[PATCH 1/1] drm/amdkfd: Add IPC API
Bhardwaj, Rajneesh
rajneesh.bhardwaj at amd.com
Tue Jul 14 16:41:51 UTC 2020
Hi Felix
While the detailed review for this is already going on, you might want
to consider below hunk if you happen to send v2.
--- a/drivers/gpu/drm/amd/amdkfd/kfd_ipc.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_ipc.c
@@ -100,11 +100,10 @@ void kfd_ipc_obj_put(struct kfd_ipc_obj **obj)
}
}
-int kfd_ipc_init(void)
+void kfd_ipc_init(void)
{
mutex_init(&kfd_ipc_handles.lock);
hash_init(kfd_ipc_handles.handles);
- return 0;
}
static int kfd_import_dmabuf_create_kfd_bo(struct kfd_dev *dev,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
index dacc131792cf..2dac77cd7b0c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
@@ -57,9 +57,7 @@ static int kfd_init(void)
if (err < 0)
goto err_topology;
- err = kfd_ipc_init();
- if (err < 0)
- goto err_ipc;
+ kfd_ipc_init();
Thanks
Rajneesh
On 7/13/2020 11:14 PM, Felix Kuehling wrote:
> [CAUTION: External Email]
>
> This allows exporting and importing buffers. The API generates handles
> that can be used with the HIP IPC API, i.e. big numbers rather than
> file descriptors.
>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 5 +
> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 56 +++-
> drivers/gpu/drm/amd/amdkfd/Makefile | 3 +-
> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 74 ++---
> drivers/gpu/drm/amd/amdkfd/kfd_ipc.c | 263 ++++++++++++++++++
> drivers/gpu/drm/amd/amdkfd/kfd_ipc.h | 55 ++++
> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 +
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 +
> include/uapi/linux/kfd_ioctl.h | 62 ++++-
> 9 files changed, 488 insertions(+), 40 deletions(-)
> create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_ipc.c
> create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_ipc.h
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> index 3f2b695cf19e..0f8dc4c4f924 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> @@ -49,6 +49,7 @@ struct kfd_bo_va_list {
> struct kgd_mem {
> struct mutex lock;
> struct amdgpu_bo *bo;
> + struct kfd_ipc_obj *ipc_obj;
> struct list_head bo_va_list;
> /* protected by amdkfd_process_info.lock */
> struct ttm_validate_buffer validate_list;
> @@ -240,9 +241,13 @@ int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
>
> int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
> struct dma_buf *dmabuf,
> + struct kfd_ipc_obj *ipc_obj,
> uint64_t va, void *vm,
> struct kgd_mem **mem, uint64_t *size,
> uint64_t *mmap_offset);
> +int amdgpu_amdkfd_gpuvm_export_ipc_obj(struct kgd_dev *kgd, void *vm,
> + struct kgd_mem *mem,
> + struct kfd_ipc_obj **ipc_obj);
>
> void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
> void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index c408936e8f98..cd5f23c0c2ca 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -29,6 +29,7 @@
> #include "amdgpu_vm.h"
> #include "amdgpu_amdkfd.h"
> #include "amdgpu_dma_buf.h"
> +#include "kfd_ipc.h"
> #include <uapi/linux/kfd_ioctl.h>
>
> /* BO flag to indicate a KFD userptr BO */
> @@ -1353,6 +1354,9 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
> *size = 0;
> }
>
> + /* Unreference the ipc_obj if applicable */
> + kfd_ipc_obj_put(&mem->ipc_obj);
> +
> /* Free the BO*/
> drm_gem_object_put_unlocked(&mem->bo->tbo.base);
> mutex_destroy(&mem->lock);
> @@ -1656,6 +1660,7 @@ int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
>
> int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
> struct dma_buf *dma_buf,
> + struct kfd_ipc_obj *ipc_obj,
