The HSR duplicate discard algorithm had even more basic problems than the
described for PRP in the previous patch. It relied only on the last
received sequence number to decide if a new frame should be forwarded to
any port. This does not work correctly in any case where frames are
received out of order. The linked bug report claims that this can even
happen with perfectly fine links due to the order in which incoming frames
are processed (which can be unexpected on multi-core systems). The issue
also occasionally shows up in the HSR selftests. The main reason is that
the sequence number that was last forwarded to the master port may have
skipped a number which will in turn never be delivered to the host.
As the problem (we accidentally skip over a sequence number that has not
been received but will be received in the future) is similar to PRP, we can
apply a similar solution. The duplicate discard algorithm based on the
"sparse bitmap" works well for HSR if it is extended to track one bitmap
for each port (A, B, master, interlink). To do this, change the sequence
number blocks to contain a flexible array member as the last member that
can keep chunks for as many bitmaps as we need. This design makes it easy
to reuse the same algorithm in a potential PRP RedBox implementation.
The duplicate discard algorithm functions are modified to deal with
sequence number blocks of different sizes and to correctly use the array of
bitmap chunks. There is a notable speciality for HSR: the port type has a
special port type NONE with value 0. This leads to the number of port types
being 5 instead of actually 4. To save memory, remove the NONE port from
the bitmap (by subtracting 1) when setting up the block buffer and when
accessing the bitmap chunks in the array.
Removing the old algorithm allows us to get rid of a few fields that are
not needed any more: time_out and seq_out for each port. We can also remove
some functions that were only necessary for the previous duplicate discard
algorithm.
The removal of seq_out is possible despite its previous usage in
hsr_register_frame_in: it was used to prevent updates to time_in when
"invalid" sequence numbers were received. With the new duplicate discard
algorithm, time_in has no relevance for the expiry of sequence numbers
anymore. They will expire based on the timestamps in the sequence number
blocks after at most 400ms. There is no need that a node "re-registers" to
"resume communication": after 400ms, all sequence numbers are accepted
again. Also, according to the IEC 62439-3:2021, all nodes are supposed to
send no traffic for 500ms after boot to lead exactly to this expiry of seen
sequence numbers. time_in is still used for pruning nodes from the node
table after no traffic has been received for 60sec. Pruning is only needed
if the node is really gone and has not been sending any traffic for that
period.
seq_out was also used to report the last incoming sequence number from a
node through netlink. I am not sure how useful this value is to userspace
at all, but added getting it from the sequence number blocks. This number
can be outdated after node merging until a new block has been added.
Update the KUnit test for the PRP duplicate discard so that the node
allocation matches and expectations on the removed fields are removed.
Reported-by: Yoann Congal <yoann.congal@smile.fr>
Closes: https://lore.kernel.org/netdev/7d221a07-8358-4c0b-a09c-3b029c052245@smile.fr/
Signed-off-by: Felix Maurer <fmaurer@redhat.com>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://patch.msgid.link/36dc3bc5bdb7e68b70bb5ef86f53ca95a3f35418.1770299429.git.fmaurer@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add tests where one link has different rates of packet loss or reorders
packets. PRP should still be able to recover from these link faults and
show no packet loss. However, it is acceptable to receive some level of
duplicate packets. This matches the current specification (IEC
62439-3:2021) of the duplicate discard algorithm that requires it to be
"designed such that it never rejects a legitimate frame, while occasional
acceptance of a duplicate can be tolerated." The rate of acceptable
duplicates in this test is intentionally high (10%) to make the test
stable, the values I observed in the worst test cases (20% loss) are around
5% duplicates.
The duplicates occur because of the 10ms ping interval in the test. As
blocks expire after 400ms based on the timestamp of the first received
sequence number in the block, every approx. 40th will lead to a new, clean
block being used where the sequence number hasn't been seen before. As this
occurs on both nodes in the test (for requests and replies), we observe
around 20 duplicate frames.
