On 21.01.25 09:18, Mounesh Badiger wrote:
Hi,
Hi,
Can linux-mm group confirm, if huge pages of size 2MB cannot be migrated
or remapped when user-space uses them for IO with the bypassing kernel?
We need to VA->PA intact once it is mapped.
We want to reserve huge pages and map them using hugetlbfs into the
userspace. Use /proc/<pid>/pagemap to get VA->PA mapping and use PA
for IO.
A kernel that supports CONFIG_MIGRATION will be able to migrate 2 MiB
hugetlb folios on most architectures (e.g., x86-64). Such migration
might be triggered by CMA allocations or memory offlining.
IIRC, ordinary page compaction should not be migrating 2 MB hugetlb
folios, because they consume a full pageblock and there is, therefore,
no value in trying to compact such pageblocks.
So short answer: they can get migrated, but migration might not happen
frequently.
To block migration, you could long-term pin these pages from getting
migrated either using vmsplice() or iouring fixed buffers. (the latter
is preferred).
--
Cheers,
David / dhildenb
