Skip to content

Commit

Permalink
Better error handling during force shutdown
Browse files Browse the repository at this point in the history
Only allow `ForceShutdown` to be raised in a thread during specific areas of the
connection-processing cycle (marked by `with_force_shutdown` blocks),
to ensure that the raised error is always rescued and handled cleanly.
Fixes an issue where the `force_shutdown_after: 0` option throws
uncaught exceptions from the threadpool on shutdown.
  • Loading branch information
wjordan committed Sep 23, 2020
1 parent 7d4df29 commit fcfe784
Show file tree
Hide file tree
Showing 5 changed files with 76 additions and 46 deletions.
1 change: 1 addition & 0 deletions History.md
Expand Up @@ -5,6 +5,7 @@

* Bugfixes
* Prevent connections from entering Reactor after shutdown begins (#2377)
* Better error handling during force shutdown (#2271)

* Refactor
* Change Events#ssl_error signature from (error, peeraddr, peercert) to (error, ssl_socket) (#2375)
Expand Down
20 changes: 14 additions & 6 deletions lib/puma/server.rb
Expand Up @@ -232,7 +232,9 @@ def run(background=true)
if @queue_requests
process_now = client.eagerly_finish
else
client.finish(@first_data_timeout)
@thread_pool.with_force_shutdown do
client.finish(@first_data_timeout)
end
process_now = true
end
rescue MiniSSL::SSLError => e
Expand All @@ -244,7 +246,7 @@ def run(background=true)
client.close

@events.parse_error e, client
rescue ConnectionError, EOFError => e
rescue ConnectionError, EOFError, ThreadPool::ForceShutdown => e
client.close

@events.connection_error e, client
Expand Down Expand Up @@ -422,7 +424,11 @@ def process_client(client, buffer)
check_for_more_data = false
end

unless client.reset(check_for_more_data)
next_request_ready = @thread_pool.with_force_shutdown do
client.reset(check_for_more_data)
end

unless next_request_ready
@shutdown_mutex.synchronize do
return unless @queue_requests
close_socket = false
Expand All @@ -435,7 +441,7 @@ def process_client(client, buffer)
end

# The client disconnected while we were reading data
rescue ConnectionError
rescue ConnectionError, ThreadPool::ForceShutdown
# Swallow them. The ensure tries to close +client+ down

# SSL handshake error
Expand Down Expand Up @@ -638,7 +644,9 @@ def handle_request(req, lines)

begin
begin
status, headers, res_body = @app.call(env)
status, headers, res_body = @thread_pool.with_force_shutdown do
@app.call(env)
end

return :async if req.hijacked

Expand Down Expand Up @@ -936,7 +944,7 @@ def graceful_shutdown

if @thread_pool
if timeout = @options[:force_shutdown_after]
@thread_pool.shutdown timeout.to_i
@thread_pool.shutdown timeout.to_f
else
@thread_pool.shutdown
end
Expand Down
22 changes: 20 additions & 2 deletions lib/puma/thread_pool.rb
Expand Up @@ -62,6 +62,8 @@ def initialize(min, max, *extra, &block)
end

@clean_thread_locals = false
@force_shutdown = false
@shutdown_mutex = Mutex.new
end

attr_reader :spawned, :trim_requested, :waiting
Expand Down Expand Up @@ -322,6 +324,19 @@ def auto_reap!(timeout=5)
@reaper.start!
end

# Allows ThreadPool::ForceShutdown to be raised within the
# provided block if the thread is forced to shutdown during execution.
def with_force_shutdown
t = Thread.current
@shutdown_mutex.synchronize do
raise ForceShutdown if @force_shutdown
t[:with_force_shutdown] = true
end
yield
ensure
t[:with_force_shutdown] = false
end

# Tell all threads in the pool to exit and wait for them to finish.
# Wait +timeout+ seconds then raise +ForceShutdown+ in remaining threads.
# Next, wait an extra +grace+ seconds then force-kill remaining threads.
Expand Down Expand Up @@ -356,8 +371,11 @@ def shutdown(timeout=-1)
join.call(timeout)

# If threads are still running, raise ForceShutdown and wait to finish.
threads.each do |t|
t.raise ForceShutdown
@shutdown_mutex.synchronize do
@force_shutdown = true
threads.each do |t|
t.raise ForceShutdown if t[:with_force_shutdown]
end
end
join.call(SHUTDOWN_GRACE_TIME)

Expand Down
67 changes: 33 additions & 34 deletions test/test_puma_server.rb
Expand Up @@ -949,66 +949,65 @@ def assert_does_not_allow_http_injection(app, opts = {})
end