Signed-off-by: Felix Maurer <fmaurer@redhat.com>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://patch.msgid.link/7b36506d3a80e53786fe56526cf6046c74dfeee1.1770299429.git.fmaurer@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The PRP duplicate discard algorithm does not work reliably with certain
link faults. Especially with packet loss on one link, the duplicate discard
algorithm drops valid packets which leads to packet loss on the PRP
interface where the link fault should in theory be perfectly recoverable by
PRP. This happens because the algorithm opens the drop window on the lossy
link, covering received and lost sequence numbers. If the other, non-lossy
link receives the duplicate for a lost frame, it is within the drop window
of the lossy link and therefore dropped.
Since IEC 62439-3:2012, a node has one sequence number counter for frames
it sends, instead of one sequence number counter for each destination.
Therefore, a node can not expect to receive contiguous sequence numbers
from a sender. A missing sequence number can be totally normal (if the
sender intermittently communicates with another node) or mean a frame was
lost.
The algorithm, as previously implemented in commit 05fd00e5e7 ("net: hsr:
Fix PRP duplicate detection"), was part of IEC 62439-3:2010 (HSRv0/PRPv0)
but was removed with IEC 62439-3:2012 (HSRv1/PRPv1). Since that, no
algorithm is specified but up to implementers. It should be "designed such
that it never rejects a legitimate frame, while occasional acceptance of a
duplicate can be tolerated" (IEC 62439-3:2021).
For the duplicate discard algorithm, this means that 1) we need to track
the sequence numbers individually to account for non-contiguous sequence
numbers, and 2) we should always err on the side of accepting a duplicate
than dropping a valid frame.
The idea of the new algorithm is to store the seen sequence numbers in a
bitmap. To keep the size of the bitmap in control, we store it as a "sparse
bitmap" where the bitmap is split into blocks and not all blocks exist at
the same time. The sparse bitmap is implemented using an xarray that keeps
the references to the individual blocks and a backing ring buffer that
stores the actual blocks. New blocks are initialized in the buffer and
added to the xarray as needed when new frames arrive. Existing blocks are
removed in two conditions:
1. The block found for an arriving sequence number is old and therefore not
relevant to the duplicate discard algorithm anymore, i.e., it has been
added more than the entry forget time ago. In this case, the block is
removed from the xarray and marked as forgotten (by setting its
timestamp to 0).
2. Space is needed in the ring buffer for a new block. In this case, the
block is removed from the xarray, if it hasn't already been forgotten
(by 1.). Afterwards, the new block is initialized in its place.
This has the nice property that we can reliably track sequence numbers on
low traffic situations (where they expire based on their timestamp) and
more quickly forget sequence numbers in high traffic situations before they
potentially wrap over and repeat before they are expired.
When nodes are merged, the blocks are merged as well. The timestamp of a
merged block is set to the minimum of the two timestamps to never keep
around a seen sequence number for too long. The bitmaps are or'd to mark
all seen sequence numbers as seen.
All of this still happens under seq_out_lock, to prevent concurrent
access to the blocks.
The KUnit test for the algorithm is updated as well. The updates are done
in a way to match the original intends pretty closely. Currently, there is
much knowledge about the actual algorithm baked into the tests (especially
the expectations) which may need some redesign in the future.
Reported-by: Steffen Lindner <steffen.lindner@de.abb.com>
Signed-off-by: Felix Maurer <fmaurer@redhat.com>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Tested-by: Steffen Lindner <steffen.lindner@de.abb.com>
Link: https://patch.msgid.link/8ce15a996099df2df5b700969a39e7df400e8dbb.1770299429.git.fmaurer@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a test case that can support different types of faulty links for all
protocol versions (HSRv0, HSRv1, PRPv1). It starts with a baseline with
fully functional links. The first faulty case is one link being cut during
the ping. This test uses a different function for ping that sends more
packets in shorter intervals to stress the duplicate detection algorithms a
bit more and allow for future tests with other link faults (packet loss,
reordering, etc.).
