commit b30973f8 (node-aware skb allocation) spread a wrong habit of
allocating net drivers skbs on a given memory node : The one closest to
the NIC hardware. This is wrong because as soon as we try to scale
network stack, we need to use many cpus to handle traffic and hit
slub/slab management on cross-node allocations/frees when these cpus
have to alloc/free skbs bound to a central node.

skb allocated in RX path are ephemeral, they have a very short
lifetime : Extra cost to maintain NUMA affinity is too expensive. What
appeared as a nice idea four years ago is in fact a bad one.

In 2010, NIC hardwares are multiqueue, or we use RPS to spread the load,
and two 10Gb NIC might deliver more than 28 million packets per second,
needing all the available cpus.

Cost of cross-node handling in network and vm stacks outperforms the
small benefit hardware had when doing its DMA transfert in its 'local'
memory node at RX time. Even trying to differentiate the two allocations
done for one skb (the sk_buff on local node, the data part on NIC
hardware node) is not enough to bring good performance.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Acked-by: NTom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The patch: "gro: fix different skb headrooms" in its part:
"2) allocate a minimal skb for head of frag_list" is buggy. The copied
skb has p->data set at the ip header at the moment, and skb_gro_offset
is the length of ip + tcp headers. So, after the change the length of
mac header is skipped. Later skb_set_mac_header() sets it into the
NET_SKB_PAD area (if it's long enough) and ip header is misaligned at
NET_SKB_PAD + NET_IP_ALIGN offset. There is no reason to assume the
original skb was wrongly allocated, so let's copy it as it was.

bugzilla : https://bugzilla.kernel.org/show_bug.cgi?id=16626
fixes commit: 3d3be433Reported-by: NPlamen Petrov <pvp-lsts@fs.uni-ruse.bg>
Signed-off-by: NJarek Poplawski <jarkao2@gmail.com>
CC: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: NEric Dumazet <eric.dumazet@gmail.com>
Tested-by: NPlamen Petrov <pvp-lsts@fs.uni-ruse.bg>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

pskb_expand_head() blindly takes references on fragments before calling
skb_release_data(), potentially releasing these references.

We can add a fast path, avoiding these atomic operations, if we own the
last reference on skb->head.

Based on a previous patch from David
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

__alloc_skb() uses a memset() to clear all the beginning of skb,
including bitfields contained in 'flags1' & 'flags2'.

We dont need any more to use kmemcheck_annotate_bitfield() on these
fields. However, we still need it for the clone part, which is not
cleared.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Packets entering GRO might have different headrooms, even for a given
flow (because of implementation details in drivers, like copybreak).
We cant force drivers to deliver packets with a fixed headroom.

1) fix skb_segment()

skb_segment() makes the false assumption headrooms of fragments are same
than the head. When CHECKSUM_PARTIAL is used, this can give csum_start
errors, and crash later in skb_copy_and_csum_dev()

2) allocate a minimal skb for head of frag_list

skb_gro_receive() uses netdev_alloc_skb(headroom + skb_gro_offset(p)) to
allocate a fresh skb. This adds NET_SKB_PAD to a padding already
provided by netdevice, depending on various things, like copybreak.

Use alloc_skb() to allocate an exact padding, to reduce cache line
needs:
NET_SKB_PAD + NET_IP_ALIGN

bugzilla : https://bugzilla.kernel.org/show_bug.cgi?id=16626

Many thanks to Plamen Petrov, testing many debugging patches !
With help of Jarek Poplawski.
Reported-by: NPlamen Petrov <pvp-lsts@fs.uni-ruse.bg>
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
CC: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

(skb->data - skb->head) can be changed by skb_headroom(skb)

Remove some uses of NET_SKBUFF_DATA_USES_OFFSET, using
(skb_end_pointer(skb) - skb->head) or
(skb_tail_pointer(skb) - skb->head) : compiler does the right thing,
and this is more readable for us ;)

(struct skb_shared_info *) casts in pskb_expand_head() to help memcpy()
to use aligned moves.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

SKBs can be "fragmented" in two ways, via a page array (called
skb_shinfo(skb)->frags[]) and via a list of SKBs (called
skb_shinfo(skb)->frag_list).

