Add fanout mode PACKET_FANOUT_CBPF that accepts a classic BPF program
to select a socket.
This avoids having to keep adding special case fanout modes. One
example use case is application layer load balancing. The QUIC
protocol, for instance, encodes a connection ID in UDP payload.
Also add socket option SOL_PACKET/PACKET_FANOUT_DATA that updates data
associated with the socket group. Fanout mode PACKET_FANOUT_CBPF is the
only user so far.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Acked-by: Alexei Starovoitov <ast@plumgrid.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Destruction of the po->rollover must be delayed until there are no
more packets in flight that can access it. The field is destroyed in
packet_release, before synchronize_net. Delay using rcu.
Fixes: 0648ab70af ("packet: rollover prepare: per-socket state")
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rollover indicates exceptional conditions. Export a counter to inform
socket owners of this state.
If no socket with sufficient room is found, rollover fails. Also count
these events.
Finally, also count when flows are rolled over early thanks to huge
flow detection, to validate its correctness.
Tested:
Read counters in bench_rollover on all other tests in the patchset
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Migrate flows from a socket to another socket in the fanout group not
only when the socket is full. Start migrating huge flows early, to
divert possible 4-tuple attacks without affecting normal traffic.
Introduce fanout_flow_is_huge(). This detects huge flows, which are
defined as taking up more than half the load. It does so cheaply, by
storing the rxhashes of the N most recent packets. If over half of
these are the same rxhash as the current packet, then drop it. This
only protects against 4-tuple attacks. N is chosen to fit all data in
a single cache line.
Tested:
Ran bench_rollover for 10 sec with 1.5 Mpps of single flow input.
lpbb5:/export/hda3/willemb# ./bench_rollover -l 1000 -r -s
cpu rx rx.k drop.k rollover r.huge r.failed
0 14 14 0 0 0 0
1 20 20 0 0 0 0
2 16 16 0 0 0 0
3 6168824 6168824 0 4867721 4867721 0
4 4867741 4867741 0 0 0 0
5 12 12 0 0 0 0
6 15 15 0 0 0 0
7 17 17 0 0 0 0
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rollover has to call packet_rcv_has_room on sockets in the fanout
group to find a socket to migrate to. This operation is expensive
especially if the packet sockets use rings, when a lock has to be
acquired.
Avoid pounding on the lock by all sockets by temporarily marking a
socket as "under memory pressure" when such pressure is detected.
While set, only the socket owner may call packet_rcv_has_room on the
socket. Once it detects normal conditions, it clears the flag. The
socket is not used as a victim by any other socket in the meantime.
Under reasonably balanced load, each socket writer frequently calls
packet_rcv_has_room and clears its own pressure field. As a backup
for when the socket is rarely written to, also clear the flag on
reading (packet_recvmsg, packet_poll) if this can be done cheaply
(i.e., without calling packet_rcv_has_room). This is only for
edge cases.
Tested:
Ran bench_rollover: a process with 8 sockets in a single fanout
group, each pinned to a single cpu that receives one nic recv
interrupt. RPS and RFS are disabled. The benchmark uses packet
rx_ring, which has to take a lock when determining whether a
socket has room.
Sent 3.5 Mpps of UDP traffic with sufficient entropy to spread
uniformly across the packet sockets (and inserted an iptables
rule to drop in PREROUTING to avoid protocol stack processing).
Without this patch, all sockets try to migrate traffic to
neighbors, causing lock contention when searching for a non-
empty neighbor. The lock is the top 9 entries.
