Unlike talloc, we don't have a count on everything, just on array
allocations.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

You don't need write access to the context to attach a destructor;
it's meta.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

$ for i in `seq 10`; do ./samba-allocs talloc.dump --tal; done | stats.c --trim-outliers
Before:
Read 25998 nodes
Tal time:                2857988-2981158(2904324)ns
Tal_free time:           2417170-2582088(2470931)ns
Single tal_free time:    1551654-1609583(1575984)ns

After:
Read 25998 nodes
Tal time:                2780974-2901323(2839352)ns
Tal_free time:           2241725-2355309(2285165)ns
Single tal_free time:    1549908-1612485(1566609)ns
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

We sanity check tal headers by ensuring that the pointers are in the
bounds of things we've allocated.  But the first one we check is the
prop ptr, which may also be a literal: this is_literal() dereferences
the pointer, which means we usually crash here if it's not a tal
object.

Move that last, and we have far more success with our sanity checking.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

All the effort trying to keep the header size down to 2 pointers turns
out to be wasted.  In addition, getting the parent of a tal pointer is
now much faster.

./samba-allocs talloc.dump --tal-size:
Before:
	Virtual size = 9633792, RSS = 3952640
After:
	Virtual size = 9793536, RSS = 3948544

And we're much faster now, esp. on free.

./samba-allocs talloc.dump --tal:
Before:
	Tal time:                2718068ns
	Tal_free time:           3360258ns
	Single tal_free time:    1667412ns
After:
	Tal time:                2788287ns
	Tal_free time:           2290167ns
	Single tal_free time:    1566998ns
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>


Header from folded patch 'tal-fix-set-destroying-bit.patch':

tal: fix destroying bit usage.

Actually looking at a destroying contexts' parent didn't work, and a couple
of places didn't filter out the destroying bit.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

If we see an ccan include line in an example, we add that to the
dependencies to try to build the example, but we didn't handle nested
modules.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

ccanlint would abort with 'Malformed line 53' if there was a bad header.
That's very poor, and deeply unhelpful.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

We need ccan/take here.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>


Header from folded patch 'path-talloc-take.patch':

tal/path: accept take() for arguments.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Tal-enhanced string routines, copied from str_talloc (which I also wrote).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Our first nested module; easy because noone else relies on it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Removing this assumption should allow nested modules.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Far more effective to just skip valgrind and gcov if you want fast
results; that's what we do for the summary already.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Clumsier than using wildcards, but that breaks down with nested modules
anyway.  In the longer term, we want ccantool to do the building, but for
the moment make is a necessary evil.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

In particular, we might need the C files for foreach and err.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Firstly, we should always ask before altering files.  Secondly, we are
not in the module directory: we need to use m->info_file->fullname.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Don't risk reading partially-constructed files.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

This makes it convenient to do:

	ptr = tal_free(ptr);
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

This makes error handling much more convenient, since take is usually
used for chaining functions.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

TAL_TAKE is awkward to use, particularly on functions which take multiple
arguments.  Instead, record annotations for pointers due to be freed in
the callee.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Sometimes you want to duplicate and enlarge in one step: particularly with
TAL_TAKE.  It's also typesafe, unlike tal_memdup().
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Instead of trying to force people to use the return value, pass a pointer.
This makes it easier if you want to handle failure: no overwriting the old
pointer!
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Always good form to have cleanup functions preserve errno.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Rusty forgets a git add.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

We call remove_node() in tal_steal, then re-use the node.  So it's important
that we always unlink the group property from the node for that case.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Don't open-code the list_head_init: left over from performance
debugging.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

First, we need to update the bounds for the group property allocated
with the children property, just in case it's at the end of the
allocations.

Second, we need to allow the tal_check() code to accept the pointer to
the null parent's group property.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

This is based on a very simple dump of S4's talloc tree, which we try
to duplicate using tal.  The benchmarks are simply to allocate all the
nodes, free all the nodes individually, and a top-level free.

Size results (32-bit x86):
	$ ./samba-allocs talloc.dump --talloc-size
	Read 25998 nodes
	Virtual size = 10469376, RSS = 4759552
	$ ./samba-allocs talloc.dump --tal-size
	Read 25998 nodes
	Virtual size = 9629696, RSS = 3948544

Speed results:
	$ ./samba-allocs talloc.dump
	Read 25998 nodes
	Malloc time:             1912082ns
	Free time:               1384892ns
	Talloc time:             2851531ns
	talloc_free time:        2133801ns
	Single talloc_free time: 1696298ns
	Tal time:                2686952ns
	Tal_free time:           3316153ns
	Single tal_free time:    1639407ns
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

We modify it a bit (to do more allocations before freeing).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>