Starting from and extending the mechanism introduced in #12110 we can specially handle the 3 main special Consul RPC endpoints that react to many config entries in a single blocking query in Connect:
- `DiscoveryChain.Get`
- `ConfigEntry.ResolveServiceConfig`
- `Intentions.Match`
All of these will internally watch for many config entries, and at least one of those will likely be not found in any given query. Because these are blends of multiple reads the exact solution from #12110 isn't perfectly aligned, but we can tweak the approach slightly and regain the utility of that mechanism.
### No Config Entries Found
In this case, despite looking for many config entries none may be found at all. Unlike #12110 in this scenario we do not return an empty reply to the caller, but instead synthesize a struct from default values to return. This can be handled nearly identically to #12110 with the first 1-2 replies being non-empty payloads followed by the standard spurious wakeup suppression mechanism from #12110.
### No Change Since Last Wakeup
Once a blocking query loop on the server has completed and slept at least once, there is a further optimization we can make here to detect if any of the config entries that were present at specific versions for the prior execution of the loop are identical for the loop we just woke up for. In that scenario we can return a slightly different internal sentinel error and basically externally handle it similar to #12110.
This would mean that even if 20 discovery chain read RPC handling goroutines wakeup due to the creation of an unrelated config entry, the only ones that will terminate and reply with a blob of data are those that genuinely have new data to report.
### Extra Endpoints
Since this pattern is pretty reusable, other key config-entry-adjacent endpoints used by `agent/proxycfg` also were updated:
- `ConfigEntry.List`
- `Internal.IntentionUpstreams` (tproxy)
Many places in consul already treated node names case insensitively.
The state store indexes already do it, but there are a few places that
did a direct byte comparison which have now been corrected.
One place of particular consideration is ensureCheckIfNodeMatches
which is executed during snapshot restore (among other places). If a
node check used a slightly different casing than the casing of the node
during register then the snapshot restore here would deterministically
fail. This has been fixed.
Primary approach:
git grep -i "node.*[!=]=.*node" -- ':!*_test.go' ':!docs'
git grep -i '\[[^]]*member[^]]*\]
git grep -i '\[[^]]*\(member\|name\|node\)[^]]*\]' -- ':!*_test.go' ':!website' ':!ui' ':!agent/proxycfg/testing.go:' ':!*.md'
There are some cross-config-entry relationships that are enforced during
"graph validation" at persistence time that are required to be
maintained. This means that config entries may form a digraph at times.
Config entry replication procedes in a particular sorted order by kind
and name.
Occasionally there are some fixups to these digraphs that end up
replicating in the wrong order and replicating the leaves
(ingress-gateway) before the roots (service-defaults) leading to
replication halting due to a graph validation error related to things
like mismatched service protocol requirements.
This PR changes replication to give each computed change (upsert/delete)
a fair shot at being applied before deciding to terminate that round of
replication in error. In the case where we've simply tried to do the
operations in the wrong order at least ONE of the outstanding requests
will complete in the right order, leading the subsequent round to have
fewer operations to do, with a smaller likelihood of graph validation
errors.
This does not address all scenarios, but for scenarios where the edits
are being applied in the wrong order this should avoid replication
halting.
Fixes#9319
The scenario that is NOT ADDRESSED by this PR is as follows:
1. create: service-defaults: name=new-web, protocol=http
2. create: service-defaults: name=old-web, protocol=http
3. create: service-resolver: name=old-web, redirect-to=new-web
4. delete: service-resolver: name=old-web
5. update: service-defaults: name=old-web, protocol=grpc
6. update: service-defaults: name=new-web, protocol=grpc
7. create: service-resolver: name=old-web, redirect-to=new-web
If you shutdown dc2 just before (4) and turn it back on after (7)
replication is impossible as there is no single edit you can make to
make forward progress.
Otherwise when the query times out we might incorrectly send a value for
the reply, when we should send an empty reply.
Also document errNotFound and how to handle the result in that case.
By using the query results as state.
Blocking queries are efficient when the query matches some results,
because the ModifyIndex of those results, returned as queryMeta.Mindex,
will never change unless the items themselves change.
Blocking queries for non-existent items are not efficient because the
queryMeta.Index can (and often does) change when other entities are
written.
