11 files changed, 247 insertions, 34 deletions

diff --git a/usr/src/uts/intel/io/vmm/io/vatpit.c b/usr/src/uts/intel/io/vmm/io/vatpit.c
index b6218935db..90b10507b3 100644
--- a/usr/src/uts/intel/io/vmm/io/vatpit.c
+++ b/usr/src/uts/intel/io/vmm/io/vatpit.c
@@ -58,6 +58,7 @@ __FBSDID("$FreeBSD$");
 
 #define	VATPIT_LOCK(vatpit)		mutex_enter(&((vatpit)->lock))
 #define	VATPIT_UNLOCK(vatpit)		mutex_exit(&((vatpit)->lock))
+#define	VATPIT_LOCKED(vatpit)		MUTEX_HELD(&((vatpit)->lock))
 
 #define	TIMER_SEL_MASK		0xc0
 #define	TIMER_RW_MASK		0x30
@@ -170,6 +171,13 @@ vatpit_callout_handler(void *a)
 
 	if (c->mode == TIMER_RATEGEN || c->mode == TIMER_SQWAVE) {
 		pit_timer_start_cntr0(vatpit);
+	} else {
+		/*
+		 * For non-periodic timers, clear the time target to distinguish
+		 * between a fired timer (thus a zero value) and a pending one
+		 * awaiting VM resumption (holding a non-zero value).
+		 */
+		c->time_target = 0;
 	}
 
 	(void) vatpic_pulse_irq(vatpit->vm, 0);
@@ -180,6 +188,16 @@ done:
 }
 
 static void
+vatpit_callout_reset(struct vatpit *vatpit)
+{
+	struct channel *c = &vatpit->channel[0];
+
+	ASSERT(VATPIT_LOCKED(vatpit));
+	callout_reset_hrtime(&c->callout, c->time_target,
+	    vatpit_callout_handler, &c->callout_arg, C_ABSOLUTE);
+}
+
+static void
 pit_timer_start_cntr0(struct vatpit *vatpit)
 {
 	struct channel *c = &vatpit->channel[0];
@@ -206,8 +224,7 @@ pit_timer_start_cntr0(struct vatpit *vatpit)
 		    hrt_freq_interval(PIT_8254_FREQ, c->total_target);
 	}
 
-	callout_reset_hrtime(&c->callout, c->time_target,
-	    vatpit_callout_handler, &c->callout_arg, C_ABSOLUTE);
+	vatpit_callout_reset(vatpit);
 }
 
 static uint16_t
@@ -497,6 +514,29 @@ vatpit_localize_resources(struct vatpit *vatpit)
 	}
 }
 
+void
+vatpit_pause(struct vatpit *vatpit)
+{
+	struct channel *c = &vatpit->channel[0];
+
+	VATPIT_LOCK(vatpit);
+	callout_stop(&c->callout);
+	VATPIT_UNLOCK(vatpit);
+}
+
+void
+vatpit_resume(struct vatpit *vatpit)
+{
+	struct channel *c = &vatpit->channel[0];
+
+	VATPIT_LOCK(vatpit);
+	ASSERT(!callout_active(&c->callout));
+	if (c->time_target != 0) {
+		vatpit_callout_reset(vatpit);
+	}
+	VATPIT_UNLOCK(vatpit);
+}
+
 static int
 vatpit_data_read(void *datap, const vmm_data_req_t *req)
 {
@@ -526,7 +566,7 @@ vatpit_data_read(void *datap, const vmm_data_req_t *req)
 		    (src->ol_sel ? (1 << 3) : 0) |
 		    (src->fr_sel ? (1 << 4) : 0);
 		/* Only channel 0 has the timer configured */
-		if (i == 0) {
+		if (i == 0 && src->time_target != 0) {
 			chan->vac_time_target =
 			    vm_normalize_hrtime(vatpit->vm, src->time_target);
 		} else {
@@ -605,8 +645,10 @@ vatpit_data_write(void *datap, const vmm_data_req_t *req)
 		out->time_target = time_target;
 		out->time_loaded = time_target -
 		    hrt_freq_interval(PIT_8254_FREQ, out->initial);
-		callout_reset_hrtime(callout, out->time_target,
-		    vatpit_callout_handler, &out->callout_arg, C_ABSOLUTE);
+
+		if (!vm_is_paused(vatpit->vm)) {
+			vatpit_callout_reset(vatpit);
+		}
 	}
 	VATPIT_UNLOCK(vatpit);
 
