For a connected socket we can precompute the flow hash for setting
in skb->hash on output. This is a performance advantage over
calculating the skb->hash for every packet on the connection. The
computation is done using the common hash algorithm to be consistent
with computations done for packets of the connection in other states
where thers is no socket (e.g. time-wait, syn-recv, syn-cookies).
This patch adds sk_txhash to the sock structure. inet_set_txhash and
ip6_set_txhash functions are added which are called from points in
TCP and UDP where socket moves to established state.
skb_set_hash_from_sk is a function which sets skb->hash from the
sock txhash value. This is called in UDP and TCP transmit path when
transmitting within the context of a socket.
Tested: ran super_netperf with 200 TCP_RR streams over a vxlan
interface (in this case skb_get_hash called on every TX packet to
create a UDP source port).
Before fix:
95.02% CPU utilization
154/256/505 90/95/99% latencies
1.13042e+06 tps
Time in functions:
0.28% skb_flow_dissect
0.21% __skb_get_hash
After fix:
94.95% CPU utilization
156/254/485 90/95/99% latencies
1.15447e+06
Neither __skb_get_hash nor skb_flow_dissect appear in perf
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sk_dst_cache has __rcu annotation, so we need a cast to avoid
following sparse error :
include/net/sock.h:1774:19: warning: incorrect type in initializer (different address spaces)
include/net/sock.h:1774:19: expected struct dst_entry [noderef] <asn:4>*__ret
include/net/sock.h:1774:19: got struct dst_entry *dst
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Fixes: 7f50236153 ("ipv4: irq safe sk_dst_[re]set() and ipv4_sk_update_pmtu() fix")
Signed-off-by: David S. Miller <davem@davemloft.net>
When an UDP application switches from AF_INET to AF_INET6 sockets, we
have a small performance degradation for IPv4 communications because of
extra cache line misses to access ipv6only information.
This can also be noticed for TCP listeners, as ipv6_only_sock() is also
used from __inet_lookup_listener()->compute_score()
This is magnified when SO_REUSEPORT is used.
Move ipv6only into struct sock_common so that it is available at
no extra cost in lookups.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We have two different ways to handle changes to sk->sk_dst
First way (used by TCP) assumes socket lock is owned by caller, and use
no extra lock : __sk_dst_set() & __sk_dst_reset()
Another way (used by UDP) uses sk_dst_lock because socket lock is not
always taken. Note that sk_dst_lock is not softirq safe.
These ways are not inter changeable for a given socket type.
ipv4_sk_update_pmtu(), added in linux-3.8, added a race, as it used
the socket lock as synchronization, but users might be UDP sockets.
Instead of converting sk_dst_lock to a softirq safe version, use xchg()
as we did for sk_rx_dst in commit e47eb5dfb2 ("udp: ipv4: do not use
sk_dst_lock from softirq context")
In a follow up patch, we probably can remove sk_dst_lock, as it is
only used in IPv6.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Fixes: 9cb3a50c5f ("ipv4: Invalidate the socket cached route on pmtu events if possible")
Signed-off-by: David S. Miller <davem@davemloft.net>
When IP route cache had been removed in linux-3.6, we broke assumption
that dst entries were all freed after rcu grace period. DST_NOCACHE
dst were supposed to be freed from dst_release(). But it appears
we want to keep such dst around, either in UDP sockets or tunnels.
In sk_dst_get() we need to make sure dst refcount is not 0
before incrementing it, or else we might end up freeing a dst
twice.
DST_NOCACHE set on a dst does not mean this dst can not be attached
to a socket or a tunnel.
Then, before actual freeing, we need to observe a rcu grace period
to make sure all other cpus can catch the fact the dst is no longer
usable.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dormando <dormando@rydia.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Define separate fields in the sock structure for configuring disabling
checksums in both TX and RX-- sk_no_check_tx and sk_no_check_rx.
The SO_NO_CHECK socket option only affects sk_no_check_tx. Also,
removed UDP_CSUM_* defines since they are no longer necessary.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sk_net_capable - The common case, operations that are safe in a network namespace.
sk_capable - Operations that are not known to be safe in a network namespace
sk_ns_capable - The general case for special cases.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several spots in the kernel perform a sequence like:
skb_queue_tail(&sk->s_receive_queue, skb);
sk->sk_data_ready(sk, skb->len);
But at the moment we place the SKB onto the socket receive queue it
can be consumed and freed up. So this skb->len access is potentially
to freed up memory.