> uint64_t va, void *vm,
> struct kgd_mem **mem, uint64_t *size,
> uint64_t *mmap_offset)
> @@ -1692,15 +1697,18 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
>
> INIT_LIST_HEAD(&(*mem)->bo_va_list);
> mutex_init(&(*mem)->lock);
> -
> - (*mem)->alloc_flags =
> - ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
> - KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
> - | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
> - | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
> + if (bo->kfd_bo)
> + (*mem)->alloc_flags = bo->kfd_bo->alloc_flags;
> + else
> + (*mem)->alloc_flags =
> + ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
> + KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
> + | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
> + | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
>
> drm_gem_object_get(&bo->tbo.base);
> (*mem)->bo = bo;
> + (*mem)->ipc_obj = ipc_obj;
> (*mem)->va = va;
> (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
> AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
> @@ -1713,6 +1721,42 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
> return 0;
> }
>
> +int amdgpu_amdkfd_gpuvm_export_ipc_obj(struct kgd_dev *kgd, void *vm,
> + struct kgd_mem *mem,
> + struct kfd_ipc_obj **ipc_obj)
> +{
> + struct amdgpu_device *adev = NULL;
> + struct dma_buf *dmabuf;
> + int r = 0;
> +
> + if (!kgd || !vm || !mem)
> + return -EINVAL;
> +
> + adev = get_amdgpu_device(kgd);
> + mutex_lock(&mem->lock);
> +
> + if (mem->ipc_obj) {
> + *ipc_obj = mem->ipc_obj;
> + goto unlock_out;
> + }
> +
> + dmabuf = amdgpu_gem_prime_export(&mem->bo->tbo.base, 0);
> + if (IS_ERR(dmabuf)) {
> + r = PTR_ERR(dmabuf);
> + goto unlock_out;
> + }
> +
> + r = kfd_ipc_store_insert(dmabuf, &mem->ipc_obj);
> + if (r)
> + dma_buf_put(dmabuf);
> + else
> + *ipc_obj = mem->ipc_obj;
> +
> +unlock_out:
> + mutex_unlock(&mem->lock);
> + return r;
> +}
> +
> /* Evict a userptr BO by stopping the queues if necessary
> *
> * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
> diff --git a/drivers/gpu/drm/amd/amdkfd/Makefile b/drivers/gpu/drm/amd/amdkfd/Makefile
> index e1e4115dcf78..815d9a9e7a3c 100644
> --- a/drivers/gpu/drm/amd/amdkfd/Makefile
> +++ b/drivers/gpu/drm/amd/amdkfd/Makefile
> @@ -54,7 +54,8 @@ AMDKFD_FILES := $(AMDKFD_PATH)/kfd_module.o \
> $(AMDKFD_PATH)/kfd_dbgdev.o \
> $(AMDKFD_PATH)/kfd_dbgmgr.o \
> $(AMDKFD_PATH)/kfd_smi_events.o \
> - $(AMDKFD_PATH)/kfd_crat.o
> + $(AMDKFD_PATH)/kfd_crat.o \
> + $(AMDKFD_PATH)/kfd_ipc.o
>
> ifneq ($(CONFIG_AMD_IOMMU_V2),)
> AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> index e9b96ad3d9a5..e7d15fa02b5e 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> @@ -38,6 +38,7 @@
> #include "kfd_priv.h"
> #include "kfd_device_queue_manager.h"
> #include "kfd_dbgmgr.h"
> +#include "kfd_ipc.h"
> #include "amdgpu_amdkfd.h"
> #include "kfd_smi_events.h"
>
> @@ -1691,53 +1692,58 @@ static int kfd_ioctl_import_dmabuf(struct file *filep,
> struct kfd_process *p, void *data)
> {
> struct kfd_ioctl_import_dmabuf_args *args = data;
> - struct kfd_process_device *pdd;
> - struct dma_buf *dmabuf;
> struct kfd_dev *dev;
> - int idr_handle;
> - uint64_t size;
> - void *mem;
> int r;
>
> dev = kfd_device_by_id(args->gpu_id);
> if (!dev)
> return -EINVAL;
>
> - dmabuf = dma_buf_get(args->dmabuf_fd);
> - if (IS_ERR(dmabuf))
> - return PTR_ERR(dmabuf);
> + r = kfd_ipc_import_dmabuf(dev, p, args->dmabuf_fd,
> + args->va_addr, &args->handle, NULL);
> + if (r)
> + pr_err("Failed to import dmabuf\n");
>
