net: r8169: add device tree based LED configuration for r8125

[mingzhangqun]

Summary

Read r8125 LED register/value pairs from the PCIe controller device tree realtek,led-data property, instead of hardcoding in the shared driver path.

The driver searches DT nodes for realtek,led-data and writes each reg/value pair via RTL_W16.

Tested on Seeed reComputer RK3588 DevKit

• Both r8125 Ethernet ports: LED registers correctly written
• No impact on other boards (property absent = default behavior)

[HeyMeco]

This is a better approach using device tree for board-specific LED config. A few suggestions to improve efficiency and safety:

1. Avoid Repeated Memory Allocation

The function allocates and parses on every call (WoL + hardware init). Cache the config once during probe instead:

// In rtl8169_private struct:
u32 *led_config;
int led_config_count;

// Parse once during probe, then reuse in rtl8125_apply_dt_led_config()

This eliminates the kmalloc_array() call from the hot path.

2. Eliminate Redundant Calls

Currently called twice during init—from both rtl_wol_enable_rx() and rtl_hw_start_8125_common(). Call it once from a single location like rtl_hw_start_8125().

3. Improve Type Safety (Minor)

for (i = 0; i < count; i += 2)
    RTL_W16(tp, (unsigned int)led_data[i], (u16)led_data[i + 1]);

Use unsigned int for register addresses instead of int cast.

---

[mingzhangqun]

@HeyMeco Thanks for the review! All three suggestions applied in the latest push:

1. Cache config at probe — struct rtl8169_private now holds led_data / led_data_count. DT parsing (rtl8125_parse_dt_led_config()) runs once from rtl_init_one(), and rtl_remove_one() frees the buffer. The hot-path rtl8125_apply_dt_led_config() only iterates the cached array — no more kmalloc_array() per call.

2. Single call site — removed the call from rtl_wol_enable_rx(). LED config is now applied only from rtl_hw_start_8125_common().

3. Type safety — dropped the (int) cast; register addresses flow as u32 straight into RTL_W16().

Patch updated to v2 (21bb15e). The diff vs the previous revision is essentially the parse/apply split + the struct fields + the probe/remove wiring.

[mingzhangqun]

Superseded by #497 (branch recreated due to CI runner failure). All discussion migrated.