The condition implementation is faster, and removes the undocumented
<wink> restriction of the former implementation that a signal/broadcast
could not be done by a signal/broadcast'er that had the condition mutex
locked.
And the test case is more fun: It implements a parallelized quicksort,
and runs a number of those in parallel, along with randomizing the input
arrays at the same time. It creates about 200 threads before it's done,
and may have as many as 100 simultaneously active. If that's too severe
for your system (it's nothing special here, but then we're in that
business <ahem>), reduce NSORTS.
BUG: The attached module worms around this, but the following self-
contained little program usually freezes with most threads reporting
Unhandled exception in thread:
Traceback (innermost last):
File "importbug.py", line 6
x = whrandom.randint(1,3)
AttributeError: randint
Here's the program; it doesn't use anything from the attached module:
import thread
def task():
global N
import whrandom
x = whrandom.randint(1,3)
a.acquire()
N = N - 1
if N == 0: done.release()
a.release()
a = thread.allocate_lock()
done = thread.allocate_lock()
N = 10
done.acquire()
for i in range(N):
thread.start_new_thread(task, ())
done.acquire()
print 'done'
Sticking an acquire/release pair around the 'import' statement makes the
problem go away.
I believe that what happens is:
1) The first thread to hit the import atomically reaches, and executes
most of, get_module. In particular, it finds Lib/whrandom.pyc,
installs its name in sys.modules, and executes
v = eval_code(co, d, d, d, (object *)NULL);
to initialize the module.
2) eval_code "ticker"-slices the 1st thread out, and gives another thread
a chance. When this 2nd thread hits the same 'import', import_module
finds 'whrandom' in sys.modules, so just proceeds.
3) But the 1st thread is still "in the middle" of executing whrandom.pyc.
So the 2nd thread has a good chance of trying to look up 'randint'
before the 1st thread has placed it in whrandom's dict.
4) The more threads there are, the more likely that at least one of them
will do this before the 1st thread finishes the import work.
If that's right, a perhaps not-too-bad workaround would be to introduce a
static "you can't interrupt this thread" flag in ceval.c, check it before
giving up interpreter_lock, and have IMPORT_NAME set it & restore (plain
clearing would not work) it around its call to import_module. To its
credit, there's something wonderfully perverse about fixing a race via an
unprotected static <grin>.
