openwrt/staging/blogic.git
15 years agonetns xfrm: trivial netns propagations
Alexey Dobriyan [Wed, 26 Nov 2008 01:29:47 +0000 (17:29 -0800)]
netns xfrm: trivial netns propagations

Take netns from xfrm_state or xfrm_policy.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: propagate netns into bydst/bysrc/byspi hash functions
Alexey Dobriyan [Wed, 26 Nov 2008 01:29:21 +0000 (17:29 -0800)]
netns xfrm: propagate netns into bydst/bysrc/byspi hash functions

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns policy hash resizing work
Alexey Dobriyan [Wed, 26 Nov 2008 01:28:57 +0000 (17:28 -0800)]
netns xfrm: per-netns policy hash resizing work

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns policy counts
Alexey Dobriyan [Wed, 26 Nov 2008 01:24:15 +0000 (17:24 -0800)]
netns xfrm: per-netns policy counts

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_policy_bydst hash
Alexey Dobriyan [Wed, 26 Nov 2008 01:23:48 +0000 (17:23 -0800)]
netns xfrm: per-netns xfrm_policy_bydst hash

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns inexact policies
Alexey Dobriyan [Wed, 26 Nov 2008 01:23:26 +0000 (17:23 -0800)]
netns xfrm: per-netns inexact policies

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_policy_byidx hashmask
Alexey Dobriyan [Wed, 26 Nov 2008 01:22:58 +0000 (17:22 -0800)]
netns xfrm: per-netns xfrm_policy_byidx hashmask

Per-netns hashes are independently resizeable.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_policy_byidx hash
Alexey Dobriyan [Wed, 26 Nov 2008 01:22:35 +0000 (17:22 -0800)]
netns xfrm: per-netns xfrm_policy_byidx hash

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns policy list
Alexey Dobriyan [Wed, 26 Nov 2008 01:22:11 +0000 (17:22 -0800)]
netns xfrm: per-netns policy list

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: add struct xfrm_policy::xp_net
Alexey Dobriyan [Wed, 26 Nov 2008 01:21:45 +0000 (17:21 -0800)]
netns xfrm: add struct xfrm_policy::xp_net

Again, to avoid complications with passing netns when not necessary.
Again, ->xp_net is set-once field, once set it never changes.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns km_waitq
Alexey Dobriyan [Wed, 26 Nov 2008 01:21:01 +0000 (17:21 -0800)]
netns xfrm: per-netns km_waitq

Disallow spurious wakeups in __xfrm_lookup().

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns state GC work
Alexey Dobriyan [Wed, 26 Nov 2008 01:20:36 +0000 (17:20 -0800)]
netns xfrm: per-netns state GC work

State GC is per-netns, and this is part of it.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns state GC list
Alexey Dobriyan [Wed, 26 Nov 2008 01:20:11 +0000 (17:20 -0800)]
netns xfrm: per-netns state GC list

km_waitq is going to be made per-netns to disallow spurious wakeups
in __xfrm_lookup().

To not wakeup after every garbage-collected xfrm_state (which potentially
can be from different netns) make state GC list per-netns.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_hash_work
Alexey Dobriyan [Wed, 26 Nov 2008 01:19:07 +0000 (17:19 -0800)]
netns xfrm: per-netns xfrm_hash_work

All of this is implicit passing which netns's hashes should be resized.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_state counts
Alexey Dobriyan [Wed, 26 Nov 2008 01:18:39 +0000 (17:18 -0800)]
netns xfrm: per-netns xfrm_state counts

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_state_hmask
Alexey Dobriyan [Wed, 26 Nov 2008 01:18:12 +0000 (17:18 -0800)]
netns xfrm: per-netns xfrm_state_hmask

Since hashtables are per-netns, they can be independently resized.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_state_byspi hash
Alexey Dobriyan [Wed, 26 Nov 2008 01:17:47 +0000 (17:17 -0800)]
netns xfrm: per-netns xfrm_state_byspi hash

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_state_bysrc hash
Alexey Dobriyan [Wed, 26 Nov 2008 01:17:24 +0000 (17:17 -0800)]
netns xfrm: per-netns xfrm_state_bysrc hash

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_state_bydst hash
Alexey Dobriyan [Wed, 26 Nov 2008 01:16:58 +0000 (17:16 -0800)]
netns xfrm: per-netns xfrm_state_bydst hash

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: per-netns xfrm_state_all list
Alexey Dobriyan [Wed, 26 Nov 2008 01:16:11 +0000 (17:16 -0800)]
netns xfrm: per-netns xfrm_state_all list

This is done to get
a) simple "something leaked" check
b) cover possible DoSes when other netns puts many, many xfrm_states
   onto a list.
c) not miss "alien xfrm_state" check in some of list iterators in future.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: add struct xfrm_state::xs_net
Alexey Dobriyan [Wed, 26 Nov 2008 01:15:16 +0000 (17:15 -0800)]
netns xfrm: add struct xfrm_state::xs_net

To avoid unnecessary complications with passing netns around.

* set once, very early after allocating
* once set, never changes

For a while create every xfrm_state in init_net.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns xfrm: add netns boilerplate
Alexey Dobriyan [Wed, 26 Nov 2008 01:14:31 +0000 (17:14 -0800)]
netns xfrm: add netns boilerplate

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoxfrm: initialise xfrm_policy_gc_work statically
Alexey Dobriyan [Wed, 26 Nov 2008 01:13:59 +0000 (17:13 -0800)]
xfrm: initialise xfrm_policy_gc_work statically

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years ago3c523: fix warning in drivers/net/3c523.c
Ingo Molnar [Wed, 26 Nov 2008 01:02:20 +0000 (17:02 -0800)]
3c523: fix warning in drivers/net/3c523.c

fix warning:

  drivers/net/3c523.c:582: warning: ‘cleanup_card’ defined but not used

No code changed:

md5:
   ebe4a1b27d3f21b0b12a78c58463b0d7  3c523.o.before.asm
   ebe4a1b27d3f21b0b12a78c58463b0d7  3c523.o.after.asm

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodepca: fix warning in drivers/net/depca.c
Ingo Molnar [Wed, 26 Nov 2008 01:00:39 +0000 (17:00 -0800)]
depca: fix warning in drivers/net/depca.c

fix warning:

  drivers/net/depca.c: In function ‘depca_eisa_probe’:
  drivers/net/depca.c:1564: warning: ‘mem_start’ may be used uninitialized in this function

this seems to be a real bug - depca_eisa_probe() does not check
for failure. Add it, symmetric to depca_isa_probe().

