[PATCH] spi: dw: use "smp_mb()" to avoid sending spi data error

[PATCH] spi: dw: use "smp_mb()" to avoid sending spi data error

[Xinwei Kong]

this patch will add memory barrier to ensure this "struct dw_spi *dws"
to complete data setting before enabling this SPI hardware interrupt.
eg:
it will fix to this following low possibility error in testing environment
which using SPI control to connect TPM Modules

kernel: tpm tpm0: Operation Timed out
kernel: tpm tpm0: tpm_relinquish_locality: : error -1

Signed-off-by: fengsheng <fengsheng5@huawei.com>
Signed-off-by: Xinwei Kong <kong.kongxinwei@hisilicon.com>
---
 drivers/spi/spi-dw.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/spi/spi-dw.c b/drivers/spi/spi-dw.c
index a92aa5c..d0d77a2 100644
--- a/drivers/spi/spi-dw.c
+++ b/drivers/spi/spi-dw.c
@@ -288,6 +288,8 @@ static int dw_spi_transfer_one(struct spi_controller *master,
 	dws->rx_end = dws->rx + transfer->len;
 	dws->len = transfer->len;
 
+	smp_mb();
+
 	spi_enable_chip(dws, 0);
 
 	/* Handle per transfer options for bpw and speed */
-- 
2.7.4

Re: [PATCH] spi: dw: use "smp_mb()" to avoid sending spi data error

[Mark Brown]

[-- Attachment #1: Type: text/plain, Size: 911 bytes --]

On Thu, Dec 26, 2019 at 05:21:28PM +0800, Xinwei Kong wrote:
> this patch will add memory barrier to ensure this "struct dw_spi *dws"
> to complete data setting before enabling this SPI hardware interrupt.
> eg:
> it will fix to this following low possibility error in testing environment
> which using SPI control to connect TPM Modules

> --- a/drivers/spi/spi-dw.c
> +++ b/drivers/spi/spi-dw.c
> @@ -288,6 +288,8 @@ static int dw_spi_transfer_one(struct spi_controller *master,
>  	dws->rx_end = dws->rx + transfer->len;
>  	dws->len = transfer->len;
>  
> +	smp_mb();
> +
>  	spi_enable_chip(dws, 0);

I'd be much more comfortable here if I understood what this was
supposed to be syncing - what exactly gets flushed here and why
is a memory barrier enough to ensure it's synced?  A comment in
the code would be especially good so anyone modifying the code
understands this in future.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH] spi: dw: use "smp_mb()" to avoid sending spi data error

[kongxinwei]

On 2019/12/27 8:22, Mark Brown wrote:
> On Thu, Dec 26, 2019 at 05:21:28PM +0800, Xinwei Kong wrote:
>> this patch will add memory barrier to ensure this "struct dw_spi *dws"
>> to complete data setting before enabling this SPI hardware interrupt.
>> eg:
>> it will fix to this following low possibility error in testing environment
>> which using SPI control to connect TPM Modules
> 
>> --- a/drivers/spi/spi-dw.c
>> +++ b/drivers/spi/spi-dw.c
>> @@ -288,6 +288,8 @@ static int dw_spi_transfer_one(struct spi_controller *master,
>>  	dws->rx_end = dws->rx + transfer->len;
>>  	dws->len = transfer->len;
>>  
>> +	smp_mb();
>> +
>>  	spi_enable_chip(dws, 0);
> 
> I'd be much more comfortable here if I understood what this was
> supposed to be syncing - what exactly gets flushed here and why
> is a memory barrier enough to ensure it's synced?  A comment in
> the code would be especially good so anyone modifying the code
> understands this in future.
> 
Because of out-of-order execution about some CPU architecture,
In this debug stage we find Completing spi interrupt enable ->
prodrucing TXEI interrupt -> running "interrupt_transfer" function
will prior to set "dw->rx and dws->rx_end" data, so it will result
in SPI sending error

Re: [PATCH] spi: dw: use "smp_mb()" to avoid sending spi data error

[Mark Brown]

[-- Attachment #1: Type: text/plain, Size: 810 bytes --]

On Sat, Dec 28, 2019 at 04:31:53PM +0800, kongxinwei wrote:

> > I'd be much more comfortable here if I understood what this was
> > supposed to be syncing - what exactly gets flushed here and why
> > is a memory barrier enough to ensure it's synced?  A comment in
> > the code would be especially good so anyone modifying the code
> > understands this in future.

> Because of out-of-order execution about some CPU architecture,
> In this debug stage we find Completing spi interrupt enable ->
> prodrucing TXEI interrupt -> running "interrupt_transfer" function
> will prior to set "dw->rx and dws->rx_end" data, so it will result
> in SPI sending error

Could you update the commit message to say that, and ideally also
add a comment saying something like "Ensure dw->rx and dw->rx_end
are visible" please?

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH] spi: dw: use "smp_mb()" to avoid sending spi data error

[kongxinwei]

On 2020/1/3 9:04, Mark Brown wrote:
> On Sat, Dec 28, 2019 at 04:31:53PM +0800, kongxinwei wrote:
> 
>>> I'd be much more comfortable here if I understood what this was
>>> supposed to be syncing - what exactly gets flushed here and why
>>> is a memory barrier enough to ensure it's synced?  A comment in
>>> the code would be especially good so anyone modifying the code
>>> understands this in future.
> 
>> Because of out-of-order execution about some CPU architecture,
>> In this debug stage we find Completing spi interrupt enable ->
>> prodrucing TXEI interrupt -> running "interrupt_transfer" function
>> will prior to set "dw->rx and dws->rx_end" data, so it will result
>> in SPI sending error
> 
> Could you update the commit message to say that, and ideally also
> add a comment saying something like "Ensure dw->rx and dw->rx_end
> are visible" please?
> 
OK, i WILL update it.