paul chipchase 227893a8f9 Fix a potential deadlock when loading from packfiles ...

This was originally fixed in CL 11458603 but it did not fix the issue entirely and in any case I was not happy with the manual reset event approach.
The problem with the manual reset event was that there is no actual safe point at which we can reset the event without risking some requesting thread missing it and waiting forever, it is just a lot rarer than the original thread lock.

- Each FChunkRequest now can be given an event that is triggered, so the correct thread is always woken up if waiting.
- Since the signals are 1-1 we no longer risk having threads never being woken.
- We only wake the thread that has a request to process so don't risk flooding the cpu.
- Removed FChunkCacheWorker::SetupDecryptionKey as it was not implemented/used.
- Removed FChunkCacheWorker::ChunkRequestAvailable and its associated methods ::WaitForNextChunk/::FlushRemainingChunkCompletionEvents.

In testing this version runs slightly faster on loading tests (0.2seconds on average), although just removing the events entirely and running a busy loop with a yield ran slightly faster on machines with less cores (4 cores with hyperthreading or less) albiet with a much higher variation and so potentially could end up a lot worse.

#rb Graeme.Thornton
#jira UE-89165


#ushell-cherrypick of 11724338 by paul.chipchase

#ROBOMERGE-SOURCE: CL 11724753 in //UE4/Release-4.25/... via CL 11724760
#ROBOMERGE-BOT: RELEASE (Release-4.25Plus -> Main) (v656-11643781)

[CL 11724764 by paul chipchase in Main branch]

2020-02-28 05:04:42 -05:00

16 KiB

Raw Blame History

View Raw