Re: REGRESSION WITH BISECT: v6.5-rc6 TPM patch breaks S3 on some Intel systems

From: Todd Brandt
Date: Fri Aug 18 2023 - 16:09:01 EST


On Fri, 2023-08-18 at 13:11 -0500, Mario Limonciello wrote:
> On 8/18/2023 13:07, Jarkko Sakkinen wrote:
> > On Fri Aug 18, 2023 at 8:57 PM EEST, Mario Limonciello wrote:
> > > On 8/18/2023 12:53, Jarkko Sakkinen wrote:
> > > > On Fri Aug 18, 2023 at 8:21 PM EEST, Mario Limonciello wrote:
> > > > > On 8/18/2023 12:00, Jarkko Sakkinen wrote:
> > > > > > On Fri Aug 18, 2023 at 4:58 AM EEST, Limonciello, Mario
> > > > > > wrote:
> > > > > > >
> > > > > > >
> > > > > > > On 8/17/2023 5:33 PM, Jarkko Sakkinen wrote:
> > > > > > > > On Fri Aug 18, 2023 at 1:25 AM EEST, Todd Brandt wrote:
> > > > > > > > > On Fri, 2023-08-18 at 00:47 +0300, Jarkko Sakkinen
> > > > > > > > > wrote:
> > > > > > > > > > On Fri Aug 18, 2023 at 12:09 AM EEST, Todd Brandt
> > > > > > > > > > wrote:
> > > > > > > > > > > While testing S3 on 6.5.0-rc6 we've found that 5
> > > > > > > > > > > systems are seeing
> > > > > > > > > > > a
> > > > > > > > > > > crash and reboot situation when S3 suspend is
> > > > > > > > > > > initiated. To
> > > > > > > > > > > reproduce
> > > > > > > > > > > it, this call is all that's required "sudo
> > > > > > > > > > > sleepgraph -m mem
> > > > > > > > > > > -rtcwake
> > > > > > > > > > > 15".
> > > > > > > > > >
> > > > > > > > > > 1. Are there logs available?
> > > > > > > > > > 2. Is this the test case:
> > > > > > > > > > https://pypi.org/project/sleepgraph/ (never
> > > > > > > > > > used it before).
> > > > > > > > >
> > > > > > > > > There are no dmesg logs because the S3 crash wipes
> > > > > > > > > them out. Sleepgraph
> > > > > > > > > isn't actually necessary to activate it, just an S3
> > > > > > > > > suspend "echo mem >
> > > > > > > > > /sys/power/state".
> > > > > > > > >
> > > > > > > > > So far it appears to only have affected test systems,
> > > > > > > > > not production
> > > > > > > > > hardware, and none of them have TPM chips, so I'm
> > > > > > > > > beginning to wonder
> > > > > > > > > if this patch just inadvertently activated a bug
> > > > > > > > > somewhere else in the
> > > > > > > > > kernel that happens to affect test hardware.
> > > > > > > > >
> > > > > > > > > I'll continue to debug it, this isn't an emergency as
> > > > > > > > > so far I haven't
> > > > > > > > > seen it in production hardware.
> > > > > > > >
> > > > > > > > OK, I'll still see if I could reproduce it just in
> > > > > > > > case.
> > > > > > > >
> > > > > > > > BR, Jarkko
> > > > > > >
> > > > > > > I'd like to better understand what kind of TPM
> > > > > > > initialization path has
> > > > > > > run. Does the machine have some sort of TPM that failed
> > > > > > > to fully
> > > > > > > initialize perhaps?
> > > > > > >
> > > > > > > If you can't share a full bootup dmesg, can you at least
> > > > > > > share
> > > > > > >
> > > > > > > # dmesg | grep -i tpm
> > > > > >
> > > > > > It would be more useful perhaps to get full dmesg output
> > > > > > after power on
> > > > > > and before going into suspend.
> > > > > >
> > > > > > Also ftrace filter could be added to the kernel command-
> > > > > > line:
> > > > > >
> > > > > > ftrace=function ftrace_filter=tpm*
> > > > > >
> > > > > > After bootup:
> > > > > >
> > > > > > mount -t tracefs nodev /sys/kernel/tracing
> > > > > > cat /sys/kernel/tracing/trace
> > > > > >
> > > > > > BR, Jarkko
> > > > >
> > > > > Todd and I have gone back and forth a little bit on the
> > > > > bugzilla
> > > > > (https://bugzilla.kernel.org/show_bug.cgi?id=217804), and it
> > > > > seems that
> > > > > this isn't an S3 problem - it's a probing problem.
> > > > >
> > > > > [ 1.132521] tpm_crb: probe of INTC6001:00 failed with
> > > > > error 378
> > > > >
> > > > > That error 378 specifically matches TPM2_CC_GET_CAPABILITY,
> > > > > which is the
> > > > > same command that was being requested. This leads me to
> > > > > believe the TPM
> > > > > isn't ready at the time of probing.
> > > > >
> > > > > In this case one solution is we could potentially ignore
> > > > > failures for
> > > > > that tpm2_get_tpm_pt() call, but I think we should first
> > > > > understand why
> > > > > it doesn't work at probing time for this TPM to ensure the
> > > > > actual quirk
> > > > > isn't built on a house of cards.
> > > >
> > > > Given that there is nothing known broken (at the moment) in
> > > > production,
> > > > I think the following might be a reasonable change.
> > > >
> > > > BR, Jarkko
> > > >
> > >
> > > Yeah that would prevent it.
> > >
> > > Here's a simpler change that I think should work too though:
> > > diff --git a/drivers/char/tpm/tpm_crb.c
> > > b/drivers/char/tpm/tpm_crb.c
> > > index 9eb1a18590123..b0e9931fe436c 100644
> > > --- a/drivers/char/tpm/tpm_crb.c
> > > +++ b/drivers/char/tpm/tpm_crb.c
> > > @@ -472,8 +472,7 @@ static int crb_check_flags(struct tpm_chip
> > > *chip)
> > > if (ret)
> > > return ret;
> > >
> > > - ret = tpm2_get_tpm_pt(chip, TPM2_PT_MANUFACTURER, &val,
> > > NULL);
> > > - if (ret)
> > > + if (tpm2_get_tpm_pt(chip, TPM2_PT_MANUFACTURER, &val,
> > > NULL))
> > > goto release;
> > >
> > > if (val == 0x414D4400U /* AMD */)
> > >
> > > I think Todd needs to check whether TPM works with that or not
> > > though.
> >
> > Hmm... I'm sorry if I have a blind spot now but what is that
> > changing?
> >
> > BR, Jarkko
>
> It throws away the error code if it fails for some reason.
> Todd just checked it works too. I'll drop it on the M/L for review.

Yea, instead of returning the error code, it just goes to "release"
with the ret value uninitialized.

So the issue is apparently that Intel TPM sometimes does not properly
returning the manufacturer code in early boot. This patch just
inadvertently triggered the real issue. We'll investigate this to see
if there's a solution, but for now this patch works great.

Thanks Mario and Jarkko!