drm/amdkfd: Rework kfd_locked handling

Currently, even if kfd_locked is set, a process is first
created and then removed to work around a race condition
in updating kfd_locked flag. Rework kfd_locked handling to
ensure no processes is created if kfd_locked is set. This
is achieved by updating kfd_locked under kfd_processes_mutex.
With this there is no need for kfd_locked to be an atomic
counter. Instead, it can be a regular integer.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c

@@ -146,13 +146,6 @@ static int kfd_open(struct inode *inode, struct file *filep)
 	if (IS_ERR(process))
 		return PTR_ERR(process);
 
-	if (kfd_is_locked()) {
-		dev_dbg(kfd_device, "kfd is locked!\n"
-				"process %d unreferenced", process->pasid);
-		kfd_unref_process(process);
-		return -EAGAIN;
-	}
-
 	/* filep now owns the reference returned by kfd_create_process */
 	filep->private_data = process;
 

drivers/gpu/drm/amd/amdkfd/kfd_device.c

@@ -42,7 +42,7 @@
  * once locked, kfd driver will stop any further GPU execution.
  * create process (open) will return -EAGAIN.
  */
-static atomic_t kfd_locked = ATOMIC_INIT(0);
+static int kfd_locked;
 
 #ifdef CONFIG_DRM_AMDGPU_CIK
 extern const struct kfd2kgd_calls gfx_v7_kfd2kgd;
@@ -880,7 +880,9 @@ int kgd2kfd_post_reset(struct kfd_dev *kfd)
 			return ret;
 	}
 
-	atomic_dec(&kfd_locked);
+	mutex_lock(&kfd_processes_mutex);
+	--kfd_locked;
+	mutex_unlock(&kfd_processes_mutex);
 
 	for (i = 0; i < kfd->num_nodes; i++) {
 		node = kfd->nodes[i];
@@ -893,21 +895,27 @@ int kgd2kfd_post_reset(struct kfd_dev *kfd)
 
 bool kfd_is_locked(void)
 {
-	return  (atomic_read(&kfd_locked) > 0);
+	lockdep_assert_held(&kfd_processes_mutex);
+	return  (kfd_locked > 0);
 }
 
 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
 {
 	struct kfd_node *node;
 	int i;
+	int count;
 
 	if (!kfd->init_complete)
 		return;
 
 	/* for runtime suspend, skip locking kfd */
 	if (!run_pm) {
+		mutex_lock(&kfd_processes_mutex);
+		count = ++kfd_locked;
+		mutex_unlock(&kfd_processes_mutex);
+
 		/* For first KFD device suspend all the KFD processes */
-		if (atomic_inc_return(&kfd_locked) == 1)
+		if (count == 1)
 			kfd_suspend_all_processes();
 	}
 
@@ -933,7 +941,10 @@ int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
 
 	/* for runtime resume, skip unlocking kfd */
 	if (!run_pm) {
-		count = atomic_dec_return(&kfd_locked);
+		mutex_lock(&kfd_processes_mutex);
+		count = --kfd_locked;
+		mutex_unlock(&kfd_processes_mutex);
+
 		WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
 		if (count == 0)
 			ret = kfd_resume_all_processes();

drivers/gpu/drm/amd/amdkfd/kfd_priv.h

@@ -201,6 +201,8 @@ extern int amdgpu_no_queue_eviction_on_vm_fault;
 /* Enable eviction debug messages */
 extern bool debug_evictions;
 
+extern struct mutex kfd_processes_mutex;
+
 enum cache_policy {
 	cache_policy_coherent,
 	cache_policy_noncoherent

drivers/gpu/drm/amd/amdkfd/kfd_process.c

@@ -50,7 +50,7 @@ struct mm_struct;
  * Unique/indexed by mm_struct*
  */
 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
-static DEFINE_MUTEX(kfd_processes_mutex);
+DEFINE_MUTEX(kfd_processes_mutex);
 
 DEFINE_SRCU(kfd_processes_srcu);
 
@@ -818,6 +818,12 @@ struct kfd_process *kfd_create_process(struct file *filep)
 	 */
 	mutex_lock(&kfd_processes_mutex);
 
+	if (kfd_is_locked()) {
+		mutex_unlock(&kfd_processes_mutex);
+		pr_debug("KFD is locked! Cannot create process");
+		return ERR_PTR(-EINVAL);
+	}
+
 	/* A prior open of /dev/kfd could have already created the process. */
 	process = find_process(thread, false);
 	if (process) {