Clean up the KVM clock mess somewhat so that it is either based on the guest
TSC ("master clock" mode), or on the host CLOCK_MONOTONIC_RAW in cases where
the TSC isn't usable.
Eliminate the third variant where it was based directly on the *host* TSC,
due to bugs in e.g. __get_kvmclock().
Kill off the last vestiges of the KVM clock being based on CLOCK_MONOTONIC
instead of CLOCK_MONOTONIC_RAW and thus being subject to NTP skew.
Fix up migration support to allow the KVM clock to be saved/restored as an
arithmetic function of the guest TSC, since that's what it actually is in
the *common* case so it can be migrated precisely. Or at least to within
±1 ns which is good enough, as discussed in
https://lore.kernel.org/kvm/[email protected]
In v2 of this series, TSC synchronization is improved and simplified a bit
too, and we allow masterclock mode to be used even when the guest TSCs are
out of sync, as long as they're running at the same *rate*. The different
*offset* shouldn't matter.
And the kvm_get_time_scale() function annoyed me by being entirely opaque,
so I studied it until my brain hurt and then added some comments.
In v2 I also dropped the commits which were removing the periodic clock
syncs. Those are going to be needed still but *only* for non-masterclock
mode, which I'll do next. Along with ensuring that a masterclock update
while already in masterclock mode doesn't jump the clock, and just does
the same as KVM_SET_CLOCK_GUEST does to preserve it.
Needs a *lot* more testing. I think I'm almost done refactoring the code,
so should focus on building up the tests next.
(I do still hate that we're abusing KVM_GET_CLOCK just to get the tuple
of {host_tsc, CLOCK_REALTIME} without even *caring* about the eponymous
KVM clock. Especially as this information is (a) fundamentally what the
vDSO gettimeofday() exposes to us anyway, (b) using CLOCK_REALTIME not
TAI, (c) not available on other platforms, for example for migrating
the Arm arch counter.)
David Woodhouse (13):
KVM: x86/xen: Do not corrupt KVM clock in kvm_xen_shared_info_init()
KVM: x86: Improve accuracy of KVM clock when TSC scaling is in force
KVM: x86: Explicitly disable TSC scaling without CONSTANT_TSC
KVM: x86: Add KVM_VCPU_TSC_SCALE and fix the documentation on TSC migration
KVM: x86: Avoid NTP frequency skew for KVM clock on 32-bit host
KVM: x86: Fix KVM clock precision in __get_kvmclock()
KVM: x86: Fix software TSC upscaling in kvm_update_guest_time()
KVM: x86: Simplify and comment kvm_get_time_scale()
KVM: x86: Remove implicit rdtsc() from kvm_compute_l1_tsc_offset()
KVM: x86: Improve synchronization in kvm_synchronize_tsc()
KVM: x86: Kill cur_tsc_{nsec,offset,write} fields
KVM: x86: Allow KVM master clock mode when TSCs are offset from each other
KVM: x86: Factor out kvm_use_master_clock()
Jack Allister (2):
KVM: x86: Add KVM_[GS]ET_CLOCK_GUEST for accurate KVM clock migration
KVM: selftests: Add KVM/PV clock selftest to prove timer correction
Documentation/virt/kvm/api.rst | 37 ++
Documentation/virt/kvm/devices/vcpu.rst | 115 +++-
arch/x86/include/asm/kvm_host.h | 15 +-
arch/x86/include/uapi/asm/kvm.h | 6 +
arch/x86/kvm/svm/svm.c | 3 +-
arch/x86/kvm/vmx/vmx.c | 2 +-
arch/x86/kvm/x86.c | 687 +++++++++++++++-------
arch/x86/kvm/xen.c | 4 +-
include/uapi/linux/kvm.h | 3 +
tools/testing/selftests/kvm/Makefile | 1 +
tools/testing/selftests/kvm/x86_64/pvclock_test.c | 192 ++++++
11 files changed, 822 insertions(+), 243 deletions(-)
From: David Woodhouse <[email protected]>
There is no reason why the KVM clock cannot be in masterclock mode when
the TSCs are not in sync, as long as they are at the same *frequency*.
Running at a different frequency would lead to a systemic skew between
the clock(s) as observed by different vCPUs due to arithmetic precision
in the scaling. So that should indeed force the clock to be based on the
host's CLOCK_MONOTONIC_RAW instead of being in masterclock mode where it
is defined by the (or 'a') guest TSC.
But when the vCPUs merely have a different TSC *offset*, that's not a
problem. The offset is applied to that vCPU's kvmclock->tsc_timestamp
field, and it all comes out in the wash.
So, remove ka->nr_vcpus_matched_tsc and replace it with a new field
ka->all_vcpus_matched_tsc which is not only changed to a boolean, but
also now tracks that the *frequency* matches, not the precise offset.
Using a *count* was always racy because a new vCPU could be being
created *while* kvm_track_tsc_matching() was running and comparing with
kvm->online_vcpus. That variable is only atomic with respect to itself.
In particular, kvm_arch_vcpu_create() runs before kvm->online_vcpus is
incremented for the new vCPU, and kvm_arch_vcpu_postcreate() runs later.
Repurpose kvm_track_tsc_matching() to be called from kvm_set_tsc_khz(),
and kill the cur_tsc_generation/last_tsc_generation fields which tracked
the precise TSC matching.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/include/asm/kvm_host.h | 6 +-
arch/x86/kvm/x86.c | 133 ++++++++++++++++----------------
2 files changed, 70 insertions(+), 69 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 7d06f389a607..52dbb2d7690b 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -906,7 +906,6 @@ struct kvm_vcpu_arch {
u64 tsc_offset_adjustment;
u64 this_tsc_nsec;
u64 this_tsc_write;
- u64 this_tsc_generation;
bool tsc_catchup;
bool tsc_always_catchup;
s8 virtual_tsc_shift;
@@ -1354,11 +1353,10 @@ struct kvm_arch {
u32 last_tsc_khz;
u64 last_tsc_offset;
u64 last_tsc_scaling_ratio;
- u64 cur_tsc_generation;
- int nr_vcpus_matched_tsc;
+ bool all_vcpus_matched_tsc;
- u32 default_tsc_khz;
bool user_set_tsc;
+ u32 default_tsc_khz;
seqcount_raw_spinlock_t pvclock_sc;
bool use_master_clock;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 92e81bfca25a..d6e4469f531a 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2493,40 +2493,6 @@ static int set_tsc_khz(struct kvm_vcpu *vcpu, u32 user_tsc_khz, bool scale)
return 0;
}
-static int kvm_set_tsc_khz(struct kvm_vcpu *vcpu, u32 user_tsc_khz)
-{
- u32 thresh_lo, thresh_hi;
- int use_scaling = 0;
-
- /* tsc_khz can be zero if TSC calibration fails */
- if (user_tsc_khz == 0) {
- /* set tsc_scaling_ratio to a safe value */
- kvm_vcpu_write_tsc_multiplier(vcpu, kvm_caps.default_tsc_scaling_ratio);
- return -1;
- }
-
- /* Compute a scale to convert nanoseconds in TSC cycles */
- kvm_get_time_scale(user_tsc_khz * 1000LL, NSEC_PER_SEC,
- &vcpu->arch.virtual_tsc_shift,
- &vcpu->arch.virtual_tsc_mult);
- vcpu->arch.virtual_tsc_khz = user_tsc_khz;
-
- /*
- * Compute the variation in TSC rate which is acceptable
- * within the range of tolerance and decide if the
- * rate being applied is within that bounds of the hardware
- * rate. If so, no scaling or compensation need be done.
- */
- thresh_lo = adjust_tsc_khz(tsc_khz, -tsc_tolerance_ppm);
- thresh_hi = adjust_tsc_khz(tsc_khz, tsc_tolerance_ppm);
- if (user_tsc_khz < thresh_lo || user_tsc_khz > thresh_hi) {
- pr_debug("requested TSC rate %u falls outside tolerance [%u,%u]\n",
- user_tsc_khz, thresh_lo, thresh_hi);
- use_scaling = 1;
- }
- return set_tsc_khz(vcpu, user_tsc_khz, use_scaling);
-}
-
static u64 compute_guest_tsc(struct kvm_vcpu *vcpu, s64 kernel_ns)
{
s64 delta = kernel_ns - vcpu->arch.this_tsc_nsec;
@@ -2557,14 +2523,34 @@ static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
#ifdef CONFIG_X86_64
struct kvm_arch *ka = &vcpu->kvm->arch;
struct pvclock_gtod_data *gtod = &pvclock_gtod_data;
+ struct kvm_vcpu *v;
+ unsigned long i;
+ bool matched = true;
+
+ lockdep_assert_held(&vcpu->kvm->arch.tsc_write_lock);
+
+ /*
+ * When called via kvm_arch_vcpu_create() for a new vCPU, the count
+ * in kvm->online_vcpus will not yet have been incremented and the
+ * kvm_for_each_vcpu() loop will not iterate over 'vcpu'.
+ *
+ * The right thing still happens in that case, because a match will
+ * be found only if the new vCPU's TSC is at the same rate as *all*
+ * the existing vCPUs' TSCs.
+ */
+ kvm_for_each_vcpu(i, v, vcpu->kvm) {
+ if (v->arch.virtual_tsc_khz != vcpu->arch.virtual_tsc_khz) {
+ matched = false;
+ break;
+ }
+ }
+ ka->all_vcpus_matched_tsc = matched;
/*
* To use the masterclock, the host clocksource must be based on TSC
- * and all vCPUs must have matching TSCs. Note, the count for matching
- * vCPUs doesn't include the reference vCPU, hence "+1".
+ * and all vCPUs must have matching TSC frequencies.
*/
- bool use_master_clock = (ka->nr_vcpus_matched_tsc + 1 ==
- atomic_read(&vcpu->kvm->online_vcpus)) &&
+ bool use_master_clock = ka->all_vcpus_matched_tsc &&
gtod_is_based_on_tsc(gtod->clock.vclock_mode);
/*
@@ -2574,12 +2560,51 @@ static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
if ((ka->use_master_clock != use_master_clock))
kvm_make_request(KVM_REQ_MASTERCLOCK_UPDATE, vcpu);
- trace_kvm_track_tsc(vcpu->vcpu_id, ka->nr_vcpus_matched_tsc,
+ trace_kvm_track_tsc(vcpu->vcpu_id, ka->all_vcpus_matched_tsc,
atomic_read(&vcpu->kvm->online_vcpus),
ka->use_master_clock, gtod->clock.vclock_mode);
#endif
}
+static int kvm_set_tsc_khz(struct kvm_vcpu *vcpu, u32 user_tsc_khz)
+{
+ u32 thresh_lo, thresh_hi;
+ int use_scaling = 0;
+
+ /* tsc_khz can be zero if TSC calibration fails */
+ if (user_tsc_khz == 0) {
+ /* set tsc_scaling_ratio to a safe value */
+ kvm_vcpu_write_tsc_multiplier(vcpu, kvm_caps.default_tsc_scaling_ratio);
+ return -1;
+ }
+
+ /* Compute a scale to convert nanoseconds in TSC cycles */
+ kvm_get_time_scale(user_tsc_khz * 1000LL, NSEC_PER_SEC,
+ &vcpu->arch.virtual_tsc_shift,
+ &vcpu->arch.virtual_tsc_mult);
+
+ raw_spin_lock_irq(&vcpu->kvm->arch.tsc_write_lock);
+ vcpu->arch.virtual_tsc_khz = user_tsc_khz;
+ kvm_track_tsc_matching(vcpu);
+ raw_spin_unlock_irq(&vcpu->kvm->arch.tsc_write_lock);
+
+ /*
+ * Compute the variation in TSC rate which is acceptable
+ * within the range of tolerance and decide if the
+ * rate being applied is within that bounds of the hardware
+ * rate. If so, no scaling or compensation need be done.
+ */
+ thresh_lo = adjust_tsc_khz(tsc_khz, -tsc_tolerance_ppm);
+ thresh_hi = adjust_tsc_khz(tsc_khz, tsc_tolerance_ppm);
+ if (user_tsc_khz < thresh_lo || user_tsc_khz > thresh_hi) {
+ pr_debug("requested TSC rate %u falls outside tolerance [%u,%u]\n",
+ user_tsc_khz, thresh_lo, thresh_hi);
+ use_scaling = 1;
+ }
+ return set_tsc_khz(vcpu, user_tsc_khz, use_scaling);
+}
+
+
/*
* Multiply tsc by a fixed point number represented by ratio.
*
@@ -2725,27 +2750,6 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
vcpu->arch.last_guest_tsc = tsc;
kvm_vcpu_write_tsc_offset(vcpu, offset);
-
- if (!matched) {
- /*
- * We split periods of matched TSC writes into generations.
