Maxim Anisimov [Mon, 29 Jun 2026 13:14:20 +0000 (16:14 +0300)]
mediatek: fix nmbm max-reserved-blocks for Keenetic/Netcraze devices
The devicetree set mediatek,bmt-max-reserved-blocks to 64, but the
factory firmware uses 256. With the lower value NMBM fails to locate
its signature on some boards, and the device does not boot.
Restore the factory value of 256 to match stock firmware and fix
boot on affected units.
Vendor nmbm max blocks value:
https://github.com/keenetic/kernel-49/blob/master/drivers/mtd/nmbm/nmbm-mtd.c#L31
MAX_RX_PKT_LEN value 2 selects 1552-byte frames, not jumbo mode.
Select value 3 so the configured 2 KiB frames are actually accepted by
the switch.
Tested on a ZBT-WE826 (MT7620). With value 2, 1534-byte IPv4
packets pass while 1535-byte packets are dropped. With value 3,
1792-byte and 1800-byte DF IPv4 packets pass.
Ran a treewide trailing-whitespace cleanup for *.c, *.h, *.dts and *.dtsi.
Only files under target/linux/lantiq/... and target/linux/qualcommax/...
required changes, so the diff is limited to those paths.
Done by executing commands:
find . -type f | grep "\.c$" | xargs sed -i 's/[ \t]\+$//'
find . -type f | grep "\.h$" | xargs sed -i 's/[ \t]\+$//'
find . -type f | grep "\.dts$" | xargs sed -i 's/[ \t]\+$//'
find . -type f | grep "\.dtsi$" | xargs sed -i 's/[ \t]\+$//'
ipq40xx: ipqess: recover TX queue on watchdog timeout
The current ipqess TX timeout handler only logs the timeout. If TX
completion interrupts are lost or left masked, descriptors can remain
pending indefinitely and the master Ethernet device stays wedged until a
reboot.
Log the TX ring indexes and interrupt state to make future failures
useful, then force the TX completion NAPI path to run and re-kick the TX
producer index. This gives the driver a chance to reclaim pending TX
descriptors and wake the queue instead of reporting the same watchdog
forever.
The ipqess driver is carried locally in OpenWrt and is not present in
upstream Linux. Move the driver sources out of the source-adding kernel
patch and into target/linux/ipq40xx/files/ so the local source is easier
to maintain.
Keep the kernel patch only for the upstream-style Kconfig, Makefile and
MAINTAINERS wiring, and refresh the dependent ipqess patches.
Use the void-returning platform driver .remove callback.
Georgi Valkov [Wed, 15 Jul 2026 14:16:29 +0000 (17:16 +0300)]
kernel: set MACOSX_DEPLOYMENT_TARGET to host version
Fix a warning interpreted as error when building the Linux kernel
on macOS host, by setting MACOSX_DEPLOYMENT_TARGET to the version
of the host. Fixes:
make[5]: Entering directory '/Volumes/test/openwrt/build_dir/target-x86_64_musl/linux-x86_64/linux-6.18.38'
HOSTCC scripts/mod/modpost.o
scripts/mod/modpost.c:1719:9: error: 'strchrnul' is only available on macOS 15.4 or newer [-Werror,-Wunguarded-availability-new]
1719 | sep = strchrnul(namespace, ',');
| ^~~~~~~~~
Jonas Jelonek [Thu, 16 Jul 2026 07:04:51 +0000 (07:04 +0000)]
realtek: fix fan control for Plasma Cloud PSX28/ESX28
Commit 82ddc472d722 ("realtek: dts: convert to upstream switch
notation") adjusted the switch description in the device tree to match
upstream. This was a correct change, however it broke the fan control
script for Plasma Cloud PSX28/ESX28 devices because it uses a
full-fledged platform device path including the switch node (because the
I2C is a child of the switch):
Since the mentioned change, this path doesn't exist anymore. Instead of
just changing the intermediate parts to match the DT structure again, do
not rely on that at all. Instead, use
/sys/bus/i2c/devices/4-002e/hwmon/
as the path which relies less on the DT structure and still points to
the same device. This makes fan control on those device work again,
reducing the fan speed during boot which has been broken since the
commit.
Fixes: 82ddc472d722 ("realtek: dts: convert to upstream switch notation") Fixes: 2b5555c1950d ("realtek: rtl931x: Add support for Plasma Cloud PSX28 Switch") Link: https://github.com/openwrt/openwrt/pull/24248 Signed-off-by: Jonas Jelonek <jelonek.jonas@gmail.com>
Robert Marko [Thu, 16 Jul 2026 12:27:41 +0000 (14:27 +0200)]
airoha: an7583: fix AN7583 PCS port count
The AN7583 PCS match data does not initialize num_port. This leaves it at
zero, so probe allocates a zero-sized ports array and a later PCS lookup
returns a pointer derived from NULL.
Set the number of ports for each AN7583 PCS instance. Also reject
out-of-range PCS indices defensively.
Fixes kernel crashing on boot once ethernet starts probing on AN7583.
