Re: [PATCH v3 5/5] alloc_tag: config to store page allocation tag refs in page flags

[Suren Baghdasaryan <surenb@google.com>]

On Mon, Oct 21, 2024 at 2:13 AM David Hildenbrand <david@redhat.com> wrote:
>
>
>
> Am 21.10.24 um 09:26 schrieb Michal Hocko:
> > On Fri 18-10-24 14:57:26, Suren Baghdasaryan wrote:
> >> On Fri, Oct 18, 2024 at 10:45 AM Suren Baghdasaryan <surenb@google.com> wrote:
> >>>
> >>> On Fri, Oct 18, 2024 at 10:08 AM Michal Hocko <mhocko@suse.com> wrote:
> >>>
> >>> 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.
> >
> > This makes much more sense!
> >
> >> 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.
> >
> > I do not want to bikeshed about this but to me it would make more sense
> > to have an extension paramater to mem_profiling and call it something
> > like compress or similar so that page flags are not really carved into
> > naming. The docuemntation then can explain that the copression cannot be
> > always guaranteed and it might fail so this is more of a optimistic and
> > potentially failing optimization that might need to be dropped in some
> > usege scenarios.
>
> Maybe we can reuse the existing parameter (e.g., tristate). Only makes sense if
> we don't expect too many other modes though :)

Yeah, I thought about adding new values to "mem_profiling" but it's a
bit complicated. Today it's a tristate:

mem_profiling=0|1|never

0/1 means we disable/enable memory profiling by default but the user
can enable it at runtime using a sysctl. This means that we enable
page_ext at boot even when it's set to 0.
"never" means we do not enable page_ext, memory profiling is disabled
and sysctl to enable it will not be exposed. Used when a distribution
has CONFIG_MEM_ALLOC_PROFILING=y but the user does not use it and does
not want to waste memory on enabling page_ext.

I can add another option like "pgflags" but then it also needs to
specify whether we should enable or disable profiling by default
(similar to 0|1 for page_ext mode). IOW we will need to encode also
the default state we want. Something like this:

mem_profiling=0|1|never|pgflags_on|pgflags_off

Would this be acceptable?


>
> >
> >> 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.
> >
> > Yes this seems to be a safe default.
>
> Agreed.
>
> >
> >> 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.
>
> Sounds reasonable to me.
>
> --
> Cheers,
>
> David / dhildenb
>