| Age | Commit message (Collapse) | Author |
|---|
|
macb_pad_and_fcs() appends the FCS in place when the skb has tailroom.
A shared skb, as pktgen sends in clone_skb mode, grows by one FCS per
transmit. BQL then completes more bytes than were queued and
dql_completed() hits its BUG_ON.
On a Raspberry Pi CM5 (RP1 GEM) pktgen with clone_skb 1000 burst 32 at
60 bytes kills the box within seconds.
Copy shared skbs before appending the FCS. Clearing IFF_TX_SKB_SHARING
would also fix it but makes pktgen refuse clone_skb on macb.
Fixes: 653e92a9175e ("net: macb: add support for padding and fcs computation")
Signed-off-by: Nicolai Buchwitz <nb@tipi-net.de>
Link: https://patch.msgid.link/20261006-nb-macb-shared-skb-net-v1-2-a80641479041@tipi-net.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
macb_pad_and_fcs() replaces or extends the skb before the ring space
check. On NETDEV_TX_BUSY the stack requeues an skb that is already freed
or grown.
Check the ring first, using the padded length for the descriptor count.
Nonlinear skbs always take the copy path so the count can assume a
linear skb.
Fixes: 653e92a9175e ("net: macb: add support for padding and fcs computation")
Signed-off-by: Nicolai Buchwitz <nb@tipi-net.de>
Link: https://patch.msgid.link/20261006-nb-macb-shared-skb-net-v1-1-a80641479041@tipi-net.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
sparx5_tc_matchall_replace() allocates a struct sparx5_mall_entry for
every offloaded matchall filter and adds it to sparx5->mall_entries.
sparx5_tc_matchall_destroy() removes the entry from the list, but never
frees it, so the entry of every deleted mirror and goto matchall filter
is leaked.
Free the entry after unlinking it.
The leak was discovered by an AI code review agent, and reproduced with
kmemleak on a lan969x EV board (EV23X71A) by repeatedly adding and
deleting matchall mirror and goto filters. With the fix, kmemleak no
longer reports the leak.
Cc: stable@vger.kernel.org
Fixes: 1ede4acf045c ("net: sparx5: add bookkeeping code for matchall rules")
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20261007-sparx5-matchall-kfree-net-v1-1-c8918685b337@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
mlxsw_sp_flower_tmplt_create() parses a flow_cls_offload template
into a stack-local struct mlxsw_sp_acl_rule_info purely to compute
rulei.values.elusage. Parsing can acquire port range registers via
mlxsw_sp_flower_parse_ports_range(), but since this rulei never goes
through mlxsw_sp_acl_rulei_destroy(), those registers were never
released, including through chain template deletion.
Factor out of mlxsw_sp_acl_rulei_destroy() the code to actually
release the necessary resources and call from
mlxsw_sp_flower_tmplt_create() to plug the leak.
The issue was found during a review of Wentao Liang's patch referenced
below.
Fixes: fe22f7410527 ("mlxsw: spectrum_flower: Add ability to match on port ranges")
Reported-by: Wentao Liang <vulab@iscas.ac.cn>
Closes: https://lore.kernel.org/netdev/20260917113236.2149095-1-vulab@iscas.ac.cn/
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://patch.msgid.link/95339cf970ce78e1a74aa86ffcc104ecf7c7762c.1791294384.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
mlx5e_poll_ico_cq() requires sq->cc to be updated only after
mlx5_cqwq_update_db_record(), otherwise a CQ overrun may occur.
The current implementation updates sq->cc before the CQ doorbell
record, violating this ordering requirement.
Update the CQ doorbell record first and use dma_wmb() before updating
sq->cc. This ensures that the CQ space is released to the device
before the corresponding ICOSQ consumer index is updated by software.
Fixes: fd9b4be8002c ("net/mlx5e: RX, Support multiple outstanding UMR posts")
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Reviewed-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20261006105820.257208-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Commit 730ff06d3f5c ("net: mana: Use page pool fragments for RX buffers
instead of full pages to improve memory efficiency.") started handing
out RX buffers with zero headroom so that two buffers fit into one page
at the default MTU.
The MANA TX path, however, stores the per scatter-gather entry DMA
mappings in `struct mana_skb_head` at skb->head, and mana_start_xmit()
therefore calls skb_cow_head(skb, MANA_HEADROOM). The port advertises
this requirement as ndev->needed_headroom = MANA_HEADROOM.
As a result every packet that is received and then forwarded out of a
MANA port fails the skb_cow() in ip_forward() and gets reallocated and
copied by pskb_expand_head(). This is invisible to a plain RX or TX
workload, but it puts a full skb reallocation plus memcpy on the hot
path of every single forwarded packet, which is exactly what a
router/NVA workload does.
Restore the headroom. Note that reserving MANA_HEADROOM (232) is not
enough: ip_forward() asks for LL_RESERVED_SPACE(dev), which rounds
hard_header_len + needed_headroom up to HH_DATA_MOD and is 256 bytes on
ethernet. Use LL_RESERVED_SPACE() directly so the value keeps tracking
both constants. Also, since LL_RESERVED_SPACE() tracks MANA_HEADROOM,
it grows with MAX_SKB_FRAGS and for MAX_SKB_FRAGS >= 19 it is greater
than 256, so we have to account for that by using the headroom the RX
queue actually uses (instead of assuming XDP_PACKET_HEADROOM) and
turning MANA_XDP_MTU_MAX into MANA_XDP_MTU_MAX(ndev) (note that at the
default CONFIG_MAX_SKB_FRAGS=17 they are equivalent).
At the default MTU on a 4K page this means a buffer no longer fits twice
into a page (SKB_DATA_ALIGN(1500 + MANA_RXBUF_PAD + 256) = 2112), so the
frag-vs-single decision is now made by computing the real buffer size
instead of comparing the MTU against PAGE_SIZE / 2. The page_pool
fragment path is still used wherever at least two buffers genuinely fit,
e.g. on 16K and 64K page sizes.
Measured on an Azure VM with a MANA NIC acting as a forwarding NVA (UDP,
1400 byte payload, 4 streams, 8 Gbps offered, only the forwarding
node's kernel differs), 8 runs each, median:
forwarded pps throughput
before 272,830 3.06 Gbps
after 390,560 4.37 Gbps (+43%)
perf on the forwarding node, same workload:
memset_orig __pi_memcpy pskb_expand_head
before 10.07% 3.96% present
after 0.94% 0.64% gone
Cc: stable@vger.kernel.org
Fixes: 730ff06d3f5c ("net: mana: Use page pool fragments for RX buffers instead of full pages to improve memory efficiency.")
Signed-off-by: Hamza Mahfooz <hamzamahfooz@linux.microsoft.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20261003013647.2051416-1-hamzamahfooz@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Javier reports seeing packet loss issues related to the re-enablement
of K1; disabling it makes the issues go away. Add the reported system
to have K1 disabled by default.
Reproduction steps (from the Link):
Plug the cable and ping the default gateway on the LAN. No suspend/resume
involved, machine freshly booted and on AC power. Nothing e1000e-related in
dmesg apart from the link up/down messages; no "Hardware Unit Hang".
Default (K1 enabled):
$ ping -c 30 192.168.10.1
--- 192.168.10.1 ping statistics ---
30 packets transmitted, 18 received, 40% packet loss, time 29734ms
rtt min/avg/max/mdev = 0.211/0.315/0.384/0.046 ms
The lost packets are spread over the run (seq 6, 8, 9, 11, 13, 14, 16, 19,
20, 24, 25, 30), not a single burst. DHCP on this link also takes 1-2
minutes to get a lease, consistent with incoming packets being dropped.
