When a packet enters OVS datapath and there is no flow to handle it,
packet goes to userspace through a MISS upcall. With per-CPU upcall
dispatch mechanism, we're using the current CPU id to select the
Netlink PID on which to send this packet. This allows us to send
packets from the same traffic flow through the same handler.
The handler will process the packet, install required flow into the
kernel and re-inject the original packet via OVS_PACKET_CMD_EXECUTE.
While handling OVS_PACKET_CMD_EXECUTE, however, we may hit a
recirculation action that will pass the (likely modified) packet
through the flow lookup again. And if the flow is not found, the
packet will be sent to userspace again through another MISS upcall.
However, the handler thread in userspace is likely running on a
different CPU core, and the OVS_PACKET_CMD_EXECUTE request is handled
in the syscall context of that thread. So, when the time comes to
send the packet through another upcall, the per-CPU dispatch will
choose a different Netlink PID, and this packet will end up processed
by a different handler thread on a different CPU.
The process continues as long as there are new recirculations, each
time the packet goes to a different handler thread before it is sent
out of the OVS datapath to the destination port. In real setups the
number of recirculations can go up to 4 or 5, sometimes more.
There is always a chance to re-order packets while processing upcalls,
because userspace will first install the flow and then re-inject the
original packet. So, there is a race window when the flow is already
installed and the second packet can match it and be forwarded to the
destination before the first packet is re-injected. But the fact that
packets are going through multiple upcalls handled by different
userspace threads makes the reordering noticeably more likely, because
we not only have a race between the kernel and a userspace handler
(which is hard to avoid), but also between multiple userspace handlers.
For example, let's assume that 10 packets got enqueued through a MISS
upcall for handler-1, it will start processing them, will install the
flow into the kernel and start re-injecting packets back, from where
they will go through another MISS to handler-2. Handler-2 will install
the flow into the kernel and start re-injecting the packets, while
handler-1 continues to re-inject the last of the 10 packets, they will
hit the flow installed by handler-2 and be forwarded without going to
the handler-2, while handler-2 still re-injects the first of these 10
packets. Given multiple recirculations and misses, these 10 packets
may end up completely mixed up on the output from the datapath.
Let's allow userspace to specify on which Netlink PID the packets
should be upcalled while processing OVS_PACKET_CMD_EXECUTE.
This makes it possible to ensure that all the packets are processed
by the same handler thread in the userspace even with them being
upcalled multiple times in the process. Packets will remain in order
since they will be enqueued to the same socket and re-injected in the
same order. This doesn't eliminate re-ordering as stated above, since
we still have a race between kernel and the userspace thread, but it
allows to eliminate races between multiple userspace threads.
Userspace knows the PID of the socket on which the original upcall is
received, so there is no need to send it up from the kernel.
Solution requires storing the value somewhere for the duration of the
packet processing. There are two potential places for this: our skb
extension or the per-CPU storage. It's not clear which is better,
so just following currently used scheme of storing this kind of things
along the skb. We still have a decent amount of space in the cb.
Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
Acked-by: Flavio Leitner <fbl@sysclose.org>
Acked-by: Eelco Chaudron <echaudro@redhat.com>
Acked-by: Aaron Conole <aconole@redhat.com>
Link: https://patch.msgid.link/20250702155043.2331772-1-i.maximets@ovn.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pull networking fixes from Paolo Abeni:
"Including fixes from Bluetooth.
Current release - new code bugs:
- eth:
- txgbe: fix the issue of TX failure
- ngbe: specify IRQ vector when the number of VFs is 7
Previous releases - regressions:
- sched: always pass notifications when child class becomes empty
- ipv4: fix stat increase when udp early demux drops the packet
- bluetooth: prevent unintended pause by checking if advertising is active
- virtio: fix error reporting in virtqueue_resize
- eth:
- virtio-net:
- ensure the received length does not exceed allocated size
- fix the xsk frame's length check
- lan78xx: fix WARN in __netif_napi_del_locked on disconnect
Previous releases - always broken:
- bluetooth: mesh: check instances prior disabling advertising
- eth:
- idpf: convert control queue mutex to a spinlock
- dpaa2: fix xdp_rxq_info leak
- amd-xgbe: align CL37 AN sequence as per databook"
* tag 'net-6.16-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (38 commits)
vsock/vmci: Clear the vmci transport packet properly when initializing it
dt-bindings: net: sophgo,sg2044-dwmac: Drop status from the example
net: ngbe: specify IRQ vector when the number of VFs is 7
net: wangxun: revert the adjustment of the IRQ vector sequence
net: txgbe: request MISC IRQ in ndo_open
virtio_net: Enforce minimum TX ring size for reliability
virtio_net: Cleanup '2+MAX_SKB_FRAGS'
virtio_ring: Fix error reporting in virtqueue_resize
virtio-net: xsk: rx: fix the frame's length check
virtio-net: use the check_mergeable_len helper
virtio-net: remove redundant truesize check with PAGE_SIZE
virtio-net: ensure the received length does not exceed allocated size
net: ipv4: fix stat increase when udp early demux drops the packet
net: libwx: fix the incorrect display of the queue number
amd-xgbe: do not double read link status
net/sched: Always pass notifications when child class becomes empty
nui: Fix dma_mapping_error() check
rose: fix dangling neighbour pointers in rose_rt_device_down()
enic: fix incorrect MTU comparison in enic_change_mtu()
amd-xgbe: align CL37 AN sequence as per databook
...
