FYI, a new 'lockstat' command, version 1.4.6, is now available at:
http://oss.sgi.com/projects/lockmeter/download
that contains a few more fixes to the reporting of write-locks (the "write"
side of rwlock_t). Earlier versions of lockstat misreported the wait-time
data for the write_lock behavior of rwlock_t's -- miscalculating the mean
wait-times, and miscalculating the percentage of waits relative to the total
number of write-lock requests. Lockstat tries to be clever and to
differentiate waiting because a lock is owned by another write_lock() owner,
vs. waiting because a lock is owned by a write_lock() or a read_lock() owner,
and this clever calculation wasn't clever enough to be correct.
Some other small fixes were done to clean up the lockstat output when using
"-c" (to see cpu-specific lock statistics). Specifically, some NaN values were
being calculated (using a divisor of zero) and printed.
Spinlock_t reporting is unchanged, and read_lock() for rwlock_t is unchanged.
Those appear to be correct.
John Hawkes
hawkes@xxxxxxxxxxxx
|