Furthermore, the skb->len can be modified by the consumer so it is
possible that the value isn't accurate.
And finally, no actual implementation of this callback actually uses
the length argument. And since nobody actually cared about it's
value, lots of call sites pass arbitrary values in such as '0' and
even '1'.
So just remove the length argument from the callback, that way there
is no confusion whatsoever and all of these use-after-free cases get
fixed as a side effect.
Based upon a patch by Eric Dumazet and his suggestion to audit this
issue tree-wide.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch basically does two things, i) removes the extern keyword
from the include/linux/filter.h file to be more consistent with the
rest of Joe's changes, and ii) moves filter accounting into the filter
core framework.
Filter accounting mainly done through sk_filter_{un,}charge() take
care of the case when sockets are being cloned through sk_clone_lock()
so that removal of the filter on one socket won't result in eviction
as it's still referenced by the other.
These functions actually belong to net/core/filter.c and not
include/net/sock.h as we want to keep all that in a central place.
It's also not in fast-path so uninlining them is fine and even allows
us to get rd of sk_filter_release_rcu()'s EXPORT_SYMBOL and a forward
declaration.
Joint work with Alexei Starovoitov.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Cc: Pavel Emelyanov <xemul@parallels.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>
Conflicts:
drivers/net/usb/r8152.c
drivers/net/xen-netback/netback.c
Both the r8152 and netback conflicts were simple overlapping
changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
We do not need to switch the net_ns if the target net_ns the same
as the current one, so here we add a pre-check of net_ns to avoid
this as David suggested.
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The original documentation was very unclear.
The code fix is presumably related to the formerly unclear
documentation: SOCK_TIMESTAMPING_RX_SOFTWARE has no effect on
__sock_recv_timestamp's behavior, so calling __sock_recv_ts_and_drops
from sock_recv_ts_and_drops if only SOCK_TIMESTAMPING_RX_SOFTWARE is
set is pointless. This should have no user-observable effect.
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pablo Neira Ayuso says:
====================
netfilter/IPVS updates for net-next
The following patchset contains Netfilter updates for your net-next tree,
they are:
* Add full port randomization support. Some crazy researchers found a way
to reconstruct the secure ephemeral ports that are allocated in random mode
by sending off-path bursts of UDP packets to overrun the socket buffer of
the DNS resolver to trigger retransmissions, then if the timing for the
DNS resolution done by a client is larger than usual, then they conclude
that the port that received the burst of UDP packets is the one that was
opened. It seems a bit aggressive method to me but it seems to work for
them. As a result, Daniel Borkmann and Hannes Frederic Sowa came up with a
new NAT mode to fully randomize ports using prandom.
* Add a new classifier to x_tables based on the socket net_cls set via
cgroups. These includes two patches to prepare the field as requested by
Zefan Li. Also from Daniel Borkmann.
* Use prandom instead of get_random_bytes in several locations of the
netfilter code, from Florian Westphal.
* Allow to use the CTA_MARK_MASK in ctnetlink when mangling the conntrack
mark, also from Florian Westphal.
* Fix compilation warning due to unused variable in IPVS, from Geert
Uytterhoeven.
* Add support for UID/GID via nfnetlink_queue, from Valentina Giusti.
* Add IPComp extension to x_tables, from Fan Du.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Namespace related cleaning
* make cred_to_ucred static
* remove unused sock_rmalloc function
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
While we're at it and introduced CGROUP_NET_CLASSID, lets also make
NETPRIO_CGROUP more consistent with the rest of cgroups and rename it
into CONFIG_CGROUP_NET_PRIO so that for networking, we now have
CONFIG_CGROUP_NET_{PRIO,CLASSID}. This not only makes the CONFIG
option consistent among networking cgroups, but also among cgroups
CONFIG conventions in general as the vast majority has a prefix of
CONFIG_CGROUP_<SUBSYS>.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Zefan Li <lizefan@huawei.com>
Cc: cgroups@vger.kernel.org
Acked-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
In file included from net/socket.c:99:0:
include/net/sock.h: In function ‘sock_rps_record_flow’:
include/net/sock.h:849:30: error: ‘const struct sock’ has no member named ‘sk_rxhash’
include/net/sock.h: In function ‘sock_rps_reset_flow’:
include/net/sock.h:854:29: error: ‘const struct sock’ has no member named ‘sk_rxhash’
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds sock_rps_record_flow_hash and sock_rps_reset_flow_hash
which take a hash value as an argument and sets the sock_flow_table
accordingly. This allows the table to be populated in cases where flow
is being tracked outside of a sock structure.
