Implement send inline feature which copies packet data directly into
work queue entries (WQEs) for improved latency. The maximum packet
size and the minimum number of Tx queues to qualify for inline send
are user-configurable.
This feature is effective when HW causes a performance bottleneck.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Replacing the variable countdown (which depends on the number of
descriptors) with a fixed relative threshold known at compile time improves
performance by reducing the TX queue structure footprint and the amount of
code to manage completions during a burst.
Completions are now requested at most once per burst after threshold is
reached.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Vasily Philipov <vasilyf@mellanox.com>
Mini (compressed) completion queue entries (CQEs) are returned by the
NIC when PCI back pressure is detected, in which case the first CQE64
contains common packet information followed by a number of CQE8
providing the rest, followed by a matching number of empty CQE64
entries to be used by software for decompression.
Before decompression:
0 1 2 6 7 8
+-------+ +---------+ +-------+ +-------+ +-------+ +-------+
| CQE64 | | CQE64 | | CQE64 | | CQE64 | | CQE64 | | CQE64 |
|-------| |---------| |-------| |-------| |-------| |-------|
| ..... | | cqe8[0] | | | . | | | | | ..... |
| ..... | | cqe8[1] | | | . | | | | | ..... |
| ..... | | ....... | | | . | | | | | ..... |
| ..... | | cqe8[7] | | | | | | | | ..... |
+-------+ +---------+ +-------+ +-------+ +-------+ +-------+
After decompression:
0 1 ... 8
+-------+ +-------+ +-------+
| CQE64 | | CQE64 | | CQE64 |
|-------| |-------| |-------|
| ..... | | ..... | . | ..... |
| ..... | | ..... | . | ..... |
| ..... | | ..... | . | ..... |
| ..... | | ..... | | ..... |
+-------+ +-------+ +-------+
This patch does not perform the entire decompression step as it would be
really expensive, instead the first CQE64 is consumed and an internal
context is maintained to interpret the following CQE8 entries directly.
Intermediate empty CQE64 entries are handed back to HW without further
processing.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Vasily Philipov <vasilyf@mellanox.com>
The latest version of Mellanox OFED exposes hardware definitions necessary
to implement data path operation bypassing Verbs. Update the minimum
version requirement to MLNX_OFED >= 3.3 and clean up compatibility checks
for previous releases.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
To keep the data path as efficient as possible, move fields only useful to
the control path into new structure rxq_ctrl.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
To keep the data path as efficient as possible, move fields only useful to
the control path into new structure txq_ctrl.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Inline TX will be fully managed by the PMD after Verbs is bypassed in the
data path. Remove the current code until then.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This is done in preparation of bypassing Verbs entirely for the data path
as a performance improvement. RX scatter cannot be maintained during the
transition and will be reimplemented later.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This is done in preparation of bypassing Verbs entirely for the data path
as a performance improvement. TX gather cannot be maintained during the
transition and will be reimplemented later.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Except for the first time when memory registration occurs, the lkey is
always cached. Since memory registration is slow and performs system calls,
performance can be improved by moving that code to its own function outside
of the data path so only the lookup code is left in the original inlined
function.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Introduce rte_mempool_populate_default() which allocates
mempool objects in several memzones.
The mempool header is now always allocated in a specific memzone
(not with its objects). Thanks to this modification, we can remove
many specific behavior that was required when hugepages are not
enabled in case we are using rte_mempool_xmem_create().
This change requires to update how kni and mellanox drivers lookup for
mbuf memory. For now, this will only work if there is only one memory
chunk (like today), but we could make use of rte_mempool_mem_iter() to
support more memory chunks.
We can also remove RTE_MEMPOOL_OBJ_NAME that is not required anymore for
the lookup, as memory chunks are referenced by the mempool.
Note that rte_mempool_create() is still broken (it was the case before)
when there is no hugepages support (rte_mempool_create_xmem() has to be
used). This is fixed in next commit.
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
This commit removes the const qualifier for the mempool in
rte_mempool_walk() callback prototype.
Indeed, most functions that can be done on a mempool require a non-const
mempool pointer, except the dump and the audit. Therefore, the
mempool_walk() is more useful if the mempool pointer is not const.
This is required by next commit where the mellanox drivers use
rte_mempool_walk() to iterate the mempools, then rte_mempool_obj_iter()
to iterate the objects in each mempool.
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
VLAN insertion can be done in hardware when supported in Verbs. A software
fallback is provided otherwise. The software implementation is also used
when multi-packet send is enabled on a queue, as both features are mutually
exclusive.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Secondary processes are expected to use queues and other resources
allocated by the primary, however Verbs resources can only be shared
between processes when inherited through fork().
This limitation can be worked around for TX by configuring separate queues
from secondary processes.
Signed-off-by: Or Ami <ora@mellanox.com>
The first and last memory pool elements are usually cache-aligned but not
page-aligned, particularly when using huge pages.
