Running woken runners in the next iteration of the event loop prevents
unexpected recursion, unexpected tracebacks, and is generally more
predictable.
The pattern is borrowed from util.promise, where we're now doing the same.
No scheduler set by default, so it will error (we plan to initialize it in
util.startup).
We wanted to avoid a hard dependency on util.timer (which in turn depends on
network backends, etc.), and we didn't add timer.sleep() because we didn't
want to add a hard dependency on util.async for things that don't need it.
Could not reproduce locally but it complained in CI that
> spec/util_promise_spec.lua:676: Cannot spy on type 'nil', only on functions or callable elements
Makes GET /metrics work regardless of the value of the Host header.
This was not possible before 220468f7a103 which added the ability to
have global HTTP modules, which was partly why the per-host part was
kept while the module lived among the community modules.
Another reason was to provide some semblance of security by obscurity by
only loading the module on e.g. a 'localhost' VirtualHost or Component.
The new IP limits obsoletes this.
Taking advantage of the new callbacks added in dcf38ac6a38c and
9c450185bac1 avoids extra timers, extra syscalls and sending the `<r>`
in its own TCP segment, improving efficiency.
The function was too large to comprehend! Breaking it up helps
readability and reuse.
The timer round rip is only to avoid ordering weirdness when sending
from inside a stanza filter. No need when handling <r> and <a>
CSI interactions both boiled down to sending an <r> immediately.
module:add_timer() creates a timer that stops working after the module
is reloaded or unloaded, in order to prevent leaks. However, when the
timers control vital session behavior, breakage occurs after reload.
E.g. sessions would stop requesting acks and stop responding to acks.
Unstoppable stoppable timer compat not needed since 26f54b462601 ca 0.11.0
module:hook_stanza was renamed in 2012
No idea what was going on with the indentation and such
These are RECOMMENDED in XEP-0215 so most likely a mistake if they are
left out.
Of the two REQUIRED fields, 'host' falls back to module.host and 'type'
rejects the whole item.
This version has several changes from the earlier mod_prometheus:
- Conversion of metrics into the text-based OpenMetrics format is moved to
util.openmetrics
- Support for IP-based access control
- Compatibility with earlier Prosody versions removed
Both in the flushing and active states the right thing to do is skip
directly to returning the data. Also in any unknown state, like if the
filter is somehow left behind on module unload.
Statistics from my server shows a high rate of very short buffer hold
times, most of which are the result of replies to pings or other iq
traffic, or mod_smacks acks and ack requests just after a flush was
completed.
This grace period should eliminate noise and quick flipping between
flushing and inactive mode.
Previously it would complain about lack of an AAAA record for
proxy65_target even in an IPv6-less environment.
Thanks to libertas for unintentionally calling attention to this.
ff4e34c448a4 broke the way net.http.server streams downloads from disk
because it made writes from the ondrain callback no longer reset the
want-write flag, causing the download to halt.
Writes from the predrain handler still must not trigger anything but
additions to the buffer, since it is about to do all the socket writing
already.
Since it changes the state to "flushing" while doing just that.
Attempting to remove the filters from a session that does not have them
should be a safe noop.
E.g. mod_smacks could queue an <r>, which would be more likely to be
included in the same write and TCP segment as the previously buffered
data, reducing syscalls and network packets needing to be sent.
Matthew Wild