Nalla Pradeep [Fri, 29 Jan 2021 12:45:09 +0000 (04:45 -0800)]
net/octeontx_ep: add Rx path
Function to deliver packets from DROQ to application is added. It also
fills DROQ with receive buffers timely such that device can fill them
with incoming packets.
Nalla Pradeep [Fri, 29 Jan 2021 12:45:07 +0000 (04:45 -0800)]
net/octeontx_ep: set up IQ and OQ registers
Configuring hardware registers with command queue (IQ) and driver output
queue (OQ) parameters.
List of parameters configured for IQ after making sure it is idle
1. Base address
2. Instruction size
3. Disabling interrupts for fastpath
List of parameters configured for OQ after making sure it is idle
1. Base address
2. Output buffer size
3. Clear output queue doorbell
4. Disable interrupts for fastpath
Nalla Pradeep [Fri, 29 Jan 2021 12:45:06 +0000 (04:45 -0800)]
net/octeontx_ep: add Tx queue setup and release
Transmit queue setup involves allocating memory for the command queue
considering tx descriptor count and initializing data structure
representing the queue. Transmit queue release function frees the
command queue.
Nalla Pradeep [Fri, 29 Jan 2021 12:45:05 +0000 (04:45 -0800)]
net/octeontx_ep: add Rx queue setup and release
Receive queue setup involves allocating memory for the queue,
initializing data structure representing the queue and filling queue
with receive buffers of rx descriptor count. Receive queues are referred
as droq. Hardware fills the receive buffers in queue with the packet.
In receive queue release, receive buffers are freed along with the
receive queue.
Igor Chauskin [Tue, 26 Jan 2021 18:32:26 +0000 (19:32 +0100)]
net/ena: prevent double doorbell
Add per-tx-ring flag for packets that were pushed to HW but await
doorbell. That is to prevent a situation when a doorbell is sent due to
reaching Tx burst threshold and next send fails (e.g., due to queue
full). In such case we shouldn't send another doorbell because there are
no actual packets waiting for transmission.
Fixes: c7519ea5eb8d ("net/ena: call additional doorbells if needed") Cc: stable@dpdk.org Signed-off-by: Igor Chauskin <igorch@amazon.com> Reviewed-by: Michal Krawczyk <mk@semihalf.com>
Igor Chauskin [Tue, 26 Jan 2021 18:32:25 +0000 (19:32 +0100)]
net/ena: fix Tx SQ free space assessment
Before starting transmission of Tx burst, the driver checked the
available space in the sq and limited the number of packets for
transmission accordingly.
The calculation was incorrect for fragmented packets and potentially had
significantly limited the length of Tx bursts.
This patch removes the assessment and pushes packets to the sq as long
as the burst is not exhausted and space is available in the sq.
Correct evaluation of the required space isn't possible before the burst
because it depends on the number of segments of each packet.
This patch adds per-packet space evaluation for each packet before
attempting to process it. In case there is not enough queue space, the
burst will just stop without error.
Fixes: 2061fe41f212 ("net/ena: linearize Tx mbuf") Cc: stable@dpdk.org Signed-off-by: Igor Chauskin <igorch@amazon.com> Reviewed-by: Michal Krawczyk <mk@semihalf.com>
Michal Krawczyk [Tue, 26 Jan 2021 18:32:24 +0000 (19:32 +0100)]
net/ena: validate Rx req ID upon acquiring descriptor
Instead of verifying the Rx descriptor each time it's being used in the
driver code, now the verification happens on the HAL side.
This simplifies code a lot as instead of doing 2 validations, only
single one is needed. The driver have to check the rc value returned
by the ena_com upon reading the Rx descriptor and trigger the reset
if needed. It was previously the responsibility of the
validate_rx_req_id() function.
As part of the change, the version of the driver was bumped to v2.2.1.
Fixes: 2061fe41f212 ("net/ena: linearize Tx mbuf") Cc: stable@dpdk.org Signed-off-by: Ido Segev <idose@amazon.com> Signed-off-by: Michal Krawczyk <mk@semihalf.com> Reviewed-by: Igor Chauskin <igorch@amazon.com>
Heinrich Kuhn [Mon, 25 Jan 2021 15:25:44 +0000 (17:25 +0200)]
net/nfp: read chip model from PluDevice register
For newer smartNIC NVRAM versions the chip model should be read from the
PluDevice register as it provides the authoritative chip model/revision.
