Date:   Sat, 8 Oct 2016 20:06:51 +0300
From:   Jarkko Sakkinen <jarkko.sakkinen@...ux.intel.com>
To:     "Winkler, Tomas" <tomas.winkler@...el.com>
Cc:     "tpmdd-devel@...ts.sourceforge.net" 
        <tpmdd-devel@...ts.sourceforge.net>,
        Jason Gunthorpe <jgunthorpe@...idianresearch.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 2/4] tmp/tpm_crb: fix Intel PTT hw bug during idle state

On Sat, Oct 08, 2016 at 04:56:39PM +0000, Winkler, Tomas wrote:
> 
> > 
> > On Sat, Oct 08, 2016 at 02:27:06PM +0000, Winkler, Tomas wrote:
> > > >
> > > > On Sat, Oct 08, 2016 at 02:59:37PM +0300, Tomas Winkler wrote:
> > > > > From: "Winkler, Tomas" <tomas.winkler@...el.com>
> > > > >
> > > > > There is a HW bug in Skylake, Kabylake, and Broxton PCH Intel PTT
> > > > > device, where most of the registers in the control area except
> > > > > START, REQUEST, CANCEL, and LOC_CTRL lost retention when the
> > > > > device is in the
> > > > idle state.
> > > > > Hence we need to bring the device to ready state before accessing
> > > > > the other registers. The fix brings device to ready state before
> > > > > trying to read command and response buffer addresses in order to
> > > > > remap the for
> > > > access.
> > > > >
> > > > > Signed-off-by: Tomas Winkler <tomas.winkler@...el.com>
> > > > > ---
> > > > > V2: cmd read need to be called also before crb_init as this will
> > > > > run self test.
> > > > > V3: resend.
> > > > > V4: add Kabylake to the list of effected platforms
> > > > >
> > > > >  drivers/char/tpm/tpm_crb.c | 47
> > > > > ++++++++++++++++++++++++++++++++++++++--------
> > > > >  1 file changed, 39 insertions(+), 8 deletions(-)
> > > > >
> > > > > diff --git a/drivers/char/tpm/tpm_crb.c
> > > > > b/drivers/char/tpm/tpm_crb.c index 0f3b3f3d12d3..4eb96b85c653
> > > > > 100644
> > > > > --- a/drivers/char/tpm/tpm_crb.c
> > > > > +++ b/drivers/char/tpm/tpm_crb.c
> > > > > @@ -319,6 +319,7 @@ static int crb_map_io(struct acpi_device
> > > > > *device,
> > > > struct crb_priv *priv,
> > > > >  	struct list_head resources;
> > > > >  	struct resource io_res;
> > > > >  	struct device *dev = &device->dev;
> > > > > +	u32 pa_high, pa_low;
> > > > >  	u64 cmd_pa;
> > > > >  	u32 cmd_size;
> > > > >  	u64 rsp_pa;
> > > > > @@ -346,12 +347,27 @@ static int crb_map_io(struct acpi_device
> > > > > *device,
> > > > struct crb_priv *priv,
> > > > >  	if (IS_ERR(priv->cca))
> > > > >  		return PTR_ERR(priv->cca);
> > > > >
> > > > > -	cmd_pa = ((u64) ioread32(&priv->cca->cmd_pa_high) << 32) |
> > > > > -		  (u64) ioread32(&priv->cca->cmd_pa_low);
> > > > > +	/*
> > > > > +	 * PTT HW bug w/a: wake up the device to access
> > > > > +	 * possibly not retained registers.
> > > > > +	 */
> > > > > +	ret = crb_cmd_ready(dev, priv);
> > > > > +	if (ret)
> > > > > +		return ret;
> > > > > +
> > > > > +	pa_high = ioread32(&priv->cca->cmd_pa_high);
> > > > > +	pa_low  = ioread32(&priv->cca->cmd_pa_low);
> > > > > +	cmd_pa = ((u64)pa_high << 32) | pa_low;
> > > > >  	cmd_size = ioread32(&priv->cca->cmd_size);
> > > > > +
> > > > > +	dev_dbg(dev, "cmd_hi = %X cmd_low = %X cmd_size %X\n",
> > > > > +		pa_high, pa_low, cmd_size);
> > > > > +
> > > > >  	priv->cmd = crb_map_res(dev, priv, &io_res, cmd_pa, cmd_size);
> > > > > -	if (IS_ERR(priv->cmd))
> > > > > -		return PTR_ERR(priv->cmd);
> > > > > +	if (IS_ERR(priv->cmd)) {
> > > > > +		ret = PTR_ERR(priv->cmd);
> > > > > +		goto out;
> > > > > +	}
> > > > >
> > > > >  	memcpy_fromio(&rsp_pa, &priv->cca->rsp_pa, 8);
> > > > >  	rsp_pa = le64_to_cpu(rsp_pa);
> > > > > @@ -359,7 +375,8 @@ static int crb_map_io(struct acpi_device
> > > > > *device, struct crb_priv *priv,
> > > > >
> > > > >  	if (cmd_pa != rsp_pa) {
> > > > >  		priv->rsp = crb_map_res(dev, priv, &io_res, rsp_pa, rsp_size);
> > > > > -		return PTR_ERR_OR_ZERO(priv->rsp);
> > > > > +		ret = PTR_ERR_OR_ZERO(priv->rsp);
> > > > > +		goto out;
> > > > >  	}
> > > > >
> > > > >  	/* According to the PTP specification, overlapping command and
> > > > > response @@ -367,12 +384,18 @@ static int crb_map_io(struct
> > > > > acpi_device
> > > > *device, struct crb_priv *priv,
> > > > >  	 */
> > > > >  	if (cmd_size != rsp_size) {
> > > > >  		dev_err(dev, FW_BUG "overlapping command and response
> > > > buffer sizes are not identical");
> > > > > -		return -EINVAL;
> > > > > +		ret = -EINVAL;
> > > > > +		goto out;
> > > > >  	}
> > > > > +
> > > > >  	priv->cmd_size = cmd_size;
> > > > >
> > > > >  	priv->rsp = priv->cmd;
> > > > > -	return 0;
> > > > > +
> > > > > +out:
> > > > > +	crb_go_idle(dev, priv);
> > > > > +
> > > > > +	return ret;
> > > > >  }
> > > > >
> > > > >  static int crb_acpi_add(struct acpi_device *device) @@ -416,7
> > > > > +439,15 @@ static int crb_acpi_add(struct acpi_device *device)
> > > > >  	if (rc)
> > > > >  		return rc;
> > > > >
> > > > > -	return crb_init(device, priv);
> > > > > +	rc  = crb_cmd_ready(dev, priv);
> > > > > +	if (rc)
> > > > > +		return rc;
> > > >
> > > > I cannot find any valid reason why crb_map_io calls crb_go_idle in
> > > > the except in the case of a failure. This is something I complained earlier.
> > >
> > > Haven't I've already explained that?  Each flow has to be enclosed by
> > > cmdReady and goIdle.  There is nothing different here from the other
> > > flows in that matter.  The assumption here is that we are starting in
> > > the idle state bug because of the HW bug we cannot access the
> > > registers. So the whole w/o is enclosed in the crb_map_io.  After that
> > > we are starting from new, from the idle state.
> > 
> > But why you have to do crb_go_idle upon returning from crb_map_io()?
> 
> As written above it's is closed code... cmdReady  <perform the task>   goIdle. 

