Re: [PATCH V3 2/6] i2c: qup: Add V2 tags support

From: Sricharan R
Date: Wed Apr 15 2015 - 10:45:14 EST


Hi Ivan,

Sorry resending again, because wrapping seemed to be
somehow wrong in my previous response.

On 04/15/2015 02:19 PM, Ivan T. Ivanov wrote:

Hi Sricharan,

On Wed, 2015-04-15 at 12:09 +0530, Sricharan R wrote:

+/* frequency definitions for high speed and max speed */
+#define I2C_QUP_CLK_FAST_FREQ 1000000

This is fast mode, if I am not mistaken.

ya, up to 1MHZ is fast and up to 3.4MHZ is HS.
We use this macro to check if the desired freq is in HS range.
The above comment can be changed though to make it clear.

My point was that this is neither high nor max speed.

I see that QUP specs defines up to 1MHZ as fast+ speed.
+
/* Status, Error flags */
#define I2C_STATUS_WR_BUFFER_FULL BIT(0)
#define I2C_STATUS_BUS_ACTIVE BIT(8)
@@ -102,6 +119,15 @@
#define RESET_BIT 0x0
#define ONE_BYTE 0x1

+#define QUP_I2C_MX_CONFIG_DURING_RUN BIT(31)

Could you explain what is this for?

This is a new feature in the V2 version of the controller,
to support multiple i2c sub transfers without having
a 'stop' bit in-between. Without this the i2c controller
inserts a 'stop' on the bus when the WR/RD count registers
reaches zero and are configured for the next transfer. So setting
this bit when the controller is in 'RUN' state, avoids sending the
'stop' during RUN state.
I can add this comment in the patch.

And how v2 of this patch was worked if you introduce this bit now?
Bit is also not used by downstream driver, AFICS?

The one of the reason for this and the bam support patches in
this series was to support multiple transfers of i2c_msgs without
a 'stop' inbetween them. So without that the driver sends a 'stop'
between each sub-transfer. The v2 of this series supported that only in
bam mode and not in non-bam mode. That was a bug. I saw the difference
in behavior between bam and non-bam mode while testing with a hdmi adv
bridge recently. So added the support in this series. This was neede to
bringup the HDMI adv bridge device. I could have added it as a
new patch to make it more explicit.


+
+struct qup_i2c_block {
+ int blocks;

count

+ u8 *block_tag_len;

just tag_len,

+ int *block_data_len;

just data_len,

+ int block_pos;

just pos?

+};
+
struct qup_i2c_dev {
struct device*dev;
void __iomem*base;
@@ -115,9 +141,17 @@ struct qup_i2c_dev {
int in_fifo_sz;
int out_blk_sz;
int in_blk_sz;
-
+ struct qup_i2c_blockblk;
unsigned longone_byte_t;

+ int is_hs;
+ bool use_v2_tags;
+
+ int tx_tag_len;
+ int rx_tag_len;
+ u8 *tags;
+ int tags_pos;
+
struct i2c_msg*msg;
/* Current posion in user message buffer */
int pos;
@@ -263,10 +297,19 @@ static int qup_i2c_wait_ready(struct qup_i2c_dev *qup, int op, bool val,
}
}

Is this better tag related fields grouping?


struct qup_i2c_tag_block {

u8 tag[2];
// int tag_len; this is alway 2, or?
int data_len; // this could be zero, right?
};

There is a struct for qup_i2c_block which has tag and data len. Not sure
what change you suggest above ? Also with V2 transfers tag len need
not be 2. It can be more than that based on the data len.

The point is that: I will like to see better grouping of
related variables. Now they are spread all over. Would it
be possible to also take care for tx_tag_len, rx_tag_len,
tags, tags_pos.

For this and the above, i will change the groupings and split this patch
in to couple of smaller ones.

