Input packet processing for local sockets involves two major demuxes.
One for the route and one for the socket.
But we can optimize this down to one demux for certain kinds of local
sockets.
Currently we only do this for established TCP sockets, but it could
at least in theory be expanded to other kinds of connections.
If a TCP socket is established then it's identity is fully specified.
This means that whatever input route was used during the three-way
handshake must work equally well for the rest of the connection since
the keys will not change.
Once we move to established state, we cache the receive packet's input
route to use later.
Like the existing cached route in sk->sk_dst_cache used for output
packets, we have to check for route invalidations using dst->obsolete
and dst->ops->check().
Early demux occurs outside of a socket locked section, so when a route
invalidation occurs we defer the fixup of sk->sk_rx_dst until we are
actually inside of established state packet processing and thus have
the socket locked.
Signed-off-by: David S. Miller <davem@davemloft.net>
Don't pretend that inet_protos[] and inet6_protos[] are hashes, thay
are just a straight arrays. Remove all unnecessary hash masking.
Document MAX_INET_PROTOS.
Use RAW_HTABLE_SIZE when appropriate.
Reported-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adding casts of objects to the same type is unnecessary
and confusing for a human reader.
For example, this cast:
int y;
int *p = (int *)&y;
I used the coccinelle script below to find and remove these
unnecessary casts. I manually removed the conversions this
script produces of casts with __force and __user.
@@
type T;
T *p;
@@
- (T *)p
+ p
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Name them in a "backward compatible" manner, i.e. reuse or not
are still 1 and 0 respectively. The reuse value of 2 means that
the socket with it will forcibly reuse everyone else's port.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix checkpatch errors of the following type:
* ERROR: "foo * bar" should be "foo *bar"
* ERROR: "(foo*)" should be "(foo *)"
Signed-off-by: Daniel Baluta <dbaluta@ixiacom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove all #inclusions of asm/system.h preparatory to splitting and killing
it. Performed with the following command:
perl -p -i -e 's!^#\s*include\s*<asm/system[.]h>.*\n!!' `grep -Irl '^#\s*include\s*<asm/system[.]h>' *`
Signed-off-by: David Howells <dhowells@redhat.com>
Add #define pr_fmt(fmt) as appropriate.
Add "IPv4: ", "TCP: ", and "IPsec: " to appropriate files.
Standardize on "UDPLite: " for appropriate uses.
Some prefixes were previously "UDPLITE: " and "UDP-Lite: ".
Add KBUILD_MODNAME ": " to icmp and gre.
Remove embedded prefixes as appropriate.
Add missing "\n" to pr_info in gre.c.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use a more current kernel messaging style.
Convert a printk block to print_hex_dump.
Coalesce formats, align arguments.
Use %s, __func__ instead of embedding function names.
Some messages that were prefixed with <foo>_close are
now prefixed with <foo>_fini. Some ah4 and esp messages
are now not prefixed with "ip ".
The intent of this patch is to later add something like
#define pr_fmt(fmt) "IPv4: " fmt.
to standardize the output messages.
Text size is trivially reduced. (x86-32 allyesconfig)
$ size net/ipv4/built-in.o*
text data bss dec hex filename
887888 31558 249696 1169142 11d6f6 net/ipv4/built-in.o.new
887934 31558 249800 1169292 11d78c net/ipv4/built-in.o.old
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, it is not easily possible to get TOS/DSCP value of packets from
an incoming TCP stream. The mechanism is there, IP_PKTOPTIONS getsockopt
with IP_RECVTOS set, the same way as incoming TTL can be queried. This is
not actually implemented for TOS, though.
This patch adds this functionality, both for IPv4 (IP_PKTOPTIONS) and IPv6
(IPV6_2292PKTOPTIONS). For IPv4, like in the IP_RECVTTL case, the value of
the TOS field is stored from the other party's ACK.
This is needed for proxies which require DSCP transparency. One such example
is at http://zph.bratcheda.org/.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows each namespace to independently set up
its levels for tcp memory pressure thresholds. This patch
alone does not buy much: we need to make this values
per group of process somehow. This is achieved in the
patches that follows in this patchset.
