From: Evgeny Pinchuk <evgeny.pinchuk@xxxxxxx>
Introducing IOCTL implementation for controlling exclusive access to performace counters.
The exclusive access is per GPU device.
Signed-off-by: Evgeny Pinchuk <evgeny.pinchuk@xxxxxxx>
Signed-off-by: Oded Gabbay <oded.gabbay@xxxxxxx>
---
drivers/gpu/hsa/radeon/kfd_chardev.c | 61 ++++++++++++++++++++++++++++++++++++
drivers/gpu/hsa/radeon/kfd_device.c | 2 ++
drivers/gpu/hsa/radeon/kfd_priv.h | 5 +++
drivers/gpu/hsa/radeon/kfd_process.c | 8 +++--
include/uapi/linux/kfd_ioctl.h | 12 +++++++
5 files changed, 86 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/hsa/radeon/kfd_chardev.c b/drivers/gpu/hsa/radeon/kfd_chardev.c
index 80b702e..b39df68 100644
--- a/drivers/gpu/hsa/radeon/kfd_chardev.c
+++ b/drivers/gpu/hsa/radeon/kfd_chardev.c
@@ -387,6 +387,59 @@ static int kfd_ioctl_get_process_apertures(struct file *filp, struct kfd_process
return 0;
}
+static long
+kfd_ioctl_pmc_acquire_access(struct file *filp, struct kfd_process *p, void __user *arg)
+{
+ struct kfd_ioctl_pmc_acquire_access_args args;
+ struct kfd_dev *dev;
+ int err = -EBUSY;
+
+ if (copy_from_user(&args, arg, sizeof(args)))
+ return -EFAULT;
+
+ dev = radeon_kfd_device_by_id(args.gpu_id);
+ if (dev == NULL)
+ return -EINVAL;
+
+ spin_lock(&dev->pmc_access_lock);
+ if (dev->pmc_locking_process == NULL) {
+ dev->pmc_locking_process = p;
+ dev->pmc_locking_trace = args.trace_id;
+ err = 0;
+ } else if (dev->pmc_locking_process == p && dev->pmc_locking_trace == args.trace_id) {
+ /* Same trace already has an access. Returning success */
+ err = 0;
+ }
+
+ spin_unlock(&dev->pmc_access_lock);
+
+ return err;
+}
+
+static long
+kfd_ioctl_pmc_release_access(struct file *filp, struct kfd_process *p, void __user *arg)
+{
+ struct kfd_ioctl_pmc_release_access_args args;
+ struct kfd_dev *dev;
+ int err = -EINVAL;
+
+ if (copy_from_user(&args, arg, sizeof(args)))
+ return -EFAULT;
+
+ dev = radeon_kfd_device_by_id(args.gpu_id);
+ if (dev == NULL)
+ return -EINVAL;
+
+ spin_lock(&dev->pmc_access_lock);
+ if (dev->pmc_locking_process == p && dev->pmc_locking_trace == args.trace_id) {
+ dev->pmc_locking_process = NULL;
+ dev->pmc_locking_trace = 0;
+ err = 0;
+ }
+ spin_unlock(&dev->pmc_access_lock);
+
+ return err;
+}
static long
kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
@@ -427,6 +480,14 @@ kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
err = kfd_ioctl_update_queue(filep, process, (void __user *)arg);
break;
+ case KFD_IOC_PMC_ACQUIRE_ACCESS:
+ err = kfd_ioctl_pmc_acquire_access(filep, process, (void __user *) arg);
+ break;
+
+ case KFD_IOC_PMC_RELEASE_ACCESS:
+ err = kfd_ioctl_pmc_release_access(filep, process, (void __user *) arg);
+ break;
+
default:
dev_err(kfd_device,
"unknown ioctl cmd 0x%x, arg 0x%lx)\n",
diff --git a/drivers/gpu/hsa/radeon/kfd_device.c b/drivers/gpu/hsa/radeon/kfd_device.c
index c602e16..9af812b 100644
--- a/drivers/gpu/hsa/radeon/kfd_device.c
+++ b/drivers/gpu/hsa/radeon/kfd_device.c
@@ -185,6 +185,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
return false;
}
+ spin_lock_init(&kfd->pmc_access_lock);
+
kfd->init_complete = true;
dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
kfd->pdev->device);
diff --git a/drivers/gpu/hsa/radeon/kfd_priv.h b/drivers/gpu/hsa/radeon/kfd_priv.h
index 049671b..e6d4993 100644
--- a/drivers/gpu/hsa/radeon/kfd_priv.h
+++ b/drivers/gpu/hsa/radeon/kfd_priv.h
@@ -135,6 +135,11 @@ struct kfd_dev {
/* QCM Device instance */
struct device_queue_manager *dqm;
+
+ /* Performance counters exclusivity lock */
+ spinlock_t pmc_access_lock;
+ struct kfd_process *pmc_locking_process;
+ uint64_t pmc_locking_trace;
};
/* KGD2KFD callbacks */
diff --git a/drivers/gpu/hsa/radeon/kfd_process.c b/drivers/gpu/hsa/radeon/kfd_process.c
index f967c15..9bb5cab 100644
--- a/drivers/gpu/hsa/radeon/kfd_process.c
+++ b/drivers/gpu/hsa/radeon/kfd_process.c
@@ -96,9 +96,13 @@ static void free_process(struct kfd_process *p)
BUG_ON(p == NULL);
- /* doorbell mappings: automatic */
-
list_for_each_entry_safe(pdd, temp, &p->per_device_data, per_device_list) {
+ spin_lock(&pdd->dev->pmc_access_lock);
+ if (pdd->dev->pmc_locking_process == p) {
+ pdd->dev->pmc_locking_process = NULL;
+ pdd->dev->pmc_locking_trace = 0;
+ }
+ spin_unlock(&pdd->dev->pmc_access_lock);
amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
list_del(&pdd->per_device_list);
kfree(pdd);
diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index d58231d..509c4a0 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -103,6 +103,16 @@ struct kfd_ioctl_get_process_apertures_args {
uint8_t num_of_nodes; /* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS]*/
};
+struct kfd_ioctl_pmc_acquire_access_args {
+ uint32_t gpu_id; /* to KFD */
+ uint64_t trace_id; /* to KFD */
+};
+
+struct kfd_ioctl_pmc_release_access_args {
+ uint32_t gpu_id; /* to KFD */
+ uint64_t trace_id; /* to KFD */
+};
+
#define KFD_IOC_MAGIC 'K'
#define KFD_IOC_GET_VERSION _IOR(KFD_IOC_MAGIC, 1, struct kfd_ioctl_get_version_args)
@@ -112,6 +122,8 @@ struct kfd_ioctl_get_process_apertures_args {
#define KFD_IOC_GET_CLOCK_COUNTERS _IOWR(KFD_IOC_MAGIC, 5, struct kfd_ioctl_get_clock_counters_args)
#define KFD_IOC_GET_PROCESS_APERTURES _IOR(KFD_IOC_MAGIC, 6, struct kfd_ioctl_get_process_apertures_args)
#define KFD_IOC_UPDATE_QUEUE _IOW(KFD_IOC_MAGIC, 7, struct kfd_ioctl_update_queue_args)
+#define KFD_IOC_PMC_ACQUIRE_ACCESS _IOW(KFD_IOC_MAGIC, 12, struct kfd_ioctl_pmc_acquire_access_args)
+#define KFD_IOC_PMC_RELEASE_ACCESS _IOW(KFD_IOC_MAGIC, 13, struct kfd_ioctl_pmc_release_access_args)
#pragma pack(pop)
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
