Jim Wallace wrote:
>> 2. Why does the first example need the loop? The second pass
>> isn't actually needed to make it deadlock, is it?
> Not to get the deadlock, but to have it succeed, it must do the second
> pass
Okay, I just didn't understand that you want it to show both states,
success and failure.
There's an implied solution here that I wouldn't want to present in a
MySQL++ example, because that implies an endorsement. The examples are
supposed to show good programming style and acceptable idioms. "I'm
doing it like the examples" should never be a defense for bad code.
In a real program, looping is no fix for this problem, because you can't
know that a second try will always fix it. The lock could be reacquired
by another program by the time you loop back around. And putting the
whole thing in a while (1) or "for (i = 0; i < bignum; ++i)" loop isn't
any good either.
I think we don't want the example to succeed. It's intentionally
demonstrating a programming bug, so it should fail.
> problem is that deadlock2 hangs waiting for the lock 1. Once deadlock1
> tries to get lock 2 (which deadlock2 has), it immediately gets the
> deadlock exception, and rolls back, releasing lock1 so deadlock2
> completes ok. And after deadlock2 completes, deadlock1 can retry and
> succeed.
Okay, so it's like with regular mutexes, where the order of acquisition
matters. Do it in the same order in all programs, and it can't
deadlock, yes?
If we get rid of the loop and maybe add another pause point, is the only
difference the order of lock acquisition? If so, you could just add an
option to examples/cmdline.cpp to switch between the modes. It could be
just a generic "mode switch" (-m?) taking an integer that means
something only to some programs. To examples/deadlock, -m1 says to
acquire lock1 first, and -m2 lock2. All other programs can ignore it
for now. The value can go in a global variable, like dtest_mode.