[PATCH 13/25] drm/amdkfd: Implement GPU reset handlers in KFD

[Felix Kuehling]

From: Shaoyun Liu <Shaoyun.Liu at amd.com>

Lock KFD and evict existing queues on reset. Notify user mode by
signaling hw_exception events.

Signed-off-by: Shaoyun Liu <Shaoyun.Liu at amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c |  3 +++
 drivers/gpu/drm/amd/amdkfd/kfd_device.c  | 43 +++++++++++++++++++++++++++++---
 drivers/gpu/drm/amd/amdkfd/kfd_events.c  | 27 ++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_events.h  |  1 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h    |  4 +++
 5 files changed, 75 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index f64c555..21d0989 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -122,6 +122,9 @@ static int kfd_open(struct inode *inode, struct file *filep)
 	if (IS_ERR(process))
 		return PTR_ERR(process);
 
+	if (kfd_is_locked())
+		return -EAGAIN;
+
 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
 		process->pasid, process->is_32bit_user_mode);
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index a8226d8..9f63ac3 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -30,7 +30,13 @@
 #include "kfd_iommu.h"
 
 #define MQD_SIZE_ALIGNED 768
-static atomic_t kfd_device_suspended = ATOMIC_INIT(0);
+
+/*
+ * kfd_locked is used to lock the kfd driver during suspend or reset
+ * once locked, kfd driver will stop any further GPU execution.
+ * create process (open) will return -EAGAIN.
+ */
+static atomic_t kfd_locked = ATOMIC_INIT(0);
 
 #ifdef KFD_SUPPORT_IOMMU_V2
 static const struct kfd_device_info kaveri_device_info = {
@@ -516,21 +522,52 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd)
 
 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
 {
+	if (!kfd->init_complete)
+		return 0;
+	kgd2kfd_suspend(kfd);
+
+	/* hold dqm->lock to prevent further execution*/
+	dqm_lock(kfd->dqm);
+
+	kfd_signal_reset_event(kfd);
 	return 0;
 }
 
+/*
+ * Fix me. KFD won't be able to resume existing process for now.
+ * We will keep all existing process in a evicted state and
+ * wait the process to be terminated.
+ */
+
 int kgd2kfd_post_reset(struct kfd_dev *kfd)
 {
+	int ret, count;
+
+	if (!kfd->init_complete)
+		return 0;
+
+	dqm_unlock(kfd->dqm);
+
+	ret = kfd_resume(kfd);
+	if (ret)
+		return ret;
+	count = atomic_dec_return(&kfd_locked);
+	WARN_ONCE(count != 0, "KFD reset ref. error");
 	return 0;
 }
 
+bool kfd_is_locked(void)
+{
+	return  (atomic_read(&kfd_locked) > 0);
+}
+
 void kgd2kfd_suspend(struct kfd_dev *kfd)
 {
 	if (!kfd->init_complete)
 		return;
 
 	/* For first KFD device suspend all the KFD processes */
-	if (atomic_inc_return(&kfd_device_suspended) == 1)
+	if (atomic_inc_return(&kfd_locked) == 1)
 		kfd_suspend_all_processes();
 
 	kfd->dqm->ops.stop(kfd->dqm);
@@ -549,7 +586,7 @@ int kgd2kfd_resume(struct kfd_dev *kfd)
 	if (ret)
 		return ret;
 
-	count = atomic_dec_return(&kfd_device_suspended);
+	count = atomic_dec_return(&kfd_locked);
 	WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
 	if (count == 0)
 		ret = kfd_resume_all_processes();
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_events.c
index b58a0e6..820133c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c
@@ -1000,3 +1000,30 @@ void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
 	mutex_unlock(&p->event_mutex);
 	kfd_unref_process(p);
 }
+
+void kfd_signal_reset_event(struct kfd_dev *dev)
+{
+	struct kfd_hsa_hw_exception_data hw_exception_data;
+	struct kfd_process *p;
+	struct kfd_event *ev;
+	unsigned int temp;
+	uint32_t id, idx;
+
+	/* Whole gpu reset caused by GPU hang and memory is lost */
+	memset(&hw_exception_data, 0, sizeof(hw_exception_data));
+	hw_exception_data.gpu_id = dev->id;
+	hw_exception_data.memory_lost = 1;
+
+	idx = srcu_read_lock(&kfd_processes_srcu);
+	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+		mutex_lock(&p->event_mutex);
+		id = KFD_FIRST_NONSIGNAL_EVENT_ID;
+		idr_for_each_entry_continue(&p->event_idr, ev, id)
+			if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
+				ev->hw_exception_data = hw_exception_data;
+				set_event(ev);
+			}
+		mutex_unlock(&p->event_mutex);
+	}
+	srcu_read_unlock(&kfd_processes_srcu, idx);
+}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.h b/drivers/gpu/drm/amd/amdkfd/kfd_events.h
index abca5bf..c7ac6c7 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_events.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.h
@@ -66,6 +66,7 @@ struct kfd_event {
 	/* type specific data */
 	union {
 		struct kfd_hsa_memory_exception_data memory_exception_data;
+		struct kfd_hsa_hw_exception_data hw_exception_data;
 	};
 };
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 4bc8d5a..2e03d6c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -975,10 +975,14 @@ int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
 				struct kfd_vm_fault_info *info);
 
+void kfd_signal_reset_event(struct kfd_dev *dev);
+
 void kfd_flush_tlb(struct kfd_process_device *pdd);
 
 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
 
+bool kfd_is_locked(void);
+
 /* Debugfs */
 #if defined(CONFIG_DEBUG_FS)
 
-- 
2.7.4