The block layer in Linux may choose to merge adjacent block IO requests.
When Linux is running as a Xen guest, the default merging algorithm is
replaced with a Xen-specific one. When Linux is running as an x86 PV
guest, some BIO's are erroneously merged, corrupting the data stream
to/from the block device.
This can result in incorrect access to an uncontrolled adjacent frame.
A buggy or malicious guest can cause Linux to read or write incorrect
memory when processing a block stream. This could leak information from
other guests in the system or from Xen itself, or be used to DoS or
escalate privilege within the system.
All x86 Xen systems using pvops Linux in a backend role (either as
dom0, or as a disk device driver domain) are affected. This includes
upstream Linux versions 2.6.37 and later. Systems using the older
classic-linux fork are not affected.
All PV x86 domains doing block IO on behalf of a guest, including dom0
and any PV driver domains, are vulnerable. (Any HVM driver domains
running are not vulnerable.) This includes Xen vbd backends such as
blkback, but also direct IO performed for the guest via eg qemu.
ARM systems are not affected.
The vulnerability is only exposed if the underlying block device has
request merging enabled. See Mitigation.
The vulnerability is only exposed to configurations which use grant
mapping as a transport mechanism for the block data. Configurations
which use exclusively grant copy are not vulnerable.
Disable bio merges on all relevant underlying backend block devices.
echo 2 > /sys/block/nvme0n1/queue/nomerges
Name: the Xen project
Upstream: Jan H. Schönherr (Amazon)
Created xen tracking bugs for this issue:
Affects: fedora-all [bug 1481765]