> - mutex_lock(&p->mutex);
> + return r;
> +}
>
> - pdd = kfd_bind_process_to_device(dev, p);
> - if (IS_ERR(pdd)) {
> - r = PTR_ERR(pdd);
> - goto err_unlock;
> - }
> +static int kfd_ioctl_ipc_export_handle(struct file *filep,
> + struct kfd_process *p,
> + void *data)
> +{
> + struct kfd_ioctl_ipc_export_handle_args *args = data;
> + struct kfd_dev *dev;
> + int r;
> +
> + dev = kfd_device_by_id(args->gpu_id);
> + if (!dev)
> + return -EINVAL;
>
> - r = amdgpu_amdkfd_gpuvm_import_dmabuf(dev->kgd, dmabuf,
> - args->va_addr, pdd->vm,
> - (struct kgd_mem **)&mem, &size,
> - NULL);
> + r = kfd_ipc_export_as_handle(dev, p, args->handle, args->share_handle);
> if (r)
> - goto err_unlock;
> + pr_err("Failed to export IPC handle\n");
>
> - idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
> - if (idr_handle < 0) {
> - r = -EFAULT;
> - goto err_free;
> - }
> + return r;
> +}
>
> - mutex_unlock(&p->mutex);
> +static int kfd_ioctl_ipc_import_handle(struct file *filep,
> + struct kfd_process *p,
> + void *data)
> +{
> + struct kfd_ioctl_ipc_import_handle_args *args = data;
> + struct kfd_dev *dev = NULL;
> + int r;
>
> - args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
> + dev = kfd_device_by_id(args->gpu_id);
> + if (!dev)
> + return -EINVAL;
>
> - return 0;
> + r = kfd_ipc_import_handle(p, args->share_handle, args->va_addr,
> + &args->handle, &args->gpu_id,
> + &args->mmap_offset);
> + if (r)
> + pr_err("Failed to import IPC handle\n");
>
> -err_free:
> - amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL);
> -err_unlock:
> - mutex_unlock(&p->mutex);
> return r;
> }
>
> @@ -1853,6 +1859,12 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
>
> AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
> kfd_ioctl_smi_events, 0),
> +
> + AMDKFD_IOCTL_DEF(AMDKFD_IOC_IPC_IMPORT_HANDLE,
> + kfd_ioctl_ipc_import_handle, 0),
> +
> + AMDKFD_IOCTL_DEF(AMDKFD_IOC_IPC_EXPORT_HANDLE,
> + kfd_ioctl_ipc_export_handle, 0),
> };
>
> #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_ipc.c b/drivers/gpu/drm/amd/amdkfd/kfd_ipc.c
> new file mode 100644
> index 000000000000..3de8d7826f07
> --- /dev/null
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_ipc.c
> @@ -0,0 +1,263 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/slab.h>
> +#include <linux/random.h>
> +
> +#include "kfd_ipc.h"
> +#include "kfd_priv.h"
> +#include "amdgpu_amdkfd.h"
> +
> +#define KFD_IPC_HASH_TABLE_SIZE_SHIFT 4
> +#define KFD_IPC_HASH_TABLE_SIZE_MASK ((1 << KFD_IPC_HASH_TABLE_SIZE_SHIFT) - 1)
> +
> +static struct kfd_ipc_handles {
> + DECLARE_HASHTABLE(handles, KFD_IPC_HASH_TABLE_SIZE_SHIFT);
> + struct mutex lock;
> +} kfd_ipc_handles;
> +
> +/* Since, handles are random numbers, it can be used directly as hashing key.
> + * The least 4 bits of the handle are used as key. However, during import all
> + * 128 bits of the handle are checked to prevent handle snooping.
> + */
> +#define HANDLE_TO_KEY(sh) ((*(uint64_t *)sh) & KFD_IPC_HASH_TABLE_SIZE_MASK)
> +
> +int kfd_ipc_store_insert(struct dma_buf *dmabuf, struct kfd_ipc_obj **ipc_obj)
> +{
> + struct kfd_ipc_obj *obj;
> +
> + obj = kmalloc(sizeof(*obj), GFP_KERNEL);
> + if (!obj)
> + return -ENOMEM;
> +
> + /* The initial ref belongs to the allocator process.
> + * The IPC object store itself does not hold a ref since
> + * there is no specific moment in time where that ref should
> + * be dropped, except "when there are no more userspace processes
> + * holding a ref to the object". Therefore the removal from IPC
> + * storage happens at ipc_obj release time.