Signed-off-by: Robert Marko <robert.marko@sartura.hr>
Felix Fietkau [Mon, 13 Jul 2026 09:33:44 +0000 (11:33 +0200)]
ucode: update to Git HEAD (2026-07-09)
20a2ae469089 io: support older linux ioctls 16929698e2d3 fs: implement ioctl for macOS 395bb888bd69 io: implement ioctl for macOS 8d7d15ee1a08 compiler: add support for shorthand method declarations 7d9febd13305 nl80211: add mesh peer link enum macros to constants list e851bb64df91 lib: avoid allocating print buffer before NULL check in uc_error_message_indent 5f3a7b87d962 compiler: prevent unbounded recompilation on failed module imports e2493b577250 compiler: scope named function expression names to their own body c3f0d27d812e tests: add regression test for named function expression scoping 738b55299966 vm: release owned value when property assignment fails 1ffd7c5a1011 vm: fix out-of-bounds read in single-name module import a255256db106 vm: fix SIGFPE on INT64_MIN division and modulo by -1 3f4b313d7d78 vm: do not assign to undeclared variables in strict mode e523682670a2 vm: release upvalue reference when resolving it on stack push 6a837090bea3 vm: release module scope reference in dynamic import 395a0a16b4f9 vm: guard error context capture against empty callframe stack 04f103bb2270 vm: respect exponent parity in integer exponentiation bd814a4b2ae3 vm: use UC_ARRAY instead of json_type_array in object spread 5a79aa56c15e vm: release resource type prototypes after the final GC 9c5f16c8e33a lexer: do not treat /*/ as a complete block comment a018067fd12e lexer: preserve NUL bytes in regular expression literals 65d41a1929de lexer: do not consume a sign into hexadecimal number literals 05f9bf9e5d53 lexer: fix source position of ternary question mark token 7bca646e5f88 lib: preserve embedded NUL bytes in reverse() 6fc93bf5bd95 lib: preserve embedded NUL bytes in lc() and uc() a51f4843bce5 lib: avoid passing a signed char to isxdigit() in hex() e8af70e4375c compiler: fix use-after-free of shorthand method name 467fb4406a47 compiler: declare leading variable in counting for loop initializer d256e153799b rtnl: fix stack buffer overflow parsing IFLA_LINKINFO f741ac0a144c rtnl: fix stack buffer overflow parsing RTA_MULTIPATH nexthops c41310f7c2df rtnl: fix inverted address check in multipath nexthop parsing
Felix Fietkau [Fri, 10 Jul 2026 13:06:14 +0000 (15:06 +0200)]
wifi-scripts: don't swap handler config under a running setup script
A reload during an in-flight setup ran wdev_config_init() immediately,
replacing handler_config/handler_data while the script was still running, so
its set_data notifications resolved against the new (renumbered) interface
table and link up/down calls became asymmetric. Defer the re-init in
check() while a handler run is in progress and re-apply from the setup and
teardown callbacks once it completes.
Felix Fietkau [Fri, 10 Jul 2026 13:03:54 +0000 (15:03 +0200)]
wifi-scripts: only manage wdev interfaces recorded in the statefile
wdev.uc adopted every operstate-down netdev on the phy and then removed any
not in its new config. hostapd BSS netdevs are legitimately operstate-down
while stopped (co-located STA down, or start_disabled), so a reload deleted
them and hostapd, seeing an unchanged config, never recreated them. Drop the
adopt-existing step; the statefile already tracks the interfaces wdev.uc
owns.
Felix Fietkau [Fri, 10 Jul 2026 13:02:20 +0000 (15:02 +0200)]
wifi-scripts: give the OWE-transition BSS its own MAC address
The transition BSS is a copy of the main BSS with the MAC already set, so
iface.prepare() was a no-op and both BSSes emitted the same bssid; with a
fixed or random MAC the duplicate was never reassigned and the second
interface failed to come up. Clear the address (and stale flags) and
prepare a fresh one, plumbing the phy/num_global/macaddr_base through
iface_setup().
append() space-joins list options into one line, which breaks hostapd
options that are parsed one entry per line: anqp_elem, nai_realm,
venue_name, venue_url, hs20_conn_capab and the radius req_attr lists (a
parse error there fails the whole phy). Add append_list() to emit one line
per element and route those options through it. vendor_elements is instead
a single concatenated hex blob (join ''), and iface_interworking() no
longer overwrites domain_name with the 3GPP cell-net data (and uses ';').
Felix Fietkau [Fri, 10 Jul 2026 12:53:29 +0000 (14:53 +0200)]
wifi-scripts: use country_code for iw reg set in setup_phy
validate() converts the country alias to country_code and deletes country
before setup_phy() runs, so the iw reg set branch tested a null
config.country and never ran, leaving wdev-only radios (adhoc/mesh/monitor)
on the world regdomain. Use config.country_code.
Felix Fietkau [Fri, 10 Jul 2026 12:52:38 +0000 (14:52 +0200)]
hostapd: ucode: keep comment lines out of the config comparison data
The generator emits a '# Setup interface: <ifname>' comment before each
interface, so the first one landed in config.radio.data and made the radio
comparison depend on the first BSS's ifname: renaming it forced a full
radio restart instead of the rename path. Skip comment lines when parsing
radio.data/bss.data.
