amdgpu libraries that use dmabuf fd to share GPU memory between
processes close the dmabuf fds immediately after using them.
However, it is possible that checkpoint of a process catches one
of the dmabuf fds open. In that case, the amdgpu plugin needs
to handle it.
The checkpoint of the dmabuf fd does require the device file
it was exported from to have already been dumped
To identify which device this dmabuf fd was exprted from, attempt
to import it on each device, then record the dmabuf handle
it imports as. This handle can be used to restore it.
Signed-off-by: David Francis <David.Francis@amd.com>
Buffer objects held by the amdgpu drm driver are checkpointed with
the new BO_INFO and MAPPING_INFO ioctls/ioctl options. Handling
is in amdgpu_plugin_drm.h
Handling of imported buffer objects may require dmabuf fds to be
transferred between processes. These occur over fdstore, with the
handle-fstore id relationships kept in shread memory. There is a
new plugin callback: RESTORE_INIT to create the shared memory.
During checkpoint, track shared buffer objects, so that buffer objects
that are shared across processes can be identified.
During restore, track which buffer objects have been restored. Retry
restore of a drm file if a buffer object is imported and the
original has not been exported yet. Skip buffer objects that have
already been completed or cannot be completed in the current restore.
So drm code can use sdma_copy_bo, that function no longer requires
kfd bo structs
Update the protobuf messages with new amdgpu drm information.
Signed-off-by: David Francis <David.Francis@amd.com>
Add a new compilation unit to host symbols and methods that will be
needed to C&R DRM devices. Refactor code that indicates support for
C&R and checkpoints KFD and DRM devices
Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu@amd.com>
Store BO contents directly to file (1 per GPU) instead of using
protobuf.
Bug Fix:
Fixes an issue where we could not handle BOs bigger than 4GB because
protobuf has an internal limit of 4GB for the Bytes structure.
Performance Improvements:
This significantly reduces CR duration on multi-GPU systems as it allows
reading and writing to disk in parallel. During checkpoint, instead of
waiting for all the BO contents to be read from the one protobuf file,
we can now start writing the BO contents as soon as the first BO is read
from disk. During restore, we can start writing BO contents to disk
after the first BO from VRAM. This also reduces the peak amount of
system memory used as we only need to keep 1 BO content in memory per
GPU at a time instead of all the BO contents.
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
Libhsakmt(thunk) uses a shared memory file in /dev/shm/hsakmt_shared_mem
and its semaphore in /dev/shm/hsakmt_shared_mem. Adding a check during
checkpoint to see if these two files exist. If they exist then the
plugin will try to restore them during restore.
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
The device topology on the restore node can be different from the
topology on the checkpointed node. The GPUs on the restore node may
have different gpu_ids, minor number. or some GPUs may have different
properties as checkpointed node. During restore, the CRIU plugin
determines the target GPUs to avoid restore failures caused by trying
to restore a process on a gpu that is different.
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
Parse local system topology in /sys/class/kfd/kfd/topology/nodes/ and
store properties for each gpu in the CRIU image files. The gpu
properties can then be used later during restore to make the process is
restored on gpu's with similar properties.
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
To support Checkpoint Restore with AMDGPUs for ROCm workloads, introduce
a new plugin to assist CRIU with the help of AMD KFD kernel driver. This
initial commit just provides the basic framework to build up further
capabilities. Like CRIU, the amdgpu plugin also uses protobuf to
serialize
and save the amdkfd data which is mostly VRAM contents with some
metadata.
We generate a data file "amdgpu-kfd-<id>.img" during the dump stage. On restore
this file is read and extracted to re-create various types of buffer
objects that belonged to the previously checkpointed process. Upon
restore the mmap page offset within a device file might change so we use
the new hook to update and adjust the mmap offsets for newly created
target process. This is needed for sys_mmap call in pie restorer phase.
Support for queues and events is added in future patches of this series.
With the current implementation (amdgpu_plugin), we support:
- Only compute workloads such (Non Gfx) are supported
- GPU visible inside a container
- AMD GPU Gfx 9 Family
- Pytorch Benchmarks such as BERT Base
amdgpu plugin dependes on libdrm and libdrm_amdgpu which are typically
installed with libdrm-dev package. We build amdgpu_plugin only when the
dependencies are met on the target system and when user intends to
install the amdgpu plugin and not by default with criu build.
Suggested-by: Felix Kuehling <felix.kuehling@amd.com>
Co-authored-by: David Yat Sin <david.yatsin@amd.com>
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@amd.com>
