On 10/02/18 13:07, H. Peter Anvin wrote:
> Hi Johan et al,
>
> I'm trying to figure out what the status is of BOTHER and termios2 is in
> various architectures, and I saw these patches:
>
> 1cee38f0363a88db374e50b232ca17b9a4c12fa0
> fefe287e4bf6ee23a5d9422a0a49de5389acc712
>
> As you might know glibc never implemented this stuff, and I'm trying to fix
> that. One of their questions is whether there are architectures on which
> capability does *NOT* exist at this point.
>
> If you could help straighten this out it would be great.
>
> Specifically, are there any platforms:
>
> 1. where simply stuffing values into c_ispeed or c_ospeed and setting BOTHER
> can be used to set the baud rate, even for the "standard" baud rates
> (obviously breaking if the kernel is too old -- how old is that?)
>
> 2. where c_ispeed and c_ospeed aren't set, even if the value in c_cflags isn't
> BOTHER?
>
> 3.
>
> Also:
>
> 4. I note alpha -- apparently as the only architecture -- doesn't define
> BOTHER, even though it has B* constants that aren't the same as the
> corresponding baud rates. It does, however, have c_[io]speed in its
> legacy struct termios. Do you happen to know how this
> is supposed to work on Alpha? There are claims that Alpha doesn't need it,
> yet I see absolutely no support code for actually setting arbitary
>
> In fact, as far as I can tell, we could actually end up overrunning the
> baud_table[] array on Alpha, as the limit check is only done on cbaud &
> CBAUDEX, and CBAUDEX on Alpha is zero, yet Alpha has 32 possible values for
> c_cflags & CBAUD, and the array is only 31 entries long.
>
> Should we just
>
> #define BOTHER 037
>
> on Alpha, which ought to solve both problems, or is there a solution that
> works even on legacy kernels?
>
> On a final note, <uapi/asm/termbits.h> isn't usable in building a libc, since
> it pollutes the namespace. I would like to propose:
>
> a. Break out the constants into a separate uapi file;
> b. Rename the exported structures __kernel_termios, __kernel_termios2 etc,
> for the users which have explicit ioctls and need these structures.
> This is a common problem in the headers, which bites extra hard with
> names that are tags and not typedefs. The only real solution is to use
> a very hidden namespace like __kernel_ and use #define if needed.
> c. [Optionally] use an #ifdef to add these defines or, perhaps better for
> legacy users, *not* define them. Alternatively a libc that doesn't want
> them could #undef them.
>
Ping on this? I'm getting traction withe the glibc people and I want to
nail this down as soon as possible.
-hpa
