Eric Dumazet [Fri, 13 Dec 2013 21:51:23 +0000 (13:51 -0800)]
tcp: refine TSO splits
While investigating performance problems on small RPC workloads,
I noticed linux TCP stack was always splitting the last TSO skb
into two parts (skbs). One being a multiple of MSS, and a small one
with the Push flag. This split is done even if TCP_NODELAY is set,
or if no small packet is in flight.
Example with request/response of 4K/4K
IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 65245936525001>
IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 65245936525001>
IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 65245936525001>
IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 65250016524593>
IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 65250016524593>
IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 65250016524593>
IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 65245936525001>
IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 65245936525001>
IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 65245936525001>
We can avoid this split by including the Nagle tests at the right place.
Note : If some NIC had trouble sending TSO packets with a partial
last segment, we would have hit the problem in GRO/forwarding workload already.
tcp_minshall_update() is moved to tcp_output.c and is updated as we might
feed a TSO packet with a partial last segment.
This patch tremendously improves performance, as the traffic now looks
like :
IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 68342776834685>
IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 68342776834685>
IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 68346866834277>
IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 68346866834277>
IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 68342796834686>
IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 68342796834686>
IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 68346876834279>
IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 68346876834279>
IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 68342806834687>
IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 68342806834687>
IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 68346886834280>
IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 68346886834280>
Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher :
2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet
scheduler.
Many thanks to Neal for review and ideas.
Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@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>
These function to manipulate multiple addresses are not used anywhere
in current net-next tree. Some out of tree code maybe using these but
too bad; they should submit their code upstream..
Also, make __hw_addr_flush local since only used by dev_addr_lists.c
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Tue, 17 Dec 2013 20:08:17 +0000 (15:08 -0500)]
Merge branch 'for-davem' of git://git.kernel.org/pub/scm/linux/kernel/git/linville/wireless-next
John W. Linville says:
====================
Please pull this batch of updates for the 3.14 stream...
For the Bluetooth bits, Gustavo says:
"This is the first batch of patches intended for 3.14. There is
nothing big here. Most of the code are refactors, clean up, small
fixes, plus some new device id support."
And...
"More patches to 3.14. Here we have the support for Low Energy
Connection Oriented Channels (LE CoC). Basically, as the name says,
this adds supports for connection oriented channels in the same way
we already have them for BR/EDR connections so profiles/protocols
that work on top of BR/EDR can now work on LE plus a plenty of new
possibilities for LE."
For the ath10k bits, Kalle says:
"Janusz and Marek implemented DFS support to ath10k, but the code is
not enabled yet due to missing cfg80211/mac80211 patches (it will be
enabled in the next pull request). Michal did some device reset fixes
and made it possible for ath10k to share an interrupt with another
device. And lots of smaller fixes from different people."
For the iwlwifi bits, Emmanuel says:
"I have here a big rework of the rate control by Eyal. This is obviously
the biggest part of this batch.
I also have enhancement of protection flags by Avri and a few bits for
WoWLAN by Eliad and Luca. Johannes cleans up the debugfs plus a few
fixes. I provided a few things for Bluetooth coexistence.
Besides this we have an implementation for low priority scan."
Along with all that, there are big batches of updates to mwifiex and
ath9k, Jeff Kirsher's FSF address fix patches, and a handful of other
bits here and there.
Please let me know if there are problems!
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This is v2 of the ethernet PHY power optimization patches to reduce
power consumption of network PHYs with link that are either unused or
the corresponding netdev is down.
Compared to the last version, this patch set drops a patch to disable
unused PHYs after late initcall, as it is not compatible with a modular
mdio bus [1]. I'll investigate different ways to have a modular mdio bus
driver get notified when driver loading is done.
Again, a branch with v2 applied to v3.13-rc2 can also be found at
https://github.com/shesselba/linux-dove.git topic/ethphy-power-v2
When phydev is going to HALTED state, we can try to suspend it to
safe more power. phy_suspend helper will check if PHY can be suspended,
so just call it when entering HALTED state.
Signed-off-by: Sebastian Hesselbarth <sebastian.hesselbarth@gmail.com> Acked-by: Mugunthan V N <mugunthanvnm@ti.com> Reviewed-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
This ensures PHYs are resumed on attach and suspended on detach.
Signed-off-by: Sebastian Hesselbarth <sebastian.hesselbarth@gmail.com> Acked-by: Mugunthan V N <mugunthanvnm@ti.com> Reviewed-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
This adds helper functions to resume and suspend a given phy_device
by calling the corresponding driver callbacks if available.
Signed-off-by: Sebastian Hesselbarth <sebastian.hesselbarth@gmail.com> Acked-by: Mugunthan V N <mugunthanvnm@ti.com> Reviewed-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Marvell PHYs support generic PHY suspend/resume, so provide those
callbacks to all marvell specific drivers.
Signed-off-by: Sebastian Hesselbarth <sebastian.hesselbarth@gmail.com> Acked-by: Mugunthan V N <mugunthanvnm@ti.com> Reviewed-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
When using phydev, it should be phy_start/phy_stop'ed properly. This
driver doesn't do that, so add the corresponding calls to port_start/
stop respectively.
Signed-off-by: Sebastian Hesselbarth <sebastian.hesselbarth@gmail.com> Acked-by: Mugunthan V N <mugunthanvnm@ti.com> Reviewed-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
wangweidong [Fri, 13 Dec 2013 05:51:04 +0000 (13:51 +0800)]
sctp: Reorder 'struc association' members to reduce its size
Members of 'struct association' are not in appropriate order to
reuse compiler added padding on 64bit architectures. In this patch
we reorder those struct members and help reduce the size of the
structure from 2776 bytes to 2720 bytes on 64 bit architectures.
Signed-off-by: Wang Weidong <wangweidong1@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Tue, 17 Dec 2013 19:31:17 +0000 (14:31 -0500)]
Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/net-next
Jeff Kirsher says:
====================
Intel Wired LAN Driver Updates
This series contains updates to i40e only (again).
Jesse provides a fix for when tx_rings structure is NULL and we do not want
to panic. Then refactors the flow control set up and disables L2 flow control
by default. Provides some trivial fixes as well as prevent compiler warnings.
Then to align to similar behaviour in ixgbe, use the total number of CPUs in
the system to suggest the number of transmit and receive queue pairs.
Shannon provides a i40e ethtool fix to get some more reasonable information
reports back out to the ethtool. In addition, fixes PF reset after offline
test, where it reorders the test to put the register test last as it is the
only one that needs a reset, and we wait to trigger the reset until after we
clear the testing bit. Lastly provides basic support for handling suspend
and resume for now, later on Wake-On-LAN support will be added.
Anjali provides changes to tell the stack about our actual number of queues
in order for RFS/RPS/XFS to work correctly. Then provides several patches to
implement dynamically changing the queue count for the main VSI. Adds
basic support for get/set channels for RSS so that the number of receive and
transmit queue pair can be changed via ethtool. Cleans up the use of
rtnl_lock in the reset patch since it runs from a work time.
Neerav Parikh cleans up the VF interface to remove FCoE code as this
feature will not be supported on VF interfaces.
v2:
- submitted patch 1 to net (since it was a fix needed for net), so dropped
from this series (this patch will get added to net-next when Dave syncs
his trees)
- Dropped patches 4 & 11 from previous submission because of feedback
received from Ben Hutchings and Sergei Shtylyov.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Tue, 17 Dec 2013 19:27:44 +0000 (14:27 -0500)]
Merge branch 'ovs_hash'
Francesco Fusco says:
====================
ovs: introduce arch-specific fast hashing improvements
From: Daniel Borkmann <dborkman@redhat.com>
We are introducing a fast hash function (see patch1) that can be
used in the context of OpenVSwitch to reduce the hashing footprint
(patch2). For details, please see individual patches!
v1->v2:
- Make hash generic and place it under lib
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Francesco Fusco [Thu, 12 Dec 2013 15:09:06 +0000 (16:09 +0100)]
net: ovs: use CRC32 accelerated flow hash if available
Currently OVS uses jhash2() for calculating flow hashes in its
internal flow_hash() function. The performance of the flow_hash()
function is critical, as the input data can be hundreds of bytes
long.
OVS is largely deployed in x86_64 based datacenters. Therefore,
we argue that the performance critical fast path of OVS should
exploit underlying CPU features in order to reduce the per packet
processing costs. We replace jhash2 with the hash implementation
provided by the kernel hash lib, which exploits the crc32l
instruction to achieve high performance
Our patch greatly reduces the hash footprint from ~200 cycles of
jhash2() to around ~90 cycles in case of ovs_flow_hash_crc()
(measured with rdtsc over maximum length flow keys on an i7 Intel
CPU).
Additionally, we wrote a microbenchmark to stress the flow table
performance. The benchmark inserts random flows into the flow
hash and then performs lookups. Our hash deployed on a CRC32
capable CPU reduces the lookup for 1000 flows, 100 masks from
~10,100us to ~6,700us, for example.
Thus, simply use the newly introduced arch_fast_hash2() as a
drop-in replacement.
Signed-off-by: Francesco Fusco <ffusco@redhat.com> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: Thomas Graf <tgraf@redhat.com> Acked-by: Jesse Gross <jesse@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Francesco Fusco [Thu, 12 Dec 2013 15:09:05 +0000 (16:09 +0100)]
lib: introduce arch optimized hash library
We introduce a new hashing library that is meant to be used in
the contexts where speed is more important than uniformity of the
hashed values. The hash library leverages architecture specific
implementation to achieve high performance and fall backs to
jhash() for the generic case.
On Intel-based x86 architectures, the library can exploit the crc32l
instruction, part of the Intel SSE4.2 instruction set, if the
instruction is supported by the processor. This implementation
is twice as fast as the jhash() implementation on an i7 processor.
Additional architectures, such as Arm64 provide instructions for
accelerating the computation of CRC, so they could be added as well
in follow-up work.
Signed-off-by: Francesco Fusco <ffusco@redhat.com> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: Thomas Graf <tgraf@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: David S. Miller <davem@davemloft.net>
wangweidong [Thu, 12 Dec 2013 01:36:42 +0000 (09:36 +0800)]
tipc: change lock_sock order in connect()
Instead of reaquiring the socket lock and taking the normal exit
path when a connection times out, we bail out early with a
return -ETIMEDOUT.
Reviewed-by: Jon Maloy <jon.maloy@ericsson.com> Reviewed-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Wang Weidong <wangweidong1@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
wangweidong [Thu, 12 Dec 2013 01:36:41 +0000 (09:36 +0800)]
tipc: Use <linux/uaccess.h> instead of <asm/uaccess.h>
As warned by checkpatch.pl, use #include <linux/uaccess.h>
instead of <asm/uaccess.h>
Reviewed-by: Jon Maloy <jon.maloy@ericsson.com> Reviewed-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Wang Weidong <wangweidong1@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
wangweidong [Thu, 12 Dec 2013 01:36:40 +0000 (09:36 +0800)]
tipc: kill unnecessary goto's
Remove a number of needless 'goto exit' in send_stream
when the socket is in an unconnected state.
This patch is cosmetic and does not alter the operation of
TIPC in any way.
Reviewed-by: Jon Maloy <jon.maloy@ericsson.com> Reviewed-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Wang Weidong <wangweidong1@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
wangweidong [Thu, 12 Dec 2013 01:36:39 +0000 (09:36 +0800)]
tipc: remove unnecessary variables and conditions
We remove a number of unnecessary variables and branches
in TIPC. This patch is cosmetic and does not change the
operation of TIPC in any way.
Reviewed-by: Jon Maloy <jon.maloy@ericsson.com> Reviewed-by: Erik Hugne <erik.hugne@ericsson.com> Signed-off-by: Wang Weidong <wangweidong1@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
This patch is the first in a 3 series patchset to implement
dynamically changing the queue count for the main VSI.
This patch starts by adding a reinit flow. This flow is designed
to be able to change just the queue count and not the number of
interrupt vectors that the device originally came up with.
Jesse Brandeburg [Tue, 26 Nov 2013 10:49:17 +0000 (10:49 +0000)]
i40e: use same number of queues as CPUs
The current driver default sets the number of transmit/receive
queue pairs based on the current node's CPU count.
A better method is to use the total number of CPUs in the system
to suggest the number of queue pairs, which aligns better with
the behavior of ixgbe, and also with the expectations of the
kernel XPS and other subsystems in the stack.
Shannon Nelson [Tue, 26 Nov 2013 10:49:12 +0000 (10:49 +0000)]
i40e: fix pf reset after offline test
When the ethtool testing starts it sets the I40E_TESTING state
bit, which blocks new netdev opens so that things don't get
confused, while the testing might be messing with register and
other things. Unfortunately, that was keeping the PF resets
after the register test from working correctly because the netdev
would not get reopened. This patch reorders the tests to put the
register test last as it is the only one that needs a reset, and
we wait to trigger the reset until after we clear the
I40E_TESTING bit.
Change-Id: Ieaa18d74264250ac336b0656b490125ee8a22d2a Signed-off-by: Shannon Nelson <shannon.nelson@intel.com> Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com> Tested-by: Kavindya Deegala <kavindya.s.deegala@intel.com> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Jerry Chu [Mon, 16 Dec 2013 02:48:07 +0000 (18:48 -0800)]
net-ipv6: Fix alleged compiler warning in ipv6_exthdrs_len()
It was reported that Commit 299603e8370a93dd5d8e8d800f0dff1ce2c53d36
("net-gro: Prepare GRO stack for the upcoming tunneling support")
triggered a compiler warning in ipv6_exthdrs_len():
net/ipv6/ip6_offload.c: In function ‘ipv6_gro_complete’:
net/ipv6/ip6_offload.c:178:24: warning: ‘optlen’ may be used uninitialized in this function [-Wmaybe-u
opth = (void *)opth + optlen;
^
net/ipv6/ip6_offload.c:164:22: note: ‘optlen’ was declared here
int len = 0, proto, optlen;
^
Note that there was no real bug here - optlen was never uninitialized
before use. (Was the version of gcc I used smarter to not complain?)
Reported-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Signed-off-by: H.K. Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Sun, 15 Dec 2013 03:33:45 +0000 (22:33 -0500)]
Merge branch 'for-davem' of git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next
Ben Hutchings says:
====================
1. Change PTP clock name to 'sfc'.
2. Complete support for hardware timestamping and PTP clock on the
SFC9100 family.
3. Various cleanups for the PTP code.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 299603e8370a93dd5d8e8d800f0dff1ce2c53d36 ("net-gro: Prepare GRO
stack for the upcoming tunneling support") used an uninitialized variable
which leads to the following compiler warning:
net/ipv6/ip6_offload.c: In function ‘ipv6_gro_complete’:
net/ipv6/ip6_offload.c:178:24: warning: ‘optlen’ may be used uninitialized in this function [-Wmaybe-uninitialized]
opth = (void *)opth + optlen;
^
net/ipv6/ip6_offload.c:164:22: note: ‘optlen’ was declared here
int len = 0, proto, optlen;
^
Fix it up.
Cc: Jerry Chu <hkchu@google.com> Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Sat, 14 Dec 2013 06:58:19 +0000 (01:58 -0500)]
Merge branch 'bonding_rcu'
Ding Tianhong says:
====================
bonding: rebuild the lock use for bond monitor
Now the bond slave list is not protected by bond lock, only by RTNL,
but the monitor still use the bond lock to protect the slave list,
it is useless, according to the Veaceslav's opinion, there were
three way to fix the protect problem:
1. add bond_master_upper_dev_link() and bond_upper_dev_unlink()
in bond->lock, but it is unsafe to call call_netdevice_notifiers()
in write lock.
2. remove unused bond->lock for monitor function, only use the exist
rtnl lock(), it will take performance loss in fast path.
3. use RCU to protect the slave list, of course, performance is better,
but in slow path, it is ignored.
obviously the solution 1 is not fit here, I will consider the 2 and 3
solution. My principle is simple, if in fast path, RCU is better,
otherwise in slow path, both is well, but according to the Jay Vosburgh's
opinion, the monitor will loss performace if use RTNL to protect the all
slave list, so remove the bond lock and replace with RCU.
The second problem is the curr_slave_lock for bond, it is too old and
unwanted in many place, because the curr_active_slave would only be
changed in 3 place:
1. enslave slave.
2. release slave.
3. change active slave.
all above were already holding bond lock, RTNL and curr_slave_lock
together, it is tedious and no need to add so mach lock, when change
the curr_active_slave, you have to hold the RTNL and curr_slave_lock
together, and when you read the curr_active_slave, RTNL or curr_slave_lock,
any one of them is no problem.
for the stability, I did not change the logic for the monitor,
all change is clear and simple, I have test the patch set for lockdep,
it work well and stability.
v2. accept the Jay Vosburgh's opinion, remove the RTNL and replace with RCU,
also add some rcu function for bond use, so the patch set reach 10.
v3. accept the Nikolay Aleksandrov's opinion, remove no needed bond_has_slave_rcu(),
add protection for several 3ad mode handler functions and current_arp_slave.
rebuild the bond_first_slave_rcu(), make it more clear.
v4. because the struct netdev_adjacent should not be exist in netdevice.h, so I have
to make a new function to support micro bond_first_slave_rcu().
also add a new patch to simplify the bond_resend_igmp_join_requests_delayed().
v5. according the Jay Vosburgh's opinion, in patch 2 and 6, the calling of notify
peer is hardly to happen with the bond_xxx_commit() when the monitoring is running,
so the performance impact about make two round trips to one trip on RTNL is minimal,
no need to do that,the reason is very clear, so modify the patch 2 and 6, recover
the notify peer in RTNL alone.
====================
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:20:26 +0000 (10:20 +0800)]
bonding: rebuild the bond_resend_igmp_join_requests_delayed()
The bond_resend_igmp_join_requests_delayed() and
bond_resend_igmp_join_requests() should be integrated,
because the bond_resend_igmp_join_requests_delayed() did
nothing except bond_resend_igmp_join_requests().
The bond igmp_retrans could only be changed in bond_change_active_slave
and here, bond_change_active_slave will be called in RTNL and curr_slave_lock,
the bond_resend_igmp_join_requests already hold RTNL, so no need
to free RTNL and hold curr_slave_lock again, it may be a small optimization,
so move the igmp_retrans in RTNL and remove the curr_slave_lock.
Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:20:22 +0000 (10:20 +0800)]
bonding: remove unwanted lock for bond_store_primaryxxx()
The bond_select_active_slave() will not release and acquire
bond lock, so it is no need to read the bond lock for them,
and the bond_store_primaryxxx() is already in RTNL, so remove the
unwanted lock.
Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:20:17 +0000 (10:20 +0800)]
bonding: remove unwanted lock for bond_option_active_slave_set()
The bond_option_active_slave_set() is always called in RTNL,
the RTNL could protect bond slave list, so remove the unwanted
bond lock.
Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:20:12 +0000 (10:20 +0800)]
bonding: add RCU for bond_3ad_state_machine_handler()
The bond_3ad_state_machine_handler() use the bond lock to protect
the bond slave list and slave port together, but it is not enough,
the bond slave list was link and unlink in RTNL, not bond lock,
so I add RCU to protect the slave list from leaving.
The bond lock is still used here, because when the slave has been
removed from the list by the time the state machine runs, it appears
to be possible for both function to manupulate the same aggregator->lag_ports
by finding the aggregator via two different ports that are both members of
that aggregator (i.e., port A of the agg is being unbound, and port B
of the agg is runing its state machine).
If I remove the bond lock, there are nothing to mutex changes
to aggregator->lag_ports between bond_3ad_state_machine_handler and
bond_3ad_unbind_slave, So the bond lock is the simplest way to protect
aggregator->lag_ports.
There was a lot of function need RCU protect, I have two choice
to make the function in RCU-safe, (1) create new similar functions
and make the bond slave list in RCU. (2) modify the existed functions
and make them in read-side critical section, because the RCU
read-side critical sections may be nested.
I choose (2) because it is no need to create more similar functions.
The nots in the function is still too old, clean up the nots.
Suggested-by: Nikolay Aleksandrov <nikolay@redhat.com> Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:20:07 +0000 (10:20 +0800)]
bonding: remove unwanted lock for bond enslave and release
The bond_change_active_slave() and bond_select_active_slave()
do't need bond lock anymore, so remove the unwanted bond lock
for these two functions.
The bond_select_active_slave() will release and acquire
curr_slave_lock, so the curr_slave_lock need to protect
the function.
In bond enslave and bond release, the bond slave list is also
protected by RTNL, so bond lock is no need to exist, remove
the lock and clean the functions.
Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:20:02 +0000 (10:20 +0800)]
bonding: rebuild the lock use for bond_activebackup_arp_mon()
The bond_activebackup_arp_mon() use the bond lock for read to
protect the slave list, it is no effect, and the RTNL is only
called for bond_ab_arp_commit() and peer notify, for the performance
better, use RCU to replace with the bond lock, to the bond slave
list need to called in RCU, add a new bond_first_slave_rcu()
to get the first slave in RCU protection.
In bond_ab_arp_probe(), the bond->current_arp_slave may changd
if bond release slave, just like:
bond_ab_arp_probe() bond_release()
cpu 0 cpu 1
...
if (bond->current_arp_slave...) ...
... bond->current_arp_slave = NULl
bond->current_arp_slave->dev->name ...
So the current_arp_slave need to dereference in the section.
Suggested-by: Nikolay Aleksandrov <nikolay@redhat.com> Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:19:55 +0000 (10:19 +0800)]
bonding: create bond_first_slave_rcu()
The bond_first_slave_rcu() will be used to instead of bond_first_slave()
in rcu_read_lock().
According to the Jay Vosburgh's suggestion, the struct netdev_adjacent
should hide from users who wanted to use it directly. so I package a
new function to get the first slave of the bond.
Suggested-by: Nikolay Aleksandrov <nikolay@redhat.com> Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:19:50 +0000 (10:19 +0800)]
bonding: rebuild the lock use for bond_loadbalance_arp_mon()
The bond_loadbalance_arp_mon() use the bond lock to protect the
bond slave list, it is no effect, so I could use RTNL or RCU to
replace it, considering the performance impact, the RCU is more
better here, so the bond lock replace with the RCU.
The bond_select_active_slave() need RTNL and curr_slave_lock
together, but there is no RTNL lock here, so add a rtnl_rtylock.
Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:19:45 +0000 (10:19 +0800)]
bonding: rebuild the lock use for bond_alb_monitor()
The bond_alb_monitor use bond lock to protect the bond slave list,
it is no effect here, we need to use RTNL or RCU to replace bond lock,
the bond_alb_monitor will called 10 times one second, RTNL may loss
performance here, so I replace bond lock with RCU to protect the
bond slave list, also the RTNL is preserved, the logic of the monitor
did not changed.
Suggested-by: Nikolay Aleksandrov <nikolay@redhat.com> Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:19:39 +0000 (10:19 +0800)]
bonding: rebuild the lock use for bond_mii_monitor()
The bond_mii_monitor() still use bond lock to protect bond slave list,
it is no effect, I have 2 way to fix the problem, move the RTNL to the
top of the function, or add RCU to protect the bond slave list,
according to the Jay Vosburgh's opinion, 10 times one second is a
truely big performance loss if use RTNL to protect the whole monitor,
so I would take the advice and use RCU to protect the bond slave list.
The bond_has_slave() will not protect by anything, there will no things
happen if the slave list is be changed, unless the bond was free, but
it will not happened before the monitor, the bond will closed before
be freed.
The peers notify for the bond will calling curr_active_slave, so
derefence the slave to make sure we will accessing the same slave
if the curr_active_slave changed, as the rcu dereference need in
read-side critical sector and bond_change_active_slave() will call
it with no RCU hold, so add peer notify in rcu_read_lock which
will be nested in monitor.
Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
dingtianhong [Fri, 13 Dec 2013 02:19:32 +0000 (10:19 +0800)]
bonding: remove the no effect lock for bond_select_active_slave()
The bond slave list was no longer protected by bond lock and only
protected by RTNL or RCU, so anywhere that use bond lock to protect
slave list is meaningless.
remove the release and acquire bond lock for bond_select_active_slave().
The curr_active_slave could only be changed in 3 place:
all above place were holding bond lock, RTNL and curr_slave_lock
together, it is tedious and meaningless, obviously bond lock is no
need here, but RTNL or curr_slave_lock is needed, so if you want
to access active slave, you have to choose one lock, RTNL or
curr_slave_lock, if RTNL is exist, no need to add curr_slave_lock,
otherwise curr_slave_lock is better, because of the performance.
there are several place calling bond_select_active_slave() and
bond_change_active_slave(), the next step I will clean these place
and remove the no effect lock.
there are some document changed together when update the function.
Suggested-by: Jay Vosburgh <fubar@us.ibm.com> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet [Thu, 12 Dec 2013 23:41:56 +0000 (15:41 -0800)]
pkt_sched: set root qdisc before change() in attach_default_qdiscs()
After commit 95dc19299f74 ("pkt_sched: give visibility to mq slave
qdiscs") we call disc_list_add() while the device qdisc might be
the noop_qdisc one.
This shows up as duplicates in "tc qdisc show", as all inactive devices
point to noop_qdisc.
Fix this by setting dev->qdisc to the new qdisc before calling
ops->change() in attach_default_qdiscs()
Add a WARN_ON_ONCE() to catch any future similar problem.
Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Sat, 14 Dec 2013 06:11:22 +0000 (01:11 -0500)]
Merge branch 'for-davem' of git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next
Ben Hutchings says:
====================
An assortment of changes for Linux 3.14:
1. Merge the sfc fixes that you have already merged into net.git.
(The branch point for those was such that this does not bring in any
other changes.)
2. Reduce log level for a generally useless warning message, from
Robert Stonehouse.
3. Include BISTs in ethtool offline self-test for EF10 and recover from
BISTs initiated through other functions, from Jon Cooper.
4. Improve a sanity check on RX completions.
5. Avoid incrementing RX dropped count while the interface is down, from
Jon Cooper.
6. Improve hardware sensor naming and log messages, from Edward Cree.
7. Log all unexpected errors returned by firmware, from Edward Cree.
8. Expose another NVRAM partition to userland.
9. Some refactoring of the PTP code in preparation for EF10 support.
10. Various minor cleanups.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Sat, 14 Dec 2013 06:08:14 +0000 (01:08 -0500)]
Merge branch 'bonding_netlink'
Scott Feldman says:
====================
bonding: add more netlink attributes
v2:
Addressed v1 review comments. In particular, Jay's concern about
current sysfs ordering limitations carrying over to iproute. Netlink
attributes are processed in a priority order in
bond_netlink.c:bond_changelink(). Lower priority attributes can't undo
higher priority attributes when attempting to set both with iproute
command. For example, this command will fail:
ip link add bond1 type bond mode active-backup miimon 10 arp_interval 10
Because we're trying to create a new bond to use incompatible miimon
and ARP interval attributes. However, if attributes are applied
one-at-a-time, previously applied attributes can be overridden:
ip link add bond1 type bond mode active-backup miimon 10
ip link set dev bond1 type bond arp_interval 10
These two commands succeed. The bond is first created to use miimon.
Next, the bond is converted to use ARP interval, which undoes miimon.
v1:
Following Jiri Pirko's lead, add more bonding netlink attributes. Sending
matching iproute2 patch separately. sysfs access to attributes is
retained.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_ARP_ALL_TARGETS to allow get/set of bonding parameter
arp_all_targets via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_ARP_VALIDATE to allow get/set of bonding parameter
arp_validate via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_ARP_IP_TARGET to allow get/set of bonding parameter
arp_ip_target via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_ARP_INTERVAL to allow get/set of bonding parameter
arp_interval via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_USE_CARRIER to allow get/set of bonding parameter
use_carrier via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_DOWNDELAY to allow get/set of bonding parameter
downdelay via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_UPDELAY to allow get/set of bonding parameter
updelay via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Add IFLA_BOND_MIIMON to allow get/set of bonding parameter
miimon via netlink.
Signed-off-by: Scott Feldman <sfeldma@cumulusnetworks.com> Signed-off-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Add support to netconf to show changes to proxy-arp status on a per
interface basis via netlink in a manner similar to forwarding
and reverse path state.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Ben Hutchings [Thu, 5 Dec 2013 21:28:42 +0000 (21:28 +0000)]
sfc: Enable PTP clock and timestamping for all functions on EF10
The SFC9100 family has only one clock per controller, shared by all
functions. Therefore only create a clock device under the primary
function, and make all other functions refer to the primary's clock
device.
Since PTP functionality is limited to port 0 and PF 0 on the earlier
SFN[56]322F boards, and we also set the primary flag for that
function, we can make the creation of a clock device conditional only
on this flag.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Fri, 18 Oct 2013 18:21:45 +0000 (19:21 +0100)]
sfc: Associate primary and secondary functions of controller
The primary function of an EF10 controller will share its clock
device with other functions in the same domain (which we call
secondary functions). To this end, we need to associate functions
on the same controller.
We do not control probe order, so allow primary and secondary
functions to appear in any order. Maintain global lists of all
primary functions and of unassociated secondary functions,
and a list of secondary functions on each primary function.
Use the VPD serial number to tell whether functions are part of the
same controller. VPD will not be readable by virtual functions, so
this may need to be revisited later.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Jon Cooper [Mon, 18 Nov 2013 12:54:41 +0000 (12:54 +0000)]
sfc: Add RX packet timestamping for EF10
The EF10 firmware can optionally insert RX timestamps in the packet
prefix. These only include the clock minor value. We must also
enable periodic time sync events on each event queue which provide
the high bits of the clock value.
[bwh: Combined and rebased several changes.
Added the above description and some sanity checks for inline vs
separate timestamps.
Changed efx_rx_skb_attach_timestamp() to read the packet prefix
from the skb head area.] Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Thu, 28 Nov 2013 18:58:11 +0000 (18:58 +0000)]
sfc: Copy RX prefix into skb head area in efx_rx_mk_skb()
We can potentially pull the entire packet contents into the head area
and then free the page it was in. In order to read an inline
timestamp safely, we need to copy the prefix into the head area as
well.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Laurence Evans [Wed, 4 Dec 2013 23:47:56 +0000 (23:47 +0000)]
sfc: Add support for SFC9100 timestamp format
The clock minor tick on the SFC9100 family is 2^-27 s, not 1 ns.
There are also various pipeline delays which we need to correct for
when interpreting timestamps.
We query the firmware for the clock format and corrections at run-time.
[bwh: Combined and rebased several changes] Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Wed, 16 Oct 2013 17:32:39 +0000 (18:32 +0100)]
sfc: Do not use MAC address as clock name
We'll be sharing clocks between multiple functions with their own MAC
addresses. The name field is now documented as 'A short "friendly
name" to identify the clock ...' and '... not meant to be a unique
id.' So use the name 'sfc'.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Tue, 15 Oct 2013 16:54:56 +0000 (17:54 +0100)]
sfc: Remove dependency of PTP on having a dedicated channel
We need a dedicated channel on Siena to ensure we can match up
the separate RX and timestamp events for each PTP packet. We won't
do this for EF10 as timestamps are delivered inline.
Pass a channel index of 0 to MC_CMD_PTP_OP_ENABLE when there is no
dedicated channel.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Wed, 9 Oct 2013 13:17:27 +0000 (14:17 +0100)]
sfc: Return EBUSY for filter insertion on EF10, matching Falcon/Siena
The MC firmware will return error MC_CMD_ERR_ENOSPC if filter
insertion fails due to lack of resources. The net driver's filter
implementation for Falcon-architecture returns EBUSY. They should
behave consistently, so for EF10 change ENOSPC to EBUSY.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Tue, 8 Oct 2013 16:33:20 +0000 (17:33 +0100)]
sfc: Fold efx_flush_all() into efx_stop_port() and update comments
efx_flush_all() is a really misleading name - it has nothing to do
with e.g. flushing DMA queues. Since it's called immediately after
efx_stop_port() and is highly dependent on what that does, combine
the two functions.
Update comments to explain what this is doing a little better.