-static void qup_i2c_set_write_mode(struct qup_i2c_dev *qup, struct i2c_msg *msg)
+static void qup_i2c_set_write_mode(struct qup_i2c_dev *qup, struct i2c_msg *msg,
+ int run)

And 'run' stands for?
'run' just says whether the controller is in 'RUN' or 'RESET' state.
I can change it to is_run_st to make it clear.
{
- /* Number of entries to shift out, including the start */
- int total = msg->len + 1;
+ /* Total Number of entries to shift out, including the tags */
+ int total;
+
+ if (qup->use_v2_tags) {
+ total = msg->len + qup->tx_tag_len;
+ if (run)
+ total |= QUP_I2C_MX_CONFIG_DURING_RUN;

What?

This means, if the controller is in 'RUN' state, for
reconfiguring the RD/WR counts this bit has to be set to avoid
'stop' bits.

I don't buy it, sorry. Problem with v1 of the tags is that controller
can not read more than 256 bytes without automatically generate STOP
at the end, because transfer length specified with QUP_TAG_REC tag
is 8 bits wide. There is no limitation for writes with v1 tags,
because controller is explicitly instructed when to send out STOP.

correct till this.

For v2 of the tags, writes behaves more or less the same, and the
good news are that there is new read tag QUP_TAG_V2_DATARD, which
did't generate STOP when specified amount of data is read, still
up to 256 bytes in chunk. Read transfers with size more than 256
could be formed in following way:

V2_START | Slave Address | V2_DATARD | countx | ... | V2_DATARD_STOP | county.

The above is true for a single subtransfer for reading/writing
more than > 256 bytes. But for doing WRITE, READ kind of sub
transfers once the first sub transfer (write) is over, and
while re-configuring the _COUNT registers for the next transfers,
'a stop' between is inserted.
To avoid this,the MX_CONFIG_DURING_RUN bit should be set while in
'RUN' state, before re-configuring for the next sub transfer.
The qupspec calls this as bus 'lock-unlock' feature.

I tested this and it fixed the wrong data reads in the case of the hdmi
adv bridge reads. It was doing 2 i2c_msgs transfers for each i2c_xfer
call. The data read was wrong without this patch in v1 mode (default)
and even in v2 mode without the lock-unlock even though we do not send
a 'STOP' tag command between sub-transfers.



-static int qup_i2c_write_one(struct qup_i2c_dev *qup, struct i2c_msg *msg)
+static void qup_i2c_create_tag_v2(struct qup_i2c_dev *qup,
+ struct i2c_msg *msg, int last)
+{
+ u16 addr = (msg->addr << 1) | ((msg->flags & I2C_M_RD) == I2C_M_RD);
+ int len = 0, prev_len = 0;
+ int blocks = 0;
+ int rem;
+ int block_len = 0;
+ int data_len;
+
+ qup->blk.block_pos = 0;
+ qup->pos = 0;
+ qup->blk.blocks = (msg->len + QUP_READ_LIMIT - 1) / QUP_READ_LIMIT;
+ rem = msg->len % QUP_READ_LIMIT;
+
+ /* 2 tag bytes for each block + 2 extra bytes for first block */
+ qup->tags = kzalloc((qup->blk.blocks * 2) + 2, GFP_KERNEL);
+ qup->blk.block_tag_len = kzalloc(qup->blk.blocks, GFP_KERNEL);
+ qup->blk.block_data_len = kzalloc(sizeof(*qup->blk.block_data_len) *
+ qup->blk.blocks, GFP_KERNEL);

Whouldn't be easy to kcalloc struct qup_i2c_tag_block?


Allocations could fail and memory is leaking here.

ya correct, will change this. Freeing is done at the end of xfer
function, but will have to check for allocation failure.

Do you see how memory leak here?

if the first allocation pass and second one fails, and i will have
to move the kfree calls at the end of i2c_xfer inside the loop.

+
+ while (blocks < qup->blk.blocks) {

+ blocks++;

Looks like 'for' cycle to me.

sorry, not getting it ?

This 'while' loop is old plain 'for'.

ok
+ }
+
+ qup->tx_tag_len = len;
+
+ if (msg->flags & I2C_M_RD)
+ qup->rx_tag_len = (qup->blk.blocks * 2);

This will need some explanation.

ok, will add it.