> + */
> + kref_init(&obj->ref);
> + obj->dmabuf = dmabuf;
> + get_random_bytes(obj->share_handle, sizeof(obj->share_handle));
> +
> + mutex_lock(&kfd_ipc_handles.lock);
> + hlist_add_head(&obj->node,
> + &kfd_ipc_handles.handles[HANDLE_TO_KEY(obj->share_handle)]);
> + mutex_unlock(&kfd_ipc_handles.lock);
> +
> + if (ipc_obj)
> + *ipc_obj = obj;
> +
> + return 0;
> +}
> +
> +static void ipc_obj_release(struct kref *r)
> +{
> + struct kfd_ipc_obj *obj;
> +
> + obj = container_of(r, struct kfd_ipc_obj, ref);
> +
> + mutex_lock(&kfd_ipc_handles.lock);
> + hash_del(&obj->node);
> + mutex_unlock(&kfd_ipc_handles.lock);
> +
> + dma_buf_put(obj->dmabuf);
> + kfree(obj);
> +}
> +
> +static struct kfd_ipc_obj *ipc_obj_get(struct kfd_ipc_obj *obj)
> +{
> + if (kref_get_unless_zero(&obj->ref))
> + return obj;
> + return NULL;
> +}
> +
> +void kfd_ipc_obj_put(struct kfd_ipc_obj **obj)
> +{
> + if (*obj) {
> + kref_put(&(*obj)->ref, ipc_obj_release);
> + *obj = NULL;
> + }
> +}
> +
> +int kfd_ipc_init(void)
> +{
> + mutex_init(&kfd_ipc_handles.lock);
> + hash_init(kfd_ipc_handles.handles);
> + return 0;
> +}
> +
> +static int kfd_import_dmabuf_create_kfd_bo(struct kfd_dev *dev,
> + struct kfd_process *p,
> + struct dma_buf *dmabuf,
> + struct kfd_ipc_obj *ipc_obj,
> + uint64_t va_addr, uint64_t *handle,
> + uint64_t *mmap_offset)
> +{
> + int r;
> + void *mem;
> + uint64_t size;
> + int idr_handle;
> + struct kfd_process_device *pdd = NULL;
> +
> + if (!handle)
> + return -EINVAL;
> +
> + if (!dev)
> + return -EINVAL;
> +
> + mutex_lock(&p->mutex);
> +
> + pdd = kfd_bind_process_to_device(dev, p);
> + if (IS_ERR(pdd)) {
> + r = PTR_ERR(pdd);
> + goto err_unlock;
> + }
> +
> + r = amdgpu_amdkfd_gpuvm_import_dmabuf(dev->kgd, dmabuf, ipc_obj,
> + va_addr, pdd->vm,
> + (struct kgd_mem **)&mem, &size,
> + mmap_offset);
> + if (r)
> + goto err_unlock;
> +
> + idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
> + if (idr_handle < 0) {
> + r = -EFAULT;
> + goto err_free;
> + }
> +
> + mutex_unlock(&p->mutex);
> +
> + *handle = MAKE_HANDLE(dev->id, idr_handle);
> +
> + return 0;
> +
> +err_free:
> + amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem,
> + NULL);
> +err_unlock:
> + mutex_unlock(&p->mutex);
> + return r;
> +}
> +
> +int kfd_ipc_import_dmabuf(struct kfd_dev *dev, struct kfd_process *p,
> + int dmabuf_fd, uint64_t va_addr,
> + uint64_t *handle, uint64_t *mmap_offset)
> +{
> + int r;
> + struct dma_buf *dmabuf = dma_buf_get(dmabuf_fd);
> +
> + if (!dmabuf)
> + return -EINVAL;
> +
> + r = kfd_import_dmabuf_create_kfd_bo(dev, p, dmabuf, NULL,
> + va_addr, handle, mmap_offset);
> + dma_buf_put(dmabuf);
> + return r;
> +}
> +
> +int kfd_ipc_import_handle(struct kfd_process *p,
> + uint32_t *share_handle, uint64_t va_addr,
> + uint64_t *handle, uint32_t *gpu_id,
> + uint64_t *mmap_offset)
> +{
> + struct kfd_dev *dev;
> + int r;
> + struct kfd_ipc_obj *entry, *found = NULL;
> +
> + mutex_lock(&kfd_ipc_handles.lock);
> + /* Convert the user provided handle to hash key and search only in that
> + * bucket
> + */
> + hlist_for_each_entry(entry,
> + &kfd_ipc_handles.handles[HANDLE_TO_KEY(share_handle)], node) {
> + if (!memcmp(entry->share_handle, share_handle,
> + sizeof(entry->share_handle))) {
> + found = ipc_obj_get(entry);
> + break;
> + }
> + }
> + mutex_unlock(&kfd_ipc_handles.lock);
> +
> + if (!found)
> + return -EINVAL;
> +
> + pr_debug("Found ipc_dma_buf: %p\n", found->dmabuf);
> +
> + dev = kfd_device_by_id(found->gpu_id);
> + if (!dev)
> + return -ENODEV;
> +
