Re: Assertion failure through vring_split_desc_read

[Laurent Vivier]

On 11/05/2020 05:51, Alexander Bulekov wrote:
> Hello,
> While fuzzing, I found an input that triggers an assertion failure
> through virtio-rng -> vring_split_desc_read. Maybe this is related to:
> Message-ID: <address@hidden> 
> Assertion failure through virtio_lduw_phys_cached
> 
> #8 0x7fe6a9acf091 in __assert_fail 
> /build/glibc-GwnBeO/glibc-2.30/assert/assert.c:101:3
> #9 0x564cbe7d96fd in address_space_read_cached include/exec/memory.h:2423:5
> #10 0x564cbe7e79c5 in vring_split_desc_read hw/virtio/virtio.c:236:5
> #11 0x564cbe7e84ce in virtqueue_split_read_next_desc hw/virtio/virtio.c:929:5
> #12 0x564cbe78f86b in virtqueue_split_get_avail_bytes 
> hw/virtio/virtio.c:1009:18
> #13 0x564cbe78ab22 in virtqueue_get_avail_bytes hw/virtio/virtio.c:1208:9
> #14 0x564cc08aade1 in get_request_size hw/virtio/virtio-rng.c:40:5
> #15 0x564cc08aa20b in virtio_rng_process hw/virtio/virtio-rng.c:115:12
> #16 0x564cc08a8c48 in virtio_rng_set_status hw/virtio/virtio-rng.c:172:5
> #17 0x564cbe7a50be in virtio_set_status hw/virtio/virtio.c:1876:9
> #18 0x564cc08d1b8f in virtio_pci_common_write hw/virtio/virtio-pci.c:1245:9
> 
> I can reproduce it in a qemu 5.0 build using these qtest commands:
> https://paste.debian.net/plain/1146089
> (not including them here, as some are quite long)
> 
> wget https://paste.debian.net/plain/1146089 -O qtest-trace; 
> ~/Development/qemu/build/i386-softmmu/qemu-system-i386 -M pc-q35-5.0  -device 
> virtio-rng-pci,addr=04.0 -display none -nodefaults -nographic -qtest stdio < 
> qtest-trace

Nice work.

If I use directly "curl https://paste.debian.net/plain/1146089 |
qemu-system-i386 ..." I have only a warning:

    qemu-system-i386: Guest moved used index from 0 to 496

I've added some traces.

The assert is triggered because addr (0xffff0) >= cache->len (0x11d0).

addr is 2nd argument of:

    address_space_read_cached(cache, i * sizeof(VRingDesc),
                              desc, sizeof(VRingDesc));

and "i" appears to be "65535".

In virtqueue_split_read_next_desc(), the value is checked not to be
greater than max.

But max is 268345360...

"max" is provided by virtqueue_split_get_avail_bytes(), it's originally
vq->vring.num (255), but in the case of VRING_DESC_F_INDIRECT max is
updated to "desc.len / sizeof(VRingDesc)". desc.len is 4293525760.

It is set from max to cache->len by address_space_cache_init() but this
value seems to be truncated in the next iteration of the loop (where we
have the assert()).

I'm investigating why we have that.

Thanks,
Laurent