Use netlink extack to report errors instead of sending them
to the kernel log with pr_info(). The error message can them be seen
with tc commands; and avoids log spam.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Link: https://patch.msgid.link/20260509171123.307549-4-stephen@networkplumber.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The current layout of struct netem_sched_data can be improved
by optimizing cache locality, compacting data types (use u8
for enum) and eliminating unused elements.
Reorganize the struct as follows:
- Cacheline 0 holds the tfifo state (t_root/t_head/t_tail/t_len),
counter, and the unconditional enqueue scalars
latency/jitter/rate/gap/loss.
- Cacheline 1 holds the remaining zero-check scalars
(duplicate/reorder/corrupt/ecn), all five crndstate correlation
structures, and loss_model.
- Cacheline 2 holds prng, delay_dist, the slot dequeue state,
slot_dist, and the inner classful qdisc pointer.
- Rate-shaping fields, q->limit (config-only; the fast path reads
sch->limit), and the CLG Markov state move to the warm tail.
- tc_netem_slot slot_config and qdisc_watchdog (only consulted on
slot reschedule and watchdog wake) move to the cold tail.
Also reorder struct clgstate to place the u8 state member after the
u32 transition probabilities. This removes the 3-byte interior hole
without changing the struct's size.
Should have no functional change.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Link: https://patch.msgid.link/20260509171123.307549-2-stephen@networkplumber.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Eric Dumazet says:
====================
net/sched: refine fq_codel memory limits
Packets that are associated with local sockets sk_wmem_alloc
do not really need additional memory control.
First patch makes is_skb_wmem() available to modules.
Second patch uses is_skb_wmem() in fq_codel.
====================
Link: https://patch.msgid.link/20260512094859.3673997-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit 95b58430ab ("fq_codel: add memory limitation per queue")
claimed that the 32Mb default was "reasonable even for heavy duty usages."
In practice, this is not the case.
Packets that are associated with local sockets sk_wmem_alloc
do not really need additional memory control.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@toke.dk>
Link: https://patch.msgid.link/20260512094859.3673997-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Following patch will use is_skb_wmem() from fq_codel.
Provide __sock_wfree() only if CONFIG_INET=y
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@toke.dk>
Link: https://patch.msgid.link/20260512094859.3673997-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Tariq Toukan says:
====================
net/mlx5e: improve RSS indirection table sizing and resizing
This series by Yael improves mlx5e RSS indirection table handling around
channel count changes and large RSS configurations.
The series:
* removes the XOR8-specific channel count limitation,
* advertises the maximum supported RSS indirection table size,
* fixes resizing of non-default RSS contexts,
* allows resizing configured default RSS contexts during channel
changes,
* and increases the default RSS spread factor from 2x to 4x to improve
traffic distribution for large channel counts.
Together, these changes make RSS table sizing more flexible and robust,
while improving load balancing behavior on large systems.
====================
Link: https://patch.msgid.link/20260511172719.330490-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Increase the RQT uniform spread factor from 2 to 4 so that each channel
gets more indirection table entries and traffic is spread more evenly.
For num_channels > 64 imbalance drops from up to ~50% to up to ~25%.
For 64 or fewer channels the 256 entry minimum already provides at least
4x coverage and the table size is unchanged by this commit.
This satisfies the minimum 4x coverage requirement validated by the
generic RSS selftest commit 9e3d4dae98 ("selftests: drv-net: rss:
validate min RSS table size").
The 4x spread factor is best-effort and the table size is always capped by
the device's log_max_rqt_size capability.
Signed-off-by: Yael Chemla <ychemla@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260511172719.330490-6-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
mlx5e_ethtool_set_channels() rejected channel count changes that
required a different RQT size when the default context indirection
table was user-configured. This restriction was introduced by
commit ee3572409f ("net/mlx5e: RSS, Block changing channels number
when RXFH is configured").
Lift the restriction. Validate the resize upfront with
ethtool_rxfh_indir_can_resize(), then fold or unfold the table
in-place via ethtool_rxfh_indir_resize() inside state_lock, before
mlx5e_safe_switch_params(), so the preactivate callback sees the
correct table content when it programs the HW.
