After Rx or Tx cleanup update completion queue head by calling
ena_com_update_dev_comp_head().
Fixes: 1daff5260f ("net/ena: use unmasked head and tail")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
The Rx drops cannot be acquired using the older API. Now, it must be
read in keep alive message.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
ENA PMD is having it's own custom statistics counters. They are exposed
to the application by using the xstats DPDK API.
The deprecated and unused statistics are removed, together with old API.
Signed-off-by: Solganik Alexander <sashas@lightbitslabs.com>
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
Those counters provide information regards sent/received bytes and
packets per queue.
Signed-off-by: Solganik Alexander <sashas@lightbitslabs.com>
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
When wrong req_id is detected some previous mbufs could be used for
receiving different segments of received packets. In such cases chained
mbufs will be twice returned to pool.
To prevent it chained mbuf is now freed just after error detection.
To simplify cleaning, pointers taken for Rx ring are set to NULL.
As after ena_rx_queue_release_bufs and ena_tx_queue_release_bufs queues
are not used updating of next_to_clean pointer is not necessary.
Fixes: c203497667 ("net/ena: add Rx out of order completion")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Use empty_rx_reqs instead of empty_tx_reqs.
As those two variables are part of union this not cause
any failure, but for consistency should be changed.
Fixes: c203497667 ("net/ena: add Rx out of order completion")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
The PMD was not passing RSS offloads values although it was supporting
the RSS. To allow application to probe the PMD for RSS support, the
missing information was added.
Fixes: 1173fca25a ("ena: add polling-mode driver")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Only PMD_*_LOG is adding new line character to the log message.
All printouts were adjusted for consistency.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Reset function should return the port to initial state, in which no Tx
and Rx queues are setup. Then application should reconfigure the queues.
According to DPDK documentation the rte_eth_dev_reset() itself is a
generic function which only does some hardware reset operations through
calling dev_unint() and dev_init().
ena_com_dev_reset which perform NIC registers reset should be called
during stop.
Fixes: 2081d5e2e9 ("net/ena: add reset routine")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
If start function fails, previously created queues have to be removed.
ena_queue_restart_all() and ena_queue_restart() are renamed to
ena_queue_start_all() and ena_queue_start().
ena_free_io_queues_all() is renamed to ena_queue_stop_all().
Fixes: df238f84c0 ("net/ena: recreate HW IO rings on start and stop")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Before sending next packet, check if calling doorbell is needed.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Some of ENA devices supports 8k Rx rings. Maximum supported size is
received upon device initialization.
As ENA_DEFAULT_RING_SIZE_RX macro is upper limit, it needs to be
adjusted.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
LLQ (Low Latency Queue) is the feature that allows pushing header
directly to the device through PCI before even DMA is triggered.
It reduces latency, because device can start preparing packet before
payload is sent through DMA.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
When invalid req_id is received, the reset should be handled by the
application, as it is indicating invalid rings state, so further Rx
is not making any sense.
Fixes: c203497667 ("net/ena: add Rx out of order completion")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
The device now allows driver to reconfigure Tx and Rx queues depth
independently. Moreover, maximum size for Tx and Rx can be different.
Those maximum values are received from the device.
After reset, previous ring configuration is restored.
If number of descriptor is set to RTE_ETH_DEV_FALLBACK_RX_RINGSIZE
or RTE_ETH_DEV_FALLBACK_TX_RINGSIZE, the maximum value is restored.
Remove checks, if provided number is not too big, as this is done in
generic functions (rte_eth_rx_queue_setup and rte_eth_tx_queue_setup).
Maximum number of segments is being set for Rx packets and provided to
ena_com_rx_pkt() for validation.
Unused definitions were removed.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Whenever the driver will receive too many descriptors from the device,
it should trigger the device reset with reset reason set to
ENA_REGS_RESET_TOO_MANY_RX_DESCS.
Fixes: 241da076b1 ("net/ena: adjust error checking and cleaning")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
The new ena_com allows the number of CPUs to be passed to the device in
the host info structure.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
ena_com is the communication layer provided by the vendor and common to
all ENA drivers.
This patch updates it to version from 2018.09.26.
It adds support for ENAv2 device together with LLQ feature, adds
doorbell optimization and reconfiguration of HW queues depth
independently.
The driver was adjusted to the new changes in the HAL.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
Since below commit, several tx_prep functions are broken, they fail to
pass supported Tx offload features check:
PKT_TX_IPVx must be set when any PKT_TX_L4 checksum is requested,
but these values are not present in the mask of supported Tx offloads
of several drivers that advertise PKT_TX_L4_MASK.
So any packet sent to those drivers with a L4 checksum request and
one of PKT_TX_IPVx bit set is rejected by the tx prepare function.
Fixes: 1037ed842c ("mbuf: fix Tx offload mask")
Cc: stable@dpdk.org
Signed-off-by: Didier Pallard <didier.pallard@6wind.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
rx_buffer_info should be refill not linearly, but out of order.
IDs should be taken from empty_rx_reqs array.
rx_refill_buffer is introduced to temporary storage
bulk of mbufs taken from pool.
In case of error unused mbufs are put back to pool.
Fixes: c203497667 ("net/ena: add Rx out of order completion")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
When queues are stopped release Tx buffers.
During start initialize array of empty Tx/Rx reqs with default values.
Fixes: df238f84c0 ("net/ena: recreate HW IO rings on start and stop")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
The ethdev flag RTE_ETH_DEV_CLOSE_REMOVE is set for drivers
having migrated to the new behaviour of rte_eth_dev_close().
As any other flag, it can be useful to know about its value
as soon as the port is probed.
Unfortunately, it was set inside the close operation,
just before being erased by memset() in rte_eth_dev_release_port().
The flag assignment is moved to the probing stage, so it can
be checked by the application in order to anticipate the behaviour.
Fixes: 42603bbdb5 ("net/mlx5: release port on close")
Fixes: 6c99085d97 ("net/vmxnet3: fix hot-unplug")
Fixes: 4d7877fde2 ("net/ena: remove resources when port is being closed")
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Luca Boccassi <bluca@debian.org>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
The new API introduced in 18.11 is suggesting, that the driver should
release all it's resources at the dev_close routine.
All resources previously released in uninit routine during PCI removal,
are now being released at the dev_close and the PMD is indicating that
it is supporting API changes by setting RTE_ETH_DEV_CLOSE_REMOVE flag.
As the device is not allocating MAC addresses dynamically, it is setting
mac_addrs field to NULL, so it wouldn't be released by the
rte_eth_dev_release_port().
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
The driver was passing to the mbuf Rx queue ID instead of hash received
from the device. Now, the RSS hash from the Rx descriptor is being set.
Fixes: 1173fca25a ("ena: add polling-mode driver")
Cc: stable@dpdk.org
Signed-off-by: Stewart Allen <allenste@amazon.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
On the start the driver was refilling all Rx buffs, but the old ones
were not released. That way running start/stop for a few times was
causing device to run out of descriptors.
To fix the issue, IO rings are now being destroyed on stop, and
recreated on start. That way the device is not losing any descriptors.
Furthermore, there was also memory leak for the Rx mbufs, which were
created on start and not destroyed on stop.
Fixes: eb0ef49dd5 ("net/ena: add stop and uninit routines")
Cc: stable@dpdk.org
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
With the enabling for hotplug on multi-process,
rte_eth_dev_pci_generic_remove can be used to detach the device from
a secondary process also. But we need to take care of the uninit callback
parameter to make sure it handles the secondary case correctly.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
A constructor is usually declared with RTE_INIT* macros.
As it is a static function, no need to declare before its definition.
The macro is used directly in the function definition.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
When the number of rx queues is 0 (what can be when application does
not receive) failed with SIGFPE.
