Modify reader/writer lock to avoid starvation of writer. The previous
implementation would cause a writer to get starved if readers kept
acquiring the lock. The new version uses an additional bit to indicate
that a writer is waiting and which keeps readers from starving the
writer.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Morten Brørup <mb@smartsharesystems.com>
DPDK libraries should never call rte_exit on failure, so change the
function return type of rte_metrics_init to "int" to allow returning an
error code to the application rather than exiting the whole app on init
failure.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Morten Brørup <mb@smartsharesystems.com>
Library functions should not cause the app to exit or panic. Replace the
existing panic call in the EAL remote launch functions with an error
code return instead.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Deprecation notice targeted for v22.11 of event vector has been
merged in the following commits, remove deprecation notices.
Fixes: 0fbb55efa5 ("eventdev: add element offset to event vector")
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Replace *u64s with u64s in rte_event_vector structure as
the *ptrs already serves the purpose of holding pointers
and the intention of u64s is to hold array of uint64_t
values.
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Add `rte` prefix to stop flush callback function pointer
declaration to avoid conflicts with application functions,
``eventdev_stop_flush_t`` is renamed to
``rte_eventdev_stop_flush_t``.
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
This patch addresses an issue of enqueuing more than
max_enq_depth and not able to dequeuing events equal
to max_cq_depth in a single call of rte_event_enqueue_burst
and rte_event_dequeue_burst.
Apply fix for restricting enqueue of events to max_enq_depth
so that in a single rte_event_enqueue_burst() call at most
max_enq_depth events are enqueued.
Also set per port and domain history list sizes based on
cq_depth. This results in dequeuing correct number of
events as set by max_cq_depth.
Fixes: f3cad285bb ("event/dlb2: add infos get and configure")
Cc: stable@dpdk.org
Signed-off-by: Abdullah Sevincer <abdullah.sevincer@intel.com>
This commit implements the changes required for using suggested
port type hint feature. Each port uses different credit quanta
based on port type specified using port configuration flags.
Each port has separate quanta defined in dlb2_priv.h
Producer and consumer ports will need larger quanta value to reduce number
of credit calls they make. Workers can use small quanta as they mostly
work out of locally cached credits and don't request/return credits often.
Signed-off-by: Abdullah Sevincer <abdullah.sevincer@intel.com>
If producer thread is only acting as a bridge between NIC and DLB, then
performance can be greatly improved by bypassing the fence instruction.
DLB enqueue API calls memory fence once per enqueue burst. If producer
thread is just reading from NIC and sending to DLB without updating
the read buffers or buffer headers OR producer is not writing
to data structures with dependencies on the enqueue write order, then
fencing can be safely disabled.
Signed-off-by: Abdullah Sevincer <abdullah.sevincer@intel.com>
For best performance, applications running on certain cores should use
the DLB device locally available on the same tile along with other
resources. To allocate optimal resources, probing is done for each
producer port (PP) for a given CPU and the best performing ports are
allocated to producers. The CPU used for probing is either the first
core of producer coremask (if present) or the second core of EAL
coremask. This will be extended later to probe for all CPUs in the
producer coremask or EAL coremask.
Producer coremask can be passed along with the BDF of the DLB devices.
"-a xx:y.z,producer_coremask=<core_mask>"
Applications also need to pass RTE_EVENT_PORT_CFG_HINT_PRODUCER during
rte_event_port_setup() for producer ports for optimal port allocation.
For optimal load balancing ports that map to one or more QIDs in common
should not be in numerical sequence. The port->QID mapping is application
dependent, but the driver interleaves port IDs as much as possible to
reduce the likelihood of sequential ports mapping to the same QID(s).
Hence, DLB uses an initial allocation of Port IDs to maximize the
average distance between an ID and its immediate neighbors. Using
the initialport allocation option can be passed through devarg
"default_port_allocation=y(or Y)".
When events are dropped by workers or consumers that use LDB ports,
completions are sent which are just ENQs and may impact the latency.
To address this, probing is done for LDB ports as well. Probing is
done on ports per 'cos'. When default cos is used, ports will be
allocated from best ports from the best 'cos', else from best ports of
the specific cos.
Signed-off-by: Abdullah Sevincer <abdullah.sevincer@intel.com>
During cleanup `rte_event_port_quiesce` should be called
irrespective of whether an event has been dequeued or not
to flush any prefetched events.
Fixes: 7da008df0c ("app/eventdev: use port quiescing")
Cc: stable@dpdk.org
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
The user has the right to supply pools with excessively large
buffers, regardless of the maximum supported Rx packet length
reported by the adapter. However, in this PMD, on EF10 boards,
a Rx descriptor has only 14 bits to specify the buffer length.
