[dm-devel] [PATCH] dm-raid: add RAID discard support

[Martin K. Petersen]

>>>>> "Neil" == NeilBrown  <neilb at suse.de> writes:

Neil,

Several people have pointed out the issues with the reliability of
discard_zeroes_data in the past. My objection with Heinz' patch is
purely that we are perpetuating an interface that we have come to learn
that we can't entirely trust.

Neil> If it is a "hint", then why isn't it
Neil> "discard_sometimes_zeroes_data"?

Because at the time we put it in it was expected to be a solid
guarantee. However, as it turns out the spec is vaguely worded and only
says that the device must return zeroes if a block has been successfully
discarded. But since discard itself is only a hint the device is free to
silently ignore all or parts of the request. *sigh*

There were several approaches being entertained in both T10 and T13
about how to tighten this up, including a FORCE bit that would cause the
device to return an error if parts of the request were being ignored. I
was hoping that we would be able to use that. However, the consensus at
the T10 meeting this summer was that WRITE SAME is the only reliable way
to ensure that zeroed blocks are returned. And the spec was amended to
clarify that.

Neil> If a read from a discarded region doesn't reliably return zeros,
Neil> then raid5 cannot support discard.  Should I turn of discard
Neil> support in raid5?

I've been on the fence about this for a while.

The good news is that while the spec has been vague and many battles
fought in the standards bodies, most vendors are reasonably sane when it
comes to implementing this. We haven't had any corruption horror stories
from the field that I'm aware of. The first couple of generations of
SSDs were a travesty in this department but current devices behave
reasonably well. Hardware RAID vendors generally solve the RAID5 SSD
problem by explicitly whitelisting drives that are known to do the right
thing. Sadly we don't have a good device list. And for most drives the
discards only get dropped in rare corner cases so it's not trivial to
write a test tool to generate such a list.

I was really hoping we'd get better guarantees from the standards folks
this summer. But as the spec is currently written switching to a WRITE
SAME bias in our discard heuristics may have the opposite effect of what
we desire. That's because the spec mandates that if a device can not
discard the blocks in the request it must write zeroes to them. IOW, we
get the determinism at the expense of potentially ending up writing the
blocks in question, thus increasing write wear. Probably not what we
want for the discard use case.

I am in the process of tweaking the discard code to adhere to the August
27th T10 SBC-4 draft. The net takeaway is that if you depend on reading
zeroes back you should be using blkdev_issue_zeroout() and not
blkdev_issue_discard(). raid5.c effectively needs to translate discard
requests from above to zeroouts below.

I am torn between making blkdev_issue_zeroout() unprovision if possible
or having an explicit allocate flag. Some users only care about getting
zeroes back but others use blkdev_issue_zeroout() to preallocate blocks
(which is the current behavior).

Right now we have:

  discard		Unprovision, may or may not return zeroes
  zeroout		Provision, return zeroes

We'll have to move to a model where we have a third variant:

  discard		Unprovision, may or may not return zeroes
  zeroout, deallocate	Unprovision if possible, return zeroes
  zeroout, allocate     Provision, return zeroes

The zeroout, deallocate variant would be enabled if the device supports
WRITE SAME w/UNMAP and has the LBPRZ flag set (and is not our libata
SATL).

Discard is the bane of my existence :(

-- 
Martin K. Petersen	Oracle Linux Engineering