Re: [PATCH v6 00/10] Intel Cache Allocation Technology

[Thomas Gleixner]

Gentlemen!

After more than two years of tinkering and real engineering, we finaly have
skinned the CAT!

That was the most amazing review journey I ever made as a maintainer. Just
a few statistics:

Design variants: 6

  6 different approaches for a user space interface. 3 of them have been
  actually implemented.

  Unfortunately the real interface discussion happened way after the first
  rounds of patches had been sent and reviewed. See below.

Patchsets:	21

  21 patch sets were posted. These can be split into two generations.

  Gen1	   	16   Oct 2014 - Dec 2015

  Gen2		5    Jul 2016 - Oct 2016

LKML-Mails:	1216

  That's the number of mails related to this project sent to LKML,
  according to my archive. About 1/3 of those mails are the postings of
  the patchsets alone.

  I cannot tell how many offlist mails have been sent around in total on
  this matter, but at least in my personal mail are close to hundred.

Beers: 	       Uncountable

  This applies to both the number of beers consumed and the number of beers
  owed.

I'm pretty happy with the final outcome of these patches and I want to say
thanks to everyone!

I know that I've been a pain in the neck for some of you due to my
pedantery about the details, but getting this wrong would have been a major
disaster. If I offended someone personally in course of the sometimes
heated discussions, then I offer my excuses.

Some lessons can be learned from this endeavour:

   1) Chip vendors should give access to the full documentation early

   2) Reviewers should never trust patch submitters, that they have read
      the documentation correctly and came to the right conclusions how to
      handle such a facility.

   3) User space interface discussions should be done upfront with a full
      explanation of the inner workings of such a facility and full
      documentation available.

Anything else is just the usual churn of patch submissions, which are
handled by the submitters with different effectiveness levels.

That said, all which needs to be done now is proper testing and a massive
exposure of the user space interface to fuzzers. I've implemented my share
of string parsers in the past and as careful as I was, there was always a
hole in them.

If any of the involved folks are at KS/LPC then I suggest we get together
at a bar during the week and drown the skinned CAT with the appropriate
beverages. The first round of drinks is my shout.

Thanks,

	tglx