Re: [RESEND PATCH v2 07/15] ASoC: qcom: q6asm: Add support to memory map and unmap

[Bjorn Andersson <bjorn.andersson@linaro.org>]

On Thu 14 Dec 09:33 PST 2017, srinivas.kandagatla@linaro.org wrote:

> From: Srinivas Kandagatla <srinivas.kandagatla@linaro.org>
> 
> This patch adds support to memory map and unmap regions commands in
> q6asm module.
> 
> Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@linaro.org>
> ---
>  sound/soc/qcom/qdsp6/q6asm.c | 343 ++++++++++++++++++++++++++++++++++++++++++-
>  sound/soc/qcom/qdsp6/q6asm.h |   5 +
>  2 files changed, 347 insertions(+), 1 deletion(-)
> 
> diff --git a/sound/soc/qcom/qdsp6/q6asm.c b/sound/soc/qcom/qdsp6/q6asm.c
> index 9cc583afef4d..4be92441f524 100644
> --- a/sound/soc/qcom/qdsp6/q6asm.c
> +++ b/sound/soc/qcom/qdsp6/q6asm.c
> @@ -14,9 +14,46 @@
>  #include "q6asm.h"
>  #include "common.h"
>  
> +#define ASM_CMD_SHARED_MEM_MAP_REGIONS		0x00010D92
> +#define ASM_CMDRSP_SHARED_MEM_MAP_REGIONS	0x00010D93
> +#define ASM_CMD_SHARED_MEM_UNMAP_REGIONS	0x00010D94
> +
>  #define TUN_READ_IO_MODE		0x0004	/* tunnel read write mode */
>  #define SYNC_IO_MODE			0x0001
>  #define ASYNC_IO_MODE			0x0002
> +#define ASM_SHIFT_GAPLESS_MODE_FLAG	31
> +#define ADSP_MEMORY_MAP_SHMEM8_4K_POOL	3
> +
> +struct avs_cmd_shared_mem_map_regions {
> +	struct apr_hdr hdr;
> +	u16 mem_pool_id;
> +	u16 num_regions;
> +	u32 property_flag;
> +} __packed;
> +
> +struct avs_shared_map_region_payload {
> +	u32 shm_addr_lsw;
> +	u32 shm_addr_msw;
> +	u32 mem_size_bytes;
> +} __packed;
> +
> +struct avs_cmd_shared_mem_unmap_regions {
> +	struct apr_hdr hdr;
> +	u32 mem_map_handle;
> +} __packed;
> +
> +struct audio_buffer {
> +	dma_addr_t phys;
> +	uint32_t used;
> +	uint32_t size;		/* size of buffer */
> +};
> +
> +struct audio_port_data {
> +	struct audio_buffer *buf;
> +	uint32_t max_buf_cnt;

This seems to denote the number of audio_buffers in the buf array, so
I'm not sure about the meaning of "max_".

> +	uint32_t dsp_buf;
> +	uint32_t mem_map_handle;
> +};
>  
>  struct audio_client {
>  	int session;
> @@ -27,6 +64,8 @@ struct audio_client {
>  	uint64_t time_stamp;
>  	struct apr_device *adev;
>  	struct mutex cmd_lock;
> +	/* idx:1 out port, 0: in port */
> +	struct audio_port_data port[2];
>  	wait_queue_head_t cmd_wait;
>  	int perf_mode;
>  	int stream_id;
> @@ -86,6 +125,260 @@ static void q6asm_session_free(struct audio_client *ac)
>  	mutex_unlock(&a->session_lock);
>  }
>  
> +static inline void q6asm_add_mmaphdr(struct audio_client *ac,
> +				     struct apr_hdr *hdr, u32 pkt_size,
> +				     bool cmd_flg, u32 token)

cmd_flg is true in both callers, so this function ends up simply
assigning hdr_field, pkt_size and token. Inlining these assignments
would make for cleaner call sites as well.

> +{
> +	hdr->hdr_field = APR_SEQ_CMD_HDR_FIELD;
> +	hdr->src_port = 0;
> +	hdr->dest_port = 0;
> +	hdr->pkt_size = pkt_size;
> +	if (cmd_flg)
> +		hdr->token = token;
> +}
> +
> +static inline void q6asm_add_hdr(struct audio_client *ac, struct apr_hdr *hdr,

This is unused.

