Bug 1272522

A domain's xenoprofile state contains an array of per-vcpu information, which is allocated once in the lifetime of a domain in response to that domain using the XENOPROF_get_buffer hypercall on itself or by a domain with the privilege to profile a target domain using the XENOPROF_set_passive hypercall. This array is leaked on domain teardown. This memory leak could over time exhaust the host's memory.

The following parties can mount a denial of service attack affecting the whole system:

- A malicious guest administrator via XENOPROF_get_buffer.
- A domain given suitable privilege over another domain via XENOPROF_set_passive (this would usually be a domain being used to profile another domain, eg with the xenoprof tool).

The ability to also restart or create suitable domains is also required to fully exploit the issue. Without this the leak is limited to a small multiple of the maximum number of vcpus for the domain. The maximum leak is 128kbytes per domain (re)boot.

Versions of Xen from 4.0 onwards are vulnerable. The XENOPROF hypercalls are only implemented on x86. ARM is therefore not vulnerable.

Mitigation:

On systems where the guest kernel is controlled by the host rather than guest administrator, running only kernels (in the target and profiling domain respectively) which do not call these hypercalls will also prevent untrusted guest users from exploiting this issue. However untrusted guest administrators can still trigger it unless further steps are taken to prevent them from loading code into the kernel (e.g. by disabling loadable modules etc) or from using other mechanisms which allow them to run code at kernel privilege. The leak is small. Preventing the creation of large numbers of new domains, and limiting the number of times an existing domain can be rebooted, can reduce the impact of this vulnerability.