A typical qdisc setup is the following :
bond0 : bonding device, using HTB hierarchy
eth1/eth2 : slaves, multiqueue NIC, using MQ + FQ qdisc
XPS allows to spread packets on specific tx queues, based on the cpu
doing the send.
Problem is that dequeues from bond0 qdisc can happen on random cpus,
due to the fact that qdisc_run() can dequeue a batch of packets.
CPUA -> queue packet P1 on bond0 qdisc, P1->ooo_okay=1
CPUA -> queue packet P2 on bond0 qdisc, P2->ooo_okay=0
CPUB -> dequeue packet P1 from bond0
enqueue packet on eth1/eth2
CPUC -> dequeue packet P2 from bond0
enqueue packet on eth1/eth2 using sk cache (ooo_okay is 0)
get_xps_queue() then might select wrong queue for P1, since current cpu
might be different than CPUA.
P2 might be sent on the old queue (stored in sk->sk_tx_queue_mapping),
if CPUC runs a bit faster (or CPUB spins a bit on qdisc lock)
Effect of this bug is TCP reorders, and more generally not optimal
TX queue placement. (A victim bulk flow can be migrated to the wrong TX
queue for a while)
To fix this, we have to record sender cpu number the first time
dev_queue_xmit() is called for one tx skb.
We can union napi_id (used on receive path) and sender_cpu,
granted we clear sender_cpu in skb_scrub_packet() (credit to Willem for
this union idea)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Nandita Dukkipati <nanditad@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The flows are hashed on the sending node address, which allows us
to spread out the TIPC link processing to RPS enabled cores. There
is no point to include the destination address in the hash as that
will always be the same for all inbound links. We have experimented
with a 3-tuple hash over [srcnode, sport, dport], but this showed to
give slightly lower performance because of increased lock contention
when the same link was handled by multiple cores.
Signed-off-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Erik Hugne <erik.hugne@ericsson.com>
Reviewed-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch addresses a kernel unaligned access bug seen on a sparc64 system
with an igb adapter. Specifically the __skb_flow_get_ports was returning a
be32 pointer which was then having the value directly returned.
In order to prevent this it is actually easier to simply not populate the
ports or address values when an skb is not present. In this case the
assumption is that the data isn't needed and rather than slow down the
faster aligned accesses by making them have to assume the unaligned path on
architectures that don't support efficent unaligned access it makes more
sense to simply switch off the bits that were copying the source and
destination address/port for the case where we only care about the protocol
types and lengths which are normally 16 bit fields anyway.
Reported-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch updates some of the flow_dissector api so that it can be used to
parse the length of ethernet buffers stored in fragments. Most of the
changes needed were to __skb_get_poff as it needed to be updated to support
sending a linear buffer instead of a skb.
I have split __skb_get_poff into two functions, the first is skb_get_poff
and it retains the functionality of the original __skb_get_poff. The other
function is __skb_get_poff which now works much like __skb_flow_dissect in
relation to skb_flow_dissect in that it provides the same functionality but
works with just a data buffer and hlen instead of needing an skb.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Acked-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes: commit 690e36e726 (net: Allow raw buffers to be passed into the flow dissector)
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes: commit 690e36e726 (net: Allow raw buffers to be passed into the flow dissector)
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace open codings of (((u64) <x> * <y>) >> 32) with reciprocal_scale().
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Drivers, and perhaps other entities we have not yet considered,
sometimes want to know how deep the protocol headers go before
deciding how large of an SKB to allocate and how much of the packet to
place into the linear SKB area.
For example, consider a driver which has a device which DMAs into
pools of pages and then tells the driver where the data went in the
DMA descriptor(s). The driver can then build an SKB and reference
most of the data via SKB fragments (which are page/offset/length
triplets).
However at least some of the front of the packet should be placed into
the linear SKB area, which comes before the fragments, so that packet
processing can get at the headers efficiently. The first thing each
protocol layer is going to do is a "pskb_may_pull()" so we might as
well aggregate as much of this as possible while we're building the
SKB in the driver.