Pull xfs fixes from Carlos Maiolino:
- Fix umount hang with unflushable inodes (and add new tracepoint used
for debugging this)
- Fix ABBA deadlock in xfs_reclaim_inode() vs xfs_ifree_cluster()
- Fix dquot buffer pin deadlock
* tag 'xfs-fixes-6.16-rc5' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
xfs: add FALLOC_FL_ALLOCATE_RANGE to supported flags mask
xfs: fix unmount hang with unflushable inodes stuck in the AIL
xfs: factor out stale buffer item completion
xfs: rearrange code in xfs_buf_item.c
xfs: add tracepoints for stale pinned inode state debug
xfs: avoid dquot buffer pin deadlock
xfs: catch stale AGF/AGF metadata
xfs: xfs_ifree_cluster vs xfs_iflush_shutdown_abort deadlock
xfs: actually use the xfs_growfs_check_rtgeom tracepoint
xfs: Improve error handling in xfs_mru_cache_create()
xfs: move xfs_submit_zoned_bio a bit
xfs: use xfs_readonly_buftarg in xfs_remount_rw
xfs: remove NULL pointer checks in xfs_mru_cache_insert
xfs: check for shutdown before going to sleep in xfs_select_zone
Thomas Gleixner says:
====================
ptp: Provide support for auxiliary clocks for PTP_SYS_OFFSET_EXTENDED
This is a follow up to the V1 series, which can be found here:
https://lore.kernel.org/all/20250626124327.667087805@linutronix.de
to address the merge logistics problem, which I created myself.
Changes vs. V1:
- Make patch 1, which provides the timestamping function temporarily
define CLOCK_AUX* if undefined so that it can be merged independently,
- Add a missing check for CONFIG_POSIX_AUX_CLOCK in the PTP IOCTL
- Picked up tags
Merge logistics if agreed on:
1) Patch #1 is applied to the tip tree on top of plain v6.16-rc1 and
tagged
2) That tag is merged into tip:timers/ptp and the temporary CLOCK_AUX
define is removed in a subsequent commit
3) Network folks merge the tag and apply patches #2 + #3
So the only fallout from this are the extra merges in both trees and the
cleanup commit in the tip tree. But that way there are no dependencies and
no duplicate commits with different SHAs.
Thoughts?
Due to the above constraints there is no branch offered to pull from right
now. Sorry for the inconveniance. Should have thought about that earlier.
====================
Link: https://patch.msgid.link/20250701130923.579834908@linutronix.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Allow ioctl(PTP_SYS_OFFSET_EXTENDED*) to select CLOCK_AUX clock ids for
generating the pre and post hardware readout timestamps.
Aside of adding these clocks to the clock ID validation, this also requires
to check the timestamp to be valid, i.e. the seconds value being greater
than or equal zero. This is necessary because AUX clocks can be
asynchronously enabled or disabled, so there is no way to validate the
availability upfront.
The same could have been achieved by handing the return value of
ktime_get_aux_ts64() all the way down to the IOCTL call site, but that'd
require to modify all existing ptp::gettimex64() callbacks and their inner
call chains. The timestamp check achieves the same with less churn and less
complicated code all over the place.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Link: https://patch.msgid.link/20250701132628.491315452@linutronix.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The inlined ptp_read_system_[pre|post]ts() switch cases expand to a copious
amount of text in drivers, e.g. ~500 bytes in e1000e. Adding auxiliary
clock support to the inlines would increase it further.
Replace the inline switch case with a call to ktime_get_clock_ts64(), which
reduces the code size in drivers and allows to access auxiliary clocks once
they are enabled in the IOCTL parameter filter.
