On 11/30/2016 09:36 AM, Lars-Peter Clausen wrote: > On 11/30/2016 09:30 AM, Lars-Peter Clausen wrote: >> On 11/30/2016 09:22 AM, Jiada Wang wrote: >>> From: Andreas Pape <ap...@de.adit-jv.com> >>> >>> Setting pointer and afterwards check for wrap around leads >>> to the possibility of returning the inconsistent pointer position. >>> This patch increments buffer pointer atomically to avoid this issue. >> >> Makes sense. >> >>> >>> Signed-off-by: Andreas Pape <ap...@de.adit-jv.com> >>> Signed-off-by: Jiada Wang <jiada_w...@mentor.com> >>> --- >>> sound/core/pcm_dmaengine.c | 8 +++++--- >>> 1 file changed, 5 insertions(+), 3 deletions(-) >>> >>> diff --git a/sound/core/pcm_dmaengine.c b/sound/core/pcm_dmaengine.c >>> index 8eb58c7..6f6da11 100644 >>> --- a/sound/core/pcm_dmaengine.c >>> +++ b/sound/core/pcm_dmaengine.c >>> @@ -139,12 +139,14 @@ >>> EXPORT_SYMBOL_GPL(snd_dmaengine_pcm_set_config_from_dai_data); >>> >>> static void dmaengine_pcm_dma_complete(void *arg) >>> { >>> + unsigned int new_pos; >>> struct snd_pcm_substream *substream = arg; >>> struct dmaengine_pcm_runtime_data *prtd = substream_to_prtd(substream); >>> >>> - prtd->pos += snd_pcm_lib_period_bytes(substream); >>> - if (prtd->pos >= snd_pcm_lib_buffer_bytes(substream)) >>> - prtd->pos = 0; >>> + new_pos = prtd->pos + snd_pcm_lib_period_bytes(substream); >>> + if (new_pos >= snd_pcm_lib_buffer_bytes(substream)) >>> + new_pos = 0; >>> + prtd->pos = new_pos; >> >> But to really make it atomic I think this needs READ_ONCE/WRITE_ONCE. > > And the access to prtd->pos in snd_dmaengine_pcm_pointer_no_residue() should > also use READ_ONCE(). It is very unlikely that the code gets mis-compiled to > generate more than one access, but having READ_ONCE() acts as a annotation > that makes it explicit that this is data that can be updated concurrently > without further synchronization.
Having given this some additional thoughts, I think a READ_ONCE() in dmaengine_pcm_dma_complete() is not necessary. dmaengine_pcm_dma_complete() is the only writer of prtd->pos and it is not running concurrently to itself. So we'll always observe consistent state, even if the compiler decides to issue multiple reads. The WRITE_ONCE() is required to make sure that the prtd->pos state stays consistent to concurrent readers. And the READ_ONCE() in snd_dmaengine_pcm_pointer_no_residue() is required to make sure that consistent state is observed from concurrent writers.