Signed-off-by: Yael Chemla <ychemla@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260511172719.330490-5-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When the channel count changes and the RQT size changes with it, a
problem arise for non-default RSS contexts. The driver-side indirection
table grows actual_table_size without filling the new entries; stale
entries from a prior larger configuration may be re-exposed, causing
mlx5e_calc_indir_rqns() to WARN on an out-of-range index.
Replace mlx5e_rss_params_indir_modify_actual_size() with
mlx5e_rss_ctx_resize(), which fills new entries by replicating
the existing pattern, matching what ethtool_rxfh_ctxs_resize() does
for the same case. And restrict the loop to non-default contexts.
Call ethtool_rxfh_ctxs_can_resize() before acquiring state_lock to
validate that all non-default contexts can be resized, and
ethtool_rxfh_ctxs_resize() after releasing it to fold or unfold their
indirection tables. Both functions acquire rss_lock internally and
cannot be called under state_lock. RTNL, held by all set_channels
callers, serialises context creation and deletion making the pre-lock
check safe.
Guard both ethtool calls on mlx5e_rx_res_rss_cnt() > 1: skip the
validation and resize when no non-default contexts exist. This
naturally covers representors and IPoIB, which share
mlx5e_ethtool_set_channels() but cannot have non-default RSS contexts.
Signed-off-by: Yael Chemla <ychemla@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260511172719.330490-4-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Set rxfh_indir_space to the maximum indirection table size the driver
can support: the next power of two above MLX5E_MAX_NUM_CHANNELS times
MLX5E_UNIFORM_SPREAD_RQT_FACTOR.
Without this, ethtool_rxfh_ctxs_can_resize() returns -EINVAL, blocking
non-default RSS contexts from tracking indirection table size changes
when the channel count changes.
Signed-off-by: Yael Chemla <ychemla@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260511172719.330490-3-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
mlx5e_ethtool_set_channels() and mlx5e_rxfh_hfunc_check() rejected
channel counts that would produce an indirection table larger than 256
entries when the XOR8 hash function was active. This check was
introduced in commit 49e6c93870 ("net/mlx5e: RSS, Block XOR hash
with over 128 channels").
XOR8 yields an 8-bit hash, so in practice only up to 256 entries in the
indirection table can be reached due to limited entropy. However, this
does not provide a strong justification for prohibiting larger
indirection tables. Remove the limitation.
Signed-off-by: Yael Chemla <ychemla@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260511172719.330490-2-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Breno Leitao says:
====================
netpoll: move out netconsole-specific functions
netpoll and netconsole were created together and their code has
been intermixed in net/core/netpoll.c for decades. The result is
that netpoll exposes two send-side interfaces:
* a generic "give me an sk_buff" path used by every stacked-device
driver (bonding, team, vlan, bridge, macvlan, dsa),
* a second path that takes raw bytes and builds a UDP/IP/Ethernet
packet -- exclusively for netconsole.
The packet builder, an skb pool allocator, and several
netconsole-specific helpers all live next to the generic plumbing even
though no other consumer ever touches them.
Worse, every netpoll user pays for that overlap: struct netpoll carries
an skb_pool and a refill work_struct that only netconsole's find_skb()
ever reads from, and net-core has to review unrelated changes (TTL, hop
limit, IP ID generation, source MAC selection, pool sizing) just because
they happen to be coded inside netpoll.
This is a waste of memory for something useless.
This series splits the netconsole-specific code out:
* netpoll_send_udp() and its private helpers (push_ipv6, push_ipv4,
push_eth, push_udp, netpoll_udp_checksum, find_skb) move into
drivers/net/netconsole.c, leaving netpoll with a single skb-only
send interface that is the same for every user.
The moves are one function per patch for reviewability; helpers are
temporarily EXPORT_SYMBOL_GPL'd while netpoll_send_udp() is still in
netpoll calling them, then those exports are dropped together once
netpoll_send_udp() itself moves.
The only new permanent export is zap_completion_queue(), needed because
find_skb() still drains the per-CPU TX completion queue before
allocating.
struct netpoll is unchanged in this series; making the pool itself
netconsole-private (and reclaiming the skb_pool / refill_wq fields for
the rest of netpoll's users) is the natural follow-up, once this patchset
lands.
====================
Link: https://patch.msgid.link/20260512-netconsole_split-v2-0-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
find_skb() is the netconsole-specific entry into the netpoll skb
pool: every other netpoll consumer (bonding, team, vlan, bridge,
macvlan, dsa) builds its own sk_buff and never touches the pool.
