From: Peter Dimov (pdimov_at_[hidden])
Date: 2023-06-02 00:35:02
Gavin Lambert wrote:
> On 2/06/2023 11:43, Peter Dimov wrote:
> > Gavin Lambert wrote:
> >> FWIW though, atomic_* on shared_ptr is entirely banned in my codebase
> >> precisely because of its spinlock implementation.
> >
> > Which part is wrong?
>
> Caused massive performance issues in my codebase. Not really a fault of
> your implementation per se (other than it's a spinlock at all), just a
> consequence of priority inversion plus too many collisions in the
> spinlock pool. Probably not something that would be a problem for most
> apps; I just have some special quirks.
Priority inversion shouldn't be an issue because the SmartPtr spinlock
now goes to sp_thread_sleep almost immediately (after just two tries.)
The spinlock pool size should probably be increased. 41 is only good for
about six cores and this isn't enough in 2023.