PTP TX hardware timestamp timeout on Raspberry Pi CM5 (regression from 6.6 to 6.12)

[zjzhao-eda]

Describe the bug

Basic Information

• Hardware: Raspberry Pi CM5 + Official CM5 IO Board
• Network Port: eth0 (Cadence GEM MAC + BCM54213PE PHY)
• Regression Scope: Kernel 6.6.74 (Working) → 6.12+ (Failing)

Test Matrix

Kernel	Provider	Result
rpi-6.6.y	PHY timestamp (idx 0)	PASS
rpi-6.12.y	PHY timestamp (idx 0)	FAIL
rpi-6.18.y	PHY timestamp (idx 0)	FAIL

Symptom

1. Test side

pi@raspberrypi:~ $ sudo ptp4l -i eth0 -m
ptp4l[70.739]: selected /dev/ptp0 as PTP clock
ptp4l[70.745]: port 1: INITIALIZING to LISTENING on INIT_COMPLETE
ptp4l[70.745]: port 0: INITIALIZING to LISTENING on INIT_COMPLETE
ptp4l[77.867]: port 1: LISTENING to MASTER on ANNOUNCE_RECEIPT_TIMEOUT_EXPIRES
ptp4l[77.867]: selected local clock 2ccf67.fffe.feb41d as best master
ptp4l[77.867]: port 1: assuming the grand master role
ptp4l[78.868]: timed out while polling for tx timestamp
ptp4l[78.868]: increasing tx_timestamp_timeout may correct this issue, but it is likely caused by a driver bug
ptp4l[78.868]: port 1: send sync failed
ptp4l[78.868]: port 1: MASTER to FAULTY on FAULT_DETECTED (FT_UNSPECIFIED)

2. Monitoring end

pi@raspberrypi:~ $ sudo tcpdump -ni eth0 udp port 319 or udp port 320
tcpdump: verbose output suppressed, use -v[v]... for full protocol decode
listening on eth0, link-type EN10MB (Ethernet), snapshot length 262144 bytes
10:11:14.984307 IP 192.168.8.4.320 > 224.0.1.129.320: PTPv2, v1 compat : no, msg type : announce msg, length : 64, domain : 0, reserved1 : 0, Flags [timescale], NS correction : 0, sub NS correction : 0, reserved2 : 0, clock identity : 0x2ccf67fffefeb41d, port id : 1, seq id : 0, control : 5 (Other), log message interval : 1, originTimeStamp : 0 seconds 0 nanoseconds, origin cur utc :37, rsvd : 0, gm priority_1 : 128, gm clock class : 248, gm clock accuracy : 254, gm clock variance : 65535, gm priority_2 : 128, gm clock id : 0x2ccf67fffefeb41d, steps removed : 0, time source : 0xa0
10:11:15.983322 IP 192.168.8.4.319 > 224.0.1.129.319: PTPv2, v1 compat : no, msg type : sync msg, length : 44, domain : 0, reserved1 : 0, Flags [two step], NS correction : 0, sub NS correction : 0, reserved2 : 0, clock identity : 0x2ccf67fffefeb41d, port id : 1, seq id : 0, control : 0 (Sync), log message interval : 0, originTimeStamp : 0 seconds, 0 nanoseconds

Reproduction (minimal)

1. Install the following tools：

sudo apt install tcpdump  linuxptp -y

2. Perform the following operations in a terminal：

sudo ptp4l -i eth0 -m

3. Open a new terminal and perform the following operations：

sudo tcpdump -ni eth0 udp port 319 or udp port 320

Analysis

1. drivers/net/phy/bcm-phy-ptp.c is functionally identical between rpi-6.6.y and rpi-6.18.y . The core TX timestamp logic ( bcm_ptp_txtstamp , bcm_ptp_do_aux_work , bcm_ptp_get_tstamp ) is unchanged.
2. EEE is NOT the cause. ethtool --set-eee eth0 eee off does not resolve the issue.
3. IRQ vs Poll mode irrelevant. Both modes fail on the newer kernels.
4. Likely suspects (the regression was introduced between Linux 6.6 and 6.12)

Impact

This breaks PTP hardware timestamping via the PHY, which is required for industrial use cases requiring sub-100ns synchronization precision (motion control, IEC 61850, 5G fronthaul, TSN 802.1AS). MAC timestamps provide a lower-precision workaround (~50-200ns RGMII jitter vs <10ns for PHY timestamps).

Steps to reproduce the behaviour

1. Install the following tools：

sudo apt install tcpdump  linuxptp -y

2. Perform the following operations in a terminal：

sudo ptp4l -i eth0 -m

3. Open a new terminal and perform the following operations：

sudo tcpdump -ni eth0 udp port 319 or udp port 320

Device (s)

Raspberry Pi CM5

System

• Hardware: Raspberry Pi CM5 + Official CM5 IO Board
• Network Port: eth0 (Cadence GEM MAC + BCM54213PE PHY)
• Regression Scope: Kernel 6.6.74 (Working) → 6.12+ (Failing)

Logs

No response

Additional context

No response