With netpoll_send_udp() (its only caller) now living in netconsole,
find_skb() can join it.
Move find_skb() into drivers/net/netconsole.c as a file-static
helper, drop EXPORT_SYMBOL_GPL(find_skb) and remove its prototype
from include/linux/netpoll.h. find_skb() drains TX completions via
netpoll_zap_completion_queue(), which is already exported in the
NETDEV_INTERNAL namespace, so netconsole picks up
MODULE_IMPORT_NS("NETDEV_INTERNAL") to consume it.
The skb pool's lifecycle (np->skb_pool, np->refill_wq, refill_skbs(),
refill_skbs_work_handler(), skb_pool_flush()) stays in netpoll: it
is initialised in __netpoll_setup() and torn down in
__netpoll_cleanup(), both of which remain netpoll's responsibility.
The refill work queued via schedule_work(&np->refill_wq) from the
moved find_skb() runs refill_skbs_work_handler() in netpoll without
any further plumbing.
This is pure code motion: the function body is unchanged and its
sole caller (netpoll_send_udp(), already moved by an earlier patch)
keeps invoking it the same way. Pre-existing concerns about
find_skb() running from NMI/printk context (zap_completion_queue()
re-entry, skb_pool spinlocks, GFP_ATOMIC allocation, fallback skb
sizing vs. MAX_SKB_SIZE, PREEMPT_RT semantics of __kfree_skb()) are
inherited as-is and are not addressed here; they predate this
series and are out of scope. Fixing them is left for follow-up
work.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-9-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
zap_completion_queue() drains the per-CPU softnet completion queue.
Rename it with the netpoll_ prefix shared by the rest of the
subsystem's public API, and promote it from file-static to
EXPORT_SYMBOL_NS_GPL in the NETDEV_INTERNAL namespace so the upcoming
netconsole-side find_skb() can call it once the function moves out.
A forward declaration is added to include/linux/netpoll.h, and the
old file-static forward declaration is dropped.
No functional change.
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-8-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
netpoll_udp_checksum() computes the UDP checksum for netconsole's
packets. Move it into drivers/net/netconsole.c as a file-static
helper; drop its EXPORT_SYMBOL_GPL and remove the prototype from
include/linux/netpoll.h.
This was the last csum_ipv6_magic() consumer in net/core/netpoll.c,
so drop the now-stale <net/ip6_checksum.h> include there. Pull it
into netconsole.c so the moved code keeps building.
It was also the last udp_hdr() consumer in net/core/netpoll.c. The
file no longer needs anything from <net/udp.h> (the UDP socket-layer
helpers); MAX_SKB_SIZE only needs struct udphdr, which is provided
by the lighter <linux/udp.h>. Swap the include accordingly.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-7-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
push_udp() builds the UDP header (and triggers the checksum) for
netconsole's UDP packets. Move it into drivers/net/netconsole.c as
a file-static helper; drop its EXPORT_SYMBOL_GPL and remove the
prototype from include/linux/netpoll.h.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-6-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
push_eth() builds the Ethernet header for netconsole's UDP packets.
Move it into drivers/net/netconsole.c as a file-static helper; drop
its EXPORT_SYMBOL_GPL and remove the prototype from
include/linux/netpoll.h.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-5-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
push_ipv4() builds the IPv4 header for netconsole's UDP packets.
Move it into drivers/net/netconsole.c as a file-static helper; drop
its EXPORT_SYMBOL_GPL and remove the prototype from
include/linux/netpoll.h.
put_unaligned() is no longer used in net/core/netpoll.c, so drop
the now-stale <linux/unaligned.h> include from there. Pull it into
netconsole.c so the moved code keeps building.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-4-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
push_ipv6() builds the IPv6 header for netconsole's UDP packets.
Its only caller, netpoll_send_udp(), now lives in netconsole, so
the helper can move there as a file-static function. Drop its
EXPORT_SYMBOL_GPL and remove the prototype from
include/linux/netpoll.h.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-3-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Move netpoll_send_udp() from net/core/netpoll.c into
drivers/net/netconsole.c as a static helper, drop EXPORT_SYMBOL(),
and remove the prototype from include/linux/netpoll.h.
netconsole was the only in-tree caller of this entry point. Every
other netpoll consumer (bonding, team, vlan, bridge, macvlan, dsa)
already builds its own sk_buff and hands it to netpoll_send_skb(),
so the netpoll send-side interface is now skb-only.
