Re: [Qemu-devel] [PATCH v2] ide: ahci: add check before calling dma_memory_unmap

[John Snow <jsnow@redhat.com>]

On 01/28/2016 02:48 PM, P J P wrote:
> From: Prasad J Pandit <pjp@fedoraproject.org>
> 
> When IDE AHCI emulation uses Frame Information Structures(FIS)
> engine for data transfer, the mapped FIS buffer address is stored
> in a static 'bounce.buffer'. When a request is made to map another
> memory region, address_space_map() returns NULL because
> 'bounce.buffer' is in_use. It leads to a null pointer dereference
> error while doing 'dma_memory_unmap'. Add a check to avoid it.
> 
> Reported-by: Zuozhi fzz <zuozhi.fzz@alibaba-inc.com>
> Signed-off-by: Prasad J Pandit <pjp@fedoraproject.org>
> ---
>  hw/ide/ahci.c | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
> 
> Update as per review
>   -> https://lists.gnu.org/archive/html/qemu-devel/2016-01/msg05715.html
> 
> diff --git a/hw/ide/ahci.c b/hw/ide/ahci.c
> index 17f1cbd..f413a59 100644
> --- a/hw/ide/ahci.c
> +++ b/hw/ide/ahci.c
> @@ -661,9 +661,11 @@ static bool ahci_map_fis_address(AHCIDevice *ad)
>  
>  static void ahci_unmap_fis_address(AHCIDevice *ad)
>  {
> -    dma_memory_unmap(ad->hba->as, ad->res_fis, 256,
> -                     DMA_DIRECTION_FROM_DEVICE, 256);
> -    ad->res_fis = NULL;
> +    if (ad->res_fis) {
> +        dma_memory_unmap(ad->hba->as, ad->res_fis, 256,
> +                         DMA_DIRECTION_FROM_DEVICE, 256);
> +        ad->res_fis = NULL;
> +    }
>  }
>  
>  static bool ahci_map_clb_address(AHCIDevice *ad)
> @@ -677,9 +679,11 @@ static bool ahci_map_clb_address(AHCIDevice *ad)
>  
>  static void ahci_unmap_clb_address(AHCIDevice *ad)
>  {
> -    dma_memory_unmap(ad->hba->as, ad->lst, 1024,
> -                     DMA_DIRECTION_FROM_DEVICE, 1024);
> -    ad->lst = NULL;
> +    if (ad->lst) {
> +        dma_memory_unmap(ad->hba->as, ad->lst, 1024,
> +                         DMA_DIRECTION_FROM_DEVICE, 1024);
> +        ad->lst = NULL;
> +    }
>  }
>  
>  static void ahci_write_fis_sdb(AHCIState *s, NCQTransferState *ncq_tfs)
> 

This is fine for now as it protects us against doing something stupid in
a mechanical fashion, but I still wonder under which case we are
"starting" the FIS or CLB engines without getting a valid address. (The
unmap should only be happening when we /stop/ the engines, which implies
they were started -- which points to a bug somewhere else, too.)

Do you have a reproducer for the original issue?