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 as a multipler/offset 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 * Because all other devices are slaved off of the passage of time in the guest,
51 * it is (2) above that drives the firing of other timers in the guest,
52 * including timer devices such as the Programmable Interrupt Timer (PIT).
59 static int handle_cpufreq_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) {
60 struct vm_time * time_state = &(info->time_state);
62 info->vm_regs.rbx = time_state->guest_cpu_freq;
64 PrintDebug("Guest request cpu frequency: return %ld\n", (long)info->vm_regs.rbx);
71 int v3_start_time(struct guest_info * info) {
72 /* We start running with guest_time == host_time */
73 uint64_t t = v3_get_host_time(&info->time_state);
75 PrintDebug("Starting initial guest time as %llu\n", t);
76 info->time_state.last_update = t;
77 info->time_state.initial_time = t;
78 info->yield_start_cycle = t;
82 // If the guest is supposed to run slower than the host, yield out until
83 // the host time is appropriately far along;
84 int v3_adjust_time(struct guest_info * info) {
85 struct vm_time * time_state = &(info->time_state);
87 if (time_state->host_cpu_freq == time_state->guest_cpu_freq) {
88 time_state->guest_host_offset = 0;
90 uint64_t guest_time, guest_elapsed, desired_elapsed;
91 uint64_t host_time, target_host_time;
92 guest_time = v3_get_guest_time(time_state);
93 guest_elapsed = (guest_time - time_state->initial_time);
94 desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq;
96 target_host_time = time_state->initial_time + desired_elapsed;
97 host_time = v3_get_host_time(time_state);
98 PrintDebug("Core %d: Yielding %Lu cycles for guest frequency mismatch "
99 "(%Lu cycles elapsed in guest, %Lu in host).\n",
100 info->cpu_id, target_host_time - host_time,
101 guest_elapsed, host_time - time_state->initial_time);
103 host_time = v3_get_host_time(time_state);
104 while (host_time < target_host_time) {
106 host_time = v3_get_host_time(time_state);
109 PrintDebug("Core %d: done adjusting time at host time %Lu.\n",
110 info->cpu_id, host_time);
111 time_state->guest_host_offset = guest_time - host_time;
117 int v3_add_timer(struct guest_info * info, struct vm_timer_ops * ops,
118 void * private_data) {
119 struct vm_timer * timer = NULL;
120 timer = (struct vm_timer *)V3_Malloc(sizeof(struct vm_timer));
121 V3_ASSERT(timer != NULL);
124 timer->private_data = private_data;
126 list_add(&(timer->timer_link), &(info->time_state.timers));
127 info->time_state.num_timers++;
132 int v3_remove_timer(struct guest_info * info, struct vm_timer * timer) {
133 list_del(&(timer->timer_link));
134 info->time_state.num_timers--;
140 void v3_update_timers(struct guest_info * info) {
141 struct vm_timer * tmp_timer;
142 uint64_t old_time = info->time_state.last_update;
145 info->time_state.last_update = v3_get_guest_time(&info->time_state);
146 cycles = info->time_state.last_update - old_time;
148 list_for_each_entry(tmp_timer, &(info->time_state.timers), timer_link) {
149 tmp_timer->ops->update_timer(info, cycles, info->time_state.guest_cpu_freq, tmp_timer->private_data);
154 * Handle full virtualization of the time stamp counter. As noted
155 * above, we don't store the actual value of the TSC, only the guest's
156 * offset from the host TSC. If the guest write's the to TSC, we handle
157 * this by changing that offset.
160 int v3_rdtsc(struct guest_info * info) {
161 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
162 info->vm_regs.rdx = tscval >> 32;
163 info->vm_regs.rax = tscval & 0xffffffffLL;
167 int v3_handle_rdtsc(struct guest_info * info) {
170 info->vm_regs.rax &= 0x00000000ffffffffLL;
171 info->vm_regs.rdx &= 0x00000000ffffffffLL;
178 int v3_rdtscp(struct guest_info * info) {
180 /* First get the MSR value that we need. It's safe to futz with
181 * ra/c/dx here since they're modified by this instruction anyway. */
182 info->vm_regs.rcx = TSC_AUX_MSR;
183 ret = v3_handle_msr_read(info);
185 info->vm_regs.rcx = info->vm_regs.rax;
187 /* Now do the TSC half of the instruction, which may hit the normal
188 * TSC hook if it exists */
189 ret = v3_rdtsc(info);
196 int v3_handle_rdtscp(struct guest_info * info) {
200 info->vm_regs.rax &= 0x00000000ffffffffLL;
201 info->vm_regs.rcx &= 0x00000000ffffffffLL;
202 info->vm_regs.rdx &= 0x00000000ffffffffLL;
210 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
211 struct v3_msr *msr_val, void *priv) {
212 struct vm_time * time_state = &(info->time_state);
214 V3_ASSERT(msr_num == TSC_AUX_MSR);
215 msr_val->lo = time_state->tsc_aux.lo;
216 msr_val->hi = time_state->tsc_aux.hi;
221 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
222 struct v3_msr msr_val, void *priv) {
223 struct vm_time * time_state = &(info->time_state);
225 V3_ASSERT(msr_num == TSC_AUX_MSR);
226 time_state->tsc_aux.lo = msr_val.lo;
227 time_state->tsc_aux.hi = msr_val.hi;
232 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
233 struct v3_msr *msr_val, void *priv) {
234 uint64_t time = v3_get_guest_tsc(&info->time_state);
236 V3_ASSERT(msr_num == TSC_MSR);
237 msr_val->hi = time >> 32;
238 msr_val->lo = time & 0xffffffffLL;
243 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
244 struct v3_msr msr_val, void *priv) {
245 struct vm_time * time_state = &(info->time_state);
246 uint64_t guest_time, new_tsc;
247 V3_ASSERT(msr_num == TSC_MSR);
248 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
249 guest_time = v3_get_guest_time(time_state);
250 time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
257 static int init_vm_time(struct v3_vm_info *vm_info) {
261 PrintDebug("Installing TSC MSR hook.\n");
262 ret = v3_hook_msr(vm_info, TSC_MSR,
263 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
265 PrintDebug("Installing TSC_AUX MSR hook.\n");
267 ret = v3_hook_msr(vm_info, TSC_AUX_MSR, tsc_aux_msr_read_hook,
268 tsc_aux_msr_write_hook, NULL);
272 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
273 ret = v3_register_hypercall(vm_info, TIME_CPUFREQ_HCALL,
274 handle_cpufreq_hcall, NULL);
278 void v3_init_time(struct guest_info * info) {
279 struct vm_time * time_state = &(info->time_state);
280 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
281 static int one_time = 0;
284 time_state->host_cpu_freq = V3_CPU_KHZ();
285 khz = v3_cfg_val(cfg_tree, "khz");
287 time_state->guest_cpu_freq = atoi(khz);
288 PrintDebug("Core %d CPU frequency requested at %d khz.\n",
289 info->cpu_id, time_state->guest_cpu_freq);
292 if (!khz || time_state->guest_cpu_freq > time_state->host_cpu_freq) {
293 time_state->guest_cpu_freq = time_state->host_cpu_freq;
295 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);
297 time_state->initial_time = 0;
298 time_state->last_update = 0;
299 time_state->guest_host_offset = 0;
300 time_state->tsc_guest_offset = 0;
302 INIT_LIST_HEAD(&(time_state->timers));
303 time_state->num_timers = 0;
305 time_state->tsc_aux.lo = 0;
306 time_state->tsc_aux.hi = 0;
309 init_vm_time(info->vm_info);