- * For each generation, we track the original measured
- * nanosecond time, offset, and write, so if TSCs are in
- * sync, we can match exact offset, and if not, we can match
- * exact software computation in compute_guest_tsc()
- *
- * These values are tracked in kvm->arch.cur_xxx variables.
- */
- kvm->arch.cur_tsc_generation++;
- kvm->arch.nr_vcpus_matched_tsc = 0;
- } else if (vcpu->arch.this_tsc_generation != kvm->arch.cur_tsc_generation) {
- kvm->arch.nr_vcpus_matched_tsc++;
- }
-
- /* Keep track of which generation this VCPU has synchronized to */
- vcpu->arch.this_tsc_generation = kvm->arch.cur_tsc_generation;
-
- kvm_track_tsc_matching(vcpu);
}
static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
@@ -3072,11 +3076,9 @@ static void pvclock_update_vm_gtod_copy(struct kvm *kvm)
struct kvm_arch *ka = &kvm->arch;
uint64_t last_tsc_hz;
int vclock_mode;
- bool host_tsc_clocksource, vcpus_matched;
+ bool host_tsc_clocksource;
lockdep_assert_held(&kvm->arch.tsc_write_lock);
- vcpus_matched = (ka->nr_vcpus_matched_tsc + 1 ==
- atomic_read(&kvm->online_vcpus));
/*
* If the host uses TSC clock, then passthrough TSC as stable
@@ -3086,7 +3088,8 @@ static void pvclock_update_vm_gtod_copy(struct kvm *kvm)
&ka->master_kernel_ns,
&ka->master_cycle_now);
- ka->use_master_clock = host_tsc_clocksource && vcpus_matched
+ ka->use_master_clock = host_tsc_clocksource
+ && ka->all_vcpus_matched_tsc
&& !ka->backwards_tsc_observed
&& !ka->boot_vcpu_runs_old_kvmclock;
@@ -3124,7 +3127,7 @@ static void pvclock_update_vm_gtod_copy(struct kvm *kvm)
vclock_mode = pvclock_gtod_data.clock.vclock_mode;
trace_kvm_update_master_clock(ka->use_master_clock, vclock_mode,
- vcpus_matched);
+ ka->all_vcpus_matched_tsc);
#endif
}
--
2.44.0
From: David Woodhouse <[email protected]>
The kvm_guest_time_update() function scales the host TSC frequency to
the guest's using kvm_scale_tsc() and the v->arch.l1_tsc_scaling_ratio
scaling ratio previously calculated for that vCPU. Then calcuates the
scaling factors for the KVM clock itself based on that guest TSC
frequency.
However, it uses kHz as the unit when scaling, and then multiplies by
1000 only at the end.
With a host TSC frequency of 3000MHz and a guest set to 2500MHz, the
result of kvm_scale_tsc() will actually come out at 2,499,999kHz. So
the KVM clock advertised to the guest is based on a frequency of
2,499,999,000 Hz.
By using Hz as the unit from the beginning, the KVM clock would be based
on a more accurate frequency of 2,499,999,999 Hz in this example.
Fixes: 78db6a503796 ("KVM: x86: rewrite handling of scaled TSC for kvmclock")
Signed-off-by: David Woodhouse <[email protected]>
Reviewed-by: Paul Durrant <[email protected]>
---
arch/x86/include/asm/kvm_host.h | 2 +-
arch/x86/kvm/x86.c | 17 +++++++++--------
arch/x86/kvm/xen.c | 2 +-
3 files changed, 11 insertions(+), 10 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 01c69840647e..8440c4081727 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -887,7 +887,7 @@ struct kvm_vcpu_arch {
gpa_t time;
struct pvclock_vcpu_time_info hv_clock;
- unsigned int hw_tsc_khz;
+ unsigned int hw_tsc_hz;
struct gfn_to_pfn_cache pv_time;
/* set guest stopped flag in pvclock flags field */
bool pvclock_set_guest_stopped_request;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 2d2619d3eee4..23281c508c27 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -3215,7 +3215,8 @@ static void kvm_setup_guest_pvclock(struct kvm_vcpu *v,
static int kvm_guest_time_update(struct kvm_vcpu *v)
{
- unsigned long flags, tgt_tsc_khz;
+ unsigned long flags;
+ uint64_t tgt_tsc_hz;
unsigned seq;
struct kvm_vcpu_arch *vcpu = &v->arch;
struct kvm_arch *ka = &v->kvm->arch;
@@ -3252,8 +3253,8 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
/* Keep irq disabled to prevent changes to the clock */
local_irq_save(flags);
- tgt_tsc_khz = get_cpu_tsc_khz();
- if (unlikely(tgt_tsc_khz == 0)) {
+ tgt_tsc_hz = get_cpu_tsc_khz() * 1000LL;
+ if (unlikely(tgt_tsc_hz == 0)) {
local_irq_restore(flags);
kvm_make_request(KVM_REQ_CLOCK_UPDATE, v);
return 1;
@@ -3288,14 +3289,14 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
/* With all the info we got, fill in the values */
if (kvm_caps.has_tsc_control)
- tgt_tsc_khz = kvm_scale_tsc(tgt_tsc_khz,
- v->arch.l1_tsc_scaling_ratio);
+ tgt_tsc_hz = kvm_scale_tsc(tgt_tsc_hz,
+ v->arch.l1_tsc_scaling_ratio);
- if (unlikely(vcpu->hw_tsc_khz != tgt_tsc_khz)) {
- kvm_get_time_scale(NSEC_PER_SEC, tgt_tsc_khz * 1000LL,
+ if (unlikely(vcpu->hw_tsc_hz != tgt_tsc_hz)) {
+ kvm_get_time_scale(NSEC_PER_SEC, tgt_tsc_hz,
&vcpu->hv_clock.tsc_shift,
&vcpu->hv_clock.tsc_to_system_mul);
- vcpu->hw_tsc_khz = tgt_tsc_khz;
+ vcpu->hw_tsc_hz = tgt_tsc_hz;
kvm_xen_update_tsc_info(v);
}
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 5a83a8154b79..014048c22652 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -2273,7 +2273,7 @@ void kvm_xen_update_tsc_info(struct kvm_vcpu *vcpu)
entry = kvm_find_cpuid_entry_index(vcpu, function, 2);
if (entry)
- entry->eax = vcpu->arch.hw_tsc_khz;
+ entry->eax = vcpu->arch.hw_tsc_hz / 1000;
}
void kvm_xen_init_vm(struct kvm *kvm)
--
2.44.0
From: David Woodhouse <[email protected]>
Let the callers pass the host TSC value in as an explicit parameter.
This leaves some fairly obviously stupid code, which using this function
to compare the guest TSC at some *other* time, with the newly-minted TSC
value from rdtsc(). Unless it's being used to measure *elapsed* time,
that isn't very sensible.
In this case, "obviously stupid" is an improvement over being non-obviously
so.
No functional change intended.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/kvm/x86.c | 14 ++++++++------
1 file changed, 8 insertions(+), 6 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index ef3cd6113037..ea59694d712a 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2601,11 +2601,12 @@ u64 kvm_scale_tsc(u64 tsc, u64 ratio)
return _tsc;
}
-static u64 kvm_compute_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 target_tsc)
+static u64 kvm_compute_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 host_tsc,
+ u64 target_tsc)
{
u64 tsc;
- tsc = kvm_scale_tsc(rdtsc(), vcpu->arch.l1_tsc_scaling_ratio);
+ tsc = kvm_scale_tsc(host_tsc, vcpu->arch.l1_tsc_scaling_ratio);
return target_tsc - tsc;
}
@@ -2758,7 +2759,7 @@ static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
bool synchronizing = false;
raw_spin_lock_irqsave(&kvm->arch.tsc_write_lock, flags);
- offset = kvm_compute_l1_tsc_offset(vcpu, data);
+ offset = kvm_compute_l1_tsc_offset(vcpu, rdtsc(), data);
ns = get_kvmclock_base_ns();
elapsed = ns - kvm->arch.last_tsc_nsec;
@@ -2809,7 +2810,7 @@ static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
} else {
u64 delta = nsec_to_cycles(vcpu, elapsed);
data += delta;
- offset = kvm_compute_l1_tsc_offset(vcpu, data);
+ offset = kvm_compute_l1_tsc_offset(vcpu, rdtsc(), data);
}
matched = true;
}
@@ -4024,7 +4025,8 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
if (msr_info->host_initiated) {
kvm_synchronize_tsc(vcpu, &data);
} else {
- u64 adj = kvm_compute_l1_tsc_offset(vcpu, data) - vcpu->arch.l1_tsc_offset;
+ u64 adj = kvm_compute_l1_tsc_offset(vcpu, rdtsc(), data) -
+ vcpu->arch.l1_tsc_offset;
adjust_tsc_offset_guest(vcpu, adj);
vcpu->arch.ia32_tsc_adjust_msr += adj;
}
@@ -5098,7 +5100,7 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
mark_tsc_unstable("KVM discovered backwards TSC");
if (kvm_check_tsc_unstable()) {
- u64 offset = kvm_compute_l1_tsc_offset(vcpu,
+ u64 offset = kvm_compute_l1_tsc_offset(vcpu, rdtsc(),
vcpu->arch.last_guest_tsc);
kvm_vcpu_write_tsc_offset(vcpu, offset);
vcpu->arch.tsc_catchup = 1;
--
2.44.0
From: David Woodhouse <[email protected]>
KVM does make an attempt to cope with non-constant TSC, and has notifiers
to handle host TSC frequency changes. However, it *only* adjusts the KVM
clock, and doesn't adjust TSC frequency scaling when the host changes.
This is presumably because non-constant TSCs were fixed in hardware long
before TSC scaling was implemented, so there should never be real CPUs
which have TSC scaling but *not* CONSTANT_TSC.
Such a combination could potentially happen in some odd L1 nesting
environment, but it isn't worth trying to support it. Just make the
dependency explicit.
Signed-off-by: David Woodhouse <[email protected]>
Reviewed-by: Paul Durrant <[email protected]>
---
arch/x86/kvm/svm/svm.c | 3 ++-
arch/x86/kvm/vmx/vmx.c | 2 +-
2 files changed, 3 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 0f3b59da0d4a..4d3ec1c3231e 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -5202,7 +5202,8 @@ static __init int svm_hardware_setup(void)
kvm_enable_efer_bits(EFER_FFXSR);
if (tsc_scaling) {
- if (!boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
+ if (!boot_cpu_has(X86_FEATURE_TSCRATEMSR) ||
+ !boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) {
tsc_scaling = false;
} else {
pr_info("TSC scaling supported\n");
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index 6780313914f8..bee830adf744 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -8428,7 +8428,7 @@ __init int vmx_hardware_setup(void)
if (!enable_apicv || !cpu_has_vmx_ipiv())
enable_ipiv = false;
- if (cpu_has_vmx_tsc_scaling())
+ if (cpu_has_vmx_tsc_scaling() && boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
kvm_caps.has_tsc_control = true;
kvm_caps.max_tsc_scaling_ratio = KVM_VMX_TSC_MULTIPLIER_MAX;
--
2.44.0
From: David Woodhouse <[email protected]>
The KVM clock is an interesting thing. It is defined as "nanoseconds
since the guest was created", but in practice it runs at two *different*
rates — or three different rates, if you count implementation bugs.
Definition A is that it runs synchronously with the CLOCK_MONOTONIC_RAW
of the host, with a delta of kvm->arch.kvmclock_offset.
But that version doesn't actually get used in the common case, where the
host has a reliable TSC and the guest TSCs are all running at the same
rate and in sync with each other, and kvm->arch.use_master_clock is set.
In that common case, definition B is used: There is a reference point in
time at kvm->arch.master_kernel_ns (again a CLOCK_MONOTONIC_RAW time),
and a corresponding host TSC value kvm->arch.master_cycle_now. This
fixed point in time is converted to guest units (the time offset by
kvmclock_offset and the TSC Value scaled and offset to be a guest TSC
value) and advertised to the guest in the pvclock structure. While in
this 'use_master_clock' mode, the fixed point in time never needs to be
changed, and the clock runs precisely in time with the guest TSC, at the
rate advertised in the pvclock structure.
The third definition C is implemented in kvm_get_wall_clock_epoch() and
__get_kvmclock(), using the master_cycle_now and master_kernel_ns fields
but converting the *host* TSC cycles directly to a value in nanoseconds
instead of scaling via the guest TSC.
