100% cpu utilization

[Jeremy Allison]

On Tue, Nov 06, 2001 at 10:40:50AM -0500, Scott Moomaw wrote:
> I've just watched the system climb to 100% cpu utilization again and may
> have caught some helpful info.  We're running the latest CVS on a Solaris
> 8 x86 box.  Looking at things this time, it may be that the select
> processing in sys_select is overwhelming the system under heavier loads.
> It appears that Solaris uses poll() as the underlying method of
> implementing the select() call.  This morning, I saw approximately 170
> smbd processes active when the problem occurred.  Every one that I was
> able to get a pstack from was in a state similar to this.
> 
> 22808:  /usr/local/samba/bin/smbd -D
>  dfa19095 poll     (80478c0, 3, ea60)
>  dfa2bee5 select   () + 1f5
>  08132133 sys_select (d, 8047a2c, 8047a24, 80a12f9) + c7
>  080a1395 receive_message_or_smb (8210991, 10040, ea60, 80a25e0) + 169
>  080a261a smbd_process (dfbed1e8, 8047b00, 8047be8, 210, dfa0d67f, dfa0d6a3, dfbe137f, 8047b00, ...) + 11e
>  0806ee59 main     (2, 8047b44, 8047b50) + 6d9
>  0806d94f _start   () + 57
> 
> Top reports extremely high kernel utilization of the CPU.  Anyone have
> knowledge of the efficiency of Solaris's poll() system call?  It seems
> like the number of processes in a poll() state is more than the system can
> handle.  The system is a dual P400 and didn't exhibit this symptom until
> after our Win2k/Samba 2.2.x upgrades this summer.
> 
> Your thoughts???

I don't know if you saw my posting to Samba Technical,
but I may finally have figured out this one.

I committed a fix to HEAD and 2.2 CVS.

If you could test the 2.2 tree out I'd appreciate
it....

Thanks,

	Jeremy.