Demo/threads/bug.py

   1 # The following self-contained little program usually freezes with most
   2 # threads reporting
   3 #
   4 # Unhandled exception in thread:
   5 # Traceback (innermost last):
   6 #   File "importbug.py", line 6
   7 #     x = whrandom.randint(1,3)
   8 # AttributeError: randint
   9 #
  10 # Here's the program; it doesn't use anything from the attached module:
  11
  12 import thread
  13
  14 def task():
  15     global N
  16     import whrandom
  17     x = whrandom.randint(1,3)
  18     a.acquire()
  19     N = N - 1
  20     if N == 0: done.release()
  21     a.release()
  22
  23 a = thread.allocate_lock()
  24 done = thread.allocate_lock()
  25 N = 10
  26
  27 done.acquire()
  28 for i in range(N):
  29     thread.start_new_thread(task, ())
  30 done.acquire()
  31 print 'done'
  32
  33
  34 # Sticking an acquire/release pair around the 'import' statement makes the
  35 # problem go away.
  36 #
  37 # I believe that what happens is:
  38 #
  39 # 1) The first thread to hit the import atomically reaches, and executes
  40 #    most of, get_module.  In particular, it finds Lib/whrandom.pyc,
  41 #    installs its name in sys.modules, and executes
  42 #
  43 #         v = eval_code(co, d, d, d, (object *)NULL);
  44 #
  45 #    to initialize the module.
  46 #
  47 # 2) eval_code "ticker"-slices the 1st thread out, and gives another thread
  48 #    a chance.  When this 2nd thread hits the same 'import', import_module
  49 #    finds 'whrandom' in sys.modules, so just proceeds.
  50 #
  51 # 3) But the 1st thread is still "in the middle" of executing whrandom.pyc.
  52 #    So the 2nd thread has a good chance of trying to look up 'randint'
  53 #    before the 1st thread has placed it in whrandom's dict.
  54 #
  55 # 4) The more threads there are, the more likely that at least one of them
  56 #    will do this before the 1st thread finishes the import work.
  57 #
  58 # If that's right, a perhaps not-too-bad workaround would be to introduce a
  59 # static "you can't interrupt this thread" flag in ceval.c, check it before
  60 # giving up interpreter_lock, and have IMPORT_NAME set it & restore (plain
  61 # clearing would not work) it around its call to import_module.  To its
  62 # credit, there's something wonderfully perverse about fixing a race via an
  63 # unprotected static <grin>.
  64 #
  65 # as-with-most-other-things-(pseudo-)parallel-programming's-more-fun-
  66 #    in-python-too!-ly y'rs  - tim
  67 #
  68 # Tim Peters   tim@ksr.com
  69 # not speaking for Kendall Square Research Corp