+
+static void qup_i2c_issue_xfer_v2(struct qup_i2c_dev *qup, struct i2c_msg *msg)
+{
+ u32 data_len = 0, tag_len;
+
+ tag_len = qup->blk.block_tag_len[qup->blk.block_pos];
+
+ if (!(msg->flags & I2C_M_RD))
+ data_len = qup->blk.block_data_len[qup->blk.block_pos];
+
+ qup_i2c_send_data(qup, tag_len, qup->tags, data_len, msg->buf);

This assumes that writes are up to 256 bytes, and tags and data blocks
are completely written to FIFO buffer, right?

Yes, since we send/read data in blocks (256 bytes).

How deep is the FIFO? Is it guaranteed that "the whole" write
or read data, including tags will fit in.

Write/read fifo functions (also for V1) currently wait for the
fifo full and empty flags conditions. This will ensure that the
fifo full and empty conditions are taken care before writing
and reading. But this can be improved by waiting for
block_write/read flags to improve the throughput.


+
+static int qup_i2c_write_one(struct qup_i2c_dev *qup, struct i2c_msg *msg,
+ int run, int last)
{
unsigned long left;
int ret;
@@ -329,13 +501,20 @@ static int qup_i2c_write_one(struct qup_i2c_dev *qup, struct i2c_msg *msg)
qup->msg = msg;
qup->pos = 0;

+ if (qup->use_v2_tags)
+ qup_i2c_create_tag_v2(qup, msg, last);
+ else
+ qup->blk.blocks = 0;
+
enable_irq(qup->irq);

- qup_i2c_set_write_mode(qup, msg);
+ qup_i2c_set_write_mode(qup, msg, run);

- ret = qup_i2c_change_state(qup, QUP_RUN_STATE);
- if (ret)
- goto err;
+ if (!run) {
+ ret = qup_i2c_change_state(qup, QUP_RUN_STATE);

To run away, or not?

Means, if the controller is not in RUN state, put it in to 'RUN'
state.

And what is the problem if controller is put in PAUSED state, FIFO
filled with data and then RUN again, like in v2 of this patch?

This function is not entered with controller in PAUSED state
Only in Reset state (for the first transfer) and Run state for
the subsequent sub-transfers. The reason for having this 'run'
variable was that while using the lock-unlock feature, the
controller should not be put in to run-reset-run state
in-between transfers.

<snip>


-static void qup_i2c_read_fifo(struct qup_i2c_dev *qup, struct i2c_msg *msg)
+static void qup_i2c_read_fifo_v1(struct qup_i2c_dev *qup, struct i2c_msg *msg)
{
- u32 opflags;
u32 val = 0;
int idx;
+ int len = msg->len + qup->rx_tag_len;

Is this intentional?

Yes, since i changed the loop around qup_i2c_read_one to count for
blocks.

No, this doesn't make any sense. v1 tags didn't use rx_tag_len, and
I don't see how this is related to counting blocks, in this function.

sorry. Did confuse my response last time. This +rx_tag_len is
not needed here and infact it is '0' in case of v1.
So no effect.
<snip>


- qup_i2c_issue_read(qup, msg);
+ qup_i2c_issue_read(qup, msg);

- ret = qup_i2c_change_state(qup, QUP_RUN_STATE);
- if (ret)
- goto err;
+ ret = qup_i2c_change_state(qup, QUP_RUN_STATE);
+ if (ret)
+ goto err;

- do {
left = wait_for_completion_timeout(&qup->xfer, HZ);
if (!left) {
writel(1, qup->base + QUP_SW_RESET);
@@ -481,7 +726,8 @@ static int qup_i2c_read_one(struct qup_i2c_dev *qup, struct i2c_msg *msg)
}

qup_i2c_read_fifo(qup, msg);
- } while (qup->pos < msg->len);
+ qup->blk.block_pos++;

This should not be incremented, unless we really have read this block.

We are reading it and only then incrementing it.

But read FIFO function could exit because of timeout, which is
not checked. The same is true for write operations.

Somehow this patch looks overly complex, I don't believe
that it should be.

Ivan.


ok, the read and write fifo timing out error checks should be
corrected once i check the error code for qup_i2c_wait_ready
function.

I will split this patch in to couple of smaller patches and make
sure it does not look complex.

Regards,
Sricharan
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/