RE: [PATCH rtw-next v3] wifi: rtw88: usb: do not log transfers lost to a mode switch

Mehmet Fide <[email protected]> wrote:
> An RTL8822BU or RTL8822CU is asked to come back as a USB 3 device by
> rtw_usb_switch_mode_new(). The chip powers off its MAC and leaves the bus
> while the last control transfers of that sequence are still in flight, so
> they complete with -EPROTO and the driver reports them as errors:
> 
>   rtw_8822bu 1-1:1.0: Firmware version 27.2.0, H2C version 13
>   rtw_8822bu 1-1:1.0: write register 0xc4 failed with -71
>   usb 1-1: USB disconnect, device number 2
>   usbcore: registered new interface driver rtw_8822bu
>   rtw_8822bu 1-1:1.0: Firmware version 27.2.0, H2C version 13
> 
> Register 0xc4 is REG_PAD_CTRL2 and the access losing the race is the
> rtw_write32_set() that ends the switch sequence, a few milliseconds before
> the disconnect. Which transfer gets caught varies from boot to boot: 0xc4
> is in the "always on" section, so every write to it is followed by a second
> one from rtw_usb_reg_sec(), and sometimes that is the one that fails:
> 
>   rtw_8822bu 1-1:1.0: rtw_usb_reg_sec: reg 0x4e0, usb write 1 fail, status: -71
> 
> Nothing is wrong here. The device re-enumerates, probes again and registers
> normally, which is why rtw_usb_probe() already treats a non-zero return
> from rtw_usb_switch_mode() as "Not a fail". On a USB 2 only port the
> switch can never succeed, so the message returns on every boot and
> everyone using such a port has to work out that it is harmless.
> 
> Mark the window in which the chip is expected to leave the bus and skip the
> error reports for transfers that fall into it. The mark is set in
> rtw_usb_switch_mode(), so it covers both the new and the old switch
> sequence, and it is dropped again unless a switch was really started. It is
> never cleared after that because the switch always ends in a re-probe with
> a fresh struct rtw_usb.
> 
> Tested with an RTL8822BU (0x7392:0xb822) on a USB 2 root port of a TI AM62,
> where the message appears exactly once per boot. With the patch both lines
> are gone while the disconnect, the re-enumeration and the second firmware
> load are unchanged.
> 
> Signed-off-by: Mehmet Fide <[email protected]>

[...]

> @@ -1189,10 +1191,21 @@ static int rtw_usb_switch_mode(struct rtw_dev *rtwdev)
>                 return 0;
>         }
> 
> +       /* From here the chip may power off its MAC and re-enumerate, so it can
> +        * leave the bus while a control transfer is still in flight. Keep the
> +        * mark only if a switch was really started.
> +        */

First line of comment block should be empty.

But I think the comment in usb.h along declaration is enough. 


> +       rtwusb->switching_mode = true;
> +
>         if (rtw_usb3_chip_old(id))
> -               return rtw_usb_switch_mode_old(rtwdev);
> +               ret = rtw_usb_switch_mode_old(rtwdev);
>         else
> -               return rtw_usb_switch_mode_new(rtwdev);
> +               ret = rtw_usb_switch_mode_new(rtwdev);
> +
> +       if (ret != 1)
> +               rtwusb->switching_mode = false;

Shouldn't it set to false unconditionally?
Regarding normal read/write (not for switching mode) gets wrong, I think
it needs to throw an error.

I also think you might add/use a flag to 'enum rtw_flags'.

> +
> +       return ret;
>  }
> 
>  #define USB_REG_PAGE   0xf4
> diff --git a/drivers/net/wireless/realtek/rtw88/usb.h b/drivers/net/wireless/realtek/rtw88/usb.h
> index 9b695b688..9d2825368 100644
> --- a/drivers/net/wireless/realtek/rtw88/usb.h
> +++ b/drivers/net/wireless/realtek/rtw88/usb.h
> @@ -85,6 +85,9 @@ struct rtw_usb {
>         struct sk_buff_head rx_free_queue;
>         struct work_struct rx_work;
>         struct work_struct rx_urb_work;
> +
> +       /* the chip is re-enumerating, control transfers are expected to fail */
> +       bool switching_mode;
>  };
> 
>  static inline struct rtw_usb_tx_data *rtw_usb_get_tx_data(struct sk_buff *skb)
> --
> 2.54.0