| Message ID | 20240912122148.12159-2-phil@nwl.cc |
|---|---|
| State | Accepted |
| Headers | show |
| Series | Dynamic hook interface binding | expand |
Phil Sutter <phil@nwl.cc> wrote: > Documentation of list_del_rcu() warns callers to not immediately free > the deleted list item. While it seems not necessary to use the > RCU-variant of list_del() here in the first place, doing so seems to > require calling kfree_rcu() on the deleted item as well. > > Fixes: 3f0465a9ef02 ("netfilter: nf_tables: dynamically allocate hooks per net_device in flowtables") > Signed-off-by: Phil Sutter <phil@nwl.cc> > --- > net/netfilter/nf_tables_api.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c > index b6547fe22bd8..2982f49b6d55 100644 > --- a/net/netfilter/nf_tables_api.c > +++ b/net/netfilter/nf_tables_api.c > @@ -9180,7 +9180,7 @@ static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable) > flowtable->data.type->setup(&flowtable->data, hook->ops.dev, > FLOW_BLOCK_UNBIND); > list_del_rcu(&hook->list); > - kfree(hook); > + kfree_rcu(hook, rcu); > } > kfree(flowtable->name); > module_put(flowtable->data.type->owner); AFAICS its safe to use list_del() everywhere, I can't find a single instance where the hooks are iterated without mutex serialization. nf_tables_flowtable_destroy() is called after the hook has been unregisted (detached from nf_hook list) and rcu grace period elapsed.
On Thu, Sep 12, 2024 at 03:32:55PM +0200, Florian Westphal wrote: > Phil Sutter <phil@nwl.cc> wrote: > > Documentation of list_del_rcu() warns callers to not immediately free > > the deleted list item. While it seems not necessary to use the > > RCU-variant of list_del() here in the first place, doing so seems to > > require calling kfree_rcu() on the deleted item as well. > > > > Fixes: 3f0465a9ef02 ("netfilter: nf_tables: dynamically allocate hooks per net_device in flowtables") > > Signed-off-by: Phil Sutter <phil@nwl.cc> > > --- > > net/netfilter/nf_tables_api.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c > > index b6547fe22bd8..2982f49b6d55 100644 > > --- a/net/netfilter/nf_tables_api.c > > +++ b/net/netfilter/nf_tables_api.c > > @@ -9180,7 +9180,7 @@ static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable) > > flowtable->data.type->setup(&flowtable->data, hook->ops.dev, > > FLOW_BLOCK_UNBIND); > > list_del_rcu(&hook->list); > > - kfree(hook); > > + kfree_rcu(hook, rcu); > > } > > kfree(flowtable->name); > > module_put(flowtable->data.type->owner); > > AFAICS its safe to use list_del() everywhere, I can't find a single > instance where the hooks are iterated without mutex serialization. > > nf_tables_flowtable_destroy() is called after the hook has been > unregisted (detached from nf_hook list) and rcu grace period elapsed. Yes, I didn't find a caller which didn't synchronize_rcu() before calling it. Same applies to chain hooks, right? I'd drop all the _rcu() calls and give it a try, but the resulting race conditions may be hard to trigger. Cheers, Phil
Phil Sutter <phil@nwl.cc> wrote: > > nf_tables_flowtable_destroy() is called after the hook has been > > unregisted (detached from nf_hook list) and rcu grace period elapsed. > > Yes, I didn't find a caller which didn't synchronize_rcu() before > calling it. Same applies to chain hooks, right? Sigh, there is nft_flowtable_find_dev() which iterates the nft_hook list from packet path. So the syncrhonize_rcu is irrelevant as long as the entry is linked up and this patch is correct as-is. list_del_rcu(&hook->list); kfree(hook); is illegal, and I think this should add a helper that unlinks and then frees the entry via kfree_rcu and converts all instances of this pattern.
On Thu, Sep 12, 2024 at 03:32:55PM +0200, Florian Westphal wrote: > Phil Sutter <phil@nwl.cc> wrote: > > Documentation of list_del_rcu() warns callers to not immediately free > > the deleted list item. While it seems not necessary to use the > > RCU-variant of list_del() here in the first place, doing so seems to > > require calling kfree_rcu() on the deleted item as well. > > > > Fixes: 3f0465a9ef02 ("netfilter: nf_tables: dynamically allocate hooks per net_device in flowtables") > > Signed-off-by: Phil Sutter <phil@nwl.cc> > > --- > > net/netfilter/nf_tables_api.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c > > index b6547fe22bd8..2982f49b6d55 100644 > > --- a/net/netfilter/nf_tables_api.c > > +++ b/net/netfilter/nf_tables_api.c > > @@ -9180,7 +9180,7 @@ static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable) > > flowtable->data.type->setup(&flowtable->data, hook->ops.dev, > > FLOW_BLOCK_UNBIND); > > list_del_rcu(&hook->list); > > - kfree(hook); > > + kfree_rcu(hook, rcu); This looks correct to me. > > } > > kfree(flowtable->name); > > module_put(flowtable->data.type->owner); > > AFAICS its safe to use list_del() everywhere, I can't find a single > instance where the hooks are iterated without mutex serialization. Netlink dump path is lockless. nft_dump_basechain_hook() is missing list_for_each_entry_rcu() for list iteration, that needs a fix. nf_tables_fill_flowtable_info() does use list_for_each_entry_rcu(). > nf_tables_flowtable_destroy() is called after the hook has been > unregisted (detached from nf_hook list) and rcu grace period elapsed.
On Mon, Sep 16, 2024 at 02:00:59AM +0200, Pablo Neira Ayuso wrote: > On Thu, Sep 12, 2024 at 03:32:55PM +0200, Florian Westphal wrote: > > Phil Sutter <phil@nwl.cc> wrote: > > > Documentation of list_del_rcu() warns callers to not immediately free > > > the deleted list item. While it seems not necessary to use the > > > RCU-variant of list_del() here in the first place, doing so seems to > > > require calling kfree_rcu() on the deleted item as well. > > > > > > Fixes: 3f0465a9ef02 ("netfilter: nf_tables: dynamically allocate hooks per net_device in flowtables") > > > Signed-off-by: Phil Sutter <phil@nwl.cc> > > > --- > > > net/netfilter/nf_tables_api.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c > > > index b6547fe22bd8..2982f49b6d55 100644 > > > --- a/net/netfilter/nf_tables_api.c > > > +++ b/net/netfilter/nf_tables_api.c > > > @@ -9180,7 +9180,7 @@ static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable) > > > flowtable->data.type->setup(&flowtable->data, hook->ops.dev, > > > FLOW_BLOCK_UNBIND); > > > list_del_rcu(&hook->list); > > > - kfree(hook); > > > + kfree_rcu(hook, rcu); > > This looks correct to me. > > > > } > > > kfree(flowtable->name); > > > module_put(flowtable->data.type->owner); > > > > AFAICS its safe to use list_del() everywhere, I can't find a single > > instance where the hooks are iterated without mutex serialization. > > Netlink dump path is lockless. > > nft_dump_basechain_hook() is missing list_for_each_entry_rcu() for > list iteration, that needs a fix. > > nf_tables_fill_flowtable_info() does use list_for_each_entry_rcu(). I'd suggest to start by sending fixes for nf.git to address these two issues. > > nf_tables_flowtable_destroy() is called after the hook has been > > unregisted (detached from nf_hook list) and rcu grace period elapsed.
On Thu, Sep 12, 2024 at 02:21:33PM +0200, Phil Sutter wrote: > Documentation of list_del_rcu() warns callers to not immediately free > the deleted list item. While it seems not necessary to use the > RCU-variant of list_del() here in the first place, doing so seems to > require calling kfree_rcu() on the deleted item as well. LGTM, I plan to take this into nf.git Thanks. > Fixes: 3f0465a9ef02 ("netfilter: nf_tables: dynamically allocate hooks per net_device in flowtables") > Signed-off-by: Phil Sutter <phil@nwl.cc> > --- > net/netfilter/nf_tables_api.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c > index b6547fe22bd8..2982f49b6d55 100644 > --- a/net/netfilter/nf_tables_api.c > +++ b/net/netfilter/nf_tables_api.c > @@ -9180,7 +9180,7 @@ static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable) > flowtable->data.type->setup(&flowtable->data, hook->ops.dev, > FLOW_BLOCK_UNBIND); > list_del_rcu(&hook->list); > - kfree(hook); > + kfree_rcu(hook, rcu); > } > kfree(flowtable->name); > module_put(flowtable->data.type->owner); > -- > 2.43.0 >
diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c index b6547fe22bd8..2982f49b6d55 100644 --- a/net/netfilter/nf_tables_api.c +++ b/net/netfilter/nf_tables_api.c @@ -9180,7 +9180,7 @@ static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable) flowtable->data.type->setup(&flowtable->data, hook->ops.dev, FLOW_BLOCK_UNBIND); list_del_rcu(&hook->list); - kfree(hook); + kfree_rcu(hook, rcu); } kfree(flowtable->name); module_put(flowtable->data.type->owner);
Documentation of list_del_rcu() warns callers to not immediately free the deleted list item. While it seems not necessary to use the RCU-variant of list_del() here in the first place, doing so seems to require calling kfree_rcu() on the deleted item as well. Fixes: 3f0465a9ef02 ("netfilter: nf_tables: dynamically allocate hooks per net_device in flowtables") Signed-off-by: Phil Sutter <phil@nwl.cc> --- net/netfilter/nf_tables_api.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)