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);
87 static int handle_rdhtsc_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) {
88 struct vm_core_time * time_state = &(info->time_state);
90 info->vm_regs.rbx = v3_get_host_time(time_state);
92 // PrintDebug("Guest request host TSC: return %ld\n", (long)info->vm_regs.rbx);
99 int v3_start_time(struct guest_info * info) {
100 /* We start running with guest_time == host_time */
101 uint64_t t = v3_get_host_time(&info->time_state);
103 info->time_state.initial_host_time = t;
104 info->yield_start_cycle = t;
106 info->time_state.last_update = 0;
107 info->time_state.guest_cycles = 0;
108 PrintDebug("Starting time for core %d at host time %llu/guest time %llu.\n",
109 info->vcpu_id, t, info->time_state.guest_cycles);
115 host_to_guest_cycles(struct guest_info * info, sint64_t host_cycles) {
116 struct vm_core_time * core_time_state = &(info->time_state);
117 uint32_t cl_num, cl_denom;
119 cl_num = core_time_state->clock_ratio_num;
120 cl_denom = core_time_state->clock_ratio_denom;
122 return (host_cycles * cl_num) / cl_denom;
127 guest_to_host_cycles(struct guest_info * info, sint64_t guest_cycles) {
128 struct vm_core_time * core_time_state = &(info->time_state);
129 uint32_t cl_num, cl_denom;
131 cl_num = core_time_state->clock_ratio_num;
132 cl_denom = core_time_state->clock_ratio_denom;
134 return (guest_cycles * cl_denom) / cl_num;
138 int v3_advance_time(struct guest_info * info, uint64_t *host_cycles)
140 uint64_t guest_cycles;
142 if (info->time_state.flags & VM_TIME_SLAVE_HOST) {
143 struct v3_time *vm_ts = &(info->vm_info->time_state);
144 uint64_t ht = v3_get_host_time(&info->time_state);
145 uint64_t host_elapsed = ht - info->time_state.initial_host_time;
146 uint64_t dilated_elapsed = (host_elapsed * vm_ts->td_num) / vm_ts->td_denom;
147 uint64_t guest_elapsed = host_to_guest_cycles(info, dilated_elapsed);
148 guest_cycles = guest_elapsed - v3_get_guest_time(&info->time_state);
149 } else if (host_cycles) {
150 guest_cycles = host_to_guest_cycles(info, *host_cycles);
155 info->time_state.guest_cycles += guest_cycles;
160 struct v3_timer * v3_add_timer(struct guest_info * info,
161 struct v3_timer_ops * ops,
162 void * private_data) {
163 struct v3_timer * timer = NULL;
164 timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer));
165 V3_ASSERT(timer != NULL);
168 timer->private_data = private_data;
170 list_add(&(timer->timer_link), &(info->time_state.timers));
171 info->time_state.num_timers++;
176 int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) {
177 list_del(&(timer->timer_link));
178 info->time_state.num_timers--;
184 void v3_update_timers(struct guest_info * info) {
185 struct vm_core_time *time_state = &info->time_state;
186 struct v3_timer * tmp_timer;
188 uint64_t old_time = time_state->last_update;
190 time_state->last_update = v3_get_guest_time(time_state);
191 cycles = (sint64_t)(time_state->last_update - old_time);
193 PrintError("Cycles appears to have rolled over - old time %lld, current time %lld.\n",
194 old_time, time_state->last_update);
198 //PrintDebug("Updating timers with %lld elapsed cycles.\n", cycles);
199 list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) {
200 tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data);
206 * Handle full virtualization of the time stamp counter. As noted
207 * above, we don't store the actual value of the TSC, only the guest's
208 * offset from monotonic guest's time. If the guest writes to the TSC, we
209 * handle this by changing that offset.
211 * Possible TODO: Proper hooking of TSC read/writes?