No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Acked-by: John Stultz <jstultz@google.com>
Link: https://patch.msgid.link/20250701132628.426168092@linutronix.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Base implementation for PTP with a temporary CLOCK_AUX* workaround to
allow integration of depending changes into the networking tree.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The timer which ensures that no more than 3 LACPDUs are transmitted in
a second rearms itself every 333ms regardless of whether an LACPDU is
transmitted when the timer expires. This causes LACPDU tx to be delayed
until the next expiration of the timer, which effectively aligns LACPDUs
to ~333ms boundaries. This results in a variable amount of jitter in the
timing of periodic LACPDUs.
Change this to only rearm the timer when an LACPDU is actually sent,
allowing tx at any point after the timer has expired.
Signed-off-by: Seth Forshee (DigitalOcean) <sforshee@kernel.org>
Reviewed-by: Carlos Bilbao <carlos.bilbao@kernel.org>
Link: https://patch.msgid.link/20250625-fix-lacpdu-jitter-v1-1-4d0ee627e1ba@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
PTP implements an inline switch case for taking timestamps from various
POSIX clock IDs, which already consumes quite some text space. Expanding it
for auxiliary clocks really becomes too big for inlining.
Provide a out of line version.
The function invalidates the timestamp in case the clock is invalid. The
invalidation allows to implement a validation check without the need to
propagate a return value through deep existing call chains.
Due to merge logistics this temporarily defines CLOCK_AUX[_LAST] if
undefined, so that the plain branch, which does not contain any of the core
timekeeper changes, can be pulled into the networking tree as prerequisite
for the PTP side changes. These temporary defines are removed after that
branch is merged into the tip::timers/ptp branch. That way the result in
-next or upstream in the next merge window has zero dependencies.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Acked-by: John Stultz <jstultz@google.com>
Link: https://lore.kernel.org/all/20250701132628.357686408@linutronix.de
Jiawen Wu says:
====================
Fix IRQ vectors
The interrupt vector order was adjusted by [1]commit 937d46ecc5 ("net:
wangxun: add ethtool_ops for channel number") in Linux-6.8. Because at
that time, the MISC interrupt acts as the parent interrupt in the GPIO
IRQ chip. When the number of Rx/Tx ring changes, the last MISC
interrupt must be reallocated. Then the GPIO interrupt controller would
be corrupted. So the initial plan was to adjust the sequence of the
interrupt vectors, let MISC interrupt to be the first one and do not
free it.
Later, irq_domain was introduced in [2]commit aefd013624 ("net: txgbe:
use irq_domain for interrupt controller") to avoid this problem.
However, the vector sequence adjustment was not reverted. So there is
still one problem that has been left unresolved.
Due to hardware limitations of NGBE, queue IRQs can only be requested
on vector 0 to 7. When the number of queues is set to the maximum 8,
the PCI IRQ vectors are allocated from 0 to 8. The vector 0 is used by
MISC interrupt, and althrough the vector 8 is used by queue interrupt,
it is unable to receive packets. This will cause some packets to be
dropped when RSS is enabled and they are assigned to queue 8.
This patch set fix the above problems.
[1] https://git.kernel.org/netdev/net-next/c/937d46ecc5f9
[2] https://git.kernel.org/netdev/net-next/c/aefd013624a1
====================
Link: https://patch.msgid.link/20250701063030.59340-1-jiawenwu@trustnetic.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
For NGBE devices, the queue number is limited to be 1 when SRIOV is
enabled. In this case, IRQ vector[0] is used for MISC and vector[1] is
used for queue, based on the previous patches. But for the hardware
design, the IRQ vector[1] must be allocated for use by the VF[6] when
the number of VFs is 7. So the IRQ vector[0] should be shared for PF
MISC and QUEUE interrupts.
+-----------+----------------------+
| Vector | Assigned To |
+-----------+----------------------+
| Vector 0 | PF MISC and QUEUE |
| Vector 1 | VF 6 |
| Vector 2 | VF 5 |
| Vector 3 | VF 4 |
| Vector 4 | VF 3 |
| Vector 5 | VF 2 |
| Vector 6 | VF 1 |
| Vector 7 | VF 0 |
+-----------+----------------------+
Minimize code modifications, only adjust the IRQ vector number for this
case.
Fixes: 877253d2cb ("net: ngbe: add sriov function support")
Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
Reviewed-by: Larysa Zaremba <larysa.zaremba@intel.com>
Link: https://patch.msgid.link/20250701063030.59340-4-jiawenwu@trustnetic.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Due to hardware limitations of NGBE, queue IRQs can only be requested
on vector 0 to 7. When the number of queues is set to the maximum 8,
the PCI IRQ vectors are allocated from 0 to 8. The vector 0 is used by
MISC interrupt, and althrough the vector 8 is used by queue interrupt,
it is unable to receive packets. This will cause some packets to be
dropped when RSS is enabled and they are assigned to queue 8.