To avoid potential problems, use this information accordingly.
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Such information is useful to client drivers which deal with
large Rx pool buffers (16-bit wide data count) and thus need
to avoid overflow when setting EF10's 14-bit wide data count.
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Maximum data count of a Tx descriptor is advertised to users,
however, this value is mistakenly derived from the Rx define.
Use the Tx one instead. The resulting value will be the same.
Fixes: 1e43fe3cb4 ("net/sfc/base: separate limitations on Tx DMA descriptors")
Cc: stable@dpdk.org
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
As rte_flow_action_handle_create/destroy/update() have their own
asynchronous rte_flow_async_action_handle_create/destroy/update()
version functions to accelerate the indirect action operations in
queue based flow engine. Currently, the asynchronous version query
function for indirect action was missing.
Add rte_flow_async_action_handle_query() function corresponding
to rte_flow_action_handle_query(). The new asynchronous version
function enables enqueue the query to the hardware similar as
asynchronous flow management does and returns immediately to free
the CPU for other tasks. Application can get the query results from
rte_flow_pull() when the hardware completes its work.
Signed-off-by: Suanming Mou <suanmingm@nvidia.com>
Acked-by: Ori Kam <orika@nvidia.com>
In queue based async flow engine, in order to optimize the flow
insertion rate, PMD can use the hints from application to have
resources pre-allocate during initialization phase for actions
such as count/meter/aging.
This commit adds the connection tracking action hints.
Signed-off-by: Suanming Mou <suanmingm@nvidia.com>
Acked-by: Ori Kam <orika@nvidia.com>
Applications may use rte_eth_rx_queue_count() in the Rx stage of the
dataplane, so only check the function parameters if built with
RTE_ETHDEV_DEBUG_RX.
Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
Acked-by: Ferruh Yigit <ferruh.yigit@xilinx.com>
Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Introduce GTP PSC QFI modify field and add testpmd CLI command support.
An example for copying GTP QFI field using modify_field action:
modify_field op set dst_type meta src_type gtp_psc_qfi width 8
An example of setting GTP QFI field value to 0x1f using modify_field
action:
modify_field op set dst_type gtp_psc_qfi src_type value src_value 1f
width 8
Signed-off-by: Dariusz Sosnowski <dsosnowski@nvidia.com>
Acked-by: Ori Kam <orika@nvidia.com>
Saving the current segments of the packet, when the next segment data is
not ready.
Fixes: 965b3127d4 ("net/axgbe: support scattered Rx")
Cc: stable@dpdk.org
Signed-off-by: Bhagyada Modali <bhagyada.modali@amd.com>
Acked-by: Chandubabu Namburu <chandu@amd.com>
Reset the EOP in the failure scenario and also after the last segment.
Remove updating the packet length explicitly as it is done in chaining.
Fixes: 965b3127d4 ("net/axgbe: support scattered Rx")
Cc: stable@dpdk.org
Signed-off-by: Bhagyada Modali <bhagyada.modali@amd.com>
Acked-by: Chandubabu Namburu <chandu@amd.com>
Added changes to base rule install mechanism.
If action type is IPsec and multi channel is set,
then base rule will not be merged.
Signed-off-by: Kiran Kumar K <kirankumark@marvell.com>
Reviewed-by: Satheesh Paul <psatheesh@marvell.com>
In vector mode, DF bit is not programmed correctly, as the
return value of vsetq_lane_u64() is ignored, which actually
contains the updated value, leading HW to free mbufs though
NIX_TX_OFFLOAD_MBUF_NOFF_F flag is set.
Hence, save return value of vsetq_lane_u64() appropriately so
that DF bit is programmed correctly.
Fixes: 862e281287 ("net/cnxk: add vector Tx for CN9K")
Fixes: f71b7dbbf0 ("net/cnxk: add vector Tx for CN10K")
Cc: stable@dpdk.org
Signed-off-by: Hanumanth Pothula <hpothula@marvell.com>
While configuring NIX, local variables 'nb_rxq' and 'nb_txq'
are declared as 8bit variables, leading to an integer overflow
when an application sends Rxq/Txq value greater than 255.
Hence, declare local variables, 'nb_rxq' and 'nb_txq' as 16bit
variable. Also, during the cleanup, make sure PFC tree is not created.
Fixes: b75e0aca84 ("net/cnxk: add device configuration operation")
Cc: stable@dpdk.org
Signed-off-by: Hanumanth Pothula <hpothula@marvell.com>
If the length of data_len in mbuf is less than 17 or
greater than the maximum frame size, it is illegal.
