On Fri, Oct 18, 2024 at 10:45 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Fri, Oct 18, 2024 at 10:08 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> >
> > On Fri 18-10-24 09:04:24, Suren Baghdasaryan wrote:
> > > On Fri, Oct 18, 2024 at 6:03 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > > >
> > > > On Tue 15-10-24 08:58:59, Suren Baghdasaryan wrote:
> > > > > On Tue, Oct 15, 2024 at 8:42 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > > > [...]
> > > > > > Right, I think what John is concerned about (and me as well) is that
> > > > > > once a new feature really needs a page flag, there will be objection
> > > > > > like "no you can't, we need them for allocation tags otherwise that
> > > > > > feature will be degraded".
> > > > >
> > > > > I do understand your concern but IMHO the possibility of degrading a
> > > > > feature should not be a reason to always operate at degraded capacity
> > > > > (which is what we have today). If one is really concerned about
> > > > > possible future regression they can set
> > > > > CONFIG_PGALLOC_TAG_USE_PAGEFLAGS=n and keep what we have today. That's
> > > > > why I'm strongly advocating that we do need
> > > > > CONFIG_PGALLOC_TAG_USE_PAGEFLAGS so that the user has control over how
> > > > > this scarce resource is used.
> > > >
> > > > I really do not think users will know how/why to setup this and I wouldn't
> > > > even bother them thinking about that at all TBH.
> > > >
> > > > This is an implementation detail. It is fine to reuse unused flags space
> > > > as a storage as a performance optimization but why do you want users to
> > > > bother with that? Why would they ever want to say N here?
> > >
> > > In this patch you can find a couple of warnings that look like this:
> > >
> > > pr_warn("With module %s there are too many tags to fit in %d page flag
> > > bits. Memory profiling is disabled!\n", mod->name,
> > > NR_UNUSED_PAGEFLAG_BITS);
> > > emitted when we run out of page flag bits during a module loading,
> > >
> > > pr_err("%s: alignment %lu is incompatible with allocation tag
> > > indexing, disable CONFIG_PGALLOC_TAG_USE_PAGEFLAGS", mod->name,
> > > align);
> > > emitted when the arch-specific section alignment is incompatible with
> > > alloc_tag indexing.
> >
> > You are asking users to workaround implementation issue by configuration
> > which sounds like a really bad idea. Why cannot you make the fallback
> > automatic?
>
> Automatic fallback is possible during boot, when we decide whether to
> enable page extensions or not. So, if during boot we decide to disable
> page extensions and use page flags, we can't go back and re-enable
> page extensions after boot is complete. Since there is a possibility
> that we run out of page flags at runtime when we load a new module,
> this leaves this case when we can't reference the module tags and we
> can't fall back to page extensions, so we have to disable memory
> profiling.
> I could keep page extensions always on just in case this happens but
> that's a lot of memory waste to handle a rare case...
After thinking more about this, I suggest a couple of changes that
IMHO would make configuration simpler:
1. Change the CONFIG_PGALLOC_TAG_USE_PAGEFLAGS to an early boot
parameter. Today we have a "mem_profiling" parameter to enable/disable
memory profiling. I suggest adding "mem_profiling_use_pgflags" to
switch the current behavior of using page extensions to use page
flags. We keep the current behavior of using page extensions as
default (mem_profiling_use_pgflags=0) because it always works even
though it has higher overhead.
2. No auto-fallback. If mem_profiling_use_pgflags=1 and we don't have
enough page flags (at boot time or later when we load a module), we
simply disable memory profiling with a warning.
I think this would be less confusing for users and will avoid David's
example when performance is unexpectedly degraded. The default option
will work for everyone as it does today and advanced users who want to
minimize the overhead can set mem_profiling_use_pgflags=1 to check if
that works for their system.
WDYT?
>
> >
> > --
> > Michal Hocko
> > SUSE Labs
