Contributed by Peter N. M. Hansteen on from the locks in the pool dept.
I showed up at t2k17 with a couple hold-over diffs from e2k17 that weren't stable then and hadn't gotten much better since, so after a red-eye through Chicago I arrived in the hackroom, fired up my laptop and synced trees.
Meanwhile, people trickled in and the best part of hackathons, the conversations and "what do you think about this?" chats started. Theo introduced me to Todd Mortimer (mortimer@), who's been hacking on clang to implement RETGUARD for C programs. Over the hackathon we discussed a few loose ends that cropped up and what the correct behavior should be for them as well as the mechanics of avoiding 0xc3 bytes (the RET opcode) embedded in the middle of other multi-byte x86 machine code. Fun stuff.
Martin (mpi@) and I had a conversation about the desirability of being able to sleep while holding netlock and pretty much came down on "oof, the scheduler does need work before the underlying issue driving this question can be resolved enough to answer it". :-(
After some final hammering I got in an enhancement to pool(9) to let a pool use (sleeping) rwlocks instead of (spinning) mutexes and then immediately used that for the per-CPU pool cache pool as well as the futex pool. Further pools are likely to be converted as well kernel upper-level locking changes are made.
Speaking of, a larger diff I had been working on for said upper-level locking was still suffering deadlock issues so I took a stab at narrowing it down to just a lock for the process tree, mostly mirroring the FreeBSD proctree_lock. That appears to be holding up much better and I just have some code arrangement issues around sys_ptrace() before that'll go out for final review.
Then most of the way through the week, Bob (beck@) vocally complained that life would be easier for libressl if we had some version of pthread_once() and the pthread mutex routines in libc. This would make some other stuff easier too (c.f. /usr/X11R6/lib/libpthread-stubs.*) and the TIB work over the last couple years has basically eliminated the runtime costs of doing so, so I spent most the rest of the hackathon finding the right place to draw a line through libpthread and move everything on the one side of the line into libc. That code seems pretty stable and the xenocara and ports people seem to like—or at least accept—the effects, so it will almost certainly go in with the next libc bump.
Lots of other random conversations, hacking, meals, and beer. Many thanks to Ken (krw@) and local conspirators for another excellent Toronto hackathon!
(Comments are closed)