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.h>
22 #include <palacios/vmm_time.h>
23 #include <palacios/vm_guest.h>
25 #ifndef V3_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);
90 info->time_state.enter_time = 0;
91 info->time_state.exit_time = t;
92 info->time_state.last_update = t;
93 info->time_state.initial_time = t;
94 info->yield_start_cycle = t;
99 int v3_offset_time( struct guest_info * info, sint64_t offset )
101 struct vm_time * time_state = &(info->time_state);
102 // PrintDebug("Adding additional offset of %lld to guest time.\n", offset);
103 time_state->guest_host_offset += offset;
107 // Control guest time in relation to host time so that the two stay
108 // appropriately synchronized to the extent possible.
109 int v3_adjust_time(struct guest_info * info) {
110 struct vm_time * time_state = &(info->time_state);
111 uint64_t host_time, target_host_time;
112 uint64_t guest_time, target_guest_time, old_guest_time;
113 uint64_t guest_elapsed, host_elapsed, desired_elapsed;
115 /* Compute the target host time given how much time has *already*
116 * passed in the guest */
117 guest_time = v3_get_guest_time(time_state);
118 guest_elapsed = (guest_time - time_state->initial_time);
119 desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
120 target_host_time = time_state->initial_time + desired_elapsed;
122 /* Now, let the host run while the guest is stopped to make the two
124 host_time = v3_get_host_time(time_state);
125 old_guest_time = v3_get_guest_time(time_state);
127 while (target_host_time > host_time) {
129 host_time = v3_get_host_time(time_state);
132 guest_time = v3_get_guest_time(time_state);
134 // We do *not* assume the guest timer was paused in the VM. If it was
135 // this offseting is 0. If it wasn't we need this.
136 v3_offset_time(info, (sint64_t)old_guest_time - (sint64_t)guest_time);
138 /* Now the host may have gotten ahead of the guest because
139 * yielding is a coarse grained thing. Figure out what guest time
140 * we want to be at, and use the use the offsetting mechanism in
141 * the VMM to make the guest run forward. We limit *how* much we skew
142 * it forward to prevent the guest time making large jumps,
144 host_elapsed = host_time - time_state->initial_time;
145 desired_elapsed = (host_elapsed * time_state->guest_cpu_freq) / time_state->host_cpu_freq;
146 target_guest_time = time_state->initial_time + desired_elapsed;
148 if (guest_time < target_guest_time) {
149 uint64_t max_skew, desired_skew, skew;
151 if (time_state->enter_time) {
152 max_skew = (time_state->exit_time - time_state->enter_time) / 10;
157 desired_skew = target_guest_time - guest_time;
158 skew = desired_skew > max_skew ? max_skew : desired_skew;
159 /* PrintDebug("Guest %llu cycles behind where it should be.\n",
161 PrintDebug("Limit on forward skew is %llu. Skewing forward %llu.\n",
164 v3_offset_time(info, skew);
170 /* Called immediately upon entry in the the VMM */
172 v3_time_exit_vm( struct guest_info * info )
174 struct vm_time * time_state = &(info->time_state);
176 time_state->exit_time = v3_get_host_time(time_state);
181 /* Called immediately prior to entry to the VM */
183 v3_time_enter_vm( struct guest_info * info )
185 struct vm_time * time_state = &(info->time_state);
186 uint64_t guest_time, host_time;
188 guest_time = v3_get_guest_time(time_state);
189 host_time = v3_get_host_time(time_state);
190 time_state->enter_time = host_time;
191 time_state->guest_host_offset = guest_time - host_time;
193 // Because we just modified the offset - shouldn't matter as this should be
194 // the last time-related call prior to entering the VMM, but worth it
196 time_state->exit_time = host_time;
203 struct v3_timer * v3_add_timer(struct guest_info * info,
204 struct v3_timer_ops * ops,
205 void * private_data) {
206 struct v3_timer * timer = NULL;
207 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
208 V3_ASSERT(timer != NULL);
211 timer->private_data = private_data;
213 list_add(&(timer->timer_link), &(info->time_state.timers));
214 info->time_state.num_timers++;
219 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
220 list_del(&(timer->timer_link));
221 info->time_state.num_timers--;
227 void v3_update_timers(struct guest_info * info) {
228 struct vm_time *time_state = &info->time_state;
229 struct v3_timer * tmp_timer;
230 uint64_t old_time = info->time_state.last_update;
233 time_state->last_update = v3_get_guest_time(time_state);
234 cycles = time_state->last_update - old_time;
236 list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
237 tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
242 * Handle full virtualization of the time stamp counter. As noted
243 * above, we don't store the actual value of the TSC, only the guest's
244 * offset from monotonic guest's time. If the guest writes to the TSC, we
245 * handle this by changing that offset.