Also update an related and erroneous comment in efx_start_port().
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Edward Cree [Fri, 31 May 2013 17:36:12 +0000 (18:36 +0100)]
sfc: Log all unexpected MCDI errors
Split each of efx_mcdi_rpc, efx_mcdi_rpc_finish, and efx_mcdi_rpc_async into
a normal and a _quiet version; made the former log MCDI errors with
netif_err (and include the raw MCDI error code), and the latter never log
them at all. Changed various callers; any where some errors are expected
(but others are not) call the _quiet version and then if necessary log the
MCDI error themselves. Said logging is done by new efx_mcdi_display_error.
Callers of efx_mcdi_rpc*_quiet functions which may want to log the error
need to ensure that their outbuf is big enough to hold an MCDI error; to
this end, they now use MCDI_DECLARE_BUF_OUT_OR_ERR, which always allocates
at least 8 bytes.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Jon Cooper [Mon, 30 Sep 2013 16:36:50 +0000 (17:36 +0100)]
sfc: Correct RX dropped count for drops while interface is down
We don't directly control RX ingress on Siena or any later
controllers, and so we cannot prevent packets from entering the RX
datapath while the RX queues are not set up. This results in
the hardware incrementing RX_NODESC_DROP_CNT, but it's not an
error and we should not include it in error stats.
When bringing an interface up or down, pull (or wait for) stats and
count the number of packets that were dropped while the interface was
down. Subtract this from the reported RX dropped count.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Ben Hutchings [Tue, 24 Sep 2013 22:21:57 +0000 (23:21 +0100)]
sfc: Tighten the check for RX merged completion events
The addition of RX event merging support means we don't reliably
detect dropped RX events now. Currently we will only detect them if
the previous event for the RX queue had the CONT bit set.
Only accept RX completion events as merged if the
GET_CAPABILITIES_OUT_RX_BATCHING bit is set in datapath_caps (which it
won't be for the low-latency datapath) and the CONT bit is not set on
the event.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Jon Cooper [Mon, 16 Sep 2013 13:18:51 +0000 (14:18 +0100)]
sfc: Add MC BISTs to ethtool offline self test on EF10
To run BISTs the MC goes down in to a special mode where it will only
respond to MCDI from the testing PF, and TX, RX and event queues are
torn down. Other PFs get a message as it goes down to tell them it's
going down.
When the other PFs get this message, they check the soft status
register to tell when the MC has rebooted after BIST mode and they can
start recovery.
[bwh: Convert the test result to 1 or -1 as for earlier NICs] Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Olof Johansson [Thu, 12 Dec 2013 08:39:37 +0000 (00:39 -0800)]
net: eth: 8390: remove section warning in etherh.c
Commit c45f812f0280 ('8390 : Replace ei_debug with msg_enable/NETIF_MSG_*
feature') ended up moving the printout of version[] from something that
will be compiled out due to defines, to something that is now evaluated
at runtime.
That means that what always used to be an access to an __initdata string
from non-__init code started showing up as a section mismatch when it
didn't before.
All other 8390 versions skip __initdata on the version string, and
starting to annotate the whole chain of callers with __init seems like
more churn than it's worth on this driver, so remove it from etherh.c as well.
Fixes: c45f812f0280 ('8390 : Replace ei_debug with msg_enable/NETIF_MSG_* feature') Signed-off-by: Olof Johansson <olof@lixom.net> Signed-off-by: David S. Miller <davem@davemloft.net>
Jerry Chu [Thu, 12 Dec 2013 04:53:45 +0000 (20:53 -0800)]
net-gro: Prepare GRO stack for the upcoming tunneling support
This patch modifies the GRO stack to avoid the use of "network_header"
and associated macros like ip_hdr() and ipv6_hdr() in order to allow
an arbitary number of IP hdrs (v4 or v6) to be used in the
encapsulation chain. This lays the foundation for various IP
tunneling support (IP-in-IP, GRE, VXLAN, SIT,...) to be added later.
With this patch, the GRO stack traversing now is mostly based on
skb_gro_offset rather than special hdr offsets saved in skb (e.g.,
skb->network_header). As a result all but the top layer (i.e., the
the transport layer) must have hdrs of the same length in order for
a pkt to be considered for aggregation. Therefore when adding a new
encap layer (e.g., for tunneling), one must check and skip flows
(e.g., by setting NAPI_GRO_CB(p)->same_flow to 0) that have a
different hdr length.
Note that unlike the network header, the transport header can and
will continue to be set by the GRO code since there will be at
most one "transport layer" in the encap chain.
Signed-off-by: H.K. Jerry Chu <hkchu@google.com> Suggested-by: Eric Dumazet <edumazet@google.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Thu, 12 Dec 2013 18:38:46 +0000 (13:38 -0500)]
Merge branch 'macvtap_capture'
Vlad Yasevich says:
====================
Add packet capture support on macvtap device
Change from RFC:
- moved to the rx_handler approach.
This series adds support for packet capturing on macvtap device.
The initial approach was to simply export the capturing code as
a function from the core network. While simple, it was not
a very architecturally clean approach.
The new appraoch is to provide macvtap with its rx_handler which can
is attached to the macvtap device itself. Macvlan will simply requeue
the packet with an updated skb->dev. BTW, macvlan layer already does this
for macvlan devices. So, now macvtap and macvlan have almost the
same exact input path.
I've toyed with short-circuting the input path for macvtap by returning
RX_HANDLER_ANOTHER, but that just made the code more complicated and
didn't provide any kind of measurable gain (at least according to
netperf and perf runs on the host).