The helpers it depends on (find_skb(), push_ipv6(), push_ipv4(),
push_udp(), push_eth(), netpoll_udp_checksum()) were exposed in
the previous patches and stay in net/core/netpoll.c for now.
Subsequent patches move each of them into netconsole one at a time
and drop the corresponding EXPORT_SYMBOL_GPL.
Pull <linux/ip.h>, <linux/ipv6.h> and <linux/udp.h> into netconsole.c
so the moved code can name the header structures.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-2-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Promote each from file-static to EXPORT_SYMBOL_GPL and forward-
declare them in include/linux/netpoll.h so netconsole can call
them once netpoll_send_udp() moves out.
These exports are kept until the end of the series, when
al of them move into netconsole.
No functional change.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260512-netconsole_split-v2-1-1191d14ad66d@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
simple_strtoul() is deprecated as it has no error checking. Replace it
with kstrtouint() which returns an error code on invalid input, and add
appropriate error handling.
Also add a NULL check before parsing flags, since strsep() can set id
to NULL if the input has fewer tokens than expected.
Preserve the original behavior for a trailing colon by checking *id
before parsing flags, so an empty string results in flags = 0 rather
than an error.
Signed-off-by: Sajal Gupta <sajal2005gupta@gmail.com>
Link: https://patch.msgid.link/20260509095518.2640-1-sajal2005gupta@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ipvlan_process_multicast() runs from process context, there is no
risk of stack overflow if we call netif_receive_skb() instead
of netif_rx().
This avoids some overhead adding/removing skbs to/from a per-cpu
backlog and raising/processing NET_RX softirqs.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260512042019.3300975-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Simon Schippers says:
====================
tun/tap & vhost-net: apply qdisc backpressure on full ptr_ring to reduce TX drops
This patch series deals with tun/tap & vhost-net which drop incoming
SKBs whenever their internal ptr_ring buffer is full. Instead, with this
patch series, the associated netdev queue is stopped - but only when a
qdisc is attached. If no qdisc is present the existing behavior is
preserved. The XDP transmit path is not affected. This patch series
touches tun/tap and vhost-net, as they share common logic and must be
updated together. Modifying only one of them would break the other.
By applying proper backpressure, this change allows the connected qdisc to
operate correctly, as reported in [1], and significantly improves
performance in real-world scenarios, as demonstrated in our paper [2]. For
example, we observed a 36% TCP throughput improvement for an OpenVPN
connection between Germany and the USA.
Synthetic pktgen benchmarks indicate a slight regression, and packet
loss is reduced to near zero. Pktgen benchmarks are provided per commit,
with the final commit showing the overall performance.
[1] Link: https://unix.stackexchange.com/questions/762935/traffic-shaping-ineffective-on-tun-device
[2] Link: https://cni.etit.tu-dortmund.de/storages/cni-etit/r/Research/Publications/2025/Gebauer_2025_VTCFall/Gebauer_VTCFall2025_AuthorsVersion.pdf
====================
Link: https://patch.msgid.link/20260510151529.43895-1-simon.schippers@tu-dortmund.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This commit prevents tail-drop when a qdisc is present and the ptr_ring
becomes full. Once the ring reaches capacity after a produce attempt,
the netdev queue is stopped instead of dropping subsequent packets.
If no qdisc is present, the previous tail-drop behavior is preserved.
If producing an entry fails anyway due to a race, tun_net_xmit() drops
the packet. Such races are expected because LLTX is enabled and the
transmit path operates without the usual locking.
The __tun_wake_queue() function of the consumer races with the producer
for waking/stopping the netdev queue, which could result in a stalled
queue. Therefore, an smp_mb__after_atomic() is introduced that pairs
with the smp_mb() of the consumer. It follows the principle of store
buffering described in tools/memory-model/Documentation/recipes.txt:
- The producer in tun_net_xmit() first sets __QUEUE_STATE_DRV_XOFF,
followed by an smp_mb__after_atomic() (= smp_mb()), and then reads the
ring with __ptr_ring_check_produce().
- The consumer in __tun_wake_queue() first writes zero to the ring in
__ptr_ring_consume(), followed by an smp_mb(), and then reads the queue
status with netif_tx_queue_stopped().