> +				 uint32_t pkt_size, bool cmd_flg,
> +				 uint32_t stream_id)
> +{
> +	hdr->hdr_field = APR_SEQ_CMD_HDR_FIELD;
> +	hdr->src_svc = ac->adev->svc_id;
> +	hdr->src_domain = APR_DOMAIN_APPS;
> +	hdr->dest_svc = APR_SVC_ASM;
> +	hdr->dest_domain = APR_DOMAIN_ADSP;
> +	hdr->src_port = ((ac->session << 8) & 0xFF00) | (stream_id);
> +	hdr->dest_port = ((ac->session << 8) & 0xFF00) | (stream_id);
> +	hdr->pkt_size = pkt_size;
> +	if (cmd_flg)
> +		hdr->token = ac->session;
> +}
> +
> +static int __q6asm_memory_unmap(struct audio_client *ac,
> +				phys_addr_t buf_add, int dir)
> +{
> +	struct avs_cmd_shared_mem_unmap_regions mem_unmap;

If you name this "cmd" you will declutter below code a bit.

> +	struct q6asm *a = dev_get_drvdata(ac->dev->parent);
> +	int rc;
> +
> +	if (!a)
> +		return -ENODEV;

Does this NULL check add any real value?

> +
> +	q6asm_add_mmaphdr(ac, &mem_unmap.hdr, sizeof(mem_unmap), true,
> +			  ((ac->session << 8) | dir));
> +	a->mem_state = -1;
> +
> +	mem_unmap.hdr.opcode = ASM_CMD_SHARED_MEM_UNMAP_REGIONS;
> +	mem_unmap.mem_map_handle = ac->port[dir].mem_map_handle;
> +
> +	if (mem_unmap.mem_map_handle == 0) {

Start the function by checking for !ac->port[dir].mem_map_handle

> +		dev_err(ac->dev, "invalid mem handle\n");
> +		return -EINVAL;
> +	}
> +
> +	rc = apr_send_pkt(a->adev, (uint32_t *) &mem_unmap);
> +	if (rc < 0)
> +		return rc;
> +
> +	rc = wait_event_timeout(a->mem_wait, (a->mem_state >= 0),
> +				5 * HZ);
> +	if (!rc) {
> +		dev_err(ac->dev, "CMD timeout for memory_unmap 0x%x\n",
> +			mem_unmap.mem_map_handle);
> +		return -ETIMEDOUT;
> +	} else if (a->mem_state > 0) {
> +		return adsp_err_get_lnx_err_code(a->mem_state);
> +	}
> +	ac->port[dir].mem_map_handle = 0;

Does all errors indicate that the region is left mapped?

> +
> +	return 0;
> +}
> +
> +/**
> + * q6asm_unmap_memory_regions() - unmap memory regions in the dsp.
> + *
> + * @dir: direction of audio stream
> + * @ac: audio client instanace
> + *
> + * Return: Will be an negative value on failure or zero on success
> + */
> +int q6asm_unmap_memory_regions(unsigned int dir, struct audio_client *ac)
> +{
> +	struct audio_port_data *port;
> +	int cnt = 0;
> +	int rc = 0;
> +
> +	mutex_lock(&ac->cmd_lock);
> +	port = &ac->port[dir];
> +	if (!port->buf) {
> +		mutex_unlock(&ac->cmd_lock);
> +		return 0;

Put a label right before the mutex_unlock below and return rc instead of
0, then you can replace these two lines with "goto unlock"

> +	}
> +	cnt = port->max_buf_cnt - 1;

What if we mapped 1 period? Why shouldn't we enter the unmap path?

> +	if (cnt >= 0) {
> +		rc = __q6asm_memory_unmap(ac, port->buf[dir].phys, dir);
> +		if (rc < 0) {
> +			dev_err(ac->dev, "%s: Memory_unmap_regions failed %d\n",
> +				__func__, rc);

Most of the code paths through __q6asm_memory_unmap() will print an
error, make this consistent and print the warning once.

> +			mutex_unlock(&ac->cmd_lock);
> +			return rc;

Same here.

> +		}
> +	}
> +
> +	port->max_buf_cnt = 0;
> +	kfree(port->buf);
> +	port->buf = NULL;
> +	mutex_unlock(&ac->cmd_lock);

I think however that if you rearrange this function somewhat you can
start off by doing:

	mutex_lock(&ac->cmd_lock);
	port = &ac->port[dir];

	bufs = port->buf;
	cnt = port->max_buf_cnt;

	port->max_buf_cnt = 0;
	port->buf = NULL;
	mutex_unlock(&ac->cmd_lock);

And then you can do the rest without locks.

