Author: Armin Rigo <[email protected]>
Branch:
Changeset: r239:fe90ba781e8a
Date: 2013-06-22 18:30 +0200
http://bitbucket.org/pypy/stmgc/changeset/fe90ba781e8a/
Log: demo_random shows another livelock with this spinloop. Added it to
the comment, revert again to always abort.
diff --git a/c4/et.c b/c4/et.c
--- a/c4/et.c
+++ b/c4/et.c
@@ -675,23 +675,23 @@
{
if (v != d->my_lock) // not locked by me: conflict
{
- /* A case that can occur: two threads A and B are both
+ /* It's delicate here to do a spinloop rather than
+ just aborting.
+
+ A case that can occur: two threads A and B are both
committing, thread A locked object a, thread B
locked object b, and then thread A tries to
validate the reads it did on object b and
vice-versa. In this case both threads cannot
commit, but if they both enter the SpinLoop()
- above, then they will livelock.
+ here, then they will livelock.
- But this might lead both threads to cancel by
- reaching this point. For now we attempt to be
- more clever and let one of the threads commit
- anyway (the choice of which one looks random).
+ Another case: thread A might be blocked in this
+ spinloop, while thread B is blocked in the
+ SpinLoop(SPLP_LOCKED_COMMIT) below.
+
+ For now we always abort.
*/
- if (d->my_lock < v) {
- SpinLoop(SPLP_LOCKED_VALIDATE);
- goto retry;
- }
fprintf(stderr, "validation failed: "
"%p is locked by another thread\n", R);
return 0;
_______________________________________________
pypy-commit mailing list
[email protected]
http://mail.python.org/mailman/listinfo/pypy-commit
