Re: [PATCH v3] ath10k: Fix crash during rmmod when probe firmware fails

[Mohammed Shafi Shajakhan <mohammed@codeaurora.org>]

Hi Kalle,

the change suggested by you helps, and the device probe, scan
is successful as well. Still good to have this change part of your
basic sanity and regression testing !

regards,
shafi

On Wed, Jan 25, 2017 at 01:46:28PM +0000, Valo, Kalle wrote:
> Kalle Valo <kvalo@qca.qualcomm.com> writes:
> 
> > Mohammed Shafi Shajakhan <mohammed@qti.qualcomm.com> writes:
> >
> >> From: Mohammed Shafi Shajakhan <mohammed@qti.qualcomm.com>
> >>
> >> This fixes the below crash when ath10k probe firmware fails,
> >> NAPI polling tries to access a rx ring resource which was never
> >> allocated, fix this by disabling NAPI right away once the probe
> >> firmware fails by calling 'ath10k_hif_stop'. Its good to note
> >> that the error is never propogated to 'ath10k_pci_probe' when
> >> ath10k_core_register fails, so calling 'ath10k_hif_stop' to cleanup
> >> PCI related things seems to be ok
> >>
> >> BUG: unable to handle kernel NULL pointer dereference at (null)
> >> IP:  __ath10k_htt_rx_ring_fill_n+0x19/0x230 [ath10k_core]
> >> __ath10k_htt_rx_ring_fill_n+0x19/0x230 [ath10k_core]
> >>
> >> Call Trace:
> >>
> >> [<ffffffffa113ec62>] ath10k_htt_rx_msdu_buff_replenish+0x42/0x90
> >> [ath10k_core]
> >> [<ffffffffa113f393>] ath10k_htt_txrx_compl_task+0x433/0x17d0
> >> [ath10k_core]
> >> [<ffffffff8114406d>] ? __wake_up_common+0x4d/0x80
> >> [<ffffffff811349ec>] ? cpu_load_update+0xdc/0x150
> >> [<ffffffffa119301d>] ? ath10k_pci_read32+0xd/0x10 [ath10k_pci]
> >> [<ffffffffa1195b17>] ath10k_pci_napi_poll+0x47/0x110 [ath10k_pci]
> >> [<ffffffff817863af>] net_rx_action+0x20f/0x370
> >>
> >> Reported-by: Ben Greear <greearb@candelatech.com>
> >> Fixes: 3c97f5de1f28 ("ath10k: implement NAPI support")
> >> Signed-off-by: Mohammed Shafi Shajakhan <mohammed@qti.qualcomm.com>
> >
> > Is there an easy way to reproduce this bug? I don't see it on my x86
> > laptop with qca988x and I call rmmod all the time. I would like to test
> > this myself.
> >
> >> --- a/drivers/net/wireless/ath/ath10k/core.c
> >> +++ b/drivers/net/wireless/ath/ath10k/core.c
> >> @@ -2164,6 +2164,7 @@ static int ath10k_core_probe_fw(struct ath10k *ar)
> >>  	ath10k_core_free_firmware_files(ar);
> >>  
> >>  err_power_down:
> >> +	ath10k_hif_stop(ar);
> >>  	ath10k_hif_power_down(ar);
> >>  
> >>  	return ret;
> >
> > This breaks the symmetry, we should not be calling ath10k_hif_stop() if
> > we haven't called ath10k_hif_start() from the same function. This can
> > just create a bigger mess later, for example with other bus support like
> > sdio or usb. In theory it should enough that we call
> > ath10k_hif_power_down() and pci.c does the rest correctly "behind the
> > scenes".
> >
> > I investigated this a bit and I think the real cause is that we call
> > napi_enable() from ath10k_pci_hif_power_up() and napi_disable() from
> > ath10k_pci_hif_stop(). Does anyone remember why?
> >
> > I was expecting that we would call napi_enable()/napi_disable() either
> > in ath10k_hif_power_up/down() or ath10k_hif_start()/stop(), but not
> > mixed like it's currently.
> 
> So below is something I was thinking of, now napi_enable() is called
> from ath10k_hif_start() and napi_disable() from ath10k_hif_stop(). Would
> that work?
> 
> --- a/drivers/net/wireless/ath/ath10k/pci.c
> +++ b/drivers/net/wireless/ath/ath10k/pci.c
> @@ -1648,6 +1648,8 @@ static int ath10k_pci_hif_start(struct ath10k *ar)
>  
>  	ath10k_dbg(ar, ATH10K_DBG_BOOT, "boot hif start\n");
>  
> +	napi_enable(&ar->napi);
> +
>  	ath10k_pci_irq_enable(ar);
>  	ath10k_pci_rx_post(ar);
>  
> @@ -2532,7 +2534,6 @@ static int ath10k_pci_hif_power_up(struct ath10k *ar)
>  		ath10k_err(ar, "could not wake up target CPU: %d\n", ret);
>  		goto err_ce;
>  	}
> -	napi_enable(&ar->napi);
>  
>  	return 0;
> 
> -- 
> Kalle Valo