As the link fault tests now cover the cut link for HSR and PRP, it can be
removed from the hsr_ping test. Note that the removed cut link test did not
really test the fault because do_ping_long takes about 1sec while the link
is only cut after a 3sec sleep.
Signed-off-by: Felix Maurer <fmaurer@redhat.com>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://patch.msgid.link/dad52276e2c349ecb96168bef7e3001bf7becc81.1770299429.git.fmaurer@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Previously the hsr_ping test only checked that all nodes in a VLAN are
reachable (using do_ping). Update the test to also check that there is no
packet loss and no duplicate packets by running the same tests for VLANs as
without VLANs (including using do_ping_long). This also adds tests for IPv6
over VLAN. To unify the test code, the topology without VLANs now uses IP
addresses from dead:beef:0::/64 to align with the 100.64.0.0/24 range for
IPv4. Error messages are updated across the board to make it easier to find
what actually failed.
Also update the VLAN test to only run in VLAN 2, as there is no need to
check if ping really works with VLAN IDs 2, 3, 4, and 5. This lowers the
number of long ping tests on VLANs to keep the overall test runtime in
bounds.
It's still necessary to bump the test timeout a bit, though: a ping long
tests takes 1sec, do_ping_tests performs 12 of them, do_link_problem_tests
6, and the VLAN tests again 12. With some buffer for setup and waiting and
for two protocol versions, 90sec timeout seems reasonable.
Signed-off-by: Felix Maurer <fmaurer@redhat.com>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://patch.msgid.link/e3ded0e2547b5f720524b62fabeb96debc579697.1770299429.git.fmaurer@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Wei Fang says:
====================
net: fec: improve XDP copy mode and add AF_XDP zero-copy support
This patch set optimizes the XDP copy mode logic as follows.
1. Separate the processing of RX XDP frames from fec_enet_rx_queue(),
and adds a separate function fec_enet_rx_queue_xdp() for handling XDP
frames.
2. For TX XDP packets, using the batch sending method to avoid frequent
MMIO writes.
3. Use the switch statement to check the tx_buf type instead of the
if...else... statement, making the cleanup logic of TX BD ring cleared
and more efficient.
We compared the performance of XDP copy mode before and after applying
this patch set, and the results show that the performance has improved.
Before applying this patch set.
root@imx93evk:~# ./xdp-bench tx eth0
Summary 396,868 rx/s 0 err,drop/s
Summary 396,024 rx/s 0 err,drop/s
root@imx93evk:~# ./xdp-bench drop eth0
Summary 684,781 rx/s 0 err/s
Summary 675,746 rx/s 0 err/s
root@imx93evk:~# ./xdp-bench pass eth0
Summary 208,552 rx/s 0 err,drop/s
Summary 208,654 rx/s 0 err,drop/s
root@imx93evk:~# ./xdp-bench redirect eth0 eth0
eth0->eth0 311,210 rx/s 0 err,drop/s 311,208 xmit/s
eth0->eth0 310,808 rx/s 0 err,drop/s 310,809 xmit/s
After applying this patch set.
root@imx93evk:~# ./xdp-bench tx eth0
Summary 425,778 rx/s 0 err,drop/s
Summary 426,042 rx/s 0 err,drop/s
root@imx93evk:~# ./xdp-bench drop eth0
Summary 698,351 rx/s 0 err/s
Summary 701,882 rx/s 0 err/s
root@imx93evk:~# ./xdp-bench pass eth0
Summary 210,348 rx/s 0 err,drop/s
Summary 210,016 rx/s 0 err,drop/s
root@imx93evk:~# ./xdp-bench redirect eth0 eth0
eth0->eth0 354,407 rx/s 0 err,drop/s 354,401 xmit/s
eth0->eth0 350,381 rx/s 0 err,drop/s 350,389 xmit/s
This patch set also addes the AF_XDP zero-copy support, and we tested
the performance on i.MX93 platform with xdpsock tool. The following is
the performance comparison of copy mode and zero-copy mode. It can be
seen that the performance of zero-copy mode is better than that of copy
mode.
