Re: Is it possible to employ epoch to simplify managing prison lifecycle

> On 23 Dec 2022, at 15:27, Mateusz Guzik <mjguzik@gmail.com> wrote:
> 
> On 12/23/22, Alexander V. Chernikov <melifaro@ipfw.ru> wrote:
>> 
>> 
>>> On 16 Dec 2022, at 16:29, Mateusz Guzik <mjguzik@gmail.com> wrote:
>>> 
>>> On 12/16/22, Zhenlei Huang <zlei.huang@gmail.com> wrote:
>>>> Hi,
>>>> 
>>>> While hacking `sys/kern/kern_jail.c` I got lost.
>>>> 
>>>> There're lots of ref / unref and flags to prevent visit invalid prison
>>>> while
>>>> concurrent modification is possible and some refs looks weird.
>>>> 
>>>> Is it possible to employ epoch(9) to simplify managing of prison
>>>> lifecycle
>>>> ?
>>>> 
>>> 
>>> Some of the ref/unref cycles are probably avoidable to begin with, but
>>> ultimately the thing to do here is to employ per-cpu reference
>>> counting, if at all needed.
>>> 
>>> I have a wip patch to provide such a mechanism, it may or may not land
>>> this month.
>> That would be nice. I’d love to convert nextops refcounting to that one.
>> Do you envision similar semantics as Linux percpu_ref? I mean, does one need
>> to explicitly mark “not in active use” stage?
> 
> There *something* needed to disable per-cpu operation, otherwise how
> can you ever know if the count is 0, apart from going over all cpus
> every time, which defeats the point.
Ack, sounds reasonable. Happy to test the KPI once it’s available.
> 
> More specifically, I have a on/off switch for said per-cpu op. This is
> modeled after what I did for counters in vfs, see vfs_ref et al.
> 
> -- 
> Mateusz Guzik <mjguzik gmail.com>
>