With K1 disabled at runtime:
# ethtool --set-priv-flags enp0s31f6 disable-k1 on
$ ethtool --show-priv-flags enp0s31f6
Private flags for enp0s31f6:
s0ix-enabled: on
disable-k1 : on
$ ping -c 30 192.168.10.1
--- 192.168.10.1 ping statistics ---
30 packets transmitted, 30 received, 0% packet loss, time 29713ms
rtt min/avg/max/mdev = 0.169/0.251/0.468/0.054 ms
$ ping -c 30 192.168.10.1
--- 192.168.10.1 ping statistics ---
30 packets transmitted, 30 received, 0% packet loss, time 29677ms
rtt min/avg/max/mdev = 0.101/0.222/0.360/0.064 ms
I have only compile tested this, but Javier has tested that disabling K1
on his system (Lenovo ThinkPad P14s Gen 5 (Intel), I219-LM [8086:550a])
resolves his issues.
Cc: stable@vger.kernel.org
Fixes: 578294b8b60d ("e1000e: Reconfigure PLL clock gate timeout and re-enable K1 on Meteor Lake")
Reported-by: Javier Herrera <javier.herrera@afronta.com>
Link: https://lore.kernel.org/intel-wired-lan/CAKRuCQLWEcci9hPgNUmvq44Bd2H30T5_vP=exMjsknXMN5Yh9A@mail.gmail.com/
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20261001222443.3500206-7-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Stefan Fleischmann reported Intel IOMMU DMA Read faults on BCM57412
NetXtreme-E NICs when transmitting packets on VLAN/macvlan interfaces:
DMAR: [DMA Read NO_PASID] Request device [18:00.0] fault addr 0xfc499000
[fault reason 0x06] PTE Read access is not set
bnxt_en 0000:18:00.0 eno1np0: Abandoning msg {0xb4 0x41a} len: 0 due to firmware status: 0x2000001
...
NETDEV WATCHDOG: eno1np0 (bnxt_en): transmit queue 0 timed out
The fault address (0xfc499000) is on an exact 4KB page boundary,
pointing to a DMA read buffer overrun.
In bnxt_start_xmit(), packets smaller than BNXT_MIN_PKT_SIZE (52 bytes),
such as 42-byte untagged ARP frames, are padded:
if (length < BNXT_MIN_PKT_SIZE) {
pad = BNXT_MIN_PKT_SIZE - length;
if (skb_pad(skb, pad))
goto tx_kick_pending;
length = BNXT_MIN_PKT_SIZE;
}
mapping = dma_map_single(&pdev->dev, skb->data, len, DMA_TO_DEVICE);
...
dma_unmap_len_set(tx_buf, len, len);
However, 'len' was initialized earlier to skb_headlen(skb) (e.g. 42 bytes)
and is left unadjusted after padding. Consequently, dma_map_single() and
dma_unmap_len_set() map and track only 42 bytes.
Later, the hardware TX buffer descriptor is programmed with the padded length:
txbd->tx_bd_len_flags_type =
cpu_to_le32(((len + pad) << TX_BD_LEN_SHIFT) | flags |
TX_BD_FLAGS_PACKET_END);
The NIC DMA engine is thus instructed to read 52 bytes from a region where
only 42 bytes were DMA-mapped. If skb->data ends near the boundary of a 4KB
page (within 'pad' bytes of the next page), the hardware DMA read overruns
into the unmapped adjacent page, triggering an IOMMU fault.
This issue was exposed after commit 447cbe95ebb9 ("vlan: fix skb_under_panic
and races when toggling HW VLAN offload") because reserving extra VLAN
headroom rounded LL_RESERVED_SPACE from 48 up to 64 bytes, shifting skb->data
offsets and potentially causing small frames to land right against page
boundaries.
Fix this by using skb_put_padto(skb, BNXT_MIN_PKT_SIZE) in
bnxt_start_xmit(), before skb_shinfo(skb)->nr_frags is sampled,
because padding might linearize the skb.
This ensures skb->len and skb_headlen(skb) reflect the padded size so
that dma_map_single() maps the full buffer and the descriptor length is
consistent. This also removes the temporary 'pad' variable and masking logic.
The padding is done after the SW USO branch, otherwise
bnxt_sw_udp_gso_xmit() would account the padding as UDP payload
and send it in extra segments.
Note that SW USO segments are still not padded: with IPv4, a segment
carrying less than 10 bytes of UDP payload (small gso_size, or a short
last segment) is sent as a frame shorter than BNXT_MIN_PKT_SIZE.
This is a separate issue, left for a followup patch.
Fixes: c0c050c58d84 ("bnxt_en: New Broadcom ethernet driver.")
Cc: stable@vger.kernel.org
Reported-by: Stefan Fleischmann <sfle@kth.se>
Closes: https://lore.kernel.org/netdev/20261004122616.56714cbd@nargothrond/
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20261007051747.455273-1-edumazet@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Currently the driver zeroes the BARs only when fatal PCIe errors
are reported so that pci_restore_state() restores it. However
firmware handles both fatal and non-fatal errors the same way when
it sees the slot reset resulting from the PCI_ERS_RESULT_NEED_RESET
return code from the driver. This means that we must re-write the
BARs post recovery even during non-fatal errors. Otherwise we will
see that every MMIO access returns all-ones and the firmware appears
dead.
Zero-out the BARs during PCIe error recovery regardless of type of
PCIe error, and make the wait after the hot reset unconditional.
Disable memory decode and bus mastering before rewriting the BARs so
the device doesn't decode a half-updated address, bailing out if
config space is still inaccessible. Defer pci_enable_device() until
after the BAR rewrite and restore, so the device isn't re-enabled
while its BARs are still being rewritten, then re-enable the device
and re-assert bus mastering. Guard the same Command register cleanup
on the re-enable failure path against an inaccessible device.
Skip re-enabling the device in bnxt_io_slot_reset() if it is already
enabled, so enable_cnt does not go unbalanced. A concurrent
bnxt_fw_reset_task() can also be re-enabling the same device in its
ENABLE_DEV state, so guard that call the same way.
Fixes: f75d9a0aa967 ("bnxt_en: Re-write PCI BARs after PCI fatal error.")
Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Reviewed-by: Scott Branden <scott.branden@broadcom.com>
Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://patch.msgid.link/20261005204246.3822563-4-michael.chan@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Fix and strengthen the FLR sequence when initializing in the kdump
kernel. If the NIC is behind a PCIe switch in synthetic (smart)
mode, the switch may need to see that the BARs have been initialized
before it will pass mem read/write TLPs to the NIC.
On a Dell system with a PEX89144 PCIe switch, echo c > /proc/sysrq-trigger
will trigger fatal AER without this patch:
bnxt_en 0000:67:00.0: enabling device (0000 -> 0002)
[Hardware Error]: Hardware error from APEI Generic Hardware Error Source: 5
[Hardware Error]: event severity: recoverable
[Hardware Error]: Error 0, type: fatal
[Hardware Error]: section_type: PCIe error
[Hardware Error]: port_type: 5, upstream switch port
...
Add a new bnxt_kdump_reset() to do the expanded FLR sequence in the
kdump kernel. We now disable bus master and memory, save the PCI
state, do the FLR, clear the BARs, and restore the PCI state. The
BARs have to be cleared to ensure that they get re-initialized and
visible to the PCIe switch.
