[PATCH 6/6] use EPOLLEXCLUSIVE on Linux 4.5+

[Eric Wong <bofh@yhbt.net>]

While the capabilities of epoll cannot be fully exploited given
our primitive design; avoiding thundering herd wakeups on larger
SMP machines while below 100% utilization is possible with
Linux 4.5+.

With this change, only one worker wakes up per-connect(2)
(instead of all of them via select(2)), avoiding the thundering
herd effect when the system is mostly idle.

Saturated instances should not notice the difference if they
rarely had multiple workers sleeping in select(2).  This change
benefits non-saturated instances.

With 2 parallel clients and 8 workers on a nominally (:P)
8-core CPU (AMD FX-8320), the uconnect.perl test script
invocation showed a reduction from ~3.4s to ~2.5s when
reading an 11-byte response body:

  echo worker_processes 8 >u.conf.rb
  bs=11 ruby -I lib -I test/ruby-2.5.5/ext/unicorn_http/ bin/unicorn \
    test/benchmark/dd.ru -E none -l /tmp/u.sock -c u.conf.rb
  time perl -I lib -w test/benchmark/uconnect.perl \
    -n 100000 -c 2 /tmp/u.sock

Times improve less as "-c" increases for uconnect.perl (system
noise and timings are inconsistent).  The benefit of this change
should be more noticeable on systems with more workers (and
more cores).

I wanted to use EPOLLET (Edge-Triggered) to further reduce
syscalls, here, (similar to the old select()-avoidance bet) but
that would've either added too much complexity to deduplicate
wakeup sources, or run into the same starvation problem we
solved in April 2020[1].

Since the kernel already has the complexity and deduplication
built-in for Level-Triggered epoll support, we'll just let the
kernel deal with it.

Note: do NOT take this as an example of how epoll should be used
in a sophisticated server.  unicorn is primitive by design and
cannot use threads nor handle multiple clients at once, thus it
it only uses epoll in this extremely limited manner.

Linux 4.5+ users will notice a regression of one extra epoll FD
per-worker and at least two epoll watches, so
/proc/sys/fs/epoll/max_user_watches may need to be changed along
with RLIMIT_NOFILE.

This change has also been tested on Linux 3.10.x (CentOS 7.x)
and FreeBSD 11.x to ensure compatibility with systems without
EPOLLEXCLUSIVE.

Various EPOLLEXCLUSIVE discussions over the years:
  https://yhbt.net/lore/lkml/?q=s:EPOLLEXCLUSIVE+d:..20211001&x=t&o=-1

[1] https://yhbt.net/unicorn-public/CAMBWrQ=Yh42MPtzJCEO7XryVknDNetRMuA87irWfqVuLdJmiBQ@mail.gmail.com/
---
 ext/unicorn_http/epollexclusive.h | 125 ++++++++++++++++++++++++++++++
 ext/unicorn_http/extconf.rb       |   1 +
 ext/unicorn_http/unicorn_http.rl  |   3 +
 lib/unicorn/http_server.rb        |  17 +++-
 lib/unicorn/select_waiter.rb      |   6 ++
 t/test-lib.sh                     |   3 +-
 test/unit/test_waiter.rb          |  34 ++++++++
 7 files changed, 184 insertions(+), 5 deletions(-)
 create mode 100644 ext/unicorn_http/epollexclusive.h
 create mode 100644 lib/unicorn/select_waiter.rb
 create mode 100644 test/unit/test_waiter.rb