> +
> +	return 0;
> +}
> +EXPORT_SYMBOL_GPL(q6asm_unmap_memory_regions);
> +
> +static int __q6asm_memory_map_regions(struct audio_client *ac, int dir,
> +				      uint32_t period_sz, uint32_t periods,

period_sz is typical size_t material.

> +				      bool is_contiguous)
> +{
> +	struct avs_cmd_shared_mem_map_regions *mmap_regions = NULL;

Calling this "cmd" would declutter the function.

> +	struct avs_shared_map_region_payload *mregions = NULL;
> +	struct q6asm *a = dev_get_drvdata(ac->dev->parent);
> +	struct audio_port_data *port = NULL;
> +	struct audio_buffer *ab = NULL;
> +	void *mmap_region_cmd = NULL;

No need to initialize this.

Also, this really is a avs_cmd_shared_mem_map_regions with some extra
data at the end, not a void *.

> +	void *payload = NULL;
> +	int rc = 0;
> +	int i = 0;
> +	int cmd_size = 0;

Most of these can be left uninitialized.

> +	uint32_t num_regions;
> +	uint32_t buf_sz;
> +
> +	if (!a)
> +		return -ENODEV;
> +	num_regions = is_contiguous ? 1 : periods;
> +	buf_sz = is_contiguous ? (period_sz * periods) : period_sz;
> +	buf_sz = PAGE_ALIGN(buf_sz);
> +
> +	cmd_size = sizeof(*mmap_regions) + (sizeof(*mregions) * num_regions);
> +
> +	mmap_region_cmd = kzalloc(cmd_size, GFP_KERNEL);
> +	if (!mmap_region_cmd)
> +		return -ENOMEM;
> +
> +	mmap_regions = (struct avs_cmd_shared_mem_map_regions *)mmap_region_cmd;
> +	q6asm_add_mmaphdr(ac, &mmap_regions->hdr, cmd_size, true,
> +			  ((ac->session << 8) | dir));
> +	a->mem_state = -1;
> +
> +	mmap_regions->hdr.opcode = ASM_CMD_SHARED_MEM_MAP_REGIONS;
> +	mmap_regions->mem_pool_id = ADSP_MEMORY_MAP_SHMEM8_4K_POOL;
> +	mmap_regions->num_regions = num_regions;
> +	mmap_regions->property_flag = 0x00;
> +
> +	payload = ((u8 *) mmap_region_cmd +
> +		   sizeof(struct avs_cmd_shared_mem_map_regions));

mmap_region_cmd is void *, so no need to type cast.


> +
> +	mregions = (struct avs_shared_map_region_payload *)payload;

Payload is void *, so no need to type cast. But there's also no benefit
of having "payload", as this line can be written as:

	mregions = mmap_region_cmd + sizeof(*mmap_regions);


But adding a flexible array member to the avs_cmd_shared_mem_map_regions
struct would make things even clearer, in particular you would be able
to read the struct definition and see that there's a payload following
the command.

> +
> +	ac->port[dir].mem_map_handle = 0;

Isn't this already 0?

> +	port = &ac->port[dir];
> +
> +	for (i = 0; i < num_regions; i++) {
> +		ab = &port->buf[i];
> +		mregions->shm_addr_lsw = lower_32_bits(ab->phys);
> +		mregions->shm_addr_msw = upper_32_bits(ab->phys);
> +		mregions->mem_size_bytes = buf_sz;

Here you're dereferencing port->buf without holding cmd_lock.

> +		++mregions;
> +	}
> +
> +	rc = apr_send_pkt(a->adev, (uint32_t *) mmap_region_cmd);
> +	if (rc < 0)
> +		goto fail_cmd;
> +
> +	rc = wait_event_timeout(a->mem_wait, (a->mem_state >= 0),
> +				5 * HZ);
> +	if (!rc) {
> +		dev_err(ac->dev, "timeout. waited for memory_map\n");
> +		rc = -ETIMEDOUT;
> +		goto fail_cmd;
> +	}
> +
> +	if (a->mem_state > 0) {
> +		rc = adsp_err_get_lnx_err_code(a->mem_state);
> +		goto fail_cmd;
> +	}
> +	rc = 0;
> +fail_cmd:
> +	kfree(mmap_region_cmd);
> +	return rc;
> +}
> +
> +/**
> + * q6asm_map_memory_regions() - map memory regions in the dsp.
> + *
> + * @dir: direction of audio stream

This sounds boolean, perhaps worth mentioning here if true means rx or
tx.

