2 * This file is part of the Palacios Virtual Machine Monitor developed
3 * by the V3VEE Project with funding from the United States National
4 * Science Foundation and the Department of Energy.
6 * The V3VEE Project is a joint project between Northwestern University
7 * and the University of New Mexico. You can find out more at
10 * Copyright (c) 2008, Jack Lange <jarusl@cs.northwestern.edu>
11 * Copyright (c) 2008, The V3VEE Project <http://www.v3vee.org>
12 * All rights reserved.
14 * Author: Jack Lange <jarusl@cs.northwestern.edu>
15 * Patrick G. Bridges <bridges@cs.unm.edu>
17 * This is free software. You are permitted to use,
18 * redistribute, and modify it as specified in the file "V3VEE_LICENSE".
21 #include <palacios/vmm_time.h>
22 #include <palacios/vmm.h>
23 #include <palacios/vm_guest.h>
25 #ifndef CONFIG_DEBUG_TIME
27 #define PrintDebug(fmt, args...)
32 * Time handling in VMMs is challenging, and Palacios uses the highest
33 * resolution, lowest overhead timer on modern CPUs that it can - the
34 * processor timestamp counter (TSC). Note that on somewhat old processors
35 * this can be problematic; in particular, older AMD processors did not
36 * have a constant rate timestamp counter in the face of power management
37 * events. However, the latest Intel and AMD CPUs all do (should...) have a
38 * constant rate TSC, and Palacios relies on this fact.
40 * Basically, Palacios keeps track of three quantities as it runs to manage
41 * the passage of time:
42 * (1) The host timestamp counter - read directly from HW and never written
43 * (2) A monotonic guest timestamp counter used to measure the progression of
44 * time in the guest. This is computed using an offsets from (1) above.
45 * (3) The actual guest timestamp counter (which can be written by
46 * writing to the guest TSC MSR - MSR 0x10) from the monotonic guest TSC.
47 * This is also computed as an offset from (2) above when the TSC and
48 * this offset is updated when the TSC MSR is written.
50 * The value used to offset the guest TSC from the host TSC is the *sum* of all
51 * of these offsets (2 and 3) above
53 * Because all other devices are slaved off of the passage of time in the guest,
54 * it is (2) above that drives the firing of other timers in the guest,
55 * including timer devices such as the Programmable Interrupt Timer (PIT).
58 * (1) Add support for temporarily skewing guest time off of where it should
59 * be to support slack simulation of guests. The idea is that simulators
60 * set this skew to be the difference between how much time passed for a
61 * simulated feature and a real implementation of that feature, making
62 * pass at a different rate from real time on this core. The VMM will then
63 * attempt to move this skew back towards 0 subject to resolution/accuracy
64 * constraints from various system timers.
66 * The main effort in doing this will be to get accuracy/resolution
67 * information from each local timer and to use this to bound how much skew
68 * is removed on each exit.
