Re: [PATCH 8/8] serial: tegra: Correct error handling on DMA setup

[Jon Hunter <jonathanh@nvidia.com>]

On 20/05/15 10:51, Jon Hunter wrote:
> 
> On 20/05/15 04:57, Alexandre Courbot wrote:
> 
> [snip]
> 
>> There may still be a leak (that is not related to your patch) in the
>> RX path though:
>>
>>     dma_buf = dma_alloc_coherent(...);
>>     ret = dmaengine_slave_config(...);
>>     if (ret < 0) {
>>         ...
>>         goto scrub;
>>     }
>>
>>     tup->rx_dma_buf_virt = dma_buf;
>>     tup->rx_dma_buf_phys = dma_phys;
>>
>> scrub:
>>     dma_release_channel(dma_chan);
>>     return ret;
>>
>> It seems that if dmaengine_slave_config() fails, then the result of
>> dma_alloc_coherent() remains purely local to the function and is never
>> freed. Or am I missing something again?
> 
> Yes, you are right here. Actually, this would be applicable to
> both RX and TX paths because in the TX path the buffer would
> not be unmapped. I will generated a separate patch for fix
> this. How does the following look?
> 
> diff --git a/drivers/tty/serial/serial-tegra.c b/drivers/tty/serial/serial-tegra.c
> index 3b63f103f0c9..cf0133ae762d 100644
> --- a/drivers/tty/serial/serial-tegra.c
> +++ b/drivers/tty/serial/serial-tegra.c
> @@ -999,6 +999,12 @@ static int tegra_uart_dma_channel_allocate(struct tegra_uart_port *tup,
>                         dma_release_channel(dma_chan);
>                         return -ENOMEM;
>                 }
> +               dma_sconfig.src_addr = tup->uport.mapbase;
> +               dma_sconfig.src_addr_width = DMA_SLAVE_BUSWIDTH_1_BYTE;
> +               dma_sconfig.src_maxburst = 4;
> +               tup->rx_dma_chan = dma_chan;
> +               tup->rx_dma_buf_virt = dma_buf;
> +               tup->rx_dma_buf_phys = dma_phys;
>         } else {
>                 dma_phys = dma_map_single(tup->uport.dev,
>                         tup->uport.state->xmit.buf, UART_XMIT_SIZE,
> @@ -1009,39 +1015,23 @@ static int tegra_uart_dma_channel_allocate(struct tegra_uart_port *tup,
>                         return -ENOMEM;
>                 }
>                 dma_buf = tup->uport.state->xmit.buf;
> -       }
> -
> -       if (dma_to_memory) {
> -               dma_sconfig.src_addr = tup->uport.mapbase;
> -               dma_sconfig.src_addr_width = DMA_SLAVE_BUSWIDTH_1_BYTE;
> -               dma_sconfig.src_maxburst = 4;
> -       } else {
>                 dma_sconfig.dst_addr = tup->uport.mapbase;
>                 dma_sconfig.dst_addr_width = DMA_SLAVE_BUSWIDTH_1_BYTE;
>                 dma_sconfig.dst_maxburst = 16;
> +               tup->tx_dma_chan = dma_chan;
> +               tup->tx_dma_buf_virt = dma_buf;
> +               tup->tx_dma_buf_phys = dma_phys;
>         }
>  
>         ret = dmaengine_slave_config(dma_chan, &dma_sconfig);
>         if (ret < 0) {
>                 dev_err(tup->uport.dev,
>                         "Dma slave config failed, err = %d\n", ret);
> -               goto scrub;
> +               tegra_uart_dma_channel_free(tup, dma_to_memory);
> +               return ret;
>         }
>  
> -       if (dma_to_memory) {
> -               tup->rx_dma_chan = dma_chan;
> -               tup->rx_dma_buf_virt = dma_buf;
> -               tup->rx_dma_buf_phys = dma_phys;
> -       } else {
> -               tup->tx_dma_chan = dma_chan;
> -               tup->tx_dma_buf_virt = dma_buf;
> -               tup->tx_dma_buf_phys = dma_phys;
> -       }
>         return 0;
> -
> -scrub:
> -       tegra_uart_dma_channel_free(tup, dma_to_memory);
> -       return ret;
>  }

Posted the patch here [1].

Jon

[1] https://lkml.org/lkml/2015/5/20/347