Felix Fietkau [Fri, 10 Jul 2026 12:51:32 +0000 (14:51 +0200)]
hostapd: ucode: filter MLD loops by phy in phy_set_state/phy_status
The per-wiphy radio-mask loops iterated all MLDs without checking the phy,
so a stop/start or status query for one wiphy touched MLDs on another wiphy
with the same radio index (clearing present bits, tearing links down, or
reporting a foreign MLD). Skip MLDs that do not belong to the phy, matching
mld_update_phy().
Felix Fietkau [Fri, 10 Jul 2026 12:50:23 +0000 (14:50 +0200)]
wifi-scripts: quote dpp_connector in the station config
dpp_connector was emitted unquoted, but wpa_supplicant parses it as a
string and hex-decodes an unquoted value; a base64url connector then failed
to parse and the network block was rejected. Emit it via the quoted-string
list. dpp_csign/dpp_netaccesskey stay raw, as they are hex.
Felix Fietkau [Fri, 10 Jul 2026 12:49:49 +0000 (14:49 +0200)]
hostapd: ucode: unregister ucode/ubus objects when deleting an unstarted BSS
hostapd_free_hapd_data() skips hostapd_ucode_free_bss()/hostapd_ubus_free_bss()
for a bss that never started, so bss.delete() on such a bss freed hapd while
the ucode registry entry and ubus object still referenced it. Call both
explicitly before os_free(); they are idempotent for the started case.
Felix Fietkau [Fri, 10 Jul 2026 12:48:10 +0000 (14:48 +0200)]
hostapd: ucode: initialise conf in iface.add_bss
The early error path jumps to out: before conf is assigned, where
hostapd_config_free(conf) frees an indeterminate pointer. Initialise it to
NULL (hostapd_config_free tolerates NULL).
Felix Fietkau [Fri, 10 Jul 2026 12:47:35 +0000 (14:47 +0200)]
hostapd: ucode: pass phy to the bss_remove callback
hostapd_ucode_free_bss() pushed only two arguments, but the bss_remove
handler (like bss_add/bss_reload) takes three (phy, name, obj), so name
received the bss resource: the DPP hook was never removed and the removal
event was malformed. Push the phy as well.
Felix Fietkau [Fri, 10 Jul 2026 12:46:58 +0000 (14:46 +0200)]
wifi-scripts: use the base phy name for config_get_macaddr_list
phydev.name is already radio-suffixed, and the hostapd handler re-applies
the suffix, so on multi-radio phys the lookup missed and the returned MAC
list was always empty, letting wdev.uc reuse a MAC hostapd had reserved for
an AP BSS. Pass the base phy and radio index.
Felix Fietkau [Fri, 10 Jul 2026 12:46:37 +0000 (14:46 +0200)]
wifi-scripts: pass device config to parse_encryption for station mode
setup_sta() called parse_encryption() without the device config, so the
sae-compat branch dereferenced a null dev_config.band and threw, aborting
the whole radio setup. Pass the device config through.
Felix Fietkau [Fri, 10 Jul 2026 12:46:09 +0000 (14:46 +0200)]
wifi-scripts: map the configured pairwise cipher for station mode
Only an exact GCMP cipher was mapped to the network block; other explicit
ciphers (CCMP-256, GCMP-256, TKIP variants) left pairwise unset, so
wpa_supplicant offered its default set and could not associate with an AP
restricted to the configured cipher. Pass wpa_pairwise through, keeping the
GCMP special case that also pins group.
Felix Fietkau [Fri, 10 Jul 2026 12:45:38 +0000 (14:45 +0200)]
wifi-scripts: do not treat WDS-station VLAN subinterfaces as the base wdev
The hotplug match /(.+)\.sta.+/ also matched VLAN subinterfaces of a WDS
station (e.g. wlan0.sta1.100 -> wlan0), so the tagged subinterface was added
into the vif base networks, breaking VLAN separation. Anchor the pattern to
the plain station suffix.
Felix Fietkau [Fri, 10 Jul 2026 12:45:10 +0000 (14:45 +0200)]
wifi-scripts: pass multicast_to_unicast as an integer
handle_link() passed the option as a ucode boolean, but the netifd
device_set binding reads it with prop_get_int (UC_INTEGER only) and ignored
a boolean, so the option was a no-op. Convert it to 0/1.
Felix Fietkau [Fri, 10 Jul 2026 12:44:51 +0000 (14:44 +0200)]
wifi-scripts: reset the retry counter on successful setup
retry was only reset on config change/start, never on a successful setup,
so occasional transient setup failures accumulated over a device lifetime
and latched retry_setup_failed after three cumulative failures. Reset it
when the device comes up.
Felix Fietkau [Fri, 10 Jul 2026 12:44:25 +0000 (14:44 +0200)]
wifi-scripts: fix VLAN key for service-injected wifi-iface
Service data VLANs were stored under vif_data.vlans, but every consumer
reads vif.vlan, so per-VLAN config from a procd service was silently
dropped. Use the singular key, matching the adjacent station handling.
Felix Fietkau [Fri, 10 Jul 2026 12:44:07 +0000 (14:44 +0200)]
wifi-scripts: return the interface-creation result from wdev_add
wdev_add() dropped the wdev_create() result, so every caller's error check
was dead and an nl80211 NEW_INTERFACE failure was silently ignored (hostapd
and wpa_supplicant then proceeded on a missing netdev with stale MLD
bookkeeping). Return the result.
Felix Fietkau [Fri, 10 Jul 2026 12:43:53 +0000 (14:43 +0200)]
hostapd: ucode: fix dead reload-all check in mld_set_config
new_config was tested against the freshly emptied new_mld object, so it was
always false and the reload-all-interfaces block never ran. Test the
incoming config instead, so the first MLD config re-adds the spliced-out
MLD BSSes.
Felix Fietkau [Fri, 10 Jul 2026 12:43:40 +0000 (14:43 +0200)]
hostapd: ucode: force restart on unresolvable static BSSID conflict
The guard meant to bail to a full restart read `!mac_idx < 0`, which parses
as `(!mac_idx) < 0` and is always false, so a static BSSID colliding with
the first (or a foreign-owned) BSS fell through and reassigned the wrong
BSS's address. Use `mac_idx <= 0`.
Felix Fietkau [Fri, 10 Jul 2026 12:21:45 +0000 (14:21 +0200)]
wifi-scripts: fix extended-feature bit test
device_extended_features() used bitwise OR instead of AND, so every
extended-feature query returned true regardless of the driver bitmap:
ftm_responder and radar_background were always reported present.
Felix Fietkau [Fri, 10 Jul 2026 12:19:48 +0000 (14:19 +0200)]
wifi-scripts: do not report a disabled wifi-device as up
A disabled wifi-device is intentionally still run through the setup path
(with no interfaces) to reliably tear down hostapd/wpa_supplicant and to
handle disabled/enabled flapping, so it reaches state "up" and status
reported up: true alongside disabled: true. Exclude disabled devices from
the up state in status instead.
Felix Fietkau [Fri, 10 Jul 2026 11:59:30 +0000 (13:59 +0200)]
hostapd: ucode: clear start_disabled when reloading a started BSS
On a radio shared with a STA every AP BSS is generated with
start_disabled=1. A smart reload that takes the full bss.set_config() path
restarts the BSS but never cleared start_disabled, so hostapd_start_beacon()
skipped beaconing and the AP went silently quiet until an unrelated
apsta_state event. Clear it when the BSS was already started, matching the
add_bss path.
Felix Fietkau [Fri, 10 Jul 2026 11:58:55 +0000 (13:58 +0200)]
wifi-scripts: emit a raw 64-hex PSK unquoted for station mode
A 64-character hex PSK was always quoted, but wpa_supplicant rejects a
quoted passphrase longer than 63 characters, so the network block failed
and the STA never associated while netifd still reported the radio up. Emit
the key unquoted when it is a 64-character raw PMK, matching the AP path.
The sae_password branch keeps quoting, as SAE passwords are arbitrary
strings.
Felix Fietkau [Fri, 10 Jul 2026 11:58:33 +0000 (13:58 +0200)]
wifi-scripts: honour a cancelled setup when the handler reports up
stop() and destroy() set cancel_setup while a setup handler is running, but
nothing read it, so a device brought down (or removed) mid-setup still went
up and stayed up, holding its interface links. Check the flag (and
autostart/delete) in wdev_mark_up and tear the device down instead of
completing the up transition.
Felix Fietkau [Fri, 10 Jul 2026 11:58:10 +0000 (13:58 +0200)]
hostapd: ucode: fix per-station PSK list parsing in sta_auth
Two defects handling the psk array returned by an auth handler: strlen()
ran before the string was validated (strlen(NULL) on a non-string entry),
and the hex-PMK branch tested the array length instead of the entry length.
A 64-character passphrase therefore reached the passphrase branch and its
memcpy of str_len + 1 overflowed passphrase[MAX_PASSPHRASE_LEN + 1] by one
byte. Validate the type first and branch on the string length.
Felix Fietkau [Fri, 10 Jul 2026 11:57:49 +0000 (13:57 +0200)]
wifi-scripts: fix undeclared variable in service wifi-device injection
The wifi-device injection path used object shorthand `name` where no such
variable is in scope (the loop binds radio/config), so under the strict VM
config_init threw and the whole wireless configuration for that reload was
skipped. Use the radio key as the device name.
Felix Fietkau [Fri, 10 Jul 2026 11:57:16 +0000 (13:57 +0200)]
hostapd: ucode: handle null config in iface_set_config again
config_reset and shutdown call iface_set_config(name) with no config, but
the null-config guard was dropped in the MLO rework, so config.phy
dereferenced null and threw: config_reset reset nothing and shutdown left
hostapd-created interfaces behind. Restore the guard, removing the stale
config entry and tearing the interface down via iface_config_remove.
Felix Fietkau [Fri, 10 Jul 2026 11:13:32 +0000 (13:13 +0200)]
hostapd: ucode: restart on config-initiated channel switch failure
iface.switch_channel() only reports whether a CSA was started, not whether
it completed, so a config-driven channel switch that the driver never
finishes would leave the AP stuck without recovery. After issuing the CSA,
arm a timer and, if the switch has not completed by the deadline
(csa_in_progress still set), fall back to a full interface restart. The
timer is armed only on the config-initiated path, so ubus- and
apsta-triggered channel switches keep their own recovery.
Add an iface.csa_in_progress() ucode method wrapping hostapd_csa_in_progress().
Felix Fietkau [Fri, 10 Jul 2026 10:31:37 +0000 (12:31 +0200)]
hostapd: ucode: apply channel/width changes via CSA instead of restart
A config change that only alters the channel or channel width previously
failed the radio comparison in iface_reload_config() and forced a full
interface restart, disconnecting every client. Classify the radio diff:
apply a channel/width-only change through iface.switch_channel() (CSA) so
clients stay associated, keep the full restart for any other radio change.
When a co-located STA on the same band owns the runtime channel
(#channel_owned_by_sta), the change is a no-op on the AP side, which keeps
following the STA via apsta_state. MLD APs and DFS targets fall back to a
full restart, as a single iface-level CSA cannot handle them.
Felix Fietkau [Fri, 10 Jul 2026 10:26:55 +0000 (12:26 +0200)]
wifi-scripts: emit frequency and STA-channel markers in hostapd config
Emit #frequency (the target channel frequency, already computed during phy
setup) and #channel_owned_by_sta (set when a co-located STA/mesh/adhoc on
the same band dictates the runtime channel) into the generated config.
The hostapd daemon uses these to decide whether a channel change can be
applied via CSA instead of a full restart.
Felix Fietkau [Fri, 10 Jul 2026 10:26:31 +0000 (12:26 +0200)]
hostapd: ucode: parse frequency and STA-channel radio markers
Lift the #frequency and #channel_owned_by_sta markers (emitted by the
config generator) onto config.radio and keep them out of radio.data so
they do not perturb the radio comparison. These feed the config-driven
channel switch handling added next.
Felix Fietkau [Fri, 10 Jul 2026 10:25:58 +0000 (12:25 +0200)]
hostapd: ucode: compute 6 GHz segment centre in iface_freq_info
On 6 GHz there is no HT Operation IE, so a 0 secondary-channel offset made
freq_info() return without a segment centre frequency, leaving wide-channel
switches (80/160/320 MHz) on the stale centre. Pass a null offset for wide
6 GHz channels so the centre is auto-derived.
Felix Fietkau [Fri, 10 Jul 2026 10:23:19 +0000 (12:23 +0200)]
hostapd: ucode: fix CSA bandwidth for 320 MHz and 80+80 channels
switch_channel computed the CSA bandwidth as 40 << oper_chwidth, which
mismaps oper_chwidth 3 (80+80, must be 80 MHz plus center_freq2) to 320
and 9 (320 MHz) to 20480. Map the operating channel width to the actual
MHz bandwidth explicitly.
ath79: fix u-boot-env nvmem-layout on enterasys,ws-ap3705i
The u-boot-env0 partition on this board stores a redundant U-Boot
environment (4-byte CRC32 + 1-byte flag + data), with the flag byte
differing between the u-boot-env0/u-boot-env1 copies (0x00 vs 0x01).
The actual environment size used by U-Boot for the CRC calculation is
0x1000 bytes, not the full 0x10000 byte partition.
The DTS declared this as the plain, non-redundant "u-boot,env" layout
with no env-size, so the kernel's u-boot-env nvmem-layout driver
miscomputes the CRC and fails to probe:
u-boot-env-layout ...nvmem-layout: Invalid calculated CRC32: 0xbc490a68 (expected: 0xfae56b95)
u-boot-env-layout ...nvmem-layout: probe with driver u-boot-env-layout failed with error -22
Because the probe fails, the macaddr_uboot_ethaddr nvmem cell is never
created, so eth0's nvmem-cells reference never resolves and the
ag71xx-legacy platform device is stuck in "deferred probe pending"
forever - the device boots with no functioning Ethernet interface at
all.
Switch to "u-boot,env-redundant-count" (matching the redundant-env
layout already used on other boards from this vendor lineage, e.g.
qca9557_extreme-networks_ws-ap3805i.dts) and set env-size to the
verified 0x1000, which fixes CRC validation and lets eth0 probe
successfully.
Verified on real hardware: before the fix, eth0 never comes up
(confirmed against the official 25.12.5 release, matching CRC values
exactly); after the fix, eth0 links up normally and dmesg is free of
any CRC/nvmem-layout errors.
Jonas Jelonek [Tue, 14 Jul 2026 13:16:42 +0000 (13:16 +0000)]
realtek: pcs: rtl930x: gate WDIG 10G-mode bit on real baud rate
This bit was set unconditionally for any 10G-capable SerDes regardless
of hw_mode, including SGMII/1000BASEX/2500BASEX running on 10G-capable
silicon. The vendor SDK's dal_longan_construct_mac_default_10gmedia_fiber
gates it on the port's declared protocol class instead, applying it to
PHY-attached XSGMII/USXGMII as well as genuine fiber - the "medium to
fiber" naming doesn't match its actual condition.
Gate on speed == RTPCS_SDS_PLL_SPD_10000 instead, matching the real
10G-class protocol family, and actively clear it for 1G/2.5G modes so a
prior 10G-class configuration doesn't leave it stuck across a runtime
downshift. 0 also matches this register's confirmed hardware reset
default on an unconfigured SerDes.
Jonas Jelonek [Fri, 10 Jul 2026 13:38:15 +0000 (13:38 +0000)]
realtek: pcs: rtl930x: gate TX tuning on attachment and baud rate
TX amp/pre-emphasis tuning depends on the SerDes' real baud rate class
rather than hw_mode, so key config_attachment's dispatch on that instead.
At 1G/2.5G, attachment doesn't change meaningful: even a non-adaptive
fiber receiver needs little compensation on a short trace, so
PHY-attached and fiber links already use the same config there. At 10G,
real channel loss is high enough that it matters: PHY-attached links
equalize on the PHY's own far-end receiver and need less amp drive,
while genuine fiber/DAC links, which have no adaptive receiver
downstream, need the full config. Real DAC-vs-fiber detection doesn't
exist yet, so both of those still share one config for now.
Make use of the PLL speed enum for now, though it is a bit blurry and
may not represent the real baud rate in some cases, e.g. QSGMII (1.25G
PLL with multiplier -> 5G).
Jonas Jelonek [Fri, 10 Jul 2026 09:13:40 +0000 (09:13 +0000)]
realtek: pcs: rtl930x: refactor TX config
Replace the on-stack, hardcoded pre/main/post-amp locals with static
const rtpcs_sds_tx_config data, one struct per speed class (1g, 2g5,
10g), instead of duplicating the same literals inline on every call.
Move the decision logic about which values are applied to the
config_media function. The tx_config should only be about what is set,
decision logic is handled one layer above. This also prepares further
refactoring.
Jonas Jelonek [Mon, 13 Jul 2026 10:50:21 +0000 (10:50 +0000)]
realtek: pcs: rtl93xx: commit attachment only after successful config
sds->attachment was written right after selection, before
config_attachment() ran, so a failed config call would leave it
reflecting a value that was never actually applied to hardware. Move
the assignment after the call succeeds, matching how sds->hw_mode is
only committed after set_hw_mode() succeeds.
Jonas Jelonek [Fri, 10 Jul 2026 13:53:01 +0000 (13:53 +0000)]
realtek: pcs: rtl93xx: rename media to attachment
"media" doesn't fit well once PHY sits alongside FIBER/DAC_SHORT/
DAC_LONG as a value: a PHY isn't a transmission medium, whereas what
actually matters here is what electrically terminates the SerDes lane
on the other end. Rename the enum, struct field, ops vtable member,
and both variants' config functions accordingly. Calling it 'attachment'
better fits what it actually is, doesn't confuse with the variants and
also fits to the PMA (Physical Medium Attachment) sublayer that lives
below PCS.
Jonas Jelonek [Fri, 10 Jul 2026 13:45:43 +0000 (13:45 +0000)]
realtek: pcs: rtl93xx: rename MEDIA_PCB to MEDIA_PHY
Every SerDes lane runs over some PCB trace before reaching the port,
even fiber/DAC ones, so PCB doesn't describe what actually
distinguishes this case: the SerDes terminates directly into a PHY
chip rather than a fiber module or DAC cable.
Jonas Jelonek [Fri, 10 Jul 2026 13:36:19 +0000 (13:36 +0000)]
realtek: pcs: rtl930x: classify SGMII as PHY-attached media
SGMII always terminates in a PHY, whether on-board or embedded in an
SFP copper module, so it belongs with QSGMII/XSGMII/USXGMII rather
than the direct fiber/DAC modes. The PHY equalizes on its own far-end
receiver in either case, so no media-dependent TX tuning is needed.
Jonas Jelonek [Mon, 13 Jul 2026 12:17:50 +0000 (12:17 +0000)]
realtek: pcs: rtl930x: only apply TX config for 10G SerDes
So far, there was only a misleading and confusing gate to avoid applying
the TX config to SerDes which do not need it or where it even breaks
some modes. This is achieved by not applying any configuration for
QSGMII mode. While this fits the real-world cases we know so far where
QSGMII is only used on SerDes 0/1 but never on others, it implicates
that this is true too for QSGMII on 10G SerDes. Looking at the SDK, this
assumption doesn't hold.
To fix this, drop out of config_media in case a non-10G SerDes is
delivered. The SDK does the same in various ways, ending up with the
config just being applied for SerDes 2 - 9.
Jonas Jelonek [Fri, 10 Jul 2026 08:56:57 +0000 (08:56 +0000)]
realtek: pcs: rtl930x: use cmu_page getter for tx_config
Derive the CMU page from rtpcs_93xx_sds_get_cmu_page() instead of
hardcoding it per hw_mode case. Also change the function to return
int and propagate errors instead of silently doing nothing.
Jonas Jelonek [Fri, 10 Jul 2026 08:48:48 +0000 (08:48 +0000)]
realtek: pcs: rtl93xx: make cmu_page getter generic
Rename rtpcs_931x_sds_cmu_page_get() to rtpcs_93xx_sds_get_cmu_page()
and move it next to the other 93xx-shared helpers, so RTL930X can
reuse it instead of duplicating the same hw_mode-to-CMU-page switch.
Also fix the wrong -ENOTSUPP return to -EOPNOTSUPP.
Jonas Jelonek [Sun, 3 May 2026 22:13:44 +0000 (22:13 +0000)]
realtek: use realtek-pse-mcu kernel driver for PoE on XMG1915-10EP
Instead of relying on the userspace implementation, make this device the
first user of the new realtek-pse kernel driver (kmod-pse-realtek-mcu-uart)
based on the PSE-PD framework to interface PoE functionality of that
switch. Control of PoE functionality happens via ethtool / netifd now.
Drop selecting realtek-poe for this device.
This doesn't include any support for PoE-related LEDs.
Jonas Jelonek [Wed, 1 Jul 2026 10:47:27 +0000 (10:47 +0000)]
realtek: pse: add patch for sysfs detection control
Add a patch which modifies the Realtek PSE MCU driver backport to add a
sysfs control to control detection mode per port. Some older,
bad-designed devices aren't detected properly in standard mode on some
switches, though vendor firmware powers them properly. Testing shows
that enabling the detection of legacy PDs by allowing wider resistance,
more input capacitance and using a wider detection window, solves this
issue.
The legacy detection can be activated by writing 1 to
portN/detection_legacy in the device's sysfs, e.g.
Jonas Jelonek [Sun, 3 May 2026 20:44:35 +0000 (20:44 +0000)]
realtek: add Realtek PSE MCU driver
Add pending patches which add a PSE driver for the PSE setup found on
most of Realtek-based switches. An MCU with a Realtek-defined firmware
and protocol fronts one or more PSE chips (from Realtek or Broadcom) as
a management controller.
The driver provides both I2C/SMBus and UART communication, which varies
among our supported switches. There is no need for complicated userspace
handling anymore, the PSE setup is defined in the device tree and
interfaced in userspace via ethtool, netlink and netifd support.
Those patches have progressed far enough upstream, the bindings are
reviewed. Thus, we can keep this intermediate version downstream to open
it up earlier for usage. Only little changes are expected til the final
version lands.
Jonas Jelonek [Sun, 3 May 2026 20:43:57 +0000 (20:43 +0000)]
generic: pse-pd: add patches for module-based PSE drivers
Add a backport patch and pending patches needed by upcoming PSE drivers
which are built as modules and probed after the MAC/PHY.
The net effect for module-built PSE controllers: attachment to PHYs
happens via the lifecycle notifier rather than via probe-time
-EPROBE_DEFER coupling, so the MDIO/DSA probe no longer sees any
PSE-originated -EPROBE_DEFER and the probe-retry storm is gone.
airoha: configure PCIe 2-lanes mode for eMMC EAGLE RFB board
The EAGLE PCIe WiFi card require PCIe0 to be in 2-lanes mode to correctly
work. Add the missing property to enable this mode and restore correctly
functionality of the WiFi card.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
airoha: replace PCIe x2-mode patch with new pending version
The current x2-mode patch is mostly an hack ported from Airoha SDK. Replace
with in favor of a new version posted upstream that better implement
handling of reset and configuring x2-mode. This now use the standard
num-lanes property and dedicated PERSTOUT resets exported by the clk
driver.
All the DTS that used x2-mode are updated.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Robert Marko [Mon, 13 Jul 2026 09:07:26 +0000 (11:07 +0200)]
qualcommax: pin EDMA processing to dedicated CPUs
IPQ60xx and IPQ807x use threaded NAPI for EDMA. The NAPI threads
inherit CPU0 affinity, and running both there can eventually starve the
RX fill ring under sustained traffic.
Add an IPQ60xx smp_affinity service and extend the existing IPQ807x
service to discover the EDMA IRQs and NAPI IDs at runtime. Place RX on
CPU1 and TX on CPU2. Retry EDMA setup asynchronously because its NAPI
threads may appear after the init service runs.
Provide the same UCI enable and logging controls on both subtargets.
IPQ50xx uses a different Ethernet driver and is left unchanged.
Robert Marko [Mon, 13 Jul 2026 08:39:43 +0000 (10:39 +0200)]
qualcommax: qca-edma: enable threaded NAPI by default
Run the EDMA RX and TX NAPI poll loops in dedicated kernel threads by
default. This allows the threads and their interrupts to be placed on
separate CPUs and avoids overloading a single softirq context.
If thread creation fails, dev_set_threaded() leaves NAPI in softirq mode
and the driver reports the failure.
A malformed RX preheader can occur repeatedly after the descriptor ring
overruns. Rate-limit the warning to prevent excessive console output
from making the receive stall worse.
Decode the little-endian preheader metadata before printing it and use
consistent hexadecimal formatting.
qualcommax: ipq8072: assign NVMEM MAC addresses for Linksys MX5300
While the base NVMEM cell (hw_mac_addr) is already defined in the device tree, it was not being utilized by the actual interfaces.
This patch completes the kernel-level MAC assignment by:
1. Correcting the label-mac-device alias to point to the WAN port (&swport5), ensuring the system reports the sticker MAC address correctly.
2. Assigning the base MAC address (offset 0) to the WAN port (&swport5).
3. Assigning the base MAC address (offset 1) to the LAN ports (&swport1 - &swport4).
4. Configuring the proper NVMEM cell references for the PCIe-attached QCA9984 radio (offset 2) and the built-in ATH11K radio (offset 3).
This eliminates the need for post-boot MAC address manipulation and ensures robust, predictable interface initialization.
