2021-04-01 09:03:21

by Srinivas Kandagatla

[permalink] [raw]
Subject: [PATCH v2] soundwire: qcom: wait for fifo space to be available before read/write

If we write registers very fast we can endup in a situation where some
of the writes will be dropped without any notice.

So wait for the fifo space to be available before reading/writing the
soundwire registers.

Signed-off-by: Srinivas Kandagatla <[email protected]>
---

Changes since v1:
merged some of the loop code to make it simple as suggested by Pierre
updated error code and comments as suggested by Vinod


drivers/soundwire/qcom.c | 66 ++++++++++++++++++++++++++++++++++++++++
1 file changed, 66 insertions(+)

diff --git a/drivers/soundwire/qcom.c b/drivers/soundwire/qcom.c
index 6affa3cd4039..5fd4a99cc8ac 100644
--- a/drivers/soundwire/qcom.c
+++ b/drivers/soundwire/qcom.c
@@ -24,6 +24,8 @@
#define SWRM_COMP_CFG_IRQ_LEVEL_OR_PULSE_MSK BIT(1)
#define SWRM_COMP_CFG_ENABLE_MSK BIT(0)
#define SWRM_COMP_PARAMS 0x100
+#define SWRM_COMP_PARAMS_WR_FIFO_DEPTH GENMASK(14, 10)
+#define SWRM_COMP_PARAMS_RD_FIFO_DEPTH GENMASK(19, 15)
#define SWRM_COMP_PARAMS_DOUT_PORTS_MASK GENMASK(4, 0)
#define SWRM_COMP_PARAMS_DIN_PORTS_MASK GENMASK(9, 5)
#define SWRM_INTERRUPT_STATUS 0x200
@@ -51,6 +53,8 @@
#define SWRM_CMD_FIFO_CMD 0x308
#define SWRM_CMD_FIFO_FLUSH 0x1
#define SWRM_CMD_FIFO_STATUS 0x30C
+#define SWRM_RD_CMD_FIFO_CNT_MASK GENMASK(20, 16)
+#define SWRM_WR_CMD_FIFO_CNT_MASK GENMASK(12, 8)
#define SWRM_CMD_FIFO_CFG_ADDR 0x314
#define SWRM_CONTINUE_EXEC_ON_CMD_IGNORE BIT(31)
#define SWRM_RD_WR_CMD_RETRIES 0x7
@@ -104,6 +108,7 @@
#define SWR_BROADCAST_CMD_ID 0x0F
#define SWR_MAX_CMD_ID 14
#define MAX_FIFO_RD_RETRY 3
+#define SWR_OVERFLOW_RETRY_COUNT 30

struct qcom_swrm_port_config {
u8 si;
@@ -147,6 +152,8 @@ struct qcom_swrm_ctrl {
int (*reg_read)(struct qcom_swrm_ctrl *ctrl, int reg, u32 *val);
int (*reg_write)(struct qcom_swrm_ctrl *ctrl, int reg, int val);
u32 slave_status;
+ u32 wr_fifo_depth;
+ u32 rd_fifo_depth;
};

struct qcom_swrm_data {
@@ -238,6 +245,55 @@ static u32 swrm_get_packed_reg_val(u8 *cmd_id, u8 cmd_data,
return val;
}

+static int swrm_wait_for_rd_fifo_avail(struct qcom_swrm_ctrl *swrm)
+{
+ u32 fifo_outstanding_data, value;
+ int fifo_retry_count = SWR_OVERFLOW_RETRY_COUNT;
+
+ do {
+ /* Check for fifo underflow during read */
+ swrm->reg_read(swrm, SWRM_CMD_FIFO_STATUS, &value);
+ fifo_outstanding_data = FIELD_GET(SWRM_RD_CMD_FIFO_CNT_MASK, value);
+
+ /* Check if read data is available in read fifo */
+ if (fifo_outstanding_data > 0)
+ return 0;
+
+ usleep_range(500, 510);
+ } while (fifo_retry_count--);
+
+ if (fifo_outstanding_data == 0) {
+ dev_err_ratelimited(swrm->dev, "%s err read underflow\n", __func__);
+ return -EIO;
+ }
+
+ return 0;
+}
+
+static int swrm_wait_for_wr_fifo_avail(struct qcom_swrm_ctrl *swrm)
+{
+ u32 fifo_outstanding_cmds, value;
+ int fifo_retry_count = SWR_OVERFLOW_RETRY_COUNT;
+
+ do {
+ /* Check for fifo overflow during write */
+ swrm->reg_read(swrm, SWRM_CMD_FIFO_STATUS, &value);
+ fifo_outstanding_cmds = FIELD_GET(SWRM_WR_CMD_FIFO_CNT_MASK, value);
+
+ /* Check for space in write fifo before writing */
+ if (fifo_outstanding_cmds < swrm->wr_fifo_depth)
+ return 0;
+
+ usleep_range(500, 510);
+ } while (fifo_retry_count--);
+
+ if (fifo_outstanding_cmds == swrm->wr_fifo_depth) {
+ dev_err_ratelimited(swrm->dev, "%s err write overflow\n", __func__);
+ return -EIO;
+ }
+
+ return 0;
+}

static int qcom_swrm_cmd_fifo_wr_cmd(struct qcom_swrm_ctrl *swrm, u8 cmd_data,
u8 dev_addr, u16 reg_addr)
@@ -256,6 +312,9 @@ static int qcom_swrm_cmd_fifo_wr_cmd(struct qcom_swrm_ctrl *swrm, u8 cmd_data,
dev_addr, reg_addr);
}

+ if (swrm_wait_for_wr_fifo_avail(swrm))
+ return SDW_CMD_FAIL_OTHER;
+
/* Its assumed that write is okay as we do not get any status back */
swrm->reg_write(swrm, SWRM_CMD_FIFO_WR_CMD, val);

@@ -295,6 +354,9 @@ static int qcom_swrm_cmd_fifo_rd_cmd(struct qcom_swrm_ctrl *swrm,
/* wait for FIFO RD CMD complete to avoid overflow */
usleep_range(250, 255);

+ if (swrm_wait_for_rd_fifo_avail(swrm))
+ return SDW_CMD_FAIL_OTHER;
+
do {
swrm->reg_read(swrm, SWRM_CMD_FIFO_RD_FIFO_ADDR, &cmd_data);
rval[0] = cmd_data & 0xFF;
@@ -586,6 +648,10 @@ static int qcom_swrm_init(struct qcom_swrm_ctrl *ctrl)
SWRM_INTERRUPT_STATUS_RMSK);
}
ctrl->slave_status = 0;
+ ctrl->reg_read(ctrl, SWRM_COMP_PARAMS, &val);
+ ctrl->rd_fifo_depth = FIELD_GET(SWRM_COMP_PARAMS_RD_FIFO_DEPTH, val);
+ ctrl->wr_fifo_depth = FIELD_GET(SWRM_COMP_PARAMS_WR_FIFO_DEPTH, val);
+
return 0;
}

--
2.21.0


2021-04-01 18:08:13

by Srinivas Kandagatla

[permalink] [raw]
Subject: Re: [PATCH v2] soundwire: qcom: wait for fifo space to be available before read/write



On 01/04/2021 15:36, Pierre-Louis Bossart wrote:
>
>
> On 4/1/21 4:00 AM, Srinivas Kandagatla wrote:
>> If we write registers very fast we can endup in a situation where some
>> of the writes will be dropped without any notice.
>>
>> So wait for the fifo space to be available before reading/writing the
>> soundwire registers.
>
> Out of curiosity, do you actually need to do a check in the read case as
> well?

Yes, This is just to make sure the read command is finished and fifo is
ready with data.

If not we will be reading quickly an empty fifo!

>
> The commit message talks about writes getting dropped, is the opposite
> also a problem?

Its highly likely, for safety I have added support for both write and
read waits in this patch.


--srini
>

2021-04-01 18:53:08

by Pierre-Louis Bossart

[permalink] [raw]
Subject: Re: [PATCH v2] soundwire: qcom: wait for fifo space to be available before read/write



On 4/1/21 4:00 AM, Srinivas Kandagatla wrote:
> If we write registers very fast we can endup in a situation where some
> of the writes will be dropped without any notice.
>
> So wait for the fifo space to be available before reading/writing the
> soundwire registers.

Out of curiosity, do you actually need to do a check in the read case as
well?

The commit message talks about writes getting dropped, is the opposite
also a problem?

2021-04-06 13:19:37

by Vinod Koul

[permalink] [raw]
Subject: Re: [PATCH v2] soundwire: qcom: wait for fifo space to be available before read/write

On 01-04-21, 10:00, Srinivas Kandagatla wrote:
> If we write registers very fast we can endup in a situation where some
> of the writes will be dropped without any notice.
>
> So wait for the fifo space to be available before reading/writing the
> soundwire registers.

Applied, thanks

--
~Vinod