Part of supporting this is that we don't have an SKB yet, so we want
to be able to let the flow dissector operate on a raw buffer in order
to compute the offset of the end of the headers.
So now we have a __skb_flow_dissect() which takes an explicit data
pointer and length.
Signed-off-by: David S. Miller <davem@davemloft.net>
Add sw_hash flag to skbuff to indicate that skb->hash was computed
from flow_dissector. This flag is checked in skb_get_hash to avoid
repeatedly trying to compute the hash (ie. in the case that no L4 hash
can be computed).
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements the receive side to support RFC 6438 which is to
use the flow label as an ECMP hash. If an IPv6 flow label is set
in a packet we can use this as input for computing an L4-hash. There
should be no need to parse any transport headers in this case.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Call standard function to get a packet hash instead of taking this from
skb->sk->sk_hash or only using skb->protocol.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the hash computation located in __skb_get_hash to be a separate
function which takes flow_keys as input. This will allow flow hash
computation in other contexts where we only have addresses and ports.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb_flow_dissect() dissects only transport header type in ip_proto. It dose not
give any information about IPv4 or IPv6.
This patch adds new member, n_proto, to struct flow_keys. Which records the
IP layer type. i.e IPv4 or IPv6.
This can be used in netdev->ndo_rx_flow_steer driver function to dissect flow.
Adding new member to flow_keys increases the struct size by around 4 bytes.
This causes BUILD_BUG_ON(sizeof(qcb->data) < sz); to fail in
qdisc_cb_private_validate()
So increase data size by 4
Signed-off-by: Govindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The packet hash can be considered a property of the packet, not just
on RX path.
This patch changes name of rxhash and l4_rxhash skbuff fields to be
hash and l4_hash respectively. This includes changing uses of the
field in the code which don't call the access functions.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Mahesh Bandewar <maheshb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The use of __constant_<foo> has been unnecessary for quite awhile now.
Make these uses consistent with the rest of the kernel.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to allow users to invoke netdev_cap_txqueue, it needs to
be moved into netdevice.h header file. While at it, also add kernel
doc header to document the API.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a new argument for ndo_select_queue() callback that passes a
fallback handler. This gets invoked through netdev_pick_tx();
fallback handler is currently __netdev_pick_tx() as most drivers
invoke this function within their customized implementation in
case for skbs that don't need any special handling. This fallback
handler can then be replaced on other call-sites with different
queue selection methods (e.g. in packet sockets, pktgen etc).
This also has the nice side-effect that __netdev_pick_tx() is
then only invoked from netdev_pick_tx() and export of that
function to modules can be undone.
Suggested-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the tx queue were selected implicitly in ndo_dfwd_start_xmit(). The
will cause several issues:
- NETIF_F_LLTX were removed for macvlan, so txq lock were done for macvlan
instead of lower device which misses the necessary txq synchronization for
lower device such as txq stopping or frozen required by dev watchdog or
control path.
- dev_hard_start_xmit() was called with NULL txq which bypasses the net device
watchdog.
- dev_hard_start_xmit() does not check txq everywhere which will lead a crash
when tso is disabled for lower device.
Fix this by explicitly introducing a new param for .ndo_select_queue() for just
selecting queues in the case of l2 forwarding offload. netdev_pick_tx() was also
extended to accept this parameter and dev_queue_xmit_accel() was used to do l2
forwarding transmission.
With this fixes, NETIF_F_LLTX could be preserved for macvlan and there's no need
to check txq against NULL in dev_hard_start_xmit(). Also there's no need to keep
a dedicated ndo_dfwd_start_xmit() and we can just reuse the code of
dev_queue_xmit() to do the transmission.
In the future, it was also required for macvtap l2 forwarding support since it
provides a necessary synchronization method.