To see if there was a performance regression, I ran 1, 2 and 4 netperf
STREAM and MAERTS tests agains the VM from both remote host and another
guest on the same system. The command ran was
netperf -H $host -t $test -l 20 -i 10 -I 95 -c -C
The numbers I was getting with the new code were consistently very
slightly (1-2%) better then the old code. I don't consider this
an improvement, but it's not a regression! :)
Running 'perf record' on the host didn't show any new hot spots
and cpu utilization stayed about the same. This was better
then I expected from simply looking at the code.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Vlad Yasevich [Wed, 11 Dec 2013 18:27:11 +0000 (13:27 -0500)]
macvlan: Remove custom recieve and forward handlers
Since now macvlan and macvtap use the same receive and
forward handlers, we can remove them completely and use
netif_rx and dev_forward_skb() directly.
Signed-off-by: Vlad Yasevich <vyasevic@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Vlad Yasevich [Wed, 11 Dec 2013 18:27:10 +0000 (13:27 -0500)]
macvtap: Add support of packet capture on macvtap device.
Macvtap device currently doesn not allow a user to capture
traffic on due to the fact that it steals the packets
from the network stack before the skb->dev is set correctly
on the receive side, and that use uses macvlan transmit
path directly on the send side. As a result, we never
get a change to give traffic to the taps while the correct
device is set in the skb.
This patch makes macvtap device behave almost exaclty like
macvlan. On the send side, we switch to using dev_queue_xmit().
On the receive side, to deliver packets to macvtap, we now
use netif_rx and dev_forward_skb just like macvlan. The only
differnce now is that macvtap has its own rx_handler which is
attached to the macvtap netdev. It is here that we now steal
the packet and provide it to the socket.
As a result, we can now capture traffic on the macvtap device:
tcpdump -i macvtap0
It also gives us the abilit to add tc actions to the macvtap
device and actually utilize different bandwidth management
queues on output.
Signed-off-by: Vlad Yasevich <vyasevic@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Thu, 12 Dec 2013 01:28:41 +0000 (20:28 -0500)]
Merge branch 'bpf'
Daniel Borkmann says:
====================
bpf/filter updates
This set adds just two minimal helper tools that complement the
already available bpf_jit_disasm and complete BPF tooling; plus
it adds and an extensive documentation update of filter.txt.
Please see individual descriptions for details.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Borkmann [Wed, 11 Dec 2013 22:43:45 +0000 (23:43 +0100)]
filter: doc: improve BPF documentation
This patch significantly updates the BPF documentation and describes
its internal architecture, Linux extensions, and handling of the
kernel's BPF and JIT engine, plus documents how development can be
facilitated with the help of bpf_dbg, bpf_asm, bpf_jit_disasm.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Borkmann [Wed, 11 Dec 2013 22:43:44 +0000 (23:43 +0100)]
filter: bpf_asm: add minimal bpf asm tool
There are a couple of valid use cases for a minimal low-level bpf asm
like tool, for example, using/linking to libpcap is not an option, the
required BPF filters use Linux extensions that are not supported by
libpcap's compiler, a filter might be more complex and not cleanly
implementable with libpcap's compiler, particular filter codes should
be optimized differently than libpcap's internal BPF compiler does,
or for security audits of emitted BPF JIT code for prepared set of BPF
instructions resp. BPF JIT compiler development in general.
Then, in such cases writing such a filter in low-level syntax can be
an good alternative, for example, xt_bpf and cls_bpf users might have
requirements that could result in more complex filter code, or one that
cannot be expressed with libpcap (e.g. different return codes in
cls_bpf for flowids on various BPF code paths).
Moreover, BPF JIT implementors may wish to manually write test cases
in order to verify the resulting JIT image, and thus need low-level
access to BPF code generation as well. Therefore, complete the available
toolchain for BPF with this small bpf_asm helper tool for the tools/net/
directory. These 3 complementary minimal helper tools round up and
facilitate BPF development.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Borkmann [Wed, 11 Dec 2013 22:43:43 +0000 (23:43 +0100)]
filter: bpf_dbg: add minimal bpf debugger
This patch adds a minimal BPF debugger that "emulates" the kernel's
BPF engine (w/o extensions) and allows for single stepping (forwards
and backwards through BPF code) or running with >=1 breakpoints through
selected or all packets from a pcap file with a provided user filter
in order to facilitate verification of a BPF program. When a breakpoint
is being hit, it dumps all register contents, decoded instructions and
in case of branches both decoded branch targets as well as other useful
information.
Having this facility is in particular useful to verify BPF programs
against given test traffic *before* attaching to a live system.
With the general availability of cls_bpf, xt_bpf, socket filters,
team driver and e.g. PTP code, all BPF users, quite often a single
more complex BPF program is being used. Reasons for a more complex
BPF program are primarily to optimize execution time for making a
verdict when multiple simple BPF programs are combined into one in
order to prevent parsing same headers multiple times. In particular,
for cls_bpf that can have various return paths for encoding flowids,
and xt_bpf to come to a fw verdict this can be the case.
Therefore, as this can result in more complex and harder to debug
code, it would be very useful to have this minimal tool for testing
purposes. It can also be of help for BPF JIT developers as filters
are "test attached" to the kernel on a temporary socket thus
triggering a JIT image dump when enabled. The tool uses an interactive
libreadline shell with auto-completion and history support.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
projects / ~andy / linux / log