1. MAC swap L2 forwarding
1.1 Zero-copy mode
root@imx93evk:~# ./xdpsock -i eth0 -l -z
sock0@eth0:0 l2fwd xdp-drv
pps pkts 1.00
rx 414715 415455
tx 414715 415455
1.2 Copy mode
root@imx93evk:~# ./xdpsock -i eth0 -l -c
sock0@eth0:0 l2fwd xdp-drv
pps pkts 1.00
rx 356396 356609
tx 356396 356609
2. TX only
2.1 Zero-copy mode
root@imx93evk:~# ./xdpsock -i eth0 -t -s 64 -z
sock0@eth0:0 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 1119573 1126720
2.2 Copy mode
root@imx93evk:~# ./xdpsock -i eth0 -t -s 64 -c
sock0@eth0:0 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 406864 407616
====================
Link: https://patch.msgid.link/20260205085742.2685134-1-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This patch adds AF_XDP zero-copy support for both TX and RX on the FEC
driver. It introduces new functions for XSK buffer allocation, RX/TX
queue processing in zero-copy mode, and XSK pool setup/teardown.
For RX, fec_alloc_rxq_buffers_zc() is added to allocate RX buffers from
XSK pool. And fec_enet_rx_queue_xsk() is used to process the frames from
the RX queue which is bound to the AF_XDP socket. Similar to the copy
mode, the zero-copy mode also supports XDP_TX, XDP_PASS, XDP_DROP and
XDP_REDIRECT actions. In addition, fec_enet_xsk_tx_xmit() is similar to
fec_enet_xdp_tx_xmit() and is used to handle XDP_TX action in zero-copy
mode.
For TX, there are two cases, one is the frames from the AF_XDP socket,
so fec_enet_xsk_xmit() is added to directly transmit the frames from
the socket and the buffer type is marked as FEC_TXBUF_T_XSK_XMIT. The
other one is the frames from the RX queue (XDP_TX action), the buffer
type is marked as FEC_TXBUF_T_XSK_TX. Therefore, fec_enet_tx_queue()
could correctly clean the TX queue base on the buffer type.
Also, some tests have been done on the i.MX93-EVK board with the xdpsock
tool, the following are the results.
Env: i.MX93 connects to a packet generator, the link speed is 1Gbps, and
flow-control is off. The RX packet size is 64 bytes including FCS. Only
one RX queue (CPU) is used to receive frames.
1. MAC swap L2 forwarding
1.1 Zero-copy mode
root@imx93evk:~# ./xdpsock -i eth0 -l -z
sock0@eth0:0 l2fwd xdp-drv
pps pkts 1.00
rx 414715 415455
tx 414715 415455
1.2 Copy mode
root@imx93evk:~# ./xdpsock -i eth0 -l -c
sock0@eth0:0 l2fwd xdp-drv
pps pkts 1.00
rx 356396 356609
tx 356396 356609
2. TX only
2.1 Zero-copy mode
root@imx93evk:~# ./xdpsock -i eth0 -t -s 64 -z
sock0@eth0:0 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 1119573 1126720
2.2 Copy mode
root@imx93evk:~# ./xdpsock -i eth0 -t -s 64 -c
sock0@eth0:0 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 406864 407616
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-16-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
To support AF_XDP zero-copy mode in the subsequent patch, the following
adjustments have been made to fec_tx_queue().
1. Change the parameters of fec_tx_queue().
2. Some variables are initialized at the time of declaration, and the
order of local variables is updated to follow the reverse xmas tree
style.
3. Remove the variable xdpf and add the variable tx_buf.
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-15-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently, the buffers of RX queue are allocated from the page pool. In
the subsequent patches to support XDP zero copy, the RX buffers will be
allocated from the UMEM. Therefore, extract fec_alloc_rxq_buffers_pp()
from fec_enet_alloc_rxq_buffers() and we will add another helper to
allocate RX buffers from UMEM for the XDP zero copy mode. In addition,
fec_alloc_rxq_buffers_pp() only initializes bdp->bufaddr and does not
initialize other fields of bdp, because these will be initialized in
fec_enet_bd_init().
