Jason Xing [Sat, 2 May 2026 20:07:20 +0000 (23:07 +0300)]
xsk: avoid skb leak in XDP_TX_METADATA case
Fix it by explicitly adding kfree_skb() before returning back to its
caller.
How to reproduce it in virtio_net:
1. the current skb is the first one (which means no frag and xs->skb is
NULL) and users enable metadata feature.
2. xsk_skb_metadata() returns a error code.
3. the caller xsk_build_skb() clears skb by using 'skb = NULL;'.
4. there is no chance to free this skb anymore.
Closes: https://lore.kernel.org/all/20260415085204.3F87AC19424@smtp.kernel.org/ Fixes: 30c3055f9c0d ("xsk: wrap generic metadata handling onto separate function") Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com> Link: https://patch.msgid.link/20260502200722.53960-7-kerneljasonxing@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jason Xing [Sat, 2 May 2026 20:07:19 +0000 (23:07 +0300)]
xsk: prevent CQ desync when freeing half-built skbs in xsk_build_skb()
Once xsk_skb_init_misc() has been called on an skb, its destructor is
set to xsk_destruct_skb(), which submits the descriptor address(es) to
the completion queue and advances the CQ producer. If such an skb is
subsequently freed via kfree_skb() along an error path - before the
skb has ever been handed to the driver - the destructor still runs and
submits a bogus, half-initialized address to the CQ.
Postpone the init phase when we believe the allocation of first frag is
successfully completed. Before this init, skb can be safely freed by
kfree_skb().
Closes: https://lore.kernel.org/all/20260419045822.843BFC2BCAF@smtp.kernel.org/ Fixes: c30d084960cf ("xsk: avoid overwriting skb fields for multi-buffer traffic") Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com> Link: https://patch.msgid.link/20260502200722.53960-6-kerneljasonxing@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jason Xing [Sat, 2 May 2026 20:07:18 +0000 (23:07 +0300)]
xsk: fix use-after-free of xs->skb in xsk_build_skb() free_err path
When xsk_build_skb() processes multi-buffer packets in copy mode, the
first descriptor stores data into the skb linear area without adding
any frags, so nr_frags stays at 0. The caller then sets xs->skb = skb
to accumulate subsequent descriptors.
If a continuation descriptor fails (e.g. alloc_page returns NULL with
-EAGAIN), we jump to free_err where the condition:
if (skb && !skb_shinfo(skb)->nr_frags)
kfree_skb(skb);
evaluates to true because nr_frags is still 0 (the first descriptor
used the linear area, not frags). This frees the skb while xs->skb
still points to it, creating a dangling pointer. On the next transmit
attempt or socket close, xs->skb is dereferenced, causing a
use-after-free or double-free.
Fix by using a !xs->skb check to handle first frag situation, ensuring
we only free skbs that were freshly allocated in this call
(xs->skb is NULL) and never free an in-progress multi-buffer skb that
the caller still references.
Closes: https://lore.kernel.org/all/20260415082654.21026-4-kerneljasonxing@gmail.com/ Fixes: 6b9c129c2f93 ("xsk: remove @first_frag from xsk_build_skb()") Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com> Link: https://patch.msgid.link/20260502200722.53960-5-kerneljasonxing@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jason Xing [Sat, 2 May 2026 20:07:17 +0000 (23:07 +0300)]
xsk: handle NULL dereference of the skb without frags issue
When a first descriptor (xs->skb == NULL) triggers -EOVERFLOW in
xsk_build_skb_zerocopy() (e.g., MAX_SKB_FRAGS exceeded), the
free_err -EOVERFLOW handler unconditionally dereferences xs->skb
via xsk_inc_num_desc(xs->skb) and xsk_drop_skb(xs->skb), causing
a NULL pointer dereference.
Fix this by guarding the existing xsk_inc_num_desc()/xsk_drop_skb()
calls with an xs->skb check (for the continuation case), and add
an else branch for the first-descriptor case that manually cancels
the one reserved CQ slot and increments invalid_descs by one to
account for the single invalid descriptor.
Fixes: cf24f5a5feea ("xsk: add support for AF_XDP multi-buffer on Tx path") Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com> Link: https://patch.msgid.link/20260502200722.53960-4-kerneljasonxing@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jason Xing [Sat, 2 May 2026 20:07:16 +0000 (23:07 +0300)]
xsk: free the skb when hitting the upper bound MAX_SKB_FRAGS
Fix it by explicitly adding kfree_skb() before returning back to its
caller.
How to reproduce it in virtio_net:
1. the current skb is the first one (which means xs->skb is NULL) and
hit the limit MAX_SKB_FRAGS.
2. xsk_build_skb_zerocopy() returns -EOVERFLOW.
3. the caller xsk_build_skb() clears skb by using 'skb = NULL;'. This
is why bug can be triggered.
4. there is no chance to free this skb anymore.
Note that if in this case the xs->skb is not NULL, xsk_build_skb() will
call xsk_drop_skb(xs->skb) to do the right thing.
Fixes: cf24f5a5feea ("xsk: add support for AF_XDP multi-buffer on Tx path") Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com> Link: https://patch.msgid.link/20260502200722.53960-3-kerneljasonxing@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jason Xing [Sat, 2 May 2026 20:07:15 +0000 (23:07 +0300)]
xsk: reject sw-csum UMEM binding to IFF_TX_SKB_NO_LINEAR devices
skb_checksum_help() is a common helper that writes the folded
16-bit checksum back via skb->data + csum_start + csum_offset,
i.e. it relies on the skb's linear head and fails (with WARN_ONCE
and -EINVAL) when skb_headlen() is 0.
AF_XDP generic xmit takes two very different paths depending on the
netdev. Drivers that advertise IFF_TX_SKB_NO_LINEAR (e.g. virtio_net)
skip the "copy payload into a linear head" step on purpose as a
performance optimisation: xsk_build_skb_zerocopy() only attaches UMEM
pages as frags and never calls skb_put(), so skb_headlen() stays 0
for the whole skb. For these skbs there is simply no linear area for
skb_checksum_help() to write the csum into - the sw-csum fallback is
structurally inapplicable.
The patch tries to catch this and reject the combination with error at
setup time. Rejecting at bind() converts this silent per-packet failure
into a synchronous, actionable -EOPNOTSUPP at setup time. HW csum and
launch_time metadata on IFF_TX_SKB_NO_LINEAR drivers are unaffected
because they do not call skb_checksum_help().
Without the patch, every descriptor carrying 'XDP_TX_METADATA |
XDP_TXMD_FLAGS_CHECKSUM' produces:
1) a WARN_ONCE "offset (N) >= skb_headlen() (0)" from skb_checksum_help(),
2) sendmsg() returning -EINVAL without consuming the descriptor
(invalid_descs is not incremented),
3) a wedged TX ring: __xsk_generic_xmit() does not advance the
consumer on non-EOVERFLOW errors, so the next sendmsg() re-reads
the same descriptor and re-hits the same WARN until the socket
is closed.
Closes: https://lore.kernel.org/all/20260419045822.843BFC2BCAF@smtp.kernel.org/#t Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Jason Xing <kernelxing@tencent.com> Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com> Fixes: 30c3055f9c0d ("xsk: wrap generic metadata handling onto separate function") Link: https://patch.msgid.link/20260502200722.53960-2-kerneljasonxing@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jakub Kicinski [Wed, 6 May 2026 02:13:12 +0000 (19:13 -0700)]
Merge branch 'net-mlx5-fixes-for-socket-direct'
Tariq Toukan says:
====================
net/mlx5: Fixes for Socket-Direct
This series fixes several race conditions and bugs in the mlx5
Socket-Direct (SD) single netdev flow.
Patch 1 serializes mlx5_sd_init()/mlx5_sd_cleanup() with
mlx5_devcom_comp_lock() and tracks the SD group state on the primary
device, preventing concurrent or duplicate bring-up/tear-down.
Patch 2 fixes the debugfs "multi-pf" directory being stored on the
calling device's sd struct instead of the primary's, which caused
memory leaks and recreation errors when cleanup ran from a different PF.
Patch 3 fixes a race where a secondary PF could access the primary's
auxiliary device after it had been unbound, by holding the primary's
device lock while operating on its auxiliary device.
Patch 4 fixes missing cleanup on ETH probe errors. The analogous gap on
the resume path requires introducing sd_suspend/resume APIs that only
destroy FW resources and is left for a follow-up series.
====================
Shay Drory [Mon, 4 May 2026 18:02:06 +0000 (21:02 +0300)]
net/mlx5e: SD, Fix race condition in secondary device probe/remove
When utilizing Socket-Direct single netdev functionality the driver
resolves the actual auxiliary device using mlx5_sd_get_adev(). However,
the current implementation returns the primary ETH auxiliary device
without holding the device lock, leading to a potential race condition
where the ETH device could be unbound or removed concurrently during
probe, suspend, resume, or remove operations.[1]
Fix this by introducing mlx5_sd_put_adev() and updating
mlx5_sd_get_adev() so that secondaries devices would get a ref and
acquire the device lock of the returned auxiliary device. After the lock
is acquired, a second devcom check is needed[2].
In addition, update The callers to pair the get operation with the new
put operation, ensuring the lock is held while the auxiliary device is
being operated on and released afterwards.
The "primary" designation is determined once in sd_register(). It's set
before devcom is marked ready, and it never changes after that.
In Addition, The primary path never locks a secondary: When the primary
device invoke mlx5_sd_get_adev(), it sees dev == primary and returns.
no additional lock is taken.
Therefore lock ordering is always: secondary_lock -> primary_lock. The
reverse never happens, so ABBA deadlock is impossible.
Shay Drory [Mon, 4 May 2026 18:02:05 +0000 (21:02 +0300)]
net/mlx5e: SD, Fix missing cleanup on probe error
When _mlx5e_probe() fails, the preceding successful mlx5_sd_init() is
not undone. Auxiliary bus probe failure skips binding, so mlx5e_remove()
is never called for that adev and the matching mlx5_sd_cleanup() never
runs - leaking the per-dev SD struct.
Call mlx5_sd_cleanup() on the probe error path to balance
mlx5_sd_init().
A similar gap exists on the resume path: mlx5_sd_init() and
mlx5_sd_cleanup() are currently bundled with both probe/remove and
suspend/resume, even though only the FW alias state actually needs to
follow the suspend/resume lifecycle - the sd struct allocation and
devcom membership are software state that should track the full bound
lifetime. As a result, a failed resume can leave a still-bound device
with sd == NULL, which mlx5_sd_get_adev() can't distinguish from a
non-SD device. Fixing this requires sd_suspend/resume APIs which will
only destroy FW resources and is left for a follow-up series.
