Re: [PATCH] tpm: Revert "tpm: fix invalid locking in NONBLOCKING mode"

[Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>]

On Tue, 2020-05-26 at 12:38 -0700, James Bottomley wrote:
> On Tue, 2020-05-26 at 19:23 +0000, Mario.Limonciello@dell.com wrote:
> > > On Tue, 2020-05-26 at 13:32 -0500, Mario Limonciello wrote:
> > > > This reverts commit d23d12484307b40eea549b8a858f5fffad913897.
> > > > 
> > > > This commit has caused regressions for the XPS 9560 containing
> > > > a Nuvoton TPM.
> > > 
> > > Presumably this is using the tis driver?
> > 
> > Correct.
> > 
> > > > As mentioned by the reporter all TPM2 commands are failing with:
> > > >   ERROR:tcti:src/tss2-tcti/tcti-
> > > > device.c:290:tcti_device_receive()
> > > >   Failed to read response from fd 3, got errno 1: Operation not
> > > > permitted
> > > > 
> > > > The reporter bisected this issue back to this commit which was
> > > > backported to stable as commit 4d6ebc4.
> > > 
> > > I think the problem is request_locality ... for some inexplicable
> > > reason a failure there returns -1, which is EPERM to user space.
> > > 
> > > That seems to be a bug in the async code since everything else
> > > gives a ESPIPE error if tpm_try_get_ops fails ... at least no-one
> > > assumes it gives back a sensible return code.
> > > 
> > > What I think is happening is that with the patch the TPM goes
> > > through a quick sequence of request, relinquish, request,
> > > relinquish and it's the third request which is failing (likely
> > > timing out).  Without the patch, the patch there's only one
> > > request,relinquish cycle because the ops are held while the async
> > > work is executed.  I have a vague recollection that there is a
> > > problem with too many locality request in quick succession, but
> > > I'll defer to Jason, who I think understands the intricacies of
> > > localities better than I do.
> > 
> > Thanks, I don't pretend to understand the nuances of this particular
> > code, but I was hoping that the request to revert got some attention
> > since Alex's kernel Bugzilla and message a few months ago to linux
> > integrity weren't.
> > 
> > > If that's the problem, the solution looks simple enough: just move
> > > the ops get down because the priv state is already protected by the
> > > buffer mutex
> > 
> > Yeah, if that works for Alex's situation it certainly sounds like a
> > better solution than reverting this patch as this patch actually does
> > fix a problem reported by Jeffrin originally.
> > 
> > Could you propose a specific patch that Alex and Jeffrin can perhaps
> > both try?
> 
> Um, what's wrong with the one I originally attached and which you quote
> below?  It's only compile tested, but I think it will work, if the
> theory is correct.

Please send a legit patch, thanks.

/Jarkko