On Fri 18-10-24 10:45:39, Suren Baghdasaryan 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.
Right, I do understand (I guess) the challenge. I am just arguing that
it makes really no sense to tell user to recompile the kernel with a
CONFIG_FOO to workaround this limitation. Please note that many users of
this feature will simply use a precompiled (e.g. distribution) kernels.
Once you force somebody to recompile with
CONFIG_PGALLOC_TAG_USE_PAGEFLAGS=n they are not going back to a more
memory optimal implementation.
Just my 2cents
--
Michal Hocko
SUSE Labs
