On 08.04.24 10:07, Yosry Ahmed wrote:
[..]
Note that totalram_pages can also change during memory onlining and
offlining. For that you need a memory notifier that also calls that
refresh function. It's simple enough, though, check out the code
around register_memory_notifier() in drivers/xen/balloon.c.
Good point, I completely missed this. It seems like totalram_pages can
actually change from contexts other than memory hotplug as well,
specifically through adjust_managed_page_count(), and mostly through
ballooning drivers. Do we trigger the notifiers in this case? I can't
find such logic.
Things like virtio-balloon never online/offline memory and would never
call it.
I see calls to adjust_managed_page_count() from
drivers/virtio/virtio_balloon.c, but I don't understand enough to know
what they are doing.
Essentially fake removing/adding pages. :)
Things like virtio-mem sometimes will online/offline memory and would
sometimes call it (but not always). Things like the Hyper-V balloon and
XEN balloon never offline memory, and would only call it when onlining
memory.
Thanks for the details.
It seems like in this case the actual amount of memory does not
change, but the drivers take it away from the system. It makes some
sense to me that the zswap limits do not change in this case,
especially that userspace just sets those limits as a percentage of
total memory. I wouldn't expect userspace to take ballooning into
account here.
For virtio-mem, it does change ("actual amount of memory"). For
virtio-balloon, it's tricky. When using virtio-balloon for VM resizing,
it would similarly change. When using it for pure memory overcommit, it
depends on whatever the policy in the hypervisor is ... might be that
under memory pressure that memory is simply given back to the VM.
That's good to know, it seems like we need to take these into account,
and not just because the users may happen to change zswap limits while
they are onlining/offlining memory.
Yes. Likely other parts of the system would want to know when available
memory changes (especially, if it's quite significant).
However, it would be a behavioral change from today where we always
rely on totalram_pages(). Also, if userspace happens to change the
limit when a driver is holding a big chunk of memory away from
totalram_pages, then the limit would be constantly underestimated.
I do not have enough familiarity with memory ballooning to know for
sure if this is okay. How much memory can memory ballooning add/remove
from totalram_pages(), and is it usually transient or does it stick
around for a while?
Also CCing David here.
It can be a lot. Especially with the Hyper-V balloon (but also on
environments where other forms of memory hotunplug are not supported),
memory ballooning can be used to unplug memory. So that memory can be
gone for good and it can end up being quite a lot of memory.
The clean thing to do would be to have a way for other subsystems to get
notified on any totalram_pages() changes, so they can adjust accordingly.
Yeah I agree. I imagined that register_memory_notifier() would be the
way to do that. Apparently it is only effective for memory hotplug.
Yes. Right now it's always called under the memory hotplug lock. We
could reuse it for this fake hotplug/unplug as well, but we'd have to
clarify how exactly we expect this to interact with the existing
notifications+notifiers.
From your description, it sounds like the ballooning drivers may have
a very similar effect to memory hotplug, but I don't see
memory_notify() being called in these paths.
Right, the existing notifications (MEM_ONLINE etc.) are called when
memory blocks change their state. That's not what the fake
hotplug/unplug does, that operates on much smaller ranges.
Do we need a separate notifier chain for totalram_pages() updates?
Good question. I actually might have the requirement to notify some arch
code (s390x) from virtio-mem when fake adding/removing memory, and
already wondered how to best wire that up.
Maybe we can squeeze that into the existing notifier chain, but needs a
bit of thought.
--
Cheers,
David / dhildenb
