The reproduction steps:
1. create a tun interface
2. enable l2 bearer
3. TIPC_NL_UDP_GET_REMOTEIP with media name set to tun
tipc: Started in network mode
tipc: Node identity 8af312d38a21, cluster identity 4711
tipc: Enabled bearer <eth:syz_tun>, priority 1
Oops: general protection fault
KASAN: null-ptr-deref in range
CPU: 1 UID: 1000 PID: 559 Comm: poc Not tainted 6.16.0-rc1+ #117 PREEMPT
Hardware name: QEMU Ubuntu 24.04 PC
RIP: 0010:tipc_udp_nl_dump_remoteip+0x4a4/0x8f0
the ub was in fact a struct dev.
when bid != 0 && skip_cnt != 0, bearer_list[bid] may be NULL or
other media when other thread changes it.
fix this by checking media_id.
Fixes: 832629ca5c ("tipc: add UDP remoteip dump to netlink API")
Signed-off-by: Haixia Qu <hxqu@hillstonenet.com>
Reviewed-by: Tung Nguyen <tung.quang.nguyen@est.tech>
Link: https://patch.msgid.link/20250617055624.2680-1-hxqu@hillstonenet.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
NIX block can receive packets from multiple links such as
MAC (RPM), LBK and CPT.
-----------------
RPM --| NIX |
-----------------
|
|
LBK
Each link supports multiple channels for example RPM link supports
16 channels. In case of link oversubsribe, NIX will assert backpressure
on receive channels.
The previous patch considered a single channel per link, resulting in
backpressure not being enabled on the remaining channels
Fixes: a7ef63dbd5 ("octeontx2-af: Disable backpressure between CPT and NIX")
Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250617063403.3582210-1-hkelam@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Tony Nguyen says:
====================
Intel Wired LAN Driver Updates 2025-06-17 (ice, e1000e)
For ice:
Krishna Kumar modifies aRFS match criteria to correctly identify
matching filters.
Grzegorz fixes a memory leak in eswitch legacy mode.
For e1000e:
Vitaly sets clock frequency on some Nahum systems which may misreport
their value.
* '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue:
e1000e: set fixed clock frequency indication for Nahum 11 and Nahum 13
ice: fix eswitch code memory leak in reset scenario
net: ice: Perform accurate aRFS flow match
====================
Link: https://patch.msgid.link/20250617172444.1419560-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit under fixes switched to uAPI generation from the YAML
spec. A number of custom defines were left behind, mostly
for commands very hard to express in YAML spec.
Among what was left behind was the name and version of
the generic netlink family. Problem is that the codegen
always outputs those values so we ended up with a duplicated,
differently named set of defines.
Provide naming info in YAML and remove the incorrect defines.
Fixes: 8d0580c6eb ("ethtool: regenerate uapi header from the spec")
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Link: https://patch.msgid.link/20250617202240.811179-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
After the following commit from 2024:
commit e37ab73736 ("tcp: fix to allow timestamp undo if no retransmits were sent")
...there was buggy behavior where TCP connections without SACK support
could easily see erroneous undo events at the end of fast recovery or
RTO recovery episodes. The erroneous undo events could cause those
connections to suffer repeated loss recovery episodes and high
retransmit rates.
The problem was an interaction between the non-SACK behavior on these
connections and the undo logic. The problem is that, for non-SACK
connections at the end of a loss recovery episode, if snd_una ==
high_seq, then tcp_is_non_sack_preventing_reopen() holds steady in
CA_Recovery or CA_Loss, but clears tp->retrans_stamp to 0. Then upon
the next ACK the "tcp: fix to allow timestamp undo if no retransmits
were sent" logic saw the tp->retrans_stamp at 0 and erroneously
concluded that no data was retransmitted, and erroneously performed an
undo of the cwnd reduction, restoring cwnd immediately to the value it
had before loss recovery. This caused an immediate burst of traffic
and build-up of queues and likely another immediate loss recovery
episode.
This commit fixes tcp_packet_delayed() to ignore zero retrans_stamp
values for non-SACK connections when snd_una is at or above high_seq,
because tcp_is_non_sack_preventing_reopen() clears retrans_stamp in
this case, so it's not a valid signal that we can undo.
Note that the commit named in the Fixes footer restored long-present
behavior from roughly 2005-2019, so apparently this bug was present
for a while during that era, and this was simply not caught.
Fixes: e37ab73736 ("tcp: fix to allow timestamp undo if no retransmits were sent")
Reported-by: Eric Wheeler <netdev@lists.ewheeler.net>
Closes: https://lore.kernel.org/netdev/64ea9333-e7f9-0df-b0f2-8d566143acab@ewheeler.net/
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Co-developed-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In vcc_sendmsg(), we account skb->truesize to sk->sk_wmem_alloc by
atm_account_tx().
It is expected to be reverted by atm_pop_raw() later called by
vcc->dev->ops->send(vcc, skb).
However, vcc_sendmsg() misses the same revert when copy_from_iter_full()
fails, and then we will leak a socket.
Let's factorise the revert part as atm_return_tx() and call it in
the failure path.
Note that the corresponding sk_wmem_alloc operation can be found in
alloc_tx() as of the blamed commit.