247 * Possible TODO: Proper hooking of TSC read/writes?
250 int v3_rdtsc(struct guest_info * info) {
251 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
253 info->vm_regs.rdx = tscval >> 32;
254 info->vm_regs.rax = tscval & 0xffffffffLL;
259 int v3_handle_rdtsc(struct guest_info * info) {
262 info->vm_regs.rax &= 0x00000000ffffffffLL;
263 info->vm_regs.rdx &= 0x00000000ffffffffLL;
270 int v3_rdtscp(struct guest_info * info) {
272 /* First get the MSR value that we need. It's safe to futz with
273 * ra/c/dx here since they're modified by this instruction anyway. */
274 info->vm_regs.rcx = TSC_AUX_MSR;
275 ret = v3_handle_msr_read(info);
281 info->vm_regs.rcx = info->vm_regs.rax;
283 /* Now do the TSC half of the instruction */
284 ret = v3_rdtsc(info);
294 int v3_handle_rdtscp(struct guest_info * info) {
295 PrintDebug("Handling virtual RDTSCP call.\n");
299 info->vm_regs.rax &= 0x00000000ffffffffLL;
300 info->vm_regs.rcx &= 0x00000000ffffffffLL;
301 info->vm_regs.rdx &= 0x00000000ffffffffLL;
308 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
309 struct v3_msr *msr_val, void *priv) {
310 struct vm_time * time_state = &(info->time_state);
312 V3_ASSERT(msr_num == TSC_AUX_MSR);
314 msr_val->lo = time_state->tsc_aux.lo;
315 msr_val->hi = time_state->tsc_aux.hi;
320 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
321 struct v3_msr msr_val, void *priv) {
322 struct vm_time * time_state = &(info->time_state);
324 V3_ASSERT(msr_num == TSC_AUX_MSR);
326 time_state->tsc_aux.lo = msr_val.lo;
327 time_state->tsc_aux.hi = msr_val.hi;
332 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
333 struct v3_msr *msr_val, void *priv) {
334 uint64_t time = v3_get_guest_tsc(&info->time_state);
336 V3_ASSERT(msr_num == TSC_MSR);
338 msr_val->hi = time >> 32;
339 msr_val->lo = time & 0xffffffffLL;
344 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
345 struct v3_msr msr_val, void *priv) {
346 struct vm_time * time_state = &(info->time_state);
347 uint64_t guest_time, new_tsc;
349 V3_ASSERT(msr_num == TSC_MSR);
351 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
352 guest_time = v3_get_guest_time(time_state);
353 time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
359 int v3_init_time_vm(struct v3_vm_info * vm) {
362 PrintDebug("Installing TSC MSR hook.\n");
363 ret = v3_hook_msr(vm, TSC_MSR,
364 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
370 PrintDebug("Installing TSC_AUX MSR hook.\n");
371 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
372 tsc_aux_msr_write_hook, NULL);
378 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
379 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
380 handle_cpufreq_hcall, NULL);
385 void v3_deinit_time_vm(struct v3_vm_info * vm) {
386 v3_unhook_msr(vm, TSC_MSR);
387 v3_unhook_msr(vm, TSC_AUX_MSR);
389 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
392 void v3_init_time_core(struct guest_info * info) {
393 struct vm_time * time_state = &(info->time_state);
394 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
397 time_state->host_cpu_freq = V3_CPU_KHZ();
398 khz = v3_cfg_val(cfg_tree, "khz");
401 time_state->guest_cpu_freq = atoi(khz);
402 PrintDebug("Core %d CPU frequency requested at %d khz.\n",
403 info->cpu_id, time_state->guest_cpu_freq);
406 if ( (khz == NULL) ||
407 (time_state->guest_cpu_freq <= 0) ||
408 (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) {
410 time_state->guest_cpu_freq = time_state->host_cpu_freq;
413 PrintDebug("Core %d CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
415 time_state->guest_cpu_freq,
416 time_state->host_cpu_freq);
418 time_state->initial_time = 0;
419 time_state->last_update = 0;
420 time_state->guest_host_offset = 0;
421 time_state->tsc_guest_offset = 0;
423 INIT_LIST_HEAD(&(time_state->timers));
424 time_state->num_timers = 0;
426 time_state->tsc_aux.lo = 0;
427 time_state->tsc_aux.hi = 0;
431 void v3_deinit_time_core(struct guest_info * core) {
432 struct vm_time * time_state = &(core->time_state);
433 struct v3_timer * tmr = NULL;
434 struct v3_timer * tmp = NULL;
436 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
437 v3_remove_timer(core, tmr);