Since skb_has_frags() tests the latter, it's name is confusing
since it sounds more like it's testing the former.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch removes the abstraction introduced by the union skb_shared_tx in
the shared skb data.

The access of the different union elements at several places led to some
confusion about accessing the shared tx_flags e.g. in skb_orphan_try().

    http://marc.info/?l=linux-netdev&m=128084897415886&w=2Signed-off-by: NOliver Hartkopp <socketcan@hartkopp.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Move frags[] at the end of struct skb_shared_info, and make
pskb_expand_head() copy only the used part of it instead of whole array.

This should avoid kmemcheck warnings and speedup pskb_expand_head() as
well, avoiding a lot of cache misses.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It should only be adjusted if ip_summed == CHECKSUM_PARTIAL.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make pskb_expand_head() check ip_summed to make sure csum_start is really
csum_start and not csum before adjusting it.

This fixes a bug I encountered using a Sun Quad-Fast Ethernet card and VLANs.
On my configuration, the sunhme driver produces skbs with differing amounts
of headroom on receive depending on the packet size.  See line 2030 of
drivers/net/sunhme.c; packets smaller than RX_COPY_THRESHOLD have 52 bytes
of headroom but packets larger than that cutoff have only 20 bytes.

When these packets reach the VLAN driver, vlan_check_reorder_header()
calls skb_cow(), which, if the packet has less than NET_SKB_PAD (== 32) bytes
of headroom, uses pskb_expand_head() to make more.

Then, pskb_expand_head() needs to adjust a lot of offsets into the skb,
including csum_start.  Since csum_start is a union with csum, if the packet
has a valid csum value this will corrupt it, which was the effect I observed.
The sunhme hardware computes receive checksums, so the skbs would be created
by the driver with ip_summed == CHECKSUM_COMPLETE and a valid csum field, and
then pskb_expand_head() would corrupt the csum field, leading to an "hw csum
error" message later on, for example in icmp_rcv() for pings larger than the
sunhme RX_COPY_THRESHOLD.

On the basis of the comment at the beginning of include/linux/skbuff.h,
I believe that the csum_start skb field is only meaningful if ip_csummed is
CSUM_PARTIAL, so this patch makes pskb_expand_head() adjust it only in that
case to avoid corrupting a valid csum value.

Please see my more in-depth disucssion of tracking down this bug for
more details if you like:

http://puellavulnerata.livejournal.com/112186.html
http://puellavulnerata.livejournal.com/112567.html
http://puellavulnerata.livejournal.com/112891.html
http://puellavulnerata.livejournal.com/113096.html
http://puellavulnerata.livejournal.com/113591.html

I am not subscribed to this list, so please CC me on replies.
Signed-off-by: NAndrea Shepard <andrea@persephoneslair.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

CodingStyle cleanups

EXPORT_SYMBOL should immediately follow the symbol declaration.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

No need to copy rxhash again in __skb_clone()
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

deliver_no_wcard is not being set in skb_copy_header.
In the skb_cloned case it is not being cleared and
may cause the skb to be dropped when the loopback device
pushes it back up the stack.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <eric.dumazet@gmail.com>
Tested-by: NMarkus Trippelsdorf <markus@trippelsdorf.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Correct sk_forward_alloc handling for error_queue would need to use a
backlog of frames that softirq handler could not deliver because socket
is owned by user thread. Or extend backlog processing to be able to
process normal and error packets.

Another possibility is to not use mem charge for error queue, this is
what I implemented in this patch.