diff --git a/usr/src/uts/intel/io/vmm/io/vatpit.h b/usr/src/uts/intel/io/vmm/io/vatpit.h
index bee3a88293..ffa1587cc9 100644
--- a/usr/src/uts/intel/io/vmm/io/vatpit.h
+++ b/usr/src/uts/intel/io/vmm/io/vatpit.h
@@ -28,6 +28,9 @@
  *
  * $FreeBSD$
  */
+/*
+ * Copyright 2022 Oxide Computer Company
+ */
 
 #ifndef _VATPIT_H_
 #define	_VATPIT_H_
@@ -45,5 +48,7 @@ int vatpit_nmisc_handler(void *arg, bool in, uint16_t port, uint8_t bytes,
     uint32_t *eax);
 
 void vatpit_localize_resources(struct vatpit *);
+void vatpit_pause(struct vatpit *);
+void vatpit_resume(struct vatpit *);
 
 #endif	/* _VATPIT_H_ */
diff --git a/usr/src/uts/intel/io/vmm/io/vhpet.c b/usr/src/uts/intel/io/vmm/io/vhpet.c
index d946dbe691..c34714c740 100644
--- a/usr/src/uts/intel/io/vmm/io/vhpet.c
+++ b/usr/src/uts/intel/io/vmm/io/vhpet.c
@@ -31,6 +31,7 @@
 
 /*
  * Copyright 2018 Joyent, Inc.
+ * Copyright 2022 Oxide Computer Company
  */
 
 #include <sys/cdefs.h>
@@ -96,6 +97,7 @@ struct vhpet {
 
 #define	VHPET_LOCK(vhp)		mutex_enter(&((vhp)->lock))
 #define	VHPET_UNLOCK(vhp)	mutex_exit(&((vhp)->lock))
+#define	VHPET_LOCKED(vhp)	MUTEX_HELD(&((vhp)->lock))
 
 static void vhpet_start_timer(struct vhpet *vhpet, int n, uint32_t counter,
     hrtime_t now);
@@ -271,43 +273,45 @@ vhpet_adjust_compval(struct vhpet *vhpet, int n, uint32_t counter)
 }
 
 static void
-vhpet_handler(void *a)
+vhpet_handler(void *arg)
 {
-	int n;
-	uint32_t counter;
-	hrtime_t now;
-	struct vhpet *vhpet;
-	struct callout *callout;
-	struct vhpet_callout_arg *arg;
-
-	arg = a;
-	vhpet = arg->vhpet;
-	n = arg->timer_num;
-	callout = &vhpet->timer[n].callout;
+	const struct vhpet_callout_arg *vca = arg;
+	struct vhpet *vhpet = vca->vhpet;
+	const int n = vca->timer_num;
+	struct callout *callout = &vhpet->timer[n].callout;
 
 	VHPET_LOCK(vhpet);
 
-	if (callout_pending(callout))		/* callout was reset */
-		goto done;
-
-	if (!callout_active(callout))		/* callout was stopped */
-		goto done;
+	if (callout_pending(callout) || !callout_active(callout)) {
+		VHPET_UNLOCK(vhpet);
+		return;
+	}
 
 	callout_deactivate(callout);
+	ASSERT(vhpet_counter_enabled(vhpet));
 
-	if (!vhpet_counter_enabled(vhpet))
-		panic("vhpet(%p) callout with counter disabled", vhpet);
+	if (vhpet_periodic_timer(vhpet, n)) {
+		hrtime_t now;
+		uint32_t counter = vhpet_counter(vhpet, &now);
 
