Previously, the transport was able to reset via the retry loop,
or via the event closures calling resetTransport. This meant
a very large amount of synchronization was necessary: one
reset meant the other had to not reset; state had to be kept
at the addrconn; and very subtle interactions were hard to
reason about.
This change removes the ability for event closures to directly
reset the transport. Instead, they signal to to the retry
loop about the event, and the retry loop is always the single
place that retries occur.
This also allows us to refactor the address switching logic
into a much simpler for loop inside the retry loop instead of
using addrConn state to keep track of an index.
Possible settings of this environment variable:
- "hybrid" (default; removed after the 1.17 release): do not wait for handshake before considering a connection ready, but wait before considering successful.
- "on" (default after the 1.17 release): wait for handshake before considering a connection ready/successful.
- "off": do not wait for handshake before considering a connection ready/successful.
This setting will be completely removed after the 1.18 release, and "on" will be the only supported behavior.
This allows the initial RPC(s) an opportunity to apply settings from the service config; without this change we would still block, but only after observing the current service config settings.
internal: fix GO_AWAY deadlock
A deadlock can occur when a GO_AWAY is followed by a connection closure. This
happens because onClose needlessly closes the current ac.transport: if a
GO_AWAY already occured, and the transport was already reset, then the later
closure (of the original address) sets ac.transport - which is now healthy -
to nil.
The manifestation of this problem is that picker_wrapper spins forever trying
to use a READY connection whose ac.transport is nil.
internal: fix client send preface problems
This CL fixes three problems:
- In clientconn_state_transitions_test.go, sometimes tests would flake because there's not enough buffer to send client side settings, causing the connection to unpredictably enter TRANSIENT FAILURE. Each time we set up a server to send SETTINGS, we should also set up the server to read. This allows the client to successfully send its SETTINGS, unflaking the test.
- In clientconn.go, we incorrectly transitioned into TRANSIENT FAILURE when creating an http2client returned an error. This should be handled in the outer resetTransport main reset loop. The reason this became a problem is that the outer resetTransport has very specific conditions around when to transition into TRANSIENT FAILURE that the egregious transition did not have. So, it could transition into TRANSIENT FAILURE after failing to dial, even if it was trying to connect to a non-final address in the list of addresses.
- In clientconn.go, we incorrectly stay in CONNECTING after `createTransport` when a server sends its connection preface but the client is not able to send its connection preface. This CL causes the addrconn to correctly enter TRANSIENT FAILURE when `createTransport` fails, even if a server preface was received. It does so by making ac.successfulHandshake to consider both server preface received as well as client preface sent.
Closing `ClientConn` sets `balancerWrapper` to nil.
If service config switches balancer, the new balancer will be notified of the existing addresses.
When these two happens together, there's a chance that a method will be called on the nil `balancerWrapper`. This change adds a check to make sure that never happens.
fixes#2367
internal: fix onClose state transitions
When onClose occurs during WaitForHandshake, it should immediately
exit createTransport instead of leaving CONNECTING and entering READY.
Furthermore, when any onClose happens, the state should change to
TRANSIENT FAILURE.
Fixes#2340Fixes#2341
Also fixes an unreported bug in which entering READY causes a
Dial call to end prematurely, instead of blocking until a READY
transport is found.
Google default creds is a combo of ALTS, TLS and OAuth2. The right set of creds will be picked to use based on environment.
This PR contains:
- A new `creds.Bundle` type
- changes to use it in ClientConn and transport
- dial option to set the bundle for a ClientConn
- balancer options and NewSubConnOption to set it for SubConn
- Google default creds implementation by @cesarghali
- grpclb changes to use different creds mode for different servers
- interop client changes for google default creds testing
This fixes a race in ac.tearDown and ac.resetTransport. If ac.resetTransport
is in backoff when ac.tearDown occurs, there's a race between the state
changing to Shutdown and ac.resetTransport calling ac.createTransport.
This fixes it by returning when ac.resetTransport encounters an error
during ac.nextAddr (specifically ac.ctx.Error()). It also fixes it by
making sure that ac.tearDown changes state to Shutdown before canceling
the context.
Both fixes were implemented because they both seem to be valuable
standalone additions: the former makes ac.resetTransport more
understandable and less dependent on behavior happening elsewhere,
and the latter makes ac.tearDown more correct.
Finally, TestDialParseTargetUnknownScheme had its buffer removed; the
buffer was likely added a while ago to assuage this issue. It should
not be necessary anymore.
internal: remove transportMonitor, replace with callbacks
This refactors the internal http2 transport to use callbacks instead
of continuously monitoring the transport in a separate goroutine. This
has several advantages:
- Less goroutines.
- Less complexity: synchronous callbacks are much easier to reason to
reason about than asynchronous monitoring goroutines.
- Callbacks: these provide definitive locations for monitoring the
creation and closure of a transport, paving the way for GracefulStop.
This CL also consolidates all the logic about backoff and iterating
through the list of addresses into a single method.
This is a breaking change, but the transport package was never intended for use outside of grpc. Any current users that we are aware of are incorrect or have a preferred alternative.
This PR splits out grpclb from grpc. I have made the PR in several commits so you can see more clearly the steps that happened.
There are a few possibly contentious points that I would like to make clear up front:
* grpclb will no longer autoload as a load balancer. I think this is okay, as service config is not widely (at all?) used, and I believe this is the only way to access it.
* `internal` is used more, as a way of having code shared between packages without exposing types
* ConnectivityStateEvaluator, as used by grpclb, is no longer thread safe. I believe there is an outer mutex that guards access, but I want to point out this subtle change up here.
All but one tests pass with this, due to another cyclic dependency. I can fix this, but it is a little more widely scoped (such as exposing grpc.server and grpc.errorDesc in the internal package). This PR is a nearly-passing sample of that last step to get this working.
PTAL @menghanl @dfawley
