On Wednesday 02 February 2011 18:37:18 Linus Lüssing wrote:
From: Sven Eckelmann sven@narfation.org
Was:
<TODO: write a long monologue about every problem we have or could have or maybe never had and would have when we not have it>
Signed-off-by: Sven Eckelmann sven@narfation.org
So after some more discussions with Marek and Sven, it looks like we have to use the rcu protected macros rcu_dereference() and rcu_assign_pointer() for the bat_priv->curr_gw and curr_gw->orig_node.
Changes here also include moving the kref_get() from unicast_send_skb() into gw_get_selected(). The orig_node could have been freed already at the time the kref_get() was called in unicast_send_skb().
Some things that are still not that clear to me:
gw_election():
- can the if-block before gw_deselect() be ommited, we had a nullpointer check for curr_gw just a couple of lines before during the rcu-lock.
I thought that this if block should be moved to gw_select. And your gw_select still has the bug that the bat_priv->curr_gw isn't set to NULL when new_gw_node is NULL.
gw_deselet():
- is the refcount at this time always 1 for gw_node, can the null
pointer check + a rcu_dereference be ommited? (at least that's what it looks like when comparing to the rcuref.txt example)
Why can't it be NULL? And _always_ use rcu_dereference. What example tells you that it isn't needed? None of the examples has any kind of rcu pointer in it (just el as pointer which is stored in a struct were the pointer inside the struct is rcu protected).
gw_get_selected():
- Probably the orig_node's refcounting has to be made atomic, too?
This part is still a little bit ugly and I cannot give you an easy answer. Just think about following: * Hash list is a bunch of rcu protected lists * pointer to originator is stored inside a bucket (list elements inside the hash) * hash bucket wants to get removed - call_rcu; reference count of the originator is decremented immediately * (!!!! lots of reordering of read and write commands inside the cpu!!!! - aren't we happy about the added complexity which tries to hide the memory latency?) * the originator was removed, the bucket which is removed in the call_rcu still points to the removed originator * a parallel running operation tries to find a originator, the rcu list iterator gets the to-be-deleted bucket to the originator * the pointer to the already removed originator inside the bucket is dereferenced, data is read/written -> Kernel Oops
Does this sound scary? At least it could be used in some horror movies (and I would watch them).
But that is the other problem I currently have with the state of batman-adv in trunk - and I think I forget to tell you about it after the release of v2011.0.0.
So, a good idea would be the removal of the buckets for the hash. Usage of "struct hlist_node" inside the hash elements should be a good starting point. But think about the problem that the different hashes could have the same element. So you need for each distinct hash an extra "struct hlist_node" inside the element which should be part of the hash. The hash_add (and related) functions don't get the actual pointer to the element, but the pointer to the correct "struct hlist_node" inside the element/struct. The comparison and hashing function would also receive "struct hlist_node" as parameter and must get the pointer to the element using the container_of macro.
@@ -171,7 +172,7 @@ struct bat_priv { struct delayed_work hna_work; struct delayed_work orig_work; struct delayed_work vis_work;
struct gw_node *curr_gw;
struct gw_node *curr_gw; /* rcu protected pointer */ struct vis_info *my_vis_info;};
Sry, but I have to say that: FAIL ;)
I think it should look that way:
struct gw_node *curr_gw;
struct gw_node __rcu *curr_gw;
Best regards, Sven