perf record -a -g sleep 5
- 17.82% bench_rollover [kernel.kallsyms] [k] _raw_spin_lock
- _raw_spin_lock
- 99.00% spin_lock
+ 81.77% packet_rcv_has_room.isra.41
+ 18.23% tpacket_rcv
+ 0.84% packet_rcv_has_room.isra.41
+ 5.20% ksoftirqd/6 [kernel.kallsyms] [k] _raw_spin_lock
+ 5.15% ksoftirqd/1 [kernel.kallsyms] [k] _raw_spin_lock
+ 5.14% ksoftirqd/2 [kernel.kallsyms] [k] _raw_spin_lock
+ 5.12% ksoftirqd/7 [kernel.kallsyms] [k] _raw_spin_lock
+ 5.12% ksoftirqd/5 [kernel.kallsyms] [k] _raw_spin_lock
+ 5.10% ksoftirqd/4 [kernel.kallsyms] [k] _raw_spin_lock
+ 4.66% ksoftirqd/0 [kernel.kallsyms] [k] _raw_spin_lock
+ 4.45% ksoftirqd/3 [kernel.kallsyms] [k] _raw_spin_lock
+ 1.55% bench_rollover [kernel.kallsyms] [k] packet_rcv_has_room.isra.41
On net-next with this patch, this lock contention is no longer a
top entry. Most time is spent in the actual read function. Next up
are other locks:
+ 15.52% bench_rollover bench_rollover [.] reader
+ 4.68% swapper [kernel.kallsyms] [k] memcpy_erms
+ 2.77% swapper [kernel.kallsyms] [k] packet_lookup_frame.isra.51
+ 2.56% ksoftirqd/1 [kernel.kallsyms] [k] memcpy_erms
+ 2.16% swapper [kernel.kallsyms] [k] tpacket_rcv
+ 1.93% swapper [kernel.kallsyms] [k] mlx4_en_process_rx_cq
Looking closer at the remaining _raw_spin_lock, the cost of probing
in rollover is now comparable to the cost of taking the lock later
in tpacket_rcv.
- 1.51% swapper [kernel.kallsyms] [k] _raw_spin_lock
- _raw_spin_lock
+ 33.41% packet_rcv_has_room
+ 28.15% tpacket_rcv
+ 19.54% enqueue_to_backlog
+ 6.45% __free_pages_ok
+ 2.78% packet_rcv_fanout
+ 2.13% fanout_demux_rollover
+ 2.01% netif_receive_skb_internal
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace rollover state per fanout group with state per socket. Future
patches will add fields to the new structure.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Having to say
> #ifdef CONFIG_NET_NS
> struct net *net;
> #endif
in structures is a little bit wordy and a little bit error prone.
Instead it is possible to say:
> typedef struct {
> #ifdef CONFIG_NET_NS
> struct net *net;
> #endif
> } possible_net_t;
And then in a header say:
> possible_net_t net;
Which is cleaner and easier to use and easier to test, as the
possible_net_t is always there no matter what the compile options.
Further this allows read_pnet and write_pnet to be functions in all
cases which is better at catching typos.
This change adds possible_net_t, updates the definitions of read_pnet
and write_pnet, updates optional struct net * variables that
write_pnet uses on to have the type possible_net_t, and finally fixes
up the b0rked users of read_pnet and write_pnet.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
af_packet can currently overwrite kernel memory by out of bound
accesses, because it assumed a [new] block can always hold one frame.
This is not generally the case, even if most existing tools do it right.
This patch clamps too long frames as API permits, and issue a one time
error on syslog.
[ 394.357639] tpacket_rcv: packet too big, clamped from 5042 to 3966. macoff=82
In this example, packet header tp_snaplen was set to 3966,
and tp_len was set to 5042 (skb->len)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Fixes: f6fb8f100b ("af-packet: TPACKET_V3 flexible buffer implementation.")
Acked-by: Daniel Borkmann <dborkman@redhat.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In PF_PACKET's packet mmap(), we can avoid using one atomic_inc()
and one atomic_dec() call in skb destructor and use a percpu
reference count instead in order to determine if packets are
still pending to be sent out. Micro-benchmark with [1] that has
been slightly modified (that is, protcol = 0 in socket(2) and
bind(2)), example on a rather crappy testing machine; I expect
it to scale and have even better results on bigger machines:
./packet_mm_tx -s7000 -m7200 -z700000 em1, avg over 2500 runs:
With patch: 4,022,015 cyc
Without patch: 4,812,994 cyc
time ./packet_mm_tx -s64 -c10000000 em1 > /dev/null, stable:
With patch:
real 1m32.241s
user 0m0.287s
sys 1m29.316s
Without patch:
real 1m38.386s
user 0m0.265s
sys 1m35.572s
In function tpacket_snd(), it is okay to use packet_read_pending()
since in fast-path we short-circuit the condition already with
ph != NULL, since we have next frames to process. In case we have
MSG_DONTWAIT, we also do not execute this path as need_wait is
false here anyway, and in case of _no_ MSG_DONTWAIT flag, it is
okay to call a packet_read_pending(), because when we ever reach
that path, we're done processing outgoing frames anyway and only
look if there are skbs still outstanding to be orphaned. We can
stay lockless in this percpu counter since it's acceptable when we
reach this path for the sum to be imprecise first, but we'll level
out at 0 after all pending frames have reached the skb destructor
eventually through tx reclaim. When people pin a tx process to
particular CPUs, we expect overflows to happen in the reference
counter as on one CPU we expect heavy increase; and distributed
through ksoftirqd on all CPUs a decrease, for example. As
David Laight points out, since the C language doesn't define the
result of signed int overflow (i.e. rather than wrap, it is
allowed to saturate as a possible outcome), we have to use
unsigned int as reference count. The sum over all CPUs when tx
is complete will result in 0 again.