Fixes: 381978d28317 ("net/mlx5e: Create single netdev per SD group") Signed-off-by: Shay Drory <shayd@nvidia.com> Signed-off-by: Tariq Toukan <tariqt@nvidia.com> Link: https://patch.msgid.link/20260504180206.268568-4-tariqt@nvidia.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Shay Drory [Mon, 4 May 2026 18:02:04 +0000 (21:02 +0300)]
net/mlx5: SD, Keep multi-pf debugfs entries on primary
mlx5_sd_init() creates the "multi-pf" debugfs directory under the
primary device debugfs root, but stored the dentry in the calling
device's sd struct. When sd_cleanup() run on a different PF,
this leads to using the wrong sd->dfs for removing entries, which
results in memory leak and an error in when re-creating the SD.[1]
Fix it by explicitly storing the debugfs dentry in the primary
device sd struct and use it for all per-group files.
[1]
debugfs: 'multi-pf' already exists in '0000:08:00.1'
Shay Drory [Mon, 4 May 2026 18:02:03 +0000 (21:02 +0300)]
net/mlx5: SD: Serialize init/cleanup
mlx5_sd_init() / mlx5_sd_cleanup() may run from multiple PFs in the same
Socket-Direct group. This can cause the SD bring-up/tear-down sequence
to be executed more than once or interleaved across PFs.
Protect SD init/cleanup with mlx5_devcom_comp_lock() and track the SD
group state on the primary device. Skip init if the primary is already
UP, and skip cleanup unless the primary is UP.
The state check on cleanup is needed because sd_register() drops the
devcom comp lock between marking the comp ready and assigning
primary_dev on each peer. A concurrent cleanup that acquires the lock
in this window would observe devcom_is_ready==true while primary_dev
is still NULL (causing mlx5_sd_get_primary() to return NULL) or while
the FW alias setup performed by mlx5_sd_init()'s body has not yet run
(causing sd_cmd_unset_primary() to dereference a NULL tx_ft). Gate the
cleanup body on primary_sd->state == MLX5_SD_STATE_UP, which is set
only at the very end of mlx5_sd_init() under the same comp lock - so
observing UP guarantees primary_dev, secondaries[], tx_ft, and dfs are
all populated. Also bail explicitly if mlx5_sd_get_primary() returns
NULL, in case state is checked on a peer whose primary_dev hasn't been
assigned yet.
In addition, move mlx5_devcom_comp_set_ready(false) from sd_unregister()
into the cleanup's locked section, including the !primary and
state != UP early-exit paths, so the device cannot unregister and free
its struct mlx5_sd while devcom is still marked ready. A concurrent
init acquiring the devcom lock will now observe devcom is no longer
ready and bail out immediately.
Fixes: 381978d28317 ("net/mlx5e: Create single netdev per SD group") Signed-off-by: Shay Drory <shayd@nvidia.com> Signed-off-by: Tariq Toukan <tariqt@nvidia.com> Link: https://patch.msgid.link/20260504180206.268568-2-tariqt@nvidia.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The reason for failure is the existence of TX keys, which are removed by
the PSP dev unregistration happening in:
profile->cleanup() -> mlx5e_psp_unregister() -> mlx5e_psp_cleanup()
-> psp_dev_unregister()
...but this isn't invoked in the devlink reload flow, only when changing
the NIC profile (e.g. when transitioning to switchdev mode) or on dev
teardown.
Move PSP device registration into mlx5e_nic_enable(), and unregistration
into the corresponding mlx5e_nic_disable(). These functions are called
during netdev attach/detach after RX & TX are set up.
This ensures that the keys will be gone by the time the PD is destroyed.
Cosmin Ratiu [Mon, 4 May 2026 18:10:58 +0000 (21:10 +0300)]
net/mlx5e: psp: Fix invalid access on PSP dev registration fail
priv->psp->psp is initialized with the PSP device as returned by
psp_dev_create(). This could also return an error, in which case a
future psp_dev_unregister() will result in unpleasantness.
Avoid that by using a local variable and only saving the PSP device when
registration succeeds.
In case psp_dev_create() fails, priv->psp and steering structs are left
in place, but they will be inert. The unchecked access of priv->psp in
mlx5e_psp_offload_handle_rx_skb() won't happen because without a PSP
device, there can be no SAs added and therefore no packets will be
successfully decrypted and be handed off to the SW handler.
Pavitra Jha [Fri, 1 May 2026 11:07:12 +0000 (07:07 -0400)]
net: wwan: t7xx: validate port_count against message length in t7xx_port_enum_msg_handler
t7xx_port_enum_msg_handler() uses the modem-supplied port_count field as
a loop bound over port_msg->data[] without checking that the message buffer
contains sufficient data. A modem sending port_count=65535 in a 12-byte
buffer triggers a slab-out-of-bounds read of up to 262140 bytes.
Add a sizeof(*port_msg) check before accessing the port message header
fields to guard against undersized messages.
Add a struct_size() check after extracting port_count and before the loop.
In t7xx_parse_host_rt_data(), guard the rt_feature header read with a
remaining-buffer check before accessing data_len, validate feat_data_len
against the actual remaining buffer to prevent OOB reads and signed
integer overflow on offset.
Pass msg_len from both call sites: skb->len at the DPMAIF path after
skb_pull(), and the validated feat_data_len at the handshake path.
Eric Dumazet [Mon, 4 May 2026 16:38:42 +0000 (16:38 +0000)]
net/sched: sch_fq_codel: annotate data-races from fq_codel_dump_class_stats()
fq_codel_dump_class_stats() acquires qdisc spinlock only when requested
to follow flow->head chain.
As we did in sch_cake recently, add the missing READ_ONCE()/WRITE_ONCE()
annotations.
Fixes: edb09eb17ed8 ("net: sched: do not acquire qdisc spinlock in qdisc/class stats dump") Signed-off-by: Eric Dumazet <edumazet@google.com> Reviewed-by: Jamal Hadi Salim <jhs@mojatatu.com> Link: https://patch.msgid.link/20260504163842.1162001-1-edumazet@google.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
I added missing Fixes tag. The original description:
Since commit 041ee6f3727a ("kthread: Rely on HK_TYPE_DOMAIN for preferred
affinity management"), the HK_TYPE_KTHREAD housekeeping cpumask may no
longer be correct in showing the actual CPU affinity of kthreads that
have no predefined CPU affinity. As the ipvs networking code is still
using HK_TYPE_KTHREAD, we need to make HK_TYPE_KTHREAD reflect the
reality.
This patch series makes HK_TYPE_KTHREAD an alias of HK_TYPE_DOMAIN
and uses RCU to protect access to the HK_TYPE_KTHREAD housekeeping
cpumask.
Julian plans to post a nf-next patch to limit the connections by using
"conn_max" sysctl. With Simon Horman, they agreed that this is an old
problem that we do not have a limit of connections and it is not a
stopper for this patchset.
* tag 'nf-26-05-05' of git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf:
sched/isolation: Make HK_TYPE_KTHREAD an alias of HK_TYPE_DOMAIN
ipvs: Guard access of HK_TYPE_KTHREAD cpumask with RCU
ipvs: fix shift-out-of-bounds in ip_vs_rht_desired_size
ipvs: fix races around est_mutex and est_cpulist
ipvs: do not leak dest after get from dest trash
ipvs: fix the spin_lock usage for RT build
ipvs: fix races around the conn_lfactor and svc_lfactor sysctl vars
ipvs: fixes for the new ip_vs_status info
====================
Pavan Chebbi [Mon, 4 May 2026 08:36:11 +0000 (14:06 +0530)]
bnxt_en: Use absolute target ns from ptp_clock_request
There is no need to calculate the target PHC cycles required
to make phase adjustment on the PPS OUT signal. This is because
the application supplies absolute n_sec value in the future and
is already the actual desired target value.
Remove the unnecessary code.
Fixes: 9e518f25802c ("bnxt_en: 1PPS functions to configure TSIO pins") Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Cc: Richard Cochran <richardcochran@gmail.com> Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com> Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev> Tested-by: Vadim Fedorenko <vadim.fedorenko@linux.dev> Link: https://patch.msgid.link/20260504083611.1383776-5-pavan.chebbi@broadcom.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Kalesh AP [Mon, 4 May 2026 08:36:10 +0000 (14:06 +0530)]
bnxt_en: Check return value of bnxt_hwrm_vnic_cfg
When the bnxt RDMA driver is loaded, it calls bnxt_register_dev().
As part of this, driver sends HWRM_VNIC_CFG firmware command
to configure the VNIC to operate in dual VNIC mode. Currently
the driver ignores the result of this firmware command. The RDMA
driver must know the result since it affects its functioning.
Check return value of call to bnxt_hwrm_vnic_cfg() in
bnxt_register_dev() and return failure on error.
Fixes: a588e4580a7e ("bnxt_en: Add interface to support RDMA driver.") Reviewed-by: Michael Chan <michael.chan@broadcom.com> Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com> Link: https://patch.msgid.link/20260504083611.1383776-4-pavan.chebbi@broadcom.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Michael Chan [Mon, 4 May 2026 08:36:09 +0000 (14:06 +0530)]
bnxt_en: Set bp->max_tpa according to what the FW supports
Fix the logic to set bp->max_tpa no higher than what the FW supports.
On P5 chips, some older FW sets max_tpa very low so we override it to
prevent performance regressions with the older FW.
Fixes: 79632e9ba386 ("bnxt_en: Expand bnxt_tpa_info struct to support 57500 chips.") Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Reviewed-by: Colin Winegarden <colin.winegarden@broadcom.com> Reviewed-by: Rukhsana Ansari <rukhsana.ansari@broadcom.com> Signed-off-by: Michael Chan <michael.chan@broadcom.com> Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com> Link: https://patch.msgid.link/20260504083611.1383776-3-pavan.chebbi@broadcom.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Michael Chan [Mon, 4 May 2026 08:36:08 +0000 (14:06 +0530)]
bnxt_en: Delay for 5 seconds after AER DPC for all chips
The FW on all chips is requiring a 5-second delay after Downstream
Port Containment (DPC) AER. The previously added 900 msec delay was
not long enough in all cases because the chip's CRS (Configuration
Request Retry Status) mechanism is not always reliable.
Fixes: d5ab32e9b02d ("bnxt_en: Add delay to handle Downstream Port Containment (DPC) AER") Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by: Michael Chan <michael.chan@broadcom.com> Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com> Link: https://patch.msgid.link/20260504083611.1383776-2-pavan.chebbi@broadcom.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Alyssa Ross [Sun, 3 May 2026 19:25:16 +0000 (21:25 +0200)]
ipv6: default IPV6_SIT to m
This basically defaulted to m until recently, since IPV6 defaulted to
m. Since IPV6 was changed to a boolean with a default of y, IPV6_SIT
started defaulting to built-in as well. This results in a surprise
sit0 device by default for defconfig (and defconfig-derived config)
users at boot. For me, this broke an (admittedly non-robust) script.