Signed-off-by: Glauber Costa <glommer@parallels.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
CC: David S. Miller <davem@davemloft.net>
CC: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
v2: add couple missing conversions in drivers
split unexporting netdev_fix_features()
implemented %pNF
convert sock::sk_route_(no?)caps
Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reading /proc/net/snmp on a machine with a lot of cpus is very expensive
(can be ~88000 us).
This is because ICMPMSG MIB uses 4096 bytes per cpu, and folding values
for all possible cpus can read 16 Mbytes of memory.
ICMP messages are not considered as fast path on a typical server, and
eventually few cpus handle them anyway. We can afford an atomic
operation instead of using percpu data.
This saves 4096 bytes per cpu and per network namespace.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ipv4: compat_ioctl is local to af_inet.c, make it static
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit d0733d2e29 (Check for mistakenly passed in non-IPv4 address)
added regression on legacy apps that use bind() with AF_UNSPEC family.
Relax the check, but make sure the bind() is done on INADDR_ANY
addresses, as AF_UNSPEC has probably no sane meaning for other
addresses.
Bugzilla reference : https://bugzilla.kernel.org/show_bug.cgi?id=42012
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Reported-and-bisected-by: Rene Meier <r_meier@freenet.de>
CC: Marcus Meissner <meissner@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hi,
Reinhard Max also pointed out that the error should EAFNOSUPPORT according
to POSIX.
The Linux manpages have it as EINVAL, some other OSes (Minix, HPUX, perhaps BSD) use
EAFNOSUPPORT. Windows uses WSAEFAULT according to MSDN.
Other protocols error values in their af bind() methods in current mainline git as far
as a brief look shows:
EAFNOSUPPORT: atm, appletalk, l2tp, llc, phonet, rxrpc
EINVAL: ax25, bluetooth, decnet, econet, ieee802154, iucv, netlink, netrom, packet, rds, rose, unix, x25,
No check?: can/raw, ipv6/raw, irda, l2tp/l2tp_ip
Ciao, Marcus
Signed-off-by: Marcus Meissner <meissner@suse.de>
Cc: Reinhard Max <max@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Le jeudi 16 juin 2011 à 23:38 -0400, David Miller a écrit :
> From: Ben Hutchings <bhutchings@solarflare.com>
> Date: Fri, 17 Jun 2011 00:50:46 +0100
>
> > On Wed, 2011-06-15 at 04:15 +0200, Eric Dumazet wrote:
> >> @@ -1594,6 +1594,7 @@ int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
> >> goto discard;
> >>
> >> if (nsk != sk) {
> >> + sock_rps_save_rxhash(nsk, skb->rxhash);
> >> if (tcp_child_process(sk, nsk, skb)) {
> >> rsk = nsk;
> >> goto reset;
> >>
> >
> > I haven't tried this, but it looks reasonable to me.
> >
> > What about IPv6? The logic in tcp_v6_do_rcv() looks very similar.
>
> Indeed ipv6 side needs the same fix.
>
> Eric please add that part and resubmit. And in fact I might stick
> this into net-2.6 instead of net-next-2.6
>
OK, here is the net-2.6 based one then, thanks !
[PATCH v2] net: rfs: enable RFS before first data packet is received
First packet received on a passive tcp flow is not correctly RFS
steered.
One sock_rps_record_flow() call is missing in inet_accept()
But before that, we also must record rxhash when child socket is setup.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Tom Herbert <therbert@google.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Jamal Hadi Salim <hadi@cyberus.ca>
Signed-off-by: David S. Miller <davem@conan.davemloft.net>
SNMP mibs use two percpu arrays, one used in BH context, another in USER
context. With increasing number of cpus in machines, and fact that ipv6
uses per network device ipstats_mib, this is consuming a lot of memory
if many network devices are registered.
commit be281e554e (ipv6: reduce per device ICMP mib sizes) shrinked
percpu needs for ipv6, but we can reduce memory use a bit more.
With recent percpu infrastructure (irqsafe_cpu_inc() ...), we no longer
need this BH/USER separation since we can update counters in a single
x86 instruction, regardless of the BH/USER context.