# Perform a server shutdown while requests are pending (one in app-server response, one still sending client request).
def shutdown_requests(app_delay: 2, request_delay: 1, post: false, response:, **options)
def shutdown_requests(s1_complete: true, s1_response: nil, post: false, s2_response: nil, **options)
@server = Puma::Server.new @app, @events, options
server_run app: ->(_) {
sleep app_delay
mutex = Mutex.new
app_finished = ConditionVariable.new
server_run app: ->(env) {
path = env['REQUEST_PATH']
mutex.synchronize do
app_finished.signal
app_finished.wait(mutex) if path == '/s1'
end
[204, {}, []]
}

s1 = send_http "GET / HTTP/1.1\r\n\r\n"
s1 = nil
s2 = send_http post ?
"POST / HTTP/1.1\r\nHost: test.com\r\nContent-Type: text/plain\r\nContent-Length: 5\r\n\r\nhi!" :
"GET / HTTP/1.1\r\n"
sleep 0.1

"POST /s2 HTTP/1.1\r\nHost: test.com\r\nContent-Type: text/plain\r\nContent-Length: 5\r\n\r\nhi!" :
"GET /s2 HTTP/1.1\r\n"
mutex.synchronize do
s1 = send_http "GET /s1 HTTP/1.1\r\n\r\n"
app_finished.wait(mutex)
app_finished.signal if s1_complete
end
@server.stop
sleep request_delay
Thread.pass until @server.instance_variable_get(:@thread_pool).instance_variable_get(:@shutdown)

s2 << "\r\n"
assert_match(s1_response, s1.gets) if s1_response

assert_match(/204/, s1.gets)
# Send s2 after shutdown begins
s2 << "\r\n" unless IO.select([s2], nil, nil, 0.1)

assert IO.select([s2], nil, nil, app_delay), 'timeout waiting for response'
assert IO.select([s2], nil, nil, 10), 'timeout waiting for response'
s2_result = begin
s2.gets
rescue Errno::ECONNABORTED, Errno::ECONNRESET
# Some platforms raise errors instead of returning a response/EOF when a TCP connection is aborted.
post ? '408' : nil
end

if response
assert_match response, s2_result
if s2_response
assert_match s2_response, s2_result
else
assert_nil s2_result
end
end

# Shutdown should allow pending requests to complete.
# Shutdown should allow pending requests and app-responses to complete.
def test_shutdown_requests
shutdown_requests response: /204/
shutdown_requests response: /204/, queue_requests: false
end

# Requests stuck longer than `first_data_timeout` should have connection closed (408 w/pending POST body).
def test_shutdown_data_timeout
shutdown_requests request_delay: 3, first_data_timeout: 2, response: nil
shutdown_requests request_delay: 3, first_data_timeout: 2, response: nil, queue_requests: false
shutdown_requests request_delay: 3, first_data_timeout: 2, response: /408/, post: true
opts = {s1_response: /204/, s2_response: /204/}
shutdown_requests(**opts)
shutdown_requests(**opts, queue_requests: false)
end

# Requests still pending after `force_shutdown_after` should have connection closed (408 w/pending POST body).
# App-responses still pending should return 503 (uncaught Puma::ThreadPool::ForceShutdown exception).
def test_force_shutdown
shutdown_requests request_delay: 4, response: nil, force_shutdown_after: 3
shutdown_requests request_delay: 4, response: nil, force_shutdown_after: 3, queue_requests: false
shutdown_requests request_delay: 4, response: /408/, force_shutdown_after: 3, post: true
end

# App-responses still pending during `force_shutdown_after` should return 503
# (uncaught Puma::ThreadPool::ForceShutdown exception).
def test_force_shutdown_app
shutdown_requests app_delay: 3, response: /503/, force_shutdown_after: 3
shutdown_requests app_delay: 3, response: /503/, force_shutdown_after: 3, queue_requests: false
opts = {s1_complete: false, s1_response: /503/, s2_response: nil, force_shutdown_after: 0}
shutdown_requests(**opts)
shutdown_requests(**opts, queue_requests: false)
shutdown_requests(**opts, post: true, s2_response: /408/)
end

def test_http11_connection_header_queue
Expand Down
12 changes: 8 additions & 4 deletions test/test_thread_pool.rb
Expand Up @@ -222,8 +222,10 @@ def test_force_shutdown_immediately

pool = mutex_pool(0, 1) do
begin
pool.signal
sleep
pool.with_force_shutdown do
pool.signal
sleep
end
rescue Puma::ThreadPool::ForceShutdown
rescued = true
end
Expand All @@ -248,8 +250,10 @@ def test_shutdown_with_grace
rescued = []
pool = mutex_pool(2, 2) do
begin
pool.signal
sleep
pool.with_force_shutdown do
pool.signal
sleep
end
rescue Puma::ThreadPool::ForceShutdown
rescued << Thread.current
sleep
Expand Down

0 comments on commit fcfe784

Please sign in to comment.