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-14-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The tx_buf has three types: FEC_TXBUF_T_SKB, FEC_TXBUF_T_XDP_NDO and
FEC_TXBUF_T_XDP_TX. Currently, the driver uses 'if...else...' statements
to check the type and perform the corresponding processing. This is very
detrimental to future expansion. To support AF_XDP zero-copy mode, two
new types will be added in the future, continuing to use 'if...else...'
would be a very bad coding style. So the 'if...else...' statements in
the current driver are replaced with switch statements.
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-11-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The tx_buf pointer will not NULL when its type is FEC_TXBUF_T_XDP_NDO or
FEC_TXBUF_T_XDP_TX. If the type is FEC_TXBUF_T_SKB, dev_kfree_skb_any()
will do NULL pointer check. So it is unnecessary to do NULL pointer check
in fec_enet_bd_init() and fec_enet_tx_queue().
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-10-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Remove fec_enet_xdp_get_tx_queue() from fec_enet_xdp_tx_xmit() and add
the tx_qid parameter to it. Then, calculate the TX queue ID for XDP_TX
frames in fec_enet_rx_queue_xdp(). This way, the TX queue ID only needs
to be calculated once for XDP_TX frames during each NAPI polling. And
since the number of RX queues and TX queues in FEC is generally equal,
the RX queue ID can be directly used as the TX queue ID. In exceptional
cases, fec_enet_xdp_get_tx_queue() is used to calculate the TX queue ID.
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-8-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently, the processing of XDP path packets and protocol stack packets
are both mixed in fec_enet_rx_queue(), which makes the logic somewhat
confusing and debugging more difficult. Furthermore, some logic is not
needed by each other. Such as the kernel path does not need to call
xdp_init_buff(), XDP path does not support swap_buffer(), because
fec_enet_bpf() returns "-EOPNOTSUPP" for those platforms which need
swap_buffer()), and so on. This prevents XDP from achieving its maximum
performance. Therefore, XDP path packets processing has been separated
from fec_enet_rx_queue() by adding the fec_enet_rx_queue_xdp() function
to optimize XDP path logic and improve XDP performance.
The XDP performance on the iMX93 platform was compared before and after
applying this patch. Detailed results are as follows and we can see the
performance has been improved.
Env: i.MX93, packet size 64 bytes including FCS, only single core and RX
BD ring are used to receive packets, flow-control is off.
Before the patch is applied:
xdp-bench tx eth0
Summary 396,868 rx/s 0 err,drop/s
Summary 396,024 rx/s 0 err,drop/s
xdp-bench drop eth0
Summary 684,781 rx/s 0 err/s
Summary 675,746 rx/s 0 err/s
xdp-bench pass eth0
Summary 208,552 rx/s 0 err,drop/s
Summary 208,654 rx/s 0 err,drop/s
xdp-bench redirect eth0 eth0
eth0->eth0 311,210 rx/s 0 err,drop/s 311,208 xmit/s
eth0->eth0 310,808 rx/s 0 err,drop/s 310,809 xmit/s
After the patch is applied:
xdp-bench tx eth0
Summary 409,975 rx/s 0 err,drop/s
Summary 411,073 rx/s 0 err,drop/s
xdp-bench drop eth0
Summary 700,681 rx/s 0 err/s
Summary 698,102 rx/s 0 err/s
xdp-bench pass eth0
Summary 211,356 rx/s 0 err,drop/s
Summary 210,629 rx/s 0 err,drop/s
xdp-bench redirect eth0 eth0
eth0->eth0 320,351 rx/s 0 err,drop/s 320,348 xmit/s
eth0->eth0 318,988 rx/s 0 err,drop/s 318,988 xmit/s
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-7-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This patch has made the following adjustments to fec_enet_rx_queue().
1. The function parameters are modified to maintain the same style as
subsequently added XDP-related interfaces.
2. Some variables are initialized at the time of declaration, and the
order of local variables is updated to follow the reverse xmas tree
style.
