Re: [PATCH] receive-pack: plug minor memory leak in unpack()

>>
>> I took a look at a few of them:
>
> I took a brief look, too.
>
> I had hoped we could just all it "env" and everybody would be happy
> using it instead of bare pointers. But quite a few callers assign
> "local_repo_env" to to child_process.env. We could copy it into an
> argv_array, of course, but that really feels like working around the
> interface. So I think we would prefer to keep both forms available.

> That raises the question: what should it be called? The "argv_array"
> form of "argv" is called "args". The more I see it, the more I hate that
> name, as the two are easily confused. We could have:
>
>    const char **argv;
>    struct argv_array argv_array;
>    const char **env;
>    struct argv_array env_array;
>
> though "argv_array" is a lot to type when you have a string of
> argv_array_pushf() calls (which are already by themselves kind of
> verbose). Maybe that's not too big a deal, though.

>>   - daemon.c::handle() uses a static set of environment variables
>>     that are not built with argv_array().  Same for argv.
>
> Most of the callers you mentioned are good candidates. This one is
> tricky.
>
> The argv array gets malloc'd and set up by the parent git-daemon
> process. Then each time we get a client, we create a new struct
> child_process that references it. So using the managed argv-array would
> actually be a bit more complicated (and not save any memory; we just
> always point to the single copy for each child).
>
> For the environment, we build it in a function-local buffer, point the
> child_process's env field at it, start the child, and then copy the
> child_process into our global list of children. When we notice a child
> is dead (by linearly going through the list with waitpid), we free the
> list entry. So there are a few potentially bad things here:
>
>    1. We memcpy the child_process to put it on the list. Which does work,
>       though it feels a little like we are violating the abstraction
>       barrier.
>
>    2. The child_process in the list points to the local "env" buffer that
>       is no longer valid. There's no bug because we don't ever look at
>       it. Moving to a managed env would fix that. But I have to wonder if
>       we even want to be keeping the "struct child_process" around in the
>       first place (all we really care about is the pid).
>
>    3. If we do move to a managed env, then we expect it to get cleaned up
>       in finish_command. But we never call that; we just free the list
>       memory containing the child_process. We would want to call
>       finish_command. Except that we will have reaped the process already
>       with our call to waitpid() from check_dead_children. So we'd need a
>       new call to do just the cleanup without the wait, I guess.
>
>    4. For every loop on the listen socket, we call waitpid() for each
>       living child, which is a bit wasteful. We'd probably do better to
>       call waitpid(0, &status, WNOHANG), and then look up the resulting
>       pid in a hashmap or sorted list when we actually see something that
>       died. I don't know that this is a huge problem in practice. We use
>       git-daemon pretty heavily on our backend servers at GitHub, and it
>       seems to use about 5% of a CPU constantly on each machine. Which is
>       kind of lame, given that it isn't doing anything at all, but is
>       hardly earth-shattering.
>
> So I'm not sure if it is worth converting to a managed env. There's a
> bit of ugliness, but none of it is causing any problems, and doing so
> opens a can of worms. The most interesting thing to fix (to me, anyway)
> is number 4, but that has nothing to do with the env in the first place.
> :)