Do time-driven purge of expired cache entries

[jserv]

An idle mount only reclaimed expired TTL-cache entries when a later read happened to sweep them (the opportunistic sweep in dnsfs_cache_lookup) or under memory pressure.

The timer callback runs in softirq context and does the I6 minimum: it sets an atomic purge_pending flag and wakes the resolver kthread, with no cache walk, allocation, or sleeping. The kthread reaps every entry whose TTL has
elapsed and that is not mid-refresh, under query_lock in process context, then re-arms while entries remain. Arming happens from the cache-store paths, guarded on the kthread's existence and made idempotent via timer_pending.

The sweep skips refreshing entries, and DNSFS_PURGE_INTERVAL_SEC (5s) sits above the gap between an entry's expiry and a demand read, so an entry read soon after expiry is still served stale; entries idle past the interval are reclaimed. A bounded stale window, documented at the constant.

Teardown is ordered: dnsfs_free_config calls timer_shutdown_sync after kthread_stop (the only side that re-arms), so no sweep can fire or re-arm before the cache it walks is drained. timer_setup runs in dnsfs_cache_init.

The softirq/process split uses an atomic flag rather than taking the process-context query_queue_lock in the timer, avoiding a lock-context inversion.

---

Summary by cubic

Add a time-driven purge for expired cache entries so idle mounts don’t retain dead records. This bounds staleness and reduces memory use without relying on reads or memory pressure.

• New Features
  • Added a cache timer that sets a purge_pending flag in softirq and wakes the resolver thread; the kthread purges expired, non-refreshing entries under lock, then re-arms while the cache is non-empty (arming happens on store paths; idempotent via timer_pending).
  • Bounded stale window with DNSFS_PURGE_INTERVAL_SEC=5: reads shortly after expiry are still served stale; idle entries past the interval are reclaimed.
  • Ordered teardown: timer_shutdown_sync runs after kthread_stop, ensuring no purge fires while the cache is being drained.
  • New thorough test test_purge_timer validates that an expired entry is purged without a read and that the next read incurs resolver latency (real miss).

^{Written for commit e4835d8. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}