the 32-bit syscall list.
If you're a 32-bit process, you don't get to use io_uring.  Would
any real users actually care about that?

We could go one step farther and declare that we're done adding *any*
new compat syscalls :)

Would you also stop adding system calls to native 32-bit systems then?

On memory constrained systems (less than 2GB a.t.m.), there is still a
strong demand for running 32-bit user space, but all of the recent Arm
cores (after Cortex-A55) dropped the ability to run 32-bit kernels, so
that compat mode may eventually become the primary way to run
Linux on cheap embedded systems.

I don't think there is any chance we can realistically take away io_uring
from the 32-bit ABI any more now.

Can't it just run requests from 32bit apps in a kernel thread that has
the 'in_compat_syscall' flag set?
Not that i recall seeing the code where it saves the 'compat' nature
of any requests.

It is already completely f*cked if you try to pass the command ring
to a child process - it uses the wrong 'mm'.

And how so? io_uring uses mm of a submitter. The exception is SQPOLL
mode, but it requires CAP_SYS_ADMIN or CAP_SYS_NICE anyway.

I suspect there are some really horrid security holes in that area.

	David.

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

-- 
Pavel Begunkov