=> Following the aforementioned principle, it is impossible for the
producer to see a full ring (and therefore not wake the queue on the
re-check) while the consumer simultaneously fails to see a stopped
queue (and therefore also does not wake it).
Benchmarks:
The benchmarks show a slight regression in raw transmission performance
when using two sending threads. Packet loss also occurs only in the
two-thread sending case; no packet loss was observed with a single
sending thread.
Test setup:
AMD Ryzen 5 5600X at 4.3 GHz, 3200 MHz RAM, isolated QEMU threads;
Average over 50 runs @ 100,000,000 packets. SRSO and spectre v2
mitigations disabled.
Note for tap+vhost-net:
XDP drop program active in VM -> ~2.5x faster; slower for tap due to
more syscalls (high utilization of entry_SYSRETQ_unsafe_stack in perf)
+--------------------------+--------------+----------------+----------+
| 1 thread | Stock | Patched with | diff |
| sending | | fq_codel qdisc | |
+------------+-------------+--------------+----------------+----------+
| TAP | Received | 1.132 Mpps | 1.123 Mpps | -0.8% |
| +-------------+--------------+----------------+----------+
| | Lost/s | 3.765 Mpps | 0 pps | |
+------------+-------------+--------------+----------------+----------+
| TAP | Received | 3.857 Mpps | 3.901 Mpps | +1.1% |
| +-------------+--------------+----------------+----------+
| +vhost-net | Lost/s | 0.802 Mpps | 0 pps | |
+------------+-------------+--------------+----------------+----------+
+--------------------------+--------------+----------------+----------+
| 2 threads | Stock | Patched with | diff |
| sending | | fq_codel qdisc | |
+------------+-------------+--------------+----------------+----------+
| TAP | Received | 1.115 Mpps | 1.081 Mpps | -3.0% |
| +-------------+--------------+----------------+----------+
| | Lost/s | 8.490 Mpps | 391 pps | |
+------------+-------------+--------------+----------------+----------+
| TAP | Received | 3.664 Mpps | 3.555 Mpps | -3.0% |
| +-------------+--------------+----------------+----------+
| +vhost-net | Lost/s | 5.330 Mpps | 938 pps | |
+------------+-------------+--------------+----------------+----------+
Co-developed-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Simon Schippers <simon.schippers@tu-dortmund.de>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Link: https://patch.msgid.link/20260510151529.43895-5-simon.schippers@tu-dortmund.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch moves the check for available free space for a new entry into
a separate function. Existing callers that only check for a non-zero
return value are unaffected; __ptr_ring_produce() now returns -EINVAL
for a zero-size ring and -ENOSPC when full, whereas before both cases
returned -ENOSPC. The new helper allows callers to determine in advance
whether subsequent __ptr_ring_produce() calls will succeed. This
information can, for example, be used to temporarily stop producing until
__ptr_ring_check_produce() indicates that space is available again.
Co-developed-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Simon Schippers <simon.schippers@tu-dortmund.de>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Link: https://patch.msgid.link/20260510151529.43895-4-simon.schippers@tu-dortmund.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add tun_wake_queue() to tun.c and export it for use by vhost-net. The
function validates that the file belongs to a tun/tap device and that
the tfile exists, dereferences the tun_struct under RCU, and delegates
to __tun_wake_queue().
vhost_net_buf_produce() now calls tun_wake_queue() after a successful
batched consume of the ring to allow the netdev subqueue to be woken up.
The point is to allow the queue to be stopped when it gets full, which
is required for traffic shaping - implemented by the following
"avoid ptr_ring tail-drop when a qdisc is present".
Without the corresponding queue stopping, this patch alone causes no
throughput regression for a tap+vhost-net setup sending to a qemu VM:
3.857 Mpps to 3.891 Mpps.
Details: AMD Ryzen 5 5600X at 4.3 GHz, 3200 MHz RAM, isolated QEMU
threads, XDP drop program active in VM, pktgen sender; Avg over
50 runs @ 100,000,000 packets. SRSO and spectre v2 mitigations disabled.
