|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH 02/10] x86: split populating of struct vcpu_time_info into a separate function
On 17.01.2023 21:19, Andrew Cooper wrote:
> On 19/10/2022 8:39 am, Jan Beulich wrote:
>> This is to facilitate subsequent re-use of this code.
>>
>> While doing so add const in a number of places, extending to
>> gtime_to_gtsc() and then for symmetry also its inverse function.
>>
>> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
>
> Acked-by: Andrew Cooper <andrew.cooper@xxxxxxxxxx>
Thanks.
>> ---
>> I was on the edge of also folding the various is_hvm_domain() into a
>> function scope boolean, but then wasn't really certain that this
>> wouldn't open up undue speculation opportunities.
>
> I can't see anything interesting under here speculation wise.
> Commentary inline.
My interpretation of those comments is that the suggested conversion
would be okay-ish (as in not making things worse), but since you didn't
provide an explicit answer I thought I'd better ask for confirmation
before possibly making a patch to that effect.
Jan
>> --- a/xen/arch/x86/include/asm/time.h
>> +++ b/xen/arch/x86/include/asm/time.h
>> @@ -52,8 +52,8 @@ uint64_t cf_check acpi_pm_tick_to_ns(uin
>> uint64_t tsc_ticks2ns(uint64_t ticks);
>>
>> uint64_t pv_soft_rdtsc(const struct vcpu *v, const struct cpu_user_regs
>> *regs);
>> -u64 gtime_to_gtsc(struct domain *d, u64 time);
>> -u64 gtsc_to_gtime(struct domain *d, u64 tsc);
>> +uint64_t gtime_to_gtsc(const struct domain *d, uint64_t time);
>> +uint64_t gtsc_to_gtime(const struct domain *d, uint64_t tsc);
>>
>> int tsc_set_info(struct domain *d, uint32_t tsc_mode, uint64_t elapsed_nsec,
>> uint32_t gtsc_khz, uint32_t incarnation);
>> --- a/xen/arch/x86/time.c
>> +++ b/xen/arch/x86/time.c
>> @@ -1373,18 +1373,14 @@ uint64_t tsc_ticks2ns(uint64_t ticks)
>> return scale_delta(ticks, &t->tsc_scale);
>> }
>>
>> -static void __update_vcpu_system_time(struct vcpu *v, int force)
>> +static void collect_time_info(const struct vcpu *v,
>> + struct vcpu_time_info *u)
>> {
>> - const struct cpu_time *t;
>> - struct vcpu_time_info *u, _u = {};
>> - struct domain *d = v->domain;
>> + const struct cpu_time *t = &this_cpu(cpu_time);
>> + const struct domain *d = v->domain;
>> s_time_t tsc_stamp;
>>
>> - if ( v->vcpu_info == NULL )
>> - return;
>> -
>> - t = &this_cpu(cpu_time);
>> - u = &vcpu_info(v, time);
>> + memset(u, 0, sizeof(*u));
>>
>> if ( d->arch.vtsc )
>> {
>> @@ -1392,7 +1388,7 @@ static void __update_vcpu_system_time(st
>>
>> if ( is_hvm_domain(d) )
>> {
>> - struct pl_time *pl = v->domain->arch.hvm.pl_time;
>> + const struct pl_time *pl = d->arch.hvm.pl_time;
>
> A PV guest could in in principle use...
>
>>
>> stime += pl->stime_offset + v->arch.hvm.stime_offset;
>
> ... this pl->stime_offset as the second deference of a whatever happens
> to sit under d->arch.hvm.pl_time in the pv union.
>
> In the current build of Xen I have to hand, that's
> d->arch.pv.mapcache.{epoch,tlbflush_timestamp}, the combination of which
> doesn't seem like it can be steered into being a legal pointer into Xen.
>
>> if ( stime >= 0 )
>> @@ -1403,27 +1399,27 @@ static void __update_vcpu_system_time(st
>> else
>> tsc_stamp = gtime_to_gtsc(d, stime);
>>
>> - _u.tsc_to_system_mul = d->arch.vtsc_to_ns.mul_frac;
>> - _u.tsc_shift = d->arch.vtsc_to_ns.shift;
>> + u->tsc_to_system_mul = d->arch.vtsc_to_ns.mul_frac;
>> + u->tsc_shift = d->arch.vtsc_to_ns.shift;
>> }
>> else
>> {
>> if ( is_hvm_domain(d) && hvm_tsc_scaling_supported )
>
> On the other hand, this is isn't safe. There's no protection of the &&
> calculation, but...
>
>> {
>> tsc_stamp = hvm_scale_tsc(d, t->stamp.local_tsc);
>
> ... this path is the only path subject to speculative type confusion,
> and all it does is read d->arch.hvm.tsc_scaling_ratio, so is
> appropriately protected in this instance.
>
> Also, all an attacker could do is encode the scaling ratio alongside
> t->stamp.local_tsc (unpredictable) in the calculation for the duration
> of the speculative window, with no way I can see to dereference the result.
>
>
>> - _u.tsc_to_system_mul = d->arch.vtsc_to_ns.mul_frac;
>> - _u.tsc_shift = d->arch.vtsc_to_ns.shift;
>> + u->tsc_to_system_mul = d->arch.vtsc_to_ns.mul_frac;
>> + u->tsc_shift = d->arch.vtsc_to_ns.shift;
>> }
>> else
>> {
>> tsc_stamp = t->stamp.local_tsc;
>> - _u.tsc_to_system_mul = t->tsc_scale.mul_frac;
>> - _u.tsc_shift = t->tsc_scale.shift;
>> + u->tsc_to_system_mul = t->tsc_scale.mul_frac;
>> + u->tsc_shift = t->tsc_scale.shift;
>> }
>> }
>>
>> - _u.tsc_timestamp = tsc_stamp;
>> - _u.system_time = t->stamp.local_stime;
>> + u->tsc_timestamp = tsc_stamp;
>> + u->system_time = t->stamp.local_stime;
>>
>> /*
>> * It's expected that domains cope with this bit changing on every
>> @@ -1431,10 +1427,21 @@ static void __update_vcpu_system_time(st
>> * or if it further requires monotonicity checks with other vcpus.
>> */
>> if ( clocksource_is_tsc() )
>> - _u.flags |= XEN_PVCLOCK_TSC_STABLE_BIT;
>> + u->flags |= XEN_PVCLOCK_TSC_STABLE_BIT;
>>
>> if ( is_hvm_domain(d) )
>> - _u.tsc_timestamp += v->arch.hvm.cache_tsc_offset;
>> + u->tsc_timestamp += v->arch.hvm.cache_tsc_offset;
>
> This path is subject to type confusion on v->arch.{pv,hvm}, with a PV
> guest able to encode the value of v->arch.pv.ctrlreg[5] into the
> timestamp. But again, no way to dereference the result.
>
>
> I really don't think there's enough flexibility here for even a
> perfectly-timed attacker to abuse.
>
> ~Andrew
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |