Message-ID: <CAJuCfpGNYgx0GW4suHRzmxVH28RGRnFBvFC6WO+F8BD4HDqxXA@mail.gmail.com>
Date: Tue, 3 Sep 2024 11:19:48 -0700
From: Suren Baghdasaryan <surenb@...gle.com>
To: Andrew Morton <akpm@...ux-foundation.org>
Cc: kent.overstreet@...ux.dev, corbet@....net, arnd@...db.de, 
	mcgrof@...nel.org, rppt@...nel.org, paulmck@...nel.org, thuth@...hat.com, 
	tglx@...utronix.de, bp@...en8.de, xiongwei.song@...driver.com, 
	ardb@...nel.org, david@...hat.com, vbabka@...e.cz, mhocko@...e.com, 
	hannes@...xchg.org, roman.gushchin@...ux.dev, dave@...olabs.net, 
	willy@...radead.org, liam.howlett@...cle.com, pasha.tatashin@...een.com, 
	souravpanda@...gle.com, keescook@...omium.org, dennis@...nel.org, 
	jhubbard@...dia.com, yuzhao@...gle.com, vvvvvv@...gle.com, 
	rostedt@...dmis.org, iamjoonsoo.kim@....com, rientjes@...gle.com, 
	minchan@...gle.com, kaleshsingh@...gle.com, linux-doc@...r.kernel.org, 
	linux-kernel@...r.kernel.org, linux-arch@...r.kernel.org, linux-mm@...ck.org, 
	linux-modules@...r.kernel.org, kernel-team@...roid.com
Subject: Re: [PATCH v2 6/6] alloc_tag: config to store page allocation tag
 refs in page flags

On Sun, Sep 1, 2024 at 10:16 PM Andrew Morton <akpm@...ux-foundation.org> wrote:
>
> On Sun,  1 Sep 2024 21:41:28 -0700 Suren Baghdasaryan <surenb@...gle.com> wrote:
>
> > Add CONFIG_PGALLOC_TAG_USE_PAGEFLAGS to store allocation tag
> > references directly in the page flags. This removes dependency on
> > page_ext and results in better performance for page allocations as
> > well as reduced page_ext memory overhead.
> > CONFIG_PGALLOC_TAG_REF_BITS controls the number of bits required
> > to be available in the page flags to store the references. If the
> > number of page flag bits is insufficient, the build will fail and
> > either CONFIG_PGALLOC_TAG_REF_BITS would have to be lowered or
> > CONFIG_PGALLOC_TAG_USE_PAGEFLAGS should be disabled.
> >
> > ...
> >
> > +config PGALLOC_TAG_USE_PAGEFLAGS
> > +     bool "Use pageflags to encode page allocation tag reference"
> > +     default n
> > +     depends on MEM_ALLOC_PROFILING
> > +     help
> > +       When set, page allocation tag references are encoded inside page
> > +       flags, otherwise they are encoded in page extensions.
> > +
> > +       Setting this flag reduces memory and performance overhead of memory
> > +       allocation profiling but also limits how many allocations can be
> > +       tagged. The number of bits is set by PGALLOC_TAG_USE_PAGEFLAGS and
> > +       they must fit in the page flags field.
>
> Again.  Please put yourself in the position of one of the all-minus-two
> people in this world who aren't kernel-memory-profiling-developers.
> How the heck are they to decide whether or not to enable this?  OK, 59%
> of them are likely to say "yes" because reasons.  But then what?  How
> are they to determine whether it was the correct choice for them?  If
> we don't tell them, who will?

Fair point. I think one would want to enable this config unless there
aren't enough unused bits if the page flags to address all page
allocation tags. That last part of determining how many bits we need
is a bit tricky.
If we put aside loadable modules for now, there are 3 cases:

1. The number of unused page flag bits is enough to address all page
allocations.
2. The number of unused page flag bits is enough if we push
last_cpupid out of page flags. In that case we get the warning at
https://elixir.bootlin.com/linux/v6.11-rc6/source/mm/mm_init.c#L124.
3. The number of unused page flag bits is not enough even if we push
last_cpupid out of page flags. In that case we get the  "Not enough
bits in page flags" build time error.

So, maybe I should make this option "default y" when
CONFIG_MEM_ALLOC_PROFILING=y and let the user disable it if they hit
case #3 or (case #2 and performance hit is unacceptable)?

For loadable modules, if we hit the limit when loading a module at
runtime, we could issue a warning and disable allocation tagging via
the static key. Another option is to fail to load the module with a
proper warning but that IMO would be less appealing.

>
> >  config PGALLOC_TAG_REF_BITS
> >       int "Number of bits for page allocation tag reference (10-64)"
> >       range 10 64
> > -     default "64"
> > +     default "16" if PGALLOC_TAG_USE_PAGEFLAGS
> > +     default "64" if !PGALLOC_TAG_USE_PAGEFLAGS
> >       depends on MEM_ALLOC_PROFILING
> >       help
> >         Number of bits used to encode a page allocation tag reference.
> > @@ -1011,6 +1027,13 @@ config PGALLOC_TAG_REF_BITS
> >         Smaller number results in less memory overhead but limits the number of
> >         allocations which can be tagged (including allocations from modules).
> >
> > +       If PGALLOC_TAG_USE_PAGEFLAGS is set, the number of requested bits should
> > +       fit inside the page flags.
>
> What does "should fit" mean?  "It is your responsibility to make it
> fit"?  "We think it will fit but we aren't really sure"?

This is the case #3 I described above, the user will get a "Not enough
bits in page flags" build time error. If we stick with this config, I
can clarify that in this description.

>
> > +       If PGALLOC_TAG_USE_PAGEFLAGS is not set, the number of bits used to store
> > +       a reference is rounded up to the closest basic type. If set higher than 32,
> > +       a direct pointer to the allocation tag is stored for performance reasons.
> > +
>
> We shouldn't be offering things like this to our users.  If we cannot decide, how
> can they?

Thinking about the ease of use, the CONFIG_PGALLOC_TAG_REF_BITS is the
hardest one to set. The user does not know how many page allocations
are there. I think I can simplify this by trying to use all unused
page flag bits for addressing the tags. Then, after compilation we can
follow the rules I mentioned before:
- If the available bits are not enough to address all kernel page
allocations, we issue an error. The user should disable
CONFIG_PGALLOC_TAG_USE_PAGEFLAGS.
- If there are enough unused bits but we have to push last_cpupid out
of page flags, we issue a warning and continue. The user can disable
CONFIG_PGALLOC_TAG_USE_PAGEFLAGS if last_cpupid has to stay in page
flags.
- If we run out of addressing space during module loading, we disable
allocation tagging and continue. The user should disable
CONFIG_PGALLOC_TAG_USE_PAGEFLAGS.

This leaves one outstanding case:
- If we run out of addressing space during module loading but we would
not run out of space if we pushed last_cpupid out of page flags during
compilation.
In this case I would want the user to have an option to request a
larger addressing space for page allocation tags at compile time.
Maybe I can keep CONFIG_PGALLOC_TAG_REF_BITS for such explicit
requests for a larger space? This would limit the use of
CONFIG_PGALLOC_TAG_REF_BITS to this case only. In all other cases the
number of bits would be set automatically. WDYT?

Powered by blists - more mailing lists