Re: [PATCH V12 3/5] i2c: tegra: Add DMA support
From: Dmitry Osipenko
Date: Wed Feb 06 2019 - 09:13:29 EST
06.02.2019 17:01, Thierry Reding ÐÐÑÐÑ:
> On Wed, Feb 06, 2019 at 01:51:12PM +0000, Sowjanya Komatineni wrote:
>>
>>>>>>> That's odd because it suggests that DMA actually completed, but
>>>>>>> the message didn't.
>>>>>>>
>>>>>>> I'm not sure I understand how that could happen.
>>>>>>>
>>>>>>> What's also weird above is that there doesn't seem to be a DMA
>>>>>>> that is started for that particular message. Or is the timeout
>>>>>>> message a response to the prior transfer (length 10)? Seems like
>>>>>>> that should not be possible because we get the "transfer complete" message.
>>>>>>
>>>>>> Wait, those are actually different instances of the I2C
>>>>>> controller, so the relevant log entries are these:
>>>>>>
>>>>>> [ 0.945445] tegra-i2c 7000d000.i2c: starting DMA for length: 16
>>>>>> [ 0.945456] tegra-i2c 7000d000.i2c: unmasked irq: 0c
>>>>>> ...
>>>>>> [ 1.049224] tegra-i2c 7000d000.i2c: i2c transfer timed out
>>>>>>
>>>>>> And these don't happen if you use higher burst sizes or before the
>>>>>> DMA series?
>>>>>
>>>>> I tried to enforce DMA without changing bursts:
>>>>>
>>>>> ----------
>>>>> diff --git a/drivers/i2c/busses/i2c-tegra.c
>>>>> b/drivers/i2c/busses/i2c-tegra.c index c538ed5f8e2c..5d1c54ce7800
>>>>> 100644
>>>>> --- a/drivers/i2c/busses/i2c-tegra.c
>>>>> +++ b/drivers/i2c/busses/i2c-tegra.c
>>>>> @@ -6,6 +6,8 @@
>>>>> * Author: Colin Cross <ccross@xxxxxxxxxxx>
>>>>> */
>>>>>
>>>>> +#define DEBUG
>>>>> +
>>>>> #include <linux/clk.h>
>>>>> #include <linux/delay.h>
>>>>> #include <linux/dmaengine.h>
>>>>> @@ -1046,8 +1048,7 @@ static int tegra_i2c_xfer_msg(struct tegra_i2c_dev *i2c_dev,
>>>>> xfer_size = msg->len + I2C_PACKET_HEADER_SIZE;
>>>>>
>>>>> xfer_size = ALIGN(xfer_size, BYTES_PER_FIFO_WORD);
>>>>> - i2c_dev->is_curr_dma_xfer = (xfer_size > I2C_PIO_MODE_MAX_LEN) &&
>>>>> - i2c_dev->dma_buf;
>>>>> + i2c_dev->is_curr_dma_xfer = !!i2c_dev->dma_buf;
>>>>> tegra_i2c_config_fifo_trig(i2c_dev, xfer_size);
>>>>> dma = i2c_dev->is_curr_dma_xfer;
>>>>> ----------
>>>>>
>>>>> Here is the log with this change:
>>>>>
>>>>> [ 0.760796] tegra_rtc 7000e000.rtc: registered as rtc1
>>>>> [ 0.760850] tegra_rtc 7000e000.rtc: Tegra internal Real Time Clock
>>>>> [ 0.761050] i2c /dev entries driver
>>>>> [ 0.918928] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 0.918940] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.919040] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 0.919050] tegra-i2c 7000c000.i2c: starting DMA for length: 8
>>>>> [ 0.919059] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.919322] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 0.919335] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 0.919343] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.919440] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 0.919448] tegra-i2c 7000c000.i2c: starting DMA for length: 112
>>>>> [ 0.919456] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.922818] tegra-i2c 7000c000.i2c: transfer complete: 11 0 0
>>>>> [ 0.922829] atmel_mxt_ts 0-004c: Family: 160 Variant: 0 Firmware V1.0.AA Objects: 18
>>>>> [ 0.922886] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 0.922895] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.922993] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 0.923002] tegra-i2c 7000c000.i2c: starting DMA for length: 224
>>>>> [ 0.923011] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.933253] tegra-i2c 7000c000.i2c: transfer complete: 11 0 0
>>>>> [ 0.933287] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 0.933297] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.933478] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 0.933487] tegra-i2c 7000c000.i2c: starting DMA for length: 12
>>>>> [ 0.933496] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 0.945120] tegra-i2c 7000d000.i2c: starting DMA for length: 16
>>>>> [ 0.945130] tegra-i2c 7000d000.i2c: unmasked irq: 0c
>>>>> [ 1.038917] tegra-i2c 7000c000.i2c: DMA transfer timeout
>>>>> [ 1.038982] atmel_mxt_ts 0-004c: __mxt_read_reg: i2c transfer failed (-110)
>>>>> [ 1.039000] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 1.039006] atmel_mxt_ts 0-004c: Failed to read T44 and T5 (-110)
>>
>>
>>
>>>>> [ 1.039009] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 1.039148] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 1.039157] tegra-i2c 7000c000.i2c: starting DMA for length: 4
>>>>> [ 1.039166] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 1.039304] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 1.039340] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 1.039349] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 1.039535] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 1.039544] tegra-i2c 7000c000.i2c: starting DMA for length: 12
>>>>> [ 1.039552] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 1.040055] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 1.040083] tegra-i2c 7000c000.i2c: starting DMA for length: 16
>>>>> [ 1.040092] tegra-i2c 7000c000.i2c: unmasked irq: 0c
>>>>> [ 1.040301] tegra-i2c 7000c000.i2c: transfer complete: 10 0 0
>>>>> [ 1.048934] tegra-i2c 7000d000.i2c: i2c transfer timed out
>>>>
>>>> This log shows DMA transfer timeout for atmel read.
>>>> Do you see issue if you donât enforce dma all time and let it choose PIO Vs DMA?
>>>>
>>>
>>> No, there are no timeout errors in PIO mode. I could post full log with the PIO-only mode if you want.
>>>
>> Yes please post full log of PIO only and DMA only
>> Also, in above log, lots of DMA transfer went thru fine except this
>> one transfer where DMA timed out.
>> DMA completion doesnt happen on DMA submit during this particular
>> transaction timeframe for some reason.
>
> Dmitry, can you provide some background information on the setup that
> you're using? I vaguely remember something about i2c@7000d000 being
> special on older Tegra devices. Perhaps there's something additional
> that we need to take into account?
>
> Aha! Look at tegra_i2c_reg_addr() which special-cases I2C_TX_FIFO on
> the DVC variant of the controller, which is exactly the one that you
> are seeing the timeouts on and it's exactly the register that we use
> as destination for the DMA transfer.
>
> I think we need to do something like this:
>
> slv_config.dst_addr = i2c_dev->base_phys + tegra_i2c_reg_addr(i2c_dev, I2C_TX_FIFO);
>
> to make sure we pass the right FIFO register address to DMA.
I'm running it on T20 at the moment. I don't know about the DVC quirks, but sounds plausible. Please let me know if you need any other info.
Here the full logs:
PIO-only: http://dpaste.com/2MX3EAN.txt
DMA-only: http://dpaste.com/3VXZSSM.txt
Note the "Failed to read T44 and T5 (-110)" error that happens in DMA-only. And it happens on i2c@7000c000, DVC is i2c@7000d000.