[SKIP SOFCI-TEST] set of PCM delay handling fixes

sound/soc/sof/ipc4-pcm.c

@@ -19,12 +19,14 @@
  * struct sof_ipc4_timestamp_info - IPC4 timestamp info
  * @host_copier: the host copier of the pcm stream
  * @dai_copier: the dai copier of the pcm stream
- * @stream_start_offset: reported by fw in memory window (converted to frames)
- * @stream_end_offset: reported by fw in memory window (converted to frames)
+ * @stream_start_offset: reported by fw in memory window (converted to
+ *                       frames at host_copier sampling rate)
+ * @stream_end_offset: reported by fw in memory window (converted to
+ *                     frames at host_copier sampling rate)
  * @llp_offset: llp offset in memory window
- * @boundary: wrap boundary should be used for the LLP frame counter
  * @delay: Calculated and stored in pointer callback. The stored value is
- *	   returned in the delay callback.
+ *         returned in the delay callback. Expressed in frames at host copier
+ *         sampling rate.
  */
 struct sof_ipc4_timestamp_info {
 	struct sof_ipc4_copier *host_copier;
@@ -33,7 +35,6 @@ struct sof_ipc4_timestamp_info {
 	u64 stream_end_offset;
 	u32 llp_offset;
 
-	u64 boundary;
 	snd_pcm_sframes_t delay;
 };
 
@@ -48,6 +49,16 @@ struct sof_ipc4_pcm_stream_priv {
 	bool chain_dma_allocated;
 };
 
+/*
+ * Modulus to use to compare host and link position counters. The sampling
+ * rates may be different, so the raw hardware counters will wrap
+ * around at different times. To calculate differences, use
+ * DELAY_BOUNDARY as a common modulus. This value must be smaller than
+ * the wrap-around point of any hardware counter, and larger than any
+ * valid delay measurement.
+ */
+#define DELAY_BOUNDARY		U32_MAX
+
 static inline struct sof_ipc4_timestamp_info *
 sof_ipc4_sps_to_time_info(struct snd_sof_pcm_stream *sps)
 {
@@ -1049,6 +1060,35 @@ static int sof_ipc4_pcm_hw_params(struct snd_soc_component *component,
 	return 0;
 }
 
+static u64 sof_ipc4_frames_dai_to_host(struct sof_ipc4_timestamp_info *time_info, u64 value)
+{
+	u64 dai_rate, host_rate;
+
+	if (!time_info->dai_copier || !time_info->host_copier)
+		return value;
+
+	/*
+	 * copiers do not change sampling rate, so we can use the
+	 * out_format independently of stream direction
+	 */
+	dai_rate = time_info->dai_copier->data.out_format.sampling_frequency;
+	host_rate = time_info->host_copier->data.out_format.sampling_frequency;
+
+	if (!dai_rate || !host_rate || dai_rate == host_rate)
+		return value;
+
+	/* take care not to overflow u64, rates can be up to 768000 */
+	if (value > U32_MAX) {
+		value = div64_u64(value, dai_rate);
+		value *= host_rate;
+	} else {
+		value *= host_rate;
+		value = div64_u64(value, dai_rate);
+	}
+
+	return value;
+}
+
 static int sof_ipc4_get_stream_start_offset(struct snd_sof_dev *sdev,
 					    struct snd_pcm_substream *substream,
 					    struct snd_sof_pcm_stream *sps,
@@ -1068,18 +1108,24 @@ static int sof_ipc4_get_stream_start_offset(struct snd_sof_dev *sdev,
 		return -EINVAL;
 	} else if (host_copier->data.gtw_cfg.node_id == SOF_IPC4_CHAIN_DMA_NODE_ID) {
 		/*
-		 * While the firmware does not supports time_info reporting for
+		 * While the firmware does not support time_info reporting for
 		 * streams using ChainDMA, it is granted that ChainDMA can only
 		 * be used on Host+Link pairs where the link position is
 		 * accessible from the host side.
 		 *
 		 * Enable delay calculation in case of ChainDMA via host
 		 * accessible registers.
 		 *
-		 * The ChainDMA uses 2x 1ms ping-pong buffer, dai side starts
-		 * when 1ms data is available
+		 * The ChainDMA prefills the link DMA with a preamble
+		 * of zero samples. Set the stream start offset based
+		 * on size of the preamble (driver provided fifo size
+		 * multiplied by 2.5). We add 1ms of margin as the FW
+		 * will align the buffer size to DMA hardware
+		 * alignment that is not known to host.
 		 */
-		time_info->stream_start_offset = substream->runtime->rate / MSEC_PER_SEC;
+		int pre_ms = SOF_IPC4_CHAIN_DMA_BUF_SIZE_MS * 5 / 2 + 1;
+
+		time_info->stream_start_offset = pre_ms * substream->runtime->rate / MSEC_PER_SEC;
 		goto out;
 	}
 