And it's idiomatic to have such a parameter later in the list, it would
probably be more natural to read the call sight if the order was:

q6asm_map_memory_regions(ac, phys, periods, size, true);

> + * @ac: audio client instanace
> + * @phys: physcial address that needs mapping.
> + * @period_sz: audio period size
> + * @periods: number of periods
> + *
> + * Return: Will be an negative value on failure or zero on success
> + */
> +int q6asm_map_memory_regions(unsigned int dir, struct audio_client *ac,
> +			     dma_addr_t phys,
> +			     unsigned int period_sz, unsigned int periods)

period_sz could with benefit be of type size_t.

> +{
> +	struct audio_buffer *buf;
> +	int cnt;
> +	int rc;
> +
> +	if (ac->port[dir].buf) {
> +		dev_err(ac->dev, "Buffer already allocated\n");
> +		return 0;
> +	}
> +
> +	mutex_lock(&ac->cmd_lock);

I believe this lock should cover above check.

> +
> +	buf = kzalloc(((sizeof(struct audio_buffer)) * periods), GFP_KERNEL);

Loose a few of those parenthesis and use *buf in the sizeof.

> +	if (!buf) {
> +		mutex_unlock(&ac->cmd_lock);
> +		return -ENOMEM;
> +	}
> +
> +
> +	ac->port[dir].buf = buf;
> +
> +	buf[0].phys = phys;
> +	buf[0].used = dir ^ 1;

Why would this be called "used", and it's probably cleaner to just use
!!dir.

> +	buf[0].size = period_sz;
> +	cnt = 1;
> +	while (cnt < periods) {

cnt goes from 1 to periods and is incremented 1 each step, this would be
more succinct as a for loop.

> +		if (period_sz > 0) {
> +			buf[cnt].phys = buf[0].phys + (cnt * period_sz);
> +			buf[cnt].used = dir ^ 1;
> +			buf[cnt].size = period_sz;
> +		}
> +		cnt++;
> +	}
> +
> +	ac->port[dir].max_buf_cnt = periods;
> +	mutex_unlock(&ac->cmd_lock);

The only possible purpose of taking cmd_lock here is to protect
ac->port[dir].buf, but 

> +
> +	rc = __q6asm_memory_map_regions(ac, dir, period_sz, periods, 1);

The last parameter should be "true".

> +	if (rc < 0) {
> +		dev_err(ac->dev,
> +			"CMD Memory_map_regions failed %d for size %d\n", rc,
> +			period_sz);
> +
> +
> +		ac->port[dir].max_buf_cnt = 0;
> +		kfree(buf);
> +		ac->port[dir].buf = NULL;

These operations are done without holding cmd_lock.

> +
> +		return rc;
> +	}
> +
> +	return 0;
> +}
> +EXPORT_SYMBOL_GPL(q6asm_map_memory_regions);
> +
>  /**
>   * q6asm_audio_client_free() - Freee allocated audio client
>   *
> @@ -117,8 +410,10 @@ static struct audio_client *q6asm_get_audio_client(struct q6asm *a,
>  
>  static int q6asm_srvc_callback(struct apr_device *adev, struct apr_client_data *data)
>  {
> -	struct q6asm *q6asm = dev_get_drvdata(&adev->dev);
> +	struct q6asm *a, *q6asm = dev_get_drvdata(&adev->dev);
>  	struct audio_client *ac = NULL;
> +	struct audio_port_data *port;
> +	uint32_t dir = 0;
>  	uint32_t sid = 0;
>  	uint32_t *payload;
>  
> @@ -135,6 +430,52 @@ static int q6asm_srvc_callback(struct apr_device *adev, struct apr_client_data *
>  		return 0;
>  	}
>  
> +	a = dev_get_drvdata(ac->dev->parent);
> +	if (data->opcode == APR_BASIC_RSP_RESULT) {

This is a case in below switch statement.

> +		switch (payload[0]) {
> +		case ASM_CMD_SHARED_MEM_MAP_REGIONS:
> +		case ASM_CMD_SHARED_MEM_UNMAP_REGIONS:
> +			if (payload[1] != 0) {
> +				dev_err(ac->dev,
> +					"cmd = 0x%x returned error = 0x%x sid:%d\n",
> +					payload[0], payload[1], sid);
> +				a->mem_state = payload[1];
> +			} else {
> +				a->mem_state = 0;

Just assign a->mem_state = payload[1] outside the conditional, as it
will be the same value.

> +			}
> +
> +			wake_up(&a->mem_wait);
> +			break;
> +		default:
> +			dev_err(&adev->dev, "command[0x%x] not expecting rsp\n",
> +				 payload[0]);
> +			break;
> +		}
> +		return 0;
> +	}

Regards,
Bjorn