* redistribute, and modify it as specified in the file "V3VEE_LICENSE".
*/
-#include <palacios/vmm_time.h>
#include <palacios/vmm.h>
+#include <palacios/vmm_time.h>
#include <palacios/vm_guest.h>
-#ifndef CONFIG_DEBUG_TIME
+#ifndef V3_CONFIG_DEBUG_TIME
#undef PrintDebug
#define PrintDebug(fmt, args...)
#endif
* (1) Add support for temporarily skewing guest time off of where it should
* be to support slack simulation of guests. The idea is that simulators
* set this skew to be the difference between how much time passed for a
- * simulated feature and a real implementation of that feature, making
+ * simulated feature and a real implementation of that feature, making time
* pass at a different rate from real time on this core. The VMM will then
* attempt to move this skew back towards 0 subject to resolution/accuracy
* constraints from various system timers.
* The main effort in doing this will be to get accuracy/resolution
* information from each local timer and to use this to bound how much skew
* is removed on each exit.
+ *
+ * (2) Look more into sychronizing the offsets *across* virtual and physical
+ * cores so that multicore guests stay mostly in sync.
+ *
+ * (3) Look into using the AMD TSC multiplier feature and adding explicit time
+ * dilation support to time handling.
*/
static int handle_cpufreq_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) {
- struct vm_time * time_state = &(info->time_state);
+ struct vm_core_time * time_state = &(info->time_state);
info->vm_regs.rbx = time_state->guest_cpu_freq;
uint64_t t = v3_get_host_time(&info->time_state);
PrintDebug("Starting initial guest time as %llu\n", t);
+
+ info->time_state.enter_time = 0;
+ info->time_state.exit_time = t;
info->time_state.last_update = t;
info->time_state.initial_time = t;
info->yield_start_cycle = t;
+
return 0;
}
-// If the guest is supposed to run slower than the host, yield out until
-// the host time is appropriately far along;
-int v3_adjust_time(struct guest_info * info) {
- struct vm_time * time_state = &(info->time_state);
+int v3_pause_time( struct guest_info * info )
+{
+ struct vm_core_time * time_state = &(info->time_state);
+ if (time_state->pause_time != 0) {
+ PrintError("Attempted to pause time when time already paused.\n");
+ return -1;
+ }
+ time_state->pause_time = v3_get_host_time( time_state );
+
+ return 0;
+}
+
+int v3_resume_time( struct guest_info * info )
+{
+ struct vm_core_time * time_state = &(info->time_state);
+ uint64_t host_time, guest_time;
+ sint64_t offset;
+
+ if (time_state->pause_time == 0) {
+ PrintError("Attempted to resume time when time not paused.\n");
+ return -1;
+ }
+
+ host_time = v3_get_host_time(time_state);
+ guest_time = v3_compute_guest_time(time_state, host_time);
+ offset = (sint64_t)guest_time - (sint64_t)host_time;
+ time_state->guest_host_offset = offset;
+ time_state->pause_time = 0;
+
+ return 0;
+}
+
+int v3_offset_time( struct guest_info * info, sint64_t offset )
+{
+ struct vm_core_time * time_state = &(info->time_state);
+ PrintDebug("Adding additional offset of %lld to guest time.\n", offset);
+ time_state->guest_host_offset += offset;
+ return 0;
+}
+
+#ifdef V3_CONFIG_TIME_DILATION
+static uint64_t compute_target_host_time(struct guest_info * info, uint64_t guest_time)
+{
+ struct vm_core_time * time_state = &(info->time_state);
+ uint64_t guest_elapsed, desired_elapsed;
+
+ guest_elapsed = (guest_time - time_state->initial_time);
+ desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
+ return time_state->initial_time + desired_elapsed;
+}
+
+static uint64_t compute_target_guest_time(struct guest_info *info)
+{
+ struct vm_core_time * time_state = &(info->time_state);
+ uint64_t host_elapsed, desired_elapsed;
+
+ host_elapsed = v3_get_host_time(time_state) - time_state->initial_time;
+ desired_elapsed = (host_elapsed * time_state->guest_cpu_freq) / time_state->host_cpu_freq;
- if (time_state->host_cpu_freq == time_state->guest_cpu_freq) {
- time_state->guest_host_offset = 0;
- } else {
- uint64_t guest_time, guest_elapsed, desired_elapsed;
- uint64_t host_time, target_host_time;
+ return time_state->initial_time + desired_elapsed;
- guest_time = v3_get_guest_time(time_state);
+}
- /* Compute what host time this guest time should correspond to. */
- guest_elapsed = (guest_time - time_state->initial_time);
- desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
- target_host_time = time_state->initial_time + desired_elapsed;
