aboutsummaryrefslogtreecommitdiff
path: root/sys/amd64/vmm/vmm.c
diff options
context:
space:
mode:
Diffstat (limited to 'sys/amd64/vmm/vmm.c')
-rw-r--r--sys/amd64/vmm/vmm.c177
1 files changed, 148 insertions, 29 deletions
diff --git a/sys/amd64/vmm/vmm.c b/sys/amd64/vmm/vmm.c
index 2ac076551165..f7c59847140b 100644
--- a/sys/amd64/vmm/vmm.c
+++ b/sys/amd64/vmm/vmm.c
@@ -562,9 +562,9 @@ vm_alloc_vcpu(struct vm *vm, int vcpuid)
}
void
-vm_slock_vcpus(struct vm *vm)
+vm_lock_vcpus(struct vm *vm)
{
- sx_slock(&vm->vcpus_init_lock);
+ sx_xlock(&vm->vcpus_init_lock);
}
void
@@ -990,6 +990,54 @@ save_guest_fpustate(struct vcpu *vcpu)
static VMM_STAT(VCPU_IDLE_TICKS, "number of ticks vcpu was idle");
+/*
+ * Invoke the rendezvous function on the specified vcpu if applicable. Return
+ * true if the rendezvous is finished, false otherwise.
+ */
+static bool
+vm_rendezvous(struct vcpu *vcpu)
+{
+ struct vm *vm = vcpu->vm;
+ int vcpuid;
+
+ mtx_assert(&vcpu->vm->rendezvous_mtx, MA_OWNED);
+ KASSERT(vcpu->vm->rendezvous_func != NULL,
+ ("vm_rendezvous: no rendezvous pending"));
+
+ /* 'rendezvous_req_cpus' must be a subset of 'active_cpus' */
+ CPU_AND(&vm->rendezvous_req_cpus, &vm->rendezvous_req_cpus,
+ &vm->active_cpus);
+
+ vcpuid = vcpu->vcpuid;
+ if (CPU_ISSET(vcpuid, &vm->rendezvous_req_cpus) &&
+ !CPU_ISSET(vcpuid, &vm->rendezvous_done_cpus)) {
+ VMM_CTR0(vcpu, "Calling rendezvous func");
+ (*vm->rendezvous_func)(vcpu, vm->rendezvous_arg);
+ CPU_SET(vcpuid, &vm->rendezvous_done_cpus);
+ }
+ if (CPU_CMP(&vm->rendezvous_req_cpus,
+ &vm->rendezvous_done_cpus) == 0) {
+ VMM_CTR0(vcpu, "Rendezvous completed");
+ CPU_ZERO(&vm->rendezvous_req_cpus);
+ vm->rendezvous_func = NULL;
+ wakeup(&vm->rendezvous_func);
+ return (true);
+ }
+ return (false);
+}
+
+static void
+vcpu_wait_idle(struct vcpu *vcpu)
+{
+ KASSERT(vcpu->state != VCPU_IDLE, ("vcpu already idle"));
+
+ vcpu->reqidle = 1;
+ vcpu_notify_event_locked(vcpu, false);
+ VMM_CTR1(vcpu, "vcpu state change from %s to "
+ "idle requested", vcpu_state2str(vcpu->state));
+ msleep_spin(&vcpu->state, &vcpu->mtx, "vmstat", hz);
+}
+
static int
vcpu_set_state_locked(struct vcpu *vcpu, enum vcpu_state newstate,
bool from_idle)
@@ -1004,13 +1052,8 @@ vcpu_set_state_locked(struct vcpu *vcpu, enum vcpu_state newstate,
* ioctl() operating on a vcpu at any point.
*/
if (from_idle) {
- while (vcpu->state != VCPU_IDLE) {
- vcpu->reqidle = 1;
- vcpu_notify_event_locked(vcpu, false);
- VMM_CTR1(vcpu, "vcpu state change from %s to "
- "idle requested", vcpu_state2str(vcpu->state));
- msleep_spin(&vcpu->state, &vcpu->mtx, "vmstat", hz);
- }
+ while (vcpu->state != VCPU_IDLE)
+ vcpu_wait_idle(vcpu);
} else {
KASSERT(vcpu->state != VCPU_IDLE, ("invalid transition from "
"vcpu idle state"));
@@ -1062,6 +1105,95 @@ vcpu_set_state_locked(struct vcpu *vcpu, enum vcpu_state newstate,
return (0);
}
+/*
+ * Try to lock all of the vCPUs in the VM while taking care to avoid deadlocks
+ * with vm_smp_rendezvous().
+ *
+ * The complexity here suggests that the rendezvous mechanism needs a rethink.
