Re: [PATCH v4 2/3] ALSA: compress_offload: Add SNDRV_COMPRESS_TSTAMP64 ioctl

[Takashi Iwai <tiwai@suse.de>]

On Tue, 05 Aug 2025 06:47:59 +0200,
Vinod Koul wrote:
> 
> On 01-08-25, 10:27, Joris Verhaegen wrote:
> > The previous patch introduced the internal infrastructure for handling
> > 64-bit timestamps. This patch exposes this capability to user-space.
> > 
> > Define the new ioctl command SNDRV_COMPRESS_TSTAMP64, which allows
> > applications to fetch the overflow-safe struct snd_compr_tstamp64.
> > 
> > The ioctl dispatch table is updated to handle the new command by
> > calling a new snd_compr_tstamp64 handler, while the legacy path is
> > renamed to snd_compr_tstamp32 for clarity.
> > 
> > This patch bumps the SNDRV_COMPRESS_VERSION to 0.4.0.
> > 
> > Reviewed-by: Miller Liang <millerliang@google.com>
> > Tested-by: Joris Verhaegen <verhaegen@google.com>
> > Signed-off-by: Joris Verhaegen <verhaegen@google.com>
> > ---
> >  include/uapi/sound/compress_offload.h |  5 +++--
> >  sound/core/compress_offload.c         | 19 +++++++++++++------
> >  2 files changed, 16 insertions(+), 8 deletions(-)
> > 
> > diff --git a/include/uapi/sound/compress_offload.h b/include/uapi/sound/compress_offload.h
> > index abd0ea3f86ee..70b8921601f9 100644
> > --- a/include/uapi/sound/compress_offload.h
> > +++ b/include/uapi/sound/compress_offload.h
> > @@ -13,8 +13,7 @@
> >  #include <sound/asound.h>
> >  #include <sound/compress_params.h>
> >  
> > -
> > -#define SNDRV_COMPRESS_VERSION SNDRV_PROTOCOL_VERSION(0, 3, 0)
> > +#define SNDRV_COMPRESS_VERSION SNDRV_PROTOCOL_VERSION(0, 4, 0)
> >  /**
> >   * struct snd_compressed_buffer - compressed buffer
> >   * @fragment_size: size of buffer fragment in bytes
> > @@ -208,6 +207,7 @@ struct snd_compr_task_status {
> >   * Note: only codec params can be changed runtime and stream params cant be
> >   * SNDRV_COMPRESS_GET_PARAMS: Query codec params
> >   * SNDRV_COMPRESS_TSTAMP: get the current timestamp value
> > + * SNDRV_COMPRESS_TSTAMP64: get the current timestamp value in 64 bit format
> >   * SNDRV_COMPRESS_AVAIL: get the current buffer avail value.
> >   * This also queries the tstamp properties
> >   * SNDRV_COMPRESS_PAUSE: Pause the running stream
> > @@ -230,6 +230,7 @@ struct snd_compr_task_status {
> >  						 struct snd_compr_metadata)
> >  #define SNDRV_COMPRESS_TSTAMP		_IOR('C', 0x20, struct snd_compr_tstamp)
> >  #define SNDRV_COMPRESS_AVAIL		_IOR('C', 0x21, struct snd_compr_avail)
> > +#define SNDRV_COMPRESS_TSTAMP64		_IOR('C', 0x22, struct snd_compr_tstamp64)
> >  #define SNDRV_COMPRESS_PAUSE		_IO('C', 0x30)
> >  #define SNDRV_COMPRESS_RESUME		_IO('C', 0x31)
> >  #define SNDRV_COMPRESS_START		_IO('C', 0x32)
> > diff --git a/sound/core/compress_offload.c b/sound/core/compress_offload.c
> > index d3164aa07158..445220fdb6a0 100644
> > --- a/sound/core/compress_offload.c
> > +++ b/sound/core/compress_offload.c
> > @@ -736,18 +736,23 @@ snd_compr_set_metadata(struct snd_compr_stream *stream, unsigned long arg)
> >  	return retval;
> >  }
> >  
> > -static inline int
> > -snd_compr_tstamp(struct snd_compr_stream *stream, unsigned long arg)
> > +static inline int snd_compr_tstamp(struct snd_compr_stream *stream,
> > +				   unsigned long arg, bool is_32bit)
> >  {
> >  	struct snd_compr_tstamp64 tstamp64 = { 0 };
> >  	struct snd_compr_tstamp tstamp32 = { 0 };
> > +	const void *copy_from = &tstamp64;
> > +	size_t copy_size = sizeof(tstamp64);
> >  	int ret;
> >  
> >  	ret = snd_compr_update_tstamp(stream, &tstamp64);
> >  	if (ret == 0) {
> > -		snd_compr_tstamp32_from_64(&tstamp32, &tstamp64);
> > -		ret = copy_to_user((struct snd_compr_tstamp __user *)arg,
> > -				   &tstamp32, sizeof(tstamp32)) ?
> > +		if (is_32bit) {
> > +			snd_compr_tstamp32_from_64(&tstamp32, &tstamp64);
> > +			copy_from = &tstamp32;
> > +			copy_size = sizeof(tstamp32);
> > +		}
> 
> Most of the applications and people would be 32bit right now and we
> expect this to progressively change, but then this imposes a penalty as
> default path is 64 bit, since we expect this ioctl to be called very
> frequently, should we do this optimization for 64bit here?

Through a quick glance over the patch, I don't think you'll hit the
significant performance loss.  It's merely a few bytes of extra copies
before copy_to_user(), after all.  But, of course, it'd be more
convincing if anyone can test and give the actual numbers.


thanks,

Takashi