Charles-François Natali <neolo...@free.fr> added the comment:

Here's a patch based on the second approach.

----------
Added file: http://bugs.python.org/file22490/heap_gc_deadlock_lockless.diff

_______________________________________
Python tracker <rep...@bugs.python.org>
<http://bugs.python.org/issue12352>
_______________________________________
diff -r fca745bc70be Lib/multiprocessing/heap.py
--- a/Lib/multiprocessing/heap.py       Sat Jun 25 16:31:06 2011 +0200
+++ b/Lib/multiprocessing/heap.py       Sun Jun 26 19:57:56 2011 +0200
@@ -101,6 +101,8 @@
         self._stop_to_block = {}
         self._allocated_blocks = set()
         self._arenas = []
+        # list of pending blocks to free - see free() comment below
+        self._pending_free_blocks = []
 
     @staticmethod
     def _roundup(n, alignment):
@@ -175,15 +177,39 @@
 
         return start, stop
 
+    def _free_pending_blocks(self):
+        # Free all the blocks in the pending list - called with the lock held.
+        while True:
+            try:
+                block = self._pending_free_blocks.pop()
+            except IndexError:
+                break
+            self._free(block)
+            self._allocated_blocks.remove(block)
+
     def free(self, block):
         # free a block returned by malloc()
+        # Since free() can be called asynchronously by the GC, it could happen
+        # that it's called while self._lock is held: in that case,
+        # self._lock.acquire() would deadlock (issue #12352). To avoid that, a
+        # trylock is used instead, and if the lock can't be acquired
+        # immediately, the block is added to a list of blocks to be freed
+        # synchronously sometimes later from malloc() or free(), by calling
+        # _free_pending_blocks() (appending and retrieving from a list is not
+        # strictly thread-safe but under cPython it's atomic thanks to the 
GIL).
         assert os.getpid() == self._lastpid
-        self._lock.acquire()
-        try:
-            self._allocated_blocks.remove(block)
-            self._free(block)
-        finally:
-            self._lock.release()
+        if not self._lock.acquire(0):
+            # can't acquire the lock, add it to the list of pending blocks to
+            # free
+            self._pending_free_blocks.append(block)
+        else:
+            # we hold the lock
+            try:
+                self._free_pending_blocks()
+                self._allocated_blocks.remove(block)
+                self._free(block)
+            finally:
+                self._lock.release()
 
     def malloc(self, size):
         # return a block of right size (possibly rounded up)
@@ -191,6 +217,7 @@
         if os.getpid() != self._lastpid:
             self.__init__()                     # reinitialize after fork
         self._lock.acquire()
+        self._free_pending_blocks()
         try:
             size = self._roundup(max(size,1), self._alignment)
             (arena, start, stop) = self._malloc(size)
diff -r fca745bc70be Lib/test/test_multiprocessing.py
--- a/Lib/test/test_multiprocessing.py  Sat Jun 25 16:31:06 2011 +0200
+++ b/Lib/test/test_multiprocessing.py  Sun Jun 26 19:57:56 2011 +0200
@@ -1737,7 +1737,31 @@
             (narena, nstart, nstop) = all[i+1][:3]
             self.assertTrue((arena != narena and nstart == 0) or
                             (stop == nstart))
-
+                            
+    def test_free_from_gc(self):
+        # Check that freeing of blocks by the garbage collector doesn't 
deadlock
+        # (issue #12352).
+        # Make sure the GC is enabled, and set lower collection thresholds to
+        # make collections more frequent (and increase the probability of
+        # deadlock).
+        if gc.isenabled():
+            thresholds = gc.get_threshold()
+            self.addCleanup(gc.set_threshold, *thresholds)
+        else:
+            gc.enable()
+            self.addCleanup(gc.disable)
+        gc.set_threshold(10)
+
+        # perform numerous block allocations, with cyclic references to make
+        # sure objects are collected asynchronously by the gc
+        for i in range(5000):
+            a = multiprocessing.heap.BufferWrapper(1)
+            b = multiprocessing.heap.BufferWrapper(1)
+            # circular references
+            a.buddy = b
+            b.buddy = a
+
+ 
 #
 #
 #
_______________________________________________
Python-bugs-list mailing list
Unsubscribe: 
http://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

Reply via email to