Use lightweight transmit handler which handles non-offloaded Tx data
path. We get CPU utilization improvement of ~8%.
Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
Use light weight receive packet handler for non-LRO and
non-scatter packets to improve CPU utilization in receive data path.
We achieved ~18% CPU cycles improvement using this handler.
Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
As per HW design of 100Gb mode, device internally uses 2 engines
(eng0 and eng1), and both engines need to be configured symmetrically.
Based on this requirement, driver design chose an approach
to allow user to allocate only even number of queues and split
those queues on both engines equally.
This approach puts a limitation on number of queues to be allocated -
i.e. user can't configure odd number of queues on 100Gb mode.
OVS configures DPDK port with 1 rxq and 1 txq, which causes initialization
of qede port to fail.
Issue is fixed by changing the implementation of queue allocation and
assignment to hw engines only for 100Gb devices and allowing user to
configure odd number queues.
New approach works as below -
- Create 'struct qede_fastpath_cmt' to hold hw queue pair of both engines
and provide it to rte_ethdev's Rx/Tx queues structure.
- So ethdev will see only one queue for underlying queue pair created for
hw engine pair.
- Install separate Rx/Tx data path handlers for 100Gb mode and regular mode
- Rx/Tx handlers for 100Gb mode will split packet processing across both
engines by providing hw queue structures from 'struct qede_fastpath_cmt'
passed by Rx/Tx callbacks to respective engines.
Fixes: 2af14ca79c ("net/qede: support 100G")
Cc: stable@dpdk.org
Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
Add 'RTE_' prefix to defines:
- rename ETHER_ADDR_LEN as RTE_ETHER_ADDR_LEN.
- rename ETHER_TYPE_LEN as RTE_ETHER_TYPE_LEN.
- rename ETHER_CRC_LEN as RTE_ETHER_CRC_LEN.
- rename ETHER_HDR_LEN as RTE_ETHER_HDR_LEN.
- rename ETHER_MIN_LEN as RTE_ETHER_MIN_LEN.
- rename ETHER_MAX_LEN as RTE_ETHER_MAX_LEN.
- rename ETHER_MTU as RTE_ETHER_MTU.
- rename ETHER_MAX_VLAN_FRAME_LEN as RTE_ETHER_MAX_VLAN_FRAME_LEN.
- rename ETHER_MAX_VLAN_ID as RTE_ETHER_MAX_VLAN_ID.
- rename ETHER_MAX_JUMBO_FRAME_LEN as RTE_ETHER_MAX_JUMBO_FRAME_LEN.
- rename ETHER_MIN_MTU as RTE_ETHER_MIN_MTU.
- rename ETHER_LOCAL_ADMIN_ADDR as RTE_ETHER_LOCAL_ADMIN_ADDR.
- rename ETHER_GROUP_ADDR as RTE_ETHER_GROUP_ADDR.
- rename ETHER_TYPE_IPv4 as RTE_ETHER_TYPE_IPv4.
- rename ETHER_TYPE_IPv6 as RTE_ETHER_TYPE_IPv6.
- rename ETHER_TYPE_ARP as RTE_ETHER_TYPE_ARP.
- rename ETHER_TYPE_VLAN as RTE_ETHER_TYPE_VLAN.
- rename ETHER_TYPE_RARP as RTE_ETHER_TYPE_RARP.
- rename ETHER_TYPE_QINQ as RTE_ETHER_TYPE_QINQ.
- rename ETHER_TYPE_ETAG as RTE_ETHER_TYPE_ETAG.
- rename ETHER_TYPE_1588 as RTE_ETHER_TYPE_1588.
- rename ETHER_TYPE_SLOW as RTE_ETHER_TYPE_SLOW.
- rename ETHER_TYPE_TEB as RTE_ETHER_TYPE_TEB.
- rename ETHER_TYPE_LLDP as RTE_ETHER_TYPE_LLDP.
- rename ETHER_TYPE_MPLS as RTE_ETHER_TYPE_MPLS.
- rename ETHER_TYPE_MPLSM as RTE_ETHER_TYPE_MPLSM.
- rename ETHER_VXLAN_HLEN as RTE_ETHER_VXLAN_HLEN.
- rename ETHER_ADDR_FMT_SIZE as RTE_ETHER_ADDR_FMT_SIZE.
- rename VXLAN_GPE_TYPE_IPV4 as RTE_VXLAN_GPE_TYPE_IPV4.