sock_rps_record_flow and sock_rps_reset_flow call this function
where the hash is taken from sk_rxhash.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
kill memcg_tcp_enter_memory_pressure. The only function of
memcg_tcp_enter_memory_pressure was to reduce deal with the
unnecessary abstraction that was tcp_memcontrol. Now that struct
tcp_memcontrol is gone remove this unnecessary function, the
unnecessary function pointer, and modify sk_enter_memory_pressure to
set this field directly, just as sk_leave_memory_pressure cleas this
field directly.
This fixes a small bug I intruduced when killing struct tcp_memcontrol
that caused memcg_tcp_enter_memory_pressure to never be called and
thus failed to ever set cg_proto->memory_pressure.
Remove the cg_proto enter_memory_pressure function as it now serves
no useful purpose.
Don't test cg_proto->memory_presser in sk_leave_memory_pressure before
clearing it. The test was originally there to ensure that the pointer
was non-NULL. Now that cg_proto is not a pointer the pointer does not
matter.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2e685cad57 (tcp_memcontrol: Kill struct tcp_memcontrol) falsly modified
the access to memory_pressure of sk->sk_prot->memory_pressure. The patch
did modify the memory_pressure-field of struct cg_proto, but not the one
of struct proto.
So, the access to sk_prot->memory_pressure should not be changed.
Acked-by: Eric Dumazet <edumazet@google.com>
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Christoph Paasch <christoph.paasch@uclouvain.be>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
What sk_reset_txq() does is just calls function sk_tx_queue_reset(),
and sk_reset_txq() is used only in sock.h, by dst_negative_advice().
Let dst_negative_advice() calls sk_tx_queue_reset() directly so we
can remove unneeded sk_reset_txq().
Signed-off-by: ZHAO Gang <gamerh2o@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace the pointers in struct cg_proto with actual data fields and kill
struct tcp_memcontrol as it is not fully redundant.
This removes a confusing, unnecessary layer of abstraction.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP listener refactoring, part 4 :
To speed up inet lookups, we moved IPv4 addresses from inet to struct
sock_common
Now is time to do the same for IPv6, because it permits us to have fast
lookups for all kind of sockets, including upcoming SYN_RECV.
Getting IPv6 addresses in TCP lookups currently requires two extra cache
lines, plus a dereference (and memory stall).
inet6_sk(sk) does the dereference of inet_sk(__sk)->pinet6
This patch is way bigger than its IPv4 counter part, because for IPv4,
we could add aliases (inet_daddr, inet_rcv_saddr), while on IPv6,
it's not doable easily.
inet6_sk(sk)->daddr becomes sk->sk_v6_daddr
inet6_sk(sk)->rcv_saddr becomes sk->sk_v6_rcv_saddr
And timewait socket also have tw->tw_v6_daddr & tw->tw_v6_rcv_saddr
at the same offset.
We get rid of INET6_TW_MATCH() as INET6_MATCH() is now the generic
macro.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP listener refactoring, part 3 :
Our goal is to hash SYN_RECV sockets into main ehash for fast lookup,
and parallel SYN processing.
Current inet_ehash_bucket contains two chains, one for ESTABLISH (and
friend states) sockets, another for TIME_WAIT sockets only.
As the hash table is sized to get at most one socket per bucket, it
makes little sense to have separate twchain, as it makes the lookup
slightly more complicated, and doubles hash table memory usage.
If we make sure all socket types have the lookup keys at the same
offsets, we can use a generic and faster lookup. It turns out TIME_WAIT
and ESTABLISHED sockets already have common lookup fields for IPv4.
