X-Git-Url: http://v3vee.org/palacios/gitweb/gitweb.cgi?p=palacios.git;a=blobdiff_plain;f=palacios%2Fsrc%2Fpalacios%2Fvmm_time.c;h=dff562b84f3f876a4a8db028573375be5c747026;hp=fa8f5cde8755b15191f1f4683ccab4c82e4b9c01;hb=88a3605446744969abe6f193a7bc20e62d5aa555;hpb=f88a692d094459f0326c5c891df5ea81b5476ba6 diff --git a/palacios/src/palacios/vmm_time.c b/palacios/src/palacios/vmm_time.c index fa8f5cd..dff562b 100644 --- a/palacios/src/palacios/vmm_time.c +++ b/palacios/src/palacios/vmm_time.c @@ -41,15 +41,14 @@ * the passage of time: * (1) The host timestamp counter - read directly from HW and never written * (2) A monotonic guest timestamp counter used to measure the progression of - * time in the guest. This is computed using an offsets from (1) above. + * time in the guest. This is stored as an absolute number of cycles elapsed + * and is updated on guest entry and exit; it can also be updated explicitly + * in the monitor at times * (3) The actual guest timestamp counter (which can be written by * writing to the guest TSC MSR - MSR 0x10) from the monotonic guest TSC. * This is also computed as an offset from (2) above when the TSC and * this offset is updated when the TSC MSR is written. * - * The value used to offset the guest TSC from the host TSC is the *sum* of all - * of these offsets (2 and 3) above - * * Because all other devices are slaved off of the passage of time in the guest, * it is (2) above that drives the firing of other timers in the guest, * including timer devices such as the Programmable Interrupt Timer (PIT). @@ -58,7 +57,7 @@ * (1) Add support for temporarily skewing guest time off of where it should * be to support slack simulation of guests. The idea is that simulators * set this skew to be the difference between how much time passed for a - * simulated feature and a real implementation of that feature, making + * simulated feature and a real implementation of that feature, making time * pass at a different rate from real time on this core. The VMM will then * attempt to move this skew back towards 0 subject to resolution/accuracy * constraints from various system timers. @@ -66,184 +65,110 @@ * The main effort in doing this will be to get accuracy/resolution * information from each local timer and to use this to bound how much skew * is removed on each exit. + * + * (2) Look more into sychronizing the offsets *across* virtual and physical + * cores so that multicore guests stay mostly in sync. + * + * (3) Look into using the AMD TSC multiplier feature and adding explicit time + * dilation support to time handling. */ static int handle_cpufreq_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) { - struct vm_time * time_state = &(info->time_state); + struct vm_core_time * time_state = &(info->time_state); info->vm_regs.rbx = time_state->guest_cpu_freq; - PrintDebug("Guest request cpu frequency: return %ld\n", (long)info->vm_regs.rbx); + PrintDebug(info->vm_info, info, "Guest request cpu frequency: return %ld\n", (long)info->vm_regs.rbx); return 0; } +static int handle_rdhtsc_hcall(struct guest_info * info, uint_t hcall_id, void * priv_data) { + struct vm_core_time * time_state = &(info->time_state); + info->vm_regs.rbx = v3_get_host_time(time_state); -int v3_start_time(struct guest_info * info) { - /* We start running with guest_time == host_time */ - uint64_t t = v3_get_host_time(&info->time_state); - - PrintDebug("Starting initial guest time as %llu\n", t); - - info->time_state.enter_time = 0; - info->time_state.exit_time = t; - info->time_state.last_update = t; - info->time_state.initial_time = t; - info->yield_start_cycle = t; - - return 0; -} - -int v3_offset_time( struct guest_info * info, sint64_t offset ) -{ - struct vm_time * time_state = &(info->time_state); -// PrintDebug("Adding additional offset of %lld to guest time.\n", offset); - time_state->guest_host_offset += offset; - return 0; -} - -static uint64_t compute_target_host_time(struct guest_info * info) -{ - struct vm_time * time_state = &(info->time_state); - uint64_t guest_elapsed, desired_elapsed; + // PrintDebug(info->vm_info, info, "Guest request host TSC: return %ld\n", (long)info->vm_regs.rbx); - guest_elapsed = (v3_get_guest_time(time_state) - time_state->initial_time); - desired_elapsed = (guest_elapsed * time_state->host_cpu_freq) / time_state->guest_cpu_freq; - return time_state->initial_time + desired_elapsed; + return 0; } -static uint64_t compute_target_guest_time(struct guest_info *info) -{ - struct vm_time * time_state = &(info->time_state); - uint64_t host_elapsed, desired_elapsed; - - host_elapsed = v3_get_host_time(time_state) - time_state->initial_time; - desired_elapsed = (host_elapsed * time_state->guest_cpu_freq) / time_state->host_cpu_freq; - return time_state->initial_time + desired_elapsed; - -} - -/* Yield time in the host to deal with a guest that wants to run slower than - * the native host cycle frequency */ -static int yield_host_time(struct guest_info * info) { - struct vm_time * time_state = &(info->time_state); - uint64_t host_time, target_host_time; - uint64_t guest_time, old_guest_time; - - /* Compute the target host time given how much time has *already* - * passed in the guest */ - target_host_time = compute_target_host_time(info); - - /* Now, let the host run while the guest is stopped to make the two - * sync up. Note that this doesn't assume that guest time is stopped; - * the offsetting in the next step will change add an offset to guest - * time to account for the time paused even if the geust isn't - * usually paused in the VMM. */ - host_time = v3_get_host_time(time_state); - old_guest_time = v3_get_guest_time(time_state); - - while (target_host_time > host_time) { - v3_yield(info); - host_time = v3_get_host_time(time_state); - } - guest_time = v3_get_guest_time(time_state); +int v3_start_time(struct guest_info * info) { + /* We start running with guest_time == host_time */ + uint64_t t = v3_get_host_time(&info->time_state); - /* We do *not* assume the guest timer was paused in the VM. If it was - * this offseting is 0. If it wasn't, we need this. */ - v3_offset_time(info, (sint64_t)old_guest_time - (sint64_t)guest_time); + info->time_state.initial_host_time = t; + info->yield_start_cycle = t; + info->time_state.last_update = 0; + info->time_state.guest_cycles = 0; + PrintDebug(info->vm_info, info, "Starting time for core %d at host time %llu/guest time %llu.\n", + info->vcpu_id, t, info->time_state.guest_cycles); + v3_yield(info,-1); return 0; } -static int skew_guest_time(struct guest_info * info) { - struct vm_time * time_state = &(info->time_state); - uint64_t target_guest_time, guest_time; - /* Now the host may have gotten ahead of the guest because - * yielding is a coarse grained thing. Figure out what guest time - * we want to be at, and use the use the offsetting mechanism in - * the VMM to make the guest run forward. We limit *how* much we skew - * it forward to prevent the guest time making large jumps, - * however. */ - target_guest_time = compute_target_guest_time(info); - guest_time = v3_get_guest_time(time_state); +static sint64_t +host_to_guest_cycles(struct guest_info * info, sint64_t host_cycles) { + struct vm_core_time * core_time_state = &(info->time_state); + uint32_t cl_num, cl_denom; - if (guest_time < target_guest_time) { - sint64_t max_skew, desired_skew, skew; + cl_num = core_time_state->clock_ratio_num; + cl_denom = core_time_state->clock_ratio_denom; - if (time_state->enter_time) { - /* Limit forward skew to 10% of the amount the guest has - * run since we last could skew time */ - max_skew = ((sint64_t)guest_time - (sint64_t)time_state->enter_time) / 10; - } else { - max_skew = 0; - } - - desired_skew = (sint64_t)target_guest_time - (sint64_t)guest_time; - skew = desired_skew > max_skew ? max_skew : desired_skew; - PrintDebug("Guest %lld cycles behind where it should be.\n", - desired_skew); - PrintDebug("Limit on forward skew is %lld. Skewing forward %lld.\n", - max_skew, skew); - - v3_offset_time(info, skew); - } - - return 0; + return (host_cycles * cl_num) / cl_denom; } -// Control guest time in relation to host time so that the two stay -// appropriately synchronized to the extent possible. -int v3_adjust_time(struct guest_info * info) { +/* +static sint64_t +guest_to_host_cycles(struct guest_info * info, sint64_t guest_cycles) { + struct vm_core_time * core_time_state = &(info->time_state); + uint32_t cl_num, cl_denom; - /* First deal with yielding if we want to slow down the guest */ - yield_host_time(info); + cl_num = core_time_state->clock_ratio_num; + cl_denom = core_time_state->clock_ratio_denom; - /* Now, if the guest is too slow, (either from excess yielding above, - * or because the VMM is doing something that takes a long time to emulate) - * allow guest time to jump forward a bit */ - skew_guest_time(info); - - return 0; + return (guest_cycles * cl_denom) / cl_num; } +*/ -/* Called immediately upon entry in the the VMM */ -int -v3_time_exit_vm( struct guest_info * info ) +int v3_advance_time(struct guest_info * info, uint64_t *host_cycles) { - struct vm_time * time_state = &(info->time_state); + uint64_t guest_cycles; + + if (info->time_state.flags & VM_TIME_SLAVE_HOST) { + struct v3_time *vm_ts = &(info->vm_info->time_state); + uint64_t ht = v3_get_host_time(&info->time_state); + uint64_t host_elapsed = ht - info->time_state.initial_host_time; + uint64_t dilated_elapsed = (host_elapsed * vm_ts->td_num) / vm_ts->td_denom; + uint64_t guest_elapsed = host_to_guest_cycles(info, dilated_elapsed); + guest_cycles = guest_elapsed - v3_get_guest_time(&info->time_state); + } else if (host_cycles) { + guest_cycles = host_to_guest_cycles(info, *host_cycles); + } else { + guest_cycles = 0; + } - time_state->exit_time = v3_get_host_time(time_state); + info->time_state.guest_cycles += guest_cycles; return 0; -} - -/* Called immediately prior to entry to the VM */ -int -v3_time_enter_vm( struct guest_info * info ) -{ - struct vm_time * time_state = &(info->time_state); - uint64_t guest_time, host_time; - - host_time = v3_get_host_time(time_state); - guest_time = v3_get_guest_time(time_state); - time_state->enter_time = host_time; - time_state->guest_host_offset = (sint64_t)guest_time - (sint64_t)host_time; - - return 0; -} - +} - struct v3_timer * v3_add_timer(struct guest_info * info, struct v3_timer_ops * ops, void * private_data) { struct v3_timer * timer = NULL; timer = (struct v3_timer *)V3_Malloc(sizeof(struct v3_timer)); - V3_ASSERT(timer != NULL); + + if (!timer) { + PrintError(info->vm_info, info, "Cannot allocate in adding a timer\n"); + return NULL; + } + + V3_ASSERT(info->vm_info, info,timer != NULL); timer->ops = ops; timer->private_data = private_data; @@ -263,21 +188,26 @@ int v3_remove_timer(struct guest_info * info, struct v3_timer * timer) { } void v3_update_timers(struct guest_info * info) { - struct vm_time *time_state = &info->time_state; + struct vm_core_time *time_state = &info->time_state; struct v3_timer * tmp_timer; - uint64_t old_time = info->time_state.last_update; sint64_t cycles; + uint64_t old_time = time_state->last_update; time_state->last_update = v3_get_guest_time(time_state); - cycles = time_state->last_update - old_time; - V3_ASSERT(cycles >= 0); + cycles = (sint64_t)(time_state->last_update - old_time); + if (cycles < 0) { + PrintError(info->vm_info, info, "Cycles appears to have rolled over - old time %lld, current time %lld.