Folio Restructuring for Huge Pages in guest_memfd

on YouTube ↗17m

" summary

Adds HugeTLB and THP support to guest_memfd by making folio restructuring (split/merge) orthogonal to THP and HugeTLB, using existing helpers like folio_split() and considering xas_split_alloc() limits for large orders. Implements per-page tracking by splitting down from 1G to PMD then to pages using a logarithmic split (reduces 262,144 allocations to ~19 steps), preserves HVO optimization, removes host mappings on share→private, and merges folios in place so the original allocator can reclaim the folio intact.

AI-written summary · about

" chapters (10)

1. 0:00Introduction and ask
2. 1:46Guest MFD overview
3. 3:12Why restructure folios
4. 6:03HTLB extension needs
5. 8:31PUD size folio support
6. 9:35Proposal for orthogonal restructuring
7. 11:04Challenges with in-place split
8. 13:13Ref counting and folio design
9. 14:38Optimal split strategy
10. 16:46Summary and Q&A