Maxime Coquelin [Tue, 29 Jun 2021 16:11:29 +0000 (18:11 +0200)]
vhost: fix missing cache logging NUMA realloc
When the guest allocates virtqueues on a different NUMA node
than the one the Vhost metadata are allocated, both the Vhost
device struct and the virtqueues struct are reallocated.
However, reallocating the log cache on the new NUMA node was
not done. This patch fixes this by reallocating it if it has
been allocated already, which means a live-migration is
on-going.
Fixes: 1818a63147fb ("vhost: move dirty logging cache out of virtqueue") Cc: stable@dpdk.org Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com> Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Maxime Coquelin [Tue, 29 Jun 2021 16:11:28 +0000 (18:11 +0200)]
vhost: fix missing guest pages table NUMA realloc
When the guest allocates virtqueues on a different NUMA node
than the one the Vhost metadata are allocated, both the Vhost
device struct and the virtqueues struct are reallocated.
However, reallocating the guest pages table was missing, which
likely causes at least one cross-NUMA accesses for every burst
of packets.
This patch reallocates this table on the same NUMA node as the
other metadata.
Fixes: e246896178e6 ("vhost: get guest/host physical address mappings") Cc: stable@dpdk.org Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com> Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Maxime Coquelin [Tue, 29 Jun 2021 16:11:27 +0000 (18:11 +0200)]
vhost: fix missing memory table NUMA realloc
When the guest allocates virtqueues on a different NUMA node
than the one the Vhost metadata are allocated, both the Vhost
device struct and the virtqueues struct are reallocated.
However, reallocating the Vhost memory table was missing, which
likely causes at least one cross-NUMA accesses for every burst
of packets.
This patch reallocates this table on the same NUMA node as the
other metadata.
Fixes: 552e8fd3d2b4 ("vhost: simplify memory regions handling") Cc: stable@dpdk.org Reported-by: David Marchand <david.marchand@redhat.com> Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com> Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Dapeng Yu [Mon, 21 Jun 2021 07:23:53 +0000 (15:23 +0800)]
net/i40e: fix multi-process shared data
The rte_eth_devices array is not in share memory, it should not be
referenced by i40e_adapter which is shared by primary and secondary.
Any process set i40e_adapter->eth_dev will corrupt another process's
context.
The patch removed the field "eth_dev" from i40e_adapter.
Now, when the data paths try to access the rte_eth_dev_data instance,
they should replace adapter->eth_dev->data with adapter->pf.dev_data.
Feifei Wang [Wed, 30 Jun 2021 06:40:36 +0000 (14:40 +0800)]
net/i40e: improve vector Tx performance
For i40e vector Tx path, if tx_offload is set as FAST_FREE_MBUF mode,
no mbuf fast free operations are executed. To fix this, add mbuf fast
free mode for vector Tx path.
Furthermore, for i40e vector Tx path, if implement FAST_FREE_MBUF mode,
it means per-queue all mbufs come from the same mempool and have
refcnt = 1. Thus we can use bulk free of the buffers when mbuf fast free
mode is enabled.
For vector path in arm platform:
In n1sdp, performance is improved by 18.4%;
In thunderx2, performance is improved by 23%.
For vector path in x86 platform:
No performance changes.
Suggested-by: Ruifeng Wang <ruifeng.wang@arm.com> Signed-off-by: Feifei Wang <feifei.wang2@arm.com> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Haiyue Wang [Tue, 29 Jun 2021 16:40:33 +0000 (00:40 +0800)]
net/ice: add option for setting HW debug mask
The HW debug mask is always zero, so user can't enable the related debug
function like ICE_DBG_XXX etc, add the devarg 'hw_debug_mask' to set the
debug mask log output at runtime.
Signed-off-by: Haiyue Wang <haiyue.wang@intel.com> Acked-by: Qi Zhang <qi.z.zhang@intel.com>
Beilei Xing [Thu, 24 Jun 2021 07:29:56 +0000 (15:29 +0800)]
net/i40e: fix flow director input set conflict
Currently, there'll be conflict error when running
the following commands:
1. flow create 0 ingress
pattern eth / ipv4 / udp src is 32 / end
actions queue index 2 / end
2. flow destroy 0 rule 0
3. flow create 0 ingress
pattern eth / ipv4 / udp dst is 32 / end
actions queue index 2 / end
This patch fixes the input set conflict issue.
