Upcoming work to instrument the rate of RPC requests by consumer (and eventually
rate limit) requires that we thread the `RPCContext` through all RPC
handlers so that we can access the underlying connection. This changeset adds
the context to everywhere we intend to initially support it and intentionally
excludes streaming RPCs and client RPCs.
To improve the ergonomics of adding the context everywhere its needed and to
clarify the requirements of dynamic vs static handlers, I've also done a good
bit of refactoring here:
* canonicalized the RPC handler fields so they're as close to identical as
possible without introducing unused fields (i.e. I didn't add loggers if the
handler doesn't use them already).
* canonicalized the imports in the handler files.
* added a `NewExampleEndpoint` function for each handler that ensures we're
constructing the handlers with the required arguments.
* reordered the registration in server.go to match the order of the files (to
make it easier to see if we've missed one), and added a bunch of commentary
there as to what the difference between static and dynamic handlers is.
A test flake revealed a bug in the CSI unpublish workflow, where an unpublish
that comes from a client that's successfully done the node-unpublish step will
not have the claim checkpointed if the controller-unpublish step fails. This
will result in a delay in releasing the volume claim until the next GC.
This changeset also ensures we're using a new snapshot after each write to raft,
and fixes two timing issues in test where either the volume watcher can
unpublish before the unpublish RPC is sent or we don't wait long enough in
resource-restricted environements like GHA.
A Nomad user reported problems with CSI volumes associated with failed
allocations, where the Nomad server did not send a controller unpublish RPC.
The controller unpublish is skipped if other non-terminal allocations on the
same node claim the volume. The check has a bug where the allocation belonging
to the claim being freed was included in the check incorrectly. During a normal
allocation stop for job stop or a new version of the job, the allocation is
terminal. But allocations that fail are not yet marked terminal at the point in
time when the client sends the unpublish RPC to the server.
For CSI plugins that support controller attach/detach, this means that the
controller will not be able to detach the volume from the allocation's host and
the replacement claim will fail until a GC is run. This changeset fixes the
conditional so that the claim's own allocation is not included, and makes the
logic easier to read. Include a test case covering this path.
Also includes two minor extra bugfixes:
* Entities we get from the state store should always be copied before
altering. Ensure that we copy the volume in the top-level unpublish workflow
before handing off to the steps.
* The list stub object for volumes in `nomad/structs` did not match the stub
object in `api`. The `api` package also did not include the current
readers/writers fields that are expected by the UI. True up the two objects and
add the previously undocumented fields to the docs.
If the node has been GC'd or is down, we can't send it a node
unpublish. The CSI spec requires that we don't send the controller
unpublish before the node unpublish, but in the case where a node is
gone we can't know the final fate of the node unpublish step.
The `csi_hook` on the client will unpublish if the allocation has
stopped and if the host is terminated there's no mount for the volume
anyways. So we'll now assume that the node has unpublished at its
end. If it hasn't, any controller unpublish will potentially hang or
error and need to be retried.
The volume watcher design was based on deploymentwatcher and drainer,
but has an important difference: we don't want to maintain a goroutine
for the lifetime of the volume. So we stop the volumewatcher goroutine
for a volume when that volume has no more claims to free. But the
shutdown races with updates on the parent goroutine, and it's possible
to drop updates. Fortunately these updates are picked up on the next
core GC job, but we're most likely to hit this race when we're
replacing an allocation and that's the time we least want to wait.
Wait until the volume has "settled" before stopping this goroutine so
that the race between shutdown and the parent goroutine sending on
`<-updateCh` is pushed to after the window we most care about quick
freeing of claims.
* Fixes a resource leak when volumewatchers are no longer needed. The
volume is nil and can't ever be started again, so the volume's
`watcher` should be removed from the top-level `Watcher`.
* De-flakes the GC job test: the test throws an error because the
claimed node doesn't exist and is unreachable. This flaked instead of
failed because we didn't correctly wait for the first pass through the
volumewatcher.
Make the GC job wait for the volumewatcher to reach the quiescent
timeout window state before running the GC eval under test, so that
we're sure the GC job's work isn't being picked up by processing one
of the earlier claims. Update the claims used so that we're sure the
GC pass won't hit a node unpublish error.
* Adds trace logging to unpublish operations
In #12112 and #12113 we solved for the problem of races in releasing
volume claims, but there was a case that we missed. During a node
drain with a controller attach/detach, we can hit a race where we call
controller publish before the unpublish has completed. This is
discouraged in the spec but plugins are supposed to handle it
safely. But if the storage provider's API is slow enough and the
plugin doesn't handle the case safely, the volume can get "locked"
into a state where the provider's API won't detach it cleanly.
Check the claim before making any external controller publish RPC
calls so that Nomad is responsible for the canonical information about
whether a volume is currently claimed.
This has a couple side-effects that also had to get fixed here:
* Changing the order means that the volume will have a past claim
without a valid external node ID because it came from the client, and
this uncovered a separate bug where we didn't assert the external node
ID was valid before returning it. Fallthrough to getting the ID from
the plugins in the state store in this case. We avoided this
originally because of concerns around plugins getting lost during node
drain but now that we've fixed that we may want to revisit it in
future work.
