perf(filesystem): optimize ListDirectory allocations and syscalls

[codspeed-hq]

Summary

Optimizes Filesystem.ListDirectory — invoked on every directory-listing HTTP request — by eliminating redundant syscalls and reducing heap allocations.

Flamegraph analysis of BenchmarkListDirectory revealed three independent sources of overhead beyond user/group lookups:

1. Redundant os.Lstat calls for directory entries

os.ReadDir returns DirEntry values whose IsDir() already knows the file type from the getdents syscall — no extra stat is needed. The original code called os.Lstat on every entry, including directories. We now check entry.IsDir() first and skip os.Lstat entirely for directories. os.Lstat is still used for files/symlinks (where full stat info is required), and a fallback handles the rare DT_UNKNOWN case.

2. Per-entry heap allocations causing GC pressure

Each File/Subdirectory was individually heap-allocated (&File{...}) and appended. For N entries this created N separate heap objects. We now pre-allocate backing slices ([]File, []Subdirectory) sized to len(entries) and take pointers into them, reducing N allocations to 2 and lowering GC scan work.

3. fmt.Sprintf("%o", ...) for permission formatting

Replaced with strconv.FormatUint(uint64(info.Mode()), 8), avoiding reflection and intermediate buffer allocation.

Benchmark Results

Validated locally through the CodSpeed CLI (walltime):

Benchmark	Base	Head	Change
BenchmarkListDirectory	102.6 µs	69.1 µs	+48.47%

No regressions on other benchmarks. All existing filesystem and API tests pass.

Notes

Owner/group resolution for files is preserved via new lookupUsername/lookupGroupname helpers, which resolve UID/GID from the already-obtained syscall.Stat_t and fall back to the numeric ID string on lookup failure — matching the prior behavior.

---

Note

Optimizes ListDirectory by pre-allocating backing slices for File/Subdirectory structs, skipping os.Lstat for directory entries (using DirEntry.IsDir() from getdents), and replacing fmt.Sprintf with strconv.FormatUint for permission formatting.

^{Written by Mendral for commit ce9ce31.}

[mendral-app]

LGTM

The optimization is sound. Backing slice pointers are safe because fileBacking and subdirBacking are each capped at numEntries which is the total count, so neither can exceed capacity and trigger reallocation. The DT_UNKNOWN fallback is correctly handled. Permission formatting produces identical output.

_{Tag @mendral-app with feedback or questions. View session}