core dumps running in bhyve

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

core dumps running in bhyve

Chuck Tuffli-2
Using the latest bhyve, I'm seeing core dumps in the guest when running:
    nvmecontrol identify nvme0
against the emulated NVMe drive. The location of the core dump changes
from run to run, but I suspect the root cause is a memory corruption
caused by the transfer of the Identify data (4KB) back to the guest.
This transfer of data is actually a memcpy to an address returned from
vm_map_gpa() based on the physical address provided by the guest.

Based on the signature of one of the core dumps, I modified
nvmecontrol to always pass a 4KB aligned buffer to the driver instead
of the (typically) unaligned address of the structure on the stack.
With this change, nvmecontrol in the guest no longer core dumps. What
I don't understand is why this changes the behavior. Do the addresses
passed to vm_map_gpa() need to be page aligned? Or did moving the
memory location from the stack to the heap merely mitigate what is
corrupted?

Thoughts?

--chuck
_______________________________________________
[hidden email] mailing list
https://lists.freebsd.org/mailman/listinfo/freebsd-emulation
To unsubscribe, send any mail to "[hidden email]"
Reply | Threaded
Open this post in threaded view
|

Re: core dumps running in bhyve

Fabian Freyer
CCing freebsd-virtualization@, because they might know more about this.

Am 25.12.2018 um 02:24 schrieb Chuck Tuffli:

> Using the latest bhyve, I'm seeing core dumps in the guest when running:
>      nvmecontrol identify nvme0
> against the emulated NVMe drive. The location of the core dump changes
> from run to run, but I suspect the root cause is a memory corruption
> caused by the transfer of the Identify data (4KB) back to the guest.
> This transfer of data is actually a memcpy to an address returned from
> vm_map_gpa() based on the physical address provided by the guest.
>
> Based on the signature of one of the core dumps, I modified
> nvmecontrol to always pass a 4KB aligned buffer to the driver instead
> of the (typically) unaligned address of the structure on the stack.
> With this change, nvmecontrol in the guest no longer core dumps. What
> I don't understand is why this changes the behavior. Do the addresses
> passed to vm_map_gpa() need to be page aligned?

AFAIK vm_map_gpa maps a page, so yes, it needs to be 4k-aligned.

> Or did moving the
> memory location from the stack to the heap merely mitigate what is
> corrupted?
>
> Thoughts?
>
> --chuck
> _______________________________________________
> [hidden email] mailing list
> https://lists.freebsd.org/mailman/listinfo/freebsd-hackers
> To unsubscribe, send any mail to "[hidden email]"
>
_______________________________________________
[hidden email] mailing list
https://lists.freebsd.org/mailman/listinfo/freebsd-emulation
To unsubscribe, send any mail to "[hidden email]"