diff --git a/ext/unicorn_http/epollexclusive.h b/ext/unicorn_http/epollexclusive.h
new file mode 100644
index 0000000..2d2a589
--- /dev/null
+++ b/ext/unicorn_http/epollexclusive.h
@@ -0,0 +1,125 @@
+/*
+ * This is only intended for use inside a unicorn worker, nowhere else.
+ * EPOLLEXCLUSIVE somewhat mitigates the thundering herd problem for
+ * mostly idle processes since we can't use blocking accept4.
+ * This is NOT intended for use with multi-threaded servers, nor
+ * single-threaded multi-client ("C10K") servers or anything advanced
+ * like that.  This use of epoll is only appropriate for a primitive,
+ * single-client, single-threaded servers like unicorn that need to
+ * support SIGKILL timeouts and parent death detection.
+ */
+#if defined(HAVE_EPOLL_CREATE1)
+#  include <sys/epoll.h>
+#  include <errno.h>
+#  include <ruby/io.h>
+#  include <ruby/thread.h>
+#endif /* __linux__ */
+
+#if defined(EPOLLEXCLUSIVE) && defined(HAVE_EPOLL_CREATE1)
+#  define USE_EPOLL (1)
+#else
+#  define USE_EPOLL (0)
+#endif
+
+#if USE_EPOLL
+/*
+ * :nodoc:
+ * returns IO object if EPOLLEXCLUSIVE works and arms readers
+ */
+static VALUE prep_readers(VALUE cls, VALUE readers)
+{
+	long i;
+	int epfd = epoll_create1(EPOLL_CLOEXEC);
+	VALUE epio;
+
+	if (epfd < 0) rb_sys_fail("epoll_create1");
+
+	epio = rb_funcall(cls, rb_intern("for_fd"), 1, INT2NUM(epfd));
+
+	Check_Type(readers, T_ARRAY);
+	for (i = 0; i < RARRAY_LEN(readers); i++) {
+		int rc;
+		struct epoll_event e;
+		rb_io_t *fptr;
+		VALUE io = rb_ary_entry(readers, i);
+
+		e.data.u64 = i; /* the reason readers shouldn't change */
+
+		/*
+		 * I wanted to use EPOLLET here, but maintaining our own
+		 * equivalent of ep->rdllist in Ruby-space doesn't fit
+		 * our design at all (and the kernel already has it's own
+		 * code path for doing it).  So let the kernel spend
+		 * cycles on maintaining level-triggering.
+		 */
+		e.events = EPOLLEXCLUSIVE | EPOLLIN;
+		io = rb_io_get_io(io);
+		GetOpenFile(io, fptr);
+		rc = epoll_ctl(epfd, EPOLL_CTL_ADD, fptr->fd, &e);
+		if (rc < 0) rb_sys_fail("epoll_ctl");
+	}
+	return epio;
+}
+#endif /* USE_EPOLL */
+
+#if USE_EPOLL
+struct ep_wait {
+	struct epoll_event *events;
+	rb_io_t *fptr;
+	int maxevents;
+	int timeout_msec;
+};
+
+static void *do_wait(void *ptr) /* runs w/o GVL */
+{
+	struct ep_wait *epw = ptr;
+
+	return (void *)(long)epoll_wait(epw->fptr->fd, epw->events,
+				epw->maxevents, epw->timeout_msec);
+}
+
+/* :nodoc: */
+/* readers must not change between prepare_readers and get_readers */
+static VALUE
+get_readers(VALUE epio, VALUE ready, VALUE readers, VALUE timeout_msec)
+{
+	struct ep_wait epw;
+	long i, n;
+	VALUE buf;
+
+	Check_Type(ready, T_ARRAY);
+	Check_Type(readers, T_ARRAY);
+	epw.maxevents = RARRAY_LENINT(readers);
+	buf = rb_str_buf_new(sizeof(struct epoll_event) * epw.maxevents);
+	epw.events = (struct epoll_event *)RSTRING_PTR(buf);
+	epio = rb_io_get_io(epio);
+	GetOpenFile(epio, epw.fptr);
+
+	epw.timeout_msec = NUM2INT(timeout_msec);
+	n = (long)rb_thread_call_without_gvl(do_wait, &epw, RUBY_UBF_IO, NULL);
+	if (n < 0) {
+		if (errno != EINTR) rb_sys_fail("epoll_wait");
+		n = 0;
+	}
+	/* Linux delivers events in order received */
+	for (i = 0; i < n; i++) {
+		struct epoll_event *ev = &epw.events[i];
+		VALUE obj = rb_ary_entry(readers, ev->data.u64);
+
+		if (RTEST(obj))
+			rb_ary_push(ready, obj);
+	}
+	rb_str_resize(buf, 0);
+	rb_gc_force_recycle(buf);
+	return Qfalse;
+}
+#endif /* USE_EPOLL */
+
+static void init_epollexclusive(VALUE mUnicorn)
+{
+#if USE_EPOLL
+	VALUE cWaiter = rb_define_class_under(mUnicorn, "Waiter", rb_cIO);
+	rb_define_singleton_method(cWaiter, "prep_readers", prep_readers, 1);
+	rb_define_method(cWaiter, "get_readers", get_readers, 3);
+#endif
+}
diff --git a/ext/unicorn_http/extconf.rb b/ext/unicorn_http/extconf.rb
index 46070a7..13dec41 100644
--- a/ext/unicorn_http/extconf.rb
+++ b/ext/unicorn_http/extconf.rb
@@ -38,4 +38,5 @@
   message("no, needs Ruby 2.6+\n")
 end
 
+have_func('epoll_create1', %w(sys/epoll.h))
 create_makefile("unicorn_http")
diff --git a/ext/unicorn_http/unicorn_http.rl b/ext/unicorn_http/unicorn_http.rl
index e934a32..605b23f 100644
--- a/ext/unicorn_http/unicorn_http.rl
+++ b/ext/unicorn_http/unicorn_http.rl
@@ -12,6 +12,7 @@
 #include "common_field_optimization.h"
 #include "global_variables.h"
 #include "c_util.h"
+#include "epollexclusive.h"
 
 void init_unicorn_httpdate(void);
 