It happens when the application is also requesting ETH_MQ_RX_RSS_FLAG
in the rte_dev->data->dev_conf.rxmode.mq_mode.
Fixed adding zero rx queues check.
Signed-off-by: Daria Kolistratova <daria.kolistratova@intel.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Write combining (WC) increases NIC performance by making better
utilization of PCI bus. ENA PMD may make usage of this feature.
To enable it load igb_uio driver with wc_activate set to 1.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Link speed should is not limited to 10Gb/s and it shouldn't be hardcoded.
They link speed is set to none instead and the applications shouldn't
rely on this value when using ENA PMD.
Fixes: 1173fca25a ("ena: add polling-mode driver")
Cc: stable@dpdk.org
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
During initializing Tx queues update Non-Uniform Memory Access
configuration in NIC firmware.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Adjust error checking and cleaning to Linux driver:
* add checking if MTU is to small,
* fix error messages (mismatched Rx and Tx),
* return error received from base driver or proper error
code instead of -1,
* in case of error release occupied resources,
* in case of Rx error trigger NIC reset.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Validate Tx req id during clearing completed packets.
If id is wrong, trigger NIC reset.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Move configuration of IO queue numbers to separate function and take
into consideration max number of IO completion queues.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Enable AENQ FATAL_ERROR and WARNING callbacks by setting flags
in aenq_groups. They are handled by "unimplemented handler".
If unimplemented handler is called, error is logged.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
In function ena_infos_get driver provides information about minimal
and maximal number of Rx and Tx descriptors.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Function ena_check_and_linearize_mbuf check Tx mbuf for number of
segments and linearize (defragment) it if necessary. It is called
before sending each packet.
Information about maximum number of segments is stored per each ring.
Maximum number of segments supported by NIC is taken from ENA COM in
ena_calc_queue_size function and stored in adapter structure.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
Add variable wd_state to make driver functional without keep alive
AENQ handler.
The watchdog will be executed only if the aenq group has keep alive
enabled.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
The admin queue can stop responding or became inactive due to unexpected
behaviour of the device. In that case, the whole device should be
restarted.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
Keep alive is executing AENQ interrupt periodically. It allows to check
health of the device and trigger reset event if the device will stop
responding.
To check for the state of the device, the DPDK application must call
rte_timer_manage().
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
Reset routine can be used by the DPDK application to reset the device in
case of receiving RTE_ETH_EVENT_INTR_RESET from the PMD.
The reset event is not triggered by the driver, yet. It will be added in
next commits to enable error recovery in case of misfunctioning of the
device.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
There is no need to check for restart all queues. It is sufficient to
check only previously initialized queues.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
When ENA notifications are provided ena_notification handler is called.
It checks if received value is not corrupted and if necessary it
reports proper warnings.
Data received from NIC is parsed in ena_update_hints. Fields for
storing those information was added to ena_adapter structure.
ENA notification are enabled by setting ENA_ADMIN_NOTIFICATION flag in
aenq_groups.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Acked-by: Michal Krawczyk <mk@semihalf.com>
To make the LSC interrupt working, the AENQ must be configured properly
in the ENA device. The AENQ interrupt is common for all maintenance
interrupts - the proper handler is then executed depending on the
received descriptor.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
Lack of uninit routine could lead to memory leak.
Stop was added to fulfill allowed PMD operations.
Checks for the PMD states in the start and configure routine were
removed, as the upper layer is already checking for them.
The interrupt setup was moved from start to init function.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
The ENA device is able to send MSI-X when it will complete an command
when polling mode is deactivated. Further, the same interrupt handler
will be used for the AENQ handling - services of the ENA device, that
allows to implement watchdog or LSC handler.
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
The legacy LLQ should no longer be supported by the drivers, as this API
is deprecated. Because of that, it was removed from the driver.
Signed-off-by: Rafal Kozik <rk@semihalf.com>
Signed-off-by: Michal Krawczyk <mk@semihalf.com>