One might naïvely think that all three definitions are identical, since
CLOCK_MONOTONIC_RAW is not skewed by NTP frequency corrections; all
three are just the result of counting the host TSC at a known frequency,
or the scaled guest TSC at a known precise fraction of the host's
frequency. The problem is with arithmetic precision, and the way that
frequency scaling is done in a division-free way by multiplying by a
scale factor, then shifting right. In practice, all three ways of
calculating the KVM clock will suffer a systemic drift from each other.
Eventually, definition C should just be eliminated. Commit 451a707813ae
("KVM: x86/xen: improve accuracy of Xen timers") worked around it for
the specific case of Xen timers, which are defined in terms of the KVM
clock and suffered from a continually increasing error in timer expiry
times. That commit notes that get_kvmclock_ns() is non-trivial to fix
and says "I'll come back to that", which remains true.
Definitions A and B do need to coexist, the former to handle the case
where the host or guest TSC is suboptimally configured. But KVM should
be more careful about switching between them, and the discontinuity in
guest time which could result.
In particular, KVM_REQ_MASTERCLOCK_UPDATE will take a new snapshot of
time as the reference in master_kernel_ns and master_cycle_now, yanking
the guest's clock back to match definition A at that moment.
When invoked from in 'use_master_clock' mode, kvm_update_masterclock()
should probably *adjust* kvm->arch.kvmclock_offset to account for the
drift, instead of yanking the clock back to defintion A. But in the
meantime there are a bunch of places where it just doesn't need to be
invoked at all.
To start with: there is no need to do such an update when a Xen guest
populates the shared_info page. This seems to have been a hangover from
the very first implementation of shared_info which automatically
populated the vcpu_info structures at their default locations, but even
then it should just have raised KVM_REQ_CLOCK_UPDATE on each vCPU
instead of using KVM_REQ_MASTERCLOCK_UPDATE. And now that userspace is
expected to explicitly set the vcpu_info even in its default locations,
there's not even any need for that either.
Fixes: 629b5348841a1 ("KVM: x86/xen: update wallclock region")
Signed-off-by: David Woodhouse <[email protected]>
Reviewed-by: Paul Durrant <[email protected]>
---
arch/x86/kvm/xen.c | 2 --
1 file changed, 2 deletions(-)
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index f65b35a05d91..5a83a8154b79 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -98,8 +98,6 @@ static int kvm_xen_shared_info_init(struct kvm *kvm)
wc->version = wc_version + 1;
read_unlock_irq(&gpc->lock);
- kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
-
out:
srcu_read_unlock(&kvm->srcu, idx);
return ret;
--
2.44.0
From: David Woodhouse <[email protected]>
There was some confusion in kvm_update_guest_time() when software needs
to advance the guest TSC.
In master clock mode, there are two points of time which need to be taken
into account. First there is the master clock reference point, stored in
kvm->arch.master_kernel_ns (and associated host TSC ->master_cycle_now).
Secondly, there is the time *now*, at the point kvm_update_guest_time()
is being called.
With software TSC upscaling, the guest TSC is getting further and further
ahead of the host TSC as time elapses. So at time "now", the guest TSC
should be further ahead of the host, than it was at master_kernel_ns.
The adjustment in kvm_update_guest_time() was not taking that into
account, and was only advancing the guest TSC by the appropriate amount
for master_kernel_ns, *not* the current time.
Fix it to calculate them both correctly.
Since the KVM clock reference point in master_kernel_ns might actually
be *earlier* than the reference point used for the guest TSC
(vcpu->last_tsc_nsec), this might lead to a negative delta. Fix the
compute_guest_tsc() function to cope with negative numbers, which
then means there is no need to force a master clock update when the
guest TSC is written.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/kvm/x86.c | 73 +++++++++++++++++++++++++++++++++++-----------
1 file changed, 56 insertions(+), 17 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 89918ba266cd..e09dc44978ea 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2491,10 +2491,19 @@ static int kvm_set_tsc_khz(struct kvm_vcpu *vcpu, u32 user_tsc_khz)
static u64 compute_guest_tsc(struct kvm_vcpu *vcpu, s64 kernel_ns)
{
- u64 tsc = pvclock_scale_delta(kernel_ns-vcpu->arch.this_tsc_nsec,
- vcpu->arch.virtual_tsc_mult,
- vcpu->arch.virtual_tsc_shift);
- tsc += vcpu->arch.this_tsc_write;
+ s64 delta = kernel_ns - vcpu->arch.this_tsc_nsec;
+ u64 tsc = vcpu->arch.this_tsc_write;
+
+ /* pvclock_scale_delta cannot cope with negative deltas */
+ if (delta >= 0)
+ tsc += pvclock_scale_delta(delta,
+ vcpu->arch.virtual_tsc_mult,
+ vcpu->arch.virtual_tsc_shift);
+ else
+ tsc -= pvclock_scale_delta(-delta,
+ vcpu->arch.virtual_tsc_mult,
+ vcpu->arch.virtual_tsc_shift);
+
return tsc;
}
@@ -2505,7 +2514,7 @@ static inline bool gtod_is_based_on_tsc(int mode)
}
#endif
-static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu, bool new_generation)
+static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
{
#ifdef CONFIG_X86_64
struct kvm_arch *ka = &vcpu->kvm->arch;
@@ -2522,12 +2531,9 @@ static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu, bool new_generation)
/*
* Request a masterclock update if the masterclock needs to be toggled
- * on/off, or when starting a new generation and the masterclock is
- * enabled (compute_guest_tsc() requires the masterclock snapshot to be
- * taken _after_ the new generation is created).
+ * on/off.
*/
- if ((ka->use_master_clock && new_generation) ||
- (ka->use_master_clock != use_master_clock))
+ if ((ka->use_master_clock != use_master_clock))
kvm_make_request(KVM_REQ_MASTERCLOCK_UPDATE, vcpu);
trace_kvm_track_tsc(vcpu->vcpu_id, ka->nr_vcpus_matched_tsc,
@@ -2705,7 +2711,7 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
vcpu->arch.this_tsc_nsec = kvm->arch.cur_tsc_nsec;
vcpu->arch.this_tsc_write = kvm->arch.cur_tsc_write;
- kvm_track_tsc_matching(vcpu, !matched);
+ kvm_track_tsc_matching(vcpu);
}
static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
@@ -3300,8 +3306,6 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
kernel_ns = get_kvmclock_base_ns();
}
- tsc_timestamp = kvm_read_l1_tsc(v, host_tsc);
-
/*
* We may have to catch up the TSC to match elapsed wall clock
* time for two reasons, even if kvmclock is used.
@@ -3313,11 +3317,46 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
* very slowly.
*/
if (vcpu->tsc_catchup) {
- u64 tsc = compute_guest_tsc(v, kernel_ns);
- if (tsc > tsc_timestamp) {
- adjust_tsc_offset_guest(v, tsc - tsc_timestamp);
- tsc_timestamp = tsc;
+ uint64_t now_host_tsc, now_guest_tsc;
+ int64_t adjustment;
+
+ /*
+ * First, calculate what the guest TSC should be at the
+ * time (kernel_ns) which will be placed in the hvclock.
+ * This may be the *current* time, or it may be the time
+ * of the master clock reference. This is 'tsc_timestamp'.
+ */
+ tsc_timestamp = compute_guest_tsc(v, kernel_ns);
+
+ now_guest_tsc = tsc_timestamp;
+ now_host_tsc = host_tsc;
+
+#ifdef CONFIG_X86_64
+ /*
+ * If the master clock was used, calculate what the guest
+ * TSC should be *now* in order to advance to that.
+ */
+ if (use_master_clock) {
+ int64_t now_kernel_ns;
+
+ if (!kvm_get_time_and_clockread(&now_kernel_ns,
+ &now_host_tsc)) {
+ now_kernel_ns = get_kvmclock_base_ns();
+ now_host_tsc = rdtsc();
+ }
+ now_guest_tsc = compute_guest_tsc(v, now_kernel_ns);
}
+#endif
+ /*
+ * Calculate the delta between what the guest TSC *should* be,
+ * and what it actually is according to kvm_read_l1_tsc().
+ */
+ adjustment = now_guest_tsc - kvm_read_l1_tsc(v, now_host_tsc);
+
+ if (adjustment > 0)
+ adjust_tsc_offset_guest(v, adjustment);
+ } else {
+ tsc_timestamp = kvm_read_l1_tsc(v, host_tsc);
}
local_irq_restore(flags);
--
2.44.0
From: David Woodhouse <[email protected]>
Commit 53fafdbb8b21 ("KVM: x86: switch KVMCLOCK base to monotonic raw
clock") did so only for 64-bit hosts, by capturing the boot offset from
within the existing clocksource notifier update_pvclock_gtod().
That notifier was added in commit 16e8d74d2da9 ("KVM: x86: notifier for
clocksource changes") but only on x86_64, because its original purpose
was just to disable the "master clock" mode which is only supported on
x86_64.
Now that the notifier is used for more than disabling master clock mode,
(well, OK, more than a decade later but clocks are hard), enable it for
the 32-bit build too so that get_kvmclock_base_ns() can be unaffected by
NTP sync on 32-bit too.
Signed-off-by: David Woodhouse <[email protected]>
Reviewed-by: Paul Durrant <[email protected]>
---
arch/x86/kvm/x86.c | 18 ++++++------------
1 file changed, 6 insertions(+), 12 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 00a7c1188dec..44b3d2a0da5b 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2245,7 +2245,6 @@ static int do_set_msr(struct kvm_vcpu *vcpu, unsigned index, u64 *data)
return kvm_set_msr_ignored_check(vcpu, index, *data, true);
}
-#ifdef CONFIG_X86_64
struct pvclock_clock {
int vclock_mode;
u64 cycle_last;
@@ -2303,13 +2302,6 @@ static s64 get_kvmclock_base_ns(void)
/* Count up from boot time, but with the frequency of the raw clock. */
return ktime_to_ns(ktime_add(ktime_get_raw(), pvclock_gtod_data.offs_boot));
}
-#else
-static s64 get_kvmclock_base_ns(void)
-{
- /* Master clock not used, so we can just use CLOCK_BOOTTIME. */
- return ktime_get_boottime_ns();
-}
-#endif
static void kvm_write_wall_clock(struct kvm *kvm, gpa_t wall_clock, int sec_hi_ofs)
{
@@ -9819,6 +9811,7 @@ static void pvclock_irq_work_fn(struct irq_work *w)
}
static DEFINE_IRQ_WORK(pvclock_irq_work, pvclock_irq_work_fn);
+#endif
/*
* Notification about pvclock gtod data update.
@@ -9826,26 +9819,26 @@ static DEFINE_IRQ_WORK(pvclock_irq_work, pvclock_irq_work_fn);
static int pvclock_gtod_notify(struct notifier_block *nb, unsigned long unused,
void *priv)
{
- struct pvclock_gtod_data *gtod = &pvclock_gtod_data;
struct timekeeper *tk = priv;
update_pvclock_gtod(tk);
+#ifdef CONFIG_X86_64
/*
* Disable master clock if host does not trust, or does not use,
* TSC based clocksource. Delegate queue_work() to irq_work as
* this is invoked with tk_core.seq write held.
*/
- if (!gtod_is_based_on_tsc(gtod->clock.vclock_mode) &&
+ if (!gtod_is_based_on_tsc(pvclock_gtod_data.clock.vclock_mode) &&
atomic_read(&kvm_guest_has_master_clock) != 0)
irq_work_queue(&pvclock_irq_work);
+#endif
return 0;
}
static struct notifier_block pvclock_gtod_notifier = {
.notifier_call = pvclock_gtod_notify,
};
-#endif
static inline void kvm_ops_update(struct kvm_x86_init_ops *ops)
{
@@ -9984,9 +9977,10 @@ int kvm_x86_vendor_init(struct kvm_x86_init_ops *ops)
if (pi_inject_timer == -1)
pi_inject_timer = housekeeping_enabled(HK_TYPE_TIMER);
-#ifdef CONFIG_X86_64
+
pvclock_gtod_register_notifier(&pvclock_gtod_notifier);
+#ifdef CONFIG_X86_64
if (hypervisor_is_type(X86_HYPER_MS_HYPERV))
set_hv_tscchange_cb(kvm_hyperv_tsc_notifier);
#endif
--
2.44.0
From: David Woodhouse <[email protected]>
When synchronizing to an existing TSC (either by explicitly writing zero,
or the legacy hack where the TSC is written within one second's worth of
the previously written TSC), the last_tsc_write and last_tsc_nsec values
were being misrecorded by __kvm_synchronize_tsc(). The *unsynchronized*
value of the TSC (perhaps even zero) was bring recorded, along with the
current time at which kvm_synchronize_tsc() was called. This could cause
*subsequent* writes to fail to synchronize correctly.
