Re: [PATCH 1/9] kernel: add a PF_FORCE_COMPAT flag

[Al Viro]

On Sun, Sep 20, 2020 at 03:55:47PM +0200, Arnd Bergmann wrote:
> On Sun, Sep 20, 2020 at 12:09 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> > On Fri, Sep 18, 2020 at 05:16:15PM +0200, Christoph Hellwig wrote:
> > > On Fri, Sep 18, 2020 at 02:58:22PM +0100, Al Viro wrote:
> > > > Said that, why not provide a variant that would take an explicit
> > > > "is it compat" argument and use it there?  And have the normal
> > > > one pass in_compat_syscall() to that...
> > >
> > > That would help to not introduce a regression with this series yes.
> > > But it wouldn't fix existing bugs when io_uring is used to access
> > > read or write methods that use in_compat_syscall().  One example that
> > > I recently ran into is drivers/scsi/sg.c.
> >
> > So screw such read/write methods - don't use them with io_uring.
> > That, BTW, is one of the reasons I'm sceptical about burying the
> > decisions deep into the callchain - we don't _want_ different
> > data layouts on read/write depending upon the 32bit vs. 64bit
> > caller, let alone the pointer-chasing garbage that is /dev/sg.
> 
> Would it be too late to limit what kind of file descriptors we allow
> io_uring to read/write on?
> 
> If io_uring can get changed to return -EINVAL on trying to
> read/write something other than S_IFREG file descriptors,
> that particular problem space gets a lot simpler, but this
> is of course only possible if nobody actually relies on it yet.

S_IFREG is almost certainly too heavy as a restriction.  Looking through
the stuff sensitive to 32bit/64bit, we seem to have
	* /dev/sg - pointer-chasing horror
	* sysfs files for efivar - different layouts for compat and native,
shitty userland ABI design (
struct efi_variable {
        efi_char16_t  VariableName[EFI_VAR_NAME_LEN/sizeof(efi_char16_t)];
        efi_guid_t    VendorGuid;
        unsigned long DataSize;
        __u8          Data[1024];
        efi_status_t  Status;
        __u32         Attributes;
} __attribute__((packed));
) is the piece of crap in question; 'DataSize' is where the headache comes
from.  Regular files, BTW...
	* uhid - character device, milder pointer-chasing horror.  Trouble
comes from this:
/* Obsolete! Use UHID_CREATE2. */
struct uhid_create_req {
        __u8 name[128];
        __u8 phys[64];
        __u8 uniq[64];
        __u8 __user *rd_data;
        __u16 rd_size;

        __u16 bus;
        __u32 vendor;
        __u32 product;
        __u32 version;
        __u32 country;
} __attribute__((__packed__));
and suggested replacement doesn't do any pointer-chasing (rd_data is an
embedded array in the end of struct uhid_create2_req).
	* evdev, uinput - bitness-sensitive layout, due to timestamps
	* /proc/bus/input/devices - weird crap with printing bitmap, different
_text_ layouts seen by 32bit and 64bit readers.  Binary structures are PITA,
but with sufficient effort you can screw the text just as hard...  Oh, and it's
a regular file.
	* similar in sysfs analogue

And AFAICS, that's it for read/write-related method instances.