So revert the adjustment of the MISC IRQ location, to make it be the
last one in IRQ vectors.
Fixes: 937d46ecc5 ("net: wangxun: add ethtool_ops for channel number")
Cc: stable@vger.kernel.org
Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
Reviewed-by: Larysa Zaremba <larysa.zaremba@intel.com>
Link: https://patch.msgid.link/20250701063030.59340-3-jiawenwu@trustnetic.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Laurent Vivier says:
====================
virtio: Fixes for TX ring sizing and resize error reporting
This patch series contains two fixes and a cleanup for the virtio subsystem.
The first patch fixes an error reporting bug in virtio_ring's
virtqueue_resize() function. Previously, errors from internal resize
helpers could be masked if the subsequent re-enabling of the virtqueue
succeeded. This patch restores the correct error propagation, ensuring that
callers of virtqueue_resize() are properly informed of underlying resize
failures.
The second patch does a cleanup of the use of '2+MAX_SKB_FRAGS'
The third patch addresses a reliability issue in virtio_net where the TX
ring size could be configured too small, potentially leading to
persistently stopped queues and degraded performance. It enforces a
minimum TX ring size to ensure there's always enough space for at least one
maximally-fragmented packet plus an additional slot.
====================
Link: https://patch.msgid.link/20250521092236.661410-1-lvivier@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The `tx_may_stop()` logic stops TX queues if free descriptors
(`sq->vq->num_free`) fall below the threshold of (`MAX_SKB_FRAGS` + 2).
If the total ring size (`ring_num`) is not strictly greater than this
value, queues can become persistently stopped or stop after minimal
use, severely degrading performance.
A single sk_buff transmission typically requires descriptors for:
- The virtio_net_hdr (1 descriptor)
- The sk_buff's linear data (head) (1 descriptor)
- Paged fragments (up to MAX_SKB_FRAGS descriptors)
This patch enforces that the TX ring size ('ring_num') must be strictly
greater than (MAX_SKB_FRAGS + 2). This ensures that the ring is
always large enough to hold at least one maximally-fragmented packet
plus at least one additional slot.
Reported-by: Lei Yang <leiyang@redhat.com>
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Link: https://patch.msgid.link/20250521092236.661410-4-lvivier@redhat.com
Tested-by: Lei Yang <leiyang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The virtqueue_resize() function was not correctly propagating error codes
from its internal resize helper functions, specifically
virtqueue_resize_packet() and virtqueue_resize_split(). If these helpers
returned an error, but the subsequent call to virtqueue_enable_after_reset()
succeeded, the original error from the resize operation would be masked.
Consequently, virtqueue_resize() could incorrectly report success to its
caller despite an underlying resize failure.
This change restores the original code behavior:
if (vdev->config->enable_vq_after_reset(_vq))
return -EBUSY;
return err;
Fix: commit ad48d53b5b ("virtio_ring: separate the logic of reset/enable from virtqueue_resize")
Cc: xuanzhuo@linux.alibaba.com
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Link: https://patch.msgid.link/20250521092236.661410-2-lvivier@redhat.com
Tested-by: Lei Yang <leiyang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When calling buf_to_xdp, the len argument is the frame data's length
without virtio header's length (vi->hdr_len). We check that len with
xsk_pool_get_rx_frame_size() + vi->hdr_len
to ensure the provided len does not larger than the allocated chunk
size. The additional vi->hdr_len is because in virtnet_add_recvbuf_xsk,
we use part of XDP_PACKET_HEADROOM for virtio header and ask the vhost
to start placing data from
hard_start + XDP_PACKET_HEADROOM - vi->hdr_len
not
hard_start + XDP_PACKET_HEADROOM
But the first buffer has virtio_header, so the maximum frame's length in
the first buffer can only be
xsk_pool_get_rx_frame_size()
not
xsk_pool_get_rx_frame_size() + vi->hdr_len
like in the current check.
This commit adds an additional argument to buf_to_xdp differentiate
between the first buffer and other ones to correctly calculate the maximum
frame's length.
Cc: stable@vger.kernel.org
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Fixes: a4e7ba7027 ("virtio_net: xsk: rx: support recv small mode")
Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
Link: https://patch.msgid.link/20250630151315.86722-2-minhquangbui99@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Bui Quang Minh says:
====================
virtio-net: fixes for mergeable XDP receive path
This series contains fixes for XDP receive path in virtio-net
- Patch 1: add a missing check for the received data length with our
allocated buffer size in mergeable mode.
