Use only the agent token for deregistration during anti-entropy
The previous behavior had the agent attempt to use the "service" token
(i.e. from the `token` field in a service definition file), and if that
was not set then it would use the agent token.
The previous behavior was problematic because, if the service token had
been deleted, the deregistration request would fail. The agent would
retry the deregistration during each anti-entropy sync, and the
situation would never resolve.
The new behavior is to only/always use the agent token for service and
check deregistration during anti-entropy. This approach is:
* Simpler: No fallback logic to try different tokens
* Faster (slightly): No time spent attempting the service token
* Correct: The agent token is able to deregister services on that
agent's node, because:
* node:write permissions allow deregistration of services/checks on
that node.
* The agent token must have node:write permission, or else the agent
is not be able to (de)register itself into the catalog
Co-authored-by: Vesa Hagström <weeezes@gmail.com>
Although the metric is defined, there is no code which ever sets its
value - the code in question is genuinely asymmetric - there are 3 types
of object for which registration can be tracked, but only 2 for which
deregistration can be tracked.
Having this type live in the agent/consul package makes it difficult to
put anything that relies on token resolution (e.g. the new gRPC services)
in separate packages without introducing import cycles.
For example, if package foo imports agent/consul for the ACLResolveResult
type it means that agent/consul cannot import foo to register its service.
We've previously worked around this by wrapping the ACLResolver to
"downgrade" its return type to an acl.Authorizer - aside from the
added complexity, this also loses the resolved identity information.
In the future, we may want to move the whole ACLResolver into the
acl/resolver package. For now, putting the result type there at least,
fixes the immediate import cycle issues.
* add config watcher to the config package
* add logging to watcher
* add test and refactor to add WatcherEvent.
* add all API calls and fix a bug with recreated files
* add tests for watcher
* remove the unnecessary use of context
* Add debug log and a test for file rename
* use inode to detect if the file is recreated/replaced and only listen to create events.
* tidy ups (#1535)
* tidy ups
* Add tests for inode reconcile
* fix linux vs windows syscall
* fix linux vs windows syscall
* fix windows compile error
* increase timeout
* use ctime ID
* remove remove/creation test as it's a use case that fail in linux
* fix linux/windows to use Ino/CreationTime
* fix the watcher to only overwrite current file id
* fix linter error
* fix remove/create test
* set reconcile loop to 200 Milliseconds
* fix watcher to not trigger event on remove, add more tests
* on a remove event try to add the file back to the watcher and trigger the handler if success
* fix race condition
* fix flaky test
* fix race conditions
* set level to info
* fix when file is removed and get an event for it after
* fix to trigger handler when we get a remove but re-add fail
* fix error message
* add tests for directory watch and fixes
* detect if a file is a symlink and return an error on Add
* rename Watcher to FileWatcher and remove symlink deref
* add fsnotify@v1.5.1
* fix go mod
* do not reset timer on errors, rename OS specific files
* rename New func
* events trigger on write and rename
* add missing test
* fix flaking tests
* fix flaky test
* check reconcile when removed
* delete invalid file
* fix test to create files with different mod time.
* back date file instead of sleeping
* add watching file in agent command.
* fix watcher call to use new API
* add configuration and stop watcher when server stop
* add certs as watched files
* move FileWatcher to the agent start instead of the command code
* stop watcher before replacing it
* save watched files in agent
* add add and remove interfaces to the file watcher
* fix remove to not return an error
* use `Add` and `Remove` to update certs files
* fix tests
* close events channel on the file watcher even when the context is done
* extract `NotAutoReloadableRuntimeConfig` is a separate struct
* fix linter errors
* add Ca configs and outgoing verify to the not auto reloadable config
* add some logs and fix to use background context
* add tests to auto-config reload
* remove stale test
* add tests to changes to config files
* add check to see if old cert files still trigger updates
* rename `NotAutoReloadableRuntimeConfig` to `StaticRuntimeConfig`
* fix to re add both key and cert file. Add test to cover this case.
