On Mon, Sep 14, 2015 at 08:32:36AM +0200, Thomas Huth wrote: > On 14/09/15 04:15, David Gibson wrote: > > On Fri, Sep 11, 2015 at 11:17:01AM +0200, Thomas Huth wrote: > >> The PAPR interface defines a hypercall to pass high-quality > >> hardware generated random numbers to guests. Recent kernels can > >> already provide this hypercall to the guest if the right hardware > >> random number generator is available. But in case the user wants > >> to use another source like EGD, or QEMU is running with an older > >> kernel, we should also have this call in QEMU, so that guests that > >> do not support virtio-rng yet can get good random numbers, too. > >> > >> This patch now adds a new pseude-device to QEMU that either > >> directly provides this hypercall to the guest or is able to > >> enable the in-kernel hypercall if available. The in-kernel > >> hypercall can be enabled with the use-kvm property, e.g.: > >> > >> qemu-system-ppc64 -device spapr-rng,use-kvm=true > >> > >> For handling the hypercall in QEMU instead, a RngBackend is required > >> since the hypercall should provide "good" random data instead of > >> pseudo-random (like from a "simple" library function like rand() > >> or g_random_int()). Since there are multiple RngBackends available, > >> the user must select an appropriate backend via the "backend" > >> property of the device, e.g.: > >> > >> qemu-system-ppc64 -object rng-random,filename=/dev/hwrng,id=rng0 \ > >> -device spapr-rng,backend=rng0 ... > >> > >> See http://wiki.qemu-project.org/Features-Done/VirtIORNG for > >> other example of specifying RngBackends. > ... > >> + > >> +#include "qemu/error-report.h" > >> +#include "sysemu/sysemu.h" > >> +#include "sysemu/device_tree.h" > >> +#include "sysemu/rng.h" > >> +#include "hw/ppc/spapr.h" > >> +#include "kvm_ppc.h" > >> + > >> +#define SPAPR_RNG(obj) \ > >> + OBJECT_CHECK(sPAPRRngState, (obj), TYPE_SPAPR_RNG) > >> + > >> +typedef struct sPAPRRngState { > >> + /*< private >*/ > >> + DeviceState ds; > >> + RngBackend *backend; > >> + bool use_kvm; > >> +} sPAPRRngState; > >> + > >> +typedef struct HRandomData { > >> + QemuSemaphore sem; > >> + union { > >> + uint64_t v64; > >> + uint8_t v8[8]; > >> + } val; > >> + int received; > >> +} HRandomData; > >> + > >> +/* Callback function for the RngBackend */ > >> +static void random_recv(void *dest, const void *src, size_t size) > >> +{ > >> + HRandomData *hrdp = dest; > >> + > >> + if (src && size > 0) { > >> + assert(size + hrdp->received <= sizeof(hrdp->val.v8)); > >> + memcpy(&hrdp->val.v8[hrdp->received], src, size); > >> + hrdp->received += size; > >> + } > >> + > >> + qemu_sem_post(&hrdp->sem); > > > > I'm assuming qemu_sem_post() includes the necessary memory barrier to > > make sure the requesting thread actually sees the data. > > Not sure whether I fully got your point here... both callback function > and main thread are calling an extern C-function, so the compiler should > not assume that the memory stays the same in the main thread...?
I'm not talking about a compiler barrier: the callback will likely be
invoked on a different CPU from the vcpu thread that invoked H_RANDOM,
so on a weakly ordered arch like Power we need a real CPU memory barrier.
> Anyway, I've tested the hypercall by implementing it in SLOF and calling
> it a couple of times there to see that all bits in the result behave
> randomly, so for me this is working fine.
Right, I'd be almost certain anyway that qemu_sem_post() (actually
likely the pthreads functions it invokes) will include the necessary
barriers to stop accesses leaking outside the locked region.