- Patch 2: remove a redundant truesize check with PAGE_SIZE in mergeable
mode
- Patch 3: make the current repeated code use the check_mergeable_len to
check for received data length in mergeable mode
====================
Link: https://patch.msgid.link/20250630144212.48471-1-minhquangbui99@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Mark Bloch says:
====================
Support rate management on traffic classes in devlink and mlx5
This patch series extends the devlink-rate API to support traffic class
(TC) bandwidth management, enabling more granular control over traffic
shaping and rate limiting across multiple TCs. The API now allows users
to specify bandwidth proportions for different traffic classes in a
single command. This is particularly useful for managing Enhanced
Transmission Selection (ETS) for groups of Virtual Functions (VFs),
allowing precise bandwidth allocation across traffic classes.
Additionally the series refines the QoS handling in net/mlx5 to support
TC arbitration and bandwidth management on vports and rate nodes.
Discussions on traffic class shaping in net-shapers began in V5 [1],
where we discussed with maintainers whether net-shapers should support
traffic classes and how this could be implemented.
Later, after further conversations with Paolo Abeni and Simon Horman,
Cosmin provided an update [2], confirming that net-shapers' tree-based
hierarchy aligns well with traffic classes when treated as distinct
subsets of netdev queues. Since mlx5 enforces a 1:1 mapping between TX
queues and traffic classes, this approach seems feasible, though some
open questions remain regarding queue reconfiguration and certain mlx5
scheduling behaviors.
Building on that discussion, Cosmin has now shared a concrete
implementation plan on the netdev mailing list [3]. The plan, developed
in collaboration with Paolo and Simon, outlines how net-shapers can be
extended to support the same use cases currently covered by
devlink-rate, with the eventual goal of aligning both and simplifying
the shaping infrastructure in the kernel.
This work was presented at Netdev 0x19 in Zagreb [4].
There we presented how TC scheduling is enforced in mlx5 hardware,
which led to discussions on the mailing list.
A summary of how things work:
Classification means labeling a packet with a traffic class based on
the packet's DSCP or VLAN PCP field, then treating packets with
different traffic classes differently during transmit processing.
In a virtualized setup, VFs are untrusted and do not control
classification or shaping. Classification is done by the hardware using
a prio-to-TC mapping set by the hypervisor. VFs only select which send
queue to use and are expected to respect the classification logic by
sending each traffic class on its dedicated queue. As stated in the
net-shapers plan [3], each transmit queue should carry only a single
traffic class. Mixing classes in a single queue can lead to HOL
blocking.
In the mlx5 implementation, if the queue used does not match the
classified traffic class, the hardware moves the queue to the correct
TC scheduler. This movement is not a reclassification; it’s a necessary
enforcement step to ensure traffic class isolation is maintained.
Extend devlink-rate API to support rate management on TCs:
- devlink: Extend the devlink rate API to support traffic class
bandwidth management
Introduce a no-op implementation:
- net/mlx5: Add no-op implementation for setting tc-bw on rate objects
Add support for enabling and disabling TC QoS on vports and nodes:
- net/mlx5: Add support for setting tc-bw on nodes
- net/mlx5: Add traffic class scheduling support for vport QoS
Support for setting tc-bw on rate objects:
- net/mlx5: Manage TC arbiter nodes and implement full support for
tc-bw
[1]
https://lore.kernel.org/netdev/20241204220931.254964-1-tariqt@nvidia.com/
[2]
https://lore.kernel.org/netdev/67df1a562614b553dcab043f347a0d7c5393ff83.camel@nvidia.com/
[3]
https://lore.kernel.org/netdev/d9831d0c940a7b77419abe7c7330e822bbfd1cfb.camel@nvidia.com/T/
[4]
https://netdevconf.info/0x19/sessions/talk/optimizing-bandwidth-allocation-with-ets-and-traffic-classes.html
====================
Link: https://patch.msgid.link/20250629142138.361537-1-mbloch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This test suite validates the functionality of the devlink-rate API for
traffic class (TC) bandwidth allocation. It ensures that bandwidth can
be distributed between different traffic classes as configured, and
verifies that explicit TC-to-queue mapping is required for the
allocation to be effective.
The first test (test_no_tc_mapping_bandwidth) is marked as expected
failure on mlx5, since the hardware automatically enforces traffic
class separation by dynamically moving queues to the correct TC
scheduler, even without explicit TC-to-queue mapping configuration.