Since it is the kdump kernel, we make every effort to continue in
the best possible way even if pci_save_state() or pcie_flr() returns
error. After FLR, we poll for an additional 5 seconds before
aborting in case the device is not properly returning CRS. This is
similar to the 5-second wait in bnxt_io_slot_reset().
Fixes: 8743db4a9acf ("bnxt_en: Issue PCIe FLR in kdump kernel to cleanup pending DMAs.")
Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://patch.msgid.link/20261005204246.3822563-3-michael.chan@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
In bnxt_io_slot_reset(), we clear the 6 BAR registers. Add a helper
function to do that. The helper will be used again in the next patch.
Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Reviewed-by: Joe Damato <joe@dama.to>
Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://patch.msgid.link/20261005204246.3822563-2-michael.chan@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When the revised suspend/resume sequence was introduced this led to an edge
case where the TX is left disabled in the following sequence.
1. phy link is down, so UMAC is held in reset and then network interface
is WoL enabled
2. Enter suspend, bcmgenet_wol_power_down_cfg() enables UMAC_RX since MAC
is in SW_RESET
4. Enter resume, UMAC_RX is left enabled. Since we only enable UMAC_TX
and UMAC_RX in SW_RESET. The UMAC_TX is never enabled again on link up.
The issue was hit in production and I verified the sequence keeps
the TX disabled in my standalone test. I tested on an internal
development board with bcm77122, but should be reproducible
on any genet HW that supports power management.
Fixes: 254f3239dd07 ("net: bcmgenet: revise suspend/resume")
Signed-off-by: Justin Chen <justin.chen@broadcom.com>
Reviewed-by: Florian Fainelli <florian.fainelli@broadcom.com>
Reviewed-by: Nicolai Buchwitz <nb@tipi-net.de>
Link: https://patch.msgid.link/20260930230634.3132729-1-justin.chen@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
This reverts commit ad0ae7aefa7a ("net/mlx5: E-Switch, preserve max tx
speed on vport state modification").
mlx5_modify_vport_admin_state() and mlx5_esw_adj_vport_modify() query
the vport's current max_tx_speed before modifying its state, and
write that value back, to avoid resetting it to 0 as a side effect
of an unrelated admin-state change.
That's unnecessary: max_tx_speed is optional in MODIFY_VPORT_STATE -
FW skips writing it whenever it's 0, treating that as "not provided"
rather than "reset to zero". Leaving it unset already preserves FW's
current value, with no query needed.
Fixes: ad0ae7aefa7a ("net/mlx5: E-Switch, preserve max tx speed on vport state modification")
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20261004083531.216988-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Currently the base non-PTP-capable lan969x variants don't forward any
Ethernet frames:
$ ip link set eth10 up
$ ip addr add 10.0.0.54/24 dev eth10
$ ping 10.0.0.1
PING 10.0.0.1 (10.0.0.1): 56 data bytes
--- 10.0.0.1 ping statistics ---
7 packets transmitted, 0 packets received, 100% packet loss
$ cat /proc/interrupts | grep fdma
20: 0 GIC-0 120 Level sparx5-fdma
$ ip -s link show dev eth10
RX: bytes packets errors dropped missed mcast
20828 102 0 0 0 72
TX: bytes packets errors dropped carrier collsns
0 0 0 0 0 0
Testing showed that starting the domain 0 TOD counter gets networking
working again, but start all three for consistency with the PTP variants.
Fixes: 207966787b71 ("net: sparx5: add feature support")
Suggested-by: Daniel Machon <daniel.machon@microchip.com>
Signed-off-by: Quentin Freimanis <quentin@q-lab.dev>
Reviewed-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20261005052826.14251-1-quentin@q-lab.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
This reverts commit 232d49dd4b40a666283de9e722899f088ed581b2.
Anirudh reports a regression from this change.
It is not appropriate for net this late in the release cycle
in the first place, so let's revert and revisit in net-next.
Reported-by: Anirudh Srinivasan <asrinivasan@oss.tenstorrent.com>
Link: https://lore.kernel.org/CAEev2e8+fr1tJnSErD+2TtLDw5fGJbQOMcM5eefMnLyFUR+m7w@mail.gmail.com
Fixes: 232d49dd4b40 ("net: stmmac: propagate PTP addend and system time programming errors")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
CPU accesses QDMA via the bus. When multiple modules are using the bus
simultaneously, CPU access to QDMA may encounter bus timeouts and fails,
resulting in QDMA configuration failures and potentially causing packet
transmission issues. In order to mitigate the issue, introduce a retry
mechanism to airoha_qdma_set_trtcm_param routine in order to ensure the
configuration is correctly applied to the hardware.
It's enough to try a second time for the TRTCM config to be actually
applied to make sure we are not in the middle of bucket handling as it does
operate on fixed time slot internally.
Fixes: ef1ca9271313b ("net: airoha: Add sched HTB offload support")
Signed-off-by: Leto Liu (刘涛) <Leto.Liu@airoha.com>
[ improve commit description, drop memory block, apply better loop logic]
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Link: https://patch.msgid.link/20260930065855.46171-1-ansuelsmth@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When disabling XSK pools, GVE calls the unlocked versions of
napi_disable and napi_enable. However, the netdev lock has already been
acquired before ndo_bpf is called because GVE supports queue management
ops. Calling the unlocked versions of napi_disable/enable results in a
deadlock when attempting to disable XSK pools, as the thread attempts to
re-acquire a lock it already holds.
Update the NAPI calls to use the locked versions.
This issue was found in production.
Fixes: 606048cbd834 ("net: designate XSK pool pointers in queues as "ops protected"")
Cc: stable@vger.kernel.org
Reviewed-by: Harshitha Ramamurthy <hramamurthy@google.com>
Reviewed-by: Jordan Rhee <jordanrhee@google.com>
Signed-off-by: Joshua Washington <joshwash@google.com>
Link: https://patch.msgid.link/20260930214735.3288780-4-joshwash@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When the error bit is set in the RX completion descriptor, the buf_state
and its attached buffer should be freed. In the case of AF_XDP ZC, the
XSK buffer was not freed, leading to a leak.
This issue was caught by LLM.
Fixes: c1fffc5d66a7 ("gve: implement DQO RX datapath and control path for AF_XDP zero-copy")
Cc: stable@vger.kernel.org
Reviewed-by: Jordan Rhee <jordanrhee@google.com>
Reviewed-by: Tim Hostetler <thostet@google.com>
Signed-off-by: Joshua Washington <joshwash@google.com>
Link: https://patch.msgid.link/20260930214735.3288780-3-joshwash@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
GVE does not free XSK buffers when resetting ring state as a part of
stopping queues. This causes all XSK buffers which are posted to the
NIC to be leaked.
Free XSK buffers attached to an allocated buf_state when stopping rings.
This issue was hit in production.
Fixes: c1fffc5d66a7 ("gve: implement DQO RX datapath and control path for AF_XDP zero-copy")
Cc: stable@vger.kernel.org
Reviewed-by: Tim Hostetler <thostet@google.com>
Reviewed-by: Jordan Rhee <jordanrhee@google.com>
Signed-off-by: Joshua Washington <joshwash@google.com>
Link: https://patch.msgid.link/20260930214735.3288780-2-joshwash@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Currently, hardware VLAN stripping is broken for 802.1ad tags. vlan_rx_hw()
hardcodes ETH_P_8021Q when putting the hardware tag into the skb, rather
than using the actual protocol from the packet. Because of this, packets
that contain a 802.1ad outer tag are incorrectly passed up the stack as
having an 802.1Q tag. This causes QinQ ping between two hosts to fail.
vlan_rx_hw() is shared by dwxgmac2 and dwmac4: on dwxgmac2 the tag type
is available in the RDES3 write-back descriptor (the ET_LT field), so the
outer tag type can be determined based on that info. However, dwmac4
doesn't seem to provide the tag type. The Length/Type field in RDES3 only
indicates whether the packet is single or double-tagged, not which tag
type was stripped.
