Re: [PATCH v4 14/14] tty: gunyah: Add tty console driver for RM Console Services

[Elliot Berman]

On 10/10/2022 1:23 PM, Greg Kroah-Hartman wrote:
> On Sun, Oct 09, 2022 at 01:59:21PM -0700, Elliot Berman wrote:
> > On 10/7/2022 12:40 AM, Greg Kroah-Hartman wrote:
> > > On Thu, Oct 06, 2022 at 10:59:51PM -0700, Elliot Berman wrote:
> > > > "GH" is the shorthand we've been using for "Gunyah". I didn't find
> > > > documentation for dynamically assigned char devices, but if it exists, I can
> > > > add entry for ttyGH.
> > >
> > > Why use a new name at all?  Why not stick with the existing tty names
> > > and device numbers?
> >
> > I can use hvc framework, although driver-level buffering is needed on
> > both the get_chars/put_chars paths because:
>
> I'm not asking about the framework (although that is a good question,
> you need to document why this has to be new.)  I'm asking why pick a new
> name?  You will not have a name conflict in your system with this device
> with any other tty name right?

That's correct, I didn't see any other instances of "ttyGH" in kernel.

> >   - get_chars wants to poll for characters, but Gunyah will push
> >     characters to Linux
> >   - put_chars can be called in atomic context in the printk console path.
> >     Gunyah RM calls can sleep, so we add to buffer and queue work to
> >     write the characters.
> >
> > I also chose to use new tty driver because the Gunyah hypervisor call to
> > open the console (gh_rm_console_open) can only be done after starting the
> > VM. Gunyah will only forward characters sent from the other VM to Linux
> > after the gh_rm_console_open call is made. When launching a VM, users would
> > want to open console before VM starts so they can get startup messages from
> > the VM. I planned to use the carrier_raised() to hold
> > tty_port_block_until_ready until the VM is started and the
> > gh_rm_console_open() happens.
>
> I'm sorry, but I don't understand this.
>
> Why is this all a new api at all?  What about the virtio api?  Why not
> just use that instead?

We want to support virtual machines and Virtual Machine Managers (the 
userspace component) that don't support virtio. Qualcomm has some 
lightweight VMs that have a Gunyah driver stack but no virtio stack. 
Further, implementing a simple userspace VMM to launch a Linux kernel is 
much easier with the Gunyah console as no device paravirtualization is 
needed to get some output from Linux. I don't anticipate these VMs to be 
commonplace, but they do exist.

Thanks,
Elliot