Test output on mlx5:
1..2
# Created VF interface: eth5
# Created VLAN eth5.101 on eth5 with tc 3 and IP 198.51.100.2
# Created VLAN eth5.102 on eth5 with tc 4 and IP 198.51.100.10
# Set representor eth4 up and added to bridge
# Bandwidth check results without TC mapping:
# TC 3: 0.19 Gbits/sec
# TC 4: 0.76 Gbits/sec
# Total bandwidth: 0.95 Gbits/sec
# TC 3 percentage: 20.0%
# TC 4 percentage: 80.0%
ok 1 devlink_rate_tc_bw.test_no_tc_mapping_bandwidth # XFAIL Bandwidth matched 80/20 split without TC mapping
# Created VF interface: eth5
# Created VLAN eth5.101 on eth5 with tc 3 and IP 198.51.100.2
# Created VLAN eth5.102 on eth5 with tc 4 and IP 198.51.100.10
# Set representor eth4 up and added to bridge
# Bandwidth check results with TC mapping:
# TC 3: 0.21 Gbits/sec
# TC 4: 0.78 Gbits/sec
# Total bandwidth: 0.98 Gbits/sec
# TC 3 percentage: 21.1%
# TC 4 percentage: 78.9%
# Bandwidth is distributed as 80/20 with TC mapping
ok 2 devlink_rate_tc_bw.test_tc_mapping_bandwidth
# Totals: pass:1 fail:0 xfail:1 xpass:0 skip:0 error:0
Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/20250629142138.361537-9-mbloch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce support for managing Traffic Class (TC) arbiter nodes and
associated vports TC nodes within the E-Switch QoS hierarchy. This
patch adds support for the new scheduling node type,
`SCHED_NODE_TYPE_VPORTS_TC_TSAR`, and implements full support for
setting tc-bw on both vports and nodes.
Key changes include:
- Introduced the new scheduling node type,
`SCHED_NODE_TYPE_VPORTS_TC_TSAR`, for managing vports within the TC
arbiter node.
- New helper functions for creating and destroying vports TC nodes
under the TC arbiter.
- Updated the minimum rate normalization function to skip nodes of type
`SCHED_NODE_TYPE_VPORTS_TC_TSAR`. Vports TC TSARs have bandwidth
shares configured on them but not minimum rates, so their `min_rate`
cannot be normalized.
- Implementation of `esw_qos_tc_arbiter_scheduling_setup()` and
`esw_qos_tc_arbiter_scheduling_teardown()` for initializing and
cleaning up TC arbiter scheduling elements. These functions now fully
support tc-bw configuration on TC arbiter nodes.
- Introduced a new helper `esw_qos_calculate_tc_bw_divider()` to
compute the total TC bandwidth share, which is used as a divider for
normalizing each TC's share.
- Added `esw_qos_tc_arbiter_get_bw_shares()` and
`esw_qos_set_tc_arbiter_bw_shares()` to handle the settings of
bandwidth shares for vports traffic class TSARs.
- `esw_qos_set_tc_arbiter_bw_shares()` normalizes each TC share based
on the total and the firmware's maximum allowed TSAR bandwidth share.
- Refactored `mlx5_esw_devlink_rate_node_tc_bw_set()` and
`mlx5_esw_devlink_rate_leaf_tc_bw_set()` to fully support configuring
tc-bw on devlink rate nodes and vports, respectively.
- Refactored `mlx5_esw_qos_node_update_parent()` to ensure that tc-bw
configuration remains compatible with setting a parent on a rate
node, preserving level hierarchy functionality.
- Refactored `esw_qos_calc_bw_share()` to generalize its input so it
can be used for both minimum rate and bandwidth share calculations.
Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/20250629142138.361537-8-mbloch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce support for traffic class (TC) scheduling on vports by
allowing the vport to own multiple TC scheduling nodes. This patch
enables more granular control of QoS by defining three distinct QoS
states for vports, each providing unique scheduling behavior:
1. Regular QoS: The `sched_node` represents the vport directly,
handling QoS as a single scheduling entity.
2. TC QoS on the vport: The `sched_node` acts as a TC arbiter, enabling
TC scheduling directly on the vport.
3. TC QoS on the parent node: The `sched_node` functions as a rate
limiter, with TC arbitration enabled at the parent level, associating
multiple scheduling nodes with each vport.
Key changes include:
- Added support for new scheduling elements, vport traffic class and
rate limiter.
- New helper functions for creating, destroying, and restoring vport TC
scheduling nodes, handling transitions between regular QoS and TC
arbitration states.
- Updated `esw_qos_vport_enable()` and `esw_qos_vport_disable()` to
support both regular QoS and TC arbitration states, ensuring consistent
transitions between scheduling modes.