@@ -1099,14 +1145,13 @@ static int sof_ipc4_get_stream_start_offset(struct snd_sof_dev *sdev,
 	time_info->stream_end_offset = ppl_reg.stream_end_offset;
 	do_div(time_info->stream_end_offset, dai_sample_size);
 
+	/* convert to host frame time */
+	time_info->stream_start_offset =
+		sof_ipc4_frames_dai_to_host(time_info, time_info->stream_start_offset);
+	time_info->stream_end_offset =
+		sof_ipc4_frames_dai_to_host(time_info, time_info->stream_end_offset);
+
 out:
-	/*
-	 * Calculate the wrap boundary need to be used for delay calculation
-	 * The host counter is in bytes, it will wrap earlier than the frames
-	 * based link counter.
-	 */
-	time_info->boundary = div64_u64(~((u64)0),
-					frames_to_bytes(substream->runtime, 1));
 	/* Initialize the delay value to 0 (no delay) */
 	time_info->delay = 0;
 
@@ -1149,6 +1194,8 @@ static int sof_ipc4_pcm_pointer(struct snd_soc_component *component,
 
 	/* For delay calculation we need the host counter */
 	host_cnt = snd_sof_pcm_get_host_byte_counter(sdev, component, substream);
+
+	/* Store the original value to host_ptr */
 	host_ptr = host_cnt;
 
 	/* convert the host_cnt to frames */
@@ -1167,6 +1214,8 @@ static int sof_ipc4_pcm_pointer(struct snd_soc_component *component,
 		sof_mailbox_read(sdev, time_info->llp_offset, &llp, sizeof(llp));
 		dai_cnt = ((u64)llp.reading.llp_u << 32) | llp.reading.llp_l;
 	}
+
+	dai_cnt = sof_ipc4_frames_dai_to_host(time_info, dai_cnt);
 	dai_cnt += time_info->stream_end_offset;
 
 	/* In two cases dai dma counter is not accurate
@@ -1200,8 +1249,9 @@ static int sof_ipc4_pcm_pointer(struct snd_soc_component *component,
 		dai_cnt -= time_info->stream_start_offset;
 	}
 
-	/* Wrap the dai counter at the boundary where the host counter wraps */
-	div64_u64_rem(dai_cnt, time_info->boundary, &dai_cnt);
+	/* Convert to a common base before comparisons */
+	div64_u64_rem(dai_cnt, DELAY_BOUNDARY, &dai_cnt);
+	div64_u64_rem(host_cnt, DELAY_BOUNDARY, &host_cnt);
 
 	if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) {
 		head_cnt = host_cnt;
@@ -1211,14 +1261,17 @@ static int sof_ipc4_pcm_pointer(struct snd_soc_component *component,
 		tail_cnt = host_cnt;
 	}
 
-	if (head_cnt < tail_cnt) {
-		time_info->delay = time_info->boundary - tail_cnt + head_cnt;
-		goto out;
-	}
+	if (unlikely(head_cnt < tail_cnt))
+		time_info->delay = DELAY_BOUNDARY - tail_cnt + head_cnt;
+	else
+		time_info->delay = head_cnt - tail_cnt;
 
-	time_info->delay =  head_cnt - tail_cnt;
+	if (time_info->delay > (DELAY_BOUNDARY >> 1)) {
+		dev_dbg_ratelimited(sdev->dev, "inaccurate delay, host %llu dai_cnt %llu",
+				    host_cnt, dai_cnt);
+		time_info->delay = 0;
+	}
 
-out:
 	/*
 	 * Convert the host byte counter to PCM pointer which wraps in buffer
 	 * and it is in frames

sound/soc/sof/ipc4-topology.c

@@ -33,7 +33,6 @@ MODULE_PARM_DESC(ipc4_ignore_cpc,
 
 #define SOF_IPC4_GAIN_PARAM_ID  0
 #define SOF_IPC4_TPLG_ABI_SIZE 6
-#define SOF_IPC4_CHAIN_DMA_BUF_SIZE_MS 2
 
 static DEFINE_IDA(alh_group_ida);
 static DEFINE_IDA(pipeline_ida);

sound/soc/sof/ipc4-topology.h

@@ -263,6 +263,8 @@ struct sof_ipc4_dma_stream_ch_map {
 #define SOF_IPC4_DMA_METHOD_HDA   1
 #define SOF_IPC4_DMA_METHOD_GPDMA 2 /* defined for consistency but not used */
 
+#define SOF_IPC4_CHAIN_DMA_BUF_SIZE_MS 2
+
 /**
  * struct sof_ipc4_dma_config: DMA configuration
  * @dma_method: HDAudio or GPDMA