= ApplicationPool algorithm


== Introduction

For efficiency reasons, Passenger keeps a pool spawned Rails/Ruby applications.
Please read the C++ API documentation for the ApplicationPool class for a full
introduction. This document describes an algorithm for managing the pool, in a
high-level way.

The algorithm should strive to keep spawning to a minimum.


== Definitions

=== Vocabulary

- "Application root":
  The toplevel directory in which an application is contained. For Rails
  application, this is the same as RAILS_ROOT, i.e. the directory that contains
  "app/", "public/", etc. For a Rack application, this is the directory that
  contains "config.ru".

- "Active application process":
  An application process that has more than 0 sessions.

=== Types

Most of the types that we use in this document are pretty standard. But we
explicitly define some special types:

- list<SomeType>
  A doubly linked list which contains elements of type SomeType. It supports
  all the usual list operations that one can expect from a linked list, like
  add_to_back(), etc.
  
  We assume that operations that insert an element into the list return an
  iterator object. An iterator object is an opaque object which represents a
  specific position in the list; it probably contains the links to the previous
  and the next iterator, as well as a reference to the actual list element,
  depending on the list implementation.
  
  The following operations deserve special mention:
  * remove(iterator)
    Removes the specified element from the list, as represented by the given
    iterator. This operation can be done in O(1) time.
  
  * move_to_front(iterator)
    Moves the specified element - as represented by the given iterator - to
    the front of the list. This operation can be done in O(1) time.
  
- Group
  A compound type (class) which contains information about an application root,
  such as the application processes that have been spawned for this application
  root.
  
  A Group has the following members:
  * name (string):
    This group's key in the _groups_ map of the application pool.

  * app_root (string):
    This group's application root. Note that it is *not* guaranteed that all
    ProcessInfo objects in _processes_ have the same application root.

  * processes (list<ProcessInfo>):
    A list of ProcessInfo objects.
    
    Invariant:
        processes is non-empty.
        for all 0 <= i < processes.size() - 1:
           processes[i].group_name == name
           if processes[i].process is active:
              processes[i + 1].process is active

  * size (unsigned integer):
    The number of items in _processes_.
    Invariant:
        if !detached:
           size == processes.size()
  
  * max_requests (unsigned integer):
    The maximum number of requests that each application process in
    this group may process. After having processed this
    many requests, the application process will be shut down.
    A value of 0 indicates that there is no maximum.

  * min_processes (unsigned integer):
    The minimum number of processes that the cleaner thread should keep in
    this group. Defaults to 0.

  * spawning (boolean): Whether a background thread is currently spawning
    a new process for this group.

  * spawner_thread: A handle to the background thread that is currently
    spawning a new process. Only valid if _spawning_ is true.

  * detached (boolean): If true, then it indicates that this Group is
    no longer accessible via _groups_.
    Set to false by the constructor.
    Invariant:
       (detached) == (This Group is accessible via _groups_.)
       if detached:
          for all process_info objects p that have once been in this.processes:
              p.detached

  * environment (string): Does nothing. Data is stored in memory for analytics
    purposes.

- ProcessInfo
  A compound type (class) which contains a reference to an application process
  object, as well as various metadata, such as iterators for various linked
  lists. These iterators make it possible to perform actions on the linked
  lists in O(1) time.
  
  A ProcessInfo has the following members:
  * process - A process object, representing an application process.
  * group_name (string) - The name of the group that this ProcessInfo belongs
    to.
  * identifier (string) - A key that uniquely identifies this ProcessInfo in
    this application pool. This key allows external processes to refer to a
    specific ProcessInfo object without knowing its memory pointer. It's set
    to a random string by ProcessInfo's constructor.
  * start_time (timestamp with milisecond resolution) - The time at which this
    application process was started. It's set to the current time by the
    constructor.
  * processed_requests (integer) - The number of requests processed by this
    application instance so far. Set to 0 by the constructor.
  * last_used (time) - The last time a session for this application process
    was opened or closed.
  * sessions (integer) - The number of open sessions for this application
    process. It's set to 0 by the constructor.
    Invariant:
       (sessions == 0 && !detached) == (This ProcessInfo is in inactive_apps.)
  * iterator - The iterator for this ProcessInfo in the linked list
    groups[process_info.group_name].processes
  * ia_iterator - The iterator for this ProcessInfo in the linked list
    inactive_apps. This iterator is only valid if this ProcessInfo really is
    in that list.
  * detached (boolean) - If true, then it indicates that this ProcessInfo is
    no longer accessible via _groups_; this implies that it's no longer
    contained in its associated Group's _processes_ member, a), and that
    _iterator_ and _ia_iterator_ are no longer valid.
    Set to false by the constructor.
    Invariant:
       (detached) == (This ProcessInfo is accessible via _groups_.)

