* the passage of time:
* (1) The host timestamp counter - read directly from HW and never written
* (2) A monotonic guest timestamp counter used to measure the progression of
- * time in the guest. This is computed as a multipler/offset from (1) above
+ * time in the guest. This is computed using an offsets from (1) above.
* (3) The actual guest timestamp counter (which can be written by
* writing to the guest TSC MSR - MSR 0x10) from the monotonic guest TSC.
* This is also computed as an offset from (2) above when the TSC and
* this offset is updated when the TSC MSR is written.
*
+ * The value used to offset the guest TSC from the host TSC is the *sum* of all
+ * of these offsets (2 and 3) above
+ *
* Because all other devices are slaved off of the passage of time in the guest,
* it is (2) above that drives the firing of other timers in the guest,
* including timer devices such as the Programmable Interrupt Timer (PIT).
*
- *
- *
+ * Future additions:
+ * (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
+ * 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.
*/
static int handle_cpufreq_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) {
struct vm_time * time_state = &(info->time_state);
- info->vm_regs.rbx = time_state->cpu_freq;
+ info->vm_regs.rbx = time_state->guest_cpu_freq;
PrintDebug("Guest request cpu frequency: return %ld\n", (long)info->vm_regs.rbx);
PrintDebug("Starting initial guest time as %llu\n", t);
info->time_state.last_update = t;
- info->time_state.pause_time = t;
+ info->time_state.initial_time = t;
info->yield_start_cycle = t;
return 0;
}
-int v3_pause_time(struct guest_info * info) {
- V3_ASSERT(info->time_state.pause_time == 0);
- info->time_state.pause_time = v3_get_guest_time(&info->time_state);
- PrintDebug("Time paused at guest time as %llu\n",
- info->time_state.pause_time);
- 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_resume_time(struct guest_info * info) {
- uint64_t t = v3_get_host_time(&info->time_state);
- V3_ASSERT(info->time_state.pause_time != 0);
- info->time_state.time_offset =
- (sint64_t)info->time_state.pause_time - (sint64_t)t;
- info->time_state.pause_time = 0;
- PrintDebug("Time resumed paused at guest time as %llu "
- "offset %lld from host time.\n", t, info->time_state.time_offset);
+ 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;
+ 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 until that host time is reached */
+ 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);
+ }
+
+ time_state->guest_host_offset = (sint64_t)guest_time - (sint64_t)host_time;
+ }
return 0;
}
return 0;
}
-
int v3_remove_timer(struct guest_info * info, struct vm_timer * timer) {
list_del(&(timer->timer_link));
info->time_state.num_timers--;
cycles = info->time_state.last_update - old_time;
list_for_each_entry(tmp_timer, &(info->time_state.timers), timer_link) {
- tmp_timer->ops->update_timer(info, cycles, info->time_state.cpu_freq, tmp_timer->private_data);
+ tmp_timer->ops->update_timer(info, cycles, info->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
- * offset from the host TSC. If the guest write's the to TSC, we handle
- * this by changing that offset.
+ * offset from monotonic guest's time. If the guest writes to the TSC, we
+ * handle this by changing that offset.
+ *
+ * Possible TODO: Proper hooking of TSC read/writes?
*/
int v3_rdtsc(struct guest_info * info) {
if (ret) return ret;
info->vm_regs.rcx = info->vm_regs.rax;
- /* Now do the TSC half of the instruction, which may hit the normal
- * TSC hook if it exists */
+ /* Now do the TSC half of the instruction */
ret = v3_rdtsc(info);
if (ret) return ret;
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_time_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
+ time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
return 0;
}
+static int init_vm_time(struct v3_vm_info *vm_info) {
+ int ret;
+
+ PrintDebug("Installing TSC MSR hook.\n");
+ ret = v3_hook_msr(vm_info, TSC_MSR,
+ tsc_msr_read_hook, tsc_msr_write_hook, NULL);
+
+ PrintDebug("Installing TSC_AUX MSR hook.\n");
+ if (ret) return ret;
+ ret = v3_hook_msr(vm_info, TSC_AUX_MSR, tsc_aux_msr_read_hook,
+ tsc_aux_msr_write_hook, NULL);
+ if (ret) return ret;
+
+ PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
+ ret = v3_register_hypercall(vm_info, TIME_CPUFREQ_HCALL,
+ handle_cpufreq_hcall, NULL);
+ return ret;
+}
+
void v3_init_time(struct guest_info * info) {
struct vm_time * time_state = &(info->time_state);
+ v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
+ static int one_time = 0;
+ char *khz;
+
+ time_state->host_cpu_freq = V3_CPU_KHZ();
+ khz = v3_cfg_val(cfg_tree, "khz");
+ 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);
+ }
+
+ if (!khz || 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, time_state->guest_cpu_freq, time_state->host_cpu_freq);
- time_state->cpu_freq = V3_CPU_KHZ();
-
- time_state->pause_time = 0;
+ time_state->initial_time = 0;
time_state->last_update = 0;
- time_state->time_offset = 0;
- time_state->time_div = 1;
- time_state->time_mult = 1;
- time_state->tsc_time_offset = 0;
+ time_state->guest_host_offset = 0;
+ time_state->tsc_guest_offset = 0;
INIT_LIST_HEAD(&(time_state->timers));
time_state->num_timers = 0;
time_state->tsc_aux.lo = 0;
time_state->tsc_aux.hi = 0;
- /* does init_time get called once, or once *per core*??? */
- v3_hook_msr(info->vm_info, TSC_MSR,
- tsc_msr_read_hook, tsc_msr_write_hook, NULL);
- v3_hook_msr(info->vm_info, TSC_AUX_MSR, tsc_aux_msr_read_hook,
- tsc_aux_msr_write_hook, NULL);
-
- v3_register_hypercall(info->vm_info, TIME_CPUFREQ_HCALL, handle_cpufreq_hcall, NULL);
+ if (!one_time) {
+ init_vm_time(info->vm_info);
+ one_time = 1;
+ }
}
+
+
+
+
+
+
+
+
+