Hardware performance can be improved significantly by registering memory
regions starting and ending on page boundaries.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Avoid dereferencing pointers twice to get to fast Verbs functions by
storing them directly in RX/TX queue structures.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Allows HW to strip the 802.1Q header from incoming frames and report it
through the mbuf structure.
This feature requires MLNX_OFED >= 3.2.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Upcoming flow director support will reuse this function to generate filter
rules.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Until now, broadcast frames were handled like unicast. Moving the related
flow to the special flows table frees up the related unicast MAC entry.
The same method is used to handle IPv6 multicast frames.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Merge redundant code by adding a static initialization table to manage
promiscuous and allmulticast (special) flows.
New function priv_rehash_flows() implements the logic to enable/disable
relevant flows in one place from any context.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Pre-registering mbuf memory pools when creating TX queues avoids costly
registrations later in the data path.
Fixes: 2e22920b85 ("mlx5: support non-scattered Tx and Rx")
Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Seen with GCC < 4.6:
error: unknown field ‘tcp_udp’ specified in initializer
error: extra brace group at end of initializer
Static initialization of anonymous structs/unions is a C11 feature
properly supported only since GCC 4.6.
Work around compilation errors with older versions by expanding
struct ibv_exp_flow_spec into struct hash_rxq_init.
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Normal flows do not currently provide IPv6 support.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Only a single flow per hash RX queue is needed in promiscuous mode.
Disable others to free up hardware resources.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
DPDK expects to have an RSS hash key per flow type (IPv4, IPv6, UDPv4,
etc.), to handle this the PMD must keep a table of hash keys to be able
to reconfigure the queues at each start/stop call.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
First implementation of rss_hash_update and rss_hash_conf_get, those
functions still lack in functionality but are usable to change the RSS
hash key. For now, the PMD does not handle an indirection table for
each kind of flow (IPv4, IPv6, etc.), the same RSS hash key is used
for all protocols. This situation explains why the rss_hash_conf_get
returns the RSS hash key for all DPDK supported protocols and why the
hash key is set for all of them too.
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
All hash RX QPs currently use the same flow steering rule (L2 MAC filtering)
regardless of their type (TCP, UDP, IPv4, IPv6), which prevents them from
being dispatched properly. This is fixed by adding flow information to the
hash RX queue initialization data and generating specific flow steering
rules for each of them.
Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
The default hash RX queue handles packets that are not matched by more
specific types and requires its own indirection table of size 1 to work
properly.
This commit implements support for multiple indirection tables by grouping
their layout and properties in a static initialization table.
Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
The new Verbs RSS API is lower-level than the previous one and much more
flexible but requires RX queues to use Work Queues (WQs) internally instead
of Queue Pairs (QPs), which are grouped in an indirection table used by a
new kind of hash RX QPs.
Hash RX QPs and the indirection table together replace the parent RSS QP
while WQs are mostly similar to child QPs.
RSS hash key is not configurable yet.
Summary of changes:
- Individual DPDK RX queues do not store flow properties anymore, this info
is now part of the hash RX queues.
- All functions affecting the parent queue when RSS is enabled or the basic
queues otherwise are modified to affect hash RX queues instead.
- Hash RX queues are also used when a single DPDK RX queue is configured (no
RSS) to remove that special case.
- Hash RX queues and indirection table are created/destroyed when device
is started/stopped in addition to create/destroy flows.
- Contrary to QPs, WQs are moved to the "ready" state before posting RX
buffers, otherwise they are ignored.
- Resource domain information is added to WQs for better performance.
- CQs are not resized anymore when switching between non-SG and SG modes as
it does not work correctly with WQs. Use the largest possible size
instead, since CQ size does not have to be the same as the number of
elements in the RX queue. This also applies to the maximum number of
outstanding WRs in a WQ (max_recv_wr).
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Or Ami <ora@mellanox.com>
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Removing this structure reduces the size of SG and non-SG RX queue elements
significantly to improve performance.
An nice side effect is that the mbuf pointer is now fully stored in
struct rxq_elt instead of relying on the WR ID data offset hack.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Or Ami <ora@mellanox.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
This is the same implementation as mlx4.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
All MAC RX flows must be updated with VLAN information when configuring a
VLAN filter.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Depending on the MTU and whether jumbo frames are enabled, RX queues may
switch between SG and non-SG modes for better performance.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
A dedicated RX callback is added to handle scattered buffers. For better
performance, it is only used when jumbo frames are enabled and MTU is larger
than a single mbuf.
On the TX path, scattered buffers are also handled in a separate function.
When there are more than MLX5_PMD_SGE_WR_N segments in a given mbuf, the
remaining segments are linearized in the last SGE entry.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
This commit adds support for MAC flow steering rules mandatory for the RX
path as well as the related callbacks to add/remove MAC addresses.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Didier Pallard <didier.pallard@6wind.com>
RSS implementation with parent/child QPs comes from mlx4 and is temporary.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