72 static int handle_cpufreq_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) {
73 struct vm_time * time_state = &(info->time_state);
75 info->vm_regs.rbx = time_state->guest_cpu_freq;
77 PrintDebug("Guest request cpu frequency: return %ld\n", (long)info->vm_regs.rbx);
84 int v3_start_time(struct guest_info * info) {
85 /* We start running with guest_time == host_time */
86 uint64_t t = v3_get_host_time(&info->time_state);
88 PrintDebug("Starting initial guest time as %llu\n", t);
89 #ifdef CONFIG_TIME_HIDE_VM_COST
90 info->time_state.pause_time = t;
92 info->time_state.pause_time = 0;
94 info->time_state.last_update = t;
95 info->time_state.initial_time = t;
96 info->yield_start_cycle = t;
100 // If the guest is supposed to run slower than the host, yield out until
101 // the host time is appropriately far along;
102 int v3_adjust_time(struct guest_info * info) {
103 struct vm_time * time_state = &(info->time_state);
105 if (time_state->host_cpu_freq != time_state->guest_cpu_freq) {
106 uint64_t guest_time, host_time, target_host_time;
107 sint64_t guest_elapsed, desired_elapsed;
109 guest_time = v3_get_guest_time(time_state);
111 /* Compute what host time this guest time should correspond to. */
112 guest_elapsed = (guest_time - time_state->initial_time);
113 desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
114 target_host_time = time_state->initial_time + desired_elapsed;
116 /* Yield until that host time is reached */
117 host_time = v3_get_host_time(time_state);
119 if (host_time < target_host_time) {
120 PrintDebug("Yielding until host time (%llu) greater than target (%llu).\n", host_time, target_host_time);
123 while (host_time < target_host_time) {
125 host_time = v3_get_host_time(time_state);
128 #ifndef CONFIG_TIME_HIDE_VM_COST
129 // XXX This should turn into a target offset we want to move towards XXX
130 time_state->guest_host_offset =
131 (sint64_t)guest_time - (sint64_t)host_time;
139 v3_pause_time( struct guest_info * info )
141 struct vm_time * time_state = &(info->time_state);
142 if (time_state->pause_time == 0) {
143 time_state->pause_time = v3_get_host_time(time_state);
144 // PrintDebug("Pausing at host time %llu.\n", time_state->pause_time);
146 PrintError("Palacios timekeeping paused when already paused.\n");
152 v3_restart_time( struct guest_info * info )
154 struct vm_time * time_state = &(info->time_state);
156 if (time_state->pause_time) {
157 sint64_t pause_diff = (v3_get_host_time(time_state) - time_state->pause_time);
158 time_state->guest_host_offset -= pause_diff;
159 time_state->pause_time = 0;
160 // PrintDebug("Resuming time after %lld cycles with offset %lld.\n", pause_diff, time_state->guest_host_offset);
162 PrintError( "Palacios time keeping restarted when not paused.");
168 int v3_offset_time( struct guest_info * info, sint64_t offset )
170 struct vm_time * time_state = &(info->time_state);
171 // PrintDebug("Adding additional offset of %lld to guest time.\n", offset);
172 time_state->guest_host_offset += offset;
176 struct v3_timer * v3_add_timer(struct guest_info * info,
177 struct v3_timer_ops * ops,
178 void * private_data) {
179 struct v3_timer * timer = NULL;
180 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
181 V3_ASSERT(timer != NULL);
184 timer->private_data = private_data;
186 list_add(&(timer->timer_link), &(info->time_state.timers));
187 info->time_state.num_timers++;
192 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
193 list_del(&(timer->timer_link));
194 info->time_state.num_timers--;
200 void v3_update_timers(struct guest_info * info) {
201 struct vm_time *time_state = &info->time_state;
202 struct v3_timer * tmp_timer;
203 uint64_t old_time = info->time_state.last_update;
206 time_state->last_update = v3_get_guest_time(time_state);
207 cycles = time_state->last_update - old_time;
209 // PrintDebug("Updating timer for %lld elapsed cycles (pt=%llu, offset=%lld).\n",
210 // cycles, time_state->pause_time, time_state->guest_host_offset);
212 list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
213 tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
218 * Handle full virtualization of the time stamp counter. As noted
219 * above, we don't store the actual value of the TSC, only the guest's
220 * offset from monotonic guest's time. If the guest writes to the TSC, we
221 * handle this by changing that offset.
223 * Possible TODO: Proper hooking of TSC read/writes?