Co-developed-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Simon Schippers <simon.schippers@tu-dortmund.de>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Link: https://patch.msgid.link/20260510151529.43895-3-simon.schippers@tu-dortmund.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce tun_ring_consume() that wraps ptr_ring_consume() and calls
__tun_wake_queue(). The latter wakes the stopped netdev subqueue once
half of the ring capacity has been consumed, tracked via the new
cons_cnt field in tun_file. As a safety net, the queue is also woken on
the last consumed entry if it leaves the ring empty. The point is to
allow the queue to be stopped when it gets full, which is required for
traffic shaping - implemented by the following "avoid ptr_ring tail-drop
when a qdisc is present".
Some implementation details:
- tun_ring_recv() replaces ptr_ring_consume() with tun_ring_consume()
to properly wake the queue.
- __tun_detach() locks the tx_ring.consumer_lock to avoid races with
the consumer on the queue_index.
- The ptr_ring_consume() call in tun_queue_purge() is not replaced with
tun_ring_consume(). Instead, within the same tx_ring.consumer_lock
in __tun_detach(), the netdev queue is woken for the ntfile taking
it over, to avoid a possible stall. This does not matter for
tun_detach_all(), as it is called during device teardown and no tfile
takes over any queue.
- Reset cons_cnt in tun_attach() so the half-ring wake threshold is
valid for the new ring size after ptr_ring_resize().
- tun_queue_resize() wakes all queues after resizing with the proper
tx_ring.consumer_lock and resets the cons_cnt to avoid a possible
stale queue.
- The aforementioned upcoming patch explains the pairing of the smp_mb()
of __tun_wake_queue().
Without the corresponding queue stopping, this patch alone causes no
regression for a tap setup sending to a qemu VM: 1.132 Mpps
to 1.134 Mpps.
Details: AMD Ryzen 5 5600X at 4.3 GHz, 3200 MHz RAM, isolated QEMU
threads, pktgen sender; Avg over 50 runs @ 100,000,000 packets;
SRSO and spectre v2 mitigations disabled.
Co-developed-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
Signed-off-by: Simon Schippers <simon.schippers@tu-dortmund.de>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Link: https://patch.msgid.link/20260510151529.43895-2-simon.schippers@tu-dortmund.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Ivan Vecera says:
====================
dpll: rework fractional frequency offset reporting
Rework how the fractional frequency offset (FFO) is reported in
the DPLL subsystem.
Both fractional-frequency-offset (PPM) and
fractional-frequency-offset-ppt (PPT) attributes are now present at
the top level of a pin and inside each pin-parent-device nest. They
carry the same measurement at different precisions.
Introduce enum dpll_ffo_type and struct dpll_ffo_param to distinguish
FFO contexts: DPLL_FFO_PORT_RXTX_RATE for the RX vs TX symbol rate
offset at the top level, and DPLL_FFO_PIN_DEVICE for the pin vs
parent DPLL offset in the nest. Drivers declare which types they
support via the supported_ffo bitmask in dpll_pin_ops; the core only
calls ffo_get for opted-in types.
Patch 1 adds the type-safe FFO API, updates the YAML spec, netlink
handling, and documentation, and converts mlx5 and zl3073x drivers.
Patch 2 implements the nested FFO for zl3073x using the
dpll_df_offset_x register with ref_ofst=1, providing 2^-48
resolution. The old per-reference frequency measurement is removed
as it was redundant with measured-frequency.
====================
Link: https://patch.msgid.link/20260511155816.99936-1-ivecera@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace the per-reference frequency offset measurement (which was
redundant with measured-frequency) with a direct read of the DPLL's
delta frequency offset vs its tracked input reference.
The new implementation uses the dpll_df_offset_x register with
ref_ofst=1 via the dpll_df_read_x semaphore mechanism. This
provides 2^-48 resolution (~3.5 fE) and reports the actual
frequency difference between the DPLL and its active input.
Switch supported_ffo from DPLL_FFO_PORT_RXTX_RATE to
DPLL_FFO_PIN_DEVICE so FFO is reported only in the per-parent
context for the active input pin.
Use atomic64_t for freq_offset to prevent torn reads on 32-bit
architectures between the periodic worker and netlink callbacks.
Rewrite ffo_check to compare the cached df_offset converted to PPT
instead of using the old per-reference measurement. Remove the
ref_ffo_update periodic measurement and the ref ffo field since
they are no longer needed.