- Introduced a `sched_nodes` array under `vport->qos` to store multiple
TC scheduling nodes per vport, enabling finer control over per-TC QoS.
- Enhanced `esw_qos_vport_update_parent()` to handle transitions between
the three QoS states based on the current and new parent node types.
This patch lays the groundwork for future support for configuring tc-bw
on vports. Although the infrastructure is in place, full support for
tc-bw is not yet implemented; attempts to set tc-bw on vports will
return `-EOPNOTSUPP`.
No functional changes are introduced at this stage.
Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/20250629142138.361537-7-mbloch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce support for enabling and disabling Traffic Class (TC)
arbitration for existing devlink rate nodes. This patch adds support
for a new scheduling node type, `SCHED_NODE_TYPE_TC_ARBITER_TSAR`.
Key changes include:
- New helper functions for transitioning existing rate nodes to TC
arbiter nodes and vice versa. These functions handle the allocation
of TC arbiter nodes, copying of child nodes, and restoring vport QoS
settings when TC arbitration is disabled.
- Implementation of `mlx5_esw_devlink_rate_node_tc_bw_set()` to manage
tc-bw configuration on nodes.
- Introduced stubs for `esw_qos_tc_arbiter_scheduling_setup()` and
`esw_qos_tc_arbiter_scheduling_teardown()`, which will be extended in
future patches to provide full support for tc-bw on devlink rate
objects.
- Validation functions for tc-bw settings, allowing graceful handling
of unsupported traffic class bandwidth configurations.
- Updated `__esw_qos_alloc_node()` to insert the new node into the
parent’s children list only if the parent is not NULL. For the root
TSAR, the new node is inserted directly after the allocation call.
- Don't allow `tc-bw` configuration for nodes containing non-leaf
children.
This patch lays the groundwork for future support for configuring tc-bw
on devlink rate nodes. Although the infrastructure is in place, full
support for tc-bw is not yet implemented; attempts to set tc-bw on
nodes will return `-EOPNOTSUPP`.
No functional changes are introduced at this stage.
Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/20250629142138.361537-6-mbloch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce support for specifying relative bandwidth shares between
traffic classes (TC) in the devlink-rate API. This new option allows
users to allocate bandwidth across multiple traffic classes in a
single command.
This feature provides a more granular control over traffic management,
especially for scenarios requiring Enhanced Transmission Selection.
Users can now define a relative bandwidth share for each traffic class.
For example, assigning share values of 20 to TC0 (TCP/UDP) and 80 to TC5
(RoCE) will result in TC0 receiving 20% and TC5 receiving 80% of the
total bandwidth. The actual percentage each class receives depends on
the ratio of its share value to the sum of all shares.
Example:
DEV=pci/0000:08:00.0
$ devlink port function rate add $DEV/vfs_group tx_share 10Gbit \
tx_max 50Gbit tc-bw 0:20 1:0 2:0 3:0 4:0 5:80 6:0 7:0
$ devlink port function rate set $DEV/vfs_group \
tc-bw 0:20 1:0 2:0 3:0 4:0 5:20 6:60 7:0
Example usage with ynl:
./tools/net/ynl/cli.py --spec Documentation/netlink/specs/devlink.yaml \
--do rate-set --json '{
"bus-name": "pci",
"dev-name": "0000:08:00.0",
"port-index": 1,
"rate-tc-bws": [
{"rate-tc-index": 0, "rate-tc-bw": 50},
{"rate-tc-index": 1, "rate-tc-bw": 50},
{"rate-tc-index": 2, "rate-tc-bw": 0},
{"rate-tc-index": 3, "rate-tc-bw": 0},
{"rate-tc-index": 4, "rate-tc-bw": 0},
{"rate-tc-index": 5, "rate-tc-bw": 0},
{"rate-tc-index": 6, "rate-tc-bw": 0},
{"rate-tc-index": 7, "rate-tc-bw": 0}
]
}'
./tools/net/ynl/cli.py --spec Documentation/netlink/specs/devlink.yaml \
--do rate-get --json '{
"bus-name": "pci",
"dev-name": "0000:08:00.0",
"port-index": 1
}'
output for rate-get:
{'bus-name': 'pci',
'dev-name': '0000:08:00.0',
'port-index': 1,
'rate-tc-bws': [{'rate-tc-bw': 50, 'rate-tc-index': 0},
{'rate-tc-bw': 50, 'rate-tc-index': 1},
{'rate-tc-bw': 0, 'rate-tc-index': 2},
{'rate-tc-bw': 0, 'rate-tc-index': 3},
{'rate-tc-bw': 0, 'rate-tc-index': 4},
{'rate-tc-bw': 0, 'rate-tc-index': 5},
{'rate-tc-bw': 0, 'rate-tc-index': 6},
{'rate-tc-bw': 0, 'rate-tc-index': 7}],
'rate-tx-max': 0,
'rate-tx-priority': 0,
'rate-tx-share': 0,
'rate-tx-weight': 0,
'rate-type': 'leaf'}
Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/20250629142138.361537-3-mbloch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We used to do twice copy_from_iter() to copy virtio-net and packet
separately. This introduce overheads for userspace access hardening as
well as SMAP (for x86 it's stac/clac). So this patch tries to use one
copy_from_iter() to copy them once and move the virtio-net header
afterwards to reduce overheads.