$ git blame -L:alloc_tx net/atm/common.c c55fa3cccbc2c~
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Reported-by: Simon Horman <horms@kernel.org>
Closes: https://lore.kernel.org/netdev/20250614161959.GR414686@horms.kernel.org/
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250616182147.963333-3-kuni1840@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Vladimir Oltean says:
====================
ptp_vclock fixes
While I was intending to test something else related to PTP in net-next
I noticed any time I would run ptp4l on an interface, the kernel would
print "ptp: physical clock is free running" and ptp4l would exit with an
error code.
I then found Jeongjun Park's patch and subsequent explanation provided
to Jakub's question, specifically related to the code which introduced
the breakage I am seeing.
https://lore.kernel.org/netdev/CAO9qdTEjQ5414un7Yw604paECF=6etVKSDSnYmZzZ6Pg3LurXw@mail.gmail.com/
I had to look at the original issue that prompted Jeongjun Park's patch,
and provide an alternative fix for it. Patch 1/2 in this set contains a
logical revert plus the alternative fix, squashed into one.
Patch 2/2 fixes another issue which was confusing during debugging/
characterization, namely: "ok, the kernel clearly thinks that any
physical clock is free-running after this change (despite there being no
vclocks), but why would ptp4l fail to create the clock altogether? Why
not just fail to adjust it?"
By reverting (locally) Jeongjun Park's commit, I could reproduce
the reported lockdep splat using the commands from patch 1/2's commit
message, and this goes away with the reworked implementation.
====================
Link: https://patch.msgid.link/20250613174749.406826-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There is a bug in ptp_clock_adjtime() which makes it refuse the
operation even if we just want to read the current clock dialed
frequency, not modify anything (tx->modes == 0). That should be possible
even if the clock is free-running. For context, the kernel UAPI is the
same for getting and setting the frequency of a POSIX clock.
For example, ptp4l errors out at clock_create() -> clockadj_get_freq()
-> clock_adjtime() time, when it should logically only have failed on
actual adjustments to the clock, aka if the clock was configured as
slave. But in master mode it should work.
This was discovered when examining the issue described in the previous
commit, where ptp_clock_freerun() returned true despite n_vclocks being
zero.
Fixes: 73f37068d5 ("ptp: support ptp physical/virtual clocks conversion")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20250613174749.406826-3-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
What is broken
--------------
ptp4l, and any other application which calls clock_adjtime() on a
physical clock, is greeted with error -EBUSY after commit 87f7ce260a
("ptp: remove ptp->n_vclocks check logic in ptp_vclock_in_use()").
Explanation for the breakage
----------------------------
The blamed commit was based on the false assumption that
ptp_vclock_in_use() callers already test for n_vclocks prior to calling
this function.
This is notably incorrect for the code path below, in which there is, in
fact, no n_vclocks test:
ptp_clock_adjtime()
-> ptp_clock_freerun()
-> ptp_vclock_in_use()
The result is that any clock adjustment on any physical clock is now
impossible. This is _despite_ there not being any vclock over this
physical clock.
$ ptp4l -i eno0 -2 -P -m
ptp4l[58.425]: selected /dev/ptp0 as PTP clock
[ 58.429749] ptp: physical clock is free running
ptp4l[58.431]: Failed to open /dev/ptp0: Device or resource busy
failed to create a clock
$ cat /sys/class/ptp/ptp0/n_vclocks
0
The patch makes the ptp_vclock_in_use() function say "if it's not a
virtual clock, then this physical clock does have virtual clocks on
top".
Then ptp_clock_freerun() uses this information to say "this physical
clock has virtual clocks on top, so it must stay free-running".
Then ptp_clock_adjtime() uses this information to say "well, if this
physical clock has to be free-running, I can't do it, return -EBUSY".
Simply put, ptp_vclock_in_use() cannot be simplified so as to remove the
test whether vclocks are in use.
What did the blamed commit intend to fix
----------------------------------------
The blamed commit presents a lockdep warning stating "possible recursive
locking detected", with the n_vclocks_store() and ptp_clock_unregister()
functions involved.
The recursive locking seems this:
n_vclocks_store()
-> mutex_lock_interruptible(&ptp->n_vclocks_mux) // 1
-> device_for_each_child_reverse(..., unregister_vclock)
-> unregister_vclock()
-> ptp_vclock_unregister()
-> ptp_clock_unregister()
-> ptp_vclock_in_use()
-> mutex_lock_interruptible(&ptp->n_vclocks_mux) // 2
The issue can be triggered by creating and then deleting vclocks:
$ echo 2 > /sys/class/ptp/ptp0/n_vclocks
$ echo 0 > /sys/class/ptp/ptp0/n_vclocks
But note that in the original stack trace, the address of the first lock
is different from the address of the second lock. This is because at
step 1 marked above, &ptp->n_vclocks_mux is the lock of the parent
(physical) PTP clock, and at step 2, the lock is of the child (virtual)
PTP clock. They are different locks of different devices.
In this situation there is no real deadlock, the lockdep warning is
caused by the fact that the mutexes have the same lock class on both the
parent and the child. Functionally it is fine.
