From 603fe20c0b8c05bc1cca8d01caadbd7060518141 Mon Sep 17 00:00:00 2001 From: thedarkone Date: Thu, 14 May 2015 03:14:53 +0200 Subject: AR::ConPool - remove synchronization around connection cache. Renamed `@reserved_connections` -> `@thread_cached_conns`. New name clearly conveys the purpose of the cache, which is to speed-up `#connection` method. The new `@thread_cached_conns` now also uses `Thread` objects as keys (instead of previously `Thread.current.object_id`). Since there is no longer any synchronization around `@thread_cached_conns`, `disconnect!` and `clear_reloadable_connections!` methods now pre-emptively obtain ownership (via `checkout`) of all existing connections, before modifying internal data structures. A private method `release` has been renamed `thread_conn_uncache` to clear-up its purpose. Fixed some brittle `thread.status == "sleep"` tests (threads can go into sleep even without locks). --- .../abstract/connection_pool.rb | 382 ++++++++++++++++----- .../lib/active_record/connection_handling.rb | 2 +- activerecord/test/cases/connection_pool_test.rb | 136 +++++++- 3 files changed, 436 insertions(+), 84 deletions(-) diff --git a/activerecord/lib/active_record/connection_adapters/abstract/connection_pool.rb b/activerecord/lib/active_record/connection_adapters/abstract/connection_pool.rb index 1d550c0b39..5bfb4d779c 100644 --- a/activerecord/lib/active_record/connection_adapters/abstract/connection_pool.rb +++ b/activerecord/lib/active_record/connection_adapters/abstract/connection_pool.rb @@ -1,7 +1,6 @@ require 'thread' require 'thread_safe' require 'monitor' -require 'set' module ActiveRecord # Raised when a connection could not be obtained within the connection @@ -10,6 +9,12 @@ module ActiveRecord class ConnectionTimeoutError < ConnectionNotEstablished end + # Raised when a pool was unable to get ahold of all its connections + # to perform a "group" action such as +ConnectionPool#disconnect!+ + # or +ConnectionPool#clear_reloadable_connections!+. + class ExclusiveConnectionTimeoutError < ConnectionTimeoutError + end + module ConnectionAdapters # Connection pool base class for managing Active Record database # connections. @@ -63,6 +68,15 @@ module ActiveRecord # connection at the end of a thread or a thread dies unexpectedly. # Regardless of this setting, the Reaper will be invoked before every # blocking wait. (Default nil, which means don't schedule the Reaper). + # + #-- + # Synchronization policy: + # * all public methods can be called outside +synchronize+ + # * access to these i-vars needs to be in +synchronize+: + # * @connections + # * @now_connecting + # * private methods that require being called in a +synchronize+ blocks + # are now explicitly documented class ConnectionPool # Threadsafe, fair, FIFO queue. Meant to be used by ConnectionPool # with which it shares a Monitor. But could be a generic Queue. @@ -191,12 +205,72 @@ module ActiveRecord end end + # Adds the ability to turn a basic fair FIFO queue into one + # biased to some thread. + module BiasableQueue # :nodoc: + class BiasedConditionVariable # :nodoc: + # semantics of condition variables guarantee that +broadcast+, +broadcast_on_biased+, + # +signal+ and +wait+ methods are only called while holding a lock + def initialize(lock, other_cond, preferred_thread) + @real_cond = lock.new_cond + @other_cond = other_cond + @preferred_thread = preferred_thread + @num_waiting_on_real_cond = 0 + end + + def broadcast + broadcast_on_biased + @other_cond.broadcast + end + + def broadcast_on_biased + @num_waiting_on_real_cond = 0 + @real_cond.broadcast + end + + def signal + if @num_waiting_on_real_cond > 0 + @num_waiting_on_real_cond -= 1 + @real_cond + else + @other_cond + end.signal + end + + def wait(timeout) + if Thread.current == @preferred_thread + @num_waiting_on_real_cond += 1 + @real_cond + else + @other_cond + end.wait(timeout) + end + end + + def with_a_bias_for(thread) + previous_cond = nil + new_cond = nil + synchronize do + previous_cond = @cond + @cond = new_cond = BiasedConditionVariable.new(@lock, @cond, thread) + end + yield + ensure + synchronize do + @cond = previous_cond if previous_cond + new_cond.broadcast_on_biased if new_cond # wake up any remaining sleepers + end + end + end + # Connections must be leased while holding the main pool mutex. This is # an internal subclass that also +.leases+ returned connections