From: Rajneesh Bhardwaj <rajneesh.bhardwaj@xxxxxxx>
KFD buffer objects do not associate a GEM handle with them so cannot
directly be used with libdrm to initiate a system dma (sDMA) operation
to speedup the checkpoint and restore operation so export them as dmabuf
objects and use with libdrm helper (amdgpu_bo_import) to further process
the sdma command submissions.
With sDMA, we see huge improvement in checkpoint and restore operations
compared to the generic pci based access via host data path.
Suggested-by: Felix Kuehling <felix.kuehling@xxxxxxx>
Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@xxxxxxx>
Signed-off-by: David Yat Sin <david.yatsin@xxxxxxx>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 57 ++++++++++++++++++++++++
1 file changed, 57 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 2d79bb6082dc..1a1039cc7f55 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -35,6 +35,7 @@
#include <linux/mman.h>
#include <linux/ptrace.h>
#include <linux/dma-buf.h>
+#include <linux/fdtable.h>
#include <asm/processor.h>
#include "kfd_priv.h"
#include "kfd_device_queue_manager.h"
@@ -43,6 +44,7 @@
#include "amdgpu_amdkfd.h"
#include "kfd_smi_events.h"
#include "amdgpu_object.h"
+#include "amdgpu_dma_buf.h"
static long kfd_ioctl(struct file *, unsigned int, unsigned long);
static int kfd_open(struct inode *, struct file *);
@@ -1899,6 +1901,33 @@ uint64_t get_process_num_bos(struct kfd_process *p)
return num_of_bos;
}
+static int criu_get_prime_handle(struct drm_gem_object *gobj, int flags,
+ u32 *shared_fd)
+{
+ struct dma_buf *dmabuf;
+ int ret;
+
+ dmabuf = amdgpu_gem_prime_export(gobj, flags);
+ if (IS_ERR(dmabuf)) {
+ ret = PTR_ERR(dmabuf);
+ pr_err("dmabuf export failed for the BO\n");
+ return ret;
+ }
+
+ ret = dma_buf_fd(dmabuf, flags);
+ if (ret < 0) {
+ pr_err("dmabuf create fd failed, ret:%d\n", ret);
+ goto out_free_dmabuf;
+ }
+
+ *shared_fd = ret;
+ return 0;
+
+out_free_dmabuf:
+ dma_buf_put(dmabuf);
+ return ret;
+}
+
static int criu_dump_bos(struct kfd_process *p, struct kfd_ioctl_criu_dumper_args *args)
{
struct kfd_criu_bo_bucket *bo_buckets;
@@ -1968,6 +1997,14 @@ static int criu_dump_bos(struct kfd_process *p, struct kfd_ioctl_criu_dumper_arg
goto exit;
}
}
+ if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+ ret = criu_get_prime_handle(&dumper_bo->tbo.base,
+ bo_bucket->alloc_flags &
+ KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
+ &bo_bucket->dmabuf_fd);
+ if (ret)
+ goto exit;
+ }
if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
KFD_MMAP_GPU_ID(pdd->dev->id);
@@ -1997,6 +2034,11 @@ static int criu_dump_bos(struct kfd_process *p, struct kfd_ioctl_criu_dumper_arg
}
exit:
+ while (ret && bo_index--) {
+ if (bo_buckets[bo_index].alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
+ close_fd(bo_buckets[bo_index].dmabuf_fd);
+ }
+
kvfree(bo_buckets);
return ret;
}
@@ -2243,6 +2285,7 @@ static int criu_restore_bos(struct kfd_process *p, struct kfd_ioctl_criu_restore
struct kfd_criu_bo_priv_data *bo_priv;
struct kfd_dev *dev;
struct kfd_process_device *pdd;
+ struct kgd_mem *kgd_mem;
void *mem;
u64 offset;
int idr_handle;
@@ -2390,6 +2433,16 @@ static int criu_restore_bos(struct kfd_process *p, struct kfd_ioctl_criu_restore
}
pr_debug("map memory was successful for the BO\n");
+ /* create the dmabuf object and export the bo */
+ kgd_mem = (struct kgd_mem *)mem;
+ if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+ ret = criu_get_prime_handle(&kgd_mem->bo->tbo.base,
+ DRM_RDWR,
+ &bo_bucket->dmabuf_fd);
+ if (ret)
+ goto exit;
+ }
+
} /* done */
/* Flush TLBs after waiting for the page table updates to complete */
@@ -2414,6 +2467,10 @@ static int criu_restore_bos(struct kfd_process *p, struct kfd_ioctl_criu_restore
ret = -EFAULT;
exit:
+ while (ret && i--) {
+ if (bo_buckets[i].alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
+ close_fd(bo_buckets[i].dmabuf_fd);
+ }
kvfree(objects);
return ret;
}
--
2.17.1
