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 static uint64_t compute_target_host_time(struct guest_info * info)
109 struct vm_time * time_state = &(info->time_state);
110 uint64_t guest_elapsed, desired_elapsed;
112 guest_elapsed = (v3_get_guest_time(time_state) - time_state->initial_time);
113 desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
114 return time_state->initial_time + desired_elapsed;
117 static uint64_t compute_target_guest_time(struct guest_info *info)
119 struct vm_time * time_state = &(info->time_state);
120 uint64_t host_elapsed, desired_elapsed;
122 host_elapsed = v3_get_host_time(time_state) - time_state->initial_time;
123 desired_elapsed = (host_elapsed * time_state->guest_cpu_freq) / time_state->host_cpu_freq;
125 return time_state->initial_time + desired_elapsed;
129 /* Yield time in the host to deal with a guest that wants to run slower than
130 * the native host cycle frequency */
131 static int yield_host_time(struct guest_info * info) {
132 struct vm_time * time_state = &(info->time_state);
133 uint64_t host_time, target_host_time;
134 uint64_t guest_time, old_guest_time;
136 /* Compute the target host time given how much time has *already*
137 * passed in the guest */
138 target_host_time = compute_target_host_time(info);
140 /* Now, let the host run while the guest is stopped to make the two
141 * sync up. Note that this doesn't assume that guest time is stopped;
142 * the offsetting in the next step will change add an offset to guest
143 * time to account for the time paused even if the geust isn't
144 * usually paused in the VMM. */
145 host_time = v3_get_host_time(time_state);
146 old_guest_time = v3_get_guest_time(time_state);
148 while (target_host_time > host_time) {
150 host_time = v3_get_host_time(time_state);
153 guest_time = v3_get_guest_time(time_state);
155 /* We do *not* assume the guest timer was paused in the VM. If it was
156 * this offseting is 0. If it wasn't, we need this. */
157 v3_offset_time(info, (sint64_t)old_guest_time - (sint64_t)guest_time);
162 static int skew_guest_time(struct guest_info * info) {
163 struct vm_time * time_state = &(info->time_state);
164 uint64_t target_guest_time, guest_time;
165 /* Now the host may have gotten ahead of the guest because
166 * yielding is a coarse grained thing. Figure out what guest time
167 * we want to be at, and use the use the offsetting mechanism in
168 * the VMM to make the guest run forward. We limit *how* much we skew
169 * it forward to prevent the guest time making large jumps,
171 target_guest_time = compute_target_guest_time(info);
172 guest_time = v3_get_guest_time(time_state);
174 if (guest_time < target_guest_time) {
175 uint64_t max_skew, desired_skew, skew;
177 if (time_state->enter_time) {
178 /* Limit forward skew to 10% of the amount the guest has
179 * run since we last could skew time */
180 max_skew = (guest_time - time_state->enter_time) / 10;
185 desired_skew = target_guest_time - guest_time;
186 skew = desired_skew > max_skew ? max_skew : desired_skew;
187 PrintDebug("Guest %llu cycles behind where it should be.\n",
189 PrintDebug("Limit on forward skew is %llu. Skewing forward %llu.\n",
192 v3_offset_time(info, skew);
198 // Control guest time in relation to host time so that the two stay
199 // appropriately synchronized to the extent possible.
200 int v3_adjust_time(struct guest_info * info) {
202 /* First deal with yielding if we want to slow down the guest */
203 yield_host_time(info);
205 /* Now, if the guest is too slow, (either from excess yielding above,
206 * or because the VMM is doing something that takes a long time to emulate)
207 * allow guest time to jump forward a bit */
208 skew_guest_time(info);
213 /* Called immediately upon entry in the the VMM */
215 v3_time_exit_vm( struct guest_info * info )
217 struct vm_time * time_state = &(info->time_state);
219 time_state->exit_time = v3_get_host_time(time_state);
224 /* Called immediately prior to entry to the VM */
226 v3_time_enter_vm( struct guest_info * info )
228 struct vm_time * time_state = &(info->time_state);
229 uint64_t guest_time, host_time;
231 guest_time = v3_get_guest_time(time_state);
232 host_time = v3_get_host_time(time_state);
233 time_state->enter_time = host_time;
234 time_state->guest_host_offset = guest_time - host_time;
241 struct v3_timer * v3_add_timer(struct guest_info * info,
242 struct v3_timer_ops * ops,
243 void * private_data) {
244 struct v3_timer * timer = NULL;
245 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
246 V3_ASSERT(timer != NULL);
249 timer->private_data = private_data;
251 list_add(&(timer->timer_link), &(info->time_state.timers));
252 info->time_state.num_timers++;
257 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
258 list_del(&(timer->timer_link));
259 info->time_state.num_timers--;
265 void v3_update_timers(struct guest_info * info) {
266 struct vm_time *time_state = &info->time_state;
267 struct v3_timer * tmp_timer;
268 uint64_t old_time = info->time_state.last_update;
271 time_state->last_update = v3_get_guest_time(time_state);
272 cycles = time_state->last_update - old_time;
273 V3_ASSERT(cycles >= 0);
275 // V3_Print("Updating timers with %lld elapsed cycles.\n", cycles);
276 list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
277 tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
282 * Handle full virtualization of the time stamp counter. As noted
283 * above, we don't store the actual value of the TSC, only the guest's
284 * offset from monotonic guest's time. If the guest writes to the TSC, we
285 * handle this by changing that offset.
