Race in dogpile handling

[peter-mogensen]

Hi... I thought I should report that I've experienced what seem like a corner case in the handling of requests blocked by dogpile effect handling.

Due to the polling of C_DE_ITEM_SLEEP_TIME intervals it's possible that the entry being waited on to appear in the cache and get evicted again between two polls - making the blocked request wait for the full grace_ttl.
This of course all depends on load and cache tuning.