as-with-most-other-things-(pseudo-)parallel-programming's-more-fun-
in-python-too!-ly y'rs - tim
Tim Peters tim@ksr.com
not speaking for Kendall Square Research Corp
# Defines classes that provide synchronization objects. Note that use of
# this module requires that your Python support threads.
#
# condition() # a POSIX-like condition-variable object
# barrier(n) # an n-thread barrier
# event() # an event object
#
# CONDITIONS
#
# A condition object is created via
# import this_module
# your_condition_object = this_module.condition()
#
# Methods:
# .acquire()
# acquire the lock associated with the condition
# .release()
# release the lock associated with the condition
# .wait()
# block the thread until such time as some other thread does a
# .signal or .broadcast on the same condition, and release the
# lock associated with the condition. The lock associated with
# the condition MUST be in the acquired state at the time
# .wait is invoked.
# .signal()
# wake up exactly one thread (if any) that previously did a .wait
# on the condition; that thread will awaken with the lock associated
# with the condition in the acquired state. If no threads are
# .wait'ing, this is a nop. If more than one thread is .wait'ing on
# the condition, any of them may be awakened.
# .broadcast()
# wake up all threads (if any) that are .wait'ing on the condition;
# the threads are woken up serially, each with the lock in the
# acquired state, so should .release() as soon as possible. If no
# threads are .wait'ing, this is a nop.
#
# Note that if a thread does a .wait *while* a signal/broadcast is
# in progress, it's guaranteeed to block until a subsequenct
# signal/broadcast.
#
# Secret feature: `broadcast' actually takes an integer argument,
# and will wake up exactly that many waiting threads (or the total
# number waiting, if that's less). Use of this is dubious, though,
# and probably won't be supported if this form of condition is
# reimplemented in C.
#
# DIFFERENCES FROM POSIX
#
# + A separate mutex is not needed to guard condition data. Instead, a
# condition object can (must) be .acquire'ed and .release'ed directly.
# This eliminates a common error in using POSIX conditions.
#
# + Because of implementation difficulties, a POSIX `signal' wakes up
# _at least_ one .wait'ing thread. Race conditions make it difficult
# to stop that. This implementation guarantees to wake up only one,
# but you probably shouldn't rely on that.
#
# PROTOCOL
#
# Condition objects are used to block threads until "some condition" is
# true. E.g., a thread may wish to wait until a producer pumps out data
# for it to consume, or a server may wish to wait until someone requests
# its services, or perhaps a whole bunch of threads want to wait until a
# preceding pass over the data is complete. Early models for conditions
# relied on some other thread figuring out when a blocked thread's
# condition was true, and made the other thread responsible both for
# waking up the blocked thread and guaranteeing that it woke up with all
# data in a correct state. This proved to be very delicate in practice,
# and gave conditions a bad name in some circles.
#
# The POSIX model addresses these problems by making a thread responsible
# for ensuring that its own state is correct when it wakes, and relies
# on a rigid protocol to make this easy; so long as you stick to the
# protocol, POSIX conditions are easy to "get right":
#
# A) The thread that's waiting for some arbitrarily-complex condition
# (ACC) to become true does:
#
# condition.acquire()
# while not (code to evaluate the ACC):
# condition.wait()
# # That blocks the thread, *and* releases the lock. When a
# # condition.signal() happens, it will wake up some thread that
# # did a .wait, *and* acquire the lock again before .wait
# # returns.
# #
# # Because the lock is acquired at this point, the state used
# # in evaluating the ACC is frozen, so it's safe to go back &
# # reevaluate the ACC.
#
# # At this point, ACC is true, and the thread has the condition
# # locked.
# # So code here can safely muck with the shared state that
# # went into evaluating the ACC -- if it wants to.
# # When done mucking with the shared state, do
# condition.release()
#
# B) Threads that are mucking with shared state that may affect the
# ACC do:
#
# condition.acquire()
# # muck with shared state
# condition.release()
# if it's possible that ACC is true now:
# condition.signal() # or .broadcast()
#
# Note: You may prefer to put the "if" clause before the release().
# That's fine, but do note that anyone waiting on the signal will
# stay blocked until the release() is done (since acquiring the
# condition is part of what .wait() does before it returns).
#
# TRICK OF THE TRADE
#
# With simpler forms of conditions, it can be impossible to know when
# a thread that's supposed to do a .wait has actually done it. But
# because this form of condition releases a lock as _part_ of doing a
# wait, the state of that lock can be used to guarantee it.
#
# E.g., suppose thread A spawns thread B and later wants to wait for B to
# complete:
#
# In A: In B:
#
# B_done = condition() ... do work ...
# B_done.acquire() B_done.acquire(); B_done.release()
# spawn B B_done.signal()
# ... some time later ... ... and B exits ...
# B_done.wait()
#
# Because B_done was in the acquire'd state at the time B was spawned,
# B's attempt to acquire B_done can't succeed until A has done its
# B_done.wait() (which releases B_done). So B's B_done.signal() is
# guaranteed to be seen by the .wait(). Without the lock trick, B
# may signal before A .waits, and then A would wait forever.
#
# BARRIERS
#
# A barrier object is created via
# import this_module
# your_barrier = this_module.barrier(num_threads)
#
# Methods:
# .enter()
# the thread blocks until num_threads threads in all have done
# .enter(). Then the num_threads threads that .enter'ed resume,
# and the barrier resets to capture the next num_threads threads
# that .enter it.
#
# EVENTS
#
# An event object is created via
# import this_module
# your_event = this_module.event()
#
# An event has two states, `posted' and `cleared'. An event is
# created in the cleared state.
#
# Methods:
#
# .post()
# Put the event in the posted state, and resume all threads
# .wait'ing on the event (if any).
#
# .clear()
# Put the event in the cleared state.
#
# .is_posted()
# Returns 0 if the event is in the cleared state, or 1 if the event
# is in the posted state.
#
# .wait()
# If the event is in the posted state, returns immediately.
# If the event is in the cleared state, blocks the calling thread
# until the event is .post'ed by another thread.
#
# Note that an event, once posted, remains posted until explicitly
# cleared. Relative to conditions, this is both the strength & weakness
# of events. It's a strength because the .post'ing thread doesn't have to
# worry about whether the threads it's trying to communicate with have
# already done a .wait (a condition .signal is seen only by threads that
# do a .wait _prior_ to the .signal; a .signal does not persist). But
# it's a weakness because .clear'ing an event is error-prone: it's easy
# to mistakenly .clear an event before all the threads you intended to
# see the event get around to .wait'ing on it. But so long as you don't
# need to .clear an event, events are easy to use safely.
import thread