This method of reading the chip model is backwards compatible with
legacy NVRAM versions too.
Since the model number is purely used for reporting purposes, follow the
hardware team convention of subtracting 0x10 from the PluDevice register
to obtain the chip model/revision number.
Fixes: c7e9729da6b5 ("net/nfp: support CPP") Cc: stable@dpdk.org Signed-off-by: Heinrich Kuhn <heinrich.kuhn@netronome.com> Signed-off-by: Simon Horman <simon.horman@netronome.com> Reviewed-by: Louis Peens <louis.peens@netronome.com>
Igor Russkikh [Thu, 21 Jan 2021 18:05:30 +0000 (19:05 +0100)]
app/testpmd: support sending cloned packets in flowgen
When testing high performance numbers, it is often that CPU performance
limits the max values device can reach (both in pps and in gbps)
Here instead of recreating each packet separately, we use clones counter
to resend the same mbuf to the line multiple times.
PMDs handle that transparently due to reference counting inside of mbuf.
Reaching max PPS on small packet sizes helps here:
Some data from our 2 port x 50G device. Using 2*6 tx queues, 64b packets,
PowerEdge R7525, AMD EPYC 7452:
Steve Yang [Thu, 28 Jan 2021 12:07:08 +0000 (12:07 +0000)]
app/testpmd: fix setting maximum packet length
"port config all max-pkt-len" command fails because it doesn't set the
'DEV_RX_OFFLOAD_JUMBO_FRAME' offload flag properly.
Commit in the fixes line moved the 'DEV_RX_OFFLOAD_JUMBO_FRAME' offload
flag update from 'cmd_config_max_pkt_len_parsed()' to 'init_config()'.
'init_config()' function is only called during testpmd startup, but the
flag status needs to be calculated whenever 'max_rx_pkt_len' changes.
The issue can be reproduced as [1], where the 'max-pkt-len' reduced and
'DEV_RX_OFFLOAD_JUMBO_FRAME' offload flag should be cleared but it
didn't.
Adding the 'update_jumbo_frame_offload()' helper function to update
'DEV_RX_OFFLOAD_JUMBO_FRAME' offload flag and 'max_rx_pkt_len'. This
function is called both by 'init_config()' and
'cmd_config_max_pkt_len_parsed()'.
Default 'max-pkt-len' value set to zero, 'update_jumbo_frame_offload()'
updates it to "RTE_ETHER_MTU + PMD specific Ethernet overhead" when it
is zero.
If '--max-pkt-len=N' argument provided, it will be used instead.
And with each "port config all max-pkt-len" command, the
'DEV_RX_OFFLOAD_JUMBO_FRAME' offload flag, 'max-pkt-len' and MTU is
updated.
[1]
--------------------------------------------------------------------------
dpdk-testpmd -c 0xf -n 4 -- -i --max-pkt-len=9000 --tx-offloads=0x8000
--rxq=4 --txq=4 --disable-rss
testpmd> set verbose 3
testpmd> port stop all
testpmd> port config all max-pkt-len 1518
testpmd> port start all
// Got fail error info without this patch
Configuring Port 0 (socket 1)
Ethdev port_id=0 rx_queue_id=0, new added offloads 0x800 must be
within per-queue offload capabilities 0x0 in rte_eth_rx_queue_setup()
Fail to configure port 0 rx queues //<-- Fail error info;
--------------------------------------------------------------------------
Bugzilla ID: 625 Fixes: 761c4d66900f ("app/testpmd: fix max Rx packet length for VLAN packets") Cc: stable@dpdk.org Signed-off-by: Steve Yang <stevex.yang@intel.com> Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com> Acked-by: Lance Richardson <lance.richardson@broadcom.com> Acked-by: Wisam Jaddo <wisamm@nvidia.com> Acked-by: Xiaoyun Li <xiaoyun.li@intel.com> Tested-by: Bo Chen <box.c.chen@intel.com>
Chengchang Tang [Fri, 22 Jan 2021 10:18:47 +0000 (18:18 +0800)]
net/hns3: fix interrupt resources in Rx interrupt mode
For Kunpeng930, the NIC engine support 1280 tqps being taken over by
a PF. In this case, a maximum of 1281 interrupt resources are also
supported in this PF. To support the maximum number of queues, several
patches are made. But the interrupt related modification are missing.