This commit reduces the churn of these queries by using a different
comparison for "has changed". Instead of using the modified index, we
use the existence of the results. If the previous result was "not found"
and the new result is still "not found", we know we can ignore the
modified index and continue to block.
This is done by setting the minQueryIndex to the returned
queryMeta.Index, which prevents the query from returning before a state
change is observed.
This test shows how blocking queries are not efficient when the query
returns no results. The test fails with 100+ calls instead of the
expected 2.
This test is still a bit flaky because it depends on the timing of the
writes. It can sometimes return 3 calls.
A future commit should fix this and make blocking queries even more
optimal for not-found results.
Follow the Go convention of accepting a small interface that documents
the methods used by the function.
Clarify the rules for implementing a query function passed to
blockingQuery.
This will both save on unnecessary raft operations as well as
unnecessarily incrementing the raft modify index of config entries
subject to no-op updates.
This commit syncs ENT changes to the OSS repo.
Original commit details in ENT:
```
commit 569d25f7f4578981c3801e6e067295668210f748
Author: FFMMM <FFMMM@users.noreply.github.com>
Date: Thu Feb 10 10:23:33 2022 -0800
Vendor fork net rpc (#1538)
* replace net/rpc w consul-net-rpc/net/rpc
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
* replace msgpackrpc and go-msgpack with fork from mono repo
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
* gofmt all files touched
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
```
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
* First phase of refactoring PermissionDeniedError
Add extended type PermissionDeniedByACLError that captures information
about the accessor, particular permission type and the object and name
of the thing being checked.
It may be worth folding the test and error return into a single helper
function, that can happen at a later date.
Signed-off-by: Mark Anderson <manderson@hashicorp.com>
Due to timing, a transparent proxy could have two upstreams to dial
directly with the same address.
For example:
- The orders service can dial upstreams shipping and payment directly.
- An instance of shipping at address 10.0.0.1 is deregistered.
- Payments is scaled up and scheduled to have address 10.0.0.1.
- The orders service receives the event for the new payments instance
before seeing the deregistration for the shipping instance. At this
point two upstreams have the same passthrough address and Envoy will
reject the listener configuration.
To disambiguate this commit considers the Raft index when storing
passthrough addresses. In the example above, 10.0.0.1 would only be
associated with the newer payments service instance.
This commit makes two changes to the validation.
Previously we would call this validation in GenerateRoot, which happens
both on initialization (when a follower becomes leader), and when a
configuration is updated. We only want to do this validation during
config update so the logic was moved to the UpdateConfiguration
function.
Previously we would compare the config values against the actual cert.
This caused problems when the cert was created manually in Vault (not
created by Consul). Now we compare the new config against the previous
config. Using a already created CA cert should never error now.
Adding the key bit and types to the config should only error when
the previous values were not the defaults.
These two tests require debug logging enabled, because they look for log lines.
Also switched to testify assertions because the previous errors were not clear.
This test found a bug in the secondary. We were appending the root cert
to the PEM, but that cert was already appended. This was failing
validation in Vault here:
https://github.com/hashicorp/vault/blob/sdk/v0.3.0/sdk/helper/certutil/types.go#L329
Previously this worked because self signed certs have the same
SubjectKeyID and AuthorityKeyID. So having the same self-signed cert
repeated doesn't fail that check.
However with an intermediate that is not self-signed, those values are
different, and so we fail the check. A test I added in a previous commit
should show that this continues to work with self-signed root certs as
well.
This is safer than embedding two interface because there are a number of
places where we check the concrete type. If we check the concrete type
on the top-level interface it will fail. So instead expose the
ACLIdentity from a method.
This change allows us to remove one of the last remaining duplicate
resolve token methods (Server.ResolveToken).
With this change we are down to only 2, where the second one also
handles setting the default EnterpriseMeta from the token.
Now that ACLResolver is embedded we don't need ResolveTokenToIdentity on
Client and Server.