* review suggestion
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
* add check to static runtime config changes
* fix test
* add changelog file
* fix review comments
* Apply suggestions from code review
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
* update flag description
Co-authored-by: FFMMM <FFMMM@users.noreply.github.com>
* fix compilation error
* add static runtime config support
* fix test
* fix review comments
* fix log test
* Update .changelog/12329.txt
Co-authored-by: Dan Upton <daniel@floppy.co>
* transfer tests to runtime_test.go
* fix filewatcher Replace to not deadlock.
* avoid having lingering locks
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
* split ReloadConfig func
* fix warning message
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
* convert `FileWatcher` into an interface
* fix compilation errors
* fix tests
* extract func for adding and removing files
Co-authored-by: Ashwin Venkatesh <ashwin@hashicorp.com>
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
Co-authored-by: FFMMM <FFMMM@users.noreply.github.com>
Co-authored-by: Daniel Upton <daniel@floppy.co>
The race detector noticed this initially in `TestAgentConfigWatcherSidecarProxy` but it is not restricted to just tests.
The two main changes here were:
- ensure that before we mutate the internal `agent/local` representation of a Service (for tags or VIPs) we clone those fields
- ensure that there's no function argument joint ownership between the caller of a function and the local state when calling `AddService`, `AddCheck`, and related using `copystructure` for now.
* clone the service under lock to avoid a data race
* add change log
* create a struct and copy the pointer to mutate it to avoid a data race
* fix failing test
* revert added space
* add comments, to clarify the data race.
Error messages related to service and check operations previously included
the following substrings:
- service %q
- check %q
From this error message, it isn't clear that the expected field is the ID for
the entity, not the name. For example, if the user has a service named test,
the error message would read 'Unknown service "test"'. This is misleading -
a service with that *name* does exist, but not with that *ID*.
The substrings above have been modified to make it clear that ID is needed,
not name:
- service with ID %q
- check with ID %q
Only default to the user token and agent token for the sync. Change the
exported methods to only return the stored tokens associated with a
specific check or service.
When de-registering in anti-entropy sync, when there is no service or
check token.
The agent token will fall back to the default (aka user) token if no agent
token is set, so the existing behaviour still works, but it will prefer
the agent token over the user token if both are set.
ref: https://www.consul.io/docs/agent/options#acl_tokens
The agent token seems more approrpiate in this case, since this is an
"internal operation", not something initiated by the user.
These types are used as values (not pointers) in other structs. Using a pointer receiver causes
problems when the value is printed. fmt will not call the String method if it is passed a value
and the String method has a pointer receiver. By using a value receiver the correct string is printed.
Also remove some unused methods.
In current implementation of Consul, check alias cannot determine
if a service exists or not. Because a service without any check
is semantically considered as passing, so when no healthchecks
are found for an agent, the check was considered as passing.
But this make little sense as the current implementation does not
make any difference between:
* a non-existing service (passing)
* a service without any check (passing as well)
In order to make it work, we have to ensure that when a check did
not find any healthcheck, the service does indeed exists. If it
does not, lets consider the check as failing.
Previously this happened to be using the method on the Server/Client that was meant to allow the ACLResolver to locally resolve tokens. On Servers that had tokens (primary or secondary dc + token replication) this function would lookup the token from raft and return the ACLIdentity. On clients this was always a noop. We inadvertently used this function instead of creating a new one when we added logging accessor ids for permission denied RPC requests.
With this commit, a new method is used for resolving the identity properly via the ACLResolver which may still resolve locally in the case of being on a server with tokens but also supports remote token resolution.
The Init method provided the same functionality as the New constructor.
The constructor is both more widely used, and more idiomatic, so remove
the Init method.
This change is in preparation for fixing printing of these IDs.
This fixes#7020.
There are two problems this PR solves:
* if the node info changes it is highly likely to get service and check registration permission errors unless those service tokens have node:write. Hopefully services you register don’t have this permission.
* the timer for a full sync gets reset for every partial sync which means that many partial syncs are preventing a full sync from happening
Instead of syncing node info last, after services and checks, and possibly saving one RPC because it is included in every service sync, I am syncing node info first. It is only ever going to be a single RPC that we are only doing when node info has changed. This way we are guaranteed to sync node info even when something goes wrong with services or checks which is more likely because there are more syncs happening for them.