287 * Possible TODO: Proper hooking of TSC read/writes?
290 int v3_rdtsc(struct guest_info * info) {
291 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
293 info->vm_regs.rdx = tscval >> 32;
294 info->vm_regs.rax = tscval & 0xffffffffLL;
299 int v3_handle_rdtsc(struct guest_info * info) {
302 info->vm_regs.rax &= 0x00000000ffffffffLL;
303 info->vm_regs.rdx &= 0x00000000ffffffffLL;
310 int v3_rdtscp(struct guest_info * info) {
312 /* First get the MSR value that we need. It's safe to futz with
313 * ra/c/dx here since they're modified by this instruction anyway. */
314 info->vm_regs.rcx = TSC_AUX_MSR;
315 ret = v3_handle_msr_read(info);
321 info->vm_regs.rcx = info->vm_regs.rax;
323 /* Now do the TSC half of the instruction */
324 ret = v3_rdtsc(info);
334 int v3_handle_rdtscp(struct guest_info * info) {
335 PrintDebug("Handling virtual RDTSCP call.\n");
339 info->vm_regs.rax &= 0x00000000ffffffffLL;
340 info->vm_regs.rcx &= 0x00000000ffffffffLL;
341 info->vm_regs.rdx &= 0x00000000ffffffffLL;
348 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
349 struct v3_msr *msr_val, void *priv) {
350 struct vm_time * time_state = &(info->time_state);
352 V3_ASSERT(msr_num == TSC_AUX_MSR);
354 msr_val->lo = time_state->tsc_aux.lo;
355 msr_val->hi = time_state->tsc_aux.hi;
360 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
361 struct v3_msr msr_val, void *priv) {
362 struct vm_time * time_state = &(info->time_state);
364 V3_ASSERT(msr_num == TSC_AUX_MSR);
366 time_state->tsc_aux.lo = msr_val.lo;
367 time_state->tsc_aux.hi = msr_val.hi;
372 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
373 struct v3_msr *msr_val, void *priv) {
374 uint64_t time = v3_get_guest_tsc(&info->time_state);
376 V3_ASSERT(msr_num == TSC_MSR);
378 msr_val->hi = time >> 32;
379 msr_val->lo = time & 0xffffffffLL;
384 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
385 struct v3_msr msr_val, void *priv) {
386 struct vm_time * time_state = &(info->time_state);
387 uint64_t guest_time, new_tsc;
389 V3_ASSERT(msr_num == TSC_MSR);
391 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
392 guest_time = v3_get_guest_time(time_state);
393 time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
399 int v3_init_time_vm(struct v3_vm_info * vm) {
402 PrintDebug("Installing TSC MSR hook.\n");
403 ret = v3_hook_msr(vm, TSC_MSR,
404 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
410 PrintDebug("Installing TSC_AUX MSR hook.\n");
411 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
412 tsc_aux_msr_write_hook, NULL);
418 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
419 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
420 handle_cpufreq_hcall, NULL);
425 void v3_deinit_time_vm(struct v3_vm_info * vm) {
426 v3_unhook_msr(vm, TSC_MSR);
427 v3_unhook_msr(vm, TSC_AUX_MSR);
429 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
432 void v3_init_time_core(struct guest_info * info) {
433 struct vm_time * time_state = &(info->time_state);
434 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
437 time_state->host_cpu_freq = V3_CPU_KHZ();
438 khz = v3_cfg_val(cfg_tree, "khz");
441 time_state->guest_cpu_freq = atoi(khz);
442 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n",
443 info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq);
446 if ( (khz == NULL) ||
447 (time_state->guest_cpu_freq <= 0) ||
448 (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) {
450 time_state->guest_cpu_freq = time_state->host_cpu_freq;
453 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
454 info->pcpu_id, info->vcpu_id,
455 time_state->guest_cpu_freq,
456 time_state->host_cpu_freq);
458 time_state->initial_time = 0;
459 time_state->last_update = 0;
460 time_state->guest_host_offset = 0;
461 time_state->tsc_guest_offset = 0;
463 INIT_LIST_HEAD(&(time_state->timers));
464 time_state->num_timers = 0;
466 time_state->tsc_aux.lo = 0;
467 time_state->tsc_aux.hi = 0;
471 void v3_deinit_time_core(struct guest_info * core) {
472 struct vm_time * time_state = &(core->time_state);
473 struct v3_timer * tmr = NULL;
474 struct v3_timer * tmp = NULL;
476 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
477 v3_remove_timer(core, tmr);