[ INET_TW_MATCH() is no longer needed ]
I'll provide a follow-up to factorize IPv6 lookup as well, to remove
INET6_TW_MATCH()
This way, SYN_RECV pseudo sockets will be supported the same.
A new sock_gen_put() helper is added, doing either a sock_put() or
inet_twsk_put() [ and will support SYN_RECV later ].
Note this helper should only be called in real slow path, when rcu
lookup found a socket that was moved to another identity (freed/reused
immediately), but could eventually be used in other contexts, like
sock_edemux()
Before patch :
dmesg | grep "TCP established"
TCP established hash table entries: 524288 (order: 11, 8388608 bytes)
After patch :
TCP established hash table entries: 524288 (order: 10, 4194304 bytes)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
include/linux/netdevice.h
net/core/sock.c
Trivial merge issues.
Removal of "extern" for functions declaration in netdevice.h
at the same time "const" was added to an argument.
Two parallel line additions in net/core/sock.c
Signed-off-by: David S. Miller <davem@davemloft.net>
The removal of the routing cache introduced a performance regression for
some UDP workloads since a dst lookup must be done for each packet.
This change caches the dst per socket in a similar manner to what we do
for TCP by implementing early_demux.
For UDP multicast we can only cache the dst if there is only one
receiving socket on the host. Since caching only works when there is
one receiving socket we do the multicast socket lookup using RCU.
For UDP unicast we only demux sockets with an exact match in order to
not break forwarding setups. Additionally since the hash chains may be
long we only check the first socket to see if it is a match and not
waste extra time searching the whole chain when we might not find an
exact match.
Benchmark results from a netperf UDP_RR test:
Before 87961.22 transactions/s
After 89789.68 transactions/s
Benchmark results from a fio 1 byte UDP multicast pingpong test
(Multicast one way unicast response):
Before 12.97us RTT
After 12.63us RTT
Signed-off-by: Shawn Bohrer <sbohrer@rgmadvisors.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP listener refactoring, part 2 :
We can use a generic lookup, sockets being in whatever state, if
we are sure all relevant fields are at the same place in all socket
types (ESTABLISH, TIME_WAIT, SYN_RECV)
This patch removes these macros :
inet_addrpair, inet_addrpair, tw_addrpair, tw_portpair
And adds :
sk_portpair, sk_addrpair, sk_daddr, sk_rcv_saddr
Then, INET_TW_MATCH() is really the same than INET_MATCH()
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/usb/qmi_wwan.c
drivers/net/wireless/brcm80211/brcmfmac/dhd_bus.h
include/net/netfilter/nf_conntrack_synproxy.h
include/net/secure_seq.h
The conflicts are of two varieties:
1) Conflicts with Joe Perches's 'extern' removal from header file
function declarations. Usually it's an argument signature change
or a function being added/removed. The resolutions are trivial.
2) Some overlapping changes in qmi_wwan.c and be.h, one commit adds
a new value, another changes an existing value. That sort of
thing.
Signed-off-by: David S. Miller <davem@davemloft.net>
Warning(include/net/sock.h:411): No description found for parameter
'sk_max_pacing_rate'
Lets please "make htmldocs" and kbuild bot.
Reported-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use of RCU api makes vxlan code easier to understand. It also
fixes bug due to missing ACCESS_ONCE() on sk_user_data dereference.
In rare case without ACCESS_ONCE() compiler might omit vs on
sk_user_data dereference.
Compiler can use vs as alias for sk->sk_user_data, resulting in
multiple sk_user_data dereference in rcu read context which
could change.
CC: Jesse Gross <jesse@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As mentioned in commit afe4fd0624 ("pkt_sched: fq: Fair Queue packet
scheduler"), this patch adds a new socket option.
SO_MAX_PACING_RATE offers the application the ability to cap the
rate computed by transport layer. Value is in bytes per second.
u32 val = 1000000;
setsockopt(sockfd, SOL_SOCKET, SO_MAX_PACING_RATE, &val, sizeof(val));
To be effectively paced, a flow must use FQ packet scheduler.
Note that a packet scheduler takes into account the headers for its
computations. The effective payload rate depends on MSS and retransmits
if any.
I chose to make this pacing rate a SOL_SOCKET option instead of a
TCP one because this can be used by other protocols.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Steinar H. Gunderson <sesse@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are a mix of function prototypes with and without extern
in the kernel sources. Standardize on not using extern for
function prototypes.
Function prototypes don't need to be written with extern.
extern is assumed by the compiler. Its use is as unnecessary as
using auto to declare automatic/local variables in a block.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After hearing many people over past years complaining against TSO being
bursty or even buggy, we are proud to present automatic sizing of TSO
packets.
One part of the problem is that tcp_tso_should_defer() uses an heuristic
relying on upcoming ACKS instead of a timer, but more generally, having
big TSO packets makes little sense for low rates, as it tends to create
micro bursts on the network, and general consensus is to reduce the
buffering amount.
This patch introduces a per socket sk_pacing_rate, that approximates
the current sending rate, and allows us to size the TSO packets so
that we try to send one packet every ms.
This field could be set by other transports.
Patch has no impact for high speed flows, where having large TSO packets
makes sense to reach line rate.
For other flows, this helps better packet scheduling and ACK clocking.
This patch increases performance of TCP flows in lossy environments.
A new sysctl (tcp_min_tso_segs) is added, to specify the
minimal size of a TSO packet (default being 2).
A follow-up patch will provide a new packet scheduler (FQ), using
sk_pacing_rate as an input to perform optional per flow pacing.
This explains why we chose to set sk_pacing_rate to twice the current
rate, allowing 'slow start' ramp up.
sk_pacing_rate = 2 * cwnd * mss / srtt
v2: Neal Cardwell reported a suspect deferring of last two segments on
initial write of 10 MSS, I had to change tcp_tso_should_defer() to take
into account tp->xmit_size_goal_segs
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Van Jacobson <vanj@google.com>
Cc: Tom Herbert <therbert@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adding paged frags skbs to af_unix sockets introduced a performance
regression on large sends because of additional page allocations, even
if each skb could carry at least 100% more payload than before.
We can instruct sock_alloc_send_pskb() to attempt high order
allocations.
Most of the time, it does a single page allocation instead of 8.
I added an additional parameter to sock_alloc_send_pskb() to
let other users to opt-in for this new feature on followup patches.
Tested:
Before patch :
$ netperf -t STREAM_STREAM
STREAM STREAM TEST
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
2304 212992 212992 10.00 46861.15
After patch :
$ netperf -t STREAM_STREAM
STREAM STREAM TEST
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
2304 212992 212992 10.00 57981.11
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eliezer renames several *ll_poll to *busy_poll, but forgets
CONFIG_NET_LL_RX_POLL, so in case of confusion, rename it too.
Cc: Eliezer Tamir <eliezer.tamir@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Cong Wang <amwang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 547669d483 ("tcp: xps: fix reordering issues") added
unexpected reorders in case netem is used in a MQ setup for high
performance test bed.
ETH=eth0
tc qd del dev $ETH root 2>/dev/null
tc qd add dev $ETH root handle 1: mq
for i in `seq 1 32`
do
tc qd add dev $ETH parent 1:$i netem delay 100ms
done
As all tcp packets are orphaned by netem, TCP stack believes it can
set skb->ooo_okay on all packets.
In order to allow producers to send more packets, we want to
keep sk_wmem_alloc from reaching sk_sndbuf limit.
We can do that by accounting one byte per skb in netem queues,
so that TCP stack is not fooled too much.
Tested:
With above MQ/netem setup, scaling number of concurrent flows gives
linear results and no reorders/retransmits
lpq83:~# for n in 1 10 20 30 40 50 60 70 80 90 100
do echo -n "n:$n " ; ./super_netperf $n -H 10.7.7.84; done
n:1 198.46
n:10 2002.69
n:20 4000.98
n:30 6006.35
n:40 8020.93
n:50 10032.3
n:60 12081.9
n:70 13971.3
n:80 16009.7
n:90 17117.3
n:100 17425.5
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Idea of this patch is to add optional limitation of number of
unsent bytes in TCP sockets, to reduce usage of kernel memory.