Fix that by resetting {data, ns} to the previous values before passing
them to __kvm_synchronize_tsc() when synchronization is detected. Except
in the case where the TSC is unstable and *has* to be synthesised from
the host clock, in which case attempt to create a nsec/tsc pair which is
on the correct line.
Furthermore, there were *three* different TSC reads used for calculating
the "current" time, all slightly different from each other. Fix that by
using kvm_get_time_and_clockread() where possible and using the same
host_tsc value in all cases.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/kvm/x86.c | 32 ++++++++++++++++++++++++++++----
1 file changed, 28 insertions(+), 4 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index ea59694d712a..6ec43f39bdb0 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -201,6 +201,10 @@ module_param(eager_page_split, bool, 0644);
static bool __read_mostly mitigate_smt_rsb;
module_param(mitigate_smt_rsb, bool, 0444);
+#ifdef CONFIG_X86_64
+static bool kvm_get_time_and_clockread(s64 *kernel_ns, u64 *tsc_timestamp);
+#endif
+
/*
* Restoring the host value for MSRs that are only consumed when running in
* usermode, e.g. SYSCALL MSRs and TSC_AUX, can be deferred until the CPU
@@ -2753,14 +2757,22 @@ static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
{
u64 data = user_value ? *user_value : 0;
struct kvm *kvm = vcpu->kvm;
- u64 offset, ns, elapsed;
+ u64 offset, host_tsc, ns, elapsed;
unsigned long flags;
bool matched = false;
bool synchronizing = false;
+#ifdef CONFIG_X86_64
+ if (!kvm_get_time_and_clockread(&ns, &host_tsc))
+#endif
+ {
+ ns = get_kvmclock_base_ns();
+ host_tsc = rdtsc();
+ }
+
raw_spin_lock_irqsave(&kvm->arch.tsc_write_lock, flags);
- offset = kvm_compute_l1_tsc_offset(vcpu, rdtsc(), data);
- ns = get_kvmclock_base_ns();
+
+ offset = kvm_compute_l1_tsc_offset(vcpu, host_tsc, data);
elapsed = ns - kvm->arch.last_tsc_nsec;
if (vcpu->arch.virtual_tsc_khz) {
@@ -2805,12 +2817,24 @@ static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
*/
if (synchronizing &&
vcpu->arch.virtual_tsc_khz == kvm->arch.last_tsc_khz) {
+ /*
+ * If synchronizing, the "last written" TSC value/time recorded
+ * by __kvm_synchronize_tsc() should not change (i.e. should
+ * be precisely the same as the existing generation)...
+ */
+ data = kvm->arch.last_tsc_write;
+
if (!kvm_check_tsc_unstable()) {
offset = kvm->arch.cur_tsc_offset;
+ ns = kvm->arch.cur_tsc_nsec;
} else {
+ /*
+ * ... unless the TSC is unstable and has to be
+ * synthesised from the host clock in nanoseconds.
+ */
u64 delta = nsec_to_cycles(vcpu, elapsed);
data += delta;
- offset = kvm_compute_l1_tsc_offset(vcpu, rdtsc(), data);
+ offset = kvm_compute_l1_tsc_offset(vcpu, host_tsc, data);
}
matched = true;
}
--
2.44.0
From: David Woodhouse <[email protected]>
Both kvm_track_tsc_matching() and pvclock_update_vm_gtod_copy() make a
decision about whether the KVM clock should be in master clock mode.
They use *different* criteria for the decision though. This isn't really
a problem; it only has the potential to cause unnecessary invocations of
KVM_REQ_MASTERCLOCK_UPDATE if the masterclock was disabled due to TSC
going backwards, or the guest using the old MSR. But it isn't pretty.
Factor the decision out to a single function. And document the historical
reason why it's disabled for guests that use the old MSR_KVM_SYSTEM_TIME.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/kvm/x86.c | 27 +++++++++++++++++++++++----
1 file changed, 23 insertions(+), 4 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index d6e4469f531a..680b39f17851 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2518,6 +2518,27 @@ static inline bool gtod_is_based_on_tsc(int mode)
}
#endif
+static bool kvm_use_master_clock(strut kvm *kvm)
+{
+ struct kvm_arch *ka = &kvm->arch;
+
+ /*
+ * The 'old kvmclock' check is a workaround (from 2015) for a
+ * SUSE 2.6.16 kernel that didn't boot if the system_time in
+ * its kvmclock was too far behind the current time. So the
+ * mode of just setting the reference point and allowing time
+ * to proceed linearly from there makes it fail to boot.
+ * Despite that being kind of the *point* of the way the clock
+ * is exposed to the guest. By coincidence, the offending
+ * kernels used the old MSR_KVM_SYSTEM_TIME, which was moved
+ * only because it resided in the wrong number range. So the
+ * workaround is activated for *all* guests using the old MSR.
+ */
+ return ka->all_vcpus_matched_tsc &&
+ !ka->backwards_tsc_observed &&
+ !ka->boot_vcpu_runs_old_kvmclock;
+}
+
static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
{
#ifdef CONFIG_X86_64
@@ -2550,7 +2571,7 @@ static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
* To use the masterclock, the host clocksource must be based on TSC
* and all vCPUs must have matching TSC frequencies.
*/
- bool use_master_clock = ka->all_vcpus_matched_tsc &&
+ bool use_master_clock = kvm_use_master_clock(kvm) &&
gtod_is_based_on_tsc(gtod->clock.vclock_mode);
/*
@@ -3089,9 +3110,7 @@ static void pvclock_update_vm_gtod_copy(struct kvm *kvm)
&ka->master_cycle_now);
ka->use_master_clock = host_tsc_clocksource
- && ka->all_vcpus_matched_tsc
- && !ka->backwards_tsc_observed
- && !ka->boot_vcpu_runs_old_kvmclock;
+ && kvm_use_master_clock(kvm);
/*
* When TSC scaling is in use (which can thankfully only happen
--
2.44.0
From: David Woodhouse <[email protected]>
These pointlessly duplicate of the last_tsc_{nsec,offset,write} values.
The only place they were used was where the TSC is stable and a new vCPU
is being synchronized to the previous setting, in which case the 'last_'
value is definitely identical.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/include/asm/kvm_host.h | 3 ---
arch/x86/kvm/x86.c | 9 ++-------
2 files changed, 2 insertions(+), 10 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index b01c1d000fff..7d06f389a607 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1354,9 +1354,6 @@ struct kvm_arch {
u32 last_tsc_khz;
u64 last_tsc_offset;
u64 last_tsc_scaling_ratio;
- u64 cur_tsc_nsec;
- u64 cur_tsc_write;
- u64 cur_tsc_offset;
u64 cur_tsc_generation;
int nr_vcpus_matched_tsc;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 6ec43f39bdb0..92e81bfca25a 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2737,9 +2737,6 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
* These values are tracked in kvm->arch.cur_xxx variables.
*/
kvm->arch.cur_tsc_generation++;
- kvm->arch.cur_tsc_nsec = ns;
- kvm->arch.cur_tsc_write = tsc;
- kvm->arch.cur_tsc_offset = offset;
kvm->arch.nr_vcpus_matched_tsc = 0;
} else if (vcpu->arch.this_tsc_generation != kvm->arch.cur_tsc_generation) {
kvm->arch.nr_vcpus_matched_tsc++;
@@ -2747,8 +2744,6 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
/* Keep track of which generation this VCPU has synchronized to */
vcpu->arch.this_tsc_generation = kvm->arch.cur_tsc_generation;
- vcpu->arch.this_tsc_nsec = kvm->arch.cur_tsc_nsec;
- vcpu->arch.this_tsc_write = kvm->arch.cur_tsc_write;
kvm_track_tsc_matching(vcpu);
}
@@ -2825,8 +2820,8 @@ static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
data = kvm->arch.last_tsc_write;
if (!kvm_check_tsc_unstable()) {
- offset = kvm->arch.cur_tsc_offset;
- ns = kvm->arch.cur_tsc_nsec;
+ offset = kvm->arch.last_tsc_offset;
+ ns = kvm->arch.last_tsc_nsec;
} else {
/*
* ... unless the TSC is unstable and has to be
--
2.44.0
From: David Woodhouse <[email protected]>
When in 'master clock mode' (i.e. when host and guest TSCs are behaving
sanely and in sync), the KVM clock is defined in terms of the guest TSC.
When TSC scaling is used, calculating the KVM clock directly from *host*
TSC cycles leads to a systemic drift from the values calculated by the
guest from its TSC.
Commit 451a707813ae ("KVM: x86/xen: improve accuracy of Xen timers")
had a simple workaround for the specific case of Xen timers, as it had an
actual vCPU to hand and could use its scaling information. That commit
noted that it was broken for the general case of get_kvmclock_ns(), and
said "I'll come back to that".
Since __get_kvmclock() is invoked without a specific CPU, it needs to
be able to find or generate the scaling values required to perform the
correct calculation.
Thankfully, TSC scaling can only happen with X86_FEATURE_CONSTANT_TSC,
so it isn't as complex as it might have been.
In __kvm_synchronize_tsc(), note the current vCPU's scaling ratio in
kvm->arch.last_tsc_scaling_ratio. That is only protected by the
tsc_write_lock, so in pvclock_update_vm_gtod_copy(), copy it into a
separate kvm->arch.master_tsc_scaling_ratio so that it can be accessed
using the kvm->arch.pvclock_sc seqcount lock. Also generate the mul and
shift factors to convert to nanoseconds for the corresponding KVM clock,
just as kvm_guest_time_update() would.
In __get_kvmclock(), which runs within a seqcount retry loop, use those
values to convert host to guest TSC and then to nanoseconds. Only fall
back to using get_kvmclock_base_ns() when not in master clock mode.
There was previously a code path in __get_kvmclock() which looked like
it could set KVM_CLOCK_TSC_STABLE without KVM_CLOCK_REALTIME, perhaps
even on 32-bit hosts. In practice that could never happen as the
ka->use_master_clock flag couldn't be set on 32-bit, and even on 64-bit
hosts it would never be set when the system clock isn't TSC-based. So
that code path is now removed.
The kvm_get_wall_clock_epoch() function had the same problem; make it
just call get_kvmclock() and subtract kvmclock from wallclock, with
the same fallback as before.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/include/asm/kvm_host.h | 4 +
arch/x86/kvm/x86.c | 151 ++++++++++++++++----------------
2 files changed, 79 insertions(+), 76 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 8440c4081727..b01c1d000fff 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1353,6 +1353,7 @@ struct kvm_arch {
u64 last_tsc_write;
u32 last_tsc_khz;
u64 last_tsc_offset;
+ u64 last_tsc_scaling_ratio;
u64 cur_tsc_nsec;
u64 cur_tsc_write;
u64 cur_tsc_offset;
@@ -1364,6 +1365,9 @@ struct kvm_arch {
seqcount_raw_spinlock_t pvclock_sc;
bool use_master_clock;
+ s8 master_tsc_shift;
+ u32 master_tsc_mul;
+ u64 master_tsc_scaling_ratio;
u64 master_kernel_ns;
u64 master_cycle_now;
struct delayed_work kvmclock_update_work;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 44b3d2a0da5b..89918ba266cd 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2674,6 +2674,7 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
kvm->arch.last_tsc_nsec = ns;
kvm->arch.last_tsc_write = tsc;
kvm->arch.last_tsc_khz = vcpu->arch.virtual_tsc_khz;
+ kvm->arch.last_tsc_scaling_ratio = vcpu->arch.l1_tsc_scaling_ratio;
kvm->arch.last_tsc_offset = offset;
vcpu->arch.last_guest_tsc = tsc;
@@ -3009,6 +3010,7 @@ static void pvclock_update_vm_gtod_copy(struct kvm *kvm)
{
#ifdef CONFIG_X86_64
struct kvm_arch *ka = &kvm->arch;
+ uint64_t last_tsc_hz;
int vclock_mode;
bool host_tsc_clocksource, vcpus_matched;
@@ -3028,6 +3030,35 @@ static void pvclock_update_vm_gtod_copy(struct kvm *kvm)
&& !ka->backwards_tsc_observed
&& !ka->boot_vcpu_runs_old_kvmclock;
+ /*
+ * When TSC scaling is in use (which can thankfully only happen
+ * with X86_FEATURE_CONSTANT_TSC), the host must calculate the
+ * KVM clock precisely as the guest would, by scaling through
+ * the guest TSC frequency. Otherwise, differences in arithmetic
+ * precision lead to systemic drift between the guest's and the
+ * host's idea of the time.
+ */
+ if (kvm_caps.has_tsc_control) {
+ /*
+ * Copy from the field protected solely by ka->tsc_write_lock,
+ * to the field protected by the ka->pvclock_sc seqlock.
