On Fri, Sep 18, 2020 at 3:44 PM Christoph Hellwig [off-list ref] wrote:

On Fri, Sep 18, 2020 at 02:40:12PM +0100, Al Viro wrote:

quoted

quoted

    /* Vector 0x110 is LINUX_32BIT_SYSCALL_TRAP */
-   return pt_regs_trap_type(current_pt_regs()) == 0x110;
+   return pt_regs_trap_type(current_pt_regs()) == 0x110 ||
+           (current->flags & PF_FORCE_COMPAT);

Can't say I like that approach ;-/  Reasoning about the behaviour is much
harder when it's controlled like that - witness set_fs() shite...

I don't particularly like it either.  But do you have a better idea
how to deal with io_uring vs compat tasks?

Do we need to worry about something other than the compat_iovec
struct for now? Regarding the code in io_import_iovec(), it would
seem that can easily be handled by exposing an internal helper.
Instead of

#ifdef CONFIG_COMPAT
     if (req->ctx->compat)
            return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
iovec, iter);
#endif
        return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);

This could do

    __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec,
                     iter, req->ctx->compat);

With the normal import_iovec() becoming a trivial wrapper around
the same thing:

ssize_t import_iovec(int type, const struct iovec __user * uvector,
                 unsigned nr_segs, unsigned fast_segs,
                 struct iovec **iov, struct iov_iter *i)
{
     return __import_iovec(type, uvector, nr_segs, fast_segs, iov,
              i, in_compat_syscall());
}


         Arnd