Proposed alternative solution
-----------------------------
We must reintroduce the body of ptp_vclock_in_use() mostly as it was
structured prior to the blamed commit, but avoid the lockdep warning.
Based on the fact that vclocks cannot be nested on top of one another
(ptp_is_attribute_visible() hides n_vclocks for virtual clocks), we
already know that ptp->n_vclocks is zero for a virtual clock. And
ptp->is_virtual_clock is a runtime invariant, established at
ptp_clock_register() time and never changed. There is no need to
serialize on any mutex in order to read ptp->is_virtual_clock, and we
take advantage of that by moving it outside the lock.
Thus, virtual clocks do not need to acquire &ptp->n_vclocks_mux at
all, and step 2 in the code walkthrough above can simply go away.
We can simply return false to the question "ptp_vclock_in_use(a virtual
clock)".
Other notes
-----------
Releasing &ptp->n_vclocks_mux before ptp_vclock_in_use() returns
execution seems racy, because the returned value can become stale as
soon as the function returns and before the return value is used (i.e.
n_vclocks_store() can run any time). The locking requirement should
somehow be transferred to the caller, to ensure a longer life time for
the returned value, but this seems out of scope for this severe bug fix.
Because we are also fixing up the logic from the original commit, there
is another Fixes: tag for that.
Fixes: 87f7ce260a ("ptp: remove ptp->n_vclocks check logic in ptp_vclock_in_use()")
Fixes: 73f37068d5 ("ptp: support ptp physical/virtual clocks conversion")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20250613174749.406826-2-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The commit under the Fixes tag below which updates the VNICs' RSS
and MRU during .ndo_queue_start(), needs to be extended to cover any
non-default RSS contexts which have their own VNICs. Without this
step, packets that are destined to a non-default RSS context may be
dropped after .ndo_queue_start().
We further optimize this scheme by updating the VNIC only if the
RX ring being restarted is in the RSS table of the VNIC. Updating
the VNIC (in particular setting the MRU to 0) will momentarily stop
all traffic to all rings in the RSS table. Any VNIC that has the
RX ring excluded from the RSS table can skip this step and avoid the
traffic disruption.
Note that this scheme is just an improvement. A VNIC with multiple
rings in the RSS table will still see traffic disruptions to all rings
in the RSS table when one of the rings is being restarted. We are
working on a FW scheme that will improve upon this further.
Fixes: 5ac066b7b0 ("bnxt_en: Fix queue start to update vnic RSS table")
Reported-by: David Wei <dw@davidwei.uk>
Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20250613231841.377988-4-michael.chan@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Before the commit under the Fixes tag below, bnxt_ulp_stop() and
bnxt_ulp_start() were always invoked in pairs. After that commit,
the new bnxt_ulp_restart() can be invoked after bnxt_ulp_stop()
has been called. This may result in the RoCE driver's aux driver
.suspend() method being invoked twice. The 2nd bnxt_re_suspend()
call will crash when it dereferences a NULL pointer:
(NULL ib_device): Handle device suspend call
BUG: kernel NULL pointer dereference, address: 0000000000000b78
PGD 0 P4D 0
Oops: Oops: 0000 [#1] SMP PTI
CPU: 20 UID: 0 PID: 181 Comm: kworker/u96:5 Tainted: G S 6.15.0-rc1 #4 PREEMPT(voluntary)
Tainted: [S]=CPU_OUT_OF_SPEC
Hardware name: Dell Inc. PowerEdge R730/072T6D, BIOS 2.4.3 01/17/2017
Workqueue: bnxt_pf_wq bnxt_sp_task [bnxt_en]
RIP: 0010:bnxt_re_suspend+0x45/0x1f0 [bnxt_re]
Code: 8b 05 a7 3c 5b f5 48 89 44 24 18 31 c0 49 8b 5c 24 08 4d 8b 2c 24 e8 ea 06 0a f4 48 c7 c6 04 60 52 c0 48 89 df e8 1b ce f9 ff <48> 8b 83 78 0b 00 00 48 8b 80 38 03 00 00 a8 40 0f 85 b5 00 00 00
RSP: 0018:ffffa2e84084fd88 EFLAGS: 00010246
RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000001
RDX: 0000000000000000 RSI: ffffffffb4b6b934 RDI: 00000000ffffffff
RBP: ffffa1760954c9c0 R08: 0000000000000000 R09: c0000000ffffdfff
R10: 0000000000000001 R11: ffffa2e84084fb50 R12: ffffa176031ef070
R13: ffffa17609775000 R14: ffffa17603adc180 R15: 0000000000000000
FS: 0000000000000000(0000) GS:ffffa17daa397000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000b78 CR3: 00000004aaa30003 CR4: 00000000003706f0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
<TASK>
bnxt_ulp_stop+0x69/0x90 [bnxt_en]
bnxt_sp_task+0x678/0x920 [bnxt_en]
? __schedule+0x514/0xf50
process_scheduled_works+0x9d/0x400
worker_thread+0x11c/0x260
? __pfx_worker_thread+0x10/0x10
kthread+0xfe/0x1e0
? __pfx_kthread+0x10/0x10
ret_from_fork+0x2b/0x40
? __pfx_kthread+0x10/0x10
ret_from_fork_asm+0x1a/0x30
Check the BNXT_EN_FLAG_ULP_STOPPED flag and do not proceed if the flag
is already set. This will preserve the original symmetrical
bnxt_ulp_stop() and bnxt_ulp_start().
