From mboxrd@z Thu Jan 1 00:00:00 1970 From: keith.busch@intel.com (Keith Busch) Date: Mon, 5 Feb 2018 12:34:16 -0700 Subject: [PATCH] NVMe-CLI Fix command failures in regress script In-Reply-To: References: <20180202195021.8560-1-jeff.lien@wdc.com> <20180202202930.GK24417@localhost.localdomain> <20180202213829.GL24417@localhost.localdomain> <20180205165704.GA31110@localhost.localdomain> Message-ID: <20180205193416.GC31110@localhost.localdomain> On Mon, Feb 05, 2018@07:00:58PM +0000, Jeffrey Lien wrote: > Ok, that makes sense. I should have thought to try that; sorry for the trouble. > > I did find 2 other problems/failures when using the block device though: > 1) The flush command has a default namespace id of All ID's (0xffffffff) which causes it to fail. So to fix that either: > * The default ns id has to change for the flush command or > * The "-n 1" parm needs to be added to the flush command in the regress script I will change nvme-cli to use the block handle NSID no problem, but isn't Flush the only IO command that is required to support the broadcast NSID? I know the wording was a bit vague on using this value and there was discussion on the technical reflector about it, but I haven't had time to read into where they're going with it. > 2) The --block-count=0 setting in the read/write commands fails when the nvme drive is formatted with a lba data size of 512 bytes. Setting it to 8 seems to work for both 512 and 4k bytes lba data size. Oops, the regress script has hard-coded an expected 4k LBA format. We need it to discover and be re-active to the format instead.