* plumbing request context to expiration manager
* moar context
* address feedback
* only using active context for revoke prefix
* using active context for revoke commands
* cancel tidy on active context
* address feedback
* core: Cancel context before taking state lock
* Create active context outside of postUnseal
* Attempt to drain requests before canceling context
* fix test
* Add request timeouts in normal request path and to expirations
* Add ability to adjust default max request duration
* Some test fixes
* Ensure tests have defaults set for max request duration
* Add context cancel checking to inmem/file
* Fix tests
* Fix tests
* Set default max request duration to basically infinity for this release for BC
* Address feedback
* Tackle #4929 a different way
This turns c.sealed into an atomic, which allows us to call sealInternal
without a lock. By doing so we can better control lock grabbing when a
condition causing the standby loop to get out of active happens. This
encapsulates that logic into two distinct pieces (although they could
be combined into one), and makes lock guarding more understandable.
* Re-add context canceling to the non-HA version of sealInternal
* Return explicitly after stopCh triggered
* This changes the way policies are reported in audit logs.
Previously, only policies tied to tokens would be reported. This could
make it difficult to perform after-the-fact analysis based on both the
initial response entry and further requests. Now, the full set of
applicable policies from both the token and any derived policies from
Identity are reported.
To keep things consistent, token authentications now also return the
full set of policies in api.Secret.Auth responses, so this both makes it
easier for users to understand their actual full set, and it matches
what the audit logs now report.
* Remove a lot of deferred functions in the request path.
There is an interesting benchmark at https://www.reddit.com/r/golang/comments/3h21nk/simple_micro_benchmark_to_measure_the_overhead_of/
It shows that defer actually adds quite a lot of overhead -- maybe 100ns
per call but we defer a *lot* of functions in the request path. So this
removes some of the ones in request handling, ha, barrier, router, and
physical cache.
One meta-note: nearly every metrics function is in a defer which means
every metrics call we add could add a non-trivial amount of time, e.g.
for every 10 extra metrics statements we add 1ms to a request. I don't
know how to solve this right now without doing what I did in some of
these cases and putting that call into a simple function call that then
goes before each return.
* Simplify barrier defer cleanup
Taking inspiration from
https://github.com/golang/go/issues/17604#issuecomment-256384471
suggests that taking the address of a stack variable for use in atomics
works (at least, the race detector doesn't complain) but is doing it
wrong.
The only other change is a change in Leader() detecting if HA is enabled
to fast-path out. This value never changes after NewCore, so we don't
need to grab the read lock to check it.
