Introduce new CSAL ftl bdev #1771
Conversation
|
Discussion question: The |
MaisenbacherD
force-pushed
the
csal-support
branch
from
b4e6c94
to
6374141
Compare
| sudo PCI_ALLOWED="<PCI-ADDRESS>" ./spdk-rs/spdk/scripts/setup.sh reset | ||
| ``` | ||
|
|
||
| Please do not submit pull requests with active cargo test cases that require PCIe devices to be present. |
There was a problem hiding this comment.
Maybe in the future we can have a feature flag for PCIe tests which can be run on a dedicated test system :)
Btw, can we add virtual PCIe devices with qemu? That could be a way of running those tests without having spare pcie nvme devices
There was a problem hiding this comment.
Do you mean a rust feature flag, such that PCIe support can be selected at compile time?
It makes sense to me! I briefly thought about adding a feature flag because of the failing test case but went with the ignore attribute for simplicity.
I can rework this to be a rust feature flag within this PR, however, if I attempt this (https://doc.rust-lang.org/cargo/reference/features.html) I get the following error:
error: failed to parse manifest at `/home/dennis/src/mayastor/Cargo.toml`
Caused by:
this virtual manifest specifies a `features` section, which is not allowed
Do you know what needs to be done here?
A QEMU NVMe emulated device via PCIe would work for testing. One just needs to make sure that the cache device is configured with ms=8 https://qemu-project.gitlab.io/qemu/system/devices/nvme.html#metadata.
There was a problem hiding this comment.
Do you know what needs to be done here?
The feature flag would need to be added to the io-engine crate, see some examples:
Line 12 in 2825e5e
I'll leave it up to you, fine to keep ignore for now as well.
A QEMU NVMe emulated device via PCIe would work for testing. One just needs to make sure that the cache device is configured with ms=8 https://qemu-project.gitlab.io/qemu/system/devices/nvme.html#metadata.
Maybe something we may be able to do here then. We run the python BDD tests in a vm, maybe we can configure it with an emulated PCIe device and be able to run these tests then?
Alternatively we can setup VM's on demand from the tests themselves? With nix we can setup test vm's easily, example: https://github.com/openebs/mayastor-extensions/blob/develop/tests/helm/test.nix
But I don't think there's emulated nvme support there...
There was a problem hiding this comment.
I now introduced a new feature flag :)
Yes we should definitely look into running the test in an emulated QEMU environment or something similar.
That on demand VM setup you linked looks interesting. It would take a bit time to play with it and figure out the capabilities.
Should we address the CI setup within this PR?
There was a problem hiding this comment.
We can do in another PR, let's get this one merged :)
There was a problem hiding this comment.
I will talk with you outside of this PR to discuss this :)
MaisenbacherD
force-pushed
the
csal-support
branch
from
6374141
to
c88d9a2
Compare
MaisenbacherD
force-pushed
the
csal-support
branch
from
c88d9a2
to
12e974e
Compare
|
bors try |
tryBuild failed: |
The SPDK ftl bdev allows to create a layered device with a (fast) cache device for buffering writes that get eventually flushed out sequentially to a base device. SDPK ftl is also known as the Cloud Storage Acceleration Layer (CSAL). This kind of device potentially enables the use of emerging storage interfaces like Zoned Namespace (ZNS) or Flexible Data Placement (FDP) capable NVMe devices. Up to this point, those NVMe command sets are not yet supported in SPDK upstream. However, the acceleration aspect of a fast cache device already adds value. With future support of new devices for SDPK ftl, Mayastor would already be capable of utilizing those features simply by upgrading SDPK. For now, the `ftl_mount_fs` test cases are hidden behind the `nvme-pci-tests` flag because PCIe devices are required for this test until SPDK v24.09 is picked up, which introduces variable sector size emulation for ftl devices. To run these tests use the following: ``` RUST_LOG=TRACE cargo test --features nvme-pci-tests -- --test-threads 1 \ --test ftl_mount_fs --nocapture ``` This patch introduces a new ftl device uri scheme: ``` ftl:///$ftl_device_name?bbdev=$bbdev_uri_percent_encoded&cbdev=$cbdev_uri_percent_encoded ``` The bbdev_uri and cbdev_uri need to use percent encoding on '?' (= '%3F') and '&' (= '%26') segment dividers. This is needed so we can successfully parse the ftl uri. Optionally, the sub-uris may also be fully percent-encoded. With SPDK v24.05 only PCIe devices will work where the cache device is formatted to 4KiB+64B LBA format and the base device to 4KiB LBA format. From SPDK v24.09 on, any device with a block size of 4KiB will work for both cache and base device. CSAL paper reference: https://dl.acm.org/doi/10.1145/3627703.3629566 Co-authored-by: Indraneel M <[email protected]> Signed-off-by: Dennis Maisenbacher <[email protected]>
Signed-off-by: Dennis Maisenbacher <[email protected]>
Init SPDK tracing when `MayastorEnvironment.num_enties` is set to a positive integer. This can be helpfull when developing new features. Traces can be found in `/dev/shm` and have the pid or if not negative the `MayastorEnvironment.shm_id` as a suffix. Further information about traces and how to read the captured traces can be found here: https://spdk.io/doc/nvmf_tgt_tracepoints.html Signed-off-by: Dennis Maisenbacher <[email protected]>
MaisenbacherD
force-pushed
the
csal-support
branch
from
12e974e
to
1dac4b4
Compare
|
Addressed the linter errors :) |
For this PR to be merged the PR openebs/spdk-rs#67
for spdk-rs needs to be resolved and the submodule reference must be
updated accordingly. :)
The SPDK ftl bdev allows to create a layered device with a (fast) cache
device for buffering writes that get eventually flushed out sequentially
to a base device. SDPK ftl is also known as the Cloud Storage
Acceleration Layer (CSAL).
This kind of device potentially enables the use of emerging storage
interfaces like Zoned Namespace (ZNS) or Flexible Data Placement (FDP)
capable NVMe devices. Up to this point, those NVMe command sets are not
yet supported in SPDK upstream. However, the acceleration aspect of a
fast cache device already adds value.
With future support of new devices for SDPK ftl, Mayastor would already
be capable of utilizing those features simply by upgrading SDPK.
For now, the
ftl_mount_fstest cases are hidden behind theignoreattribute because PCIe devices are required for this test until
SPDK v24.09 is picked up, which introduces variable sector size emulation
for ftl devices.
To run these tests use the following:
This patch introduces a new ftl device uri scheme:
<dev-uri-nested-encoding>can be any already valid device uri where '?'are replaced with '!' and '&' are replaced with '|'.
With SPDK v24.05 only PCIe devices will work where the cache device is
formatted to 4KiB+64B LBA format and the base device to 4KiB LBA format.
From SPDK v24.09 on, any device with a block size of 4KiB will work for
both cache and base device.
CSAL paper reference:
https://dl.acm.org/doi/10.1145/3627703.3629566