Other arches than x86 might need to disable irq in their
irqsafe_cpu_inc() implementation : If this happens to be a problem, we
can make SNMP_ARRAY_SZ arch dependent, but a previous poll
( https://lkml.org/lkml/2011/3/17/174 ) to arch maintainers did not
raise strong opposition.
Only on 32bit arches, we need to disable BH for 64bit counters updates
done from USER context (currently used for IP MIB)
This also reduces vmlinux size :
1) x86_64 build
$ size vmlinux.before vmlinux.after
text data bss dec hex filename
7853650 1293772 1896448 11043870 a8841e vmlinux.before
7850578 1293772 1896448 11040798 a8781e vmlinux.after
2) i386 build
$ size vmlinux.before vmlinux.afterpatch
text data bss dec hex filename
6039335 635076 3670016 10344427 9dd7eb vmlinux.before
6037342 635076 3670016 10342434 9dd022 vmlinux.afterpatch
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Andi Kleen <andi@firstfloor.org>
CC: Ingo Molnar <mingo@elte.hu>
CC: Tejun Heo <tj@kernel.org>
CC: Christoph Lameter <cl@linux-foundation.org>
CC: Benjamin Herrenschmidt <benh@kernel.crashing.org
CC: linux-arch@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
Check against mistakenly passing in IPv6 addresses (which would result
in an INADDR_ANY bind) or similar incompatible sockaddrs.
Signed-off-by: Marcus Meissner <meissner@suse.de>
Cc: Reinhard Max <max@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds IPPROTO_ICMP socket kind. It makes it possible to send
ICMP_ECHO messages and receive the corresponding ICMP_ECHOREPLY messages
without any special privileges. In other words, the patch makes it
possible to implement setuid-less and CAP_NET_RAW-less /bin/ping. In
order not to increase the kernel's attack surface, the new functionality
is disabled by default, but is enabled at bootup by supporting Linux
distributions, optionally with restriction to a group or a group range
(see below).
Similar functionality is implemented in Mac OS X:
http://www.manpagez.com/man/4/icmp/
A new ping socket is created with
socket(PF_INET, SOCK_DGRAM, PROT_ICMP)
Message identifiers (octets 4-5 of ICMP header) are interpreted as local
ports. Addresses are stored in struct sockaddr_in. No port numbers are
reserved for privileged processes, port 0 is reserved for API ("let the
kernel pick a free number"). There is no notion of remote ports, remote
port numbers provided by the user (e.g. in connect()) are ignored.
Data sent and received include ICMP headers. This is deliberate to:
1) Avoid the need to transport headers values like sequence numbers by
other means.
2) Make it easier to port existing programs using raw sockets.
ICMP headers given to send() are checked and sanitized. The type must be
ICMP_ECHO and the code must be zero (future extensions might relax this,
see below). The id is set to the number (local port) of the socket, the
checksum is always recomputed.
ICMP reply packets received from the network are demultiplexed according
to their id's, and are returned by recv() without any modifications.
IP header information and ICMP errors of those packets may be obtained
via ancillary data (IP_RECVTTL, IP_RETOPTS, and IP_RECVERR). ICMP source
quenches and redirects are reported as fake errors via the error queue
(IP_RECVERR); the next hop address for redirects is saved to ee_info (in
network order).
socket(2) is restricted to the group range specified in
"/proc/sys/net/ipv4/ping_group_range". It is "1 0" by default, meaning
that nobody (not even root) may create ping sockets. Setting it to "100
100" would grant permissions to the single group (to either make
/sbin/ping g+s and owned by this group or to grant permissions to the
"netadmins" group), "0 4294967295" would enable it for the world, "100
4294967295" would enable it for the users, but not daemons.
The existing code might be (in the unlikely case anyone needs it)
extended rather easily to handle other similar pairs of ICMP messages
(Timestamp/Reply, Information Request/Reply, Address Mask Request/Reply
etc.).
Userspace ping util & patch for it:
http://openwall.info/wiki/people/segoon/ping
For Openwall GNU/*/Linux it was the last step on the road to the
setuid-less distro. A revision of this patch (for RHEL5/OpenVZ kernels)
is in use in Owl-current, such as in the 2011/03/12 LiveCD ISOs:
http://mirrors.kernel.org/openwall/Owl/current/iso/
Initially this functionality was written by Pavel Kankovsky for
Linux 2.4.32, but unfortunately it was never made public.
