Expand the use of mempool registration to MR management for other
drivers.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Add global shared MR cache as a field of common device structure.
Move MR management to use this global cache for all drivers.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
To support SubFunction based on auxiliary bus, common driver supports
new bus-agnostic driver.
This patch migrates net driver to new common driver.
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
This patch separates Tx function declarations to different header file
in preparation for removing their implementation from the source file
and as an optional preparation for Tx cleanup.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
This patch separates Rx function implementations to different source
file as an optional preparation step for further consolidation of Rx
burst functions.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The mlx5_rxtx.c file contains a lot of Tx burst functions, each of those
is performance-optimized for the specific set of requested offloads.
These ones are generated on the basis of the template function and it
takes significant time to compile, just due to a large number of giant
functions generated in the same file and this compilation is not being
done in parallel with using multithreading.
Therefore we can split the mlx5_rxtx.c file into several separate files
to allow different functions to be compiled simultaneously.
In this patch, we separate Rx function declarations to different header
file in preparation for removing them from the source file and as an
optional preparation step for further consolidation of Rx burst
functions.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The Rx metadata might use the metadata register C0 to keep the
values. The same register C0 might be used by kernel for source
vport value handling, kernel uses upper half of the register,
leaving the lower half for application usage.
In the extended metadata mode 1 (dv_xmeta_en devarg is
assigned with value 1) the metadata width is 16 bits only,
the Rx datapath code fetched the entire 32-bit value of the
metadata register and presented one to application. The patch
provides data masking depending on the chosen metadata mode.
Fixes: 6c55b622a9 ("net/mlx5: set dynamic flow metadata in Rx queues")
Cc: stable@dpdk.org
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Verbs cannot be used to configure newly introduced miniCQE formats for
Flow Tag and L3/L4 Header compression. Support for these formats has
been added to the DevX configuration only. And the RX queue descriptor
has been updated with the CQE compression format information only as
well. But the datapath relies on this info no matter which method is
used for Rx queues configuration. Set proper CQE compression format
information in the Verbs configuration to fix the miniCQE parsing logic.
Fixes: 54c2d46b16 ("net/mlx5: support flow tag and packet header miniCQEs")
Cc: stable@dpdk.org
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
This patch adds support of the mbuf fast free offload to the
transmit datapath. This offload allows freeing the mbufs on
transmit completion in the most efficient way. It requires
the all mbufs were allocated from the same pool, have
the reference counter value as 1, and have no any externally
attached buffers.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
The mlx5 PMD supports packet data inlining by pushing data
to the transmit descriptor. If packet is short enough and all
data are inline, the mbuf is not needed for data send anymore
and can be freed.
The mbuf free was performed in the most inner loop building
the transmit descriptors. This patch postpones the mbuf free
transaction to the tx_burst routine exit, optimizing the loop
and allowing the bulk freeing for the multiple mbufs in single
pool API call.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Functions mlx5_check_mprq_support(), mlx5_rxq_mprq_enabled(),
mlx5_mprq_enabled() are moved from source file mlx5_rxq.c to header file
mlx5_rxtx.h and their type is updated to 'static __rte_always_inline'.
Previously the functions were declared as 'inline' in the source file
which was reported as 'unresolved external symbol' error by some Windows
linkers.
Signed-off-by: Ophir Munk <ophirmu@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
The shared RSS action update was not operational due to lack
of kernel driver support of TIR object modification.
This commit introduces the workaround to support shared RSS
action modify using an indirect queue table update instead of
touching TIR object directly.
Limitations: the only supported RSS property to update is queues, the
rest of the properties ignored.
Fixes: d2046c09aa ("net/mlx5: support shared action for RSS")
Signed-off-by: Andrey Vesnovaty <andreyv@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
CQE compression allows us to save the PCI bandwidth and improve
the performance by compressing several CQEs together to a miniCQE.
But the miniCQE size is only 8 bytes and this limits the ability
to successfully keep the compression session in case of various
traffic patterns.
The current miniCQE format only keeps the compression session alive
in case of uniform traffic with the Hash RSS as the only difference.
There are requests to keep the compression session in case of tagged
traffic by RTE Flow Mark Id and mixed UDP/TCP and IPv4/IPv6 traffic.
Add 2 new miniCQE formats in order to achieve the best performance
for these traffic patterns: Flow Tag and Packet Header miniCQEs.