Note: this reverts commit 29030374
(net: fix sk_forward_alloc corruptions), since we dont need to lock
socket anymore.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As David found out, sock_queue_err_skb() should be called with socket
lock hold, or we risk sk_forward_alloc corruption, since we use non
atomic operations to update this field.

This patch adds bh_lock_sock()/bh_unlock_sock() pair to three spots.
(BH already disabled)

1) skb_tstamp_tx() 
2) Before calling ip_icmp_error(), in __udp4_lib_err() 
3) Before calling ipv6_icmp_error(), in __udp6_lib_err()
Reported-by: NAnton Blanchard <anton@samba.org>
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NChangli Gao <xiaosuo@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds F_GETPIPE_SZ and F_SETPIPE_SZ fcntl() actions for
growing and shrinking the size of a pipe and adjusts pipe.c and splice.c
(and relay and network splice) usage to work with these larger (or smaller)
pipes.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

When GRO produces fraglist entries, and the resulting skb hits
an interface that is incapable of TSO but capable of FRAGLIST,
we end up producing a bogus packet with gso_size non-zero.

This was reported in the field with older versions of KVM that
did not set the TSO bits on tuntap.

This patch fixes that.
Reported-by: NIgor Zhang <yugzhang@redhat.com>
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Use low order bit of skb->_skb_dst to tell dst is not refcounted.

Change _skb_dst to _skb_refdst to make sure all uses are catched.

skb_dst() returns the dst, regardless of noref bit set or not, but
with a lockdep check to make sure a noref dst is not given if current
user is not rcu protected.

New skb_dst_set_noref() helper to set an notrefcounted dst on a skb.
(with lockdep check)

skb_dst_drop() drops a reference only if skb dst was refcounted.

skb_dst_force() helper is used to force a refcount on dst, when skb
is queued and not anymore RCU protected.

Use skb_dst_force() in __sk_add_backlog(), __dev_xmit_skb() if
!IFF_XMIT_DST_RELEASE or skb enqueued on qdisc queue, in
sock_queue_rcv_skb(), in __nf_queue().

Use skb_dst_force() in dev_requeue_skb().

Note: dst_use_noref() still dirties dst, we might transform it
later to do one dirtying per jiffies.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

With following patch I can reach maximum rate of my pktgen+udpsink
simulator :
- 'old' machine : dual quad core E5450  @3.00GHz
- 64 UDP rx flows (only differ by destination port)
- RPS enabled, NIC interrupts serviced on cpu0
- rps dispatched on 7 other cores. (~130.000 IPI per second)
- SLAB allocator (faster than SLUB in this workload)
- tg3 NIC
- 1.080.000 pps without a single drop at NIC level.

Idea is to add two prefetchw() calls in __alloc_skb(), one to prefetch
first sk_buff cache line, the second to prefetch the shinfo part.

Also using one memset() to initialize all skb_shared_info fields instead
of one by one to reduce number of instructions, using long word moves.

All skb_shared_info fields before 'dataref' are cleared in 
__alloc_skb().
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In commit 6be8ac2f ("[NET]: uninline skb_pull, de-bloats a lot")
we uninlined skb_pull.

But in some critical paths it makes sense to inline this thing
and it helps performance significantly.

Create an skb_pull_inline() so that we can do this in a way that
serves also as annotation.

Based upon a patch by Eric Dumazet.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There is no need to export skb_under_panic() and skb_over_panic() in
skbuff.c, since these methods are used only in skbuff.c ; this patch
removes these two exports. It also marks these functions as 'static'
and removeS the extern declarations of them from
include/linux/skbuff.h
Signed-off-by: NRami Rosen <ramirose@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch implements software receive side packet steering (RPS).  RPS
distributes the load of received packet processing across multiple CPUs.

Problem statement: Protocol processing done in the NAPI context for received
packets is serialized per device queue and becomes a bottleneck under high
packet load.  This substantially limits pps that can be achieved on a single
queue NIC and provides no scaling with multiple cores.

