On 28-03-2022 20:41, Amadeusz Sławiński wrote:
On 3/28/2022 8:14 AM, Sameer Pujar wrote:
For DPCM links, the order of hw_param() call depends on the sequence of BE connection to FE. It is possible that one BE link can provide clock to another BE link. In such cases consumer BE DAI, to get the rate set by provider BE DAI, can use the standard clock functions only if provider has already set the appropriate rate during its hw_param() stage.
Above sentence seems to suggest that consumer can set clock only after provider has started, but code in this patch seems to do it the other way around?
This patch makes provider calls to happen first.
Presently the order is fixed and does not depend on the provider and consumer relationships. So the clock rates need to be known ahead of hw_param() stage.
This patch tweaks the hw_param() order by connecting the provider BEs late to a FE. With this hw_param() calls for provider BEs happen first and then followed by consumer BEs. The consumers can use the standard clk_get_rate() function to get the rate of the clock they depend on.
I'm bit confused by " With this hw_param() calls for provider BEs happen first and then followed by consumer BEs. "
Aren't consumers started first and provider second? Code and previous sentence "connecting the provider BEs late to a FE" confuse me.
The dpcm_be_connect() call adds the new connection to a list using list_add() which would be a stack. When dpcm_be_connect() is deferred for provider BEs, these occupy top of the stack. When operating on this list during hw_params() stage, the provider call happen first. Is this part clear now? I can rephrase the comments/commit message for more clarity.
Overall I don't exactly understand correct order of events after reading commit message and patch...
Consider there are two BEs (BE1 and BE2) and "BE1<==>BE2" can be an I2S interface for example. I am trying to get following sequence.
1. When BE1 is provider and BE2 is consumer, the call sequence expected is : hw_params(BE1) -> hw_params(BE2).
2. When BE2 is provider and BE1 is consumer, the call sequence expected is : hw_params(BE2) -> hw_params(BE1).
Idea is to make use of standard clock functions for rate info. Provider can use clk_set_rate() and consumer can clk_get_rate().
Signed-off-by: Sameer Pujar spujar@nvidia.com
TODO: * The FE link is not considered in this. For Tegra it is fine to call hw_params() for FE at the end. But systems, which want to apply this tweak for FE as well, have to extend this tweak to FE. * Also only DPCM is considered here. If normal links require such tweak, it needs to be extended.
sound/soc/soc-pcm.c | 60 ++++++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 59 insertions(+), 1 deletion(-)
diff --git a/sound/soc/soc-pcm.c b/sound/soc/soc-pcm.c index 9a95468..5829514 100644 --- a/sound/soc/soc-pcm.c +++ b/sound/soc/soc-pcm.c @@ -1442,6 +1442,29 @@ static int dpcm_prune_paths(struct snd_soc_pcm_runtime *fe, int stream, return prune; }
+static bool defer_dpcm_be_connect(struct snd_soc_pcm_runtime *rtd) +{ + struct snd_soc_dai *dai; + int i;
+ if (!(rtd->dai_link->dai_fmt & SND_SOC_DAIFMT_FORMAT_MASK)) + return false;
+ if ((rtd->dai_link->dai_fmt & SND_SOC_DAIFMT_CLOCK_PROVIDER_MASK) == + SND_SOC_DAIFMT_CBC_CFC) {
+ for_each_rtd_cpu_dais(rtd, i, dai) {
+ if (!snd_soc_dai_is_dummy(dai)) + return true; + } + }
+ return false; +}
+#define MAX_CLK_PROVIDER_BE 10
Not sure about this define, it adds unnecessary limitation on max clock number, can't you just run same loop twice while checking defer_dpcm_be_connect() first time and !defer_dpcm_be_connect() second time? defer_dpcm_be_connect() function name may need a bit of adjustment (rtd_is_clock_consumer() maybe?), but it gets rid of the limit.
or do something like following instead of copy pasting loop twice:
rename original dpcm_add_paths() to _dpcm_add_paths() and add additional argument and check somewhere inline: static int dpcm_add_paths(struct snd_soc_pcm_runtime *fe, int stream, struct snd_soc_dapm_widget_list **list_, bool clock_consumer) { ...
// with renamed defer_dpcm_be_connect if (clock_consumer ^ !rtd_is_clock_consumer()) continue;
... }
static int dpcm_add_paths(struct snd_soc_pcm_runtime *fe, int stream, struct snd_soc_dapm_widget_list **list_) { int ret;
/* start clock consumer BEs */ ret = _dpcm_add_paths(*fe, stream, **list_, true); if (ret) return ret;
/* start clock provider BEs */ ret = _dpcm_add_paths(*fe, stream, **list_, false);
return ret; }
Thanks for the suggestion. I will check if loop copy can be avoided.
static int dpcm_add_paths(struct snd_soc_pcm_runtime *fe, int stream, struct snd_soc_dapm_widget_list **list_) { @@ -1449,7 +1472,8 @@ static int dpcm_add_paths(struct snd_soc_pcm_runtime *fe, int stream, struct snd_soc_dapm_widget_list *list = *list_; struct snd_soc_pcm_runtime *be; struct snd_soc_dapm_widget *widget; - int i, new = 0, err; + struct snd_soc_pcm_runtime *prov[MAX_CLK_PROVIDER_BE]; + int i, new = 0, err, count = 0;
/* Create any new FE <--> BE connections */ for_each_dapm_widgets(list, i, widget) { @@ -1489,6 +1513,40 @@ static int dpcm_add_paths(struct snd_soc_pcm_runtime *fe, int stream, (be->dpcm[stream].state != SND_SOC_DPCM_STATE_CLOSE)) continue;
+ /* Connect clock provider BEs at the end */ + if (defer_dpcm_be_connect(be)) { + if (count >= MAX_CLK_PROVIDER_BE) { + dev_err(fe->dev, "ASoC: too many clock provider BEs\n"); + return -EINVAL; + }
+ prov[count++] = be; + continue; + }
+ /* newly connected FE and BE */ + err = dpcm_be_connect(fe, be, stream); + if (err < 0) { + dev_err(fe->dev, "ASoC: can't connect %s\n", + widget->name); + break; + } else if (err == 0) /* already connected */ + continue;
+ /* new */ + dpcm_set_be_update_state(be, stream, SND_SOC_DPCM_UPDATE_BE); + new++; + }
+ /* + * Now connect clock provider BEs. A late connection means, + * these BE links appear first in the list maintained by FE + * and hw_param() call for these happen first.
Let's stick to ALSA terminology, hw_params() please, same in commit message.
Sorry about this. I will fix it.