source: trunk/src/gcc/libobjc/THREADS@ 2101

This file describes in little detail the modifications to the
Objective-C runtime needed to make it thread safe. 

First off, kudos to Galen Hunt who is the author of this great work.

If you have an comments or just want to know where to
send me money to express your undying gratitude for threading the
Objective-C runtime you can reach Galen at:

        gchunt@cs.rochester.edu

Any questions, comments, bug reports, etc. should send email either to the
GCC bug account or to:

        Scott Christley <scottc@net-community.com>

* Sarray Threading:

The most critical component of the Objective-C runtime is the sparse array
structure (sarray).  Sarrays store object selectors and implementations.  
Following in the tradition of the Objective-C runtime, my threading
support assumes that fast message dispatching is far more important
than *ANY* and *ALL* other operations.  The message dispatching thus
uses *NO* locks on any kind.  In fact, if you look in sarray.h, you
will notice that the message dispatching has not been modified.
Instead, I have modified the sarray management functions so that all
updates to the sarray data structure can be made in parallel will
message dispatching.  

To support concurrent message dispatching, no dynamically allocated
sarray data structures are freed while more than one thread is
operational.  Sarray data structures that are no longer in use are
kept in a linked list of garbage and are released whenever the program
is operating with a single thread.  The programmer can also flush the 
garbage list by calling sarray_remove_garbage when the programmer can
ensure that no message dispatching is taking place concurrently.  The
amount of un-reclaimed sarray garbage should normally be extremely
small in a real program as sarray structures are freed only when using
the "poseAs" functionality and early in program initialization, which
normally occurs while the program is single threaded.

******************************************************************************
* Static Variables:

The following variables are either statically or globally defined. This list 
does not include variables which are internal to implementation dependent 
versions of thread-*.c.

The following threading designations are used:
        SAFE   : Implicitly thread safe.
        SINGLE : Must only be used in single thread mode.
        MUTEX  : Protected by single global mutex objc_runtime_mutex.
        UNUSED : Not used in the runtime.

Variable Name:                  Usage:  Defined:        Also used in:
===========================     ======  ============    =====================
__objc_class_hash               MUTEX   class.c
__objc_class_links_resolved     UNUSED  class.c         runtime.h
__objc_class_number             MUTEX   class.c
__objc_dangling_categories      UNUSED  init.c
__objc_module_list              MUTEX   init.c
__objc_selector_array           MUTEX   selector.c
__objc_selector_hash            MUTEX   selector.c
__objc_selector_max_index       MUTEX   selector.c      sendmsg.c runtime.h
__objc_selector_names           MUTEX   selector.c
__objc_thread_exit_status       SAFE    thread.c
__objc_uninstalled_dtable       MUTEX   sendmsg.c       selector.c
_objc_load_callback             SAFE    init.c          objc-api.h
_objc_lookup_class              SAFE    class.c         objc-api.h
_objc_object_alloc              SINGLE  objects.c       objc-api.h
_objc_object_copy               SINGLE  objects.c       objc-api.h
_objc_object_dispose            SINGLE  objects.c       objc-api.h
frwd_sel                        SAFE2   sendmsg.c
idxsize                         MUTEX   sarray.c        sendmsg.c sarray.h
initialize_sel                  SAFE2   sendmsg.c
narrays                         MUTEX   sarray.c        sendmsg.c sarray.h
nbuckets                        MUTEX   sarray.c        sendmsg.c sarray.h
nindices                        MUTEX   sarray.c        sarray.h
previous_constructors           SAFE1   init.c
proto_class                     SAFE1   init.c
unclaimed_categories            MUTEX   init.c
unclaimed_proto_list            MUTEX   init.c
uninitialized_statics           MUTEX   init.c

Notes:
1) Initialized once in unithread mode.
2) Initialized value will always be same, guaranteed by lock on selector 
   hash table.


******************************************************************************
* Frontend/Backend design:

The design of the Objective-C runtime thread and mutex functions utilizes a
frontend/backend implementation.

The frontend, as characterized by the files thr.h and thr.c, is a set
of platform independent structures and functions which represent the
user interface.  Objective-C programs should use these structures and
functions for their thread and mutex work if they wish to maintain a
high degree of portability across platforms.

The backend is composed of a file with the necessary code to map the ObjC
thread and mutex to a platform specific implementation.  For example, the
file thr-solaris.c contains the implementation for Solaris.

If you are compiling libobjc as part of GCC, the thr-objc.c backend is
always used; this backend uses GCC's gthread code.  The thread system
is automatically configured when GCC is configured.  Important: make
sure you configure GCC using `--enable-threads' if you want threads !
  
