talloc: Other minor issues/queries
Jeremy Allison
jra at samba.org
Wed Oct 21 00:11:13 UTC 2020
On Tue, Oct 20, 2020 at 03:10:14PM -0500, Arran Cudbard-Bell via samba-technical wrote:
> Just to keep this separate from the memlimit/pools discussion.
>
> - I've noticed talloc_realloc_size clears the name/type of the chunk be
> allocated. This isn't mentioned in the documentation. Is clearing the
> name/type of the chunk expected behaviour?
Doesn't it add the name of the program text that called the realloc ?
> - talloc_set_memlimit is marked as deprecated in the
> current HEAD version of talloc, with a note to use cgroups.
> cgroups wouldn't offer equivalent functionality for us. Could I humbly
> request memlimits not be deprecated? :)
OK, I think memlimits add horrible extra complexity to
save applications some bookkeeping.
I want to *reduce* complexity in talloc.
First fix I want to do after the memlimit changes and tests
have gone in is to remove the code paths around ALWAYS_REALLOC.
We never compile with it in test, and we have tests that
depend on it *not* being set, so it really needs to go.
Second, I'd like to remove the memlimit code. Can you be
*really* clear as to what you are using them for, and
can we find some other way to do this ?
If we can do that, then.....
> - Is there any appetite for functions which return chunks of a specific
> alignment? i.e. talloc_aligned, talloc_aligned_array,
> talloc_aligned_pool.
> I've had to fudge some horrible variants of the above for local use.
> We use aligned memory for the following purposes:
> - Page aligned pools for protecting memory areas with mprotect
> to prevent runtime modifications.
> - 8 byte aligned buffers that can have the start and end regions
> easily poisoned with ASAN.
> - Cache line aligned ring buffers, with the cache line
> alignment done to reduce contention where the producer and
> consumer are close.
Adding this, more interesting functionality becomes
a lot easier to handle and manage complexity wise.
Just my 2cents :-).
More information about the samba-technical
mailing list