+/* Yield time in the host to deal with a guest that wants to run slower than
+ * the native host cycle frequency */
+static int yield_host_time(struct guest_info * info) {
+ struct vm_core_time * time_state = &(info->time_state);
+ uint64_t host_time, target_host_time;
+ uint64_t guest_time, old_guest_time;
- /* Yield until that host time is reached */
+ /* Now, let the host run while the guest is stopped to make the two
+ * sync up. Note that this doesn't assume that guest time is stopped;
+ * the offsetting in the next step will change add an offset to guest
+ * time to account for the time paused even if the geust isn't
+ * usually paused in the VMM. */
+ host_time = v3_get_host_time(time_state);
+ old_guest_time = v3_compute_guest_time(time_state, host_time);
+ target_host_time = compute_target_host_time(info, old_guest_time);
+
+ while (target_host_time > host_time) {
+ v3_yield(info);
host_time = v3_get_host_time(time_state);
- while (host_time < target_host_time) {
- v3_yield(info);
- host_time = v3_get_host_time(time_state);
+ }
+
+ guest_time = v3_compute_guest_time(time_state, host_time);
+
+ /* We do *not* assume the guest timer was paused in the VM. If it was
+ * this offseting is 0. If it wasn't, we need this. */
+ v3_offset_time(info, (sint64_t)(old_guest_time - guest_time));
+
+ return 0;
+}
+
+static int skew_guest_time(struct guest_info * info) {
+ struct vm_core_time * time_state = &(info->time_state);
+ uint64_t target_guest_time, guest_time;
+ /* Now the host may have gotten ahead of the guest because
+ * yielding is a coarse grained thing. Figure out what guest time
+ * we want to be at, and use the use the offsetting mechanism in
+ * the VMM to make the guest run forward. We limit *how* much we skew
+ * it forward to prevent the guest time making large jumps,
+ * however. */
+ target_guest_time = compute_target_guest_time(info);
+ guest_time = v3_get_guest_time(time_state);
+
+ if (guest_time < target_guest_time) {
+ sint64_t max_skew, desired_skew, skew;
+
+ if (time_state->enter_time) {
+ /* Limit forward skew to 10% of the amount the guest has
+ * run since we last could skew time */
+ max_skew = (sint64_t)(guest_time - time_state->enter_time) / 10;
+ } else {
+ max_skew = 0;
}
- time_state->guest_host_offset = (sint64_t)guest_time - (sint64_t)host_time;
+ desired_skew = (sint64_t)(target_guest_time - guest_time);
+ skew = desired_skew > max_skew ? max_skew : desired_skew;
+ PrintDebug("Guest %lld cycles behind where it should be.\n",
+ desired_skew);
+ PrintDebug("Limit on forward skew is %lld. Skewing forward %lld.\n",
+ max_skew, skew);
+
+ v3_offset_time(info, skew);
}
+
return 0;
}
+#endif /* V3_CONFIG_TIME_DILATION */
+
+// Control guest time in relation to host time so that the two stay
+// appropriately synchronized to the extent possible.
+int v3_adjust_time(struct guest_info * info) {
+
+#ifdef V3_CONFIG_TIME_DILATION
+ /* First deal with yielding if we want to slow down the guest */
+ yield_host_time(info);
+ /* Now, if the guest is too slow, (either from excess yielding above,
+ * or because the VMM is doing something that takes a long time to emulate)
+ * allow guest time to jump forward a bit */
+ skew_guest_time(info);
+#endif
+ return 0;
+}
+
+/* Called immediately upon entry in the the VMM */
+int
+v3_time_exit_vm( struct guest_info * info, uint64_t * guest_cycles )
+{
+ struct vm_core_time * time_state = &(info->time_state);
+
+ time_state->exit_time = v3_get_host_time(time_state);
+
+#ifdef V3_CONFIG_TIME_DILATION
+ v3_pause_time( info );
+#endif
+
+ return 0;
+}
+
+/* Called immediately prior to entry to the VM */
+int
+v3_time_enter_vm( struct guest_info * info )
+{
+ struct vm_core_time * time_state = &(info->time_state);
+ uint64_t host_time;
+
+ host_time = v3_get_host_time(time_state);
+ time_state->enter_time = host_time;
+#ifdef V3_CONFIG_TIME_DILATION
+ v3_resume_time( info );
+#else
+ time_state->guest_host_offset = 0;
+#endif
+
+ return 0;
+}
+
+
+
struct v3_timer * v3_add_timer(struct guest_info * info,
struct v3_timer_ops * ops,
void * private_data) {
}
void v3_update_timers(struct guest_info * info) {
+ struct vm_core_time *time_state = &info->time_state;
struct v3_timer * tmp_timer;
+ sint64_t cycles;
uint64_t old_time = info->time_state.last_update;
- uint64_t cycles;
- info->time_state.last_update = v3_get_guest_time(&info->time_state);
- cycles = info->time_state.last_update - old_time;
+ time_state->last_update = v3_get_guest_time(time_state);
+ cycles = (sint64_t)(time_state->last_update - old_time);
+ V3_ASSERT(cycles >= 0);
- list_for_each_entry(tmp_timer, &(info->time_state.timers), timer_link) {
- tmp_timer->ops->update_timer(info, cycles, info->time_state.guest_cpu_freq, tmp_timer->private_data);
+ // V3_Print("Updating timers with %lld elapsed cycles.\n", cycles);
+ list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
+ tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
}
}
+
/*
* Handle full virtualization of the time stamp counter. As noted
* above, we don't store the actual value of the TSC, only the guest's
int v3_rdtsc(struct guest_info * info) {
uint64_t tscval = v3_get_guest_tsc(&info->time_state);
+
info->vm_regs.rdx = tscval >> 32;
info->vm_regs.rax = tscval & 0xffffffffLL;
+
return 0;
}
* ra/c/dx here since they're modified by this instruction anyway. */
info->vm_regs.rcx = TSC_AUX_MSR;
ret = v3_handle_msr_read(info);
- if (ret) return ret;
+
+ if (ret != 0) {
+ return ret;
+ }
+
info->vm_regs.rcx = info->vm_regs.rax;
/* Now do the TSC half of the instruction */
ret = v3_rdtsc(info);
- if (ret) return ret;
-
+
+ if (ret != 0) {
+ return ret;
+ }
+
return 0;
}
int v3_handle_rdtscp(struct guest_info * info) {
+ PrintDebug("Handling virtual RDTSCP call.\n");
v3_rdtscp(info);
-
+
info->vm_regs.rax &= 0x00000000ffffffffLL;
info->vm_regs.rcx &= 0x00000000ffffffffLL;
info->vm_regs.rdx &= 0x00000000ffffffffLL;
static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
struct v3_msr *msr_val, void *priv) {
- struct vm_time * time_state = &(info->time_state);
+ struct vm_core_time * time_state = &(info->time_state);
V3_ASSERT(msr_num == TSC_AUX_MSR);
+
msr_val->lo = time_state->tsc_aux.lo;
msr_val->hi = time_state->tsc_aux.hi;
static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
struct v3_msr msr_val, void *priv) {
- struct vm_time * time_state = &(info->time_state);
+ struct vm_core_time * time_state = &(info->time_state);
V3_ASSERT(msr_num == TSC_AUX_MSR);
+
time_state->tsc_aux.lo = msr_val.lo;
time_state->tsc_aux.hi = msr_val.hi;
uint64_t time = v3_get_guest_tsc(&info->time_state);
V3_ASSERT(msr_num == TSC_MSR);
+
msr_val->hi = time >> 32;
msr_val->lo = time & 0xffffffffLL;
static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
struct v3_msr msr_val, void *priv) {
- struct vm_time * time_state = &(info->time_state);
+ struct vm_core_time * time_state = &(info->time_state);
uint64_t guest_time, new_tsc;
+
V3_ASSERT(msr_num == TSC_MSR);
+
new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
guest_time = v3_get_guest_time(time_state);
time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
handle_cpufreq_hcall, NULL);
+ PrintDebug("Setting base time dilation factor.\n");
+ vm->time_state.td_mult = 1;
+
return ret;
}
v3_unhook_msr(vm, TSC_MSR);
v3_unhook_msr(vm, TSC_AUX_MSR);
- // v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
+ v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
}
void v3_init_time_core(struct guest_info * info) {
- struct vm_time * time_state = &(info->time_state);
+ struct vm_core_time * time_state = &(info->time_state);
v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
char * khz = NULL;
if (khz) {
time_state->guest_cpu_freq = atoi(khz);
- PrintDebug("Core %d CPU frequency requested at %d khz.\n",
- info->cpu_id, time_state->guest_cpu_freq);
- }
+ PrintDebug("Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n",
+ info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq);
+ }
- if ((khz == NULL) || (time_state->guest_cpu_freq > time_state->host_cpu_freq)) {
+ if ( (khz == NULL) ||
+ (time_state->guest_cpu_freq <= 0) ||
+ (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) {
+
time_state->guest_cpu_freq = time_state->host_cpu_freq;
}
- PrintDebug("Core %d CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
- info->cpu_id,
+ PrintDebug("Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
+ info->pcpu_id, info->vcpu_id,
time_state->guest_cpu_freq,
time_state->host_cpu_freq);
time_state->last_update = 0;
time_state->guest_host_offset = 0;
time_state->tsc_guest_offset = 0;
+ time_state->enter_time = 0;
+ time_state->exit_time = 0;
+ time_state->pause_time = 0;
+
INIT_LIST_HEAD(&(time_state->timers));
time_state->num_timers = 0;
time_state->tsc_aux.lo = 0;
time_state->tsc_aux.hi = 0;
-
-
}
void v3_deinit_time_core(struct guest_info * core) {
- struct vm_time * time_state = &(core->time_state);
+ struct vm_core_time * time_state = &(core->time_state);
struct v3_timer * tmr = NULL;
struct v3_timer * tmp = NULL;
}
}
-
-
-
-
-
-