> + r = kfd_import_dmabuf_create_kfd_bo(dev, p, found->dmabuf, found,
> + va_addr, handle, mmap_offset);
> + if (r)
> + goto error_unref;
> +
> + *gpu_id = found->gpu_id;
> +
> + return r;
> +
> +error_unref:
> + kfd_ipc_obj_put(&found);
> + return r;
> +}
> +
> +int kfd_ipc_export_as_handle(struct kfd_dev *dev, struct kfd_process *p,
> + uint64_t handle, uint32_t *ipc_handle)
> +{
> + struct kfd_process_device *pdd = NULL;
> + struct kfd_ipc_obj *ipc_obj;
> + struct kgd_mem *mem;
> + int r;
> +
> + if (!dev || !ipc_handle)
> + return -EINVAL;
> +
> + mutex_lock(&p->mutex);
> + pdd = kfd_bind_process_to_device(dev, p);
> + if (IS_ERR(pdd)) {
> + mutex_unlock(&p->mutex);
> + pr_err("Failed to get pdd\n");
> + return PTR_ERR(pdd);
> + }
> +
> + mem = kfd_process_device_translate_handle(pdd, GET_IDR_HANDLE(handle));
> + mutex_unlock(&p->mutex);
> +
> + if (!mem) {
> + pr_err("Failed to get bo");
> + return -EINVAL;
> + }
> +
> + r = amdgpu_amdkfd_gpuvm_export_ipc_obj(dev->kgd, pdd->vm, mem,
> + &ipc_obj);
> + if (r)
> + return r;
> +
> + ipc_obj->gpu_id = dev->id;
> + memcpy(ipc_handle, ipc_obj->share_handle,
> + sizeof(ipc_obj->share_handle));
> +
> + return r;
> +}
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_ipc.h b/drivers/gpu/drm/amd/amdkfd/kfd_ipc.h
> new file mode 100644
> index 000000000000..9450a667918e
> --- /dev/null
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_ipc.h
> @@ -0,0 +1,55 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + *
> + */
> +
> +#ifndef KFD_IPC_H_
> +#define KFD_IPC_H_
> +
> +#include <linux/types.h>
> +#include <linux/dma-buf.h>
> +
> +/* avoid including kfd_priv.h */
> +struct kfd_dev;
> +struct kfd_process;
> +
> +struct kfd_ipc_obj {
> + struct hlist_node node;
> + struct kref ref;
> + struct dma_buf *dmabuf;
> + uint32_t share_handle[4];
> + uint32_t gpu_id;
> +};
> +
> +int kfd_ipc_import_handle(struct kfd_process *p,
> + uint32_t *share_handle, uint64_t va_addr,
> + uint64_t *handle, uint32_t *gpu_id,
> + uint64_t *mmap_offset);
> +int kfd_ipc_import_dmabuf(struct kfd_dev *kfd, struct kfd_process *p,
> + int dmabuf_fd, uint64_t va_addr,
> + uint64_t *handle, uint64_t *mmap_offset);
> +int kfd_ipc_export_as_handle(struct kfd_dev *dev, struct kfd_process *p,
> + uint64_t handle, uint32_t *ipc_handle);
> +
> +int kfd_ipc_store_insert(struct dma_buf *dmabuf, struct kfd_ipc_obj **ipc_obj);
> +void kfd_ipc_obj_put(struct kfd_ipc_obj **obj);
> +
> +#endif /* KFD_IPC_H_ */
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> index f4b7f7e6c40e..0946d5692692 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> @@ -52,6 +52,10 @@ static int kfd_init(void)
> if (err < 0)
> goto err_topology;
>
> + err = kfd_ipc_init();
> + if (err < 0)
> + goto err_ipc;
> +
> err = kfd_process_create_wq();
> if (err < 0)
> goto err_create_wq;
> @@ -66,6 +70,7 @@ static int kfd_init(void)
> return 0;
>
> err_create_wq:
> +err_ipc:
> kfd_topology_shutdown();
> err_topology:
> kfd_chardev_exit();
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 51ba2020732e..1588b2b45a32 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -321,6 +321,8 @@ struct kfd_dev {
> spinlock_t smi_lock;
> };
>
> +struct kfd_ipc_obj;
> +
> enum kfd_mempool {
> KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
> KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
> @@ -1087,6 +1089,9 @@ int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
>
> bool kfd_is_locked(void);
>
> +/* IPC Support */
> +int kfd_ipc_init(void);
> +
> /* Compute profile */
> void kfd_inc_compute_active(struct kfd_dev *dev);
