Re: [RFC PATCH 2/3] hw/sd/sdhci: Prohibit DMA accesses to devices

[Jason Wang]

On Thu, Dec 16, 2021 at 4:57 AM Philippe Mathieu-Daudé
<philmd@redhat.com> wrote:
>
> From: Philippe Mathieu-Daudé <f4bug@amsat.org>
>
> The issue reported by OSS-Fuzz produces the following backtrace:
>
>   ==447470==ERROR: AddressSanitizer: heap-buffer-overflow
>   READ of size 1 at 0x61500002a080 thread T0
>       #0 0x71766d47 in sdhci_read_dataport hw/sd/sdhci.c:474:18
>       #1 0x7175f139 in sdhci_read hw/sd/sdhci.c:1022:19
>       #2 0x721b937b in memory_region_read_accessor softmmu/memory.c:440:11
>       #3 0x72171e51 in access_with_adjusted_size softmmu/memory.c:554:18
>       #4 0x7216f47c in memory_region_dispatch_read1 softmmu/memory.c:1424:16
>       #5 0x7216ebb9 in memory_region_dispatch_read softmmu/memory.c:1452:9
>       #6 0x7212db5d in flatview_read_continue softmmu/physmem.c:2879:23
>       #7 0x7212f958 in flatview_read softmmu/physmem.c:2921:12
>       #8 0x7212f418 in address_space_read_full softmmu/physmem.c:2934:18
>       #9 0x721305a9 in address_space_rw softmmu/physmem.c:2962:16
>       #10 0x7175a392 in dma_memory_rw_relaxed include/sysemu/dma.h:89:12
>       #11 0x7175a0ea in dma_memory_rw include/sysemu/dma.h:132:12
>       #12 0x71759684 in dma_memory_read include/sysemu/dma.h:152:12
>       #13 0x7175518c in sdhci_do_adma hw/sd/sdhci.c:823:27
>       #14 0x7174bf69 in sdhci_data_transfer hw/sd/sdhci.c:935:13
>       #15 0x7176aaa7 in sdhci_send_command hw/sd/sdhci.c:376:9
>       #16 0x717629ee in sdhci_write hw/sd/sdhci.c:1212:9
>       #17 0x72172513 in memory_region_write_accessor softmmu/memory.c:492:5
>       #18 0x72171e51 in access_with_adjusted_size softmmu/memory.c:554:18
>       #19 0x72170766 in memory_region_dispatch_write softmmu/memory.c:1504:16
>       #20 0x721419ee in flatview_write_continue softmmu/physmem.c:2812:23
>       #21 0x721301eb in flatview_write softmmu/physmem.c:2854:12
>       #22 0x7212fca8 in address_space_write softmmu/physmem.c:2950:18
>       #23 0x721d9a53 in qtest_process_command softmmu/qtest.c:727:9
>
> A DMA descriptor is previously filled in RAM. An I/O access to the
> device (frames #22 to #16) start the DMA engine (frame #13). The
> engine fetch the descriptor and execute the request, which itself
> accesses the SDHCI I/O registers (frame #1 and #0), triggering a
> re-entrancy issue.
>
> Fix by prohibit transactions from the DMA to devices. The DMA engine
> is thus restricted to memories.
>
> Reported-by: OSS-Fuzz (Issue 36391)
> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/451
> Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
> ---
>  hw/sd/sdhci.c | 5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/hw/sd/sdhci.c b/hw/sd/sdhci.c
> index fe2f21f0c37..0e5e988927e 100644
> --- a/hw/sd/sdhci.c
> +++ b/hw/sd/sdhci.c
> @@ -741,6 +741,7 @@ static void sdhci_do_adma(SDHCIState *s)
>  {
>      unsigned int begin, length;
>      const uint16_t block_size = s->blksize & BLOCK_SIZE_MASK;
> +    const MemTxAttrs attrs = { .memory = true };
>      ADMADescr dscr = {};
>      MemTxResult res;
>      int i;
> @@ -794,7 +795,7 @@ static void sdhci_do_adma(SDHCIState *s)
>                      res = dma_memory_write(s->dma_as, dscr.addr,
>                                             &s->fifo_buffer[begin],
>                                             s->data_count - begin,
> -                                           MEMTXATTRS_UNSPECIFIED);
> +                                           attrs);
>                      if (res != MEMTX_OK) {
>                          break;
>                      }
> @@ -823,7 +824,7 @@ static void sdhci_do_adma(SDHCIState *s)
>                      res = dma_memory_read(s->dma_as, dscr.addr,
>                                            &s->fifo_buffer[begin],
>                                            s->data_count - begin,
> -                                          MEMTXATTRS_UNSPECIFIED);
> +                                          attrs);
>                      if (res != MEMTX_OK) {
>                          break;
>                      }

I wonder how we can fix this for other devices, as this seems to be a
known issue for many years. We've received many reports from the
networking side.

It looks like this patch simply forbids p2p which is probably not the
case for other devices.

I remember there's ideas like using bh from Paolo or detecting
reentrancy in the memory core, both of them seems more general than
this?

Thanks

> --
> 2.33.1
>