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoatlx: fix warning in drivers/net/atlx/atl2.c
Ingo Molnar [Wed, 26 Nov 2008 01:00:05 +0000 (17:00 -0800)]
atlx: fix warning in drivers/net/atlx/atl2.c

fix this warning:

  drivers/net/atlx/atl2.c: In function ‘atl2_request_irq’:
  drivers/net/atlx/atl2.c:644: warning: unused variable ‘err’

'err' is unused in the !CONFIG_PCI_MSI case.

Instead of further increasing the #ifdeffery in this function,
restructure the code a bit and get rid of the #ifdef. This
relies on the fact that pci_enable_msi() will always fail in
the !CONFIG_PCI_MSI case.

There should be no change in driver behavior.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agobluetooth: fix warning in net/bluetooth/rfcomm/sock.c
Ingo Molnar [Wed, 26 Nov 2008 00:59:21 +0000 (16:59 -0800)]
bluetooth: fix warning in net/bluetooth/rfcomm/sock.c

fix this warning:

  net/bluetooth/rfcomm/sock.c: In function ‘rfcomm_sock_ioctl’:
  net/bluetooth/rfcomm/sock.c:795: warning: unused variable ‘sk’

perhaps BT_DEBUG() should be improved to do printf format checking
instead of the #ifdef, but that looks quite intrusive: each bluetooth
.c file undefines the macro.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agosunrpc: fix warning in net/sunrpc/xprtrdma/verbs.c
Ingo Molnar [Wed, 26 Nov 2008 00:58:42 +0000 (16:58 -0800)]
sunrpc: fix warning in net/sunrpc/xprtrdma/verbs.c

fix this warning:

  net/sunrpc/xprtrdma/verbs.c: In function ‘rpcrdma_conn_upcall’:
  net/sunrpc/xprtrdma/verbs.c:279: warning: unused variable ‘addr’

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoax25: fix warning in net/ax25/sysctl_net_ax25.c
Ingo Molnar [Wed, 26 Nov 2008 00:58:19 +0000 (16:58 -0800)]
ax25: fix warning in net/ax25/sysctl_net_ax25.c

fix this warning:

  net/ax25/sysctl_net_ax25.c:27: warning: ‘min_ds_timeout’ defined but not used
  net/ax25/sysctl_net_ax25.c:27: warning: ‘max_ds_timeout’ defined but not used

These are only used in the CONFIG_AX25_DAMA_SLAVE case.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agomlx4: fix warning in drivers/net/mlx4/mcg.c
Ingo Molnar [Wed, 26 Nov 2008 00:57:59 +0000 (16:57 -0800)]
mlx4: fix warning in drivers/net/mlx4/mcg.c

fix warning:

  drivers/net/mlx4/mcg.c: In function ‘mlx4_multicast_attach’:
  drivers/net/mlx4/mcg.c:217: warning: integer overflow in expression

there was no real danger of overflow here though.

md5:
   db8eb55620f886c03854a2abb2ce6c3f  mcg.o.before.asm
   db8eb55620f886c03854a2abb2ce6c3f  mcg.o.after.asm

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodccp: fix warning in net/dccp/options.c
Ingo Molnar [Wed, 26 Nov 2008 00:57:30 +0000 (16:57 -0800)]
dccp: fix warning in net/dccp/options.c

this warning:

  net/dccp/options.c: In function ‘dccp_parse_options’:
  net/dccp/options.c:67: warning: ‘value’ may be used uninitialized in this function

is a bogus GCC warning. The compiler does not recognize the relation
between "value" and "mandatory" variables: the code flow can ever reach
the "out_invalid_option:" label if 'mandatory' is set to 1, and when
'mandatory' is non-zero, we'll always have 'value' initialized.

Help out the compiler by annotating the variable.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agocassini: fix warning in drivers/net/cassini.c
Ingo Molnar [Wed, 26 Nov 2008 00:57:05 +0000 (16:57 -0800)]
cassini: fix warning in drivers/net/cassini.c

this warning:

  drivers/net/cassini.c: In function ‘cas_rx_ringN’:
  drivers/net/cassini.c:2350: warning: ‘skb’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between cas_rx_process_pkt() and 'skb'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agomlx4: fix error path in drivers/net/mlx4/en_rx.c
Ingo Molnar [Wed, 26 Nov 2008 00:53:32 +0000 (16:53 -0800)]
mlx4: fix error path in drivers/net/mlx4/en_rx.c

this warning:

  drivers/net/mlx4/en_rx.c: In function ‘mlx4_en_activate_rx_rings’:
  drivers/net/mlx4/en_rx.c:412: warning: ‘err’ may be used uninitialized in this function

Triggers because 'err' is uninitialized in the following input
conditions: priv->rx_ring_num is zero and mlx4_en_fill_rx_buffers()
fails.

But even if ->rx_ring_num is nonzero, 'err' will be zero if
mlx4_en_fill_rx_buffers() fails and mlx4_en_activate_rx_rings() returns
success - incorrectly.

