FYI, I really don't think this primitive is a good idea. In the
concept of non-overwritable storage (NAND, SMR drives) the entire
concept of a one-shoot 'provisioning' that will guarantee later writes
are always possible is simply bogus.
On Fri, May 19 2023 at 12:09P -0400,
Christoph Hellwig <[email protected]> wrote:
> FYI, I really don't think this primitive is a good idea. In the
> concept of non-overwritable storage (NAND, SMR drives) the entire
> concept of a one-shoot 'provisioning' that will guarantee later writes
> are always possible is simply bogus.
Valid point for sure, such storage shouldn't advertise support (and
will return -EOPNOTSUPP).
But the primitive still has utility for other classes of storage.
On Fri, May 19, 2023 at 10:41:31AM -0400, Mike Snitzer wrote:
> On Fri, May 19 2023 at 12:09P -0400,
> Christoph Hellwig <[email protected]> wrote:
>
> > FYI, I really don't think this primitive is a good idea. In the
> > concept of non-overwritable storage (NAND, SMR drives) the entire
> > concept of a one-shoot 'provisioning' that will guarantee later writes
> > are always possible is simply bogus.
>
> Valid point for sure, such storage shouldn't advertise support (and
> will return -EOPNOTSUPP).
>
> But the primitive still has utility for other classes of storage.
Yet the thing people are wanting to us filesystem developers to use
this with is thinly provisioned storage that has snapshot
capability. That, by definition, is non-overwritable storage. These
are the use cases people are asking filesystes to gracefully handle
and report errors when the sparse backing store runs out of space.
e.g. journal writes after a snapshot is taken on a busy filesystem
are always an overwrite and this requires more space in the storage
device for the write to succeed. ENOSPC from the backing device for
journal IO is a -fatal error-. Hence if REQ_PROVISION doesn't
guarantee space for overwrites after snapshots, then it's not
actually useful for solving the real world use cases we actually
need device-level provisioning to solve.
It is not viable for filesystems to have to reprovision space for
in-place metadata overwrites after every snapshot - the filesystem
may not even know a snapshot has been taken! And it's not feasible
for filesystems to provision on demand before they modify metadata
because we don't know what metadata is going to need to be modified
before we start modifying metadata in transactions. If we get ENOSPC
from provisioning in the middle of a dirty transcation, it's all
over just the same as if we get ENOSPC during metadata writeback...
Hence what filesystems actually need is device provisioned space to
be -always over-writable- without ENOSPC occurring. Ideally, if we
provision a range of the block device, the block device *must*
guarantee all future writes to that LBA range succeeds. That
guarantee needs to stand until we discard or unmap the LBA range,
and for however many writes we do to that LBA range.
e.g. If the device takes a snapshot, it needs to reprovision the
potential COW ranges that overlap with the provisioned LBA range at
snapshot time. e.g. by re-reserving the space from the backing pool
for the provisioned space so if a COW occurs there is space
guaranteed for it to succeed. If there isn't space in the backing
pool for the reprovisioning, then whatever operation that triggers
the COW behaviour should fail with ENOSPC before doing anything
else....
Software devices like dm-thin/snapshot should really only need to
keep a persistent map of the provisioned space and refresh space
reservations for used space within that map whenever something that
triggers COW behaviour occurs. i.e. a snapshot needs to reset the
provisioned ranges back to "all ranges are freshly provisioned"
before the snapshot is started. If that space is not available in
the backing pool, then the snapshot attempt gets ENOSPC....
That means filesystems only need to provision space for journals and
fixed metadata at mkfs time, and they only need issue a
REQ_PROVISION bio when they first allocate over-write in place
metadata. We already have online discard and/or fstrim for releasing
provisioned space via discards.
This will require some mods to filesystems like ext4 and XFS to
issue REQ_PROVISION and fail gracefully during metadata allocation.
However, doing so means that we can actually harden filesystems
against sparse block device ENOSPC errors by ensuring they will
never occur in critical filesystem structures....
-Dave.
--
Dave Chinner
[email protected]
On Fri, May 19 2023 at 7:07P -0400,
Dave Chinner <[email protected]> wrote:
> On Fri, May 19, 2023 at 10:41:31AM -0400, Mike Snitzer wrote:
> > On Fri, May 19 2023 at 12:09P -0400,
> > Christoph Hellwig <[email protected]> wrote:
> >
> > > FYI, I really don't think this primitive is a good idea. In the
> > > concept of non-overwritable storage (NAND, SMR drives) the entire
> > > concept of a one-shoot 'provisioning' that will guarantee later writes
> > > are always possible is simply bogus.
> >
> > Valid point for sure, such storage shouldn't advertise support (and
> > will return -EOPNOTSUPP).
> >
> > But the primitive still has utility for other classes of storage.
>
> Yet the thing people are wanting to us filesystem developers to use
> this with is thinly provisioned storage that has snapshot
> capability. That, by definition, is non-overwritable storage. These
> are the use cases people are asking filesystes to gracefully handle
> and report errors when the sparse backing store runs out of space.
DM thinp falls into this category but as you detailed it can be made
to work reliably. To carry that forward we need to first establish
the REQ_PROVISION primitive (with this series).
Follow-on associated dm-thinp enhancements can then serve as reference
for how to take advantage of XFS's ability to operate reliably of
thinly provisioned storage.
> e.g. journal writes after a snapshot is taken on a busy filesystem
> are always an overwrite and this requires more space in the storage
> device for the write to succeed. ENOSPC from the backing device for
> journal IO is a -fatal error-. Hence if REQ_PROVISION doesn't
> guarantee space for overwrites after snapshots, then it's not
> actually useful for solving the real world use cases we actually
> need device-level provisioning to solve.
>
> It is not viable for filesystems to have to reprovision space for
> in-place metadata overwrites after every snapshot - the filesystem
> may not even know a snapshot has been taken! And it's not feasible
> for filesystems to provision on demand before they modify metadata
> because we don't know what metadata is going to need to be modified
> before we start modifying metadata in transactions. If we get ENOSPC
> from provisioning in the middle of a dirty transcation, it's all
> over just the same as if we get ENOSPC during metadata writeback...
>
> Hence what filesystems actually need is device provisioned space to
> be -always over-writable- without ENOSPC occurring. Ideally, if we
> provision a range of the block device, the block device *must*
> guarantee all future writes to that LBA range succeeds. That
> guarantee needs to stand until we discard or unmap the LBA range,
> and for however many writes we do to that LBA range.
>
> e.g. If the device takes a snapshot, it needs to reprovision the
> potential COW ranges that overlap with the provisioned LBA range at
> snapshot time. e.g. by re-reserving the space from the backing pool
> for the provisioned space so if a COW occurs there is space
> guaranteed for it to succeed. If there isn't space in the backing
> pool for the reprovisioning, then whatever operation that triggers
> the COW behaviour should fail with ENOSPC before doing anything
> else....
Happy to implement this in dm-thinp. Each thin block will need a bit
to say if the block must be REQ_PROVISION'd at time of snapshot (and
the resulting block will need the same bit set).
Walking all blocks of a thin device and triggering REQ_PROVISION for
each will obviously make thin snapshot creation take more time.
I think this approach is better than having a dedicated bitmap hooked
off each thin device's metadata (with bitmap being copied and walked
at the time of snapshot). But we'll see... I'll get with Joe to
discuss further.
> Software devices like dm-thin/snapshot should really only need to
> keep a persistent map of the provisioned space and refresh space
> reservations for used space within that map whenever something that
> triggers COW behaviour occurs. i.e. a snapshot needs to reset the
> provisioned ranges back to "all ranges are freshly provisioned"
> before the snapshot is started. If that space is not available in
> the backing pool, then the snapshot attempt gets ENOSPC....
>
> That means filesystems only need to provision space for journals and
> fixed metadata at mkfs time, and they only need issue a
> REQ_PROVISION bio when they first allocate over-write in place
> metadata. We already have online discard and/or fstrim for releasing
> provisioned space via discards.
>
> This will require some mods to filesystems like ext4 and XFS to
> issue REQ_PROVISION and fail gracefully during metadata allocation.
> However, doing so means that we can actually harden filesystems
> against sparse block device ENOSPC errors by ensuring they will
> never occur in critical filesystem structures....
Yes, let's finally _do_ this! ;)
Mike
On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> On Fri, May 19 2023 at 7:07P -0400,
> Dave Chinner <[email protected]> wrote:
>
> > On Fri, May 19, 2023 at 10:41:31AM -0400, Mike Snitzer wrote:
> > > On Fri, May 19 2023 at 12:09P -0400,
> > > Christoph Hellwig <[email protected]> wrote:
> > >
> > > > FYI, I really don't think this primitive is a good idea. In the
> > > > concept of non-overwritable storage (NAND, SMR drives) the entire
> > > > concept of a one-shoot 'provisioning' that will guarantee later writes
> > > > are always possible is simply bogus.
> > >
> > > Valid point for sure, such storage shouldn't advertise support (and
> > > will return -EOPNOTSUPP).
> > >
> > > But the primitive still has utility for other classes of storage.
> >
> > Yet the thing people are wanting to us filesystem developers to use
> > this with is thinly provisioned storage that has snapshot
> > capability. That, by definition, is non-overwritable storage. These
> > are the use cases people are asking filesystes to gracefully handle
> > and report errors when the sparse backing store runs out of space.
>
> DM thinp falls into this category but as you detailed it can be made
> to work reliably. To carry that forward we need to first establish
> the REQ_PROVISION primitive (with this series).
>
> Follow-on associated dm-thinp enhancements can then serve as reference
> for how to take advantage of XFS's ability to operate reliably of
> thinly provisioned storage.
>
> > e.g. journal writes after a snapshot is taken on a busy filesystem
> > are always an overwrite and this requires more space in the storage
> > device for the write to succeed. ENOSPC from the backing device for
> > journal IO is a -fatal error-. Hence if REQ_PROVISION doesn't
> > guarantee space for overwrites after snapshots, then it's not
> > actually useful for solving the real world use cases we actually
> > need device-level provisioning to solve.
> >
> > It is not viable for filesystems to have to reprovision space for
> > in-place metadata overwrites after every snapshot - the filesystem
> > may not even know a snapshot has been taken! And it's not feasible
> > for filesystems to provision on demand before they modify metadata
> > because we don't know what metadata is going to need to be modified
> > before we start modifying metadata in transactions. If we get ENOSPC
> > from provisioning in the middle of a dirty transcation, it's all
> > over just the same as if we get ENOSPC during metadata writeback...
> >
> > Hence what filesystems actually need is device provisioned space to
> > be -always over-writable- without ENOSPC occurring. Ideally, if we
> > provision a range of the block device, the block device *must*
> > guarantee all future writes to that LBA range succeeds. That
> > guarantee needs to stand until we discard or unmap the LBA range,
> > and for however many writes we do to that LBA range.
> >
> > e.g. If the device takes a snapshot, it needs to reprovision the
> > potential COW ranges that overlap with the provisioned LBA range at
> > snapshot time. e.g. by re-reserving the space from the backing pool
> > for the provisioned space so if a COW occurs there is space
> > guaranteed for it to succeed. If there isn't space in the backing
> > pool for the reprovisioning, then whatever operation that triggers
> > the COW behaviour should fail with ENOSPC before doing anything
> > else....
>
> Happy to implement this in dm-thinp. Each thin block will need a bit
> to say if the block must be REQ_PROVISION'd at time of snapshot (and
> the resulting block will need the same bit set).
>
> Walking all blocks of a thin device and triggering REQ_PROVISION for
> each will obviously make thin snapshot creation take more time.
>
> I think this approach is better than having a dedicated bitmap hooked
> off each thin device's metadata (with bitmap being copied and walked
> at the time of snapshot). But we'll see... I'll get with Joe to
> discuss further.
>
Hi Mike,
If you recall our most recent discussions on this topic, I was thinking
about the prospect of reserving the entire volume at mount time as an
initial solution to this problem. When looking through some of the old
reservation bits we prototyped years ago, it occurred to me that we have
enough mechanism to actually prototype this.
So FYI, I have some hacky prototype code that essentially has the
filesystem at mount time tell dm it's using the volume and expects all
further writes to succeed. dm-thin acquires reservation for the entire
range of the volume for which writes would require block allocation
(i.e., holes and shared dm blocks) or otherwise warns that the fs cannot
be "safely" mounted.
The reservation pool associates with the thin volume (not the
filesystem), so if a snapshot is requested from dm, the snapshot request
locates the snapshot origin and if it's currently active, increases the
reservation pool to account for outstanding blocks that are about to
become shared, or otherwise fails the snapshot with -ENOSPC. (I suspect
discard needs similar treatment, but I hadn't got to that yet.). If the
fs is not active, there is nothing to protect and so the snapshot
proceeds as normal.
This seems to work on my simple, initial tests for protecting actively
mounted filesystems from dm-thin -ENOSPC. This definitely needs a sanity
check from dm-thin folks, however, because I don't know enough about the
broader subsystem to reason about whether it's sufficiently correct. I
just managed to beat the older prototype code into submission to get it
to do what I wanted on simple experiments.
Thoughts on something like this? I think the main advantage is that it
significantly reduces the requirements on the fs to track individual
allocations. It's basically an on/off switch from the fs perspective,
doesn't require any explicit provisioning whatsoever (though it can be
done to improve things in the future) and in fact could probably be tied
to thin volume activation to be made completely filesystem agnostic.
Another advantage is that it requires no on-disk changes, no breaking
COWs up front during snapshots, etc.
The disadvantages are that it's space inefficient wrt to thin pool free
space, but IIUC this is essentially what userspace management layers
(such as Stratis) are doing today, they just put restrictions up front
at volume configuration/creation time instead of at runtime. There also
needs to be some kind of interface between the fs and dm. I suppose we
could co-opt provision and discard primitives with a "reservation"
modifier flag to get around that in a simple way, but that sounds
potentially ugly. TBH, the more I think about this the more I think it
makes sense to reserve on volume activation (with some caveats to allow
a read-only mode, explicit bypass, etc.) and then let the
cross-subsystem interface be dictated by granularity improvements...
... since I also happen to think there is a potentially interesting
development path to make this sort of reserve pool configurable in terms
of size and active/inactive state, which would allow the fs to use an
emergency pool scheme for managing metadata provisioning and not have to
track and provision individual metadata buffers at all (dealing with
user data is much easier to provision explicitly). So the space
inefficiency thing is potentially just a tradeoff for simplicity, and
filesystems that want more granularity for better behavior could achieve
that with more work. Filesystems that don't would be free to rely on the
simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
protection with very minimal changes.
That's getting too far into the weeds on the future bits, though. This
is essentially 99% a dm-thin approach, so I'm mainly curious if there's
sufficient interest in this sort of "reserve mode" approach to try and
clean it up further and have dm guys look at it, or if you guys see any
obvious issues in what it does that makes it potentially problematic, or
if you would just prefer to go down the path described above...
Brian
> > Software devices like dm-thin/snapshot should really only need to
> > keep a persistent map of the provisioned space and refresh space
> > reservations for used space within that map whenever something that
> > triggers COW behaviour occurs. i.e. a snapshot needs to reset the
> > provisioned ranges back to "all ranges are freshly provisioned"
> > before the snapshot is started. If that space is not available in
> > the backing pool, then the snapshot attempt gets ENOSPC....
> >
> > That means filesystems only need to provision space for journals and
> > fixed metadata at mkfs time, and they only need issue a
> > REQ_PROVISION bio when they first allocate over-write in place
> > metadata. We already have online discard and/or fstrim for releasing
> > provisioned space via discards.
> >
> > This will require some mods to filesystems like ext4 and XFS to
> > issue REQ_PROVISION and fail gracefully during metadata allocation.
