* 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
#define PrintDebug(fmt, args...)
#endif
+/* Overview
+ *
+ * Time handling in VMMs is challenging, and Palacios uses the highest
+ * resolution, lowest overhead timer on modern CPUs that it can - the
+ * processor timestamp counter (TSC). Note that on somewhat old processors
+ * this can be problematic; in particular, older AMD processors did not
+ * have a constant rate timestamp counter in the face of power management
+ * events. However, the latest Intel and AMD CPUs all do (should...) have a
+ * constant rate TSC, and Palacios relies on this fact.
+ *
+ * Basically, Palacios keeps track of three quantities as it runs to manage
+ * 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 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);
-void v3_init_time(struct guest_info * info) {
- struct vm_time * time_state = &(info->time_state);
-
- time_state->cpu_freq = V3_CPU_KHZ();
-
- time_state->pause_time = 0;
- time_state->last_update = 0;
- time_state->host_offset = 0;
- time_state->offset_sum = 0;
-
- INIT_LIST_HEAD(&(time_state->timers));
- time_state->num_timers = 0;
-
- v3_register_hypercall(info->vm_info, TIME_CPUFREQ_HCALL, handle_cpufreq_hcall, NULL);
-}
-
int v3_start_time(struct guest_info * info) {
/* We start running with guest_time == host_time */
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.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);
+int v3_offset_time( struct guest_info * info, sint64_t offset )
+{
+ struct vm_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;
}
-/* Use a control-theoretic approach, specifically a PI control approach,
- * to adjust host_offset towards 0. Overall control documentation in
- * palacios/docs/time_control.tex Control parameters are P and I,
- * broken into rational numbers
- */
+// 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) {
+ struct vm_time * time_state = &(info->time_state);
+ uint64_t host_time, target_host_time;
+ uint64_t guest_time, target_guest_time, old_guest_time;
+ uint64_t guest_elapsed, host_elapsed, desired_elapsed;
+
+ /* Compute the target host time given how much time has *already*
+ * passed in the guest */
+ guest_time = v3_get_guest_time(time_state);
+ 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;
+
+ /* Now, let the host run while the guest is stopped to make the two
+ * sync up. */
+ host_time = v3_get_host_time(time_state);
+ old_guest_time = v3_get_guest_time(time_state);
+
+ while (target_host_time > host_time) {
+ v3_yield(info);
+ host_time = v3_get_host_time(time_state);
+ }
-/* These numbers need to be actually determined by pole placement work. They're
- * just blind guesses for now, which is a really bad idea. :) */
-#define P_NUM 1
-#define P_DENOM 2
-#define I_NUM 1
-#define I_DENOM 20
-
-void adjust_time_offset(struct guest_info * info) {
- /* Set point for control: Desired offset = 0;
- * Error = host_offset - 0 = host_offset */
-
- sint64_t adjust;
-
- /* Update the integral of the errror */
- info->time_state.offset_sum += info->time_state.host_offset;
- adjust = (P_NUM * info->time_state.host_offset) / P_DENOM +
- (I_NUM * info->time_state.offset_sum) / I_DENOM;
-
- /* We may want to constrain *adjust* because of
- * resolution/accuracy constraints. Explore that later. */
- info->time_state.host_offset -= adjust;
- return;
-}
-
-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.host_offset =
- (sint64_t)info->time_state.pause_time - (sint64_t)t;
-#ifdef CONFIG_TIME_TSC_OFFSET_ADJUST
- adjust_time_offset(info);
-#endif
- 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.host_offset);
+ guest_time = v3_get_guest_time(time_state);
+
+ // 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 - (sint64_t)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. */
+ host_elapsed = host_time - time_state->initial_time;
+ desired_elapsed = (host_elapsed * time_state->guest_cpu_freq) / time_state->host_cpu_freq;
+ target_guest_time = time_state->initial_time + desired_elapsed;
+
+ if (guest_time < target_guest_time) {
+ uint64_t max_skew, desired_skew, skew;
+
+ if (time_state->enter_time) {
+ max_skew = (time_state->exit_time - time_state->enter_time) / 10;
+ } else {
+ max_skew = 0;
+ }
+
+ desired_skew = target_guest_time - guest_time;
+ skew = desired_skew > max_skew ? max_skew : desired_skew;
+/* PrintDebug("Guest %llu cycles behind where it should be.\n",
+ desired_skew);
+ PrintDebug("Limit on forward skew is %llu. Skewing forward %llu.\n",
+ max_skew, skew); */
+
+ v3_offset_time(info, skew);
+ }
+
+ return 0;
+}
+
+/* Called immediately upon entry in the the VMM */
+int
+v3_time_exit_vm( struct guest_info * info )
+{
+ struct vm_time * time_state = &(info->time_state);
+
+ time_state->exit_time = v3_get_host_time(time_state);
return 0;
}
-int v3_add_timer(struct guest_info * info, struct vm_timer_ops * ops,
- void * private_data) {
- struct vm_timer * timer = NULL;
- timer = (struct vm_timer *)V3_Malloc(sizeof(struct vm_timer));
+/* Called immediately prior to entry to the VM */
+int
+v3_time_enter_vm( struct guest_info * info )
+{
+ struct vm_time * time_state = &(info->time_state);
+ uint64_t guest_time, host_time;
+
+ guest_time = v3_get_guest_time(time_state);
+ host_time = v3_get_host_time(time_state);
+ time_state->enter_time = host_time;
+ time_state->guest_host_offset = guest_time - host_time;
+
+ // Because we just modified the offset - shouldn't matter as this should be
+ // the last time-related call prior to entering the VMM, but worth it
+ // just in case.