The BUG_ON() in tpacket_destruct_skb() we can remove as well. It
can _only_ be set from inside tpacket_snd() path and we made sure
to increase tx_ring.pending in any case before we called po->xmit(skb).
So testing for tx_ring.pending == 0 is not too useful. Instead, it
would rather have been useful to test if lower layers didn't orphan
the skb so that we're missing ring slots being put back to
TP_STATUS_AVAILABLE. But such a bug will be caught in user space
already as we end up realizing that we do not have any
TP_STATUS_AVAILABLE slots left anymore. Therefore, we're all set.
Btw, in case of RX_RING path, we do not make use of the pending
member, therefore we also don't need to use up any percpu memory
here. Also note that __alloc_percpu() already returns a zero-filled
percpu area, so initialization is done already.
[1] http://wiki.ipxwarzone.com/index.php5?title=Linux_packet_mmap
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces a PACKET_QDISC_BYPASS socket option, that
allows for using a similar xmit() function as in pktgen instead
of taking the dev_queue_xmit() path. This can be very useful when
PF_PACKET applications are required to be used in a similar
scenario as pktgen, but with full, flexible packet payload that
needs to be provided, for example.
On default, nothing changes in behaviour for normal PF_PACKET
TX users, so everything stays as is for applications. New users,
however, can now set PACKET_QDISC_BYPASS if needed to prevent
own packets from i) reentering packet_rcv() and ii) to directly
push the frame to the driver.
In doing so we can increase pps (here 64 byte packets) for
PF_PACKET a bit:
# CPUs -- QDISC_BYPASS -- qdisc path -- qdisc path[**]
1 CPU == 1,509,628 pps -- 1,208,708 -- 1,247,436
2 CPUs == 3,198,659 pps -- 2,536,012 -- 1,605,779
3 CPUs == 4,787,992 pps -- 3,788,740 -- 1,735,610
4 CPUs == 6,173,956 pps -- 4,907,799 -- 1,909,114
5 CPUs == 7,495,676 pps -- 5,956,499 -- 2,014,422
6 CPUs == 9,001,496 pps -- 7,145,064 -- 2,155,261
7 CPUs == 10,229,776 pps -- 8,190,596 -- 2,220,619
8 CPUs == 11,040,732 pps -- 9,188,544 -- 2,241,879
9 CPUs == 12,009,076 pps -- 10,275,936 -- 2,068,447
10 CPUs == 11,380,052 pps -- 11,265,337 -- 1,578,689
11 CPUs == 11,672,676 pps -- 11,845,344 -- 1,297,412
[...]
20 CPUs == 11,363,192 pps -- 11,014,933 -- 1,245,081
[**]: qdisc path with packet_rcv(), how probably most people
seem to use it (hopefully not anymore if not needed)
The test was done using a modified trafgen, sending a simple
static 64 bytes packet, on all CPUs. The trick in the fast
"qdisc path" case, is to avoid reentering packet_rcv() by
setting the RAW socket protocol to zero, like:
socket(PF_PACKET, SOCK_RAW, 0);
Tradeoffs are documented as well in this patch, clearly, if
queues are busy, we will drop more packets, tc disciplines are
ignored, and these packets are not visible to taps anymore. For
a pktgen like scenario, we argue that this is acceptable.
The pointer to the xmit function has been placed in packet
socket structure hole between cached_dev and prot_hook that
is hot anyway as we're working on cached_dev in each send path.
Done in joint work together with Jesper Dangaard Brouer.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Salam reported a use after free bug in PF_PACKET that occurs when
we're sending out frames on a socket bound device and suddenly the
net device is being unregistered. It appears that commit 827d9780
introduced a possible race condition between {t,}packet_snd() and
packet_notifier(). In the case of a bound socket, packet_notifier()
can drop the last reference to the net_device and {t,}packet_snd()
might end up suddenly sending a packet over a freed net_device.
To avoid reverting 827d9780 and thus introducing a performance
regression compared to the current state of things, we decided to
hold a cached RCU protected pointer to the net device and maintain
it on write side via bind spin_lock protected register_prot_hook()
and __unregister_prot_hook() calls.