- PoolOptions
  A structure containing additional information used by the spawn manager's
  spawning process, as well as by the get() function.
  
  A PoolOptions has at least the following members:
  * app_group_name (string) - A name which is used to group application
    processes together.
  * max_requests (unsigned integer) - The maximum number of requests that the
    application process may process. After having processed this many requests,
    the application process will be shut down. A value of 0 indicates that there
    is no maximum.
  * min_processes (unsigned integer) - The minimum number of processes for the
    current group that the cleaner thread should keep around.
  * use_global_queue (boolean) - Whether to use a global queue for all
    application processes, or a queue that's private to the application process.
    The users guide explains this feature in more detail.
  * restart_dir (string) - The directory in which the algorithm should look for
    restart.txt and always_restart.txt. The existance and modification times of
    these files tell the algorithm whether an application should be restarted.
  * environment (string) - The environment (RAILS_ENV/RACK_ENV) in which the app
    should run.

=== Special functions

- spawn(app_root, options)
  Spawns a new application process at the given application root with the given
  spawn options. Throws an exception if something went wrong. This function is
  thread-safe. Note that application process initialization can take an arbitrary
  amount of time.

=== Instance variables

The algorithm requires the following instance variables for storing state
information:

- lock: mutex
  This lock is used for implementing thread-safetiness. We assume that it
  is non-recursive, i.e. if a thread locks a mutex that it has already locked,
  then it will result in a deadlock.

- groups: map[string => Group]
  Maps an application root to its Group object. This map contains all
  application processes in the pool.
  
  Invariant:
     for all values g in groups:
        !g.detached
        g.size <= count
	for all i in g.processes:
           !i.detached
     (sum of all g.size in groups) == count
  
- max: integer
  The maximum number of ProcessInfo objects that may exist in the pool.

- max_per_app: integer
  The maximum number of ProcessInfo objects that may be simultaneously alive
  for a single Group.

- count: integer
  The current number of ProcessInfo objects in the pool.
  Since 'max' can be set dynamically during the life time of an application
  pool, 'count > max' is possible.

- active: integer
  The number of application processes in the pool that are active.
  Invariant:
     active <= count

- inactive_apps: list<ProcessInfo>
  A linked list of ProcessInfo objects. All application processes in this list
  are inactive.
  
  Invariant:
     inactive_apps.size() == count - active
     for all x in inactive_apps:
        x can be accessed from _groups_.
        x.sessions == 0

- waiting_on_global_queue: integer
  If global queuing mode is enabled, then when get() is waiting for a backend
  process to become idle, this variable will be incremented. When get() is done
  waiting, this variable will be decremented.


== Class relations

Here's an UML diagram in ASCII art:

[ProcessInfo] 1..* --------+
                           |
                           |
                           
                           1
[ApplicationPool]       [Group]
       1                  0..*

       |                   |
       +-------------------+


== Algorithm in pseudo code

# Thread-safetiness notes:
# - All wait commands are to unlock the lock during waiting.


# Connect to an existing application process, or spawn a new application process
# and connect to that if necessary.
# 'app_root' refers to an application root.
# 'options' is an object of type 'PoolOptions', which contains additional
# information which may be relevant for spawning.
#
# Returns a Session object, representing a single HTTP request/response pair.
function get(app_root, options):
	MAX_ATTEMPTS = 10
	attempt = 0
	while (true):
		attempt++
		lock.synchronize:
			process_info, group = checkout_without_lock(app_root, options)
		try:
			return process_info.process.connect()
		on exception:
			# The app process seems to have crashed.
			# So we remove this process from our data
			# structures.
			lock.synchronize:
				detach_without_lock(process_info.identifier)
				process_info.sessions--
			if (attempt == MAX_ATTEMPTS):
				propagate exception


# Detach the process with the given identifier from the pool's data structures.
function detach(identifier):
	lock.synchronize:
		return detach_without_lock(identifier)