The existing rxq_cqe_comp_en devarg is modified to specify the
desired miniCQE format. Specifying 2 selects Flow Tag format
for better compression rate in case of RTE Flow Mark traffic.
Specifying 3 selects Checksum format (existing format for MPRQ).
Specifying 4 selects L3/L4 Header format for better compression
rate in case of mixed TCP/UDP and IPv4/IPv6 traffic.
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The rte_atomic API is deprecated and needs to be replaced with
C11 atomic builtins. Use the relaxed ordering and explicit
memory barrier for Clock Queue and timestamps synchronization.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
This commit applies the cache linked list to Rx queue to make it thread
safe.
Signed-off-by: Suanming Mou <suanmingm@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
When create shared RSS hrxq, the hrxq will be created directly, no hrxq
will be reused.
In this case, add the shared RSS hrxq to the queue list is redundant.
And it also hurts the generic queue lookup.
This commit avoids add the shared RSS hrxq to the queue list.
Signed-off-by: Suanming Mou <suanmingm@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
In the current implementation of the single port hairpin, there is
an implicit splitting process for actions. When inserting a hairpin
flow, all the actions will be included with the ingress attribute.
The flow engine will check and decide which actions should be moved
into the TX flow part, e.g., encapsulation, VLAN push.
In some NICs, some actions can only be done in one direction. Since
the hairpin flow will be split into two parts, such validation will
be skipped.
With the hairpin explicit TX flow mode, no splitting is needed any
more. The hairpin flow may have no big difference from a standard
flow (except the queue). The application should take full charge of
the actions and the flow engine should validate the hairpin flow in
the same way as other flows.
In the meanwhile, a new internal API is added to get the hairpin
configuration. This will bypass the useless atomic operation to save
the CPU cycles.
Signed-off-by: Bing Zhao <bingz@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
In order to support hairpin between two ports, mlx5 PMD needs to
implement the functions and provide them as the function pointers.
The bind and unbind functions are executed per port pairs. All the
hairpin queues between the two ports should have the same attributes
during queues setup. Different configurations among queue pairs from
the same ports are not supported. It is allowed that two ports only
have one direction hairpin.
In order to set up the connection between two queues, peer Rx queue
HW information must be fetched via the internal RTE API and the queue
information could be used to modify the SQ object. Then the RQ object
will be modified with the Tx queue HW information. The reverse
operation is not supported right now.
When disconnecting the queues pair, SQ and RQ object should be reset
without any peer HW information. The unbinding operation will try to
disconnect all Tx queues from the port from the Rx queues of the peer
port.
Tx explicit mode attribute will be saved and used when creating a
hairpin flow.
Signed-off-by: Bing Zhao <bingz@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The routine to provide Rx queue setup with specifying
extended receiving buffer description is added.
It allows application to specify desired segment
lengths, data position offsets in the buffer
and dedicated memory pool for each segment.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Implement modification for hashed table of Rx queue object (see
mlx5_hrxq_modify()). This implementation relies on the capability to
modify TIR object via DevX API, i.e. current implementation doesn't
support verbs HW object operations. The functionality to modify hashed
table of Rx queue object is prerequisite to implement
rete_flow_shared_action_update() for shared RSS action in mlx5 PMD.
Signed-off-by: Andrey Vesnovaty <andreyv@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
MPRQ (Multi-Packet Rx Queue) processes one packet at a time using
simple scalar instructions. MPRQ works by posting a single large buffer
(consisted of multiple fixed-size strides) in order to receive multiple
packets at once on this buffer. A Rx packet is then copied to a
user-provided mbuf or PMD attaches the Rx packet to the mbuf by the
pointer to an external buffer.
There is an opportunity to speed up the packet receiving by processing
4 packets simultaneously using SIMD (single instruction, multiple data)
extensions. Allocate mbufs in batches for every MPRQ buffer and process
the packets in groups of 4 until all the strides are exhausted. Then
switch to another MPRQ buffer and repeat the process over again.
The vectorized MPRQ burst routine is engaged automatically in case
the mprq_en=1 devarg is specified and the vectorization is not disabled
explicitly by providing rx_vec_en=0 devarg. There is a limitation:
LRO is not supported and scalar MPRQ is selected if it is on.
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The HW objects of the Tx queue is created/destroyed in the device
start\stop stage while the ethdev configurations for the Tx queue
starts from the tx_queue_setup stage.