+ */
+ ka->master_tsc_scaling_ratio = ka->last_tsc_scaling_ratio ? :
+ kvm_caps.default_tsc_scaling_ratio;
+
+ /*
+ * Calculate the scaling factors precisely the same way
+ * that kvm_guest_time_update() does.
+ */
+ last_tsc_hz = kvm_scale_tsc(tsc_khz * 1000LL,
+ ka->master_tsc_scaling_ratio);
+ kvm_get_time_scale(NSEC_PER_SEC, last_tsc_hz,
+ &ka->master_tsc_shift, &ka->master_tsc_mul);
+ } else if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) {
+ kvm_get_time_scale(NSEC_PER_SEC, tsc_khz * 1000LL,
+ &ka->master_tsc_shift, &ka->master_tsc_mul);
+ }
+
if (ka->use_master_clock)
atomic_set(&kvm_guest_has_master_clock, 1);
@@ -3100,36 +3131,49 @@ static unsigned long get_cpu_tsc_khz(void)
static void __get_kvmclock(struct kvm *kvm, struct kvm_clock_data *data)
{
struct kvm_arch *ka = &kvm->arch;
- struct pvclock_vcpu_time_info hv_clock;
+
+#ifdef CONFIG_X86_64
+ uint64_t cur_tsc_khz = 0;
+ struct timespec64 ts;
/* both __this_cpu_read() and rdtsc() should be on the same cpu */
get_cpu();
- data->flags = 0;
if (ka->use_master_clock &&
- (static_cpu_has(X86_FEATURE_CONSTANT_TSC) || __this_cpu_read(cpu_tsc_khz))) {
-#ifdef CONFIG_X86_64
- struct timespec64 ts;
+ (cur_tsc_khz = get_cpu_tsc_khz()) &&
+ !kvm_get_walltime_and_clockread(&ts, &data->host_tsc))
+ cur_tsc_khz = 0;
- if (kvm_get_walltime_and_clockread(&ts, &data->host_tsc)) {
- data->realtime = ts.tv_nsec + NSEC_PER_SEC * ts.tv_sec;
- data->flags |= KVM_CLOCK_REALTIME | KVM_CLOCK_HOST_TSC;
- } else
-#endif
- data->host_tsc = rdtsc();
-
- data->flags |= KVM_CLOCK_TSC_STABLE;
- hv_clock.tsc_timestamp = ka->master_cycle_now;
- hv_clock.system_time = ka->master_kernel_ns + ka->kvmclock_offset;
- kvm_get_time_scale(NSEC_PER_SEC, get_cpu_tsc_khz() * 1000LL,
- &hv_clock.tsc_shift,
- &hv_clock.tsc_to_system_mul);
- data->clock = __pvclock_read_cycles(&hv_clock, data->host_tsc);
- } else {
- data->clock = get_kvmclock_base_ns() + ka->kvmclock_offset;
+ put_cpu();
+
+ if (cur_tsc_khz) {
+ uint64_t tsc_cycles;
+ uint32_t mul;
+ int8_t shift;
+
+ tsc_cycles = data->host_tsc - ka->master_cycle_now;
+
+ if (kvm_caps.has_tsc_control)
+ tsc_cycles = kvm_scale_tsc(tsc_cycles,
+ ka->master_tsc_scaling_ratio);
+
+ if (static_cpu_has(X86_FEATURE_CONSTANT_TSC)) {
+ mul = ka->master_tsc_mul;
+ shift = ka->master_tsc_shift;
+ } else {
+ kvm_get_time_scale(NSEC_PER_SEC, cur_tsc_khz * 1000LL,
+ &shift, &mul);
+ }
+ data->clock = ka->master_kernel_ns + ka->kvmclock_offset +
+ pvclock_scale_delta(tsc_cycles, mul, shift);
+ data->realtime = ts.tv_nsec + NSEC_PER_SEC * ts.tv_sec;
+ data->flags = KVM_CLOCK_REALTIME | KVM_CLOCK_HOST_TSC | KVM_CLOCK_TSC_STABLE;
+ return;
}
+#endif
- put_cpu();
+ data->clock = get_kvmclock_base_ns() + ka->kvmclock_offset;
+ data->flags = 0;
}
static void get_kvmclock(struct kvm *kvm, struct kvm_clock_data *data)
@@ -3330,68 +3374,23 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
* that and kvmclock, but even that would be subject to change over
* time.
*
- * Attempt to calculate the epoch at a given moment using the *same*
- * TSC reading via kvm_get_walltime_and_clockread() to obtain both
- * wallclock and kvmclock times, and subtracting one from the other.
+ * Use get_kvmclock() to obtain a simultaneous reading of wallclock
+ * and kvmclock times from the *same* TSC reading, and subtract one
+ * from the other.
*
* Fall back to using their values at slightly different moments by
* calling ktime_get_real_ns() and get_kvmclock_ns() separately.
*/
uint64_t kvm_get_wall_clock_epoch(struct kvm *kvm)
{
-#ifdef CONFIG_X86_64
- struct pvclock_vcpu_time_info hv_clock;
- struct kvm_arch *ka = &kvm->arch;
- unsigned long seq, local_tsc_khz;
- struct timespec64 ts;
- uint64_t host_tsc;
-
- do {
- seq = read_seqcount_begin(&ka->pvclock_sc);
-
- local_tsc_khz = 0;
- if (!ka->use_master_clock)
- break;
-
- /*
- * The TSC read and the call to get_cpu_tsc_khz() must happen
- * on the same CPU.
- */
- get_cpu();
-
- local_tsc_khz = get_cpu_tsc_khz();
-
- if (local_tsc_khz &&
- !kvm_get_walltime_and_clockread(&ts, &host_tsc))
- local_tsc_khz = 0; /* Fall back to old method */
-
- put_cpu();
-
- /*
- * These values must be snapshotted within the seqcount loop.
- * After that, it's just mathematics which can happen on any
- * CPU at any time.
- */
- hv_clock.tsc_timestamp = ka->master_cycle_now;
- hv_clock.system_time = ka->master_kernel_ns + ka->kvmclock_offset;
+ struct kvm_clock_data data;
- } while (read_seqcount_retry(&ka->pvclock_sc, seq));
+ get_kvmclock(kvm, &data);
- /*
- * If the conditions were right, and obtaining the wallclock+TSC was
- * successful, calculate the KVM clock at the corresponding time and
- * subtract one from the other to get the guest's epoch in nanoseconds
- * since 1970-01-01.
- */
- if (local_tsc_khz) {
- kvm_get_time_scale(NSEC_PER_SEC, local_tsc_khz * NSEC_PER_USEC,
- &hv_clock.tsc_shift,
- &hv_clock.tsc_to_system_mul);
- return ts.tv_nsec + NSEC_PER_SEC * ts.tv_sec -
- __pvclock_read_cycles(&hv_clock, host_tsc);
- }
-#endif
- return ktime_get_real_ns() - get_kvmclock_ns(kvm);
+ if (data.flags & KVM_CLOCK_REALTIME)
+ return data.realtime - data.clock;
+ else
+ return ktime_get_real_ns() - data.clock;
}
/*
--
2.44.0
From: David Woodhouse <[email protected]>
Commit 3ae13faac400 ("KVM: x86: pass kvm_get_time_scale arguments in hertz")
made this function take 64-bit values in Hz rather than 32-bit kHz. Thus
making it entrely pointless to shadow its arguments into local 64-bit
variables. Just use scaled_hz and base_hz directly.
Also rename the 'tps32' variable to 'base32', having utterly failed to
think of any reason why it might have been called that in the first place.
This could probably have been eliminated too, but it helps to make the
code clearer and *might* just help a naïve 32-bit compiler realise that it
doesn't need to do full 64-bit shifts.
Having taken the time to reverse-engineer the function, add some comments
explaining it.
No functional change intended.
Signed-off-by: David Woodhouse <[email protected]>
---
arch/x86/kvm/x86.c | 60 ++++++++++++++++++++++++++++++++++++----------
1 file changed, 47 insertions(+), 13 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index e09dc44978ea..ef3cd6113037 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2375,32 +2375,66 @@ static uint32_t div_frac(uint32_t dividend, uint32_t divisor)
return dividend;
}
+/*
+ * Calculate scaling factors to be applied with pvclock_scale_delta().
+ *
+ * The output of this function is a fixed-point factor which is used to
+ * scale a tick count at base_hz, to a tick count at scaled_hz, within
+ * the limitations of the Xen/KVM pvclock ABI.
+ *
+ * Mathematically, the factor is (*pmultiplier) >> (32 - *pshift).
+ *
+ * Working backwards, the div_frac() function divides (dividend << 32) by
+ * the given divisor, in other words giving dividend/divisor in the form
+ * of a 32-bit fixed-point fraction in the range 0 to 0x0.FFFFFFFF, which
+ * is (*pmultiplier >> 32).
+ *
+ * The rest of the function is shifting the scaled_hz and base_hz left or
+ * right as appropriate to ensure maximal precision within the constraints.
+ *
+ * The first constraint is that the result of the division *must* be less
+ * than 1, which means the dividend (derived from scaled_hz) must be greater
+ * than the divisor (derived from base_hz).
+ *
+ * The second constraint is that for optimal precision, the dividend (scaled)
+ * shouldn't be more than twice the divisor (base) — i.e. the top bit ought
+ * to be set in the resulting *pmultiplier.
+ */
static void kvm_get_time_scale(uint64_t scaled_hz, uint64_t base_hz,
s8 *pshift, u32 *pmultiplier)
{
- uint64_t scaled64;
int32_t shift = 0;
- uint64_t tps64;
- uint32_t tps32;
+ uint32_t base32;
- tps64 = base_hz;
- scaled64 = scaled_hz;
- while (tps64 > scaled64*2 || tps64 & 0xffffffff00000000ULL) {
- tps64 >>= 1;
+ /*
+ * Start by shifting the base_hz right until it fits in 32 bits, and
+ * is lower than double the target rate. This introduces a negative
+ * shift value which would result in pvclock_scale_delta() shifting
+ * the actual tick count right before performing the multiplication.
+ */
+ while (base_hz > scaled_hz*2 || base_hz & 0xffffffff00000000ULL) {
+ base_hz >>= 1;
shift--;
}
- tps32 = (uint32_t)tps64;
- while (tps32 <= scaled64 || scaled64 & 0xffffffff00000000ULL) {
- if (scaled64 & 0xffffffff00000000ULL || tps32 & 0x80000000)
- scaled64 >>= 1;
+ /* Now the shifted base_hz fits in 32 bits, copy it to base32 */
+ base32 = (uint32_t)base_hz;
+
+ /*
+ * Next, shift the scaled_hz right until it fits in 32 bits, and ensure
+ * that the shifted base_hz is not larger (so that the result of the
+ * final division also fits in 32 bits).
+ */
+ while (base32 <= scaled_hz || scaled_hz & 0xffffffff00000000ULL) {
+ if (scaled_hz & 0xffffffff00000000ULL || base32 & 0x80000000)
+ scaled_hz >>= 1;
else
- tps32 <<= 1;
+ base32 <<= 1;
shift++;
}
*pshift = shift;
- *pmultiplier = div_frac(scaled64, tps32);
+ *pmultiplier = div_frac(scaled_hz, base32);
}
#ifdef CONFIG_X86_64
--
2.44.0
From: Jack Allister <[email protected]>
In the common case (where kvm->arch.use_master_clock is true), the KVM
clock is defined as a simple arithmetic function of the guest TSC, based on
a reference point stored in kvm->arch.master_kernel_ns and
kvm->arch.master_cycle_now.
The existing KVM_[GS]ET_CLOCK functionality does not allow for this
relationship to be precisely saved and restored by userspace. All it can
currently do is set the KVM clock at a given UTC reference time, which is
necessarily imprecise.
So on live update, the guest TSC can remain cycle accurate at precisely the
same offset from the host TSC, but there is no way for userspace to restore
the KVM clock accurately.
Even on live migration to a new host, where the accuracy of the guest time-
keeping is fundamentally limited by the accuracy of wallclock
synchronization between the source and destination hosts, the clock jump
experienced by the guest's TSC and its KVM clock should at least be
*consistent*. Even when the guest TSC suffers a discontinuity, its KVM
clock should still remain the *same* arithmetic function of the guest TSC,
and not suffer an *additional* discontinuity.
To allow for accurate migration of the KVM clock, add per-vCPU ioctls which
save and restore the actual PV clock info in pvclock_vcpu_time_info.
The restoration in KVM_SET_CLOCK_GUEST works by creating a new reference
point in time just as kvm_update_masterclock() does, and calculating the
corresponding guest TSC value. This guest TSC value is then passed through
the user-provided pvclock structure to generate the *intended* KVM clock
value at that point in time, and through the *actual* KVM clock calculation.
