This reverts commit 4465b46900.
Conflicts:
net/ipv4/fib_frontend.c
As reported by Ben Greear, this causes regressions:
> Change 4465b46900 caused rules
> to stop matching the input device properly because the
> FLOWI_FLAG_MATCH_ANY_IIF is always defined in ip_dev_find().
>
> This breaks rules such as:
>
> ip rule add pref 512 lookup local
> ip rule del pref 0 lookup local
> ip link set eth2 up
> ip -4 addr add 172.16.0.102/24 broadcast 172.16.0.255 dev eth2
> ip rule add to 172.16.0.102 iif eth2 lookup local pref 10
> ip rule add iif eth2 lookup 10001 pref 20
> ip route add 172.16.0.0/24 dev eth2 table 10001
> ip route add unreachable 0/0 table 10001
>
> If you had a second interface 'eth0' that was on a different
> subnet, pinging a system on that interface would fail:
>
> [root@ct503-60 ~]# ping 192.168.100.1
> connect: Invalid argument
Reported-by: Ben Greear <greearb@candelatech.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Special care is taken inside sk_port_alloc to avoid overwriting
skc_node/skc_nulls_node. We should also avoid overwriting
skc_bind_node/skc_portaddr_node.
The patch fixes the following crash:
BUG: unable to handle kernel paging request at fffffffffffffff0
IP: [<ffffffff812ec6dd>] udp4_lib_lookup2+0xad/0x370
[<ffffffff812ecc22>] __udp4_lib_lookup+0x282/0x360
[<ffffffff812ed63e>] __udp4_lib_rcv+0x31e/0x700
[<ffffffff812bba45>] ? ip_local_deliver_finish+0x65/0x190
[<ffffffff812bbbf8>] ? ip_local_deliver+0x88/0xa0
[<ffffffff812eda35>] udp_rcv+0x15/0x20
[<ffffffff812bba45>] ip_local_deliver_finish+0x65/0x190
[<ffffffff812bbbf8>] ip_local_deliver+0x88/0xa0
[<ffffffff812bb2cd>] ip_rcv_finish+0x32d/0x6f0
[<ffffffff8128c14c>] ? netif_receive_skb+0x99c/0x11c0
[<ffffffff812bb94b>] ip_rcv+0x2bb/0x350
[<ffffffff8128c14c>] netif_receive_skb+0x99c/0x11c0
Signed-off-by: Leonard Crestez <lcrestez@ixiacom.com>
Signed-off-by: Octavian Purdila <opurdila@ixiacom.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After commit c1f19b51d1 (net: support time stamping in phy devices.),
kernel might crash if CONFIG_NETWORK_PHY_TIMESTAMPING=y and
skb_defer_rx_timestamp() handles a packet without an ethernet header.
Fixes kernel bugzilla #24102
Reference: https://bugzilla.kernel.org/show_bug.cgi?id=24102
Reported-and-tested-by: Andrew Watts <akwatts@ymail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pavel Emelyanov tried to fix a race between sk_filter_(de|at)tach and
sk_clone() in commit 47e958eac2
Problem is we can have several clones sharing a common sk_filter, and
these clones might want to sk_filter_attach() their own filters at the
same time, and can overwrite old_filter->rcu, corrupting RCU queues.
We can not use filter->rcu without being sure no other thread could do
the same thing.
Switch code to a more conventional ref-counting technique : Do the
atomic decrement immediately and queue one rcu call back when last
reference is released.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We forgot to use __GFP_HIGHMEM in several __vmalloc() calls.
In ceph, add the missing flag.
In fib_trie.c, xfrm_hash.c and request_sock.c, using vzalloc() is
cleaner and allows using HIGHMEM pages as well.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix kernel-doc warning for sk_filter_rcu_release():
Warning(net/core/filter.c:586): missing initial short description on line:
* sk_filter_rcu_release: Release a socket filter by rcu_head
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: netdev@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
netif_set_real_num_rx_queues() can decrement and increment
the number of rx queues. For example ixgbe does this as
features and offloads are toggled. Presumably this could
also happen across down/up on most devices if the available
resources changed (cpu offlined).
The kobject needs to be zero'd in this case so that the
state is not preserved across kobject_put()/kobject_init_and_add().
