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.initial_host_time = t;
94 info->yield_start_cycle = t;
96 info->time_state.last_update = 0;
97 info->time_state.guest_cycles = 0;
98 PrintDebug("Starting time for core %d at host time %llu/guest time %llu.\n",
99 info->vcpu_id, t, info->time_state.guest_cycles);
105 host_to_guest_cycles(struct guest_info * info, sint64_t host_cycles) {
106 struct vm_core_time * core_time_state = &(info->time_state);
107 uint32_t cl_num, cl_denom;
109 cl_num = core_time_state->clock_ratio_num;
110 cl_denom = core_time_state->clock_ratio_denom;
112 return (host_cycles * cl_num) / cl_denom;
117 guest_to_host_cycles(struct guest_info * info, sint64_t guest_cycles) {
118 struct vm_core_time * core_time_state = &(info->time_state);
119 uint32_t cl_num, cl_denom;
121 cl_num = core_time_state->clock_ratio_num;
122 cl_denom = core_time_state->clock_ratio_denom;
124 return (guest_cycles * cl_denom) / cl_num;
128 int v3_advance_time(struct guest_info * info, uint64_t *host_cycles)
130 uint64_t guest_cycles;
132 if (info->flags & VM_TIME_SLAVE_HOST) {
133 struct v3_time *vm_ts = &(info->vm_info->time_state);
134 uint64_t ht = v3_get_host_time(&info->time_state);
135 uint64_t host_elapsed = ht - info->time_state.initial_host_time;
136 uint64_t dilated_elapsed = (host_elapsed * vm_ts->td_num) / vm_ts->td_denom;
137 uint64_t guest_elapsed = host_to_guest_cycles(info, dilated_elapsed);
138 guest_cycles = guest_elapsed - v3_get_guest_time(&info->time_state);
139 } else if (host_cycles) {
140 guest_cycles = host_to_guest_cycles(info, *host_cycles);
145 info->time_state.guest_cycles += guest_cycles;
150 struct v3_timer * v3_add_timer(struct guest_info * info,
151 struct v3_timer_ops * ops,
152 void * private_data) {
153 struct v3_timer * timer = NULL;
154 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
155 V3_ASSERT(timer != NULL);
158 timer->private_data = private_data;
160 list_add(&(timer->timer_link), &(info->time_state.timers));
161 info->time_state.num_timers++;
166 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
167 list_del(&(timer->timer_link));
168 info->time_state.num_timers--;
174 void v3_update_timers(struct guest_info * info) {
175 struct vm_core_time *time_state = &info->time_state;
176 struct v3_timer * tmp_timer;
178 uint64_t old_time = time_state->last_update;
180 time_state->last_update = v3_get_guest_time(time_state);
181 cycles = (sint64_t)(time_state->last_update - old_time);
183 PrintError("Cycles appears to have rolled over - old time %lld, current time %lld.\n",
184 old_time, time_state->last_update);
188 //PrintDebug("Updating timers with %lld elapsed cycles.\n", cycles);
189 list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
190 tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
196 * Handle full virtualization of the time stamp counter. As noted
197 * above, we don't store the actual value of the TSC, only the guest's
198 * offset from monotonic guest's time. If the guest writes to the TSC, we
199 * handle this by changing that offset.
201 * Possible TODO: Proper hooking of TSC read/writes?
204 int v3_rdtsc(struct guest_info * info) {
205 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
207 info->vm_regs.rdx = tscval >> 32;
208 info->vm_regs.rax = tscval & 0xffffffffLL;
213 int v3_handle_rdtsc(struct guest_info * info) {
216 info->vm_regs.rax &= 0x00000000ffffffffLL;
217 info->vm_regs.rdx &= 0x00000000ffffffffLL;
224 int v3_rdtscp(struct guest_info * info) {
226 /* First get the MSR value that we need. It's safe to futz with
227 * ra/c/dx here since they're modified by this instruction anyway. */
228 info->vm_regs.rcx = TSC_AUX_MSR;
229 ret = v3_handle_msr_read(info);
235 info->vm_regs.rcx = info->vm_regs.rax;
237 /* Now do the TSC half of the instruction */
238 ret = v3_rdtsc(info);
248 int v3_handle_rdtscp(struct guest_info * info) {
249 PrintDebug("Handling virtual RDTSCP call.\n");
253 info->vm_regs.rax &= 0x00000000ffffffffLL;
254 info->vm_regs.rcx &= 0x00000000ffffffffLL;
255 info->vm_regs.rdx &= 0x00000000ffffffffLL;
262 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
263 struct v3_msr *msr_val, void *priv) {
264 struct vm_core_time * time_state = &(info->time_state);
266 V3_ASSERT(msr_num == TSC_AUX_MSR);
268 msr_val->lo = time_state->tsc_aux.lo;
269 msr_val->hi = time_state->tsc_aux.hi;
274 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
275 struct v3_msr msr_val, void *priv) {
276 struct vm_core_time * time_state = &(info->time_state);
278 V3_ASSERT(msr_num == TSC_AUX_MSR);
280 time_state->tsc_aux.lo = msr_val.lo;
281 time_state->tsc_aux.hi = msr_val.hi;
286 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
287 struct v3_msr *msr_val, void *priv) {
288 uint64_t time = v3_get_guest_tsc(&info->time_state);
290 V3_ASSERT(msr_num == TSC_MSR);
292 msr_val->hi = time >> 32;
293 msr_val->lo = time & 0xffffffffLL;