3. Replace variable cbd_bufaddr with dma.
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-6-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The FEC of some platforms supports RX FIFO shift-16, it means the actual
frame data starts at bit 16 of the first word read from RX FIFO aligning
the Ethernet payload on a 32-bit boundary. The MAC writes two additional
bytes in front of each frame received into the RX FIFO. Currently, the
fec_enet_rx_queue() updates the data_start, sub_len and the rx_bytes
statistics by checking whether FEC_QUIRK_HAS_RACC is set. This makes the
code less concise, so rx_shift is added to represent the number of extra
bytes padded in front of the RX frame. Furthermore, when adding separate
RX handling functions for XDP copy mode and zero copy mode in the future,
it will no longer be necessary to check FEC_QUIRK_HAS_RACC to update the
corresponding variables.
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Link: https://patch.msgid.link/20260205085742.2685134-4-wei.fang@nxp.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Daniel Golle says:
====================
net: dsa: mxl-gsw1xx: setup polarities and validate chip
Now that common PHY properties make it easy to configure the SerDes RX
and TX polarities, use that for the SGMII/1000Base-X/2500Base-X port of
the MaxLinear GSW1xx switches.
Also, validate hardware in probe() function to make sure the switch is
actually present and MDIO communication works properly.
====================
Link: https://patch.msgid.link/cover.1769916962.git.daniel@makrotopia.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
No check for actually present hardware is being performed in the probe
function of the mxl-gsw1xx switch driver. So even if the switch isn't
present at the configured MDIO bus address the driver wrongly tells the
user that a "GSWIP version 0 mod 0" was found, outputting errors about
PHY capabilities not matching.
Read and validate the chip MANU_ID and PNUM_ID registers and output
information while probing, but return an error and abort probing in case
the hardware is not actually present.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Link: https://patch.msgid.link/3194d3d3bb0b51f08755d392e1fdf7bb6dc49608.1769916962.git.daniel@makrotopia.org
Reviewed-by: Vladimir Oltean <olteanv@gmail.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Configure SerDes (port 4) RX and TX polarities using the newly
introduced generic properties. The polarities are described at the port
level which equals the polarities of the external pins of the chip.
Note that the RX lane is inverted internally and the vendor driver
simply always sets bit GSW1XX_SGMII_PHY_RX0_CFG2_INVERT unconditionally
to end up with the correct (ie. as documented in datasheets) polarity at
the external pins.
In this sense, PHY_POLARITY_NORMAL denotes normal polarity for pins as
documented for the MRQFN 105-pin package (GSW120, GSW125, GSW140, GSW141
and GSW145 all use the same package and have identical pin layouts
except for TP port 2 and 3 being N/C on GSW12x):
pin B18 (TX0_P) positive signal of the differential SGMII data output pair
pin B19 (TX0_M) negative signal of the differential SGMII data output pair
pin B20 (RX0_P) positive signal of the differential SGMII data input pair
pin B21 (RX0_M) negative signal of the differential SGMII data input pair
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Link: https://patch.msgid.link/8bf79b3476e23673fceffbe2bc9d6abc13d132e5.1769916962.git.daniel@makrotopia.org
Reviewed-by: Vladimir Oltean <olteanv@gmail.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Jakub Kicinski says:
====================
net: stats, tools, driver tests for HW GRO [part]
Add miscellaneous pieces related to production use of HW-GRO:
- report standard stats from drivers (bnxt included here,
Gal recently posted patches for mlx5 which is great)
- CLI tool for calculating HW GRO savings / effectiveness
====================
Link: https://patch.msgid.link/20260207003509.3927744-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Extend ynltool to compute HW GRO savings metric - how many
packets has HW GRO been able to save the kernel from seeing.
Note that this definition does not actually take into account
whether the segments were or weren't eligible for HW GRO.
If a machine is receiving all-UDP traffic - new metric will show
HW-GRO savings of 0%. Conversely since the super-packet still
counts as a received packet, savings of 100% is not achievable.