> > However, doing so means that we can actually harden filesystems
> > against sparse block device ENOSPC errors by ensuring they will
> > never occur in critical filesystem structures....
>
> Yes, let's finally _do_ this! ;)
>
> Mike
>
On Tue, May 23 2023 at 10:05P -0400,
Brian Foster <[email protected]> wrote:
> On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > On Fri, May 19 2023 at 7:07P -0400,
> > Dave Chinner <[email protected]> wrote:
> >
...
> > > e.g. If the device takes a snapshot, it needs to reprovision the
> > > potential COW ranges that overlap with the provisioned LBA range at
> > > snapshot time. e.g. by re-reserving the space from the backing pool
> > > for the provisioned space so if a COW occurs there is space
> > > guaranteed for it to succeed. If there isn't space in the backing
> > > pool for the reprovisioning, then whatever operation that triggers
> > > the COW behaviour should fail with ENOSPC before doing anything
> > > else....
> >
> > Happy to implement this in dm-thinp. Each thin block will need a bit
> > to say if the block must be REQ_PROVISION'd at time of snapshot (and
> > the resulting block will need the same bit set).
> >
> > Walking all blocks of a thin device and triggering REQ_PROVISION for
> > each will obviously make thin snapshot creation take more time.
> >
> > I think this approach is better than having a dedicated bitmap hooked
> > off each thin device's metadata (with bitmap being copied and walked
> > at the time of snapshot). But we'll see... I'll get with Joe to
> > discuss further.
> >
>
> Hi Mike,
>
> If you recall our most recent discussions on this topic, I was thinking
> about the prospect of reserving the entire volume at mount time as an
> initial solution to this problem. When looking through some of the old
> reservation bits we prototyped years ago, it occurred to me that we have
> enough mechanism to actually prototype this.
>
> So FYI, I have some hacky prototype code that essentially has the
> filesystem at mount time tell dm it's using the volume and expects all
> further writes to succeed. dm-thin acquires reservation for the entire
> range of the volume for which writes would require block allocation
> (i.e., holes and shared dm blocks) or otherwise warns that the fs cannot
> be "safely" mounted.
>
> The reservation pool associates with the thin volume (not the
> filesystem), so if a snapshot is requested from dm, the snapshot request
> locates the snapshot origin and if it's currently active, increases the
> reservation pool to account for outstanding blocks that are about to
> become shared, or otherwise fails the snapshot with -ENOSPC. (I suspect
> discard needs similar treatment, but I hadn't got to that yet.). If the
> fs is not active, there is nothing to protect and so the snapshot
> proceeds as normal.
>
> This seems to work on my simple, initial tests for protecting actively
> mounted filesystems from dm-thin -ENOSPC. This definitely needs a sanity
> check from dm-thin folks, however, because I don't know enough about the
> broader subsystem to reason about whether it's sufficiently correct. I
> just managed to beat the older prototype code into submission to get it
> to do what I wanted on simple experiments.
Feel free to share what you have.
But my initial gut on the approach is: why even use thin provisioning
at all if you're just going to reserve the entire logical address
space of each thin device?
> Thoughts on something like this? I think the main advantage is that it
> significantly reduces the requirements on the fs to track individual
> allocations. It's basically an on/off switch from the fs perspective,
> doesn't require any explicit provisioning whatsoever (though it can be
> done to improve things in the future) and in fact could probably be tied
> to thin volume activation to be made completely filesystem agnostic.
> Another advantage is that it requires no on-disk changes, no breaking
> COWs up front during snapshots, etc.
I'm just really unclear on the details without seeing it.
You shared a roll-up of the code we did from years ago so I can kind
of imagine the nature of the changes. I'm concerned about snapshots,
and the implicit need to compound the reservation for each snapshot.
> The disadvantages are that it's space inefficient wrt to thin pool free
> space, but IIUC this is essentially what userspace management layers
> (such as Stratis) are doing today, they just put restrictions up front
> at volume configuration/creation time instead of at runtime. There also
> needs to be some kind of interface between the fs and dm. I suppose we
> could co-opt provision and discard primitives with a "reservation"
> modifier flag to get around that in a simple way, but that sounds
> potentially ugly. TBH, the more I think about this the more I think it
> makes sense to reserve on volume activation (with some caveats to allow
> a read-only mode, explicit bypass, etc.) and then let the
> cross-subsystem interface be dictated by granularity improvements...
It just feels imprecise to the point of being both excessive and
nebulous.
thin devices, and snapshots of them, can be active without associated
filesystem mounts being active. It just takes a single origin volume
to be mounted, with a snapshot active, to force thin blocks' sharing
to be broken.
> ... since I also happen to think there is a potentially interesting
> development path to make this sort of reserve pool configurable in terms
> of size and active/inactive state, which would allow the fs to use an
> emergency pool scheme for managing metadata provisioning and not have to
> track and provision individual metadata buffers at all (dealing with
> user data is much easier to provision explicitly). So the space
> inefficiency thing is potentially just a tradeoff for simplicity, and
> filesystems that want more granularity for better behavior could achieve
> that with more work. Filesystems that don't would be free to rely on the
> simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> protection with very minimal changes.
>
> That's getting too far into the weeds on the future bits, though. This
> is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> sufficient interest in this sort of "reserve mode" approach to try and
> clean it up further and have dm guys look at it, or if you guys see any
> obvious issues in what it does that makes it potentially problematic, or
> if you would just prefer to go down the path described above...
The model that Dave detailed, which builds on REQ_PROVISION and is
sticky (by provisioning same blocks for snapshot) seems more useful to
me because it is quite precise. That said, it doesn't account for
hard requirements that _all_ blocks will always succeed. I'm really
not sure we need to go to your extreme (even though stratis
has.. difference is they did so as a crude means to an end because the
existing filesystem code can easily get caught out by -ENOSPC at
exactly the wrong time).
Mike
> > > Software devices like dm-thin/snapshot should really only need to
> > > keep a persistent map of the provisioned space and refresh space
> > > reservations for used space within that map whenever something that
> > > triggers COW behaviour occurs. i.e. a snapshot needs to reset the
> > > provisioned ranges back to "all ranges are freshly provisioned"
> > > before the snapshot is started. If that space is not available in
> > > the backing pool, then the snapshot attempt gets ENOSPC....
> > >
> > > That means filesystems only need to provision space for journals and
> > > fixed metadata at mkfs time, and they only need issue a
> > > REQ_PROVISION bio when they first allocate over-write in place
> > > metadata. We already have online discard and/or fstrim for releasing
> > > provisioned space via discards.
> > >
> > > This will require some mods to filesystems like ext4 and XFS to
> > > issue REQ_PROVISION and fail gracefully during metadata allocation.
> > > However, doing so means that we can actually harden filesystems
> > > against sparse block device ENOSPC errors by ensuring they will
> > > never occur in critical filesystem structures....
> >
> > Yes, let's finally _do_ this! ;)
> >
> > Mike
> >
>
> --
> dm-devel mailing list
> [email protected]
> https://listman.redhat.com/mailman/listinfo/dm-devel
>
On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > ... since I also happen to think there is a potentially interesting
> > development path to make this sort of reserve pool configurable in terms
> > of size and active/inactive state, which would allow the fs to use an
> > emergency pool scheme for managing metadata provisioning and not have to
> > track and provision individual metadata buffers at all (dealing with
> > user data is much easier to provision explicitly). So the space
> > inefficiency thing is potentially just a tradeoff for simplicity, and
> > filesystems that want more granularity for better behavior could achieve
> > that with more work. Filesystems that don't would be free to rely on the
> > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > protection with very minimal changes.
> >
> > That's getting too far into the weeds on the future bits, though. This
> > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > sufficient interest in this sort of "reserve mode" approach to try and
> > clean it up further and have dm guys look at it, or if you guys see any
> > obvious issues in what it does that makes it potentially problematic, or
> > if you would just prefer to go down the path described above...
>
> The model that Dave detailed, which builds on REQ_PROVISION and is
> sticky (by provisioning same blocks for snapshot) seems more useful to
> me because it is quite precise. That said, it doesn't account for
> hard requirements that _all_ blocks will always succeed.
Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
but I don't think we'd ever need a hard guarantee from the block
device that every write bio issued from the filesystem will succeed
without ENOSPC.
If the block device can provide a guarantee that a provisioned LBA
range is always writable, then everything else is a filesystem level
optimisation problem and we don't have to involve the block device
in any way. All we need is a flag we can ready out of the bdev at
mount time to determine if the filesystem should be operating with
LBA provisioning enabled...
e.g. If we need to "pre-provision" a chunk of the LBA space for
filesystem metadata, we can do that ahead of time and track the
pre-provisioned range(s) in the filesystem itself.
In XFS, That could be as simple as having small chunks of each AG
reserved to metadata (e.g. start with the first 100MB) and limiting
all metadata allocation free space searches to that specific block
range. When we run low on that space, we pre-provision another 100MB
chunk and then allocate all metadata out of that new range. If we
start getting ENOSPC to pre-provisioning, then we reduce the size of
the regions and log low space warnings to userspace. If we can't
pre-provision any space at all and we've completely run out, we
simply declare ENOSPC for all incoming operations that require
metadata allocation until pre-provisioning succeeds again.
This is built entirely on the premise that once proactive backing
device provisioning fails, the backing device is at ENOSPC and we
have to wait for that situation to go away before allowing new data
to be ingested. Hence the block device really doesn't need to know
anything about what the filesystem is doing and vice versa - The
block dev just says "yes" or "no" and the filesystem handles
everything else.
It's worth noting that XFS already has a coarse-grained
implementation of preferred regions for metadata storage. It will
currently not use those metadata-preferred regions for user data
unless all the remaining user data space is full. Hence I'm pretty
sure that a pre-provisioning enhancment like this can be done
entirely in-memory without requiring any new on-disk state to be
added.
Sure, if we crash and remount, then we might chose a different LBA
region for pre-provisioning. But that's not really a huge deal as we
could also run an internal background post-mount fstrim operation to
remove any unused pre-provisioning that was left over from when the
system went down.
Further, managing shared pool exhaustion doesn't require a
reservation pool in the backing device and for the filesystems to
request space from it. Filesystems already have their own reserve
pools via pre-provisioning. If we want the filesystems to be able to
release that space back to the shared pool (e.g. because the shared
backing pool is critically short on space) then all we need is an
extension to FITRIM to tell the filesystem to also release internal
pre-provisioned reserves.
Then the backing pool admin (person or automated daemon!) can simply
issue a trim on all the filesystems in the pool and spce will be
returned. Then filesystems will ask for new pre-provisioned space
when they next need to ingest modifications, and the backing pool
can manage the new pre-provisioning space requests directly....
Hence I think if we get the basic REQ_PROVISION overwrite-in-place
guarantees defined and implemented as previously outlined, then we
don't need any special coordination between the fs and block devices
to avoid fatal ENOSPC issues with sparse and/or snapshot capable
block devices...
As a bonus, if we can implement the guarantees in dm-thin/-snapshot
and have a filesystem make use of it, then we also have a reference
implementation to point at device vendors and standards
associations....
Cheers,
Dave.
--
Dave Chinner
[email protected]
On Tue, May 23 2023 at 8:40P -0400,
Dave Chinner <[email protected]> wrote:
> On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > ... since I also happen to think there is a potentially interesting
> > > development path to make this sort of reserve pool configurable in terms
> > > of size and active/inactive state, which would allow the fs to use an
> > > emergency pool scheme for managing metadata provisioning and not have to
> > > track and provision individual metadata buffers at all (dealing with
> > > user data is much easier to provision explicitly). So the space
> > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > filesystems that want more granularity for better behavior could achieve
> > > that with more work. Filesystems that don't would be free to rely on the
> > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > protection with very minimal changes.
> > >
> > > That's getting too far into the weeds on the future bits, though. This
> > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > sufficient interest in this sort of "reserve mode" approach to try and
> > > clean it up further and have dm guys look at it, or if you guys see any
> > > obvious issues in what it does that makes it potentially problematic, or
> > > if you would just prefer to go down the path described above...
> >
> > The model that Dave detailed, which builds on REQ_PROVISION and is
> > sticky (by provisioning same blocks for snapshot) seems more useful to
> > me because it is quite precise. That said, it doesn't account for
> > hard requirements that _all_ blocks will always succeed.
>
> Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> but I don't think we'd ever need a hard guarantee from the block
> device that every write bio issued from the filesystem will succeed
> without ENOSPC.
>
> If the block device can provide a guarantee that a provisioned LBA
> range is always writable, then everything else is a filesystem level
> optimisation problem and we don't have to involve the block device
> in any way. All we need is a flag we can ready out of the bdev at
> mount time to determine if the filesystem should be operating with
> LBA provisioning enabled...
>
> e.g. If we need to "pre-provision" a chunk of the LBA space for
> filesystem metadata, we can do that ahead of time and track the
> pre-provisioned range(s) in the filesystem itself.
>
> In XFS, That could be as simple as having small chunks of each AG
> reserved to metadata (e.g. start with the first 100MB) and limiting
> all metadata allocation free space searches to that specific block
> range. When we run low on that space, we pre-provision another 100MB
> chunk and then allocate all metadata out of that new range. If we
> start getting ENOSPC to pre-provisioning, then we reduce the size of
> the regions and log low space warnings to userspace. If we can't
> pre-provision any space at all and we've completely run out, we
> simply declare ENOSPC for all incoming operations that require
> metadata allocation until pre-provisioning succeeds again.
This is basically saying the same thing but:
It could be that the LBA space is fragmented and so falling back to
the smallest region size (that matches the thinp block size) would be
the last resort? Then if/when thinp cannot even service allocating a
new free thin block, dm-thinp will transition to out-of-data-space
mode.
> This is built entirely on the premise that once proactive backing
> device provisioning fails, the backing device is at ENOSPC and we
> have to wait for that situation to go away before allowing new data
> to be ingested. Hence the block device really doesn't need to know
> anything about what the filesystem is doing and vice versa - The
> block dev just says "yes" or "no" and the filesystem handles
> everything else.
Yes.
> It's worth noting that XFS already has a coarse-grained
> implementation of preferred regions for metadata storage. It will
> currently not use those metadata-preferred regions for user data
> unless all the remaining user data space is full. Hence I'm pretty
> sure that a pre-provisioning enhancment like this can be done
> entirely in-memory without requiring any new on-disk state to be
> added.
>
> Sure, if we crash and remount, then we might chose a different LBA
> region for pre-provisioning. But that's not really a huge deal as we
> could also run an internal background post-mount fstrim operation to
> remove any unused pre-provisioning that was left over from when the
> system went down.
This would be the FITRIM with extension you mention below? Which is a
filesystem interface detail? So dm-thinp would _not_ need to have new
state that tracks "provisioned but unused" block? Nor would the block
layer need an extra discard flag for a new class of "provisioned"
blocks.
If XFS tracked this "provisioned but unused" state, dm-thinp could
just discard the block like its told. Would be nice to avoid dm-thinp
needing to track "provisioned but unused".
That said, dm-thinp does still need to know if a block was provisioned
(given our previous designed discussion, to allow proper guarantees
from this interface at snapshot time) so that XFS and other
filesystems don't need to re-provision areas they already
pre-provisioned.