Also, inside bnxt_ulp_start(), clear the BNXT_EN_FLAG_ULP_STOPPED
flag after taking the mutex to avoid any race condition. And for
symmetry, only proceed in bnxt_ulp_start() if the
BNXT_EN_FLAG_ULP_STOPPED is set.
Fixes: 3c163f35bd ("bnxt_en: Optimize recovery path ULP locking in the driver")
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Co-developed-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250613231841.377988-2-michael.chan@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Marc Kleine-Budde says:
====================
pull-request: can 2025-06-17
The patch is by Brett Werling, and fixes the power regulator retrieval
during probe of the tcan4x5x glue code for the m_can driver.
* tag 'linux-can-fixes-for-6.16-20250617' of git://git.kernel.org/pub/scm/linux/kernel/git/mkl/linux-can:
can: tcan4x5x: fix power regulator retrieval during probe
openvswitch: Allocate struct ovs_pcpu_storage dynamically
====================
Link: https://patch.msgid.link/20250617155123.2141584-1-mkl@pengutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
skb_ensure_writable should succeed when it's trying to write to the
header of the unreadable skbs, so it doesn't need an unconditional
skb_frags_readable check. The preceding pskb_may_pull() call will
succeed if write_len is within the head and fail if we're trying to
write to the unreadable payload, so we don't need an additional check.
Removing this check restores DSCP functionality with unreadable skbs as
it's called from dscp_tg.
Cc: willemb@google.com
Cc: asml.silence@gmail.com
Fixes: 65249feb6b ("net: add support for skbs with unreadable frags")
Signed-off-by: Mina Almasry <almasrymina@google.com>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Link: https://patch.msgid.link/20250615200733.520113-1-almasrymina@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While transmitting XDP frames for XDP_TX, page_pool is
used to get the DMA buffers (already mapped to the pages)
and need to be freed/reycled once the transmission is complete.
This need not be explicitly done by the driver as this is handled
more gracefully by the xdp driver while returning the xdp frame.
__xdp_return() frees the XDP memory based on its memory type,
under which page_pool memory is also handled. This change fixes
the transmit queue timeout while running XDP_TX.
logs:
[ 309.069682] icssg-prueth icssg1-eth eth2: NETDEV WATCHDOG: CPU: 0: transmit queue 0 timed out 45860 ms
[ 313.933780] icssg-prueth icssg1-eth eth2: NETDEV WATCHDOG: CPU: 0: transmit queue 0 timed out 50724 ms
[ 319.053656] icssg-prueth icssg1-eth eth2: NETDEV WATCHDOG: CPU: 0: transmit queue 0 timed out 55844 ms
...
Fixes: 62aa3246f4 ("net: ti: icssg-prueth: Add XDP support")
Signed-off-by: Meghana Malladi <m-malladi@ti.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://patch.msgid.link/20250616063319.3347541-1-m-malladi@ti.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
On some systems with Nahum 11 and Nahum 13 the value of the XTAL clock in
the software STRAP is incorrect. This causes the PTP timer to run at the
wrong rate and can lead to synchronization issues.
The STRAP value is configured by the system firmware, and a firmware
update is not always possible. Since the XTAL clock on these systems
always runs at 38.4MHz, the driver may ignore the STRAP and just set
the correct value.
Fixes: cc23f4f0b6 ("e1000e: Add support for Meteor Lake")
Signed-off-by: Vitaly Lifshits <vitaly.lifshits@intel.com>
Tested-by: Mor Bar-Gabay <morx.bar.gabay@intel.com>
Reviewed-by: Gil Fine <gil.fine@linux.intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
This patch fixes an issue seen in a large-scale deployment under heavy
incoming pkts where the aRFS flow wrongly matches a flow and reprograms the
NIC with wrong settings. That mis-steering causes RX-path latency spikes
and noisy neighbor effects when many connections collide on the same
hash (some of our production servers have 20-30K connections).
set_rps_cpu() calls ndo_rx_flow_steer() with flow_id that is calculated by
hashing the skb sized by the per rx-queue table size. This results in
multiple connections (even across different rx-queues) getting the same
hash value. The driver steer function modifies the wrong flow to use this
rx-queue, e.g.: Flow#1 is first added:
Flow#1: <ip1, port1, ip2, port2>, Hash 'h', q#10
Later when a new flow needs to be added:
Flow#2: <ip3, port3, ip4, port4>, Hash 'h', q#20
The driver finds the hash 'h' from Flow#1 and updates it to use q#20. This
results in both flows getting un-optimized - packets for Flow#1 goes to
q#20, and then reprogrammed back to q#10 later and so on; and Flow #2
programming is never done as Flow#1 is matched first for all misses. Many
flows may wrongly share the same hash and reprogram rules of the original
flow each with their own q#.
Tested on two 144-core servers with 16K netperf sessions for 180s. Netperf
clients are pinned to cores 0-71 sequentially (so that wrong packets on q#s
72-143 can be measured). IRQs are set 1:1 for queues -> CPUs, enable XPS,
enable aRFS (global value is 144 * rps_flow_cnt).
Test notes about results from ice_rx_flow_steer():
---------------------------------------------------
1. "Skip:" counter increments here:
if (fltr_info->q_index == rxq_idx ||
arfs_entry->fltr_state != ICE_ARFS_ACTIVE)
goto out;
2. "Add:" counter increments here:
ret = arfs_entry->fltr_info.fltr_id;
INIT_HLIST_NODE(&arfs_entry->list_entry);
3. "Update:" counter increments here:
/* update the queue to forward to on an already existing flow */
Runtime comparison: original code vs with the patch for different
rps_flow_cnt values.
+-------------------------------+--------------+--------------+
| rps_flow_cnt | 512 | 2048 |
+-------------------------------+--------------+--------------+
| Ratio of Pkts on Good:Bad q's | 214 vs 822K | 1.1M vs 980K |
| Avoid wrong aRFS programming | 0 vs 310K | 0 vs 30K |
| CPU User | 216 vs 183 | 216 vs 206 |
| CPU System | 1441 vs 1171 | 1447 vs 1320 |
| CPU Softirq | 1245 vs 920 | 1238 vs 961 |
| CPU Total | 29 vs 22.7 | 29 vs 24.9 |
| aRFS Update | 533K vs 59 | 521K vs 32 |
| aRFS Skip | 82M vs 77M | 7.2M vs 4.5M |
+-------------------------------+--------------+--------------+
A separate TCP_STREAM and TCP_RR with 1,4,8,16,64,128,256,512 connections
showed no performance degradation.
Some points on the patch/aRFS behavior:
1. Enabling full tuple matching ensures flows are always correctly matched,
even with smaller hash sizes.
2. 5-6% drop in CPU utilization as the packets arrive at the correct CPUs
and fewer calls to driver for programming on misses.
3. Larger hash tables reduces mis-steering due to more unique flow hashes,
but still has clashes. However, with larger per-device rps_flow_cnt, old
flows take more time to expire and new aRFS flows cannot be added if h/w
limits are reached (rps_may_expire_flow() succeeds when 10*rps_flow_cnt
pkts have been processed by this cpu that are not part of the flow).
Fixes: 28bf26724f ("ice: Implement aRFS")
Signed-off-by: Krishna Kumar <krikku@gmail.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Rinitha S <sx.rinitha@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
PERCPU_MODULE_RESERVE defines the maximum size that can by used for the
per-CPU data size used by modules. This is 8KiB.
Commit 035fcdc4d2 ("openvswitch: Merge three per-CPU structures into
one") restructured the per-CPU memory allocation for the module and
moved the separate alloc_percpu() invocations at module init time to a
static per-CPU variable which is allocated by the module loader.
The size of the per-CPU data section for openvswitch is 6488 bytes which
is ~80% of the available per-CPU memory. Together with a few other
modules it is easy to exhaust the available 8KiB of memory.
Allocate ovs_pcpu_storage dynamically at module init time.
Reported-by: Gal Pressman <gal@nvidia.com>
Closes: https://lore.kernel.org/all/c401e017-f8db-4f57-a1cd-89beb979a277@nvidia.com
Fixes: 035fcdc4d2 ("openvswitch: Merge three per-CPU structures into one")
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Aaron Conole <aconole@redhat.com>
Link: https://patch.msgid.link/20250613123629.-XSoQTCu@linutronix.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
pldmfw calls crc32 code and depends on it being enabled, else
there is a link error as follows. So PLDMFW should select CRC32.
lib/pldmfw/pldmfw.o: In function `pldmfw_flash_image':
pldmfw.c:(.text+0x70f): undefined reference to `crc32_le_base'
This problem was introduced by commit b8265621f4 ("Add pldmfw library
for PLDM firmware update").
It manifests as of commit d69ea414c9 ("ice: implement device flash
update via devlink").
And is more likely to occur as of commit 9ad19171b6 ("lib/crc: remove
unnecessary prompt for CONFIG_CRC32 and drop 'default y'").
Found by chance while exercising builds based on tinyconfig.
Fixes: b8265621f4 ("Add pldmfw library for PLDM firmware update")
Signed-off-by: Simon Horman <horms@kernel.org>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://patch.msgid.link/20250613-pldmfw-crc32-v1-1-f3fad109eee6@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some stress/negative firmware testing around devcmd(s) returning
EAGAIN found that the done bit could get out of sync in the
firmware when it wasn't cleared in a retry case.
While here, change the type of the local done variable to a bool
to match the return type from ionic_dev_cmd_done().
Fixes: ec8ee71473 ("ionic: stretch heartbeat detection")
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250609212827.53842-1-shannon.nelson@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pull networking fixes from Jakub Kicinski:
"Including fixes from bluetooth and wireless.
