fdisk(8) vs gpart(8), and gnop

Justin T. Gibbs gibbs at scsiguy.com
Mon Jun 2 17:15:53 UTC 2014


On Jun 2, 2014, at 9:49 AM, Mark Felder <feld at FreeBSD.org> wrote:

> On 2014-06-02 10:02, Nathan Whitehorn wrote:
>> My bigger concern is this pool upgrade one -- what if someone puts in
>> a 4K disk in the future?
> 
> This is a concern of mine, and I sort of wish we did 4k by default and forced people to override if they want 512b or something else.

Adding a 4k sectored device is fine.  You just need to use it in a new top-level vdev in the pool.

If you are at the point where you can’t get new or compatible warrantee replacements for the drives that may fail in your existing pool, you should be migrating your data to new devices anyway.  Mixing devices with different performance characteristics within a TLV can lead to pessimal behavior.  I don’t think that ZFS should jump through large hoops to try and make this work well.  Instead, we should encourage the use of similar devices within a TLV (guidance that the installer has sufficient information to provide*) and the system should be optimized assuming this is how it will be used.

I certainly *do not* want FreeBSD to automatically inflate the ashift used on my pools.  Doing so is an attempt to guess why I chose the devices I did at pool creation time and my strategy for retiring them in the future.  The current proposal guesses wrong for me and the products I help build.  I’d bet it will be wrong more times than right.

—
Justin

*) Using the tools already in FreeBSD it is quite easy to group devices by transport type, capacity, logical block size, physical block size, and, for at least SCSI transports, media rotational speed.  We do this in Spectra’s ZFS appliance so users have to work really hard to mix devices that they shouldn’t.


More information about the freebsd-fs mailing list