> void kfd_dec_compute_active(struct kfd_dev *dev);
> diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
> index f738c3b53f4e..90ff334e2b5b 100644
> --- a/include/uapi/linux/kfd_ioctl.h
> +++ b/include/uapi/linux/kfd_ioctl.h
> @@ -29,9 +29,10 @@
> /*
> * - 1.1 - initial version
> * - 1.3 - Add SMI events support
> + * - 1.4 - Add IPC export/import
> */
> #define KFD_IOCTL_MAJOR_VERSION 1
> -#define KFD_IOCTL_MINOR_VERSION 3
> +#define KFD_IOCTL_MINOR_VERSION 4
>
> struct kfd_ioctl_get_version_args {
> __u32 major_version; /* from KFD */
> @@ -464,6 +465,57 @@ enum kfd_mmio_remap {
> KFD_MMIO_REMAP_HDP_REG_FLUSH_CNTL = 4,
> };
>
> +/* Export IPC handle
> + *
> + * @handle[in]: buffer handle of the BO to export
> + * @gpu_id[in]: GPU ID where @handle was allocated
> + * @share_handle[out]: share handle that can be used with
> + * @kfd_ioctl_ipc_import_handle_args
> + *
> + * @share_handle is a 128 bit random number generated with
> + * @get_random_bytes. This number should be very hard to guess.
> + * Knowledge of the @share_handle implies authorization to access
> + * the shared memory. User mode should treat it like a secret key.
> + * It can be used to import this BO in a different process context
> + * for IPC buffer sharing. The handle will be valid as long as the
> + * underlying BO exists. If the same BO is exported multiple times,
> + * the same handle will be returned.
> + *
> + * Return 0 on success, negative errno on errors.
> + */
> +struct kfd_ioctl_ipc_export_handle_args {
> + __u64 handle; /* to KFD */
> + __u32 share_handle[4]; /* from KFD */
> + __u32 gpu_id; /* to KFD */
> + __u32 pad;
> +};
> +
> +/* Import IPC handle
> + *
> + * @share_handle[in]: share handle from @kfd_ioctl_ipc_export_handle_args
> + * @va_addr[in]: virtual address at which to import the BO
> + * @handle[out]: buffer handle of the imported BO
> + * @gpu_id[out]: device in which the shared BO was created
> + * @mmap_offset[out]: mmap offset for CPU-mapping the BO
> + *
> + * @handle represents a new reference to the shared BO. This reference
> + * must be released with kfd_ioctl_free_memory_of_gpu_args.
> + *
> + * The BO can be mapped for GPU access with @kfd_ioctl_map_memory_to_gpu_args.
> + *
> + * It can be mapped for CPU access using the @mmap_offset.
> + *
> + * Return 0 on success, negative errno on errors.
> + */
> +struct kfd_ioctl_ipc_import_handle_args {
> + __u64 handle; /* from KFD */
> + __u64 va_addr; /* to KFD */
> + __u64 mmap_offset; /* from KFD */
> + __u32 share_handle[4]; /* to KFD */
> + __u32 gpu_id; /* from KFD */
> + __u32 pad;
> +};
> +
> #define AMDKFD_IOCTL_BASE 'K'
> #define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr)
> #define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type)
> @@ -564,7 +616,13 @@ enum kfd_mmio_remap {
> #define AMDKFD_IOC_SMI_EVENTS \
> AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
>
> +#define AMDKFD_IOC_IPC_IMPORT_HANDLE \
> + AMDKFD_IOWR(0x20, struct kfd_ioctl_ipc_import_handle_args)
> +
> +#define AMDKFD_IOC_IPC_EXPORT_HANDLE \
> + AMDKFD_IOWR(0x21, struct kfd_ioctl_ipc_export_handle_args)
> +
> #define AMDKFD_COMMAND_START 0x01
> -#define AMDKFD_COMMAND_END 0x20
> +#define AMDKFD_COMMAND_END 0x22
>
> #endif
> --
> 2.27.0
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=02%7C01%7Crajneesh.bhardwaj%40amd.com%7Cb0e43edbbeb24faacc6408d827a41995%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637302933054506951&sdata=w2QSn0fIBNtNFeyHwG0ArxbxGAlOJBLyK1CGicPUylU%3D&reserved=0
More information about the amd-gfx
mailing list