Re: [PATCH RFC] mm/memory_hotplug: Introduce memory block types

On 26.11.18 13:30, David Hildenbrand wrote:

On 23.11.18 19:06, Michal Suchánek wrote:

quoted

On Fri, 23 Nov 2018 12:13:58 +0100
David Hildenbrand [off-list ref] wrote:

quoted

On 28.09.18 17:03, David Hildenbrand wrote:

quoted

How to/when to online hotplugged memory is hard to manage for
distributions because different memory types are to be treated differently.
Right now, we need complicated udev rules that e.g. check if we are
running on s390x, on a physical system or on a virtualized system. But
there is also sometimes the demand to really online memory immediately
while adding in the kernel and not to wait for user space to make a
decision. And on virtualized systems there might be different
requirements, depending on "how" the memory was added (and if it will
eventually get unplugged again - DIMM vs. paravirtualized mechanisms).

On the one hand, we have physical systems where we sometimes
want to be able to unplug memory again - e.g. a DIMM - so we have to online
it to the MOVABLE zone optionally. That decision is usually made in user
space.

On the other hand, we have memory that should never be onlined
automatically, only when asked for by an administrator. Such memory only
applies to virtualized environments like s390x, where the concept of
"standby" memory exists. Memory is detected and added during boot, so it
can be onlined when requested by the admininistrator or some tooling.
Only when onlining, memory will be allocated in the hypervisor.

But then, we also have paravirtualized devices (namely xen and hyper-v
balloons), that hotplug memory that will never ever be removed from a
system right now using offline_pages/remove_memory. If at all, this memory
is logically unplugged and handed back to the hypervisor via ballooning.

For paravirtualized devices it is relevant that memory is onlined as
quickly as possible after adding - and that it is added to the NORMAL
zone. Otherwise, it could happen that too much memory in a row is added
(but not onlined), resulting in out-of-memory conditions due to the
additional memory for "struct pages" and friends. MOVABLE zone as well
as delays might be very problematic and lead to crashes (e.g. zone
imbalance).

Therefore, introduce memory block types and online memory depending on
it when adding the memory. Expose the memory type to user space, so user
space handlers can start to process only "normal" memory. Other memory
block types can be ignored. One thing less to worry about in user space.
  

So I was looking into alternatives.

1. Provide only "normal" and "standby" memory types to user space. This
way user space can make smarter decisions about how to online memory.
Not really sure if this is the right way to go.


2. Use device driver information (as mentioned by Michal S.).

The problem right now is that there are no drivers for memory block
devices. The "memory" subsystem has no drivers, so the KOBJ_ADD uevent
will not contain a "DRIVER" information and we ave no idea what kind of
memory block device we hold in our hands.

$ udevadm info -q all -a /sys/devices/system/memory/memory0

  looking at device '/devices/system/memory/memory0':
    KERNEL=="memory0"
    SUBSYSTEM=="memory"
    DRIVER==""
    ATTR{online}=="1"
    ATTR{phys_device}=="0"
    ATTR{phys_index}=="00000000"
    ATTR{removable}=="0"
    ATTR{state}=="online"
    ATTR{valid_zones}=="none"


If we would provide "fake" drivers for the memory block devices we want
to treat in a special way in user space (e.g. standby memory on s390x),
user space could use that information to make smarter decisions.

Adding such drivers might work. My suggestion would be to let ordinary
DIMMs be without a driver for now and only special case standby memory
and eventually paravirtualized memory devices (XEN and Hyper-V).

Any thoughts?

If we are going to fake the driver information we may as well add the
type attribute and be done with it.

I think the problem with the patch was more with the semantic than the
attribute itself.

What is normal, paravirtualized, and standby memory?

I can understand DIMM device, baloon device, or whatever mechanism for
adding memory you might have.

I can understand "memory designated as standby by the cluster
administrator".

However, DIMM vs baloon is orthogonal to standby and should not be
conflated into one property.

paravirtualized means nothing at all in relationship to memory type and
the desired online policy to me.

Right, so with whatever we come up, it should allow to make a decision
in user space about
- if memory is to be onlined automatically

And I will think about if we really should model standby memory. Maybe
it is really better to have in user space something like (as Dan noted)

if (isS390x() && type == "dimm") {
	/* don't online, on s390x system DIMMs are standby memory */
}

The we could have in addition

if (type == "balloon") {
	/*
	 * Balloon will not be unplugged by offlining the whole block at
	 * once, online as !movable.
	 */
}

But I'll have to think about the wording / types etc. (I neither like
"dimm" nor "balloon").

-- 

Thanks,

David / dhildenb