[PATCH] Fix queue stats for sockets with SO_REUSEPORT

[PATCH] Fix queue stats for sockets with SO_REUSEPORT

[Dale Hamel]

SO_REUSEPORT was introduced in 2013, see https://lwn.net/Articles/542629/.

However, the current logic for calculating socket queue backlog stats
is from 2011, predating the advent of SO_REUSEPORT.

The current strategy was thus written before the notion that the mapping of
INET_ADDR:socket could potentially be 1:N instead of always 1:1.

This causes raindrops to provide invalid socket backlog values when
SO_REUSEPORT is used, which Unicorn supports since 2013. The current behaviour
will set the queue metric to the queue depth of the last inet_diag_msg
it processes matching the INET_ADDR. In practice, this will result in
the backlog being off by a factor N, assuming relatively even
distribution of requests to sockets, which the kernel should guarantee
through consistent hashing.

The fix here is to accumulate the socket queue depth as we iterate
over the the socket diagnostics, rather than reset it each time.

I have a provided a test, but it only checks the queues, not the
accept metrics, as those are not affected by this bug, and it is
not possible to know which of the listeners will be dispatched the
request by the kernel, and thus which should call accept.
---
 ext/raindrops/linux_inet_diag.c               |  4 +-
 test/test_linux_reuseport_tcp_listen_stats.rb | 51 +++++++++++++++++++
 2 files changed, 53 insertions(+), 2 deletions(-)
 create mode 100644 test/test_linux_reuseport_tcp_listen_stats.rb

diff --git a/ext/raindrops/linux_inet_diag.c b/ext/raindrops/linux_inet_diag.c
index cabd427..2a2360c 100644
--- a/ext/raindrops/linux_inet_diag.c
+++ b/ext/raindrops/linux_inet_diag.c
@@ -306,7 +306,7 @@ static void table_set_queued(st_table *table, struct inet_diag_msg *r)
 {
 	struct listen_stats *stats = stats_for(table, r);
 	stats->listener_p = 1;
-	stats->queued = r->idiag_rqueue;
+	stats->queued += r->idiag_rqueue;
 }
 
 /* inner loop of inet_diag, called for every socket returned by netlink */
@@ -328,7 +328,7 @@ static inline void r_acc(struct nogvl_args *args, struct inet_diag_msg *r)
 		if (args->table)
 			table_set_queued(args->table, r);
 		else
-			args->stats.queued = r->idiag_rqueue;
+			args->stats.queued += r->idiag_rqueue;
 	}
 	/*
 	 * we wont get anything else because of the idiag_states filter
diff --git a/test/test_linux_reuseport_tcp_listen_stats.rb b/test/test_linux_reuseport_tcp_listen_stats.rb
new file mode 100644
index 0000000..c977c43
--- /dev/null
+++ b/test/test_linux_reuseport_tcp_listen_stats.rb
@@ -0,0 +1,51 @@
+# -*- encoding: binary -*-
+require "./test/rack_unicorn"
+require 'test/unit'
+require 'socket'
+require 'raindrops'
+$stderr.sync = $stdout.sync = true
+
+class TestLinuxReuseportTcpListenStats < Test::Unit::TestCase
+  include Raindrops::Linux
+  include Unicorn::SocketHelper
+  TEST_ADDR = ENV['UNICORN_TEST_ADDR'] || '127.0.0.1'
+  DEFAULT_BACKLOG = 10
+
+  def setup
+    @socks = []
+  end
+
+  def teardown
+    @socks.each { |io| io.closed? or io.close }
+  end
+
+  def new_socket_server(**kwargs)
+    s = new_tcp_server TEST_ADDR, kwargs[:port] || 0, kwargs
+    s.listen(kwargs[:backlog] || DEFAULT_BACKLOG)
+    @socks << s
+    [ s, s.addr[1] ]
+  end
+
+  def new_client(port)
+    s = TCPSocket.new("127.0.0.1", port)
+    @socks << s
+    s
+  end
+
+  def test_reuseport_queue_stats
+    listeners = 10
+    _, port = new_socket_server(reuseport: true)
+    addr = "#{TEST_ADDR}:#{port}"
+    listeners.times do
+      new_socket_server(reuseport: true, port: port)
+    end
+
+    listeners.times do |i|
+      all = Raindrops::Linux.tcp_listener_stats
+      assert_equal [0, i], all[addr].to_a
+      new_client(port)
+      all = Raindrops::Linux.tcp_listener_stats
+      assert_equal [0, i+1], all[addr].to_a
+    end
+  end
+end if RUBY_PLATFORM =~ /linux/
-- 
2.34.1

