Re: [PATCH] TOMOYO: Add garbage collector support. (v2)

["Serge E. Hallyn" <serge@hallyn.com>]

Quoting Tetsuo Handa (penguin-kernel@I-love.SAKURA.ne.jp):
> Hello.
> 
> Serge E. Hallyn wrote:
> > So yes, you might be able to get more review of your patch
> > if you split it up into:
> > 
> > 	1. move allocations outside of semaphore
> > 	2. add proper refcounting
> > 	3. get rid of ->is_deleted
> 
> I'm ready to post part (1) and part (2).
> But I'm not sure whether part (3) is worth to implement or not.
>
> If we don't use is_deleted flag, we will need two "struct list_head"s,
> one is for readers and writers, the other is for GC.
> 
> struct something {
> 	struct {
> 		struct list_head list;
> 		atomic_t users;
> 	} entry;
> 	int data;
> 	bool is_deleted;
> } *p;
> 
> If we remove p from the list using list_del() (or list_del_rcu()) instead of
> setting p->is_deleted to true, we have to somehow remember the element pointed
> by p so that that element is reachable from the GC's list.
> We can't wait for p->entry.users to become 0 inside writer function (which is
> called only when modifying lists). Only GC function (which is called
> repeatedly) can check whether p->entry.users == 0 or not because the element
> pointed by p is long-lived. We need to remember the location of the element
> pointed by p .

?

Can't you just move the deleted item from the live list to some dead
list, using the same ->list list_head?

> We can allocate memory when adding p to GC's list, but allocating memory at
> list_del() (or list_del_rcu()) time could fail. It is not a good thing that
> removing an element from the reader's list fails because of -ENOMEM. We already
> have a way (i.e. is_deleted flag) to avoid -ENOMEM.
> 
> We can add another "struct list_head" to each element which is used by GC.
> But that approach consumes more memory than is_deleted flag.

?

See above.

-serge