On 03/08/2018 12:50 AM, Paolo Bonzini wrote:
The NBD spec states that since trim requests can affect disk contents,
then they should allow for FUA semantics just like writes for ensuring
the disk has settled before returning. As bdrv_[co_]pdiscard() does
not (yet?) support a flags argument, we can't pass FUA down the block
layer stack, and must therefore emulate it with a flush at the NBD
layer.
TRIM requests should not need FUA since they're just advisory. On
the other hand, WRITE ZEROES requests need to support FUA.
It looks like qemu's NBD implementation properly supports FUA on WRITE
ZEROES requests. The block layer in bdrv_co_do_pwrite_zeroes checks for
FUA support for both an actual zero request to the driver, as well as
any fallback write requests, and does a followup flush if any driver
request did not support BDRV_REQ_FUA. Even when BDRV_REQ_MAY_UNMAP is
passed through to the driver and allows the zeroes to be written by a
trim, then either that trim must also honor FUA semantics
(block/nbd{,-client}.c does this), or the driver can't advertise FUA
support on zeroes (block/iscsi.c does this) and the block layer fallback
kicks in. So I think we're fine on that front.
Still, while you argue that TRIM is advisory (which I agree), if it does
nothing, then you've (implicitly) honored FUA (that transaction didn't
affect persistent storage, so you didn't have to wait any longer for
anything to land); but if it DOES change the disk contents, then waiting
for that change to land IS worth supporting, hence why the NBD protocol
requires the FUA flag to be honored on trim.
--
Eric Blake, Principal Software Engineer
Red Hat, Inc. +1-919-301-3266
Virtualization: qemu.org | libvirt.org
