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->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.pause_time = t;
78 info->yield_start_cycle = t;
82 int v3_pause_time(struct guest_info * info) {
83 V3_ASSERT(info->time_state.pause_time == 0);
84 info->time_state.pause_time = v3_get_guest_time(&info->time_state);
85 PrintDebug("Time paused at guest time as %llu\n",
86 info->time_state.pause_time);
90 int v3_resume_time(struct guest_info * info) {
91 uint64_t t = v3_get_host_time(&info->time_state);
92 V3_ASSERT(info->time_state.pause_time != 0);
93 info->time_state.time_offset =
94 (sint64_t)info->time_state.pause_time - (sint64_t)t;
95 info->time_state.pause_time = 0;
96 PrintDebug("Time resumed paused at guest time as %llu "
97 "offset %lld from host time.\n", t, info->time_state.time_offset);
102 int v3_add_timer(struct guest_info * info, struct vm_timer_ops * ops,
103 void * private_data) {
104 struct vm_timer * timer = NULL;
105 timer = (struct vm_timer *)V3_Malloc(sizeof(struct vm_timer));
106 V3_ASSERT(timer != NULL);
109 timer->private_data = private_data;
111 list_add(&(timer->timer_link), &(info->time_state.timers));
112 info->time_state.num_timers++;
118 int v3_remove_timer(struct guest_info * info, struct vm_timer * timer) {
119 list_del(&(timer->timer_link));
120 info->time_state.num_timers--;
126 void v3_update_timers(struct guest_info * info) {
127 struct vm_timer * tmp_timer;
128 uint64_t old_time = info->time_state.last_update;
131 info->time_state.last_update = v3_get_guest_time(&info->time_state);
132 cycles = info->time_state.last_update - old_time;
134 list_for_each_entry(tmp_timer, &(info->time_state.timers), timer_link) {
135 tmp_timer->ops->update_timer(info, cycles, info->time_state.cpu_freq, tmp_timer->private_data);
141 * Handle full virtualization of the time stamp counter. As noted
142 * above, we don't store the actual value of the TSC, only the guest's
143 * offset from the host TSC. If the guest write's the to TSC, we handle
144 * this by changing that offset.
147 int v3_rdtsc(struct guest_info * info) {
148 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
149 info->vm_regs.rdx = tscval >> 32;
150 info->vm_regs.rax = tscval & 0xffffffffLL;
154 int v3_handle_rdtsc(struct guest_info * info) {
157 info->vm_regs.rax &= 0x00000000ffffffffLL;
158 info->vm_regs.rdx &= 0x00000000ffffffffLL;
165 int v3_rdtscp(struct guest_info * info) {
167 /* First get the MSR value that we need. It's safe to futz with
168 * ra/c/dx here since they're modified by this instruction anyway. */
169 info->vm_regs.rcx = TSC_AUX_MSR;
170 ret = v3_handle_msr_read(info);
172 info->vm_regs.rcx = info->vm_regs.rax;
174 /* Now do the TSC half of the instruction, which may hit the normal
175 * TSC hook if it exists */
176 ret = v3_rdtsc(info);
183 int v3_handle_rdtscp(struct guest_info * info) {
187 info->vm_regs.rax &= 0x00000000ffffffffLL;
188 info->vm_regs.rcx &= 0x00000000ffffffffLL;
189 info->vm_regs.rdx &= 0x00000000ffffffffLL;
196 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
197 struct v3_msr *msr_val, void *priv) {
198 struct vm_time * time_state = &(info->time_state);
200 V3_ASSERT(msr_num == TSC_AUX_MSR);
201 msr_val->lo = time_state->tsc_aux.lo;
202 msr_val->hi = time_state->tsc_aux.hi;
207 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
208 struct v3_msr msr_val, void *priv) {
209 struct vm_time * time_state = &(info->time_state);
211 V3_ASSERT(msr_num == TSC_AUX_MSR);
212 time_state->tsc_aux.lo = msr_val.lo;
213 time_state->tsc_aux.hi = msr_val.hi;
218 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
219 struct v3_msr *msr_val, void *priv) {
220 uint64_t time = v3_get_guest_tsc(&info->time_state);
222 V3_ASSERT(msr_num == TSC_MSR);
223 msr_val->hi = time >> 32;
224 msr_val->lo = time & 0xffffffffLL;
229 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
230 struct v3_msr msr_val, void *priv) {
231 struct vm_time * time_state = &(info->time_state);
232 uint64_t guest_time, new_tsc;
233 V3_ASSERT(msr_num == TSC_MSR);
234 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
235 guest_time = v3_get_guest_time(time_state);
236 time_state->tsc_time_offset = (sint64_t)new_tsc - (sint64_t)guest_time;
242 void v3_init_time(struct guest_info * info) {
243 struct vm_time * time_state = &(info->time_state);
245 time_state->cpu_freq = V3_CPU_KHZ();
247 time_state->pause_time = 0;
248 time_state->last_update = 0;
249 time_state->time_offset = 0;
250 time_state->time_div = 1;
251 time_state->time_mult = 1;
252 time_state->tsc_time_offset = 0;
254 INIT_LIST_HEAD(&(time_state->timers));
255 time_state->num_timers = 0;
257 time_state->tsc_aux.lo = 0;
258 time_state->tsc_aux.hi = 0;
260 /* does init_time get called once, or once *per core*??? */
261 v3_hook_msr(info->vm_info, TSC_MSR,
262 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
263 v3_hook_msr(info->vm_info, TSC_AUX_MSR, tsc_aux_msr_read_hook,
264 tsc_aux_msr_write_hook, NULL);
266 v3_register_hypercall(info->vm_info, TIME_CPUFREQ_HCALL, handle_cpufreq_hcall, NULL);