\n", + old_time, time_state->last_update); + return; + } - // V3_Print("Updating timers with %lld elapsed cycles.\n", cycles); + //PrintDebug(info->vm_info, info, "Updating timers with %lld elapsed cycles.\n", cycles); list_for_each_entry(tmp_timer, &(time_state->timers), timer_link) { tmp_timer->ops->update_timer(info, cycles, time_state->guest_cpu_freq, tmp_timer->private_data); } } + /* * Handle full virtualization of the time stamp counter. As noted * above, we don't store the actual value of the TSC, only the guest's @@ -297,6 +227,7 @@ int v3_rdtsc(struct guest_info * info) { } int v3_handle_rdtsc(struct guest_info * info) { + PrintDebug(info->vm_info, info, "Handling virtual RDTSC call.\n"); v3_rdtsc(info); info->vm_regs.rax &= 0x00000000ffffffffLL; @@ -332,7 +263,7 @@ int v3_rdtscp(struct guest_info * info) { int v3_handle_rdtscp(struct guest_info * info) { - PrintDebug("Handling virtual RDTSCP call.\n"); + PrintDebug(info->vm_info, info, "Handling virtual RDTSCP call.\n"); v3_rdtscp(info); @@ -347,9 +278,9 @@ int v3_handle_rdtscp(struct guest_info * info) { static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num, struct v3_msr *msr_val, void *priv) { - struct vm_time * time_state = &(info->time_state); + struct vm_core_time * time_state = &(info->time_state); - V3_ASSERT(msr_num == TSC_AUX_MSR); + V3_ASSERT(info->vm_info, info,msr_num == TSC_AUX_MSR); msr_val->lo = time_state->tsc_aux.lo; msr_val->hi = time_state->tsc_aux.hi; @@ -359,9 +290,9 @@ static int tsc_aux_msr_read_hook(struct guest_info *info, uint_t msr_num, static int tsc_aux_msr_write_hook(struct guest_info *info, uint_t msr_num, struct v3_msr msr_val, void *priv) { - struct vm_time * time_state = &(info->time_state); + struct vm_core_time * time_state = &(info->time_state); - V3_ASSERT(msr_num == TSC_AUX_MSR); + V3_ASSERT(info->vm_info, info,msr_num == TSC_AUX_MSR); time_state->tsc_aux.lo = msr_val.lo; time_state->tsc_aux.hi = msr_val.hi; @@ -373,7 +304,8 @@ static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num, struct v3_msr *msr_val, void *priv) { uint64_t time = v3_get_guest_tsc(&info->time_state); - V3_ASSERT(msr_num == TSC_MSR); + PrintDebug(info->vm_info, info, "Handling virtual TSC MSR read call.\n"); + V3_ASSERT(info->vm_info, info,msr_num == TSC_MSR); msr_val->hi = time >> 32; msr_val->lo = time & 0xffffffffLL; @@ -383,23 +315,82 @@ static int tsc_msr_read_hook(struct guest_info *info, uint_t msr_num, static int tsc_msr_write_hook(struct guest_info *info, uint_t msr_num, struct v3_msr msr_val, void *priv) { - struct vm_time * time_state = &(info->time_state); + struct vm_core_time * time_state = &(info->time_state); uint64_t guest_time, new_tsc; - V3_ASSERT(msr_num == TSC_MSR); + PrintDebug(info->vm_info, info, "Handling virtual TSC MSR write call.