On Mon, Aug 22, 2016 at 8:25 PM, Damien Le Moal [off-list ref] wrote:

Shaun,

On 8/23/16 09:22, Shaun Tancheff wrote:

quoted

On Mon, Aug 22, 2016 at 6:57 PM, Damien Le Moal [off-list ref] wrote:

quoted

Also you may note that in my patch to get Host Aware working
with the zone cache I do not include the runt zone in the cache.

Why not ? The RB-tree will handle it just fine (the insert and lookup
code as Hannes had them was not relying on a constant zone size).

A good point. I didn't pay too much attention while brining this
forward. I think a few of my hacks may be pointless now. I'll
try to rework it and get rid of the runt check.

quoted

So as it sits I need this fallback otherwise doing blkdiscard over
the whole device ends in a error, as well as mkfs.f2fs et. al.

Got it, but I do not see a problem with including it. I have not checked
the code, but the split of a big discard call into "chunks" should be
already handling the last chunk and make sure that the operation does
not exceed the device capacity (in any case, that's easy to fix in the
sd_zbc_setup_discard code).

Yes I agree the split of big discards does handle the last chunk correctly.

quoted

quoted

Some 10TB host managed disks out there have 1% conventional zone space,
that is 100GB of capacity. When issuing a "reset all", doing a write
same in these zones will take forever... If the user really wants zeroes
in those zones, let it issue a zeroout.

I think that it would a better choice to simply not report
discard_zeroes_data as true and do nothing for conventional zones reset.

I think that would be unfortunate for Host Managed but I think it's
the right choice for Host Aware at this time. So either we base
it on disk type or we have some other config flag added to sysfs.

I do not see any difference between host managed and host aware. Both
define the same behavior for reset, and both end up in a NOP for
conventional zone reset (no data "erasure" required by the standard).
For write pointer zones, reading unwritten LBAs returns the
initialization pattern, with the exception of host-managed disks with
the URSWRZ bit set to 0. But that case is covered in sd.c, so the
behavior is consistent across all models. So why forcing data zeroing
when the standards do not mandate it ?

Well you do have point.
It appears to be only mkfs and similar tools that are really utilizing
discard zeros data at the moment.

I did a quick test:

mkfs -t ext4 -b 4096 -g 32768 -G 32  \
 -E lazy_itable_init=0,lazy_journal_init=0,offset=0,num_backup_sb=0,packed_meta_blocks=1,discard
  \
 -O flex_bg,extent,sparse_super2

   - discard zeroes data true - 3 minutess
   - discard zeroes data false - 6 minutes
So for the smaller conventional space on the current HA drive
there is some advantage to enabling discard zeroes data.

However for a larger conventional space you are correct the overall
impact is worse performance.

For some reason I had been assuming that some file systems
used or relied on discard zeroes data during normal operation.
Now that I am looking for that I don't seem to be finding any
evidence of it, so aside from mkfs I don't have as good an
argument discard zeroes data as I though I did.

Regards,
Shaun