On 31.10.23 12:55, Ryan Roberts wrote:
On 31/10/2023 11:50, Ryan Roberts wrote:
On 06/10/2023 21:06, David Hildenbrand wrote:
[...]
Change 2: sysfs interface.
If we call it THP, it shall go under "/sys/kernel/mm/transparent_hugepage/", I
agree.
What we expose there and how, is TBD. Again, not a friend of "orders" and
bitmaps at all. We can do better if we want to go down that path.
Maybe we should take a look at hugetlb, and how they added support for multiple
sizes. What *might* make sense could be (depending on which values we actually
support!)
/sys/kernel/mm/transparent_hugepage/hugepages-64kB/
/sys/kernel/mm/transparent_hugepage/hugepages-128kB/
/sys/kernel/mm/transparent_hugepage/hugepages-256kB/
/sys/kernel/mm/transparent_hugepage/hugepages-512kB/
/sys/kernel/mm/transparent_hugepage/hugepages-1024kB/
/sys/kernel/mm/transparent_hugepage/hugepages-2048kB/
Each one would contain an "enabled" and "defrag" file. We want something minimal
first? Start with the "enabled" option.
enabled: always [global] madvise never
Initially, we would set it for PMD-sized THP to "global" and for everything else
to "never".
Hi David,
I've just started coding this, and it occurs to me that I might need a small
clarification here; the existing global "enabled" control is used to drive
decisions for both anonymous memory and (non-shmem) file-backed memory. But the
proposed new per-size "enabled" is implicitly only controlling anon memory (for
now).
1) Is this potentially confusing for the user? Should we rename the per-size
controls to "anon_enabled"? Or is it preferable to jsut keep it vague for now so
we can reuse the same control for file-backed memory in future?
2) The global control will continue to drive the file-backed memory decision
(for now), even when hugepages-2048kB/enabled != "global"; agreed?
Thanks,
Ryan
Also, an implementation question:
hugepage_vma_check() doesn't currently care whether enabled="never" for DAX VMAs
(although it does honour MADV_NOHUGEPAGE and the prctl); It will return true
regardless. Is that by design? It couldn't fathom any reasoning from the commit log:
The whole DAX "hugepage" and THP mixup is just plain confusing. We're
simply using PUD/PMD mappings of DAX memory, and PMD/PTE- remap when
required (VMA split I assume, COW).
It doesn't result in any memory waste, so who really cares how it's
mapped? Apparently we want individual processes to just disable PMD/PUD
mappings of DAX using the prctl and madvise. Maybe there are good reasons.
Looks like a design decision, probably some legacy leftovers.
--
Cheers,
David / dhildenb