\n"); + V3_ASSERT(info->vm_info, info,msr_num == TSC_MSR); new_tsc = (((uint64_t)msr_val.hi) << 32) | (uint64_t)msr_val.lo; guest_time = v3_get_guest_time(time_state); - time_state->tsc_guest_offset = (sint64_t)new_tsc - (sint64_t)guest_time; + time_state->tsc_guest_offset = (sint64_t)(new_tsc - guest_time); return 0; } +static int +handle_time_configuration(struct v3_vm_info * vm, v3_cfg_tree_t *cfg) { + char *source, *dilation, *tsc; -int v3_init_time_vm(struct v3_vm_info * vm) { - int ret; + vm->time_state.flags = V3_TIME_SLAVE_HOST; + vm->time_state.td_num = vm->time_state.td_denom = 1; + + if (!cfg) return 0; + + source = v3_cfg_val(cfg, "source"); + if (source) { + if (strcasecmp(source, "none") == 0) { + vm->time_state.flags &= ~V3_TIME_SLAVE_HOST; + } else if (strcasecmp(source, "host") != 0) { + PrintError(vm, VCORE_NONE, "Unknown time source for VM core time management.\n"); + } else { + PrintDebug(vm, VCORE_NONE,"VM time slaved to host TSC.\n"); + } + } + + // Should we make a separate TSC device that handles this sort of thing? + tsc = v3_cfg_val(cfg, "tsc"); + if (tsc) { + if (strcasecmp(tsc, "host") == 0) { + if (!(vm->time_state.flags & V3_TIME_SLAVE_HOST)) { + PrintError(vm, VCORE_NONE, "WARNING: Guest TSC set to passthrough host TSC, but guest time not slaved to host time."); + } + vm->time_state.flags |= V3_TIME_TSC_PASSTHROUGH; + } else if (!source || (strcasecmp(source, "guest") != 0)) { + PrintError(vm, VCORE_NONE, "ERROR: Unknown TSC configuration in time configuration.\n"); + } + } - PrintDebug("Installing TSC MSR hook.\n"); + dilation = v3_cfg_val(cfg, "dilation"); + if (dilation) { + if (!(vm->time_state.flags & VM_TIME_SLAVE_HOST)) { + PrintError(vm, VCORE_NONE, "Time dilation only valid when slaved to host time.\n"); + } else { + uint32_t num = 1, denom = 1; + denom = atoi(dilation); + if ((num > 0) && (denom > 0)) { + vm->time_state.td_num = num; + vm->time_state.td_denom = denom; + } + } + if ((vm->time_state.td_num != 1) + || (vm->time_state.td_denom != 1)) { + V3_Print(vm, VCORE_NONE, "Time dilated from host time by a factor of %d/%d" + " in guest.\n", vm->time_state.td_denom, + vm->time_state.td_num); + } else { + PrintError(vm, VCORE_NONE,"Time dilation specifier in configuration did not" + " result in actual time dilation in VM.\n"); + } + } + return 0; +} + +int v3_init_time_vm(struct v3_vm_info * vm) { + v3_cfg_tree_t * cfg_tree = vm->cfg_data->cfg; + int ret=0; + + PrintDebug(vm, VCORE_NONE, "Installing TSC MSR hook.\n"); ret = v3_hook_msr(vm, TSC_MSR, tsc_msr_read_hook, tsc_msr_write_hook, NULL); @@ -407,7 +398,7 @@ int v3_init_time_vm(struct v3_vm_info * vm) { return ret; } - PrintDebug("Installing TSC_AUX MSR hook.\n"); + PrintDebug(vm, VCORE_NONE, "Installing TSC_AUX MSR hook.\n"); ret = v3_hook_msr(vm, TSC_AUX_MSR, tsc_aux_msr_read_hook, tsc_aux_msr_write_hook, NULL); @@ -415,10 +406,24 @@ int v3_init_time_vm(struct v3_vm_info * vm) { return ret; } - PrintDebug("Registering TIME_CPUFREQ hypercall.\n"); + PrintDebug(vm, VCORE_NONE, "Registering TIME_CPUFREQ hypercall.\n"); ret = v3_register_hypercall(vm, TIME_CPUFREQ_HCALL, handle_cpufreq_hcall, NULL); + if (ret!=0) { + return ret; + } + + PrintDebug(vm, VCORE_NONE, "Registering TIME_RDHTSC hypercall.\n"); + ret = v3_register_hypercall(vm, TIME_RDHTSC_HCALL, + handle_rdhtsc_hcall, NULL); + + if (ret!=0) { + return ret; + } + + handle_time_configuration(vm, v3_cfg_subtree(cfg_tree, "time")); + return ret; } @@ -429,8 +434,39 @@ void v3_deinit_time_vm(struct v3_vm_info * vm) { v3_remove_hypercall(vm, TIME_CPUFREQ_HCALL); } +static uint32_t +gcd ( uint32_t a, uint32_t b ) +{ + uint32_t c; + while ( a != 0 ) { + c = a; a = b%a; b = c; + } + return b; +} + +static int compute_core_ratios(struct guest_info * info, + uint32_t hostKhz, uint32_t guestKhz) +{ + struct vm_core_time * time_state = &(info->time_state); + uint32_t khzGCD; + + /* Compute these using the GCD() of the guest and host CPU freq. + * If the GCD is too small, make it "big