Re: [PATCH v4 3/4] tpm: add SNP SVSM vTPM driver

[Jarkko Sakkinen <jarkko@kernel.org>]

On Thu, Mar 27, 2025 at 11:03:07AM +0100, Stefano Garzarella wrote:
> On Wed, Mar 26, 2025 at 09:30:53PM +0200, Jarkko Sakkinen wrote:
> > On Mon, Mar 24, 2025 at 11:46:48AM +0100, Stefano Garzarella wrote:
> > > From: Stefano Garzarella <sgarzare@redhat.com>
> > > 
> > > Add driver for the vTPM defined by the AMD SVSM spec [1].
> > > 
> > > The specification defines a protocol that a SEV-SNP guest OS can use to
> > > discover and talk to a vTPM emulated by the Secure VM Service Module (SVSM)
> > > in the guest context, but at a more privileged level (VMPL0).
> > > 
> > > The new tpm-svsm platform driver uses two functions exposed by x86/sev
> > > to verify that the device is actually emulated by the platform and to
> > > send commands and receive responses.
> > > 
> > > The device cannot be hot-plugged/unplugged as it is emulated by the
> > > platform, so we can use module_platform_driver_probe(). The probe
> > > function will only check whether in the current runtime configuration,
> > > SVSM is present and provides a vTPM.
> > > 
> > > This device does not support interrupts and sends responses to commands
> > > synchronously. In order to have .recv() called just after .send() in
> > > tpm_try_transmit(), the .status() callback returns 0, and both
> > > .req_complete_mask and .req_complete_val are set to 0.
> > > 
> > > [1] "Secure VM Service Module for SEV-SNP Guests"
> > >     Publication # 58019 Revision: 1.00
> > > 
> > > Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
> > > ---
> > > v4:
> > > - moved "asm" includes after the "linux" includes [Tom]
> > > - allocated buffer separately [Tom/Jarkko/Jason]
> > > v3:
> > > - removed send_recv() ops and followed the ftpm driver implementing .status,
> > >   .req_complete_mask, .req_complete_val, etc. [Jarkko]
> > > - removed link to the spec because those URLs are unstable [Borislav]
> > > ---
> > >  drivers/char/tpm/tpm_svsm.c | 155 ++++++++++++++++++++++++++++++++++++
> > >  drivers/char/tpm/Kconfig    |  10 +++
> > >  drivers/char/tpm/Makefile   |   1 +
> > >  3 files changed, 166 insertions(+)
> > >  create mode 100644 drivers/char/tpm/tpm_svsm.c
> > > 
> > > diff --git a/drivers/char/tpm/tpm_svsm.c b/drivers/char/tpm/tpm_svsm.c
> > > new file mode 100644
> > > index 000000000000..1281ff265927
> > > --- /dev/null
> > > +++ b/drivers/char/tpm/tpm_svsm.c
> > > @@ -0,0 +1,155 @@
> > > +// SPDX-License-Identifier: GPL-2.0-only
> > > +/*
> > > + * Copyright (C) 2025 Red Hat, Inc. All Rights Reserved.
> > > + *
> > > + * Driver for the vTPM defined by the AMD SVSM spec [1].
> > > + *
> > > + * The specification defines a protocol that a SEV-SNP guest OS can use to
> > > + * discover and talk to a vTPM emulated by the Secure VM Service Module (SVSM)
> > > + * in the guest context, but at a more privileged level (usually VMPL0).
> > > + *
> > > + * [1] "Secure VM Service Module for SEV-SNP Guests"
> > > + *     Publication # 58019 Revision: 1.00
> > > + */
> > > +
> > > +#include <linux/module.h>
> > > +#include <linux/kernel.h>
> > > +#include <linux/platform_device.h>
> > > +#include <linux/tpm_svsm.h>
> > > +
> > > +#include <asm/sev.h>
> > > +
> > > +#include "tpm.h"
> > > +
> > > +struct tpm_svsm_priv {
> > > +	void *buffer;
> > > +	u8 locality;
> > > +};
> > > +
> > > +static int tpm_svsm_send(struct tpm_chip *chip, u8 *buf, size_t len)
> > > +{
> > > +	struct tpm_svsm_priv *priv = dev_get_drvdata(&chip->dev);
> > > +	int ret;
> > > +
> > > +	ret = svsm_vtpm_cmd_request_fill(priv->buffer, priv->locality, buf, len);
> > > +	if (ret)
> > > +		return ret;
> > > +
> > > +	/*
> > > +	 * The SVSM call uses the same buffer for the command and for the
> > > +	 * response, so after this call, the buffer will contain the response
> > > +	 * that can be used by .recv() op.
> > > +	 */
> > > +	return snp_svsm_vtpm_send_command(priv->buffer);
> > > +}
> > > +
> > > +static int tpm_svsm_recv(struct tpm_chip *chip, u8 *buf, size_t len)
> > > +{
> > > +	struct tpm_svsm_priv *priv = dev_get_drvdata(&chip->dev);
> > > +
> > > +	/*
> > > +	 * The internal buffer contains the response after we send the command
> > > +	 * to SVSM.
> > > +	 */
> > > +	return svsm_vtpm_cmd_response_parse(priv->buffer, buf, len);
> > > +}
> > > +
> > > +static void tpm_svsm_cancel(struct tpm_chip *chip)
> > > +{
> > > +	/* not supported */
> > > +}
> > > +
> > > +static u8 tpm_svsm_status(struct tpm_chip *chip)
> > > +{
> > > +	return 0;
> > > +}
> > > +
> > > +static bool tpm_svsm_req_canceled(struct tpm_chip *chip, u8 status)
> > > +{
> > > +	return false;
> > > +}
> > > +
> > > +static struct tpm_class_ops tpm_chip_ops = {
> > > +	.flags = TPM_OPS_AUTO_STARTUP,
> > > +	.recv = tpm_svsm_recv,
> > > +	.send = tpm_svsm_send,
> > > +	.cancel = tpm_svsm_cancel,
> > > +	.status = tpm_svsm_status,
> > > +	.req_complete_mask = 0,
> > > +	.req_complete_val = 0,
> > > +	.req_canceled = tpm_svsm_req_canceled,
> > 
> > If this was bundled with the patch set, this would short a lot:
> > 
> > https://lore.kernel.org/linux-integrity/20250326161838.123606-1-jarkko@kernel.org/T/#u
> > 
> > So maybe for v5? Including this patch does not take send_recv()
> > out of consideration, it is just smart thing to do in all cases.
> > 
> > It would be probably easiest to roll out my patch together with
> > rest of the patch set.
> 
> Yeah, I agree. I'll include it in this series and adapt this patch on top of
> it.

Yeah, and you could simplify to goal in the other patch set: it's about
avoiding double-copy of a buffer.

It's a totally legit argument that we can measure. So in the end this
will help out landing that too because it takes away the extra cruft
and streamlines the goal.

> 
> Thanks,
> Stefano
> 
> 

BR, Jarkko