Add a dedicated tokio blocking runtime for fsync/fdatasync operations,
controlled by RUSTFS_RUNTIME_FSYNC_BLOCKING_THREADS (default 0 = off).
When configured with >1 threads, fsync operations are dispatched to a
separate blocking pool so device-bound fsync does not starve read
operations (pread/stat/open) on the main blocking pool.
Changed call sites:
- run_file_sync_blocking (batch fdatasync)
- fsync_dir (directory fsync)
- fsync_open_dst_dir_group (dst dir group fsync)
- run_blocking_namespace_file_sync_operation_with_global (namespace+fsync)
Unchanged (stay on main pool):
- run_blocking_namespace_operation (rename/metadata)
- local.rs spawn_blocking calls (pread/stat/open)
Default behavior is unchanged (0 = no isolation).
Co-Authored-By: heihutu <heihutu@gmail.com>
houseme
e4c7e12098