Moving ResolveTokenAndDefaultMeta to ACLResolver removes the duplicate
implementation.
set -euo pipefail
unset CDPATH
cd "$(dirname "$0")"
for f in $(git grep '\brequire := require\.New(' | cut -d':' -f1 | sort -u); do
echo "=== require: $f ==="
sed -i '/require := require.New(t)/d' $f
# require.XXX(blah) but not require.XXX(tblah) or require.XXX(rblah)
sed -i 's/\brequire\.\([a-zA-Z0-9_]*\)(\([^tr]\)/require.\1(t,\2/g' $f
# require.XXX(tblah) but not require.XXX(t, blah)
sed -i 's/\brequire\.\([a-zA-Z0-9_]*\)(\(t[^,]\)/require.\1(t,\2/g' $f
# require.XXX(rblah) but not require.XXX(r, blah)
sed -i 's/\brequire\.\([a-zA-Z0-9_]*\)(\(r[^,]\)/require.\1(t,\2/g' $f
gofmt -s -w $f
done
for f in $(git grep '\bassert := assert\.New(' | cut -d':' -f1 | sort -u); do
echo "=== assert: $f ==="
sed -i '/assert := assert.New(t)/d' $f
# assert.XXX(blah) but not assert.XXX(tblah) or assert.XXX(rblah)
sed -i 's/\bassert\.\([a-zA-Z0-9_]*\)(\([^tr]\)/assert.\1(t,\2/g' $f
# assert.XXX(tblah) but not assert.XXX(t, blah)
sed -i 's/\bassert\.\([a-zA-Z0-9_]*\)(\(t[^,]\)/assert.\1(t,\2/g' $f
# assert.XXX(rblah) but not assert.XXX(r, blah)
sed -i 's/\bassert\.\([a-zA-Z0-9_]*\)(\(r[^,]\)/assert.\1(t,\2/g' $f
gofmt -s -w $f
done
Remove some unnecessary comments around query_blocking metric. The only
line that needs any comments in the atomic decrement.
Cleanup the block and return comments and logic. The old comment about
AbandonCh may have been relevant before, but it is expected behaviour
now.
The logic was simplified by inverting the err condition.
This helps keep the logic in blockingQuery more focused. In the future we
may have a separate struct for RPC queries which may allow us to move this
off of Server.
This safeguard should be safe to apply in general. We are already
applying it to non-blocking queries that call blockingQuery, so it
should be fine to apply it to others.
To remove the TODO, and make it more readable.
In general this reduces the scope of variables, making them easier to reason about.
It also introduces more early returns so that we can see the flow from the structure
of the function.
`newIntermediate` is always equal to `needsNewIntermediate`, so we can
remove the extra variable and use the original directly.
Also remove the `activeRoot.ID != newActiveRoot.ID` case from an if,
because that case is already checked above, and `needsNewIntermediate` will
already be true in that case.
This condition now reads a lot better:
> Persist a new root if we did not have one before, or if generated a new intermediate.
In the previous commit the single use of this storedRoot was removed.
In this commit the original objective is completed. The
Provider.ActiveRoot is being removed because
1. the secondary should get the active root from the Consul primary DC,
not the provider, so that secondary DCs do not need to communicate
with a provider instance in a different DC.
2. so that the Provider.ActiveRoot interface can be changed without
impacting other code paths.
This method had only one caller, which always looked for the active
root. This commit moves the lookup into the method to reduce the logic
in the one caller.
This is being done in preparation for a larger change. Keeping this
separate so it is easier to see.
The `storedRootID != primaryRoots.ActiveRootID` is being removed because
these can never be different.
The `storedRootID` comes from `provider.ActiveRoot`, the
`primaryRoots.ActiveRootID` comes from the store `CARoot` from the
primary. In both cases the source of the data is the primary DC.
Technically they could be different if someone modified the provider
outside of Consul, but that would break many things, so is not a
supported flow.
If these were out of sync because of ordering of events then the
secondary will soon receive an update to `primaryRoots` and everything
will be sorted out again.
ActiveRoot should not be called from the secondary DC, because there
should not be a requirement to run the same Vault instance in a
secondary DC. SignIntermediate is called in a secondary DC, so it should
not call ActiveRoot
We would also like to change the interface of ActiveRoot so that we can
support using an intermediate cert as the primary CA in Consul. In
preparation for making that change I am reducing the number of calls to
ActiveRoot, so that there are fewer code paths to modify when the
interface changes.
