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 stored as an absolute number of cycles elapsed
45 * and is updated on guest entry and exit; it can also be updated explicitly
46 * in the monitor at times
47 * (3) The actual guest timestamp counter (which can be written by
48 * writing to the guest TSC MSR - MSR 0x10) from the monotonic guest TSC.
49 * This is also computed as an offset from (2) above when the TSC and
50 * this offset is updated when the TSC MSR is written.
52 * Because all other devices are slaved off of the passage of time in the guest,
53 * it is (2) above that drives the firing of other timers in the guest,
54 * including timer devices such as the Programmable Interrupt Timer (PIT).
57 * (1) Add support for temporarily skewing guest time off of where it should
58 * be to support slack simulation of guests. The idea is that simulators
59 * set this skew to be the difference between how much time passed for a
60 * simulated feature and a real implementation of that feature, making time
61 * pass at a different rate from real time on this core. The VMM will then
62 * attempt to move this skew back towards 0 subject to resolution/accuracy
63 * constraints from various system timers.
65 * The main effort in doing this will be to get accuracy/resolution
66 * information from each local timer and to use this to bound how much skew
67 * is removed on each exit.
69 * (2) Look more into sychronizing the offsets *across* virtual and physical
70 * cores so that multicore guests stay mostly in sync.
72 * (3) Look into using the AMD TSC multiplier feature and adding explicit time
73 * dilation support to time handling.
77 static int handle_cpufreq_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) {
78 struct vm_core_time * time_state = &(info->time_state);
80 info->vm_regs.rbx = time_state->guest_cpu_freq;
82 PrintDebug("Guest request cpu frequency: return %ld\n", (long)info->vm_regs.rbx);
89 int v3_start_time(struct guest_info * info) {
90 /* We start running with guest_time == host_time */
91 uint64_t t = v3_get_host_time(&info->time_state);
93 info->time_state.vm_enter_host_time = 0;
94 info->time_state.vm_pause_host_time = t;
95 info->time_state.initial_host_time = t;
96 info->yield_start_cycle = t;
98 info->time_state.last_update = 0;
99 info->time_state.guest_cycles = 0;
100 PrintDebug("Starting time for core %d at host time %llu/guest time %llu.\n",
101 info->vcpu_id, t, info->time_state.guest_cycles);
106 int v3_offset_time( struct guest_info * info, sint64_t offset )
108 struct vm_core_time * time_state = &(info->time_state);
109 if (info->vm_info->time_state.follow_host_time) {
110 PrintError("Cannot offset guest time passage while slaved to host clock.\n");
113 time_state->guest_cycles += offset;
118 int v3_skip_time(struct guest_info * info) {
119 if (info->vm_info->time_state.follow_host_time) {
120 PrintError("Cannot skip guest time passage while slaved to host clock.\n");
123 info->time_state.vm_pause_host_time = v3_get_host_time(&info->time_state);
128 static sint64_t host_to_guest_cycles(struct guest_info * info, sint64_t host_cycles) {
129 return (host_cycles * info->time_state.clock_ratio_num) / info->time_state.clock_ratio_denom;
132 int v3_time_advance_cycles(struct guest_info * info, uint64_t *host_cycles)
134 uint64_t t = v3_get_host_time(&info->time_state);
136 info->time_state.vm_pause_host_time = t;
138 if (info->vm_info->time_state.follow_host_time) {
139 /* How many guest cycles should have elapsed? */
140 sint64_t host_elapsed = t - info->time_state.initial_host_time;
141 sint64_t guest_elapsed = host_to_guest_cycles(info, host_elapsed);
143 info->time_state.guest_cycles = guest_elapsed;
145 uint64_t guest_cycles;
147 guest_cycles = host_to_guest_cycles(info, *host_cycles);
149 guest_cycles = host_to_guest_cycles(info, (sint64_t)(t - info->time_state.vm_pause_host_time));
151 info->time_state.guest_cycles += guest_cycles;
157 int v3_advance_time(struct guest_info * info) {
158 return v3_time_advance_cycles(info, NULL);
161 /* Called immediately upon entry in the the VMM */
163 v3_time_exit_vm( struct guest_info * info, uint64_t * host_cycles )
165 return v3_time_advance_cycles(info, host_cycles);
168 /* Called immediately prior to entry to the VM */
170 v3_time_enter_vm( struct guest_info * info )
172 struct vm_core_time * time_state = &(info->time_state);
173 uint64_t host_time = v3_get_host_time(&info->time_state);
175 time_state->vm_enter_host_time = host_time;
181 struct v3_timer * v3_add_timer(struct guest_info * info,
182 struct v3_timer_ops * ops,
183 void * private_data) {
184 struct v3_timer * timer = NULL;
185 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
186 V3_ASSERT(timer != NULL);
189 timer->private_data = private_data;
191 list_add(&(timer->timer_link), &(info->time_state.timers));
192 info->time_state.num_timers++;
197 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
198 list_del(&(timer->timer_link));
199 info->time_state.num_timers--;
205 void v3_update_timers(struct guest_info * info) {
206 struct vm_core_time *time_state = &info->time_state;
207 struct v3_timer * tmp_timer;
209 uint64_t old_time = time_state->last_update;
211 time_state->last_update = v3_get_guest_time(time_state);
212 cycles = (sint64_t)(time_state->last_update - old_time);
214 PrintError("Cycles appears to have rolled over - old time %lld, current time %lld.\n",
215 old_time, time_state->last_update);
219 PrintDebug("Updating timers with %lld elapsed cycles.\n", cycles);
220 list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
221 tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
227 * Handle full virtualization of the time stamp counter. As noted
228 * above, we don't store the actual value of the TSC, only the guest's
229 * offset from monotonic guest's time. If the guest writes to the TSC, we
230 * handle this by changing that offset.