Perfect HW-GRO on a machine with 4k MTU and 64kB super-frames
would show ~93.75% savings. With 1.5k MTU we may see up to
~97.8% savings (if my math is right).
Example after 10 sec of iperf on a freshly booted machine
with 1.5k MTU:
$ ynltool qstats show
eth0 rx-packets: 40681280 rx-bytes: 61575208437
rx-alloc-fail: 0 rx-hw-gro-packets: 1225133
rx-hw-gro-wire-packets: 40656633
$ ynltool qstats hw-gro
eth0: 96.9% savings
None of the NICs I have access to can report "missed" HW-GRO
opportunities so computing a true "effectiveness" metric
is not possible. One could also argue that effectiveness metric
is inferior in environments where we control both senders and
receivers, the savings metrics will capture both regressions
in receiver's HW GRO effectiveness but also regressions in senders
sending smaller TSO trains. And we care about both. The main
downside is that it's hard to tell at a glance how well the NIC
is doing because the savings will be dependent on traffic patterns.
Reviewed-by: Petr Machata <petrm@nvidia.com>
Link: https://patch.msgid.link/20260207003509.3927744-4-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Count and report HW-GRO stats as seen by the kernel.
The device stats for GRO seem to not reflect the reality,
perhaps they count sessions which did not actually result
in any aggregation. Also they count wire packets, so we
have to count super-frames, anyway.
Reviewed-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20260207003509.3927744-2-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Alice Mikityanska says:
====================
BIG TCP without HBH in IPv6
Resubmitting after the grace period.
This series is part 1 of "BIG TCP for UDP tunnels". Due to the number of
patches, I'm splitting it into two logical parts:
* Remove hop-by-hop header for BIG TCP IPv6 to align with BIG TCP IPv4.
* Fix up things that prevent BIG TCP from working with UDP tunnels.
The current BIG TCP IPv6 code inserts a hop-by-hop extension header with
32-bit length of the packet. When the packet is encapsulated, and either
the outer or the inner protocol is IPv6, or both are IPv6, there will be
1 or 2 HBH headers that need to be dealt with. The issues that arise:
1. The drivers don't strip it, and they'd all need to know the structure
of each tunnel protocol in order to strip it correctly, also taking into
account all combinations of IPv4/IPv6 inner/outer protocols.
2. Even if (1) is implemented, it would be an additional performance
penalty per aggregated packet.
3. The skb_gso_validate_network_len check is skipped in
ip6_finish_output_gso when IP6SKB_FAKEJUMBO is set, but it seems that it
would make sense to do the actual validation, just taking into account
the length of the HBH header. When the support for tunnels is added, it
becomes trickier, because there may be one or two HBH headers, depending
on whether it's IPv6 in IPv6 or not.
At the same time, having an HBH header to store the 32-bit length is not
strictly necessary, as BIG TCP IPv4 doesn't do anything like this and
just restores the length from skb->len. The same thing can be done for
BIG TCP IPv6. Removing HBH from BIG TCP would allow to simplify the
implementation significantly, and align it with BIG TCP IPv4, which has
been a long-standing goal.
====================
Link: https://patch.msgid.link/20260205133925.526371-1-alice.kernel@fastmail.im
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
BIG TCP IPv6 inserts a hop-by-hop extension header to indicate the real
IPv6 payload length when it doesn't fit into the 16-bit field in the
IPv6 header itself. While it helps tools parse the packet, it also
requires every driver that supports TSO and BIG TCP to remove this
8-byte extension header. It might not sound that bad until we try to
apply it to tunneled traffic. Currently, the drivers don't attempt to
strip HBH if skb->encapsulation = 1. Moreover, trying to do so would
require dissecting different tunnel protocols and making corresponding
adjustments on case-by-case basis, which would slow down the fastpath
(potentially also requiring adjusting checksums in outer headers).
At the same time, BIG TCP IPv4 doesn't insert any extra headers and just
calculates the payload length from skb->len, significantly simplifying
implementing BIG TCP for tunnels.
Stop inserting HBH when building BIG TCP GSO SKBs.