Testpmd + vhost_net shows 10% improvement from 5.45Mpps to 6.0Mpps.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Link: https://patch.msgid.link/20250701010352.74515-2-jasowang@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Willem de Bruijn says:
====================
preserve MSG_ZEROCOPY with forwarding
Avoid false positive copying of zerocopy skb frags when entering the
ingress path if the skb is not queued locally but forwarded.
Patch 1 for more details and feature.
Patch 2 converts the existing selftest to a pass/fail test and adds
coverage for this new feature.
====================
Link: https://patch.msgid.link/20250630194312.1571410-1-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Zerocopy skbs are converted to regular copy skbs when data is queued
to a local socket. This happens in the existing test with a sender and
receiver communicating over a veth device.
Zerocopy skbs are sent without copying if egressing a device. Verify
that this behavior is maintained even in the common container setup
where data is forwarded over a veth to the physical device.
Update msg_zerocopy.sh to
1. Have a dummy network device to simulate a physical device.
2. Have forwarding enabled between veth and dummy.
3. Add a tx-only test that sends out dummy via the forwarding path.
4. Verify the exitcode of the sender, which signals zerocopy success.
As dummy drops all packets, this cannot be a TCP connection. Test
the new case with unconnected UDP only.
Update msg_zerocopy.c to
- Accept an argument whether send with zerocopy is expected.
- Return an exitcode whether behavior matched that expectation.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250630194312.1571410-3-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
MSG_ZEROCOPY data must be copied before data is queued to local
sockets, to avoid indefinite timeout until memory release.
This test is performed by skb_orphan_frags_rx, which is called when
looping an egress skb to packet sockets, error queue or ingress path.
To preserve zerocopy for skbs that are looped to ingress but are then
forwarded to an egress device rather than delivered locally, defer
this last check until an skb enters the local IP receive path.
This is analogous to existing behavior of skb_clear_delivery_time.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250630194312.1571410-2-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a new test to ensure that when the transport changes a null pointer
dereference does not occur. The bug was reported upstream [1] and fixed
with commit 2cb7c756f6 ("vsock/virtio: discard packets if the
transport changes").
KASAN: null-ptr-deref in range [0x0000000000000060-0x0000000000000067]
CPU: 2 UID: 0 PID: 463 Comm: kworker/2:3 Not tainted
Workqueue: vsock-loopback vsock_loopback_work
RIP: 0010:vsock_stream_has_data+0x44/0x70
Call Trace:
virtio_transport_do_close+0x68/0x1a0
virtio_transport_recv_pkt+0x1045/0x2ae4
vsock_loopback_work+0x27d/0x3f0
process_one_work+0x846/0x1420
worker_thread+0x5b3/0xf80
kthread+0x35a/0x700
ret_from_fork+0x2d/0x70
ret_from_fork_asm+0x1a/0x30
Note that this test may not fail in a kernel without the fix, but it may
hang on the client side if it triggers a kernel oops.
This works by creating a socket, trying to connect to a server, and then
executing a second connect operation on the same socket but to a
different CID (0). This triggers a transport change. If the connect
operation is interrupted by a signal, this could cause a null-ptr-deref.
Since this bug is non-deterministic, we need to try several times. It
is reasonable to assume that the bug will show up within the timeout
period.
If there is a G2H transport loaded in the system, the bug is not
triggered and this test will always pass. This is because
`vsock_assign_transport`, when using CID 0, like in this case, sets
vsk->transport to `transport_g2h` that is not NULL if a G2H transport is
available.
[1]https://lore.kernel.org/netdev/Z2LvdTTQR7dBmPb5@v4bel-B760M-AORUS-ELITE-AX/
Suggested-by: Hyunwoo Kim <v4bel@theori.io>
Suggested-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: Luigi Leonardi <leonardi@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Link: https://patch.msgid.link/20250630-test_vsock-v5-2-2492e141e80b@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
