Re: [RFC PATCH v2 3/6] tpm: add send_recv() ops in tpm_class_ops

[Jarkko Sakkinen <jarkko@kernel.org>]

On Wed, Mar 05, 2025 at 10:04:25AM +0100, Stefano Garzarella wrote:
> On Tue, Mar 04, 2025 at 10:21:55PM +0200, Jarkko Sakkinen wrote:
> > On Tue, Mar 04, 2025 at 06:56:02PM +0200, Jarkko Sakkinen wrote:
> > > On Mon, 2025-03-03 at 17:21 +0100, Stefano Garzarella wrote:
> > > > On Sat, Mar 01, 2025 at 03:45:10AM +0200, Jarkko Sakkinen wrote:
> > > > > On Fri, Feb 28, 2025 at 06:07:17PM +0100, Stefano Garzarella wrote:
> > > > > > +	int (*send_recv)(struct tpm_chip *chip, u8 *buf, size_t
> > > > > > buf_len,
> > > > > > +			 size_t to_send);
> > > > >
> > > > > Please describe the meaning and purpose of to_send.
> > > >
> > > > Sure, I'll add in the commit description.
> > > 
> > > It's always a command, right? So better be more concerete than
> > > "to_send", e.g. "cmd_len".
> 
> Right!
> 
> > > 
> > > I'd do instead:
> > > 
> > > if (!chip->send)
> > > 	goto out_recv;
> > > 
> > > And change recv into:
> > > 
> > > int (*recv)(struct tpm_chip *chip, u8 *buf, size_t buf_len,
> > > 	    cmd_len);
> > 
> > I think I went here over the top, and *if* we need a new callback
> > putting send_recv would be fine. Only thing I'd take from this is to
> > rename to_len as cmd_len.
> 
> Got it.
> 
> > 
> > However, I don't think there are strong enough reasons to add complexity
> > to the callback interface with the basis of this single driver. You
> > should deal with this internally inside the driver instead.
> > 
> > So do something along the lines of, e.g.:
> > 
> > 1. Create dummy send() copying the command to internal
> >   buffer.
> > 2. Create ->status() returning zero, and set req_complete_mask and
> >   req_complete_val to zero.
> > 3. Performan transaction in recv().
> > 
> > How you split send_recv() between send() and recv() is up to you. This
> > was merely showing that we don't need send_recv() desperately.
> 
> We did something similar in v1 [1], but instead of your point 2, we just set
> `chip->flags |= TPM_CHIP_FLAG_IRQ;` in the probe() after we allocated the
> chip.
> 
> Jason suggested the send_recv() ops [2], which I liked, but if you prefer to
> avoid that, I can restore what we did in v1 and replace the
> TPM_CHIP_FLAG_IRQ hack with your point 2 (or use TPM_CHIP_FLAG_IRQ if you
> think it is fine).
> 
> @Jarkko, @Jason, I don't have a strong preference about it, so your choice
> :-)

I'd say, unless you have actual identified blocker, please go with
a driver where the complexity is managed within the driver.

> 
> Thanks,
> Stefano
> 
> [1] https://lore.kernel.org/linux-integrity/20241210143423.101774-2-sgarzare@redhat.com/
> [2] https://lore.kernel.org/linux-integrity/CAGxU2F51EoqDqi6By6eBa7qT+VT006DJ9+V-PANQ6GQrwVWt_Q@mail.gmail.com/


BR, Jarkko