On Sat, Jul 8, 2017 at 12:08 AM, Thor Thayer
<thor.thayer@xxxxxxxxxxxxxxx> wrote:
> On 07/07/2017 11:25 AM, Andy Shevchenko wrote:
>> On Mon, Jun 19, 2017 at 11:36 PM, <thor.thayer@xxxxxxxxxxxxxxx> wrote:
>>> + while (bytes_to_transfer-- > 0) {
>>> + *idev->buf++ = readl(idev->base + ALTR_I2C_RX_DATA);
>>> + if (idev->msg_len == 1)
>>> + altr_i2c_stop(idev);
>>> + else
>>> + writel(0, idev->base + ALTR_I2C_TFR_CMD);
>>> +
>>> + idev->msg_len--;
>>> + }
>>
>>
>> Move out invariant from the loop (and I see a bug, you might go over
>> boundaries).
>>
>> while (bytes_to_transfer-- > 0) {
>> *idev->buf++ = readl(idev->base + ALTR_I2C_RX_DATA);
>> if (idev->msg_len-- == 1)
>> break;
>> writel(0, idev->base + ALTR_I2C_TFR_CMD);
>> }
>>
>> altr_i2c_stop(idev);
>>
> I see your point on the boundary.
Actually I didn't notice min() call above. So, you may ignore that part.
> However your change is slightly different
> from what I'm trying to do.
I don't see how it differs to what you wrote.
> I think you assumed the alt_i2c_stop() call can cause a stop condition. This
> soft IP can't send just a start or a stop condition by itself - both of
> these conditions need to be paired with a byte.
OK.
> The other subtle side effect is the start condition + byte write is the
> first write which is why the last write is skipped.
> I need to send a byte with a stop condition on the last expected byte
> (idev->msg_len == 1) while this change would send it after the FIFO is empty
> or after (msg_len == 1).
Consider the corner case when msg_len _is_ 1 at the beginning.
Then continue with 2. I really didn't see the difference in two
snippets. Perhaps you have a bug there.
> Your version is cleaner so I'll just add the alt_i2c_stop(idev) call inside
> the (msg_len == 1) condition and before the break.
>>> +static int altr_i2c_fill_tx_fifo(struct altr_i2c_dev *idev)
>>> +{
>>> + size_t tx_fifo_avail = idev->fifo_size - readl(idev->base +
>>> +
>>> ALTR_I2C_TC_FIFO_LVL);
>>> + int bytes_to_transfer = min(tx_fifo_avail, idev->msg_len);
>>> + int ret = idev->msg_len - bytes_to_transfer;
>>> +
>>> + while (bytes_to_transfer-- > 0) {
>>> + if (idev->msg_len == 1)
>>> + writel(ALTR_I2C_TFR_CMD_STO | *idev->buf++,
>>> + idev->base + ALTR_I2C_TFR_CMD);
>>> + else
>>> + writel(*idev->buf++, idev->base +
>>> ALTR_I2C_TFR_CMD);
>>> + idev->msg_len--;
>>> + }
>>
>>
>> Ditto.
>>
> See above but I will move the msg_len-- inside the condition check like you
> had.
Ditto.
>>> +static int altr_i2c_wait_for_core_idle(struct altr_i2c_dev *idev)
>>> +{
>>> + unsigned long timeout = jiffies +
>>> msecs_to_jiffies(ALTR_I2C_TIMEOUT);
>>> +
>>> + do {
>>> + if (time_after(jiffies, timeout)) {
>>> + dev_err(idev->dev, "Core Idle timeout\n");
>>> + return -ETIMEDOUT;
>>> + }
>>> + } while (readl(idev->base + ALTR_I2C_STATUS) &
>>> ALTR_I2C_STAT_CORE);
>>> +
>>> + return 0;
>>> +}
>> readl_poll_timeout[_atomic]() please.
You ignored some of my comments including this one. Why? Can you go
again through my rreview and answer the rest?
>>> + if (of_property_read_u32(np, "fifo-size", &idev->fifo_size))
>>> + idev->fifo_size = ALTR_I2C_DFLT_FIFO_SZ;
>> Shouldn't be possible to auto detect?
>>
> I agree. That would have been SO much better but the hardware designers
> released this without capturing the size in a register - they come from a
> bare-metal project perspective. Since the FIFO size is configurable in the
> FPGA from 4 to 256 levels deep, I need to capture this with the device tree.
You may do it manually, right? There are examples for similar cases
like writing the offset into FIFO until it returns the written value
and FIFO maximum possible size is not achieved.
--
With Best Regards,
Andy Shevchenko
