While playing with mlx4 hardware timestamping of RX packets, I found
that some packets were received by TCP stack with a ~200 ms delay...
Since the timestamp was provided by the NIC, and my probe was added
in tcp_v4_rcv() while in BH handler, I was confident it was not
a sender issue, or a drop in the network.
This would happen with a very low probability, but hurting RPC
workloads.
A NAPI driver normally arms the IRQ after the napi_complete_done(),
after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab
it.
Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit
while IRQ are not disabled, we might have later an IRQ firing and
finding this bit set, right before napi_complete_done() clears it.
This can happen with busy polling users, or if gro_flush_timeout is
used. But some other uses of napi_schedule() in drivers can cause this
as well.
thread 1 thread 2 (could be on same cpu, or not)
// busy polling or napi_watchdog()
napi_schedule();
...
napi->poll()
device polling:
read 2 packets from ring buffer
Additional 3rd packet is
available.
device hard irq
// does nothing because
NAPI_STATE_SCHED bit is owned by thread 1
napi_schedule();
napi_complete_done(napi, 2);
rearm_irq();
Note that rearm_irq() will not force the device to send an additional
IRQ for the packet it already signaled (3rd packet in my example)
This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep()
can set if it could not grab NAPI_STATE_SCHED
Then napi_complete_done() properly reschedules the napi to make sure
we do not miss something.
Since we manipulate multiple bits at once, use cmpxchg() like in
sk_busy_loop() to provide proper transactions.
In v2, I changed napi_watchdog() to use a relaxed variant of
napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point.
In v3, I added more details in the changelog and clears
NAPI_STATE_MISSED in busy_poll_stop()
In v4, I added the ideas given by Alexander Duyck in v3 review
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Steffen Klassert says:
====================
pull request (net-next): ipsec-next 2017-02-16
1) Make struct xfrm_input_afinfo const, nothing writes to it.
From Florian Westphal.
2) Remove all places that write to the afinfo policy backend
and make the struct const then.
From Florian Westphal.
3) Prepare for packet consuming gro callbacks and add
ESP GRO handlers. ESP packets can be decapsulated
at the GRO layer then. It saves a round through
the stack for each ESP packet.
Please note that this has a merge coflict between commit
63fca65d08 ("net: add confirm_neigh method to dst_ops")
from net-next and
3d7d25a68e ("xfrm: policy: remove garbage_collect callback")
a2817d8b27 ("xfrm: policy: remove family field")
from ipsec-next.
The conflict can be solved as it is done in linux-next.
Please pull or let me know if there are problems.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a skb_gro_flush_final helper to prepare for consuming
skbs in call_gro_receive. We will extend this helper to not
touch the skb if the skb is consumed by a gro callback with
a followup patch. We need this to handle the upcomming IPsec
ESP callbacks as they reinject the skb to the napi_gro_receive
asynchronous. The handler is used in all gro_receive functions
that can call the ESP gro handlers.
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
garp_port is only used in net/802/garp.c which is only compiled with
CONFIG_GARP enabled. Same goes for mrp_port which is only used in
net/802/mrp.c with CONFIG_MRP enabled.
Only include the two members in struct net_device if their respective
CONFIG_* is enabled. This saves a few bytes in struct net_device in case
CONFIG_GARP or CONFIG_MRP are not enabled.
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
The stack must not pass packets to device drivers that are shorter
than the minimum link layer header length.
Previously, packet sockets would drop packets smaller than or equal
to dev->hard_header_len, but this has false positives. Zero length
payload is used over Ethernet. Other link layer protocols support
variable length headers. Support for validation of these protocols
removed the min length check for all protocols.
Introduce an explicit dev->min_header_len parameter and drop all
packets below this value. Initially, set it to non-zero only for
Ethernet and loopback. Other protocols can follow in a patch to
net-next.
Fixes: 9ed988cd59 ("packet: validate variable length ll headers")
Reported-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 18bfb924f0 ("net: introduce default neigh_construct/destroy
ndo calls for L2 upper devices") we added these ndos to stacked devices
such as team and bond, so that calls will be propagated to mlxsw.
However, previous commit removed the reliance on these ndos and no new
users of these ndos have appeared since above mentioned commit. We can
therefore safely remove this dead code.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All __napi_complete() callers have been converted to
use the more standard napi_complete_done(),
we can now remove this NAPI method for good.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We added generic support for busy polling in NAPI layer in linux-4.5
No network driver uses ndo_busy_poll() anymore, we can get rid
of the pointer in struct net_device_ops, and its use in sk_busy_loop()
Saves NETIF_F_BUSY_POLL features bit.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit cdba756f58 ("net: move ndo_features_check() close to
ndo_start_xmit()") inadvertently moved the doc comment for
.ndo_fix_features instead of .ndo_features_check. Fix the comment
ordering.
Fixes: cdba756f58 ("net: move ndo_features_check() close to ndo_start_xmit()")
Signed-off-by: Dimitris Michailidis <dmichail@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For reporting things that may or may not be serious, depending on some
condition, netif_cond_dbg will check the condition and print the report
at either dbg (if the condition is true) or the specified level.
