In ice_set_ringparam, tx_rings and xdp_rings are allocated before
rx_rings. If the allocation of rx_rings fails, the code jumps to
the done label leaking both tx_rings and xdp_rings. Furthermore, if
the setup of an individual Rx ring fails during the loop, the code jumps
to the free_tx label which releases tx_rings but leaks xdp_rings.
Fix this by introducing a free_xdp label and updating the error paths to
ensure both xdp_rings and tx_rings are properly freed if rx_rings
allocation or setup fails.
Compile tested only. Issue found using a prototype static analysis tool
and code review.
Fixes: fcea6f3da5 ("ice: Add stats and ethtool support")
Fixes: efc2214b60 ("ice: Add support for XDP")
Signed-off-by: Zilin Guan <zilin@seu.edu.cn>
Reviewed-by: Paul Menzel <pmenzel@molgen.mpg.de>
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Tested-by: Rinitha S <sx.rinitha@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
The referenced commit came from a misunderstanding of the FW LLDP filter
AQ (Admin Queue) command due to the error in the internal documentation.
Contrary to the assumptions in the original commit, VFs can be added and
deleted from this filter without any problems. Introduced dev_info message
proved to be useful, so reverting the whole commit does not make sense.
Without this fix, trusted VFs do not receive LLDP traffic, if there is an
AQ LLDP filter on PF. When trusted VF attempts to add an LLDP multicast
MAC address, the following message can be seen in dmesg on host:
ice 0000:33:00.0: Failed to add Rx LLDP rule on VSI 20 error: -95
Revert checking VSI type when adding LLDP filter through AQ.
Fixes: 4d5a1c4e6d ("ice: do not add LLDP-specific filter if not necessary")
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Signed-off-by: Larysa Zaremba <larysa.zaremba@intel.com>
Tested-by: Rafal Romanowski <rafal.romanowski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
I was debugging a NIC driver when I noticed that when I enable
threaded busypoll, bpftrace hangs when starting up. dmesg showed:
rcu_tasks_wait_gp: rcu_tasks grace period number 85 (since boot) is 10658 jiffies old.
rcu_tasks_wait_gp: rcu_tasks grace period number 85 (since boot) is 40793 jiffies old.
rcu_tasks_wait_gp: rcu_tasks grace period number 85 (since boot) is 131273 jiffies old.
rcu_tasks_wait_gp: rcu_tasks grace period number 85 (since boot) is 402058 jiffies old.
INFO: rcu_tasks detected stalls on tasks:
00000000769f52cd: .N nvcsw: 2/2 holdout: 1 idle_cpu: -1/64
task:napi/eth2-8265 state:R running task stack:0 pid:48300 tgid:48300 ppid:2 task_flags:0x208040 flags:0x00004000
Call Trace:
<TASK>
? napi_threaded_poll_loop+0x27c/0x2c0
? __pfx_napi_threaded_poll+0x10/0x10
? napi_threaded_poll+0x26/0x80
? kthread+0xfa/0x240
? __pfx_kthread+0x10/0x10
? ret_from_fork+0x31/0x50
? __pfx_kthread+0x10/0x10
? ret_from_fork_asm+0x1a/0x30
</TASK>
The cause is that in threaded busypoll, the main loop is in
napi_threaded_poll rather than napi_threaded_poll_loop, where the
latter rarely iterates more than once within its loop. For
rcu_softirq_qs_periodic inside napi_threaded_poll_loop to report its
qs state, the last_qs must be 100ms behind, and this can't happen
because napi_threaded_poll_loop rarely iterates in threaded busypoll,
and each time napi_threaded_poll_loop is called last_qs is reset to
latest jiffies.
This patch changes so that in threaded busypoll, last_qs is saved
in the outer napi_threaded_poll, and whether busy_poll_last_qs
is NULL indicates whether napi_threaded_poll_loop is called for
busypoll. This way last_qs would not reset to latest jiffies on
each invocation of napi_threaded_poll_loop.
Fixes: c18d4b190a ("net: Extend NAPI threaded polling to allow kthread based busy polling")
Cc: stable@vger.kernel.org
Signed-off-by: YiFei Zhu <zhuyifei@google.com>
Reviewed-by: Samiullah Khawaja <skhawaja@google.com>
Link: https://patch.msgid.link/20260227221937.1060857-1-zhuyifei@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
syzbot reported a circular locking dependency in rds_tcp_tune() where
sk_net_refcnt_upgrade() is called while holding the socket lock:
======================================================
WARNING: possible circular locking dependency detected
======================================================
kworker/u10:8/15040 is trying to acquire lock:
ffffffff8e9aaf80 (fs_reclaim){+.+.}-{0:0},
at: __kmalloc_cache_noprof+0x4b/0x6f0
but task is already holding lock:
ffff88805a3c1ce0 (k-sk_lock-AF_INET6){+.+.}-{0:0},
at: rds_tcp_tune+0xd7/0x930
The issue occurs because sk_net_refcnt_upgrade() performs memory
allocation (via get_net_track() -> ref_tracker_alloc()) while the
socket lock is held, creating a circular dependency with fs_reclaim.
Fix this by moving sk_net_refcnt_upgrade() outside the socket lock
critical section. This is safe because the fields modified by the
sk_net_refcnt_upgrade() call (sk_net_refcnt, ns_tracker) are not
accessed by any concurrent code path at this point.
v2:
- Corrected fixes tag
- check patch line wrap nits
- ai commentary nits
Reported-by: syzbot+2e2cf5331207053b8106@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=2e2cf5331207053b8106
Fixes: 3a58f13a88 ("net: rds: acquire refcount on TCP sockets")
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260227202336.167757-1-achender@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Vimlesh Kumar says:
====================
avoid compiler and IQ/OQ reordering
Utilize READ_ONCE and WRITE_ONCE APIs to prevent compiler
optimization and reordering. Ensure IO queue OUT/IN_CNT
registers are flushed. Relocate IQ/OQ IN/OUT_CNTS updates
to occur before NAPI completion, and replace napi_complete
with napi_complete_done.