while # still in queue's critical section (queue synchronizes with the same # +@lock+ as the main pool) so that a returned connection is already # leased and there is no need to re-enter synchronized block. class ConnectionLeasingQueue < Queue # :nodoc: + include BiasableQueue + private def internal_poll(timeout) conn = super @@ -253,8 +327,17 @@ module ActiveRecord # default max pool size to 5 @size = (spec.config[:pool] && spec.config[:pool].to_i) || 5 - # The cache of reserved connections mapped to threads - @reserved_connections = ThreadSafe::Cache.new(:initial_capacity => @size) + # The cache of threads mapped to reserved connections, the sole purpose + # of the cache is to speed-up +connection+ method, it is not the authoritative + # registry of which thread owns which connection, that is tracked by + # +connection.owner+ attr on each +connection+ instance. + # The invariant works like this: if there is mapping of +thread => conn+, + # then that +thread+ does indeed own that +conn+, however an absence of a such + # mapping does not mean that the +thread+ doesn't own the said connection, in + # that case +conn.owner+ attr should be consulted. + # Access and modification of +@thread_cached_conns+ does not require + # synchronization. + @thread_cached_conns = ThreadSafe::Cache.new(:initial_capacity => @size) @connections = [] @automatic_reconnect = true @@ -264,6 +347,9 @@ module ActiveRecord # currently in the process of independently establishing connections to the DB. @now_connecting = 0 + # A boolean toggle that allows/disallows new connections. + @new_cons_enabled = true + @available = ConnectionLeasingQueue.new self end @@ -271,41 +357,45 @@ module ActiveRecord # #checkout to obtain one if necessary. # # #connection can be called any number of times; the connection is - # held in a hash keyed by the thread id. + # held in a cache keyed by a thread. def connection - # this is correctly done double-checked locking - # (ThreadSafe::Cache's lookups have volatile semantics) - @reserved_connections[current_connection_id] || synchronized_connection_retrieval + @thread_cached_conns[connection_cache_key(Thread.current)] ||= checkout end # Is there an open connection that is being used for the current thread? + # + # This method only works for connections that have been abtained through + # #connection or #with_connection methods, connections obtained through + # #checkout will not be detected by #active_connection? def active_connection? - synchronize do - @reserved_connections.fetch(current_connection_id) { - return false - }.in_use? - end + @thread_cached_conns[connection_cache_key(Thread.current)] end # Signal that the thread is finished with the current connection. # #release_connection releases the connection-thread association # and returns the connection to the pool. - def release_connection(with_id = current_connection_id) - synchronize do - conn = @reserved_connections.delete(with_id) - checkin conn if conn + # + # This method only works for connections that have been obtained through + # #connection or #with_connection methods, connections obtained through + # #checkout will not be automatically released. + def release_connection(owner_thread = Thread.current) + if conn = @thread_cached_conns.delete(connection_cache_key(owner_thread)) + checkin conn end end - # If a connection already exists yield it to the block. If no connection + # If a connection obtained through #connection or #with_connection methods + # already exists yield it to the block. If no such connection # exists checkout a connection, yield it to the block, and checkin the # connection when finished. def with_connection - connection_id = current_connection_id - fresh_connection = true unless active_connection? - yield connection + unless conn = @thread_cached_conns[connection_cache_key(Thread.current)] + conn = connection + fresh_connection = true + end + yield conn ensure - release_connection(connection_id) if fresh_connection + release_connection if fresh_connection end # Returns true if a connection has already been opened. @@ -314,32 +404,81 @@ module ActiveRecord end # Disconnects all connections in the pool, and clears the pool. - def disconnect! - synchronize do - @reserved_connections.clear - @connections.each do |conn| - checkin conn - conn.disconnect! + # + # Raises: + # - +ExclusiveConnectionTimeoutError+ if unable to gain