Signed-off-by: Alice Mikityanska <alice@isovalent.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260205133925.526371-3-alice.kernel@fastmail.im
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The next commits will transition away from using the hop-by-hop
extension header to encode packet length for BIG TCP. Add wrappers
around ip6->payload_len that return the actual value if it's non-zero,
and calculate it from skb->len if payload_len is set to zero (and a
symmetrical setter).
The new helpers are used wherever the surrounding code supports the
hop-by-hop jumbo header for BIG TCP IPv6, or the corresponding IPv4 code
uses skb_ip_totlen (e.g., in include/net/netfilter/nf_tables_ipv6.h).
No behavioral change in this commit.
Signed-off-by: Alice Mikityanska <alice@isovalent.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260205133925.526371-2-alice.kernel@fastmail.im
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Ivan Vecera says:
====================
dpll: zl3073x: Include current frequency in supported frequencies list
This series ensures that the current operating frequency of a DPLL pin
is always reported in its supported frequencies list.
Problem:
When a ZL3073x DPLL pin is registered, its supported frequencies are
read from the firmware node's "supported-frequencies-hz" property.
However, if the firmware node is missing, or doesn't include the
current operating frequency, the pin reports a frequency that isn't
in its supported list. This inconsistency can confuse userspace tools
that expect the current frequency to be among the supported values.
Solution:
Always include the current pin frequency as the first entry in the
supported frequencies list, followed by any additional frequencies
from the firmware node (with duplicates filtered out).
Patch 1 refactors the output pin frequency calculation into a reusable
helper function zl3073x_dev_output_pin_freq_get(), which mirrors the
existing zl3073x_dev_ref_freq_get() for input pins.
Patch 2 modifies zl3073x_pin_props_get() to obtain the current
frequency early and place it at index 0 of the supported frequencies
array, ensuring it is always present regardless of firmware node
contents.
====================
Link: https://patch.msgid.link/20260205154350.3180465-1-ivecera@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Ensure the current pin frequency is always present in the list of
supported frequencies reported to userspace. Previously, if the
firmware node was missing or didn't include the current operating
frequency in the supported-frequencies-hz property, the pin would
report a frequency that wasn't in its supported list.
Get the current frequency early in zl3073x_pin_props_get():
- For input pins: use zl3073x_dev_ref_freq_get()
- For output pins: use zl3073x_dev_output_pin_freq_get()
Place the current frequency at index 0 of the supported frequencies
array, then append frequencies from the firmware node (if present),
skipping any duplicate of the current frequency.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260205154350.3180465-3-ivecera@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce zl3073x_dev_output_pin_freq_get() helper function to compute
the output pin frequency based on synthesizer frequency, output divisor,
and signal format. For N-div signal formats, the N-pin frequency is
additionally divided by esync_n_period.
Add zl3073x_out_is_ndiv() helper to check if an output is configured
in N-div mode (2_NDIV or 2_NDIV_INV signal formats).
Refactor zl3073x_dpll_output_pin_frequency_get() callback to use the
new helper, reducing code duplication and enabling reuse of the
frequency calculation logic in other contexts.
This is a preparatory change for adding current frequency to the
supported frequencies list in pin properties.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260205154350.3180465-2-ivecera@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, both the icssg-prueth and icssg-prueth-sr1 drivers create
a dedicated 'emac->cmd_wq' workqueue.
In the icssg-prueth-sr1 driver, this workqueue is not utilized at all.
In the icssg-prueth driver, the workqueue is only used to execute the
actual processing of ndo_set_rx_mode. However, creating a dedicated
workqueue for such a simple use case is unnecessary. To simplify the
code, switch to using the system default workqueue instead.
Signed-off-by: Kevin Hao <haokexin@gmail.com>
Tested-by: Meghana Malladi <m-malladi@ti.com>
Reviewed-by: MD Danish Anwar <danishanwar@ti.com>
Link: https://patch.msgid.link/20260205-icssg-prueth-workqueue-v2-1-cf5cf97efb37@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