So it's best to keep the error code uptodate on mlx4_en_fill_rx_buffers()
calls as well.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoz85230: fix warning in drivers/net/wan/z85230.c
Ingo Molnar [Wed, 26 Nov 2008 00:53:08 +0000 (16:53 -0800)]
z85230: fix warning in drivers/net/wan/z85230.c

this warning:

  drivers/net/wan/z85230.c: In function ‘z8530_interrupt’:
  drivers/net/wan/z85230.c:713: warning: ‘intr’ may be used uninitialized in this function

is clearly bogus - annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agosis900: fix warning in drivers/net/sis900.c
Ingo Molnar [Wed, 26 Nov 2008 00:52:13 +0000 (16:52 -0800)]
sis900: fix warning in drivers/net/sis900.c

this warning:

  drivers/net/sis900.c: In function ‘sis900_timer’:
  drivers/net/sis900.c:1280: warning: ‘speed’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between sis900_read_mode(), 'speed' and 'duplex'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agofix warning in fs/dlm/netlink.c
Ingo Molnar [Wed, 26 Nov 2008 00:51:45 +0000 (16:51 -0800)]
fix warning in fs/dlm/netlink.c

this warning:

  fs/dlm/netlink.c: In function ‘dlm_timeout_warn’:
  fs/dlm/netlink.c:131: warning: ‘send_skb’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between prepare_data() and send_skb.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodsa: fix warning in net/dsa/mv88e6060.c
Ingo Molnar [Wed, 26 Nov 2008 00:51:13 +0000 (16:51 -0800)]
dsa: fix warning in net/dsa/mv88e6060.c

this warning:

  net/dsa/mv88e6060.c: In function ‘mv88e6060_poll_link’:
  net/dsa/mv88e6060.c:225: warning: ‘port_status’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between 'link' and 'port_status'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodsa: fix warning in net/dsa/mv88e6xxx.c
Ingo Molnar [Wed, 26 Nov 2008 00:50:49 +0000 (16:50 -0800)]
dsa: fix warning in net/dsa/mv88e6xxx.c

this warning:

  net/dsa/mv88e6xxx.c: In function ‘mv88e6xxx_poll_link’:
  net/dsa/mv88e6xxx.c:361: warning: ‘port_status’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between 'link' and 'port_status'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoipv6: fix warning in net/ipv6/ip6_flowlabel.c
Ingo Molnar [Wed, 26 Nov 2008 00:50:30 +0000 (16:50 -0800)]
ipv6: fix warning in net/ipv6/ip6_flowlabel.c

this warning:

  net/ipv6/ip6_flowlabel.c: In function ‘ipv6_flowlabel_opt’:
  net/ipv6/ip6_flowlabel.c:467: warning: ‘err’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between fl_create() and 'err'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agopkt_sched: fix warning in net/sched/sch_hfsc.c
Ingo Molnar [Wed, 26 Nov 2008 00:50:02 +0000 (16:50 -0800)]
pkt_sched: fix warning in net/sched/sch_hfsc.c

this warning:

  net/sched/sch_hfsc.c: In function ‘hfsc_enqueue’:
  net/sched/sch_hfsc.c:1577: warning: ‘err’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between hfsc_classify(), 'cl' and 'err'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agosunrpc: fix warning in net/sunrpc/xprtrdma/svc_rdma_transport.c
Ingo Molnar [Wed, 26 Nov 2008 00:49:37 +0000 (16:49 -0800)]
sunrpc: fix warning in net/sunrpc/xprtrdma/svc_rdma_transport.c

this warning:

  net/sunrpc/xprtrdma/svc_rdma_transport.c: In function ‘svc_rdma_accept’:
  net/sunrpc/xprtrdma/svc_rdma_transport.c:830: warning: ‘dma_mr_acc’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) flow connection
between need_dma_mr and dma_mr_acc.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoqla3xxx: fix warning in drivers/net/qla3xxx.c
Ingo Molnar [Wed, 26 Nov 2008 00:49:07 +0000 (16:49 -0800)]
qla3xxx: fix warning in drivers/net/qla3xxx.c

this warning:

  drivers/net/qla3xxx.c: In function ‘ql3xxx_probe’:
  drivers/net/qla3xxx.c:3912: warning: ‘pci_using_dac’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between 'pci_using_dac' and 'err'.

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoniu: fix another warning in drivers/net/niu.c
Ingo Molnar [Wed, 26 Nov 2008 00:48:42 +0000 (16:48 -0800)]
niu: fix another warning in drivers/net/niu.c

this warning:

  drivers/net/niu.c: In function ‘esr_reset’:
  drivers/net/niu.c:741: warning: ‘reset’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between:

 - esr_read_reset() and 'reset'

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoniu: fix warnings in drivers/net/niu.c
Ingo Molnar [Wed, 26 Nov 2008 00:48:12 +0000 (16:48 -0800)]
niu: fix warnings in drivers/net/niu.c

these warnings:

  drivers/net/niu.c: In function ‘serdes_init_niu_1g_serdes’:
  drivers/net/niu.c:451: warning: ‘sig’ may be used uninitialized in this function
  drivers/net/niu.c: In function ‘serdes_init_niu_10g_serdes’:
  drivers/net/niu.c:550: warning: ‘sig’ may be used uninitialized in this function

triggers because GCC does not recognize that the max_retry loop
always initializes 'sig', due to max_retry != 0.

Annotate them.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agos2io: fix warning in drivers/net/s2io.c
Ingo Molnar [Wed, 26 Nov 2008 00:47:35 +0000 (16:47 -0800)]
s2io: fix warning in drivers/net/s2io.c

this warning:

  drivers/net/s2io.c: In function ‘rx_intr_handler’:
  drivers/net/s2io.c:7369: warning: ‘lro’ may be used uninitialized in this function

triggers because GCC does not recognize the (correct) error flow
between:

 - s2io_club_tcp_session()and 'lro'

Annotate it.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetns: filter out uevent not belonging to init_net
Daniel Lezcano [Wed, 26 Nov 2008 00:46:37 +0000 (16:46 -0800)]
netns: filter out uevent not belonging to init_net

This patch will filter out the uevent not related to the init_net.
Without this patch if a network device is created in a network
namespace with the same name as one network device belonging to the
initial network namespace (eg. eth0), when the network namespace
will die and the network device will be destroyed, an event will
be sent and catched by the udevd daemon. That will result to have
the real network device to be shutdown because the udevd/uevent are
not namespace aware.

Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com>
Acked-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoixgbe: Naming interrupt vectors
Robert Olsson [Wed, 26 Nov 2008 00:43:52 +0000 (16:43 -0800)]
ixgbe: Naming interrupt vectors

Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Acked-by: Peter P Waskiewicz Jr <peter.p.waskiewicz.jr@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoniu: Naming interrupt vectors.
Robert Olsson [Wed, 26 Nov 2008 00:41:57 +0000 (16:41 -0800)]
niu: Naming interrupt vectors.