Then kvm->arch.kvmclock_offset is adjusted to eliminate for the difference.
Where kvm->arch.use_master_clock is false (because the host TSC is
unreliable, or the guest TSCs are configured strangely), the KVM clock
is *not* defined as a function of the guest TSC so KVM_GET_CLOCK_GUEST
returns an error. In this case, as documented, userspace shall use the
legacy KVM_GET_CLOCK ioctl. The loss of precision is acceptable in this
case since the clocks are imprecise in this mode anyway.
On *restoration*, if kvm->arch.use_master_clock is false, an error is
returned for similar reasons and userspace shall fall back to using
KVM_SET_CLOCK. This does mean that, as documented, userspace needs to use
*both* KVM_GET_CLOCK_GUEST and KVM_GET_CLOCK and send both results with the
migration data (unless the intent is to refuse to resume on a host with bad
TSC).
(It may have been possible to make KVM_SET_CLOCK_GUEST "good enough" in the
non-masterclock mode, as that mode is necessarily imprecise anyway. The
explicit fallback allows userspace to deliberately fail migration to a host
with misbehaving TSC where master clock mode wouldn't be active.)
Co-developed-by: David Woodhouse <[email protected]>
Signed-off-by: Jack Allister <[email protected]>
Signed-off-by: David Woodhouse <[email protected]>
Reviewed-by: Paul Durrant <[email protected]>
CC: Dongli Zhang <[email protected]>
---
Documentation/virt/kvm/api.rst | 37 ++++++++
arch/x86/kvm/x86.c | 156 +++++++++++++++++++++++++++++++++
include/uapi/linux/kvm.h | 3 +
3 files changed, 196 insertions(+)
diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
index f0b76ff5030d..758f6fc08fe5 100644
--- a/Documentation/virt/kvm/api.rst
+++ b/Documentation/virt/kvm/api.rst
@@ -6352,6 +6352,43 @@ a single guest_memfd file, but the bound ranges must not overlap).
See KVM_SET_USER_MEMORY_REGION2 for additional details.
+4.143 KVM_GET_CLOCK_GUEST
+----------------------------
+
+:Capability: none
+:Architectures: x86_64
+:Type: vcpu ioctl
+:Parameters: struct pvclock_vcpu_time_info (out)
+:Returns: 0 on success, <0 on error
+
+Retrieves the current time information structure used for KVM/PV clocks,
+in precisely the form advertised to the guest vCPU, which gives parameters
+for a direct conversion from a guest TSC value to nanoseconds.
+
+When the KVM clock not is in "master clock" mode, for example because the
+host TSC is unreliable or the guest TSCs are oddly configured, the KVM clock
+is actually defined by the host CLOCK_MONOTONIC_RAW instead of the guest TSC.
+In this case, the KVM_GET_CLOCK_GUEST ioctl returns -EINVAL.
+
+4.144 KVM_SET_CLOCK_GUEST
+----------------------------
+
+:Capability: none
+:Architectures: x86_64
+:Type: vcpu ioctl
+:Parameters: struct pvclock_vcpu_time_info (in)
+:Returns: 0 on success, <0 on error
+
+Sets the KVM clock (for the whole VM) in terms of the vCPU TSC, using the
+pvclock structure as returned by KVM_GET_CLOCK_GUEST. This allows the precise
+arithmetic relationship between guest TSC and KVM clock to be preserved by
+userspace across migration.
+
+When the KVM clock is not in "master clock" mode, and the KVM clock is actually
+defined by the host CLOCK_MONOTONIC_RAW, this ioctl returns -EINVAL. Userspace
+may choose to set the clock using the less precise KVM_SET_CLOCK ioctl, or may
+choose to fail, denying migration to a host whose TSC is misbehaving.
+
5. The kvm_run structure
========================
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 23281c508c27..42abce7b4fc9 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -5868,6 +5868,154 @@ static int kvm_vcpu_ioctl_enable_cap(struct kvm_vcpu *vcpu,
}
}
+#ifdef CONFIG_X86_64
+static int kvm_vcpu_ioctl_get_clock_guest(struct kvm_vcpu *v, void __user *argp)
+{
+ struct pvclock_vcpu_time_info *hv_clock = &v->arch.hv_clock;
+
+ /*
+ * If KVM_REQ_CLOCK_UPDATE is already pending, or if the hv_clock has
+ * never been generated at all, call kvm_guest_time_update() to do so.
+ * Might as well use the PVCLOCK_TSC_STABLE_BIT as the check for ever
+ * having been written.
+ */
+ if (kvm_check_request(KVM_REQ_CLOCK_UPDATE, v) ||
+ !(hv_clock->flags & PVCLOCK_TSC_STABLE_BIT)) {
+ if (kvm_guest_time_update(v))
+ return -EINVAL;
+ }
+
+ /*
+ * PVCLOCK_TSC_STABLE_BIT is set in use_master_clock mode where the
+ * KVM clock is defined in terms of the guest TSC. Otherwise, it is
+ * is defined by the host CLOCK_MONOTONIC_RAW, and userspace should
+ * use the legacy KVM_[GS]ET_CLOCK to migrate it.
+ */
+ if (!(hv_clock->flags & PVCLOCK_TSC_STABLE_BIT))
+ return -EINVAL;
+
+ if (copy_to_user(argp, hv_clock, sizeof(*hv_clock)))
+ return -EFAULT;
+
+ return 0;
+}
+
+/*
+ * Reverse the calculation in the hv_clock definition.
+ *
+ * time_ns = ( (cycles << shift) * mul ) >> 32;
+ * (although shift can be negative, so that's bad C)
+ *
+ * So for a single second,
+ * NSEC_PER_SEC = ( ( FREQ_HZ << shift) * mul ) >> 32
+ * NSEC_PER_SEC << 32 = ( FREQ_HZ << shift ) * mul
+ * ( NSEC_PER_SEC << 32 ) / mul = FREQ_HZ << shift
+ * ( NSEC_PER_SEC << 32 ) / mul ) >> shift = FREQ_HZ
+ */
+static uint64_t hvclock_to_hz(uint32_t mul, int8_t shift)
+{
+ uint64_t tm = NSEC_PER_SEC << 32;
+
+ /* Maximise precision. Shift right until the top bit is set */
+ tm <<= 2;
+ shift += 2;
+
+ /* While 'mul' is even, increase the shift *after* the division */
+ while (!(mul & 1)) {
+ shift++;
+ mul >>= 1;
+ }
+
+ tm /= mul;
+
+ if (shift > 0)
+ return tm >> shift;
+ else
+ return tm << -shift;
+}
+
+static int kvm_vcpu_ioctl_set_clock_guest(struct kvm_vcpu *v, void __user *argp)
+{
+ struct pvclock_vcpu_time_info user_hv_clock;
+ struct kvm *kvm = v->kvm;
+ struct kvm_arch *ka = &kvm->arch;
+ uint64_t curr_tsc_hz, user_tsc_hz;
+ uint64_t user_clk_ns;
+ uint64_t guest_tsc;
+ int rc = 0;
+
+ if (copy_from_user(&user_hv_clock, argp, sizeof(user_hv_clock)))
+ return -EFAULT;
+
+ if (!user_hv_clock.tsc_to_system_mul)
+ return -EINVAL;
+
+ user_tsc_hz = hvclock_to_hz(user_hv_clock.tsc_to_system_mul,
+ user_hv_clock.tsc_shift);
+
+
+ kvm_hv_request_tsc_page_update(kvm);
+ kvm_start_pvclock_update(kvm);
+ pvclock_update_vm_gtod_copy(kvm);
+
+ /*
+ * If not in use_master_clock mode, do not allow userspace to set
+ * the clock in terms of the guest TSC. Userspace should either
+ * fail the migration (to a host with suboptimal TSCs), or should
+ * knowingly restore the KVM clock using KVM_SET_CLOCK instead.
+ */
+ if (!ka->use_master_clock) {
+ rc = -EINVAL;
+ goto out;
+ }
+
+ curr_tsc_hz = get_cpu_tsc_khz() * 1000LL;
+ if (unlikely(curr_tsc_hz == 0)) {
+ rc = -EINVAL;
+ goto out;
+ }
+
+ if (kvm_caps.has_tsc_control)
+ curr_tsc_hz = kvm_scale_tsc(curr_tsc_hz,
+ v->arch.l1_tsc_scaling_ratio);
+
+ /*
+ * The scaling factors in the hv_clock do not depend solely on the
+ * TSC frequency *requested* by userspace. They actually use the
+ * host TSC frequency that was measured/detected by the host kernel,
+ * scaled by kvm_scale_tsc() with the vCPU's l1_tsc_scaling_ratio.
+ *
+ * So a sanity check that they *precisely* match would have false
+ * negatives. Allow for a discrepancy of 1 kHz either way.
+ */
+ if (user_tsc_hz < curr_tsc_hz - 1000 ||
+ user_tsc_hz > curr_tsc_hz + 1000) {
+ rc = -ERANGE;
+ goto out;
+ }
+
+ /*
+ * The call to pvclock_update_vm_gtod_copy() has created a new time
+ * reference point in ka->master_cycle_now and ka->master_kernel_ns.
+ *
+ * Calculate the guest TSC at that moment, and the corresponding KVM
+ * clock value according to user_hv_clock. The value according to the
+ * current hv_clock will of course be ka->master_kernel_ns since no
+ * TSC cycles have elapsed.
+ *
+ * Adjust ka->kvmclock_offset to the delta, so that both definitions
+ * of the clock give precisely the same reading at the reference time.
+ */
+ guest_tsc = kvm_read_l1_tsc(v, ka->master_cycle_now);
+ user_clk_ns = __pvclock_read_cycles(&user_hv_clock, guest_tsc);
+ ka->kvmclock_offset = user_clk_ns - ka->master_kernel_ns;
+
+out:
+ kvm_end_pvclock_update(kvm);
+ return rc;
+}
+#endif
+
long kvm_arch_vcpu_ioctl(struct file *filp,
unsigned int ioctl, unsigned long arg)
{
@@ -6256,6 +6404,14 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
srcu_read_unlock(&vcpu->kvm->srcu, idx);
break;
}
+#ifdef CONFIG_X86_64
+ case KVM_SET_CLOCK_GUEST:
+ r = kvm_vcpu_ioctl_set_clock_guest(vcpu, argp);
+ break;
+ case KVM_GET_CLOCK_GUEST:
+ r = kvm_vcpu_ioctl_get_clock_guest(vcpu, argp);
+ break;
+#endif
#ifdef CONFIG_KVM_HYPERV
case KVM_GET_SUPPORTED_HV_CPUID:
r = kvm_ioctl_get_supported_hv_cpuid(vcpu, argp);
diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
index 2190adbe3002..0d306311e4d6 100644
--- a/include/uapi/linux/kvm.h
+++ b/include/uapi/linux/kvm.h
@@ -1548,4 +1548,7 @@ struct kvm_create_guest_memfd {
__u64 reserved[6];
};
+#define KVM_SET_CLOCK_GUEST _IOW(KVMIO, 0xd5, struct pvclock_vcpu_time_info)
+#define KVM_GET_CLOCK_GUEST _IOR(KVMIO, 0xd6, struct pvclock_vcpu_time_info)
+
#endif /* __LINUX_KVM_H */
--
2.44.0
From: Jack Allister <[email protected]>
A VM's KVM/PV clock has an inherent relationship to its TSC (guest). When
either the host system live-updates or the VM is live-migrated this pairing
of the two clock sources should stay the same.
In reality this is not the case without some correction taking place. Two
new IOCTLs (KVM_GET_CLOCK_GUEST/KVM_SET_CLOCK_GUEST) can be utilized to
perform a correction on the PVTI (PV time information) structure held by
KVM to effectively fixup the kvmclock_offset prior to the guest VM resuming
in either a live-update/migration scenario.
This test proves that without the necessary fixup there is a perceived
change in the guest TSC & KVM/PV clock relationship before and after a LU/
LM takes place.
The following steps are made to verify there is a delta in the relationship
and that it can be corrected:
1. PVTI is sampled by guest at boot (let's call this PVTI0).
2. Induce a change in PVTI data (KVM_REQ_MASTERCLOCK_UPDATE).
3. PVTI is sampled by guest after change (PVTI1).
4. Correction is requested by usermode to KVM using PVTI0.
5. PVTI is sampled by guest after correction (PVTI2).
The guest the records a singular TSC reference point in time and uses it to
calculate 3 KVM clock values utilizing the 3 recorded PVTI prior. Let's
call each clock value CLK[0-2].
In a perfect world CLK[0-2] should all be the same value if the KVM clock
& TSC relationship is preserved across the LU/LM (or faked in this test),
however it is not.