232 * Possible TODO: Proper hooking of TSC read/writes?
235 int v3_rdtsc(struct guest_info * info) {
236 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
238 info->vm_regs.rdx = tscval >> 32;
239 info->vm_regs.rax = tscval & 0xffffffffLL;
244 int v3_handle_rdtsc(struct guest_info * info) {
247 info->vm_regs.rax &= 0x00000000ffffffffLL;
248 info->vm_regs.rdx &= 0x00000000ffffffffLL;
255 int v3_rdtscp(struct guest_info * info) {
257 /* First get the MSR value that we need. It's safe to futz with
258 * ra/c/dx here since they're modified by this instruction anyway. */
259 info->vm_regs.rcx = TSC_AUX_MSR;
260 ret = v3_handle_msr_read(info);
266 info->vm_regs.rcx = info->vm_regs.rax;
268 /* Now do the TSC half of the instruction */
269 ret = v3_rdtsc(info);
279 int v3_handle_rdtscp(struct guest_info * info) {
280 PrintDebug("Handling virtual RDTSCP call.\n");
284 info->vm_regs.rax &= 0x00000000ffffffffLL;
285 info->vm_regs.rcx &= 0x00000000ffffffffLL;
286 info->vm_regs.rdx &= 0x00000000ffffffffLL;
293 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
294 struct v3_msr *msr_val, void *priv) {
295 struct vm_core_time * time_state = &(info->time_state);
297 V3_ASSERT(msr_num == TSC_AUX_MSR);
299 msr_val->lo = time_state->tsc_aux.lo;
300 msr_val->hi = time_state->tsc_aux.hi;
305 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
306 struct v3_msr msr_val, void *priv) {
307 struct vm_core_time * time_state = &(info->time_state);
309 V3_ASSERT(msr_num == TSC_AUX_MSR);
311 time_state->tsc_aux.lo = msr_val.lo;
312 time_state->tsc_aux.hi = msr_val.hi;
317 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
318 struct v3_msr *msr_val, void *priv) {
319 uint64_t time = v3_get_guest_tsc(&info->time_state);
321 V3_ASSERT(msr_num == TSC_MSR);
323 msr_val->hi = time >> 32;
324 msr_val->lo = time & 0xffffffffLL;
329 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
330 struct v3_msr msr_val, void *priv) {
331 struct vm_core_time * time_state = &(info->time_state);
332 uint64_t guest_time, new_tsc;
334 V3_ASSERT(msr_num == TSC_MSR);
336 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
337 guest_time = v3_get_guest_time(time_state);
338 time_state->tsc_guest_offset = (sint64_t)(new_tsc - guest_time);
344 int v3_init_time_vm(struct v3_vm_info * vm) {
347 PrintDebug("Installing TSC MSR hook.\n");
348 ret = v3_hook_msr(vm, TSC_MSR,
349 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
355 PrintDebug("Installing TSC_AUX MSR hook.\n");
356 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
357 tsc_aux_msr_write_hook, NULL);
363 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
364 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
365 handle_cpufreq_hcall, NULL);
367 vm->time_state.td_num = 1;
368 vm->time_state.td_denom = 1;
369 PrintDebug("Setting base time dilation factor to %d/%d.\n",
370 vm->time_state.td_num, vm->time_state.td_denom);
372 vm->time_state.follow_host_time = 1;
373 PrintDebug("Locking guest time to host time.\n");
377 void v3_deinit_time_vm(struct v3_vm_info * vm) {
378 v3_unhook_msr(vm, TSC_MSR);
379 v3_unhook_msr(vm, TSC_AUX_MSR);
381 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
384 void v3_init_time_core(struct guest_info * info) {
385 struct vm_core_time * time_state = &(info->time_state);
386 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
389 time_state->host_cpu_freq = V3_CPU_KHZ();
390 khz = v3_cfg_val(cfg_tree, "khz");
393 time_state->guest_cpu_freq = atoi(khz);
394 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n",
395 info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq);
398 if ( (khz == NULL) ||
399 (time_state->guest_cpu_freq <= 0) ||
400 (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) {
402 time_state->guest_cpu_freq = time_state->host_cpu_freq;
405 /* Compute these using the GCD() of the guest and host CPU freq.
406 * If the GCD is too small, make it "big enough" */
407 time_state->clock_ratio_num = 1;
408 time_state->clock_ratio_denom = 1;
410 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
411 info->pcpu_id, info->vcpu_id,
412 time_state->guest_cpu_freq,
413 time_state->host_cpu_freq);
415 time_state->guest_cycles = 0;
416 time_state->tsc_guest_offset = 0;
417 time_state->last_update = 0;
419 time_state->initial_host_time = 0;
420 time_state->vm_enter_host_time = 0;
421 time_state->vm_pause_host_time = 0;
423 time_state->time_flags = 0; // XXX need to set trap TSC flag or not wisely
425 INIT_LIST_HEAD(&(time_state->timers));
426 time_state->num_timers = 0;
428 time_state->tsc_aux.lo = 0;
429 time_state->tsc_aux.hi = 0;
433 void v3_deinit_time_core(struct guest_info * core) {
434 struct vm_core_time * time_state = &(core->time_state);
435 struct v3_timer * tmr = NULL;
436 struct v3_timer * tmp = NULL;
438 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
439 v3_remove_timer(core, tmr);