Regular Client doesn't care much about connections, because HTTP client's Do
method can reuse old ones or create additional ones on the fly. So one request
can fail and the next one easily succeed. WSClient is different, it works via
a single connection and if it breaks, it breaks forever for this
client. Callers will get some error on every request afterwards and it'd be
nice for this error to be the same so that API users could detect
disconnection this way too.
Related to nspcc-dev/neofs-node#2325.
Signed-off-by: Roman Khimov <roman@nspcc.ru>
It should be this way because we can't provide some features golangci-lint
require us to do due to the need to support not only the latest Go version,
but also two versions below.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
go.uber.org/atomic deprecated CAS methods in version 1.10 (that introduced
CompareAndSwap), so we need to fix it.
Signed-off-by: Roman Khimov <roman@nspcc.ru>
We were using _very_ old version, so there is a number of changes, including:
* fix for unlimited reads
* UTF-8 check for HTTP requests
Signed-off-by: Roman Khimov <roman@nspcc.ru>
If the contract was deployed then cache must be initialized after
in-memory data reset. If the contract isn't active yet, then no
cache will be initialized on deploy (i.e. on call to Initialize()
method by native Management).
Close#2984.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
Make the contracts cache initialization unified. The order of cache
iniitialization is not important and Nottary contract is added to the
bc.contracts.Contracts wrt P2PSigExtensions setting, thus no functional
changes, just refactoring for future applications.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
Initialize Prometheus metrics on node start where appropriate and review
the usage of the following metrics:
```
anna@kiwi:~/Documents/GitProjects/nspcc-dev/neo-go$ find | grep prometheus.go
./pkg/network/prometheus.go
./pkg/core/stateroot/prometheus.go
./pkg/core/prometheus.go
./pkg/services/rpcsrv/prometheus.go
./pkg/services/metrics/prometheus.go
```
Close#2970.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
Close#2894.
It should be noted that the subscriber's channel is being removed from the
list of receivers and closed, but it is still *in the list of subscribers*
and no unsubscription is performed by WSClient. Which means that RPC server
keeps sending notifications to WSClient and WSClient keeps dropping them
(because there's no receiver for this subscription and it's OK, WSClient
can handle this and this behaviour is documented). However, it's still the
caller's duty to call Unsubscribe() method for this subscription.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
All Notary contract witnesses in incomplete transaction (both main
and fallback) may either have invocation scripts pushing dummy signature
on stack or be empty, both ways are OK. Notary actor keeps main tx's
Notary witness empty and keeps fallback tx's Notary witness filled
with dummy signature.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
The reader is about to exit and it will close legacy c.Notifications, but it
will leave subscription channels at the same time. This is wrong since these
channels will no longer receive any new events, game over.
Signed-off-by: Roman Khimov <roman@nspcc.ru>
wsReader() closes c.done first and then goes over the list of
c.respChannels. Technically this means that any of the two can be taken in
this select.
Signed-off-by: Roman Khimov <roman@nspcc.ru>
If `StartWhenSynchronized` is unset in config, `node` command runs RPC
service instantly. Previously there was a ground for deadlock. Command
started RPC server synchronously. According to server implementation, it
sends all internal failures to the parameterized error channel. Deadlock
occured because main routine didn't scan the channel.
Run `rpcsrv.Server.Start` in a separate go-routine in `startServer`.
This prevents potential deadlock caused by writing into unread channel.
Fixes#2896.
Signed-off-by: Leonard Lyubich <leonard@morphbits.io>
Previously RPC server shutdown procedure listened to the execution
channel and stopped at the first element that arrived in the queue. This
could lead to the following problems:
* stopper could steal the execution result from subscriber
* stopper didn't wait for other subscription actions to complete
Add dedicated channel to `Server` for subscription routine. Close the
channel on `handleSubEvents` return and wait for signal in `Shutdown`.
Signed-off-by: Leonard Lyubich <leonard@morphbits.io>
Previously RPC server could never be shut down completely due to
some start precondition failure (in particular, inability to serve HTTP
on any configured endpoint). The problem was caused by next facts:
* start method ran subscription routine after HTTP init succeeded only
* stop method blocked waiting for the subscription routine to return
Run `handleSubEvents` routine on fresh `Start` unconditionally. With
this change, `Shutdown` method won't produce deadlock since
`handleSubEvents` closes wait channel.
Refs #2896.
Signed-off-by: Leonard Lyubich <leonard@morphbits.io>
There is an existing problem with RPC server shutdown freeze after start
failure due to some init actions (at least HTTP listen) described in
#2896.
Add dedicated unit test which checks that `Shutdown` returns within 5s
after `Start` method encounters internal problems.
Signed-off-by: Leonard Lyubich <leonard@morphbits.io>
Move them to the core/network packages, close#2950. The name of
mempool's unsorted transactions metrics has been changed along the
way to match the core's metrics naming convention.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
I've carefully checked the way how new service can be added to the
Blockchain instance or to be removed from it. Current implemention
of SetNotary and SetOracle methods doesn't contain dangerous code,
and native contracts have atomic values everywhere where service
is stored.
Current implementation of Notary, Oracle and StateRoot services'
reload/disabling/enabling on SIGUSR1 is safe and doesn't require
any adjustment.
This commit closes#2944, it's not a bug in the code, it's just
stale documentation.
Signed-off-by: Anna Shaleva <shaleva.ann@nspcc.ru>
