It was recently observed at [1] that during the folio unmapping stage of migration, when the PTEs are cleared, a racing thread faulting on that folio may increase the refcount of the folio, sleep on the folio lock (the migration path has the lock), and migration ultimately fails when asserting the actual refcount against the expected.
Migration is a best effort service; the unmapping and the moving phase are wrapped around loops for retrying. The refcount of the folio is currently being asserted during the move stage; if it fails, we retry. But, if a racing thread changes the refcount, and ends up sleeping on the folio lock (which is mostly the case), there is no way the refcount would be decremented; as a result, this renders the retrying useless. In the first patch, we make the refcount check also during the unmap stage; if it fails, we restore the original state of the PTE, drop the folio lock, let the system make progress, and retry unmapping again. This improves the probability of migration winning the race.
Given that migration is a best-effort service, it is wrong to fail the test for just a single failure; hence, fail the test after 100 consecutive failures (where 100 is still a subjective choice).
[1] https://lore.kernel.org/all/20240801081657.1386743-1-dev.jain@arm.com/
Dev Jain (2): mm: Retry migration earlier upon refcount mismatch selftests/mm: Do not fail test for a single migration failure
mm/migrate.c | 9 +++++++++ tools/testing/selftests/mm/migration.c | 17 +++++++++++------ 2 files changed, 20 insertions(+), 6 deletions(-)