talloc quiz, and: dangling talloc references and inconsistencies in the talloc model

[tridge at samba.org]

Hi Sam,

 > Also, the test I supplied called "test_dangling_loop" shows a bug,
 > described just above the test definition.

I don't see it as a bug. The test is an extension of a simple
loop. The simple loop 


                             p1
                            /  \     
                            |   |
                            |   |    
                             \ /
                             p2

The diagram is supposed to show that p1 is a parent of p2, but p2 is a
parent of p1. So what do we do if you talloc_free(p1) ? If we follow
your rules then the free will do nothing. You can't get rid of that
loop.

So (quite deliberately) I added code so that a talloc_free() on a loop
like that does free the loop. That seemed to be the only sensible
thing to do. 

Viewed using the print format that talloc_report_full() uses, the
memory looks like this:

full talloc report on 'null_context' (total     51 bytes in   7 blocks)
    root                           contains     51 bytes in   6 blocks (ref 0) 0x20380c0
        r1                             contains     25 bytes in   2 blocks (ref 0) 0x20381e0
            reference to: p2
        p1                             contains     26 bytes in   3 blocks (ref 1) 0x2038120
            p2                             contains     25 bytes in   2 blocks (ref 1) 0x2038180
                reference to: p1

then we free p1, and we end up with this:

full talloc report on 'null_context' (total      2 bytes in   4 blocks)
    root                           contains      2 bytes in   3 blocks (ref 0) 0x20380c0
        r1                             contains      2 bytes in   2 blocks (ref 0) 0x20381e0
            p2                             contains      1 bytes in   1 blocks (ref 0) 0x2038180

I think that is quite sensible. We had a loop, which got removed, but
one part of the loop had an external reference, and that part of the
loop got saved, as a child of the pointer that referenced it.

 > I tried to consider if there was any way (without my full blown
 > solution) ....

What is your proposed full blown solution? I may well have just missed
it in the various emails. Do you have a working set of code which
would replace talloc, or a testsuite showing how you think it might be
used?  That might help me understand what you are pushing for.

Cheers, Tridge