This patch is fixing a race condition that may cause setting
count_pending to -1, which results in unwanted big bulk of arp messages
(in case of "notify peers").
Consider following scenario:
count_pending == 2
CPU0 CPU1
team_notify_peers_work
atomic_dec_and_test (dec count_pending to 1)
schedule_delayed_work
team_notify_peers
atomic_add (adding 1 to count_pending)
team_notify_peers_work
atomic_dec_and_test (dec count_pending to 1)
schedule_delayed_work
team_notify_peers_work
atomic_dec_and_test (dec count_pending to 0)
schedule_delayed_work
team_notify_peers_work
atomic_dec_and_test (dec count_pending to -1)
Fix this race by using atomic_dec_if_positive - that will prevent
count_pending running under 0.
Fixes: fc423ff00df3a1955441 ("team: add peer notification")
Fixes: 492b200efdd20b8fcfd ("team: add support for sending multicast rejoins")
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Large receive offloading is known to cause problems if received packets are passed to other host. Therefore the kernel disables it by calling dev_disable_lro() whenever a network device is enslaved in a bridge or forwarding is enabled for it (or globally). For virtual devices we need to disable LRO on the underlying physical device (which is actually receiving the packets). Current dev_disable_lro() code handles this propagation for a vlan (including 802.1ad nested vlan), macvlan or a vlan on top of a macvlan. It doesn't handle other stacked devices and their combinations, in particular propagation from a bond to its slaves which often causes problems in virtualization setups. As we now have generic data structures describing the upper-lower device relationship, dev_disable_lro() can be generalized to disable LRO also for all lower devices (if any) once it is disabled for the device itself. For bonding and teaming devices, it is necessary to disable LRO not only on current slaves at the moment when dev_disable_lro() is called but also on any slave (port) added later. v2: use lower device links for all devices (including vlan and macvlan) Signed-off-by: Michal Kubecek <mkubecek@suse.cz> Acked-by: Veaceslav Falico <vfalico@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Testing xmit_more support with netperf and connected UDP sockets, I found strange dst refcount false sharing. Current handling of IFF_XMIT_DST_RELEASE is not optimal. Dropping dst in validate_xmit_skb() is certainly too late in case packet was queued by cpu X but dequeued by cpu Y The logical point to take care of drop/force is in __dev_queue_xmit() before even taking qdisc lock. As Julian Anastasov pointed out, need for skb_dst() might come from some packet schedulers or classifiers. This patch adds new helper to cleanly express needs of various drivers or qdiscs/classifiers. Drivers that need skb_dst() in their ndo_start_xmit() should call following helper in their setup instead of the prior : dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; -> netif_keep_dst(dev); Instead of using a single bit, we use two bits, one being eventually rebuilt in bonding/team drivers. The other one, is permanent and blocks IFF_XMIT_DST_RELEASE being rebuilt in bonding/team. Eventually, we could add something smarter later. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Julian Anastasov <ja@ssi.bg> Signed-off-by: David S. Miller <davem@davemloft.net>
When team_notify_peers and team_mcast_rejoin are called, they both reset
their respective .count_pending atomic variable. Then when the actual
worker function is executed, the variable is atomically decremented.
This pattern introduces a potential race condition where the
.count_pending rolls over and the worker function keeps rescheduling
until .count_pending decrements to zero again:
THREAD 1 THREAD 2
======== ========
team_notify_peers(teamX)
atomic_set count_pending = 1
schedule_delayed_work
team_notify_peers(teamX)
atomic_set count_pending = 1
team_notify_peers_work
atomic_dec_and_test
count_pending = 0
(return)
schedule_delayed_work
team_notify_peers_work
atomic_dec_and_test
count_pending = -1
schedule_delayed_work
(repeat until count_pending = 0)
Instead of assigning a new value to .count_pending, use atomic_add to
tack-on the additional desired worker function invocations.
