Linus Torvalds (torvalds.delete@this.linux-foundation.org) on April 9, 2021 3:03 pm wrote:
>
> Honestly, back when Intel was the undisputed price/performance leader, they
> could afford to play games, knowing that they had the best game in town.
>
> That's not true any more, and Intel had better get their act together.

Side note: Intel has lost of lot of goodwill among kernel engineers over the last several years. It's not just that they aren't the performance leader any more, it's also a matter of "you've screwed us over too many times and it's not been a good experience".

Of course, most of it has been about Meltdown and friends, but TSX has been another of those "this just causes us pain for no gain" thing.

Even when the kernel itself doesn't use TSX, some of the errata ended up having microcode workarounds with things like clobbering one of the performance counters. So on some hardware, you have the choice of "allow TSX in user space, lose one performance counter" or "force TSX to always fail, get access to all the performance counters". And that's something that the kernel has to then deal with.

That's independent of how people also found a use for TSX as a way to make some of the CPU information leaks easier, and the workarounds for that.

So TSX bugs caused lots of pain for the user space that tried to use it (see various complaints about that in this long thread), but it's also been painful for the kernel even when the kernel didn't use it.

Guess why people are just so unenthusiastic about the whole pile of steaming garbage?

Linus