From: Evgeny Pinchuk evgeny.pinchuk@amd.com
Introducing IOCTL implementation for controlling exclusive access to performace counters. The exclusive access is per GPU device.
Signed-off-by: Evgeny Pinchuk evgeny.pinchuk@amd.com Signed-off-by: Oded Gabbay oded.gabbay@amd.com --- 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)