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 info->time_state.last_update = t;
90 info->time_state.initial_time = t;
91 info->yield_start_cycle = t;
95 // If the guest is supposed to run slower than the host, yield out until
96 // the host time is appropriately far along;
97 int v3_adjust_time(struct guest_info * info) {
98 struct vm_time * time_state = &(info->time_state);
100 if (time_state->host_cpu_freq == time_state->guest_cpu_freq) {
101 time_state->guest_host_offset = 0;
103 uint64_t guest_time, guest_elapsed, desired_elapsed;
104 uint64_t host_time, target_host_time;
106 guest_time = v3_get_guest_time(time_state);
108 /* Compute what host time this guest time should correspond to. */
109 guest_elapsed = (guest_time - time_state->initial_time);
110 desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
111 target_host_time = time_state->initial_time + desired_elapsed;
113 /* Yield until that host time is reached */
114 host_time = v3_get_host_time(time_state);
116 while (host_time < target_host_time) {
118 host_time = v3_get_host_time(time_state);
121 time_state->guest_host_offset = (sint64_t)guest_time - (sint64_t)host_time;
127 struct v3_timer * v3_add_timer(struct guest_info * info,
128 struct v3_timer_ops * ops,
129 void * private_data) {
130 struct v3_timer * timer = NULL;
131 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
132 V3_ASSERT(timer != NULL);
135 timer->private_data = private_data;
137 list_add(&(timer->timer_link), &(info->time_state.timers));
138 info->time_state.num_timers++;
143 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
144 list_del(&(timer->timer_link));
145 info->time_state.num_timers--;
151 void v3_update_timers(struct guest_info * info) {
152 struct v3_timer * tmp_timer;
153 uint64_t old_time = info->time_state.last_update;
156 info->time_state.last_update = v3_get_guest_time(&info->time_state);
157 cycles = info->time_state.last_update - old_time;
159 list_for_each_entry(tmp_timer, &(info->time_state.timers), timer_link) {
160 tmp_timer->ops->update_timer(info, cycles, info->time_state.guest_cpu_freq, tmp_timer->private_data);
165 * Handle full virtualization of the time stamp counter. As noted
166 * above, we don't store the actual value of the TSC, only the guest's
167 * offset from monotonic guest's time. If the guest writes to the TSC, we
168 * handle this by changing that offset.
170 * Possible TODO: Proper hooking of TSC read/writes?
173 int v3_rdtsc(struct guest_info * info) {
174 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
175 info->vm_regs.rdx = tscval >> 32;
176 info->vm_regs.rax = tscval & 0xffffffffLL;
180 int v3_handle_rdtsc(struct guest_info * info) {
183 info->vm_regs.rax &= 0x00000000ffffffffLL;
184 info->vm_regs.rdx &= 0x00000000ffffffffLL;
191 int v3_rdtscp(struct guest_info * info) {
193 /* First get the MSR value that we need. It's safe to futz with
194 * ra/c/dx here since they're modified by this instruction anyway. */
195 info->vm_regs.rcx = TSC_AUX_MSR;
196 ret = v3_handle_msr_read(info);
202 info->vm_regs.rcx = info->vm_regs.rax;
204 /* Now do the TSC half of the instruction */
205 ret = v3_rdtsc(info);
215 int v3_handle_rdtscp(struct guest_info * info) {
219 info->vm_regs.rax &= 0x00000000ffffffffLL;
220 info->vm_regs.rcx &= 0x00000000ffffffffLL;
221 info->vm_regs.rdx &= 0x00000000ffffffffLL;
228 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
229 struct v3_msr *msr_val, void *priv) {
230 struct vm_time * time_state = &(info->time_state);
232 V3_ASSERT(msr_num == TSC_AUX_MSR);
234 msr_val->lo = time_state->tsc_aux.lo;
235 msr_val->hi = time_state->tsc_aux.hi;
240 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
241 struct v3_msr msr_val, void *priv) {
242 struct vm_time * time_state = &(info->time_state);
244 V3_ASSERT(msr_num == TSC_AUX_MSR);
246 time_state->tsc_aux.lo = msr_val.lo;
247 time_state->tsc_aux.hi = msr_val.hi;
252 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
253 struct v3_msr *msr_val, void *priv) {
254 uint64_t time = v3_get_guest_tsc(&info->time_state);
256 V3_ASSERT(msr_num == TSC_MSR);
258 msr_val->hi = time >> 32;
259 msr_val->lo = time & 0xffffffffLL;
264 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
265 struct v3_msr msr_val, void *priv) {
266 struct vm_time * time_state = &(info->time_state);
267 uint64_t guest_time, new_tsc;
269 V3_ASSERT(msr_num == TSC_MSR);
271 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
272 guest_time = v3_get_guest_time(time_state);
273 time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
279 int v3_init_time_vm(struct v3_vm_info * vm) {
282 PrintDebug("Installing TSC MSR hook.\n");
283 ret = v3_hook_msr(vm, TSC_MSR,
284 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
290 PrintDebug("Installing TSC_AUX MSR hook.\n");
291 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
292 tsc_aux_msr_write_hook, NULL);
298 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
299 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
300 handle_cpufreq_hcall, NULL);
305 void v3_deinit_time_vm(struct v3_vm_info * vm) {
306 v3_unhook_msr(vm, TSC_MSR);
307 v3_unhook_msr(vm, TSC_AUX_MSR);
309 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
312 void v3_init_time_core(struct guest_info * info) {
313 struct vm_time * time_state = &(info->time_state);
314 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
317 time_state->host_cpu_freq = V3_CPU_KHZ();
318 khz = v3_cfg_val(cfg_tree, "khz");
321 time_state->guest_cpu_freq = atoi(khz);
322 PrintDebug("Core %d CPU frequency requested at %d khz.\n",
323 info->cpu_id, time_state->guest_cpu_freq);
326 if ((khz == NULL) || (time_state->guest_cpu_freq > time_state->host_cpu_freq)) {
327 time_state->guest_cpu_freq = time_state->host_cpu_freq;
330 PrintDebug("Core %d CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
332 time_state->guest_cpu_freq,
333 time_state->host_cpu_freq);
335 time_state->initial_time = 0;
336 time_state->last_update = 0;
337 time_state->guest_host_offset = 0;
338 time_state->tsc_guest_offset = 0;
340 INIT_LIST_HEAD(&(time_state->timers));
341 time_state->num_timers = 0;
343 time_state->tsc_aux.lo = 0;
344 time_state->tsc_aux.hi = 0;
350 void v3_deinit_time_core(struct guest_info * core) {
351 struct vm_time * time_state = &(core->time_state);
352 struct v3_timer * tmr = NULL;
353 struct v3_timer * tmp = NULL;
355 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
356 v3_remove_timer(core, tmr);