| Message ID | 20221201152931.47913-1-yishaih@nvidia.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Add migration PRE_COPY support for mlx5 driver | expand |
On Thu, Dec 01, 2022 at 05:29:17PM +0200, Yishai Hadas wrote: > > Jason Gunthorpe (1): > vfio: Extend the device migration protocol with PRE_COPY > > Shay Drory (3): > net/mlx5: Introduce ifc bits for pre_copy > vfio/mlx5: Fallback to STOP_COPY upon specific PRE_COPY error > vfio/mlx5: Enable MIGRATION_PRE_COPY flag > > Yishai Hadas (10): > vfio/mlx5: Enforce a single SAVE command at a time > vfio/mlx5: Refactor PD usage > vfio/mlx5: Refactor MKEY usage > vfio/mlx5: Refactor migration file state > vfio/mlx5: Refactor to use queue based data chunks > vfio/mlx5: Introduce device transitions of PRE_COPY > vfio/mlx5: Introduce SW headers for migration states > vfio/mlx5: Introduce vfio precopy ioctl implementation > vfio/mlx5: Consider temporary end of stream as part of PRE_COPY > vfio/mlx5: Introduce multiple loads This looks OK to me now, the logic is clear Reviewed-by: Jason Gunthorpe <jgg@nvidia.com> Thanks, Jason
> From: Yishai Hadas <yishaih@nvidia.com> > Sent: Thursday, December 1, 2022 11:29 PM > > In mlx5 driver we could gain with this series about 20-30 percent > improvement > in the downtime compared to the previous code when PRE_COPY wasn't > supported. > Curious to see more data here. what is the workload/configuration? What is the size of the full state and downtime w/o PRECOPY? with PRECOPY what is the size of initial/middle/final states?
On 02/12/2022 10:57, Tian, Kevin wrote: >> From: Yishai Hadas <yishaih@nvidia.com> >> Sent: Thursday, December 1, 2022 11:29 PM >> >> In mlx5 driver we could gain with this series about 20-30 percent >> improvement >> in the downtime compared to the previous code when PRE_COPY wasn't >> supported. >> > Curious to see more data here. > > what is the workload/configuration? We tested with multiple workloads which were varied by the number of allocated resources, number of VFs on the VM, busy or idle device depending on some traffic that runs in the background, etc. > > What is the size of the full state and downtime w/o PRECOPY? It depends on the amount of the allocated resources that were already opened upon the migration, and the other workloads parameters as mentioned above. The downtime gain was mainly achieved by sending the initial/middle states having the metadata without regard to the size. > > with PRECOPY what is the size of initial/middle/final states? Generally saying, the initial state may include metadata on the current state, the middle states may hold 'diff' compared to the initial/previous ones and in most cases may be smaller, the final state holds the data itself and may be larger. Yishai
This series adds migration PRE_COPY uAPIs and their implementation as part of mlx5 driver. The uAPIs follow some discussion that was done in the mailing list [1] in this area. By the time the patches were sent, there was no driver implementation for the uAPIs, now we have it for mlx5 driver. The optional PRE_COPY state opens the saving data transfer FD before reaching STOP_COPY and allows the device to dirty track the internal state changes with the general idea to reduce the volume of data transferred in the STOP_COPY stage. While in PRE_COPY the device remains RUNNING, but the saving FD is open. A new ioctl VFIO_MIG_GET_PRECOPY_INFO is provided to allow userspace to query the progress of the precopy operation in the driver with the idea it will judge to move to STOP_COPY once the initial data set is transferred, and possibly after the dirty size has shrunk appropriately. User space can detect whether PRE_COPY is supported for a given device by checking the VFIO_MIGRATION_PRE_COPY flag once using the VFIO_DEVICE_FEATURE_MIGRATION ioctl. Extra details exist as part of the specific uAPI patch from the series. Finally, we come with mlx5 implementation based on its device specification for PRE_COPY. To support PRE_COPY, mlx5 driver is transferring multiple states (images) of the device. e.g.: the source VF can save and transfer multiple states, and the target VF will load them by that order. The device is saving three kinds of states: 1) Initial state - when the device moves to PRE_COPY state. 