@@ -1017,5 +1018,7 @@ void Init_unicorn_http(void)
   id_clear = rb_intern("clear");
 #endif
   id_is_chunked_p = rb_intern("is_chunked?");
+
+  init_epollexclusive(mUnicorn);
 }
 #undef SET_GLOBAL
diff --git a/lib/unicorn/http_server.rb b/lib/unicorn/http_server.rb
index 7f33f98..21f2a05 100644
--- a/lib/unicorn/http_server.rb
+++ b/lib/unicorn/http_server.rb
@@ -685,7 +685,6 @@ def init_worker_process(worker)
     LISTENERS.each { |sock| sock.close_on_exec = true }
 
     worker.user(*user) if user.kind_of?(Array) && ! worker.switched
-    self.timeout /= 2.0 # halve it for select()
     @config = nil
     build_app! unless preload_app
     @after_fork = @listener_opts = @orig_app = nil
@@ -705,11 +704,22 @@ def reopen_worker_logs(worker_nr)
     exit!(77) # EX_NOPERM in sysexits.h
   end
 
+  def prep_readers(readers)
+    wtr = Unicorn::Waiter.prep_readers(readers)
+    @timeout *= 500 # to milliseconds for epoll, but halved
+    wtr
+  rescue
+    require_relative 'select_waiter'
+    @timeout /= 2.0 # halved for IO.select
+    Unicorn::SelectWaiter.new
+  end
+
   # runs inside each forked worker, this sits around and waits
   # for connections and doesn't die until the parent dies (or is
   # given a INT, QUIT, or TERM signal)
   def worker_loop(worker)
     readers = init_worker_process(worker)
+    waiter = prep_readers(readers)
     reopen = false
 
     # this only works immediately if the master sent us the signal
@@ -722,8 +732,7 @@ def worker_loop(worker)
     begin
       reopen = reopen_worker_logs(worker.nr) if reopen
       worker.tick = time_now.to_i
-      tmp = ready.dup
-      while sock = tmp.shift
+      while sock = ready.shift
         # Unicorn::Worker#kgio_tryaccept is not like accept(2) at all,
         # but that will return false
         if client = sock.kgio_tryaccept
@@ -735,7 +744,7 @@ def worker_loop(worker)
 
       # timeout so we can .tick and keep parent from SIGKILL-ing us
       worker.tick = time_now.to_i
-      ret = IO.select(readers, nil, nil, @timeout) and ready = ret[0]
+      waiter.get_readers(ready, readers, @timeout)
     rescue => e
       redo if reopen && readers[0]
       Unicorn.log_error(@logger, "listen loop error", e) if readers[0]
diff --git a/lib/unicorn/select_waiter.rb b/lib/unicorn/select_waiter.rb
new file mode 100644
index 0000000..cb84aab
--- /dev/null
+++ b/lib/unicorn/select_waiter.rb
@@ -0,0 +1,6 @@
+# fallback for non-Linux and Linux <4.5 systems w/o EPOLLEXCLUSIVE
+class Unicorn::SelectWaiter # :nodoc:
+  def get_readers(ready, readers, timeout) # :nodoc:
+    ret = IO.select(readers, nil, nil, timeout) and ready.replace(ret[0])
+  end
+end
diff --git a/t/test-lib.sh b/t/test-lib.sh
index 7f97958..e70d0c6 100644
--- a/t/test-lib.sh
+++ b/t/test-lib.sh
@@ -94,7 +94,8 @@ check_stderr () {
 	set +u
 	_r_err=${1-${r_err}}
 	set -u
-	if grep -v $T $_r_err | grep -i Error
+	if grep -v $T $_r_err | grep -i Error | \
+		grep -v NameError.*Unicorn::Waiter
 	then
 		die "Errors found in $_r_err"
 	elif grep SIGKILL $_r_err
diff --git a/test/unit/test_waiter.rb b/test/unit/test_waiter.rb
new file mode 100644
index 0000000..0995de2
--- /dev/null
+++ b/test/unit/test_waiter.rb
@@ -0,0 +1,34 @@
+require 'test/unit'
+require 'unicorn'
+require 'unicorn/select_waiter'
+class TestSelectWaiter < Test::Unit::TestCase
+
+  def test_select_timeout # n.b. this is level-triggered
+    sw = Unicorn::SelectWaiter.new
+    IO.pipe do |r,w|
+      sw.get_readers(ready = [], [r], 0)
+      assert_equal [], ready
+      w.syswrite '.'
+      sw.get_readers(ready, [r], 1000)
+      assert_equal [r], ready
+      sw.get_readers(ready, [r], 0)
+      assert_equal [r], ready
+    end
+  end
+
+  def test_linux # ugh, also level-triggered, unlikely to change
+    IO.pipe do |r,w|
+      wtr = Unicorn::Waiter.prep_readers([r])
+      wtr.get_readers(ready = [], [r], 0)
+      assert_equal [], ready
+      w.syswrite '.'
+      wtr.get_readers(ready = [], [r], 1000)
+      assert_equal [r], ready
+      wtr.get_readers(ready = [], [r], 1000)
+      assert_equal [r], ready, 'still ready (level-triggered :<)'
+      assert_nil wtr.close
+    end
+  rescue SystemCallError => e
+    warn "#{e.message} (#{e.class})"
+  end if Unicorn.const_defined?(:Waiter)
+end