This PR removes the complicated session management introduced in https://github.com/juanfont/headscale/pull/1791 which kept track of the sessions in a map, in addition to the channel already kept track of in the notifier.
Instead of trying to close the mapsession, it will now be replaced by the new one and closed after so all new updates goes to the right place.
The map session serve function is also split into a streaming and a non-streaming version for better readability.
RemoveNode in the notifier will not remove a node if the channel is not matching the one that has been passed (e.g. it has been replaced with a new one).
A new tuning parameter has been added to added to set timeout before the notifier gives up to send an update to a node.
Add a keep alive resetter so we wait with sending keep alives if a node has just received an update.
In addition it adds a bunch of env debug flags that can be set:
- `HEADSCALE_DEBUG_HIGH_CARDINALITY_METRICS`: make certain metrics include per node.id, not recommended to use in prod.
- `HEADSCALE_DEBUG_PROFILING_ENABLED`: activate tracing
- `HEADSCALE_DEBUG_PROFILING_PATH`: where to store traces
- `HEADSCALE_DEBUG_DUMP_CONFIG`: calls `spew.Dump` on the config object startup
- `HEADSCALE_DEBUG_DEADLOCK`: enable go-deadlock to dump goroutines if it looks like a deadlock has occured, enabled in integration tests.
Signed-off-by: Kristoffer Dalby <kristoffer@tailscale.com>
This commit restructures the map session in to a struct
holding the state of what is needed during its lifetime.
For streaming sessions, the event loop is structured a
bit differently not hammering the clients with updates
but rather batching them over a short, configurable time
which should significantly improve cpu usage, and potentially
flakyness.
The use of Patch updates has been dialed back a little as
it does not look like its a 100% ready for prime time. Nodes
are now updated with full changes, except for a few things
like online status.
Signed-off-by: Kristoffer Dalby <kristoffer@tailscale.com>
This commits removes the locks used to guard data integrity for the
database and replaces them with Transactions, turns out that SQL had
a way to deal with this all along.
This reduces the complexity we had with multiple locks that might stack
or recurse (database, nofitifer, mapper). All notifications and state
updates are now triggered _after_ a database change.
Signed-off-by: Kristoffer Dalby <kristoffer@tailscale.com>
This commits adds a test to verify that nodes get updated if a node in
their network expires.
Signed-off-by: Kristoffer Dalby <kristoffer@tailscale.com>
This commit makes the initial SSH test a bit simpler:
- Use the same pattern/functions for all clients as other tests
- Only test within _one_ namespace/user to confirm the base case
- Use retry function, same as taildrop, there is some funky going on
there...
Signed-off-by: Kristoffer Dalby <kristoffer@tailscale.com>