Preserve the behaviour of most configs by avoiding building this
module, that's probably overall seldom used compared to IPv6 as a
whole, into the kernel.
Fixes: 309b905deee59 ("ipv6: convert CONFIG_IPV6 to built-in only and clean up Kconfigs") Signed-off-by: Alyssa Ross <hi@alyssa.is> Reviewed-by: Fernando Fernandez Mancera <fmancera@suse.de> Link: https://patch.msgid.link/20260503192515.290900-2-hi@alyssa.is Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Dipayaan Roy [Fri, 1 May 2026 02:47:12 +0000 (19:47 -0700)]
net: mana: Fix crash from unvalidated SHM offset read from BAR0 during FLR
During Function Level Reset recovery, the MANA driver reads
hardware BAR0 registers that may temporarily contain garbage values.
The SHM (Shared Memory) offset read from GDMA_REG_SHM_OFFSET is used
to compute gc->shm_base, which is later dereferenced via readl() in
mana_smc_poll_register(). If the hardware returns an unaligned or
out-of-range value, the driver must not blindly use it, as this would
propagate the hardware error into a kernel crash.
The following crash was observed on an arm64 Hyper-V guest running
kernel 6.17.0-3013-azure during VF reset recovery triggered by HWC
timeout.
From the crash signature x20 = ffff8000a200001b, this address
ends in 0x1b which is not 4-byte aligned, so the 'ldr w1, [x20]'
instruction (readl) triggers the arm64 alignment fault (FSC = 0x21).
The root cause is in mana_gd_init_vf_regs(), which computes:
The offset is used without any validation. The same problem exists
in mana_gd_init_pf_regs() for sriov_base_off and sriov_shm_off.
Fix this by validating all offsets before use:
- VF: check shm_off is within BAR0, properly aligned to 4 bytes
(readl requirement), and leaves room for the full 256-bit
(32-byte) SMC aperture.
- PF: check sriov_base_off is within BAR0, aligned to 8 bytes
(readq requirement), and leaves room to safely read the
sriov_shm_off register at sriov_base_off + GDMA_PF_REG_SHM_OFF.
Then check sriov_shm_off leaves room for the full SMC aperture.
All arithmetic uses subtraction rather than addition to avoid
integer overflow on garbage values.
Define SMC_APERTURE_SIZE (32 bytes, derived from the 256-bit aperture
width)
Return -EPROTO on invalid values. The existing recovery path in
mana_serv_reset() already handles -EPROTO by falling through to PCI
device rescan, giving the hardware another chance to present valid
register values after reset.
Nan Li [Fri, 1 May 2026 01:08:44 +0000 (09:08 +0800)]
net/rds: handle zerocopy send cleanup before the message is queued
A zerocopy send can fail after user pages have been pinned but before
the message is attached to the sending socket.
The purge path currently infers zerocopy state from rm->m_rs, so an
unqueued message can be cleaned up as if it owned normal payload pages.
However, zerocopy ownership is really determined by the presence of
op_mmp_znotifier, regardless of whether the message has reached the
socket queue.
Capture op_mmp_znotifier up front in rds_message_purge() and use it as
the cleanup discriminator. If the message is already associated with a
socket, keep the existing completion path. Otherwise, drop the pinned
page accounting directly and release the notifier before putting the
payload pages.
This keeps early send failure cleanup consistent with the zerocopy
lifetime rules without changing the normal queued completion path.
Fixes: 0cebaccef3ac ("rds: zerocopy Tx support.") Cc: stable@kernel.org Reported-by: Yuan Tan <yuantan098@gmail.com> Reported-by: Yifan Wu <yifanwucs@gmail.com> Reported-by: Juefei Pu <tomapufckgml@gmail.com> Reported-by: Xin Liu <bird@lzu.edu.cn> Co-developed-by: Xiao Liu <lx24@stu.ynu.edu.cn> Signed-off-by: Xiao Liu <lx24@stu.ynu.edu.cn> Signed-off-by: Nan Li <tonanli66@gmail.com> Signed-off-by: Ren Wei <n05ec@lzu.edu.cn> Reviewed-by: Allison Henderson <achender@kernel.org> Link: https://patch.msgid.link/d2ea98a6313d5467bac00f7c9fef8c7acddb9258.1777550074.git.tonanli66@gmail.com Signed-off-by: Paolo Abeni <pabeni@redhat.com>
====================
openvswitch: fix self-deadlock on release of tunnel vports
Two patches - the fix for the actual bug and the selftest that reproduces it.
I missed the self-deadlock in the original patch that introduced the issue,
because testing required code modification in the ovs-vswitchd to force it to
use legacy tunnel ports. I thought I made the change correctly, but apparently
something went wrong and the tests were run with the standard LWT infra instead.
The selftest added in this patch set will at least prevent this kind of mistakes
in the future.
I mentioned, however, that these tunnel vports are legacy and not actually used
by ovs-vswitchd. RTM_NEWLINK + COLLECT_METADATA is used in conjunction with the
standard OVS_VPORT_TYPE_NETDEV instead since 2017. The code to use the legacy
tunnels still exists in ovs-vswitchd however, but only as a fallback for older
kernels and we're planning to remove it in the next release. I'll be sending an
RFC to remove support for these legacy tunnel types from the kernel, as they
serve no real purpose today and only increase the uAPI surface for CVEs, but
we need to fix the known bugs for stable versions.
selftests: openvswitch: add tests for tunnel vport refcounting
There were a few issues found with the tunnel vport types around the
vport destruction code. Add some basic tests, so at least we know that
they can be properly added and removed without obvious issues.
The test creates OVS datapath, adds a non-LWT tunnel port, makes sure
they are created, and then removes the datapath and waits for all the
ports to be gone.
The dpctl script had a few bugs in the none-lwt tunnel creation code,
so fixing them as well to make the testing possible:
- The type of the --lwt option changed in order to properly disable it.
- Removed byte order conversion for the port numbers, as the value
supposed to be in the host order.
- Added missing 'gre' choice for the tunnel type.
openvswitch: vport: fix self-deadlock on release of tunnel ports
vports are used concurrently and protected by RCU, so netdev_put()
must happen after the RCU grace period. So, either in an RCU call or
after the synchronize_net(). The rtnl_delete_link() must happen under
RTNL and so can't be executed in RCU context. Calling synchronize_net()
while holding RTNL is not a good idea for performance and system
stability under load in general, so calling netdev_put() in RCU call
is the right solution here.
However,
when the device is deleted, rtnl_unlock() will call netdev_run_todo()
and block until all the references are gone. In the current code this
means that we never reach the call_rcu() and the vport is never freed
and the reference is never released, causing a self-deadlock on device
removal.
Fix that by moving the rcu_call() before the rtnl_unlock(), so the
scheduled RCU callback will be executed when synchronize_net() is
called from the rtnl_unlock()->netdev_run_todo() while the RTNL itself
is already released.
openvswitch: vport: fix race between tunnel creation and linking
When a tunnel vport is created it first creates the tunnel device, e.g.,
with geneve_dev_create_fb(), then it calls ovs_netdev_link() to take a
reference and link it to the device that represents openvswitch datapath.
The creation of the device is happening under RTNL, but then RTNL is
released and re-acquired to find the device by name. It is technically
possible for the tunnel device to be re-named or deleted within that
window while RTNL is not held, and some other device created in its
place. This will cause a non-tunnel device to be referenced in the
vport and tunnel-specific functions used on it, e.g. vxlan_get_options()
that directly casts the private netdev data into a struct vxlan_dev
causing an invalid memory access:
BUG: KASAN: slab-use-after-free in vxlan_get_options+0x323/0x3a0
vxlan_get_options+0x323/0x3a0
ovs_vport_cmd_new+0x6e3/0xd30
Fix that by taking a reference to the just created device before
releasing RTNL. This ensures that the device in the vport is always
the one that was just created. The search by name is only needed
for a standard vport-netdev that links pre-existing devices, so that
functionality and device type checks are moved to netdev_create().
It is also awkward that ovs_netdev_link() takes ownership of the vport
and destroys it on failure. It doesn't know the type of the port it is
dealing with, so we need to pass down the indicator that it's a tunnel,
so the link can be properly deleted on failure.
It's possible to refactor the logic to make the ovs_netdev_link() do
only the linking part and let the callers perform a proper destruction,
but it will be much more code for each legacy tunnel port type, so it
is not worth it for the bug fix.
Fixes: 614732eaa12d ("openvswitch: Use regular VXLAN net_device device") Reported-by: Yuan Tan <tanyuan98@outlook.com> Reported-by: Yifan Wu <yifanwucs@gmail.com> Reported-by: Juefei Pu <tomapufckgml@gmail.com> Reported-by: Xin Liu <bird@lzu.edu.cn> Reported-by: Yang Yang <n05ec@lzu.edu.cn> Signed-off-by: Ilya Maximets <i.maximets@ovn.org> Acked-by: Eelco Chaudron <echaudro@redhat.com> Link: https://patch.msgid.link/20260430213349.407991-1-i.maximets@ovn.org Signed-off-by: Paolo Abeni <pabeni@redhat.com>
====================
net: mana: Fix mana_destroy_rxq() cleanup for partial RXQ init
When mana_create_rxq() fails partway through initialization (e.g. the
hardware rejects the WQ object creation), the error path calls
mana_destroy_rxq() to tear down a partially-initialized RXQ.
This exposed multiple issues in mana_destroy_rxq() path, as it assumed
the RXQ was always fully initialized, leading to multiple issues:
1. xdp_rxq_info_unreg() was called on an unregistered xdp_rxq,
triggering a WARN_ON ("Driver BUG") in net/core/xdp.c.
2. mana_destroy_wq_obj() was called with INVALID_MANA_HANDLE,
sending a bogus destroy command to the hardware.
3. mana_deinit_cq() was called twice — once inside mana_destroy_rxq()
and again in mana_create_rxq()'s error path — causing a
use-after-free since mana_destroy_rxq() frees the rxq first.
This was observed during ethtool ring parameter changes when the
hardware returned an error creating the RXQ. This series makes
mana_destroy_rxq() safe to call at any stage of RXQ initialization
by guarding each teardown step, and removes the redundant cleanup
in mana_create_rxq().