So, in RX interrupt mode, a large number of queues will be aggregated
into one interrupt due to insufficient interrupts. It will lead to
waste of interrupt resources and reduces usability.
To utilize all these interrupt resources, related IMP command has been
extended. And, the I/O address of the extended interrupt resources are
different from the existing ones. So, a function used for calculating
the address offset has been added.
Lijun Ou [Fri, 22 Jan 2021 10:18:45 +0000 (18:18 +0800)]
net/hns3: refactor reset event report function
Here encapsulate the process code of the imp reset report and
global reset report into function in order to reduce the
complexity of the hns3_check_event_cause function.
Lijun Ou [Fri, 22 Jan 2021 10:18:44 +0000 (18:18 +0800)]
net/hns3: extract common checks for flow director
When parse flow director with all types, it needs to judge the spec
of item and mask of item for all packet types. The judgement is the
same for all types. Therefore, we move it into the concentrated
location.
Lijun Ou [Fri, 22 Jan 2021 10:18:42 +0000 (18:18 +0800)]
net/hns3: refactor flow checks into own functions
Here moves some judgement conditions to a separated function
for parsing IPv4 hdr and TCP hdr in hns3_parse_normal function.
Also, move the check of the selected input tuple of RSS to a
separated functions named hns3_rss_input_tuple_supported
in order to enhance scalability and complexity.
Huisong Li [Fri, 22 Jan 2021 10:18:40 +0000 (18:18 +0800)]
net/hns3: move queue stats to xstats
One of the hot discussions in community recently was moving queue stats
to xstats. In this solution, a temporary
'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' device flag is created to implement
the smooth switch. And the first half of this work has been completed in
the ethdev framework. Now driver needs to remove the flag from the
driver initialization process and does the rest of work.
For better readability and reasonability, per-queue stats also should be
cleared when rte_eth_stats is cleared. Otherwise, the sum of one item in
per-queue stats may be greater than corresponding item in rte_eth_stats.
Signed-off-by: Huisong Li <lihuisong@huawei.com> Signed-off-by: Lijun Ou <oulijun@huawei.com>
Huisong Li [Fri, 22 Jan 2021 10:18:39 +0000 (18:18 +0800)]
net/hns3: encapsulate DFX stats in datapath
pkt_len_errors and l2_errors in Rx datapath indicate that driver
needs to discard received packets. And driver does not discard
packets for l3/l4/ol3/ol4_csum_errors in Rx datapath and others
stats in Tx datapath. Therefore, it is necessary for improving
code readability and maintainability to encapsulate error stats
and dfx stats.
Signed-off-by: Huisong Li <lihuisong@huawei.com> Signed-off-by: Lijun Ou <oulijun@huawei.com>
Jiawei Wang [Tue, 26 Jan 2021 03:49:42 +0000 (05:49 +0200)]
app/testpmd: fix sample action for RSS with no key
When the RSS with null key was set in sample actions list, it
caused the segmentation fault since the RSS key pointer was
NULL while did the memory copy.
This patch adds the RSS key NULL pointer checking before copying
to fix the segmentation fault issue.
Fixes: 11b1b0eddade ("app/testpmd: support RSS in sample action") Signed-off-by: Jiawei Wang <jiaweiw@nvidia.com> Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Dapeng Yu [Tue, 26 Jan 2021 09:52:11 +0000 (17:52 +0800)]
net/i40e: fix register setting for hash enable
The original code causes wrong value to be set into PFQF_HENA
register because unnecessary calling to get translated pctype
value for X722 NIC. As a result RSS is not working.
In order to optimize number of hops in the steering tables there
was the hint field of source vport in the matcher. If this hint
was provided for the wire port the actual vport match, used by
kernel, was not set in the matcher metadata register field.
This could led to not creating the rule in hardware and E-Switch
malfunction if kernel used non-zero metadata value for the wire
vport.
This patch always sets the match for the vport metadata register
value besides the hint.
Add support for new MODIFY_FIELD action to the Mellanox PMD.
This is the generic API that allows to manipulate any packet
header field by copying data from another packet field or
mark, metadata, tag, or immediate value (or pointer to it).
Since the API is generic and covers a lot of action under its
umbrella it makes sense to implement all the mechanics gradually
in order to move to this API for any packet field manipulations
in the future. This is the first step of RTE flows consolidation.