These illegal packets will lead to Tx/Rx hang and
can't recover automatically.
This patch check those illegal packets and protect
Tx/Rx from hanging.
Fixes: 17c7d0f9d6 ("net/ice: support basic Rx/Tx")
Cc: stable@dpdk.org
Signed-off-by: Kevin Liu <kevinx.liu@intel.com>
Acked-by: Qi Zhang <qi.z.zhang@intel.com>
If the length of data_len in mbuf is less than 17 or
greater than the maximum frame size, it is illegal.
These illegal packets will lead to Tx/Rx hang and
can't recover automatically.
This patch check those illegal packets and protect
Tx/Rx from hanging.
Fixes: a2b29a7733 ("net/avf: enable basic Rx Tx")
Cc: stable@dpdk.org
Signed-off-by: Kevin Liu <kevinx.liu@intel.com>
Acked-by: Qi Zhang <qi.z.zhang@intel.com>
If a meter's profile is changed after meter disabled, there's an issue
that will fail when deleting the old profile.
This patch fixes this by adding the correct process to decrease the old
profile's reference count when changing profile.
Fixes: 63ffeb2ff2 ("net/mlx5: support meter profile update")
Cc: stable@dpdk.org
Signed-off-by: Shun Hao <shunh@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Currently, when flow usese meter hierarchy, a tag action is always applied
to set the first meter's meter id, so as to update the first meter's drop
count. But it's not considered if first meter doesn't have drop count.
This patch fixes it, that in hierarchy, if the first meter doesn't have
drop count, no need to add the meter id tag action. No change for
non-hierarchy meter.
Fixes: e8146c63 ("net/mlx5: support represented port item in flow rules")
Cc: stable@dpdk.org
Signed-off-by: Shun Hao <shunh@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
If there's no param in represented_port item, it will be treated as
matching all ports by default. But there's some limitation when using it
with meter hierarchy.
This patch adds the limitation that when matching all ports, the meter
hierarchy should not contain any meter having drop count.
Fixes: e8146c63 ("net/mlx5: support represented port item in flow rules")
Cc: stable@dpdk.org
Signed-off-by: Shun Hao <shunh@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
There is a new item type represented_port, and currently it will fail
when using meter hierarchy in flow using represented_port match.
This patch fixes this fail by adding support for represented_port item
in meter hierarchy flow split.
Fixes: e8146c63 ("net/mlx5: support represented port item in flow rules")
Cc: stable@dpdk.org
Signed-off-by: Shun Hao <shunh@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
The driver splits the flow with sample action into two sub-flows,
sub prefix flow and sub suffix flow.
In the case of tunnel flow including a decap action, the driver should
translate the inner as outer for actions coming after the decap action.
In the case of flow splitting, the packet layers, used to detect the
attributes, are inherited from the prefix flow to the suffix flow but
the driver wrongly didn't handle the decap adjustment and the inner
layers didn't shift to the outer.
This patch adjusts the inherited layers in case of decap.
Fixes: 6e77151286 ("net/mlx5: fix match information in meter")
Cc: stable@dpdk.org
Signed-off-by: Jiawei Wang <jiaweiw@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
If hardware descriptor (WQE) length exceeds one the HW can handle,
the Tx queue failure occurs. PMD does the length check but there was
a bug - the length limit was expressed in 16B units (WQEBB segments),
while the calculated WQE length and limit were in 64B units (WQEBBs).
Fix the condition to avoid subsequent Tx queue failure.
Fixes: 18a1c20 ("net/mlx5: implement Tx burst template")
Cc: stable@dpdk.org
Signed-off-by: Raja Zidane <rzidane@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The hardware descriptor (WQE) length field is 6 bits wide
and we have the native limitation for the overall descriptor
length. To improve the PCIe bandwidth the packet data can be
inline into descriptor. If PMD was configured to inline large
amount of data it happened there was no enough space remaining
in the descriptor to specify all the packet data segments and
PMD rejected problematic packets.
The patch tries to adjust the inline data length conservatively
and allows to avoid error occurring.
Fixes: 18a1c20044 ("net/mlx5: implement Tx burst template")
Fixes: e2259f93ef ("net/mlx5: fix Tx when inlining is impossible")
Cc: stable@dpdk.org
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Reviewed-by: Dmitry Kozlyuk <dkozlyuk@nvidia.com>
The mlx5 PMD can inline packet data into transmitting descriptor (WQE)
and free mbuf immediately as data no longer needed, for non-inline
packets the mbuf pointer should be stored in elts array for coming
freeing on send completion. There was an optimization on storing
pointers in batch and there was missed storing mbuf for single
packet if non-inline was explicitly requested by flag.