ownership of all + # connections in the pool within a timeout interval (default duration is + # +spec.config[:checkout_timeout] * 2+ seconds). + def disconnect(raise_on_acquisition_timeout = true) + with_exclusively_acquired_all_connections(raise_on_acquisition_timeout) do + synchronize do + @connections.each do |conn| + checkin conn + conn.disconnect! + end + @connections = [] + @available.clear end - @connections = [] - @available.clear end end - # Clears the cache which maps classes. - def clear_reloadable_connections! - synchronize do - @reserved_connections.clear - @connections.each do |conn| - checkin conn - conn.disconnect! if conn.requires_reloading? - end - @connections.delete_if(&:requires_reloading?) - @available.clear - @connections.each do |conn| - @available.add conn + # Disconnects all connections in the pool, and clears the pool. + # + # The pool first tries to gain ownership of all connections, if unable to + # do so within a timeout interval (default duration is + # +spec.config[:checkout_timeout] * 2+ seconds), the pool is forcefully + # disconneted wihout any regard for other connection owning threads. + def disconnect! + disconnect(false) + end + + # Clears the cache which maps classes and re-connects connections that + # require reloading. + # + # Raises: + # - +ExclusiveConnectionTimeoutError+ if unable to gain ownership of all + # connections in the pool within a timeout interval (default duration is + # +spec.config[:checkout_timeout] * 2+ seconds). + def clear_reloadable_connections(raise_on_acquisition_timeout = true) + num_new_conns_required = 0 + + with_exclusively_acquired_all_connections(raise_on_acquisition_timeout) do + synchronize do + @connections.each do |conn| + checkin conn + conn.disconnect! if conn.requires_reloading? + end + @connections.delete_if(&:requires_reloading?) + + @available.clear + + if @connections.size < @size + # because of the pruning done by this method, we might be running + # low on connections, while threads stuck in queue are helpless + # (not being able to establish new connections for themselves), + # see also more detailed explanation in +remove+ + num_new_conns_required = num_waiting_in_queue - @connections.size + end + + @connections.each do |conn| + @available.add conn + end end end + + bulk_make_new_connections(num_new_conns_required) if num_new_conns_required > 0 + end + + # Clears the cache which maps classes and re-connects connections that + # require reloading. + # + # The pool first tries to gain ownership of all connections, if unable to + # do so within a timeout interval (default duration is + # +spec.config[:checkout_timeout] * 2+ seconds), the pool forcefully + # clears the cache and reloads connections without any regard for other + # connection owning threads. + def clear_reloadable_connections! + clear_reloadable_connections(false) end # Check-out a database connection from the pool, indicating that you want @@ -356,8 +495,8 @@ module ActiveRecord # # Raises: # - ConnectionTimeoutError: no connection can be obtained from the pool. - def checkout - checkout_and_verify(acquire_connection) + def checkout(checkout_timeout = @checkout_timeout) + checkout_and_verify(acquire_connection(checkout_timeout)) end # Check-in a database connection back into the pool, indicating that you @@ -367,14 +506,12 @@ module ActiveRecord # calling +checkout+ on this pool. def checkin(conn) synchronize do - owner = conn.owner + remove_connection_from_thread_cache conn conn.run_callbacks :checkin do conn.expire end - release conn, owner - @available.add conn end end @@ -385,11 +522,11 @@ module ActiveRecord needs_new_connection = false synchronize do + remove_connection_from_thread_cache conn + @connections.delete conn @available.delete conn - release conn, conn.owner - # @available.any_waiting? => true means that prior to removing this # conn, the pool was at its max size (@connections.size == @size) # this would mean that any threads stuck waiting in the queue wouldn't @@ -405,12 +542,9 @@ module ActiveRecord # This is intentionally done outside of the synchronized section as we # would like not to hold the main mutex while checking out new connections, # thus there is some chance that needs_new_connection information is now - # stale, we can live with that (try_to_checkout_new_connection will make + # stale, we can live with that (bulk_make_new_connections will