George Moussalem [Mon, 13 Jul 2026 05:48:04 +0000 (09:48 +0400)]
qualcommax: ipq50xx: enable in-band-status for MXL GPY115C PHY on Linksys MX6200
The MXL GPY115C PHY requires in-band-status to be set properly detect
link status and communicate phy capabilities. So enable in-band-status
in the Linksys MX6200 device tree.
ramips: ethernet: ralink: refine DSA tag offload handling
DSA copies the driver features to slave device, including offload
capabilities. Once a packet is sent through a DSA slave interface,
according to its features, the kernel does not calculate checksums,
expecting that the HW will fill the gaps. DSA adds the defined DSA
tag and sends the tagged packet through the master device.
Ethertype DSA tags expect the driver to calculate checksum based on the
csum_start/csum_offset. However, mtk_soc_eth does not use that info.
It checks the network header and decides if the HW can manage that
packet, unaware that mac layer now contains an extra DSA tag. When
that tag is the Mediatek CPU tag, offload will work as expected.
When it is an incompatible DSA tag or if DSA is stacking two incompatible
DSA tags, the driver will still count on the HW offload. In this case,
packets go to the network with an incorrect checksum.
Before this change, tag_ops->proto == DSA_TAG_PROTO_MTK was used,
which disabled offload for every non-Mediatek tag, including the
RTL8367S switch's RTL8_4 tag that this fix specifically targets.
Replace the hardcoded whitelist with a blacklist scoped to the
RTL8367S: only disable offload for DSA_TAG_PROTO_RTL8_4. The RTL8_4T
(the trailing-tag variant) is left untouched because its rtl8_4t tagger
already checksums the frame in software before appending the tag.
Tags for other DSA-capable switches are left untouched by this fix
and are not claimed to be offload-safe.
Fixes: 3c0a73b4d202 ("ramips: ethernet: ralink: fix offload with diff dsa tag") Signed-off-by: Mieczyslaw Nalewaj <namiltd@yahoo.com> Link: https://github.com/openwrt/openwrt/pull/24064 Signed-off-by: Robert Marko <robimarko@gmail.com>
mediatek: ethernet: fix offload with incompatible dsa tag
DSA copies the driver features to slave device, including offload
capabilities. Once a packet is sent through a DSA slave interface,
according to its features, the kernel does not calculate checksums,
expecting that the HW will fill the gaps. DSA adds the defined DSA
tag and sends the tagged packet through the master device.
Ethertype DSA tags expect the driver to calculate checksum based on the
csum_start/csum_offset. However, mtk_eth_soc does not use that info.
mtk_eth_soc checks the network header and decides if the HW can manage
that packet, unaware that MAC layer now contains an extra DSA tag. When
that tag is the Mediatek CPU tag, offload will work as expected.
When it is an incompatible DSA tag or if DSA is stacking two incompatible
DSA tags, the driver will still count on the HW offload. In this case,
packets are sent to the network with an incorrect checksum.
This patch adds an extra check that disables offloading only for the
RTL8367S switch's RTL8_4 tag, which is known to break MTK checksum
offload. The RTL8_4T (the trailing-tag variant) is not affected because
its rtl8_4t tagger already checksums the frame in software before appending
the tag, so ip_summed is no longer CHECKSUM_PARTIAL by the time this driver
sees it.
This approach avoids an overly conservative whitelist which would
disable offload for every non-Mediatek tag, regardless of whether
that tag is actually incompatible with MTK checksum offload.
Tags for other DSA-capable switches are left untouched by this fix
and are not claimed to be offload-safe.
Signed-off-by: Luiz Angelo Daros de Luca <luizluca@gmail.com> Signed-off-by: Schneider Azima <Schneider-Azima12@protonmail.com> Signed-off-by: Mieczyslaw Nalewaj <namiltd@yahoo.com> Link: https://github.com/openwrt/openwrt/pull/23996 Signed-off-by: Robert Marko <robimarko@gmail.com>
Michael Pratt [Sat, 4 Jul 2026 07:21:49 +0000 (03:21 -0400)]
tools: gnulib: rename macro file for cond module
It was reported that cond.m4 in gnulib is a name clash with
cond.m4 provided by Automake, where they are for completely
different purposes instead of different versions of the same macros.
A quick survey of all the macro files in the build directory reveals that
this is the only case where the gnulib copy is signficantly smaller
than the rest of the copies of the same macro name
distributed in the rest of the build system,
and the only one that name clashes with Automake.
A previous fix added a prefix to all macros from gnulib,
but the name must match how it is described in the respective modules files
as a functional requirement to build certain tools for certain (older) hosts,
so patch the problematic module instead of renaming all macros from gnulib.
Ref: c820f097e0be ("tools: gnulib: install .m4 file with gl_ prefix")
Ref: 78a8cfb57772 ("tools: gnulib: fix broken install of .m4 files") Reported-by: Christian Marangi <ansuelsmth@gmail.com> Signed-off-by: Michael Pratt <mcpratt@pm.me> Link: https://github.com/openwrt/openwrt/pull/24136 Signed-off-by: Robert Marko <robimarko@gmail.com>
Shiji Yang [Tue, 27 May 2025 16:32:56 +0000 (00:32 +0800)]
lantiq: use gpiod API for PCIe GPIO reset
This is the recommended way for the OF based platform. According to
the original patch, set GPIO to low level to assert the reset, set
GPIO to high level to deassert. Hence, adjust the dts GPIO polarity
to active-low.
projects / thirdparty / openwrt.git / log