-	counter = vhpet_counter(vhpet, &now);
-	vhpet_start_timer(vhpet, n, counter, now);
+		vhpet_start_timer(vhpet, n, counter, now);
+	} else {
+		/*
+		 * Zero out the expiration time to distinguish a fired timer
+		 * from one which is held due to a VM pause.
+		 */
+		vhpet->timer[n].callout_expire = 0;
+	}
 	vhpet_timer_interrupt(vhpet, n);
-done:
+
 	VHPET_UNLOCK(vhpet);
 }
 
 static void
 vhpet_stop_timer(struct vhpet *vhpet, int n, hrtime_t now)
 {
+	ASSERT(VHPET_LOCKED(vhpet));
+
 	callout_stop(&vhpet->timer[n].callout);
 
 	/*
@@ -320,6 +324,7 @@ vhpet_stop_timer(struct vhpet *vhpet, int n, hrtime_t now)
 	if (vhpet->timer[n].callout_expire < now) {
 		vhpet_timer_interrupt(vhpet, n);
 	}
+	vhpet->timer[n].callout_expire = 0;
 }
 
 static void
@@ -327,6 +332,8 @@ vhpet_start_timer(struct vhpet *vhpet, int n, uint32_t counter, hrtime_t now)
 {
 	struct vhpet_timer *timer = &vhpet->timer[n];
 
+	ASSERT(VHPET_LOCKED(vhpet));
+
 	if (timer->comprate != 0)
 		vhpet_adjust_compval(vhpet, n, counter);
 	else {
@@ -739,6 +746,34 @@ vhpet_localize_resources(struct vhpet *vhpet)
 	}
 }
 
+void
+vhpet_pause(struct vhpet *vhpet)
+{
+	VHPET_LOCK(vhpet);
+	for (uint_t i = 0; i < VHPET_NUM_TIMERS; i++) {
+		struct vhpet_timer *timer = &vhpet->timer[i];
+
+		callout_stop(&timer->callout);
+	}
+	VHPET_UNLOCK(vhpet);
+}
+
+void
+vhpet_resume(struct vhpet *vhpet)
+{
+	VHPET_LOCK(vhpet);
+	for (uint_t i = 0; i < VHPET_NUM_TIMERS; i++) {
+		struct vhpet_timer *timer = &vhpet->timer[i];
+
+		if (timer->callout_expire != 0) {
+			callout_reset_hrtime(&timer->callout,
+			    timer->callout_expire, vhpet_handler,
+			    &timer->arg, C_ABSOLUTE);
+		}
+	}
+	VHPET_UNLOCK(vhpet);
+}
+
 static int
 vhpet_data_read(void *datap, const vmm_data_req_t *req)
 {
@@ -762,7 +797,7 @@ vhpet_data_read(void *datap, const vmm_data_req_t *req)
 		timer_out->vht_msi = timer->msireg;
 		timer_out->vht_comp_val = timer->compval;
 		timer_out->vht_comp_rate = timer->comprate;
-		if (callout_pending(&timer->callout)) {
+		if (timer->callout_expire != 0) {
 			timer_out->vht_time_target =
 			    vm_normalize_hrtime(vhpet->vm,
 			    timer->callout_expire);
@@ -895,10 +930,15 @@ vhpet_data_write(void *datap, const vmm_data_req_t *req)
 		 * from the HPET would come along for the ride.
 		 */
 
-		/* TODO: properly configure timer */
 		if (timer_src->vht_time_target != 0) {
 			timer->callout_expire = vm_denormalize_hrtime(vhpet->vm,
 			    timer_src->vht_time_target);
+
+			if (!vm_is_paused(vhpet->vm)) {
+				callout_reset_hrtime(&timer->callout,
+				    timer->callout_expire, vhpet_handler,
+				    &timer->arg, C_ABSOLUTE);
+			}
 		} else {
 			timer->callout_expire = 0;
 		}
diff --git a/usr/src/uts/intel/io/vmm/io/vhpet.h b/usr/src/uts/intel/io/vmm/io/vhpet.h
index 0ea0a6b15a..8729b44424 100644
--- a/usr/src/uts/intel/io/vmm/io/vhpet.h
+++ b/usr/src/uts/intel/io/vmm/io/vhpet.h
@@ -31,6 +31,7 @@
 