Cc: John Fastabend <john.r.fastabend@intel.com>
Cc: Neil Horman <nhorman@tuxdriver.com>
Cc: e1000-devel@lists.sourceforge.net
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Changing name of function as part of making the hash in skbuff to be
generic property, not just for receive path.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
By moving code around, we avoid :
1) A reload of iph->ihl (bit field, so needs a mask)
2) A conditional test (replaced by a conditional mov on x86)
Fast path loads iph->protocol anyway.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/emulex/benet/be.h
drivers/net/netconsole.c
net/bridge/br_private.h
Three mostly trivial conflicts.
The net/bridge/br_private.h conflict was a function signature (argument
addition) change overlapping with the extern removals from Joe Perches.
In drivers/net/netconsole.c we had one change adjusting a printk message
whilst another changed "printk(KERN_INFO" into "pr_info(".
Lastly, the emulex change was a new inline function addition overlapping
with Joe Perches's extern removals.
Signed-off-by: David S. Miller <davem@davemloft.net>
We don't validate iph->ihl which may lead a dead loop if we meet a IPIP
skb whose iph->ihl is zero. Fix this by failing immediately when iph->ihl
is evil (less than 5).
This issue were introduced by commit ec5efe7946
(rps: support IPIP encapsulation).
Cc: Eric Dumazet <edumazet@google.com>
Cc: Petr Matousek <pmatouse@redhat.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We also can defer the initialization of hashrnd in flow_dissector
to its first use. Since net_get_random_once is irq safe now we don't
have to audit the call paths if one of this functions get called by an
interrupt handler.
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Factor out the code that extracts the ports from skb_flow_dissect and
add a new function skb_flow_get_ports which can be re-used.
Suggested-by: Veaceslav Falico <vfalico@redhat.com>
Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Veaceslav Falico <vfalico@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 8ed781668d ("flow_keys: include thoff into flow_keys for
later usage"), we missed that existing code was using nhoff as a
temporary variable that could not always contain transport header
offset.
This is not a problem for TCP/UDP because port offset (@poff)
is 0 for these protocols.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <dborkman@redhat.com>
Cc: Nikolay Aleksandrov <nikolay@redhat.com>
Acked-by: Nikolay Aleksandrov <nikolay@redhat.com>
Acked-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit 416186fbf8 ("net: Split core bits of netdev_pick_tx
into __netdev_pick_tx") added a bug that disables caching of queue
index in the socket.
This is the source of packet reorders for TCP flows, and
again this is happening more often when using FQ pacing.
Old code was doing
if (queue_index != old_index)
sk_tx_queue_set(sk, queue_index);
Alexander renamed the variables but forgot to change sk_tx_queue_set()
2nd parameter.
if (queue_index != new_index)
sk_tx_queue_set(sk, queue_index);
This means we store -1 over and over in sk->sk_tx_queue_mapping
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Acked-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/stmicro/stmmac/stmmac_platform.c
net/bridge/br_multicast.c
net/ipv6/sit.c
The conflicts were minor:
1) sit.c changes overlap with change to ip_tunnel_xmit() signature.
2) br_multicast.c had an overlap between computing max_delay using
msecs_to_jiffies and turning MLDV2_MRC() into an inline function
with a name using lowercase instead of uppercase letters.
3) stmmac had two overlapping changes, one which conditionally allocated
and hooked up a dma_cfg based upon the presence of the pbl OF property,
and another one handling store-and-forward DMA made. The latter of
which should not go into the new of_find_property() basic block.
Signed-off-by: David S. Miller <davem@davemloft.net>
commit 8728c544a9 ("net: dev_pick_tx() fix") and commit
b6fe83e952 ("bonding: refine IFF_XMIT_DST_RELEASE capability")
are quite incompatible : Queue selection is disabled because skb
dst was dropped before entering bonding device.
This causes major performance regression, mainly because TCP packets
for a given flow can be sent to multiple queues.