 A patch to put names on the niu interrupt vectors according the syntax below.
 This is needed to assign correct affinity.

 > So on a multiqueue card with 2 RX queues and 2 TX queues we'd
 > have names like:
 >
 >  eth0-rx-0
 >  eth0-rx-1
 >  eth0-tx-0
 >  eth0-tx-1

Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Tested-by: Jesper Dangaard Brouer <jdb@comx.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: skb_shift cannot cache frag ptrs past pskb_expand_head
Ilpo Järvinen [Tue, 25 Nov 2008 21:57:01 +0000 (13:57 -0800)]
tcp: skb_shift cannot cache frag ptrs past pskb_expand_head

Since pskb_expand_head creates copy of the shared area we
cannot keep any frag ptr past de-cloning. This fixes the
tcpdump recvfrom -EFAULT problem.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agopkt_sched: sch_api: Remove qdisc_list_lock
Jarek Poplawski [Tue, 25 Nov 2008 21:56:06 +0000 (13:56 -0800)]
pkt_sched: sch_api: Remove qdisc_list_lock

After implementing qdisc->ops->peek() there is no more calling
qdisc_tree_decrease_qlen() without rtnl_lock(), so qdisc_list_lock
added by commit: f6e0b239a2657ea8cb67f0d83d0bfdbfd19a481b "pkt_sched:
Fix qdisc list locking" can be removed.

Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: udp_unhash() can test if sk is hashed
Eric Dumazet [Tue, 25 Nov 2008 21:55:15 +0000 (13:55 -0800)]
net: udp_unhash() can test if sk is hashed

Impact: Optimization

Like done in inet_unhash(), we can avoid taking a chain lock if
socket is not hashed in udp_unhash()

Triggered by close(socket(AF_INET, SOCK_DGRAM, 0));

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: Make sure BHs are disabled in sock_prot_inuse_add()
Eric Dumazet [Tue, 25 Nov 2008 21:53:27 +0000 (13:53 -0800)]
net: Make sure BHs are disabled in sock_prot_inuse_add()

prot->destroy is not called with BH disabled. So we must add
explicit BH disable around call to sock_prot_inuse_add()
in sctp_destroy_sock()

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: tcp_limit_reno_sacked can become static
Ilpo Järvinen [Tue, 25 Nov 2008 21:45:29 +0000 (13:45 -0800)]
tcp: tcp_limit_reno_sacked can become static

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoRevert "hso: Fix crashes on close."
David S. Miller [Tue, 25 Nov 2008 11:53:09 +0000 (03:53 -0800)]
Revert "hso: Fix crashes on close."

This reverts commit 4a3e818181e1baf970e9232ca8b747e233176b87.

On request from Alan Cox.

Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoRevert "hso: Fix free of mutexes still in use."
David S. Miller [Tue, 25 Nov 2008 11:52:46 +0000 (03:52 -0800)]
Revert "hso: Fix free of mutexes still in use."

This reverts commit 52429eb216385fdc6969c0112ba8b46cffefaaef.

On request from Alan Cox.

Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoRevert "hso: Add TIOCM ioctl handling."
David S. Miller [Tue, 25 Nov 2008 11:52:17 +0000 (03:52 -0800)]
Revert "hso: Add TIOCM ioctl handling."

This reverts commit 7ea3a9ad9bf360f746a7ad6fa72511a5c359490d.

On request from Alan Cox.

Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoxfrm: remove useless forward declarations
Alexey Dobriyan [Tue, 25 Nov 2008 09:05:54 +0000 (01:05 -0800)]
xfrm: remove useless forward declarations

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoah4/ah6: remove useless NULL assignments
Alexey Dobriyan [Tue, 25 Nov 2008 09:05:09 +0000 (01:05 -0800)]
ah4/ah6: remove useless NULL assignments

struct will be kfreed in a moment, so...

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoigb: loopback bits not correctly cleared from RCTL register
Alexander Duyck [Tue, 25 Nov 2008 09:04:03 +0000 (01:04 -0800)]
igb: loopback bits not correctly cleared from RCTL register

This change forces the bits to 0 by using an &= operation with an inverted
mask of all options instead of using an |= with a value of 0.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoigb: remove unneeded bit refrence when enabling jumbo frames
Alexander Duyck [Tue, 25 Nov 2008 09:03:26 +0000 (01:03 -0800)]
igb: remove unneeded bit refrence when enabling jumbo frames

There is a reference to a Buffer Size extention bit that is unneded by
82575/82576 hardware.  Since it is not needed it should be removed from the
code.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoDCB: fix kconfig option
Jeff Kirsher [Tue, 25 Nov 2008 09:02:08 +0000 (01:02 -0800)]
DCB: fix kconfig option

Since the netlink option for DCB is necessary to actually be useful,
simplified the Kconfig option.  In addition, added useful help text for the
Kconfig option.

Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoaoe: remove private mac address format function
Harvey Harrison [Tue, 25 Nov 2008 08:40:37 +0000 (00:40 -0800)]
aoe: remove private mac address format function

Add %pm to omit the colons when printing a mac address.

Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agohso: Hook up ->reset_resume
Denis Joseph Barrow [Tue, 25 Nov 2008 08:36:10 +0000 (00:36 -0800)]
hso: Hook up ->reset_resume

Made usb_drivers reset_resume function point to hso_resume this
fixes problems a usb reset is done when the network interface
is left idle for a few minutes. Possibly reset_resume should
initialise hardware more but this works in the common case.

Signed-off-by: Denis Joseph Barrow <D.Barow@option.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agohso: Add TIOCM ioctl handling.
Denis Joseph Barrow [Tue, 25 Nov 2008 08:35:26 +0000 (00:35 -0800)]
hso: Add TIOCM ioctl handling.

Makes TIOCM ioctls for Data Carrier Detect & related functions
work like /drivers/serial/serial-core.c potentially needed
for pppd & similar user programs.

Signed-off-by: Denis Joseph Barrow <D.Barow@option.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agohso: Fix free of mutexes still in use.
Denis Joseph Barrow [Tue, 25 Nov 2008 08:33:13 +0000 (00:33 -0800)]
hso: Fix free of mutexes still in use.

A new structure hso_mutex_table had to be declared statically
& used as as hso_device mutex_lock(&serial->parent->mutex) etc
is freed in hso_serial_open & hso_serial_close by kref_put while
the mutex is still in use.

This is a substantial change but should make the driver much stabler.

Signed-off-by: Denis Joseph Barrow <D.Barow@option.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agohso: Fix URB submission -EINVAL.
Denis Joseph Barrow [Tue, 25 Nov 2008 08:30:48 +0000 (00:30 -0800)]
hso: Fix URB submission -EINVAL.

Added check for IFF_UP in hso_resume, this should eliminate -EINVAL (-22)
errors caused from urb's being submitted twice, once by hso_resume
& once in hso_net_open, if suspend/resume USB power saving  mode is enabled

Signed-off-by: Denis Joseph Barrow <D.Barow@option.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agohso: Fix crashes on close.
Denis Joseph Barrow [Tue, 25 Nov 2008 08:27:50 +0000 (00:27 -0800)]
hso: Fix crashes on close.

Moved serial_open_count in hso_serial_open to
prevent crashes owing to the serial structure being made NULL
when hso_serial_close is called even though hso_serial_open
returned -ENODEV, Alan Cox pointed out this happens,
also put in sanity check in hso_serial_close
to check for a valid serial structure which should prevent
the most reproducable crash in the driver when the hso device
is disconnected while in use.

Signed-off-by: Denis Joseph Barrow <D.Barow@option.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agohso: Add new usb device id's.
Denis Joseph Barrow [Tue, 25 Nov 2008 08:26:12 +0000 (00:26 -0800)]
hso: Add new usb device id's.

Signed-off-by: Denis Joseph Barrow <D.Barow@option.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetdev: add HAVE_NET_DEVICE_OPS
Stephen Hemminger [Tue, 25 Nov 2008 08:20:43 +0000 (00:20 -0800)]
netdev: add HAVE_NET_DEVICE_OPS

As a concession to vendors who have to deal with one source for different
kernel versions, add a HAVE_NET_DEVICE_OPS so they don't end up hard
coding ifdef against kernel version.

Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: handle shift/merge of cloned skbs too
Ilpo Järvinen [Tue, 25 Nov 2008 05:30:21 +0000 (21:30 -0800)]
tcp: handle shift/merge of cloned skbs too

This caused me to get repeatably:

  tcpdump: pcap_loop: recvfrom: Bad address

Happens occassionally when I tcpdump my for-looped test xfers:
  while [ : ]; do echo -n "$(date '+%s.%N') "; ./sendfile; sleep 20; done

Rest of the relevant commands:
  ethtool -K eth0 tso off
  tc qdisc add dev eth0 root netem drop 4%
  tcpdump -n -s0 -i eth0 -w sacklog.all

Running net-next under kvm, connection goes to the same host
(basically just out of kvm). The connection itself works ok
and data gets sent without corruption even with a large
number of tests while tcpdump fails usually within less than
5 tests.

Whether it only happens because of this change or not, I
don't know for sure but it's the only thing with which
I've seen that error. The non-cloned variant works w/o it
for much longer time. I'm yet to debug where the error
actually comes from.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: add some mibs to track collapsing
Ilpo Järvinen [Tue, 25 Nov 2008 05:27:22 +0000 (21:27 -0800)]
tcp: add some mibs to track collapsing

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: Make shifting not clear the hints
Ilpo Järvinen [Tue, 25 Nov 2008 05:26:56 +0000 (21:26 -0800)]
tcp: Make shifting not clear the hints

The earlier version was just very basic one which is "playing
safe" by always clearing the hints. However, clearing of a hint
is extremely costly operation with large windows, so it must be
avoided at all cost whenever possible, there is a way with
shifting too achieve not-clearing.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: Try to restore large SKBs while SACK processing
Ilpo Järvinen [Tue, 25 Nov 2008 05:20:15 +0000 (21:20 -0800)]
tcp: Try to restore large SKBs while SACK processing

During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.

This approach has a number of benefits:

1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
   which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
   of TCP has to be aware of this (this was not the case with
   some other approach that was, well, quite intrusive all
   around).
4) Clean_rtx_queue can release most of the pages using single
   put_page instead of previous PAGE_SIZE/mss+1 calls

In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.

TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).

I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
  1) ambiguous => no sensible measurement can be taken anyway
  2) non-ambiguous is due to reordering => having the timestamp
     of the last segment there is just skewing things more off
     than does some good since the ack got triggered by one of
     the holes (besides some substle issues that would make
     determining right hole/skb even harder problem). Anyway,
     it has nothing to do with this change then.

I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.

In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.

Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.

TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)

My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...

[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]

...To counter that, I gave up on the fifth advantage:

5) When growing previous SACK block, less allocs for new skbs
   are done, basically a new alloc is needed only when new hole
   is detected and when the previous skb runs out of frags space

...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.

With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:

                  TCPSackShifted 398
                   TCPSackMerged 877
            TCPSackShiftFallback 320
      TCPSACKCOLLAPSEFALLBACKGSO 0
  TCPSACKCOLLAPSEFALLBACKSKBBITS 0
  TCPSACKCOLLAPSEFALLBACKSKBDATA 0
    TCPSACKCOLLAPSEFALLBACKBELOW 0
    TCPSACKCOLLAPSEFALLBACKFIRST 1
 TCPSACKCOLLAPSEFALLBACKPREVBITS 318
      TCPSACKCOLLAPSEFALLBACKMSS 1
   TCPSACKCOLLAPSEFALLBACKNOHEAD 0
    TCPSACKCOLLAPSEFALLBACKSHIFT 0
          TCPSACKCOLLAPSENOOPSEQ 0
  TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
     TCPSACKCOLLAPSENOOPSMALLLEN 0
             TCPSACKCOLLAPSEHOLE 12