>
> >> +}
> >> +
> >> +/* Handler for the H_RANDOM hypercall */
> >> +static target_ulong h_random(PowerPCCPU *cpu, sPAPRMachineState *spapr,
> >> + target_ulong opcode, target_ulong *args)
> >> +{
> >> + sPAPRRngState *rngstate;
> >> + HRandomData hrdata;
> >> +
> >> + rngstate = SPAPR_RNG(object_resolve_path_type("", TYPE_SPAPR_RNG,
> >> NULL));
> >> +
> >> + if (!rngstate || !rngstate->backend) {
> >> + return H_HARDWARE;
> >> + }
> >> +
> >> + qemu_sem_init(&hrdata.sem, 0);
> >> + hrdata.val.v64 = 0;
> >> + hrdata.received = 0;
> >> +
> >> + qemu_mutex_unlock_iothread();
> >> + while (hrdata.received < 8) {
> >> + rng_backend_request_entropy(rngstate->backend, 8 -
> >> hrdata.received,
> >> + random_recv, &hrdata);
> >> + qemu_sem_wait(&hrdata.sem);
> >> + }
> >> + qemu_mutex_lock_iothread();
> >> +
> >> + qemu_sem_destroy(&hrdata.sem);
> >> + args[0] = hrdata.val.v64;
> >> +
> >> + return H_SUCCESS;
> >> +}
> >> +
> >> +static void spapr_rng_instance_init(Object *obj)
> >> +{
> >> + sPAPRRngState *rngstate = SPAPR_RNG(obj);
> >> +
> >> + if (object_resolve_path_type("", TYPE_SPAPR_RNG, NULL) != NULL) {
> >> + error_report("spapr-rng can not be instantiated twice!");
> >> + return;
> >> + }
> >> +
> >> + object_property_add_link(obj, "backend", TYPE_RNG_BACKEND,
> >> + (Object **)&rngstate->backend,
> >> + object_property_allow_set_link,
> >> + OBJ_PROP_LINK_UNREF_ON_RELEASE, NULL);
> >> + object_property_set_description(obj, "backend",
> >> + "ID of the random number generator
> >> backend",
> >> + NULL);
> >
> > Since virtio-rng does it the same way, I'm assuming there's a reason
> > this is constructed with object_propery_add() rather than listing it
> > in spapr_rng_properties, but it's not obvious what the reason is.
>
> I did not spot a macro a la "DEFINE_PROP_LINK" that could be used for
> this. Do you see a possibility to define a link that way?
No, like I say since virtio-rng does it that way I assume there's a reason.
> > More importantly, this should probably be called "rng" not "backend"
> > to match virtio-rng.
>
> Since the device is already called "spapr-rng", i.e. has "rng" in its
> name, I'd rather like to keep this as "backend" to make it clear that
> you specify the backend this way.
Hm, personally I'd weigh consistency with virtio-rng higher than the
slightly confusing name.
>
> >> +}
> >> +
> >> +static void spapr_rng_realize(DeviceState *dev, Error **errp)
> >> +{
> >> +
> >> + sPAPRRngState *rngstate = SPAPR_RNG(dev);
> >> +
> >> + if (rngstate->use_kvm) {
> >> + if (kvmppc_enable_hwrng() != 0) {
> >> + error_setg(errp, "Could not initialize in-kernel H_RANDOM
> >> call!");
> >> + }
> >> + return;
> >> + }
> >> +
> >> + if (!rngstate->backend) {
> >> + error_setg(errp, "spapr-rng needs a RNG backend!");
> >> + return;
> >> + }
> >
> > So, the logic here means you have to explicitly choose whether to use
> > the kernel implementation or the qemu imeplementation.
> >
> > It seems to me it might be useful to be able to specify "use the
> > kernel implementation if available, otherwise fall back to qemu".
>
> Right, makes sense, I'll update this logic.
Ok.
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
pgpXitL35sR4n.pgp
Description: PGP signature
