rtw88: Problem with sdio.c

Commit Message

Larry Finger May 24, 2023, 6:51 p.m. UTC

Martin,

When I applied your latest changes to my rtw88 repo, the driver for 8723ds 
broke. The dmesg log first showed things like:

[    3.603884] rtw_8723ds mmc1:0001:1: Firmware version 48.0.0, H2C version 0
[    3.615430] sunxi-mmc 4021000.mmc: unaligned scatterlist: os e80 length 2
[    3.622248] sunxi-mmc 4021000.mmc: map DMA failed
[    3.626974] rtw_8723ds mmc1:0001:1: sdio read16 failed (0x10040): -22
[    3.633435] sunxi-mmc 4021000.mmc: unaligned scatterlist: os e80 length 2
[    3.640236] sunxi-mmc 4021000.mmc: map DMA failed

There were similar messages for write16 operations.

I was able to "fix" the problem by turning off rea16/write16 operations for the 
RTW8723DS with the following patch that uses the rtw_chip_wcpu_11n() function:

         }
@@ -125,7 +125,7 @@ static u16 rtw_sdio_readw(struct rtw_dev *rtwdev, u32 addr, 
int *err_ret)
         u8 buf[2];
         int i;

-       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2))
+       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2) && !rtw_chip_wcpu_11n(rtwdev))
                 return sdio_readw(rtwsdio->sdio_func, addr, err_ret);

         for (i = 0; i < 2; i++) {

This leaves 16-bit reads and write enabled for the other chips. Alternatives 
would be to detect when this particular SDIO controller is in use, or last of 
all, add a module parameter.

Larry

Comments

Jernej Škrabec May 24, 2023, 7:18 p.m. UTC | #1

Dne sreda, 24. maj 2023 ob 20:51:53 CEST je Larry Finger napisal(a):
> Martin,
> 
> When I applied your latest changes to my rtw88 repo, the driver for 8723ds 
> broke. The dmesg log first showed things like:
> 
> [    3.603884] rtw_8723ds mmc1:0001:1: Firmware version 48.0.0, H2C version 0
> [    3.615430] sunxi-mmc 4021000.mmc: unaligned scatterlist: os e80 length 2
> [    3.622248] sunxi-mmc 4021000.mmc: map DMA failed
> [    3.626974] rtw_8723ds mmc1:0001:1: sdio read16 failed (0x10040): -22
> [    3.633435] sunxi-mmc 4021000.mmc: unaligned scatterlist: os e80 length 2
> [    3.640236] sunxi-mmc 4021000.mmc: map DMA failed
> 
> There were similar messages for write16 operations.

This was fixed in:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cb0ddaaa5db09d7d216fcbf0e68779be223a1128

Best regards,
Jernej

> 
> I was able to "fix" the problem by turning off rea16/write16 operations for the 
> RTW8723DS with the following patch that uses the rtw_chip_wcpu_11n() function:
> 
> diff --git a/sdio.c b/sdio.c
> index 1647cdc..2051c30 100644
> --- a/sdio.c
> +++ b/sdio.c
> @@ -87,7 +87,7 @@ static void rtw_sdio_writew(struct rtw_dev *rtwdev, u16 val, 
> u32 addr,
>          u8 buf[2];
>          int i;
> 
> -       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2)) {
> +       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2) && !rtw_chip_wcpu_11n(rtwdev)) {
>                  sdio_writew(rtwsdio->sdio_func, val, addr, err_ret);
>                  return;
>          }
> @@ -125,7 +125,7 @@ static u16 rtw_sdio_readw(struct rtw_dev *rtwdev, u32 addr, 
> int *err_ret)
>          u8 buf[2];
>          int i;
> 
> -       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2))
> +       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2) && !rtw_chip_wcpu_11n(rtwdev))
>                  return sdio_readw(rtwsdio->sdio_func, addr, err_ret);
> 
>          for (i = 0; i < 2; i++) {
> 
> This leaves 16-bit reads and write enabled for the other chips. Alternatives 
> would be to detect when this particular SDIO controller is in use, or last of 
> all, add a module parameter.
> 
> Larry
>

Larry Finger May 24, 2023, 8:31 p.m. UTC | #2

On 5/24/23 14:18, Jernej Škrabec wrote:
> 
> This was fixed in:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cb0ddaaa5db09d7d216fcbf0e68779be223a1128

That fix is in the kernel mainline source, but not in wireless-next, where I got 
my source. I hope it gets resolved correctly.

Larry

Kalle Valo May 27, 2023, 7:15 a.m. UTC | #3

Larry Finger <[e-mail]> writes:

> On 5/24/23 14:18, Jernej Škrabec wrote:
>>
>> This was fixed in:
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cb0ddaaa5db09d7d216fcbf0e68779be223a1128
>
> That fix is in the kernel mainline source, but not in wireless-next,
> where I got my source. I hope it gets resolved correctly.

The fix should come to wireless-next after the next (no pun intended)
net-next pull request. One way to avoid this problem of not having all
the fixes is to use the wireless-testing tree:

https://git.kernel.org/pub/scm/linux/kernel/git/wireless/wireless-testing.git/

In that tree Bob Copeland periodically pulls both wireless and
wireless-next to the latest -rc releases and that way you can easily get
all the latest wireless code from one tree. For example I use
wireless-testing as the baseline for all my ath1*k drivers testing. But
do note that the tree is rebased so the history is not stable.

Patch

diff --git a/sdio.c b/sdio.c
index 1647cdc..2051c30 100644
--- a/sdio.c
+++ b/sdio.c
@@ -87,7 +87,7 @@  static void rtw_sdio_writew(struct rtw_dev *rtwdev, u16 val, 
u32 addr,
         u8 buf[2];
         int i;

-       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2)) {
+       if (rtw_sdio_use_memcpy_io(rtwdev, addr, 2) && !rtw_chip_wcpu_11n(rtwdev)) {
                 sdio_writew(rtwsdio->sdio_func, val, addr, err_ret);
                 return;