The modify field RTE flow action supports three operations: set,
add and sub. This patch brings to live only the "set" operation.
Support is provided for any packet header field as well as
meta/tag/mark and immediate value can be used as a source.
There are few limitations for this first version of API support:
- encapsulation levels are not supported, just outermost header
can be manipulated for now.
- offsets can only be 4-bytes aligned: 32, 64 and 96 for IPv6.
- the special ITEM_START ID is not supported as we do not allow
to cross packet header field boundaries yet.
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com> Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Jiawei Wang [Tue, 26 Jan 2021 15:37:56 +0000 (17:37 +0200)]
net/mlx5: fix count actions query in sample flow
If the count action was presented in sample actions list, MLX5 PMD
created the counter resource and saved the index of counter in the
sample resource only, the counter index of flow was not updated.
This patch removes the counter index in the sampler resource and
saves it into the flow, and adds the checking to make sure only one
count be created once per flow, it's used for all sample sub flows.
Fixes: 0756228b2704 ("net/mlx5: update translate function for sample action") Cc: stable@dpdk.org Signed-off-by: Jiawei Wang <jiaweiw@nvidia.com> Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Liron Himi [Wed, 27 Jan 2021 16:09:38 +0000 (18:09 +0200)]
net/mvpp2: introduce fixup for FIFO overrun
Currently the HW is configured with only one pool which its
buffer size may be larger than the rx-fifo-size.
In that situation, frame size larger than the fifo-size
is gets dropped due to fifo overrun.
This is cause because the HW works in cut-through mode which
waits to have in the fifo at least the amount of bytes as define
in the smallest pool's buffer size.
This patch add a dummy pool which its buffer size
is very small (smaller than 64B frame). This tricks the HW and
any frame size is gets passed from the FIFO to the PP2.
Liron Himi [Wed, 27 Jan 2021 16:09:32 +0000 (18:09 +0200)]
net/mvpp2: build flow table key along with rule
Currently the key is always built in network order
regardless of user input.
Besides the fact that it is not in user order,
it prevent future support for udf header between standard headers.
This patch builds the key while parsing the given rule.
Liron Himi [Wed, 27 Jan 2021 16:09:31 +0000 (18:09 +0200)]
net/mvpp2: support generic flow pattern combinations
Currently only specific pattern combinations are supported.
This makes it hard to support additional pattern.
In addition there is no a real limitation that prevent any combination.
This patch iterate the input patterns and convert them to a mvpp2 API.
The doc doesn't describes those limitation so
just update the action list
Liron Himi [Wed, 27 Jan 2021 16:09:16 +0000 (18:09 +0200)]
net/mvpp2: remove debug log on fast-path
In case of non-ip frame the current code reached the 'default'
case which result with function call to log a msg.
Those kind of calls should not be performed on fast-path.
The performance for this kind of frames increased by 50%
Heinrich Kuhn [Tue, 19 Jan 2021 11:56:17 +0000 (13:56 +0200)]
net/nfp: free port private data on close
Free the private data of a port when the .dev_close() callback is
invoked. For NFP6000/4000 devices multiple ports may exist under a
single PF device. In this situation the PF resources will only be freed
when all the ports associated with the PF has been freed too.
PF hot plugging isn't explicitly supported for NFP6000/4000 devices but
all the private data of all the ports under the PF in question will be
freed upon device removal.
Signed-off-by: Heinrich Kuhn <heinrich.kuhn@netronome.com> Signed-off-by: Simon Horman <simon.horman@netronome.com> Reviewed-by: Louis Peens <louis.peens@netronome.com>
Heinrich Kuhn [Tue, 19 Jan 2021 11:56:16 +0000 (13:56 +0200)]
net/nfp: create separate entity for PF device
Before this change memory for the private data of all physical ports
where allocated with single rte_zmalloc() call. Specific port private
data was accessed by means of an offset into this memory. This scheme is
problematic when attempting to free only one port's private data at a
time.
To address this, a new entity is created called struct nfp_pf_dev. This
struct represents the PF device. It has a number of PF specific members.
Notably it has a pointer of type rte_eth_dev that points to the eth_dev
associated with the first physical port of the device. It also has an
array of nfp_net_hw's containing pointers to all the physical ports
under the PF.
Memory is first allocated for the PF and PF specific initialization is
attempted. Next, all the physical ports under the PF is iterated and
memory is allocated separately for the private data of each port. Port 0
is skipped during this phase because memory has already been allocated
and an eth_dev already exits for the 0th port.
