Re: srfi-18 requirements

[Julian Graham]

Hi guys,

Just wanted to see if anyone had had a chance to take a look at the
patch I attached to the last message on this thread.  Once again, I
know it's big for a patch (I'd be glad to break it into a few smaller
patches), but about half of it is Scheme code.  I'd be happy to
explain / discuss any aspect of it!


Regards,
Julian


On Oct 29, 2007 9:37 AM, Julian Graham <address@hidden> wrote:
> Hi Guilers,
>
>   Find attached a first draft of a patch to add SRFI-18 support to
> Guile.  The patch contains the necessary modifications to Guile's C
> code to support SRFI-18, which is provided as a Scheme module (also
> attached). I don't believe any breaking changes to the API are
> introduced by this code, and, in general, the only behavior that is
> significantly different is that which was previously "unspecified" or
> of dubious correctness -- e.g., unlocking mutexes from outside the
> threads that originally locked them, or leaving abandoned mutexes
> locked forever.
>
>   I realize this is kind of a big patch, so I've included some, uh,
> commentary to help you guys sift through it.  I'm working on updating
> the Texinfo documentation, but didn't want to wait to submit, as the
> scope and content of this patch could conceivably change. I'm also
> attaching some test code that might be helpful in illustrating some of
> the concepts involved in SRFI-18. Of course, I also encourage everyone
> to take a look at the SRFI itself, which is available at
> http://srfi.schemers.org/srfi-18/srfi-18.html.
>
>   Please let me know if you have any questions / comments!
>
>
> Best Regards,
> Julian
>
>
> SIGNIFICANT CHANGES TO THREADS.H
> ================================
>
> * Three members have been added to the thread data structure:
>
> ** admin_mutex, which is a lighter-weight version of the
> thread_admin_mutex, which should be used instead of thread_admin_mutex
> in situations requiring synchronous access to thread-specific
> information, such as checking whether a thread has been marked exited,
> that do not require access to the global list of threads
>
> ** mutexes, a SCM list of mutexes owned by a thread.  This is
> necessary so that threads waiting on mutexes that an exiting thread is
> abandoning can be notified
>
> ** exception, a SCM holding any uncaught exceptions thrown by the
> thread or its cleanup handler.  We need this so we can deliver
> uncaught exceptions to threads that join on a terminated thread
>
>
> NEW C FUNCTIONS
> ===============
>
> * scm_join_thread_timed (t, timeout, timeout_val): An overload of
> scm_join_thread featuring extended timeout semantics from SRFI-18
> * scm_thread_p (obj): A type predicate for threads
> * scm_lock_mutex_timed (m, abstime, thread): An overload of
> scm_lock_mutex featuring extended timeout and ownership semantics from
> SRFI-18
> * scm_unlock_mutex_timed (m, cond, abstime): An overload of
> scm_unlock_mutex featuring extended timeout semantics from SRFI-18
> * scm_mutex_state (m): Mutex state reporting from SRFI-18
> * scm_mutex_p (obj): A type predicate for mutexes
> * scm_condition_variable_p (obj): A type predicate for condition variables
>
>
> NEW SCHEME FUNCTIONS (without loading SRFI-18)
> ==============================================
>
> * join-thread thread -> join-thread thread [timeout [timeoutval]]
> * thread? obj
> * lock-mutex mutex -> lock-mutex mutex [timeout [owner]]
> * unlock-mutex mutex -> unlock-mutex mutex [cond [timeout]]
> * mutex-state mutex
> * mutex? obj
> * condition-variable? obj
>
>
> SIGNIFICANT CHANGES TO THREADS.C
> ================================
>
> * A static function, scm_to_timespec, has been added to thread.c that
> converts SCM timeouts -- either in numerical form or as (secs . usecs)
> -- to an scm_t_timespec
>
> * Because the owner of a locked mutex can be #f, unblock_from_queue
> now returns SCM_UNDEFINED when a queue is empty, in order to avoid
> ambiguity.  This is purely for consistency -- #f cannot actually be
> added to a queue
>
> * The catch calls in do_thread_exit and really_launch now include a
> handler (if the caller does not specify one already),
> exception_preserve_catch_handler, which saves exception data in the
> thread's exception field
>
> * scm_join_thread, which now calls scm_join_thread_timed, will rethrow
> any uncaught exceptions thrown by the terminated thread
>
> * fat_mutex_lock and fat_mutex_unlock now add and remove mutexes from
> a thread's list of locked mutexes. As part of thread exit, other
> threads waiting on mutexes in this list are woken up
>
> * An unlocked fat_mutex now has its owner set to SCM_UNDEFINED, not SCM_BOOL_F
>
> * fat_mutex_lock has been largely rewritten (although the behavior is
> almost exactly the same for calls that do not specify a timeout) in
> order to deal with SRFI-18's concept of mutex states. It is now a loop
> that synchronously inspects the current state of the mutex to
> determine whether it can be locked, sleeping on the mutex's condition
> variable if it cannot be. As per SRFI-18, an attempt to lock an
> abandoned mutex will succeed, although an abandoned-mutex-exception
> will be thrown. fat_mutex_lock now returns an exception instead of a
> char *, indicating the success of the lock via a passed-in int
> pointer.
>
> * fat_mutex_unlock now allows a mutex to be unlocked from any thread,
> as per SRFI-18
>
>
> SIGNIFICANT CHANGES TO SCMSIGS.C
> ================================
>
> * The scm_spawn_thread call that launches the signal delivery thread
> no longer specifies a handler. No one should call scm_spawn_thread
> with a handler, because of an already-present deadlock in 1.8.x -- in
> a multithreaded context, a guile mode thread (i.e., one that has
> locked its heap mutex) may attempt to enter a critical section in
> eval.i.c at the same time a non-guile mode thread created by
> scm_spawn_thread is within a critical section in make_jmpbuf while
> setting up a catch handler and attempts to do a GC
>