However, it may be that if thinp did track "provisioned but unused"
it'd be useful to allow snapshots to share provisioned blocks that
were never used. Meaning, we could then avoid "breaking sharing" at
snapshot-time for "provisioned but unused" blocks. But allowing this
"optimization" undercuts the gaurantee that XFS needs for thinp
storage that allows snapshots... SO, I think I answered my own
question: thinp doesnt need to track "provisioned but unused" blocks
but we must always ensure snapshots inherit provisoned blocks ;)
> Further, managing shared pool exhaustion doesn't require a
> reservation pool in the backing device and for the filesystems to
> request space from it. Filesystems already have their own reserve
> pools via pre-provisioning. If we want the filesystems to be able to
> release that space back to the shared pool (e.g. because the shared
> backing pool is critically short on space) then all we need is an
> extension to FITRIM to tell the filesystem to also release internal
> pre-provisioned reserves.
So by default FITRIM will _not_ discard provisioned blocks. Only if
a flag is used will it result in discarding provisioned blocks.
My dwelling on this is just double-checking that the
> Then the backing pool admin (person or automated daemon!) can simply
> issue a trim on all the filesystems in the pool and spce will be
> returned. Then filesystems will ask for new pre-provisioned space
> when they next need to ingest modifications, and the backing pool
> can manage the new pre-provisioning space requests directly....
>
> Hence I think if we get the basic REQ_PROVISION overwrite-in-place
> guarantees defined and implemented as previously outlined, then we
> don't need any special coordination between the fs and block devices
> to avoid fatal ENOSPC issues with sparse and/or snapshot capable
> block devices...
>
> As a bonus, if we can implement the guarantees in dm-thin/-snapshot
> and have a filesystem make use of it, then we also have a reference
> implementation to point at device vendors and standards
> associations....
Yeap.
Thanks,
Mike
On Wed, May 24, 2023 at 04:02:49PM -0400, Mike Snitzer wrote:
> On Tue, May 23 2023 at 8:40P -0400,
> Dave Chinner <[email protected]> wrote:
>
> > On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > > ... since I also happen to think there is a potentially interesting
> > > > development path to make this sort of reserve pool configurable in terms
> > > > of size and active/inactive state, which would allow the fs to use an
> > > > emergency pool scheme for managing metadata provisioning and not have to
> > > > track and provision individual metadata buffers at all (dealing with
> > > > user data is much easier to provision explicitly). So the space
> > > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > > filesystems that want more granularity for better behavior could achieve
> > > > that with more work. Filesystems that don't would be free to rely on the
> > > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > > protection with very minimal changes.
> > > >
> > > > That's getting too far into the weeds on the future bits, though. This
> > > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > > sufficient interest in this sort of "reserve mode" approach to try and
> > > > clean it up further and have dm guys look at it, or if you guys see any
> > > > obvious issues in what it does that makes it potentially problematic, or
> > > > if you would just prefer to go down the path described above...
> > >
> > > The model that Dave detailed, which builds on REQ_PROVISION and is
> > > sticky (by provisioning same blocks for snapshot) seems more useful to
> > > me because it is quite precise. That said, it doesn't account for
> > > hard requirements that _all_ blocks will always succeed.
> >
> > Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> > but I don't think we'd ever need a hard guarantee from the block
> > device that every write bio issued from the filesystem will succeed
> > without ENOSPC.
> >
> > If the block device can provide a guarantee that a provisioned LBA
> > range is always writable, then everything else is a filesystem level
> > optimisation problem and we don't have to involve the block device
> > in any way. All we need is a flag we can ready out of the bdev at
> > mount time to determine if the filesystem should be operating with
> > LBA provisioning enabled...
> >
> > e.g. If we need to "pre-provision" a chunk of the LBA space for
> > filesystem metadata, we can do that ahead of time and track the
> > pre-provisioned range(s) in the filesystem itself.
> >
> > In XFS, That could be as simple as having small chunks of each AG
> > reserved to metadata (e.g. start with the first 100MB) and limiting
> > all metadata allocation free space searches to that specific block
> > range. When we run low on that space, we pre-provision another 100MB
> > chunk and then allocate all metadata out of that new range. If we
> > start getting ENOSPC to pre-provisioning, then we reduce the size of
> > the regions and log low space warnings to userspace. If we can't
> > pre-provision any space at all and we've completely run out, we
> > simply declare ENOSPC for all incoming operations that require
> > metadata allocation until pre-provisioning succeeds again.
>
> This is basically saying the same thing but:
>
> It could be that the LBA space is fragmented and so falling back to
> the smallest region size (that matches the thinp block size) would be
> the last resort? Then if/when thinp cannot even service allocating a
> new free thin block, dm-thinp will transition to out-of-data-space
> mode.
Yes, something of that sort, though we'd probably give up if we
can't get at least megabyte scale reservations - a single
modification in XFS can modify many structures and require
allocation of a lot of new metadata, so the fileystem cut-off would
for metadata provisioning failure would be much larger than the
dm-thinp region size....
> > This is built entirely on the premise that once proactive backing
> > device provisioning fails, the backing device is at ENOSPC and we
> > have to wait for that situation to go away before allowing new data
> > to be ingested. Hence the block device really doesn't need to know
> > anything about what the filesystem is doing and vice versa - The
> > block dev just says "yes" or "no" and the filesystem handles
> > everything else.
>
> Yes.
>
> > It's worth noting that XFS already has a coarse-grained
> > implementation of preferred regions for metadata storage. It will
> > currently not use those metadata-preferred regions for user data
> > unless all the remaining user data space is full. Hence I'm pretty
> > sure that a pre-provisioning enhancment like this can be done
> > entirely in-memory without requiring any new on-disk state to be
> > added.
> >
> > Sure, if we crash and remount, then we might chose a different LBA
> > region for pre-provisioning. But that's not really a huge deal as we
> > could also run an internal background post-mount fstrim operation to
> > remove any unused pre-provisioning that was left over from when the
> > system went down.
>
> This would be the FITRIM with extension you mention below? Which is a
> filesystem interface detail?
No. We might reuse some of the internal infrastructure we use to
implement FITRIM, but that's about it. It's just something kinda
like FITRIM but with different constraints determined by the
filesystem rather than the user...
As it is, I'm not sure we'd even need it - a preiodic userspace
FITRIM would acheive the same result, so leaked provisioned spaces
would get cleaned up eventually without the filesystem having to do
anything specific...
> So dm-thinp would _not_ need to have new
> state that tracks "provisioned but unused" block?
No idea - that's your domain. :)
dm-snapshot, for certain, will need to track provisioned regions
because it has to guarantee that overwrites to provisioned space in
the origin device will always succeed. Hence it needs to know how
much space breaking sharing in provisioned regions after a snapshot
has been taken with be required...
> Nor would the block
> layer need an extra discard flag for a new class of "provisioned"
> blocks.
Right, I don't see that the discard operations need to care whether
the underlying storage is provisioned. dm-thinp and dm-snapshot can
treat REQ_OP_DISCARD as "this range is not longer in use" and do
whatever they want with them.
> If XFS tracked this "provisioned but unused" state, dm-thinp could
> just discard the block like its told. Would be nice to avoid dm-thinp
> needing to track "provisioned but unused".
>
> That said, dm-thinp does still need to know if a block was provisioned
> (given our previous designed discussion, to allow proper guarantees
> from this interface at snapshot time) so that XFS and other
> filesystems don't need to re-provision areas they already
> pre-provisioned.
Right.
I've simply assumed that dm-thinp would need to track entire
provisioned regions - used or unused - so it knows which writes to
empty or shared regions have a reservation to allow allocation to
succeed when the backing pool is otherwise empty.....
> However, it may be that if thinp did track "provisioned but unused"
> it'd be useful to allow snapshots to share provisioned blocks that
> were never used. Meaning, we could then avoid "breaking sharing" at
> snapshot-time for "provisioned but unused" blocks. But allowing this
> "optimization" undercuts the gaurantee that XFS needs for thinp
> storage that allows snapshots... SO, I think I answered my own
> question: thinp doesnt need to track "provisioned but unused" blocks
> but we must always ensure snapshots inherit provisoned blocks ;)
Sounds like a potential optimisation, but I haven't thought through
a potential snapshot device implementation that far to comment
sanely. I stopped once I got to the point where accounting tricks
count be used to guarantee space is available for breaking sharing
of used provisioned space after a snapshot was taken....
> > Further, managing shared pool exhaustion doesn't require a
> > reservation pool in the backing device and for the filesystems to
> > request space from it. Filesystems already have their own reserve
> > pools via pre-provisioning. If we want the filesystems to be able to
> > release that space back to the shared pool (e.g. because the shared
> > backing pool is critically short on space) then all we need is an
> > extension to FITRIM to tell the filesystem to also release internal
> > pre-provisioned reserves.
>
> So by default FITRIM will _not_ discard provisioned blocks. Only if
> a flag is used will it result in discarding provisioned blocks.
No. FITRIM results in discard of any unused free space in the
filesystem that matches the criteria set by the user. We don't care
if free space was once provisioned used space - we'll issue a
discard for the range regardless. The "special" FITRIM extension I
mentioned is to get filesystem metadata provisioning released;
that's completely separate to user data provisioning through
fallocate() which FITRIM will always discard if it has been freed...
IOWs, normal behaviour will be that a FITRIM ends up discarding a
mix of unprovisioned and provisioned space. Nobody will be able to
predict what mix the device is going to get at any point in time.
Also, if we turn on online discard, the block device is going to get
a constant stream of discard operations that will also be a mix of
provisioned and unprovisioned space that is not longer in use by the
filesystem.
I suspect that you need to stop trying to double guess what
operations the filesystem will use provisioning for, what it will
send discards for and when it will send discards for them.. Just
assume the device will receive a constant stream of both
REQ_PROVISION and REQ_OP_DISCARD (for both provisioned and
unprovisioned regions) operations whenver the filesystem is active
on a thinp device.....
Cheers,
Dave.
--
Dave Chinner
[email protected]
On Thu, May 25 2023 at 7:39P -0400,
Dave Chinner <[email protected]> wrote:
> On Wed, May 24, 2023 at 04:02:49PM -0400, Mike Snitzer wrote:
> > On Tue, May 23 2023 at 8:40P -0400,
> > Dave Chinner <[email protected]> wrote:
> >
> > > On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > > > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > > > ... since I also happen to think there is a potentially interesting
> > > > > development path to make this sort of reserve pool configurable in terms
> > > > > of size and active/inactive state, which would allow the fs to use an
> > > > > emergency pool scheme for managing metadata provisioning and not have to
> > > > > track and provision individual metadata buffers at all (dealing with
> > > > > user data is much easier to provision explicitly). So the space
> > > > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > > > filesystems that want more granularity for better behavior could achieve
> > > > > that with more work. Filesystems that don't would be free to rely on the
> > > > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > > > protection with very minimal changes.
> > > > >
> > > > > That's getting too far into the weeds on the future bits, though. This
> > > > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > > > sufficient interest in this sort of "reserve mode" approach to try and
> > > > > clean it up further and have dm guys look at it, or if you guys see any
> > > > > obvious issues in what it does that makes it potentially problematic, or
> > > > > if you would just prefer to go down the path described above...
> > > >
> > > > The model that Dave detailed, which builds on REQ_PROVISION and is
> > > > sticky (by provisioning same blocks for snapshot) seems more useful to
> > > > me because it is quite precise. That said, it doesn't account for
> > > > hard requirements that _all_ blocks will always succeed.
> > >
> > > Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> > > but I don't think we'd ever need a hard guarantee from the block
> > > device that every write bio issued from the filesystem will succeed
> > > without ENOSPC.
> > >
> > > If the block device can provide a guarantee that a provisioned LBA
> > > range is always writable, then everything else is a filesystem level
> > > optimisation problem and we don't have to involve the block device
> > > in any way. All we need is a flag we can ready out of the bdev at
> > > mount time to determine if the filesystem should be operating with
> > > LBA provisioning enabled...
> > >
> > > e.g. If we need to "pre-provision" a chunk of the LBA space for
> > > filesystem metadata, we can do that ahead of time and track the
> > > pre-provisioned range(s) in the filesystem itself.
> > >
> > > In XFS, That could be as simple as having small chunks of each AG
> > > reserved to metadata (e.g. start with the first 100MB) and limiting
> > > all metadata allocation free space searches to that specific block
> > > range. When we run low on that space, we pre-provision another 100MB
> > > chunk and then allocate all metadata out of that new range. If we
> > > start getting ENOSPC to pre-provisioning, then we reduce the size of
> > > the regions and log low space warnings to userspace. If we can't
> > > pre-provision any space at all and we've completely run out, we
> > > simply declare ENOSPC for all incoming operations that require
> > > metadata allocation until pre-provisioning succeeds again.
> >
> > This is basically saying the same thing but:
> >
> > It could be that the LBA space is fragmented and so falling back to
> > the smallest region size (that matches the thinp block size) would be
> > the last resort? Then if/when thinp cannot even service allocating a
> > new free thin block, dm-thinp will transition to out-of-data-space
> > mode.
>
> Yes, something of that sort, though we'd probably give up if we
> can't get at least megabyte scale reservations - a single
> modification in XFS can modify many structures and require
> allocation of a lot of new metadata, so the fileystem cut-off would
> for metadata provisioning failure would be much larger than the
> dm-thinp region size....
>
> > > This is built entirely on the premise that once proactive backing
> > > device provisioning fails, the backing device is at ENOSPC and we
> > > have to wait for that situation to go away before allowing new data
> > > to be ingested. Hence the block device really doesn't need to know
> > > anything about what the filesystem is doing and vice versa - The
> > > block dev just says "yes" or "no" and the filesystem handles
> > > everything else.
> >
> > Yes.
> >
> > > It's worth noting that XFS already has a coarse-grained
> > > implementation of preferred regions for metadata storage. It will
> > > currently not use those metadata-preferred regions for user data
> > > unless all the remaining user data space is full. Hence I'm pretty
> > > sure that a pre-provisioning enhancment like this can be done
> > > entirely in-memory without requiring any new on-disk state to be
> > > added.
> > >
> > > Sure, if we crash and remount, then we might chose a different LBA
> > > region for pre-provisioning. But that's not really a huge deal as we
> > > could also run an internal background post-mount fstrim operation to
> > > remove any unused pre-provisioning that was left over from when the
> > > system went down.
> >
> > This would be the FITRIM with extension you mention below? Which is a
> > filesystem interface detail?
>
> No. We might reuse some of the internal infrastructure we use to
> implement FITRIM, but that's about it. It's just something kinda
> like FITRIM but with different constraints determined by the
> filesystem rather than the user...
>
> As it is, I'm not sure we'd even need it - a preiodic userspace
> FITRIM would acheive the same result, so leaked provisioned spaces
> would get cleaned up eventually without the filesystem having to do
> anything specific...
>
> > So dm-thinp would _not_ need to have new
> > state that tracks "provisioned but unused" block?
>
> No idea - that's your domain. :)
>
> dm-snapshot, for certain, will need to track provisioned regions
> because it has to guarantee that overwrites to provisioned space in
> the origin device will always succeed. Hence it needs to know how
> much space breaking sharing in provisioned regions after a snapshot
> has been taken with be required...
dm-thinp offers its own much more scalable snapshot support (doesn't
use old dm-snapshot N-way copyout target).
dm-snapshot isn't going to be modified to support this level of
hardening (dm-snapshot is basically in "maintenance only" now).
But I understand your meaning: what you said is 100% applicable to
dm-thinp's snapshot implementation and needs to be accounted for in
thinp's metadata (inherent 'provisioned' flag).
> > Nor would the block
> > layer need an extra discard flag for a new class of "provisioned"
> > blocks.
>
> Right, I don't see that the discard operations need to care whether
> the underlying storage is provisioned. dm-thinp and dm-snapshot can
> treat REQ_OP_DISCARD as "this range is not longer in use" and do
> whatever they want with them.
>
> > If XFS tracked this "provisioned but unused" state, dm-thinp could
> > just discard the block like its told. Would be nice to avoid dm-thinp
> > needing to track "provisioned but unused".
