fsimpl/gofer/tmpfs: Support statx btime

[tanyifeng]

gVisor did not report valid statx btime: gofer/directfs used fstat(2), which has no btime, and tmpfs did not track creation time.

Use host statx(AT_EMPTY_PATH) for gofer/directfs metadata. Restrict host statx in gofer/sandbox seccomp filters to FD-only calls using a fixed empty path and AT_EMPTY_PATH.

Track optional STATX_BTIME availability so filesystems without btime support do not expose zero btime. For tmpfs, initialize btime at inode creation and report STATX_BTIME, matching Linux >= 5.18.

Reference: fs/stat.c:vfs_statx(), mm/shmem.c:shmem_getattr()

[ayushr2]

Curious, which workload is this impacting?

[tanyifeng]

Curious, which workload is this impacting?

The dependency is from tokio-rs/tracing:

Specifically, test_max_log_files uses max_log_files = 2, writes three hourly log files, and expects the oldest one to be pruned.

The pruning code depends on file creation time here:

https://github.com/tokio-rs/tracing/blob/908cc432a5994f6e17c8f36e13c217dc40085704/tracing-appender/src/rolling.rs#L605

let created = metadata.created().ok()?;

On Linux, Rust's Metadata::created() uses statx(STATX_BTIME). Before this fix, gVisor did not expose a valid STATX_BTIME, so metadata.created() failed, the old log file was skipped from pruning, and the test hit:

https://github.com/tokio-rs/tracing/blob/908cc432a5994f6e17c8f36e13c217dc40085704/tracing-appender/src/rolling.rs#L1097

panic!("this file should have been pruned already!");

[ayushr2]

Mostly LGTM, thanks!

[ayushr2]

FYI I imported this and fixed some things:

• directfs was missing btime update on revalidation and restore paths (lisafs was updating it correctly)
• changed directfsInode.updateMetadataFromStatLocked() => updateMetadataFromStatxLocked() and updated it to be same as lisafsInode.updateMetadataFromStatxLocked() implementation.
• Deleted dentryTimestampFromStatx and just updated dentryTimestampFromUnix to use unix.Statx_t and updated all callsites.
• Fixed newDirectfsDentry() to check unix.Statx_t.Mask field and accordingly initialize the dentry fields, analogous to newLisafsDentry().
• Some minor cosmetic changes.

Will submit that change and retain your authorship on that.