Re: [PATCH] block: switch to atomic_t for request references

On 12/6/21 1:31 AM, Peter Zijlstra wrote:

On Sun, Dec 05, 2021 at 10:53:49PM -0800, Christoph Hellwig wrote:

quoted

On Fri, Dec 03, 2021 at 08:35:40AM -0700, Jens Axboe wrote:

quoted

refcount_t is not as expensive as it used to be, but it's still more
expensive than the io_uring method of using atomic_t and just checking
for potential over/underflow.

This borrows that same implementation, which in turn is based on the
mm implementation from Linus.

If refcount_t isn't good enough for a normal kernel fast path we have
a problem.  Can we discuss that with the maintainers instead of coming
up with our home grown schemes again?

Quite; and for something that pretends to be about performance, it also
lacks any actual numbers to back that claim.

I can certainly generate that, it was already done for the two previous
similar conversions though.

The proposed implementation also doesn't do nearly as much as the
refcount_t one does.

Anyway refcount_t is just a single "lock xadd" and a few branches, where
does it go wrong? Do you have perf output to compare between them?

I'll generate that.

-- 
Jens Axboe