Mv hot threads1 #101
Mv hot threads1 #101
Conversation
…xtra logging intact
… fix ... and merged with changes already on master
… fix ... and merged with changes already on master
revert to original, fixed 20 percent for internal databasses
Mv async close
… fix ... and merged with changes already on master
… fix ... and merged with changes already on master
|
I'm reviewing this code now. Also, since most of this was previously reviewed in #97, I went spent some time reconstructing a proper change base to review this branch against to make it easier to see what's new since the last review. So, reviewing based off this changeset. |
| BGFileUnmapper2(ptr); | ||
| BGCloseInfo * ptr=new BGCloseInfo(fd_, base_, file_offset_, limit_-base_, | ||
| NULL, metadata_offset_); | ||
| BGFileUnmapper2(ptr); |
There was a problem hiding this comment.
In this case (the non-async close case), we don't ever delete the allocated BGCloseInfo because line 1209 was changed to assume that BGCloseInfo was submitted to/deallocated by the thread pool -- which isn't true in this case. I think we just need a delete ptr after the BGFileUnmapper2(ptr) call here.
There was a problem hiding this comment.
Good find. I have not rerun valgrind against recent code. Will add refine and refdec pair so that same logic works in both conditions. Refdec to zero deletes object.
|
Ran branch through the same stress testing I did in #97. As noted above, OS X had 200% CPU usage from the safety threads that were never sleeping, but worked fine otherwise. Linux was fine across the board. I reintroduced the changes that help trigger the known deadlock. With the All looks good other than the issues I mentioned in the diff comments above. |
|
Looked over changes and retested on OS X and Linux. We should fix the semaphore creation issue before 2.0 final, but no reason to hold up merging now. +1 merge. |
ghost
assigned 
This is branch contains the threading code originally developed for the eleveldb async NIF. It was previously delivered in leveldb as part of a patch in branch 1.2.4-turner. Joe B. spotted an unlikely, but real, race condition that this branch addressed via the QueueThread object.
Today only Imm to Level-0 compactions post to a hot thread pool. The goal is to move background disk writes and all background compactions to a series of specialized hot thread pools ... but that will be a different branch.