214 int v3_rdtsc(struct guest_info * info) {
215 uint64_t tscval = v3_get_guest_tsc(&info->time_state);
217 info->vm_regs.rdx = tscval >> 32;
218 info->vm_regs.rax = tscval & 0xffffffffLL;
223 int v3_handle_rdtsc(struct guest_info * info) {
224 PrintDebug("Handling virtual RDTSC call.\n");
227 info->vm_regs.rax &= 0x00000000ffffffffLL;
228 info->vm_regs.rdx &= 0x00000000ffffffffLL;
235 int v3_rdtscp(struct guest_info * info) {
237 /* First get the MSR value that we need. It's safe to futz with
238 * ra/c/dx here since they're modified by this instruction anyway. */
239 info->vm_regs.rcx = TSC_AUX_MSR;
240 ret = v3_handle_msr_read(info);
246 info->vm_regs.rcx = info->vm_regs.rax;
248 /* Now do the TSC half of the instruction */
249 ret = v3_rdtsc(info);
259 int v3_handle_rdtscp(struct guest_info * info) {
260 PrintDebug("Handling virtual RDTSCP call.\n");
264 info->vm_regs.rax &= 0x00000000ffffffffLL;
265 info->vm_regs.rcx &= 0x00000000ffffffffLL;
266 info->vm_regs.rdx &= 0x00000000ffffffffLL;
273 static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num,
274 struct v3_msr *msr_val, void *priv) {
275 struct vm_core_time * time_state = &(info->time_state);
277 V3_ASSERT(msr_num == TSC_AUX_MSR);
279 msr_val->lo = time_state->tsc_aux.lo;
280 msr_val->hi = time_state->tsc_aux.hi;
285 static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num,
286 struct v3_msr msr_val, void *priv) {
287 struct vm_core_time * time_state = &(info->time_state);
289 V3_ASSERT(msr_num == TSC_AUX_MSR);
291 time_state->tsc_aux.lo = msr_val.lo;
292 time_state->tsc_aux.hi = msr_val.hi;
297 static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num,
298 struct v3_msr *msr_val, void *priv) {
299 uint64_t time = v3_get_guest_tsc(&info->time_state);
301 PrintDebug("Handling virtual TSC MSR read call.\n");
302 V3_ASSERT(msr_num == TSC_MSR);
304 msr_val->hi = time >> 32;
305 msr_val->lo = time & 0xffffffffLL;
310 static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num,
311 struct v3_msr msr_val, void *priv) {
312 struct vm_core_time * time_state = &(info->time_state);
313 uint64_t guest_time, new_tsc;
315 PrintDebug("Handling virtual TSC MSR write call.\n");
316 V3_ASSERT(msr_num == TSC_MSR);
318 new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo;
319 guest_time = v3_get_guest_time(time_state);
320 time_state->tsc_guest_offset = (sint64_t)(new_tsc - guest_time);
326 handle_time_configuration(struct v3_vm_info * vm, v3_cfg_tree_t *cfg) {
327 char *source, *dilation, *tsc;
329 vm->time_state.flags = V3_TIME_SLAVE_HOST;
330 vm->time_state.td_num = vm->time_state.td_denom = 1;
334 source = v3_cfg_val(cfg, "source");
336 if (strcasecmp(source, "none") == 0) {
337 vm->time_state.flags &= ~V3_TIME_SLAVE_HOST;
338 } else if (strcasecmp(source, "host") != 0) {
339 PrintError("Unknown time source for VM core time management.\n");
341 PrintDebug("VM time slaved to host TSC.\n");
345 // Should we make a separate TSC device that handles this sort of thing?
346 tsc = v3_cfg_val(cfg, "tsc");
348 if (strcasecmp(tsc, "host") == 0) {
349 if (!(vm->time_state.flags & V3_TIME_SLAVE_HOST)) {
350 PrintError("WARNING: Guest TSC set to passthrough host TSC, but guest time not slaved to host time.");
352 vm->time_state.flags |= V3_TIME_TSC_PASSTHROUGH;
353 } else if (strcasecmp(source, "guest") != 0) {
354 PrintError("ERROR: Unknown TSC configuration in time configuration.\n");
358 dilation = v3_cfg_val(cfg, "dilation");
360 if (!(vm->time_state.flags & VM_TIME_SLAVE_HOST)) {
361 PrintError("Time dilation only valid when slaved to host time.\n");
363 uint32_t num = 1, denom = 1;
364 denom = atoi(dilation);
365 if ((num > 0) && (denom > 0)) {
366 vm->time_state.td_num = num;
367 vm->time_state.td_denom = denom;
370 if ((vm->time_state.td_num != 1)
371 || (vm->time_state.td_denom != 1)) {
372 V3_Print("Time dilated from host time by a factor of %d/%d"
373 " in guest.\n", vm->time_state.td_denom,
374 vm->time_state.td_num);
376 PrintError("Time dilation specifier in configuration did not"
377 " result in actual time dilation in VM.\n");
383 int v3_init_time_vm(struct v3_vm_info * vm) {
384 v3_cfg_tree_t * cfg_tree = vm->cfg_data->cfg;
387 PrintDebug("Installing TSC MSR hook.\n");
388 ret = v3_hook_msr(vm, TSC_MSR,
389 tsc_msr_read_hook, tsc_msr_write_hook, NULL);
395 PrintDebug("Installing TSC_AUX MSR hook.\n");
396 ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook,
397 tsc_aux_msr_write_hook, NULL);
403 PrintDebug("Registering TIME_CPUFREQ hypercall.\n");
404 ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL,
405 handle_cpufreq_hcall, NULL);
406 PrintDebug("Registering TIME_RDHTSC hypercall.\n");
407 ret = v3_register_hypercall(vm, TIME_RDHTSC_HCALL,
408 handle_rdhtsc_hcall, NULL);
410 handle_time_configuration(vm, v3_cfg_subtree(cfg_tree, "time"));
415 void v3_deinit_time_vm(struct v3_vm_info * vm) {
416 v3_unhook_msr(vm, TSC_MSR);
417 v3_unhook_msr(vm, TSC_AUX_MSR);
419 v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL);
423 gcd ( uint32_t a, uint32_t b )
427 c = a; a = b%a; b = c;
432 static int compute_core_ratios(struct guest_info * info,
433 uint32_t hostKhz, uint32_t guestKhz)
435 struct vm_core_time * time_state = &(info->time_state);
438 /* Compute these using the GCD() of the guest and host CPU freq.
