[illumos-Developer] MUCH better --> zil for zvol TRUNCATE (and iSCSI UNMAP) working...

[Dan McDonald]

On Wed, Mar 02, 2011 at 05:48:09PM -0500, Eric Schrock wrote:
> My only concern is that txg_wait_synced() can hurt performance, since it
> will force out a txg.  And in a real world environment, we will likely be
> freeing blocks a lot.   It doesn't appear that we do a txg_wait_synced() on
> every write for non-WCE zvols.  IS there something different about frees, or
> the semantics of DF_WAIT_SYNC?

DF_WAIT_SYNC reflects a desire by the caller to wait until data has been
fully written.  The COMSTAR caller has marked this LU as having a disabled
write cache.  My tests which use a default setting don't have this flag set,
e.g.

I've updated the code and webrev:

	http://www.kebe.com/~danmcd/webrevs/unmap/

With the changes.  I will be running one more set of tests, as my last batch
only included the now-simplified zvol_replay_truncate().  I still need to
test this newest check-WCE/check-DF_WAIT_SYNC path.

Dan