Re: [PATCH] tpm: introduce struct tpm_buf

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

On Wed, Jun 03, 2015 at 10:11:28AM -0600, Jason Gunthorpe wrote:
> On Wed, Jun 03, 2015 at 03:23:31PM +0300, Jarkko Sakkinen wrote:
> > On Tue, Jun 02, 2015 at 12:13:15PM -0600, Jason Gunthorpe wrote:
> > > On Tue, Jun 02, 2015 at 04:04:22PM +0300, Jarkko Sakkinen wrote:
> > > > +/* A string buffer type for constructing TPM commands. This is based on the
> > > > + * code in security/keys/trusted.h.
> > > > + */
> > > > +
> > > > +#define TPM_BUF_SIZE 512
> > > > +
> > > > +struct tpm_buf {
> > > > +	u8 data[TPM_BUF_SIZE];
> > > This should be u32 or u64 to guarentee correct alignment for the
> > > casting.
> > 
> > Good catch. The functions where this might cause trouble are *_length()
> > and *_tag().
> > 
> > In other places misalignment should not cause any regressions since data
> > is not directly assigned to the buffer with pointer casting.
> > 
> > I would prefer to fix by changing *_length() and *_tag() to copy the
> > value to a local variable and return that. It's a fail safe way and here
> > the performance is not an issue.
> 
> I would change the type, that is very simple and will improve
> performance of the memcpy as well.
> 
> It actually isn't a problem for the casts to tpm_input_header - any
> structure marked __packed will cause the compiler to assume that the
> entire structure is unaligned and code gen accordingly, so the cast
> will always work, but on x86 it will be more efficient if the array
> is aligned.
> 
> Ideally we could change to something like:
>  __attribute__((packed,aligned(4))) 
> 
> Instead of __packed which will help the compiler minimize unaligned
> load instructions..

I realized basically the same what you said here when I refined the
patch (read this email after sending v2) :) I decided that align to 8
bytes.

I don't think the packed attribute is needed here because now I cast
the beginning to tpm_input_header, which is a packed struct.

> Jason

/Jarkko