This change required a change to the mockCAServerDelegate we use in
tests. It was returning the RootCert for SignIntermediate, but that is
not an accurate fake of production. In production this would also be a
separate cert.
Immediately above this line we are already appending the full list of
intermediates. The `provider.ActiveIntermediate` MUST be in this list of
intermediates because it must be available to all the other non-leader
Servers. If it was not in this list of intermediates then any proxy
that received data from a non-leader would have the wrong certs.
This is being removed now because we are planning on changing the
`Provider.ActiveIntermediate` interface, and removing these extra calls ahead of
time helps make that change easier.
Using tracing and cpu profiling I found that the majority of the time in
these test cases is spent generating a private key. We really don't need
separate private keys, so we can generate only one and use it for all
cases.
With this change the test runs much faster.
Fix the name to match the function it is testing
Remove unused code
Fix the signature, instead of returning (error, string) which should be (string, error)
accept a testing.T to emit errors.
Handle the error from encode.
The only function passed to SnapshotRPC today always returns a nil error, so there's no
way to exercise this bug in practice. This change is being made for correctness so that
it doesn't become a problem in the future, if we ever pass a different function to
SnapshotRPC.
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
Co-authored-by: Chris S. Kim <ckim@hashicorp.com>
Co-authored-by: Dan Upton <daniel@floppy.co>
This query has been incorrectly querying by accessor ID since New ACLs
were added. However, the legacy token compat allowed this to continue to
work, since it made a fallback query for the anonymousToken ID.
PR #11184 removed this legacy token query, which means that the query by
accessor ID is now the only check for the anonymous token's existence.
This PR updates the GetBySecret call to use the secret ID of the token.
These helper functions actually end up hiding important setup details
that should be visible from the test case. We already have a convenient
way of setting this config when calling newTestServerWithConfig.
I suspect one problem was that we set structs.IntermediateCertRenewInterval to 1ms, which meant
that in some cases the intermediate could renew before we stored the original value.
Another problem was that the 'wait for intermediate' loop was calling the provider.ActiveIntermediate,
but the comparison needs to use the RPC endpoint to accurately represent a user request. So
changing the 'wait for' to use the state store ensures we don't race.
Also moves the patching into a separate function.
Removes the addition of ca.CertificateTimeDriftBuffer as part of calculating halfTime. This was added
in a previous commit to attempt to fix the flake, but it did not appear to fix the problem. Adding the
time here was making the tests fail when using the shared patch
function. It's not clear to me why, but there's no reason we should be
including this time in the halfTime calculation.
Use the new verifyLearfCert to show the cert verifies with intermediates
from both sources. This required using the RPC interface so that the
leaf pem was constructed correctly.
Add IndexedCARoots.Active since that is a common operation we see in a
few places.
Previously we had a couple copies that reproduced the FSM operation.
These copies introduce risk that the test does not accurately match
production.
This PR removes the test versions of the FSM operation, and exports the
real production FSM operation so that it can be used in tests.
The consul provider tests did need to change because of this. Previously
we would return a hardcoded value of 2, but in production this value is
always incremented.
Failing over to a partition is more siimilar to failing over to another
datacenter than it is to failing over to a namespace. In a future
release we should update how localities for failover are specified. We
should be able to accept a list of localities which can include both
partition and datacenter.
* Add partition fields to targets like service route destinations
* Update validation to prevent cross-DC + cross-partition references
* Handle partitions when reading config entries for disco chain
* Encode partition in compiled targets
Given that we do not allow wildcard partitions in intentions, no one ixn
can override the DefaultAllow setting. Only the default ACL policy
applies across all partitions.
This table purposefully does not index by partition/namespace. It's a
global view into all service names.
This table is intended to replace the current serviceListTxn watch in
intentionTopologyTxn. For cross-partition transparent proxying we need
to be able to calculate upstreams from intentions in any partition. This
means that the existing serviceListTxn function is insufficient since
it's scoped to a partition.
Moving away from that function is also beneficial because it watches the
main "services" table, so watchers will wake up when any instance is
registered or deregistered.