* We should make sure we're handling `FailedPrecondition` cases from
the controller plugin the same way we handle other retryable cases.
* Several tests had to be updated because they were assuming we fail
in a particular order that we're no longer doing.
Pass-through the `-secret` and `-parameter` flags to allow setting
parameters for the snapshot and overriding the secrets we've stored on
the CSI volume in the state store.
When a node is garbage collected, we assume that the volume is no
longer attached to it and ignore the `ErrUnknownNode` error. But we
used `errors.Is` to check for a wrapped error, and RPC flattens the
errors during serialization. This results in an error check that works
in automated testing but not in real clusters. Use a string contains
check instead.
CSI `CreateVolume` RPC is idempotent given that the topology,
capabilities, and parameters are unchanged. CSI volumes have many
user-defined fields that are immutable once set, and many fields that
are not user-settable.
Update the `Register` RPC so that updating a volume via the API merges
onto any existing volume without touching Nomad-controlled fields,
while validating it with the same strict requirements expected for
idempotent `CreateVolume` RPCs.
Also, clarify that this state store method is used for everything, not just
for the `Register` RPC.
The `CreateSnapshot` RPC expects a plugin ID to be set by the API, but
in the common case of the `nomad volume snapshot create` command, we
don't ask the user for the plugin ID because it's available from the
volume we're snapshotting.
Change the order of the RPC so that we get the volume first and then
use the volume's plugin ID for the plugin if the API didn't set the
value.
The `CSIPlugin.List` RPC was intended to accept a prefix to filter the
list of plugins being listed. This was being accidentally being done
in the state store instead, which contributed to incorrect filtering
behavior for plugins in the `volume plugin status` command.
Move the prefix matching into the RPC so that it calls the
prefix-matching method in the state store if we're looking for a
prefix.
Update the `plugin status command` to accept a prefix for the plugin
ID argument so that it matches the expected behavior of other commands.
When using a prefix value and the * wildcard for namespace, the endpoint
would not take the prefix value into consideration due to the order in
which the checks were executed but also the logic for retrieving volumes
from the state store.
This commit changes the order to check for a prefix first and wraps the
result iterator of the state store query in a filter to apply the
prefix.
* The volume claim GC method and volumewatcher both have logic
collecting terminal allocations that duplicates most of the logic
that's now in the state store's `CSIVolumeDenormalize` method. Copy
this logic into the state store so that all code paths have the same
view of the past claims.
* Remove logic in the volume claim GC that now lives in the state
store's `CSIVolumeDenormalize` method.
* Remove logic in the volumewatcher that now lives in the state
store's `CSIVolumeDenormalize` method.
* Remove logic in the node unpublish RPC that now lives in the state
store's `CSIVolumeDenormalize` method.
When `nomad volume create` was introduced in Nomad 1.1.0, we changed the
volume spec to take a list of capabilities rather than a single capability, to
meet the requirements of the CSI spec. When a volume is registered via `nomad
volume register`, we should be using the same fields to validate the volume
with the controller plugin.
Include the VolumeCapability.MountVolume data in
ControllerPublishVolume, CreateVolume, and ValidateVolumeCapabilities
RPCs sent to the CSI controller. The previous behavior was to only
include the MountVolume capability in the NodeStageVolume request, which
on some CSI implementations would be rejected since the Volume was not
originally provisioned with the specific mount capabilities requested.
Plugins could potentially ignore the `max_entries` field and return a list of
entries that is greater, so we slice the return value in the server RPC to
enforce these value. But page sizes less than the number of entries for the
external CSI ListVolumes and ListSnapshots RPCs could cause a panic, so fix
the boundary checking.
The plugin stub object does not include fine-grained capability checks, which
means `nomad volume status -verbose` will return ugly and verbose error
"Unimplemented" messages from the plugin if it does not support the CSI
`ListVolumes` RPC. Return a nicer error message from our RPC handler instead.
Registration of Nomad volumes previously allowed for a single volume
capability (access mode + attachment mode pair). The recent `volume create`
command requires that we pass a list of requested capabilities, but the
existing workflow for claiming volumes and attaching them on the client
assumed that the volume's single capability was correct and unchanging.
Add `AccessMode` and `AttachmentMode` to `CSIVolumeClaim`, use these fields to
set the initial claim value, and add backwards compatibility logic to handle
the existing volumes that already have claims without these fields.
In order to support new controller RPCs, we need to fingerprint volume
capabilities in more detail and perform controller RPCs only when the specific
capability is present. This fixes a bug in Ceph support where the plugin can
only suport create/delete but we assume that it also supports attach/detach.
Callers of `CSIVolumeByID` are generally assuming they should receive a single
volume. This potentially results in feasibility checking being performed
against the wrong volume if a volume's ID is a prefix substring of other
volume (for example: "test" and "testing").
Removing the incorrect prefix matching from `CSIVolumeByID` breaks prefix
matching in the command line client. Add the required elements for prefix
matching to the commands and API.