 /*
  * Copyright 2018 Joyent, Inc.
+ * Copyright 2022 Oxide Computer Company
  */
 
 #ifndef _VHPET_H_
@@ -48,5 +49,7 @@ int vhpet_mmio_read(struct vm *vm, int vcpuid, uint64_t gpa, uint64_t *val,
 int vhpet_getcap(struct vm_hpet_cap *cap);
 
 void vhpet_localize_resources(struct vhpet *vhpet);
+void vhpet_pause(struct vhpet *);
+void vhpet_resume(struct vhpet *);
 
 #endif	/* _VHPET_H_ */
diff --git a/usr/src/uts/intel/io/vmm/io/vlapic.c b/usr/src/uts/intel/io/vmm/io/vlapic.c
index e6b5f3be00..d4edbdd53b 100644
--- a/usr/src/uts/intel/io/vmm/io/vlapic.c
+++ b/usr/src/uts/intel/io/vmm/io/vlapic.c
@@ -772,6 +772,14 @@ vlapic_callout_handler(void *arg)
 			vlapic->timer_fire_when += vlapic->timer_period;
 		}
 		vlapic_callout_reset(vlapic);
+	} else {
+		/*
+		 * Clear the target time so that logic can distinguish from a
+		 * timer which has fired (where the value is zero) from one
+		 * which is held pending due to the instance being paused (where
+		 * the value is non-zero, but the callout is not pending).
+		 */
+		vlapic->timer_fire_when = 0;
 	}
 done:
 	VLAPIC_TIMER_UNLOCK(vlapic);
@@ -1708,6 +1716,25 @@ vlapic_localize_resources(struct vlapic *vlapic)
 	vmm_glue_callout_localize(&vlapic->callout);
 }
 
+void
+vlapic_pause(struct vlapic *vlapic)
+{
+	VLAPIC_TIMER_LOCK(vlapic);
+	callout_stop(&vlapic->callout);
+	VLAPIC_TIMER_UNLOCK(vlapic);
+
+}
+
+void
+vlapic_resume(struct vlapic *vlapic)
+{
+	VLAPIC_TIMER_LOCK(vlapic);
+	if (vlapic->timer_fire_when != 0) {
+		vlapic_callout_reset(vlapic);
+	}
+	VLAPIC_TIMER_UNLOCK(vlapic);
+}
+
 static int
 vlapic_data_read(void *datap, const vmm_data_req_t *req)
 {
@@ -1726,7 +1753,7 @@ vlapic_data_read(void *datap, const vmm_data_req_t *req)
 
 	out->vl_msr_apicbase = vlapic->msr_apicbase;
 	out->vl_esr_pending = vlapic->esr_pending;
-	if (callout_pending(&vlapic->callout)) {
+	if (vlapic->timer_fire_when != 0) {
 		out->vl_timer_target =
 		    vm_normalize_hrtime(vlapic->vm, vlapic->timer_fire_when);
 	} else {
@@ -1923,7 +1950,12 @@ vlapic_data_write(void *datap, const vmm_data_req_t *req)
 	if (src->vl_timer_target != 0) {
 		vlapic->timer_fire_when =
 		    vm_denormalize_hrtime(vlapic->vm, src->vl_timer_target);
-		vlapic_callout_reset(vlapic);
+
+		if (!vm_is_paused(vlapic->vm)) {
+			vlapic_callout_reset(vlapic);
+		}
+	} else {
+		vlapic->timer_fire_when = 0;
 	}
 
 	if (vlapic->ops.sync_state) {
diff --git a/usr/src/uts/intel/io/vmm/io/vlapic.h b/usr/src/uts/intel/io/vmm/io/vlapic.h
index 4fe2d79c69..386c9a0922 100644
--- a/usr/src/uts/intel/io/vmm/io/vlapic.h
+++ b/usr/src/uts/intel/io/vmm/io/vlapic.h
@@ -30,7 +30,7 @@
 
