* add $upstream_cache_status
* add '-$connection_requests' after $connection
* enable subrequest logging
$connection_requests makes it much easier to see connection reuse in the
logs and also helps to understand subrequests.
This switches to a fully custom log format instead of using a variant of
the standard combined format since we don't use any tools requiring the
logs to be a standard format. This provides a cleaner format, allows us
to freely add new fields and gets rid of legacy/redundant fields.
The redundant timestamp already provided as the syslog timestamp is
dropped along with the legacy identd field always set to a dash.
This adds the connection serial number for identifying requests coming
from the same connection. TLS version is added as a replacement for our
previous addition of the URI scheme. This also adds the total request
length and total bytes sent to the client instead of only the body bytes
sent.
There's a remaining issue fixed in mainline that's not fixed in the
current stable branch yet, but it doesn't apply unless HTTP/2 is being
used without encryption. Currently sendfile is only really used for the
backend proxy connections in practice due to TLS, and those are never
HTTP/2.
This is useless for TLSv1.3 since there's no longer any distinction in
the protocol based on whether the server is using stateless or stateful
session resumption. OpenSSL has a non-standard anti-replay mechanism for
0-RTT based on stateful session resumption but 0-RTT still ends up being
a downgrade for the TLS security properties. nginx disables that feature
since otherwise 0-RTT wouldn't work with the default stateless approach.
Since this cache is only used for TLSv1.2 when stateless resumption
isn't disabled and nearly all TLSv1.2 clients support tickets, it isn't
getting any significant use. It provides worse forward secrecy than
tickets because we implement ticket key rotation based on the expiry
time and sessions aren't actively purged from the stateful cache when
they expire. Cached session state varies in size and nginx ends up
writing errors to the log when clearing out a session fails to make room
for a new one due to it being larger. It's best to finally get rid of
this flawed approach to session resumption.
TLSv1.3 provides the option of forward secrecy for resumed sessions and
it's the only approach that's normally enabled so we don't need to worry
about this anymore once TLSv1.2 is disabled as long as we never enable
0-RTT which weakens forward secrecy and other security properties.