Suggested-by: Jon Cooper <jcooper@solarflare.com>
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 16e5cc6471 ("net: rework setup_tc ndo op to consume
general tc operand") changed the ndo_setup_tc() signature, but did not
update the comments in netdevice.h, so do that now.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Reviewed-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The network stack no longer uses the last_rx member of struct net_device
since the bonding driver switched to use its own private last_rx in
commit 9f24273837 ("bonding: use last_arp_rx in slave_last_rx()").
However, some drivers still (ab)use the field for their own purposes and
some driver just update it without actually using it.
Previously, there was an accompanying comment for the last_rx member
added in commit 4dc89133f4 ("net: add a comment on netdev->last_rx")
which asked drivers not to update is, unless really needed. However,
this commend was removed in commit f8ff080dac ("bonding: remove
useless updating of slave->dev->last_rx"), so some drivers added later
on still did update last_rx.
Remove all usage of last_rx and switch three drivers (sky2, atp and
smc91c92_cs) which actually read and write it to use their own private
copy in netdev_priv.
Compile-tested with allyesconfig and allmodconfig on x86 and arm.
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Jay Vosburgh <j.vosburgh@gmail.com>
Cc: Veaceslav Falico <vfalico@gmail.com>
Cc: Andy Gospodarek <andy@greyhouse.net>
Cc: Mirko Lindner <mlindner@marvell.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Acked-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jay Vosburgh <jay.vosburgh@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
netif_wake_subqueue() is duplicating the same thing that netif_tx_wake_queue()
does, so make it call it directly after looking up the queue from the index.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The GRO fast path caches the frag0 address. This address becomes
invalid if frag0 is modified by pskb_may_pull or its variants.
So whenever that happens we must disable the frag0 optimization.
This is usually done through the combination of gro_header_hard
and gro_header_slow, however, the IPv6 extension header path did
the pulling directly and would continue to use the GRO fast path
incorrectly.
This patch fixes it by disabling the fast path when we enter the
IPv6 extension header path.
Fixes: 78a478d0ef ("gro: Inline skb_gro_header and cache frag0 virtual address")
Reported-by: Slava Shwartsman <slavash@mellanox.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The network device operation for reading statistics is only called
in one place, and it ignores the return value. Having a structure
return value is potentially confusing because some future driver could
incorrectly assume that the return value was used.
Fix all drivers with ndo_get_stats64 to have a void function.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Oftenly, introducing side effects on packet processing on the other half
of the stack by adjusting one of TX/RX via sysctl is not desirable.
There are cases of demand for asymmetric, orthogonal configurability.
This holds true especially for nodes where RPS for RFS usage on top is
configured and therefore use the 'old dev_weight'. This is quite a
common base configuration setup nowadays, even with NICs of superior processing
support (e.g. aRFS).
A good example use case are nodes acting as noSQL data bases with a
large number of tiny requests and rather fewer but large packets as responses.
It's affordable to have large budget and rx dev_weights for the
requests. But as a side effect having this large a number on TX
processed in one run can overwhelm drivers.
This patch therefore introduces an independent configurability via sysctl to
userland.
Signed-off-by: Matthias Tafelmeier <matthias.tafelmeier@gmx.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
RFS is not commonly used, so add a jump label to avoid some conditionals
in fast path.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to support hardware offloading when the device given by the tc
rule is different from the Hardware underline device, extract the mirred
(egress) device from the tc action when a filter is added, using the new
tc_action_ops, get_dev().
Flower caches the information about the mirred device and use it for
calling ndo_setup_tc in filter change, update stats and delete.
Calling ndo_setup_tc of the mirred (egress) device instead of the
ingress device will allow a resolution between the software ingress
device and the underline hardware device.
The resolution will take place inside the offloading driver using
'egress_device' flag added to tc_to_netdev struct which is provided to
the offloading driver.
Signed-off-by: Hadar Hen Zion <hadarh@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add an IFLA_XDP_FLAGS attribute that can be passed for setting up
XDP along with IFLA_XDP_FD, which eventually allows user space to
implement typical add/replace/delete logic for programs. Right now,
calling into dev_change_xdp_fd() will always replace previous programs.
When passed XDP_FLAGS_UPDATE_IF_NOEXIST, we can handle this more
graceful when requested by returning -EBUSY in case we try to
attach a new program, but we find that another one is already
attached. This will be used by upcoming front-end for iproute2 as
well.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
sparse warns about context imbalance in any code
that uses HARD_TX_LOCK/UNLOCK - this is because it's
unable to determine that flags don't change so
lock and unlock are paired.
Seems easy enough to fix by adding __acquire/__release
calls.
With this patch af_packet.c is now sparse-clean,
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
udplite conflict is resolved by taking what 'net-next' did
which removed the backlog receive method assignment, since
it is no longer necessary.
Two entries were added to the non-priv ethtool operations
switch statement, one in 'net' and one in 'net-next, so
simple overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
