Re: [PATCH v3] dmaengine: qcom: gpi: Add GPI immediate DMA support for SPI protocol

[neil.armstrong@linaro.org]

Hi,

On 04/12/2024 14:29, Jyothi Kumar Seerapu wrote:
> 
> 
> On 12/4/2024 6:19 PM, neil.armstrong@linaro.org wrote:
>> Hi,
>>
>> On 04/12/2024 13:20, Jyothi Kumar Seerapu wrote:
>>> The DMA TRE(Transfer ring element) buffer contains the DMA
>>> buffer address. Accessing data from this address can cause
>>> significant delays in SPI transfers, which can be mitigated to
>>> some extent by utilizing immediate DMA support.
>>>
>>> QCOM GPI DMA hardware supports an immediate DMA feature for data
>>> up to 8 bytes, storing the data directly in the DMA TRE buffer
>>> instead of the DMA buffer address. This enhancement enables faster
>>> SPI data transfers.
>>>
>>> This optimization reduces the average transfer time from 25 us to
>>> 16 us for a single SPI transfer of 8 bytes length, with a clock
>>> frequency of 50 MHz.
>>>
>>> Signed-off-by: Jyothi Kumar Seerapu <quic_jseerapu@quicinc.com>
>>> ---
>>>
>>> v2-> v3:
>>>     - When to enable Immediate DMA support, control is moved to GPI driver
>>>       from SPI driver.
>>>     - Optimizations are done in GPI driver related to immediate dma changes.
>>>     - Removed the immediate dma supported changes in qcom-gpi-dma.h file
>>>       and handled in GPI driver.
>>>
>>>     Link to v2:
>>>     https://lore.kernel.org/all/20241128133351.24593-2-quic_jseerapu@quicinc.com/
>>>     https://lore.kernel.org/all/20241128133351.24593-3-quic_jseerapu@quicinc.com/
>>>
>>> v1 -> v2:
>>>     - Separated the patches to dmaengine and spi subsystems
>>>     - Removed the changes which are not required for this feature from
>>>       qcom-gpi-dma.h file.
>>>     - Removed the type conversions used in gpi_create_spi_tre.
>>>     Link to v1:
>>>     https://lore.kernel.org/lkml/20241121115201.2191-2-quic_jseerapu@quicinc.com/
>>>
>>>   drivers/dma/qcom/gpi.c | 32 +++++++++++++++++++++++++++-----
>>>   1 file changed, 27 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/drivers/dma/qcom/gpi.c b/drivers/dma/qcom/gpi.c
>>> index 52a7c8f2498f..35451d5a81f7 100644
>>> --- a/drivers/dma/qcom/gpi.c
>>> +++ b/drivers/dma/qcom/gpi.c
>>> @@ -27,6 +27,7 @@
>>>   #define TRE_FLAGS_IEOT        BIT(9)
>>>   #define TRE_FLAGS_BEI        BIT(10)
>>>   #define TRE_FLAGS_LINK        BIT(11)
>>> +#define TRE_FLAGS_IMMEDIATE_DMA    BIT(16)
>>>   #define TRE_FLAGS_TYPE        GENMASK(23, 16)
>>>   /* SPI CONFIG0 WD0 */
>>> @@ -64,6 +65,7 @@
>>>   /* DMA TRE */
>>>   #define TRE_DMA_LEN        GENMASK(23, 0)
>>> +#define TRE_DMA_IMMEDIATE_LEN    GENMASK(3, 0)
>>>   /* Register offsets from gpi-top */
>>>   #define GPII_n_CH_k_CNTXT_0_OFFS(n, k)    (0x20000 + (0x4000 * (n)) + (0x80 * (k)))
>>> @@ -1711,6 +1713,8 @@ static int gpi_create_spi_tre(struct gchan *chan, struct gpi_desc *desc,
>>>       dma_addr_t address;
>>>       struct gpi_tre *tre;
>>>       unsigned int i;
>>> +    int len;
>>> +    u8 immediate_dma;
>>
>> Should be bool
> Hi Neil, Thanks for the review.
> Sure, will change it to bool.
>>
>>>       /* first create config tre if applicable */
>>>       if (direction == DMA_MEM_TO_DEV && spi->set_config) {
>>> @@ -1763,14 +1767,32 @@ static int gpi_create_spi_tre(struct gchan *chan, struct gpi_desc *desc,
>>>       tre_idx++;
>>>       address = sg_dma_address(sgl);
>>> -    tre->dword[0] = lower_32_bits(address);
>>> -    tre->dword[1] = upper_32_bits(address);
>>> +    len = sg_dma_len(sgl);
>>> -    tre->dword[2] = u32_encode_bits(sg_dma_len(sgl), TRE_DMA_LEN);
>>> +    immediate_dma = (direction == DMA_MEM_TO_DEV) && len <= 2 * sizeof(tre->dword[0]);
>>
>> I would have added () around 2 * sizeof(tre->dword[0])
> 
> Below condition is fine or should i add brackets ?
> 
> immediate_dma = direction == DMA_MEM_TO_DEV &&
>                  len <= 2 * sizeof(tre->dword[0]);

This is what Dmitry requested, it's fine for me.

Neil

>>
>>> +
>>> +    /* Support Immediate dma for write transfers for data length up to 8 bytes */
>>> +    if (immediate_dma) {
>>> +        /*
>>> +         * For Immediate dma, data length may not always be length of 8 bytes,
>>> +         * it can be length less than 8, hence initialize both dword's with 0
>>> +         */
>>> +        tre->dword[0] = 0;
>>> +        tre->dword[1] = 0;
>>> +        memcpy(&tre->dword[0], sg_virt(sgl), len);
>>> +
>>> +        tre->dword[2] = u32_encode_bits(len, TRE_DMA_IMMEDIATE_LEN);
>>> +    } else {
>>> +        tre->dword[0] = lower_32_bits(address);
>>> +        tre->dword[1] = upper_32_bits(address);
>>> +
>>> +        tre->dword[2] = u32_encode_bits(len, TRE_DMA_LEN);
>>> +    }
>>>       tre->dword[3] = u32_encode_bits(TRE_TYPE_DMA, TRE_FLAGS_TYPE);
>>> -    if (direction == DMA_MEM_TO_DEV)
>>> -        tre->dword[3] |= u32_encode_bits(1, TRE_FLAGS_IEOT);
>>> +    tre->dword[3] |= u32_encode_bits(!!immediate_dma, TRE_FLAGS_IMMEDIATE_DMA);
>>
>> And you can drop !! if it's a bool
> Sure thanks, i will drop !! in V4.
>>
>>> +    tre->dword[3] |= u32_encode_bits(!!(direction == DMA_MEM_TO_DEV),
>>> +                     TRE_FLAGS_IEOT);
>>
>> I thingk you can drop !! here aswell, the check will return a bool
> Sure thanks, i will drop !! in V4.
>>
>>>       for (i = 0; i < tre_idx; i++)
>>>           dev_dbg(dev, "TRE:%d %x:%x:%x:%x\n", i, desc->tre[i].dword[0],
>>
>> Otherwise I like the simplification :-)
> Thanks, i will incorporate the changes in V4.
>>
>> Thanks,
>> Neil