[PATCH] Fix off by one error in test

[Dale Hamel]

It doesn't affect the test outcome, but it is not the intended behaviour of
the test.

We listen once to find out what port to use, then should bind with SO_REUSEPORT
N-1 times, but we are doing it N times, so we have one extra listener.
---
 test/test_linux_reuseport_tcp_listen_stats.rb | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/test_linux_reuseport_tcp_listen_stats.rb b/test/test_linux_reuseport_tcp_listen_stats.rb
index c977c43..73995ea 100644
--- a/test/test_linux_reuseport_tcp_listen_stats.rb
+++ b/test/test_linux_reuseport_tcp_listen_stats.rb
@@ -36,7 +36,7 @@ def test_reuseport_queue_stats
     listeners = 10
     _, port = new_socket_server(reuseport: true)
     addr = "#{TEST_ADDR}:#{port}"
-    listeners.times do
+    (listeners - 1).times do
       new_socket_server(reuseport: true, port: port)
     end
 
-- 
2.39.2

Re: [PATCH] Fix queue stats for sockets with SO_REUSEPORT

[Eric Wong]

Thanks, pushed to raindrops.git:

10d5fb25b631 Fix off by one error in test
a5902f243cd2 Fix queue stats for sockets with SO_REUSEPORT

I'm honestly shocked that these features get used;
I never convinced the org I worked for to use them :x

Anything else?  Could tag a release in a bit...

Re: [PATCH] Fix queue stats for sockets with SO_REUSEPORT

[Dale Hamel]

> I'm honestly shocked that these features get used;
> I never convinced the org I worked for to use them :x

So far we haven't actually found a production use case
for SO_REUSEPORT with Unicorn. We thought we had some
promising early results - but it turned out to be this
instrumentation error in raindrops.

At least by fixing this, efforts in the future will have the correct
metrics to base tests off of.

> Anything else?  Could tag a release in a bit...

No, that's it! My apologies for sending as two patches,
I didn't notice the bug in my test until I had already
submitted, naturally.

Thanks very much for the quick review!

Re: [PATCH] Fix queue stats for sockets with SO_REUSEPORT

[Eric Wong]

Dale Hamel <dale.hamel@shopify.com> wrote:
> > I'm honestly shocked that these features get used;
> > I never convinced the org I worked for to use them :x
> 
> So far we haven't actually found a production use case
> for SO_REUSEPORT with Unicorn. We thought we had some
> promising early results - but it turned out to be this
> instrumentation error in raindrops.

Oh, I meant the socket instrumentation part of raindrops.

I figured atomic counters were the only thing that ever got used.

Anyways I think SO_REUSEPORT was too risky due to the
possibility of connection drops, but then again I haven't used
more than 4 real cores.

> At least by fixing this, efforts in the future will have the correct
> metrics to base tests off of.

Cool.

> > Anything else?  Could tag a release in a bit...
> 
> No, that's it! My apologies for sending as two patches,
> I didn't notice the bug in my test until I had already
> submitted, naturally.

No worries.  In the future, you can also ask me to do a squash
in the diffstat area or signature area, or use [SQUASH] instead of
[PATCH] in the subject prefix.

> Thanks very much for the quick review!

np.  Easy for me as long as I never leave a terminal :>