In {t,}packet_snd() path, we access this pointer under rcu_read_lock
through packet_cached_dev_get() that holds reference to the device
to prevent it from being freed through packet_notifier() while
we're in send path. This is okay to do as dev_put()/dev_hold() are
per-cpu counters, so this should not be a performance issue. Also,
the code simplifies a bit as we don't need need_rls_dev anymore.
Fixes: 827d978037 ("af-packet: Use existing netdev reference for bound sockets.")
Reported-by: Salam Noureddine <noureddine@aristanetworks.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Salam Noureddine <noureddine@aristanetworks.com>
Cc: Ben Greear <greearb@candelatech.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, packet_sock has a struct tpacket_stats stats member for
TPACKET_V1 and TPACKET_V2 statistic accounting, and with TPACKET_V3
``union tpacket_stats_u stats_u'' was introduced, where however only
statistics for TPACKET_V3 are held, and when copied to user space,
TPACKET_V3 does some hackery and access also tpacket_stats' stats,
although everything could have been done within the union itself.
Unify accounting within the tpacket_stats_u union so that we can
remove 8 bytes from packet_sock that are there unnecessary. Note that
even if we switch to TPACKET_V3 and would use non mmap(2)ed option,
this still works due to the union with same types + offsets, that are
exposed to the user space.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There's a 4 byte hole in packet_ring_buffer structure before
prb_bdqc, that can be filled with 'pending' member, thus we can
reduce the overall structure size from 224 bytes to 216 bytes.
This also has the side-effect, that in struct packet_sock 2*4 byte
holes after the embedded packet_ring_buffer members are removed,
and overall, packet_sock can be reduced by 1 cacheline:
Before: size: 1344, cachelines: 21, members: 24
After: size: 1280, cachelines: 20, members: 24
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Changes:
v3->v2: rebase (no other changes)
passes selftest
v2->v1: read f->num_members only once
fix bug: test rollover mode + flag
Minimize packet drop in a fanout group. If one socket is full,
roll over packets to another from the group. Maintain flow
affinity during normal load using an rxhash fanout policy, while
dispersing unexpected traffic storms that hit a single cpu, such
as spoofed-source DoS flows. Rollover breaks affinity for flows
arriving at saturated sockets during those conditions.
The patch adds a fanout policy ROLLOVER that rotates between sockets,
filling each socket before moving to the next. It also adds a fanout
flag ROLLOVER. If passed along with any other fanout policy, the
primary policy is applied until the chosen socket is full. Then,
rollover selects another socket, to delay packet drop until the
entire system is saturated.
Probing sockets is not free. Selecting the last used socket, as
rollover does, is a greedy approach that maximizes chance of
success, at the cost of extreme load imbalance. In practice, with
sufficiently long queues to absorb bursts, sockets are drained in
parallel and load balance looks uniform in `top`.
To avoid contention, scales counters with number of sockets and
accesses them lockfree. Values are bounds checked to ensure
correctness.
Tested using an application with 9 threads pinned to CPUs, one socket
per thread and sufficient busywork per packet operation to limits each
thread to handling 32 Kpps. When sent 500 Kpps single UDP stream
packets, a FANOUT_CPU setup processes 32 Kpps in total without this
patch, 270 Kpps with the patch. Tested with read() and with a packet
ring (V1).
Also, passes psock_fanout.c unit test added to selftests.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tx data offset of packet mmap tx ring used to be :
(TPACKET2_HDRLEN - sizeof(struct sockaddr_ll))
The problem is that, with SOCK_RAW socket, the payload (14 bytes after
the beginning of the user data) is misaligned.
This patch allows to let the user gives an offset for it's tx data if
he desires.
Set sock option PACKET_TX_HAS_OFF to 1, then specify in each frame of
your tx ring tp_net for SOCK_DGRAM, or tp_mac for SOCK_RAW.
Signed-off-by: Paul Chavent <paul.chavent@onera.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reported value is the same reported by the FANOUT getsockoption, but
unlike it, the absent fanout setup results in absent nlattr, rather
than in nlattr with zero value. This is done so, since zero fanout
report may mean both -- no fanout, and fanout with both id and type zero.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The diag module will need to access some private packet_sock data, so
move it to a header in advance. This file will be shared between the
af_packet.c and the diag.c
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Willem de Bruijn
Eric W. Biederman
Eric Dumazet
Daniel Borkmann
Paul Chavent
Pavel Emelyanov