TCP receiver might announce a big window, and TCP sender autotuning
might allow a large amount of bytes in write queue, but this has little
performance impact if a large part of this buffering is wasted :
Write queue needs to be large only to deal with large BDP, not
necessarily to cope with scheduling delays (incoming ACKS make room
for the application to queue more bytes)
For most workloads, using a value of 128 KB or less is OK to give
applications enough time to react to POLLOUT events in time
(or being awaken in a blocking sendmsg())
This patch adds two ways to set the limit :
1) Per socket option TCP_NOTSENT_LOWAT
2) A sysctl (/proc/sys/net/ipv4/tcp_notsent_lowat) for sockets
not using TCP_NOTSENT_LOWAT socket option (or setting a zero value)
Default value being UINT_MAX (0xFFFFFFFF), meaning this has no effect.
This changes poll()/select()/epoll() to report POLLOUT
only if number of unsent bytes is below tp->nosent_lowat
Note this might increase number of sendmsg()/sendfile() calls
when using non blocking sockets,
and increase number of context switches for blocking sockets.
Note this is not related to SO_SNDLOWAT (as SO_SNDLOWAT is
defined as :
Specify the minimum number of bytes in the buffer until
the socket layer will pass the data to the protocol)
Tested:
netperf sessions, and watching /proc/net/protocols "memory" column for TCP
With 200 concurrent netperf -t TCP_STREAM sessions, amount of kernel memory
used by TCP buffers shrinks by ~55 % (20567 pages instead of 45458)
lpq83:~# echo -1 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# (super_netperf 200 -t TCP_STREAM -H remote -l 90 &); sleep 60 ; grep TCP /proc/net/protocols
TCPv6 1880 2 45458 no 208 yes ipv6 y y y y y y y y y y y y y n y y y y y
TCP 1696 508 45458 no 208 yes kernel y y y y y y y y y y y y y n y y y y y
lpq83:~# echo 131072 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# (super_netperf 200 -t TCP_STREAM -H remote -l 90 &); sleep 60 ; grep TCP /proc/net/protocols
TCPv6 1880 2 20567 no 208 yes ipv6 y y y y y y y y y y y y y n y y y y y
TCP 1696 508 20567 no 208 yes kernel y y y y y y y y y y y y y n y y y y y
Using 128KB has no bad effect on the throughput or cpu usage
of a single flow, although there is an increase of context switches.
A bonus is that we hold socket lock for a shorter amount
of time and should improve latencies of ACK processing.
lpq83:~# echo -1 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# perf stat -e context-switches ./netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3
OMNI Send TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 7.7.7.84 () port 0 AF_INET : +/-2.500% @ 99% conf.
Local Remote Local Elapsed Throughput Throughput Local Local Remote Remote Local Remote Service
Send Socket Recv Socket Send Time Units CPU CPU CPU CPU Service Service Demand
Size Size Size (sec) Util Util Util Util Demand Demand Units
Final Final % Method % Method
1651584 6291456 16384 20.00 17447.90 10^6bits/s 3.13 S -1.00 U 0.353 -1.000 usec/KB
Performance counter stats for './netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3':
412,514 context-switches
200.034645535 seconds time elapsed
lpq83:~# echo 131072 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# perf stat -e context-switches ./netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3
OMNI Send TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 7.7.7.84 () port 0 AF_INET : +/-2.500% @ 99% conf.
Local Remote Local Elapsed Throughput Throughput Local Local Remote Remote Local Remote Service
Send Socket Recv Socket Send Time Units CPU CPU CPU CPU Service Service Demand
Size Size Size (sec) Util Util Util Util Demand Demand Units
Final Final % Method % Method
1593240 6291456 16384 20.00 17321.16 10^6bits/s 3.35 S -1.00 U 0.381 -1.000 usec/KB
Performance counter stats for './netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3':
2,675,818 context-switches
200.029651391 seconds time elapsed
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-By: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several call sites use the hardcoded following condition :
sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)
Lets use a helper because TCP_NOTSENT_LOWAT support will change this
condition for TCP sockets.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch moves the private error queue delivery function from the
af_packet code to the core socket method. In this way, network layers
only needing the error queue for transmit time stamping can share common
code.
