The broken reverse-mode, selected by snps,ps-speed, is configured when
the platform provides a valid port speed and a PCS is being used.
Both these remain constant after the driver has probed, so the software
state doesn't need to be re-initialised each time stmmac_hw_setup() is
called (which is called at open and resume time.)
Move the software setup of reverse-mode to stmmac_check_pcs_mode()
which is called from the driver probe function.
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Tested-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Tested-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@bp.renesas.com>
Link: https://patch.msgid.link/E1v9P6Y-0000000Aolr-0vLH@rmk-PC.armlinux.org.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
After a lot of digging, it seems that the oddly named hw->ps member is
all about configuring the core for reverse SGMII. This member is set to
one of 0, SPEED_10, SPEED_100 or SPEED_1000 depending on
priv->plat->mac_port_sel_speed. On DT systems, this comes from the
"snps,ps-speed" DT property.
When set to a non-zero value, it:
1. Configures the MAC at initialisation time to operate at a specific
speed. However, this will be overwritten by mac_link_up() when the
link comes up (e.g. with the fixed-link parameters.)
Note that dwxgmac2 wants to also support SPEED_2500 and SPEED_10000,
but both these values are impossible.
2. It _incorrectly_ enables the transmitter (GMAC_CONFIG_TE) which
makes no sense, rather than enabling the "transmit configuration"
bit (GMAC_CONFIG_TC). Likely a typo.
3. It configures the SGMII rate adapter layer to retrieve its speed
setting from the MAC configuration register rather than the PHY.
There are two ways forward here:
a) fixing (2) so that we set GMAC_CONFIG_TC. However, we have platform
that set the "snps,ps-speed" property and that work today. Fixing
this will cause the RGMII, SGMII or SMII inband configuration to be
transmitted, which will be a functional change which could cause a
regression.
b) ripping out (1) and (2) as they are ineffective. This also has the
possibility of regressions, but the patch author believes this risk
is much lower than (a).
Therefore, this commit takes the approach in (b).
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Tested-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Tested-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@bp.renesas.com>
Link: https://patch.msgid.link/E1v9P6N-0000000Aolg-3y0a@rmk-PC.armlinux.org.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Changing the netif_carrier_*() state behind phylink's back has always
been prohibited because it messes up with phylinks state tracking, and
means that phylink no longer guarantees to call the mac_link_down()
and mac_link_up() methods at the appropriate times. This was later
documented in the sfp-phylink network driver conversion guide.
stmmac was converted to phylink in 2019, but nothing was done with the
"PCS" code. Since then, apart from the updates as part of phylink
development, nothing has happened with stmmac to improve its use of
phylink, or even to address this point.
A couple of years ago, a has_integrated_pcs boolean was added by Bart,
which later became the STMMAC_FLAG_HAS_INTEGRATED_PCS flag, to avoid
manipulating the netif_carrier_*() state. This flag is mis-named,
because whenever the stmmac is synthesized for its native SGMII, TBI
or RTBI interfaces, it has an "integrated PCS". This boolean/flag
actually means "ignore the status from the integrated PCS".
Discussing with Bart, the reasons for this are lost to the winds of
time (which is why we should always document the reasons in the commit
message.)
RGMII also has in-band status, and the dwmac cores and stmmac code
supports this but with one bug that saves the day.
When dwmac cores are synthesised for RGMII only, they do not contain
an integrated PCS, and so priv->dma_cap.pcs is clear, which prevents
(incorrectly) the "RGMII PCS" being used, meaning we don't read the
in-band status. However, a core synthesised for RGMII and also SGMII,
TBI or RTBI will have this capability bit set, thus making these
code paths reachable.