The PMD should save all the last configurations it got from the ethdev
and to apply them to the device in the dev_start operation.
Wrongly, last code added to mitigate the reference counters didn't take
into account the above rule and combined the configurations and HW
objects to be created\destroyed together.
This causes to memory leak and other memory issues.
Make sure the HW object is released in stop operation when there is no
any reference to it while the configurations stay saved.
Fixes: 17a57183c0 ("net/mlx5: mitigate Tx queue reference counters")
Signed-off-by: Matan Azrad <matan@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The HW objects of the Rx queue is created/destroyed in the device
start\stop stage while the ethdev configurations for the Rx queue
starts from the rx_queue_setup stage.
The PMD should save all the last configurations it got from the ethdev
and to apply them to the device in the dev_start operation.
Wrongly, last code added to mitigate the reference counters didn't take
into account the above rule and combined the configurations and HW
objects to be created\destroyed together.
This causes to memory leak and other memory issues.
Make sure the HW object is released in stop operation when there is no
any reference to it while the configurations stay saved.
Fixes: 24e4b650ba ("net/mlx5: mitigate Rx queue reference counters")
Signed-off-by: Matan Azrad <matan@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The mbuf timestamp is moved to a dynamic field
in order to allow removal of the deprecated static field.
The related mbuf flag is also replaced.
The dynamic offset and flag are stored in struct mlx5_rxq_data
to favor cache locality.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: David Christensen <drc@linux.vnet.ibm.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Acked-by: David Marchand <david.marchand@redhat.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
As an arrangement to Windows OS support, the Verbs operations should be
separated to another file.
By this way, the build can easily cut the unsupported Verbs APIs from
the compilation process.
Define operation structure and DevX module in addition to the existing
Linux Verbs module.
Separate Tx object creation into the Verbs/DevX modules and update the
operation structure according to the OS support and the user
configuration.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
The Tx queue structures manage 2 different reference counter per queue:
txq_ctrl reference counter and txq_obj reference counter.
There is no real need to use two different counters, it just complicates
the release functions.
Remove the txq_obj counter and use only the txq_ctrl counter.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Use C11 atomics with RELAXED ordering instead of the rte_atomic ops
which enforce unnecessary barriers on aarch64.
Signed-off-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Since the 20.08 release deprecated rte_cio_*mb APIs because these APIs
provide the same functionality as rte_io_*mb APIs on all platforms, so
remove them and use rte_io_*mb instead.
Signed-off-by: Phil Yang <phil.yang@arm.com>
Signed-off-by: Joyce Kong <joyce.kong@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: David Marchand <david.marchand@redhat.com>
Move Rx queue drop action similar resources allocations from Verbs
module to a shared location.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Move Rx hash queue object similar resources allocations from DevX and
Verbs modules to a shared location.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Move Rx indirection table object similar resources allocations from DevX
and Verbs modules to a shared location.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Separate Rx hash queue creation into both Verbs and DevX modules.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Separate Rx indirection table object creation into both Verbs and DevX
modules.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Rearrangement of WQ and CQ creation for Verbs Rx queue:
1. Rename the allocation function.
2. Reduce the number of arguments that the creation functions receive.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Rearrangement of RQ and CQ resource handling for DevX Rx queue:
1. Rename the allocation function so that it is understood that it
allocates all resources and not just the CQ or RQ.
2. Move the allocation and release of the doorbell into creation and
release functions.
3. Reduce the number of arguments that the creation functions receive.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Move Rx object similar resources allocations and debug logs from DevX
and Verbs modules to a shared location.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
As an arrangement to Windows OS support, the Verbs operations should be
separated to another file.
By this way, the build can easily cut the unsupported Verbs APIs from
the compilation process.
Define operation structure and DevX module in addition to the existing
linux Verbs module.
Separate Rx object creation into the Verbs/DevX modules and update the
operation structure according to the OS support and the user
configuration.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
The Rx queue structures manage 2 different reference counter per queue:
rxq_ctrl reference counter and rxq_obj reference counter.
There is no real need to use two different counters, it just complicates
the release functions.
Remove the rxq_obj counter and use only the rxq_ctrl counter.
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Rx HW objects can be created by both Verbs and DevX operations.
The management of the 2 types of operations are done directly in the
main flow of the object’s creations.
Some arrangements and validations were wrongly done to the irrelevant
type:
1. LRO related validations were done for Verbs type where LRO is not
supported at all.