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: make tcp_sacktag_one able to handle partial skb too
Ilpo Järvinen [Tue, 25 Nov 2008 05:14:43 +0000 (21:14 -0800)]
tcp: make tcp_sacktag_one able to handle partial skb too

This is preparatory work for SACK combiner patch which may
have to count TCP state changes for only a part of the skb
because it will intentionally avoids splitting skb to SACKed
and not sacked parts.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: Make SACK code to split only at mss boundaries
Ilpo Järvinen [Tue, 25 Nov 2008 05:13:50 +0000 (21:13 -0800)]
tcp: Make SACK code to split only at mss boundaries

Sadly enough, this adds possible divide though we try to avoid
it by checking one mss as common case.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: more aggressive skipping
Ilpo Järvinen [Tue, 25 Nov 2008 05:12:28 +0000 (21:12 -0800)]
tcp: more aggressive skipping

I knew already when rewriting the sacktag that this condition
was too conservative, change it now since it prevent lot of
useless work (especially in the sack shifter decision code
that is being added by a later patch). This shouldn't change
anything really, just save some processing regardless of the
shifter.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: move tcp_simple_retransmit to tcp_input
Ilpo Järvinen [Tue, 25 Nov 2008 05:11:55 +0000 (21:11 -0800)]
tcp: move tcp_simple_retransmit to tcp_input

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotcp: collapse more than two on retransmission
Ilpo Järvinen [Tue, 25 Nov 2008 05:03:43 +0000 (21:03 -0800)]
tcp: collapse more than two on retransmission

I always had thought that collapsing up to two at a time was
intentional decision to avoid excessive processing if 1 byte
sized skbs are to be combined for a full mtu, and consecutive
retransmissions would make the size of the retransmittee
double each round anyway, but some recent discussion made me
to understand that was not the case. Thus make collapse work
more and wait less.

It would be possible to take advantage of the shifting
machinery (added in the later patch) in the case of paged
data but that can be implemented on top of this change.

tcp_skb_is_last check is now provided by the loop.

I tested a bit (ss-after-idle-off, fill 4096x4096B xfer,
10s sleep + 4096 x 1byte writes while dropping them for
some a while with netem):

16774097:16775545(1448) ack 1 win 46
16775545:16776993(1448) ack 1 win 46
. ack 16759617 win 2399
16776993:16777217(224) ack 1 win 46
. ack 16762513 win 2399
. ack 16765409 win 2399
. ack 16768305 win 2399
. ack 16771201 win 2399
. ack 16774097 win 2399
. ack 16776993 win 2399
. ack 16777217 win 2399
16777217:16777257(40) ack 1 win 46
. ack 16777257 win 2399
16777257:16778705(1448) ack 1 win 46
16778705:16780153(1448) ack 1 win 46
FP 16780153:16781313(1160) ack 1 win 46
. ack 16778705 win 2399
. ack 16780153 win 2399
F 1:1(0) ack 16781314 win 2399

While without drop-all period I get this:

16773585:16775033(1448) ack 1 win 46
. ack 16764897 win 9367
. ack 16767793 win 9367
. ack 16770689 win 9367
. ack 16773585 win 9367
16775033:16776481(1448) ack 1 win 46
16776481:16777217(736) ack 1 win 46
. ack 16776481 win 9367
. ack 16777217 win 9367
16777217:16777218(1) ack 1 win 46
16777218:16777219(1) ack 1 win 46
16777219:16777220(1) ack 1 win 46
  ...
16777247:16777248(1) ack 1 win 46
. ack 16777218 win 9367
. ack 16777219 win 9367
  ...
. ack 16777233 win 9367
. ack 16777248 win 9367
16777248:16778696(1448) ack 1 win 46
16778696:16780144(1448) ack 1 win 46
FP 16780144:16781313(1169) ack 1 win 46
. ack 16780144 win 9367
F 1:1(0) ack 16781314 win 9367

The window seems to be 30-40 segments, which were successfully
combined into: P 16777217:16777257(40) ack 1 win 46

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: avoid a pair of dst_hold()/dst_release() in ip_push_pending_frames()
Eric Dumazet [Tue, 25 Nov 2008 00:07:50 +0000 (16:07 -0800)]
net: avoid a pair of dst_hold()/dst_release() in ip_push_pending_frames()

We can reduce pressure on dst entry refcount that slowdown UDP transmit
path on SMP machines. This pressure is visible on RTP servers when
delivering content to mediagateways, especially big ones, handling
thousand of streams. Several cpus send UDP frames to the same
destination, hence use the same dst entry.

This patch makes ip_push_pending_frames() steal the refcount its
callers had to take when filling inet->cork.dst.

This doesnt avoid all refcounting, but still gives speedups on SMP,
on UDP/RAW transmit path.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: avoid a pair of dst_hold()/dst_release() in ip_append_data()
Eric Dumazet [Mon, 24 Nov 2008 23:52:46 +0000 (15:52 -0800)]
net: avoid a pair of dst_hold()/dst_release() in ip_append_data()

We can reduce pressure on dst entry refcount that slowdown UDP transmit
path on SMP machines. This pressure is visible on RTP servers when
delivering content to mediagateways, especially big ones, handling
thousand of streams. Several cpus send UDP frames to the same
destination, hence use the same dst entry.

This patch makes ip_append_data() eventually steal the refcount its
callers had to take on the dst entry.

This doesnt avoid all refcounting, but still gives speedups on SMP,
on UDP/RAW transmit path

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: gen_estimator: Fix gen_kill_estimator() lookups
Jarek Poplawski [Mon, 24 Nov 2008 23:48:05 +0000 (15:48 -0800)]
net: gen_estimator: Fix gen_kill_estimator() lookups

gen_kill_estimator() linear lists lookups are very slow, and e.g. while
deleting a large number of HTB classes soft lockups were reported. Here
is another try to fix this problem: this time internally, with rbtree,
so similarly to Jamal's hashing idea IIRC. (Looking for next hits could
be still optimized, but it's really fast as it is.)

