Re: [PATCH 0/1] ipc/util.{c,h}: Use binary search for max_idx

[Manfred Spraul]

Hi Davidlohr,

On 4/25/21 8:07 PM, Davidlohr Bueso wrote:
> On Sun, 25 Apr 2021, Manfred Spraul wrote:
>
> > 2nd version of the patch:
> > @Andrew: Could you add the patch to your mm tree, as candidate for
> > linux-next?
> >
> > Note:
> > I have tried to remove the ids->max_idx cache entirely. Unfortunately,
> > this causes a significant slow-down of semstat(,,IPC_STAT):
> >   * no object allocated, no ipcmni_extended: +50%
> >   * no object allocated, with ipcmni_extended: +80%
> >   * 30 objects allocated, with large gaps, no ipcmni_extended:
> >           +350%
> > Thus I haven't removed ids->max_id.
>
> Right, IPC_STAT is the main usecase for max_id. But I'm not sure why
> you were looking to remove it in the first place - or was it just to
> avoid this patch altogether?
I had assumed that after removing the linear search, the lookup would be 
so fast that the cache can be removed entirely.
It would save ~20 lines of code and one int in struct ipc_ids (12 bytes 
per namespace?).

But: My assumption was wrong, the slowdown is too large.

--

    Manfred