On Thu, Nov 16, 2017 at 11:13:04AM +1000, David Gwynne wrote:
>
> > On 16 Nov 2017, at 7:23 am, Gregor Best <[email protected]> wrote:
> >
> > Hi,
> >
> > On Mon, Nov 13, 2017 at 01:47:01PM +1000, David Gwynne wrote:
> >> [...]
> >> pools maintain count of how many times they failed to provide an
> >> allocation. you can watch this with vmstat -m or systat pool.
> >> however, we could use that to populate mb_drops too.
> >> [...]
> >
> > That's certainly smarter than my idea.
>
> does it work though?
>
Now that I think about it, not quite, or at least not without reaching
around into the pools innards to lock them while reading the statistics
to get a consistent view.
The problem is that the MBSTAT_DROPS part of the `counters` data
structure never gets modified, so we'd still need to loop over all pools
in sysctl_mbstat, since this is not just about mbpool but also about the
pools for payload data.
On the other hand, going back to my initial proposal would mean
essentially duplicating functionality the pools already provide, so
that's at least a bit unelegant. Especially since it adds at least one
splnet/splx dance.
Another option would be to just say "screw it" and count the failed
allocations without acquiring any locks on the pools, maybe via atomic
operations. Sounds a bit too complicated though.
At the moment, and after a night's sleep, I think my initial proposal is
the most straightforward way to do this. That, or getting rid of the
confusing counters in `netstat -m` that stay at 0 all the time...
--
Gregor