+ */
+int
+vcpu_set_state_all(struct vm *vm, enum vcpu_state newstate)
+{
+ cpuset_t locked;
+ struct vcpu *vcpu;
+ int error, i;
+ uint16_t maxcpus;
+
+ KASSERT(newstate != VCPU_IDLE,
+ ("vcpu_set_state_all: invalid target state %d", newstate));
+
+ error = 0;
+ CPU_ZERO(&locked);
+ maxcpus = vm->maxcpus;
+
+ mtx_lock(&vm->rendezvous_mtx);
+restart:
+ if (vm->rendezvous_func != NULL) {
+ /*
+ * If we have a pending rendezvous, then the initiator may be
+ * blocked waiting for other vCPUs to execute the callback. The
+ * current thread may be a vCPU thread so we must not block
+ * waiting for the initiator, otherwise we get a deadlock.
+ * Thus, execute the callback on behalf of any idle vCPUs.
+ */
+ for (i = 0; i < maxcpus; i++) {
+ vcpu = vm_vcpu(vm, i);
+ if (vcpu == NULL)
+ continue;
+ vcpu_lock(vcpu);
+ if (vcpu->state == VCPU_IDLE) {
+ (void)vcpu_set_state_locked(vcpu, VCPU_FROZEN,
+ true);
+ CPU_SET(i, &locked);
+ }
+ if (CPU_ISSET(i, &locked)) {
+ /*
+ * We can safely execute the callback on this
+ * vCPU's behalf.
+ */
+ vcpu_unlock(vcpu);
+ (void)vm_rendezvous(vcpu);
+ vcpu_lock(vcpu);
+ }
+ vcpu_unlock(vcpu);
+ }
+ }
+
+ /*
+ * Now wait for remaining vCPUs to become idle. This may include the
+ * initiator of a rendezvous that is currently blocked on the rendezvous
+ * mutex.
+ */
+ CPU_FOREACH_ISCLR(i, &locked) {
+ if (i >= maxcpus)
+ break;
+ vcpu = vm_vcpu(vm, i);
+ if (vcpu == NULL)
+ continue;
+ vcpu_lock(vcpu);
+ while (vcpu->state != VCPU_IDLE) {
+ mtx_unlock(&vm->rendezvous_mtx);
+ vcpu_wait_idle(vcpu);
+ vcpu_unlock(vcpu);
+ mtx_lock(&vm->rendezvous_mtx);
+ if (vm->rendezvous_func != NULL)
+ goto restart;
+ vcpu_lock(vcpu);
+ }
+ error = vcpu_set_state_locked(vcpu, newstate, true);
+ vcpu_unlock(vcpu);
+ if (error != 0) {
+ /* Roll back state changes. */
+ CPU_FOREACH_ISSET(i, &locked)
+ (void)vcpu_set_state(vcpu, VCPU_IDLE, false);
+ break;
+ }
+ CPU_SET(i, &locked);
+ }
+ mtx_unlock(&vm->rendezvous_mtx);
+ return (error);
+}
+
static void
vcpu_require_state(struct vcpu *vcpu, enum vcpu_state newstate)
{
@@ -1083,36 +1215,23 @@ vcpu_require_state_locked(struct vcpu *vcpu, enum vcpu_state newstate)
static int
vm_handle_rendezvous(struct vcpu *vcpu)
{
- struct vm *vm = vcpu->vm;
+ struct vm *vm;
struct thread *td;
- int error, vcpuid;
- error = 0;
- vcpuid = vcpu->vcpuid;
td = curthread;
+ vm = vcpu->vm;
+
mtx_lock(&vm->rendezvous_mtx);
while (vm->rendezvous_func != NULL) {
- /* 'rendezvous_req_cpus' must be a subset of 'active_cpus' */
- CPU_AND(&vm->rendezvous_req_cpus, &vm->rendezvous_req_cpus, &vm->active_cpus);
-
- if (CPU_ISSET(vcpuid, &vm->rendezvous_req_cpus) &&
- !CPU_ISSET(vcpuid, &vm->rendezvous_done_cpus)) {
- VMM_CTR0(vcpu, "Calling rendezvous func");
- (*vm->rendezvous_func)(vcpu, vm->rendezvous_arg);
- CPU_SET(vcpuid, &vm->rendezvous_done_cpus);
- }
- if (CPU_CMP(&vm->rendezvous_req_cpus,
- &vm->rendezvous_done_cpus) == 0) {
- VMM_CTR0(vcpu, "Rendezvous completed");
- CPU_ZERO(&vm->rendezvous_req_cpus);
- vm->rendezvous_func = NULL;
- wakeup(&vm->rendezvous_func);
+ if (vm_rendezvous(vcpu))
break;
- }
+
VMM_CTR0(vcpu, "Wait for rendezvous completion");
mtx_sleep(&vm->rendezvous_func, &vm->rendezvous_mtx, 0,
"vmrndv", hz);
if (td_ast_pending(td, TDA_SUSPEND)) {
+ int error;
+
mtx_unlock(&vm->rendezvous_mtx);
error = thread_check_susp(td, true);
if (error != 0)
generated by cgit v1.2.3 (git 2.25.1) at 2025-10-18 21:53:49 +0000