All ping options (-b, -p, -Q, -R, -s, -t, -T, -M, -I), are tested with
the patch.
PATCH v3:
- switched to flowi4.
- minor changes to be consistent with raw sockets code.
PATCH v2:
- changed ping_debug() to pr_debug().
- removed CONFIG_IP_PING.
- removed ping_seq_fops.owner field (unused for procfs).
- switched to proc_net_fops_create().
- switched to %pK in seq_printf().
PATCH v1:
- fixed checksumming bug.
- CAP_NET_RAW may not create icmp sockets anymore.
RFC v2:
- minor cleanups.
- introduced sysctl'able group range to restrict socket(2).
Signed-off-by: Vasiliy Kulikov <segoon@openwall.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These two functions must be invoked only when the socket is locked
(because socket identity modifications are made non-atomically).
Therefore we can use the cork flow for output route lookups.
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that output route lookups update the flow with
source address selection, we can fetch it from
fl4->saddr instead of rt->rt_src
Signed-off-by: David S. Miller <davem@davemloft.net>
We lack proper synchronization to manipulate inet->opt ip_options
Problem is ip_make_skb() calls ip_setup_cork() and
ip_setup_cork() possibly makes a copy of ipc->opt (struct ip_options),
without any protection against another thread manipulating inet->opt.
Another thread can change inet->opt pointer and free old one under us.
Use RCU to protect inet->opt (changed to inet->inet_opt).
Instead of handling atomic refcounts, just copy ip_options when
necessary, to avoid cache line dirtying.
We cant insert an rcu_head in struct ip_options since its included in
skb->cb[], so this patch is large because I had to introduce a new
ip_options_rcu structure.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
These functions are used together as a unit for route resolution
during connect(). They address the chicken-and-egg problem that
exists when ports need to be allocated during connect() processing,
yet such port allocations require addressing information from the
routing code.
It's currently more heavy handed than it needs to be, and in
particular we allocate and initialize a flow object twice.
Let the callers provide the on-stack flow object. That way we only
need to initialize it once in the ip_route_connect() call.
Later, if ip_route_newports() needs to do anything, it re-uses that
flow object as-is except for the ports which it updates before the
route re-lookup.
Also, describe why this set of facilities are needed and how it works
in a big comment.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <eric.dumazet@gmail.com>
Add const qualifiers to structs iphdr, ipv6hdr and in6_addr pointers
where possible, to make code intention more obvious.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The idea here is this minimizes the number of places one has to edit
in order to make changes to how flows are defined and used.
Signed-off-by: David S. Miller <davem@davemloft.net>
SIOCGETSGCNT is not a unique ioctl value as it it maps tio SIOCPROTOPRIVATE +1,
which unfortunately means the existing infrastructure for compat networking
ioctls is insufficient. A trivial compact ioctl implementation would conflict
with:
SIOCAX25ADDUID
SIOCAIPXPRISLT
SIOCGETSGCNT_IN6
SIOCGETSGCNT
SIOCRSSCAUSE
SIOCX25SSUBSCRIP
SIOCX25SDTEFACILITIES
To make this work I have updated the compat_ioctl decode path to mirror the
the normal ioctl decode path. I have added an ipv4 inet_compat_ioctl function
so that I can have ipv4 specific compat ioctls. I have added a compat_ioctl
function into struct proto so I can break out ioctls by which kind of ip socket
I am using. I have added a compat_raw_ioctl function because SIOCGETSGCNT only
works on raw sockets. I have added a ipmr_compat_ioctl that mirrors the normal
ipmr_ioctl.
This was necessary because unfortunately the struct layout for the SIOCGETSGCNT
has unsigned longs in it so changes between 32bit and 64bit kernels.
This change was sufficient to run a 32bit ip multicast routing daemon on a
64bit kernel.
Reported-by: Bill Fenner <fenner@aristanetworks.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Quoting Ben Hutchings: we presumably won't be defining features that
can only be enabled on 64-bit architectures.
