#include <linux/gfp.h>
#include <linux/of_platform.h>
#include <linux/list.h>
+#include <linux/slab.h>
#include <sound/core.h>
#include <sound/pcm.h>
SNDRV_PCM_RATE_CONTINUOUS)
struct dma_object {
- struct list_head list;
struct snd_soc_platform_driver dai;
dma_addr_t ssi_stx_phys;
dma_addr_t ssi_srx_phys;
+ unsigned int ssi_fifo_depth;
struct ccsr_dma_channel __iomem *channel;
unsigned int irq;
bool assigned;
unsigned int irq;
struct snd_pcm_substream *substream;
dma_addr_t ssi_sxx_phys;
+ unsigned int ssi_fifo_depth;
dma_addr_t ld_buf_phys;
unsigned int current_link;
dma_addr_t dma_buf_phys;
struct fsl_dma_link_descriptor *link =
&dma_private->link[dma_private->current_link];
- /* Update our link descriptors to point to the next period */
- if (dma_private->substream->stream == SNDRV_PCM_STREAM_PLAYBACK)
- link->source_addr =
- cpu_to_be32(dma_private->dma_buf_next);
- else
- link->dest_addr =
- cpu_to_be32(dma_private->dma_buf_next);
+ /* Update our link descriptors to point to the next period. On a 36-bit
+ * system, we also need to update the ESAD bits. We also set (keep) the
+ * snoop bits. See the comments in fsl_dma_hw_params() about snooping.
+ */
+ if (dma_private->substream->stream == SNDRV_PCM_STREAM_PLAYBACK) {
+ link->source_addr = cpu_to_be32(dma_private->dma_buf_next);
+#ifdef CONFIG_PHYS_64BIT
+ link->source_attr = cpu_to_be32(CCSR_DMA_ATR_SNOOP |
+ upper_32_bits(dma_private->dma_buf_next));
+#endif
+ } else {
+ link->dest_addr = cpu_to_be32(dma_private->dma_buf_next);
+#ifdef CONFIG_PHYS_64BIT
+ link->dest_attr = cpu_to_be32(CCSR_DMA_ATR_SNOOP |
+ upper_32_bits(dma_private->dma_buf_next));
+#endif
+ }
/* Update our variables for next time */
dma_private->dma_buf_next += dma_private->period_size;
* This function is called when the codec driver calls snd_soc_new_pcms(),
* once for each .dai_link in the machine driver's snd_soc_card
* structure.
+ *
+ * snd_dma_alloc_pages() is just a front-end to dma_alloc_coherent(), which
+ * (currently) always allocates the DMA buffer in lowmem, even if GFP_HIGHMEM
+ * is specified. Therefore, any DMA buffers we allocate will always be in low
+ * memory, but we support for 36-bit physical addresses anyway.
+ *
+ * Regardless of where the memory is actually allocated, since the device can
+ * technically DMA to any 36-bit address, we do need to set the DMA mask to 36.
*/
static int fsl_dma_new(struct snd_card *card, struct snd_soc_dai *dai,
struct snd_pcm *pcm)
{
- static u64 fsl_dma_dmamask = DMA_BIT_MASK(32);
+ static u64 fsl_dma_dmamask = DMA_BIT_MASK(36);
int ret;
if (!card->dev->dma_mask)
if (!card->dev->coherent_dma_mask)
card->dev->coherent_dma_mask = fsl_dma_dmamask;
- ret = snd_dma_alloc_pages(SNDRV_DMA_TYPE_DEV, card->dev,
- fsl_dma_hardware.buffer_bytes_max,
- &pcm->streams[0].substream->dma_buffer);
- if (ret) {
- dev_err(card->dev, "can't allocate playback dma buffer\n");
- return ret;
+ /* Some codecs have separate DAIs for playback and capture, so we
+ * should allocate a DMA buffer only for the streams that are valid.
