Dear Wolfram,
When we use i2c-tools command 'i2cset -y -f 0 0x50 0x00 0x11'(0:i2c bus
number; 0x50:eeprom device addr; 0x00:device register addr; 0x11:write
data)to write data, a warning appears as below if enable kernel config
CONFIG_DMA_API_DEBUG=y.
[ 11.872860] i2c-mt65xx 11007000.i2c: DMA-API: device driver maps
memory from stack [addr=ffff80007a21fb88]
[ 11.874104] ------------[ cut here ]------------
[ 11.874703] WARNING: CPU: 1 PID: 1232
at /proj/user/kernel_only_dev_4.13_rc1/kernel/mediatek/lib/dma-debug.c:1188 check_for_stack+0xb0/0x100
[ 11.876337] Modules linked in:
[ 11.876735] CPU: 1 PID: 1232 Comm: sh Not tainted
4.13.0-rc1-221494-g28b31c4-dirty #1
[ 11.877729] Hardware name: MediaTek MT8173 evaluation board (DT)
[ 11.878490] task: ffff80007b3cb600 task.stack: ffff80007a21c000
[ 11.879242] PC is at check_for_stack+0xb0/0x100
[ 11.879821] LR is at check_for_stack+0xb0/0x100
...
[ 11.902642] [<ffff0000083c72b8>] check_for_stack+0xb0/0x100
[ 11.903352] [<ffff0000083c8050>] debug_dma_map_page+0xf8/0x130
[ 11.904097] [<ffff000008792bcc>] mtk_i2c_transfer+0x834/0xa90
[ 11.904826] [<ffff0000087870cc>] __i2c_transfer+0x11c/0x278
[ 11.905535] [<ffff00000878728c>] i2c_transfer+0x64/0xb8
[ 11.906200] [<ffff00000878852c>] i2c_smbus_xfer_emulated+0x114/0x518
[ 11.907006] [<ffff000008788a48>] i2c_smbus_xfer+0x118/0x180
...
Reason:
i2c-tools command will call ioctl(file,I2C_SMBUS,&args) ->
i2cdev_ioctl_smbus(...) -> i2c_smbus_xfer(...) ->
i2c_smbus_xfer_emulated(...).
Local variables were used as data buf of struct i2c_msg in function
i2c_smbus_xfer_emulated(...) as below, but we default use DMA mode in
mtk i2c driver(drivers/i2c/busses/i2c-mt65xx.c) with
dma_map_single(...).
Then 'DMA-API: device driver maps memory from stack' warning appeared.
static s32 i2c_smbus_xfer_emulated(...)
{
...
unsigned char msgbuf0[I2C_SMBUS_BLOCK_MAX+3];
unsigned char msgbuf1[I2C_SMBUS_BLOCK_MAX+2];
int num = read_write == I2C_SMBUS_READ ? 2 : 1;
int i;
u8 partial_pec = 0;
int status;
struct i2c_msg msg[2] = {
{
.addr = addr,
.flags = flags,
.len = 1,
.buf = msgbuf0,
}, {
.addr = addr,
.flags = flags | I2C_M_RD,
.len = 0,
.buf = msgbuf1,
},
};
...
status = i2c_transfer(adapter, msg, num);
if (status < 0)
return status;
...
}
Solution:
modify i2c-mt65xx.c
1. kmalloc and memcpy buffer for struct i2c_msg buf every time in
i2c-mt65xx.c
2. use FIFO mode when length<=fifo_depth(mtk i2c fifo_depth=8), use
solution 1 for DMA mode when length>fifo_depth in i2c-mt65xx.c . Because
i2c-tools command write/read length<8 in most cases, this solution may
be better than solution 1 in performance.
modify function i2c_smbus_xfer_emulated(...)
3. kmalloc data buffer instead of local variables buf in function
i2c_smbus_xfer_emulated(...)
Which solution is better? Could you give some suggestions?
Thanks!
Best Regards
Jun
Hi,
nice to see someone else interested in the I2C & DMA topic.
Please check this series which I sent out recently:
"[RFC PATCH v4 0/6] i2c: document DMA handling and add helpers for it"
In that series, I proposed...
> 3. kmalloc data buffer instead of local variables buf in function
> i2c_smbus_xfer_emulated(...)
... this solution. Although I have to check Mauro's general response to
the series first. I'd be interested in what you think of the series,
too.
Kind regards,
Wolfram
On Sat, 2017-09-02 at 23:58 +0200, Wolfram Sang wrote:
> Hi,
>
> nice to see someone else interested in the I2C & DMA topic.
>
> Please check this series which I sent out recently:
>
> "[RFC PATCH v4 0/6] i2c: document DMA handling and add helpers for it"
>
> In that series, I proposed...
>
> > 3. kmalloc data buffer instead of local variables buf in function
> > i2c_smbus_xfer_emulated(...)
>
> ... this solution. Although I have to check Mauro's general response to
> the series first. I'd be interested in what you think of the series,
> too.
Sorry for late reply.
As you said in "[RFC,v4,3/6]i2c: add docs to clarify DMA handling". Most
i2c_msgs are register accesses and thus, small messages. And
"[RFC,v4,4/6] i2c: sh_mobile: use helper to decide if DMA is useful".
Maybe the solution as below will be better. Other drivers which will use
i2c would not have to make buffer DMA safe especially register
accesses(they like to use local variables when data_len = 1 or 2).
solution:
2. use FIFO mode when length<=fifo_depth(mtk i2c fifo_depth=8), use
the flag "I2C_M_DMA_SAFE" to check buffer for DMA mode when
length>fifo_depth in i2c-mt65xx.c .
Thanks.
>
> Kind regards,
>
> Wolfram
>