datastore vs. t.lock2

[Matthias Andree]

Hi,

I've seen t.lock2 take time in the range of minutes on an ext3fs file
system (Linux), while I usually run tests on a reiserfs file system
where it takes in the range of seconds. I conduct these tests on a
Fujitsu MAH-3182MP U160 SCSI 7200/min drive. I straced the program and
found it to call fsync() often next to the close(), and I heard the disk
grind.

I'm asking myself two questions:

1. might reiserfs cheat again or suffer from a broken fsync()
   implementation or is it really so much faster than ext3fs for
   synchronous operations? Something to test for somebody, but beyond
   bogofilter's scope.

2. Might this have killed the Debian MIPS little endian (mipsel) buildd
   t.lock2? I can imagine that under I/O contention a slow drive, this
   test takes ages. Interesting to look at for me.

I'll change the db_close to allow a "bool nosync" be passed in which is
then used in wordlists_open when it defers execution because it cannot
obtain a lock.

This change might cause bogofilter to be more easy on the hardware and
to finish quicker by inferring less hardware-induced (mechanical)
latency times. We'll see, I'll do some benchmarks.

-- 
Matthias Andree