make # sure not to exceed the pool's @size limit). - if needs_new_connection && new_conn = try_to_checkout_new_connection - # make the new_conn available to the starving threads stuck @available Queue - checkin new_conn - end + bulk_make_new_connections(1) if needs_new_connection end # Recover lost connections for the pool. A lost connection can occur if @@ -435,22 +569,117 @@ module ActiveRecord end end + def num_waiting_in_queue # :nodoc: + @available.num_waiting + end + private - def synchronized_connection_retrieval - conn = checkout + #-- + # this is unfortunately not concurrent + def bulk_make_new_connections(num_new_conns_needed) + num_new_conns_needed.times do + # try_to_checkout_new_connection will not exceed pool's @size limit + if new_conn = try_to_checkout_new_connection + # make the new_conn available to the starving threads stuck @available Queue + checkin(new_conn) + end + end + end + + #-- + # From the discussion on Github: + # https://github.com/rails/rails/pull/14938#commitcomment-6601951 + # This hook-in method allows for easier monkey-patching fixes needed by + # JRuby users that use Fibers. + def connection_cache_key(thread) + thread + end + + # Take control of all existing connections so a "group" action such as + # reload/disconnect can be performed safely. It is no longer enough to + # wrap it in +synchronize+ because some pool's actions are allowed + # to be performed outside of the main +synchronize+ block. + def with_exclusively_acquired_all_connections(raise_on_acquisition_timeout = true) + with_new_connections_blocked do + attempt_to_checkout_all_existing_connections(raise_on_acquisition_timeout) + yield + end + end + + def attempt_to_checkout_all_existing_connections(raise_on_acquisition_timeout = true) + collected_conns = synchronize do + # account for our own connections + @connections.select {|conn| conn.owner == Thread.current} + end + + newly_checked_out = [] + timeout_time = Time.now + (@checkout_timeout * 2) + + @available.with_a_bias_for(Thread.current) do + while true + synchronize do + return if collected_conns.size == @connections.size && @now_connecting == 0 + remaining_timeout = timeout_time - Time.now + remaining_timeout = 0 if remaining_timeout < 0 + conn = checkout_for_exclusive_access(remaining_timeout) + collected_conns << conn + newly_checked_out << conn + end + end + end + rescue ExclusiveConnectionTimeoutError + # `raise_on_acquisition_timeout == false` means we are directed to ignore any + # timeouts and are expected to just give up: we've obtained as many connections + # as possible, note that in a case like that we don't return any of the + # `newly_checked_out` connections. + + if raise_on_acquisition_timeout + release_newly_checked_out = true + raise + end + rescue Exception # if something else went wrong + # this can't be a "naked" rescue, because we have should return conns + # even for non-StandardErrors + release_newly_checked_out = true + raise + ensure + if release_newly_checked_out && newly_checked_out + # releasing only those conns that were checked out in this method, conns + # checked outside this method (before it was called) are not for us to release + newly_checked_out.each {|conn| checkin(conn)} + end + end + + #-- + # Must be called in a synchronize block. + def checkout_for_exclusive_access(checkout_timeout) + checkout(checkout_timeout) + rescue ConnectionTimeoutError + # this block can't be easily moved into attempt_to_checkout_all_existing_connections's + # rescue block, because doing so would put it outside of synchronize section, without + # being in a critical section thread_report might become inaccurate + msg = "could not obtain ownership of all database connections in #{checkout_timeout} seconds" + + thread_report = [] + @connections.each do |conn| + unless conn.owner == Thread.current + thread_report << "#{conn} is owned by #{conn.owner}" + end + end + + msg << " (#{thread_report.join(', ')})" if thread_report.any? + + raise ExclusiveConnectionTimeoutError, msg + end + + def with_new_connections_blocked previous_value = nil - synchronize do # re-checking under lock for correct DCL semantics - # Cache#put_if_absent returns either `nil` (if insertion was successful, ie there was - # no previous current_connection_id mapping) or an existing value (if insertion - # failed because there already was a current_connection_id mapping) - previous_value = @reserved_connections.put_if_absent(current_connection_id, conn) - end - if previous_value # if we were too late and insertion failed - checkin(conn) - previous_value - else - conn + synchronize do + previous_value, @new_cons_enabled = @new_cons_enabled, false end + yield + ensure + synchronize { @new_cons_enabled = previous_value } end # Acquire a connection by one of 1) immediately removing one @@ -464,7 +693,7 @@ module ActiveRecord #-- # Implementation detail: the connection returned by +acquire_connection+ # will already be "+connection.lease+ -ed" to the current thread. - def acquire_connection + def acquire_connection(checkout_timeout) # NOTE: we rely on `@available.poll` and `try_to_checkout_new_connection` to # `conn.lease` the returned connection (and to do this in a `synchronized` # section), this is not the cleanest implementation, as ideally we would @@ -475,17 +704,16 @@ module ActiveRecord conn else reap - @available.poll(@checkout_timeout) + @available.poll(checkout_timeout) end end - def release(conn, owner) - thread_id = owner.object_id - - if @reserved_connections[thread_id] == conn - @reserved_connections.delete thread_id - end + #-- + # if owner_thread param is omitted, this must be called in synchronize block + def remove_connection_from_thread_cache(conn, owner_thread = conn.owner) + @thread_cached_conns.delete_pair(connection_cache_key(owner_thread), conn) end + alias_method :release, :remove_connection_from_thread_cache def new_connection Base.send(spec.adapter_method, spec.config).tap do |conn| @@ -493,10 +721,6 @@ module ActiveRecord end end - def current_connection_id #:nodoc: - Base.connection_id ||= Thread.current.object_id - end - # If the pool is not at a +@size+ limit, establish new connection. Connecting # to the DB is done outside main synchronized section. #-- @@ -507,7 +731,7 @@ module ActiveRecord # and increment @now_connecting, to prevent overstepping this pool's @size # constraint do_checkout = synchronize do - if (@connections.size + @now_connecting) < @size + if @new_cons_enabled && (@connections.size + @now_connecting) < @size @now_connecting += 1 end end diff --git a/activerecord/lib/active_record/connection_handling.rb b/activerecord/lib/active_record/connection_handling.rb index 24f5849e45..d6b661ff76 100644 --- a/activerecord/lib/active_record/connection_handling.rb +++ b/activerecord/lib/active_record/connection_handling.rb @@ -88,7 +88,7 @@ module ActiveRecord end def connection_id - ActiveRecord::RuntimeRegistry.connection_id + ActiveRecord::RuntimeRegistry.connection_id ||= Thread.current.object_id end def connection_id=(connection_id) diff --git a/activerecord/test/cases/connection_pool_test.rb b/activerecord/test/cases/connection_pool_test.rb index 3e563cd7cf..90b6fb55ef 100644 --- a/activerecord/test/cases/connection_pool_test.rb +++ b/activerecord/test/cases/connection_pool_test.rb @@ -100,7 +100,7 @@ module ActiveRecord t = Thread.new { @pool.checkout } # make sure our thread is in the timeout section - Thread.pass until t.status == "sleep" + Thread.pass until @pool.num_waiting_in_queue == 1 connection = cs.first connection.close @@ -112,7 +112,7 @@ module ActiveRecord t = Thread.new { @pool.checkout } # make sure our thread is in the timeout section - Thread.pass until t.status == "sleep" + Thread.pass until @pool.num_waiting_in_queue == 1 connection = cs.first @pool.remove connection @@ -234,7 +234,7 @@ module ActiveRecord mutex.synchronize { errors << e } end } - Thread.pass until t.status == "sleep" + Thread.pass until @pool.num_waiting_in_queue == i t end @@ -271,7 +271,7 @@ module ActiveRecord mutex.synchronize { errors << e } end } - Thread.pass until t.status == "sleep" + Thread.pass until @pool.num_waiting_in_queue == i t end @@ -390,6 +390,134 @@ module ActiveRecord connecting_threads.map(&:join) end end + + def test_non_bang_disconnect_and_clear_reloadable_connections_throw_exception_if_threads_dont_return_their_conns + @pool.checkout_timeout = 0.001 # no need to delay test suite by waiting the whole full default timeout + [:disconnect, :clear_reloadable_connections].each do |group_action_method| + @pool.with_connection do |connection| + assert_raises(ExclusiveConnectionTimeoutError) do + Thread.new { @pool.send(group_action_method) }.join + end + end + end + end + + def test_disconnect_and_clear_reloadable_connections_attempt_to_wait_for_threads_to_return_their_conns + [:disconnect, :disconnect!, :clear_reloadable_connections, :clear_reloadable_connections!].each do |group_action_method| + begin + thread = timed_join_result = nil + @pool.with_connection do |connection| + thread = Thread.new { @pool.send(group_action_method) } + + # give the other `thread` some time to get stuck in `group_action_method` + timed_join_result = thread.join(0.3) + # thread.join # => `nil` means the other thread hasn't finished running and is still waiting for us to + # release our connection + assert_nil timed_join_result + + # assert that since this is within default timeout our connection hasn't been forcefully taken away from us + assert @pool.active_connection? + end + ensure + thread.join if thread && !timed_join_result # clean up the other thread + end + end + end + + def test_bang_versions_of_disconnect_and_clear_reloadable_connections_if_unable_to_aquire_all_connections_proceed_anyway + @pool.checkout_timeout = 0.001 # no need to delay test suite by waiting the whole full default timeout + [:disconnect!, :clear_reloadable_connections!].each do |group_action_method| + @pool.with_connection do |connection| + Thread.new { @pool.send(group_action_method) }.join + # assert connection has been forcefully taken away from us + assert_not @pool.active_connection? + end + end + end + + def test_disconnect_and_clear_reloadable_connections_are_able_to_preempt_other_waiting_threads + with_single_connection_pool do |pool| + [:disconnect, :disconnect!, :clear_reloadable_connections, :clear_reloadable_connections!].each do |group_action_method| + conn = pool.connection # drain the only available connection + second_thread_done = ActiveSupport::Concurrency::Latch.new + + # create a first_thread and let it get into the FIFO queue first + first_thread = Thread.new do + pool.with_connection { second_thread_done.await } + end + + # wait for first_thread to get in queue + Thread.pass until pool.num_waiting_in_queue == 1 + + # create a different, later thread, that will attempt to do a "group action", + # but because of the group action semantics it should be able to preempt the + # first_thread when a connection is made available + second_thread = Thread.new do + pool.send(group_action_method) + second_thread_done.release + end + + # wait for second_thread to get in queue + Thread.pass until pool.num_waiting_in_queue == 2 + + # return the only available connection + pool.checkin(conn) + + # if the second_thread is not able to preempt the first_thread, + # they will temporarily (until either of them timeouts with ConnectionTimeoutError) + # deadlock and a join(2) timeout will be reached + failed = true unless second_thread.join(2) + + #--- post test clean up start + second_thread_done.release if failed + + # after `pool.disconnect()` the first thread will be left stuck in queue, no need to wait for + # it to timeout with ConnectionTimeoutError + if (group_action_method == :disconnect || group_action_method == :disconnect!) && pool.num_waiting_in_queue > 0 + pool.with_connection {} # create a new connection in case there are threads still stuck in a queue + end + + first_thread.join + second_thread.join + #--- post test clean up end + + flunk "#{group_action_method} is not able to preempt other waiting threads" if failed + end + end + end + + def test_clear_reloadable_connections_creates_new_connections_for_waiting_threads_if_necessary + with_single_connection_pool do |pool| + conn = pool.connection # drain the only available connection + def conn.requires_reloading? # make sure it gets removed from the pool by clear_reloadable_connections + true + end + + stuck_thread = Thread.new do + pool.with_connection {} + end + + # wait for stuck_thread to get in queue + Thread.pass until pool.num_waiting_in_queue == 1 + + pool.clear_reloadable_connections + + unless stuck_thread.join(2) + flunk 'clear_reloadable_connections must not let other connection waiting threads get stuck in queue' + end + + assert_equal 0, pool.num_waiting_in_queue + end + end + + private + def with_single_connection_pool + one_conn_spec = ActiveRecord::Base.connection_pool.spec.dup + one_conn_spec.config[:pool] = 1 # this is safe to do, because .dupped ConnectionSpecification also auto-dups its config + yield(pool = ConnectionPool.new(one_conn_spec)) + ensure + pool.disconnect! if pool + end end end end -- cgit v1.2.3