This partially reverts commit bde94613e6276d48a6e0be7a592ebcf9836b4aaf.
Commit bde94613e627 was aimed to slightly ( < 1%) increase performance
in the case where EMC disabled, but it avoids RSS hash calculation and
OVS has to calculate it while executing OVS_ACTION_ATTR_HASH in order
to handle balanced-tcp bonding. At the time of executing that action
there is no parsed flow, and OVS parses the packet for the second time
to calculate the hash. This happens for all packets received from the
virtual interfaces because they have no HW RSS.
Here is the example of 'perf' output for VM --> (bonded PHY) traffic:
Samples: 401K of event 'cycles', Event count (approx.): 50964771478
Overhead Shared Object Symbol
27.50% ovs-vswitchd [.] dpcls_lookup.370382
16.30% ovs-vswitchd [.] rte_vhost_dequeue_burst.9267
14.95% ovs-vswitchd [.] miniflow_extract
7.22% ovs-vswitchd [.] flow_extract
7.10% ovs-vswitchd [.] dp_netdev_input__.371002.4826
4.01% ovs-vswitchd [.] fast_path_processing.370987.4893
We can see that packet parsed twice. First time by 'miniflow_extract'
right after receiving and the second time by 'flow_extract' while
executing actions.
In this particular case calculating RSS on receive saves > 7% of the
total CPU processing time. It varies from ~7 to ~10 % depending on
scenario/traffic types.
It's better to calculate hash each time because performance
improvements of avoiding are negligible in compare with performance
drop in case of sending packets to bonded interface.
Another solution could be to pass the parsed flow explicitly through
the datapath, but this will require big code changes and will have
additional overhead for metadata updating on packet changes.
Also, this change should have small impact since SMC works well in most
cases and will be enabled/recommended by default in the future.
CC: Antonio Fischetti <antonio.fischetti@intel.com>
Fixes: bde94613e627 ("dpif-netdev: Avoid reading RSS hash when EMC is disabled.")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
It's a normal case for 'find' function, especially because this
happens for every first packet of flow that was not offloaded yet.
Should not warn about this. Dropped to DBG to avoid log trashing in
case of big number of new flows.
CC: Yuanhan Liu <yliu@fridaylinux.org>
Fixes: 241bad15d99a ("dpif-netdev: associate flow with a mark id")
Acked-by: Flavio Leitner <fbl@sysclose.org>
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
As a side effect, this also reduces a lot of log messages' severities from
ERR to WARN. They just didn't seem like messages that in general reported
anything that would prevent functioning.
Signed-off-by: Ben Pfaff <blp@ovn.org>
Fragmentation handling is added for supporting conntrack.
Both v4 and v6 are supported.
After discussion with several people, I decided to not store
configuration state in the database to be more consistent with
the kernel in future, similarity with other conntrack configuration
which will not be in the database as well and overall simplicity.
Accordingly, fragmentation handling is enabled by default.
This patch enables fragmentation tests for the userspace datapath.
Signed-off-by: Darrell Ball <dlu998@gmail.com>
Signed-off-by: Ben Pfaff <blp@ovn.org>
This is needed in a subsequent patch and may otherwise be useful.
Signed-off-by: Darrell Ball <dlu998@gmail.com>
Signed-off-by: Ben Pfaff <blp@ovn.org>
'sorted_poll_list()' uses the 'pmd->poll_list' that should be
guarded by 'pmd->port_mutex'.
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Conditional EMC insert helps a lot in scenarios with high numbers
of parallel flows, but in current implementation this option affects
all the threads and ports at once. There are scenarios where we have
different number of flows on different ports. For example, if one
of the VMs encapsulates traffic using additional headers, it will
receive large number of flows but only few flows will come out of
this VM. In this scenario it's much faster to use EMC instead of
classifier for traffic from the VM, but it's better to disable EMC
for the traffic which flows to VM.
To handle above issue introduced 'emc-enable' configurable to
enable/disable EMC on a per-port basis. Ex.:
ovs-vsctl set interface dpdk0 other_config:emc-enable=false
EMC probability kept as is and it works for all the ports with
'emc-enable=true'.
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Port rx queues that have not been statically assigned to PMDs are currently
assigned based on periodically sampled load measurements.
The assignment is performed at specific instances – port addition, port
deletion, upon reassignment request via CLI etc.
Due to change in traffic pattern over time it can cause uneven load among
the PMDs and thus resulting in lower overall throughout.
This patch enables the support of auto load balancing of PMDs based on
measured load of RX queues. Each PMD measures the processing load for each
of its associated queues every 10 seconds. If the aggregated PMD load reaches
95% for 6 consecutive intervals then PMD considers itself to be overloaded.
If any PMD is overloaded, a dry-run of the PMD assignment algorithm is
performed by OVS main thread. The dry-run does NOT change the existing
queue to PMD assignments.
If the resultant mapping of dry-run indicates an improved distribution
of the load then the actual reassignment will be performed.
The automatic rebalancing will be disabled by default and has to be
enabled via configuration option. The interval (in minutes) between
two consecutive rebalancing can also be configured via CLI, default
is 1 min.
Following example commands can be used to set the auto-lb params:
ovs-vsctl set open_vswitch . other_config:pmd-auto-lb="true"
ovs-vsctl set open_vswitch . other_config:pmd-auto-lb-rebalance-intvl="5"
Co-authored-by: Rohith Basavaraja <rohith.basavaraja@gmail.com>
Co-authored-by: Venkatesan Pradeep <venkatesan.pradeep@ericsson.com>
Signed-off-by: Rohith Basavaraja <rohith.basavaraja@gmail.com>
Signed-off-by: Venkatesan Pradeep <venkatesan.pradeep@ericsson.com>
Signed-off-by: Nitin Katiyar <nitin.katiyar@ericsson.com>
Acked-by: Kevin Traynor <ktraynor@redhat.com>
Tested-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Flow offloading thread uses concurrent hash maps which are
based on rcu protected variables. It must use them while in
active state. Working in a quiescent state could cause
segmentation faults because of possible cmap internal
structure changes.
Fixes: 02bb2824e51d ("dpif-netdev: do hw flow offload in a thread")
Acked-by: Flavio Leitner <fbl@sysclose.org>
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Data pointed by cmap node must not be freed while iterating.
ovsrcu_postpone should be used instead.
CC: Finn Christensen <fc@napatech.com>
Fixes: e8a2b5bf92bb ("netdev-dpdk: implement flow offload with rte flow")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This is the third patch in the patch-set to support dynamic rebalancing
of offloaded flows.
The dynamic rebalancing functionality is implemented in this patch. The
ukeys that are not scheduled for deletion are obtained and passed as input
to the rebalancing routine. The rebalancing is done in the context of
revalidation leader thread, after all other revalidator threads are
done with gathering rebalancing data for flows.
For each netdev that is in OOR state, a list of flows - both offloaded
and non-offloaded (pending) - is obtained using the ukeys. For each netdev
that is in OOR state, the flows are grouped and sorted into offloaded and
pending flows. The offloaded flows are sorted in descending order of
pps-rate, while pending flows are sorted in ascending order of pps-rate.
The rebalancing is done in two phases. In the first phase, we try to
offload all pending flows and if that succeeds, the OOR state on the device
is cleared. If some (or none) of the pending flows could not be offloaded,
then we start replacing an offloaded flow that has a lower pps-rate than
a pending flow, until there are no more pending flows with a higher rate
than an offloaded flow. The flows that are replaced from the device are
added into kernel datapath.
A new OVS configuration parameter "offload-rebalance", is added to ovsdb.
The default value of this is "false". To enable this feature, set the
value of this parameter to "true", which provides packets-per-second
rate based policy to dynamically offload and un-offload flows.
Note: This option can be enabled only when 'hw-offload' policy is enabled.
It also requires 'tc-policy' to be set to 'skip_sw'; otherwise, flow
offload errors (specifically ENOSPC error this feature depends on) reported
by an offloaded device are supressed by TC-Flower kernel module.
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Co-authored-by: Venkat Duvvuru <venkatkumar.duvvuru@broadcom.com>
Signed-off-by: Venkat Duvvuru <venkatkumar.duvvuru@broadcom.com>
Reviewed-by: Sathya Perla <sathya.perla@broadcom.com>
Reviewed-by: Ben Pfaff <blp@ovn.org>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
Datapath flows in dpif-netdev classifier always has exact match
mask set for vlan. We have to enable it for flow_put operation
too in order to avoid flow modification failure due to
classifier lookup with wrong hash.
Found by OFtest.
CC: Jan Scheurich <jan.scheurich@ericsson.com>
Fixes: beb75a40fdc2 ("userspace: Switching of L3 packets in L2 pipeline")
Reported-by: Ben Pfaff <blp@ovn.org>
Reported-at: https://mail.openvswitch.org/pipermail/ovs-dev/2018-September/352579.html
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ben Pfaff <blp@ovn.org>
Prior to OVS 2.9 automatic assignment of Rxqs to PMDs
(i.e. CPUs) was done by round-robin.
That was changed in OVS 2.9 to ordering the Rxqs based on
their measured processing cycles. This was to assign the
busiest Rxqs to different PMDs, improving aggregate
throughput.
For the most part the new scheme should be better, but
there could be situations where a user prefers a simple
round-robin scheme because Rxqs from a single port are
more likely to be spread across multiple PMDs, and/or
traffic is very bursty/unpredictable.
Add 'pmd-rxq-assign' config to allow a user to select
round-robin based assignment.
Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
Acked-by: Eelco Chaudron <echaudro@redhat.com>
Acked-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Added new types to the flow dump filter, and allowed multiple filter
types to be passed at once, as a comma separated list. The new types
added are:
* tc - specifies flows handled by the tc dp
* non-offloaded - specifies flows not offloaded to the HW
* all - specifies flows of all types
The type list is now fully parsed by the dpctl, and a new struct was
added to dpif which enables dpctl to define which types of dumps to
provide, rather than passing the type string and having dpif parse it.
Signed-off-by: Gavi Teitz <gavi@mellanox.com>
Acked-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
In a previous commit, the dpif_flow struct was expanded, with the
'offloaded' field being moved into a new struct which also includes a
field for the dp layer the flow is handled on. The initialization of
these fields was only done in dpif-netlink.
This completes that commit, by initializing the fields in dpif-netdev
as well. As the 'offloaded' field was previously ignored by
dpif-netdev, the attrs are initialized to the default values of
'false' for the offloaded state, and 'ovs' for the dp layer.
Fixes: d63ca5329ff9 ("dpctl: Properly reflect a rule's offloaded to HW state")
Signed-off-by: Gavi Teitz <gavi@mellanox.com>
Acked-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
dpif_netdev_meter_get() retrieved a pointer to a meter entry without
holding a lock. It's possible that another thread could have deleted
that entry between retrieving the pointer and dereferencing the pointer.
This makes the function hold the lock the entire time the meter entry is
needed.
Found by inspection.
Signed-off-by: Justin Pettit <jpettit@ovn.org>
Acked-by: Flavio Leitner <fbl@sysclose.org>
OVS reads packets in batches from a given port and packets in the
batch are subjected to potentially 3 levels of lookups to identify
the datapath megaflow entry (or flow) associated with the packet.
Each megaflow entry has a dedicated buffer in which packets that match
the flow classification criteria are collected. This buffer helps OVS
perform batch processing for all packets associated with a given flow.
Each packet in the received batch is first subjected to lookup in the
Exact Match Cache (EMC). Each EMC entry will point to a flow. If the
EMC lookup is successful, the packet is moved from the rx batch to the
per-flow buffer.
Packets that did not match any EMC entry are rearranged in the rx batch
at the beginning and are now subjected to a lookup in the megaflow cache.
Packets that match a megaflow cache entry are *appended* to the per-flow
buffer.
Packets that do not match any megaflow entry are subjected to slow-path
processing through the upcall mechanism. This cannot change the order of
packets as by definition upcall processing is only done for packets
without matching megaflow entry.
The EMC entry match fields encompass all potentially significant header
fields, typically more than specified in the associated flow's match
criteria. Hence, multiple EMC entries can point to the same flow. Given
that per-flow batching happens at each lookup stage, packets belonging
to the same megaflow can get re-ordered because some packets match EMC
entries while others do not.
The following example can illustrate the issue better. Consider
following batch of packets (labelled P1 to P8) associated with a single
TCP connection and associated with a single flow. Let us assume that
packets with just the ACK bit set in TCP flags have been received in a
prior batch also and a corresponding EMC entry exists.
1. P1 (TCP Flag: ACK)
2. P2 (TCP Flag: ACK)
3. P3 (TCP Flag: ACK)
4. P4 (TCP Flag: ACK, PSH)
5. P5 (TCP Flag: ACK)
6. P6 (TCP Flag: ACK)
7. P7 (TCP Flag: ACK)
8. P8 (TCP Flag: ACK)
The megaflow classification criteria does not include TCP flags while
the EMC match criteria does. Thus, all packets other than P4 match
the existing EMC entry and are moved to the per-flow packet batch.
Subsequently, packet P4 is moved to the same per-flow packet batch as
a result of the megaflow lookup. Though the packets have all been
correctly classified as being associated with the same flow, the
packet order has not been preserved because of the per-flow batching
performed during the EMC lookup stage. This packet re-ordering has
performance implications for TCP applications.
This patch preserves the packet ordering by performing the per-flow
batching after both the EMC and megaflow lookups are complete. As an
optimization, packets are flow-batched in emc processing till any
packet in the batch has an EMC miss.
A new flow map is maintained to keep the original order of packet
along with flow information. Post fastpath processing, packets from
flow map are *appended* to per-flow buffer.
Signed-off-by: Vishal Deep Ajmera <vishal.deep.ajmera@ericsson.com>
Co-authored-by: Venkatesan Pradeep <venkatesan.pradeep@ericsson.com>
Signed-off-by: Venkatesan Pradeep <venkatesan.pradeep@ericsson.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This patch defines the dpif interface to support conntrack
per zone limit. Basically, OVS users can use this interface
to set, delete, and get the conntrack per zone limit for various
dpif interfaces. The following patch will make use of the proposed
interface to implement the feature.
Signed-off-by: Yi-Hung Wei <yihung.wei@gmail.com>
Signed-off-by: Justin Pettit <jpettit@ovn.org>
The original intent of the API appears to be that the underlying DPIF
implementaion would choose a local meter id. However, neither of the
existing datapath meter implementations (userspace or Linux) implemented
that; they expected a valid meter id to be passed in, otherwise they
returned an error. This commit follows the existing implementations and
makes the API somewhat cleaner.
Signed-off-by: Justin Pettit <jpettit@ovn.org>
Acked-by: Ben Pfaff <blp@ovn.org>
'key.len' should be calculated before inserting to EMC, otherwise
resulting entry will match with any packet with the same hash.
CC: Yipeng Wang <yipeng1.wang@intel.com>
Fixes: 60d8ccae135f ("dpif-netdev: Add SMC cache after EMC cache")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Yipeng Wang <yipeng1.wang@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Another dpif provider will soon add support for meters, so move
some of the common sanity checks up into the dpif layer so that each
provider doesn't need to re-implement them.
Signed-off-by: Justin Pettit <jpettit@ovn.org>
Acked-by: Ben Pfaff <blp@ovn.org>
Commit ab15e70eb587 ("dpctl: Expand the flow dump type filter") will be
reverted, which this patch fixed, so it needs to be reverted as well.
This reverts commit b10ac772218afd4f296db866f6b80258e1d1ca8a.
CC: Gavi Teitz <gavi@mellanox.com>
CC: Simon Horman <simon.horman@netronome.com>
CC: Roi Dayan <roid@mellanox.com>
CC: Aaron Conole <aconole@redhat.com>
Signed-off-by: Justin Pettit <jpettit@ovn.org>
Acked-by: Ben Pfaff <blp@ovn.org>
This patch adds a signature match cache (SMC) after exact match
cache (EMC). The difference between SMC and EMC is SMC only stores
a signature of a flow thus it is much more memory efficient. With
same memory space, EMC can store 8k flows while SMC can store 1M
flows. It is generally beneficial to turn on SMC but turn off EMC
when traffic flow count is much larger than EMC size.
SMC cache will map a signature to an dp_netdev_flow index in
flow_table. Thus, we add two new APIs in cmap for lookup key by
index and lookup index by key.
For now, SMC is an experimental feature that it is turned off by
default. One can turn it on using ovsdb options.
Signed-off-by: Yipeng Wang <yipeng1.wang@intel.com>
Co-authored-by: Jan Scheurich <jan.scheurich@ericsson.com>
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Currently, the major trigger for hw flow offload is at upcall handling,
which is actually in the datapath. Moreover, the hw offload installation
and modification is not that lightweight. Meaning, if there are so many
flows being added or modified frequently, it could stall the datapath,
which could result to packet loss.
To diminish that, all those flow operations will be recorded and appended
to a list. A thread is then introduced to process this list (to do the
real flow offloading put/del operations). This could leave the datapath
as lightweight as possible.
Signed-off-by: Yuanhan Liu <yliu@fridaylinux.org>
Co-authored-by: Shahaf Shuler <shahafs@mellanox.com>
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
So that we could skip some very costly CPU operations, including but
not limiting to miniflow_extract, emc lookup, dpcls lookup, etc. Thus,
performance could be greatly improved.
A PHY-PHY forwarding with 1000 mega flows (udp,tp_src=1000-1999) and
1 million streams (tp_src=1000-1999, tp_dst=2000-2999) show more that
260% performance boost.
Note that though the heavy miniflow_extract is skipped, we still have
to do per packet checking, due to we have to check the tcp_flags.
Co-authored-by: Finn Christensen <fc@napatech.com>
Signed-off-by: Yuanhan Liu <yliu@fridaylinux.org>
Signed-off-by: Finn Christensen <fc@napatech.com>
Co-authored-by: Shahaf Shuler <shahafs@mellanox.com>
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Most modern NICs have the ability to bind a flow with a mark, so that
every packet matches such flow will have that mark present in its
descriptor.
The basic idea of doing that is, when we receives packets later, we could
directly get the flow from the mark. That could avoid some very costly
CPU operations, including (but not limiting to) miniflow_extract, emc
lookup, dpcls lookup, etc. Thus, performance could be greatly improved.
Thus, the major work of this patch is to associate a flow with a mark
id (an uint32_t number). The association in netdev datapath is done
by CMAP, while in hardware it's done by the rte_flow MARK action.
One tricky thing in OVS-DPDK is, the flow tables is per-PMD. For the
case there is only one phys port but with 2 queues, there could be 2
PMDs. In other words, even for a single mega flow (i.e. udp,tp_src=1000),
there could be 2 different dp_netdev flows, one for each PMD. That could
results to the same mega flow being offloaded twice in the hardware,
worse, we may get 2 different marks and only the last one will work.
To avoid that, a megaflow_to_mark CMAP is created. An entry will be
added for the first PMD that wants to offload a flow. For later PMDs,
it will see such megaflow is already offloaded, then the flow will not
be offloaded to HW twice.
Meanwhile, the mark to flow mapping becomes to 1:N mapping. That is
what the mark_to_flow CMAP is for. When the first PMD wants to offload
a flow, it allocates a new mark and performs the flow offload by reusing
the ->flow_put method. When it succeeds, a "mark to flow" entry will be
added. For later PMDs, it will get the corresponding mark by above
megaflow_to_mark CMAP. Then, another "mark to flow" entry will be added.
Signed-off-by: Yuanhan Liu <yliu@fridaylinux.org>
Co-authored-by: Finn Christensen <fc@napatech.com>
Signed-off-by: Finn Christensen <fc@napatech.com>
Co-authored-by: Shahaf Shuler <shahafs@mellanox.com>
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
It's always been OVS coding style to use spaces rather than tabs for
indentation, but some tabs have snuck in over time. This commit converts
them to spaces.
Signed-off-by: Ben Pfaff <blp@ovn.org>
Acked-by: Justin Pettit <jpettit@ovn.org>
Several OVS structs contain embedded named unions, like this:
struct {
...
union {
...
} u;
};
C11 standardized a feature that many compilers already implemented
anyway, where an embedded union may be unnamed, like this:
struct {
...
union {
...
};
};
This is more convenient because it allows the programmer to omit "u."
in many places. OVS already used this feature in several places. This
commit embraces it in several others.
Signed-off-by: Ben Pfaff <blp@ovn.org>
Acked-by: Justin Pettit <jpettit@ovn.org>
Tested-by: Alin Gabriel Serdean <aserdean@ovn.org>
Acked-by: Alin Gabriel Serdean <aserdean@ovn.org>
Unconditional return may cause packet leak in case of
'should_steal == true'.
Additionally, removed redundant checking for depth level.
CC: Sugesh Chandran <sugesh.chandran@intel.com>
Fixes: 7c12dfc527a5 ("tunneling: Avoid datapath-recirc by
combining recirc actions at xlate.")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokess <ian.stokes@intel.com>
When initialization of the DPDK PMD driver fails
(dpdk_eth_dev_init()), the reconfigure_datapath() function will remove
the port from dp_netdev, and the port is not used.
Now when bridge_reconfigure() is called again, no changes to the
previous failing netdev configuration are detected and therefore the
ports gets added to dp_netdev and used uninitialized. This is causing
exceptions...
The fix has two parts to it. First in netdev-dpdk.c we remember if the
DPDK port was started or not, and when calling
netdev_dpdk_reconfigure() we also try re-initialization if the port
was not already active. The second part of the change is in
dpif-netdev.c where it makes sure netdev_reconfigure() is called if
the port needs reconfiguration, as netdev_is_reconf_required() is only
true until netdev_reconfigure() is called (even if it fails).
Signed-off-by: Eelco Chaudron <echaudro@redhat.com>
Tested-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This patch enhances dpif-netdev-perf to detect iterations with
suspicious statistics according to the following criteria:
- iteration lasts longer than US_THR microseconds (default 250).
This can be used to capture events where a PMD is blocked or
interrupted for such a period of time that there is a risk for
dropped packets on any of its Rx queues.
- max vhost qlen exceeds a threshold Q_THR (default 128). This can
be used to infer virtio queue overruns and dropped packets inside
a VM, which are not visible in OVS otherwise.
Such suspicious iterations can be logged together with their iteration
statistics to be able to correlate them to packet drop or other events
outside OVS.
A new command is introduced to enable/disable logging at run-time and
to adjust the above thresholds for suspicious iterations:
ovs-appctl dpif-netdev/pmd-perf-log-set on | off
[-b before] [-a after] [-e|-ne] [-us usec] [-q qlen]
Turn logging on or off at run-time (on|off).
-b before: The number of iterations before the suspicious iteration to
be logged (default 5).
-a after: The number of iterations after the suspicious iteration to
be logged (default 5).
-e: Extend logging interval if another suspicious iteration is
detected before logging occurs.
-ne: Do not extend logging interval (default).
-q qlen: Suspicious vhost queue fill level threshold. Increase this
to 512 if the Qemu supports 1024 virtio queue length.
(default 128).
-us usec: change the duration threshold for a suspicious iteration
(default 250 us).
Note: Logging of suspicious iterations itself consumes a considerable
amount of processing cycles of a PMD which may be visible in the iteration
history. In the worst case this can lead OVS to detect another
suspicious iteration caused by logging.
If more than 100 iterations around a suspicious iteration have been
logged once, OVS falls back to the safe default values (-b 5/-a 5/-ne)
to avoid that logging itself causes continuos further logging.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This patch instruments the dpif-netdev datapath to record detailed
statistics of what is happening in every iteration of a PMD thread.
The collection of detailed statistics can be controlled by a new
Open_vSwitch configuration parameter "other_config:pmd-perf-metrics".
By default it is disabled. The run-time overhead, when enabled, is
in the order of 1%.
The covered metrics per iteration are:
- cycles
- packets
- (rx) batches
- packets/batch
- max. vhostuser qlen
- upcalls
- cycles spent in upcalls
This raw recorded data is used threefold:
1. In histograms for each of the following metrics:
- cycles/iteration (log.)
- packets/iteration (log.)
- cycles/packet
- packets/batch
- max. vhostuser qlen (log.)
- upcalls
- cycles/upcall (log)
The histograms bins are divided linear or logarithmic.
2. A cyclic history of the above statistics for 999 iterations
3. A cyclic history of the cummulative/average values per millisecond
wall clock for the last 1000 milliseconds:
- number of iterations
- avg. cycles/iteration
- packets (Kpps)
- avg. packets/batch
- avg. max vhost qlen
- upcalls
- avg. cycles/upcall
The gathered performance metrics can be printed at any time with the
new CLI command
ovs-appctl dpif-netdev/pmd-perf-show [-nh] [-it iter_len] [-ms ms_len]
[-pmd core] [dp]
The options are
-nh: Suppress the histograms
-it iter_len: Display the last iter_len iteration stats
-ms ms_len: Display the last ms_len millisecond stats
-pmd core: Display only the specified PMD
The performance statistics are reset with the existing
dpif-netdev/pmd-stats-clear command.
The output always contains the following global PMD statistics,
similar to the pmd-stats-show command:
Time: 15:24:55.270
Measurement duration: 1.008 s
pmd thread numa_id 0 core_id 1:
Cycles: 2419034712 (2.40 GHz)
Iterations: 572817 (1.76 us/it)
- idle: 486808 (15.9 % cycles)
- busy: 86009 (84.1 % cycles)
Rx packets: 2399607 (2381 Kpps, 848 cycles/pkt)
Datapath passes: 3599415 (1.50 passes/pkt)
- EMC hits: 336472 ( 9.3 %)
- Megaflow hits: 3262943 (90.7 %, 1.00 subtbl lookups/hit)
- Upcalls: 0 ( 0.0 %, 0.0 us/upcall)
- Lost upcalls: 0 ( 0.0 %)
Tx packets: 2399607 (2381 Kpps)
Tx batches: 171400 (14.00 pkts/batch)
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
If the caller provides a non-NULL qfill pointer and the netdev
implemementation supports reading the rx queue fill level, the rxq_recv()
function returns the remaining number of packets in the rx queue after
reception of the packet burst to the caller. If the implementation does
not support this, it returns -ENOTSUP instead. Reading the remaining queue
fill level should not substantilly slow down the recv() operation.
A first implementation is provided for ethernet and vhostuser DPDK ports
in netdev-dpdk.c.
This output parameter will be used in the upcoming commit for PMD
performance metrics to supervise the rx queue fill level for DPDK
vhostuser ports.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Until mow, this macro has blindly read the passed-in type's size, but
that's unnecessarily risky. This commit changes it to verify that the
passed-in type is the same size as the field and, on GCC and Clang, that
the types are compatible. It also adds a version that does not check,
for the one case where (currently) we deliberately read the wrong size,
and updates a few uses to use more precise field names.
Signed-off-by: Ben Pfaff <blp@ovn.org>
Reviewed-by: Yifeng Sun <pkusunyifeng@gmail.com>
Reviewed-by: Armando Migliaccio <armamig@gmail.com>
This patch adds support of flushing a conntrack entry specified by the
conntrack 5-tuple in dpif-netdev.
Signed-off-by: Yi-Hung Wei <yihung.wei@gmail.com>
Signed-off-by: Ben Pfaff <blp@ovn.org>
Acked-by: Darrell Ball <dlu998@gmail.com>
ofp-util had been far too large and monolithic for a long time. This
commit breaks it up into units that make some logical sense. It also
moves the pieces of ofp-parse that were specific to each unit into the
relevant unit.
Most of this commit is just moving code around.
Signed-off-by: Ben Pfaff <blp@ovn.org>
Reviewed-by: Yifeng Sun <pkusunyifeng@gmail.com>
This supports using the ct_clear action in the kernel datapath. To
preserve compatibility with current ct_clear behavior on old kernels, we
only pass this action down to the datapath if a probe reveals the
datapath actually supports it.
Signed-off-by: Eric Garver <e@erig.me>
Acked-by: William Tu <u9012063@gmail.com>
Acked-by: Flavio Leitner <fbl@sysclose.org>
Signed-off-by: Justin Pettit <jpettit@ovn.org>
It is based on the length of history that is stored about an
rxq (currently 1 min).
$ ovs-appctl dpif-netdev/pmd-rxq-show
pmd thread numa_id 0 core_id 4:
isolated : false
port: dpdkphy1 queue-id: 0 pmd usage: 70 %
port: dpdkvhost0 queue-id: 0 pmd usage: 0 %
pmd thread numa_id 0 core_id 6:
isolated : false
port: dpdkphy0 queue-id: 0 pmd usage: 64 %
port: dpdkvhost1 queue-id: 0 pmd usage: 0 %
These values are what would be used as part of rxq to pmd
assignment due to a reconfiguration event e.g. adding pmds,
adding rxqs or with the command:
ovs-appctl dpif-netdev/pmd-rxq-rebalance
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Co-authored-by: Jan Scheurich <jan.scheurich@ericsson.com>
Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
An rxq may have processing cycles counted in the current
counter when a reload happens. That could temporarily create
a small skew on the stats for an rxq. Reset the counter after
reload.
Fixes: 4809891b2e01 ("dpif-netdev: Count the rxq processing cycles for an rxq.")
Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This allows to collect packets from more than one RX burst
and send them together with a configurable intervals.
'other_config:tx-flush-interval' can be used to configure
time that a packet can wait in output batch for sending.
'tx-flush-interval' has microsecond resolution.
Tested-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Jan Scheurich <jan.scheurich@ericsson.com>
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Upcoming time-based output batching will allow to collect in a single
output batch packets from different RX queues. Lets keep the list of
RX queues for each output packet and collect cycles for them on send.
Tested-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Jan Scheurich <jan.scheurich@ericsson.com>
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Simplify the historically grown TSC cycle counting in PMD threads.
Cycles are currently counted for the following purposes:
1. Measure PMD ustilization
PMD utilization is defined as ratio of cycles spent in busy iterations
(at least one packet received or sent) over the total number of cycles.
This is already done in pmd_perf_start_iteration() and
pmd_perf_end_iteration() based on a TSC timestamp saved in current
iteration at start_iteration() and the actual TSC at end_iteration().
No dependency on intermediate cycle accounting.
2. Measure the processing load per RX queue
This comprises cycles spend on polling and processing packets received
from the rx queue and the cycles spent on delayed sending of these packets
to tx queues (with time-based batching).
The previous scheme using cycles_count_start(), cycles_count_intermediate()
and cycles-count_end() originally introduced to simplify cycle counting
and saving calls to rte_get_tsc_cycles() was rather obscuring things.
Replace by a nestable cycle_timer with with start and stop functions to
embrace a code segment to be timed. The timed code may contain arbitrary
nested cycle_timers. The duration of nested timers is excluded from the
outer timer.
The caller must ensure that each call to cycle_timer_start() is
followed by a call to cycle_timer_end(). Failure to do so will lead to
assertion failure or a memory leak.
The new cycle_timer is used to measure the processing cycles per rx queue.
This is not yet strictly necessary but will be made use of in a subsequent
commit.
All cycle count functions and data are relocated to module
dpif-netdev-perf.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off: Ian Stokes <ian.stokes@intel.com>
Add module dpif-netdev-perf to host all PMD performance-related
data structures and functions in dpif-netdev. Refactor the PMD
stats handling in dpif-netdev and delegate whatever possible into
the new module, using clean interfaces to shield dpif-netdev from
the implementation details. Accordingly, the all PMD statistics
members are moved from the main struct dp_netdev_pmd_thread into
a dedicated member of type struct pmd_perf_stats.
Include Darrel's prior refactoring of PMD stats contained in
[PATCH v5,2/3] dpif-netdev: Refactor some pmd stats:
1. The cycles per packet counts are now based on packets
received rather than packet passes through the datapath.
2. Packet counters are now kept for packets received and
packets recirculated. These are kept as separate counters for
maintainability reasons. The cost of incrementing these counters
is negligible. These new counters are also displayed to the user.
3. A display statistic is added for the average number of
datapath passes per packet. This should be useful for user
debugging and understanding of packet processing.
4. The user visible 'miss' counter is used for successful upcalls,
rather than the sum of sucessful and unsuccessful upcalls. Hence,
this becomes what user historically understands by OVS 'miss upcall'.
The user display is annotated to make this clear as well.
5. The user visible 'lost' counter remains as failed upcalls, but
is annotated to make it clear what the meaning is.
6. The enum pmd_stat_type is annotated to make the usage of the
stats counters clear.
7. The subtable lookup stats is renamed to make it clear that it
relates to masked lookups.
8. The PMD stats test is updated to handle the new user stats of
packets received, packets recirculated and average number of datapath
passes per packet.
On top of that introduce a "-pmd <core>" option to the PMD info
commands to filter the output for a single PMD.
Made the pmd-stats-show output a bit more readable by adding a blank
between colon and value.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Co-authored-by: Darrell Ball <dlu998@gmail.com>
Signed-off-by: Darrell Ball <dlu998@gmail.com>
Acked-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off: Ian Stokes <ian.stokes@intel.com>
