Fcntl hangs in 2.2.5 on Solaris

[David Collier-Brown]

Adrian LEE wrote:
>  we have a ever increasing use of this very
> large central fileserver and we've now reached the threshold
> we're this Solaris vs. Samba problem exposes itself enough to
> make for unhappy customers. 

	Ok, I have some info and a question about the
	workaround.

	The application is an ISAM-like one, using
	fcntl locks on IBM, VMS and Unix. These
	are taken on the whole file at open time
	and used to ensure consistency in the index file.

	With 50 files open and 2000 users, smbd slows to a crawl.
	On IBM, 2300 users are handles easily. Without this
	problem, the Sun will handle 6000 users.

	Sun has this open as bug number 4700402, and is 
	discussing stuff that's currently way over my head...

	In the meantime, we're trying building an smbd with 
	spinlocks. Jeremy has cautioned us that This May Be Bad (;-))

	So, let me ask some spinlock question: 
	1) I know that if one process goes down holding a 
	spinlock, the lock will be non-removable. Will it be 
	removed if all smbds go down, and specifically if 
	they go down on a system crash, or are they persistent?

	2) if they are persistent (they appear to be part of the tdb
	data structure ???) can the system be brought down and then
	recover/delete them?

	3) will they be cleaned up if a client machine goes
	down and the smbd discovers this via a keepalive?

	If we can find a workaround with acceptable behavior,
	I'll recommend isolating this set of databases on a
	single samba, dedicated to that task and using spinlocks.
	This is still a workaround, but the production system
	is at risk both ways...

--dave
-- 
David Collier-Brown,           | Always do right. This will gratify 
Performance & Engineering      | some people and astonish the rest.
Americas Customer Engineering, |                      -- Mark Twain
(905) 415-2849                 | davecb at canada.sun.com