On 01/04/2018 06:41 PM, Alexander.Steffen@xxxxxxxxxxxx wrote:
On Fri, Dec 15, 2017 at 02:04:39PM +0200, Jarkko Sakkinen wrote:
On Fri, Dec 08, 2017 at 07:46:49PM +0100, Alexander Steffen wrote:
Disclaimer: This is RFC because unfortunately I do not have the time to
bring it up to the usual standards. But from lying around on my hard disk it
definitely won't improve, so I've decided to publish it now, to get at least
some feedback or maybe for somebody else to pick it up. Since the work
began some time ago, it is probably not rebased against the current state
(but hopefully still applies).
This is a collection of all the fixes I made during the investigation of the
problems with the "ignore burstcount" change. These are mainly fixes for the
wait state handling in tpm_tis_spi, that probably was not really tested so far,
since the previous implementation avoided wait states as far as possible. It
also includes changes to tpm_tis_core, so that it follows the specification
more closely. Finally, I've included a rebased version of the original "ignore
burstcount" patch. Together, those patches pass all my tests and also each
intermediate step is fine (so that bisecting still works).
Alexander Steffen (8):
tpm_tis_core: clean up whitespace
tpm_tis_core: access single TIS registers before doing complex
transfers
tpm_tis_core: correctly wait for flags to become zero
tpm_tis_core: send all data in single operation
tpm_tis_core: use XDATA_FIFO for transfers if available
tpm_tis_spi: fix sending wrong data during wait state handling
tpm_tis_spi: release CS line when wait state handling fails
tpm_tis_spi: add delay between wait state retries
Nayna Jain (1):
tpm: ignore burstcount to improve tpm_tis send() performance
drivers/char/tpm/tpm_tis_core.c | 131 ++++++++++++++++++++---------
-----------
drivers/char/tpm/tpm_tis_core.h | 26 ++++----
drivers/char/tpm/tpm_tis_spi.c | 21 +++++--
3 files changed, 95 insertions(+), 83 deletions(-)
--
2.7.4
Might not have time to review this before holidays but will do if
I can.
Unfortunately this and your patches for broken TPMs came so late that
they have no chance to make into 4.16. Thus, I will postpone their
review after the pull request and take care of the high priority stuff
for that release.
As long the "ignore burstcount" patch does not make it into 4.16 either, that's fine. Then nothing new gets broken, it just does not get any faster.
Alex, if I am right, among following 4 patches,
Patch 1 - tpm: move wait_for_tpm_stat() to respective driver files
Patch 2 - tpm: ignore burstcount to improve tpm_tis send() performance
Patch 3 - tpm: reduce tpm polling delay in tpm_tis_core
Patch 4 - tpm: use tpm_msleep() value as max delay
it is only the burstcount patch (Patch 2) which caused the issue.
With that, Jarkko, Alex, can we just exclude this and have other three patches pulled for 4.16 ?
I tested with these other three patches, and performance improvement is from ~41 to ~14.
Thanks & Regards,
- Nayna
Alexander