enough" */ + khzGCD = gcd(hostKhz, guestKhz); + if (khzGCD < 1024) + khzGCD = 1024; + + time_state->clock_ratio_num = guestKhz / khzGCD; + time_state->clock_ratio_denom = hostKhz / khzGCD; + + time_state->ipc_ratio_num = 1; + time_state->ipc_ratio_denom = 1; + + return 0; +} + void v3_init_time_core(struct guest_info * info) { - struct vm_time * time_state = &(info->time_state); + struct vm_core_time * time_state = &(info->time_state); v3_cfg_tree_t * cfg_tree = info->core_cfg_data; char * khz = NULL; @@ -439,42 +475,73 @@ void v3_init_time_core(struct guest_info * info) { if (khz) { time_state->guest_cpu_freq = atoi(khz); - PrintDebug("Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n", + PrintDebug(info->vm_info, info, "Logical Core %d (vcpu=%d) CPU frequency requested at %d khz.\n", info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq); } if ( (khz == NULL) || - (time_state->guest_cpu_freq <= 0) || - (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) { - + (time_state->guest_cpu_freq <= 0)) { +/* || (time_state->guest_cpu_freq > time_state->host_cpu_freq) ) { */ time_state->guest_cpu_freq = time_state->host_cpu_freq; } + compute_core_ratios(info, time_state->host_cpu_freq, + time_state->guest_cpu_freq); + + time_state->flags = 0; + if (info->vm_info->time_state.flags & V3_TIME_SLAVE_HOST) { + time_state->flags |= VM_TIME_SLAVE_HOST; + } + if (info->vm_info->time_state.flags & V3_TIME_TSC_PASSTHROUGH) { + time_state->flags |= VM_TIME_TSC_PASSTHROUGH; + } - PrintDebug("Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n", + if ((time_state->clock_ratio_denom != 1) || + (time_state->clock_ratio_num != 1) || + (info->vm_info->time_state.td_num != 1) || + (info->vm_info->time_state.td_denom != 1)) { + if (time_state->flags | VM_TIME_TSC_PASSTHROUGH) { + PrintError(info->vm_info, info, "WARNING: Cannot use reqested passthrough TSC with clock or time modification also requested.\n"); + time_state->flags &= ~VM_TIME_TSC_PASSTHROUGH; + } + time_state->flags |= VM_TIME_TRAP_RDTSC; + } + + PrintDebug(info->vm_info, info, "Logical Core %d (vcpu=%d) CPU frequency set to %d KHz (host CPU frequency = %d KHz).\n", info->pcpu_id, info->vcpu_id, time_state->guest_cpu_freq, time_state->host_cpu_freq); - - time_state->initial_time = 0; - time_state->last_update = 0; - time_state->guest_host_offset = 0; + PrintDebug(info->vm_info, info, " td_mult = %d/%d, cl_mult = %u/%u, ipc_mult = %u/%u.\n", + info->vm_info->time_state.td_num, + info->vm_info->time_state.td_denom, + time_state->clock_ratio_num, time_state->clock_ratio_denom, + time_state->ipc_ratio_num, time_state->ipc_ratio_denom); + PrintDebug(info->vm_info, info, " time source = %s, tsc handling = %s\n", + (time_state->flags & VM_TIME_SLAVE_HOST) ? "host" : "none", + (time_state->flags & VM_TIME_TSC_PASSTHROUGH) ? "passthrough" + : (time_state->flags & VM_TIME_TRAP_RDTSC) ? "trapping" + : "offsettting"); + + time_state->guest_cycles = 0; time_state->tsc_guest_offset = 0; + time_state->last_update = 0; + time_state->initial_host_time = 0; INIT_LIST_HEAD(&(time_state->timers)); time_state->num_timers = 0; - + time_state->tsc_aux.lo = 0; time_state->tsc_aux.hi = 0; } void v3_deinit_time_core(struct guest_info * core) { - struct vm_time * time_state = &(core->time_state); + struct vm_core_time * time_state = &(core->time_state); struct v3_timer * tmr = NULL; struct v3_timer * tmp = NULL; - list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) { - v3_remove_timer(core, tmr); + if (*(void**)&time_state->timers) { + list_for_each_entry_safe(tmr, tmp, &(time_state->timers), timer_link) { + v3_remove_timer(core, tmr); + } } - }