On Mon, Aug 07, 2017 at 09:59:35AM -0500, Julia Cartwright wrote:
> On Fri, Aug 04, 2017 at 04:56:51PM -0500, Haris Okanovic wrote:
> > I have a latency issue using a SPI-based TPM chip with tpm_tis driver
> > from non-rt usermode application, which induces ~400 us latency spikes
> > in cyclictest (Intel Atom E3940 system, PREEMPT_RT_FULL kernel).
> > The spikes are caused by a stalling ioread8() operation, following a
> > sequence of 30+ iowrite8()s to the same address. I believe this happens
> > because the writes are cached (in cpu or somewhere along the bus), which
> > gets flushed on the first LOAD instruction (ioread*()) that follows.
> To use the ARM parlance, these accesses aren't "cached" (which would
> imply that a result could be returned to the load from any intermediate
> node in the interconnect), but instead are "bufferable".
> It is really unfortunate that we continue to run into this class of
> problem across various CPU vendors and various underlying bus
> technologies; it's the continuing curse of running an PREEMPT_RT on
> commodity hardware. RT is not easy :)
> > The enclosed change appears to fix this issue: read the TPM chip's
> > access register (status code) after every iowrite*() operation.
> Are we engaged in a game of wack-a-mole with all of the drivers which
> use this same access pattern (of which I imagine there are quite a
> I'm wondering if we should explore the idea of adding a load in the
> iowriteN()/writeX() macros (marking those accesses in which reads cause
> side effects explicitly, redirecting to a _raw() variant or something).
> Obviously that would be expensive for non-RT use cases, but for helping
> constrain latency, it may be worth it for RT.
What if we as quick resort we add tpm_tis_iowrite8() to the TPM driver.
Would be easy to move to iowrite8() if the problem is sorted out there