+ time_state->exit_time = host_time;
+
+ return 0;
+}
+
+
+
+struct v3_timer * v3_add_timer(struct guest_info * info,
+ struct v3_timer_ops * ops,
+ void * private_data) {
+ struct v3_timer * timer = NULL;
+ timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
V3_ASSERT(timer != NULL);
timer->ops = ops;
list_add(&(timer->timer_link), &(info->time_state.timers));
info->time_state.num_timers++;
- return 0;
+ return timer;
}
-
-int v3_remove_timer(struct guest_info * info, struct vm_timer * timer) {
+int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
list_del(&(timer->timer_link));
info->time_state.num_timers--;
}
void v3_update_timers(struct guest_info * info) {
- struct vm_timer * tmp_timer;
+ struct vm_time *time_state = &info->time_state;
+ struct v3_timer * tmp_timer;
uint64_t old_time = info->time_state.last_update;
- uint64_t cycles;
+ sint64_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 = 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);
+ 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
+ * 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) {
+ uint64_t tscval = v3_get_guest_tsc(&info->time_state);
+
+ info->vm_regs.rdx = tscval >> 32;
+ info->vm_regs.rax = tscval & 0xffffffffLL;
+
+ return 0;
+}
+
+int v3_handle_rdtsc(struct guest_info * info) {
+ v3_rdtsc(info);
+
+ info->vm_regs.rax &= 0x00000000ffffffffLL;
+ info->vm_regs.rdx &= 0x00000000ffffffffLL;
+
+ info->rip += 2;
+
+ return 0;
+}
+
+int v3_rdtscp(struct guest_info * info) {
+ int ret;
+ /* First get the MSR value that we need. It's safe to futz with
+ * 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 != 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 != 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;
+
+ info->rip += 3;
+
+ return 0;
+}
+
+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);
+
+ V3_ASSERT(msr_num == TSC_AUX_MSR);
+
+ msr_val->lo = time_state->tsc_aux.lo;
+ msr_val->hi = time_state->tsc_aux.hi;
+
+ return 0;
+}
+
+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);
+
+ V3_ASSERT(msr_num == TSC_AUX_MSR);
+
+ time_state->tsc_aux.lo = msr_val.lo;
+ time_state->tsc_aux.hi = msr_val.hi;
+
+ return 0;
+}
+
+static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
+ struct v3_msr *msr_val, void *priv) {
+ 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;
+
+ return 0;
+}
+
+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);
+ 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;
+
+ return 0;
+}
+
+
+int v3_init_time_vm(struct v3_vm_info * vm) {
+ int ret;
+
+ PrintDebug("Installing TSC MSR hook.\n");
+ ret = v3_hook_msr(vm, TSC_MSR,
+ tsc_msr_read_hook, tsc_msr_write_hook, NULL);
+
+ if (ret != 0) {
+ return ret;
+ }
+
+ PrintDebug("Installing TSC_AUX MSR hook.\n");
+ ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
+ tsc_aux_msr_write_hook, NULL);
+
+ if (ret != 0) {
+ return ret;
+ }
+
+ PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
+ ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
+ handle_cpufreq_hcall, NULL);
+
+ return ret;
+}
+
+void v3_deinit_time_vm(struct v3_vm_info * vm) {
+ v3_unhook_msr(vm, TSC_MSR);
+ v3_unhook_msr(vm, TSC_AUX_MSR);
+
+ v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
+}
+
+void v3_init_time_core(struct guest_info * info) {
+ struct vm_time * time_state = &(info->time_state);
+ v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
+ char * khz = NULL;
+
+ 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 == 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,
+ time_state->guest_cpu_freq,
+ time_state->host_cpu_freq);
+
+ time_state->initial_time = 0;
+ time_state->last_update = 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;
+}
+
+
+void v3_deinit_time_core(struct guest_info * core) {
+ struct vm_time * time_state = &(core->time_state);
+ struct v3_timer * tmr = NULL;
+ struct v3_timer * tmp = NULL;
+
+ list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
+ v3_remove_timer(core, tmr);
+ }
+
+}