 /*
  * Copyright 2018 Joyent, Inc.
- * Copyright 2020 Oxide Computer Company
+ * Copyright 2022 Oxide Computer Company
  */
 
 #ifndef _VLAPIC_H_
@@ -104,5 +104,7 @@ void vlapic_lvt_write_handler(struct vlapic *vlapic, uint32_t offset);
 void vlapic_self_ipi_handler(struct vlapic *vlapic, uint32_t val);
 
 void vlapic_localize_resources(struct vlapic *vlapic);
+void vlapic_pause(struct vlapic *);
+void vlapic_resume(struct vlapic *);
 
 #endif	/* _VLAPIC_H_ */
diff --git a/usr/src/uts/intel/io/vmm/io/vrtc.c b/usr/src/uts/intel/io/vmm/io/vrtc.c
index 644532f077..fd4fafdff6 100644
--- a/usr/src/uts/intel/io/vmm/io/vrtc.c
+++ b/usr/src/uts/intel/io/vmm/io/vrtc.c
@@ -970,6 +970,23 @@ vrtc_localize_resources(struct vrtc *vrtc)
 	vmm_glue_callout_localize(&vrtc->callout);
 }
 
+void
+vrtc_pause(struct vrtc *vrtc)
+{
+	VRTC_LOCK(vrtc);
+	callout_stop(&vrtc->callout);
+	VRTC_UNLOCK(vrtc);
+}
+
+void
+vrtc_resume(struct vrtc *vrtc)
+{
+	VRTC_LOCK(vrtc);
+	ASSERT(!callout_active(&vrtc->callout));
+	vrtc_callout_reset(vrtc, vrtc_freq(vrtc));
+	VRTC_UNLOCK(vrtc);
+}
+
 static int
 vrtc_data_read(void *datap, const vmm_data_req_t *req)
 {
@@ -1020,8 +1037,9 @@ vrtc_data_write(void *datap, const vmm_data_req_t *req)
 	time_t curtime = vrtc_curtime(vrtc, NULL);
 	secs_to_rtc(curtime, vrtc, 1);
 
-	/* Make sure the callout is appropriately scheduled */
-	vrtc_callout_reset(vrtc, vrtc_freq(vrtc));
+	if (!vm_is_paused(vrtc->vm)) {
+		vrtc_callout_reset(vrtc, vrtc_freq(vrtc));
+	}
 
 	VRTC_UNLOCK(vrtc);
 	return (0);
diff --git a/usr/src/uts/intel/io/vmm/io/vrtc.h b/usr/src/uts/intel/io/vmm/io/vrtc.h
index d3140c1308..e7869335eb 100644
--- a/usr/src/uts/intel/io/vmm/io/vrtc.h
+++ b/usr/src/uts/intel/io/vmm/io/vrtc.h
@@ -30,6 +30,7 @@
 
 /*
  * Copyright 2018 Joyent, Inc.
+ * Copyright 2022 Oxide Computer Company
  */
 
 #ifndef _VRTC_H_
@@ -54,5 +55,7 @@ int vrtc_data_handler(void *arg, bool in, uint16_t port, uint8_t bytes,
     uint32_t *val);
 
 void vrtc_localize_resources(struct vrtc *);
+void vrtc_pause(struct vrtc *);
+void vrtc_resume(struct vrtc *);
 
 #endif
diff --git a/usr/src/uts/intel/io/vmm/sys/vmm_kernel.h b/usr/src/uts/intel/io/vmm/sys/vmm_kernel.h
index 290044b438..c5c7d7889e 100644
--- a/usr/src/uts/intel/io/vmm/sys/vmm_kernel.h
+++ b/usr/src/uts/intel/io/vmm/sys/vmm_kernel.h
@@ -131,6 +131,10 @@ void vm_get_topology(struct vm *vm, uint16_t *sockets, uint16_t *cores,
 int vm_set_topology(struct vm *vm, uint16_t sockets, uint16_t cores,
     uint16_t threads, uint16_t maxcpus);
 