Changes v3 -> v4:
- Switch to DPLL_FFO_PIN_DEVICE, remove dpll=NULL guard
- Use atomic64_t for freq_offset (torn read on 32-bit)
Reviewed-by: Petr Oros <poros@redhat.com>
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260511155816.99936-3-ivecera@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add both fractional-frequency-offset (PPM) and
fractional-frequency-offset-ppt (PPT) attributes to the
pin-parent-device nested attribute set, alongside the existing
top-level pin attributes. Both carry the same measurement at
different precisions.
Introduce enum dpll_ffo_type and struct dpll_ffo_param to
distinguish FFO contexts: DPLL_FFO_PORT_RXTX_RATE for the RX vs
TX symbol rate offset reported at the top level, and
DPLL_FFO_PIN_DEVICE for the pin vs parent DPLL offset reported
in the pin-parent-device nest.
Add a supported_ffo bitmask to struct dpll_pin_ops so drivers
declare which FFO types they support. The core only calls ffo_get
for types the driver has opted into, eliminating the need for
per-driver NULL pointer guards. Validate at pin registration time
that supported_ffo is not set without an ffo_get callback.
Update mlx5 (DPLL_FFO_PORT_RXTX_RATE) and zl3073x
(DPLL_FFO_PORT_RXTX_RATE) drivers to use the new API.
Add documentation for both FFO types to dpll.rst.
Changes v3 -> v4:
- Replace dpll=NULL overloading with enum dpll_ffo_type and
struct dpll_ffo_param (Jakub Kicinski)
- Add supported_ffo opt-in bitmask in dpll_pin_ops for fail-close
driver validation (Jakub Kicinski)
- Add WARN_ON in dpll_pin_register for supported_ffo without
ffo_get callback
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260511155816.99936-2-ivecera@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add chip info entry for the Realtek RTL8367SB switch. This device has
chip ID 0x6367 and version 0x0010. It exposes two external interfaces:
port 6 supports MII, TMII, RMII, RGMII, SGMII and HSGMII, while port 7
supports MII, TMII, RMII and RGMII. Use the existing 8365MB-VC jam table
for initialization.
Reviewed-by: Luiz Angelo Daros de Luca <luizluca@gmail.com>
Signed-off-by: Mieczyslaw Nalewaj <namiltd@yahoo.com>
Link: https://patch.msgid.link/3c6d822b-0e85-4173-86ba-2badb140bbf1@yahoo.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Eric Dumazet says:
====================
net: use IP_OUTNOROUTES drop reason
First patch changes sk_skb_reason_drop() sock to be const.
Second and last patch add SKB_DROP_REASON_IP_OUTNOROUTES
to both tcp_v6_send_response() and inet6_csk_xmit().
====================
Link: https://patch.msgid.link/20260511072310.1094859-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace a bare kfree_skb() with a modern sk_skb_reason_drop() call,
and provide IP_OUTNOROUTES drop reason.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260511072310.1094859-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace a bare kfree_skb() with a modern sk_skb_reason_drop() call,
and provide IP_OUTNOROUTES drop reason.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260511072310.1094859-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
sk_skb_reason_drop() does not change sock parameter, make it
const so that we can call it from TCP stack without a cast
on a (const) listener socket.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260511072310.1094859-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
iproute2 can spend considerable amount of time in ll_init_map()
or ll_link_get() to dump verbose netdev attributes, contributing
to RTNL pressure.
Add RTEXT_FILTER_NAME_ONLY new flag so that rtnl_fill_ifinfo()
limits its output to:
- struct nlmsghdr
- IFLA_IFNAME
- IFLA_PROP_LIST (alternate names)
We can later avoid using RTNL when RTEXT_FILTER_NAME_ONLY
is requested, as none of these attributes need RTNL.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260511070244.971028-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
... and PCI device helpers.
The various struct pci_device_id arrays were initialized mostly by one
the PCI_DEVICE macros and then list expressions. The latter isn't easily
readable if you're not into PCI. Using named initializers is more
explicit and thus easier to parse.
Also use PCI_DEVICE* helper macros to assign .vendor, .device,
.subvendor and .subdevice where appropriate and skip explicit
assignments of 0 (which the compiler takes care of).
The secret plan is to make struct pci_device_id::driver_data an
anonymous union (similar to
https://lore.kernel.org/all/cover.1776579304.git.u.kleine-koenig@baylibre.com/)
and that requires named initializers. But it's also a nice cleanup on
its own.