A delta can be observed between CLK0-CLK1 due to KVM recalculating the PVTI
(and the inaccuracies associated with that). A delta of ~3500ns can be
observed if guest TSC scaling to half host TSC frequency is also enabled,
where as without scaling this is observed at ~180ns.
With the correction it should be possible to achieve a delta of ±1ns.
An option to enable guest TSC scaling is available via invoking the tester
with -s/--scale-tsc.
Example of the test output below:
* selftests: kvm: pvclock_test
* scaling tsc from 2999999KHz to 1499999KHz
* before=5038374946 uncorrected=5038371437 corrected=5038374945
* delta_uncorrected=3509 delta_corrected=1
Signed-off-by: Jack Allister <[email protected]>
Signed-off-by: David Woodhouse <[email protected]>
Reviewed-by: Paul Durrant <[email protected]>
CC: Dongli Zhang <[email protected]>
---
tools/testing/selftests/kvm/Makefile | 1 +
.../selftests/kvm/x86_64/pvclock_test.c | 192 ++++++++++++++++++
2 files changed, 193 insertions(+)
create mode 100644 tools/testing/selftests/kvm/x86_64/pvclock_test.c
diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile
index 871e2de3eb05..e33f56fedb0c 100644
--- a/tools/testing/selftests/kvm/Makefile
+++ b/tools/testing/selftests/kvm/Makefile
@@ -87,6 +87,7 @@ TEST_GEN_PROGS_x86_64 += x86_64/pmu_counters_test
TEST_GEN_PROGS_x86_64 += x86_64/pmu_event_filter_test
TEST_GEN_PROGS_x86_64 += x86_64/private_mem_conversions_test
TEST_GEN_PROGS_x86_64 += x86_64/private_mem_kvm_exits_test
+TEST_GEN_PROGS_x86_64 += x86_64/pvclock_test
TEST_GEN_PROGS_x86_64 += x86_64/set_boot_cpu_id
TEST_GEN_PROGS_x86_64 += x86_64/set_sregs_test
TEST_GEN_PROGS_x86_64 += x86_64/smaller_maxphyaddr_emulation_test
diff --git a/tools/testing/selftests/kvm/x86_64/pvclock_test.c b/tools/testing/selftests/kvm/x86_64/pvclock_test.c
new file mode 100644
index 000000000000..376ffb730a53
--- /dev/null
+++ b/tools/testing/selftests/kvm/x86_64/pvclock_test.c
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright © 2024, Amazon.com, Inc. or its affiliates.
+ *
+ * Tests for pvclock API
+ * KVM_SET_CLOCK_GUEST/KVM_GET_CLOCK_GUEST
+ */
+#include <asm/pvclock.h>
+#include <asm/pvclock-abi.h>
+#include <sys/stat.h>
+#include <stdint.h>
+#include <stdio.h>
+
+#include "test_util.h"
+#include "kvm_util.h"
+#include "processor.h"
+
+enum {
+ STAGE_FIRST_BOOT,
+ STAGE_UNCORRECTED,
+ STAGE_CORRECTED
+};
+
+#define KVMCLOCK_GPA 0xc0000000ull
+#define KVMCLOCK_SIZE sizeof(struct pvclock_vcpu_time_info)
+
+static void trigger_pvti_update(vm_paddr_t pvti_pa)
+{
+ /*
+ * We need a way to trigger KVM to update the fields
+ * in the PV time info. The easiest way to do this is
+ * to temporarily switch to the old KVM system time
+ * method and then switch back to the new one.
+ */
+ wrmsr(MSR_KVM_SYSTEM_TIME, pvti_pa | KVM_MSR_ENABLED);
+ wrmsr(MSR_KVM_SYSTEM_TIME_NEW, pvti_pa | KVM_MSR_ENABLED);
+}
+
+static void guest_code(vm_paddr_t pvti_pa)
+{
+ struct pvclock_vcpu_time_info *pvti_va =
+ (struct pvclock_vcpu_time_info *)pvti_pa;
+
+ struct pvclock_vcpu_time_info pvti_boot;
+ struct pvclock_vcpu_time_info pvti_uncorrected;
+ struct pvclock_vcpu_time_info pvti_corrected;
+ uint64_t cycles_boot;
+ uint64_t cycles_uncorrected;
+ uint64_t cycles_corrected;
+ uint64_t tsc_guest;
+
+ /*
+ * Setup the KVMCLOCK in the guest & store the original
+ * PV time structure that is used.
+ */
+ wrmsr(MSR_KVM_SYSTEM_TIME_NEW, pvti_pa | KVM_MSR_ENABLED);
+ pvti_boot = *pvti_va;
+ GUEST_SYNC(STAGE_FIRST_BOOT);
+
+ /*
+ * Trigger an update of the PVTI, if we calculate
+ * the KVM clock using this structure we'll see
+ * a delta from the TSC.
+ */
+ trigger_pvti_update(pvti_pa);
+ pvti_uncorrected = *pvti_va;
+ GUEST_SYNC(STAGE_UNCORRECTED);
+
+ /*
+ * The test should have triggered the correction by this
+ * point in time. We have a copy of each of the PVTI structs
+ * at each stage now.
+ *
+ * Let's sample the timestamp at a SINGLE point in time and
+ * then calculate what the KVM clock would be using the PVTI
+ * from each stage.
+ *
+ * Then return each of these values to the tester.
+ */
+ pvti_corrected = *pvti_va;
+ tsc_guest = rdtsc();
+
+ cycles_boot = __pvclock_read_cycles(&pvti_boot, tsc_guest);
+ cycles_uncorrected = __pvclock_read_cycles(&pvti_uncorrected, tsc_guest);
+ cycles_corrected = __pvclock_read_cycles(&pvti_corrected, tsc_guest);
+
+ GUEST_SYNC_ARGS(STAGE_CORRECTED, cycles_boot, cycles_uncorrected,
+ cycles_corrected, 0);
+}
+
+static void run_test(struct kvm_vm *vm, struct kvm_vcpu *vcpu)
+{
+ struct pvclock_vcpu_time_info pvti_before;
+ uint64_t before, uncorrected, corrected;
+ int64_t delta_uncorrected, delta_corrected;
+ struct ucall uc;
+ uint64_t ucall_reason;
+
+ /* Loop through each stage of the test. */
+ while (true) {
+
+ /* Start/restart the running vCPU code. */
+ vcpu_run(vcpu);
+ TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
+
+ /* Retrieve and verify our stage. */
+ ucall_reason = get_ucall(vcpu, &uc);
+ TEST_ASSERT(ucall_reason == UCALL_SYNC,
+ "Unhandled ucall reason=%lu",
+ ucall_reason);
+
+ /* Run host specific code relating to stage. */
+ switch (uc.args[1]) {
+ case STAGE_FIRST_BOOT:
+ /* Store the KVM clock values before an update. */
+ vcpu_ioctl(vcpu, KVM_GET_CLOCK_GUEST, &pvti_before);
+
+ /* Sleep for a set amount of time to increase delta. */
+ sleep(5);
+ break;
+
+ case STAGE_UNCORRECTED:
+ /* Restore the KVM clock values. */
+ vcpu_ioctl(vcpu, KVM_SET_CLOCK_GUEST, &pvti_before);
+ break;
+
+ case STAGE_CORRECTED:
+ /* Query the clock information and verify delta. */
+ before = uc.args[2];
+ uncorrected = uc.args[3];
+ corrected = uc.args[4];
+
+ delta_uncorrected = before - uncorrected;
+ delta_corrected = before - corrected;
+
+ pr_info("before=%lu uncorrected=%lu corrected=%lu\n",
+ before, uncorrected, corrected);
+
+ pr_info("delta_uncorrected=%ld delta_corrected=%ld\n",
+ delta_uncorrected, delta_corrected);
+
+ TEST_ASSERT((delta_corrected <= 1) && (delta_corrected >= -1),
+ "larger than expected delta detected = %ld", delta_corrected);
+ return;
+ }
+ }
+}
+
+static void configure_pvclock(struct kvm_vm *vm, struct kvm_vcpu *vcpu)
+{
+ unsigned int gpages;
+
+ gpages = vm_calc_num_guest_pages(VM_MODE_DEFAULT, KVMCLOCK_SIZE);
+ vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
+ KVMCLOCK_GPA, 1, gpages, 0);
+ virt_map(vm, KVMCLOCK_GPA, KVMCLOCK_GPA, gpages);
+
+ vcpu_args_set(vcpu, 1, KVMCLOCK_GPA);
+}
+
+static void configure_scaled_tsc(struct kvm_vcpu *vcpu)
+{
+ uint64_t tsc_khz;
+
+ tsc_khz = __vcpu_ioctl(vcpu, KVM_GET_TSC_KHZ, NULL);
+ pr_info("scaling tsc from %ldKHz to %ldKHz\n", tsc_khz, tsc_khz / 2);
+ tsc_khz /= 2;
+ vcpu_ioctl(vcpu, KVM_SET_TSC_KHZ, (void *)tsc_khz);
+}
+
+int main(int argc, char *argv[])
+{
+ struct kvm_vcpu *vcpu;
+ struct kvm_vm *vm;
+ bool scale_tsc;
+
+ scale_tsc = argc > 1 && (!strncmp(argv[1], "-s", 3) ||
+ !strncmp(argv[1], "--scale-tsc", 10));
+
+ TEST_REQUIRE(sys_clocksource_is_based_on_tsc());
+
+ vm = vm_create_with_one_vcpu(&vcpu, guest_code);
+
+ configure_pvclock(vm, vcpu);
+
+ if (scale_tsc)
+ configure_scaled_tsc(vcpu);
+
+ run_test(vm, vcpu);
+
+ return 0;
+}
--
2.44.0
On 4/27/2024 4:05 AM, David Woodhouse wrote:
> From: David Woodhouse <[email protected]>
>
> Both kvm_track_tsc_matching() and pvclock_update_vm_gtod_copy() make a
> decision about whether the KVM clock should be in master clock mode.
>
> They use *different* criteria for the decision though. This isn't really
> a problem; it only has the potential to cause unnecessary invocations of
> KVM_REQ_MASTERCLOCK_UPDATE if the masterclock was disabled due to TSC
> going backwards, or the guest using the old MSR. But it isn't pretty.
>
> Factor the decision out to a single function. And document the historical
> reason why it's disabled for guests that use the old MSR_KVM_SYSTEM_TIME.
>
> Signed-off-by: David Woodhouse <[email protected]>
> ---
> arch/x86/kvm/x86.c | 27 +++++++++++++++++++++++----
> 1 file changed, 23 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index d6e4469f531a..680b39f17851 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -2518,6 +2518,27 @@ static inline bool gtod_is_based_on_tsc(int mode)
> }
> #endif
>
> +static bool kvm_use_master_clock(strut kvm *kvm)
typo: 'strut' -> 'struct'
> +{
> + struct kvm_arch *ka = &kvm->arch;
> +
> + /*
> + * The 'old kvmclock' check is a workaround (from 2015) for a
> + * SUSE 2.6.16 kernel that didn't boot if the system_time in
> + * its kvmclock was too far behind the current time. So the
> + * mode of just setting the reference point and allowing time
> + * to proceed linearly from there makes it fail to boot.
> + * Despite that being kind of the *point* of the way the clock
> + * is exposed to the guest. By coincidence, the offending
> + * kernels used the old MSR_KVM_SYSTEM_TIME, which was moved
> + * only because it resided in the wrong number range. So the
> + * workaround is activated for *all* guests using the old MSR.
> + */
> + return ka->all_vcpus_matched_tsc &&
> + !ka->backwards_tsc_observed &&
> + !ka->boot_vcpu_runs_old_kvmclock;
> +}
> +
> static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
> {
> #ifdef CONFIG_X86_64
> @@ -2550,7 +2571,7 @@ static void kvm_track_tsc_matching(struct kvm_vcpu *vcpu)
> * To use the masterclock, the host clocksource must be based on TSC
> * and all vCPUs must have matching TSC frequencies.
> */
> - bool use_master_clock = ka->all_vcpus_matched_tsc &&
> + bool use_master_clock = kvm_use_master_clock(kvm) &&
'kvm' should be `vcpu->kvm`
> gtod_is_based_on_tsc(gtod->clock.vclock_mode);
On Wed, 2024-05-01 at 10:55 -0700, Chen, Zide wrote:
>
> > +static bool kvm_use_master_clock(strut kvm *kvm)
>
> typo: 'strut' -> 'struct'
Yeah, sorry about that. Apparently I forgot to *commit* the fixups for
that last-minute cleanup, before mailing the series out.