The Jetson Xavier NX uses RGMII mode to talk to its PHY, and removing
the incorrect check for priv->dma_cap.pcs reveals the theortical issue
with netif_carrier_*() manipulation is real:
dwc-eth-dwmac 2490000.ethernet eth0: Register MEM_TYPE_PAGE_POOL RxQ-0
dwc-eth-dwmac 2490000.ethernet eth0: PHY [stmmac-0:00] driver [RTL8211F Gigabit Ethernet] (irq=141)
dwc-eth-dwmac 2490000.ethernet eth0: No Safety Features support found
dwc-eth-dwmac 2490000.ethernet eth0: IEEE 1588-2008 Advanced Timestamp supported
dwc-eth-dwmac 2490000.ethernet eth0: registered PTP clock
dwc-eth-dwmac 2490000.ethernet eth0: configuring for phy/rgmii-id link mode
8021q: adding VLAN 0 to HW filter on device eth0
dwc-eth-dwmac 2490000.ethernet eth0: Adding VLAN ID 0 is not supported
Link is Up - 1000/Full
Link is Down
Link is Up - 1000/Full
This looks good until one realises that the phylink "Link" status
messages are missing, even when the RJ45 cable is reconnected. Nothing
one can do results in the interface working. The interrupt handler
(which prints those "Link is" messages) always wins over phylink's
resolve worker, meaning phylink never calls the mac_link_up() nor
mac_link_down() methods.
eth0 also sees no traffic received, and is unable to obtain a DHCP
address:
3: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP group defa
ult qlen 1000
link/ether e6:d3:6a:e6:92:de brd ff:ff:ff:ff:ff:ff
RX: bytes packets errors dropped overrun mcast
0 0 0 0 0 0
TX: bytes packets errors dropped carrier collsns
27686 149 0 0 0 0
With the STMMAC_FLAG_HAS_INTEGRATED_PCS flag set, which disables the
netif_carrier_*() manipulation then stmmac works normally:
dwc-eth-dwmac 2490000.ethernet eth0: Register MEM_TYPE_PAGE_POOL RxQ-0
dwc-eth-dwmac 2490000.ethernet eth0: PHY [stmmac-0:00] driver [RTL8211F Gigabit Ethernet] (irq=141)
dwc-eth-dwmac 2490000.ethernet eth0: No Safety Features support found
dwc-eth-dwmac 2490000.ethernet eth0: IEEE 1588-2008 Advanced Timestamp supported
dwc-eth-dwmac 2490000.ethernet eth0: registered PTP clock
dwc-eth-dwmac 2490000.ethernet eth0: configuring for phy/rgmii-id link mode
8021q: adding VLAN 0 to HW filter on device eth0
dwc-eth-dwmac 2490000.ethernet eth0: Adding VLAN ID 0 is not supported
Link is Up - 1000/Full
dwc-eth-dwmac 2490000.ethernet eth0: Link is Up - 1Gbps/Full - flow control rx/tx
and packets can be transferred.
This clearly shows that when priv->hw->pcs is set, but
STMMAC_FLAG_HAS_INTEGRATED_PCS is clear, the driver reliably fails.
Discovering whether a platform falls into this is impossible as
parsing all the dtsi and dts files to find out which use the stmmac
driver, whether any of them use RGMII or SGMII and also depends
whether an external interface is being used. The kernel likely
doesn't contain all dts files either.
The only driver that sets this flag uses the qcom,sa8775p-ethqos
compatible, and uses SGMII or 2500BASE-X.
but these are saved from this problem by the incorrect check for
priv->dma_cap.pcs.
So, we have to assume that for every other platform that uses SGMII
with stmmac is using an external PCS.
Moreover, ethtool output can be incorrect. With the full-duplex link
negotiated, ethtool reports:
Speed: 1000Mb/s
Duplex: Half
because with dwmac4, the full-duplex bit is in bit 16 of the status,
priv->xstats.pcs_duplex becomes BIT(16) for full duplex, but the
ethtool ksettings duplex member is u8 - so becomes zero. Moreover,
the supported, advertised and link partner modes are all "not
reported".
Finally, ksettings_set() won't be able to set the advertisement on
a PHY if this PCS code is activated, which is incorrect when SGMII
is used with a PHY.
Thus, remove:
1. the incorrect netif_carrier_*() manipulation.
2. the broken ethtool ksettings code.
Given that all uses of STMMAC_FLAG_HAS_INTEGRATED_PCS are now gone,
remove the flag from stmmac.h and dwmac-qcom-ethqos.c.
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Tested-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Tested-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@bp.renesas.com>
Link: https://patch.msgid.link/E1v9P5y-0000000AolC-1QWH@rmk-PC.armlinux.org.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Correct multiple kernel-doc warnings in nl802154.h:
- Fix a typo on one enum name to avoid a kernel-doc warning.
- Drop 2 enum descriptions that are no longer needed.
- Mark 2 internal enums as "private:" so that kernel-doc is not needed
for them.
Warning: nl802154.h:239 Enum value 'NL802154_CAP_ATTR_MAX_MAXBE' not described in enum 'nl802154_wpan_phy_capability_attr'
Warning: nl802154.h:239 Excess enum value '%NL802154_CAP_ATTR_MIN_CCA_ED_LEVEL' description in 'nl802154_wpan_phy_capability_attr'
Warning: nl802154.h:239 Excess enum value '%NL802154_CAP_ATTR_MAX_CCA_ED_LEVEL' description in 'nl802154_wpan_phy_capability_attr'
Warning: nl802154.h:369 Enum value '__NL802154_CCA_OPT_ATTR_AFTER_LAST' not described in enum 'nl802154_cca_opts'
Warning: nl802154.h:369 Enum value 'NL802154_CCA_OPT_ATTR_MAX' not described in enum 'nl802154_cca_opts'
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20251016035917.1148012-1-rdunlap@infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cpus serving NIC interrupts and specifically TX completions are often
trapped in also restarting a busy qdisc (because qdisc was stopped by BQL
or the driver's own flow control).
When they call netdev_tx_completed_queue() or netif_tx_wake_queue(),
they call __netif_schedule() so that the queue can be run
later from net_tx_action() (involving NET_TX_SOFTIRQ)
Quite often, by the time the cpu reaches net_tx_action(), another cpu
grabbed the qdisc spinlock from __dev_xmit_skb(), and we spend too much
time spinning on this lock.
We can detect in __netif_schedule() if a cpu is already at a specific
point in __dev_xmit_skb() where we have the guarantee the queue will
be run.
This patch gives a 13 % increase of throughput on an IDPF NIC (200Gbit),
32 TX qeues, sending UDP packets of 120 bytes.
This also helps __qdisc_run() to not force a NET_TX_SOFTIRQ
if another thread is waiting in __dev_xmit_skb()
Before:
sar -n DEV 5 5|grep eth1|grep Average
Average: eth1 1496.44 52191462.56 210.00 13369396.90 0.00 0.00 0.00 54.76
After:
sar -n DEV 5 5|grep eth1|grep Average
Average: eth1 1457.88 59363099.96 205.08 15206384.35 0.00 0.00 0.00 62.29
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20251017145334.3016097-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ksz9477_phy_errata function currently assigns the return value of
genphy_restart_aneg() to a variable and then immediately returns it
err = genphy_restart_aneg(phydev);
if (err)
return err;
return err;
This can be simplified by directly returning the function call
result, as the intermediate variable and conditional are redundant.
Signed-off-by: Alok Tiwari <alok.a.tiwari@oracle.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20251017193525.1457064-1-alok.a.tiwari@oracle.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Daniel Golle says:
====================
net: dsa: lantiq_gswip: clean up and improve VLAN handling
This series was developed by Vladimir Oltean to improve and clean up the
VLAN handling logic in the Lantiq GSWIP DSA driver.
As Vladimir currently doesn't have the availability to take care of the
submission process, we agreed that I would send the patches on his
behalf.
The series focuses on consolidating the VLAN management paths for both
VLAN-unaware and VLAN-aware bridges, simplifying internal logic, and
removing legacy or redundant code. It also fixes a number of subtle
inconsistencies regarding VLAN ID 0 handling, bridge FDB entries, and
brings the driver into shape to permit dynamic changes to the VLAN
filtering state.
Notable changes include:
- Support for bridge FDB entries on the CPU port
- Consolidation of gswip_vlan_add_unaware() and gswip_vlan_add_aware()
into a unified implementation
- Removal of legacy VLAN configuration options and redundant
assignments
- Improved handling of VLAN ID 0 and PVID behavior
- Better validation and error reporting in VLAN removal paths
- Support for dynamic VLAN filtering configuration changes
Overall, this refactor improves readability and maintainability of the
Lantiq GSWIP DSA driver. It also results in all local-termination.sh
tests now passing, and slightly improves the results of
bridge-vlan-{un,}aware.sh.
All patches have been authored by Vladimir Oltean; a small unintended
functional change in patch "net: dsa: lantiq_gswip: merge
gswip_vlan_add_unaware() and gswip_vlan_add_aware()" has been ironed out
and some of the commit descriptions were improved by me, apart from that
I'm only handling the submission and will help with follow-up
discussions or review feedback as needed.
Despite the fact that some changes here do actually fix things (in the
sense that selftests which would previously FAIL now PASS) we decided
that it would be the best for this series of patches to go via net-next.
If requested some of it can still be ported to stable kernels later on.
====================
Link: https://patch.msgid.link/cover.1760566491.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
User space can force the altering of VID 0 as it was privately set up by
this driver.
For example, when the port joins a VLAN-aware bridge,
dsa_user_manage_vlan_filtering() will set NETIF_F_HW_VLAN_CTAG_FILTER.
If the port is subsequently brought up and CONFIG_VLAN_8021Q is enabled,
the vlan_vid0_add() function will want to make sure we are capable of
accepting packets tagged with VID 0.
Generally, DSA/switchdev drivers want to suppress that bit of help from
the 8021q layer, and handle VID 0 filters themselves. The 8021q layer
might actually be even detrimential, because VLANs added through
vlan_vid_add() pass through dsa_user_vlan_rx_add_vid(), which is
documented as this:
/* This API only allows programming tagged, non-PVID VIDs */
.flags = 0,
so it will force VID 0 to be reconfigured as egress-tagged, non-PVID.
Whereas the driver configures it as PVID and egress-untagged, the exact
opposite.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Link: https://patch.msgid.link/9f68340c34b5312c3b8c6c7ecf3cfce574a3f65d.1760566491.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This driver doesn't support dynamic VLAN filtering changes, for simplicity.
It expects that on a port, either gswip_vlan_add_unaware() or
gswip_vlan_add_aware() is called, but not both.
When !br_vlan_enabled(), the configure_vlan_while_not_filtering = false
option is exactly what will prevent calls to gswip_port_vlan_add() from
being issued by DSA.
In fact, at the time these features were submitted:
https://patchwork.ozlabs.org/project/netdev/patch/20190501204506.21579-3-hauke@hauke-m.de/
"configure_vlan_while_not_filtering = false" did not even have a name,
it was implicit behaviour. It only became legacy in commit 54a0ed0df4
("net: dsa: provide an option for drivers to always receive bridge
VLANs").
Section "Bridge VLAN filtering" of Documentation/networking/switchdev.rst
describes the exact set of rules. Notably, the PVID of the port must
follow the VLAN awareness state of the bridge port. A VLAN-unaware
bridge port should not respond to the addition of a bridge VLAN with the
PVID flag. In fact, the pvid_change() test in
tools/testing/selftests/net/forwarding/bridge_vlan_unaware.sh tests
exactly this.
