Disclose Three Bugs in xhyve

Foreword

Earlier this year, I stumbled upon an interesting advisory from ZDI, which was yet to be published back then.

It said that a vulnerability exits in xhyve, and was scheduled to be disclosed in a month.

So I decided to do a code review of xhyve, to see that if I could locate the bug before the disclosure.

In the end, I find 4 bugs, all of them are technically 1-Day, and have been already fixed in the upstream or downstream projects.

Unfortunately, ZDI did not publish this advisory untill recently, so I was not sure if any one of these 4 bugs was the one mentioned in the advisory.

Turns out, I did find it, and it is a bug in ./src/pci_e82545.c:1410, the e1000 virtual device code and has been found and fixed in xhyve’s upstream project bhyve.

Since ZDI still gives CVE-2022-35867 to this bug, I guess it makes sense to do a responsible disclosure of the other 3 bugs here.

All of the other 3 bugs were found by GitHub Security Lab in HyperKit, a project based on xhyve, and disclosed in GHSL-2021-054_057 advisory.

HyperKit is a toolkit for embedding hypervisor capabilities in your application. It includes a complete hypervisor, based on xhyve/bhyve, which is optimized for lightweight virtual machines and container deployment. It is designed to be interfaced with higher-level components such as the VPNKit and DataKit.

Since the security patches in HyperKit were not upstreamed, the latest version of xhyve still has these 4 bugs.

pci_virtio_rnd vc_cfgread NULL Pointer Dereference

Virtio devices use struct virtio_consts to register handler functions for device IO.

struct virtio_consts {
	/* name of driver (for diagnostics) */
	const char *vc_name;
	/* number of virtual queues */
	int vc_nvq;
	/* size of dev-specific config regs */
	size_t vc_cfgsize;
	/* called on virtual device reset */
	void (*vc_reset)(void *);
	/* called on QNOTIFY if no VQ notify */
	void (*vc_qnotify)(void *, struct vqueue_info *);
	/* called to read config regs */
	int (*vc_cfgread)(void *, int, int, uint32_t *);
	/* called to write config regs */
	int (*vc_cfgwrite)(void *, int, int, uint32_t);
	/* called to apply negotiated features */
	void (*vc_apply_features)(void *, uint64_t);
	/* hypervisor-provided capabilities */
	uint64_t vc_hv_caps;
};

In the case of pci_virtio_rnd, a virtio entropy device emulation, vc_cfgread is initialized to NULL.

static struct virtio_consts vtrnd_vi_consts = {
	"vtrnd", /* our name */
	1, /* we support 1 virtqueue */
	0, /* config reg size */
	pci_vtrnd_reset, /* reset */
	pci_vtrnd_notify, /* device-wide qnotify */
	NULL, /* read virtio config */
	NULL, /* write virtio config */
	NULL, /* apply negotiated features */
	0, /* our capabilities */
};

Yet, in vi_pci_read(), the existance of the vc_cfgread handler is not checked.

error = (*vc->vc_cfgread)(DEV_SOFTC(vs), ((int) newoff), size, &value);

So, when a malicious guest triggers the vc_cfgread IO, a null pointer dereference will crash the guest vm process.

pci_virtio_rnd vc_cfgwrite NULL Pointer Dereference

Similarly, in vi_pci_write(), the existance of the vc_cfgwrite handler is also not checked.

error = (*vc->vc_cfgwrite)(DEV_SOFTC(vs), ((int) newoff), size,
    ((uint32_t) value));

So, when a malicious guest triggers the vc_cfgwrite IO, a null pointer dereference will crash the guest vm process.

pci_virtio_rnd Uninitialized Stack Buffer Usage

In the pci_vtrnd_notify() handler function, the return value of vq_getchain() is not checked, so if it returns due to some failed check,

iov stack variable will be left uninitialized, and iov.iov_base will be used as the buffer to store random data read form sc->vrsc_fd.

static void
pci_vtrnd_notify(void *vsc, struct vqueue_info *vq)
{
    struct iovec iov;
        ...
        vq_getchain(vq, &idx, &iov, 1, NULL);

        len = (int) read(sc->vrsc_fd, iov.iov_base, iov.iov_len);
        ...

Theoretically, this bug is exploitable. One possible way is to leave an address of the length of some buffer on the stack as iov.iov_base.

Since the data from sc->vrsc_fd is typically very large, we can overwrite the length with a larger value and achieve an OOB read/write primitive.

However, during my brief test, the stack position of iov seems to be too shallow to avoid being overwritten before pci_vtrnd_notify() is actually reached.

Afterword

At the time of writing, July 21, 2022, all of these 4 bugs have not been fixed by xhyve, and the project’s last commit was back on Oct 2, 2021.

Besides, none of the issues, open or closed, mentions any of these 4 bugs.

According to my understanding, Homebrew and MacPorts do not provide aditional security patch themselves,

so both builds from their repository are also affected.

Sometimes, maintainers of some small open source projects may not be able to do a timely fix.

Until then, xhyve users may remove the pci_virtio_rnd and e1000 device from their vm.