It does extra cmdReady/goIdle. I don't really get your explanation
here. crb_map_io() is not a utility function used after the driver
has initialized. 

crb_map_io() could just leave it ready so you don't have to do an
explicit call to crb_cmd_ready() in crb_acpi_add. I think this is
quite obvious.

> > 
> > > > A minor thing but the extra crb_cmd_ready is basically clutter to
> > > > the initialization.
> > >
> > > The extra cmdReady is here in case runtime pm is not compiled into the
> > > kernel and assumption here is that we are starting in the idle state.
> > > Please remember that unfortunately we cannot detect whether we are in
> > > ready on idle state as the status register is not retained, so
> > > everything has to be ordered properly.
> > 
> > Rather you it would be better not to use cmdReady/goIdle at all if kernel is not
> > compiled with CONFIG_PM.
> 
> ????. You have to do the initial cmdReady, and that's what the code
> does, otherwise your driver won't work at all.  CONFIG_PM is the
> software thing, it doesn't magically match the HW power state. So you
> need to do cmdReady. 

In every machine that I've tried so far the PTT has worked if
cmdReady/goIdle haven't been used at all by the OS during a
power cycle.

> Thanks
> Tomas

I'll keep the current commit (v4 version) because it is tested
quite well, no worries about that, but this is something where a
cleanup would make sense later on.

/Jarkko

Powered by blists - more mailing lists