AMD Continues Enhancing AMDGPU/AMDKFD Drivers For Checkpoint/Restore
CRIU is for Checkpoint/Restore in Userspace to be able to freeze a running container or app, preserve its state to disk, and later restore said running workload. A few years ago we saw AMD working on being able to checkpoint/restore running ROCm workloads. As seemingly the first work in a while on the matter by the AMDGPU/AMDKFD kernel drivers, there are some new CRIU elements coming for Linux 6.18.
Last week's AMDGPU pull request of additional material for Linux 6.18 adds CRIU support for GEM memory management objects. CRIU support for GEM objects is also added to the AMDKFD compute kernel driver.
As part of the CRIU support for GEM objects to AMDGPU and AMDKFD drivers, there is a new user-space API for handling it. Over in user-space is this pending CRIU pull request for DMA-BUF IPC support on AMDGPU.
In addition to this new AMD CRIU activity, this pull request brings UVD and DPM fixes for aging Southern Islands graphics cards. That work for the aging AMD Southern Islands GPUs is important as part of enhancing the GCN 1.0 and GCN 1.1 support on the AMDGPU driver as an alternative to the default Radeon DRM driver.
Plus this pull request has more updates for Cyan Skillfish, continued user-queue "UserQ" work, exposing the vBIOS build number to user-space via sysfs, eDP updates, and various other fixes.
Last week's AMDGPU pull request of additional material for Linux 6.18 adds CRIU support for GEM memory management objects. CRIU support for GEM objects is also added to the AMDKFD compute kernel driver.
As part of the CRIU support for GEM objects to AMDGPU and AMDKFD drivers, there is a new user-space API for handling it. Over in user-space is this pending CRIU pull request for DMA-BUF IPC support on AMDGPU.
In addition to this new AMD CRIU activity, this pull request brings UVD and DPM fixes for aging Southern Islands graphics cards. That work for the aging AMD Southern Islands GPUs is important as part of enhancing the GCN 1.0 and GCN 1.1 support on the AMDGPU driver as an alternative to the default Radeon DRM driver.
Plus this pull request has more updates for Cyan Skillfish, continued user-queue "UserQ" work, exposing the vBIOS build number to user-space via sysfs, eDP updates, and various other fixes.