This resolves the following error report.
ixgbe 0000:03:00.0: eth2: NIC Link is Up 10 Gbps, Flow Control: RX/TX
kobject (ffff880324b83210): tried to init an initialized object, something is seriously wrong.
Pid: 1972, comm: lldpad Not tainted 2.6.37-rc18021qaz+ #169
Call Trace:
[<ffffffff8121c940>] kobject_init+0x3a/0x83
[<ffffffff8121cf77>] kobject_init_and_add+0x23/0x57
[<ffffffff8107b800>] ? mark_lock+0x21/0x267
[<ffffffff813c6d11>] net_rx_queue_update_kobjects+0x63/0xc6
[<ffffffff813b5e0e>] netif_set_real_num_rx_queues+0x5f/0x78
[<ffffffffa0261d49>] ixgbe_set_num_queues+0x1c6/0x1ca [ixgbe]
[<ffffffffa0262509>] ixgbe_init_interrupt_scheme+0x1e/0x79c [ixgbe]
[<ffffffffa0274596>] ixgbe_dcbnl_set_state+0x167/0x189 [ixgbe]
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
nlmsg_total_size() calculates the length of a netlink message
including header and alignment. nla_total_size() calculates the
space an individual attribute consumes which was meant to be used
in this context.
Also, ensure to account for the attribute header for the
IFLA_INFO_XSTATS attribute as implementations of get_xstats_size()
seem to assume that we do so.
The addition of two message headers minus the missing attribute
header resulted in a calculated message size that was larger than
required. Therefore we never risked running out of skb tailroom.
Signed-off-by: Thomas Graf <tgraf@infradead.org>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Robin Holt tried to boot a 16TB machine and found some limits were
reached : sysctl_tcp_mem[2], sysctl_udp_mem[2]
We can switch infrastructure to use long "instead" of "int", now
atomic_long_t primitives are available for free.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Reported-by: Robin Holt <holt@sgi.com>
Reviewed-by: Robin Holt <holt@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is a possibility malicious users can get limited information about
uninitialized stack mem array. Even if sk_run_filter() result is bound
to packet length (0 .. 65535), we could imagine this can be used by
hostile user.
Initializing mem[] array, like Dan Rosenberg suggested in his patch is
expensive since most filters dont even use this array.
Its hard to make the filter validation in sk_chk_filter(), because of
the jumps. This might be done later.
In this patch, I use a bitmap (a single long var) so that only filters
using mem[] loads/stores pay the price of added security checks.
For other filters, additional cost is a single instruction.
[ Since we access fentry->k a lot now, cache it in a local variable
and mark filter entry pointer as const. -DaveM ]
Reported-by: Dan Rosenberg <drosenberg@vsecurity.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Followup of commit ef885afbf8 (net: use rcu_barrier() in
rollback_registered_many)
dst_dev_event() scans a garbage dst list that might be feeded by various
network notifiers at device dismantle time.
Its important to call dst_dev_event() after other notifiers, or we might
enter the infamous msleep(250) in netdev_wait_allrefs(), and wait one
second before calling again call_netdevice_notifiers(NETDEV_UNREGISTER,
dev) to properly remove last device references.
Use priority -10 to let dst_dev_notifier be called after other network
notifiers (they have the default 0 priority)
Reported-by: Ben Greear <greearb@candelatech.com>
Reported-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Reported-by: Octavian Purdila <opurdila@ixiacom.com>
Reported-by: Benjamin LaHaise <bcrl@kvack.org>
Tested-by: Ben Greear <greearb@candelatech.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fix a bug reported by backyes.
Right the first time pktgen's using queue_map that's not been initialized
by set_cur_queue_map(pkt_dev);
Signed-off-by: Junchang Wang <junchangwang@gmail.com>
Signed-off-by: Backyes <backyes@mail.ustc.edu.cn>
Signed-off-by: David S. Miller <davem@davemloft.net>
This should fix the following warning:
net/core/pktgen.c: In function ‘pktgen_if_write’:
net/core/pktgen.c:890: warning: comparison of distinct pointer types lacks a cast
Signed-off-by: Dmitry Torokhov <dtor@mail.ru>
Reviewed-by: Nelson Elhage <nelhage@ksplice.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In dev_pick_tx recompute the queue index if the value stored in the
socket is greater than or equal to the number of real queues for the
device. The saved index in the sock structure is not guaranteed to
be appropriate for the egress device (this could happen on a route
change or in presence of tunnelling). The result of the queue index
being bad would be to return a bogus queue (crash could prersumably
follow).