====================
Dipayaan Roy [Thu, 30 Apr 2026 03:57:54 +0000 (20:57 -0700)]
net: mana: remove double CQ cleanup in mana_create_rxq error path
In mana_create_rxq(), the error cleanup path calls mana_destroy_rxq()
followed by mana_deinit_cq(). This is incorrect for two reasons:
1. mana_destroy_rxq() already calls mana_deinit_cq() internally,
so the CQ's GDMA queue is destroyed twice.
2. mana_destroy_rxq() frees the rxq via kfree(rxq) before returning.
The subsequent mana_deinit_cq(apc, cq) then operates on freed memory
since cq points to &rxq->rx_cq, which is embedded in the
already-freed rxq structure — a use-after-free.
Remove the redundant mana_deinit_cq() call from the error path since
mana_destroy_rxq() already handles CQ cleanup. mana_deinit_cq() is
itself safe for an uninitialized CQ as it checks for a NULL gdma_cq
before proceeding.
Fixes: ca9c54d2d6a5 ("net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)") Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com> Signed-off-by: Dipayaan Roy <dipayanroy@linux.microsoft.com> Reviewed-by: Aditya Garg <gargaditya@linux.microsoft.com> Link: https://patch.msgid.link/20260430035935.1859220-4-dipayanroy@linux.microsoft.com Reviewed-by: Simon Horman <horms@kernel.org> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Dipayaan Roy [Thu, 30 Apr 2026 03:57:53 +0000 (20:57 -0700)]
net: mana: Skip WQ object destruction for uninitialized RXQ
In mana_destroy_rxq(), mana_destroy_wq_obj() is called unconditionally
even when the WQ object was never created (rxobj is still
INVALID_MANA_HANDLE). When mana_create_rxq() fails before
mana_create_wq_obj() succeeds, the error path calls mana_destroy_rxq()
which sends a bogus destroy command to the hardware:
mana 7870:00:00.0: HWC: Failed hw_channel req: 0x1d
mana 7870:00:00.0: Failed to send mana message: -71, 0x1d
mana 7870:00:00.0 eth7: Failed to destroy WQ object: -71
Guard mana_destroy_wq_obj() with an INVALID_MANA_HANDLE check so that
mana_destroy_rxq() is safe to call at any stage of RXQ initialization.
Fixes: ca9c54d2d6a5 ("net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)") Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com> Signed-off-by: Dipayaan Roy <dipayanroy@linux.microsoft.com> Link: https://patch.msgid.link/20260430035935.1859220-3-dipayanroy@linux.microsoft.com Reviewed-by: Simon Horman <horms@kernel.org> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Dipayaan Roy [Thu, 30 Apr 2026 03:57:52 +0000 (20:57 -0700)]
net: mana: check xdp_rxq registration before unreg in mana_destroy_rxq()
When mana_create_rxq() fails at mana_create_wq_obj() or any step before
xdp_rxq_info_reg() is called, the error path jumps to `out:` which calls
mana_destroy_rxq(). mana_destroy_rxq() unconditionally calls
xdp_rxq_info_unreg() on xilinx xdp_rxq that was never registered,
triggering a WARN_ON in net/core/xdp.c:
Guard the xdp_rxq_info_unreg() call with xdp_rxq_info_is_reg() so that
mana_destroy_rxq() is safe to call regardless of how far initialization
progressed.
Fixes: ed5356b53f07 ("net: mana: Add XDP support") Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com> Signed-off-by: Dipayaan Roy <dipayanroy@linux.microsoft.com> Link: https://patch.msgid.link/20260430035935.1859220-2-dipayanroy@linux.microsoft.com Reviewed-by: Simon Horman <horms@kernel.org> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Daniel Golle [Sat, 2 May 2026 10:55:02 +0000 (11:55 +0100)]
net: dsa: mt7530: fix .get_stats64 sleeping in atomic context
The .get_stats64 callback runs in atomic context, but on
MDIO-connected switches every register read acquires the MDIO bus
mutex, which can sleep:
[ 12.645973] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:609
[ 12.654442] in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 759, name: grep
[ 12.663377] preempt_count: 0, expected: 0
[ 12.667410] RCU nest depth: 1, expected: 0
[ 12.671511] INFO: lockdep is turned off.
[ 12.675441] CPU: 0 UID: 0 PID: 759 Comm: grep Tainted: G S W 7.0.0+ #0 PREEMPT
[ 12.675453] Tainted: [S]=CPU_OUT_OF_SPEC, [W]=WARN
[ 12.675456] Hardware name: Bananapi BPI-R64 (DT)
[ 12.675459] Call trace:
[ 12.675462] show_stack+0x14/0x1c (C)
[ 12.675477] dump_stack_lvl+0x68/0x8c
[ 12.675487] dump_stack+0x14/0x1c
[ 12.675495] __might_resched+0x14c/0x220
[ 12.675504] __might_sleep+0x44/0x80
[ 12.675511] __mutex_lock+0x50/0xb10
[ 12.675523] mutex_lock_nested+0x20/0x30
[ 12.675532] mt7530_get_stats64+0x40/0x2ac
[ 12.675542] dsa_user_get_stats64+0x2c/0x40
[ 12.675553] dev_get_stats+0x44/0x1e0
[ 12.675564] dev_seq_printf_stats+0x24/0xe0
[ 12.675575] dev_seq_show+0x14/0x3c
[ 12.675583] seq_read_iter+0x37c/0x480
[ 12.675595] seq_read+0xd0/0xec
[ 12.675605] proc_reg_read+0x94/0xe4
[ 12.675615] vfs_read+0x98/0x29c
[ 12.675625] ksys_read+0x54/0xdc
[ 12.675633] __arm64_sys_read+0x18/0x20
[ 12.675642] invoke_syscall.constprop.0+0x54/0xec
[ 12.675653] do_el0_svc+0x3c/0xb4
[ 12.675662] el0_svc+0x38/0x200
[ 12.675670] el0t_64_sync_handler+0x98/0xdc
[ 12.675679] el0t_64_sync+0x158/0x15c
For MDIO-connected switches, poll MIB counters asynchronously using a
delayed workqueue every second and let .get_stats64 return the cached
values under a spinlock. A mod_delayed_work() call on each read
triggers an immediate refresh so counters stay responsive when queried
more frequently.
MMIO-connected switches (MT7988, EN7581, AN7583) are not affected
because their regmap does not sleep, so they continue to read MIB
counters directly in .get_stats64.
David Carlier [Sat, 2 May 2026 14:19:45 +0000 (15:19 +0100)]
psp: strip variable-length PSP header in psp_dev_rcv()
psp_dev_rcv() unconditionally removes a fixed PSP_ENCAP_HLEN, even
when psph->hdrlen indicates that the PSP header carries optional
fields. A frame whose PSP header advertises a non-zero VC or any
extension would therefore be silently mis-decapsulated: option bytes
would spill into the inner packet head and downstream parsing would
fail on a corrupted skb.
Compute the full PSP header length from psph->hdrlen, pull the
optional bytes into the linear region, and strip the whole header
when decapsulating. Optional fields (VC, ...) are still ignored,
just discarded with the rest of the header instead of leaking.
crypt_offset and the VIRT flag are intentionally not validated here
- callers know their device's PSP implementation and can decide.
Both in-tree callers gate on hardware-validated PSP, so this is a
correctness fix rather than a reachable corruption path under
current configurations.
Fixes: 0eddb8023cee ("psp: provide decapsulation and receive helper for drivers") Reviewed-by: Willem de Bruijn <willemb@google.com> Reviewed-by: Daniel Zahka <daniel.zahka@gmail.com> Cc: stable@vger.kernel.org Signed-off-by: David Carlier <devnexen@gmail.com> Link: https://patch.msgid.link/20260502141945.14484-1-devnexen@gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
mptcp: sockopt: increase seq in mptcp_setsockopt_all_sf
mptcp_setsockopt_all_sf() was missing a call to sockopt_seq_inc(). This
is required not to cause missing synchronization for newer subflows
created later on.
This helper is called each time a socket option is set on subflows, and
future ones will need to inherit this option after their creation.
Fixes: 51c5fd09e1b4 ("mptcp: add TCP_MAXSEG sockopt support") Cc: stable@vger.kernel.org Suggested-by: Paolo Abeni <pabeni@redhat.com> Reviewed-by: Mat Martineau <martineau@kernel.org> Signed-off-by: Matthieu Baerts (NGI0) <matttbe@kernel.org> Link: https://patch.msgid.link/20260501-net-mptcp-misc-fixes-7-1-rc3-v1-4-b70118df778e@kernel.org Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Shardul Bankar [Fri, 1 May 2026 19:35:35 +0000 (21:35 +0200)]
mptcp: use MPTCP_RST_EMPTCP for ACK HMAC validation failure
When HMAC validation fails on a received ACK + MP_JOIN in
subflow_syn_recv_sock(), the subflow is reset with reason
MPTCP_RST_EPROHIBIT ("Administratively prohibited"). This is
incorrect: HMAC validation failure is an MPTCP protocol-level
error, not an administrative policy denial.
The mirror site on the client, in subflow_finish_connect(), already
uses MPTCP_RST_EMPTCP ("MPTCP-specific error") for the same kind of
HMAC failure on the SYN/ACK + MP_JOIN. Use the same reason on the
server side for symmetry and accuracy.
Shardul Bankar [Fri, 1 May 2026 19:35:34 +0000 (21:35 +0200)]
mptcp: use MPJoinSynAckHMacFailure for SynAck HMAC failure
In subflow_finish_connect(), HMAC validation of the server's HMAC
in SYN/ACK + MP_JOIN increments MPTCP_MIB_JOINACKMAC ("HMAC was
wrong on ACK + MP_JOIN") on failure. The function processes the
SYN/ACK, not the ACK; the matching MPTCP_MIB_JOINSYNACKMAC counter
("HMAC was wrong on SYN/ACK + MP_JOIN") exists but is not
incremented anywhere in the tree.
The mirror site on the server, subflow_syn_recv_sock(), already
uses JOINACKMAC correctly for ACK HMAC failure. Use JOINSYNACKMAC
at the SYN/ACK validation site so each counter reflects the packet
whose HMAC actually failed.
Markus Baier [Fri, 1 May 2026 16:39:41 +0000 (18:39 +0200)]
net: usb: asix: ax88772: re-add usbnet_link_change() in phylink callbacks
Commit e0bffe3e6894 ("net: asix: ax88772: migrate to phylink") replaced
the asix_adjust_link() PHY callback with phylink's mac_link_up() and
mac_link_down() handlers, but did not carry over the usbnet_link_change()
notification that commit 805206e66fab ("net: asix: fix "can't send until
first packet is send" issue") had added.