If you want to compile libobjc standalone, then you would need to
modify the configure.in and makefiles for it; and you need to pick an
appropriate backend file for the target platform; you make this choice
by assigning the OBJC_THREAD_FILE make variable to the basename of the
backend file.  For example, OBJC_THREAD_FILE=thr-posix would indicate
that the generic posix backend file, thr-posix.c, should be compiled
with the ObjC runtime library.  If your platform does not support
threads then you should specify the OBJC_THREAD_FILE=thr-single
backend file to compile the ObjC runtime library without thread or
mutex support; note that programs which rely upon the ObjC thread and
mutex functions will compile and link correctly but attempting to
create a thread or mutex will result in an error.
  
It is questionable whether it is really necessary to have both a
frontend and backend function for all available functionality.  On the
one hand, it provides a clear, consistent differentiation between what
is public and what is private with the downside of having the overhead
of multiple functions calls.  For example, the function to have a
thread yield the processor is objc_thread_yield; in the current
implementation this produces a function call set:

objc_thread_yield()  ->  __objc_thread_yield()  ->  system yield function

This has two extra function calls over calling the platform specific function
explicitly, but the issue is whether only the overhead of a single function
is necessary.

objc_thread_yield()  ->  system yield function

This breaks the public/private dichotomy between the frontend/backend
for the sake of efficiency.  It is possible to just use a preprocessor
define so as to eliminate the extra function call:

#define objc_thread_yield() __objc_thread_yield()

This has the undesirable effect that if objc_thread_yield is actually
turned into a function based upon future need; then ObjC programs which
access the thread functions would need to be recompiled versus just
being relinked.
 
******************************************************************************
* Threads:

The thread system attempts to create multiple threads using whatever
operating system or library thread support is available.  It does
assume that all system functions are thread safe.  Notably this means
that the system implementation of malloc and free must be thread safe.
If a system has multiple processors, the threads are configured for
full parallel processing.

* Backend initialization functions

__objc_init_thread_system(void), int
        Initialize the thread subsystem.  Called once by __objc_exec_class.
        Return -1 if error otherwise return 0.

__objc_close_thread_system(void), int
        Closes the thread subsystem, not currently guaranteed to be called.
        Return -1 if error otherwise return 0.

*****
* Frontend thread functions
* User programs should use these functions.

objc_thread_detach(SEL selector, id object, id argument), objc_thread_t
        Creates and detaches a new thread.  The new thread starts by
        sending the given selector with a single argument to the
        given object.

objc_thread_set_priority(int priority), int
        Sets a thread's relative priority within the program.  Valid
        options are:
        
        OBJC_THREAD_INTERACTIVE_PRIORITY
        OBJC_THREAD_BACKGROUND_PRIORITY
        OBJC_THREAD_LOW_PRIORITY

objc_thread_get_priority(void), int
        Query a thread's priority.

objc_thread_yield(void), void
        Yields processor to another thread with equal or higher
        priority.  It is up to the system scheduler to determine if
        the processor is taken or not.

objc_thread_exit(void), int
        Terminates a thread.  If this is the last thread executing
        then the program will terminate.

objc_thread_id(void), int
        Returns the current thread's id.

objc_thread_set_data(void *value), int
        Set a pointer to the thread's local storage.  Local storage is
        thread specific.

objc_thread_get_data(void), void *
        Returns the pointer to the thread's local storage.

*****
* Backend thread functions
* User programs should *NOT* directly call these functions.

__objc_thread_detach(void (*func)(void *arg), void *arg), objc_thread_t
        Spawns a new thread executing func, called by objc_thread_detach.
        Return NULL if error otherwise return thread id.

__objc_thread_set_priority(int priority), int
        Set the thread's priority, called by objc_thread_set_priority.
        Return -1 if error otherwise return 0.

__objc_thread_get_priority(void), int
        Query a thread's priority, called by objc_thread_get_priority.
        Return -1 if error otherwise return the priority.

__objc_thread_yield(void), void
        Yields the processor, called by objc_thread_yield.

__objc_thread_exit(void), int
        Terminates the thread, called by objc_thread_exit.
        Return -1 if error otherwise function does not return.

__objc_thread_id(void), objc_thread_t
        Returns the current thread's id, called by objc_thread_id.
        Return -1 if error otherwise return thread id.

__objc_thread_set_data(void *value), int
        Set pointer for thread local storage, called by objc_thread_set_data.
        Returns -1 if error otherwise return 0.

__objc_thread_get_data(void), void *
        Returns the pointer to the thread's local storage.
        Returns NULL if error, called by objc_thread_get_data.


******************************************************************************
* Mutexes:

Mutexes can be locked recursively.  Each locked mutex remembers
its owner (by thread id) and how many times it has been locked.  The
last unlock on a mutex removes the system lock and allows other
threads to access the mutex.