Since dwmac4 cannot report the stripped tag type, it cannot support
hardware S-Tag stripping correctly. Therefore, restrict the
NETIF_F_HW_VLAN_STAG_RX and NETIF_F_HW_VLAN_STAG_FILTER advertisement
to dwxgmac2 only.
With this, 802.1ad tags are left in place and handled by the software
VLAN path.
Fixes: 750011e239a5 ("net: stmmac: Add support for HW-accelerated VLAN stripping")
Signed-off-by: Ovidiu Panait <ovidiu.panait.rb@renesas.com>
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260928203441.34876-6-ovidiu.panait.rb@renesas.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
C-VLAN and S-VLAN tag stripping are both controlled by the EVLS bit,
so disabling rx-vlan-offload also disables S-VLAN tag stripping.
However, rx-vlan-stag-hw-parse keeps being advertised as enabled:
root@rzv2h-evk:~# ethtool -K end1 rx-vlan-offload off
root@rzv2h-evk:~# ethtool -k end1 | grep -i vlan
rx-vlan-offload: off
tx-vlan-offload: off [fixed]
rx-vlan-filter: on [fixed]
vlan-challenged: off [fixed]
tx-vlan-stag-hw-insert: off [fixed]
rx-vlan-stag-hw-parse: on [fixed]
rx-vlan-stag-filter: on [fixed]
Fix this inconsistency by making NETIF_F_HW_VLAN_STAG_RX follow
NETIF_F_HW_VLAN_CTAG_RX.
Fixes: 750011e239a5 ("net: stmmac: Add support for HW-accelerated VLAN stripping")
Signed-off-by: Ovidiu Panait <ovidiu.panait.rb@renesas.com>
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260928203441.34876-5-ovidiu.panait.rb@renesas.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The ESVL and DOVLTC bits control S-VLAN tag processing and have
nothing to do with the double VLAN feature, which only provides a way
to process an additional inner VLAN tag. However, the driver code
that handles them always refers to "double VLAN", which is unrelated
and makes the implementation confusing. The driver does not use any
of the inner VLAN tag features, and the networking core does not
support offloads for the inner tag anyway.
To reduce the confusion regarding S-Tag vs double VLAN handling,
rename double -> svlan.
No functional change intended.
Suggested-by: Joseph Steel <recv.jo@gmail.com>
Signed-off-by: Ovidiu Panait <ovidiu.panait.rb@renesas.com>
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260928203441.34876-4-ovidiu.panait.rb@renesas.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Currently, the double VLAN EDVLP bit is toggled whenever an 802.1ad VLAN
is registered. This bit enables processing of the inner VLAN tag, which
is completely unrelated to S-Tag VLAN handling.
Move EDVLP handling into vlan_set_hw_mode() instead, and keep it always
enabled, so that COE can work for packets with an inner VLAN header.
Add a dedicated callback for dwxlgmac2, as it doesn't implement the
set_hw_vlan_mode callback, like the other cores.
Suggested-by: Joseph Steel <recv.jo@gmail.com>
Signed-off-by: Ovidiu Panait <ovidiu.panait.rb@renesas.com>
Link: https://patch.msgid.link/20260928203441.34876-3-ovidiu.panait.rb@renesas.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
stmmac_vlan_update() falls back to "perfect matching" when the VLAN hash
filter is unavailable (!priv->dma_cap.vlhash). This fallback has been
unreachable in normal operation since its introduction in
commit c7ab0b8088d7 ("net: stmmac: Fallback to VLAN Perfect filtering if
HASH is not available") because the NETIF_F_HW_VLAN_{CTAG,STAG}_FILTER
features are advertised only when priv->dma_cap.vlhash is true.
The fallback is also duplicating the code in vlan_add_hw_rx_fltr(), which
is always available since stmmac_get_num_vlan() returns at least 1.
Therefore, remove it.
Signed-off-by: Ovidiu Panait <ovidiu.panait.rb@renesas.com>
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260928203441.34876-2-ovidiu.panait.rb@renesas.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
mlx5_lag_sum_devices_speed computes the LAG aggregate by summing oper
speeds across all ports. This has two bugs.
First, it relies on the assumption that a port whose carrier is down
will report an oper speed of zero and therefore not contribute to the
sum. This assumption does not always hold: when the link partner
disconnects the port transitions to DOWN state but firmware may still
report a non-zero oper speed, causing the aggregate to include a port
that is not actively carrying traffic.
Second, in active-backup mode only one port transmits at a time, so
the aggregate should reflect a single port speed rather than the sum
of all ports.
Fix this by splitting mlx5_lag_sum_devices_speed into two helpers.
mlx5_lag_get_devices_oper_speed reflects the speed currently available:
it queries the vport state of each port and skips any port that is not
UP, rather than relying on oper speed being zero.
mlx5_lag_get_devices_max_speed is state-independent and returns the
maximum achievable speed used as a fallback when speed is 0; for
active-backup it takes the maximum single-port speed instead of the sum.
The bugs were discovered during integration and reproduced on a
back-to-back ConnectX-8 setup.
The first bug (stale oper speed when link goes down) was observed: after
the link partner disconnected, the aggregate TX speed reported to vports
remained full speed because firmware still returned the last oper speed,
causing vports to be rate-limited to the old aggregate speed.
The second bug (active-backup summing all port speeds instead of the
active port's speed) was also triggered: in active-backup mode the
computed aggregate was double the expected value since both port speeds
were summed even though only one port was active.
Fixes: 28ea6036dad2 ("net/mlx5: Handle port and vport speed change events in MPESW")
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Shay Drori <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260930170406.148548-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue
Tony Nguyen says:
====================
Fix i40e/ice/iavf VF bonding after netdev lock changes
Jose Ignacio Tornos Martinez says:
This series fixes VF bonding failures introduced by commit ad7c7b2172c3
("net: hold netdev instance lock during sysfs operations").
When adding VFs to a bond immediately after setting trust mode, MAC
address changes fail with -EAGAIN, preventing bonding setup. This
affects both i40e (700-series) and ice (800-series) Intel NICs.
The core issue is lock contention: iavf_set_mac() is now called with the
netdev lock held and waits for MAC change completion while holding it.
However, both the watchdog task that sends the request and the adminq_task
that processes PF responses also need this lock, creating a deadlock where
neither can run, causing timeouts.
Additionally, setting VF trust triggers an unnecessary ~10 second VF reset
in i40e driver that delays bonding setup, even though filter
synchronization happens naturally during normal VF operation. For ice
driver, the delay is not so big, but in the same way the operation is not
necessary.
This series:
1. Eliminates unnecessary VF reset when setting trust in i40e (reset only
if revoking trust and VF has advanced features configured).
2. Fixes lock contention by polling admin queue synchronously
3. Eliminates unnecessary VF reset when setting trust in ice, (reset only
if revoking trust and VF has advanced features configured).
The key fix (patch 2/3) implements a synchronous MAC change operation
similar to the approach used for ndo_change_mtu deadlock fix:
https://lore.kernel.org/20260211191855.1532226-1-poros@redhat.com
Instead of scheduling work and waiting, it:
- Sends the virtchnl message directly (not via watchdog)
- Polls the admin queue hardware directly for responses
- Processes all messages inline (including non-MAC messages)
- Returns when complete or times out
This allows the operation to complete synchronously while holding
netdev_lock, without relying on watchdog or adminq_task.
The function can sleep for up to 2.5 seconds polling hardware, but this
is acceptable since netdev_lock is per-device and only serializes
operations on the same interface.
Testing shows VF bonding now works reliably in ~5 seconds vs 15+ seconds
before (i40e), without timeouts or errors (i40e and ice).
Tested on Intel 700-series (i40e) and 800-series (ice) dual-port NICs
with iavf driver.
Thanks to Jan Tluka <jtluka@redhat.com> and Yuying Ma <yuma@redhat.com> for
reporting the issues.
* '40GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue:
ice: skip unnecessary VF reset when setting trust
iavf: send MAC change request synchronously
i40e: skip unnecessary VF reset when setting trust
====================
Link: https://patch.msgid.link/20260928224454.483072-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
ice_eswitch_release_repr() uses metadata_dst_free() to release the
representor's metadata_dst. metadata_dst_free() directly frees the
underlying memory without checking the dst_entry refcount.
When ice_eswitch_port_start_xmit() processes a packet, it takes a
reference via dst_hold() and attaches the metadata_dst to the skb.
If the representor is torn down while packets are still queued on
the lower device (e.g. in a qdisc), the metadata_dst is freed while
references are still held.
Use dst_release() instead, which correctly decrements the refcount
and only frees the object when all references are dropped. The dst
subsystem already handles metadata_dst cleanup in dst_destroy() when
DST_METADATA is set.
Other drivers sharing this pattern (nfp, airoha, bnxt) already use
dst_release() for their metadata_dst lifecycle.
Fixes: f5396b8a663f7 ("ice: switchdev slow path")
Cc: stable@vger.kernel.org
Signed-off-by: Tristan Madani <tristan@talencesecurity.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260928230429.495442-5-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The DQL accounting state which tracks the number of bytes queued for the
NIC to transmit, namely dql->num_queued, is updated by the ICE driver
when it invokes __netdev_tx_sent_queue(). Subsequently the driver
updates the hardware queue tail allowing the NIC to begin processing the
work queue. The queue accounting update should be ordered before the NIC
begins transmitting descriptors.
The transmit path currently updates the hardware doorbell using
writel_relaxed(), which (on arm64) does not provide the same ordering
guarantees between writes to normal memory and writes to MMIO registers
that writel() does. This introduces a potential race where the NIC
begins transmitting descriptors before the dql->num_queued update is
globally visible. If the NIC finishes before the update is observed by
dql_completed(), it detects an invalid state where more bytes have been
completed than have been queued. When this happens the following BUG_ON
is triggered.
BUG_ON(count > num_queued - dql->num_completed);
This has been observed and manifests as the following crash (note that
the trace has been trimmed) in an environment with sustained network
load that uses an Intel Corporation Ethernet Controller E810-XXV for SFP
(rev 02) on an arm64 machine. Replacing writel_relaxed() with writel()
in a test kernel eliminated the crash in the user's workload, which
previously reproduced the issue reliably.
kernel BUG at lib/dynamic_queue_limits.c:99
Internal error: Oops - BUG: 00000000f2000800 [#1] SMP
pc : dql_completed+0x268/0x2a0
lr : ice_clean_tx_irq+0x1d4/0x620 [ice]
Call trace:
dql_completed+0x268/0x2a0 (P)
ice_napi_poll+0x94/0x520 [ice]
__napi_poll+0x48/0x3f0
net_rx_action+0x194/0x420
This restores the behaviour prior to commit ccde82e90946 ("ice: add E830
Earliest TxTime First Offload support"), which changed the notification
mechanism from writel() to writel_relaxed().
Link: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2161572
Fixes: ccde82e90946 ("ice: add E830 Earliest TxTime First Offload support")
Signed-off-by: Bryan Fraschetti <bryan.fraschetti@canonical.com>
Tested-by: Alexander Nowlin <alexander.nowlin@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260928230429.495442-4-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
ice_dealloc_dynamic_port() uses dyn_port->vsi->idx to erase the dynamic
port from pf->dyn_ports. However, it frees the VSI before reading the
index for the erase, resulting in a use-after-free.
Follow the reverse of the allocation order in ice_alloc_dynamic_port()
by erasing the xarray entry before freeing the VSI.
Fixes: eda69d654c7e ("ice: add basic devlink subfunctions support")
Cc: stable@vger.kernel.org
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
Reviewed-by: Marcin Szycik <marcin.szycik@linux.intel.com>
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Tested-by: Patryk Holda <patryk.holda@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260928230429.495442-3-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
In the dmaengine path the driver pre-submits RX buffers and holds
in-flight TX buffers whose SKBs are DMA-mapped by the driver and freed
only in the completion callbacks. On ndo_stop() dmaengine_terminate_sync()
aborts these descriptors without running their callbacks, and the driver
then frees only the ring shells, leaking every SKB still owned by the
engine and its DMA mapping on each ifdown. With 128 RX buffers pre-posted
per channel, the mapping leak can eventually exhaust a limited IOMMU
aperture.
Clear the slot's skb in the TX and RX callbacks so a non-NULL skb marks a
slot that still owns a live, DMA-mapped buffer, and on stop unmap and free
every such buffer.
axienet_dma_rx_cb() runs from the DMA tasklet and re-arms the RX ring on
each completion, so it can race axienet_stop(): a completion may submit a
fresh buffer after dmaengine_terminate_sync() has returned, leaving the
channel armed with a buffer the teardown then frees while the engine may
still write into it (dma_release_channel() does not stop it either). Add a
lock that axienet_dma_rx_cb() holds across the @stopping check and the
resubmit, and axienet_stop() holds to set @stopping before terminating.
Once @stopping is set no callback can arm a new buffer, and any armed just
before is aborted by the terminate, so teardown only frees buffers the
engine no longer owns.
Fixes: 6a91b846af85 ("net: axienet: Introduce dmaengine support")
Cc: stable@vger.kernel.org
Signed-off-by: Suraj Gupta <suraj.gupta2@amd.com>
Link: https://patch.msgid.link/20260928184207.2361931-1-suraj.gupta2@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
dmaengine_desc_free() is designed for DMA_CTRL_REUSE. If DMA_CTRL_REUSE
is not set, dmaengine_desc_free() will return -EPERM.
if (!dmaengine_desc_test_reuse(desc))
return -EPERM;
So calling dmaengine_desc_free() does nothing.
dmaengine_terminate_(a)sync() will free all already allocated descriptors.
Signed-off-by: Frank Li <Frank.Li@nxp.com>
Reviewed-by: Joe Damato <joe@dama.to>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
stmmac_update_subsecond_increment() ignores the error returned by
stmmac_config_addend(), and stmmac_init_tstamp_counter() discards the
addend and system time programming errors, always returning success. A
failure to program the addend (PTP_TCR_TSADDREG) or to initialize the
system time counter (PTP_TCR_TSINIT) is therefore silently swallowed,
leaving the hardware timestamp counter in a non-running or partially
configured state while the driver keeps operating as if timestamping
were up. This matters for TAPRIO/EST offloading, which derives the gate
base time from the hardware timestamp counter.
The same hooks are also called from the PHC callbacks: settime64 and
adjfine drop the error and report success to clock_settime() and
clock_adjtime(), so a dead PTP reference clock goes unnoticed by
ptp4l/phc2sys.
Return error codes from stmmac_update_subsecond_increment(),
stmmac_init_tstamp_counter(), stmmac_dl_ts_coarse_set() and the
settime64/adjfine callbacks instead of silently returning success. On
failure, roll back the partially applied configuration so the hardware
and the driver bookkeeping stay consistent, and report the reason
through the devlink extack. Also guard against a zero sub-second
increment, which would otherwise divide by zero when computing the
addend.
Reset the persistent timestamping state (hwts_tx_en, hwts_rx_en,
tstamp_config, systime_flags and tsfupdt_coarse) when (re)initializing
timestamping, so a failed init does not leave TX/RX timestamping
enabled on a counter that never started.
Fixes: cc4c9001ce31 ("net: stmmac: Switch stmmac_hwtimestamp to generic HW Interface Helpers")
Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@oss.qualcomm.com>
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
chcr_ktls_cpl_act_open_rpl() inserts into tid_list from the receive path
using xa_insert_bh(). The array is shared by the adapter's connections
and initialized with XA_FLAGS_LOCK_BH. However, chcr_ktls_dev_del() and
the chcr_ktls_dev_add() error path use xa_erase(), which takes the XArray
lock without disabling softirqs.
When cleanup runs with softirqs enabled, a receive softirq for another
connection on the same CPU can interrupt the erase and spin on the lock
held by the interrupted task. Use xa_erase_bh() at both sites to match
the insertion path.
Fixes: 65e302a9bd57 ("cxgb4/ch_ktls: Clear resources when pf4 device is removed")
Reported-by: Changyul Lee <lcy8047@gmail.com>
Assisted-by: LLM
Signed-off-by: Sang-Hoon Choi <csh0052@gmail.com>
Reviewed-by: Ayush Sawal <ayush.sawal@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
out_free_buffer_pools reuses open()'s loop index for the TX LTB
walk. After the pool unwind i is -1, so the TX buffers allocated
earlier are leaked. request_irq() failure has the same leak.
The TX-fail goto on that walk also skips dma_unmap of the filter
list (pre-existing since the LTB was added). Rewrite the TX walk
to real_num_tx_queues with a pointer check, and unmap the filter
list before those frees.
Fixes: d926793c1de9 ("ibmveth: Implement multi queue on xmit")
Cc: stable@vger.kernel.org
Signed-off-by: Mingming Cao <mmc@linux.ibm.com>
Reviewed-by: Dave Marquardt <davemarq@linux.ibm.com>
Link: https://patch.msgid.link/d14314c2b10f125f5e578782484ba3cc50564d53.1790357373.git.mmc@linux.ibm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
ibmveth_open() registers the logical LAN, then allocates the RX
buffer pools. A pool failure jumps to out_free_buffer_pools without
h_free_logical_lan(), so PHYP still owns the buffer-list page when
it is unmapped. request_irq() failure already issued the hypercall
before taking the same label.
Move that h_free to the shared post-register unwind so both paths
deregister before the buffer list is unmapped.
Fixes: d43732ce021f ("ibmveth: properly unwind on init errors")
Cc: stable@vger.kernel.org
Signed-off-by: Mingming Cao <mmc@linux.ibm.com>
Reviewed-by: Dave Marquardt <davemarq@linux.ibm.com>
Link: https://patch.msgid.link/7a56fbfb9e0674f011f4e310fbd7dfd8e67c30aa.1790357373.git.mmc@linux.ibm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
mvneta_swbm_add_rx_fragment() sets XDP_FLAGS_FRAGS_PF_MEMALLOC on the
xdp_buff when a fragment page is a pfmemalloc one (page under memory
pressure). The xdp_buff is reused for the next frame, but only the
XDP_FLAGS_HAS_FRAGS bit was cleared at frame start, so the pfmemalloc
bit leaked from one frame into the following ones. mvneta_swbm_build_skb()
propagates the flag to skb->pfmemalloc through xdp_update_skb_frags_info(),
so the skb of a subsequent fragmented frame could be wrongly marked as
pfmemalloc even if none of its pages are under pressure.
Clear all the xdp_buff flags in mvneta_swbm_rx_frame(), which is invoked
for each new frame, instead of just the XDP_FLAGS_HAS_FRAGS bit.
Fixes: ed7a58cb40bd ("net: marvell: rely on xdp_update_skb_shared_info utility routine")
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@oss.qualcomm.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://patch.msgid.link/20260929-mvneta-xdp-clear-frag-fix-v4-1-1e63b25eeed8@oss.qualcomm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
During XSK TX queue teardown, outstanding descriptors are completed
without a CQE. If one requested a TX timestamp, the completion path
passes that NULL CQE to mlx5e_xsk_fill_timestamp(), which dereferences
it.
Return zero when no CQE is available, indicating that teardown did not
produce a TX timestamp.
Fixes: ec706a860eba ("net/mlx5e: Implement AF_XDP TX timestamp and checksum offload")
Signed-off-by: Prathamesh Deshpande <prathameshdeshpande7@gmail.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260926165402.5902-1-prathameshdeshpande7@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Force EEE off on the RTL8168h/8111h (RTL_GIGA_MAC_VER_46) at PHY connect
with phy_disable_eee().
Commit 202fef9bbbf5 ("net: phy: realtek: fix EEE advertisement write on
the internal PHY MMD path") made EEE actually advertise on the generic
Realtek PHY; the write had been a no-op before. On the RTL8168h that
un-masks a latent defect: once EEE negotiates, RX silently stalls after
~7-20 minutes. The carrier stays up, no counter or dmesg moves, and only
"ip link set down/up" recovers it; disabling EEE keeps the link stable.
Root-causing the RTL8168h LPI/RX path needs hardware not available now, so
disable EEE for this version. Use phy_disable_eee() rather than dropping
the version from rtl_supports_eee(): the latter also skips
rtl_enable_tx_lpi(), whose disable branch clears the MAC TX-LPI bits (ERI
0x1b0[1:0]) on link up. rtl_hw_start_8168h_1() does not clear them (the
RTL8402/RTL8106e init does), so a warm reboot from an EEE-active state
could otherwise leave TX-LPI asserted while the PHY no longer negotiates
EEE. Keeping the version EEE-capable preserves that clear path. This also
covers the RTL8168M, which shares RTL_GIGA_MAC_VER_46.
Reported-by: Dennis Piecha <d.piecha@ymail.com>
Closes: https://lore.kernel.org/all/353419280.954808.1790290283530@mail.yahoo.com/
Fixes: 202fef9bbbf5 ("net: phy: realtek: fix EEE advertisement write on the internal PHY MMD path")
Cc: stable@vger.kernel.org
Signed-off-by: Oleksij Rempel <o.rempel@pengutronix.de>
Link: https://patch.msgid.link/20260925071040.2137178-1-o.rempel@pengutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The conversion to the dedicated RSS context operations replaced the
pointer to struct otx2_rss_ctx with an inline u32 ind_tbl[] array.
However, otx2_rss_init() still uses sizeof(*rss->ind_tbl) to set rss_size.
This now yields the size of one u32 (4), rather than the 256 entries in
the indirection table.
The RSS initialization and hardware programming loops use rss_size,
so only four entries are initialized and programmed despite the NIX LF
being allocated a 256-entry table. On an OCTEON CN102 with eight RX
queues, the table contained 0, 1, 2, 3 followed by zeros. Traffic was
concentrated on queue 0 and ethtool -X equal 8 did not correct the table.
Use ARRAY_SIZE() to count the indirection table entries. On the same
hardware, this restores the full table repeating queue numbers 0 through
7, and bidirectional multi-flow traffic increments multiple RX queues.
Runtime testing on an Asterfusion ET2500 (OCTEON CN102 A0).
Fixes: 62e01d8c4170 ("eth: otx2: migrate to the *_rxfh_context ops")
Cc: stable@vger.kernel.org
Closes: https://lore.kernel.org/netdev/361df563-e29e-43bf-8806-24bdfefed33d@nocloud.org/
Signed-off-by: Jari Seppälä <jase@nocloud.org>
Reviewed-by: Ratheesh Kannoth <rkannoth@marvell.com>
Link: https://patch.msgid.link/20260928084711.2462571-1-jase@nocloud.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When the switch uses a random base MAC address it does not mark each
port's address as random:
$ dmesg | grep "MAC addr"
sparx5-switch e00c0000.switch: MAC addr was not set, use random MAC
$ cat /sys/class/net/eth0/addr_assign_type
0
0 is NET_ADDR_PERM, so userspace thinks this is a stable address.
With a random MAC reported as NET_ADDR_PERM, systemd's
MACAddressPolicy=persistent [1] leaves it unchanged, so a DHCP client
gets a new IP address every boot and a static reservation can't be used.
Reporting it as NET_ADDR_RANDOM makes systemd replace it with a stable
address derived from the interface name and machine-id.
Fixes: f3cad2611a77 ("net: sparx5: add hostmode with phylink support")
Signed-off-by: Quentin Freimanis <quentin@q-lab.dev>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20260928045012.30178-1-quentin@q-lab.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The only call outside the KUnit tests is in
sparx5_tc_add_rule_copy(), which passes a fixed key list with no
duplicate entries. Therefore, the potential double-free is not
currently reachable in tree.
However, vcap_filter_rule_keys() is exported and does not require its
key list entries to be unique. If a future or external caller supplies
the same key at another position that the inner loop visits, it calls
list_del() and kfree() a second time on the freed field, leading to a
potential double-free.
Break the inner loop after removing a matching field. The outer safe
list traversal continues to process the remaining rule fields.
Fixes: 465a38a269e9 ("net: microchip: sparx5: Support for copying and modifying rules in the API")
Reported-by: Dan Carpenter <error27@gmail.com>
Closes: https://lore.kernel.org/all/ahs-lCAXXikkaHky@stanley.mountain/
Signed-off-by: Harshit Mogalapalli <harshit.m.mogalapalli@oracle.com>
Reviewed-by: Daniel Machon <daniel.machon@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
sparx5_ptp_init() returns early on the base lan969x variants
because they do not have the SPX5_FEATURE_PTP flag. It also returns early
when no "ptp" interrupt is described. Unbinding the driver then causes a
NULL pointer dereference when cleaning up uninitialized tx_skbs queues:
Unable to handle kernel NULL pointer dereference at virtual address
0000000000000008
Call trace:
skb_queue_purge_reason+0x68/0x120 (P)
sparx5_ptp_deinit+0x64/0xc0
mchp_sparx5_remove+0x38/0x70
platform_remove+0x20/0x30
device_remove+0x4c/0x80
Before per-port tx_skbs were added, a NULL dereference would have happened
when calling ptp_clock_unregister() on the never-registered PTP clocks.
Fix by returning early in sparx5_ptp_deinit() if PTP is not used.
Fixes: 0933bd04047c ("net: sparx5: Add support for ptp clocks")
Signed-off-by: Quentin Freimanis <quentin@q-lab.dev>
Reviewed-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20260928044542.29287-1-quentin@q-lab.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The RX rings and the shared NAPI are carried by an internal "dummy" netdev
(eth->dummy_dev). mtk_probe() creates that dummy device, and adds the
shared tx/rx NAPI to it, only after the MAC netdevs have been registered.
That order has been there since the MT7623 support was added and leaves a
window in which the netdevs are visible to the network core before the
shared NAPI and the dummy device exist.
If anything brings the first netdev up during that window (e.g. netifd
opening the WAN link the instant the interface is registered), the
first-open path runs without them. On NETSYS v2 and later SoCs
mtk_rx_alloc() -> __xdp_rxq_info_reg() then dereferences eth->dummy_dev ==
NULL:
WARNING: CPU: ... Missing net_device from driver
in net/core/xdp.c and mtk_open() fails with -ENODEV. On NETSYS v1 SoCs
(MT7621/MT7622/MT7623/MT7629) there is no page pool, so the same open
instead calls napi_enable() on a napi_struct that netif_napi_add() has not
initialised yet and dereferences a NULL napi->dev.
Allocate the dummy netdev and add the shared tx/rx NAPI to it before the
register_netdev() loop. Both the dummy-allocation failure and a
register_netdev() failure now unwind through mtk_unreg_dev(), which
unregisters the net_device notifiers and the netdevs registered so far,
stopping a concurrently-opened netdev (ndo_stop) and its NAPI before the
dummy carrier is freed.
A netdev that opened during the window can also have queued the
frame-engine reset worker (via the tx watchdog or phylink), so the probe
error path now cancels eth->pending_work before freeing the shared state,
as mtk_cleanup() already does on remove.
Fixes: 656e705243fd ("net-next: mediatek: add support for MT7623 ethernet")
Signed-off-by: Sandeep Haemoon <sandeeph@loljews.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/patch-e52c13380b0768fec63af04c80e4b487@loljews.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The Rx ring priv and index are only set in bcmgenet_init_rx_ring(),
which runs at open. Setting the Rx coalescing parameters on an interface
that was never opened, e.g. "ethtool -C eth0 rx-usecs 50", dereferences
a NULL ring->priv. The oops happens with RTNL held, so networking
stays stuck until a reboot:
Unable to handle kernel paging request at virtual address 0000000000002788
pc : bcmgenet_set_rx_coalesce.isra.0+0x10/0x88
lr : bcmgenet_set_coalesce+0xe8/0x180
Call trace:
bcmgenet_set_rx_coalesce.isra.0+0x10/0x88 (P)
__ethnl_set_coalesce.isra.0+0x490/0x570
ethnl_set_coalesce+0x48/0xc0
ethnl_default_set_doit+0xf4/0x230
genl_family_rcv_msg_doit+0xe8/0x160
genl_rcv_msg+0x220/0x2a0
netlink_rcv_skb+0x68/0x140
genl_rcv+0x40/0x60
netlink_unicast+0x338/0x3c0
netlink_sendmsg+0x19c/0x3f8
__sock_sendmsg+0x64/0xc0
__sys_sendto+0x124/0x198
__arm64_sys_sendto+0x30/0x48
Set priv and index at probe. Writing the registers while the interface
is down is harmless because open resets the MAC and
bcmgenet_init_rx_coalesce() reapplies the stored values.
It was observed on a Raspberry Pi CM4, but should apply to all other
genet devices.
Fixes: 9f4ca05827a2 ("net: bcmgenet: Add support for adaptive RX coalescing")
Signed-off-by: Nicolai Buchwitz <nb@tipi-net.de>
Reviewed-by: Florian Fainelli <florian.fainelli@broadcom.com>
Link: https://patch.msgid.link/20260928105642.856210-1-nb@tipi-net.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Since commit 014c607f86ab ("gve: add support for UDP GSO for DQO
format"), gve_prep_tso() matches shinfo->gso_type exactly. gso_type is
a bitmask, though: SKB_GSO_DODGY is set on every GSO packet that comes
from tun/tap, packet sockets or other untrusted sources, and
SKB_GSO_TCP_ECN and SKB_GSO_TCP_FIXEDID can be set as well. Such packets
hit the default case, gve_tx_add_skb_dqo() fails and gve_try_tx_skb()
drops them, counting them in tx_dropped.
