Re: [PATCH RFC 1/4] tpm: migrate struct tpm_buf to struct tpm_chip

From: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
To: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Cc: tpmdd-devel@lists.sourceforge.net,
	linux-security-module@vger.kernel.org,
	Peter Huewe <peterhuewe@gmx.de>,
	Marcel Selhorst <tpmdd@selhorst.net>,
	open list <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH RFC 1/4] tpm: migrate struct tpm_buf to struct tpm_chip
Date: Tue, 3 Jan 2017 12:13:28 -0700	[thread overview]
Message-ID: <20170103191328.GB26706@obsidianresearch.com> (raw)
In-Reply-To: <20170103005737.t2qrc32xzdnvqy4b@intel.com>

On Tue, Jan 03, 2017 at 02:57:37AM +0200, Jarkko Sakkinen wrote:
> On Mon, Jan 02, 2017 at 02:01:01PM -0700, Jason Gunthorpe wrote:
> > On Mon, Jan 02, 2017 at 03:22:07PM +0200, Jarkko Sakkinen wrote:
> > > Since there is only one thread using TPM chip at a time to transmit data
> > > we can migrate struct tpm_buf to struct tpm_chip. This makes the use of
> > > it more fail safe as the buffer is allocated from heap when the device
> > > is created and not for every transaction.
> > 
> > Eh? What? I don't think that is the case..
> > 
> > We don't serialize until we hit tramsit_cmd at which point the buffer
> > is already being used and cannot be shared between threads.
> 
> There is a regression in the patch. All functions that use 'tr_buf'
> should take tpm_mutex first and use TPM_TRANSMIT_UNLOCKED. There's
> also a similar regression in TPM space patch that I have to correct.

No, you can't steal TPM_TRANSMIT_UNLOCKED and tpm_mutex for this, that
is to allow a chain of commands to execute atomicly, so a new lock is
needed just for the tr_buf.

> > Why would the resource manager need a single global tpm buffer? That
> > seems like a big regression from where we have been going. I don't
> > think this is a good idea to go down this road.
> 
> What? 'tr_buf' is not specifically for resource manager. This commit
> makes creating TPM commands more fail-safe because there is no need
> to allocate page for every transmit.

That doesn't seem all that important, honestly. There kernel does not
fail single page allocations without a lot of duress.

> For RM decorations this is really important because I rather would have
> them fail as rarely as possible. If this would become a scalability
> issue then the granularity could be reconsidered.

Why? The RM design already seems to have the prepare/commit/abort
kind of model so it can already fail. What does it matter if the
caller can fail before getting that far?

It seems like alot of dangerous churn to introduce a new locking model
without a really good reason...

Jason