When calling the flow_free() to free the flow, we call many times
(cpu_possible_mask, eg. 128 as default) cpumask_next(). That will
take up our CPU usage if we call the flow_free() frequently.
When we put all packets to userspace via upcall, and OvS will send
them back via netlink to ovs_packet_cmd_execute(will call flow_free).
The test topo is shown as below. VM01 sends TCP packets to VM02,
and OvS forward packtets. When testing, we use perf to report the
system performance.
VM01 --- OvS-VM --- VM02
Without this patch, perf-top show as below: The flow_free() is
3.02% CPU usage.
4.23% [kernel] [k] _raw_spin_unlock_irqrestore
3.62% [kernel] [k] __do_softirq
3.16% [kernel] [k] __memcpy
3.02% [kernel] [k] flow_free
2.42% libc-2.17.so [.] __memcpy_ssse3_back
2.18% [kernel] [k] copy_user_generic_unrolled
2.17% [kernel] [k] find_next_bit
When applied this patch, perf-top show as below: Not shown on
the list anymore.
4.11% [kernel] [k] _raw_spin_unlock_irqrestore
3.79% [kernel] [k] __do_softirq
3.46% [kernel] [k] __memcpy
2.73% libc-2.17.so [.] __memcpy_ssse3_back
2.25% [kernel] [k] copy_user_generic_unrolled
1.89% libc-2.17.so [.] _int_malloc
1.53% ovs-vswitchd [.] xlate_actions
With this patch, the TCP throughput(we dont use Megaflow Cache
+ Microflow Cache) between VMs is 1.18Gbs/sec up to 1.30Gbs/sec
(maybe ~10% performance imporve).
This patch adds cpumask struct, the cpu_used_mask stores the cpu_id
that the flow used. And we only check the flow_stats on the cpu we
used, and it is unncessary to check all possible cpu when getting,
cleaning, and updating the flow_stats. Adding the cpu_used_mask to
sw_flow struct does’t increase the cacheline number.
Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the ovs_flow_stats_update(), we only use the node
var to alloc flow_stats struct. But this is not a
common case, it is unnecessary to call the numa_node_id()
everytime. This patch is not a bugfix, but there maybe
a small increase.
Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rick Farrington says:
====================
liquidio: avoid vm low memory crashes
This patchset addresses issues brought about by low memory conditions
in a VM. These conditions were not seen when the driver was exercised
normally. Rather, they were brought about through manual fault injection.
They are being included in the interest of hardening the driver against
unforeseen circumstances.
1. Fix GPF in octeon_init_droq(); zero the allocated block 'recv_buf_list'.
This prevents a GPF trying to access an invalid 'recv_buf_list[i]' entry
in octeon_droq_destroy_ring_buffers() if init didn't alloc all entries.
2. Don't dereference a NULL ptr in octeon_droq_destroy_ring_buffers().
3. For defensive programming, zero the allocated block 'oct->droq' in
octeon_setup_output_queues() and 'oct->instr_queue' in
octeon_setup_instr_queues().
change log:
V1 -> V2:
1. Corrected syntax in 'Subject' lines; no functional or code changes.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
For defensive programming, zero the allocated block 'oct->droq[0]' in
octeon_setup_output_queues() and 'oct->instr_queue[0]' in
octeon_setup_instr_queues().
Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Satanand Burla <satananda.burla@cavium.com>
Signed-off-by: Raghu Vatsavayi <raghu.vatsavayi@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Don't dereference a NULL ptr in octeon_droq_destroy_ring_buffers().
Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Satanand Burla <satananda.burla@cavium.com>
Signed-off-by: Raghu Vatsavayi <raghu.vatsavayi@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix GPF in octeon_init_droq(); zero the allocated block 'recv_buf_list'.
This prevents a GPF trying to access an invalid 'recv_buf_list[i]' entry
in octeon_droq_destroy_ring_buffers() if init didn't alloc all entries.
Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Satanand Burla <satananda.burla@cavium.com>
Signed-off-by: Raghu Vatsavayi <raghu.vatsavayi@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Added support for new firmware statistic 'tx_err_pki'.
Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Derek Chickles <derek.chickles@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Arvind Yadav says:
====================
constify net attribute_group structures.
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/sysfs.h> work with const
attribute_group. So mark the non-const structs as const.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/sysfs.h> work
with const attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
28720 985 12 29717 7415 net/.../cxgb3/cxgb3_main.o
File size After adding 'const':
text data bss dec hex filename
28848 857 12 29717 7415 net/.../cxgb3/cxgb3_main.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/netdevice.h> work
with const attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
4512 1472 0 5984 1760 drivers/net/bonding/bond_sysfs.o
File size After adding 'const':
text data bss dec hex filename
4576 1408 0 5984 1760 drivers/net/bonding/bond_sysfs.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/netdevice.h> work
with const attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
3409 948 28 4385 1121 drivers/net/arcnet/com20020-pci.o
File size After adding 'const':
text data bss dec hex filename
3473 884 28 4385 1121 drivers/net/arcnet/com20020-pci.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/sysfs.h> work
with const attribute_group. So mark the non-const structs as const.
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/sysfs.h> work
with const attribute_group. So mark the non-const structs as const.
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/sysfs.h> work
with const attribute_group. So mark the non-const structs as const.
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/sysfs.h> work
with const attribute_group. So mark the non-const structs as const.
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/netdevice.h> work
with const attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
11800 368 0 12168 2f88 drivers/net/can/janz-ican3.o
File size After adding 'const':
text data bss dec hex filename
11864 304 0 12168 2f88 drivers/net/can/janz-ican3.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/netdevice.h> work
with const attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
6164 304 0 6468 1944 drivers/net/can/at91_can.o
File size After adding 'const':
text data bss dec hex filename
6228 240 0 6468 1944 drivers/net/can/at91_can.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
attribute_group are not supposed to change at runtime. All functions
working with attribute_group provided by <linux/netdevice.h> work
with const attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
13275 928 1 14204 377c drivers/net/usb/cdc_ncm.o
File size After adding 'const':
text data bss dec hex filename
13339 864 1 14204 377c drivers/net/usb/cdc_ncm.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jiri Pirko says:
====================
mlxsw: Preparations for IPv6 UC router
Ido says:
The purpose of this set is to prepare the driver for the introduction of
IPv6 FIB offload. It's mainly composed of small and non-functional
changes, that either add the IPv6 equivalent of existing IPv4 code or
aimed at making the introduction of IPv6-specific code easier.
The first five patches enable IPv6 forwarding in the device and allow us
to configure router interfaces (RIFs) based on inet6addr notifications.
The next six patches add support for programming IPv6 neighbours into
the device's table as well as dumping their activity and updating the
kernel accordingly.
The last 11 patches extend current infrastructure to allow us to program
IPv6 routes, set catch-all IPv6 trap in case of abort and make the code
more receptive towards up-coming changes.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The number of possible prefix lengths for IPv6 is 129 and not 128.
Fixes following warning from UBSAN when /128 routes are offloaded:
UBSAN: Undefined behaviour in
drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c:2510:27 index 128 is out
of range for type 'long unsigned int [128]'
Fixes: 5e9c16cc83 ("mlxsw: spectrum_router: Implement private fib")
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These functions aren't specific to IPv4 and can be re-used for IPv6.
Drop the '4' designation from their name.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Functions that take as argument a FIB entry don't need to take FIB node
as well, as it can be extracted from the entry.
Remove unnecessary FIB node parameter.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The functions to create and destroy a nexthop group are IPv4 specific
and should be renamed accordingly, so that they won't be confused with
the IPv6 specific functions in follow-up patches.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some of the parameters stored in the FIB entry structure are specific to
IPv4 and therefore better placed in an IPv4 specific structure.
Create an IPv4 specific structure that encapsulates the common FIB entry
structure and contains IPv4 specific parameters.
In a follow-up patchset an IPv6 specific structure will be introduced.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we fail to insert a route we invoke the abort mechanism which
flushes all the tables and inserts a default route in each, so that all
packets incoming to the router will be trapped to the CPU.
Upon abort, add an IPv6 default route to the IPv6 tables.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Take advantage of previous patch and allow the RALUE register to be
called with IPv6 routes.
In order to re-use as much code as possible between IPv4 and IPv6, only
the lowest-level function that actually does the register packing is
demuxed based on the passed protocol.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update the register so that IPv6 LPM entries could be programmed to the
device's table.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A Virtual Router (VR) is an entity which corresponds to a VRF and
performs FIB lookup in an LPM tree according to the {VR, IP Proto} ->
Tree binding.
Extend the virtual router data structure towards IPv6 FIB offload.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A FIB node is an entity which stores routes sharing the same prefix and
length. The data structure itself is already family agnostic, but we
make some of its operations agnostic as well and thus re-use them for
IPv6 offload.
Instead of passing an IPv4-specific structure to fib4_node_get(), pass
general routing parameters and rename the function accordingly.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When looking up a FIB entry we shouldn't create the FIB node where it's
supposed to be linked in case the node doesn't already exist.
Instead, lookup the node and fail if it doesn't exist.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Thankfully, the neighbour subsystem is agnostic to the upper protocol
and used by both IPv4 and IPv6. By removing assumptions regarding the
neighbour type we can thus re-use much of the neighbour-related code for
both IPv4 and IPv6.
For each nexthop, store its gateway IP and for nexthop group store the
neighbour table used by its nexthops.
Use this information throughout the code and remove assumption about the
neighbour type.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The neighbours' activity is currently dumped according to the ARP
table's DELAY_PROBE time, but with the introduction of IPv6 offload we
should set the interval according to the minimum between the ARP and
ndisc tables.
Signed-off-by: Arkadi Sharshvesky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In addition to IPv4, periodically dump IPv6 neighbours and update the
kernel about them.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update the register so that the active IPv6 neighbours could be dumped
from the device's neighbour table.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As with IPv4, listen to NEIGH_UPDATE events from the ndisc table and
program relevant neighbours to the device's neighbour table.
Note that neighbours with a link-local IP address aren't programmed, as
packets with a link-local destination IP are trapped after LPM lookup
and never reach the neighbour table.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update the register, so the IPv6 neighbours could be programmed to the
device's neighbour table.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a netdev is configured with an IP address a router interface (RIF)
should be configured for it in the device. Allow configuration of RIFs
based on IPv6 address notifications as well as IPv4.
Note that the RIF exists as long as an IP address is configured on the
netdev, regardless of the address family.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Up until now we only flooded broadcast packets to the router when an L3
interface was configured on top of a bridge. However, IPv6 Neighbour
Discovery packets are trapped to the CPU inside the router and these can
be sent with a multicast address.
Flood unregistered multicast packets to the router port, so that
relevant packets could be trapped there.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Before we can start using IPv6, we need to trap certain control packets
to the CPU. Among others, these include Neighbour Discovery, DHCP and
neighbour misses.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Enable IPv6 and IPv6 forwarding on router interfaces (RIFs), so that
they will be able to receive and forward IPv6 traffic.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Before we add IPv6 constructs like traps and router interfaces, we first
need to enable IPv6 routing in the device.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Florian Westphal says:
====================
xfrm: remove flow cache
After RCU-ification of ipsec packet path there are no major scalability
issues anymore without flow cache.
We still incur a performance hit, which comes mostly from the extra xfrm
dst allocation/freeing.
The last patch in the series adds a simple percpu cache to avoid the
extra allocation if a packet matched the same policies as last one.
The main concern with this is that we will see performance drops,
especially with large numbers of policies/SAs.
However, during hallway discussions at nfws 2017 it seemed the issues
with flow caching outweight the removal downsides, and that it
might be best to just 'remove it' and see where the practical issues
(if any) will appear.
It should now be possible to also remove the genid member in the policies
as we don't hold bundles for prolonged time anymore, but I think
this change is controversial (and intrusive) enough as-is, so defer
that to a later point in time.
Changes since last rfc:
- fix build failures due to implicit interrupt.h includes
- rework last patch (pcpu cache):
* avoid xchg()
* check policies for walk.dead = 1 instead of more costly bundle_ok().
* flush pcpu bundles when sa/policies get removed, to allow module
references to go away (suggested by Ilan Tayari)
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
retain last used xfrm_dst in a pcpu cache.
On next request, reuse this dst if the policies are the same.
The cache will not help with strict RR workloads as there is no hit.
The cache packet-path part is reasonably small, the notifier part is
needed so we do not add long hangs when a device is dismantled but some
pcpu xdst still holds a reference, there are also calls to the flush
operation when userspace deletes SAs so modules can be removed
(there is no hit.
We need to run the dst_release on the correct cpu to avoid races with
packet path. This is done by adding a work_struct for each cpu and then
doing the actual test/release on each affected cpu via schedule_work_on().
Test results using 4 network namespaces and null encryption:
ns1 ns2 -> ns3 -> ns4
netperf -> xfrm/null enc -> xfrm/null dec -> netserver
what TCP_STREAM UDP_STREAM UDP_RR
Flow cache: 14644.61 294.35 327231.64
No flow cache: 14349.81 242.64 202301.72
Pcpu cache: 14629.70 292.21 205595.22
UDP tests used 64byte packets, tests ran for one minute each,
value is average over ten iterations.
'Flow cache' is 'net-next', 'No flow cache' is net-next plus this
series but without this patch.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
After rcu conversions performance degradation in forward tests isn't that
noticeable anymore.
See next patch for some numbers.
A followup patcg could then also remove genid from the policies
as we do not cache bundles anymore.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows to remove flow cache object embedded in struct xfrm_dst.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes the wrapper and renames the __xfrm_policy_lookup variant
to get rid of another place that used flow cache objects.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
XFRM_POLICY_IN/OUT/FWD are identical to FLOW_DIR_*, so gcc already
removed this function as its just returns the argument. Again, no
code change.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
after previous change oldflo and xdst are always NULL.
These branches were already removed by gcc, this doesn't change code.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of consulting flow cache, call the xfrm bundle/policy lookup
functions directly. This pretends the flow cache had no entry.
This helps to gradually remove flow cache integration,
followup commit will remove the dead code that this change adds.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
revert c386578f1c ("xfrm: Let the flowcache handle its size by default.").
Once we remove flow cache, we don't have a flow cache limit anymore.
We must not allow (virtually) unlimited allocations of xfrm dst entries.
Revert back to the old xfrm dst gc limits.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Tonghao Zhang