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A program that accidentally writes too much data to the pktgen file can overflow
the kernel stack and oops the machine. This is only triggerable by root, so
there's no security issue, but it's still an unfortunate bug.
printk() won't print more than 1024 bytes in a single call, anyways, so let's
just never copy more than that much data. We're on a fairly shallow stack, so
that should be safe even with CONFIG_4KSTACKS.
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This helps protect us from overflow issues down in the
individual protocol sendmsg/recvmsg handlers. Once
we hit INT_MAX we truncate out the rest of the iovec
by setting the iov_len members to zero.
This works because:
1) For SOCK_STREAM and SOCK_SEQPACKET sockets, partial
writes are allowed and the application will just continue
with another write to send the rest of the data.
2) For datagram oriented sockets, where there must be a
one-to-one correspondance between write() calls and
packets on the wire, INT_MAX is going to be far larger
than the packet size limit the protocol is going to
check for and signal with -EMSGSIZE.
Based upon a patch by Linus Torvalds.
Signed-off-by: David S. Miller <davem@davemloft.net>
Adds __rcu annotation to (struct fib_rule)->ctarget
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
NETIF_F_HW_CSUM indicates the ability to update an TCP/IP-style 16-bit
checksum with the checksum of an arbitrary part of the packet data,
whereas the FCoE CRC is something entirely different.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Cc: stable@kernel.org [2.6.32+]
Signed-off-by: David S. Miller <davem@davemloft.net>
dev_can_checksum() incorrectly returns true in these cases:
1. The skb has both out-of-band and in-band VLAN tags and the device
supports checksum offload for the encapsulated protocol but only with
one layer of encapsulation.
2. The skb has a VLAN tag and the device supports generic checksumming
but not in conjunction with VLAN encapsulation.
Rearrange the VLAN tag checks to avoid these.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some panic reports in fib_rules_lookup() show a rule could have a NULL
pointer as a next pointer in the rules_list.
This can actually happen because of a bug in fib_nl_newrule() : It
checks if current rule is the destination of unresolved gotos. (Other
rules have gotos to this about to be inserted rule)
Problem is it does the resolution of the gotos before the rule is
inserted in the rules_list (and has a valid next pointer)
Fix this by moving the rules_list insertion before the changes on gotos.
A lockless reader can not any more follow a ctarget pointer, unless
destination is ready (has a valid next pointer)
Reported-by: Oleg A. Arkhangelsky <sysoleg@yandex.ru>
Reported-by: Joe Buehler <aspam@cox.net>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add __rcu annotation to :
(struct sock)->sk_filter
And use appropriate rcu primitives to reduce sparse warnings if
CONFIG_SPARSE_RCU_POINTER=y
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
add __rcu annotation to (struct net)->gen, and use
rcu_dereference_protected() in net_assign_generic()
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add __rcu annotations to :
(struct netdev_rx_queue)->rps_map
(struct netdev_rx_queue)->rps_flow_table
struct rps_sock_flow_table *rps_sock_flow_table;
And use appropriate rcu primitives.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
(struct net_device)->ip6_ptr is rcu protected :
add __rcu annotation and proper rcu primitives.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Three is definitely too low, and we know from reports that GRE tunnels
stacked as deeply as 37 levels cause stack overflows, so pick some
reasonable value between those two.
Signed-off-by: David S. Miller <davem@davemloft.net>
The temporary variable "i" is needlessly initialized to zero
in two distinct cases in this file:
1) where it is set to zero and then used as an argument in an addition
before being assigned a non-zero value.