As a result, the original symptom returns: when the link comes up,
usbnet is never notified, so the RX URB submission stays dormant until
some other event (e.g. a transmitted packet triggering the status
endpoint interrupt) wakes it up.
This is reproducible with the Apple A1277 USB Ethernet Adapter
(05ac:1402, AX88772A based) on a Banana Pro using a static IPv4
configuration. After bringing the interface up, no incoming packets are
received until the first outgoing frame triggers usbnet's RX path.
Restore the link change notification, gated on a carrier transition so
the call remains idempotent if the status endpoint also reports the
change later.
Breno Leitao [Fri, 1 May 2026 09:58:41 +0000 (02:58 -0700)]
netpoll: pass buffer size to egress_dev() to avoid MAC truncation
egress_dev() formats np->dev_mac via snprintf() but receives buf as
a bare char *, so it cannot derive the buffer size from the pointer. The
size argument was hardcoded to MAC_ADDR_STR_LEN (3 * ETH_ALEN - 1 = 17),
which is silly wrong in two ways:
1) misleading kernel log output on the MAC-selected target path
(np->dev_name[0] == '\0'); for example "aa:bb:cc:dd:ee:ff doesn't
exist, aborting" was logged as "aa:bb:cc:dd:ee:f doesn't exist,
aborting".
2) the second argument of snprintf is the size of the buffer, not the
size of what you want to write.
Add a bufsz parameter to egress_dev() and pass sizeof(buf) from each
caller, matching the standard snprintf() idiom and removing the
hardcoded size from the helper.
Every caller already declares "char buf[MAC_ADDR_STR_LEN + 1]" so the
formatted MAC continues to fit.
Tested by booting with
netconsole=6665@/aa:bb:cc:dd:ee:ff,6666@10.0.0.1/00:11:22:33:44:55
on a kernel without a matching device. Pre-fix dmesg shows
"aa:bb:cc:dd:ee:f doesn't exist, aborting"; post-fix shows the full
"aa:bb:cc:dd:ee:ff doesn't exist, aborting".
unix_peek_fpl() relies on gc_in_progress not to confuse GC
by MSG_PEEK.
Let's set gc_in_progress to true in unix_gc().
Fixes: 8b90a9f819dc ("af_unix: Run GC on only one CPU.") Reported-by: Igor Ushakov <sysroot314@gmail.com> Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com> Link: https://patch.msgid.link/20260501073945.1884564-1-kuniyu@google.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Waiman Long [Thu, 30 Apr 2026 07:44:20 +0000 (10:44 +0300)]
sched/isolation: Make HK_TYPE_KTHREAD an alias of HK_TYPE_DOMAIN
Since commit 041ee6f3727a ("kthread: Rely on HK_TYPE_DOMAIN for preferred
affinity management"), kthreads default to use the HK_TYPE_DOMAIN
cpumask. IOW, it is no longer affected by the setting of the nohz_full
boot kernel parameter.
That means HK_TYPE_KTHREAD should now be an alias of HK_TYPE_DOMAIN
instead of HK_TYPE_KERNEL_NOISE to correctly reflect the current kthread
behavior. Make the change as HK_TYPE_KTHREAD is still being used in
some networking code.
Fixes: 041ee6f3727a ("kthread: Rely on HK_TYPE_DOMAIN for preferred affinity management") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Julian Anastasov <ja@ssi.bg> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Waiman Long [Thu, 30 Apr 2026 07:44:19 +0000 (10:44 +0300)]
ipvs: Guard access of HK_TYPE_KTHREAD cpumask with RCU
The ip_vs_ctl.c file and the associated ip_vs.h file are the only places
in the kernel where HK_TYPE_KTHREAD cpumask is being retrieved and used.
Now that HK_TYPE_KTHREAD/HK_TYPE_DOMAIN cpumask can be changed at run
time. We need to use RCU to guard access to this cpumask to avoid a
potential UAF problem as the returned cpumask may be freed before it
is being used.
We can replace HK_TYPE_KTHREAD by HK_TYPE_DOMAIN as they are aliases
of each other, but keeping the HK_TYPE_KTHREAD name can highlight the
fact that it is the kthread initiated by ipvs that is being controlled.
Fixes: 03ff73510169 ("cpuset: Update HK_TYPE_DOMAIN cpumask from cpuset") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Julian Anastasov <ja@ssi.bg> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Sashiko reports for races and possible crash around
the usage of est_cpulist_valid and sysctl_est_cpulist.
The problem is that we do not lock est_mutex in some
places which can lead to wrong write ordering and
as result problems when calling cpumask_weight()
and cpumask_empty().
Fix them by moving the est_max_threads read/write under
locked est_mutex. Do the same for one ip_vs_est_reload_start()
call to protect the cpumask_empty() usage of sysctl_est_cpulist.
To remove the chance of deadlock while stopping the
estimation kthreads, keep the data structure for kthread 0
even after last estimator is removed and do not hold mutexes
while stopping this task. Now we will use a new flag 'needed'
to know when kthread 0 should run. The kthreads above 0
do not use mutexes, so stop them under est_mutex because
their kthread data still can be destroyed if they do not
serve estimators. Now all kthreads will be started by
the est_reload_work to properly serialize the stop/start
for kthread 0.
Reduce the use of service_mutex in ip_vs_est_calc_phase()
because under est_mutex we can safely walk est_kt_arr to
stop the kthreads above slot 0.
As ip_vs_stop_estimator() for tot_stats should be called
under service_mutex, do it early in the netns exit path
in ip_vs_flush() to avoid locking the mutex again later.
It still should be called in ip_vs_control_net_cleanup_sysctl()
when we are called during netns init error. Use -2 for ktid
as indicator if estimator was already stopped.
Finally, fix use-after-free for kd->est_row in
ip_vs_est_calc_phase(). est->ktrow should simply switch to
a delay value while estimator is linked to est_temp_list.
syzbot reports for sleeping function called from invalid context [1].
The recently added code for resizable hash tables uses
hlist_bl bit locks in combination with spin_lock for
the connection fields (cp->lock).
Fix the following problems:
* avoid using spin_lock(&cp->lock) under locked bit lock
because it sleeps on PREEMPT_RT
* as the recent changes call ip_vs_conn_hash() only for newly
allocated connection, the spin_lock can be removed there because
the connection is still not linked to table and does not need
cp->lock protection.
* the lock can be removed also from ip_vs_conn_unlink() where we
are the last connection user.
* the last place that is fixed is ip_vs_conn_fill_cport()
where now the cp->lock is locked before the other locks to
ensure other packets do not modify the cp->flags in non-atomic
way. Here we make sure cport and flags are changed only once
if two or more packets race to fill the cport. Also, we fill
cport early, so that if we race with resizing there will be
valid cport key for the hashing. Add a warning if too many
hash table changes occur for our RCU read-side critical
section which is error condition but minor because the
connection still can expire gracefully. Still, restore the
cport to 0 to allow retransmitted packet to properly fill
the cport. Problems reported by Sashiko.
ipvs: fix races around the conn_lfactor and svc_lfactor sysctl vars
Sashiko warns that the new sysctls vars can be changed
after the hash tables are destroyed and their respective
resizing works canceled, leading to mod_delayed_work()
being called for canceled works.
Solve this in different ways. conn_tab can be present even
without services and is destroyed only on netns exit, so use
disable_delayed_work_sync() to disable the work instead of
adding more synchronization mechanisms.
As for the svc_table, it is destroyed when the services
are deleted, so we must be sure that netns exit is not
called yet (the check for 'enable') and the work is
not canceled by checking all under same mutex lock.
Also, use WRITE_ONCE when updating the sysctl vars as we
already read them with READ_ONCE.
Sashiko reports some problems for the recently added
/proc/net/ip_vs_status:
* ip_vs_status_show() as a table reader may run long after the
conn_tab and svc_table table are released. While ip_vs_conn_flush()
properly changes the conn_tab_changes counter when conn_tab is removed,
ip_vs_del_service() and ip_vs_flush() were missing such change for
the svc_table_changes counter. As result, readers like
ip_vs_dst_event() and ip_vs_status_show() may continue to use
a freed table after a cond_resched_rcu() call.
* While counting the buckets in ip_vs_status_show() make sure we
traverse only the needed number of entries in the chain. This also
prevents possible overflow of the 'count' variable.
* Add check for 'loops' to prevent infinite loops while restarting
the traversal on table change.
* While IP_VS_CONN_TAB_MAX_BITS is 20 on 32-bit platforms and
there is no risk to overflow when multiplying the number of
conn_tab buckets to 100, prefer the div_u64() helper to make
the following dividing safer.
* Use 0440 permissions for ip_vs_status to restrict the
info only to root due to the exported information for hash
distribution.
Jakub Kicinski [Wed, 29 Apr 2026 22:29:38 +0000 (15:29 -0700)]
net: tls: fix silent data drop under pipe back-pressure
tls_sw_splice_read() uses len when advancing rxm->offset / rxm->full_len
after skb_splice_bits(), rather than copied (the actual number of bytes
successfully spliced into the pipe). When the destination pipe cannot
accept all the requested bytes, splice_to_pipe() returns fewer bytes
than len, and 'len - copied' of data is effectively skipped over.
ip6_gre: Use cached t->net in ip6erspan_changelink().
After commit 5e72ce3e3980 ("net: ipv6: Use link netns in newlink() of
rtnl_link_ops"), ip6erspan_newlink() correctly resolves the per-netns
ip6gre hash via link_net. ip6erspan_changelink() was not converted in
that series and still uses dev_net(dev), which diverges from the
device's creation netns after IFLA_NET_NS_FD migration.
This re-inserts the tunnel into the wrong per-netns hash. The
original netns keeps a stale entry. When that netns is later
destroyed, ip6gre_exit_rtnl_net() walks the stale entry, producing a
slab-use-after-free reported by KASAN, followed by a kernel BUG at
net/core/dev.c (LIST_POISON1) in unregister_netdevice_many_notify().
Reachable from an unprivileged user namespace (unshare --user
--map-root-user --net).
ip6gre_changelink() earlier in the same file already uses the cached
t->net; only ip6erspan_changelink() has the wrong shape.
====================
Replace direct dequeue call with qdisc_dequeue_peeked
When sfb and red qdiscs have children (eg qfq qdisc) whose peek() callback is
qdisc_peek_dequeued(), we could get a kernel panic. When the parent of such
qdiscs (eg illustrated in patch #3 as tbf) wants to retrieve an skb from
its child (red/sfb in this case), it will do the following:
1a. do a peek() - and when sensing there's an skb the child can offer, then
- the child in this case(red/sfb) calls its child's (qfq) peek.
qfq does the right thing and will return the gso_skb queue packet.
Note: if there wasnt a gso_skb entry then qfq will store it there.
1b. invoke a dequeue() on the child (red/sfb). And herein lies the problem.
- red/sfb will call the child's dequeue() which will essentially just
try to grab something of qfq's queue.
The right thing to do in #1b is to grab the skb off gso_skb queue.
This patchset fixes that issue by changing #1b to use qdisc_dequeue_peeked()
method instead.
Patch 1 fixes the issue for red qdisc. Patch 2 fixes it for sfb.
Patch 3 adds testcases for the two setups.
====================
Victor Nogueira [Thu, 30 Apr 2026 15:29:57 +0000 (11:29 -0400)]
selftests/tc-testing: Add tests that force red and sfb to dequeue from child's gso_skb
Create 4 test cases:
- Force red to dequeue from its child's gso_skb with qfq leaf
- Force sfb to dequeue from its child's gso_skb with qfq leaf
- Force red to dequeue from its child's gso_skb with dualpi2 leaf
- Force sfb to dequeue from its child's gso_skb with dualpi2 leaf
All of them have tbf followed by red (or sfb) followed by qfq (or
dualpi2). Since tbf calls its child's peek followed by
qdisc_dequeue_peeked, it will force red/sfb to call their child's peek.
In this case, since the child (qfq/dualpi2) has qdisc_peek_dequeued as
its peek callback, the packet will be stored in its gso_skb queue. During
the subsequent call to qdisc_dequeue_peeked, red/sfb will have to dequeue
from the child's gso_skb to retrieve the packet.
Not doing so will cause a NULL ptr deref which was happening before a
recent fix.
Victor Nogueria [Thu, 30 Apr 2026 15:29:56 +0000 (11:29 -0400)]
net/sched: sch_sfb: Replace direct dequeue call with peek and qdisc_dequeue_peeked
When sfb has children (eg qfq qdisc) whose peek() callback is
qdisc_peek_dequeued(), we could get a kernel panic. When the parent of such
qdiscs (eg illustrated in patch #3 as tbf) wants to retrieve an skb from
its child (sfb in this case), it will do the following:
1a. do a peek() - and when sensing there's an skb the child can offer, then
- the child in this case(sfb) calls its child's (qfq) peek.
qfq does the right thing and will return the gso_skb queue packet.
Note: if there wasnt a gso_skb entry then qfq will store it there.
1b. invoke a dequeue() on the child (sfb). And herein lies the problem.
- sfb will call the child's dequeue() which will essentially just
try to grab something of qfq's queue.
The right thing to do in #1b is to grab the skb off gso_skb queue.
This patchset fixes that issue by changing #1b to use qdisc_dequeue_peeked()
method instead.
Fixes: e13e02a3c68d ("net_sched: SFB flow scheduler") Signed-off-by: Victor Nogueria <victor@mojatatu.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Link: https://patch.msgid.link/20260430152957.194015-3-jhs@mojatatu.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net/sched: sch_red: Replace direct dequeue call with peek and qdisc_dequeue_peeked
When red qdisc has children (eg qfq qdisc) whose peek() callback is
qdisc_peek_dequeued(), we could get a kernel panic. When the parent of such
qdiscs (eg illustrated in patch #3 as tbf) wants to retrieve an skb from
its child (red in this case), it will do the following:
1a. do a peek() - and when sensing there's an skb the child can offer, then
- the child in this case(red) calls its child's (qfq) peek.
qfq does the right thing and will return the gso_skb queue packet.
Note: if there wasnt a gso_skb entry then qfq will store it there.
1b. invoke a dequeue() on the child (red). And herein lies the problem.
- red will call the child's dequeue() which will essentially just
try to grab something of qfq's queue.
The right thing to do in #1b is to grab the skb off gso_skb queue.
This patchset fixes that issue by changing #1b to use qdisc_dequeue_peeked()
method instead.
Fixes: 77be155cba4e ("pkt_sched: Add peek emulation for non-work-conserving qdiscs.") Reported-by: Manas <ghandatmanas@gmail.com> Reported-by: Rakshit Awasthi <rakshitawasthi17@gmail.com> Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Link: https://patch.msgid.link/20260430152957.194015-2-jhs@mojatatu.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
XGBE_PTP_ACT_CLK_FREQ and XGBE_V2_PTP_ACT_CLK_FREQ were 10x too
large (500MHz/1GHz instead of 50MHz/100MHz), causing the computed
addend to overflow the 32-bit tstamp_addend. In the general case
this would result in the clock advancing at the wrong rate. For v2
(PCI), ptpclk_rate is hardcoded to 125MHz, so the addend formula
(ACT_CLK_FREQ << 32) / ptpclk_rate yields exactly 8 * 2^32, and
when stored to the 32-bit tstamp_addend the value is zero. With
addend = 0 the hardware accumulator never overflows and the PTP
clock is fully stopped. For v1 (platform), ptpclk_rate is read from
ACPI/DT so the exact overflow behavior depends on the
firmware-reported frequency.
Define the constants as NSEC_PER_SEC / SSINC so the relationship is
explicit and cannot drift out of sync.
If the driver is used in a non tdm mode priv->utdm is a NULL pointer.
Therefore we need to check this pointer first before checking si_regs.
Fixes: c19b6d246a35 ("drivers/net: support hdlc function for QE-UCC") Signed-off-by: Holger Brunck <holger.brunck@hitachienergy.com> Reviewed-by: Simon Horman <horms@kernel.org> Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Unmapping of uf_regs is done from ucc_fast_free and doesn't need to be
done explicitly. If already unmapped ucc_fast_free will crash.
Fixes: c19b6d246a35 ("drivers/net: support hdlc function for QE-UCC") Signed-off-by: Holger Brunck <holger.brunck@hitachienergy.com> Reviewed-by: Simon Horman <horms@kernel.org> Signed-off-by: Jakub Kicinski <kuba@kernel.org>
====================
ipv6: fix ECMP route failover on carrier loss
This patchset resolves an issue where established IPv6 connections are
unable to transition to alternative ECMP nexthops upon carrier loss.
Unlike IPv4, the IPv6 routing subsystem does not actively invalidate
cached destinations during a NETDEV_CHANGE event. Sockets persist
with dead routes, leading to stalled traffic or connection drops.
This series introduces a fix to trigger route invalidation by
updating the route serial number on link carrier loss and provides
a corresponding selftest to validate the failover behavior for IPv4
and IPv6.
====================
When using IPv6 ECMP routes, if a netdev listed as a nexthop experiences
a carrier change event (e.g., a bond device generating a NETDEV_CHANGE
event after its slaves go linkdown), established connections utilizing
that nexthop fail to fail over to other available nexthops. Instead,
these connections stall or drop.
This happens because the IPv6 FIB code does not invalidate the socket's
cached destination when a NETDEV_CHANGE event occurs. While
fib6_ifdown() correctly marks the nexthop with RTNH_F_LINKDOWN, it
leaves the route's serial number unchanged. As a result, sockets with a
previously cached dst do not realize the route is no longer viable and
continue to try using the non-functional nexthop.
This behavior contrasts with IPv4, which actively flushes cached
destinations on a NETDEV_CHANGE event (see fib_netdev_event() in
net/ipv4/fib_frontend.c).
Fix this by updating the route serial number in fib6_ifdown() when
setting RTNH_F_LINKDOWN. This invalidates stale cached destinations,
forcing sockets to perform a new route lookup and fail over to a
functioning nexthop.
Fixes: 51ebd3181572 ("ipv6: add support of equal cost multipath (ECMP)") Signed-off-by: Sagarika Sharma <sharmasagarika@google.com> Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com> Reviewed-by: Ido Schimmel <idosch@nvidia.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Link: https://patch.msgid.link/20260430200909.527827-2-sharmasagarika@google.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Alex Cheema [Wed, 29 Apr 2026 17:57:39 +0000 (18:57 +0100)]
net: usb: cdc_ncm: add Apple Mac USB-C direct networking quirk
Apple Silicon Macs expose two CDC NCM "private" data interfaces over
USB-C with VID:PID 0x05ac:0x1905 and product string "Mac". This is the
same protocol Apple already ships on iPhone (0x05ac:0x12a8) and iPad
(0x05ac:0x12ab) for RemoteXPC since iOS 17 -- both data interfaces lack
an interrupt status endpoint, so they rely on the FLAG_LINK_INTR-
conditional bind path introduced in commit 3ec8d7572a69 ("CDC-NCM: add
support for Apple's private interface").
The id_table currently has entries for iPhone and iPad but not for the
Mac. Without a match, cdc_ncm falls through to the generic CDC NCM
class-match entry, which uses the FLAG_LINK_INTR-having cdc_ncm_info
struct, so bind_common() fails on the missing status endpoint and no
netdev appears.
Add id_table entries for both interface numbers (0 and 2) of the Mac,
bound to the existing apple_private_interface_info driver_info.
Verified empirically on a Mac Studio M3 Ultra running macOS 26.5: when
a Mac is connected via USB-C, ioreg shows VID 0x05ac, PID 0x1905,
product string "Mac", with two NCM data interfaces at numbers 0 and 2.
The same PID is presented by all current Apple Silicon Mac models
(MacBook Pro/Air, Mac mini, Mac Studio across the M-series), mirroring
Apple's single-PID-per-family pattern from iPhone/iPad.
After this patch, plugging a Mac into a Linux host running the patched
kernel produces two enx... interfaces (one per data interface),
"ip -br link" lists them as UP, and standard userspace networking
(DHCP, NetworkManager shared mode, etc.) works without any modprobe
overrides or out-of-tree modules.
On Ethernet devices (the overwhelming majority of SR-IOV NICs)
dev->addr_len is 6, so only the first 6 bytes of broadcast[] are
written. The remaining 26 bytes retain whatever was previously on
the kernel stack. The full struct is then handed to userspace via:
leaking up to 26 bytes of uninitialised kernel stack per VF per
RTM_GETLINK request, repeatable.
The other vf_* structs in the same function are explicitly zeroed
for exactly this reason - see the memset() calls for ivi,
vf_vlan_info, node_guid and port_guid a few lines above.
vf_broadcast was simply missed when it was added.
Reachability: any unprivileged local process can open AF_NETLINK /
NETLINK_ROUTE without capabilities and send RTM_GETLINK with an
IFLA_EXT_MASK attribute carrying RTEXT_FILTER_VF. The kernel walks
each VF and emits IFLA_VF_BROADCAST, leaking 26 bytes of stack per
VF per request. Stack residue at this call site can include return
addresses and transient sensitive data; KASAN with stack
instrumentation, or KMSAN, will flag the nla_put() when reproduced.
