bozonlock

[Mark Vitale]

I'm hoping someone on the list can provide more historical background on the "bozonlock" (AFS_BOZONLOCK_ENV, vcache->pvnLock). By reading how it is used in code, it is clearly intended to only allow one thread at a time to interact with the paging subsystem for a given vcache's memory-mapped pages. This is also alluded to in the name "pvnLock" - paged vnode lock.

Some code comments (1) imply that bozonlock was originally intended for use in the Solaris port. However, sometime before IBM 1.0, it was specifically disabled for Solaris by the addition of #define AFS_NOBOZO_LOCK (2).

Currently bozonlock function (AFS_BOZONLOCK_ENV 1) is only defined for the Darwin port (and DUX).

Does anyone know why bozonlock was disabled for Solaris? Does anyone know any reason why it could/should not be reinstated for Solaris?

Thanks,
--
Mark Vitale
***@sinenomine.net

(1) from src/afs/afs_lock.c:
#ifdef AFS_BOZONLOCK_ENV
/* operations on locks that don't mind if we lock the same thing twice. I'd like to dedicate
this function to Sun Microsystems' Version 4.0 virtual memory system, without
which this wouldn't have been necessary */
void
afs_BozonLock(struct afs_bozoLock *alock, struct vcache *avc)
...


(2) Recently Andrew Deason (in commit 8017773587bf28a58e480f634fdccc287c443d3b) explained and simplified the conditional logic for the AFS bozonlock; he eliminated AFS_NOBOZO_LOCK and extended AFS_BOZONLOCK_ENV to essentially preserve the existing usage. So even though AFS_NOBOZO_LOCK no longer exists on master, Solaris is still not using the bozonlock.

[Andrew Deason]

[Copying to openafs-devel, since this possibly not Solaris-specific.]

On Thu, 3 Jul 2014 22:15:54 +0000

Post by Mark Vitale
Some code comments (1) imply that bozonlock was originally intended
for use in the Solaris port. However, sometime before IBM 1.0, it was
specifically disabled for Solaris by the addition of #define
AFS_NOBOZO_LOCK (2).

Specifically, the comments say it's for a Sun OS, and it was disabled
for SunOS 5 (i.e. Solaris). For pre-Solaris SunOS presumably it was
still active at the time, and that is what it was developed for.

I assumed it was created due to some shortcoming in the SunOS 4 apis,
which is why it is (or was) the only place we have a kind of recursive
lock.

Post by Mark Vitale
Does anyone know why bozonlock was disabled for Solaris? Does anyone
know any reason why it could/should not be reinstated for Solaris?

Well, you have to have a conversation about why you'd want it turned on
in the first place to have a meaningful conversation about this. Without
a reason to turn it on, obviously it's preferable to not have it.

But otherwise:

It serializes all page-related code for a single vnode. I'm not looking
at the code right now, but I don't believe we otherwise would do that
without the pvnLock; that alone means it's an unnecessary overhead and
makes parallel reads and other such things possibly much slower.

It also always sounded to me like a platform-specific hack, and adding
platform-specific oddities and restrictions seems like it would make the
code more complex and harder to understand. (Of course, it's pretty
confusing having those calls in solaris-specific code when they're
always disabled, but they should probably just be removed from there.)

I assume it was taken out because the justification for it in the first
place was some limitation in SunOS 4. That restriction was lifted in
Solaris, and so the pvn lock stuff was removed, and replaced with
something more reasonable.