Signed-off-by: Heinrich Kuhn <heinrich.kuhn@netronome.com> Signed-off-by: Simon Horman <simon.horman@netronome.com> Reviewed-by: Louis Peens <louis.peens@netronome.com>
Andrew Boyer [Mon, 18 Jan 2021 20:34:58 +0000 (12:34 -0800)]
net/ionic: observe endianness in Rx filter
The IONIC firmware is little-endian.
Add a new field to struct ionic_rx_filter to store the CPU-endian
match type.
Use a local variable for the VLAN when searching the hash table.
Matan Azrad [Wed, 6 Jan 2021 06:43:29 +0000 (06:43 +0000)]
vdpa/mlx5: fix configuration mutex cleanup
When the vDPA device is closed, the driver polling thread is canceled.
The polling thread locks the configuration mutex while it polls the CQs.
When the cancellation happens, it may terminate the thread inside the
critical section what remains the configuration mutex locked.
After device close, the driver may be configured again, in this case,
for example, when the first queue state is updated, the driver tries to
lock the mutex again and deadlock appears.
Initialize the mutex after the polling thread cancellation.
Joyce Kong [Tue, 26 Jan 2021 09:57:43 +0000 (17:57 +0800)]
net/virtio: fix Arm build with old compiler
In file included from ../drivers/net/virtio/virtio_rxtx_packed.c:22:0:
../drivers/net/virtio/virtio_rxtx_packed_neon.h:
In function ‘virtqueue_enqueue_batch_packed_vec’:
../drivers/net/virtio/virtio_rxtx_packed_neon.h:74:2: warning:
implicit declaration of function ‘vreinterpretq_p128_u32’
[-Wimplicit-function-declaration]
poly128_t cmp1 = vreinterpretq_p128_u32(~vceqq_u32(ref_seg, def_ref_seg));
The message shows ‘vreinterpretq_p128_u32’ intrinsic is not supported
because an old version gcc (gcc 4.8.5) is used. So fix the issue with
implementing the logic with other intrinsics.
Bugzilla ID: 621 Fixes: 530887469350 ("net/virtio: add vectorized packed ring NEON Tx") Fixes: 5971ce5e2a59 ("net/virtio: add vectorized packed ring NEON Rx") Reported-by: Subhi Masri <subhim@nvidia.com> Signed-off-by: Joyce Kong <joyce.kong@arm.com> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com> Tested-by: Subhi Masri <subhim@nvidia.com>
Cheng Jiang [Tue, 12 Jan 2021 04:38:57 +0000 (04:38 +0000)]
examples/vhost: refactor vhost data path
Change the vm2vm data path to batch enqueue for better performance.
Support latest async vhost API, refactor vhost async data path,
replace rte_atomicNN_xxx to __atomic_XXX and clean some codes.
Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com> Reviewed-by: Jiayu Hu <jiayu.hu@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Cheng Jiang [Tue, 12 Jan 2021 04:38:56 +0000 (04:38 +0000)]
examples/vhost: add ioat ring space count and check
Add ioat ring space count and check, if ioat ring space is not enough
for the next async vhost packet enqueue, then just return to prevent
enqueue failure. Add rte_ioat_completed_ops() fail handler.
Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com> Reviewed-by: Jiayu Hu <jiayu.hu@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch fixes virtio_user_dev_setup() error path,
by cleaning all resources it allocates. It introduces
virtio_user_dev_uninit_notify() that cleans all open
FDs. It implies assigning all FDs to -1 at init time.
With these changes done, virtio_user_dev_init_notify()
can be simplified.
Suggested-by: Adrian Moreno <amorenoz@redhat.com> Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com> Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Maxime Coquelin [Tue, 26 Jan 2021 10:16:32 +0000 (11:16 +0100)]
net/virtio: make server mode blocking
This patch makes the Vhost-user backend server mode
blocking at init, waiting for the client connection.
The goal is to make the driver more reliable, as without
waiting for client connection, the Virtio driver has to
assume the Vhost-user backend will support all the
features it has advertized, which could lead to undefined
behaviour.
For example, without this patch, if the user enables packed
ring Virtio feature but the backend does not support it,
the ring initialized by the driver will not be compatible
with the backend.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com> Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
git.droids-corp.org / dpdk.git / log