+ */
+
+ if (dai->driver->playback.channels_min) {
+ ret = snd_dma_alloc_pages(SNDRV_DMA_TYPE_DEV, card->dev,
+ fsl_dma_hardware.buffer_bytes_max,
+ &pcm->streams[0].substream->dma_buffer);
+ if (ret) {
+ dev_err(card->dev, "can't alloc playback dma buffer\n");
+ return ret;
+ }
}
- ret = snd_dma_alloc_pages(SNDRV_DMA_TYPE_DEV, card->dev,
- fsl_dma_hardware.buffer_bytes_max,
- &pcm->streams[1].substream->dma_buffer);
- if (ret) {
- snd_dma_free_pages(&pcm->streams[0].substream->dma_buffer);
- dev_err(card->dev, "can't allocate capture dma buffer\n");
- return ret;
+ if (dai->driver->capture.channels_min) {
+ ret = snd_dma_alloc_pages(SNDRV_DMA_TYPE_DEV, card->dev,
+ fsl_dma_hardware.buffer_bytes_max,
+ &pcm->streams[1].substream->dma_buffer);
+ if (ret) {
+ snd_dma_free_pages(&pcm->streams[0].substream->dma_buffer);
+ dev_err(card->dev, "can't alloc capture dma buffer\n");
+ return ret;
+ }
}
return 0;
else
dma_private->ssi_sxx_phys = dma->ssi_srx_phys;
+ dma_private->ssi_fifo_depth = dma->ssi_fifo_depth;
dma_private->dma_channel = dma->channel;
dma_private->irq = dma->irq;
dma_private->substream = substream;
struct device *dev = rtd->platform->dev;
/* Number of bits per sample */
- unsigned int sample_size =
+ unsigned int sample_bits =
snd_pcm_format_physical_width(params_format(hw_params));
/* Number of bytes per frame */
- unsigned int frame_size = 2 * (sample_size / 8);
+ unsigned int sample_bytes = sample_bits / 8;
/* Bus address of SSI STX register */
dma_addr_t ssi_sxx_phys = dma_private->ssi_sxx_phys;
* that offset here. While we're at it, also tell the DMA controller
* how much data to transfer per sample.
*/
- switch (sample_size) {
+ switch (sample_bits) {
case 8:
mr |= CCSR_DMA_MR_DAHTS_1 | CCSR_DMA_MR_SAHTS_1;
ssi_sxx_phys += 3;
break;
default:
/* We should never get here */
- dev_err(dev, "unsupported sample size %u\n", sample_size);
+ dev_err(dev, "unsupported sample size %u\n", sample_bits);
return -EINVAL;
}
/*
- * BWC should always be a multiple of the frame size. BWC determines
- * how many bytes are sent/received before the DMA controller checks the
- * SSI to see if it needs to stop. For playback, the transmit FIFO can
- * hold three frames, so we want to send two frames at a time. For
- * capture, the receive FIFO is triggered when it contains one frame, so
- * we want to receive one frame at a time.
+ * BWC determines how many bytes are sent/received before the DMA
+ * controller checks the SSI to see if it needs to stop. BWC should
+ * always be a multiple of the frame size, so that we always transmit
+ * whole frames. Each frame occupies two slots in the FIFO. The
+ * parameter for CCSR_DMA_MR_BWC() is rounded down the next power of two
+ * (MR[BWC] can only represent even powers of two).
+ *
+ * To simplify the process, we set BWC to the largest value that is
+ * less than or equal to the FIFO watermark. For playback, this ensures
+ * that we transfer the maximum amount without overrunning the FIFO.
+ * For capture, this ensures that we transfer the maximum amount without
+ * underrunning the FIFO.
+ *
+ * f = SSI FIFO depth
+ * w = SSI watermark value (which equals f - 2)
+ * b = DMA bandwidth count (in bytes)
+ * s = sample size (in bytes, which equals frame_size * 2)
+ *
+ * For playback, we never transmit more than the transmit FIFO
+ * watermark, otherwise we might write more data than the FIFO can hold.
