[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAAH8bW-jO=BEY3j5JVJcuVsCeUzWsgNqfg-b7EHt-w00vjP+gQ@mail.gmail.com>
Date: Tue, 29 Aug 2023 06:50:08 -0700
From: Yury Norov <yury.norov@...il.com>
To: Andy Shevchenko <andriy.shevchenko@...ux.intel.com>
Cc: Rasmus Villemoes <linux@...musvillemoes.dk>,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH 00/12] bitmap: rework bitmap_{bit,}remap()
On Tue, Aug 29, 2023 at 6:39 AM Andy Shevchenko
<andriy.shevchenko@...ux.intel.com> wrote:
>
> On Tue, Aug 29, 2023 at 09:33:29AM +0200, Rasmus Villemoes wrote:
> > On 28/08/2023 20.43, Yury Norov wrote:
> > > This series adds a test, const-time optimizaton and fixes O(N^2)
> > > complexity problem for the functions. It's based on discussion in
> > > bitmap: optimize bitmap_remap() series [1], but there's much more work
> > > here, so I decided to give it a separete run, and don't name it as v2.
> > >
> > > bitmap_remap() API has just one user in generic code, and few more in
> > > drivers, so this may look like an overkill. But the work gives ~10x
> > > better performance for a 1000-bit bitmaps, which is typical for nodemasks
> > > in major distros like Ubuntu.
> >
> > Can you find just _one_ project on Debian Code Search or elsewhere that
> > actually uses mbind(2), that could possibly ever trigger the use of that
> > bitmap_remap stuff? Also, the bitmap may be order 10, but that's just
> > because the kitchen sink distros are silly, real machines have nowhere
> > near that number of nodes, so even if mbind is used, the bitmaps
> > involved will never actually have anything beyond index ~64.
> >
> > I think this is all total overkill for non-existing problems, and when
> > it takes me 20 seconds to find the first bug, I really don't think it's
> > worth the churn. I'm not giving a thorough review on the rest of the
> > series, nor commenting on followups.
>
> I posted one patch to replace these APIs with something else, more particular
> for GPIO case(s). Have you chance to look at that? With that taking in, I'm
> fully agree on the above statement (as we lose the user of this complicated
> thingy which is a niche of the NUMA as you mentioned already).
I saw the code in a comment in some thread but not as a separate patch, and
AFAIK you said it's a work-in-progress.
Sorry if I missed your submission. Can you please send the patch or point me to
the previous submission?
Also, if after your change bitmap_remap would become a NUMA-specific, should
protect it with NUMA guards?
Thanks,
Yury
Powered by blists - more mailing lists