====================
Link: https://patch.msgid.link/20260227091402.1773833-1-vimleshk@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Utilize READ_ONCE and WRITE_ONCE APIs for IO queue Tx/Rx
variable access to prevent compiler optimization and reordering.
Additionally, ensure IO queue OUT/IN_CNT registers are flushed
by performing a read-back after writing.
The compiler could reorder reads/writes to pkts_pending, last_pkt_count,
etc., causing stale values to be used when calculating packets to process
or register updates to send to hardware. The Octeon hardware requires a
read-back after writing to OUT_CNT/IN_CNT registers to ensure the write
has been flushed through any posted write buffers before the interrupt
resend bit is set. Without this, we have observed cases where the hardware
didn't properly update its internal state.
wmb/rmb only provides ordering guarantees but doesn't prevent the compiler
from performing optimizations like caching in registers, load tearing etc.
Fixes: 1cd3b40797 ("octeon_ep_vf: add Tx/Rx processing and interrupt support")
Signed-off-by: Sathesh Edara <sedara@marvell.com>
Signed-off-by: Shinas Rasheed <srasheed@marvell.com>
Signed-off-by: Vimlesh Kumar <vimleshk@marvell.com>
Link: https://patch.msgid.link/20260227091402.1773833-5-vimleshk@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Relocate IQ/OQ IN/OUT_CNTS updates to occur before NAPI completion.
Moving the IQ/OQ counter updates before napi_complete_done ensures
1. Counter registers are updated before re-enabling interrupts.
2. Prevents a race where new packets arrive but counters aren't properly
synchronized.
Fixes: 1cd3b40797 ("octeon_ep_vf: add Tx/Rx processing and interrupt support")
Signed-off-by: Sathesh Edara <sedara@marvell.com>
Signed-off-by: Shinas Rasheed <srasheed@marvell.com>
Signed-off-by: Vimlesh Kumar <vimleshk@marvell.com>
Link: https://patch.msgid.link/20260227091402.1773833-4-vimleshk@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Utilize READ_ONCE and WRITE_ONCE APIs for IO queue Tx/Rx
variable access to prevent compiler optimization and reordering.
Additionally, ensure IO queue OUT/IN_CNT registers are flushed
by performing a read-back after writing.
The compiler could reorder reads/writes to pkts_pending, last_pkt_count,
etc., causing stale values to be used when calculating packets to process
or register updates to send to hardware. The Octeon hardware requires a
read-back after writing to OUT_CNT/IN_CNT registers to ensure the write
has been flushed through any posted write buffers before the interrupt
resend bit is set. Without this, we have observed cases where the hardware
didn't properly update its internal state.
wmb/rmb only provides ordering guarantees but doesn't prevent the compiler
from performing optimizations like caching in registers, load tearing etc.
Fixes: 37d79d0596 ("octeon_ep: add Tx/Rx processing and interrupt support")
Signed-off-by: Sathesh Edara <sedara@marvell.com>
Signed-off-by: Shinas Rasheed <srasheed@marvell.com>
Signed-off-by: Vimlesh Kumar <vimleshk@marvell.com>
Link: https://patch.msgid.link/20260227091402.1773833-3-vimleshk@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Relocate IQ/OQ IN/OUT_CNTS updates to occur before NAPI completion,
and replace napi_complete with napi_complete_done.
Moving the IQ/OQ counter updates before napi_complete_done ensures
1. Counter registers are updated before re-enabling interrupts.
2. Prevents a race where new packets arrive but counters aren't properly
synchronized.
napi_complete_done (vs napi_complete) allows for better
interrupt coalescing.
Fixes: 37d79d0596 ("octeon_ep: add Tx/Rx processing and interrupt support")
Signed-off-by: Sathesh Edara <sedara@marvell.com>
Signed-off-by: Shinas Rasheed <srasheed@marvell.com>
Signed-off-by: Vimlesh Kumar <vimleshk@marvell.com>
Link: https://patch.msgid.link/20260227091402.1773833-2-vimleshk@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Jiayuan Chen says:
====================
bonding: fix missing XDP compat check on xmit_hash_policy change
syzkaller reported a bug https://syzkaller.appspot.com/bug?extid=5a287bcdc08104bc3132
When a bond device is in 802.3ad or balance-xor mode, XDP is supported
only when xmit_hash_policy != vlan+srcmac. This constraint is enforced
in bond_option_mode_set() via bond_xdp_check(), which prevents switching
to an XDP-incompatible mode while a program is loaded. However, the
symmetric path -- changing xmit_hash_policy while XDP is loaded -- had
no such guard in bond_option_xmit_hash_policy_set().
This means the following sequence silently creates an inconsistent state:
