On 28.10.24 22:56, Daniel Gomez wrote:
On Fri Oct 25, 2024 at 10:21 PM CEST, David Hildenbrand wrote:
Sorry for the late reply!
IMHO, as I discussed with Kirill, we still need maintain compatibility
with the 'huge=' mount option. This means that if 'huge=never' is set
for tmpfs, huge page allocation will still be prohibited (which can
address Hugh's request?). However, if 'huge=' is not set, we can
allocate large folios based on the write size.
So, in order to make tmpfs behave like other filesystems, we need to
allocate large folios by default. Not setting 'huge=' is the same as
setting it to 'huge=never' as per documentation. But 'huge=' is meant to
control THP, not large folios, so it should not have a conflict here, or
else, what case are you thinking?
I think we really have to move away from "huge/thp == PMD", that's a
historical artifact. Everything else will simply be inconsistent and
confusing in the future -- and I don't see any real need for that. For
anonymous memory and anon shmem we managed the transition. (there is a
longer writeup from me about this topic, so I won't go into detail).
I think I raised this in the past, but tmpfs/shmem is just like any
other file system .. except it sometimes really isn't and behaves much
more like (swappable) anonymous memory. (or mlocked files)
There are many systems out there that run without swap enabled, or with
extremely minimal swap (IIRC until recently kubernetes was completely
incompatible with swapping). Swap can even be disabled today for shmem
using a mount option.
That's a big difference to all other file systems where you are
guaranteed to have backend storage where you can simply evict under
memory pressure (might temporarily fail, of course).
I *think* that's the reason why we have the "huge=" parameter that also
controls the THP allocations during page faults (IOW possible memory
over-allocation). Maybe also because it was a new feature, and we only
had a single THP size.
There is, of course also the "fallocate() might not free up memory if
there is an unexpected reference on the page because splitting it will
fail" problem, that even exists when not over-allocating memory in the
first place ...
So ...I don't think tmpfs behaves like other file system in some cases.
And I don't think ignoring these points is a good idea.
Assuming a system without swap, what's the difference you are concern
about between using the current tmpfs allocation method vs large folios
implementation?
As raised above, there is the interesting interaction between
fallocate(FALLOC_FL_PUNCH_HOLE) and raised refcounts, where we can fail
to reclaim memory.
shmem_fallocate()->shmem_truncate_range()->truncate_inode_pages_range()->truncate_inode_partial_folio().
It's better than it was in the past -- in the past we didn't even try
splitting, but today splitting can still fail and we'll never try
reclaiming that memory again later. This is very different to anonymous
memory where we have the deferred split queue+remember which pages where
zapped implicitly using the page tables (instead of zeroing them out and
not freeing up the memory).
It's one of the issues people ran into when using THP+shmem for backing
guest VMs along with memory ballooning. For that reason, the
recommendation still is to disable THP when using shmem for backing
guest VMs and relying on memory overcommit optimizations such as memory
balloon inflation.
Fortunately I don't maintain that code :)
If we don't want to go with the shmem_enabled toggles, we should
probably still extend the documentation to cover "all THP sizes", like
we did elsewhere.
huge=never: no THPs of any size
huge=always: THPs of any size (fault/write/etc)
huge=fadvise: like "always" but only with fadvise/madvise
huge=within_size: like "fadvise" but respect i_size
We could think about adding a "nowaste" extension and try make it the
default.
For example
"huge=always:nowaste: THPs of any size as long as we don't over-allocate
memory (write)"
This is the default behaviour in other fs too. I don't think is
necessary to make it explicit.
Please keep in mind that allocating THPs of different size during *page
faults* will have to fit into the whole picture we are creating here.
That's also what "huge=always" controls for shmem today IIRC.
I consider allocating large folios in shmem/tmpfs on the write path less
controversial than allocating them on the page fault path -- especially
as long as we stay within the size to-be-written.
I think in RHEL THP on shmem/tmpfs are disabled as default (e.g.,
shmem_enabled=never). Maybe because of some rather undesired
side-effects (maybe some are historical?): I recall issues with VMs with
THP+ memory ballooning, as we cannot reclaim pages of folios if
splitting fails). I assume most of these problematic use cases don't use
tmpfs as an ordinary file system (write()/read()), but mmap() the whole
thing.
Sadly, I don't find any information about shmem/tmpfs + THP in the RHEL
documentation; most documentation is only concerned about anon THP.
Which makes me conclude that they are not suggested as of now.
I see more issues with allocating them on the page fault path and not
having a way to disable it -- compared to allocating them on the write()
path.
I may not understand your issues. IIUC, you can disable allocating huge
pages on the page fault path by using the 'huge=never' mount option or
setting shmem_enabled=deny. No?
That's what I am saying: if there is some way to disable it that will
keep working, great.
I agree. That aligns with what I recall Hugh requested. However, I
believe if that is the way to go, we shouldn't limit it to tmpfs.
Otherwise, why should tmpfs be prevented from allocating large folios if
other filesystems in the system are allowed to allocate them?
See above. On systems without/little swap you might not want them for
shmem/tmpfs, but would happily use them elsewhere.
The "write() won't waste memory" case is really interesting, the
"fallocate cannot free the memory" still exists. A shrinker might help.
The previous implementation with large folios allocation was wrong
and was actually wasting memory by rounding up while trying to find
the order. Matthew already pointed it out [1]. So, with that fixed, we
should not end up wasting memory.
Again, we should have a clear path forward how we deal with page faults
and how this fits into the picture.
--
Cheers,
David / dhildenb
