This change is backported from 092f2e659c.
Previously, the sharded permissions cache was only accessible to the
implementation of `auth::service` in `auth/service.cc`. The intention
was that invoking `auth::service::get_permissions` on shard `k` would
query the cache on shard `k`, which would in turn depend on
`auth::service` on shard k to check for superuser status.
The problem is in `auth::service::start`.
`seastar::sharded<auth::permissions_cache>::start` is invoked with
`*this` of shard 0, causing all instances of the cache to reference the
same object.
I wasn't able to locally reproduce errors or crashes due to this bug
when I compiled a release build of Scylla. However, running a debug
build meant that the glorious `seastar::debug_shared_ptr_counter_type`
quickly saved the day with its checks that `seastar::shared_ptr` isn't
being misused.
To eliminate this problem, we move ownership of a single instance of
`auth::permissions_cache` to a single instance of `auth::service`. When
`auth::service` is sharded, so is the permissions cache.
I verified interactively that no assertions failed in debug mode with
this change.
Fixes#3296.
Tests: unit (debug, release)
Signed-off-by: Jesse Haber-Kucharsky <jhaberku@scylladb.com>
Message-Id: <280a889f551180db1c00d8a80eddf85b2ff0ac60.1521696176.git.jhaberku@scylladb.com>
c5718bf620