Reported-by: Badalian Vyacheslav <slavon@bigtelecom.ru>
Reported-by: Denys Fedoryshchenko <denys@visp.net.lb>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Acked-by: Jamal Hadi Salim <hadi@cyberus.ca>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agopkt_sched: sch_drr: fix drr_dequeue loop()
Patrick McHardy [Mon, 24 Nov 2008 23:46:08 +0000 (15:46 -0800)]
pkt_sched: sch_drr: fix drr_dequeue loop()

Jarek Poplawski points out:

If all child qdiscs of sch_drr are non-work-conserving (e.g. sch_tbf)
drr_dequeue() will busy-loop waiting for skbs instead of leaving the
job for a watchdog. Checking for list_empty() in each loop isn't
necessary either, because this can never be true except the first time.

Using non-work-conserving qdiscs as children of DRR makes no sense,
simply bail out in that case.

Reported-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoinfiniband: Kill directly reference of netdev->priv
Wang Chen [Mon, 24 Nov 2008 23:34:00 +0000 (15:34 -0800)]
infiniband: Kill directly reference of netdev->priv

This use of netdev->priv is wrong.
The right way is:
alloc_netdev() with no memory for private data.
make netdev->ml_priv to point to c2_dev.

Signed-off-by: Wang Chen <wangchen@cn.fujitsu.com>
Acked-by: Roland Dreier <rolandd@cisco.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonetdevice sbni: Convert directly reference of netdev->priv
Wang Chen [Mon, 24 Nov 2008 22:52:16 +0000 (14:52 -0800)]
netdevice sbni: Convert directly reference of netdev->priv

1. convert netdev->priv to netdev_priv().
2. make sbni_pci_probe() be static.

Signed-off-by: Wang Chen <wangchen@cn.fujitsu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotokenring/3c359.c: Prevent possible mem leak when open failed
Jirka Pirko [Mon, 24 Nov 2008 22:49:11 +0000 (14:49 -0800)]
tokenring/3c359.c: Prevent possible mem leak when open failed

Freeing previously allocated buffers in case of error.

Signed-off-by: Jirka Pirko <jirka@pirko.cz>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotokenring/3c359.c: Fix error message when allocating tx_ring
Jirka Pirko [Mon, 24 Nov 2008 22:48:25 +0000 (14:48 -0800)]
tokenring/3c359.c: Fix error message when allocating tx_ring

Pointed out by Joe Perches. Error message after tx_ring allocation check was
wrong.

Signed-off-by: Jirka Pirko <jirka@jirka.pirko.cz>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agotokenring/3c359.c: fix allocation null check
Jirka Pirko [Mon, 24 Nov 2008 22:47:53 +0000 (14:47 -0800)]
tokenring/3c359.c: fix allocation null check

Fixed typo when allocating rx_ring, tx_ring was checked for null instead.

Signed-off-by: Jirka Pirko <jirka@pirko.cz>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years ago8139too: use err.h macros
Stephen Hemminger [Mon, 24 Nov 2008 22:47:01 +0000 (14:47 -0800)]
8139too: use err.h macros

Instead of using call by reference use the PTR_ERR macros to handle
return value with error case. Compile tested only.

Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: Make sure BHs are disabled in sock_prot_inuse_add()
Eric Dumazet [Mon, 24 Nov 2008 22:05:22 +0000 (14:05 -0800)]
net: Make sure BHs are disabled in sock_prot_inuse_add()

There is still a call to sock_prot_inuse_add() in af_netlink
while in a preemptable section. Add explicit BH disable around
this call.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: Make sure BHs are disabled in sock_prot_inuse_add()
Eric Dumazet [Mon, 24 Nov 2008 08:09:29 +0000 (00:09 -0800)]
net: Make sure BHs are disabled in sock_prot_inuse_add()

The rule of calling sock_prot_inuse_add() is that BHs must
be disabled.  Some new calls were added where this was not
true and this tiggers warnings as reported by Ilpo.

Fix this by adding explicit BH disabling around those call sites,
or moving sock_prot_inuse_add() call inside an existing BH disabled
section.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoeth: Declare an optimized compare_ether_addr_64bits() function
Eric Dumazet [Mon, 24 Nov 2008 07:24:32 +0000 (23:24 -0800)]
eth: Declare an optimized compare_ether_addr_64bits() function

Linus mentioned we could try to perform long word operations, even
on potentially unaligned addresses, on x86 at least. David mentioned
the HAVE_EFFICIENT_UNALIGNED_ACCESS test to handle this on all
arches that have efficient unailgned accesses.

I tried this idea and got nice assembly on 32 bits:

158:   33 82 38 01 00 00       xor    0x138(%edx),%eax
15e:   33 8a 34 01 00 00       xor    0x134(%edx),%ecx
164:   c1 e0 10                shl    $0x10,%eax
167:   09 c1                   or     %eax,%ecx
169:   74 0b                   je     176 <eth_type_trans+0x87>

And very nice assembly on 64 bits of course (one xor, one shl)

Nice oprofile improvement in eth_type_trans(), 0.17 % instead of 0.41 %,
expected since we remove 8 instructions on a fast path.

This patch implements a compare_ether_addr_64bits() function, that
uses the CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS ifdef to efficiently
perform the 6 bytes comparison on all capable arches.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agoaxnet_cs: Fix build after net device ops ne2k conversion.
David S. Miller [Mon, 24 Nov 2008 04:01:59 +0000 (20:01 -0800)]
axnet_cs: Fix build after net device ops ne2k conversion.

Commit 4e4fd4e485ad63a9074ff09a9b53ffc7a5c594ec ("ne2k: convert to
net_device_ops") exported some ei_* symbols from the 8390 library,
but the axnet_cs driver defines local static versions of the same
functions.

Rename them to avoid the namespace conflict.

Reported by Stephen Rothwell.

Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: Make sure BHs are disabled in sock_prot_inuse_add()
David S. Miller [Mon, 24 Nov 2008 01:34:03 +0000 (17:34 -0800)]
net: Make sure BHs are disabled in sock_prot_inuse_add()

The rule of calling sock_prot_inuse_add() is that BHs must
be disabled.  Some new calls were added where this was not
true and this tiggers warnings as reported by Ilpo.

Fix this by adding explicit BH disabling around those call sites.

Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: fix tunnels in netns after ndo_ changes
Alexey Dobriyan [Mon, 24 Nov 2008 01:26:26 +0000 (17:26 -0800)]
net: fix tunnels in netns after ndo_ changes

dev_net_set() should be the very first thing after alloc_netdev().

"ndo_" changes turned simple assignment (which is OK to do before netns
assignment) into quite non-trivial operation (which is not OK, init_net was
used). This leads to incomplete initialisation of tunnel device in netns.

BUG: unable to handle kernel NULL pointer dereference at 00000004
IP: [<c02efdb5>] ip6_tnl_exit_net+0x37/0x4f
*pde = 00000000
Oops: 0000 [#1] PREEMPT DEBUG_PAGEALLOC
last sysfs file: /sys/class/net/lo/operstate

Pid: 10, comm: netns Not tainted (2.6.28-rc6 #1)
EIP: 0060:[<c02efdb5>] EFLAGS: 00010246 CPU: 0
EIP is at ip6_tnl_exit_net+0x37/0x4f
EAX: 00000000 EBX: 00000020 ECX: 00000000 EDX: 00000003
ESI: c5caef30 EDI: c782bbe8 EBP: c7909f50 ESP: c7909f48
 DS: 007b ES: 007b FS: 0000 GS: 0000 SS: 0068
Process netns (pid: 10, ti=c7908000 task=c7905780 task.ti=c7908000)
Stack:
 c03e75e0 c7390bc8 c7909f60 c0245448 c7390bd8 c7390bf0 c7909fa8 c012577a
 00000000 00000002 00000000 c0125736 c782bbe8 c7909f90 c0308fe3 c782bc04
 c7390bd4 c0245406 c084b718 c04f0770 c03ad785 c782bbe8 c782bc04 c782bc0c
Call Trace:
 [<c0245448>] ? cleanup_net+0x42/0x82
 [<c012577a>] ? run_workqueue+0xd6/0x1ae
 [<c0125736>] ? run_workqueue+0x92/0x1ae
 [<c0308fe3>] ? schedule+0x275/0x285
 [<c0245406>] ? cleanup_net+0x0/0x82
 [<c0125ae1>] ? worker_thread+0x81/0x8d
 [<c0128344>] ? autoremove_wake_function+0x0/0x33
 [<c0125a60>] ? worker_thread+0x0/0x8d
 [<c012815c>] ? kthread+0x39/0x5e
 [<c0128123>] ? kthread+0x0/0x5e
 [<c0103b9f>] ? kernel_thread_helper+0x7/0x10
Code: db e8 05 ff ff ff 89 c6 e8 dc 04 f6 ff eb 08 8b 40 04 e8 38 89 f5 ff 8b 44 9e 04 85 c0 75 f0 43 83 fb 20 75 f2 8b 86 84 00 00 00 <8b> 40 04 e8 1c 89 f5 ff e8 98 04 f6 ff 89 f0 e8 f8 63 e6 ff 5b
EIP: [<c02efdb5>] ip6_tnl_exit_net+0x37/0x4f SS:ESP 0068:c7909f48
---[ end trace 6c2f2328fccd3e0c ]---

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agonet: Convert TCP/DCCP listening hash tables to use RCU
Eric Dumazet [Mon, 24 Nov 2008 01:22:55 +0000 (17:22 -0800)]
net: Convert TCP/DCCP listening hash tables to use RCU

This is the last step to be able to perform full RCU lookups
in __inet_lookup() : After established/timewait tables, we
add RCU lookups to listening hash table.

The only trick here is that a socket of a given type (TCP ipv4,
TCP ipv6, ...) can now flight between two different tables
(established and listening) during a RCU grace period, so we
must use different 'nulls' end-of-chain values for two tables.

We define a large value :

#define LISTENING_NULLS_BASE (1U << 29)

So that slots in listening table are guaranteed to have different
end-of-chain values than slots in established table. A reader can
still detect it finished its lookup in the right chain.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodccp: Header option insertion routine for feature-negotiation
Gerrit Renker [Mon, 24 Nov 2008 00:10:23 +0000 (16:10 -0800)]
dccp: Header option insertion routine for feature-negotiation

The patch extends existing code:
 * Confirm options divide into the confirmed value plus an optional preference
   list for SP values. Previously only the preference list was echoed for SP
   values, now the confirmed value is added as per RFC 4340, 6.1;
 * length and sanity checks are added to avoid illegal memory (or NULL) access.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodccp: Support for Mandatory options
Gerrit Renker [Mon, 24 Nov 2008 00:09:11 +0000 (16:09 -0800)]
dccp: Support for Mandatory options

Support for Mandatory options is provided by this patch, which will
be used by subsequent feature-negotiation patches.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodccp: Increase the scope of variable-length htonl/ntohl functions
Gerrit Renker [Mon, 24 Nov 2008 00:07:53 +0000 (16:07 -0800)]
dccp: Increase the scope of variable-length htonl/ntohl functions

This extends the scope of two available functions,
encode|decode_value_var, to work up to 6 (8) bytes, to match maximum
requirements in the RFC.

These functions are going to be used both by general option processing
and feature negotiation code, hence declarations have been put into
feat.h.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodccp: API to query the current TX/RX CCID
Gerrit Renker [Mon, 24 Nov 2008 00:04:59 +0000 (16:04 -0800)]
dccp: API to query the current TX/RX CCID

This provides function to query the current TX/RX CCID dynamically,
without reliance on the minisock value, using dynamic information
available in the currently loaded CCID module.

This query function is then used to
 (a) provide the getsockopt part for getting/setting CCIDs via sockopts;
 (b) replace the current test for "which CCID is in use" in probe.c.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: David S. Miller <davem@davemloft.net>
15 years agodccp: Set per-connection CCIDs via socket options
Gerrit Renker [Mon, 24 Nov 2008 00:02:31 +0000 (16:02 -0800)]
dccp: Set per-connection CCIDs via socket options

With this patch, TX/RX CCIDs can now be changed on a per-connection
basis, which overrides the defaults set by the global sysctl variables
for TX/RX CCIDs.

To make full use of this facility, the remaining patches of this patch
set are needed, which track dependencies and activate negotiated
feature values.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>