Fixed in
https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/clocks
along with eliminating the global clock updates again, but this time
only when !use_master_clock.
On Sat, Apr 27, 2024 at 12:04:58PM +0100, David Woodhouse wrote:
> From: David Woodhouse <[email protected]>
>
> The KVM clock is an interesting thing. It is defined as "nanoseconds
> since the guest was created", but in practice it runs at two *different*
> rates — or three different rates, if you count implementation bugs.
>
> Definition A is that it runs synchronously with the CLOCK_MONOTONIC_RAW
> of the host, with a delta of kvm->arch.kvmclock_offset.
>
> But that version doesn't actually get used in the common case, where the
> host has a reliable TSC and the guest TSCs are all running at the same
> rate and in sync with each other, and kvm->arch.use_master_clock is set.
>
> In that common case, definition B is used: There is a reference point in
> time at kvm->arch.master_kernel_ns (again a CLOCK_MONOTONIC_RAW time),
> and a corresponding host TSC value kvm->arch.master_cycle_now. This
> fixed point in time is converted to guest units (the time offset by
> kvmclock_offset and the TSC Value scaled and offset to be a guest TSC
> value) and advertised to the guest in the pvclock structure. While in
> this 'use_master_clock' mode, the fixed point in time never needs to be
> changed, and the clock runs precisely in time with the guest TSC, at the
> rate advertised in the pvclock structure.
>
> The third definition C is implemented in kvm_get_wall_clock_epoch() and
> __get_kvmclock(), using the master_cycle_now and master_kernel_ns fields
> but converting the *host* TSC cycles directly to a value in nanoseconds
> instead of scaling via the guest TSC.
>
> One might naïvely think that all three definitions are identical, since
> CLOCK_MONOTONIC_RAW is not skewed by NTP frequency corrections; all
> three are just the result of counting the host TSC at a known frequency,
> or the scaled guest TSC at a known precise fraction of the host's
> frequency. The problem is with arithmetic precision, and the way that
> frequency scaling is done in a division-free way by multiplying by a
> scale factor, then shifting right. In practice, all three ways of
> calculating the KVM clock will suffer a systemic drift from each other.
>
> Eventually, definition C should just be eliminated. Commit 451a707813ae
> ("KVM: x86/xen: improve accuracy of Xen timers") worked around it for
> the specific case of Xen timers, which are defined in terms of the KVM
> clock and suffered from a continually increasing error in timer expiry
> times. That commit notes that get_kvmclock_ns() is non-trivial to fix
> and says "I'll come back to that", which remains true.
>
> Definitions A and B do need to coexist, the former to handle the case
> where the host or guest TSC is suboptimally configured. But KVM should
> be more careful about switching between them, and the discontinuity in
> guest time which could result.
>
> In particular, KVM_REQ_MASTERCLOCK_UPDATE will take a new snapshot of
> time as the reference in master_kernel_ns and master_cycle_now, yanking
> the guest's clock back to match definition A at that moment.
KVM_REQ_MASTERCLOCK_UPDATE stops the vcpus because:
* To avoid that problem, do not allow visibility of distinct
* system_timestamp/tsc_timestamp values simultaneously: use a master
* copy of host monotonic time values. Update that master copy
* in lockstep.
> When invoked from in 'use_master_clock' mode, kvm_update_masterclock()
> should probably *adjust* kvm->arch.kvmclock_offset to account for the
> drift, instead of yanking the clock back to defintion A.
You are likely correct...
> But in the meantime there are a bunch of places where it just doesn't need to be
> invoked at all.
>
> To start with: there is no need to do such an update when a Xen guest
> populates the shared_info page. This seems to have been a hangover from
> the very first implementation of shared_info which automatically
> populated the vcpu_info structures at their default locations, but even
> then it should just have raised KVM_REQ_CLOCK_UPDATE on each vCPU
> instead of using KVM_REQ_MASTERCLOCK_UPDATE. And now that userspace is
> expected to explicitly set the vcpu_info even in its default locations,
> there's not even any need for that either.
>
> Fixes: 629b5348841a1 ("KVM: x86/xen: update wallclock region")
> Signed-off-by: David Woodhouse <[email protected]>
> Reviewed-by: Paul Durrant <[email protected]>
> ---
> arch/x86/kvm/xen.c | 2 --
> 1 file changed, 2 deletions(-)
>
> diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> index f65b35a05d91..5a83a8154b79 100644
> --- a/arch/x86/kvm/xen.c
> +++ b/arch/x86/kvm/xen.c
> @@ -98,8 +98,6 @@ static int kvm_xen_shared_info_init(struct kvm *kvm)
> wc->version = wc_version + 1;
> read_unlock_irq(&gpc->lock);
>
> - kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
> -
> out:
> srcu_read_unlock(&kvm->srcu, idx);
> return ret;
> --
> 2.44.0
So KVM_REQ_MASTERCLOCK_UPDATE is to avoid the race above.
In what contexes is kvm_xen_shared_info_init called from again?
Not clear to me KVM_REQ_MASTERCLOCK_UPDATE is not needed (or that is
needed, for that matter...).
On Sat, 2024-05-04 at 04:42 -0300, Marcelo Tosatti wrote:
> On Sat, Apr 27, 2024 at 12:04:58PM +0100, David Woodhouse wrote:
> >
> > In particular, KVM_REQ_MASTERCLOCK_UPDATE will take a new snapshot of
> > time as the reference in master_kernel_ns and master_cycle_now, yanking
> > the guest's clock back to match definition A at that moment.
>
> KVM_REQ_MASTERCLOCK_UPDATE stops the vcpus because:
Took me a while to work that one out, btw. The fact that the
KVM_REQ_MCLOCK_INPROGRESS request is asserted but never actually
*handled*, so all it does is repeatedly kick the vCPU out and make it
spin until the request is cleared is... interesting. Likewise the way
that we set KVM_REQ_MASTERCLOCK_UPDATE on *all* vCPUs, so they *all*
call kvm_update_masterclock(), when only one of them needed to. I may
clean that up a little.
> * To avoid that problem, do not allow visibility of distinct
> * system_timestamp/tsc_timestamp values simultaneously: use a master
> * copy of host monotonic time values. Update that master copy
> * in lockstep.
Right. That comment is a lot longer than the part you cited here, and
starts with 'assuming a stable TSC across pCPUS, and a stable TSC
across vCPUs'. It's the "if (ka->use_master_clock)" case.
And yes, what it's basically saying is a special case of the fact that
if you let the KVM clock run at its "natural" rate based on the guest
TSC (definition B), but each vCPU runs at that rate from a *different*
point on the line that is definition A (the host CLOCK_MONOTONIC_RAW),
bad things will happen.
I'm OK with it stopping the vCPUs (although I'd like it to do so in a
less implicitly awful way). But when we don't need to update the
reference time at all, let's not do so.
> > When invoked from in 'use_master_clock' mode, kvm_update_masterclock()
> > should probably *adjust* kvm->arch.kvmclock_offset to account for the
> > drift, instead of yanking the clock back to defintion A.
>
> You are likely correct...
>
> > But in the meantime there are a bunch of places where it just doesn't need to be
> > invoked at all.
> >
> > To start with: there is no need to do such an update when a Xen guest
> > populates the shared_info page. This seems to have been a hangover from
> > the very first implementation of shared_info which automatically
> > populated the vcpu_info structures at their default locations, but even
> > then it should just have raised KVM_REQ_CLOCK_UPDATE on each vCPU
> > instead of using KVM_REQ_MASTERCLOCK_UPDATE. And now that userspace is
> > expected to explicitly set the vcpu_info even in its default locations,
> > there's not even any need for that either.
> >
> > Fixes: 629b5348841a1 ("KVM: x86/xen: update wallclock region")
> > Signed-off-by: David Woodhouse <[email protected]>
> > Reviewed-by: Paul Durrant <[email protected]>
> > ---
> > arch/x86/kvm/xen.c | 2 --
> > 1 file changed, 2 deletions(-)
> >
> > diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> > index f65b35a05d91..5a83a8154b79 100644
> > --- a/arch/x86/kvm/xen.c
> > +++ b/arch/x86/kvm/xen.c
> > @@ -98,8 +98,6 @@ static int kvm_xen_shared_info_init(struct kvm *kvm)
> > wc->version = wc_version + 1;
> > read_unlock_irq(&gpc->lock);
> >
> > - kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
> > -
> > out:
> > srcu_read_unlock(&kvm->srcu, idx);
> > return ret;
> > --
> > 2.44.0
>
> So KVM_REQ_MASTERCLOCK_UPDATE is to avoid the race above.
>
> In what contexes is kvm_xen_shared_info_init called from again?
>
> Not clear to me KVM_REQ_MASTERCLOCK_UPDATE is not needed (or that is
> needed, for that matter...).
We cal kvm_xen_shared_info_init() when the Xen "shared info" page is
set up. The only interesting part of that is the *wallclock* epoch.
The wallclock (just like KSR_KVM_WALL_CLOCK{,_NEW}) is *entirely* hosed
ever since the KVM clock stopped being based on CLOCK_MONOTONIC, since
that means that the value of "wallclock time minus KVM clock time"
actually *changes* as the KVM clock runs at a different rate to
wallclock time.
I'm looking at a replacement for that which uses the gtod information
to give the guest a direct mapping of guest TSC to host CLOCK_TAI. And
in doing so we can *also* indicate when live migration has potentially
disrupted the guest TSC, so any further NTP/PTP refinement that the
guest may have done for itself needs to be thrown away.
On 4/27/2024 7:05 PM, David Woodhouse wrote:
> From: David Woodhouse <[email protected]>
>
> These pointlessly duplicate of the last_tsc_{nsec,offset,write} values.
>
> The only place they were used was where the TSC is stable and a new vCPU
> is being synchronized to the previous setting, in which case the 'last_'
> value is definitely identical.
>
> Signed-off-by: David Woodhouse <[email protected]>
> ---
> arch/x86/include/asm/kvm_host.h | 3 ---
> arch/x86/kvm/x86.c | 9 ++-------
> 2 files changed, 2 insertions(+), 10 deletions(-)
>
> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> index b01c1d000fff..7d06f389a607 100644
> --- a/arch/x86/include/asm/kvm_host.h
> +++ b/arch/x86/include/asm/kvm_host.h
> @@ -1354,9 +1354,6 @@ struct kvm_arch {
> u32 last_tsc_khz;
> u64 last_tsc_offset;
> u64 last_tsc_scaling_ratio;
> - u64 cur_tsc_nsec;
> - u64 cur_tsc_write;
> - u64 cur_tsc_offset;
> u64 cur_tsc_generation;
> int nr_vcpus_matched_tsc;
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 6ec43f39bdb0..92e81bfca25a 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -2737,9 +2737,6 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
> * These values are tracked in kvm->arch.cur_xxx variables.
> */
> kvm->arch.cur_tsc_generation++;
> - kvm->arch.cur_tsc_nsec = ns;
> - kvm->arch.cur_tsc_write = tsc;
> - kvm->arch.cur_tsc_offset = offset;
> kvm->arch.nr_vcpus_matched_tsc = 0;
> } else if (vcpu->arch.this_tsc_generation != kvm->arch.cur_tsc_generation) {
> kvm->arch.nr_vcpus_matched_tsc++;
> @@ -2747,8 +2744,6 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
>
> /* Keep track of which generation this VCPU has synchronized to */
> vcpu->arch.this_tsc_generation = kvm->arch.cur_tsc_generation;
> - vcpu->arch.this_tsc_nsec = kvm->arch.cur_tsc_nsec;
> - vcpu->arch.this_tsc_write = kvm->arch.cur_tsc_write;
Do we need to track vcpu->arch.this_tsc_nsec/this_tsc_write? At least
they are still used in compute_guest_tsc() to calculate the guest tsc.
>
> kvm_track_tsc_matching(vcpu);
> }
> @@ -2825,8 +2820,8 @@ static void kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 *user_value)
> data = kvm->arch.last_tsc_write;
>
> if (!kvm_check_tsc_unstable()) {
> - offset = kvm->arch.cur_tsc_offset;
> - ns = kvm->arch.cur_tsc_nsec;
> + offset = kvm->arch.last_tsc_offset;
> + ns = kvm->arch.last_tsc_nsec;
> } else {
> /*
> * ... unless the TSC is unstable and has to be
On Fri, 2024-05-10 at 17:03 +0800, Chenyi Qiang wrote:
>
> Do we need to track vcpu->arch.this_tsc_nsec/this_tsc_write? At least
> they are still used in compute_guest_tsc() to calculate the guest
> tsc.
Ah yes, that's true. Good catch; thanks.
That should be caught by a test case which runs the guest TSC at a
higher speed than the host, but *without* hardware TSC scaling.