> >
> > That said, dm-thinp does still need to know if a block was provisioned
> > (given our previous designed discussion, to allow proper guarantees
> > from this interface at snapshot time) so that XFS and other
> > filesystems don't need to re-provision areas they already
> > pre-provisioned.
>
> Right.
>
> I've simply assumed that dm-thinp would need to track entire
> provisioned regions - used or unused - so it knows which writes to
> empty or shared regions have a reservation to allow allocation to
> succeed when the backing pool is otherwise empty.....
>
> > However, it may be that if thinp did track "provisioned but unused"
> > it'd be useful to allow snapshots to share provisioned blocks that
> > were never used. Meaning, we could then avoid "breaking sharing" at
> > snapshot-time for "provisioned but unused" blocks. But allowing this
> > "optimization" undercuts the gaurantee that XFS needs for thinp
> > storage that allows snapshots... SO, I think I answered my own
> > question: thinp doesnt need to track "provisioned but unused" blocks
> > but we must always ensure snapshots inherit provisoned blocks ;)
>
> Sounds like a potential optimisation, but I haven't thought through
> a potential snapshot device implementation that far to comment
> sanely. I stopped once I got to the point where accounting tricks
> count be used to guarantee space is available for breaking sharing
> of used provisioned space after a snapshot was taken....
>
> > > Further, managing shared pool exhaustion doesn't require a
> > > reservation pool in the backing device and for the filesystems to
> > > request space from it. Filesystems already have their own reserve
> > > pools via pre-provisioning. If we want the filesystems to be able to
> > > release that space back to the shared pool (e.g. because the shared
> > > backing pool is critically short on space) then all we need is an
> > > extension to FITRIM to tell the filesystem to also release internal
> > > pre-provisioned reserves.
> >
> > So by default FITRIM will _not_ discard provisioned blocks. Only if
> > a flag is used will it result in discarding provisioned blocks.
>
> No. FITRIM results in discard of any unused free space in the
> filesystem that matches the criteria set by the user. We don't care
> if free space was once provisioned used space - we'll issue a
> discard for the range regardless. The "special" FITRIM extension I
> mentioned is to get filesystem metadata provisioning released;
> that's completely separate to user data provisioning through
> fallocate() which FITRIM will always discard if it has been freed...
>
> IOWs, normal behaviour will be that a FITRIM ends up discarding a
> mix of unprovisioned and provisioned space. Nobody will be able to
> predict what mix the device is going to get at any point in time.
> Also, if we turn on online discard, the block device is going to get
> a constant stream of discard operations that will also be a mix of
> provisioned and unprovisioned space that is not longer in use by the
> filesystem.
>
> I suspect that you need to stop trying to double guess what
> operations the filesystem will use provisioning for, what it will
> send discards for and when it will send discards for them.. Just
> assume the device will receive a constant stream of both
> REQ_PROVISION and REQ_OP_DISCARD (for both provisioned and
> unprovisioned regions) operations whenver the filesystem is active
> on a thinp device.....
Yeah, I was getting tripped up in the weeds a bit. It's pretty
straight-forward (and like I said at the start of our subthread here:
this follow-on work, to inherit provisioned flag, can build on this
REQ_PROVISION patchset).
All said, I've now gotten this sub-thread on Joe Thornber's radar and
we've started discussing. We'll be discussing with more focus
tomorrow.
Mike
On Wed, May 24, 2023 at 10:40:34AM +1000, Dave Chinner wrote:
> On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > ... since I also happen to think there is a potentially interesting
> > > development path to make this sort of reserve pool configurable in terms
> > > of size and active/inactive state, which would allow the fs to use an
> > > emergency pool scheme for managing metadata provisioning and not have to
> > > track and provision individual metadata buffers at all (dealing with
> > > user data is much easier to provision explicitly). So the space
> > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > filesystems that want more granularity for better behavior could achieve
> > > that with more work. Filesystems that don't would be free to rely on the
> > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > protection with very minimal changes.
> > >
> > > That's getting too far into the weeds on the future bits, though. This
> > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > sufficient interest in this sort of "reserve mode" approach to try and
> > > clean it up further and have dm guys look at it, or if you guys see any
> > > obvious issues in what it does that makes it potentially problematic, or
> > > if you would just prefer to go down the path described above...
> >
> > The model that Dave detailed, which builds on REQ_PROVISION and is
> > sticky (by provisioning same blocks for snapshot) seems more useful to
> > me because it is quite precise. That said, it doesn't account for
> > hard requirements that _all_ blocks will always succeed.
>
> Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> but I don't think we'd ever need a hard guarantee from the block
> device that every write bio issued from the filesystem will succeed
> without ENOSPC.
>
The bigger picture goal that I didn't get into in my previous mail is
the "full device" reservation model is intended to be a simple, crude
reference implementation that can be enabled for any arbitrary thin
volume consumer (filesystem or application). The idea is to build that
on a simple enough reservation mechanism that any such consumer could
override it based on its own operational model. The goal is to guarantee
that a particular filesystem never receives -ENOSPC from dm-thin on
writes, but the first phase of implementing that is to simply guarantee
every block is writeable.
As a specific filesystem is able to more explicitly provision its own
allocations in a way that it can guarantee to return -ENOSPC from
dm-thin up front (rather than at write bio time), it can reduce the need
for the amount of reservation required, ultimately to zero if that
filesystem provides the ability to pre-provision all of its writes to
storage in some way or another.
I think for filesystems with complex metadata management like XFS, it's
not very realistic to expect explicit 1-1 provisioning for all metadata
changes on a per-transaction basis in the same way that can (fairly
easily) be done for data, which means a pool mechanism is probably still
needed for the metadata class of writes. Therefore, my expectation for
something like XFS is that it grows the ability to explicitly provision
data writes up front (we solved this part years ago), and then uses a
much smaller pool of reservation for the purpose of dealing with
metadata.
I think what you describe below around preprovisioned perag metadata
ranges is interesting because it _very closely_ maps conceptually to
what I envisioned the evolution of the reserve pool scheme to end up
looking like, but just implemented rather differently to use
reservations instead of specific LBA ranges.
Let me try to connect the dots and identify the differences/tradeoffs...
> If the block device can provide a guarantee that a provisioned LBA
> range is always writable, then everything else is a filesystem level
> optimisation problem and we don't have to involve the block device
> in any way. All we need is a flag we can ready out of the bdev at
> mount time to determine if the filesystem should be operating with
> LBA provisioning enabled...
>
> e.g. If we need to "pre-provision" a chunk of the LBA space for
> filesystem metadata, we can do that ahead of time and track the
> pre-provisioned range(s) in the filesystem itself.
>
> In XFS, That could be as simple as having small chunks of each AG
> reserved to metadata (e.g. start with the first 100MB) and limiting
> all metadata allocation free space searches to that specific block
> range. When we run low on that space, we pre-provision another 100MB
> chunk and then allocate all metadata out of that new range. If we
> start getting ENOSPC to pre-provisioning, then we reduce the size of
> the regions and log low space warnings to userspace. If we can't
> pre-provision any space at all and we've completely run out, we
> simply declare ENOSPC for all incoming operations that require
> metadata allocation until pre-provisioning succeeds again.
>
The more interesting aspect of this is not so much how space is
provisioned and allocated, but how the filesystem is going to consume
that space in a way that guarantees -ENOSPC is provided up front before
userspace is allowed to make modifications. You didn't really touch on
that here, so I'm going to assume we'd have something like a perag
counter of how many free blocks currently live in preprovisioned ranges,
and then an fs-wide total somewhere so a transaction has the ability to
consume these blocks at trans reservation time, the fs knows when to
preprovision more space (or go into -ENOSPC mode), etc.
Some accounting of that nature is necessary here in order to prevent the
filesystem from ever writing to unprovisioned space. So what I was
envisioning is rather than explicitly preprovision a physical range of
each AG and tracking all that, just reserve that number of arbitrarily
located blocks from dm for each AG.
The initial perag reservations can be populated at mount time,
replenished as needed in a very similar way as what you describe, and
100% released back to the thin pool at unmount time. On top of that,
there's no need to track physical preprovisioned ranges at all. Not just
for allocation purposes, but also to avoid things like having to protect
background trims from preprovisioned ranges of free space dedicated for
metadata, etc.
> This is built entirely on the premise that once proactive backing
> device provisioning fails, the backing device is at ENOSPC and we
> have to wait for that situation to go away before allowing new data
> to be ingested. Hence the block device really doesn't need to know
> anything about what the filesystem is doing and vice versa - The
> block dev just says "yes" or "no" and the filesystem handles
> everything else.
>
Yup, everything you describe about going into a simulated -ENOSPC mode
would work exactly the same. The primary difference is that the internal
provisioned space accounting in the filesystem is backed by dynamic
reservation in dm, rather than physically provisioned LBA ranges.
> It's worth noting that XFS already has a coarse-grained
> implementation of preferred regions for metadata storage. It will
> currently not use those metadata-preferred regions for user data
> unless all the remaining user data space is full. Hence I'm pretty
> sure that a pre-provisioning enhancment like this can be done
> entirely in-memory without requiring any new on-disk state to be
> added.
>
> Sure, if we crash and remount, then we might chose a different LBA
> region for pre-provisioning. But that's not really a huge deal as we
> could also run an internal background post-mount fstrim operation to
> remove any unused pre-provisioning that was left over from when the
> system went down.
>
None of this is really needed..
> Further, managing shared pool exhaustion doesn't require a
> reservation pool in the backing device and for the filesystems to
> request space from it. Filesystems already have their own reserve
> pools via pre-provisioning. If we want the filesystems to be able to
> release that space back to the shared pool (e.g. because the shared
> backing pool is critically short on space) then all we need is an
> extension to FITRIM to tell the filesystem to also release internal
> pre-provisioned reserves.
>
> Then the backing pool admin (person or automated daemon!) can simply
> issue a trim on all the filesystems in the pool and spce will be
> returned. Then filesystems will ask for new pre-provisioned space
> when they next need to ingest modifications, and the backing pool
> can manage the new pre-provisioning space requests directly....
>
This is written as to imply that the reservation pool is some big
complex thing, which makes me think there is some
confusion/miscommunication. It's basically just an in memory counter of
space that is allocated out of a shared thin pool and is held in a
specific thin volume while it is currently in use. The counter on the
volume is managed indirectly by filesystem requests and/or direct
operations on the volume (like dm snapshots).
Sure, you could replace the counter and reservation interface with
explicitly provisioned/trimmed LBA ranges that the fs can manage to
provide -ENOSPC guarantees, but then the fs has to do those various
things you've mentioned:
- Provision those ranges in the fs and change allocation behavior
accordingly.
- Do the background post-crash fitrim preprovision clean up thing.
- Distinguish between trims that are intended to return preprovisioned
space vs. those that come from userspace.
- Have some daemon or whatever (?) responsible for communicating the
need for trims in the fs to return space back to the pool.
Then this still depends on changing how dm thin snapshots work and needs
a way to deal with delayed allocation to actually guarantee -ENOSPC
protection..?
> Hence I think if we get the basic REQ_PROVISION overwrite-in-place
> guarantees defined and implemented as previously outlined, then we
> don't need any special coordination between the fs and block devices
> to avoid fatal ENOSPC issues with sparse and/or snapshot capable
> block devices...
>
This all sounds like a good amount of coordination and unnecessary
complexity to me. What I was thinking as a next phase (i.e. after
initial phase full device reservation) approach for a filesystem like
XFS would be something like this.
- Support a mount option for a configurable size metadata reservation
pool (with sane/conservative default).
- The pool is populated at mount time, else the fs goes right into
simulated -ENOSPC mode.
- Thin pool reservation consumption is controlled by a flag on write
bios that is managed by the fs (flag polarity TBD).
- All fs data writes are explicitly reserved up front in the write path.
Delalloc maps to explicit reservation, overwrites are easy and just
involve an explicit provision.
- Metadata writes are not reserved or provisioned at all. They allocate
out of the thin pool on write (if needed), just as they do today. On
an -ENOSPC metadata write error, the fs goes into simulated -ENOSPC mode
and allows outstanding metadata writes to now use the bio flag to
consume emergency reservation.
So this means that metadata -ENOSPC protection is only as reliable as
the size of the specified pool. This is by design, so the filesystem
still does not have to track provisioning, allocation or overwrites of
its own metadata usage. Users with metadata heavy workloads or who
happen to be sensitive to -ENOSPC errors can be more aggressive with
pool size, while other users might be able to get away with a smaller
pool. Users who are super paranoid and want perfection can continue to
reserve the entire device and pay for the extra storage.
Users who are not sure can test their workload in an appropriate
environment, collect some data/metrics on maximum outstanding dirty
metadata, and then use that as a baseline/minimum pool size for reliable
behavior going forward. This is also where something like Stratis can
come in to generate this sort of information, make recommendations or
implement heuristics (based on things like fs size, amount of RAM, for
e.g.) to provide sane defaults based on use case. I.e., this is
initially exposed as a userspace/tuning issue instead of a
filesystem/dm-thin hard guarantee.
Finally, if you really want to get to that last step of maximally
efficient and safe provisioning in the fs, implement a
'thinreserve=adaptive' mode in the fs that alters the acquisition and
consumption of dm-thin reserved blocks to be adaptive in nature and
promises to do it's own usage throttling against outstanding
reservation. I think this is the mode that most closely resembles your
preprovisioned range mechanism.
For example, adaptive mode could add the logic/complexity where you do
the per-ag provision thing (just using reservation instead of physical
ranges), change the transaction path to attempt to increase the
reservation pool or go into -ENOSPC mode, and flag all writes to be
satisfied from the reserve pool (because you've done the
provision/reservation up front).
At this point the "reserve pool" concept is very different and pretty
much managed entirely by the filesystem. It's just a counter of the set
of blocks the fs is anticipating to write to in the near term, but it's
built on the same underlying reservation mechanism used differently by
other filesystems. So something like ext4 could elide the need for an
adaptive mode, implement the moderate data/metadata pool mechanism and
rely on userspace tools or qualified administrators to do the sizing
correctly, while simultaneously using the same underlying mechanism that
XFS is using for finer grained provisioning.
> As a bonus, if we can implement the guarantees in dm-thin/-snapshot
> and have a filesystem make use of it, then we also have a reference
> implementation to point at device vendors and standards
> associations....
>
I think that's a ways ahead yet.. :P
Thoughts on any of the above? Does that describe enough of the big
picture? (Mike, I hope this at least addresses the whole "why even do
this?" question). I am deliberately trying to work through a progression
that starts simple and generic but actually 100% solves the problem
(even if in a dumb way), then iterates to something that addresses the
biggest drawback with the reference implementation with minimal changes
required to individual filesystems (i.e. metadata pool sizing), and
finally ends up allowing any particular filesystem to refine from there
to achieve maximal efficiency based on its own cost/benefit analysis.
Another way to look at it is... step 1 is to implement the
'thinreserve=full' mount option, which can be trivially implemented by
any filesystem with a couple function calls. Step two is to implement
'thinsreserve=N' support, which consists of a standard iomap
provisioning implementation for data and a bio tagging/error handling
approach that is still pretty simple for most filesystems to implement.
Finally, 'thinreserve=adaptive' is the filesystems best effort to
guarantee -ENOSPC safety with maximal space efficiency.
One general tradeoff with using reservations vs. preprovisioning is the
the latter can just use the provision/trim primitives to alloc/free LBA
ranges. My thought on that is those primitives could possibly be
modified to do the same sort of things with reservation as for physical
allocations. That seems fairly easy to do with bio op flags/modifiers,
though one thing I'm not sure about is how to submit a provision bio to
request a certain amount location agnostic blocks. I'd have to
investigate that more.
So in summary, I can sort of see how this problem could be solved with
this combination of physically preprovisioned ranges and changes to
dm-thin snapshot behavior and whatnot (I'm still missing how this is
supposed to handle delalloc, mostly), but I think that involves more
complexity and customization work than is really necessary. Either way,
this is a distinctly different approach to what I was thinking of
morphing the prototype bits into. So to me the relevant question is does
something like the progression that is outlined above for a block
reservation approach seem a reasonable path to ultimately be able to
accomplish the same sort of results in the fs? If so, then I'm happy to
try and push things in that direction to at least try to prove it out.
If not, then I'm also happy to just leave it alone.. ;)
Brian
> Cheers,
>
> Dave.
> --
> Dave Chinner
> [email protected]
>
On Thu, May 25, 2023 at 9:00 AM Mike Snitzer <[email protected]> wrote:
>
> On Thu, May 25 2023 at 7:39P -0400,
> Dave Chinner <[email protected]> wrote:
>
> > On Wed, May 24, 2023 at 04:02:49PM -0400, Mike Snitzer wrote:
> > > On Tue, May 23 2023 at 8:40P -0400,
> > > Dave Chinner <[email protected]> wrote:
> > >
> > > > On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > > > > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > > > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > > > > ... since I also happen to think there is a potentially interesting
> > > > > > development path to make this sort of reserve pool configurable in terms
> > > > > > of size and active/inactive state, which would allow the fs to use an
> > > > > > emergency pool scheme for managing metadata provisioning and not have to
> > > > > > track and provision individual metadata buffers at all (dealing with
> > > > > > user data is much easier to provision explicitly). So the space
> > > > > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > > > > filesystems that want more granularity for better behavior could achieve
> > > > > > that with more work. Filesystems that don't would be free to rely on the
> > > > > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > > > > protection with very minimal changes.
> > > > > >
> > > > > > That's getting too far into the weeds on the future bits, though. This
> > > > > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > > > > sufficient interest in this sort of "reserve mode" approach to try and
> > > > > > clean it up further and have dm guys look at it, or if you guys see any
> > > > > > obvious issues in what it does that makes it potentially problematic, or
> > > > > > if you would just prefer to go down the path described above...
> > > > >
> > > > > The model that Dave detailed, which builds on REQ_PROVISION and is
> > > > > sticky (by provisioning same blocks for snapshot) seems more useful to
> > > > > me because it is quite precise. That said, it doesn't account for
> > > > > hard requirements that _all_ blocks will always succeed.
> > > >
> > > > Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> > > > but I don't think we'd ever need a hard guarantee from the block
> > > > device that every write bio issued from the filesystem will succeed
> > > > without ENOSPC.
> > > >
> > > > If the block device can provide a guarantee that a provisioned LBA
> > > > range is always writable, then everything else is a filesystem level
> > > > optimisation problem and we don't have to involve the block device
> > > > in any way. All we need is a flag we can ready out of the bdev at
> > > > mount time to determine if the filesystem should be operating with
> > > > LBA provisioning enabled...
> > > >
> > > > e.g. If we need to "pre-provision" a chunk of the LBA space for
> > > > filesystem metadata, we can do that ahead of time and track the
> > > > pre-provisioned range(s) in the filesystem itself.
> > > >
> > > > In XFS, That could be as simple as having small chunks of each AG
> > > > reserved to metadata (e.g. start with the first 100MB) and limiting
> > > > all metadata allocation free space searches to that specific block
> > > > range. When we run low on that space, we pre-provision another 100MB
> > > > chunk and then allocate all metadata out of that new range. If we
> > > > start getting ENOSPC to pre-provisioning, then we reduce the size of
> > > > the regions and log low space warnings to userspace. If we can't
> > > > pre-provision any space at all and we've completely run out, we
> > > > simply declare ENOSPC for all incoming operations that require
> > > > metadata allocation until pre-provisioning succeeds again.
> > >
> > > This is basically saying the same thing but:
> > >
> > > It could be that the LBA space is fragmented and so falling back to
> > > the smallest region size (that matches the thinp block size) would be
> > > the last resort? Then if/when thinp cannot even service allocating a
> > > new free thin block, dm-thinp will transition to out-of-data-space
> > > mode.
> >
> > Yes, something of that sort, though we'd probably give up if we
> > can't get at least megabyte scale reservations - a single
> > modification in XFS can modify many structures and require
> > allocation of a lot of new metadata, so the fileystem cut-off would
> > for metadata provisioning failure would be much larger than the
> > dm-thinp region size....
> >
> > > > This is built entirely on the premise that once proactive backing
> > > > device provisioning fails, the backing device is at ENOSPC and we
> > > > have to wait for that situation to go away before allowing new data
> > > > to be ingested. Hence the block device really doesn't need to know
> > > > anything about what the filesystem is doing and vice versa - The
> > > > block dev just says "yes" or "no" and the filesystem handles
> > > > everything else.
> > >
> > > Yes.
> > >
> > > > It's worth noting that XFS already has a coarse-grained
> > > > implementation of preferred regions for metadata storage. It will
> > > > currently not use those metadata-preferred regions for user data
> > > > unless all the remaining user data space is full. Hence I'm pretty
> > > > sure that a pre-provisioning enhancment like this can be done
> > > > entirely in-memory without requiring any new on-disk state to be
> > > > added.
> > > >
> > > > Sure, if we crash and remount, then we might chose a different LBA
> > > > region for pre-provisioning. But that's not really a huge deal as we
> > > > could also run an internal background post-mount fstrim operation to
> > > > remove any unused pre-provisioning that was left over from when the
> > > > system went down.
> > >
> > > This would be the FITRIM with extension you mention below? Which is a
> > > filesystem interface detail?
> >
> > No. We might reuse some of the internal infrastructure we use to
> > implement FITRIM, but that's about it. It's just something kinda
> > like FITRIM but with different constraints determined by the
> > filesystem rather than the user...
> >
> > As it is, I'm not sure we'd even need it - a preiodic userspace
> > FITRIM would acheive the same result, so leaked provisioned spaces
> > would get cleaned up eventually without the filesystem having to do
> > anything specific...
> >
> > > So dm-thinp would _not_ need to have new
> > > state that tracks "provisioned but unused" block?
> >
> > No idea - that's your domain. :)
> >
> > dm-snapshot, for certain, will need to track provisioned regions
> > because it has to guarantee that overwrites to provisioned space in
> > the origin device will always succeed. Hence it needs to know how
> > much space breaking sharing in provisioned regions after a snapshot
> > has been taken with be required...
>
> dm-thinp offers its own much more scalable snapshot support (doesn't
> use old dm-snapshot N-way copyout target).
>
> dm-snapshot isn't going to be modified to support this level of
> hardening (dm-snapshot is basically in "maintenance only" now).
>
> But I understand your meaning: what you said is 100% applicable to
> dm-thinp's snapshot implementation and needs to be accounted for in
> thinp's metadata (inherent 'provisioned' flag).
>
A bit orthogonal: would dm-thinp need to differentiate between
user-triggered provision requests (eg. from fallocate()) vs
fs-triggered requests? I would lean towards user provisioned areas not
getting dedup'd on snapshot creation, but that would entail tracking
the state of the original request and possibly a provision request
flag (REQ_PROVISION_DEDUP_ON_SNAPSHOT) or an inverse flag
(REQ_PROVISION_NODEDUP). Possibly too convoluted...
> > > Nor would the block
> > > layer need an extra discard flag for a new class of "provisioned"
> > > blocks.
> >
> > Right, I don't see that the discard operations need to care whether
> > the underlying storage is provisioned. dm-thinp and dm-snapshot can
> > treat REQ_OP_DISCARD as "this range is not longer in use" and do
> > whatever they want with them.
> >
> > > If XFS tracked this "provisioned but unused" state, dm-thinp could
> > > just discard the block like its told. Would be nice to avoid dm-thinp
> > > needing to track "provisioned but unused".
> > >
> > > That said, dm-thinp does still need to know if a block was provisioned
> > > (given our previous designed discussion, to allow proper guarantees
> > > from this interface at snapshot time) so that XFS and other
> > > filesystems don't need to re-provision areas they already
> > > pre-provisioned.
> >
> > Right.
> >
> > I've simply assumed that dm-thinp would need to track entire
> > provisioned regions - used or unused - so it knows which writes to
> > empty or shared regions have a reservation to allow allocation to
> > succeed when the backing pool is otherwise empty.....
> >
> > > However, it may be that if thinp did track "provisioned but unused"
> > > it'd be useful to allow snapshots to share provisioned blocks that
> > > were never used. Meaning, we could then avoid "breaking sharing" at
> > > snapshot-time for "provisioned but unused" blocks. But allowing this
> > > "optimization" undercuts the gaurantee that XFS needs for thinp
> > > storage that allows snapshots... SO, I think I answered my own
> > > question: thinp doesnt need to track "provisioned but unused" blocks
> > > but we must always ensure snapshots inherit provisoned blocks ;)
> >
> > Sounds like a potential optimisation, but I haven't thought through
> > a potential snapshot device implementation that far to comment
> > sanely. I stopped once I got to the point where accounting tricks
> > count be used to guarantee space is available for breaking sharing
> > of used provisioned space after a snapshot was taken....
> >
> > > > Further, managing shared pool exhaustion doesn't require a
> > > > reservation pool in the backing device and for the filesystems to
> > > > request space from it. Filesystems already have their own reserve
> > > > pools via pre-provisioning. If we want the filesystems to be able to
> > > > release that space back to the shared pool (e.g. because the shared
> > > > backing pool is critically short on space) then all we need is an
> > > > extension to FITRIM to tell the filesystem to also release internal
> > > > pre-provisioned reserves.
> > >
> > > So by default FITRIM will _not_ discard provisioned blocks. Only if
> > > a flag is used will it result in discarding provisioned blocks.
> >
> > No. FITRIM results in discard of any unused free space in the
> > filesystem that matches the criteria set by the user. We don't care
> > if free space was once provisioned used space - we'll issue a
> > discard for the range regardless. The "special" FITRIM extension I
> > mentioned is to get filesystem metadata provisioning released;
> > that's completely separate to user data provisioning through
> > fallocate() which FITRIM will always discard if it has been freed...
> >
> > IOWs, normal behaviour will be that a FITRIM ends up discarding a
> > mix of unprovisioned and provisioned space. Nobody will be able to
> > predict what mix the device is going to get at any point in time.
> > Also, if we turn on online discard, the block device is going to get
> > a constant stream of discard operations that will also be a mix of
> > provisioned and unprovisioned space that is not longer in use by the
> > filesystem.
> >
> > I suspect that you need to stop trying to double guess what
> > operations the filesystem will use provisioning for, what it will
> > send discards for and when it will send discards for them.. Just
> > assume the device will receive a constant stream of both
> > REQ_PROVISION and REQ_OP_DISCARD (for both provisioned and
> > unprovisioned regions) operations whenver the filesystem is active
> > on a thinp device.....
>
> Yeah, I was getting tripped up in the weeds a bit. It's pretty
> straight-forward (and like I said at the start of our subthread here:
> this follow-on work, to inherit provisioned flag, can build on this
> REQ_PROVISION patchset).
>
> All said, I've now gotten this sub-thread on Joe Thornber's radar and
> we've started discussing. We'll be discussing with more focus
> tomorrow.
>
From the perspective of this patch series, I'll wait for more feedback
before sending out v8 (which would be the above patches and the
follow-on patch to pass through FALLOC_FL_UNSHARE_RANGE [1]).
[1] https://listman.redhat.com/archives/dm-devel/2023-May/054188.html
Thanks!
Sarthak
> Mike
On Thu, May 25, 2023 at 03:47:21PM -0700, Sarthak Kukreti wrote:
> On Thu, May 25, 2023 at 9:00 AM Mike Snitzer <[email protected]> wrote:
> > On Thu, May 25 2023 at 7:39P -0400,
> > Dave Chinner <[email protected]> wrote:
> > > On Wed, May 24, 2023 at 04:02:49PM -0400, Mike Snitzer wrote:
> > > > On Tue, May 23 2023 at 8:40P -0400,
> > > > Dave Chinner <[email protected]> wrote:
> > > > > It's worth noting that XFS already has a coarse-grained
> > > > > implementation of preferred regions for metadata storage. It will
> > > > > currently not use those metadata-preferred regions for user data
> > > > > unless all the remaining user data space is full. Hence I'm pretty
> > > > > sure that a pre-provisioning enhancment like this can be done
> > > > > entirely in-memory without requiring any new on-disk state to be
> > > > > added.
> > > > >
> > > > > Sure, if we crash and remount, then we might chose a different LBA
> > > > > region for pre-provisioning. But that's not really a huge deal as we
> > > > > could also run an internal background post-mount fstrim operation to
> > > > > remove any unused pre-provisioning that was left over from when the
> > > > > system went down.
> > > >
> > > > This would be the FITRIM with extension you mention below? Which is a
> > > > filesystem interface detail?
> > >
> > > No. We might reuse some of the internal infrastructure we use to
> > > implement FITRIM, but that's about it. It's just something kinda
> > > like FITRIM but with different constraints determined by the
> > > filesystem rather than the user...
> > >
> > > As it is, I'm not sure we'd even need it - a preiodic userspace
> > > FITRIM would acheive the same result, so leaked provisioned spaces
> > > would get cleaned up eventually without the filesystem having to do
> > > anything specific...
> > >
> > > > So dm-thinp would _not_ need to have new
> > > > state that tracks "provisioned but unused" block?
> > >
> > > No idea - that's your domain. :)
> > >
> > > dm-snapshot, for certain, will need to track provisioned regions
> > > because it has to guarantee that overwrites to provisioned space in
> > > the origin device will always succeed. Hence it needs to know how
> > > much space breaking sharing in provisioned regions after a snapshot
> > > has been taken with be required...
> >
> > dm-thinp offers its own much more scalable snapshot support (doesn't
> > use old dm-snapshot N-way copyout target).
> >
> > dm-snapshot isn't going to be modified to support this level of
> > hardening (dm-snapshot is basically in "maintenance only" now).
Ah, of course. Sorry for the confusion, I was kinda using
dm-snapshot as shorthand for "dm-thinp + snapshots".
> > But I understand your meaning: what you said is 100% applicable to
> > dm-thinp's snapshot implementation and needs to be accounted for in
> > thinp's metadata (inherent 'provisioned' flag).
*nod*
> A bit orthogonal: would dm-thinp need to differentiate between
> user-triggered provision requests (eg. from fallocate()) vs
> fs-triggered requests?
Why? How is the guarantee the block device has to provide to
provisioned areas different for user vs filesystem internal
provisioned space?
> I would lean towards user provisioned areas not
> getting dedup'd on snapshot creation,
<twitch>
Snapshotting is a clone operation, not a dedupe operation.
Yes, the end result of both is that you have a block shared between
multiple indexes that needs COW on the next overwrite, but the two
operations that get to that point are very different...
</pedantic mode disegaged>
> but that would entail tracking
> the state of the original request and possibly a provision request
> flag (REQ_PROVISION_DEDUP_ON_SNAPSHOT) or an inverse flag
> (REQ_PROVISION_NODEDUP). Possibly too convoluted...
Let's not try to add everyone's favourite pony to this interface
before we've even got it off the ground.
It's the simple precision of the API, the lack of cross-layer
communication requirements and the ability to implement and optimise
the independent layers independently that makes this a very
appealing solution.
We need to start with getting the simple stuff working and prove the
concept. Then once we can observe the behaviour of a working system
we can start working on optimising individual layers for efficiency
and performance....
Cheers,
Dave.
--
Dave Chinner
[email protected]
On Thu, May 25, 2023 at 6:36 PM Dave Chinner <[email protected]> wrote:
>
> On Thu, May 25, 2023 at 03:47:21PM -0700, Sarthak Kukreti wrote:
> > On Thu, May 25, 2023 at 9:00 AM Mike Snitzer <[email protected]> wrote:
> > > On Thu, May 25 2023 at 7:39P -0400,
> > > Dave Chinner <[email protected]> wrote:
> > > > On Wed, May 24, 2023 at 04:02:49PM -0400, Mike Snitzer wrote:
> > > > > On Tue, May 23 2023 at 8:40P -0400,
> > > > > Dave Chinner <[email protected]> wrote:
> > > > > > It's worth noting that XFS already has a coarse-grained
> > > > > > implementation of preferred regions for metadata storage. It will
> > > > > > currently not use those metadata-preferred regions for user data
> > > > > > unless all the remaining user data space is full. Hence I'm pretty
> > > > > > sure that a pre-provisioning enhancment like this can be done
> > > > > > entirely in-memory without requiring any new on-disk state to be
> > > > > > added.
> > > > > >
> > > > > > Sure, if we crash and remount, then we might chose a different LBA
> > > > > > region for pre-provisioning. But that's not really a huge deal as we
> > > > > > could also run an internal background post-mount fstrim operation to
> > > > > > remove any unused pre-provisioning that was left over from when the
> > > > > > system went down.
> > > > >
> > > > > This would be the FITRIM with extension you mention below? Which is a
> > > > > filesystem interface detail?
> > > >
> > > > No. We might reuse some of the internal infrastructure we use to
> > > > implement FITRIM, but that's about it. It's just something kinda
> > > > like FITRIM but with different constraints determined by the
> > > > filesystem rather than the user...
> > > >
> > > > As it is, I'm not sure we'd even need it - a preiodic userspace
> > > > FITRIM would acheive the same result, so leaked provisioned spaces
> > > > would get cleaned up eventually without the filesystem having to do
> > > > anything specific...
> > > >
> > > > > So dm-thinp would _not_ need to have new
> > > > > state that tracks "provisioned but unused" block?
> > > >
> > > > No idea - that's your domain. :)
> > > >
> > > > dm-snapshot, for certain, will need to track provisioned regions
> > > > because it has to guarantee that overwrites to provisioned space in
> > > > the origin device will always succeed. Hence it needs to know how
> > > > much space breaking sharing in provisioned regions after a snapshot
> > > > has been taken with be required...
> > >
> > > dm-thinp offers its own much more scalable snapshot support (doesn't
> > > use old dm-snapshot N-way copyout target).
> > >
> > > dm-snapshot isn't going to be modified to support this level of
> > > hardening (dm-snapshot is basically in "maintenance only" now).
>
> Ah, of course. Sorry for the confusion, I was kinda using
> dm-snapshot as shorthand for "dm-thinp + snapshots".
>
> > > But I understand your meaning: what you said is 100% applicable to
> > > dm-thinp's snapshot implementation and needs to be accounted for in
> > > thinp's metadata (inherent 'provisioned' flag).
>
> *nod*
>
> > A bit orthogonal: would dm-thinp need to differentiate between
> > user-triggered provision requests (eg. from fallocate()) vs
> > fs-triggered requests?
>
> Why? How is the guarantee the block device has to provide to
> provisioned areas different for user vs filesystem internal
> provisioned space?
>
After thinking this through, I stand corrected. I was primarily
concerned with how this would balloon thin snapshot sizes if users
potentially provision a large chunk of the filesystem but that's
putting the cart way before the horse.
Best
Sarthak
> > I would lean towards user provisioned areas not
> > getting dedup'd on snapshot creation,
>
> <twitch>
>
> Snapshotting is a clone operation, not a dedupe operation.
>
> Yes, the end result of both is that you have a block shared between
> multiple indexes that needs COW on the next overwrite, but the two
> operations that get to that point are very different...
>
> </pedantic mode disegaged>
>
> > but that would entail tracking
> > the state of the original request and possibly a provision request
> > flag (REQ_PROVISION_DEDUP_ON_SNAPSHOT) or an inverse flag
> > (REQ_PROVISION_NODEDUP). Possibly too convoluted...
>
> Let's not try to add everyone's favourite pony to this interface
> before we've even got it off the ground.
>
> It's the simple precision of the API, the lack of cross-layer
> communication requirements and the ability to implement and optimise
> the independent layers independently that makes this a very
> appealing solution.
>
> We need to start with getting the simple stuff working and prove the
> concept. Then once we can observe the behaviour of a working system
> we can start working on optimising individual layers for efficiency
> and performance....
>
> Cheers,
>
> Dave.
> --
> Dave Chinner
> [email protected]
On Thu, May 25, 2023 at 12:19:47PM -0400, Brian Foster wrote:
> On Wed, May 24, 2023 at 10:40:34AM +1000, Dave Chinner wrote:
> > On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > > ... since I also happen to think there is a potentially interesting
> > > > development path to make this sort of reserve pool configurable in terms
> > > > of size and active/inactive state, which would allow the fs to use an
> > > > emergency pool scheme for managing metadata provisioning and not have to
> > > > track and provision individual metadata buffers at all (dealing with
> > > > user data is much easier to provision explicitly). So the space
> > > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > > filesystems that want more granularity for better behavior could achieve
> > > > that with more work. Filesystems that don't would be free to rely on the
> > > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > > protection with very minimal changes.
> > > >
> > > > That's getting too far into the weeds on the future bits, though. This
> > > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > > sufficient interest in this sort of "reserve mode" approach to try and
> > > > clean it up further and have dm guys look at it, or if you guys see any
> > > > obvious issues in what it does that makes it potentially problematic, or
> > > > if you would just prefer to go down the path described above...
> > >
> > > The model that Dave detailed, which builds on REQ_PROVISION and is
> > > sticky (by provisioning same blocks for snapshot) seems more useful to
> > > me because it is quite precise. That said, it doesn't account for
> > > hard requirements that _all_ blocks will always succeed.
> >
> > Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> > but I don't think we'd ever need a hard guarantee from the block
> > device that every write bio issued from the filesystem will succeed
> > without ENOSPC.
> >
>
> The bigger picture goal that I didn't get into in my previous mail is
> the "full device" reservation model is intended to be a simple, crude
> reference implementation that can be enabled for any arbitrary thin
> volume consumer (filesystem or application). The idea is to build that
> on a simple enough reservation mechanism that any such consumer could
> override it based on its own operational model. The goal is to guarantee
> that a particular filesystem never receives -ENOSPC from dm-thin on
> writes, but the first phase of implementing that is to simply guarantee
> every block is writeable.
>
> As a specific filesystem is able to more explicitly provision its own
> allocations in a way that it can guarantee to return -ENOSPC from
> dm-thin up front (rather than at write bio time), it can reduce the need
> for the amount of reservation required, ultimately to zero if that
> filesystem provides the ability to pre-provision all of its writes to
> storage in some way or another.
>
> I think for filesystems with complex metadata management like XFS, it's
> not very realistic to expect explicit 1-1 provisioning for all metadata
> changes on a per-transaction basis in the same way that can (fairly
> easily) be done for data, which means a pool mechanism is probably still
> needed for the metadata class of writes.
I'm trying to avoid need for 1-1 provisioning and the need for a
accounting based reservation pool approach. I've tried the
reservation pool thing several times, and they've all collapsed
under the complexity of behaviour guarantees under worst case write
amplification situations.
The whole point of the LBA provisioning approach is that it
completely avoids the need to care about write amplification because
the underlying device guarantees any write to a LBA that is
provisioned will succeed. It takes care of the write amplification
problem for us, and we can make it even easier for the backing
device by aligning LBA range provision requests to device region
sizes.
> > If the block device can provide a guarantee that a provisioned LBA
> > range is always writable, then everything else is a filesystem level
> > optimisation problem and we don't have to involve the block device
> > in any way. All we need is a flag we can ready out of the bdev at
> > mount time to determine if the filesystem should be operating with
> > LBA provisioning enabled...
> >
> > e.g. If we need to "pre-provision" a chunk of the LBA space for
> > filesystem metadata, we can do that ahead of time and track the
> > pre-provisioned range(s) in the filesystem itself.
> >
> > In XFS, That could be as simple as having small chunks of each AG
> > reserved to metadata (e.g. start with the first 100MB) and limiting
> > all metadata allocation free space searches to that specific block
> > range. When we run low on that space, we pre-provision another 100MB
> > chunk and then allocate all metadata out of that new range. If we
> > start getting ENOSPC to pre-provisioning, then we reduce the size of
> > the regions and log low space warnings to userspace. If we can't
> > pre-provision any space at all and we've completely run out, we
> > simply declare ENOSPC for all incoming operations that require
> > metadata allocation until pre-provisioning succeeds again.
> >
>
> The more interesting aspect of this is not so much how space is
> provisioned and allocated, but how the filesystem is going to consume
> that space in a way that guarantees -ENOSPC is provided up front before
> userspace is allowed to make modifications.
Yeah, that's trivial with REQ_PROVISION.
If, at transaction reservation time, we don't have enough
provisioned metadata space available for the potential allocations
we'll need to make, we kick provisioning work off wait for more to
come available. If that fails and none is available, we'll get an
enospc error right there, same as if the filesystem itself has no
blocks available for allocation.
This is no different to, say, having xfs_create() fail reservation
because ENOSPC, then calling xfs_flush_inodes() to kick off an inode
cache walk to trim away all the unused post-eof allocations in
memory to free up some space we can use. When that completes,
we try the reservation again.
There's no new behaviours we need to introduce here - it's just
replication of existing behaviours and infrastructure.
> You didn't really touch on
> that here, so I'm going to assume we'd have something like a perag
> counter of how many free blocks currently live in preprovisioned ranges,
> and then an fs-wide total somewhere so a transaction has the ability to
> consume these blocks at trans reservation time, the fs knows when to
> preprovision more space (or go into -ENOSPC mode), etc.
Sure, something like that. Those are all implementation details, and
not really that complex to implement and is largely replication of
reservation infrastructure we already have.
> Some accounting of that nature is necessary here in order to prevent the
> filesystem from ever writing to unprovisioned space. So what I was
> envisioning is rather than explicitly preprovision a physical range of
> each AG and tracking all that, just reserve that number of arbitrarily
> located blocks from dm for each AG.
>
> The initial perag reservations can be populated at mount time,
> replenished as needed in a very similar way as what you describe, and
> 100% released back to the thin pool at unmount time. On top of that,
> there's no need to track physical preprovisioned ranges at all. Not just
> for allocation purposes, but also to avoid things like having to protect
> background trims from preprovisioned ranges of free space dedicated for
> metadata, etc.
That's all well and good, but reading further down the email the
breadth and depth of changes to filesystem and block device
behaviour to enable this are ... significant.
> > Further, managing shared pool exhaustion doesn't require a
> > reservation pool in the backing device and for the filesystems to
> > request space from it. Filesystems already have their own reserve
> > pools via pre-provisioning. If we want the filesystems to be able to
> > release that space back to the shared pool (e.g. because the shared
> > backing pool is critically short on space) then all we need is an
> > extension to FITRIM to tell the filesystem to also release internal
> > pre-provisioned reserves.
> >
> > Then the backing pool admin (person or automated daemon!) can simply
> > issue a trim on all the filesystems in the pool and spce will be
> > returned. Then filesystems will ask for new pre-provisioned space
> > when they next need to ingest modifications, and the backing pool
> > can manage the new pre-provisioning space requests directly....
> >
>
> This is written as to imply that the reservation pool is some big
> complex thing, which makes me think there is some
> confusion/miscommunication.
No confusion, I'm just sceptical that it will work given my
experience trying to implement reservation based solutions multiple
different ways over the past decade. They've all failed because
they collapse under either the complexity explosion or space
overhead required to handle the worst case behavioural scenarios.
At one point I calculated the worst case reservation needed ensure
log recovery will always succeeded, ignoring write amplification,
was about 16x the size of the log. If I took write amplification for
dm-thinp having 64kB blocks and each inode hitting a different
cluster in it's own dm thinp block, that write amplication hit 64x.
So for recovering a 2GB log, if dm-thinp doesn't have a reserve of
well over 100GB of pool space, there is no guarantee that log
recovery will -always- succeed.
It's worst case numbers like this which made me conclude that
reservation based approaches cannot provide guarantees that ENOSPC
will never occur. The numbers are just too large when you start
considering journals that can hold a million dirty objects,
intent chains that might require modifying hundreds of metadata
blocks across a dozen transactions before they complete, etc.
OTOH, REQ_PROVISION makes this "log recovery needs new space to be
allocated" problem go away entirely. It provides a mechanism that
ensures log recovery does not consume any new space in the backing
pool as all the overwrites it performs are to previously provisioned
metadata.....
This is just one of the many reasons why I think the REQ_PROVISION
method is far better than reservations - it solves problems that
pure runtime reservations can't.
> It's basically just an in memory counter of
> space that is allocated out of a shared thin pool and is held in a
> specific thin volume while it is currently in use. The counter on the
> volume is managed indirectly by filesystem requests and/or direct
> operations on the volume (like dm snapshots).
>
> Sure, you could replace the counter and reservation interface with
> explicitly provisioned/trimmed LBA ranges that the fs can manage to
> provide -ENOSPC guarantees, but then the fs has to do those various
> things you've mentioned:
>
> - Provision those ranges in the fs and change allocation behavior
> accordingly.
This is relatively simple - most of the allocator functionality is
already there.
> - Do the background post-crash fitrim preprovision clean up thing.
We've already decided this is not needed.
> - Distinguish between trims that are intended to return preprovisioned
> space vs. those that come from userspace.
It's about ten lines of code in xfs_trim_extents() to do this. i.e.
the free space tree walk simply skips over free extents in the
metadata provisioned region based on a flag value.
> - Have some daemon or whatever (?) responsible for communicating the
> need for trims in the fs to return space back to the pool.
Systems are already configured to run a periodic fstrim passes to do
this via systemd units. And I'm pretty sure dm-thinp has a low space
notification to userspace (via dbus?) that is already used by
userspace agents to handle "near ENOSPC" events automatically.
> Then this still depends on changing how dm thin snapshots work and needs
> a way to deal with delayed allocation to actually guarantee -ENOSPC
> protection..?
I think you misunderstand: I'm not proposing to use REQ_PROVISION
for writes the filesystem does not guarantee will succeed. Never
have, I think it makes no sense at all. If the filesystem
can return ENOSPC for an unprovisioned user data write, then the
block device can too.
> > Hence I think if we get the basic REQ_PROVISION overwrite-in-place
> > guarantees defined and implemented as previously outlined, then we
> > don't need any special coordination between the fs and block devices
> > to avoid fatal ENOSPC issues with sparse and/or snapshot capable
> > block devices...
> >
>
> This all sounds like a good amount of coordination and unnecessary
> complexity to me. What I was thinking as a next phase (i.e. after
> initial phase full device reservation) approach for a filesystem like
> XFS would be something like this.
>
> - Support a mount option for a configurable size metadata reservation
> pool (with sane/conservative default).
I want this to all to work without the user having be aware that
there filesystem is running on a sparse device.
> - The pool is populated at mount time, else the fs goes right into
> simulated -ENOSPC mode.
What are the rules of this mode?
Hmmmm.
Log recovery needs to be able to allocate new metadata (i.e. in
intent replay), so I'm guessing reservation is needed before log
recovery? But if pool reservation fails, how do we then safely
perform log recovery given the filesystem is in ENOSPC mode?