226 int v3_rdtsc(struct guest_info * info) {
227 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
228 info->vm_regs.rdx = tscval >> 32;
229 info->vm_regs.rax = tscval & 0xffffffffLL;
233 int v3_handle_rdtsc(struct guest_info * info) {
236 info->vm_regs.rax &= 0x00000000ffffffffLL;
237 info->vm_regs.rdx &= 0x00000000ffffffffLL;
244 int v3_rdtscp(struct guest_info * info) {
246 /* First get the MSR value that we need. It's safe to futz with
247 * ra/c/dx here since they're modified by this instruction anyway. */
248 info->vm_regs.rcx = TSC_AUX_MSR;
249 ret = v3_handle_msr_read(info);
255 info->vm_regs.rcx = info->vm_regs.rax;
257 /* Now do the TSC half of the instruction */
258 ret = v3_rdtsc(info);
268 int v3_handle_rdtscp(struct guest_info * info) {
269 PrintDebug("Handling virtual RDTSCP call.\n");
273 info->vm_regs.rax &= 0x00000000ffffffffLL;
274 info->vm_regs.rcx &= 0x00000000ffffffffLL;
275 info->vm_regs.rdx &= 0x00000000ffffffffLL;
282 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
283 struct v3_msr *msr_val, void *priv) {
284 struct vm_time * time_state = &(info->time_state);
286 V3_ASSERT(msr_num == TSC_AUX_MSR);
288 msr_val->lo = time_state->tsc_aux.lo;
289 msr_val->hi = time_state->tsc_aux.hi;
294 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
295 struct v3_msr msr_val, void *priv) {
296 struct vm_time * time_state = &(info->time_state);
298 V3_ASSERT(msr_num == TSC_AUX_MSR);
300 time_state->tsc_aux.lo = msr_val.lo;
301 time_state->tsc_aux.hi = msr_val.hi;
306 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
307 struct v3_msr *msr_val, void *priv) {
308 uint64_t time = v3_get_guest_tsc(&info->time_state);
310 V3_ASSERT(msr_num == TSC_MSR);
312 msr_val->hi = time >> 32;
313 msr_val->lo = time & 0xffffffffLL;
318 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
319 struct v3_msr msr_val, void *priv) {
320 struct vm_time * time_state = &(info->time_state);
321 uint64_t guest_time, new_tsc;
323 V3_ASSERT(msr_num == TSC_MSR);
325 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
326 guest_time = v3_get_guest_time(time_state);
327 time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
333 int v3_init_time_vm(struct v3_vm_info * vm) {
336 PrintDebug("Installing TSC MSR hook.\n");
337 ret = v3_hook_msr(vm, TSC_MSR,
338 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
344 PrintDebug("Installing TSC_AUX MSR hook.\n");
345 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
346 tsc_aux_msr_write_hook, NULL);
352 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
353 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
354 handle_cpufreq_hcall, NULL);
359 void v3_deinit_time_vm(struct v3_vm_info * vm) {
360 v3_unhook_msr(vm, TSC_MSR);
361 v3_unhook_msr(vm, TSC_AUX_MSR);
363 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
366 void v3_init_time_core(struct guest_info * info) {
367 struct vm_time * time_state = &(info->time_state);
368 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
371 time_state->host_cpu_freq = V3_CPU_KHZ();
372 khz = v3_cfg_val(cfg_tree, "khz");
375 time_state->guest_cpu_freq = atoi(khz);
376 PrintDebug("Core %d CPU frequency requested at %d khz.\n",
377 info->cpu_id, time_state->guest_cpu_freq);
380 if ((khz == NULL) || (time_state->guest_cpu_freq <= 0)
381 || (time_state->guest_cpu_freq > time_state->host_cpu_freq)) {
382 time_state->guest_cpu_freq = time_state->host_cpu_freq;
385 PrintDebug("Core %d CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
387 time_state->guest_cpu_freq,
388 time_state->host_cpu_freq);
390 time_state->initial_time = 0;
391 time_state->last_update = 0;
392 time_state->guest_host_offset = 0;
393 time_state->tsc_guest_offset = 0;
395 INIT_LIST_HEAD(&(time_state->timers));
396 time_state->num_timers = 0;
398 time_state->tsc_aux.lo = 0;
399 time_state->tsc_aux.hi = 0;
405 void v3_deinit_time_core(struct guest_info * core) {
406 struct vm_time * time_state = &(core->time_state);
407 struct v3_timer * tmr = NULL;
408 struct v3_timer * tmp = NULL;
410 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
411 v3_remove_timer(core, tmr);