1. Create a bond in 802.3ad mode with xmit_hash_policy=layer2+3.
2. Attach a native XDP program to the bond.
3. Change xmit_hash_policy to vlan+srcmac (no error, not checked).
Now bond->xdp_prog is set but bond_xdp_check() returns false for the
same device. When the bond is later torn down (e.g. netns deletion),
dev_xdp_uninstall() calls bond_xdp_set(dev, NULL) to remove the
program, which hits the bond_xdp_check() guard and returns -EOPNOTSUPP,
triggering a kernel WARNING:
bond1 (unregistering): Error: No native XDP support for the current bonding mode
------------[ cut here ]------------
dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)
WARNING: net/core/dev.c:10361 at dev_xdp_uninstall net/core/dev.c:10361 [inline], CPU#0: kworker/u8:22/11031
Modules linked in:
CPU: 0 UID: 0 PID: 11031 Comm: kworker/u8:22 Not tainted syzkaller #0 PREEMPT(full)
Workqueue: netns cleanup_net
RIP: 0010:dev_xdp_uninstall net/core/dev.c:10361 [inline]
RIP: 0010:unregister_netdevice_many_notify+0x1efd/0x2370 net/core/dev.c:12393
RSP: 0018:ffffc90003b2f7c0 EFLAGS: 00010293
RAX: ffffffff8971e99c RBX: ffff888052f84c40 RCX: ffff88807896bc80
RDX: 0000000000000000 RSI: 00000000ffffffa1 RDI: 0000000000000000
RBP: ffffc90003b2f930 R08: ffffc90003b2f207 R09: 1ffff92000765e40
R10: dffffc0000000000 R11: fffff52000765e41 R12: 00000000ffffffa1
R13: ffff888052f84c38 R14: 1ffff1100a5f0988 R15: ffffc9000df67000
FS: 0000000000000000(0000) GS:ffff8881254ae000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f60871d5d58 CR3: 000000006c41c000 CR4: 00000000003526f0
Call Trace:
<TASK>
ops_exit_rtnl_list net/core/net_namespace.c:187 [inline]
ops_undo_list+0x3d3/0x940 net/core/net_namespace.c:248
cleanup_net+0x56b/0x800 net/core/net_namespace.c:704
process_one_work kernel/workqueue.c:3275 [inline]
process_scheduled_works+0xaec/0x17a0 kernel/workqueue.c:3358
worker_thread+0xa50/0xfc0 kernel/workqueue.c:3439
kthread+0x388/0x470 kernel/kthread.c:467
ret_from_fork+0x51e/0xb90 arch/x86/kernel/process.c:158
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
</TASK>
Beyond the WARNING itself, when dev_xdp_install() fails during
dev_xdp_uninstall(), bond_xdp_set() returns early without calling
bpf_prog_put() on the old program. dev_xdp_uninstall() then releases
only the reference held by dev->xdp_state[], while the reference held
by bond->xdp_prog is never dropped, leaking the struct bpf_prog.
The fix refactors the core logic of bond_xdp_check() into a new helper
__bond_xdp_check_mode(mode, xmit_policy) that takes both parameters
explicitly, avoiding the need to read them from the bond struct.
bond_xdp_check() becomes a thin wrapper around it.
bond_option_xmit_hash_policy_set() then uses __bond_xdp_check_mode()
directly, passing the candidate xmit_policy before it is committed,
mirroring exactly what bond_option_mode_set() already does for mode
changes.
Patch 1 adds the kernel fix.
Patch 2 adds a selftest that reproduces the WARNING by attaching native
XDP to a bond in 802.3ad mode, then attempting to change xmit_hash_policy
to vlan+srcmac -- verifying the change is rejected with the fix applied.
====================
Link: https://patch.msgid.link/20260226080306.98766-1-jiayuan.chen@linux.dev
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a selftest to verify that changing xmit_hash_policy to vlan+srcmac
is rejected when a native XDP program is loaded on a bond in 802.3ad
mode. Without the fix in bond_option_xmit_hash_policy_set(), the change
succeeds silently, creating an inconsistent state that triggers a kernel
WARNING in dev_xdp_uninstall() when the bond is torn down.
The test attaches native XDP to a bond0 (802.3ad, layer2+3), then
attempts to switch xmit_hash_policy to vlan+srcmac and asserts the
operation fails. It also verifies the change succeeds after XDP is
detached, confirming the rejection is specific to the XDP-loaded state.
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
Link: https://patch.msgid.link/20260226080306.98766-3-jiayuan.chen@linux.dev
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
bond_option_mode_set() already rejects mode changes that would make a
loaded XDP program incompatible via bond_xdp_check(). However,
bond_option_xmit_hash_policy_set() has no such guard.
For 802.3ad and balance-xor modes, bond_xdp_check() returns false when
xmit_hash_policy is vlan+srcmac, because the 802.1q payload is usually
absent due to hardware offload. This means a user can:
1. Attach a native XDP program to a bond in 802.3ad/balance-xor mode
with a compatible xmit_hash_policy (e.g. layer2+3).
2. Change xmit_hash_policy to vlan+srcmac while XDP remains loaded.
This leaves bond->xdp_prog set but bond_xdp_check() now returning false
for the same device. When the bond is later destroyed, dev_xdp_uninstall()
calls bond_xdp_set(dev, NULL, NULL) to remove the program, which hits
the bond_xdp_check() guard and returns -EOPNOTSUPP, triggering:
WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL))
Fix this by rejecting xmit_hash_policy changes to vlan+srcmac when an
XDP program is loaded on a bond in 802.3ad or balance-xor mode.
commit 39a0876d59 ("net, bonding: Disallow vlan+srcmac with XDP")
introduced bond_xdp_check() which returns false for 802.3ad/balance-xor
modes when xmit_hash_policy is vlan+srcmac. The check was wired into
bond_xdp_set() to reject XDP attachment with an incompatible policy, but
the symmetric path -- preventing xmit_hash_policy from being changed to an
incompatible value after XDP is already loaded -- was left unguarded in
bond_option_xmit_hash_policy_set().
Note:
commit 094ee6017e ("bonding: check xdp prog when set bond mode")
later added a similar guard to bond_option_mode_set(), but
bond_option_xmit_hash_policy_set() remained unprotected.
Reported-by: syzbot+5a287bcdc08104bc3132@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/6995aff6.050a0220.2eeac1.014e.GAE@google.com/T/
Fixes: 39a0876d59 ("net, bonding: Disallow vlan+srcmac with XDP")
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
Link: https://patch.msgid.link/20260226080306.98766-2-jiayuan.chen@linux.dev
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Commit 1cc93c48b5 ("selftests/net: packetdrill: remove tests for
tcp_rcv_*big") removed the test for the reverted commit 1d2fbaad7c
("tcp: stronger sk_rcvbuf checks") but also the one for commit
9ca48d616e ("tcp: do not accept packets beyond window").
Restore the test with the necessary adaptation: expect a delayed ACK
instead of an immediate one, since tcp_can_ingest() does not fail
anymore for the last data packet.
Signed-off-by: Simon Baatz <gmbnomis@gmail.com>
Link: https://patch.msgid.link/20260301-tcp_rcv_big_endseq-v1-1-86ab7415ab58@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Extend the MAC_TCR_SS (Speed Select) register field width from 2 bits
to 3 bits to properly support all speed settings.
The MAC_TCR register's SS field encoding requires 3 bits to represent
all supported speeds:
- 0x00: 10Gbps (XGMII)
- 0x02: 2.5Gbps (GMII) / 100Mbps
- 0x03: 1Gbps / 10Mbps
- 0x06: 2.5Gbps (XGMII) - P100a only
With only 2 bits, values 0x04-0x07 cannot be represented, which breaks
2.5G XGMII mode on newer platforms and causes incorrect speed select
values to be programmed.
Fixes: 07445f3c7c ("amd-xgbe: Add support for 10 Mbps speed")
Co-developed-by: Guruvendra Punugupati <Guruvendra.Punugupati@amd.com>
Signed-off-by: Guruvendra Punugupati <Guruvendra.Punugupati@amd.com>
Signed-off-by: Raju Rangoju <Raju.Rangoju@amd.com>
Link: https://patch.msgid.link/20260226170753.250312-1-Raju.Rangoju@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When both eth interfaces with links up are added to a bridge or hsr
interface, ping fails if the link speed is not 1Gbps (e.g., 100Mbps).
The issue is seen because when switching to offload (bridge/hsr) mode,
prueth_emac_restart() restarts the firmware and clears DRAM with
memset_io(), setting all memory to 0. This includes PORT_LINK_SPEED_OFFSET
which firmware reads for link speed. The value 0 corresponds to
FW_LINK_SPEED_1G (0x00), so for 1Gbps links the default value is correct
and ping works. For 100Mbps links, the firmware needs FW_LINK_SPEED_100M
(0x01) but gets 0 instead, causing ping to fail. The function
emac_adjust_link() is called to reconfigure, but it detects no state change
(emac->link is still 1, speed/duplex match PHY) so new_state remains false
and icssg_config_set_speed() is never called to correct the firmware speed
value.
The fix resets emac->link to 0 before calling emac_adjust_link() in
prueth_emac_common_start(). This forces new_state=true, ensuring
icssg_config_set_speed() is called to write the correct speed value to
firmware memory.
Fixes: 06feac1540 ("net: ti: icssg-prueth: Fix emac link speed handling")
Signed-off-by: MD Danish Anwar <danishanwar@ti.com>
Link: https://patch.msgid.link/20260226102356.2141871-1-danishanwar@ti.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
syzkaller reported a null-ptr-deref in lec_arp_clear_vccs().
This issue can be easily reproduced using the syzkaller reproducer.
In the ATM LANE (LAN Emulation) module, the same atm_vcc can be shared by
multiple lec_arp_table entries (e.g., via entry->vcc or entry->recv_vcc).
When the underlying VCC is closed, lec_vcc_close() iterates over all
ARP entries and calls lec_arp_clear_vccs() for each matched entry.
For example, when lec_vcc_close() iterates through the hlists in
priv->lec_arp_empty_ones or other ARP tables:
1. In the first iteration, for the first matched ARP entry sharing the VCC,
lec_arp_clear_vccs() frees the associated vpriv (which is vcc->user_back)
and sets vcc->user_back to NULL.
2. In the second iteration, for the next matched ARP entry sharing the same
VCC, lec_arp_clear_vccs() is called again. It obtains a NULL vpriv from
vcc->user_back (via LEC_VCC_PRIV(vcc)) and then attempts to dereference it
via `vcc->pop = vpriv->old_pop`, leading to a null-ptr-deref crash.
Fix this by adding a null check for vpriv before dereferencing
it. If vpriv is already NULL, it means the VCC has been cleared
by a previous call, so we can safely skip the cleanup and just
clear the entry's vcc/recv_vcc pointers.
The entire cleanup block (including vcc_release_async()) is placed inside
the vpriv guard because a NULL vpriv indicates the VCC has already been
fully released by a prior iteration — repeating the teardown would
redundantly set flags and trigger callbacks on an already-closing socket.
The Fixes tag points to the initial commit because the entry->vcc path has
been vulnerable since the original code. The entry->recv_vcc path was later
added by commit 8d9f73c0ad ("atm: fix a memory leak of vcc->user_back")
with the same pattern, and both paths are fixed here.
Reported-by: syzbot+72e3ea390c305de0e259@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/68c95a83.050a0220.3c6139.0e5c.GAE@google.com/T/
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Suggested-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
Link: https://patch.msgid.link/20260225123250.189289-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit 31a7a0bbeb ("dpaa2-switch: add bounds check for if_id in IRQ
handler") introduces a range check for if_id to avoid an out-of-bounds
access. If an out-of-bounds if_id is detected, the interrupt status is
not cleared. This may result in an interrupt storm.
Clear the interrupt status after detecting an out-of-bounds if_id to avoid
the problem.
Found by an experimental AI code review agent at Google.
Fixes: 31a7a0bbeb ("dpaa2-switch: add bounds check for if_id in IRQ handler")
Cc: Junrui Luo <moonafterrain@outlook.com>
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/20260227055812.1777915-1-linux@roeck-us.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Nikhil P. Rao says:
====================
xsk: Fixes for AF_XDP fragment handling
This series fixes two issues in AF_XDP zero-copy fragment handling:
Patch 1 fixes a buffer leak caused by incorrect list node handling after
commit b692bf9a75. The list_node field is now reused for both the xskb
pool list and the buffer free list. Using list_del() instead of
list_del_init() causes list_empty() checks in xp_free() to fail, preventing
buffers from being added to the free list.
Patch 2 fixes partial packet delivery to userspace. In the zero-copy path,
if the Rx queue fills up while enqueuing fragments, the remaining fragments
are dropped, causing the application to receive incomplete packets. The fix
ensures the Rx queue has sufficient space for all fragments before starting
to enqueue them.
[1] https://lore.kernel.org/oe-kbuild-all/202602051720.YfZO23pZ-lkp@intel.com/
[2] https://lore.kernel.org/oe-kbuild-all/202602172046.vf9DtpdF-lkp@intel.com/
====================
Link: https://patch.msgid.link/20260225000456.107806-1-nikhil.rao@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
AF_XDP should ensure that only a complete packet is sent to application.
In the zero-copy case, if the Rx queue gets full as fragments are being
enqueued, the remaining fragments are dropped.
For the multi-buffer case, add a check to ensure that the Rx queue has
enough space for all fragments of a packet before starting to enqueue
them.
Fixes: 24ea50127e ("xsk: support mbuf on ZC RX")
Signed-off-by: Nikhil P. Rao <nikhil.rao@amd.com>
Link: https://patch.msgid.link/20260225000456.107806-3-nikhil.rao@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
After commit b692bf9a75 ("xsk: Get rid of xdp_buff_xsk::xskb_list_node"),
the list_node field is reused for both the xskb pool list and the buffer
free list, this causes a buffer leak as described below.
xp_free() checks if a buffer is already on the free list using
list_empty(&xskb->list_node). When list_del() is used to remove a node
from the xskb pool list, it doesn't reinitialize the node pointers.
This means list_empty() will return false even after the node has been
removed, causing xp_free() to incorrectly skip adding the buffer to the
free list.
Fix this by using list_del_init() instead of list_del() in all fragment
handling paths, this ensures the list node is reinitialized after removal,
allowing the list_empty() to work correctly.
Fixes: b692bf9a75 ("xsk: Get rid of xdp_buff_xsk::xskb_list_node")
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Nikhil P. Rao <nikhil.rao@amd.com>
Link: https://patch.msgid.link/20260225000456.107806-2-nikhil.rao@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Tony Nguyen says:
====================
Intel Wired LAN Driver Updates 2026-02-19 (idpf, ice, i40e, ixgbevf, e1000e)
For idpf:
Li Li moves the check for software marker to occur after incrementing
next to clean to avoid re-encountering the same packet. He also adds a
couple of checks to prevent NULL pointer dereferences and NULLs rss_key,
after free, in error path so that later checks are properly evaluated.
Brian Vazquez adjusts IRQ naming to have correlation with netdev naming.
Sreedevi removes validation of action type as part of ntuple rule
deletion.
For ice:
Aaron Ma breaks RDMA initialization into two steps and adjusts calls so
that VSIs are entirely configured before plugging.
Michal Schmidt fixes initialization of loopback VSI to have proper
resources allocated to allow for loopback testing to occur.
For i40e:
Thomas Gleixner fixes a leak of preempt count by replacing get_cpu()
with smp_processor_id().
For ixgbevf:
Jedrzej adds a check for mailbox version before attempting to call an
associated link state call that is supported in that mailbox version.
For e1000e:
Vitaly clears power gating feature for Panther Lake systems to avoid
packet issues.
* '200GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue:
e1000e: clear DPG_EN after reset to avoid autonomous power-gating
e1000e: introduce new board type for Panther Lake PCH
ixgbevf: fix link setup issue
i40e: Fix preempt count leak in napi poll tracepoint
ice: fix crash in ethtool offline loopback test
ice: recap the VSI and QoS info after rebuild
idpf: Fix flow rule delete failure due to invalid validation
idpf: change IRQ naming to match netdev and ethtool queue numbering
idpf: nullify pointers after they are freed
idpf: skip deallocating txq group's txqs if it is NULL
idpf: skip deallocating bufq_sets from rx_qgrp if it is NULL
idpf: increment completion queue next_to_clean in sw marker wait routine
====================
Link: https://patch.msgid.link/20260225211546.1949260-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We hit another corner case which leads to TcpExtTCPRcvQDrop
Connections which send RPCs in the 20-80kB range over loopback
experience spurious drops. The exact conditions for most of
the drops I investigated are that:
- socket exchanged >1MB of data so its not completely fresh
- rcvbuf is around 128kB (default, hasn't grown)
- there is ~60kB of data in rcvq
- skb > 64kB arrives
The sum of skb->len (!) of both of the skbs (the one already
in rcvq and the arriving one) is larger than rwnd.
My suspicion is that this happens because __tcp_select_window()
rounds the rwnd up to (1 << wscale) if less than half of
the rwnd has been consumed.
Eric suggests that given the number of Fixes we already have
pointing to 1d2fbaad7c it's probably time to give up on it,
until a bigger revamp of rmem management.
Also while we could risk tweaking the rwnd math, there are other
drops on workloads I investigated, after the commit in question,
not explained by this phenomenon.
Suggested-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/20260225122355.585fd57b@kernel.org
Fixes: 1d2fbaad7c ("tcp: stronger sk_rcvbuf checks")
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260227003359.2391017-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Let's say we bind() an UDP socket to the wildcard address with a
non-zero port, connect() it to an address, and disconnect it from
the address.
bind() sets SOCK_BINDPORT_LOCK on sk->sk_userlocks (but not
SOCK_BINDADDR_LOCK), and connect() calls udp_lib_hash4() to put
the socket into the 4-tuple hash table.
Then, __udp_disconnect() calls sk->sk_prot->rehash(sk).
It computes a new hash based on the wildcard address and moves
the socket to a new slot in the 4-tuple hash table, leaving a
garbage in the chain that no packet hits.
Let's remove such a socket from 4-tuple hash table when disconnected.
Note that udp_sk(sk)->udp_portaddr_hash needs to be udpated after
udp_hash4_dec(hslot2) in udp_unhash4().
Fixes: 78c91ae2c6 ("ipv4/udp: Add 4-tuple hash for connected socket")
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260227035547.3321327-1-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
MANA hardware requires at least one doorbell ring every 8 wraparounds
of the CQ. The driver rings the doorbell as a form of flow control to
inform hardware that CQEs have been consumed.
The NAPI poll functions mana_poll_tx_cq() and mana_poll_rx_cq() can
poll up to CQE_POLLING_BUFFER (512) completions per call. If the CQ
has fewer than 512 entries, a single poll call can process more than
4 wraparounds without ringing the doorbell. The doorbell threshold
check also uses ">" instead of ">=", delaying the ring by one extra
CQE beyond 4 wraparounds. Combined, these issues can cause the driver
to exceed the 8-wraparound hardware limit, leading to missed
completions and stalled queues.
Fix this by capping the number of CQEs polled per call to 4 wraparounds
of the CQ in both TX and RX paths. Also change the doorbell threshold
from ">" to ">=" so the doorbell is rung as soon as 4 wraparounds are
reached.
Cc: stable@vger.kernel.org
Fixes: 58a63729c9 ("net: mana: Fix doorbell out of order violation and avoid unnecessary doorbell rings")
Signed-off-by: Long Li <longli@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Link: https://patch.msgid.link/20260226192833.1050807-1-longli@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jakub reports test flakes on debug kernels:
FAIL: test_udp_gro_ct: Expected software segmentation to occur, had 23 and 17
This test assumes that the kernels nfnetlink_queue module sees N GSO
packets, segments them into M skbs and queues them to userspace for
reinjection.
Hence, if M >= N, no segmentation occurred.
However, its possible that this happens:
- nfnetlink_queue gets GSO packet
- segments that into n skbs
- userspace buffer is full, kernel drops the segmented skbs
-> "toqueue" counter incremented by 1, "fromqueue" is unchanged.
If this happens often enough in a single run, M >= N check triggers
incorrectly.
To solve this, allow the nf_queue.c test program to set the FAIL_OPEN
flag so that the segmented skbs bypass the queueing step in the kernel
if the receive buffer is full.
Also, reduce number of sending socat instances, decrease their priority
and increase nice value for the nf_queue program itself to reduce the
probability of overruns happening in the first place.
Fixes: 59ecffa399 ("selftests: netfilter: nft_queue.sh: add udp fraglist gro test case")
Reported-by: Jakub Kicinski <kuba@kernel.org>
Closes: https://lore.kernel.org/netdev/20260218184114.0b405b72@kernel.org/
Signed-off-by: Florian Westphal <fw@strlen.de>
Link: https://patch.msgid.link/20260226161920.1205-1-fw@strlen.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The gate action can be replaced while the hrtimer callback or dump path is
walking the schedule list.
Convert the parameters to an RCU-protected snapshot and swap updates under
tcf_lock, freeing the previous snapshot via call_rcu(). When REPLACE omits
the entry list, preserve the existing schedule so the effective state is
unchanged.
Fixes: a51c328df3 ("net: qos: introduce a gate control flow action")
Cc: stable@vger.kernel.org
Signed-off-by: Paul Moses <p@1g4.org>
Tested-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Link: https://patch.msgid.link/20260223150512.2251594-2-p@1g4.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In the current implementation, flushing multicast entries in MAC mode
incorrectly deletes entries for all ports instead of only the target port,
disrupting multicast traffic on other ports. The cause is adding multicast
entries by setting only host port bit, and not setting the MAC port bits.
Fix this by setting the MAC port's bit in the port mask while adding the
multicast entry. Also fix the flush logic to preserve the host port bit
during removal of MAC port and free ALE entries when mask contains only
host port.
Fixes: 5c50a856d5 ("drivers: net: ethernet: cpsw: add multicast address to ALE table")
Signed-off-by: Chintan Vankar <c-vankar@ti.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260224181359.2055322-1-c-vankar@ti.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Danielle Ratson says:
====================
bridge: Check relevant options in VLAN range grouping
The br_vlan_opts_eq_range() function determines if consecutive VLANs can
be grouped together in a range for compact netlink notifications. It
currently checks state, tunnel info, and multicast router configuration,
but misses two categories of per-VLAN options that affect the output:
1. User-visible priv_flags (neigh_suppress, mcast_enabled)
2. Port multicast context options (mcast_max_groups, mcast_n_groups)
When VLANs have different settings for these options, they are incorrectly
grouped into ranges, causing netlink notifications to report only one
VLAN's settings for the entire range.
Fix by checking priv_flags equality, but only for flags that affect netlink
output (BR_VLFLAG_NEIGH_SUPPRESS_ENABLED and BR_VLFLAG_MCAST_ENABLED),
and comparing multicast context options (mcast_max_groups, mcast_n_groups).
Add a test with four test cases for each option, to ensure that VLANs with
different values are not grouped into ranges and VLANs with matching
values are properly grouped together.
====================
Link: https://patch.msgid.link/20260225143956.3995415-1-danieller@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a new test file bridge_vlan_dump.sh with four test cases that verify
VLANs with different per-VLAN options are not incorrectly grouped into
ranges in the dump output.
The tests verify the kernel's br_vlan_opts_eq_range() function correctly
prevents VLAN range grouping when neigh_suppress, mcast_max_groups,
mcast_n_groups, or mcast_enabled options differ.
Each test verifies that VLANs with different option values appear as
individual entries rather than ranges, and that VLANs with matching
values are properly grouped together.
Example output:
$ ./bridge_vlan_dump.sh
TEST: VLAN range grouping with neigh_suppress [ OK ]
TEST: VLAN range grouping with mcast_max_groups [ OK ]
TEST: VLAN range grouping with mcast_n_groups [ OK ]
TEST: VLAN range grouping with mcast_enabled [ OK ]
Signed-off-by: Danielle Ratson <danieller@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Link: https://patch.msgid.link/20260225143956.3995415-3-danieller@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The br_vlan_opts_eq_range() function determines if consecutive VLANs can
be grouped together in a range for compact netlink notifications. It
currently checks state, tunnel info, and multicast router configuration,
but misses two categories of per-VLAN options that affect the output:
1. User-visible priv_flags (neigh_suppress, mcast_enabled)
2. Port multicast context (mcast_max_groups, mcast_n_groups)
When VLANs have different settings for these options, they are incorrectly
grouped into ranges, causing netlink notifications to report only one
VLAN's settings for the entire range.
Fix by checking priv_flags equality, but only for flags that affect netlink
output (BR_VLFLAG_NEIGH_SUPPRESS_ENABLED and BR_VLFLAG_MCAST_ENABLED),
and comparing multicast context (mcast_max_groups and mcast_n_groups).
Example showing the bugs before the fix:
$ bridge vlan set vid 10 dev dummy1 neigh_suppress on
$ bridge vlan set vid 11 dev dummy1 neigh_suppress off
$ bridge -d vlan show dev dummy1
port vlan-id
dummy1 10-11
... neigh_suppress on
$ bridge vlan set vid 10 dev dummy1 mcast_max_groups 100
$ bridge vlan set vid 11 dev dummy1 mcast_max_groups 200
$ bridge -d vlan show dev dummy1
port vlan-id
dummy1 10-11
... mcast_max_groups 100
After the fix, VLANs 10 and 11 are shown as separate entries with their
correct individual settings.
Fixes: a1aee20d5d ("net: bridge: Add netlink knobs for number / maximum MDB entries")
Fixes: 83f6d60079 ("bridge: vlan: Allow setting VLAN neighbor suppression state")
Signed-off-by: Danielle Ratson <danieller@nvidia.com>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Acked-by: Nikolay Aleksandrov <razor@blackwall.org>
Link: https://patch.msgid.link/20260225143956.3995415-2-danieller@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pull networking fixes from Paolo Abeni:
"Including fixes from IPsec, Bluetooth and netfilter
Current release - regressions:
- wifi: fix dev_alloc_name() return value check
- rds: fix recursive lock in rds_tcp_conn_slots_available
Current release - new code bugs:
- vsock: lock down child_ns_mode as write-once
Previous releases - regressions:
- core:
- do not pass flow_id to set_rps_cpu()
- consume xmit errors of GSO frames
- netconsole: avoid OOB reads, msg is not nul-terminated
- netfilter: h323: fix OOB read in decode_choice()
- tcp: re-enable acceptance of FIN packets when RWIN is 0
- udplite: fix null-ptr-deref in __udp_enqueue_schedule_skb().
- wifi: brcmfmac: fix potential kernel oops when probe fails
- phy: register phy led_triggers during probe to avoid AB-BA deadlock
- eth:
- bnxt_en: fix deleting of Ntuple filters
- wan: farsync: fix use-after-free bugs caused by unfinished tasklets
- xscale: check for PTP support properly
Previous releases - always broken:
- tcp: fix potential race in tcp_v6_syn_recv_sock()
- kcm: fix zero-frag skb in frag_list on partial sendmsg error
- xfrm:
- fix race condition in espintcp_close()
- always flush state and policy upon NETDEV_UNREGISTER event
- bluetooth:
- purge error queues in socket destructors
- fix response to L2CAP_ECRED_CONN_REQ
- eth:
- mlx5:
- fix circular locking dependency in dump
- fix "scheduling while atomic" in IPsec MAC address query
- gve: fix incorrect buffer cleanup for QPL
- team: avoid NETDEV_CHANGEMTU event when unregistering slave
- usb: validate USB endpoints"
* tag 'net-7.0-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (72 commits)
netfilter: nf_conntrack_h323: fix OOB read in decode_choice()
dpaa2-switch: validate num_ifs to prevent out-of-bounds write
net: consume xmit errors of GSO frames
vsock: document write-once behavior of the child_ns_mode sysctl
vsock: lock down child_ns_mode as write-once
selftests/vsock: change tests to respect write-once child ns mode
net/mlx5e: Fix "scheduling while atomic" in IPsec MAC address query
net/mlx5: Fix missing devlink lock in SRIOV enable error path
net/mlx5: E-switch, Clear legacy flag when moving to switchdev
net/mlx5: LAG, disable MPESW in lag_disable_change()
net/mlx5: DR, Fix circular locking dependency in dump
selftests: team: Add a reference count leak test
team: avoid NETDEV_CHANGEMTU event when unregistering slave
net: mana: Fix double destroy_workqueue on service rescan PCI path
MAINTAINERS: Update maintainer entry for QUALCOMM ETHQOS ETHERNET DRIVER
dpll: zl3073x: Remove redundant cleanup in devm_dpll_init()
selftests/net: packetdrill: Verify acceptance of FIN packets when RWIN is 0
tcp: re-enable acceptance of FIN packets when RWIN is 0
vsock: Use container_of() to get net namespace in sysctl handlers
net: usb: kaweth: validate USB endpoints
...
In decode_choice(), the boundary check before get_len() uses the
variable `len`, which is still 0 from its initialization at the top of
the function:
unsigned int type, ext, len = 0;
...
if (ext || (son->attr & OPEN)) {
BYTE_ALIGN(bs);
if (nf_h323_error_boundary(bs, len, 0)) /* len is 0 here */
return H323_ERROR_BOUND;
len = get_len(bs); /* OOB read */
When the bitstream is exactly consumed (bs->cur == bs->end), the check
nf_h323_error_boundary(bs, 0, 0) evaluates to (bs->cur + 0 > bs->end),
which is false. The subsequent get_len() call then dereferences
*bs->cur++, reading 1 byte past the end of the buffer. If that byte
has bit 7 set, get_len() reads a second byte as well.