+ * The watermark is equal to the FIFO depth minus two.
+ *
+ * For capture, two equations must hold:
+ * w > f - (b / s)
+ * w >= b / s
+ *
+ * So, b > 2 * s, but b must also be <= s * w. To simplify, we set
+ * b = s * w, which is equal to
+ * (dma_private->ssi_fifo_depth - 2) * sample_bytes.
*/
- if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK)
- mr |= CCSR_DMA_MR_BWC(2 * frame_size);
- else
- mr |= CCSR_DMA_MR_BWC(frame_size);
+ mr |= CCSR_DMA_MR_BWC((dma_private->ssi_fifo_depth - 2) * sample_bytes);
out_be32(&dma_channel->mr, mr);
link->count = cpu_to_be32(period_size);
- /* Even though the DMA controller supports 36-bit addressing,
- * for simplicity we allow only 32-bit addresses for the audio
- * buffer itself. This was enforced in fsl_dma_new() with the
- * DMA mask.
- *
- * The snoop bit tells the DMA controller whether it should tell
+ /* The snoop bit tells the DMA controller whether it should tell
* the ECM to snoop during a read or write to an address. For
* audio, we use DMA to transfer data between memory and an I/O
* device (the SSI's STX0 or SRX0 register). Snooping is only
* flush out the data for the previous period. So if you
* increased period_bytes_min to a large enough size, you might
* get more performance by not snooping, and you'll still be
- * okay.
+ * okay. You'll need to update fsl_dma_update_pointers() also.
*/
if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) {
link->source_addr = cpu_to_be32(temp_addr);
- link->source_attr = cpu_to_be32(CCSR_DMA_ATR_SNOOP);
+ link->source_attr = cpu_to_be32(CCSR_DMA_ATR_SNOOP |
+ upper_32_bits(temp_addr));
link->dest_addr = cpu_to_be32(ssi_sxx_phys);
- link->dest_attr = cpu_to_be32(CCSR_DMA_ATR_NOSNOOP);
+ link->dest_attr = cpu_to_be32(CCSR_DMA_ATR_NOSNOOP |
+ upper_32_bits(ssi_sxx_phys));
} else {
link->source_addr = cpu_to_be32(ssi_sxx_phys);
- link->source_attr = cpu_to_be32(CCSR_DMA_ATR_NOSNOOP);
+ link->source_attr = cpu_to_be32(CCSR_DMA_ATR_NOSNOOP |
+ upper_32_bits(ssi_sxx_phys));
link->dest_addr = cpu_to_be32(temp_addr);
- link->dest_attr = cpu_to_be32(CCSR_DMA_ATR_SNOOP);
+ link->dest_attr = cpu_to_be32(CCSR_DMA_ATR_SNOOP |
+ upper_32_bits(temp_addr));
}
temp_addr += period_size;
dma_addr_t position;
snd_pcm_uframes_t frames;
- if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK)
+ /* Obtain the current DMA pointer, but don't read the ESAD bits if we
+ * only have 32-bit DMA addresses. This function is typically called
+ * in interrupt context, so we need to optimize it.
+ */
+ if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) {
position = in_be32(&dma_channel->sar);
- else
+#ifdef CONFIG_PHYS_64BIT
+ position |= (u64)(in_be32(&dma_channel->satr) &
+ CCSR_DMA_ATR_ESAD_MASK) << 32;
+#endif
+ } else {
position = in_be32(&dma_channel->dar);
+#ifdef CONFIG_PHYS_64BIT
+ position |= (u64)(in_be32(&dma_channel->datr) &
+ CCSR_DMA_ATR_ESAD_MASK) << 32;
+#endif
+ }
/*
* When capture is started, the SSI immediately starts to fill its FIFO.
