fix(bt/osi): stop timer before re-arming in alarm_set (IDFGH-17864)

[anuragak021]

Summary

alarm_set() calls esp_timer_start_once / esp_timer_start_periodic without first stopping the timer handle. If the Bluedroid stack re arms an alarm that is already running (e.g. a scheduling alarm fired before the previous instance was cancelled), esp_timer_start_* returns ESP_ERR_INVALID_STATE and the alarm is silently dropped, producing the log error:

E (xxxx) BT_OSI: alarm_set failed to start timer, err 0x103

alarm_free() already uses the correct pattern it calls esp_timer_stop before esp_timer_delete. This fix applies the same stop before start in alarm_set so that a double-set reschedules the alarm rather than failing.

Change

Add esp_timer_stop(alarm->alarm_hdl) before the esp_timer_start_once / esp_timer_start_periodic call inside the static alarm_set() helper in components/bt/common/osi/alarm.c.

esp_timer_stop returns ESP_ERR_INVALID_STATE when the timer is already stopped, which is expected and harmless; the subsequent start call then proceeds normally.

Reproduction

Observed on ESP32 S3 with CONFIG_BT_ALLOCATION_FROM_SPIRAM_FIRST=y. PSRAM allocation latency shifts task scheduling enough that osi_alarm_set is called on a running timer handle during BLE GAP advertising setup, triggering the error on every boot.

Test

Build with Bluedroid enabled (CONFIG_BT_BLUEDROID_ENABLED=y, CONFIG_BT_BLE_ENABLED=y) and start BLE advertising the alarm_set failed error no longer appears.

---

Note

Low Risk
Single-line behavioral fix in BT timer scheduling aligned with existing alarm_free() usage; no auth, data, or API surface changes.

Overview
Fixes Bluedroid alarms being silently dropped when osi_alarm_set / osi_alarm_set_periodic runs on a handle that is still active. esp_timer_start_once and esp_timer_start_periodic return ESP_ERR_INVALID_STATE in that case, which surfaced as alarm_set failed to start timer, err 0x103 during BLE GAP advertising on ESP32-S3 with SPIRAM-first BT allocation.

alarm_set() in alarm.c now calls esp_timer_stop on the handle before starting the timer, matching the stop-then-operate pattern already used in alarm_free(). Re-arming reschedules the alarm instead of failing; stopping an already-idle timer is harmless (ESP_ERR_INVALID_STATE is ignored).

^{Reviewed by Cursor Bugbot for commit b7af0d4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[CLAassistant]

All committers have signed the CLA.

[anuragak021]

Reproduced consistently on ESP32-S3 custom board with CONFIG_BT_ALLOCATION_FROM_SPIRAM_FIRST=y, IDF v5.5.2. Happy to test any backport builds against release/v5.5 if that helps.