Web lists-archives.com

Re: [PATCH 1/2] packfile: close and free packs upon releasing an object store

On Wed, May 09, 2018 at 05:12:10PM -0700, Stefan Beller wrote:

> In d0b59866223 (object-store: close all packs upon clearing the object
> store, 2018-03-23), we made sure to close all packfiles on releasing
> an object store, but we also have to free the memory of the closed packs.

I know we've assumed in a few places that a "struct packed_git" will
never go away. The one that comes to mind is the mru list.

It looks like we'll be OK here:

> diff --git a/object.c b/object.c
> index 66cffaf6e51..3e64a4a26dd 100644
> --- a/object.c
> +++ b/object.c
> @@ -485,6 +485,6 @@ void raw_object_store_clear(struct raw_object_store *o)
>  	o->alt_odb_tail = NULL;
>  	INIT_LIST_HEAD(&o->packed_git_mru);
> -	close_all_packs(o);
> +	close_and_free_packs(o);
>  	o->packed_git = NULL;
>  }

because we clear the list above. But it would be dangerous for anybody
else to call close_and_free_packs(). Should that INIT_LIST_HEAD get
moved down into that function?

Probably the same applies to setting NULL here; you're left with a
dangling pointer if you just call close_and_free_packs(). Should
that helper maybe just be a static function in object.c?

Just brainstorming other places where the immutability of "struct
packed_git" might be important:

  - pack-objects keeps a pointer from each object_entry to its
    containing packed_git. That's probably OK, as you wouldn't expect to
    be able to close the object store in the middle of that operation.

  - the reachability bitmap code holds a pointer to the pack that has a
    bitmap. Probably that whole "struct bitmap_index" needs to be part
    of the object_store (arguably it should all just be _inside_ the
    packed_git, but the current implementation avoids complexity by just
    having a single bitmap-per-repo).