Current release - regressions:
- af_unix: allow passing cred for embryo without SO_PASSCRED/SO_PASSPIDFD
Current release - new code bugs:
- eth: airoha: correct enable mask for RX queues 16-31
- veth: prevent NULL pointer dereference in veth_xdp_rcv when peer
disappears under traffic
- ipv6: move fib6_config_validate() to ip6_route_add(), prevent
invalid routes
Previous releases - regressions:
- phy: phy_caps: don't skip better duplex match on non-exact match
- dsa: b53: fix untagged traffic sent via cpu tagged with VID 0
- Revert "wifi: mwifiex: Fix HT40 bandwidth issue.", it caused
transient packet loss, exact reason not fully understood, yet
Previous releases - always broken:
- net: clear the dst when BPF is changing skb protocol (IPv4 <> IPv6)
- sched: sfq: fix a potential crash on gso_skb handling
- Bluetooth: intel: improve rx buffer posting to avoid causing issues
in the firmware
- eth: intel: i40e: make reset handling robust against multiple
requests
- eth: mlx5: ensure FW pages are always allocated on the local NUMA
node, even when device is configure to 'serve' another node
- wifi: ath12k: fix GCC_GCC_PCIE_HOT_RST definition for WCN7850,
prevent kernel crashes
- wifi: ath11k: avoid burning CPU in ath11k_debugfs_fw_stats_request()
for 3 sec if fw_stats_done is not set"
* tag 'net-6.16-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (70 commits)
selftests: drv-net: rss_ctx: Add test for ntuple rules targeting default RSS context
net: ethtool: Don't check if RSS context exists in case of context 0
af_unix: Allow passing cred for embryo without SO_PASSCRED/SO_PASSPIDFD.
ipv6: Move fib6_config_validate() to ip6_route_add().
net: drv: netdevsim: don't napi_complete() from netpoll
net/mlx5: HWS, Add error checking to hws_bwc_rule_complex_hash_node_get()
veth: prevent NULL pointer dereference in veth_xdp_rcv
net_sched: remove qdisc_tree_flush_backlog()
net_sched: ets: fix a race in ets_qdisc_change()
net_sched: tbf: fix a race in tbf_change()
net_sched: red: fix a race in __red_change()
net_sched: prio: fix a race in prio_tune()
net_sched: sch_sfq: reject invalid perturb period
net: phy: phy_caps: Don't skip better duplex macth on non-exact match
MAINTAINERS: Update Kuniyuki Iwashima's email address.
selftests: net: add test case for NAT46 looping back dst
net: clear the dst when changing skb protocol
net/mlx5e: Fix number of lanes to UNKNOWN when using data_rate_oper
net/mlx5e: Fix leak of Geneve TLV option object
net/mlx5: HWS, make sure the uplink is the last destination
...
Pull pin control fixes from Linus Walleij:
- Add some missing pins on the Qualcomm QCM2290, along with a managed
resources patch that make it clean and nice
- Drop an unused function in the ST Micro driver
- Drop bouncing MAINTAINER entry
- Drop of_match_ptr() macro to rid compile warnings in the TB10x
driver
- Fix up calculation of pin numbers from base in the Sunxi driver
* tag 'pinctrl-v6.16-2' of git://git.kernel.org/pub/scm/linux/kernel/git/linusw/linux-pinctrl:
pinctrl: sunxi: dt: Consider pin base when calculating bank number from pin
pinctrl: tb10x: Drop of_match_ptr for ID table
pinctrl: MAINTAINERS: Drop bouncing Jianlong Huang
pinctrl: st: Drop unused st_gpio_bank() function
pinctrl: qcom: pinctrl-qcm2290: Add missing pins
pinctrl: qcom: switch to devm_gpiochip_add_data()
Pull ARC fixes from Vineet Gupta:
- arch_atomic64_cmpxchg relaxed variant [Jason]
- use of inbuilt swap in stack unwinder [Yu-Chun Lin]
- use of __ASSEMBLER__ in kernel headers [Thomas Huth]
* tag 'arc-6.16-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/vgupta/arc:
ARC: Replace __ASSEMBLY__ with __ASSEMBLER__ in the non-uapi headers
ARC: Replace __ASSEMBLY__ with __ASSEMBLER__ in uapi headers
ARC: unwind: Use built-in sort swap to reduce code size and improve performance
ARC: atomics: Implement arch_atomic64_cmpxchg using _relaxed
Johannes Berg says:
====================
Another quick round of updates:
- revert mwifiex HT40 that was causing issues
- many ath10k/ath11k/ath12k fixes
- re-add some iwlwifi code I lost in a merge
- use kfree_sensitive() on an error path in cfg80211
* tag 'wireless-2025-06-12' of https://git.kernel.org/pub/scm/linux/kernel/git/wireless/wireless:
wifi: cfg80211: use kfree_sensitive() for connkeys cleanup
wifi: iwlwifi: fix merge damage related to iwl_pci_resume
Revert "wifi: mwifiex: Fix HT40 bandwidth issue."