Fixes: cacb44a099 ("net/mlx5: add no-inline Tx flag")
Cc: stable@dpdk.org
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The mlx5 PMD supports send scheduling feature, it allows
to send packets at specified moment of time, to do that
PMD pushes special wait descriptor (WQE) to the hardware
queue and then pushes descriptor for packet data as usual.
If queue is close to be full or there is no enough elts
buffers to store mbufs being sent the data descriptors might
be not pushed and the orphan wait WQE (not followed by the
data) might reside in queue on tx_burst routine exit.
To avoid orphan wait WQEs there was the check for enough
free space in the queue WQE buffer and enough amount of the
free elts in queue mbuf storage. This check was incomplete
and did not cover all the cases for Enhanced Multi-Packet
Write descriptors.
Fixes: 2f827f5ea6 ("net/mlx5: support scheduling on send routine template")
Cc: stable@dpdk.org
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
ESP is one of IPSec protocols over both IPv4 and IPv6 and is considered
a tunnel layer that cannot be followed by any other layer. Taking that
into consideration, ESP is considered as a 4 layer.
Not defining ESP's priority will make it match with the same priority as
its prior IP layer, which has a layer 3 priority. This will lead to
issues in matching and will match the packet with the first matching
rule even if it doesn't have an esp layer in its pattern, disregarding
any following rules that could have an esp item and can be actually
a more accurate match since it will have a longer matching criterion.
This is fixed by defining the priority for the ESP item to have a
layer 4 priority, making the match be for the rule with the more
accurate and longer matching criteria.
Fixes: 18ca4a4ec7 ("net/mlx5: support ESP SPI match and RSS hash")
Cc: stable@dpdk.org
Signed-off-by: Bassam Zaid AlKilani <bzalkilani@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Acked-by: Raslan Darawsheh <rasland@nvidia.com>
One of the conditions to allow LRO offload is the DV configuration.
The function incorrectly checks the DV configuration before initializing
it by the user devarg; hence, LRO cannot be allowed.
This patch moves this check to mlx5_shared_dev_ctx_args_config, where DV
configuration is initialized.
Fixes: c4b8620135 ("net/mlx5: refactor to detect operation by DevX")
Cc: stable@dpdk.org
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Reported-by: Gal Shalom <galshalom@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Current PMD logs all DevX errors at error level.
DevX interface can fail queue counters allocation on some hardware
types. That is a known issue.
PMD fallbacks to Verbs API to allocate queue counters
when it detects the fault.
That DevX failure should not be logged as PMD error.
The patch provides DevX with flexible API that selects log level.
Signed-off-by: Gregory Etelson <getelson@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
FDs passed from rte_mp_msg are duplicated to the secondary process and
need to be closed.
Fixes: 9a8ab29b84 ("net/mlx5: replace IPC socket with EAL API")
Cc: stable@dpdk.org
Signed-off-by: Long Li <longli@microsoft.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
FDs passed from rte_mp_msg are duplicated to the secondary process and
need to be closed.
Fixes: 0203d33a10 ("net/mlx4: support secondary process")
Cc: stable@dpdk.org
Signed-off-by: Long Li <longli@microsoft.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
In case of higher order (greater than 99) logical cores, name was
truncated (length is restricted to 16 characters, including the
terminating null byte ('\0')) and it makes hard to follow threads.
Before this fix, this issue can be reproduced using following arguments:
--lcores=0,10@1,100@2
Then we had:
lcore-worker-10
lcore-worker-10
Signed-off-by: Abdullah Ömer Yamaç <omer.yamac@ceng.metu.edu.tr>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Checking a const pointer for alignment would emit a warning about the
const qualifier being discarded.
No need to calculate the aligned pointer; just check the last bits of the
pointer.
Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
The rte_mov256 function was missing for AVX2.
Fixes: 9144d6bcde ("eal/x86: optimize memcpy for SSE and AVX")
Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
In async_enqueue_pkts(), the failed pkts will
be freed before return, but, the failed pkts may be
retried later, it will cause use after free. So,
we free the failed pkts after retry.
Fixes: 1907ce4bae ("examples/vhost: fix retry logic on Rx path")
Cc: stable@dpdk.org
Signed-off-by: Wenwu Ma <wenwux.ma@intel.com>
Tested-by: Wei Ling <weix.ling@intel.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
If an application does not request IOMMU support, we can avoid
allocating a IOMMU pool.
This saves 112kB (IOTLB_CACHE_SIZE * sizeof(struct vhost_iotlb_entry))
per vq.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
