Repository navigation
host_stop() wedges after a UAC device was attached but never opened (S3 host, P4 sound card) #29
Description
Activity
Status 2026-10-08: not reproduced tonight. The only S3 host on the bench is the LCD-7, and its native USB port is unplugged, so nothing can host the P4 panel. What the code says about the log above, for whoever runs it next:
device_free_all -> 0x10cisn't the failure.ESP_ERR_NOT_FINISHEDis whatusb_host_device_free_all()returns whenever any device is still attached (usb/include/usb/usb_host.h:361in ESP-IDF 5.5.4: "wait for USB_HOST_LIB_EVENT_FLAGS_ALL_FREE"). With a hub in the path there's always one.- What's missing is the line after it. The ALL_FREE wait (
usbif_host.c, 100 polls of 10 ms) always printsALL_FREE seenorNOT seen, thenuninstall -> .... Neither appeared beforehost_stop()gave up at 2000 ms. - That budget is tight. Before the wait, teardown pumps events in each class driver's
_for_host_stop()close (UAC's alone is ten 10 ms polls), then a 200 ms drain. An ALL_FREE that never comes adds a full second on top. So this could be ALL_FREE not arriving for an enumerated-but-unopened device, outlastinghost_stop()'s 2 s, andhost_stop()then marking the task wedged (which is theEIOon the nextstart()). It could also be ausb_host_lib_handle_events()call that blocks.
Next run: timestamp each teardown line (
esp_timer_get_time()), and run once with the P4 straight on the S3's port and once behind the hub. That separates the two. If it's the budget, the fix is to makehost_stop()wait longer than the teardown's own worst case and let the uninstall result decide whether the host is wedged. Waits on an S3 host cabled to the P4.
2026-09-23, S3 Touch on 61ba01c hosting the P4 wearing cdc+uac. A probe started the host with
classes=("uac",), listeduac_audio.audio_devices()(one device, one stream, never opened) and calledstop():The next
host.start()in the same session then fails withOSError: [Errno 5] EIO; a hard reset clears it.device_free_all -> 0x10cis ESP_ERR_NOT_FINISHED: the library still held the device when we asked it to free everything, so the host task never reached its exit. The README's 8-of-8 clean stops were measured with devices that had been opened and closed through a class driver; this is the other path, a device that was only enumerated.