These packets do reach the driver: tcp_gso_segment() passes DODGY skbs
through unsegmented to devices that support TSO, after recomputing
gso_segs, so drivers must tolerate the bit.
This breaks virtual machines behind a tap on GCE instances using the DQO
queue formats. Every TSO packet forwarded from a guest (gso_type
SKB_GSO_TCPV4 | SKB_GSO_DODGY) is dropped, and guest uploads slow to a
crawl of retransmissions or stall entirely, while the host's own TSO
traffic (gso_type SKB_GSO_TCPV4) is unaffected. On an n4 instance
(DQO-QPL) with a Cloud Hypervisor guest, a 64 MB upload from the guest
went from a 60 s timeout at ~0.9 MB/s to 0.19 s with this change, with
tx_dropped no longer increasing.
Host TCP with ECN is hit as well: gve advertises NETIF_F_TSO_ECN, so a
TSO packet carrying CWR has SKB_GSO_TCP_ECN set and is dropped too.
Restore the bitmask test that commit 1b9f75634441 ("gve: ignore
nonrelevant GSO type bits when processing TSO headers") introduced for
the same problem, keeping the UDP GSO support.
While here, reload shinfo after skb_cow_head(). If the head was cloned,
pskb_expand_head() moves skb_shared_info to the new head, and the
pointer cached at function entry can then refer to memory that another
clone frees.
Fixes: 014c607f86ab ("gve: add support for UDP GSO for DQO format")
Signed-off-by: Hannu Varjoranta <hannu@varjosoft.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Ankit Garg <nktgrg@google.com>
Reviewed-by: Harshitha Ramamurthy <hramamurthy@google.com>
Link: https://patch.msgid.link/20260925115237.56465-1-hannu@varjosoft.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
When a received frame is larger than dma_buf_sz, the DMA scatters it
across multiple RX descriptors (rx Scatter-Gather). The secondary RX
buffer (sec_page) was only allocated and programmed when split-header
(SPH) was active, so for regular frames buffer2 was neither allocated
nor backed by a valid mapping. As soon as an incoming frame overflowed
buffer1, the DMA wrote the overflow into the unmapped secondary-buffer
address, triggering an SMMU translation fault on IOMMU-based platforms:
arm-smmu 15000000.iommu: Unhandled context fault: fsr=0x402, iova=0x00000000, fsynr=0x7f0011, cbfrsynra=0x1c90, cb=11
arm-smmu 15000000.iommu: FSR = 00000402 [Format=2 TF], SID=0x1c90
arm-smmu 15000000.iommu: FSYNR0 = 007f0011 [S1CBNDX=127 WNR PLVL=1]
Enable scatter-gather for non-SPH frames on cores that can program an
independent secondary RX buffer (GMAC4/XGMAC): allocate and mark buffer2
as valid in stmmac_init_rx_buffers() and stmmac_rx_refill(), and account
for it in the buffer length computation. Legacy cores have no set_sec_addr
op, so they keep buffer2 disabled.
Since buffer2 is handed to the DMA at page offset 0, the page pool sync
window is widened to cover both buffers in every mode: offset is set to 0
and max_len to dma_buf_sz + stmmac_rx_offset().
The FCS can straddle the buffer1/buffer2 boundary and a descriptor
boundary, so it is now stripped from the tail of the assembled frame with
pskb_trim() instead of from a single buffer, avoiding an unsigned underflow
for frames that overflow a buffer by 1..3 bytes. For single-buffer frames
the XDP program must not see the FCS, so it is removed from the XDP buffer
before the program runs.
Native XDP currently only supports single-buffer (linear) frames. An
oversized frame accepted by the MAC (jumbo enabled) is received via
buffer2; the XDP program only sees buffer1, so on a TX/REDIRECT verdict
the frame is forwarded truncated. XDP multi-buffer support to handle
this case is planned as a follow-up.
AF_XDP zero-copy RX is not covered by this change: a ZC queue still
programs buffer2 at DMA address 0 (and XGMAC has no buffer2-valid bit),
so an oversized frame overflowing buffer1 can still trigger the same
SMMU translation fault. Handling is planned as a follow-up.
Fixes: 88ebe2cf7f3f ("net: stmmac: Rework stmmac_rx()")
Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@oss.qualcomm.com>
Link: https://patch.msgid.link/20260923-stmmac-rx-sg-fix-v3-1-ed26fea7180d@oss.qualcomm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The SC resource cleanup and statistics loops were incorrectly
iterating over secy.max instead of sc.max, so the highest SC id
could be left allocated across teardown or FLR. Use sc.max as
the loop bound.
Fixes: cfc14181d497 ("octeontx2-af: cn10k: mcs: Manage the MCS block hardware resources")
Signed-off-by: Baha Mesleh <baha.mesleh@nokia.com>
Signed-off-by: Subrat Pandey <subratp@marvell.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260922102149.2078168-3-subratp@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Previously, BPID assignment under CONFIG_DCB assumed
`queue_to_pfc_map` was always initialized. For SDP VFs this leads to
invalid memory access as it was not initialized.
This patch adds a NULL check for `queue_to_pfc_map` before
dereferencing it. Also, simplifies the logic by always assigning a
default BPID first, then conditionally overriding it if CONFIG_DCB is
enabled and the map exists.
Fixes: 184fb40f731b ("octeontx2-pf: Avoid adding dcbnl_ops for LBK and SDP vf")
Signed-off-by: Geetha sowjanya <gakula@marvell.com>
Signed-off-by: Subrat Pandey <subratp@marvell.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260922102149.2078168-2-subratp@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The mlx5_handle_changeupper_event() and mlx5_handle_changeinfodata_event()
functions check for LAG masters (which includes both bonding and team
devices) but only call bonding-specific APIs. When processing team device
events, calling bond_slave_get_rcu() and bond_is_slave_inactive() on
team_port structures causes KASAN to detect an out-of-bounds memory access
since team_port is smaller than bond slave.
Fix this by wrapping the bond-specific API calls with a check for bonding
devices. This allows the function to still process LAG events for both
bonding and teams but only calls bond-specific functions when dealing with
actual bonding devices.
For mlx5_handle_changeinfodata_event(), keep the explicit bond check as
mlx5 doesn't handle state information for team ports anyway.
Tested with Mellanox ConnectX-5 on Linux 7.2.0-rc6:
- Bonding: PASS (no regressions)
- Team device: PASS (no KASAN errors)
Fixes: 54493a08e21f ("net/mlx5: Lag, record inactive state of bond device")
Suggested-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Anirudh Virdi <avirdi@redhat.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/20260924064244.94045-1-avirdi@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
In UniMAC hardware, there is a 0xC byte gap between the RX MIB counters
and the TX MIB counters, and a second 0xC byte gap between the TX MIB
counters and the RX RUNT MIB counters.
In bcm_sysport_update_mib_counters(), 'offset' was only set to
UMAC_MIB_STAT_OFFSET (0xC) for all non-RX counters, omitting the second
0xC gap for BCM_SYSPORT_STAT_RUNT counters. As a result, all 4 RUNT MIB
counters were read from unmapped gap register space.
Fix this by setting offset to 2 * UMAC_MIB_STAT_OFFSET (0x18) when
reading BCM_SYSPORT_STAT_RUNT counters.
Fixes: 80105befdb4b ("net: systemport: add Broadcom SYSTEMPORT Ethernet MAC driver")
Reviewed-by: Nicolai Buchwitz <nb@tipi-net.de>
Signed-off-by: Florian Fainelli <florian.fainelli@broadcom.com>
Link: https://patch.msgid.link/20260922232440.598918-6-florian.fainelli@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