* state: port KV and Tombstone tables to new pattern
* go fmt'ed
* handle wildcards for tombstones
* Fix graveyard ent vs oss
* fix oss compilation error
* add partition to tombstones and kv state store indexes
* refactor to use `indexWithEnterpriseIndexable`
* Apply suggestions from code review
Co-authored-by: Chris S. Kim <ckim@hashicorp.com>
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
* add `singleValueID` implementation assertions
* partition `tableSessions` table
* fix sessions to use UUID and fix prefix index
* fix oss build
* clean up unused functions
* fix oss compilation
* add a partition indexer for sessions
* Fix oss to not have partition index
* fix oss tests
* remove unused operations_ent.go and operations_oss.go func
* remove unused const
* convert `IndexID` of `session_checks` table
* convert `indexSession` of `session_checks` table
* convert `indexNodeCheck` of `session_checks` table
* partition `indexID` and `indexSession` of `tableSessionChecks`
* fix oss linter
* fix review comments
* remove partition for Checks as it's always use the session partition
* fix tests
* fix tests
* do not namespace nodeChecks index
Co-authored-by: Daniel Nephin <dnephin@hashicorp.com>
Co-authored-by: Chris S. Kim <ckim@hashicorp.com>
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
Cross port of ent #1383 "Reject non-default datacenter when making partitioned ACLs"
On the OSS side this is a minor refactor to add some more checks that are only applicable to enterprise code.
Signed-off-by: Mark Anderson <manderson@hashicorp.com>
The test added in this commit shows the problem. Previously the
SigningKeyID was set to the RootCert not the local leaf signing cert.
This same bug was fixed in two other places back in 2019, but this last one was
missed.
While fixing this bug I noticed I had the same few lines of code in 3
places, so I extracted a new function for them.
There would be 4 places, but currently the InitializeCA flow sets this
SigningKeyID in a different way, so I've left that alone for now.
While working on the CA system it is important to be able to run all the
tests related to the system, without having to wait for unrelated tests.
There are many slow and unrelated tests in agent/consul, so we need some
way to filter to only the relevant tests.
This PR renames all the CA system related tests to start with either
`TestCAMananger` for tests of internal operations that don't have RPC
endpoint, or `TestConnectCA` for tests of RPC endpoints. This allows us
to run all the test with:
go test -run 'TestCAMananger|TestConnectCA' ./agent/consul
The test naming follows an undocumented convention of naming tests as
follows:
Test[<struct name>_]<function name>[_<test case description>]
I tried to always keep Primary/Secondary at the end of the description,
and _Vault_ has to be in the middle because of our regex to run those
tests as a separate CI job.
You may notice some of the test names changed quite a bit. I did my best
to identify the underlying method being tested, but I may have been
slightly off in some cases.
As a method on the struct type this would not be safe to call without first checking
c.isIntermediateUsedToSignLeaf.
So for now, move this logic to the CAMananger, so that it is always correct.
We were not adding the local signing cert to the CARoot. This commit
fixes that bug, and also adds support for fixing existing CARoot on
upgrade.
Also update the tests for both primary and secondary to be more strict.
Check the SigningKeyID is correct after initialization and rotation.
Previously we believe it was necessary for all code that required ports
to use freeport to prevent conflicts.
https://github.com/dnephin/freeport-test shows that it is actually save
to use port 0 (`127.0.0.1:0`) as long as it is passed directly to
`net.Listen`, and the listener holds the port for as long as it is
needed.
This works because freeport explicitly avoids the ephemeral port range,
and port 0 always uses that range. As you can see from the test output
of https://github.com/dnephin/freeport-test, the two systems never use
overlapping ports.
This commit converts all uses of freeport that were being passed
directly to a net.Listen to use port 0 instead. This allows us to remove
a bit of wrapping we had around httptest, in a couple places.
In d2ab767fef21244e9fe3b9887ea70fc177912381 raftApply was changed to handle this check in
a single place, instad of having every caller check it. It looks like these few places
were missed when I did that clean up.
This commit removes the remaining resp.(error) checks, since they are all no-ops now.
This function is only ever called from operations that have already acquired the state lock, so checking
the value of state can never fail.
This change is being made in preparation for splitting out a separate type for the secondary logic. The
state can't easily be shared, so really only the expored top-level functions should acquire the 'state lock'.
R.B. Boyer