Signed-off-by: Joe Lawrence <joe.lawrence@stratus.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Fixes: fc423ff00df3a19554414ee ("team: add peer notification")
Fixes: 492b200efdd20b8fcfdac87 ("team: add support for sending multicast rejoins")
Signed-off-by: David S. Miller <davem@davemloft.net>
When one tries to add eth as a port into team and that eth is already in use by other rx_handler device (macvlan, bond, bridge, ...) a bug in team_port_add() causes that IFF_TEAM_PORT flag is set before rx_handler is registered. In between, netdev nofifier is called and team_device_event() sees IFF_TEAM_PORT and thinks that rx_handler_data pointer is set to team_port. But it isn't. Fix this by reordering rx_handler register and IFF_TEAM_PORT priv flag set so it is very similar to how bonding does this. Reported-by: Erik Hugne <erik.hugne@ericsson.com> Fixes: 3d249d4ca7 "net: introduce ethernet teaming device" Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
The variable "err" is not necessary. Return register_netdevice() directly. Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp> Acked-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
clean up names related to socket filtering and bpf in the following way:
- everything that deals with sockets keeps 'sk_*' prefix
- everything that is pure BPF is changed to 'bpf_*' prefix
split 'struct sk_filter' into
struct sk_filter {
atomic_t refcnt;
struct rcu_head rcu;
struct bpf_prog *prog;
};
and
struct bpf_prog {
u32 jited:1,
len:31;
struct sock_fprog_kern *orig_prog;
unsigned int (*bpf_func)(const struct sk_buff *skb,
const struct bpf_insn *filter);
union {
struct sock_filter insns[0];
struct bpf_insn insnsi[0];
struct work_struct work;
};
};
so that 'struct bpf_prog' can be used independent of sockets and cleans up
'unattached' bpf use cases
split SK_RUN_FILTER macro into:
SK_RUN_FILTER to be used with 'struct sk_filter *' and
BPF_PROG_RUN to be used with 'struct bpf_prog *'
__sk_filter_release(struct sk_filter *) gains
__bpf_prog_release(struct bpf_prog *) helper function
also perform related renames for the functions that work
with 'struct bpf_prog *', since they're on the same lines:
sk_filter_size -> bpf_prog_size
sk_filter_select_runtime -> bpf_prog_select_runtime
sk_filter_free -> bpf_prog_free
sk_unattached_filter_create -> bpf_prog_create
sk_unattached_filter_destroy -> bpf_prog_destroy
sk_store_orig_filter -> bpf_prog_store_orig_filter
sk_release_orig_filter -> bpf_release_orig_filter
__sk_migrate_filter -> bpf_migrate_filter
__sk_prepare_filter -> bpf_prepare_filter
API for attaching classic BPF to a socket stays the same:
sk_attach_filter(prog, struct sock *)/sk_detach_filter(struct sock *)
and SK_RUN_FILTER(struct sk_filter *, ctx) to execute a program
which is used by sockets, tun, af_packet
API for 'unattached' BPF programs becomes:
bpf_prog_create(struct bpf_prog **)/bpf_prog_destroy(struct bpf_prog *)
and BPF_PROG_RUN(struct bpf_prog *, ctx) to execute a program
which is used by isdn, ppp, team, seccomp, ptp, xt_bpf, cls_bpf, test_bpf
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 34c5bd66e5ed introduced the possibility that an
uninitialized pointer on the stack (orig_fp) can call into
sk_unattached_filter_destroy() when its value is non NULL.
Before that commit orig_fp was only destroyed in the same
block where it was assigned a valid BPF prog before. Fix it
up by initializing it to NULL.
Fixes: 34c5bd66e5ed ("net: filter: don't release unattached filter through call_rcu()")
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Pablo Neira <pablo@netfilter.org>
Cc: Alexei Starovoitov <ast@plumgrid.com>
Cc: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
sk_unattached_filter_destroy() does not always need to release the
filter object via rcu. Since this filter is never attached to the
socket, the caller should be responsible for releasing the filter
in a safe way, which may not necessarily imply rcu.
This is a short summary of clients of this function:
1) xt_bpf.c and cls_bpf.c use the bpf matchers from rules, these rules
are removed from the packet path before the filter is released. Thus,
the framework makes sure the filter is safely removed.
2) In the ppp driver, the ppp_lock ensures serialization between the
xmit and filter attachment/detachment path. This doesn't use rcu
so deferred release via rcu makes no sense.
3) In the isdn/ppp driver, it is called from isdn_ppp_release()
the isdn_ppp_ioctl(). This driver uses mutex and spinlocks, no rcu.
Thus, deferred rcu makes no sense to me either, the deferred releases
may be just masking the effects of wrong locking strategy, which
should be fixed in the driver itself.
4) In the team driver, this is the only place where the rcu
synchronization with unattached filter is used. Therefore, this
patch introduces synchronize_rcu() which is called from the
genetlink path to make sure the filter doesn't go away while packets
are still walking over it. I think we can revisit this once struct
bpf_prog (that only wraps specific bpf code bits) is in place, then
add some specific struct rcu_head in the scope of the team driver if
Jiri thinks this is needed.
Deferred rcu release for unattached filters was originally introduced
in 302d663 ("filter: Allow to create sk-unattached filters").
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>