This change doesn't introduce changes to the compiled pci_device_id
arrays. Tested on x86 and arm64.
Reviewed-by: Jijie Shao <shaojijie@huawei.com>
Acked-by: Arend van Spriel <arend.vanspriel@broadcom.com>
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Reviewed-by: Petr Machata <petrm@nvidia.com> # for mlxsw
Acked-by: Jacob Keller <jacob.e.keller@intel.com>
Acked-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: Uwe Kleine-König (The Capable Hub) <u.kleine-koenig@baylibre.com>
Forwarded: id:76da4f44d48bdde84580963862bf9616bee5c9e9.1778149923.git.u.kleine-koenig@baylibre.com (v2)
Reviewed-by: Michael Grzeschik <mgr@kernel.org>
Link: https://patch.msgid.link/20260511090023.1634387-6-u.kleine-koenig@baylibre.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
With .class_mask being zero the value of .class doesn't matter because
to check if a pci_device_id entry matches a given device the expression
(id->class ^ dev->class) & id->class_mask
is checked for being zero (see pci_match_one_device()). So drop the
useless and irritating assignment for .class to match what (I think) all
other drivers are doing that don't need to match on .class, i.e. set
both members to zero.
Signed-off-by: Uwe Kleine-König (The Capable Hub) <u.kleine-koenig@baylibre.com>
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Link: https://patch.msgid.link/20260511090023.1634387-5-u.kleine-koenig@baylibre.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Report PCS receive error counts for all supported PEF 7061, 7071, 7072 and
xRX200 PHYs.
Accumulate the vendor-specific PHY_ERRCNT read-clear counter
(SEL=RXERR) in .update_stats() and expose it as both IEEE 802.3
SymbolErrorDuringCarrier and generic rx_errors via
.get_phy_stats().
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Link: https://patch.msgid.link/20260509205933.3965832-1-olek2@wp.pl
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This test was converted from shell script to drv-net test.
The new version is flaky in dbg builds on the netdev.bots dashboard.
The previous shell script had more protections to avoid these. Added
in commit a7ee79b9c4 ("selftests: net: cope with slow env in
so_txtime.sh test").
Add the same overall protection:
- Suppress so_txtime process failure if KSFT_MACHINE_SLOW
Also relax two timeouts to reduce the number of process failures
themselves
- Increase SO_RCVTIMEO to 2 seconds
- Increase process start-up stabilization to 2 seconds
Delays were experimentally arrived at while running with vng
built with kernel/configs/debug.config
Fixes: 5c6baef388 ("selftests: drv-net: convert so_txtime to drv-net")
Reported-by: Jakub Kicinski <kuba@kernel.org>
Closes: https://lore.kernel.org/netdev/20260510174219.74aeee6d@kernel.org/
Signed-off-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260511222138.2045551-1-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
sizeof(u32) on the _RINGS_CQE_SIZE line is missing its closing
parenthesis, causing nla_total_size() to absorb the subsequent
_TX_PUSH and _RX_PUSH entries.
The resulting size estimate happens to be numerically identical
due to NLA alignment, so not treating this as a real fix.
But the nesting is wrong and misleading.
Signed-off-by: Tao Cui <cuitao@kylinos.cn>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Reviewed-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260508125412.189804-1-cuitao@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Eric Dumazet says:
====================
net/sched: prepare lockless qdisc dumps
Goal is to no longer acquire RTNL in qdisc dumps.
This series annotate data-races, and change mq and mq_prio to
no longer acquire children qdisc spinlocks.
====================
Link: https://patch.msgid.link/20260510091455.4039245-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prepare mqprio_dump_class_stats() for RTNL avoidance.
Use RCU instead of RTNL, and no longer acquire each children spinlock.
As a bonus we no longer have to release/acquire d->lock.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>
Link: https://patch.msgid.link/20260510091455.4039245-9-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prepare mqprio_dump() for RTNL avoidance.
Use RCU instead of RTNL, and no longer acquire each children spinlock.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>
Link: https://patch.msgid.link/20260510091455.4039245-8-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prepare mq_dump_common() for RTNL avoidance.
Use RCU instead of RTNL, and no longer acquire each children spinlock.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>
Link: https://patch.msgid.link/20260510091455.4039245-7-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