> - Thin pool reservation consumption is controlled by a flag on write
> bios that is managed by the fs (flag polarity TBD).
So we still need a bio flag to communicate "this IO consumes
reservation".
What are the semantics of this flag? What happens on submission
error? e.g. the bio is failed before it gets to the layer that
consumes it - how does the filesystem know that reservation was
consumed or not at completion?
How do we know when to set it for user data writes?
What happens if the device recieves a bio with this flag but there
is no reservation remaining? e.g. the filesystem or device
accounting have got out of whack?
Hmmm. On that note, what about write amplification? Or should I call
it "reservation amplification". i.e. a 4kB bio with a "consume
reservation" flag might trigger a dm-region COW or allocation and
require 512kB of dm-thinp pool space to be allocated. How much
reservation actually gets consumed, and how do we reconcile the
differences in physical consumption vs reservation consumption?
> - All fs data writes are explicitly reserved up front in the write path.
> Delalloc maps to explicit reservation, overwrites are easy and just
> involve an explicit provision.
This is the first you've mentioned an "explicit provision"
operation. Is this like REQ_PROVISION, or something else?
This seems to imply that the ->iomap_begin method has to do
explicit provisioning callouts when we get a write that lands in an
IOMAP_MAPPED extent? Or something else?
Can you describe this mechanism in more detail?
> - Metadata writes are not reserved or provisioned at all. They allocate
> out of the thin pool on write (if needed), just as they do today. On
> an -ENOSPC metadata write error, the fs goes into simulated -ENOSPC mode
> and allows outstanding metadata writes to now use the bio flag to
> consume emergency reservation.
Okay. We need two pools in the backing device? The normal free space
pool, and an emergency reservation pool?
Without reading further, this implies that the filesystem is
reliant on the emergency reservation pool being large enough that
it can write any dirty metadata it has outstanding without ENOSPC
occuring. How does the size of this emergency pool get configured?
> So this means that metadata -ENOSPC protection is only as reliable as
> the size of the specified pool. This is by design, so the filesystem
> still does not have to track provisioning, allocation or overwrites of
> its own metadata usage. Users with metadata heavy workloads or who
> happen to be sensitive to -ENOSPC errors can be more aggressive with
> pool size, while other users might be able to get away with a smaller
> pool. Users who are super paranoid and want perfection can continue to
> reserve the entire device and pay for the extra storage.
Oh. Hand tuning. :(
> Users who are not sure can test their workload in an appropriate
> environment, collect some data/metrics on maximum outstanding dirty
> metadata, and then use that as a baseline/minimum pool size for reliable
> behavior going forward. This is also where something like Stratis can
> come in to generate this sort of information, make recommendations or
> implement heuristics (based on things like fs size, amount of RAM, for
> e.g.) to provide sane defaults based on use case. I.e., this is
> initially exposed as a userspace/tuning issue instead of a
> filesystem/dm-thin hard guarantee.
Which are the same things people have been complaining about for years.
> Finally, if you really want to get to that last step of maximally
> efficient and safe provisioning in the fs, implement a
> 'thinreserve=adaptive' mode in the fs that alters the acquisition and
> consumption of dm-thin reserved blocks to be adaptive in nature and
> promises to do it's own usage throttling against outstanding
> reservation. I think this is the mode that most closely resembles your
> preprovisioned range mechanism.
>
> For example, adaptive mode could add the logic/complexity where you do
> the per-ag provision thing (just using reservation instead of physical
> ranges), change the transaction path to attempt to increase the
> reservation pool or go into -ENOSPC mode, and flag all writes to be
> satisfied from the reserve pool (because you've done the
> provision/reservation up front).
Ok, so why not just go straight to this model using REQ_PROVISION?
If we then want to move to a different "accounting only" model for
provisioning, we just change REQ_PROVISION?
But I still see the problem of write amplification accounting being
unsolved by the "filesystem accounting only" approach advocated
here. We have no idea when the backing device has snapshots taken,
we have no idea when a filesystem write IO actually consumes more
thinp blocks than filesystem blocks, etc. How does the filesystem
level reservation pool address these problems?
> Thoughts on any of the above?
I'd say it went wrong at the requirements stage, resulting in an
overly complex, over-engineered solution.
> One general tradeoff with using reservations vs. preprovisioning is the
> the latter can just use the provision/trim primitives to alloc/free LBA
> ranges. My thought on that is those primitives could possibly be
> modified to do the same sort of things with reservation as for physical
> allocations. That seems fairly easy to do with bio op flags/modifiers,
> though one thing I'm not sure about is how to submit a provision bio to
> request a certain amount location agnostic blocks. I'd have to
> investigate that more.
Sure, if the constrained LBA space aspect of the REQ_PROVISION
implementation causes issues, then we see if we can optimise away
the fixed LBA space requirement.
--
Dave Chinner
[email protected]
On Fri, May 26, 2023 at 07:37:43PM +1000, Dave Chinner wrote:
> On Thu, May 25, 2023 at 12:19:47PM -0400, Brian Foster wrote:
> > On Wed, May 24, 2023 at 10:40:34AM +1000, Dave Chinner wrote:
> > > On Tue, May 23, 2023 at 11:26:18AM -0400, Mike Snitzer wrote:
> > > > On Tue, May 23 2023 at 10:05P -0400, Brian Foster <[email protected]> wrote:
> > > > > On Mon, May 22, 2023 at 02:27:57PM -0400, Mike Snitzer wrote:
> > > > > ... since I also happen to think there is a potentially interesting
> > > > > development path to make this sort of reserve pool configurable in terms
> > > > > of size and active/inactive state, which would allow the fs to use an
> > > > > emergency pool scheme for managing metadata provisioning and not have to
> > > > > track and provision individual metadata buffers at all (dealing with
> > > > > user data is much easier to provision explicitly). So the space
> > > > > inefficiency thing is potentially just a tradeoff for simplicity, and
> > > > > filesystems that want more granularity for better behavior could achieve
> > > > > that with more work. Filesystems that don't would be free to rely on the
> > > > > simple/basic mechanism provided by dm-thin and still have basic -ENOSPC
> > > > > protection with very minimal changes.
> > > > >
> > > > > That's getting too far into the weeds on the future bits, though. This
> > > > > is essentially 99% a dm-thin approach, so I'm mainly curious if there's
> > > > > sufficient interest in this sort of "reserve mode" approach to try and
> > > > > clean it up further and have dm guys look at it, or if you guys see any
> > > > > obvious issues in what it does that makes it potentially problematic, or
> > > > > if you would just prefer to go down the path described above...
> > > >
> > > > The model that Dave detailed, which builds on REQ_PROVISION and is
> > > > sticky (by provisioning same blocks for snapshot) seems more useful to
> > > > me because it is quite precise. That said, it doesn't account for
> > > > hard requirements that _all_ blocks will always succeed.
> > >
> > > Hmmm. Maybe I'm misunderstanding the "reserve pool" context here,
> > > but I don't think we'd ever need a hard guarantee from the block
> > > device that every write bio issued from the filesystem will succeed
> > > without ENOSPC.
> > >
> >
> > The bigger picture goal that I didn't get into in my previous mail is
> > the "full device" reservation model is intended to be a simple, crude
> > reference implementation that can be enabled for any arbitrary thin
> > volume consumer (filesystem or application). The idea is to build that
> > on a simple enough reservation mechanism that any such consumer could
> > override it based on its own operational model. The goal is to guarantee
> > that a particular filesystem never receives -ENOSPC from dm-thin on
> > writes, but the first phase of implementing that is to simply guarantee
> > every block is writeable.
> >
> > As a specific filesystem is able to more explicitly provision its own
> > allocations in a way that it can guarantee to return -ENOSPC from
> > dm-thin up front (rather than at write bio time), it can reduce the need
> > for the amount of reservation required, ultimately to zero if that
> > filesystem provides the ability to pre-provision all of its writes to
> > storage in some way or another.
> >
> > I think for filesystems with complex metadata management like XFS, it's
> > not very realistic to expect explicit 1-1 provisioning for all metadata
> > changes on a per-transaction basis in the same way that can (fairly
> > easily) be done for data, which means a pool mechanism is probably still
> > needed for the metadata class of writes.
>
> I'm trying to avoid need for 1-1 provisioning and the need for a
> accounting based reservation pool approach. I've tried the
> reservation pool thing several times, and they've all collapsed
> under the complexity of behaviour guarantees under worst case write
> amplification situations.
>
> The whole point of the LBA provisioning approach is that it
> completely avoids the need to care about write amplification because
> the underlying device guarantees any write to a LBA that is
> provisioned will succeed. It takes care of the write amplification
> problem for us, and we can make it even easier for the backing
> device by aligning LBA range provision requests to device region
> sizes.
>
> > > If the block device can provide a guarantee that a provisioned LBA
> > > range is always writable, then everything else is a filesystem level
> > > optimisation problem and we don't have to involve the block device
> > > in any way. All we need is a flag we can ready out of the bdev at
> > > mount time to determine if the filesystem should be operating with
> > > LBA provisioning enabled...
> > >
> > > e.g. If we need to "pre-provision" a chunk of the LBA space for
> > > filesystem metadata, we can do that ahead of time and track the
> > > pre-provisioned range(s) in the filesystem itself.
> > >
> > > In XFS, That could be as simple as having small chunks of each AG
> > > reserved to metadata (e.g. start with the first 100MB) and limiting
> > > all metadata allocation free space searches to that specific block
> > > range. When we run low on that space, we pre-provision another 100MB
> > > chunk and then allocate all metadata out of that new range. If we
> > > start getting ENOSPC to pre-provisioning, then we reduce the size of
> > > the regions and log low space warnings to userspace. If we can't
> > > pre-provision any space at all and we've completely run out, we
> > > simply declare ENOSPC for all incoming operations that require
> > > metadata allocation until pre-provisioning succeeds again.
> > >
> >
> > The more interesting aspect of this is not so much how space is
> > provisioned and allocated, but how the filesystem is going to consume
> > that space in a way that guarantees -ENOSPC is provided up front before
> > userspace is allowed to make modifications.
>
> Yeah, that's trivial with REQ_PROVISION.
>
> If, at transaction reservation time, we don't have enough
> provisioned metadata space available for the potential allocations
> we'll need to make, we kick provisioning work off wait for more to
> come available. If that fails and none is available, we'll get an
> enospc error right there, same as if the filesystem itself has no
> blocks available for allocation.
>
> This is no different to, say, having xfs_create() fail reservation
> because ENOSPC, then calling xfs_flush_inodes() to kick off an inode
> cache walk to trim away all the unused post-eof allocations in
> memory to free up some space we can use. When that completes,
> we try the reservation again.
>
> There's no new behaviours we need to introduce here - it's just
> replication of existing behaviours and infrastructure.
>
Yes, this is just context. What I'm trying to say is the semantics for
this aspect would be the same irrespective of "guaranteed writeable
space" being implemented as a reservation or preprovisioned LBA range.
I.e., it's a limited resource that has to be managed in a way to provide
specific user visible behavior.
> > You didn't really touch on
> > that here, so I'm going to assume we'd have something like a perag
> > counter of how many free blocks currently live in preprovisioned ranges,
> > and then an fs-wide total somewhere so a transaction has the ability to
> > consume these blocks at trans reservation time, the fs knows when to
> > preprovision more space (or go into -ENOSPC mode), etc.
>
> Sure, something like that. Those are all implementation details, and
> not really that complex to implement and is largely replication of
> reservation infrastructure we already have.
>
Ack.
> > Some accounting of that nature is necessary here in order to prevent the
> > filesystem from ever writing to unprovisioned space. So what I was
> > envisioning is rather than explicitly preprovision a physical range of
> > each AG and tracking all that, just reserve that number of arbitrarily
> > located blocks from dm for each AG.
> >
> > The initial perag reservations can be populated at mount time,
> > replenished as needed in a very similar way as what you describe, and
> > 100% released back to the thin pool at unmount time. On top of that,
> > there's no need to track physical preprovisioned ranges at all. Not just
> > for allocation purposes, but also to avoid things like having to protect
> > background trims from preprovisioned ranges of free space dedicated for
> > metadata, etc.
>
> That's all well and good, but reading further down the email the
> breadth and depth of changes to filesystem and block device
> behaviour to enable this are ... significant.
>
> > > Further, managing shared pool exhaustion doesn't require a
> > > reservation pool in the backing device and for the filesystems to
> > > request space from it. Filesystems already have their own reserve
> > > pools via pre-provisioning. If we want the filesystems to be able to
> > > release that space back to the shared pool (e.g. because the shared
> > > backing pool is critically short on space) then all we need is an
> > > extension to FITRIM to tell the filesystem to also release internal
> > > pre-provisioned reserves.
> > >
> > > Then the backing pool admin (person or automated daemon!) can simply
> > > issue a trim on all the filesystems in the pool and spce will be
> > > returned. Then filesystems will ask for new pre-provisioned space
> > > when they next need to ingest modifications, and the backing pool
> > > can manage the new pre-provisioning space requests directly....
> > >
> >
> > This is written as to imply that the reservation pool is some big
> > complex thing, which makes me think there is some
> > confusion/miscommunication.
>
> No confusion, I'm just sceptical that it will work given my
> experience trying to implement reservation based solutions multiple
> different ways over the past decade. They've all failed because
> they collapse under either the complexity explosion or space
> overhead required to handle the worst case behavioural scenarios.
>
> At one point I calculated the worst case reservation needed ensure
> log recovery will always succeeded, ignoring write amplification,
> was about 16x the size of the log. If I took write amplification for
> dm-thinp having 64kB blocks and each inode hitting a different
> cluster in it's own dm thinp block, that write amplication hit 64x.
>
Ok. Can you give some examples of operations that lead to this worst
case behavior? It sounds like you're talking about inode chunk intent
initialization or some such, but I'd like to be sure I understand.
> So for recovering a 2GB log, if dm-thinp doesn't have a reserve of
> well over 100GB of pool space, there is no guarantee that log
> recovery will -always- succeed.
>
> It's worst case numbers like this which made me conclude that
> reservation based approaches cannot provide guarantees that ENOSPC
> will never occur. The numbers are just too large when you start
> considering journals that can hold a million dirty objects,
> intent chains that might require modifying hundreds of metadata
> blocks across a dozen transactions before they complete, etc.
>
> OTOH, REQ_PROVISION makes this "log recovery needs new space to be
> allocated" problem go away entirely. It provides a mechanism that
> ensures log recovery does not consume any new space in the backing
> pool as all the overwrites it performs are to previously provisioned
> metadata.....
>
Ah, I see. So this relies on the change in behavior to dm-thin
snapshots to preserve overwriteably of previously provisioned metadata
to indirectly manage log recovery -> metadata write amplification.
This is useful context and helps understand the intent of that
suggestion. I also think it calls out some of the disconnect..
> This is just one of the many reasons why I think the REQ_PROVISION
> method is far better than reservations - it solves problems that
> pure runtime reservations can't.
>
> > It's basically just an in memory counter of
> > space that is allocated out of a shared thin pool and is held in a
> > specific thin volume while it is currently in use. The counter on the
> > volume is managed indirectly by filesystem requests and/or direct
> > operations on the volume (like dm snapshots).
> >
> > Sure, you could replace the counter and reservation interface with
> > explicitly provisioned/trimmed LBA ranges that the fs can manage to
> > provide -ENOSPC guarantees, but then the fs has to do those various
> > things you've mentioned:
> >
> > - Provision those ranges in the fs and change allocation behavior
> > accordingly.
>
> This is relatively simple - most of the allocator functionality is
> already there.
>
> > - Do the background post-crash fitrim preprovision clean up thing.
>
> We've already decided this is not needed.