* Fix segfault when removing both a service and associated check
updateSyncState creates entries in the services and checks maps for
remote services/checks that are not found locally, so that we can then
make sure to delete them in our reconciliation process. However, the
values added to the map are missing key fields that the rest of the code
expects to not be nil.
* Add comment stating Check field can be nil
Deregistering a service from the catalog automatically deregisters its
checks, however the agent still performs a deregister call for each
service checks even after the service has been deregistered.
With ACLs enabled this results in logs like:
"message:consul: "Catalog.Deregister" RPC failed to server
server_ip:8300: rpc error making call: rpc error making call: Unknown
check 'check_id'"
This change removes associated checks from the agent state when
deregistering a service, which results in less calls to the servers and
supresses the error logs.
Fixes: #5396
This PR adds a proxy configuration stanza called expose. These flags register
listeners in Connect sidecar proxies to allow requests to specific HTTP paths from outside of the node. This allows services to protect themselves by only
listening on the loopback interface, while still accepting traffic from non
Connect-enabled services.
Under expose there is a boolean checks flag that would automatically expose all
registered HTTP and gRPC check paths.
This stanza also accepts a paths list to expose individual paths. The primary
use case for this functionality would be to expose paths for third parties like
Prometheus or the kubelet.
Listeners for requests to exposed paths are be configured dynamically at run
time. Any time a proxy, or check can be registered, a listener can also be
created.
In this initial implementation requests to these paths are not
authenticated/encrypted.
Read requests performed during anti antropy full sync currently target
the leader only. This generates a non-negligible load on the leader when
the DC is large enough and can be offloaded to the followers following
the "eventually consistent" policy for the agent state.
We switch the AE read calls to use stale requests with a small (2s)
MaxStaleDuration value and make sure we do not read too fast after a
write.
The observed bug was that a full restart of a consul datacenter (servers
and clients) in conjunction with a restart of a connect-flavored
application with bring-your-own-service-registration logic would very
frequently cause the envoy sidecar service check to never reflect the
aliased service.
Over the course of investigation several bugs and unfortunate
interactions were corrected:
(1)
local.CheckState objects were only shallow copied, but the key piece of
data that gets read and updated is one of the things not copied (the
underlying Check with a Status field). When the stock code was run with
the race detector enabled this highly-relevant-to-the-test-scenario field
was found to be racy.
Changes:
a) update the existing Clone method to include the Check field
b) copy-on-write when those fields need to change rather than
incrementally updating them in place.
This made the observed behavior occur slightly less often.
(2)
If anything about how the runLocal method for node-local alias check
logic was ever flawed, there was no fallback option. Those checks are
purely edge-triggered and failure to properly notice a single edge
transition would leave the alias check incorrect until the next flap of
the aliased check.
The change was to introduce a fallback timer to act as a control loop to
double check the alias check matches the aliased check every minute
(borrowing the duration from the non-local alias check logic body).
This made the observed behavior eventually go away when it did occur.
(3)
Originally I thought there were two main actions involved in the data race:
A. The act of adding the original check (from disk recovery) and its
first health evaluation.
B. The act of the HTTP API requests coming in and resetting the local
state when re-registering the same services and checks.
It took awhile for me to realize that there's a third action at work:
C. The goroutines associated with the original check and the later
checks.
The actual sequence of actions that was causing the bad behavior was
that the API actions result in the original check to be removed and
re-added _without waiting for the original goroutine to terminate_. This
means for brief windows of time during check definition edits there are
two goroutines that can be sending updates for the alias check status.
In extremely unlikely scenarios the original goroutine sees the aliased
check start up in `critical` before being removed but does not get the
notification about the nearly immediate update of that check to
`passing`.
This is interlaced wit the new goroutine coming up, initializing its
base case to `passing` from the current state and then listening for new
notifications of edge triggers.
If the original goroutine "finishes" its update, it then commits one
more write into the local state of `critical` and exits leaving the
alias check no longer reflecting the underlying check.
The correction here is to enforce that the old goroutines must terminate
before spawning the new one for alias checks.
Paul Glass