This solution queues packets early on in the receive path on the backlog queues
of other CPUs.   This allows protocol processing (e.g. IP and TCP) to be
performed on packets in parallel.   For each device (or each receive queue in
a multi-queue device) a mask of CPUs is set to indicate the CPUs that can
process packets. A CPU is selected on a per packet basis by hashing contents
of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index
into the CPU mask.  The IPI mechanism is used to raise networking receive
softirqs between CPUs.  This effectively emulates in software what a multi-queue
NIC can provide, but is generic requiring no device support.

Many devices now provide a hash over the 4-tuple on a per packet basis
(e.g. the Toeplitz hash).  This patch allow drivers to set the HW reported hash
in an skb field, and that value in turn is used to index into the RPS maps.
Using the HW generated hash can avoid cache misses on the packet when
steering it to a remote CPU.

The CPU mask is set on a per device and per queue basis in the sysfs variable
/sys/class/net/<device>/queues/rx-<n>/rps_cpus.  This is a set of canonical
bit maps for receive queues in the device (numbered by <n>).  If a device
does not support multi-queue, a single variable is used for the device (rx-0).

Generally, we have found this technique increases pps capabilities of a single
queue device with good CPU utilization.  Optimal settings for the CPU mask
seem to depend on architectures and cache hierarcy.  Below are some results
running 500 instances of netperf TCP_RR test with 1 byte req. and resp.
Results show cumulative transaction rate and system CPU utilization.

e1000e on 8 core Intel
   Without RPS: 108K tps at 33% CPU
   With RPS:    311K tps at 64% CPU

forcedeth on 16 core AMD
   Without RPS: 156K tps at 15% CPU
   With RPS:    404K tps at 49% CPU

bnx2x on 16 core AMD
   Without RPS  567K tps at 61% CPU (4 HW RX queues)
   Without RPS  738K tps at 96% CPU (8 HW RX queues)
   With RPS:    854K tps at 76% CPU (4 HW RX queues)

Caveats:
- The benefits of this patch are dependent on architecture and cache hierarchy.
Tuning the masks to get best performance is probably necessary.
- This patch adds overhead in the path for processing a single packet.  In
a lightly loaded server this overhead may eliminate the advantages of
increased parallelism, and possibly cause some relative performance degradation.
We have found that masks that are cache aware (share same caches with
the interrupting CPU) mitigate much of this.
- The RPS masks can be changed dynamically, however whenever the mask is changed
this introduces the possibility of generating out of order packets.  It's
probably best not change the masks too frequently.
Signed-off-by: NTom Herbert <therbert@google.com>

 include/linux/netdevice.h |   32 ++++-
 include/linux/skbuff.h    |    3 +
 net/core/dev.c            |  335 +++++++++++++++++++++++++++++++++++++--------
 net/core/net-sysfs.c      |  225 ++++++++++++++++++++++++++++++-
 net/core/skbuff.c         |    2 +
 5 files changed, 538 insertions(+), 59 deletions(-)
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NAlexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

To help grep games, rename iif to skb_iif
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When we've merged skb's with page frags, and subsequently receive
a trailer skb (< MSS) that is not completely non-linear (this can
occur on Intel NICs if the packet size falls below the threshold),
GRO ends up producing an illegal GSO skb with a frag_list.

This is harmless unless the skb is then forwarded through an
interface that requires software GSO, whereupon the GSO code
will BUG.

This patch detects this case in GRO and avoids merging the
trailer skb.
Reported-by: NMark Wagner <mwagner@redhat.com>
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

NAPI drivers try to recycle SKBs in their polling routine, but we
generally don't know the context in which the polling will be called,
and the skb recycling itself may require IRQs to be enabled.

This patch adds irqs_disabled() test to the skb_recycle_check()
routine, so that we'll not let the drivers hit the skb recycling
path with IRQs disabled.