When making updates to CSI plugins, the state store methods that have open
write transactions were querying the state store using the same methods used
by the CSI RPC endpoint, but these method creates their own top-level read
transactions. During concurrent plugin updates (as happens when a plugin job
is stopped), this can cause write skew in the plugin counts.
* Refactor the CSIPlugin query methods to have an implementation method that
accepts a transaction, which can be called with either a read txn or a write
txn.
* Refactor the CSIVolume query methods to have an implementation method that
accepts a transaction, which can be called with either a read txn or a write
txn.
* CSI volumes need to be "denormalized" with their plugins and (optionally)
allocations. Read-only RPC endpoints should take a snapshot so that we can
make multiple state store method calls with a consistent view.
The unpublish workflow requires that we know the mode (RW vs RO) if we want to
unpublish the node. Update the hook and the Unpublish RPC so that we mark the
claim for release in a new state but leave the mode alone. This fixes a bug
where RO claims were failing node unpublish.
The core job GC doesn't know the mode, but we don't need it for that workflow,
so add a mode specifically for GC; the volumewatcher uses this as a sentinel
to check whether claims (with their specific RW vs RO modes) need to be claimed.
The soundness guarantees of the CSI specification leave a little to be desired
in our ability to provide a 100% reliable automated solution for managing
volumes. This changeset provides a new command to bridge this gap by providing
the operator the ability to intervene.
The command doesn't take an allocation ID so that the operator doesn't have to
keep track of alloc IDs that may have been GC'd. Handle this case in the
unpublish RPC by sending the client RPC for all the terminal/nil allocs on the
selected node.
When the client-side actions of a CSI client RPC succeed but we get
disconnected during the RPC or we fail to checkpoint the claim state, we want
to be able to retry the client RPC without getting blocked by the client-side
state (ex. mount points) already having been cleaned up in previous calls.
Using the count of node claims from earlier in the `CSIVolume.Unpublish RPC
doesn't correctly account for cases where the RPC was interrupted but
checkpointed. Instead, we'll check the current allocation count and status to
determine whether we need to send a controller unpublish.
Some of the CSI RPC endpoints were missing validation that the ID or
the Volume definition was present. This could result in nonsense
`CSIVolume` structs being written to raft during registration. This
changeset corrects that bug and adds validation checks to present
nicer error messages to operators in some other cases.
The MVP for CSI in the 0.11.0 release of Nomad did not include support
for opaque volume parameters or volume context. This changeset adds
support for both.
This also moves args for ControllerValidateCapabilities into a struct.
The CSI plugin `ControllerValidateCapabilities` struct that we turn
into a CSI RPC is accumulating arguments, so moving it into a request
struct will reduce the churn of this internal API, make the plugin
code more readable, and make this method consistent with the other
plugin methods in that package.
CSI plugins can require credentials for some publishing and
unpublishing workflow RPCs. Secrets are configured at the time of
volume registration, stored in the volume struct, and then passed
around as an opaque map by Nomad to the plugins.
This changeset implements a periodic garbage collection of unused CSI
plugins. Plugins are self-cleaning when the last allocation for a
plugin is stopped, but this feature will cover any missing edge cases
and ensure that upgrades from 0.11.0 and 0.11.1 get any stray plugins
cleaned up.
The `CSIVolumeClaim` fields were added after 0.11.1, so claims made
before that may be missing the value. Repair this when we read the
volume out of the state store.
The `NodeID` field was added after 0.11.0, so we need to ensure it's
been populated during upgrades from 0.11.0.
Adds a `CSIVolumeClaim` type to be tracked as current and past claims
on a volume. Allows for a client RPC failure during node or controller
detachment without having to keep the allocation around after the
first garbage collection eval.
This changeset lays groundwork for moving the actual detachment RPCs
into a volume watching loop outside the GC eval.
Follow-up for a method missed in the refactor for #7688. The
`volAndPluginLookup` method is only ever called from the server's `CSI`
RPC and never the `ClientCSI` RPC, so move it into that scope.
The current design of `ClientCSI` RPC requires that callers in the
server know about the free-standing `nodeForControllerPlugin`
function. This makes it difficult to send `ClientCSI` RPC messages
from subpackages of `nomad` and adds a bunch of boilerplate to every
server-side caller of a controller RPC.
This changeset makes it so that the `ClientCSI` RPCs will populate and
validate the controller's client node ID if it hasn't been passed by
the caller, centralizing the logic of picking and validating
controller targets into the `nomad.ClientCSI` struct.
If a volume-claiming alloc stops and the CSI Node plugin that serves
that alloc's volumes is missing, there's no way for the allocrunner
hook to send the `NodeUnpublish` and `NodeUnstage` RPCs.
This changeset addresses this issue with a redesign of the client-side
for CSI. Rather than unmounting in the alloc runner hook, the alloc
runner hook will simply exit. When the server gets the
`Node.UpdateAlloc` for the terminal allocation that had a volume claim,
it creates a volume claim GC job. This job will made client RPCs to a
new node plugin RPC endpoint, and only once that succeeds, move on to
making the client RPCs to the controller plugin. If the node plugin is
unavailable, the GC job will fail and be requeued.