Occurences found by `grep -r` on net/, drivers/net, include/
[ Move features and vlan_features next to each other in
struct netdev, as per Eric Dumazet's suggestion -DaveM ]
Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use the macros defined for the members of flowi to clean the code up.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now cmpxchg() is available on all arches, we can use it in
build_ehash_secret() and rt_bind_peer() instead of using spinlocks.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Signed-off-by: David S. Miller <davem@davemloft.net>
a new boolean flag no_autobind is added to structure proto to avoid the autobind
calls when the protocol is TCP. Then sock_rps_record_flow() is called int the
TCP's sendmsg() and sendpage() pathes.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
----
include/net/inet_common.h | 4 ++++
include/net/sock.h | 1 +
include/net/tcp.h | 8 ++++----
net/ipv4/af_inet.c | 15 +++++++++------
net/ipv4/tcp.c | 11 +++++------
net/ipv4/tcp_ipv4.c | 3 +++
net/ipv6/af_inet6.c | 8 ++++----
net/ipv6/tcp_ipv6.c | 3 +++
8 files changed, 33 insertions(+), 20 deletions(-)
Signed-off-by: David S. Miller <davem@davemloft.net>
/proc/net/snmp and /proc/net/netstat expose SNMP counters.
Width of these counters is either 32 or 64 bits, depending on the size
of "unsigned long" in kernel.
This means user program parsing these files must already be prepared to
deal with 64bit values, regardless of user program being 32 or 64 bit.
This patch introduces 64bit snmp values for IPSTAT mib, where some
counters can wrap pretty fast if they are 32bit wide.
# netstat -s|egrep "InOctets|OutOctets"
InOctets: 244068329096
OutOctets: 244069348848
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In preparation for 64bit snmp counters for some mibs,
add an 'align' parameter to snmp_mib_init(), instead
of assuming mibs only contain 'unsigned long' fields.
Callers can use __alignof__(type) to provide correct
alignment.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
CC: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
CC: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
this patch is implementing IP_NODEFRAG option for IPv4 socket.
The reason is, there's no other way to send out the packet with user
customized header of the reassembly part.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
remove useless union keyword in rtable, rt6_info and dn_route.
Since there is only one member in a union, the union keyword isn't useful.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
(Dropped the infiniband part, because Tetsuo modified the related code,
I will send a separate patch for it once this is accepted.)
This patch introduces /proc/sys/net/ipv4/ip_local_reserved_ports which
allows users to reserve ports for third-party applications.
The reserved ports will not be used by automatic port assignments
(e.g. when calling connect() or bind() with port number 0). Explicit
port allocation behavior is unchanged.
Signed-off-by: Octavian Purdila <opurdila@ixiacom.com>
Signed-off-by: WANG Cong <amwang@redhat.com>
Cc: Neil Horman <nhorman@tuxdriver.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Idea from Eric Dumazet.
As for placement inside of struct sock, I tried to choose a place
that otherwise has a 32-bit hole on 64-bit systems.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Sparse can help us find endianness bugs, but we need to make some
cleanups to be able to more easily spot real bugs.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Define a new function to return the waitqueue of a "struct sock".
static inline wait_queue_head_t *sk_sleep(struct sock *sk)
{
return sk->sk_sleep;
}
Change all read occurrences of sk_sleep by a call to this function.
Needed for a future RCU conversion. sk_sleep wont be a field directly
available.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements receive flow steering (RFS). RFS steers
received packets for layer 3 and 4 processing to the CPU where
the application for the corresponding flow is running. RFS is an
extension of Receive Packet Steering (RPS).
The basic idea of RFS is that when an application calls recvmsg
(or sendmsg) the application's running CPU is stored in a hash
table that is indexed by the connection's rxhash which is stored in
the socket structure. The rxhash is passed in skb's received on
the connection from netif_receive_skb. For each received packet,
the associated rxhash is used to look up the CPU in the hash table,
if a valid CPU is set then the packet is steered to that CPU using
the RPS mechanisms.
The convolution of the simple approach is that it would potentially
allow OOO packets. If threads are thrashing around CPUs or multiple
threads are trying to read from the same sockets, a quickly changing
CPU value in the hash table could cause rampant OOO packets--
we consider this a non-starter.
