On Wed, Oct 04, 2023 at 02:48:45PM +0200, Florian Westphal wrote: > I also think we need to find a different strategy for the > dump-and-reset part when the reset could be interrupted > by a transaction. I think it should be possible to deal with this from userspace. The idea would be to keep the old cache. Then, from the new cache, if EINTR happened before, iterate over the list of objects in the new cache and then lookup for the old objects, then pour the stats from the old to the new objects, then release old cache. Then only one old cache is kept around in worst case. This needs a lookup function for each stateful object type on the old cache based on the handle.