wifi: ath12k: fix uaf in ath12k_core_init()
wifi: ath12k: Fix hal_reo_cmd_status kernel-doc
wifi: ath12k: fix GCC_GCC_PCIE_HOT_RST definition for WCN7850
wifi: ath11k: validate ath11k_crypto_mode on top of ath11k_core_qmi_firmware_ready
wifi: ath11k: consistently use ath11k_mac_get_fw_stats()
wifi: ath11k: move locking outside of ath11k_mac_get_fw_stats()
wifi: ath11k: adjust unlock sequence in ath11k_update_stats_event()
wifi: ath11k: move some firmware stats related functions outside of debugfs
wifi: ath11k: don't wait when there is no vdev started
wifi: ath11k: don't use static variables in ath11k_debugfs_fw_stats_process()
wifi: ath11k: avoid burning CPU in ath11k_debugfs_fw_stats_request()
wil6210: fix support for sparrow chipsets
wifi: ath10k: Avoid vdev delete timeout when firmware is already down
ath10k: snoc: fix unbalanced IRQ enable in crash recovery
====================
Link: https://patch.msgid.link/20250612082519.11447-3-johannes@sipsolutions.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Gal Pressman says:
====================
Fix ntuple rules targeting default RSS
This series addresses a regression in ethtool flow steering where rules
targeting the default RSS context (context 0) were incorrectly rejected.
The default RSS context always exists but is not stored in the rss_ctx
xarray like additional contexts. The current validation logic was
checking for the existence of context 0 in this array, causing valid
flow steering rules to be rejected.
This prevented configurations such as:
- High priority rules directing specific traffic to the default context
- Low priority catch-all rules directing remaining traffic to additional
contexts
Patch 1 fixes the validation logic to skip the existence check for
context 0.
Patch 2 adds a selftest that verifies this behavior.
v3: https://lore.kernel.org/20250609120250.1630125-1-gal@nvidia.com
v2: https://lore.kernel.org/20250225071348.509432-1-gal@nvidia.com
====================
Link: https://patch.msgid.link/20250612071958.1696361-1-gal@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add test_rss_default_context_rule() to verify that ntuple rules can
correctly direct traffic to the default RSS context (context 0).
The test creates two ntuple rules with explicit location priorities:
- A high-priority rule (loc 0) directing specific port traffic to
context 0.
- A low-priority rule (loc 1) directing all other TCP traffic to context
1.
This validates that:
1. Rules targeting the default context function properly.
2. Traffic steering works as expected when mixing default and
additional RSS contexts.
The test was written by AI, and reviewed by humans.
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Gal Pressman <gal@nvidia.com>
Link: https://patch.msgid.link/20250612071958.1696361-3-gal@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Context 0 (default context) always exists, there is no need to check
whether it exists or not when adding a flow steering rule.
The existing check fails when creating a flow steering rule for context
0 as it is not stored in the rss_ctx xarray.
For example:
$ ethtool --config-ntuple eth2 flow-type tcp4 dst-ip 194.237.147.23 dst-port 19983 context 0 loc 618
rmgr: Cannot insert RX class rule: Invalid argument
Cannot insert classification rule
An example usecase for this could be:
- A high-priority rule (loc 0) directing specific port traffic to
context 0.
- A low-priority rule (loc 1) directing all other TCP traffic to context
1.
This is a user-visible regression that was caught in our testing
environment, it was not reported by a user yet.
Fixes: de7f7582df ("net: ethtool: prevent flow steering to RSS contexts which don't exist")
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Nimrod Oren <noren@nvidia.com>
Signed-off-by: Gal Pressman <gal@nvidia.com>
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Edward Cree <ecree.xilinx@gmail.com>
Link: https://patch.msgid.link/20250612071958.1696361-2-gal@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Luiz Augusto von Dentz says:
====================
bluetooth pull request for net:
- eir: Fix NULL pointer deference on eir_get_service_data
- eir: Fix possible crashes on eir_create_adv_data
- hci_sync: Fix broadcast/PA when using an existing instance
- ISO: Fix using BT_SK_PA_SYNC to detect BIS sockets
- ISO: Fix not using bc_sid as advertisement SID
- MGMT: Fix sparse errors
* tag 'for-net-2025-06-11' of git://git.kernel.org/pub/scm/linux/kernel/git/bluetooth/bluetooth:
Bluetooth: MGMT: Fix sparse errors
Bluetooth: ISO: Fix not using bc_sid as advertisement SID
Bluetooth: ISO: Fix using BT_SK_PA_SYNC to detect BIS sockets
Bluetooth: eir: Fix possible crashes on eir_create_adv_data
Bluetooth: hci_sync: Fix broadcast/PA when using an existing instance
Bluetooth: Fix NULL pointer deference on eir_get_service_data
====================
Link: https://patch.msgid.link/20250611204944.1559356-1-luiz.dentz@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Before the cited commit, the kernel unconditionally embedded SCM
credentials to skb for embryo sockets even when both the sender
and listener disabled SO_PASSCRED and SO_PASSPIDFD.
Now, the credentials are added to skb only when configured by the
sender or the listener.
However, as reported in the link below, it caused a regression for
some programs that assume credentials are included in every skb,
but sometimes not now.
