Hello everyone, This series adds a new migration capability called "precopy initial data". The purpose of this capability is to reduce migration downtime in cases where loading of migration data in the destination can take a lot of time, such as with VFIO migration data.
The series then moves to add precopy support and precopy initial data support for VFIO migration. Precopy initial data is used by VFIO migration, but other migration users can add support for it and use it as well. === Background === Migration downtime estimation is calculated based on bandwidth and remaining migration data. This assumes that loading of migration data in the destination takes a negligible amount of time and that downtime depends only on network speed. While this may be true for RAM, it's not necessarily true for other migration users. For example, loading the data of a VFIO device in the destination might require from the device to allocate resources and prepare internal data structures which can take a significant amount of time to do. This poses a problem, as the source may think that the remaining migration data is small enough to meet the downtime limit, so it will stop the VM and complete the migration, but in fact sending and loading the data in the destination may take longer than the downtime limit. To solve this, VFIO migration uAPI defines "initial bytes" as part of its precopy stream [1]. Initial bytes can be used in various ways to improve VFIO migration performance. For example, it can be used to transfer device metadata to pre-allocate resources in the destination. However, for this to work we need to make sure that all initial bytes are sent and loaded in the destination before the source VM is stopped. The new precopy initial data migration capability helps us achieve this. It allows the source to send initial precopy data and the destination to ACK that this data has been loaded. Migration will not attempt to stop the source VM and complete the migration until this ACK is received. Note that this relies on the return path capability to communicate from the destination back to the source. === Flow of operation === To use precopy initial data, the capability must be enabled in the source. As this capability must be supported also in the destination, a handshake is performed during migration setup. The purpose of the handshake is to notify the destination that precopy initial data is used and to check if it's supported. The handshake is done in two levels. First, a general handshake is done with the destination migration code to notify that precopy initial data is used. Then, for each migration user in the source that supports precopy initial data, a handshake is done with its counterpart in the destination: If both support it, precopy initial data will be used for them. If source doesn't support it, precopy initial data will not be used for them. If source supports it and destination doesn't, migration will be failed. Assuming the handshake succeeded, migration starts to send precopy data and as part of it also the initial precopy data. Initial precopy data is just like any other precopy data and as such, migration code is not aware of it. Therefore, it's the responsibility of the migration users (such as VFIO devices) to notify their counterparts in the destination that their initial precopy data has been sent (for example, VFIO migration does it when its initial bytes reach zero). In the destination, migration code will query each migration user that supports precopy initial data and check if its initial data has been loaded. If initial data has been loaded by all of them, an ACK will be sent to the source which will now be able to complete migration when appropriate. === Test results === The below table shows the downtime of two identical migrations. In the first migration precopy initial data is disabled and in the second it is enabled. The migrated VM is assigned with a mlx5 VFIO device which has 300MB of device data to be migrated. +----------------------+-----------------------+----------+ | Precopy initial data | VFIO device data size | Downtime | +----------------------+-----------------------+----------+ | Disabled | 300MB | 1900ms | | Enabled | 300MB | 420ms | +----------------------+-----------------------+----------+ Precopy initial data gives a roughly 4.5 times improvement in downtime. The 1480ms difference is time that is used for resource allocation for the VFIO device in the destination. Without precopy initial data, this time is spent when the source VM is stopped and thus the downtime is much higher. With precopy initial data, the time is spent when the source VM is still running. === Patch breakdown === - Patches 1-5 add the precopy initial data capability. - Patches 6-7 add VFIO migration precopy support. Similar version of them was previously sent here [2]. - Patch 8 adds precopy initial data support for VFIO migration. Thanks for reviewing! [1] https://elixir.bootlin.com/linux/latest/source/include/uapi/linux/vfio.h#L1048 [2] https://lore.kernel.org/qemu-devel/20230222174915.5647-3-avih...@nvidia.com/ Avihai Horon (8): migration: Add precopy initial data capability migration: Add precopy initial data handshake migration: Add precopy initial data loaded ACK functionality migration: Enable precopy initial data capability tests: Add migration precopy initial data capability test vfio/migration: Refactor vfio_save_block() to return saved data size vfio/migration: Add VFIO migration pre-copy support vfio/migration: Add support for precopy initial data capability docs/devel/vfio-migration.rst | 35 ++++-- qapi/migration.json | 9 +- include/hw/vfio/vfio-common.h | 6 + include/migration/register.h | 13 ++ migration/migration.h | 15 +++ migration/options.h | 1 + migration/savevm.h | 1 + hw/vfio/common.c | 6 +- hw/vfio/migration.c | 218 +++++++++++++++++++++++++++++++--- migration/migration.c | 45 ++++++- migration/options.c | 16 +++ migration/savevm.c | 141 ++++++++++++++++++++++ tests/qtest/migration-test.c | 23 ++++ hw/vfio/trace-events | 4 +- migration/trace-events | 4 + 15 files changed, 504 insertions(+), 33 deletions(-) -- 2.26.3