298 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
299 struct v3_msr msr_val, void *priv) {
300 struct vm_core_time * time_state = &(info->time_state);
301 uint64_t guest_time, new_tsc;
303 V3_ASSERT(msr_num == TSC_MSR);
305 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
306 guest_time = v3_get_guest_time(time_state);
307 time_state->tsc_guest_offset = (sint64_t)(new_tsc - guest_time);
313 handle_time_configuration(struct v3_vm_info * vm, v3_cfg_tree_t *cfg) {
314 v3_cfg_tree_t * slave;
316 vm->time_state.flags = 0;
317 vm->time_state.td_num = vm->time_state.td_denom = 1;
321 slave = v3_cfg_subtree(cfg, "slave");
324 char *source = v3_cfg_val(slave, "source");
325 v3_cfg_tree_t *dilation = v3_cfg_subtree(slave, "dilation");
327 if (strcasecmp(source, "host") == 0) {
328 PrintDebug("Slaving VM guest time to host time.\n");
329 vm->time_state.flags |= V3_TIME_SLAVE_HOST;
331 PrintError("Unknown time source for slaving.\n");
334 if (dilation && (vm->time_state.flags & V3_TIME_SLAVE_HOST)) {
336 uint32_t num = 1, denom = 1;
337 if ((str1 = v3_cfg_val(dilation, "value"))) {
339 } else if ((str1 = v3_cfg_val(dilation, "num"))
340 && (str2 = v3_cfg_val(dilation, "denom"))) {
344 if ((num > 0) && (denom > 0)) {
345 vm->time_state.td_num = num;
346 vm->time_state.td_denom = denom;
348 if ((vm->time_state.td_num != 1)
349 || (vm->time_state.td_denom != 1)) {
350 V3_Print("Time dilated from host time by a factor of %d/%d"
351 " in guest.\n", denom, num);
353 PrintError("Time dilation specifier in configuration did not"
354 " result in actual time dilation in VM.\n");
361 int v3_init_time_vm(struct v3_vm_info * vm) {
362 v3_cfg_tree_t * cfg_tree = vm->cfg_data->cfg;
365 PrintDebug("Installing TSC MSR hook.\n");
366 ret = v3_hook_msr(vm, TSC_MSR,
367 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
373 PrintDebug("Installing TSC_AUX MSR hook.\n");
374 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
375 tsc_aux_msr_write_hook, NULL);
381 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
382 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
383 handle_cpufreq_hcall, NULL);
385 handle_time_configuration(vm, v3_cfg_subtree(cfg_tree, "time"));
390 void v3_deinit_time_vm(struct v3_vm_info * vm) {
391 v3_unhook_msr(vm, TSC_MSR);
392 v3_unhook_msr(vm, TSC_AUX_MSR);
394 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
398 gcd ( uint32_t a, uint32_t b )
402 c = a; a = b%a; b = c;
407 static int compute_core_ratios(struct guest_info * info,
408 uint32_t hostKhz, uint32_t guestKhz)
410 struct vm_core_time * time_state = &(info->time_state);
413 /* Compute these using the GCD() of the guest and host CPU freq.
414 * If the GCD is too small, make it "big enough" */
415 khzGCD = gcd(hostKhz, guestKhz);
419 time_state->clock_ratio_num = guestKhz / khzGCD;
420 time_state->clock_ratio_denom = hostKhz / khzGCD;
422 time_state->ipc_ratio_num = 1;
423 time_state->ipc_ratio_denom = 1;
428 void v3_init_time_core(struct guest_info * info) {
429 struct vm_core_time * time_state = &(info->time_state);
430 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
433 time_state->host_cpu_freq = V3_CPU_KHZ();
434 khz = v3_cfg_val(cfg_tree, "khz");
437 time_state->guest_cpu_freq = atoi(khz);
438 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n",
439 info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq);
442 if ( (khz == NULL) ||
443 (time_state->guest_cpu_freq <= 0) ||
444 (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) {
446 time_state->guest_cpu_freq = time_state->host_cpu_freq;
448 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
449 info->pcpu_id, info->vcpu_id,
450 time_state->guest_cpu_freq,
451 time_state->host_cpu_freq);
453 compute_core_ratios(info, time_state->host_cpu_freq,
454 time_state->guest_cpu_freq);
456 PrintDebug(" td_mult = %d/%d, cl_mult = %u/%u, ipc_mult = %u/%u.\n",
457 info->vm_info->time_state.td_num,
458 info->vm_info->time_state.td_denom,
459 time_state->clock_ratio_num, time_state->clock_ratio_denom,
460 time_state->ipc_ratio_num, time_state->ipc_ratio_denom);
461 time_state->guest_cycles = 0;
462 time_state->tsc_guest_offset = 0;
463 time_state->last_update = 0;
465 time_state->initial_host_time = 0;
467 time_state->flags = 0;
468 if (info->vm_info->time_state.flags & V3_TIME_SLAVE_HOST) {
469 time_state->flags |= VM_TIME_SLAVE_HOST;
471 if ((time_state->clock_ratio_denom != 1) ||
472 (time_state->clock_ratio_num != 1)) {
473 time_state->flags |= VM_TIME_TRAP_RDTSC;
476 INIT_LIST_HEAD(&(time_state->timers));
477 time_state->num_timers = 0;
479 time_state->tsc_aux.lo = 0;
480 time_state->tsc_aux.hi = 0;
484 void v3_deinit_time_core(struct guest_info * core) {
485 struct vm_core_time * time_state = &(core->time_state);
486 struct v3_timer * tmr = NULL;
487 struct v3_timer * tmp = NULL;
489 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
490 v3_remove_timer(core, tmr);