On Sat Feb 17 00:19:25 2024 +0000, Paul Gofman wrote:
I agree with both comments (while I don't see 50ms is something risky in such a case, doing that 1000ms also won't hurt). However, before adjusting and resending the patches for nits, I'd wait for a general comment(s) if this is going forward at all.
FWIW my own major concern with the patch is that (hopefully) upcoming ntsync changes should render it useless (as it essentially is already with esync / fsync out of tree patches). In fact, adding a facility to request APC processing from the calling thread upon server call exit (as I suggested in another comment) also won't help much as in many practically interesting cases the async completion is triggered by server fd poll without any calling thread present (named pipes with IO fully performed through server are an exception but probably less interesting perf wise than sockets).