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit eea86af6b1 ("net: sock: adapt SOCK_MIN_RCVBUF and
SOCK_MIN_SNDBUF") forgot the sk_buff alignment taken into account
in __alloc_skb() : skb->truesize = SKB_TRUESIZE(size);
While above commit fixed the sender issue, the receiver is still
dropping the second packet (on loopback device), because the receiver
socket can not really hold two skbs :
First packet truesize already is above sk_rcvbuf, so even TCP coalescing
cannot help.
On a typical 64bit build, each tcp skb truesize is 2304, instead of 2272
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <dborkman@redhat.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Tested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current situation is that SOCK_MIN_RCVBUF is 2048 + sizeof(struct sk_buff))
while SOCK_MIN_SNDBUF is 2048. Since in both cases, skb->truesize is used for
sk_{r,w}mem_alloc accounting, we should have both sizes adjusted via defining a
TCP_SKB_MIN_TRUESIZE.
Further, as Eric Dumazet points out, the minimal skb truesize in transmit path is
SKB_TRUESIZE(2048) after commit f07d960df3 ("tcp: avoid frag allocation for
small frames"), and tcp_sendmsg() tries to limit skb size to half the congestion
window, meaning we try to build two skbs at minimum. Thus, having SOCK_MIN_SNDBUF
as 2048 can hit a small regression for some applications setting to low
SO_SNDBUF / SO_RCVBUF. Note that we define a TCP_SKB_MIN_TRUESIZE, because
SKB_TRUESIZE(2048) adds SKB_DATA_ALIGN(sizeof(struct skb_shared_info)), but in
case of TCP skbs, the skb_shared_info is part of the 2048 bytes allocation for
skb->head.
The minor adaption in sk_stream_moderate_sndbuf() is to silence a warning by
using a typed max macro, as similarly done in SOCK_MIN_RCVBUF occurences, that
would appear otherwise.
Suggested-by: Eric Dumazet <eric.dumazet@gmail.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>
adds a socket option for low latency polling.
This allows overriding the global sysctl value with a per-socket one.
Unexport sysctl_net_ll_poll since for now it's not needed in modules.
Signed-off-by: Eliezer Tamir <eliezer.tamir@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adds an ndo_ll_poll method and the code that supports it.
This method can be used by low latency applications to busy-poll
Ethernet device queues directly from the socket code.
sysctl_net_ll_poll controls how many microseconds to poll.
Default is zero (disabled).
Individual protocol support will be added by subsequent patches.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Eliezer Tamir <eliezer.tamir@linux.intel.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Tested-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We have seen multiple NULL dereferences in __inet6_lookup_established()
After analysis, I found that inet6_sk() could be NULL while the
check for sk_family == AF_INET6 was true.
Bug was added in linux-2.6.29 when RCU lookups were introduced in UDP
and TCP stacks.
Once an IPv6 socket, using SLAB_DESTROY_BY_RCU is inserted in a hash
table, we no longer can clear pinet6 field.
This patch extends logic used in commit fcbdf09d96
("net: fix nulls list corruptions in sk_prot_alloc")
TCP/UDP/UDPLite IPv6 protocols provide their own .clear_sk() method
to make sure we do not clear pinet6 field.
At socket clone phase, we do not really care, as cloning the parent (non
NULL) pinet6 is not adding a fatal race.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, sock_tx_timestamp() always returns 0. The comment that
describes the sock_tx_timestamp() function wrongly says that it
returns an error when an invalid argument is passed (from commit
20d4947353, ``net: socket infrastructure for SO_TIMESTAMPING'').
Make the function void, so that we can also remove all the unneeded
if conditions that check for such a _non-existant_ error case in the
output path.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, when a socket receives something on the error queue it only wakes up
the socket on select if it is in the "read" list, that is the socket has
something to read. It is useful also to wake the socket if it is in the error
list, which would enable software to wait on error queue packets without waking
up for regular data on the socket. The main use case is for receiving
timestamped transmit packets which return the timestamp to the socket via the
error queue. This enables an application to select on the socket for the error
queue only instead of for the regular traffic.
-v2-
* Added the SO_SELECT_ERR_QUEUE socket option to every architechture specific file
* Modified every socket poll function that checks error queue
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jeffrey Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Richard Cochran <richardcochran@gmail.com>
Cc: Matthew Vick <matthew.vick@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller