Skip to content

cloud-kernel
cloud-kernel

openanolis / cloud-kernel
1 年多 前同步成功

通知 160
Star 36
Fork 7
cloud-kernel
cloud-kernel
切换分支/标签
  1. 24 4月, 2015 1 次提交
    • E
      inet: fix possible panic in reqsk_queue_unlink() · b357a364
      由 Eric Dumazet 提交于 
      [ 3897.923145] BUG: unable to handle kernel NULL pointer dereference at
 0000000000000080
[ 3897.931025] IP: [<ffffffffa9f27686>] reqsk_timer_handler+0x1a6/0x243

There is a race when reqsk_timer_handler() and tcp_check_req() call
inet_csk_reqsk_queue_unlink() on the same req at the same time.

Before commit fa76ce73 ("inet: get rid of central tcp/dccp listener
timer"), listener spinlock was held and race could not happen.

To solve this bug, we change reqsk_queue_unlink() to not assume req
must be found, and we return a status, to conditionally release a
refcount on the request sock.

This also means tcp_check_req() in non fastopen case might or not
consume req refcount, so tcp_v6_hnd_req() & tcp_v4_hnd_req() have
to properly handle this.

(Same remark for dccp_check_req() and its callers)

inet_csk_reqsk_queue_drop() is now too big to be inlined, as it is
called 4 times in tcp and 3 times in dccp.

Fixes: fa76ce73 ("inet: get rid of central tcp/dccp listener timer")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Reported-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      b357a364
  3. 14 4月, 2015 1 次提交
    • E
      tcp/dccp: get rid of central timewait timer · 789f558c
      由 Eric Dumazet 提交于 
      Using a timer wheel for timewait sockets was nice ~15 years ago when
memory was expensive and machines had a single processor.

This does not scale, code is ugly and source of huge latencies
(Typically 30 ms have been seen, cpus spinning on death_lock spinlock.)

We can afford to use an extra 64 bytes per timewait sock and spread
timewait load to all cpus to have better behavior.

Tested:

On following test, /proc/sys/net/ipv4/tcp_tw_recycle is set to 1
on the target (lpaa24)

Before patch :

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
419594

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
437171

While test is running, we can observe 25 or even 33 ms latencies.

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 20601ms
rtt min/avg/max/mdev = 0.020/0.217/25.771/1.535 ms, pipe 2

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 20702ms
rtt min/avg/max/mdev = 0.019/0.183/33.761/1.441 ms, pipe 2

After patch :

About 90% increase of throughput :

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
810442

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
800992

And latencies are kept to minimal values during this load, even
if network utilization is 90% higher :

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 19991ms
rtt min/avg/max/mdev = 0.023/0.064/0.360/0.042 ms
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      789f558c
  5. 01 4月, 2015 2 次提交
  7. 30 3月, 2015 1 次提交
  9. 25 3月, 2015 4 次提交
  11. 24 3月, 2015 2 次提交
  13. 21 3月, 2015 2 次提交
    • E
      inet: get rid of central tcp/dccp listener timer · fa76ce73
      由 Eric Dumazet 提交于 
      One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.

This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.

SYNACK are sent in huge bursts, likely to cause severe drops anyway.

This model was OK 15 years ago when memory was very tight.

We now can afford to have a timer per request sock.

Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.

With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.

This is ~100 times more what we could achieve before this patch.

Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      fa76ce73
    • E
      inet: drop prev pointer handling in request sock · 52452c54
      由 Eric Dumazet 提交于 
      When request sock are put in ehash table, the whole notion
of having a previous request to update dl_next is pointless.

Also, following patch will get rid of big purge timer,
so we want to delete a request sock without holding listener lock.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      52452c54
  15. 19 3月, 2015 2 次提交
  17. 17 3月, 2015 1 次提交
  19. 15 3月, 2015 1 次提交
  21. 13 3月, 2015 2 次提交
  23. 06 1月, 2015 1 次提交
    • D
      net: tcp: add per route congestion control · 81164413
      由 Daniel Borkmann 提交于 
      This work adds the possibility to define a per route/destination
congestion control algorithm. Generally, this opens up the possibility
for a machine with different links to enforce specific congestion
control algorithms with optimal strategies for each of them based
on their network characteristics, even transparently for a single
application listening on all links.

For our specific use case, this additionally facilitates deployment
of DCTCP, for example, applications can easily serve internal
traffic/dsts in DCTCP and external one with CUBIC. Other scenarios
would also allow for utilizing e.g. long living, low priority
background flows for certain destinations/routes while still being
able for normal traffic to utilize the default congestion control
algorithm. We also thought about a per netns setting (where different
defaults are possible), but given its actually a link specific
property, we argue that a per route/destination setting is the most
natural and flexible.

The administrator can utilize this through ip-route(8) by appending
"congctl [lock] <name>", where <name> denotes the name of a
congestion control algorithm and the optional lock parameter allows
to enforce the given algorithm so that applications in user space
would not be allowed to overwrite that algorithm for that destination.

The dst metric lookups are being done when a dst entry is already
available in order to avoid a costly lookup and still before the
algorithms are being initialized, thus overhead is very low when the
feature is not being used. While the client side would need to drop
the current reference on the module, on server side this can actually
even be avoided as we just got a flat-copied socket clone.

Joint work with Florian Westphal.
Suggested-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      81164413
  25. 23 12月, 2014 1 次提交
  27. 10 12月, 2014 1 次提交
  29. 26 11月, 2014 1 次提交
  31. 12 11月, 2014 2 次提交
    • E
      net: introduce SO_INCOMING_CPU · 2c8c56e1
      由 Eric Dumazet 提交于 
      Alternative to RPS/RFS is to use hardware support for multiple
queues.

Then split a set of million of sockets into worker threads, each
one using epoll() to manage events on its own socket pool.

Ideally, we want one thread per RX/TX queue/cpu, but we have no way to
know after accept() or connect() on which queue/cpu a socket is managed.

We normally use one cpu per RX queue (IRQ smp_affinity being properly
set), so remembering on socket structure which cpu delivered last packet
is enough to solve the problem.

After accept(), connect(), or even file descriptor passing around
processes, applications can use :

 int cpu;
 socklen_t len = sizeof(cpu);

 getsockopt(fd, SOL_SOCKET, SO_INCOMING_CPU, &cpu, &len);

And use this information to put the socket into the right silo
for optimal performance, as all networking stack should run
on the appropriate cpu, without need to send IPI (RPS/RFS).
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      2c8c56e1
    • E
      tcp: move sk_mark_napi_id() at the right place · 3d97379a
      由 Eric Dumazet 提交于 
      sk_mark_napi_id() is used to record for a flow napi id of incoming
packets for busypoll sake.
We should do this only on established flows, not on listeners.

This was 'working' by virtue of the socket cloning, but doing
this on SYN packets in unecessary cache line dirtying.

Even if we move sk_napi_id in the same cache line than sk_lock,
we are working to make SYN processing lockless, so it is desirable
to set sk_napi_id only for established flows.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      3d97379a
  33. 23 10月, 2014 1 次提交
    • S
      net: fix saving TX flow hash in sock for outgoing connections · 9e7ceb06
      由 Sathya Perla 提交于 
      The commit "net: Save TX flow hash in sock and set in skbuf on xmit"
introduced the inet_set_txhash() and ip6_set_txhash() routines to calculate
and record flow hash(sk_txhash) in the socket structure. sk_txhash is used
to set skb->hash which is used to spread flows across multiple TXQs.

But, the above routines are invoked before the source port of the connection
is created. Because of this all outgoing connections that just differ in the
source port get hashed into the same TXQ.

This patch fixes this problem for IPv4/6 by invoking the the above routines
after the source port is available for the socket.

Fixes: b73c3d0e("net: Save TX flow hash in sock and set in skbuf on xmit")
Signed-off-by: NSathya Perla <sathya.perla@emulex.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      9e7ceb06
  35. 18 10月, 2014 1 次提交
  37. 29 9月, 2014 2 次提交
    • E
      tcp: better TCP_SKB_CB layout to reduce cache line misses · 971f10ec
      由 Eric Dumazet 提交于 
      TCP maintains lists of skb in write queue, and in receive queues
(in order and out of order queues)

Scanning these lists both in input and output path usually requires
access to skb->next, TCP_SKB_CB(skb)->seq, and TCP_SKB_CB(skb)->end_seq

These fields are currently in two different cache lines, meaning we
waste lot of memory bandwidth when these queues are big and flows
have either packet drops or packet reorders.

We can move TCP_SKB_CB(skb)->header at the end of TCP_SKB_CB, because
this header is not used in fast path. This allows TCP to search much faster
in the skb lists.

Even with regular flows, we save one cache line miss in fast path.

Thanks to Christoph Paasch for noticing we need to cleanup
skb->cb[] (IPCB/IP6CB) before entering IP stack in tx path,
and that I forgot IPCB use in tcp_v4_hnd_req() and tcp_v4_save_options().
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      971f10ec
    • E
      ipv6: add a struct inet6_skb_parm param to ipv6_opt_accepted() · a224772d
      由 Eric Dumazet 提交于 
      ipv6_opt_accepted() assumes IP6CB(skb) holds the struct inet6_skb_parm
that it needs. Lets not assume this, as TCP stack might use a different
place.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      a224772d
  39. 28 9月, 2014 1 次提交
  41. 16 9月, 2014 1 次提交
    • E
      tcp: use TCP_SKB_CB(skb)->tcp_flags in input path · e11ecddf
      由 Eric Dumazet 提交于 
      Input path of TCP do not currently uses TCP_SKB_CB(skb)->tcp_flags,
which is only used in output path.

tcp_recvmsg(), looks at tcp_hdr(skb)->syn for every skb found in receive queue,
and its unfortunate because this bit is located in a cache line right before
the payload.

We can simplify TCP by copying tcp flags into TCP_SKB_CB(skb)->tcp_flags.

This patch does so, and avoids the cache line miss in tcp_recvmsg()

Following patches will
- allow a segment with FIN being coalesced in tcp_try_coalesce()
- simplify tcp_collapse() by not copying the headers.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      e11ecddf
  43. 10 9月, 2014 1 次提交
    • E
      tcp: remove dst refcount false sharing for prequeue mode · ca777eff
      由 Eric Dumazet 提交于 
      Alexander Duyck reported high false sharing on dst refcount in tcp stack
when prequeue is used. prequeue is the mechanism used when a thread is
blocked in recvmsg()/read() on a TCP socket, using a blocking model
rather than select()/poll()/epoll() non blocking one.

We already try to use RCU in input path as much as possible, but we were
forced to take a refcount on the dst when skb escaped RCU protected
region. When/if the user thread runs on different cpu, dst_release()
will then touch dst refcount again.

Commit 09316255 (tcp: force a dst refcount when prequeue packet)
was an example of a race fix.

It turns out the only remaining usage of skb->dst for a packet stored
in a TCP socket prequeue is IP early demux.

We can add a logic to detect when IP early demux is probably going
to use skb->dst. Because we do an optimistic check rather than duplicate
existing logic, we need to guard inet_sk_rx_dst_set() and
inet6_sk_rx_dst_set() from using a NULL dst.

Many thanks to Alexander for providing a nice bug report, git bisection,
and reproducer.

Tested using Alexander script on a 40Gb NIC, 8 RX queues.
Hosts have 24 cores, 48 hyper threads.

echo 0 >/proc/sys/net/ipv4/tcp_autocorking

for i in `seq 0 47`
do
  for j in `seq 0 2`
  do
     netperf -H $DEST -t TCP_STREAM -l 1000 \
             -c -C -T $i,$i -P 0 -- \
             -m 64 -s 64K -D &
  done
done

Before patch : ~6Mpps and ~95% cpu usage on receiver
After patch : ~9Mpps and ~35% cpu usage on receiver.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Reported-by: NAlexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      ca777eff
  45. 06 9月, 2014 1 次提交
  47. 15 8月, 2014 1 次提交
  49. 07 8月, 2014 1 次提交
  51. 08 7月, 2014 2 次提交
    • T
      net: Save TX flow hash in sock and set in skbuf on xmit · b73c3d0e
      由 Tom Herbert 提交于 
      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: NTom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      b73c3d0e
    • N
      tcp: switch snt_synack back to measuring transmit time of first SYNACK · 86c6a2c7
      由 Neal Cardwell 提交于 
      Always store in snt_synack the time at which the server received the
first client SYN and attempted to send the first SYNACK.

Recent commit aa27fc50 ("tcp: tcp_v[46]_conn_request: fix snt_synack
initialization") resolved an inconsistency between IPv4 and IPv6 in
the initialization of snt_synack. This commit brings back the idea
from 843f4a55 (tcp: use tcp_v4_send_synack on first SYN-ACK), which
was going for the original behavior of snt_synack from the commit
where it was added in 9ad7c049 ("tcp: RFC2988bis + taking RTT
sample from 3WHS for the passive open side") in v3.1.

In addition to being simpler (and probably a tiny bit faster),
unconditionally storing the time of the first SYNACK attempt has been
useful because it allows calculating a performance metric quantifying
how long it took to establish a passive TCP connection.
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Cc: Octavian Purdila <octavian.purdila@intel.com>
Cc: Jerry Chu <hkchu@google.com>
Acked-by: NOctavian Purdila <octavian.purdila@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      86c6a2c7
  53. 28 6月, 2014 3 次提交