Fixes: 42044b69c67d ("net/i40e: support input set selection for FDIR") Fixes: 4a072ad43442 ("net/i40e: fix flow director config after flow validate") Cc: stable@dpdk.org Signed-off-by: Beilei Xing <beilei.xing@intel.com> Tested-by: Lingli Chen <linglix.chen@intel.com>
net/iavf: check Tx packet with correct UP and queue
Add check in the Tx packet preparation function, to guarantee that the
packet with specific user priority is distributed to the correct Tx
queue according to the configured Tx queue TC mapping.
net/iavf: query QoS capabilities and set queue TC mapping
This patch added the support for VF to config the ETS-based Tx QoS,
including querying current QoS configuration from PF and config queue TC
mapping. PF QoS is configured in advance and the queried info is
provided to the user for future usage. VF queues are mapped to different
TCs in PF through virtchnl.
This patch supports the ETS-based QoS configuration. It enables the DCF
to configure bandwidth limits for each VF VSI of different TCs. A
hierarchy scheduler tree is built with port, TC and VSI nodes.
In the adminq command query port ETS function, the root node teid is
needed. However, for DCF, the root node is not initialized, which will
cause error when we refer to the variable. In this patch, we will check
whether the root node is available or not first.
common/iavf: support ETS-based QoS offload configuration
This patch adds new virtchnl opcodes and structures for QoS
configuration, which includes:
1. VIRTCHNL_VF_OFFLOAD_TC, to negotiate the capability supporting QoS
configuration. If VF and PF both have this flag, then the ETS-based QoS
offload function is supported.
2. VIRTCHNL_OP_DCF_CONFIG_BW, DCF is supposed to configure min and max
bandwidth for each VF per enabled TCs. To make the VSI node bandwidth
configuration work, DCF also needs to configure TC node bandwidth
directly.
3. VIRTCHNL_OP_GET_QOS_CAPS, VF queries current QoS configuration, such
as enabled TCs, arbiter type, up2tc and bandwidth of VSI node. The
configuration is previously set by DCB and DCF, and now is the potential
QoS capability of VF. VF can take it as reference to configure queue TC
mapping.
4. VIRTCHNL_OP_CONFIG_TC_MAP, set VF queues to TC mapping for all Tx and
Rx queues. Queues mapping to one TC should be continuous and all
allocated queues should be mapped.
Wenzhuo Lu [Tue, 29 Jun 2021 02:29:21 +0000 (10:29 +0800)]
net/ice: add AVX2 offload Rx
Add a specific path for RX AVX2.
In this path, support the HW offload features, like,
checksum, VLAN stripping, RSS hash.
This path is chosen automatically according to the
configuration.
'inline' is used, then the duplicate code is generated
by the compiler.
Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com> Acked-by: Leyi Rong <leyi.rong@intel.com>
Wenzhuo Lu [Tue, 29 Jun 2021 02:29:20 +0000 (10:29 +0800)]
net/ice: add AVX2 offload Tx
Add a specific path for TX AVX2.
In this path, support the HW offload features, like,
checksum insertion, VLAN insertion.
This path is chosen automatically according to the
configuration.
'inline' is used, then the duplicate code is generated
by the compiler.
Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com> Acked-by: Leyi Rong <leyi.rong@intel.com>
Wenjun Wu [Mon, 21 Jun 2021 08:59:23 +0000 (16:59 +0800)]
net/ice/base: fix first profile mask
Since each PF does not share the same structure space, the first
mask value should start at 0 instead of hw->pf_id * per_pf to avoid
address overflow. Otherwise, address space will overlap when
masks.first + masks.count > ICE_PROF_MASK_COUNT, and it may lead to
unexpected variable assignment, which causes segmentation fault.
Fixes: 9467486f179f ("net/ice/base: enable masking for RSS and FD field vectors") Cc: stable@dpdk.org Signed-off-by: Wenjun Wu <wenjun1.wu@intel.com> Acked-by: Qi Zhang <qi.z.zhang@intel.com>
Steve Yang [Mon, 21 Jun 2021 08:03:42 +0000 (08:03 +0000)]
net/i40e: fix available RSS hash function
i40e can support following rss hash function types: default/toeplitz,
symmetric toeplitz, and simple_xor. However, when filter engine parses
pattern action, it only supports symmetric toeplitz & default.
Add simple_xor and toeplitz hash functions support when parsing pattern
action.
Fixes: ef4c16fd9148 ("net/i40e: refactor RSS flow") Cc: stable@dpdk.org Signed-off-by: Steve Yang <stevex.yang@intel.com> Acked-by: Beilei Xing <beilei.xing@intel.com>
Tudor Cornea [Tue, 15 Jun 2021 10:38:34 +0000 (13:38 +0300)]
net/ice: fix overflow in maximum packet length config
The len variable, used in the computation of max_pkt_len could overflow,
if used to store the result of the following computation:
ICE_SUPPORT_CHAIN_NUM * rxq->rx_buf_len
Since, we could define the mbuf size to have a large value (i.e 13312),
and ICE_SUPPORT_CHAIN_NUM is defined as 5, the computation mentioned
above could potentially result in a value which might be bigger than
MAX_USHORT.