+int vm_pause_instance(struct vm *);
+int vm_resume_instance(struct vm *);
+bool vm_is_paused(struct vm *);
+
 /*
  * APIs that race against hardware.
  */
diff --git a/usr/src/uts/intel/io/vmm/vmm.c b/usr/src/uts/intel/io/vmm/vmm.c
index 6f85f13be6..44f1ee5ca2 100644
--- a/usr/src/uts/intel/io/vmm/vmm.c
+++ b/usr/src/uts/intel/io/vmm/vmm.c
@@ -220,6 +220,7 @@ struct vm {
 	struct ioport_config ioports;		/* (o) ioport handling */
 
 	bool		mem_transient;		/* (o) alloc transient memory */
+	bool		is_paused;		/* (i) instance is paused */
 };
 
 static int vmm_initialized;
@@ -728,6 +729,58 @@ vm_reinit(struct vm *vm, uint64_t flags)
 	return (0);
 }
 
+bool
+vm_is_paused(struct vm *vm)
+{
+	return (vm->is_paused);
+}
+
+int
+vm_pause_instance(struct vm *vm)
+{
+	if (vm->is_paused) {
+		return (EALREADY);
+	}
+	vm->is_paused = true;
+
+	for (uint_t i = 0; i < vm->maxcpus; i++) {
+		struct vcpu *vcpu = &vm->vcpu[i];
+
+		if (!CPU_ISSET(i, &vm->active_cpus)) {
+			continue;
+		}
+		vlapic_pause(vcpu->vlapic);
+	}
+	vhpet_pause(vm->vhpet);
+	vatpit_pause(vm->vatpit);
+	vrtc_pause(vm->vrtc);
+
+	return (0);
+}
+
+int
+vm_resume_instance(struct vm *vm)
+{
+	if (!vm->is_paused) {
+		return (EALREADY);
+	}
+	vm->is_paused = false;
+
+	vrtc_resume(vm->vrtc);
+	vatpit_resume(vm->vatpit);
+	vhpet_resume(vm->vhpet);
+	for (uint_t i = 0; i < vm->maxcpus; i++) {
+		struct vcpu *vcpu = &vm->vcpu[i];
+
+		if (!CPU_ISSET(i, &vm->active_cpus)) {
+			continue;
+		}
+		vlapic_resume(vcpu->vlapic);
+	}
+
+	return (0);
+}
+
 int
 vm_map_mmio(struct vm *vm, vm_paddr_t gpa, size_t len, vm_paddr_t hpa)
 {
diff --git a/usr/src/uts/intel/io/vmm/vmm_sol_dev.c b/usr/src/uts/intel/io/vmm/vmm_sol_dev.c
index 882d22b435..c66f1ce17a 100644
--- a/usr/src/uts/intel/io/vmm/vmm_sol_dev.c
+++ b/usr/src/uts/intel/io/vmm/vmm_sol_dev.c
@@ -481,6 +481,8 @@ vmmdev_do_ioctl(vmm_softc_t *sc, int cmd, intptr_t arg, int md,
 	case VM_MUNMAP_MEMSEG:
 	case VM_WRLOCK_CYCLE:
 	case VM_PMTMR_LOCATE:
+	case VM_PAUSE:
+	case VM_RESUME:
 		vmm_write_lock(sc);
 		lock_type = LOCK_WRITE_HOLD;
 		break;
@@ -1844,6 +1846,15 @@ vmmdev_do_ioctl(vmm_softc_t *sc, int cmd, intptr_t arg, int md,
 		break;
 	}
 
+	case VM_PAUSE: {
+		error = vm_pause_instance(sc->vmm_vm);
+		break;
+	}
+	case VM_RESUME: {
+		error = vm_resume_instance(sc->vmm_vm);
+		break;
+	}
+
 	default:
 		error = ENOTTY;
 		break;