The only problematic scenario would be that a socket starts listening
before setting the option. Then, there will be 2 types of non-small
race window, where a client can send skb without credentials, which
the peer receives as an "invalid" message (and aborts the connection
it seems ?):
Client Server
------ ------
s1.listen() <-- No SO_PASS{CRED,PIDFD}
s2.connect()
s2.send() <-- w/o cred
s1.setsockopt(SO_PASS{CRED,PIDFD})
s2.send() <-- w/ cred
or
Client Server
------ ------
s1.listen() <-- No SO_PASS{CRED,PIDFD}
s2.connect()
s2.send() <-- w/o cred
s3, _ = s1.accept() <-- Inherit cred options
s2.send() <-- w/o cred but not set yet
s3.setsockopt(SO_PASS{CRED,PIDFD})
s2.send() <-- w/ cred
It's unfortunate that buggy programs depend on the behaviour,
but let's restore the previous behaviour.
Fixes: 3f84d577b7 ("af_unix: Inherit sk_flags at connect().")
Reported-by: Jacek Łuczak <difrost.kernel@gmail.com>
Closes: https://lore.kernel.org/all/68d38b0b-1666-4974-85d4-15575789c8d4@gmail.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Tested-by: Christian Heusel <christian@heusel.eu>
Tested-by: André Almeida <andrealmeid@igalia.com>
Tested-by: Jacek Łuczak <difrost.kernel@gmail.com>
Link: https://patch.msgid.link/20250611202758.3075858-1-kuni1840@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
netdevsim supports netpoll. Make sure we don't call napi_complete()
from it, since it may not be scheduled. Breno reports hitting a
warning in napi_complete_done():
WARNING: CPU: 14 PID: 104 at net/core/dev.c:6592 napi_complete_done+0x2cc/0x560
__napi_poll+0x2d8/0x3a0
handle_softirqs+0x1fe/0x710
This is presumably after netpoll stole the SCHED bit prematurely.
Reported-by: Breno Leitao <leitao@debian.org>
Fixes: 3762ec05a9 ("netdevsim: add NAPI support")
Tested-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20250611174643.2769263-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Eric Dumazet says:
====================
net_sched: no longer use qdisc_tree_flush_backlog()
This series is based on a report from Gerrard Tai.
Essentially, all users of qdisc_tree_flush_backlog() are racy.
We must instead use qdisc_purge_queue().
====================
Link: https://patch.msgid.link/20250611111515.1983366-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Gerrard Tai reported a race condition in ETS, whenever SFQ perturb timer
fires at the wrong time.
The race is as follows:
CPU 0 CPU 1
[1]: lock root
[2]: qdisc_tree_flush_backlog()
[3]: unlock root
|
| [5]: lock root
| [6]: rehash
| [7]: qdisc_tree_reduce_backlog()
|
[4]: qdisc_put()
This can be abused to underflow a parent's qlen.
Calling qdisc_purge_queue() instead of qdisc_tree_flush_backlog()
should fix the race, because all packets will be purged from the qdisc
before releasing the lock.
Fixes: b05972f01e ("net: sched: tbf: don't call qdisc_put() while holding tree lock")
Reported-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Suggested-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250611111515.1983366-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Gerrard Tai reported a race condition in TBF, whenever SFQ perturb timer
fires at the wrong time.
The race is as follows:
CPU 0 CPU 1
[1]: lock root
[2]: qdisc_tree_flush_backlog()
[3]: unlock root
|
| [5]: lock root
| [6]: rehash
| [7]: qdisc_tree_reduce_backlog()
|
[4]: qdisc_put()
This can be abused to underflow a parent's qlen.
Calling qdisc_purge_queue() instead of qdisc_tree_flush_backlog()
should fix the race, because all packets will be purged from the qdisc
before releasing the lock.
Fixes: b05972f01e ("net: sched: tbf: don't call qdisc_put() while holding tree lock")
Reported-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Suggested-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Zhengchao Shao <shaozhengchao@huawei.com>
Link: https://patch.msgid.link/20250611111515.1983366-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Gerrard Tai reported a race condition in RED, whenever SFQ perturb timer
fires at the wrong time.
The race is as follows:
CPU 0 CPU 1
[1]: lock root
[2]: qdisc_tree_flush_backlog()
[3]: unlock root
|
| [5]: lock root
| [6]: rehash
| [7]: qdisc_tree_reduce_backlog()
|
[4]: qdisc_put()
This can be abused to underflow a parent's qlen.
Calling qdisc_purge_queue() instead of qdisc_tree_flush_backlog()
should fix the race, because all packets will be purged from the qdisc
before releasing the lock.
Fixes: 0c8d13ac96 ("net: sched: red: delay destroying child qdisc on replace")
Reported-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Suggested-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250611111515.1983366-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Gerrard Tai reported a race condition in PRIO, whenever SFQ perturb timer
fires at the wrong time.
The race is as follows:
CPU 0 CPU 1
[1]: lock root
[2]: qdisc_tree_flush_backlog()
[3]: unlock root
|
| [5]: lock root
| [6]: rehash
| [7]: qdisc_tree_reduce_backlog()
|
[4]: qdisc_put()
This can be abused to underflow a parent's qlen.
Calling qdisc_purge_queue() instead of qdisc_tree_flush_backlog()
should fix the race, because all packets will be purged from the qdisc
before releasing the lock.
Fixes: 7b8e0b6e65 ("net: sched: prio: delay destroying child qdiscs on change")
Reported-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Suggested-by: Gerrard Tai <gerrard.tai@starlabs.sg>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250611111515.1983366-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
