netfilter-devel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Pablo Neira Ayuso <pablo@netfilter.org>
To: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
Cc: netfilter-devel@vger.kernel.org
Subject: Re: [PATCH 2/8] netfilter: ipset: Prepare ipset core for RCU locking instead of rwlock per set
Date: Thu, 18 Dec 2014 19:38:33 +0100	[thread overview]
Message-ID: <20141218183833.GA3552@salvia> (raw)
In-Reply-To: <1418421489-17411-3-git-send-email-kadlec@blackhole.kfki.hu>

Hi Jozsef,

Several comments to this new round.

On Fri, Dec 12, 2014 at 10:58:03PM +0100, Jozsef Kadlecsik wrote:
> @@ -1407,9 +1407,13 @@ call_ad(struct sock *ctnl, struct sk_buff *skb, struct ip_set *set,
>  	bool eexist = flags & IPSET_FLAG_EXIST, retried = false;
>  
>  	do {
> -		write_lock_bh(&set->lock);
> +		spin_lock_bh(&set->lock);
>  		ret = set->variant->uadt(set, tb, adt, &lineno, flags, retried);
> -		write_unlock_bh(&set->lock);
> +		spin_unlock_bh(&set->lock);
> +		if (ret == -EINPROGRESS) {
> +			synchronize_rcu_bh();

Let me zoom in to code that is related to this -EINPROGRESS case.

> +			ret = 0;
> +		}
>  		retried = true;
>  	} while (ret == -EAGAIN &&
>  		 set->variant->resize &&

>From list_set_uadd():

       ...
       if (n) {
               list_set_replace(set, e, n);
               ret = -EINPROGRESS;

This EINPROGRESS is propagated via set->variant->uadt(), which results
in th synchronize_rcu_bh() from the core.

Then, let's have a look at list_set_replace():

>static inline void
>list_set_replace(struct ip_set *set, struct set_elem *e, struct set_elem *old)
>{
>       list_replace_rcu(&old->list, &e->list);
>       __list_set_del(set, old);
>}

This uses list rcu safe variant, this is good.

>static void
>__list_set_del(struct ip_set *set, struct set_elem *e)
> {
[...]
>        kfree_rcu(e, rcu);
>}

You use kfree_rcu() to defer the release of the object by when no
readers are walking over those bits, good.

But then, you don't need the synchronize_rcu(). To clarify:

1) If you release memory synchronously, you use this pattern:

  list_replace_rcu(...);
  synchronize_rcu(); <---- waits until no readers are referencing to
                           objects that we removes with the rcu safe list
                           variant
  kfree(...);

2) If you release memory asynchronously, then:

  list_replace_rcu(...);
  kfree_rcu(...);

In this case, I think you don't need to call synchronize_rcu() since
you selected approach 2).

More concerns: Let's revisit __list_set_del():

>static void
>__list_set_del(struct ip_set *set, struct set_elem *e)
> {
>        struct list_set *map = set->data;
> 
>        ip_set_put_byindex(map->net, e->id);
>        /* We may call it, because we don't have a to be destroyed
>         * extension which is used by the kernel.
>         */
>        ip_set_ext_destroy(set, e);
         ^.......................^

You may have a reader still walking on the extension area of the
element by when you call this.

>        kfree_rcu(e, rcu);
>}

The alternative to make sure that everything is released by when no
readers are accessing the object anymore is to use call_rcu():

        call_rcu(e, ipset_list_free_rcu);

Then:

static void ipset_list_free_rcu(struct rcu_head *rcu)
{
        ip_set_put_byindex(...);
        ip_set_ext_destroy(...);
        kfree(set);
}

This safely release memory with 100% guarantee no readers are accesing
the object you're destroying.

*But* you have to make sure none of those function in the callback may
sleep. Since this callback is called from the rcu softirq (interrupt
context).

I suspect (actually I would need to make a closer look) you cannot use
this pattern easily unless elements keep a pointer to the set they
belong to, to access the data you need from the callback.

Let me know,
Thanks.

  reply	other threads:[~2014-12-18 18:36 UTC|newest]

Thread overview: 12+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2014-12-12 21:58 [PATCH 0/8] ipset patches for nf-next, v3 Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 1/8] netfilter: ipset: Remove rbtree from hash:net,iface in order to run under RCU Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 2/8] netfilter: ipset: Prepare ipset core for RCU locking instead of rwlock per set Jozsef Kadlecsik
2014-12-18 18:38   ` Pablo Neira Ayuso [this message]
2014-12-21 11:52     ` Jozsef Kadlecsik
2014-12-21 12:23       ` Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 3/8] netfilter: ipset: Introduce RCU locking in the bitmap types Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 4/8] netfilter: ipset: Introduce RCU locking in the list type Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 5/8] netfilter: ipset: Introduce RCU locking in the hash types Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 6/8] netfilter: ipset: styles warned by checkpatch.pl fixed Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 7/8] netfilter: ipset: Fix parallel resizing and listing of the same set Jozsef Kadlecsik
2014-12-12 21:58 ` [PATCH 8/8] netfilter: ipset: Fix sparse warning Jozsef Kadlecsik

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20141218183833.GA3552@salvia \
    --to=pablo@netfilter.org \
    --cc=kadlec@blackhole.kfki.hu \
    --cc=netfilter-devel@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).