Sun Enterprise 4500 Server
 
Sun Enterprise 5500 Server
 
Sun Enterprise 450 Server
 
Sun Ultra 450 Workstation
 
Sun Ultra 80 Workstation
 
Sun Enterprise 3500 Server
 
Sun Enterprise 6500 Server
 
Sun Enterprise 220R Server
 
Sun Enterprise 250 Server
 
Solaris SPARC Operating System
 
Sun Ultra 60 Workstation
 
Sun Enterprise 420R Server
 

GCS>Sun Microsystems>Desktops>Workstations
 

This does not eliminate panics of this nature but it does dramatically reduce their likelihood. Affected systems include all volume class servers and desktop systems with USII processors. Early models of the Enterprise (E) class systems are also affected unless they contain mirrored cache modules. If you are unsure if your E class server has mirrored e-cache modules, please contact Sun Services and reference this InfoDoc.

The first step is to make sure you have the minimum kernel patch installed on your system so that the scrubber is available:

All versions of Solaris more recent than 8 have the scrubber already installed.

Next, to maximize the effectiveness of the scrubber the following lines need to be added to /etc/system:

set ecache_scrub_enable=1

set ecache_scan_rate=1000

set ecache_calls_a_sec=100

These settings will have a marginal performance impact (less then 1% on CPUs which run at 400mhz or more).

1. The E-cache scrubber avoids lines
being resident in cache long enough to "decay" leading to parity
errors.

2. The E-cache scrubber tries to find and displace bad clean
lines before something else does ,avoiding E-cache errors that
could lead to a system downtime.

The E-cache scrubber wakes up and inspects a number of cache
lines 100 times a second on each CPU. It inspects the entire cache
in 10 seconds.

Upon line tag inspection (Modified), parity (AFSRs) and
idle/busy flag, the scrubber decides whether to flush the cache
line. The cache line is flushed if the tag is invalid. If the cache
line is not in use, it is skipped.

* On mirrored E$ cpu modules, we additionally flush clean_good_busy and dirty_good_busy lines.



 Modify : EC_state indication of clean/dirty (non-modified/modified)

 Parity : Parity error found in E$ data

 Idle   : CPU currently running idle thread

If flushing, a physical displacement flush using PA (Physical
Address), constructed from tag, is performed.

If ecache_scrub_verbose is set ( the default on Solaris 8
108528-16 or above ), the kernel logs the E-cache lines of the
following states:

clean_bad_idle : CBI event

clean_bad_busy : CBB event

dirty_bad_idle and : DBI event

dirty_bad_busy : DBB event

One can use "kstat -n ecache_kstat"
to confirm the scrubber is active.

If ecache_scrub_panic is set, the kernel panics in addition to
logging (not recommended)

Standard E-Cache Best Practices apply to CBI CBB DBI and DBB
events - if only a single error is observered the CPU should not be
replaced ( unless it has mirrored e-cache ). See <Document: 1009200.1> for further details.

Note: In the case of DBI and DBB events the bad data cannot be
flushed from the cache, and as a result DBI/DBB events are often
immediately followed by a WP/EDP/CP event as a thread attempts to
actually use the bad data. In those cases the DBI/DBB and WP/EDP/CP
events should be considered as a single E-Cache event.

See < Solution: 201640 > for additional details regarding the scrubber.

See <Document: 1018753.1> to identify mirrored e-cache
modules.

This solution has no attachment