# Checkout a process from the application pool and mark it as being used.
# If there's no appropriate process in the pool, or if there are not
# enough processes, then one will be spawned.
#
# Returns a pair of [ProcessInfo, Group].
# All exceptions that occur are propagated.
private function checkout_without_lock(app_root, options):
	group = groups[options.app_group_name]
	
	if needs_restart(app_root, options):
		Tell spawn server to reload code for options.app_group_name.
		if (group != null):
			detach_group_without_lock(group)
			group = null
	
	if (group != null):
		# There are existing processes for this app group.
		processes = group.processes
		
		if (processes.front.sessions == 0):
			# There is an inactive process, so we use it.
			process_info = processes.front
			processes.move_to_back(process_info.iterator)
			inactive_apps.remove(process_info.ia_iterator)
			mutate_max(active + 1)
		else:
			# All existing processes are active. We either use
			# one of them now or we wait until one of them becomes
			# available. And, if we're allowed to, we spawn an
			# extra process in the background.
			if spawning_allowed(group, options) and !group.spawning:
				spawn_in_background(group, options)
			process_info = select_process(processes, options)
			if (process_info == null):
				goto beginning of function
	else:
		# There are no processes for this app group.
		if (active >= max):
			# Looks like the pool is full and all processes are busy.
			# Wait until the pool appears to have changed in such a
			# way that we can spawn a new app group, and restart
			# this function.
			new_app_group_creatable.wait
			goto beginning of function
		elsif count == max:
			# The pool is full, and not all processes are busy, but
			# we're in a though situation nevertheless: there are
			# several processes which are inactive, and none of them
			# belong to our current app group, so we must kill one
			# of them in order to free a spot in the pool. But which
			# one do we kill? We want to minimize spawning.
			#
			# It's probably a good idea to keep some kind of
			# statistics in order to decide this. We want the
			# application root that gets the least traffic to be
			# killed. But for now, we kill a random application
			# process.
			process_info = inactive_apps.pop_front
			process_info.detached = true
			group = groups[process_info.group_name]
			processes = group.processes
			processes.remove(process_info.iterator)
			if processes.empty():
				detach_group_without_lock(group)
			else:
				group.size--
			mutate_count(count - 1)
		process_info = new ProcessInfo
		process_info.process = spawn(app_root, options)
		process_info.group_name = options.app_group_name
		group = new Group
		group.name = options.app_group_name
		group.app_root = app_root
		group.size = 1
		groups[options.app_group_name] = group
		iterator = group.processes.add_to_back(process_info)
		process_info.iterator = iterator
		mutate_count(count + 1)
		mutate_active(active + 1)
		if (options.min_processes > 1) and spawning_allowed(group, options):
			spawn_in_background(group, options)
	
	group.max_requests = options.max_requests
	group.min_processes = options.min_processes
	group.environment = options.environment
	
	process_info.last_used = current_time()
	process_info.sessions++
	
	return [process_info, group]


private function mutate_active(value):
	if (value < active):
		new_app_group_creatable.notify_all
		global_queue_position_became_available.notify_all
	active = value

private function mutate_count(value):
	# No point in notifying new_app_group_creatable here;
	# if _count_ is being increased then that means the pool
	# isn't full, and nobody is waiting on
	# new_app_group_creatable.
	global_queue_position_became_available.notify_all
	count = value

private function mutate_max(value):
	if (value > max):
		new_app_group_creatable.notify_all
		# We will want any code waiting on the global queue
		# to go ahead and spawn another process.
		global_queue_position_became_available.notify_all
	max = value


private function needs_restart(app_root, options):
	if (options.restart_dir is not set):
		restart_dir = app_root + "/tmp"
	else if (options.restart_dir is an absolute path):
		restart_dir = options.restart_dir
	else:
		restart_dir = app_root + "/" + options.restart_dir
	
	return (file_exists("$restart_dir/always_restart.txt")) or
	       (we haven't seen "$restart_dir/restart.txt" before) or
	       ("$restart_dir/restart.txt" changed since the last time we checked)


private function spawning_allowed(group, options):
	return ( count < max ) and
		( (max_per_app == 0) or (group.size < max_per_app) )


# Precondition: !group.detached
private function detach_group_without_lock(group):
	for all process_info in group.processes:
		if (process_info.sessions == 0):
			inactive_apps.remove(process_info.ia_iterator)
		else:
			mutate_active(active - 1)
		group.processes.remove(process_info.iterator)
		process_info.detached = true
		mutate_count(count - 1)
	if (group.spawning):
		group.spawner_thread.interrupt_and_join
		group.spawner_thread = null
		group.spawning = false
	group.detached = true
	groups.remove(options.app_group_name)


private function select_process(processes, options):
	if options.use_global_queue:
		# So we wait until _active_ has changed, then
		# we restart this function and try again.
		waiting_on_global_queue++
		global_queue_position_became_available.wait
		waiting_on_global_queue--
		return null
	else:
		# So we connect to an already active process.
		# This connection will be put into that
		# process's private queue.
		process_info = an element in _processes_ with the smallest _session_ value
		processes.move_to_back(process_info.iterator)
		return process_info