class condition:
def __init__(self):
# the lock actually used by .acquire() and .release()
self.mutex = thread.allocate_lock()
# lock used to block threads until a signal
self.checkout = thread.allocate_lock()
self.checkout.acquire()
# internal critical-section lock, & the data it protects
self.idlock = thread.allocate_lock()
self.id = 0
self.waiting = 0 # num waiters subject to current release
self.pending = 0 # num waiters awaiting next signal
self.torelease = 0 # num waiters to release
self.releasing = 0 # 1 iff release is in progress
def acquire(self):
self.mutex.acquire()
def release(self):
self.mutex.release()
def wait(self):
mutex, checkout, idlock = self.mutex, self.checkout, self.idlock
if not mutex.locked():
raise ValueError, \
"condition must be .acquire'd when .wait() invoked"
idlock.acquire()
myid = self.id
self.pending = self.pending + 1
idlock.release()
mutex.release()
while 1:
checkout.acquire(); idlock.acquire()
if myid < self.id:
break
checkout.release(); idlock.release()
self.waiting = self.waiting - 1
self.torelease = self.torelease - 1
if self.torelease:
checkout.release()
else:
self.releasing = 0
if self.waiting == self.pending == 0:
self.id = 0
idlock.release()
mutex.acquire()
def signal(self):
self.broadcast(1)
def broadcast(self, num = -1):
if num < -1:
raise ValueError, '.broadcast called with num ' + `num`
if num == 0:
return
self.idlock.acquire()
if self.pending:
self.waiting = self.waiting + self.pending
self.pending = 0
self.id = self.id + 1
if num == -1:
self.torelease = self.waiting
else:
self.torelease = min( self.waiting,
self.torelease + num )
if self.torelease and not self.releasing:
self.releasing = 1
self.checkout.release()
self.idlock.release()
class barrier:
def __init__(self, n):
self.n = n
self.togo = n
self.full = condition()
def enter(self):
full = self.full
full.acquire()
self.togo = self.togo - 1
if self.togo:
full.wait()
else:
self.togo = self.n
full.broadcast()
full.release()
class event:
def __init__(self):
self.state = 0
self.posted = condition()
def post(self):
self.posted.acquire()
self.state = 1
self.posted.broadcast()
self.posted.release()
def clear(self):
self.posted.acquire()
self.state = 0
self.posted.release()
def is_posted(self):
self.posted.acquire()
answer = self.state
self.posted.release()
return answer
def wait(self):
self.posted.acquire()
while not self.state:
self.posted.wait()
self.posted.release()
# The rest of the file is a test case, that runs a number of parallelized
# quicksorts in parallel. If it works, you'll get about 600 lines of
# tracing output, with a line like
# test passed! 209 threads created in all
# as the last line. The content and order of preceding lines will
# vary across runs.
def _new_thread(func, *args):
global TID
tid.acquire(); id = TID = TID+1; tid.release()
io.acquire(); alive.append(id); \
print 'starting thread', id, '--', len(alive), 'alive'; \
io.release()
thread.start_new_thread( func, (id,) + args )
def _qsort(tid, a, l, r, finished):
# sort a[l:r]; post finished when done
io.acquire(); print 'thread', tid, 'qsort', l, r; io.release()
if r-l > 1:
pivot = a[l]
j = l+1 # make a[l:j] <= pivot, and a[j:r] > pivot
for i in range(j, r):
if a[i] <= pivot:
a[j], a[i] = a[i], a[j]
j = j + 1
a[l], a[j-1] = a[j-1], pivot
l_subarray_sorted = event()
r_subarray_sorted = event()
_new_thread(_qsort, a, l, j-1, l_subarray_sorted)
_new_thread(_qsort, a, j, r, r_subarray_sorted)
l_subarray_sorted.wait()
r_subarray_sorted.wait()
io.acquire(); print 'thread', tid, 'qsort done'; \
alive.remove(tid); io.release()
finished.post()
def _randarray(tid, a, finished):
io.acquire(); print 'thread', tid, 'randomizing array'; \
io.release()
for i in range(1, len(a)):
wh.acquire(); j = randint(0,i); wh.release()
a[i], a[j] = a[j], a[i]
io.acquire(); print 'thread', tid, 'randomizing done'; \
alive.remove(tid); io.release()
finished.post()
def _check_sort(a):
if a != range(len(a)):
raise ValueError, ('a not sorted', a)
def _run_one_sort(tid, a, bar, done):
# randomize a, and quicksort it
# for variety, all the threads running this enter a barrier
# at the end, and post `done' after the barrier exits
io.acquire(); print 'thread', tid, 'randomizing', a; \
io.release()
finished = event()
_new_thread(_randarray, a, finished)
finished.wait()
io.acquire(); print 'thread', tid, 'sorting', a; io.release()
finished.clear()
_new_thread(_qsort, a, 0, len(a), finished)
finished.wait()
_check_sort(a)
io.acquire(); print 'thread', tid, 'entering barrier'; \
io.release()
bar.enter()
io.acquire(); print 'thread', tid, 'leaving barrier'; \
io.release()
io.acquire(); alive.remove(tid); io.release()
bar.enter() # make sure they've all removed themselves from alive
## before 'done' is posted
bar.enter() # just to be cruel
done.post()
def test():
global TID, tid, io, wh, randint, alive
import whrandom
randint = whrandom.randint
TID = 0 # thread ID (1, 2, ...)
tid = thread.allocate_lock() # for changing TID
io = thread.allocate_lock() # for printing, and 'alive'
wh = thread.allocate_lock() # for calls to whrandom
alive = [] # IDs of active threads
NSORTS = 5
arrays = []
for i in range(NSORTS):
arrays.append( range( (i+1)*10 ) )
bar = barrier(NSORTS)
finished = event()
for i in range(NSORTS):
_new_thread(_run_one_sort, arrays[i], bar, finished)
finished.wait()
print 'all threads done, and checking results ...'
if alive:
raise ValueError, ('threads still alive at end', alive)
for i in range(NSORTS):
a = arrays[i]
if len(a) != (i+1)*10:
raise ValueError, ('length of array', i, 'screwed up')
_check_sort(a)
print 'test passed!', TID, 'threads created in all'
if __name__ == '__main__':
test()
# end of module