Hello,
On Mon, Jul 3, 2023 at 3:02 AM Samuel Thibault <samuel.thiba...@gnu.org> wrote:
> I applied this, *but* I see:
>
> * Fields in this structure are locked either by the lock on the
> * object that the page belongs to (O) or by the lock on the page
> * queues (P). [Some fields require that both locks be held to
> * change that field; holding either lock is sufficient to read.]
>
> That's not going to work at all with bitfields...
Great point. (But I don't think this patch makes it any worse.)
> We do need to sort this out by separating bitfields that are not
> protected the same way, otherwise it's no wonder if SMP fails oddly.
This is really interesting, because Mach surely was developed for SMP
from the start ("an OS for a parallel computer" was an explicit design
goal, and M in Mach stands for multi-processor), so it must have
worked for them. But indeed I see that they have separated the
bitfields in XNU, so we should too. Hopefully without blowing the
struct size back up.
Any feedback/comments on the rest of the series? Does forward entry
coalescing work for you? Any issues / noticeable performance effects?
Sergey