This can be triggered remotely by sending a crafted Q.931 SETUP message
with a User-User Information Element containing exactly 2 bytes of
PER-encoded data ({0x08, 0x00}) to port 1720 through a firewall with
the nf_conntrack_h323 helper active. The decoder fully consumes the
PER buffer before reaching this code path, resulting in a 1-2 byte
heap-buffer-overflow read confirmed by AddressSanitizer.
Fix this by checking for 2 bytes (the maximum that get_len() may read)
instead of the uninitialized `len`. This matches the pattern used at
every other get_len() call site in the same file, where the caller
checks for 2 bytes of available data before calling get_len().
Fixes: ec8a8f3c31 ("netfilter: nf_ct_h323: Extend nf_h323_error_boundary to work on bits as well")
Signed-off-by: Vahagn Vardanian <vahagn@redrays.io>
Signed-off-by: Florian Westphal <fw@strlen.de>
Link: https://patch.msgid.link/20260225130619.1248-2-fw@strlen.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The driver obtains sw_attr.num_ifs from firmware via dpsw_get_attributes()
but never validates it against DPSW_MAX_IF (64). This value controls
iteration in dpaa2_switch_fdb_get_flood_cfg(), which writes port indices
into the fixed-size cfg->if_id[DPSW_MAX_IF] array. When firmware reports
num_ifs >= 64, the loop can write past the array bounds.
Add a bound check for num_ifs in dpaa2_switch_init().
dpaa2_switch_fdb_get_flood_cfg() appends the control interface (port
num_ifs) after all matched ports. When num_ifs == DPSW_MAX_IF and all
ports match the flood filter, the loop fills all 64 slots and the control
interface write overflows by one entry.
The check uses >= because num_ifs == DPSW_MAX_IF is also functionally
broken.
build_if_id_bitmap() silently drops any ID >= 64:
if (id[i] < DPSW_MAX_IF)
bmap[id[i] / 64] |= ...
Fixes: 539dda3c5d ("staging: dpaa2-switch: properly setup switching domains")
Signed-off-by: Junrui Luo <moonafterrain@outlook.com>
Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/SYBPR01MB78812B47B7F0470B617C408AAF74A@SYBPR01MB7881.ausprd01.prod.outlook.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
udpgro_frglist.sh and udpgro_bench.sh are the flakiest tests
currently in NIPA. They fail in the same exact way, TCP GRO
test stalls occasionally and the test gets killed after 10min.
These tests use veth to simulate GRO. They attach a trivial
("return XDP_PASS;") XDP program to the veth to force TSO off
and NAPI on.
Digging into the failure mode we can see that the connection
is completely stuck after a burst of drops. The sender's snd_nxt
is at sequence number N [1], but the receiver claims to have
received (rcv_nxt) up to N + 3 * MSS [2]. Last piece of the puzzle
is that senders rtx queue is not empty (let's say the block in
the rtx queue is at sequence number N - 4 * MSS [3]).
In this state, sender sends a retransmission from the rtx queue
with a single segment, and sequence numbers N-4*MSS:N-3*MSS [3].
Receiver sees it and responds with an ACK all the way up to
N + 3 * MSS [2]. But sender will reject this ack as TCP_ACK_UNSENT_DATA
because it has no recollection of ever sending data that far out [1].
And we are stuck.
The root cause is the mess of the xmit return codes. veth returns
an error when it can't xmit a frame. We end up with a loss event
like this:
-------------------------------------------------
| GSO super frame 1 | GSO super frame 2 |
|-----------------------------------------------|
| seg | seg | seg | seg | seg | seg | seg | seg |
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
-------------------------------------------------
x ok ok <ok>| ok ok ok <x>
\\
snd_nxt
"x" means packet lost by veth, and "ok" means it went thru.
Since veth has TSO disabled in this test it sees individual segments.
Segment 1 is on the retransmit queue and will be resent.
So why did the sender not advance snd_nxt even tho it clearly did
send up to seg 8? tcp_write_xmit() interprets the return code
from the core to mean that data has not been sent at all. Since
TCP deals with GSO super frames, not individual segment the crux
of the problem is that loss of a single segment can be interpreted
as loss of all. TCP only sees the last return code for the last
segment of the GSO frame (in <> brackets in the diagram above).
Of course for the problem to occur we need a setup or a device
without a Qdisc. Otherwise Qdisc layer disconnects the protocol
layer from the device errors completely.
We have multiple ways to fix this.
1) make veth not return an error when it lost a packet.
While this is what I think we did in the past, the issue keeps
reappearing and it's annoying to debug. The game of whack
a mole is not great.
2) fix the damn return codes
We only talk about NETDEV_TX_OK and NETDEV_TX_BUSY in the
documentation, so maybe we should make the return code from
ndo_start_xmit() a boolean. I like that the most, but perhaps
some ancient, not-really-networking protocol would suffer.
3) make TCP ignore the errors
It is not entirely clear to me what benefit TCP gets from
interpreting the result of ip_queue_xmit()? Specifically once
the connection is established and we're pushing data - packet
loss is just packet loss?
4) this fix
Ignore the rc in the Qdisc-less+GSO case, since it's unreliable.
We already always return OK in the TCQ_F_CAN_BYPASS case.
In the Qdisc-less case let's be a bit more conservative and only
mask the GSO errors. This path is taken by non-IP-"networks"
like CAN, MCTP etc, so we could regress some ancient thing.
This is the simplest, but also maybe the hackiest fix?
Similar fix has been proposed by Eric in the past but never committed
because original reporter was working with an OOT driver and wasn't
providing feedback (see Link).
Link: https://lore.kernel.org/CANn89iJcLepEin7EtBETrZ36bjoD9LrR=k4cfwWh046GB+4f9A@mail.gmail.com
Fixes: 1f59533f9c ("qdisc: validate frames going through the direct_xmit path")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260223235100.108939-1-kuba@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