*****
* Frontend mutex functions
* User programs should use these functions.

objc_mutex_allocate(void), objc_mutex_t
        Allocates a new mutex.  Mutex is initially unlocked.
        Return NULL if error otherwise return mutex pointer.

objc_mutex_deallocate(objc_mutex_t mutex), int
        Free a mutex.  Before freeing the mutex, makes sure that no
        one else is using it.
        Return -1 if error otherwise return 0.

objc_mutex_lock(objc_mutex_t mutex), int
        Locks a mutex.  As mentioned earlier, the same thread may call
        this routine repeatedly.
        Return -1 if error otherwise return 0.
        
objc_mutex_trylock(objc_mutex_t mutex), int
        Attempts to lock a mutex.  If lock on mutex can be acquired 
        then function operates exactly as objc_mutex_lock.
        Return -1 if failed to acquire lock otherwise return 0.

objc_mutex_unlock(objc_mutex_t mutex), int
        Unlocks the mutex by one level.  Other threads may not acquire
        the mutex until this thread has released all locks on it.
        Return -1 if error otherwise return 0.

*****
* Backend mutex functions
* User programs should *NOT* directly call these functions.

__objc_mutex_allocate(objc_mutex_t mutex), int
        Allocates a new mutex, called by objc_mutex_allocate.
        Return -1 if error otherwise return 0.

__objc_mutex_deallocate(objc_mutex_t mutex), int
        Free a mutex, called by objc_mutex_deallocate.
        Return -1 if error otherwise return 0.

__objc_mutex_lock(objc_mutex_t mutex), int
        Locks a mutex, called by objc_mutex_lock.
        Return -1 if error otherwise return 0.
        
__objc_mutex_trylock(objc_mutex_t mutex), int
        Attempts to lock a mutex, called by objc_mutex_trylock.
        Return -1 if failed to acquire lock or error otherwise return 0.

__objc_mutex_unlock(objc_mutex_t mutex), int
        Unlocks the mutex, called by objc_mutex_unlock.
        Return -1 if error otherwise return 0.

******************************************************************************
* Condition Mutexes:

Mutexes can be locked recursively.  Each locked mutex remembers
its owner (by thread id) and how many times it has been locked.  The
last unlock on a mutex removes the system lock and allows other
threads to access the mutex.

*
* Frontend condition mutex functions
* User programs should use these functions.
*

objc_condition_allocate(void), objc_condition_t 
        Allocate a condition mutex.
        Return NULL if error otherwise return condition pointer.

objc_condition_deallocate(objc_condition_t condition), int
        Deallocate a condition. Note that this includes an implicit
        condition_broadcast to insure that waiting threads have the 
        opportunity to wake.  It is legal to dealloc a condition only
        if no other thread is/will be using it. Does NOT check for
        other threads waiting but just wakes them up.
        Return -1 if error otherwise return 0.

objc_condition_wait(objc_condition_t condition, objc_mutex_t mutex), int
        Wait on the condition unlocking the mutex until objc_condition_signal()
        or objc_condition_broadcast() are called for the same condition. The
        given mutex *must* have the depth 1 so that it can be unlocked
        here, for someone else can lock it and signal/broadcast the condition.
        The mutex is used to lock access to the shared data that make up the
        "condition" predicate.
        Return -1 if error otherwise return 0.
        
objc_condition_broadcast(objc_condition_t condition), int
        Wake up all threads waiting on this condition. It is recommended that 
        the called would lock the same mutex as the threads in
        objc_condition_wait before changing the "condition predicate"
        and make this call and unlock it right away after this call.
        Return -1 if error otherwise return 0.

objc_condition_signal(objc_condition_t condition), int
        Wake up one thread waiting on this condition.
        Return -1 if error otherwise return 0.

*
* Backend condition mutex functions
* User programs should *NOT* directly call these functions.
*

__objc_condition_allocate(objc_condition_t condition), int
        Allocate a condition mutex, called by objc_condition_allocate.
        Return -1 if error otherwise return 0.

__objc_condition_deallocate(objc_condition_t condition), int
        Deallocate a condition, called by objc_condition_deallocate.
        Return -1 if error otherwise return 0.

__objc_condition_wait(objc_condition_t condition, objc_mutex_t mutex), int
        Wait on the condition, called by objc_condition_wait.
        Return -1 if error otherwise return 0 when condition is met.
        
__objc_condition_broadcast(objc_condition_t condition), int
        Wake up all threads waiting on this condition.
        Called by objc_condition_broadcast.
        Return -1 if error otherwise return 0.

__objc_condition_signal(objc_condition_t condition), int
        Wake up one thread waiting on this condition.
        Called by objc_condition_signal.
        Return -1 if error otherwise return 0.