2. Verbs allocation arrangements were done for DevX operations where it
is not needed.
3. Doorbell destroy was considered for Verbs types where it is
irrelevant.
Adjust the aforementioned points only for the relevant types.
Fixes: e79c9be915 ("net/mlx5: support Rx hairpin queues")
Fixes: 08d1838f64 ("net/mlx5: implement CQ for Rx using DevX API")
Fixes: 17ed314c6c ("net/mlx5: allow LRO per Rx queue")
Fixes: dc9ceff73c ("net/mlx5: create advanced RxQ via DevX")
Cc: stable@dpdk.org
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Several DV-based structs of type 'struct mlx5dv_devx_XXX' are replaced
with 'void *' to enable compilation under non-Linux operating systems.
New getter functions were added to retrieve the specific fields that
were previously accessed directly.
Replaced structs:
'struct mlx5dv_pp *'
'struct mlx5dv_devx_event_channel *'
'struct mlx5dv_devx_umem *'
'struct mlx5dv_devx_uar *'
Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
The mlx5 PMD did not support queue_start and queue_stop eth_dev API
routines, queue could not be suspended and resumed during device
operation.
There is the use case when this feature is crucial for applications:
- there is the secondary process handling the queue
- secondary process crashed/aborted
- some mbufs were allocated or used by secondary application
- some mbufs were allocated by Rx queues to receive packets
- some mbufs were placed to send queue
- queue goes to undefined state
In this case there was no reliable way to recovery queue handling
by restarted secondary process but reset queue to initial state
freeing all involved resources, including buffers involved in queue
operations, reset the mbuf pools, and then reinitialize queue
to working state:
- reset mbuf pool, allocate all mbuf to initialize pool into
safe state after the crush and allow safe mbuf free calls
- stop queue, free all potentially involved mbufs
- reset mbuf pool again
- start queue, reallocate mbufs needed
This patch introduces the queue start/stop feature with some
limitations:
- hairpin queues are not supported
- it is application responsibility to synchronize start/stop
with datapath routines, rx/tx_burst must be suspended during
the queue_start/queue_stop calls
- it is application responsibility to track queue usage and
provide coordinated queue_start/queue_stop calls from
secondary and primary processes.
- Rx queues with vectorized Rx routine and engaged CQE
compression are not supported by this patch currently
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
This patch continues the work to use DevX API for different objects
creation and management.
On Rx control path, the RQ, RQT, and TIR objects can already be
created using DevX API.
This patch adds the support to create CQ for RxQ using DevX API.
The corresponding event channel is also created and utilized using
DevX API.
Signed-off-by: Dekel Peled <dekelp@mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Several source files include Verbs header files as in (1). These source
files will not compile under non-Linux operating systems. This commit
removes this inclusion in two cases:
Case 1: There is no usage of ibv_* or mlx5dv_* symbols in the source
file so the inclusion in (1) can be safely removed.
Case 2: Verbs symbols are used. Please note the inclusion in (1) already
appears in file linux/mlx5_glue.h (which represents the interface
to the rdma-core library). Therefore, replace (1) in the source file
with (2). Under non-Linux operating systems - file mlx5_glue.h will not
include (1).
(1)
#include <infiniband/verbs.h>
#include <infiniband/mlx5dv.h>
(2)
#include <mlx5_glue.h>
Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
The ConnectX-6DX supports the timestamps in various formats,
the new realtime format is introduced - the upper 32-bit word
of timestamp contains the UTC seconds and the lower 32-bit word
contains the nanoseconds. This patch detects what format is
configured in the NIC and performs the conversion accordingly.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
The application provides timestamps in Tx mbuf as clocks,
the hardware performs scheduling on Clock Queue completion index
match. This patch introduces the timestamp-to-completion-index
inline routine.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
The fields to support send scheduling on dynamic timestamp
field are introduced and initialized on device start.
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
Service routine is invoked periodically on Rearm Queue
completion interrupts, typically once per some milliseconds
(1-16) to track clock jitter and wander in robust fashion.
It performs the following:
- fetches the completed CQEs for Rearm Queue
- restarts Rearm Queue on errors
- pushes new requests to Rearm Queue to make it
continuously running and pushing cross-channel requests
to Clock Queue
- reads and caches the Clock Queue CQE to be used in datapath
- gathers statistics to estimate clock jitter and wander
- gathers Clock Queue errors statistics
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