- rename VXLAN_GPE_TYPE_IPV6 as RTE_VXLAN_GPE_TYPE_IPV6.
- rename VXLAN_GPE_TYPE_ETH as RTE_VXLAN_GPE_TYPE_ETH.
- rename VXLAN_GPE_TYPE_NSH as RTE_VXLAN_GPE_TYPE_NSH.
- rename VXLAN_GPE_TYPE_MPLS as RTE_VXLAN_GPE_TYPE_MPLS.
- rename VXLAN_GPE_TYPE_GBP as RTE_VXLAN_GPE_TYPE_GBP.
- rename VXLAN_GPE_TYPE_VBNG as RTE_VXLAN_GPE_TYPE_VBNG.
- rename ETHER_VXLAN_GPE_HLEN as RTE_ETHER_VXLAN_GPE_HLEN.
Do not update the command line library to avoid adding a dependency to
librte_net.
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Allocating replacement buffer per received packet is expensive.
Instead, process received packets first and allocate
replacement buffers in bulk later.
This improves performance by ~25% in terms of PPS on AMD
platforms.
Fixes: 2ea6f76aff ("qede: add core driver")
Cc: stable@dpdk.org
Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
Acked-by: Rasesh Mody <rmody@marvell.com>
Following commit updated the Tx offload mask
commit 1037ed842c ("mbuf: fix Tx offload mask").
So, not having PKT_TX_IPV6 and PKT_TX_IPV4 in qede PMDs supported
Tx offload mask breaks TSO support since application will fail in transmit
prepare function.
Fixes: 1037ed842c ("mbuf: fix Tx offload mask")
Cc: stable@dpdk.org
Signed-off-by: Shahed Shaikh <shahed.shaikh@cavium.com>
Tunneling offloads are represented by multi-bit values. So, feature
wise tunneling offload can only be entirely supported/unsupported
using PKT_TX_TUNNEL_MASK. Its upon PMDs to further isolate which of
the tunneling offload types are supported by respective PMD.
Using subset of bits from PKT_TX_TUNNEL_MASK to indicate supported vs
unsupported offloads can lead to undesired result.
Use PKT_TX_TUNNEL_MASK in QEDE_TX_OFFLOAD_MASK and use independent
value of supported PKT_TX_TUNNEL_* in .tx_pkt_prepare() to mark
supported tunnel offloads.
Fixes: 44346c24b7 ("net/qede: fix VXLAN tunnel Tx offload flag setting")
Cc: stable@dpdk.org
Suggested-by: Ferruh Yigit <ferruh.yigit@intel.com>
Signed-off-by: Shahed Shaikh <shahed.shaikh@cavium.com>
Reviewed-by: Rasesh Mody <rasesh.mody@cavium.com>
This patch implement eth_dev_ops->rx_descriptor_status
callback.
Walk through receive completion ring to calculate receive
descriptors used by firmware and then provide the status of
offset accordingly.
Signed-off-by: Shahed Shaikh <shahed.shaikh@cavium.com>
- HW does not include CRC in received frame when passed to host,
so no need to consider CRC length while calculating Rx buffer size.
- In scattered Rx mode, driver may allocate Rx buffer larger than
the size of mbuf because it tries to adjust the buffer size to cache
line size by ceiling it. Fix this by flooring the size instead of
ceiling.
- Consider the rule imposed by HW regarding the minimum size of Rx buffer
in scattered Rx mode -
(MTU + Maximum L2 Header Size + 2) / ETH_RX_MAX_BUFF_PER_PKT
Fixes: f6033f2497 ("net/qede: fix minimum buffer size and scatter Rx check")
CC: stable@dpdk.org
Signed-off-by: Shahed Shaikh <shahed.shaikh@cavium.com>
We were using LICENSE.qede_pmd to reference inclusion of SPDX licensing
tag from all the source file. Remove the LICENSE.qede_pmd file and
directly include SPDX tags in source files.
Signed-off-by: Rasesh Mody <rasesh.mody@cavium.com>
Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>
- We don't support QinQ offload, so removing it now.
- Fix incorrect offload flags in default rxconf
Since qede PMD does not support per queue rx offload, it
should not set default_rxconf.offload flags in .dev_infos_get().
Although these offloads are enabled by default, they are per port
and not per queue.
Fixes: 946dfd18a4 ("net/qede: convert to new Rx/Tx offloads API")
Cc: stable@dpdk.org
Signed-off-by: Shahed Shaikh <shahed.shaikh@cavium.com>
Ethdev RX/TX offloads API has changed since:
commit ce17eddefc ("ethdev: introduce Rx queue offloads API")
commit cba7f53b71 ("ethdev: introduce Tx queue offloads API")
This patch makes use of new offload flags as suggested and fix
the default RX/TX configuration. Also indent the code properly
around those changes.
Signed-off-by: Harish Patil <harish.patil@cavium.com>
Signed-off-by: Rasesh Mody <rasesh.mody@cavium.com>
This patch fixes issues related to MTU set and max_rx_pkt_len usage.
- Adjust MTU during device configuration when jumbo is enabled
- In qede_set_mtu():
Return not supported for VF as currently we do not support it.
Cache new mtu value in mtu_new for proper update.
Add check for RXQ allocation before calculating RX buffer size
if not allocated defer RX buffer size calculation till RXQ setup.
Add check for before performing device start/stop.
- Use max_rx_pkt_len appropriately
- Change QEDE_ETH_OVERHEAD macro to adjust driver specifics
Fixes: 4c4bdadfa9 ("net/qede: refactoring multi-queue implementation")
Fixes: 9a6d30ae6d ("net/qede: refactoring vport handling code")
Fixes: 1ef4c3a5c1 ("net/qede: prevent crash while changing MTU dynamically")
Cc: stable@dpdk.org
Signed-off-by: Rasesh Mody <rasesh.mody@cavium.com>
Add support for inner/outer L3/L4 TX csum offload for MPLS-in-UDP packets.
The driver checks for PKT_TX_TUNNEL_MPLSINUDP in Tx ol_flags and updates TX
BD elements with appropriate offsets/length of tunnel headers.
The pseudo csum calculation is removed from qede_xmit_prep_pkts() since its
not needed.
Signed-off-by: Harish Patil <harish.patil@cavium.com>
Signed-off-by: Rasesh Mody <rasesh.mody@cavium.com>
This patch fixes missing PKT_TX_TUNNEL_VXLAN Tx offload flag from the
supported Tx offloads and an incorrect tunnel TX BD bit setting.
Fixes: 3d4bb44116 ("net/qede: add fastpath support for VXLAN tunneling")
Cc: stable@dpdk.org
Signed-off-by: Harish Patil <harish.patil@cavium.com>
This patch does the following refactoring and cleanup:
- As part of multi-queue support a struct member called 'type' was added
in struct qede_fastpath in order to identify whether a queue is RX or
TX and take actions based on that. This was unnecessary in the first
place since pointers to RX and TX queues are already available in
rte_eth_dev->data. So all usage of fp->type is removed.
- Remove remaining additional layer of internal callbacks for RX/TX
queues and fastpath related operations from the qed_eth_ops_pass.
With this change the files qede_eth_if.[c,h] are no longer needed.
- Add new per-queue start/stop APIs instead of clubbing it all together.
- Remove multiple TXQs references (num_tc and fp->txqs) since CoS is not
supported.
- Enable sharing of the status block for each queue pair.
- Remove enum qede_dev_state and instead make use of existing port
states RTE_ETH_QUEUE_STATE_STOPPED/RTE_ETH_QUEUE_STATE_STARTED.
- Move qede_dev_start() and qede_dev_stop() to qede_ethdev.c from
qede_rxtc.c.
Signed-off-by: Harish Patil <harish.patil@cavium.com>
remove __rte_unused instances that are not required.
Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Allain Legacy <allain.legacy@windriver.com>
Acked-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
- Add a common routine to handle ETH_RX_CQE_TYPE_TPA_CONT and
ETH_RX_CQE_TYPE_TPA_END.
- Remove enum qede_agg_state since there is no need to maintain
aggregation state.
- Modify the segment chaining logic by tracking head and tail
TPA segments of each aggregation.
- Add more debug and comments.
- Mark the packet as PKT_RX_LRO.
Fixes: 29540be7ef ("net/qede: support LRO/TSO offloads")
Signed-off-by: Harish Patil <harish.patil@cavium.com>
The driver can handle dynamic MTU change without needing the port to be
stopped explicitly by the application. However, there is currently no
check to prevent I/Os from happening on a different thread while the
port is going thru' reset internally. This patch fixes this issue by
assigning RX/TX burst functions to a dummy function and also reconfigure
RX bufsize for each rx queue based on the new MTU value.