This is particularly visible when using the new FQ packet scheduler
with MQ + FQ setup on the slaves.
We can safely revert the first commit now that 416186fbf8
("net: Split core bits of netdev_pick_tx into __netdev_pick_tx")
properly caps the queue_index.
Reported-by: Xi Wang <xii@google.com>
Diagnosed-by: Xi Wang <xii@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Cc: Denys Fedorysychenko <nuclearcat@nuclearcat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Same behavior than 802.1q : finds the encapsulated protocol and
skip 32bit header.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Explicitly set proto to ETH_P_IP and jump directly to ip processing.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
__skb_get_poff() returns the offset to the payload as far as it could
be dissected. The main user is currently BPF, so that we can dynamically
truncate packets without needing to push actual payload to the user
space and instead can analyze headers only.
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In skb_flow_dissect(), we perform a dissection of a skbuff. Since we're
doing the work here anyway, also store thoff for a later usage, e.g. in
the BPF filter.
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for L2 GRE tunnels, so that RPS can be more effective.
Signed-off-by: Michael Dalton <mwdalton@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
__skb_tx_hash() and __skb_get_rxhash() are all for calculating hash
value based by some fields in skb, mostly used for selecting queues
by device drivers.
Meanwhile, net/core/dev.c is bloating.
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce ipv6_addr_hash() helper doing a XOR on all bits
of an IPv6 address, with an optimized x86_64 version.
Use it in flow dissector, as suggested by Andrew McGregor,
to reduce hash collision probabilities in fq_codel (and other
users of flow dissector)
Use it in ip6_tunnel.c and use more bit shuffling, as suggested
by David Laight, as existing hash was ignoring most of them.
Use it in sunrpc and use more bit shuffling, using hash_32().
Use it in net/ipv6/addrconf.c, using hash_32() as well.
As a cleanup, use it in net/ipv4/tcp_metrics.c
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Andrew McGregor <andrewmcgr@gmail.com>
Cc: Dave Taht <dave.taht@gmail.com>
Cc: Tom Herbert <therbert@google.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The file net/core/flow_dissector.c seems to be missing
including linux/export.h.
Signed-off-by: Jesper Dangaard Brouer <hawk@comx.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Le lundi 28 novembre 2011 à 19:06 -0500, David Miller a écrit :
> From: Dimitris Michailidis <dm@chelsio.com>
> Date: Mon, 28 Nov 2011 08:25:39 -0800
>
> >> +bool skb_flow_dissect(const struct sk_buff *skb, struct flow_keys
> >> *flow)
> >> +{
> >> + int poff, nhoff = skb_network_offset(skb);
> >> + u8 ip_proto;
> >> + u16 proto = skb->protocol;
> >
> > __be16 instead of u16 for proto?
>
> I'll take care of this when I apply these patches.
( CC trimmed )
Thanks David !
Here is a small patch to use one 64bit load/store on x86_64 instead of
two 32bit load/stores.
[PATCH net-next] flow_dissector: use a 64bit load/store
gcc compiler is smart enough to use a single load/store if we
memcpy(dptr, sptr, 8) on x86_64, regardless of
CONFIG_CC_OPTIMIZE_FOR_SIZE
In IP header, daddr immediately follows saddr, this wont change in the
future. We only need to make sure our flow_keys (src,dst) fields wont
break the rule.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We use at least two flow dissectors in network stack, with known
limitations and code duplication.
Introduce skb_flow_dissect() to factorize this, highly inspired from
existing dissector from __skb_get_rxhash()
Note : We extensively use skb_header_pointer(), this permits us to not
touch skb at all.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet
Erik Hugne
Alexander Duyck
Alexander Duyck
WANG Cong
Daniel Borkmann
David S. Miller
Tom Herbert
Govindarajulu Varadarajan
Joe Perches
Jason Wang
Hannes Frederic Sowa
Nikolay Aleksandrov
Michael Dalton
Jesper Dangaard Brouer
Eric Dumazet