}
}
-/* List of DMA nodes that we've probed */
-static LIST_HEAD(dma_list);
-
/**
* find_ssi_node -- returns the SSI node that points to his DMA channel node
*
.pointer = fsl_dma_pointer,
};
-static int __devinit fsl_soc_dma_probe(struct of_device *of_dev,
+static int __devinit fsl_soc_dma_probe(struct platform_device *pdev,
const struct of_device_id *match)
{
struct dma_object *dma;
- struct device_node *np = of_dev->dev.of_node;
+ struct device_node *np = pdev->dev.of_node;
struct device_node *ssi_np;
struct resource res;
+ const uint32_t *iprop;
int ret;
/* Find the SSI node that points to us. */
ssi_np = find_ssi_node(np);
if (!ssi_np) {
- dev_err(&of_dev->dev, "cannot find parent SSI node\n");
+ dev_err(&pdev->dev, "cannot find parent SSI node\n");
return -ENODEV;
}
ret = of_address_to_resource(ssi_np, 0, &res);
- of_node_put(ssi_np);
if (ret) {
- dev_err(&of_dev->dev, "could not determine device resources\n");
+ dev_err(&pdev->dev, "could not determine resources for %s\n",
+ ssi_np->full_name);
+ of_node_put(ssi_np);
return ret;
}
dma = kzalloc(sizeof(*dma) + strlen(np->full_name), GFP_KERNEL);
if (!dma) {
- dev_err(&of_dev->dev, "could not allocate dma object\n");
+ dev_err(&pdev->dev, "could not allocate dma object\n");
+ of_node_put(ssi_np);
return -ENOMEM;
}
dma->ssi_stx_phys = res.start + offsetof(struct ccsr_ssi, stx0);
dma->ssi_srx_phys = res.start + offsetof(struct ccsr_ssi, srx0);
- ret = snd_soc_register_platform(&of_dev->dev, &dma->dai);
+ iprop = of_get_property(ssi_np, "fsl,fifo-depth", NULL);
+ if (iprop)
+ dma->ssi_fifo_depth = *iprop;
+ else
+ /* Older 8610 DTs didn't have the fifo-depth property */
+ dma->ssi_fifo_depth = 8;
+
+ of_node_put(ssi_np);
+
+ ret = snd_soc_register_platform(&pdev->dev, &dma->dai);
if (ret) {
- dev_err(&of_dev->dev, "could not register platform\n");
+ dev_err(&pdev->dev, "could not register platform\n");
kfree(dma);
return ret;
}
dma->channel = of_iomap(np, 0);
dma->irq = irq_of_parse_and_map(np, 0);
- list_add(&dma->list, &dma_list);
+
+ dev_set_drvdata(&pdev->dev, dma);
return 0;
}
-static int __devexit fsl_soc_dma_remove(struct of_device *of_dev)
+static int __devexit fsl_soc_dma_remove(struct platform_device *pdev)
{
- struct list_head *n, *ptr;
- struct dma_object *dma;
+ struct dma_object *dma = dev_get_drvdata(&pdev->dev);
- list_for_each_safe(ptr, n, &dma_list) {
- dma = list_entry(ptr, struct dma_object, list);
- list_del_init(ptr);
-
- snd_soc_unregister_platform(&of_dev->dev);
- iounmap(dma->channel);
- irq_dispose_mapping(dma->irq);
- kfree(dma);
- }
+ snd_soc_unregister_platform(&pdev->dev);
+ iounmap(dma->channel);
+ irq_dispose_mapping(dma->irq);
+ kfree(dma);
return 0;
}
of_unregister_platform_driver(&fsl_soc_dma_driver);
}
-/* We want the DMA driver to be initialized before the SSI driver, so that
- * when the SSI driver calls fsl_soc_dma_dai_from_node(), the DMA driver
- * will already have been probed. The easiest way to do that is to make the
- * __init function called via arch_initcall().
- */
module_init(fsl_soc_dma_init);
module_exit(fsl_soc_dma_exit);
projects / ~andy / linux / blobdiff