On 20.09.22 03:03, Doug Berger wrote:
On 9/19/2022 2:00 AM, David Hildenbrand wrote:
Hi Dough,
I have some high-level questions.
Thanks for your interest. I will attempt to answer them.
Hi Doug,
sorry for the late reply, slowly catching up on mails.
MOTIVATION:
Some Broadcom devices (e.g. 7445, 7278) contain multiple memory
controllers with each mapped in a different address range within
a Uniform Memory Architecture. Some users of these systems have
How large are these areas typically?
How large are they in comparison to other memory in the system?
How is this memory currently presented to the system?
I'm not certain what is typical because these systems are highly
configurable and Broadcom's customers have different ideas about
application processing.
The 7278 device has four ARMv8 CPU cores in an SMP cluster and two
memory controllers (MEMCs). Each MEMC is capable of controlling up to
8GB of DRAM. An example 7278 system might have 1GB on each controller,
so an arm64 kernel might see 1GB on MEMC0 at 0x40000000-0x7FFFFFFF and
1GB on MEMC1 at 0x300000000-0x33FFFFFFF.
The Designated Movable Block concept introduced here has the potential
to offer useful services to different constituencies. I tried to
highlight this in my V1 patch set with the hope of attracting some
interest, but it can complicate the overall discussion, so I would like
to maybe narrow the discussion here. It may be good to keep them in mind
when assessing the overall value, but perhaps the "other opportunities"
can be covered as a follow on discussion.
The base capability described in commits 7-15 of this V1 patch set is to
allow a 'movablecore' block to be created at a particular base address
rather than solely at the end of addressable memory.
Just so we're on the same page:
Having too much ZONE_MOVABLE memory (ratio compared to !ZONE_MOVABLE
memory) is dangerous. Acceptable ratios highly depend on the target
workload. An extreme example is memory-hungry applications that end up
long-term pinning a lot of memory (e.g., VMs with SR-IO): we can run
easily out of free memory in the !ZONE_MOVABLE zones and might not want
ZONE_MOVABLE at all.
So whatever we do, this should in general not be the kernel sole
decision to make this memory any special and let ZONE_MOVABLE manage it.
It's the same with CMA. "Heavy" CMA users require special configuration:
hugetlb_cma is one prime example.
expressed the desire to locate ZONE_MOVABLE memory on each
memory controller to allow user space intensive processing to
make better use of the additional memory bandwidth.
Can you share some more how exactly ZONE_MOVABLE would help here to make
better use of the memory bandwidth?
ZONE_MOVABLE memory is effectively unusable by the kernel. It can be
used by user space applications through both the page allocator and the
Hugetlbfs. If a large 'movablecore' allocation is defined and it can
Hugetlbfs not necessarily by all architectures. Some architectures don't
support placing hugetlb pages on ZONE_MOVABLE (not migratable) and
gigantic pages are special either way.
only be located at the end of addressable memory then it will always be
located on MEMC1 of a 7278 system. This will create a tendency for user
space accesses to consume more bandwidth on the MEMC1 memory controller
and kernel space accesses to consume more bandwidth on MEMC0. A more
even distribution of ZONE_MOVABLE memory between the available memory
controllers in theory makes more memory bandwidth available to user
space intensive loads.
Sorry to be dense, is this also about different memory access latency or
just memory bandwidth?
Do these memory areas have special/different performance
characteristics? Using dedicated/fake NUMA nodes might be more in line
with what CXL and PMEM are up to.
Using ZONE_MOVABLE for that purpose feels a little bit like an abuse of
the mechanism. To be clearer what I mean:
We can place any movable allocations on ZONE_MOVABLE, including kernel
allocations. User space allocations are just one example, and int he
future we'll turn more and more allocations movable to be able to cope
with bigger ZONE_MOVABLE demands due to DAX/CXL. I once looked into
migrating user space page tables, just to give an example.
Unfortunately, the historical monotonic layout of zones would
mean that if the lowest addressed memory controller contains
ZONE_MOVABLE memory then all of the memory available from
memory controllers at higher addresses must also be in the
ZONE_MOVABLE zone. This would force all kernel memory accesses
onto the lowest addressed memory controller and significantly
reduce the amount of memory available for non-movable
allocations.
