On 11/29/2024 5:02 PM, Jyothi Kumar Seerapu wrote:
On 11/28/2024 8:53 PM, Bjorn Andersson wrote:
On Thu, Nov 28, 2024 at 07:03:50PM +0530, 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@xxxxxxxxxxx>
---
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.
drivers/dma/qcom/gpi.c | 32 +++++++++++++++++++++++++++-----
include/linux/dma/qcom-gpi-dma.h | 6 ++++++
2 files changed, 33 insertions(+), 5 deletions(-)
diff --git a/drivers/dma/qcom/gpi.c b/drivers/dma/qcom/gpi.c
index 52a7c8f2498f..4c5df696ddd8 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;
+ u8 *buf;
+ int len = 0;
First use of "len" is an assignment, so you shouldn't zero-initialize it
here.
Sure, will do it in V3.
/* 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);
+ /* Support Immediate dma for write transfers for data length up
to 8 bytes */
And what happens if the developer writing the SPI driver forgets to read
this comment and sets QCOM_GPI_IMMEDIATE_DMA for a 9 byte transfer?
In V2 patch, QCOM_GPI_IMMEDIATE_DMA is set based on
QCOM_GPI_IMMEDIATE_DMA_LEN only.
As per Hardware programming guide, immediate dma support is for up to 8
bytes only.
Need to check what is the behavior if we want to handle 9 bytes using
immediate dma feature support.
+ if ((spi->flags & QCOM_GPI_IMMEDIATE_DMA) && direction ==
DMA_MEM_TO_DEV) {
Why is this flag introduced?
If I understand the next patch, all DMA_MEM_TO_DEV transfers of <=
QCOM_GPI_IMMEDIATE_DMA_LEN can use the immediate mode, so why not move
the condition here?
Also ordering[1].
if (direction == DMA_MEM_TO_DEV && len <= 2 * sizeof(tre->dword[0]))
Sure, thanks for the suggestion.
so, instead using "QCOM_GPI_IMMEDIATE_DMA_LEN" need to use " 2 *
sizeof(tre->dword[0])" for 8 bytes length check.
[1] Compare "all transfers of length 8 or less, which are mem to device"
vs "all transfers which are mem to device, with a length of 8 or less".
The bigger "selection criteria" is the direction, then that's fine tuned
by the length query.
+ buf = sg_virt(sgl);
It's a question of style, but I think you could just put the sg_virt()
directly in the memcpy() call and avoid the extra variable.
Okay, i will directly put sg_virt() in memcpy().
- tre->dword[3] = u32_encode_bits(TRE_TYPE_DMA, TRE_FLAGS_TYPE);
- if (direction == DMA_MEM_TO_DEV)
+ /* memcpy may not always be length of 8, hence pre-fill both
dword's with 0 */
The implementation of memcpy() is always more than 8 bytes, it's buf
that might be less than 8 bytes ;)
Also you're not "pre-filling", you're "zero-initializing", or just
"initialize".
>
Sure, will correct it to use initialize instead pre-fill. i will update
it in V3.
>
That said, does it matter? Will the QUP read beyond the
TRE_DMA_IMMEDIATE_LEN bytes? If so, please put _that_ in the comment
("QUP reads beyond the provided len, so additional content needs to be
cleared", or similar)
>
Data lengths upto 8 bytes(len <=8) can be handled using immediate dma.
QUP won't handle beyond 8 bytes in immediate dma, if more than 8 bytes
then it process using the logic mentioned in else path.
Each dword is of size 4 bytes and so in this immediate dma case, dword0
and dword1 is used to hold 8 bytes data.In scenarios like data size
lesss then 8 bytes, for example if the spi data transfer size is 4bytes
then 4 bytes data will be updated into dword0 and dword1 might contain
some garbage data and so make sure to initialize both dword0 and dword1
with 0 and then memcpy the actual spi transfer data based on the length
to the dword's.
+ tre->dword[0] = 0;
+ tre->dword[1] = 0;
+ memcpy(&tre->dword[0], buf, len);
+
+ tre->dword[2] = u32_encode_bits(len, TRE_DMA_IMMEDIATE_LEN);
Does the format of tre->dword[2] really change when
TRE_FLAGS_IMMEDIATE_DMA is set, or is TRE_DMA_IMMEDIATE_LEN just a
mask to highlight that len can't be more than 4 bits?
It seems like you could drop TRE_DMA_IMMEDIATE_LEN and just use
TRE_DMA_LEN here? (But it should match what the hardware programming
guide states)
As per hardware programming guide, for Immediate dma, in dword2 length
should be 4 bytes only and so need to use TRE_DMA_IMMEDIATE_LEN .
Perhaps you could reduce the scope of this if/else then as well, as the
assignment of of dword[2] and dword[3] is mostly the same with and
without immediate mode (just the one bit to enable it)
sure, will check it and update in V3.
+
+ tre->dword[3] = u32_encode_bits(TRE_TYPE_DMA, TRE_FLAGS_TYPE);
tre->dword[3] |= u32_encode_bits(1, TRE_FLAGS_IEOT);
+ tre->dword[3] |= u32_encode_bits(1, TRE_FLAGS_IMMEDIATE_DMA);
+ } 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);
+ }
for (i = 0; i < tre_idx; i++)
dev_dbg(dev, "TRE:%d %x:%x:%x:%x\n", i, desc->tre[i].dword[0],
diff --git a/include/linux/dma/qcom-gpi-dma.h
b/include/linux/dma/qcom-gpi-dma.h
index 6680dd1a43c6..84598848d53a 100644
--- a/include/linux/dma/qcom-gpi-dma.h
+++ b/include/linux/dma/qcom-gpi-dma.h
@@ -15,6 +15,10 @@ enum spi_transfer_cmd {
SPI_DUPLEX,
};
+#define QCOM_GPI_IMMEDIATE_DMA BIT(1)
+
+#define QCOM_GPI_IMMEDIATE_DMA_LEN 8
+
/**
* struct gpi_spi_config - spi config for peripheral
*
@@ -30,6 +34,7 @@ enum spi_transfer_cmd {
* @cs: chip select toggle
* @set_config: set peripheral config
* @rx_len: receive length for buffer
+ * @flags: true for immediate dma support
Per above I think you can remove this flag, but "true for immediate DMA
support" doesn't match what you have written in the code. (Also in
general u8 shouldn't be "true")
Sure, will update in V3.
Thanks for providing the review comments.
Regards,
Bjorn
*/
struct gpi_spi_config {
u8 set_config;
@@ -44,6 +49,7 @@ struct gpi_spi_config {
u32 clk_src;
enum spi_transfer_cmd cmd;
u32 rx_len;
+ u8 flags;
};
enum i2c_op {
--
2.17.1
