Add non-mmap fallback for unsupported file systems (jffs2)

[abcfy2]

Background

ResolvesSagerNet/sing-box#4251.

On embedded devices where the cache file lives on a file system that does not support mmap (e.g. jffs2 on Asus Merlin routers, or certain OpenWrt MTD-backed partitions), bbolt.Open() fails with:

FATAL start service: initialize cache-file: invalid argument

The root cause: jffs2 rejects mmap(PROT_READ, MAP_SHARED) for files opened O_RDWR with EINVAL, because jffs2 only supports read-only shared mappings. bbolt's entire design assumes mmap, so this is fatal.

Solution

Port the mmap fallback implementation from MetaCubeX/bbolt (commit d4ec34a, authored by @wwqgtxx).

When mmap() fails with an error indicating the file system does not support memory mapping (ENOSYS, ENODEV, EOPNOTSUPP, ENOTSUP, EINVAL), bbolt transparently falls back to a heap-backed buffer that mirrors the data file via ReadAt/WriteAt. This preserves all existing read/write semantics with zero API changes — callers like sing-box need no modifications.

Changes

File	Change
mmap_fallback.go	New — heap mirror layer: mmapFallback(), munmapFallback(), copyToMmapFallback()
mmap_error_unix.go	New — isMmapUnsupported() for Linux/macOS/BSD (EINVAL, ENOSYS, ENODEV, EOPNOTSUPP, ENOTSUP)
mmap_error_windows.go	New — isMmapUnsupported() for Windows
db.go	Add mmapFallback field; mmap fail → fallback branch; munmap() dispatch; new writeAt() method that mirrors writes to heap buffer in fallback mode
tx.go	Route write path through db.writeAt() (2 call sites) so heap mirror stays consistent
bolt_openbsd.go	Skip msync() in fallback mode (heap buffer doesn't need it)

All new symbols are unexported (internal to the package). No public API changes.

Adaptations from upstream Mihomo patch

• Use root-package maxMapSize instead of common.MaxMapSize (sagernet fork keeps arch constants in bolt_<arch>.go)
• Remove Logger calls (sagernet fork has no Logger() method)
• Remove MaxSize field checks (not present in this fork)
• Add db.writeAt() method + route tx.go through it (sagernet fork lacked this abstraction; Mihomo already had it)

Verification

Unit behavior

• go build ./... passes
• go vet ./... passes

End-to-end on simulated jffs2

Tested with mtdram + jffs2 mount (the same method documented in MetaCubeX/mihomo#2922):

Control (original bbolt, no patch):

FATAL[0001] start service: initialize cache-file: invalid argument

Patched bbolt (with fallback):

INFO[0000] sing-box started (0.01s)
cache.db created (32768 bytes)

Data persistence (selector switch → restart → verify):

• Switch selector direct → block via Clash API
• Kill sing-box, keep cache.db on jffs2
• Restart → selector reads back as block ✅

The jffs2 mmap rejection (errno=22 EINVAL) was confirmed before testing:

mmap.mmap(fd, 32768, flags=mmap.MAP_SHARED, prot=mmap.PROT_READ)
# OSError: [Errno 22] Invalid argument

[abcfy2]

@nekohasekai PLS review this PR. Thanks.

After merge this PR, then upgrade dependencies of sing-box will resolve mmap issue.