On Sun, 03 Sep 2023 21:42:55 +0200, Christophe JAILLET wrote:
Le 03/09/2023 à 18:37, Takashi Iwai a écrit :
On Sun, 03 Sep 2023 17:04:47 +0200,
...
Indeed. The fix would be rather a oneliner like below, though:
Looks much better than mine :)
I let you send the patch, it is your solution.
Just for my understanding, how is snd_ump_ops used, especially .open? I've not been able to figure out where it was called.
It's called via rawmidi open (the snd_ump_endpoint is a sort of child class of snd_rawmidi).
In alloc_midi_urbs(), if usb_alloc_coherent() fails, then ctx->urb->transfer_buffer could be anything because usb_fill_xxx_urb() is not called. So there could be an edge case where your fix could still be incomplete.
Each URB is allocated in the loop via usb_alloc_urb(), and it does zero-initialize the object, hence the buffer is supposed to be NULL until it's set up via usb_fill_xxx().
thanks,
Takashi
For the start_input_streams() caller, this is fine, because the corresponding memory is kzalloc()'ed in start_input_streams() at some point, but I've not been able to check for snd_usb_midi_v2_open().
CJ
--- a/sound/usb/midi2.c +++ b/sound/usb/midi2.c @@ -265,7 +265,7 @@ static void free_midi_urbs(struct snd_usb_midi2_endpoint *ep) if (!ep) return;
- for (i = 0; i < ep->num_urbs; ++i) {
- for (i = 0; i < NUM_URBS; ++i) { ctx = &ep->urbs[i]; if (!ctx->urb) break;
That was the intended behavior of free_midi_urbs().
Takashi