2) Middle state - during PRE_COPY phase via VFIO_MIG_GET_PRECOPY_INFO, can be multiple such states. 3) Final state - when the device moves to STOP_COPY state. After moving to PRE_COPY state, the user is holding the saving FD and should use it for transferring the data from the source to the target while the VM is still running. From user point of view, it's a stream of data, however, from mlx5 driver point of view it includes multiple images/states. For that, it sets some headers with metadata on the source to be parsed on the target. At some point, user may switch the device state from PRE_COPY to STOP_COPY, this will invoke saving of the final state. As discussed earlier in the mailing list, the data that is returned as part of PRE_COPY is not required to have any bearing relative to the data size available during the STOP_COPY phase. For this, we have the VFIO_DEVICE_FEATURE_MIG_DATA_SIZE option. In mlx5 driver we could gain with this series about 20-30 percent improvement in the downtime compared to the previous code when PRE_COPY wasn't supported. The series includes some pre-patches to be ready for managing multiple images then it comes with the PRE_COPY implementation itself. The matching qemu changes can be previewed here [2]. They come on top of the v2 migration protocol patches that were sent already to the mailing list. Note: As this series includes a net/mlx5 patch, we may need to send it as a pull request format to VFIO to avoid conflicts before acceptance. [1] https://lore.kernel.org/kvm/20220302172903.1995-8-shameerali.kolothum.thodi@huawei.com/ [2] https://github.com/avihai1122/qemu/commits/mig_v2_precopy Changes from V1: https://www.spinics.net/lists/kvm/msg296475.html Patch #2: Rephrase the 'initial_bytes' meaning as was suggested by Jason. Patch #9: Fix to send header based on PRE_COPY support. Patch #13: Fix some unwind flow to call complete(). Changes from V0: https://www.spinics.net/lists/kvm/msg294247.html Drop the first 2 patches that Alex merged already. Refactor mlx5 implementation based on Jason's comments on V0, it includes the below: * Refactor the PD usage to be aligned with the migration file life cycle. * Refactor the MKEY usage to be aligned with the migration file life cycle. * Refactor the migration file state. * Use queue based data chunks to simplify the driver code. * Use the FSM model on the target to simplify the driver code. * Extend the driver pre_copy header for future use. Yishai Jason Gunthorpe (1): vfio: Extend the device migration protocol with PRE_COPY Shay Drory (3): net/mlx5: Introduce ifc bits for pre_copy vfio/mlx5: Fallback to STOP_COPY upon specific PRE_COPY error vfio/mlx5: Enable MIGRATION_PRE_COPY flag Yishai Hadas (10): vfio/mlx5: Enforce a single SAVE command at a time vfio/mlx5: Refactor PD usage vfio/mlx5: Refactor MKEY usage vfio/mlx5: Refactor migration file state vfio/mlx5: Refactor to use queue based data chunks vfio/mlx5: Introduce device transitions of PRE_COPY vfio/mlx5: Introduce SW headers for migration states vfio/mlx5: Introduce vfio precopy ioctl implementation vfio/mlx5: Consider temporary end of stream as part of PRE_COPY vfio/mlx5: Introduce multiple loads drivers/vfio/pci/mlx5/cmd.c | 409 +++++++++++++++---- drivers/vfio/pci/mlx5/cmd.h | 96 ++++- drivers/vfio/pci/mlx5/main.c | 747 ++++++++++++++++++++++++++++------ drivers/vfio/vfio_main.c | 74 +++- include/linux/mlx5/mlx5_ifc.h | 14 +- include/uapi/linux/vfio.h | 122 +++++- 6 files changed, 1242 insertions(+), 220 deletions(-)