Linux Lands Fix for Early 6.17 Regression Causing 37~43% Performance Hit

11 days ago

Copy Link

Linux 6.17 faced a significant performance regression due to an ARM64 optimization commit.
Intel's kernel test robot reported a 37% regression in stress-ng benchmarks.
Oracle engineer Lorenzo Stoakes reproduced a 43% regression on an Intel Raptor Lake system.
The issue was traced to expensive folio lookups during mremap folio PTE batch operations.
A two-line patch was merged to avoid unnecessary folio lookups, restoring performance.
The fix leverages pte_batch_hint() to determine PTE batch size without vmemmap lookups.
Performance improved from 81,503 to 138,701 realloc calls/sec, nearly restoring original levels.
The regression was quickly addressed days after the Linux 6.17-rc1 release.

Hasty Briefsbeta