The result will be that Jumbo Frames will not work properly
Dapeng Yu [Wed, 9 Jun 2021 07:05:08 +0000 (15:05 +0800)]
net/i40e: fix data path selection in secondary process
The flag use_avx2 and use_avx512 are defined as local variables, they
will not be aware by the secondary process, then wrong data path is
selected. Fix the issue by moving them into struct i40e_adapter.
Pavan Nikhilesh [Tue, 29 Jun 2021 07:44:21 +0000 (13:14 +0530)]
net/cnxk: enable PTP processing in vector Tx
Enable PTP offload in vector Tx burst function. Since, we can
no-longer use a single LMT line for burst of 4, split the LMT
into two and transmit twice.
Harman Kalra [Tue, 22 Jun 2021 20:05:53 +0000 (01:35 +0530)]
common/octeontx2: send link event to VF
Currently link event is only sent to the PF by AF as soon as it comes
up, or in case of any physical change in link. PF will broadcast
these link events to all its VFs as soon as it receives it.
But no event is sent when a new VF comes up, hence it will not have
the link status.
Adding support for sending link status to the VF once it comes up
successfully.
Signed-off-by: Harman Kalra <hkalra@marvell.com> Acked-by: Jerin Jacob <jerinj@marvell.com>
Sunil Kumar Kori [Wed, 23 Jun 2021 04:47:01 +0000 (10:17 +0530)]
net/cnxk: support multicast filter
Patch adds multicast filter support for cn9k and cn10k platforms.
CGX DMAC filter table(32 entries) is divided among all LMACs
connected to it i.e. if CGX has 4 LMACs then each LMAC can have
up to 8 filters. If CGX has 1 LMAC then it can have up to 32
filters.
Above mentioned filter table is used to install unicast and multicast
DMAC address filters. Unicast filters are installed via
rte_eth_dev_mac_addr_add API while multicast filters are installed
via rte_eth_dev_set_mc_addr_list API.
So in total, supported MAC filters are equal to DMAC filters plus
mcast filters.
Signed-off-by: Sunil Kumar Kori <skori@marvell.com>
Kiran Kumar K [Wed, 23 Jun 2021 04:46:49 +0000 (10:16 +0530)]
net/cnxk: support flow API
Adding initial version of rte_flow support for cnxk family device.
Supported rte_flow ops are flow_validate, flow_create, flow_destroy,
flow_flush, flow_query, flow_isolate.
Signed-off-by: Kiran Kumar K <kirankumark@marvell.com>
Sunil Kumar Kori [Wed, 23 Jun 2021 04:46:41 +0000 (10:16 +0530)]
net/cnxk: support Rx interrupt
Application may choose to enable/disable interrupts on Rx queues
so that application can select its processing if no packets are
available on queues for a longer period.
Signed-off-by: Sunil Kumar Kori <skori@marvell.com>
Add device start and stop operation callbacks for
CN9K and CN10K. Device stop is common for both platforms
while device start as some platform dependent portion where
the platform specific offload flags are recomputed and
the right Rx/Tx burst function is chosen.
aDD tx queue setup and release for CN9K and CN10K.
Release is common while setup is platform dependent due
to differences in fast path Tx queue structures.
Add Rx queue setup and release op for CN9K and CN10K
SoC. Release is completely common while setup is platform
dependent due to fast path Rx queue structure variation.
Fastpath is platform dependent partly due to core cacheline
size difference.
Add platform specific probe and remove callbacks for CN9K
and CN10K which use common probe and remove functions.
Register ethdev driver for CN9K and CN10K.
Harman Kalra [Wed, 23 Jun 2021 04:46:05 +0000 (10:16 +0530)]
common/cnxk: allocate LMT region in userspace
As per the new LMTST design, userspace shall allocate LMT region,
setup the DMA translation and share the IOVA with kernel via MBOX.
Kernel will convert this IOVA to physical memory and update the
LMT table entry with the same.
With this new design also shared mode (i.e. all pci funcs sharing
the LMT region allocated by primary/base pci func) is intact.
Signed-off-by: Harman Kalra <hkalra@marvell.com> Reviewed-by: Jerin Jacob <jerinj@marvell.com>
git.droids-corp.org / dpdk.git / log