[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <3cb4c839-caa4-5c6e-1672-6498124d3930@suse.cz>
Date: Mon, 27 Nov 2023 12:52:27 +0100
From: Vlastimil Babka <vbabka@...e.cz>
To: Hyeonggon Yoo <42.hyeyoo@...il.com>
Cc: sxwjean@...com, linux-mm@...ck.org, cl@...ux.com,
penberg@...nel.org, rientjes@...gle.com, iamjoonsoo.kim@....com,
akpm@...ux-foundation.org, roman.gushchin@...ux.dev,
corbet@....net, linux-doc@...r.kernel.org,
linux-kernel@...r.kernel.org,
Xiongwei Song <xiongwei.song@...driver.com>
Subject: Re: [PATCH v2] Documentation: kernel-parameters: remove
slab_max_order and noaliencache
On 11/27/23 01:26, Hyeonggon Yoo wrote:
> On Fri, Nov 24, 2023 at 8:24 PM Vlastimil Babka <vbabka@...e.cz> wrote:
>> > -
>> > noalign [KNL,ARM]
>> >
>> > noaltinstr [S390] Disables alternative instructions patching
>> > @@ -5887,12 +5883,6 @@
>> > own.
>> > For more information see Documentation/mm/slub.rst.
>> >
>> > - slab_max_order= [MM, SLAB]
>> > - Determines the maximum allowed order for slabs.
>> > - A high setting may cause OOMs due to memory
>> > - fragmentation. Defaults to 1 for systems with
>> > - more than 32MB of RAM, 0 otherwise.
>>
>> I think here we should consider the long-term plan first. It's a bit
>> unfortunate (in hindsight) SLUB brought its own prefix of parameters, even
>> if some became interchangeable aliases later (slab/slub_nomerge), some not.
>
> Good point, thank you for pointing them out!
>
>> I think it would be best to unify them, and consider the string "slub" an
>> implementation detail of the general "slab allocator" term going forward.
>
> So slab_{nomerge,min_objects,min_order, ...etc} are common to the
> concept of slab so slab_$param will be appropriate.
>
> But if we add something like slub_nocmpxchg later, it would be slub_nocmpxchg
> as it's an implementation-specific feature.
Hm but we're going to have only one implementation and I hope that won't
change again (the implementation could furter evolve, even with large
changes, but not split into separate selection again). So I'd rather use
slab_$thing even for parameters that look like implementation specific.
>> So what I'd propose is that we change all parameters to accept a
>> "slab_$param" as a primary and documented name (and the description can
>> contain just [MM] tag, no [SLAB] or [SLUB] needed), with "slub_$param" is
>> also accepted as an alias where it exists today, and there's just a note
>> that the slub_$param name is also accepted in the description of the
>> canonical parameter, not in a separate description.
>
> No reason not to do it.
>
>> Then maybe in a few
>> years we can mark the old names as deprecated and start issuing low-key
>> warnings (while still accepting them), and in 10 years maybe remove them
>> completely. Thoughts?
>
> That might be the safest way to remove a kernel parameter but should
> we remove them?
> Probably 1) allowing both slub_$param and slab_$param for general
> parameters (forever)
Well we can certainly try deprecate them in few years, just to avoid cruft.
But it's not critical in case that's rejected for some reason.
> and 2) only using slub_$param for slub-specific params would be enough?
>
>> > -
>> > slub_debug[=options[,slabs][;[options[,slabs]]...] [MM, SLUB]
>> > Enabling slub_debug allows one to determine the
>> > culprit if slab objects become corrupted. Enabling
>>
Powered by blists - more mailing lists