As a side effect, this patch actually disables skb recycling for some
[broken] drivers. E.g. gianfar driver grabs an irqsave spinlock during
TX ring processing, and then tries to recycle an skb, and that caused
the following badness:

nf_conntrack version 0.5.0 (1008 buckets, 4032 max)
------------[ cut here ]------------
Badness at kernel/softirq.c:143
NIP: c003e3c4 LR: c423a528 CTR: c003e344
...
NIP [c003e3c4] local_bh_enable+0x80/0xc4
LR [c423a528] destroy_conntrack+0xd4/0x13c [nf_conntrack]
Call Trace:
[c15d1b60] [c003e32c] local_bh_disable+0x1c/0x34 (unreliable)
[c15d1b70] [c423a528] destroy_conntrack+0xd4/0x13c [nf_conntrack]
[c15d1b80] [c02c6370] nf_conntrack_destroy+0x3c/0x70
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

mac80211 required this due to the master netdev, but now
it can put all information into skb->cb and this can go.
Signed-off-by: NJohannes Berg <johannes@sipsolutions.net>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

The skb mac_header field is sometimes NULL (or ~0u) as a sentinel
value. The places where skb is expanded add an offset which would
change this flag into an invalid pointer (or offset).
Signed-off-by: NStephen Hemminger <shemminger@vyatta.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Looking at the crash in log_martians(), one suspect is that the check for
mac header being set is not correct.  The value of mac_header defaults to
0 on allocation, therefore skb_mac_header_was_set will always be true on
platforms using NET_SKBUFF_USES_OFFSET.
Signed-off-by: NStephen Hemminger <shemminger@vyatta.com>
Acked-by: NArnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NVegard Nossum <vegard.nossum@gmail.com>

In order to handle powersave frames properly we had needed
to pass these out to the device queues again, and introduce
the skb->requeue bit. This, however, also has unnecessary
overhead by needing to 'clean up' already tried frames, and
this clean-up code is also buggy when software encryption
is used.

Instead of sending the frames via the master netdev queue
again, simply put them into the pending queue. This also
fixes a problem where frames for that particular station
could be reordered when some were still on the software
queues and older ones are re-injected into the software
queue after them.
Signed-off-by: NJohannes Berg <johannes@sipsolutions.net>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Signed-off-by: NDavid S. Miller <davem@davemloft.net>

On Thu, Jun 04, 2009 at 09:06:00PM +1000, Herbert Xu wrote:
>
> tun: Optimise handling of bogus gso->hdr_len
>
> As all current versions of virtio_net generate a value for the
> header length that's too small, we should optimise this so that
> we don't copy it twice.  This can be done by ensuring that it is
> at least as large as the place where we'll write the checksum.
>
> Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

With this applied we can strengthen the partial checksum check:

In skb_partial_csum_set we check to see if the checksum offset
is within the packet.  However, we really should check that it
is within the skb head as that's the only bit we can modify
without copying.
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Acked-by: NRusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Define three accessors to get/set dst attached to a skb

struct dst_entry *skb_dst(const struct sk_buff *skb)

void skb_dst_set(struct sk_buff *skb, struct dst_entry *dst)

void skb_dst_drop(struct sk_buff *skb)
This one should replace occurrences of :
dst_release(skb->dst)
skb->dst = NULL;

Delete skb->dst field
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch stores the two shinfo pointers in local variables
because they're used over and over again in skb_gro_receive.
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch reverses the direction of the frags array copy in
skb_gro_receive in order simplify the loop conditional.  It
also avoids touching the first element of the original frags
array.
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch stores the offset/headlen in local variables as they're
used repeatedly in skb_gro_offset.
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

gcc does a poor job at generating code for the memcpy of the frags
array in skb_gro_receive, which is the primary purpose of that
function when merging frags.  In particular, it can't utilise the
alignment information of the source and destination.  This patch
open-codes the copy so we process words instead of bytes.
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>