439 * If the GCD is too small, make it "big enough" */
440 khzGCD = gcd(hostKhz, guestKhz);
444 time_state->clock_ratio_num = guestKhz / khzGCD;
445 time_state->clock_ratio_denom = hostKhz / khzGCD;
447 time_state->ipc_ratio_num = 1;
448 time_state->ipc_ratio_denom = 1;
453 void v3_init_time_core(struct guest_info * info) {
454 struct vm_core_time * time_state = &(info->time_state);
455 v3_cfg_tree_t * cfg_tree = info->core_cfg_data;
458 time_state->host_cpu_freq = V3_CPU_KHZ();
459 khz = v3_cfg_val(cfg_tree, "khz");
462 time_state->guest_cpu_freq = atoi(khz);
463 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n",
464 info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq);
467 if ( (khz == NULL) ||
468 (time_state->guest_cpu_freq <= 0)) {
469 /* || (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) { */
470 time_state->guest_cpu_freq = time_state->host_cpu_freq;
472 compute_core_ratios(info, time_state->host_cpu_freq,
473 time_state->guest_cpu_freq);
475 time_state->flags = 0;
476 if (info->vm_info->time_state.flags & V3_TIME_SLAVE_HOST) {
477 time_state->flags |= VM_TIME_SLAVE_HOST;
479 if (info->vm_info->time_state.flags & V3_TIME_TSC_PASSTHROUGH) {
480 time_state->flags |= VM_TIME_TSC_PASSTHROUGH;
483 if ((time_state->clock_ratio_denom != 1) ||
484 (time_state->clock_ratio_num != 1) ||
485 (info->vm_info->time_state.td_num != 1) ||
486 (info->vm_info->time_state.td_denom != 1)) {
487 if (time_state->flags | VM_TIME_TSC_PASSTHROUGH) {
488 PrintError("WARNING: Cannot use reqested passthrough TSC with clock or time modification also requested.\n");
489 time_state->flags &= ~VM_TIME_TSC_PASSTHROUGH;
491 time_state->flags |= VM_TIME_TRAP_RDTSC;
494 PrintDebug("Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n",
495 info->pcpu_id, info->vcpu_id,
496 time_state->guest_cpu_freq,
497 time_state->host_cpu_freq);
498 PrintDebug(" td_mult = %d/%d, cl_mult = %u/%u, ipc_mult = %u/%u.\n",
499 info->vm_info->time_state.td_num,
500 info->vm_info->time_state.td_denom,
501 time_state->clock_ratio_num, time_state->clock_ratio_denom,
502 time_state->ipc_ratio_num, time_state->ipc_ratio_denom);
503 PrintDebug(" time source = %s, tsc handling = %s\n",
504 (time_state->flags & VM_TIME_SLAVE_HOST) ? "host" : "none",
505 (time_state->flags & VM_TIME_TSC_PASSTHROUGH) ? "passthrough"
506 : (time_state->flags & VM_TIME_TRAP_RDTSC) ? "trapping"
509 time_state->guest_cycles = 0;
510 time_state->tsc_guest_offset = 0;
511 time_state->last_update = 0;
512 time_state->initial_host_time = 0;
514 INIT_LIST_HEAD(&(time_state->timers));
515 time_state->num_timers = 0;
517 time_state->tsc_aux.lo = 0;
518 time_state->tsc_aux.hi = 0;
522 void v3_deinit_time_core(struct guest_info * core) {
523 struct vm_core_time * time_state = &(core->time_state);
524 struct v3_timer * tmr = NULL;
525 struct v3_timer * tmp = NULL;
527 list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) {
528 v3_remove_timer(core, tmr);