On 16.06.2020 14:42, Damien Le Moal wrote:
On 2020/06/16 23:16, Javier González wrote:
On 16.06.2020 12:35, Damien Le Moal wrote:
On 2020/06/16 21:24, Javier González wrote:
On 16.06.2020 14:06, Matias Bjørling wrote:
On 16/06/2020 14.00, Javier González wrote:
On 16.06.2020 13:18, Matias Bjørling wrote:
On 16/06/2020 12.41, Javier González wrote:
On 16.06.2020 08:34, Keith Busch wrote:
Add support for NVM Express Zoned Namespaces (ZNS) Command Set defined
in NVM Express TP4053. Zoned namespaces are discovered based on their
Command Set Identifier reported in the namespaces Namespace
Identification Descriptor list. A successfully discovered Zoned
Namespace will be registered with the block layer as a host managed
zoned block device with Zone Append command support. A namespace that
does not support append is not supported by the driver.
Why are we enforcing the append command? Append is optional on the
current ZNS specification, so we should not make this mandatory in the
implementation. See specifics below.
There is already general support in the kernel for the zone append
command. Feel free to submit patches to emulate the support. It is
outside the scope of this patchset.
It is fine that the kernel supports append, but the ZNS specification
does not impose the implementation for append, so the driver should not
do that either.
ZNS SSDs that choose to leave append as a non-implemented optional
command should not rely on emulated SW support, specially when
traditional writes work very fine for a large part of current ZNS use
cases.
Please, remove this virtual constraint.
The Zone Append command is mandatory for zoned block devices. Please
see https://lwn.net/Articles/818709/ for the background.
I do not see anywhere in the block layer that append is mandatory for
zoned devices. Append is emulated on ZBC, but beyond that there is no
mandatory bits. Please explain.
This is to allow a single write IO path for all types of zoned block device for
higher layers, e.g file systems. The on-going re-work of btrfs zone support for
instance now relies 100% on zone append being supported. That significantly
simplifies the file system support and more importantly remove the need for
locking around block allocation and BIO issuing, allowing to preserve a fully
asynchronous write path that can include workqueues for efficient CPU usage of
things like encryption and compression. Without zone append, file system would
either (1) have to reject these drives that do not support zone append, or (2)
implement 2 different write IO path (slower regular write and zone append). None
of these options are ideal, to say the least.
So the approach is: mandate zone append support for ZNS devices. To allow other
ZNS drives, an emulation similar to SCSI can be implemented, with that emulation
ideally combined to work for both types of drives if possible.
Enforcing QD=1 becomes a problem on devices with large zones. In
a ZNS device that has smaller zones this should not be a problem.
Let's be precise: this is not running the drive at QD=1, it is "at most one
write *request* per zone". If the FS is simultaneously using multiple block
groups mapped to different zones, you will get a total write QD > 1, and as many
reads as you want.
Would you agree that it is possible to have a write path that relies on
QD=1, where the FS / application has the responsibility for enforcing
this? Down the road this QD can be increased if the device is able to
buffer the writes.
Doing QD=1 per zone for writes at the FS layer, that is, at the BIO layer does
not work. This is because BIOs can be as large as the FS wants them to be. Such
large BIO will be split into multiple requests in the block layer, resulting in
more than one write per zone. That is why the zone write locking is at the
scheduler level, between BIO split and request dispatch. That avoids the
multiple requests fragments of a large BIO to be reordered and fail. That is
mandatory as the block layer itself can occasionally reorder requests and lower
levels such as AHCI HW is also notoriously good at reversing sequential
requests. For NVMe with multi-queue, the IO issuing process getting rescheduled
on a different CPU can result in sequential IOs being in different queues, with
the likely result of an out-of-order execution. All cases are avoided with zone
write locking and at most one write request dispatch per zone as recommended by
the ZNS specifications (ZBC and ZAC standards for SMR HDDs are silent on this).
I understand. I agree that the current FSs supporting ZNS follow this
approach and it makes sense that there is a common interface that
simplifies the FS implementation. See the comment below on the part I
believe we see things differently.
I would be OK with some FS implementations to rely on append and impose
the constraint that append has to be supported (and it would be our job
to change that), but I would like to avoid the driver rejecting
initializing the device because current FS implementations have
implemented this logic.
What is the difference between the driver rejecting drives and the FS rejecting
the same drives ? That has the same end result to me: an entire class of devices
cannot be used as desired by the user. Implementing zone append emulation avoids
the rejection entirely while still allowing the FS to have a single write IO
path, thus simplifying the code.
The difference is that users that use a raw ZNS device submitting I/O
through the kernel would still be able to use these devices. The result
would be that the ZNS SSD is recognized and initialized, but the FS
format fails.
We can agree that a number of initial customers will use these devices
raw, using the in-kernel I/O path, but without a FS on top.
Thoughts?
and note that
this emulation would require the drive to be operated with mq-deadline to enable
zone write locking for preserving write command order. While on a HDD the
performance penalty is minimal, it will likely be significant on a SSD.
Exactly my concern. I do not want ZNS SSDs to be impacted by this type
of design decision at the driver level.
But your proposed FS level approach would end up doing the exact same thing with
the same limitation and so the same potential performance impact. The block
layer generic approach has the advantage that we do not bother the higher levels
with the implementation of in-order request dispatch guarantees. File systems
are complex enough. The less complexity is required for zone support, the better.
This depends very much on how the FS / application is managing
stripping. At the moment our main use case is enabling user-space
applications submitting I/Os to raw ZNS devices through the kernel.
Can we enable this use case to start with?
Thanks,
Javier