# Preconditions:
#   !group.detached
#   !group.spawning
private function spawn_in_background(group, options):
	group.spawning = true
	group.spawner_thread = new thread(spawner_thread_callback,
		with these arguments to the thread function:
		group, options)


private function spawner_thread_callback(group, options):
	Ignore thread interruptions in this function
	while true:
		try:
			Allow thread interruptions in this block
			process = spawn(app_root, options)
		on thread interruption:
			lock.synchronize:
				group.spawning = false
				group.spawner_thread = null
				return
		on exception:
			lock.synchronize:
				if (!group.detached):
					group.spawning = false
					group.spawner_thread = null
					# We want to report the error to the browser
					# but there's no way to do this in this thread, so
					# we just remove the entire group and have the next
					# get() call spawn the process and display the
					# error.
					remove_group_without_lock(group)
				return
		
		lock.synchronize:
			if (group.detached):
				return
			else:
				process_info = new ProcessInfo
				process_info.process = process
				process_info.group_name = options.app_group_name
				process_info.iterator = group.processes.add_to_front(process_info)
				process_info.ia_iterator = inactive_apps.add_to_back(process_info)
				group.size++
				mutate_count(count + 1)
				if (group.size >= options.min_processes) or
				   (!spawning_allowed(group, options)):
					group.spawning = false
					group.spawner_thread = null
					return


private function detach_without_lock(identifier):
	for group in groups:
		processes = group.processes
		for process_info in processes:
			if process_info.identifier == identifier:
				# Found a matching process.
				process_info.detached = true
				processes.remove(process_info.iterator)
				group.size--
				if processes.empty():
					detach_group_without_lock(group)
				if process_info.sessions == 0:
					inactive_apps.remove(process_info.ia_iterator)
				else:
					mutate_active(active - 1)
				mutate_count(count - 1)
				return true
	return false


# The following function is to be called when a session has been closed.
# _process_info_ is a weak reference to the ProcessInfo that belongs to
# the process whose session has been closed; it evaluates to NULL if the
# ProcessInfo object that it belongs to has been destroyed.
callback session_has_been_closed(session, process_info):
	Convert process_info into a normal reference.
	
	# We check process_info.detached without locking. This should be safe:
	# even if a boolean update isn't atomic on the current CPU, a non-zero
	# value evaluates to true. Once true, _detached_ will never become false,
	# so instruction reorderings by the compiler or CPU won't cause any
	# damage.
	if (process_info == null) or (process_info.detached):
		return
	
	lock.synchronize:
		if process_info.detached:
			return
		
		group = groups[process_info.group_name]
		processes = group.processes
		process_info.processed++
		
		if (group.max_requests > 0) and (process_info.processed >= group.max_requests):
			# The application process has processed its maximum allowed
			# number of requests, so we shut it down.
			process_info.detached = true
			processes.remove(process_info.iterator)
			group.size--
			if processes.empty():
				detach_group_without_lock(group)
			mutate_count(count - 1)
			if (process_info.sessions == 0):
				inactive_apps.remove(process_info.ia_iterator)
			else:
				mutate_active(active - 1)
		else:
			process_info.last_used = current_time()
			process_info.sessions--
			if (process_info.sessions == 0):
				processes.move_to_front(process_info.iterator)
				process_info.ia_iterator = inactive_apps.add_to_back(process_info)
				mutate_active(active - 1)


# The following thread will be responsible for cleaning up idle application
# process, i.e. processes that haven't been used for a while.
# This can be disabled per app when setting it's maxIdleTime to 0.
thread cleaner:
	lock.synchronize:
		while true:
			# If MAX_IDLE_TIME is 0 we don't clean up any processes,
			# giving us the option to persist the processes
			# forever unless it's killed in order to free up space
			# for another process.
			if (MAX_IDLE_TIME == 0):
				Wait until the thread has been signalled to quit
				  or until MAX_IDLE_TIME changed.
				if thread has been signalled to quit:
					return
				else:
					restart loop
			else:
				Wait until MAX_IDLE_TIME seconds have passed,
				  or until the thread has been signalled to quit,
				  or until MAX_IDLE_TIME changed.
				if thread has been signalled to quit:
					return
				else if MAX_IDLE_TIME changed:
					restart loop
			
			# Invariant:
			# From this point on, MAX_IDLE_TIME > 0
			
			now = current_time()
			for all process_info in inactive_apps:
				if (now - process_info.last_used > MAX_IDLE_TIME):
					process = process_info.process
					group = groups[process_info.group_name]
					if (group.size > group.min_processes):
						processes = group.processes
						processes.remove(process_info.iterator)
						process_info.detached = true
						inactive_apps.remove(process_info.ia_iterator)
						group.size--
						mutate_count(count - 1)
						if processes.empty():
							detach_group_without_lock(group)