The lantiq_gswip driver indeed does not respond to the addition of PVID
VLANs while VLAN-unaware in the way described above, but only because of
configure_vlan_while_not_filtering. Our purpose here is to get rid of
configure_vlan_while_not_filtering, so we must add more complex logic
which follows the VLAN awareness state and walks through the Active VLAN
table entries, to find the index of the PVID register that should be
committed to hardware on each port.
As a side-effect of now having a proper implementation to assign the
PVID all the "VLAN upper: ..." tests of the local_termination.sh self-
tests which would previously all FAIL now all PASS (or XFAIL, but
that's ok).
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Tested-by: Daniel Golle <daniel@makrotopia.org>
Link: https://patch.msgid.link/47dab8a8b69ebb92624b9795b723114475d3fe4e.1760566491.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The two functions largely duplicate functionality. The differences
consist in:
- the "fid" passed to gswip_vlan_active_create(). The unaware variant
always passes -1, the aware variant passes fid = priv->vlans[i].fid,
where i is an index into priv->vlans[] for which priv->vlans[i].bridge
is equal to the given bridge.
- the "vid" is not passed to gswip_vlan_add_unaware(). It is implicitly
GSWIP_VLAN_UNAWARE_PVID (zero).
- The "untagged" is not passed to gswip_vlan_add_unaware(). It is
implicitly true. Also, the CPU port must not be a tag member of the
PVID used for VLAN-unaware bridging.
- The "pvid" is not passed to gswip_vlan_add_unaware(). It is implicitly
true.
- The GSWIP_PCE_DEFPVID(port) register is written by the aware variant
with an "idx", but with a hardcoded 0 by the unaware variant.
Merge the two functions into a single unified function without any
functional changes.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Link: https://patch.msgid.link/2be190701d4c17038ce4b8047f9fb0bdf8abdf6e.1760566491.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, the driver takes the bridge from dsa_port_bridge_dev_get(),
which only works for user ports. This is why it has to ignore FDB
entries installed on the CPU port.
Commit c26933639b ("net: dsa: request drivers to perform FDB
isolation") introduced the possibility of getting the originating bridge
from the passed dsa_db argument, so let's do that instead. This way, we
can act on the local FDB entries coming from the bridge.
Note that we do not expect FDB events for the DSA_DB_PORT database,
because this driver doesn't fulfill the dsa_switch_supports_uc_filtering()
requirements. So we can just return -EOPNOTSUPP and expect it will never
be triggered.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Link: https://patch.msgid.link/ed9d847c0356f0fec81422bdad9ebdcc6a59da79.1760566491.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Martin KaFai Lau says:
====================
pull-request: bpf-next 2025-10-16
We've added 6 non-merge commits during the last 1 day(s) which contain
a total of 18 files changed, 577 insertions(+), 38 deletions(-).
The main changes are:
1) Bypass the global per-protocol memory accounting either by setting
a netns sysctl or using bpf_setsockopt in a bpf program,
from Kuniyuki Iwashima.
* tag 'for-netdev' of https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next:
selftests/bpf: Add test for sk->sk_bypass_prot_mem.
bpf: Introduce SK_BPF_BYPASS_PROT_MEM.
bpf: Support bpf_setsockopt() for BPF_CGROUP_INET_SOCK_CREATE.
net: Introduce net.core.bypass_prot_mem sysctl.
net: Allow opt-out from global protocol memory accounting.
tcp: Save lock_sock() for memcg in inet_csk_accept().
====================
Link: https://patch.msgid.link/20251016204539.773707-1-martin.lau@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Make tcp-md5 use the MD5 library API (added in 6.18) instead of the
crypto_ahash API. This is much simpler and also more efficient:
- The library API just operates on struct md5_ctx. Just allocate this
struct on the stack instead of using a pool of pre-allocated
crypto_ahash and ahash_request objects.
- The library API accepts standard pointers and doesn't require
scatterlists. So, for hashing the headers just use an on-stack buffer
instead of a pool of pre-allocated kmalloc'ed scratch buffers.
- The library API never fails. Therefore, checking for MD5 hashing
errors is no longer necessary. Update tcp_v4_md5_hash_skb(),
tcp_v6_md5_hash_skb(), tcp_v4_md5_hash_hdr(), tcp_v6_md5_hash_hdr(),
tcp_md5_hash_key(), tcp_sock_af_ops::calc_md5_hash, and
tcp_request_sock_ops::calc_md5_hash to return void instead of int.
- The library API provides direct access to the MD5 code, eliminating
unnecessary overhead such as indirect function calls and scatterlist
management. Microbenchmarks of tcp_v4_md5_hash_skb() on x86_64 show a
speedup from 7518 to 7041 cycles (6% fewer) with skb->len == 1440, or
from 1020 to 678 cycles (33% fewer) with skb->len == 140.
Since tcp_sigpool_hash_skb_data() can no longer be used, add a function
tcp_md5_hash_skb_data() which is specialized to MD5. Of course, to the
extent that this duplicates any code, it's well worth it.
To preserve the existing behavior of TCP-MD5 support being disabled when
the kernel is booted with "fips=1", make tcp_md5_do_add() check
fips_enabled itself. Previously it relied on the error from
crypto_alloc_ahash("md5") being bubbled up. I don't know for sure that
this is actually needed, but this preserves the existing behavior.
Tested with bidirectional TCP-MD5, both IPv4 and IPv6, between a kernel
that includes this commit and a kernel that doesn't include this commit.
(Side note: please don't use TCP-MD5! It's cryptographically weak. But
as long as Linux supports it, it might as well be implemented properly.)
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Link: https://patch.msgid.link/20251014215836.115616-1-ebiggers@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Quoting Eric Dumazet:
"I do not understand the fascination with net/core/drop_monitor.c [..]
misses all the features, flexibility, scalability that 'perf',
eBPF tracing, bpftrace, .... have today."
Reword DROP_MONITOR kconfig help text to clearly state that its not
related to perf-based drop monitoring and that its safe to disable
this unless support for the older netlink-based tools is needed.
Signed-off-by: Florian Westphal <fw@strlen.de>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251016115147.18503-1-fw@strlen.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Xuanqiang Luo says:
====================
net: Avoid ehash lookup races
After replacing R/W locks with RCU in commit 3ab5aee7fe ("net: Convert
TCP & DCCP hash tables to use RCU / hlist_nulls"), a race window emerged
during the switch from reqsk/sk to sk/tw.
Now that both timewait sock (tw) and full sock (sk) reside on the same
ehash chain, it is appropriate to introduce hlist_nulls replace
operations, to eliminate the race conditions caused by this window.
Before this series of patches, I previously sent another version of the
patch, attempting to avoid the issue using a lock mechanism. However, it
seems there are some problems with that approach now, so I've switched to
the "replace" method in the current patches to resolve the issue.
For details, refer to:
https://lore.kernel.org/netdev/20250903024406.2418362-1-xuanqiang.luo@linux.dev/
Before I encountered this type of issue recently, I found there had been
several historical discussions about it. Therefore, I'm adding this
background information for those interested to reference:
1. https://lore.kernel.org/lkml/20230118015941.1313-1-kerneljasonxing@gmail.com/
2. https://lore.kernel.org/netdev/20230606064306.9192-1-duanmuquan@baidu.com/
====================
Link: https://patch.msgid.link/20251015020236.431822-1-xuanqiang.luo@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since ehash lookups are lockless, if another CPU is converting sk to tw
concurrently, fetching the newly inserted tw with tw->tw_refcnt == 0 cause
lookup failure.
The call trace map is drawn as follows:
CPU 0 CPU 1
----- -----
inet_twsk_hashdance_schedule()
spin_lock()
inet_twsk_add_node_rcu(tw, ...)
__inet_lookup_established()
(find tw, failure due to tw_refcnt = 0)
__sk_nulls_del_node_init_rcu(sk)
refcount_set(&tw->tw_refcnt, 3)
spin_unlock()
By replacing sk with tw atomically via hlist_nulls_replace_init_rcu() after
setting tw_refcnt, we ensure that tw is either fully initialized or not
visible to other CPUs, eliminating the race.
It's worth noting that we held lock_sock() before the replacement, so
there's no need to check if sk is hashed. Thanks to Kuniyuki Iwashima!
Fixes: 3ab5aee7fe ("net: Convert TCP & DCCP hash tables to use RCU / hlist_nulls")
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Jiayuan Chen <jiayuan.chen@linux.dev>
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251015020236.431822-4-xuanqiang.luo@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since ehash lookups are lockless, if one CPU performs a lookup while
another concurrently deletes and inserts (removing reqsk and inserting sk),
the lookup may fail to find the socket, an RST may be sent.
The call trace map is drawn as follows:
CPU 0 CPU 1
----- -----
inet_ehash_insert()
spin_lock()
sk_nulls_del_node_init_rcu(osk)
__inet_lookup_established()
(lookup failed)
__sk_nulls_add_node_rcu(sk, list)
spin_unlock()
As both deletion and insertion operate on the same ehash chain, this patch
introduces a new sk_nulls_replace_node_init_rcu() helper functions to
implement atomic replacement.
Fixes: 5e0724d027 ("tcp/dccp: fix hashdance race for passive sessions")
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Jiayuan Chen <jiayuan.chen@linux.dev>
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251015020236.431822-3-xuanqiang.luo@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In {tcp6,udp6,raw6}_sock, struct ipv6_pinfo is always placed at
the beginning of a new cache line because
1. __alignof__(struct tcp_sock) is 64 due to ____cacheline_aligned
of __cacheline_group_begin(tcp_sock_write_tx)
2. __alignof__(struct udp_sock) is 64 due to ____cacheline_aligned
of struct numa_drop_counters
3. in raw6_sock, struct numa_drop_counters is placed before
struct ipv6_pinfo
. struct ipv6_pinfo is 136 bytes, but the last cache line is
only used by ipv6_fl_list:
$ pahole -C ipv6_pinfo vmlinux
struct ipv6_pinfo {
...
/* --- cacheline 2 boundary (128 bytes) --- */
struct ipv6_fl_socklist * ipv6_fl_list; /* 128 8 */
/* size: 136, cachelines: 3, members: 23 */
Let's move ipv6_fl_list from struct ipv6_pinfo to struct inet_sock
to save a full cache line for {tcp6,udp6,raw6}_sock.
Now, struct ipv6_pinfo is 128 bytes, and {tcp6,udp6,raw6}_sock have
64 bytes less, while {tcp,udp,raw}_sock retain the same size.
Before:
# grep -E "^(RAW|UDP[^L\-]|TCP)" /proc/slabinfo | awk '{print $1, "\t", $4}'
RAWv6 1408
UDPv6 1472
TCPv6 2560
RAW 1152
UDP 1280
TCP 2368
After:
# grep -E "^(RAW|UDP[^L\-]|TCP)" /proc/slabinfo | awk '{print $1, "\t", $4}'
RAWv6 1344
UDPv6 1408
TCPv6 2496
RAW 1152
UDP 1280
TCP 2368
Also, ipv6_fl_list and inet_flags (SNDFLOW bit) are placed in the
same cache line.
$ pahole -C inet_sock vmlinux
...
/* --- cacheline 11 boundary (704 bytes) was 56 bytes ago --- */
struct ipv6_pinfo * pinet6; /* 760 8 */
/* --- cacheline 12 boundary (768 bytes) --- */
struct ipv6_fl_socklist * ipv6_fl_list; /* 768 8 */
unsigned long inet_flags; /* 776 8 */
Doc churn is due to the insufficient Type column (only 1 space short).
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20251014224210.2964778-1-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Théo Lebrun says:
====================
net: macb: various cleanups
Fix many oddities inside the MACB driver. They accumulated in my
work-in-progress branch while working on MACB/GEM EyeQ5 support.
Part of this series has been seen on the lkml in March then June.
See below for a semblance of a changelog.
The initial goal was to post them alongside EyeQ5 support, but that
makes for too big of a series. It'll come afterwards, with new
features (interrupt coalescing, ethtool .set_channels() and XDP mostly).
[0]: https://lore.kernel.org/lkml/20250627-macb-v2-0-ff8207d0bb77@bootlin.com/
====================
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-0-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The pattern:
entry = macb_tx_ring_wrap(bp, i);
tx_skb = &queue->tx_skb[entry];
is the exact definition of:
macb_tx_skb(queue, i);
The pattern:
entry = macb_tx_ring_wrap(bp, i);
desc = macb_tx_desc(queue, entry);
is redundant because macb_tx_desc() calls macb_tx_ring_wrap().
One explicit call to macb_tx_ring_wrap() is still required for checking
if it is the last buffer (TX_WRAP case).
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-12-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Whenever min(a, b) is used with a and b unsigned variables or literals,
`make W=2` complains. Change four min() calls into umin().
stderr extract (GCC 11.2.0, MIPS Codescape):
./include/linux/minmax.h:68:57: warning: comparison is always true due
to limited range of data type [-Wtype-limits]
68 | #define __is_nonneg(ux) statically_true((long long)(ux) >= 0)
| ^~
drivers/net/ethernet/cadence/macb_main.c:2299:26: note: in expansion of
macro ‘min’
2299 | hdrlen = min(skb_headlen(skb), bp->max_tx_length);
| ^~~
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-11-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The low 16 bits of GEM_DCFG6 tell us which queues are enabled in HW. In
theory, there could be holes in the bitfield. In practice, the macb
driver would fail if there were holes as most loops iterate upon
bp->num_queues. Only macb_init() iterated correctly.
- Drop bp->queue_mask field.
- Error out at probe if a hole is in the queue mask.
- Rely upon bp->num_queues for iteration.
- As we drop the queue_mask probe local variable, fix RCT.
- Compute queue_mask on the fly for TAPRIO using bp->num_queues.
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-10-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Introduce macb_dma64() and macb_dma_ptp() helper functions.
Many codepaths are made simpler by dropping conditional compilation.
This implies two additional changes:
- Always compile related structure definitions inside <macb.h>.
- MACB_EXT_DESC can be dropped as it is useless now.
The common case:
#ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
struct macb_dma_desc_64 *desc_64;
if (bp->hw_dma_cap & HW_DMA_CAP_64B) {
desc_64 = macb_64b_desc(bp, desc);
// ...
}
#endif
Is replaced by:
if (macb_dma64(bp)) {
struct macb_dma_desc_64 *desc_64 = macb_64b_desc(bp, desc);
// ...
}
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-9-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The function body uses a switch statement on bp->hw_dma_cap and handles
its four possible values: 0, is_64b, is_ptp, is_64b && is_ptp.
Instead, refactor by noticing that the return value is:
desc_size * MULT
with MULT = 3 if is_64b && is_ptp,
2 if is_64b || is_ptp,
1 otherwise.
MULT can be expressed as:
1 + is_64b + is_ptp
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-7-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
macb_dma_desc_get_size() does a switch on bp->hw_dma_cap and covers all
four cases: 0, 64B, PTP, 64B+PTP. It also covers the #ifndef
MACB_EXT_DESC separately, making it four codepaths.
Instead, notice the descriptor size grows with enabled features and use
plain if-statements on 64B and PTP flags.
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://patch.msgid.link/20251014-macb-cleanup-v1-6-31cd266e22cd@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
