You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
New Features
Credentials
Pre-request credential expiry check. Credentials are checked before each OSS request (60s margin) and force-refreshed if expired or near expiry, preventing the spurious 403s from a token lapsing between scheduled refreshes.
OSS Client & Networking
Path-style requests. New --path_style (default false) uses http(s)://endpoint/bucket/object instead of the virtual-hosted form, for PrivateLink endpoints without wildcard DNS.
IPv4-only resolution. New --enable_ipv6 (default true); set false to resolve both the endpoint and HTTP proxy over IPv4 only, working around broken IPv6 routes in dual-stack environments.
Caching
ETag-based page-cache retention. Page-cache validity no longer depends on mtime — if the ETag is unchanged, the cache is reused even when mtime changes. The ETag is also back-filled after a write flush, so write-then-read hits the cache instead of re-fetching from OSS.
Exclusive lock on the disk cache dir. With disk data caching on (--disk_data_cache_dir), an exclusive flock now stops multiple ossfs2 instances from sharing one cache dir and corrupting each other's data. The lock is <cache_dir>/.ossfs2.lock and data lives in <cache_dir>/cache-data/, so only the cache dir itself (not its parent) needs write permission.
Bug Fixes
Log rotation recreates a deleted log dir. If the log dir is removed externally (e.g. systemd-tmpfiles cleaning /tmp), rotation now recreates it and retries the open instead of failing permanently.
Crash on write retry near the multipart part limit. Edge case: near the 10000-part limit, a write failing with -EFBIG left residual merge state; retrying on the same handle re-merged remote data and tripped a fatal assert (SIGILL / exit 132). Retries now skip the redundant re-merge.
Appendable disk-cache "ghost pad". Only affects appendable objects with disk data caching on: once such an object grew past a size whose tail block had been cached with zero-padding, a full cache hit could return that padding as data and fail CRC. The cache is now dropped whenever the remote object identity changes, and the attribute lock became a mutex (from a spinlock) to avoid vCPU starvation on this path.