To avoid OOO packets, this solution implements two types of hash
tables: rps_sock_flow_table and rps_dev_flow_table.
rps_sock_table is a global hash table. Each entry is just a CPU
number and it is populated in recvmsg and sendmsg as described above.
This table contains the "desired" CPUs for flows.
rps_dev_flow_table is specific to each device queue. Each entry
contains a CPU and a tail queue counter. The CPU is the "current"
CPU for a matching flow. The tail queue counter holds the value
of a tail queue counter for the associated CPU's backlog queue at
the time of last enqueue for a flow matching the entry.
Each backlog queue has a queue head counter which is incremented
on dequeue, and so a queue tail counter is computed as queue head
count + queue length. When a packet is enqueued on a backlog queue,
the current value of the queue tail counter is saved in the hash
entry of the rps_dev_flow_table.
And now the trick: when selecting the CPU for RPS (get_rps_cpu)
the rps_sock_flow table and the rps_dev_flow table for the RX queue
are consulted. When the desired CPU for the flow (found in the
rps_sock_flow table) does not match the current CPU (found in the
rps_dev_flow table), the current CPU is changed to the desired CPU
if one of the following is true:
- The current CPU is unset (equal to RPS_NO_CPU)
- Current CPU is offline
- The current CPU's queue head counter >= queue tail counter in the
rps_dev_flow table. This checks if the queue tail has advanced
beyond the last packet that was enqueued using this table entry.
This guarantees that all packets queued using this entry have been
dequeued, thus preserving in order delivery.
Making each queue have its own rps_dev_flow table has two advantages:
1) the tail queue counters will be written on each receive, so
keeping the table local to interrupting CPU s good for locality. 2)
this allows lockless access to the table-- the CPU number and queue
tail counter need to be accessed together under mutual exclusion
from netif_receive_skb, we assume that this is only called from
device napi_poll which is non-reentrant.
This patch implements RFS for TCP and connected UDP sockets.
It should be usable for other flow oriented protocols.
There are two configuration parameters for RFS. The
"rps_flow_entries" kernel init parameter sets the number of
entries in the rps_sock_flow_table, the per rxqueue sysfs entry
"rps_flow_cnt" contains the number of entries in the rps_dev_flow
table for the rxqueue. Both are rounded to power of two.
The obvious benefit of RFS (over just RPS) is that it achieves
CPU locality between the receive processing for a flow and the
applications processing; this can result in increased performance
(higher pps, lower latency).
The benefits of RFS are dependent on cache hierarchy, application
load, and other factors. On simple benchmarks, we don't necessarily
see improvement and sometimes see degradation. However, for more
complex benchmarks and for applications where cache pressure is
much higher this technique seems to perform very well.
Below are some benchmark results which show the potential benfit of
this patch. The netperf test has 500 instances of netperf TCP_RR
test with 1 byte req. and resp. The RPC test is an request/response
test similar in structure to netperf RR test ith 100 threads on
each host, but does more work in userspace that netperf.
e1000e on 8 core Intel
No RFS or RPS 104K tps at 30% CPU
No RFS (best RPS config): 290K tps at 63% CPU
RFS 303K tps at 61% CPU
RPC test tps CPU% 50/90/99% usec latency Latency StdDev
No RFS/RPS 103K 48% 757/900/3185 4472.35
RPS only: 174K 73% 415/993/2468 491.66
RFS 223K 73% 379/651/1382 315.61
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With latest CONFIG_PROVE_RCU stuff, I felt more comfortable to make this
work.
sk->sk_dst_cache is currently protected by a rwlock (sk_dst_lock)
This rwlock is readlocked for a very small amount of time, and dst
entries are already freed after RCU grace period. This calls for RCU
again :)
This patch converts sk_dst_lock to a spinlock, and use RCU for readers.
__sk_dst_get() is supposed to be called with rcu_read_lock() or if
socket locked by user, so use appropriate rcu_dereference_check()
condition (rcu_read_lock_held() || sock_owned_by_user(sk))
This patch avoids two atomic ops per tx packet on UDP connected sockets,
for example, and permits sk_dst_lock to be much less dirtied.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>