kern/184048: developers handbook, 10.7 Debugging Deadlocks - must include warning on DIAGNOSTIC adding too much overhead

[Anton Shterenlikht]

>Number:         184048
>Category:       kern
>Synopsis:       developers handbook, 10.7 Debugging Deadlocks - must include warning on DIAGNOSTIC adding too much overhead
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    freebsd-bugs
>State:          open
>Quarter:        
>Keywords:       
>Date-Required:
>Class:          change-request
>Submitter-Id:   current-users
>Arrival-Date:   Mon Nov 18 09:40:00 UTC 2013
>Closed-Date:
>Last-Modified:
>Originator:     Anton Shterenlikht
>Release:        FreeBSD 11.0-CURRENT ia64
>Organization:
University of Bristol 
>Environment:
System: FreeBSD mech-cluster241.men.bris.ac.uk 11.0-CURRENT FreeBSD 11.0-CURRENT #8 r256683: Fri Oct 18 08:52:38 BST 2013 root at mech-cluster241.men.bris.ac.uk:/usr/obj/usr/src/sys/TZAV ia64


	
>Description:

http://www.freebsd.org/doc/en/books/developers-handbook/kerneldebug-deadlocks.html

My experience with ia64 -current shows that
kernel option DIAGNOSTIC adds too much overhead
to the system. It seems the extra debugging
consumes 100% of system resources and the system
can do nothing else. In particular, the symptoms are:

- kernel thread taskq, which consumes 100% CPU, e.g.:

  PID USERNAME   PRI NICE   SIZE    RES STATE   C   TIME     CPU COMMAND
    0 root         8    0     0K   416K CPU0    0 186:02 100.00% kernel{thread taskq}

- system extremely unresponsive, requiring over 1 min(!)
  to execute ls, df, ps, date, uptime, etc.

- lots of messages similar to:

Expensive timeout(9) function: 0x9ffc000000dca8b0(0) 0.036993424 s

	
>How-To-Repeat:
	
>Fix:

Removing DIAGNOSTIC gets the system back to usable.

I suggest adding a warning to the above page, something
like: "Be warned that option DIAGNOSTIC adds a lot of
extra overheads to the system, which can result in very
noticeable degradation of performance."
And then listing the above sysmptoms as examples of
degraded performance.
Or something like this.

I've discussed this with marcel@ and he agrees with me.

	

>Release-Note:
>Audit-Trail:
>Unformatted: