GFP_ATOMIC is used in snd_pcm_link() just because the kmalloc is
called inside a lock.  Since this function isn't too critical for
speed and is rarely called in practice, better to allocate the chunk
at first before spinlock and free it in error paths, so that
GFP_KERNEL can be used.
Signed-off-by: Takashi Iwai <tiwai@suse.de>