
On Wed, 2019-03-06 at 10:19 +0100, Jaroslav Kysela wrote:
Dne 06. 03. 19 v 9:46 KaiChieh Chuang napsal(a):
the dpcm get from fe_clients/be_clients may be free before use
@@ -1294,9 +1294,11 @@ void dpcm_be_disconnect(struct snd_soc_pcm_runtime *fe, int stream) #ifdef CONFIG_DEBUG_FS debugfs_remove(dpcm->debugfs_state); #endif
list_del(&dpcm->list_be); list_del(&dpcm->list_fe); kfree(dpcm);spin_lock(&fe->card->dpcm_lock);
spin_unlock(&fe->card->dpcm_lock);
The unlock might be moved before kfree(). Also, I don't see the list_add() spin lock protection in your patch.
Jaroslav
The dpcm_lock in this patch is to protect the instance of dpcm, e.g. protect dpcm not to be free while dereference from the be_clients/fe_clients.
The lock is not meant to protect the list "be_clients" and "fe_client", e.g. not meant to protect add/remove dpcm from these list.
The lock is added only at the places that may have race with dpcm_be_disconnect(), e.g. kfree(dpcm). And note that, many places that call for_each_dpcm_be/fe() cannot use spin_lock, since there are pcm callbacks which can sleep.