Improvements to out_of_band hook #2234
Conversation
This moves the hook to be executed outside of busy loop considering the thread to be truly idle at a time of execution of the hook
wjordan
force-pushed
the
improve-oob-hook
branch
from
3aa150f
to
58014f4
Compare
lib/puma/server.rb
Outdated
| end | ||
| end | ||
| private :accept_socket | ||
|
|
There was a problem hiding this comment.
Extracted this logic into a private method both to simplify the accept-loop, but more importantly to allow the to be stubbed in the test_blocks_new_connection unit-test.
There was a problem hiding this comment.
hmm, the original ECONNABORTED rescue clause had io.close not sock.close, I should change this back to the original behavior.
That said, the line is ancient, dating back to 2006 (ab3c808) (!) and never covered by any tests, so it's very likely this is 'superstitious' code that should probably be changed/removed separately anyway.
In my reading of the current code, ECONNABORTED would only ever be thrown by accept_nonblock (accept syscall) if the listener socket is closed accepts a connection that was closed while waiting on the listen queue, but in that case io would be nil causing io.close to always fail..
There was a problem hiding this comment.
I managed to figure out how to stub the accept call without needing to extract the method, so I'll move this refactoring/cleanup to a separate PR.
lib/puma/server.rb
Outdated
| io.close | ||
| rescue | ||
| Thread.current.purge_interrupt_queue if Thread.current.respond_to? :purge_interrupt_queue | ||
| pool.wait_until_not_full |
There was a problem hiding this comment.
Moving the call to wait_until_not_full before accept_socket prevents connections from being accepted until out_of_band hooks are completed.
(I'm not 100% sure this change won't have any other side-effects, but the test suite seems to think it's okay...)
An alternative would be to just add an extra pool.with_mutex {} line above accept_socket.
| @@ -224,8 +224,11 @@ def run(background=true) | |||
| @reactor.add client | |||
| end | |||
| end | |||
|
|
|||
| process_now | |||
There was a problem hiding this comment.
This return value of the ThreadPool block (process_now is true when a complete request was processed by the block) is used to determine if an OOB hook should be scheduled to fire once the thread is eventually idle.
wjordan
force-pushed
the
improve-oob-hook
branch
from
58014f4
to
bad2be5
Compare
* Don't trigger OOB on partially-queued requests * Don't accept new connections during OOB * Move tests to `TestOutOfBandServer`
wjordan
force-pushed
the
improve-oob-hook
branch
from
bad2be5
to
d416d0c
Compare
| end | ||
| app_wait = options.delete(:app_wait) | ||
| app = ->(_) do | ||
| raise 'OOB conflict' if in_oob.locked? |
There was a problem hiding this comment.
Nice, I like that approach.
|
Tests are 😍 |
Description
Followup to #2230 (extending the idea to move the OOB hook from
ServerintoThreadPool), addressing the two extra edge cases described at #2230 (comment).Also:
TestPumaServerclass into its ownTestOutOfBandServerclass(Turns out there are a ton of complicated little edge-cases to consider around what seemed like such a simple feature!)
Your checklist for this pull request
[changelog skip]the pull request title.[ci skip]to the title of the PR.#issue" to the PR description or my commit messages.