2) where it is only used in a standard/typical loop counter
For (1), simply delete assignment to zero and usages while still
zero; for (2) simply make the loop start at zero as per standard
practice as seen everywhere else in the same file.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next-2.6: (1699 commits)
bnx2/bnx2x: Unsupported Ethtool operations should return -EINVAL.
vlan: Calling vlan_hwaccel_do_receive() is always valid.
tproxy: use the interface primary IP address as a default value for --on-ip
tproxy: added IPv6 support to the socket match
cxgb3: function namespace cleanup
tproxy: added IPv6 support to the TPROXY target
tproxy: added IPv6 socket lookup function to nf_tproxy_core
be2net: Changes to use only priority codes allowed by f/w
tproxy: allow non-local binds of IPv6 sockets if IP_TRANSPARENT is enabled
tproxy: added tproxy sockopt interface in the IPV6 layer
tproxy: added udp6_lib_lookup function
tproxy: added const specifiers to udp lookup functions
tproxy: split off ipv6 defragmentation to a separate module
l2tp: small cleanup
nf_nat: restrict ICMP translation for embedded header
can: mcp251x: fix generation of error frames
can: mcp251x: fix endless loop in interrupt handler if CANINTF_MERRF is set
can-raw: add msg_flags to distinguish local traffic
9p: client code cleanup
rds: make local functions/variables static
...
Fix up conflicts in net/core/dev.c, drivers/net/pcmcia/smc91c92_cs.c and
drivers/net/wireless/ath/ath9k/debug.c as per David
The function napi_reuse_skb is only used inside core.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When there are VLANs on a VETH device, the packets being transmitted
through the VETH device may be 4 bytes bigger than MTU. A check
in dev_forward_skb did not take this into account and so dropped
these packets.
This patch is needed at least as far back as 2.6.34.7 and should
be considered for -stable.
Signed-off-by: Ben Greear <greearb@candelatech.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The function rtnl_kill_links is defined but never used.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that vlan acceleration is handled consistently regardless of usage,
it is possible to enable and disable it at will. This adds support for
Ethtool operations that change the offloading status for debugging
purposes, similar to other forms of hardware acceleration.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently each driver that is capable of vlan hardware acceleration
must be aware of the vlan groups that are configured and then pass
the stripped tag to a specialized receive function. This is
different from other types of hardware offload in that it places a
significant amount of knowledge in the driver itself rather keeping
it in the networking core.
This makes vlan offloading function more similarly to other forms
of offloading (such as checksum offloading or TSO) by doing the
following:
* On receive, stripped vlans are passed directly to the network
core, without attempting to check for vlan groups or reconstructing
the header if no group
* vlans are made less special by folding the logic into the main
receive routines
* On transmit, the device layer will add the vlan header in software
if the hardware doesn't support it, instead of spreading that logic
out in upper layers, such as bonding.
There are a number of advantages to this:
* Fixes all bugs with drivers incorrectly dropping vlan headers at once.
* Avoids having to disable VLAN acceleration when in promiscuous mode
(good for bridging since it always puts devices in promiscuous mode).
* Keeps VLAN tag separate until given to ultimate consumer, which
avoids needing to do header reconstruction as in tg3 unless absolutely
necessary.
* Consolidates common code in core networking.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently users of hardware vlan accleration need to know whether
the device supports it before generating packets. However, vlan
acceleration will soon be available in a more flexible manner so
knowing ahead of time becomes much more difficult. This adds
a software fallback path for vlan packets on devices without the
necessary offloading support, similar to other types of hardware
accleration.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is no point using RCU for dst we allocate for a very short time
(used once).
Change dst_release() to take DST_NOCACHE into account, but also change
skb_dst_set_noref() to force a refcount increment for such dst.
This is a _huge_ gain, because we dont waste memory to store xx thousand
of dsts. Instead of queueing them to RCU, we can free them instantly.
CPU caches can stay hot, re-using same memory blocks to hold temporary
dsts.
Note : remove unneeded smp_mb__before_atomic_dec(); in dst_release(),
since atomic_dec_return() implies a full memory barrier.
Stress test, 160.000.000 udp frames sent, IP route cache disabled
(DDOS).
Before:
real 0m38.091s
user 0m13.189s
sys 7m53.018s
After:
real 0m29.946s
user 0m12.157s
sys 7m40.605s
For reference, if IP route cache was enabled :
real 0m32.030s
user 0m10.521s
sys 8m15.243s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces netif_alloc_netdev_queues which is called from
register_device instead of alloc_netdev_mq. This makes TX queue
allocation symmetric with RX allocation. Also, queue locks allocation
is done in netdev_init_one_queue. Change set_real_num_tx_queues to
fail if requested number < 1 or greater than number of allocated
queues.
Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Clean up in RX queue allocation. In netif_set_real_num_rx_queues
return error on attempt to set zero queues, or requested number is
greater than number of allocated queues. In netif_alloc_rx_queues,
do BUG_ON if queue_count is zero.
Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In alloc_netdev_mq fail if requested queue_count < 1.
Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In an erlier patch I modified napi_poll so that devices with IFF_MASTER polled
the per_cpu list instead of the device list for napi. I did this because the
bonding driver has no napi instances to poll, it instead expects to check the
slave devices napi instances, which napi_poll was unaware of. Looking at this
more closely however, I now see this isn't strictly needed. As the bond driver
poll_controller calls the slaves poll_controller via netpoll_poll_dev, which
recursively calls poll_napi on each slave, allowing those napi instances to get
serviced. The earlier patch isn't at all harmfull, its just not needed, so lets
revert it to make the code cleaner. Sorry for the noise,
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Reviewed-by: WANG Cong <amwang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Usually the netpoll path, when preforming a napi poll can get away with just
polling all the napi instances of the configured device. Thats not the case for
the bonding driver however, as the napi instances which may wind up getting
flagged as needing polling after the poll_controller call don't belong to the
bonded device, but rather to the slave devices. Fix this by checking the device
in question for the IFF_MASTER flag, if set, we know we need to check the full
poll list for this cpu, rather than just the devices napi instance list.
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The bonding driver currently modifies the netpoll structure in its xmit path
while sending frames from netpoll. This is racy, as other cpus can access the
netpoll structure in parallel. Since the bonding driver points np->dev to a
slave device, other cpus can inadvertently attempt to send data directly to
slave devices, leading to improper locking with the bonding master, lost frames,
and deadlocks. This patch fixes that up.
This patch also removes the real_dev pointer from the netpoll structure as that
data is really only used by bonding in the poll_controller, and we can emulate
its behavior by check each slave for IS_UP.
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
fib_nl_delrule() calls synchronize_rcu() for no apparent reason,
while rtnl is held.
I suspect it was done to avoid an atomic_inc_not_zero() in
fib_rules_lookup(), which commit 7fa7cb7109 added anyway.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit b30973f877 (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: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We tried very hard to remove all possible dev_hold()/dev_put() pairs in
network stack, using RCU conversions.
There is still an unavoidable device refcount change for every dst we
create/destroy, and this can slow down some workloads (routers or some
app servers, mmap af_packet)
We can switch to a percpu refcount implementation, now dynamic per_cpu
infrastructure is mature. On a 64 cpus machine, this consumes 256 bytes
per device.
On x86, dev_hold(dev) code :
before
lock incl 0x280(%ebx)
after:
movl 0x260(%ebx),%eax
incl fs:(%eax)
Stress bench :
(Sending 160.000.000 UDP frames,
IP route cache disabled, dual E5540 @2.53GHz,
32bit kernel, FIB_TRIE)
Before:
real 1m1.662s
user 0m14.373s
sys 12m55.960s
After:
real 0m51.179s
user 0m15.329s
sys 10m15.942s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct dst_ops tracks number of allocated dst in an atomic_t field,
subject to high cache line contention in stress workload.
Switch to a percpu_counter, to reduce number of time we need to dirty a
central location. Place it on a separate cache line to avoid dirtying
read only fields.
Stress test :
(Sending 160.000.000 UDP frames,
IP route cache disabled, dual E5540 @2.53GHz,
32bit kernel, FIB_TRIE, SLUB/NUMA)
Before:
real 0m51.179s
user 0m15.329s
sys 10m15.942s
After:
real 0m45.570s
user 0m15.525s
sys 9m56.669s
With a small reordering of struct neighbour fields, subject of a
following patch, (to separate refcnt from other read mostly fields)
real 0m41.841s
user 0m15.261s
sys 8m45.949s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a seqlock in struct neighbour to protect neigh->ha[], and avoid
dirtying neighbour in stress situation (many different flows / dsts)
Dirtying takes place because of read_lock(&n->lock) and n->used writes.
Switching to a seqlock, and writing n->used only on jiffies changes
permits less dirtying.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
