Age | Commit message (Collapse) | Author | Files | Lines |
|
When n_rx_packets is less then 16(VEC256) or 8(VEC128), code execution
will fall into scalar path of processing packets. But with a wrong
initialization value for n_left set to zero, i in the for-loop will
equal to n_rx_packets. This leads to the bypass of required ip4 checksum
validation and byte count endianness conversion in scalar path.
Besides, refactor the code using while instead of for-loop to keep
consistency with VPP code style.
Type: fix
Fixes: bf93670c515d ("rdma: fix ipv4 checksum check in rdma-input node")
Signed-off-by: Lijian Zhang <lijian.zhang@arm.com>
Signed-off-by: Jieqiang Wang <jieqiang.wang@arm.com>
Change-Id: Ib4e8cb5202735f8b060c99caddf26035657551e1
|
|
- cqe_flags pointer should be incremented accordingly otherwise only the
first element in cqe_flags will be updated
- flag l3_ok should be set for match variable when verifying if packets
are IPv4 packets with flag l3_ok set
- mask/match variables should be converted to network byte order to
match the endianness of cqe_flags
- vector processing of checking cqe flags will set return value to
0xFFFF by mistake if packet numbers are not multiple of 16(VEC256) or
8(VEC128)
Type: fix
Signed-off-by: Jieqiang Wang <jieqiang.wang@arm.com>
Change-Id: I9fec09e449fdffbb0ace8e5a6ccfeb6869b5cac1
|
|
- fix branch prediction for checking rdma ERROR flag
- add the missing right angle bracket to help message
Type: improvement
Signed-off-by: Jieqiang Wang <jieqiang.wang@arm.com>
Reviewed-by: Lijian Zhang <lijian.zhang@arm.com>
Reviewed-by: Tianyu Li <tianyu.li@arm.com>
Change-Id: I2ce667631b3e3f60939069e2a16ddba0ff12a695
|
|
Type: refactor
Change-Id: Id10cbf52e8f2dd809080a228d8fa282308be84ac
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Current rdma input L3 validating behavior for scalar path is:
if any packet L3_OK flag matches, then unset skip_ip4_cksum.
The correct behavior should be if any packet L3_OK NOT match,
then unset skip_ip4_cksum. The logic is also different from
the vector path. This patch fixes the wrong behavior for scalar path.
Type: fix
Signed-off-by: Tianyu Li <tianyu.li@arm.com>
Change-Id: I5ca5ed3aa0c07d441f3c87b33f03ea8f7a3c9826
|
|
Change-Id: I17228bd35cc7aa0d207a16502bf1376c75675302
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
Type: fix
|
|
Change-Id: Id539d36635f0ab9625dc2fc73630be39bead09af
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
Type: improvement
|
|
Change-Id: I623617ad3c80610805dd3cf2a5f371e6677f4844
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
Type: improvement
|
|
When using classifier to filter traces, not all packets will be traced.
In that case, we should only count traced packets.
Type: fix
Change-Id: I87d1e217b580ebff8c6ade7860eb43950420ae78
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
This change leverages the striding RQ feature of
ConnectX-5 adapters to support chained buffers on
the RX path. In Striding RQ mode, WQE are SG lists
of data segments, each mapped to a vlib_buffer.
When a packet is received, it can consume one or
multiple data segments belonging to the WQE,
without wasting the whole WQE.
Change-Id: I74eba5b2c2c66538e75e046335058ba011cb27fd
Type: improvement
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
|
|
allocation, the remainder modulo 8 must be returned to the pool so that the number of kept buffers is a multiple of 8.
Type: fix
Change-Id: Ifd97b03ea220300e7e6fe81a8ff4a25060fea6c1
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
|
|
Type: fix
Change-Id: I0ff7a6f3354066c2252c1ebd0d43ac59db278bf0
Signed-off-by: Damjan Marion <damarion@cisco.com>
Signed-off-by: Elias Rudberg <elias.rudberg@bahnhof.net>
|
|
Type: feature
Change-Id: I3f287ab536a482c366ad7df47e1c04e640992ebc
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: improvement
Change-Id: If81847bc0c92d167ce03e1e94a2f8e18f8154af2
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: improvement
Change-Id: Ib9921718ec88e6dc56a38e40a140f16f86ed2e44
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: improvement
Change-Id: Ie6f2c79e0a769f70eab079c75a500b9308dd51e6
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Fix a bug that caused some input packets to be dropped due to errors of
the type 'ip4 length > l2 length'. The change is related to the second
call to the rdma_device_input_bufs() function that happens when the end
of the ring buffer is reached.
Type: fix
Change-Id: I332d69ab22242b3443a0baca6e5dd86349a54765
Signed-off-by: Elias Rudberg <elias.rudberg@bahnhof.net>
|
|
Type: feature
Change-Id: I913f08383ee1c24d610c3d2aac07cef402570e2c
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Refactor rdma driver for improved performance and prepare for raw
datapath access.
Type: refactor
Change-Id: Iae31872055a6947708ea9f430bd1dc083ea63b5a
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
Input feature arcs enable functions such as l2patch or bonding.
Type: feature
Change-Id: Ie333c9b8f7ea20f8a3a409a9a06367807e215a57
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
In case of tx success after multiple retries, the last buffers to be
enqueued will be both enqueued for tx and freed.
Type: fix
Fixes: 211ef2eb24
Change-Id: I57d218cff58b74c1f3d6dc5722624327f0821758
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
The fast path almost always has to deal with the real
pointers. Deriving the frame pointer from a frame_index requires a
load of the 32bit frame_index from memory, another 64bit load of the
heap base pointer and some calculations.
Lets store the full pointer instead and do a single 64bit load only.
This helps avoiding problems when the heap is grown and frames are
allocated below vm->heap_aligned_base.
Type: refactor
Change-Id: Ifa6e6e984aafe1e2755bff80f0a4dfcddee3623c
Signed-off-by: Andreas Schultz <andreas.schultz@travelping.com>
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Add support for multiqueue for RDMA devices.
Change-Id: I78a2481cec6747494c670cc776475828be3af55b
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
rx: add batching for WC processing and release
tx: improve batching for WC submission and processing
rdma-core: compile in release mode to remove assert()
Change-Id: I5fb8736db36b50f8b758cd688100477b67e72d80
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
RDMA ibverb is a userspace API to efficiently rx/tx packets. This is an
initial, unoptimized driver targeting Mellanox cards.
Next steps should include batching, multiqueue and additional cards.
Change-Id: I0309c7a543f75f2f9317eaf63ca502ac7a093ef9
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|