We do have code that relies on zones during boot to not overlap within a
single node.
I believe my changes address all such reliance, but if you are aware of
something I missed please let me know.
One example I'm aware of is drivers/base/memory.c:memory_block_add_nid()
/ early_node_zone_for_memory_block().
If we get it wrong, or actually have memory blocks that span multiple
zones, we can no longer offline these memory blocks. We really wanted to
avoid scanning the memmap for now and it seems to get the job done in
environments we care about.
The main objective of this patch set is therefore to allow a
block of memory to be designated as part of the ZONE_MOVABLE
zone where it will always only be used by the kernel page
allocator to satisfy requests for movable pages. The term
Designated Movable Block is introduced here to represent such a
block. The favored implementation allows modification of the
Sorry to say, but that term is rather suboptimal to describe what you
are doing here. You simply have some system RAM you'd want to have
managed by ZONE_MOVABLE, no?
That may be true, but I found it superior to the 'sticky' movable
terminology put forth by Mel Gorman ;). I'm happy to entertain
alternatives, but they may not be as easy to find as you think.
Especially the "blocks" part is confusing. Movable pageblocks? Movable
Linux memory blocks?
Note that the sticky movable *pageblocks* were a completely different
concept than simply reusing ZONE_MOVABLE for some memory ranges.
'movablecore' kernel parameter to allow specification of a base
address and support for multiple blocks. The existing
'movablecore' mechanisms are retained. Other mechanisms based on
device tree are also included in this set.
BACKGROUND:
NUMA architectures support distributing movablecore memory
across each node, but it is undesirable to introduce the
overhead and complexities of NUMA on systems that don't have a
Non-Uniform Memory Architecture.
How exactly would that look like? I think I am missing something :)
The notion would be to consider each memory controller as a separate
node, but as stated it is not desirable.
Doing it the DAX/CXL way would be to expose these memory ranges as
daxdev instead, and letting the admin decide how to online these memory
ranges when adding them to the buddy via the dax/kmem kernel module.
That could mean that your booting with memory on MC0 only, and expose
memory of MC1 via a daxdev, giving the admin the possibility do decide
to which zone the memory should be onlined too.
That would avoid most kernel code changes.
Why can't we simply designate these regions as CMA regions?
We and others have encountered significant performance issues when large
CMA regions are used. There are significant restrictions on the page
allocator's use of MIGRATE_CMA pages and the memory subsystem works very
hard to keep about half of the memory in the CMA region free. There have
been attempts to patch the CMA implementation to alter this behavior
(for example the set I referenced Mel's response to in [1]), but there
are users that desire the current behavior.
Optimizing that would be great, eventually making it configurable or
selecting the behavior based on the actual CMA area sizes.
Why do we have to start using ZONE_MOVABLE for them?
One of the "other opportunities" for Designated Movable Blocks is to
allow CMA to allocate from a DMB as an alternative. This would allow
current users to continue using CMA as they want, but would allow users
(e.g. hugetlb_cma) that are not sensitive to the allocation latency to
let the kernel page allocator make more complete use (i.e. waste less)
of the shared memory. ZONE_MOVABLE pageblocks are always MIGRATE_MOVABLE
so the restrictions placed on MIGRATE_CMA pageblocks are lifted within a
DMB.
The whole purpose of ZONE_MOVABLE is that *no* unmovable allocations end
up on it. The biggest difference to CMA is that the CMA *owner* is able
to place unmovable allocations on it.
Using ZONE_MOVABLE for unmovable allocations (hugetlb_cma) is not
acceptable as is.
Using ZONE_MOVABLE in different context and calling it DMB is very
confusing TBH.
Just a note that I described the idea of a "PREFER_MOVABLE" zone in the
past. In contrast to ZONE_MOVABLE, we cannot run into weird OOM
situations in a ZONE misconfiguration, and we'd end up placing only
movable allocations on it as long as we can. However, especially
gigantic pages could be allocated from it. It sounds kind-of more like
what you want -- and maybe in combination of daxctl to let the user
decide how to online memory ranges.
And just to make it clear again: depending on ZONE_MOVABLE == only user
space allocations is not future proof.
Thanks for your consideration,
Dough Baker ... I mean Doug Berger :).
:) Thanks Doug!
--
Thanks,
David / dhildenb