>
> > - Distinguish between trims that are intended to return preprovisioned
> > space vs. those that come from userspace.
>
> It's about ten lines of code in xfs_trim_extents() to do this. i.e.
> the free space tree walk simply skips over free extents in the
> metadata provisioned region based on a flag value.
>
> > - Have some daemon or whatever (?) responsible for communicating the
> > need for trims in the fs to return space back to the pool.
>
> Systems are already configured to run a periodic fstrim passes to do
> this via systemd units. And I'm pretty sure dm-thinp has a low space
> notification to userspace (via dbus?) that is already used by
> userspace agents to handle "near ENOSPC" events automatically.
>
Yeah, Ok. To be clear, I'm not trying to suggest any of these particular
things are complex or the whole thing is intractable or anything like
that. I'm pretty sure I understand how this can all be made to work, at
least for metadata. But I am saying this makes a bunch of customization
changes that could lead to a very XFS centric approach, may be more work
than necessary, and hasn't been described in a way that explains how it
actually (or helps) solves the broader -ENOSPC problem.
> > Then this still depends on changing how dm thin snapshots work and needs
> > a way to deal with delayed allocation to actually guarantee -ENOSPC
> > protection..?
>
> I think you misunderstand: I'm not proposing to use REQ_PROVISION
> for writes the filesystem does not guarantee will succeed. Never
> have, I think it makes no sense at all. If the filesystem
> can return ENOSPC for an unprovisioned user data write, then the
> block device can too.
>
Well, yes.. that's why I was asking. :) I still don't parse what you're
saying here.
Is this intended to prevent -ENOSPC from dm-thin for data writes, or is
that an exercise for the reader?
> > > Hence I think if we get the basic REQ_PROVISION overwrite-in-place
> > > guarantees defined and implemented as previously outlined, then we
> > > don't need any special coordination between the fs and block devices
> > > to avoid fatal ENOSPC issues with sparse and/or snapshot capable
> > > block devices...
> > >
> >
> > This all sounds like a good amount of coordination and unnecessary
> > complexity to me. What I was thinking as a next phase (i.e. after
> > initial phase full device reservation) approach for a filesystem like
> > XFS would be something like this.
> >
> > - Support a mount option for a configurable size metadata reservation
> > pool (with sane/conservative default).
>
> I want this to all to work without the user having be aware that
> there filesystem is running on a sparse device.
>
Reservation (or provision) support can certainly be autodetected.
> > - The pool is populated at mount time, else the fs goes right into
> > simulated -ENOSPC mode.
>
> What are the rules of this mode?
>
Earlier you mentioned that the filesystem would "declare -ENOSPC" when
preprovisioning starts to fail. The rules here would be exactly the
same.
> Hmmmm.
>
> Log recovery needs to be able to allocate new metadata (i.e. in
> intent replay), so I'm guessing reservation is needed before log
> recovery? But if pool reservation fails, how do we then safely
> perform log recovery given the filesystem is in ENOSPC mode?
>
> > - Thin pool reservation consumption is controlled by a flag on write
> > bios that is managed by the fs (flag polarity TBD).
>
> So we still need a bio flag to communicate "this IO consumes
> reservation".
>
> What are the semantics of this flag? What happens on submission
> error? e.g. the bio is failed before it gets to the layer that
> consumes it - how does the filesystem know that reservation was
> consumed or not at completion?
>
The semantics are to use reservation or not. If so, then it's implied
reservation exists and if not that's a bug. If reservation is not
enabled, then dm-thin processes those writes exactly as it does today
(allocates out of the pool when necessary, fails otherwise).
> How do we know when to set it for user data writes?
>
User data is always reserved before writes are accepted, so reservation
is always enabled for user data write bios.
> What happens if the device recieves a bio with this flag but there
> is no reservation remaining? e.g. the filesystem or device
> accounting have got out of whack?
>
That's a bug. Almost the same as if the filesystem were to allow a
delalloc write that can't ultimately allocate because in-core counters
become inconsistent with free space btrees (not like that hasn't
happened before ;).
Let's not get too into the weeds here as if to imply coding errors
translate to design flaws. I'd say it probably should warn, fallback
to normal pool allocation.
> Hmmm. On that note, what about write amplification? Or should I call
> it "reservation amplification". i.e. a 4kB bio with a "consume
> reservation" flag might trigger a dm-region COW or allocation and
> require 512kB of dm-thinp pool space to be allocated. How much
> reservation actually gets consumed, and how do we reconcile the
> differences in physical consumption vs reservation consumption?
>
The filesystem has no concept of the amount of reservation. Only whether
outstanding writes have been reserved or not. In this specific reference
implementation, all data writes are reserved and metadata writes are not
(until an -ENOSPC error happens and the fs decides to "declare -ENOSPC"
based on underlying volume state).
> > - All fs data writes are explicitly reserved up front in the write path.
> > Delalloc maps to explicit reservation, overwrites are easy and just
> > involve an explicit provision.
>
> This is the first you've mentioned an "explicit provision"
> operation. Is this like REQ_PROVISION, or something else?
>
> This seems to imply that the ->iomap_begin method has to do
> explicit provisioning callouts when we get a write that lands in an
> IOMAP_MAPPED extent? Or something else?
>
> Can you describe this mechanism in more detail?
>
So something I've hacked up from the older prototype is to essentially
implement a simple form of a REQ_PROVISION|REQ_RESERVE type operation.
You can think of it like REQ_PROVISION as implemented by this series,
but it doesn't actually do the COW breaking and allocation and whatnot
right away. Instead, it reserves however many blocks out of the pool
might be required to guarantee subsequent writes to the specified region
are guaranteed not to fail with -ENOSPC.
(Note that the prototype isn't currently using REQ_PROVISION. It's just
a function call at the moment. I'm just explaining the concept.)
So the idea for user data in general is something like:
iomap looks up an extent, does a "reserve provision" over it based on
the size of the write, etc. If that succeeds, then the write can proceed
to dirty pages with a guarantee that dm-thin will not -ENOSPC at
writeback time.
If the extent is a hole, then delalloc translates to location agnostic
reservation that is eventually translated to a "reserve provision" at
filesystem allocation time. Note that this does introduce an aspect of
reservation amplification due to block size differences, but this was
already addressed by the older prototype. The same 'flush inodes on
-ENOSPC' mechanism you refer to above provides a feedback mechanism to
allow outstanding reservations to flush and prevent any premature error
problems.
And that can be optimized further in various ways. For example, to
simply map outstanding delalloc extents in the fs and do the "reserve
provision" across the ultimate LBA ranges to release overprovisioned
reserves, while still deferring writeback to later. A shrinker could be
used to allow the thin pool to signal lower space conditions to active
volumes to smooth out behavior rather than waiting for an actual
-ENOSPC, etc. etc.
> > - Metadata writes are not reserved or provisioned at all. They allocate
> > out of the thin pool on write (if needed), just as they do today. On
> > an -ENOSPC metadata write error, the fs goes into simulated -ENOSPC mode
> > and allows outstanding metadata writes to now use the bio flag to
> > consume emergency reservation.
>
> Okay. We need two pools in the backing device? The normal free space
> pool, and an emergency reservation pool?
>
Err not exactly.. it's really just selective use of the bio flag that
allows reserve consumption. Always enabled on data, never on metadata,
until -ENOSPC error and the fs decides to open reserves for metadata and
attempt to allow a full quiesce.
> Without reading further, this implies that the filesystem is
> reliant on the emergency reservation pool being large enough that
> it can write any dirty metadata it has outstanding without ENOSPC
> occuring. How does the size of this emergency pool get configured?
>
> > So this means that metadata -ENOSPC protection is only as reliable as
> > the size of the specified pool. This is by design, so the filesystem
> > still does not have to track provisioning, allocation or overwrites of
> > its own metadata usage. Users with metadata heavy workloads or who
> > happen to be sensitive to -ENOSPC errors can be more aggressive with
> > pool size, while other users might be able to get away with a smaller
> > pool. Users who are super paranoid and want perfection can continue to
> > reserve the entire device and pay for the extra storage.
>
> Oh. Hand tuning. :(
>
Yes and no.. from the fs perspective it's hand tuning. From a user
perspective it can be if desired I guess, but really intended to be done
by the management software that already exists with intent to help
manage this problem. To put another way, any complex user of thin
provisioning who is concerned about this problem is already doing some
degree of tuning here to try and prevent it.
Also, an alternative to what I describe above could be for a filesystem
to implement thinreserve=N mode with a throttle to best ensure -ENOSPC
reliability at the cost of performance. It's still a tunable, but maybe
easier to turn into a heuristic. Not sure, just a random thought.
> > Users who are not sure can test their workload in an appropriate
> > environment, collect some data/metrics on maximum outstanding dirty
> > metadata, and then use that as a baseline/minimum pool size for reliable
> > behavior going forward. This is also where something like Stratis can
> > come in to generate this sort of information, make recommendations or
> > implement heuristics (based on things like fs size, amount of RAM, for
> > e.g.) to provide sane defaults based on use case. I.e., this is
> > initially exposed as a userspace/tuning issue instead of a
> > filesystem/dm-thin hard guarantee.
>
> Which are the same things people have been complaining about for years.
>
Priorities and progress. I don't think that because this isn't the
absolute perfect, most easily usable, completely efficient solution
right off the bat is a very good argument to imply it's not a feasible
approach in general. This is why I'm trying to describe the intended
progression here. Users are already dealing with this sort of thing
through Stratis, and this is a step to at least try to make things
better.
And really, the same could be said for preprovisioning until/unless it's
able to fully guarantee prevention of -ENOSPC errors for data and
metadata. That is exactly the same sort of thing "people have been
complaining about for years."
> > Finally, if you really want to get to that last step of maximally
> > efficient and safe provisioning in the fs, implement a
> > 'thinreserve=adaptive' mode in the fs that alters the acquisition and
> > consumption of dm-thin reserved blocks to be adaptive in nature and
> > promises to do it's own usage throttling against outstanding
> > reservation. I think this is the mode that most closely resembles your
> > preprovisioned range mechanism.
> >
> > For example, adaptive mode could add the logic/complexity where you do
> > the per-ag provision thing (just using reservation instead of physical
> > ranges), change the transaction path to attempt to increase the
> > reservation pool or go into -ENOSPC mode, and flag all writes to be
> > satisfied from the reserve pool (because you've done the
> > provision/reservation up front).
>
> Ok, so why not just go straight to this model using REQ_PROVISION?
>
A couple reasons.. one is I've also been trying to hack around this
problem for a while and have yet to see a full solution that can't
either be completely broken or just driven to impracticality performance
or space consumption wise.
We had an offline discussion with some of the Stratis and dm folks
fairly recently where they explained what Stratis is doing now to
mitigate this. Almost everybody agrees this approach stinks, which is
why I expected a similar first reaction to my "phase 1 full reservation"
model. The thought process there, however, is that rather than continue
to try and hack up various invasive solutions to provide such a simple
user visible behavior and ultimately not making progress, why not take
what they're doing and users are apparently already using and work to
make it better?
IOW, when thinking about the prospect of "hand tuning" above, I think
the more appropriate way to look at it is not that we're providing some
full end-user solution right off the bat. Instead, we're taking Stratis'
already existing "no overprovision" mode and starting to improve on it.
Step one lifts it into the kernel to make it dynamic (runtime
reservation vs provision time no overprovision enforcement), next steps
focus on making it more efficient while preserving the -ENOSPC safety
guarantee. No end user really needs to interact with it directly
until/unless filesystems grow the super-smart ability to do everything
automagically.
So it could very well be that this all remains an experimental feature
until "adaptive mode" can be made to work, but at least we have the
ability to solve the problem incrementally and without permanent
changes. The ability to just rip it out without having made any
permanent changes to filesystems or thin pool metadata should it just
happen to fail spectatularly is also a feature IMO. It could also be the
case that the simple sizing mechanism works well enough, and Stratis is
able to make good enough recommendations that most users are satisifed,
and there's really no need for the levels of complexity we're talking
about for adaptive mode (or preprovisioning) at all.
> If we then want to move to a different "accounting only" model for
> provisioning, we just change REQ_PROVISION?
>
AFAIU this relies on some permanent changes to dm that are not
necessarily trivial to undo..? If so, I think it's actually wiser to
move in the opposite direction. If reservation proves too broad and
expensive due to things like amplification, then move toward physical
provisioning and permanent snapshot changes to address that problem.
The benefit of this is that the reservation approach solves the
fundamental problem from the start, even if the implementation is
ultimately too impractical to be useful, so this mitigates the risk of
getting too far down the road with permanent changes to disk formats and
whatnot only to find the solution doesn't ultimately work.
> But I still see the problem of write amplification accounting being
> unsolved by the "filesystem accounting only" approach advocated
> here. We have no idea when the backing device has snapshots taken,
> we have no idea when a filesystem write IO actually consumes more
> thinp blocks than filesystem blocks, etc. How does the filesystem
> level reservation pool address these problems?
>
This is a fair concern in general, but as mentioned above, I think still
highlights a misunderstanding with the reserve metadata pool idea.
The key point I think is that metadata writes are not actually reserved.
The reserve pool exists solely to handle the -ENOSPC mode transition,
not to continuously supply ongoing metadata transactions. This means
write amplification is less of a concern: every successful FSB sized
write allocates DMB (device mapper block) sized blocks out of the thin
pool, further reducing the odds that subsequent writes to any
overlapping FSB blocks will ever require reservation for the current
active cycle of the log.
A snapshot could happen at any point, but dm-thin snapshots already call
into freeze infrastructure, which already quiesces the log. After that
point the game starts all over, all overwrites require allocation, and
the pool has to be sized large enough to acommodate that the filesystem
is able to quiesce in the event of -ENOSPC, particularly if snapshots
are in use.
So I'm not saying write amplification is not a problem. I think things like
crashing a filesystem, doing a snapshot, then running recovery could
exacerbate this problem, for example. But that's another corner case
that I don't necessarily think discredits the idea. For example, if XFS
really wanted to, it could add another pass to log recovery to do
"reserve provisions" on affected metadata before recovering anything, or
just scan and reserve provision the entire metadata allocated portion of
the fs, or refuse to proceed and require full reservation for any time
the log is dirty, etc. etc.
I think this is something that could use some examples (re: my earlier
question) to help work through whether the pool approach is sane, or if
the size would just always be too big. If not, you could still decide
that the configurable pool approach just doesn't work at all for XFS,
but track the outstanding metadata block usage somewhere (or estimate
based on existing ag buffer btree block counters), reserve that number
of blocks at mount time, and use that to guarantee all metadata block
overwrites will always succeed in the exact same way a preprovision
scheme would. A snapshot while mounted would bump the volume side
reservation appropriately or fail.
I suspect that trades off mount time performance for better snapshot
behavior, but again is just another handwavy option on the table for
consideration that doesn't preclude other fs' from possibly doing
something more simple.
Brian
> > Thoughts on any of the above?
>
> I'd say it went wrong at the requirements stage, resulting in an
> overly complex, over-engineered solution.
>
> > One general tradeoff with using reservations vs. preprovisioning is the
> > the latter can just use the provision/trim primitives to alloc/free LBA
> > ranges. My thought on that is those primitives could possibly be
> > modified to do the same sort of things with reservation as for physical
> > allocations. That seems fairly easy to do with bio op flags/modifiers,
> > though one thing I'm not sure about is how to submit a provision bio to
> > request a certain amount location agnostic blocks. I'd have to
> > investigate that more.
>
> Sure, if the constrained LBA space aspect of the REQ_PROVISION
> implementation causes issues, then we see if we can optimise away
> the fixed LBA space requirement.
>
>
> --
> Dave Chinner
> [email protected]
>