Zero the on-stack struct before the partial memcpy, matching the
existing pattern used for the other vf_* structs in the same
function.
Eric Dumazet [Thu, 30 Apr 2026 07:06:11 +0000 (07:06 +0000)]
ipmr: prevent info-leak in pmr_cache_report()
Yiming Qian reported:
<quote>
ipmr_cache_report()` allocates a report skb with `alloc_skb(128,
GFP_ATOMIC)` and appends a `struct igmphdr` using `skb_put()`. In the
non-`IGMPMSG_WHOLEPKT` path it initializes only:
- `igmp->type`
- `igmp->code`
but does not initialize:
- `igmp->csum`
- `igmp->group`
Later, `igmpmsg_netlink_event()` copies the bytes after `sizeof(struct
igmpmsg)` into the `IPMRA_CREPORT_PKT` netlink attribute and emits
`RTM_NEWCACHEREPORT` on `RTNLGRP_IPV4_MROUTE_R`.
As a result, 6 bytes of stale heap data from the skb head are
disclosed to userspace.
</quote>
Let's use skb_put_zero() instead of skb_put() to fix this bug.
The TRENDnet TUC-ET2G V2.0 is an RTL8156B based 2.5G Ethernet controller.
Add the vendor and product ID values to the driver. This makes Ethernet
work with the adapter.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl> Reviewed-by: Andrew Lunn <andrew@lunn.ch> Reviewed-by: Birger Koblitz <mail@birger-koblitz.de> Link: https://patch.msgid.link/20260430213435.21821-1-olek2@wp.pl Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jakub Kicinski [Fri, 1 May 2026 23:45:41 +0000 (16:45 -0700)]
Merge tag 'nf-26-05-01' of git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf
Pablo Neira Ayuso says:
====================
Netfilter fixes for net
The following batch contains Netfilter fixes for net:
1) Replace skb_try_make_writable() by skb_ensure_writable() in
nft_fwd_netdev and the flowtable to deal with uncloned packets
having their network header in paged fragments.
2) Drop packet if output device does not exist and ensure sufficient
headroom in nft_fwd_netdev before transmitting the skb.
3) Use the existing dup recursion counter in nft_fwd_netdev for the
neigh_xmit variant, from Weiming Shi.
4) Add .check_hooks interface to x_tables to detach the control plane
hook check based on the match/target configuration. Then, update
nft_compat to use .check_hooks from .validate path, this fixes a
lack of hook validation for several match/targets.
5) Fix incorrect .usersize in xt_CT, from Florian Westphal.
6) Fix a memleak with netdev tables in dormant state,
from Florian Westphal.
7) Several patches to check if the packet is a fragment, then skip
layer 4 inspection, for x_tables and nf_tables; as well as common
nf_socket infrastructure. The xt_hashlimit match drops fragments
to stay consistent with the existing approach when failing to parse
the layer 4 protocol header.
8) Ensure sufficient headroom in the flowtable before transmitting
the skb.
9) Fix the flowtable inline vlan approach for double-tagged vlan:
Reverse the iteration over .encap[] since it represents the
encapsulation as seen from the ingress path. Postpone pushing
layer 2 header so output device is available to calculate needed
headroom. Finally, add and use nf_flow_vlan_push() to fix it.
10) Fix flowtable inline pppoe with GSO packets. Moreover, use
FLOW_OFFLOAD_XMIT_DIRECT to fill up destination hardware
address since neighbour cache does not exist in pppoe.
11) Use skb_pull_rcsum() to decapsulate vlan and pppoe headers, for
double-tagged vlan in particular this should provide some benefits
in certain scenarios.
More notes regarding 9-11):
- sashiko is also signalling to use it for IPIP headers, but that needs
more adjustments such setting skb->protocol after removing the IPIP
header, will follow up in a separated patch.
- I plan to submit selftests to cover double-tagged-vlan. As for pppoe,
it should be possible but that would mandate a few userspace dependencies.
This has been semi-automatically tested by me and reporters describing
broken double-vlan-tagged and pppoe currently in the flowtable.
* tag 'nf-26-05-01' of git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf:
netfilter: flowtable: use skb_pull_rcsum() to pop vlan/pppoe header
netfilter: flowtable: fix inline pppoe encapsulation in xmit path
netfilter: flowtable: fix inline vlan encapsulation in xmit path
netfilter: flowtable: ensure sufficient headroom in xmit path
netfilter: xtables: fix L4 header parsing for non-first fragments
netfilter: nf_tables: skip L4 header parsing for non-first fragments
netfilter: nf_socket: skip socket lookup for non-first fragments
netfilter: nf_tables: fix netdev hook allocation memleak with dormant tables
netfilter: xt_CT: fix usersize for v1 and v2 revision
netfilter: nft_compat: run xt_check_hooks_{match,target}() from .validate
netfilter: x_tables: add .check_hooks to matches and targets
netfilter: nft_fwd_netdev: use recursion counter in neigh egress path
netfilter: nft_fwd_netdev: add device and headroom validate with neigh forwarding
netfilter: replace skb_try_make_writable() by skb_ensure_writable()
====================
netfilter: flowtable: use skb_pull_rcsum() to pop vlan/pppoe header
This adjusts the checksum, if required, after pulling the layer 2
header, either the pppoe header or the inner vlan header in the
double-tagged vlan packets.
This series tightens Marvell OcteonTX2 AF NPC support for CN20K silicon
around MCAM key typing, optional debugfs setup, defrag allocation rollback,
defrag entry relocation bookkeeping, logical MCAM clear and programming,
default-rule index handling with explicit teardown, and NIXLF reserved-slot
lookup when default rules are missing.
Patches 1 through 3 focus on AF error handling: propagate
npc_mcam_idx_2_key_type() failures through cn20k MCAM enable, config, copy,
and read paths; treat cn20k NPC debugfs nodes as optional so probe does not
fail when debugfs is unavailable; and fix defrag MCAM allocation rollback
so allocation errno is not overwritten during subbank index resolution.
Patch 4 fixes npc_defrag_move_vdx_to_free(): when an MCAM line is moved to
a new physical index, move entry2target_pffunc[] association to the new
slot, clear the old slot, and retarget the matching mcam_rules entry so
software state matches hardware after defrag.
Patches 5 through 7 refine cn20k MCAM programming: clear entries using the
logical MCAM index and resolved key width, fix bank/CFG sequencing in
npc_cn20k_config_mcam_entry(), and read action metadata from the correct
bank in npc_cn20k_read_mcam_entry().
Patches 8 through 10 complete default-rule lifecycle handling: initialize
default-rule index outputs eagerly, tear down reserved default MCAM rules
explicitly (coordinated with npc_mcam_free_all_entries()), and reject
USHRT_MAX sentinel indices from npc_get_nixlf_mcam_index() on cn20k.
====================
octeontx2-af: npc: cn20k: Reject missing default-rule MCAM indices
When cn20k default L2 rules are not installed,
npc_cn20k_dft_rules_idx_get() leaves broadcast, multicast, promiscuous, and
unicast slots at USHRT_MAX. npc_get_nixlf_mcam_index() previously returned
that sentinel as a valid MCAM index, so callers could program hardware with
an invalid index.
Return -EINVAL from the cn20k branches of npc_get_nixlf_mcam_index() when
the requested slot is still USHRT_MAX. Harden cn20k NPC MCAM entry helpers
to reject out-of-range indices before touching hardware.
Drop the early bounds check in npc_enable_mcam_entry() for cn20k so invalid
indices are validated inside npc_cn20k_enable_mcam_entry() instead of being
silently ignored.
In rvu_npc_update_flowkey_alg_idx(), treat negative MCAM indices like
out-of-range values, and only update RSS actions for promiscuous and
all-multi paths when the resolved index is non-negative.
octeontx2-af: npc: cn20k: Tear down default MCAM rules explicitly on free
npc_cn20k_dft_rules_free() used the NPC MCAM mbox "free all" path, which
does not match how cn20k tracks default-rule MCAM slots indexes.
Resolve the default-rule indices, then for each valid slot clear the bitmap
entry, drop the PF/VF map, disable the MCAM line, clear the target
function, and npc_cn20k_idx_free(). Remove any matching software mcam_rules
nodes. On hard failure from idx_free, WARN and stop so the box stays up for
analysis.
In npc_mcam_free_all_entries(), prefetch the same default-rule indices and,
on cn20k, skip bitmap clear and idx_free when the scanned entry is one of
those reserved defaults (they are released by npc_cn20k_dft_rules_free).
octeontx2-af: npc: cn20k: Initialize default-rule index outputs up front
npc_cn20k_dft_rules_idx_get() wrote USHRT_MAX into individual outputs only
on some error paths (lbk promisc lookup, VF ucast lookup, and the PF rule
walk), which could leave other caller slots stale across retries.
Set every non-NULL bcast/mcast/promisc/ucast pointer to USHRT_MAX once at
entry, then drop the duplicate assignments on failure. Successful lookups
still overwrite the relevant slot before returning.
npc_cn20k_read_mcam_entry() always reloaded action and vtag_action from
bank 0 after programming the CAM words. Use the bank returned by
npc_get_bank() for the ACTION reads as well, and read those registers once
up front so both X2 and X4 paths share the same metadata.
Return directly from the X2 keyword path now that the action fields are
already populated.
Cc: Suman Ghosh <sumang@marvell.com> Fixes: 6d1e70282f76 ("octeontx2-af: npc: cn20k: Use common APIs") Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com> Link: https://patch.msgid.link/20260429022722.1110289-8-rkannoth@marvell.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For X4 keys its loop reused the bank parameter as the loop counter, so bank
no longer reflected the caller's bank after the loop and the control flow
was hard to follow.
Program NPC_AF_CN20K_MCAMEX_BANKX_CFG_EXT directly in
npc_cn20k_config_mcam_entry(): one CFG write for X2 using the computed
bank, and one CFG write per bank inside the X4 action loop. Enable the
entry at the end with npc_cn20k_enable_mcam_entry(..., true) instead of
embedding the enable bit in bank_cfg via the removed helper.
Cc: Suman Ghosh <sumang@marvell.com> Fixes: 4e527f1e5c15 ("octeontx2-af: npc: cn20k: Add new mailboxes for CN20K silicon") Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com> Link: https://patch.msgid.link/20260429022722.1110289-7-rkannoth@marvell.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
octeontx2-af: npc: cn20k: Clear MCAM entries by index and key width
Replace the old four-argument CN20K MCAM clear with a per-bank static
helper and npc_cn20k_clear_mcam_entry() that takes a logical MCAM index,
resolves the key width via npc_mcam_idx_2_key_type(), and clears either one
bank (X2) or every bank (X4).
Call it from npc_clear_mcam_entry() on cn20k and log when key-type lookup
fails. Use the per-bank helper from npc_cn20k_config_mcam_entry() for
pre-program clears.
For loopback VFs, use the promisc MCAM index as ucast_idx when copying RSS
action for promisc, matching cn20k default-rule layout.
Cc: Suman Ghosh <sumang@marvell.com> Fixes: 6d1e70282f76 ("octeontx2-af: npc: cn20k: Use common APIs") Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com> Link: https://patch.msgid.link/20260429022722.1110289-6-rkannoth@marvell.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
npc_defrag_move_vdx_to_free() disables, copies, and enables the MCAM entry
at a new index but previously left entry2target_pffunc[] and the mcam_rules
list still keyed to the old index. Copy the target PF association to the
new slot, clear the old one, and retarget the rule entry so software state
matches the relocated hardware context.
octeontx2-af: npc: cn20k: Propagate errors in defrag MCAM alloc rollback
npc_defrag_alloc_free_slots() allocates MCAM indexes in up to two passes on
bank0 then bank1. On failure it rolls back by freeing entries already
placed in save[].
__npc_subbank_alloc() can return a negative errno while only part of the
indexes are valid. The rollback loop used rc for
npc_mcam_idx_2_subbank_idx() as well, so a successful lookup stored zero in
rc and a later __npc_subbank_free() failure could still end with return 0
when the allocation path had also left rc at zero (for example shortfall
after zero return values from the alloc helpers).
Jump to the rollback path immediately when either __npc_subbank_alloc()
call fails, preserving its errno. If both calls succeed but the total
allocated count is still less than cnt, set rc to -ENOSPC before rollback.
Use a separate err variable for npc_mcam_idx_2_subbank_idx() so a
successful lookup no longer clears a non-zero rc from the allocation phase.
octeontx2-af: npc: cn20k: Drop debugfs_create_file() error checks in init
debugfs is not intended to be checked for allocation failures the way other
kernel APIs are: callers should not fail probe or subsystem init because a
debugfs node could not be created, including when debugfs is disabled in
Kconfig. Replacing NULL checks with IS_ERR() checks is similarly wrong for
optional debugfs.
Remove dentry checks and -EFAULT returns from npc_cn20k_debugfs_init().
See:
https://staticthinking.wordpress.com/2023/07/24/
debugfs-functions-are-not-supposed-to-be-checked/
octeontx2-af: npc: cn20k: Propagate MCAM key-type errors on cn20k
npc_mcam_idx_2_key_type() can fail; callers used to ignore it and still
used kw_type when enabling, configuring, copying, and reading MCAM entries.
That could program or decode hardware with an undefined key type.
Return -EINVAL when key-type lookup fails. Return -EINVAL from
npc_cn20k_copy_mcam_entry() when src and dest key types differ instead of
failing silently.
Change npc_cn20k_{enable,config,copy,read}_mcam_entry() to return int on
success or error. Thread those errors through the cn20k MCAM write and read
mbox handlers, the cn20k baseline steer read path, NPC defrag move
(disable/copy/enable with dev_err and -EFAULT), and the DMAC update path in
rvu_npc_fs.c.
Make npc_copy_mcam_entry() return int so the cn20k branch can return
npc_cn20k_copy_mcam_entry() without a void/int mismatch, and fail
NPC_MCAM_SHIFT_ENTRY when copy fails.
Lorenzo Bianconi [Wed, 29 Apr 2026 12:02:31 +0000 (14:02 +0200)]
net: airoha: Move entries to queue head in case of DMA mapping failure in airoha_dev_xmit()
In order to respect the original descriptor order and avoid any
potential IOMMU fault or memory corruption, move pending queue entries
to the head of hw queue tx_list if the DMA mapping of current inflight
packet fails in airoha_dev_xmit routine.
Currently, request_threaded_irq() is used with a primary handler but a
NULL threaded handler, while also setting the IRQF_ONESHOT flag. This
specific combination triggers a WARNING since the commit aef30c8d569c
("genirq: Warn about using IRQF_ONESHOT without a threaded handler").
WARNING: kernel/irq/manage.c:1502 at __setup_irq+0x4fa/0x760
Fix the issue by switching to request_irq(), which is the appropriate
interface or a non-threaded interrupt handler, and removing the
unnecessary IRQF_ONESHOT flag.
Register WX_CFG_PORT_ST is a PF restricted register. When a VF is
initialized, attempting to read this register triggers an illegal
register access, which lead to a system hang.
When the device is VF, the bus function ID can be obtained directly from
the PCI_FUNC(pdev->devfn).
net: enetc: fix VSI mailbox timeout handling and DMA lifecycle
In the current VSI mailbox implementation, the VSI allocates a DMA buffer
to store the message sent to the PSI. When the PSI receives the message
request from the VSI, the hardware copies the message data from this DMA
buffer to PSI's DMA buffer for processing.
When enetc_msg_vsi_send() times out, two scenarios can occur:
1) Use-after-free: If the hardware hasn't completed message copying when
the VSI frees the buffer, the hardware may subsequently copy the data
from freed memory to PSI's DMA buffer.
2) Message race: If PSI hasn't processed the previous message when the
next message is sent, the VSI may receive the previous message's
reply, leading to incorrect handling.
To address these issues, implement the following changes:
- Check the mailbox busy status before sending a new message. If the
mailbox is in busy state, it indicates the previous message is still
being processed, so return an error immediately.
- Add the 'msg' field to struct enetc_si to preserve the DMA buffer
information. The caller of enetc_msg_vsi_send() no longer frees the
DMA buffer. Instead, defer freeing until it is safe to do so (when
mailbox is not busy on next send).
- Add cleanup in enetc_vf_remove() to free the last message buffer.
This ensures the DMA buffer remains valid during message copying and
prevents message reply mismatches.
Daniel Borkmann [Wed, 29 Apr 2026 15:46:48 +0000 (17:46 +0200)]
ipv6: Implement limits on extension header parsing
ipv6_{skip_exthdr,find_hdr}() and ip6_{tnl_parse_tlv_enc_lim,
protocol_deliver_rcu}() iterate over IPv6 extension headers until they
find a non-extension-header protocol or run out of packet data. The
loops have no iteration counter, relying solely on the packet length
to bound them. For a crafted packet with 8-byte extension headers
filling a 64KB jumbogram, this means a worst case of up to ~8k
iterations with a skb_header_pointer call each. ipv6_skip_exthdr(),
for example, is used where it parses the inner quoted packet inside
an incoming ICMPv6 error:
- icmpv6_rcv
- checksum validation
- case ICMPV6_DEST_UNREACH
- icmpv6_notify
- pskb_may_pull() <- pull inner IPv6 header
- ipv6_skip_exthdr() <- iterates here
- pskb_may_pull()
- ipprot->err_handler() <- sk lookup
The per-iteration cost of ipv6_skip_exthdr itself is generally
light, but skb_header_pointer becomes more costly on reassembled
packets: the first ~1232 bytes of the inner packet are in the skb's
linear area, but the remaining ~63KB are in the frag_list where
skb_copy_bits is needed to read data.
Initially, the idea was to add a configurable limit via a new
sysctl knob with default 8, in line with knobs from commit 47d3d7ac656a ("ipv6: Implement limits on Hop-by-Hop and Destination
options"), but two reasons eventually argued against it:
- It adds to UAPI that needs to be maintained forever, and
upcoming work is restricting extension header ordering anyway,
leaving little reason for another sysctl knob
- exthdrs_core.c is always built-in even when CONFIG_IPV6=n,
where struct net has no .ipv6 member, so the read site would
need an ifdef'd fallback to a constant anyway
Therefore, just use a constant (IP6_MAX_EXT_HDRS_CNT). All four
extension header walking functions are now bound by this limit.
Note that the check in ip6_protocol_deliver_rcu() happens right
before the goto resubmit, such that we don't have to have a test
for ipv6_ext_hdr() in the fast-path.
There's an ongoing IETF draft-iurman-6man-eh-occurrences to enforce
IPv6 extension headers ordering and occurrence. The latter also
discusses security implications. As per RFC8200 section 4.1, the
occurrence rules for extension headers provide a practical upper
bound which is 8. In order to be conservative, let's define
IP6_MAX_EXT_HDRS_CNT as 12 to leave enough room for quirky setups.
In the unlikely event that this is still not enough, then we might
need to reconsider a sysctl.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Reviewed-by: Ido Schimmel <idosch@nvidia.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Reviewed-by: Justin Iurman <justin.iurman@gmail.com> Link: https://patch.msgid.link/20260429154648.809751-1-daniel@iogearbox.net Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Robert Marko [Tue, 28 Apr 2026 13:41:01 +0000 (15:41 +0200)]
net: phy: micrel: fix LAN8814 QSGMII soft reset
LAN8814 QSGMII soft reset was moved into the probe function to avoid
triggering it for each of 4 PHY-s in the package.
However, that broke QSGMII link between the MAC and PHY on most LAN8814
PHY-s, specificaly for us on the Microchip LAN969x switch.
Reading the QSGMII status registers it was visible that lanes were only
partially synced.
It looks like the reset timing is crucial, so lets move the reset back
into the .config_init function but guard it with phy_package_init_once()
to avoid it being triggered on each of 4 PHY-s in the package.
Change the probe function to use phy_package_probe_once() for coma and PtP
setup.
netfilter: flowtable: fix inline pppoe encapsulation in xmit path
Address two issues in the inline pppoe encapsulation:
- Add needs_gso_segment flag to segment PPPoE packets in software
given that there is no GSO support for this.
- Use FLOW_OFFLOAD_XMIT_DIRECT since neighbour cache is not available
in point-to-point device, use the hardware address that is obtained
via flowtable path discovery (ie. fill_forward_path).
Fixes: 18d27bed0880 ("netfilter: flowtable: inline pppoe encapsulation in xmit path") Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
netfilter: flowtable: fix inline vlan encapsulation in xmit path
Several issues in the inline vlan support:
- The layer 2 encapsulation representation in the tuple takes encap[0] as
the outer header and encap[1] as the inner header as seen from the ingress
path. Reverse the encap loop to push first the inner then the outer vlan
header.
- Postpone pushing the layer 2 header once destination device is known.
This allows to calculate the needed hearoom via LL_RESERVED_SPACE to
accommodate the layer 2 headers.
- Add and use nf_flow_vlan_push() as suggested by Eric Woudstra, this
is a simplified version of skb_vlan_push() for egress path only.
Fixes: c653d5a78f34 ("netfilter: flowtable: inline vlan encapsulation in xmit path") Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
projects / thirdparty / kernel / linux.git / log