Fixes: 200645ac79 ("net/qede: set MTU")
Cc: stable@dpdk.org
Signed-off-by: Harish Patil <harish.patil@qlogic.com>
This patch includes slowpath configuration and fastpath changes
to support LRO and TSO. A bit of revamping is needed in order
to make use of existing packet classification schemes in Rx fastpath
and for SG element processing in Tx.
Signed-off-by: Harish Patil <harish.patil@qlogic.com>
L2 handler changes:
This is change to remove the queue-id/qzone difference for Tx queues.
It does that by mainly doing:
a. VFs queues are no longer determined by the SBs they're using.
Instead, the ecore-client needs to maintain those and choose the values
to be used by VF when initializing it.
b. Eliminate the HW-cid array in the hw-function.
To do that, have all the rx/tx functionality turn into 'handle' base -
when queue would be started the caller would get a (void*) handle,
which it would later use with ecore for configuring various
queue-related stop [update, close].
Signed-off-by: Rasesh Mody <rasesh.mody@cavium.com>
- Fix minimum RX buffer size to 1024B
- Force enable scatter/gather mode if given RX buf size is lesser than MTU
- Adjust RX buffer size to cache-line size with overhead included
Fixes: bec0228816 ("net/qede: support scatter gather")
Fixes: 2ea6f76aff ("qede: add core driver")
Signed-off-by: Harish Patil <harish.patil@qlogic.com>
This patch includes the following:
- Fix missing hash_key_size advertisement
- Fix RSS hash query function
- Update RSS offload flag
- Accept user provided RSS configuration params via rx_adv_conf
in dev_configure()
- Decouple RSS configuration from common qed_update_vport() and
instead make use of existing RSS APIs for default RSS configuration
Fixes: 6d9e26c42c ("net/qede: get RSS hash configuration")
Fixes: 9c5d0a669f ("net/qede: fix RSS")
Signed-off-by: Harish Patil <harish.patil@qlogic.com>
Some applications set port configuration params like promisc mode
before calling dev_start(). This config results in a firmware exception
since this operation internally translates to sending of VPORT-UPDATE
before VPORT-START ramrod which is considered illegal from firmware
standpoint. So the fix is to send VPORT-START ramrod sooner
in dev_configure() rather than deferring it to dev_start().
This requires a bit of reshuffling in the code to move sending of
VPORT-START from qede_start_queues() to qede_dev_configure()
and VPORT-STOP from qede_stop_queues() to qede_dev_stop().
This sequence change also exposes a flaw in the port restart
flows where the fastpath resource allocation routine qede_init_fp()
functionalities need to be split, so that appropriate action is taken
based on the current port state. Eg: Do not re-initialize the status
block in a port restart case. This change ensures port start/stop
can be paired.
A new port state QEDE_DEV_CONFIG is added to distinguish between
port started from scratch vs port requiring a reconfig (like MTU).
The function qede_config_rx_mode() is removed since the individual
port config will be replayed anyways on a restart.
Fixes: 2ea6f76aff ("qede: add core driver")
Signed-off-by: Harish Patil <harish.patil@qlogic.com>
Previous release of the qede PMD had a limitation that the
driver expects the number of tx and rx queues to be the same.
This patch fixes this issue by making appropriate changes in
control and data path.
Fixes: 2ea6f76aff ("qede: add core driver")
Signed-off-by: Sony Chacko <sony.chacko@qlogic.com>
This patch updates the base driver and incorporates necessary changes
required to bring in the new firmware 8.10.9.0.
In addition, it would allow driver to add new functionalities that might
be needed in future.
Signed-off-by: Rasesh Mody <rasesh.mody@qlogic.com>
The Qlogic Everest Driver for Ethernet(QEDE) Poll Mode Driver(PMD) is
the DPDK specific module for QLogic FastLinQ QL4xxxx 25G/40G CNA family
of adapters as well as their virtual functions (VF) in SR-IOV context.
This patch adds QEDE PMD, which interacts with base driver and
initialises the HW.
This patch content also includes:
- eth_dev_ops callbacks
- Rx/Tx support for the driver
- link default configuration
- change link property
- link up/down/update notifications
- vlan offload and filtering capability
- device/function/port statistics
- qede nic guide and updated overview.rst
Note that the follow on commits contain the code for the features mentioned
in documents but not implemented in this patch.
Signed-off-by: Harish Patil <harish.patil@qlogic.com>
Signed-off-by: Rasesh Mody <rasesh.mody@qlogic.com>
Signed-off-by: Sony Chacko <sony.chacko@qlogic.com>
