[RFC,v6,13/92] kvm: introspection: make the vCPU wait even when its jobs list is empty

Commit Message

Adalbert Lazăr Aug. 9, 2019, 3:59 p.m. UTC

Usually, the vCPU thread will run the functions from its jobs list
(unless the thread is SIGKILL-ed) and continue to guest when the
list is empty. But, there are cases when it has to wait for something
(e.g. another vCPU runs in single-step mode, or the current vCPU waits
for an event reply from the introspection tool).

In these cases, it will append a "wait job" into its own list, which
will do (a) nothing if the list is not empty or it doesn't have to wait
any longer or (b) wait (in the same wake-queue used by KVM) until it
is kicked. It should be OK if the receiving worker appends a new job in
the same time.

Signed-off-by: Adalbert Lazăr <[e-mail]>
---
 include/linux/swait.h | 11 ++++++
 virt/kvm/kvmi.c       | 80 +++++++++++++++++++++++++++++++++++++++++++
 virt/kvm/kvmi_int.h   |  2 ++
 3 files changed, 93 insertions(+)

Comments

Paolo Bonzini Aug. 13, 2019, 8:43 a.m. UTC | #1

On 09/08/19 17:59, Adalbert Lazăr wrote:
> +void kvmi_handle_requests(struct kvm_vcpu *vcpu)
> +{
> +	struct kvmi *ikvm;
> +
> +	ikvm = kvmi_get(vcpu->kvm);
> +	if (!ikvm)
> +		return;
> +
> +	for (;;) {
> +		int err = kvmi_run_jobs_and_wait(vcpu);
> +
> +		if (err)
> +			break;
> +	}
> +
> +	kvmi_put(vcpu->kvm);
> +}
> +

Using kvmi_run_jobs_and_wait from two places (here and kvmi_send_event)
is very confusing.  Does kvmi_handle_requests need to do this, or can it
just use kvmi_run_jobs?

Paolo

Patch

diff --git a/include/linux/swait.h b/include/linux/swait.h
index 73e06e9986d4..2486625e7fb4 100644
--- a/include/linux/swait.h
+++ b/include/linux/swait.h
@@ -297,4 +297,15 @@  do {									\
 	__ret;								\
 })
 
+#define __swait_event_killable(wq, condition)				\
+	___swait_event(wq, condition, TASK_KILLABLE, 0,	schedule())	\
+
+#define swait_event_killable(wq, condition)				\
+({									\
+	int __ret = 0;							\
+	if (!(condition))						\
+		__ret = __swait_event_killable(wq, condition);		\
+	__ret;								\
+})
+
 #endif /* _LINUX_SWAIT_H */
diff --git a/virt/kvm/kvmi.c b/virt/kvm/kvmi.c
index 07ebd1c629b0..3c884dc0e38c 100644
--- a/virt/kvm/kvmi.c
+++ b/virt/kvm/kvmi.c
@@ -135,6 +135,19 @@  static void kvmi_free_job(struct kvmi_job *job)
 	kmem_cache_free(job_cache, job);
 }
 
+static struct kvmi_job *kvmi_pull_job(struct kvmi_vcpu *ivcpu)
+{
+	struct kvmi_job *job = NULL;
+
+	spin_lock(&ivcpu->job_lock);
+	job = list_first_entry_or_null(&ivcpu->job_list, typeof(*job), link);
+	if (job)
+		list_del(&job->link);
+	spin_unlock(&ivcpu->job_lock);
+
+	return job;
+}
+
 static bool alloc_ivcpu(struct kvm_vcpu *vcpu)
 {
 	struct kvmi_vcpu *ivcpu;
@@ -496,6 +509,73 @@  void kvmi_destroy_vm(struct kvm *kvm)
 	wait_for_completion_killable(&kvm->kvmi_completed);
 }
 
+void kvmi_run_jobs(struct kvm_vcpu *vcpu)
+{
+	struct kvmi_vcpu *ivcpu = IVCPU(vcpu);
+	struct kvmi_job *job;
+
+	while ((job = kvmi_pull_job(ivcpu))) {
+		job->fct(vcpu, job->ctx);
+		kvmi_free_job(job);
+	}
+}
+
+static bool done_waiting(struct kvm_vcpu *vcpu)
+{
+	struct kvmi_vcpu *ivcpu = IVCPU(vcpu);
+
+	return !list_empty(&ivcpu->job_list);
+}
+
+static void kvmi_job_wait(struct kvm_vcpu *vcpu, void *ctx)
+{
+	struct swait_queue_head *wq = kvm_arch_vcpu_wq(vcpu);
+	struct kvmi_vcpu *ivcpu = IVCPU(vcpu);
+	int err;
+
+	err = swait_event_killable(*wq, done_waiting(vcpu));
+
+	if (err)
+		ivcpu->killed = true;
+}
+
+int kvmi_run_jobs_and_wait(struct kvm_vcpu *vcpu)
+{
+	struct kvmi_vcpu *ivcpu = IVCPU(vcpu);
+	int err = 0;
+
+	for (;;) {
+		kvmi_run_jobs(vcpu);
+
+		if (ivcpu->killed) {
+			err = -1;
+			break;
+		}
+
+		kvmi_add_job(vcpu, kvmi_job_wait, NULL, NULL);
+	}
+
+	return err;
+}
+
+void kvmi_handle_requests(struct kvm_vcpu *vcpu)
+{
+	struct kvmi *ikvm;
+
+	ikvm = kvmi_get(vcpu->kvm);
+	if (!ikvm)
+		return;
+
+	for (;;) {
+		int err = kvmi_run_jobs_and_wait(vcpu);
+
+		if (err)
+			break;
+	}
+
+	kvmi_put(vcpu->kvm);
+}
+
 int kvmi_cmd_control_vm_events(struct kvmi *ikvm, unsigned int event_id,
 			       bool enable)
 {
diff --git a/virt/kvm/kvmi_int.h b/virt/kvm/kvmi_int.h
index 97f91a568096..47418e9a86f6 100644
--- a/virt/kvm/kvmi_int.h
+++ b/virt/kvm/kvmi_int.h
@@ -85,6 +85,8 @@  struct kvmi_job {
 struct kvmi_vcpu {
 	struct list_head job_list;
 	spinlock_t job_lock;
+
+	bool killed;
 };
 
 #define IKVM(kvm) ((struct kvmi *)((kvm)->kvmi))