| 1 | // SPDX-License-Identifier: AGPL-3.0-or-later |
| 2 | |
| 3 | package storage |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "io/fs" |
| 10 | "os" |
| 11 | "os/exec" |
| 12 | "path/filepath" |
| 13 | "regexp" |
| 14 | "strings" |
| 15 | ) |
| 16 | |
| 17 | // RepoFS owns the on-disk layout for bare git repositories. All callers |
| 18 | // that touch repo paths route through this type so the path-validation |
| 19 | // rules live in exactly one place. |
| 20 | type RepoFS struct { |
| 21 | root string |
| 22 | } |
| 23 | |
| 24 | // NewRepoFS validates root (must be absolute, must exist, must be a |
| 25 | // directory) and returns the layer. |
| 26 | func NewRepoFS(root string) (*RepoFS, error) { |
| 27 | if root == "" { |
| 28 | return nil, errors.New("storage: repofs: root required") |
| 29 | } |
| 30 | if !filepath.IsAbs(root) { |
| 31 | return nil, fmt.Errorf("storage: repofs: root must be absolute, got %q", root) |
| 32 | } |
| 33 | abs, err := filepath.Abs(filepath.Clean(root)) |
| 34 | if err != nil { |
| 35 | return nil, fmt.Errorf("storage: repofs: clean root: %w", err) |
| 36 | } |
| 37 | info, err := os.Stat(abs) |
| 38 | if err != nil { |
| 39 | return nil, fmt.Errorf("storage: repofs: stat root: %w", err) |
| 40 | } |
| 41 | if !info.IsDir() { |
| 42 | return nil, fmt.Errorf("storage: repofs: root %q is not a directory", abs) |
| 43 | } |
| 44 | return &RepoFS{root: abs}, nil |
| 45 | } |
| 46 | |
| 47 | // Root returns the absolute root path. Useful for logging and `storage check`. |
| 48 | func (r *RepoFS) Root() string { return r.root } |
| 49 | |
| 50 | // ownerNameRE is the whitelist for owner names: lowercase ASCII letters, |
| 51 | // digits, and hyphens; cannot start or end with a hyphen; length 1..39 |
| 52 | // (matches GitHub's username constraint). |
| 53 | var ownerNameRE = regexp.MustCompile(`^[a-z0-9](?:[a-z0-9-]{0,37}[a-z0-9])?$`) |
| 54 | |
| 55 | // repoNameRE is the whitelist for repository names: lowercase ASCII |
| 56 | // letters, digits, hyphens, dots, and underscores. Can't start or end |
| 57 | // with a separator. Length 1..100 (matches GitHub). |
| 58 | var repoNameRE = regexp.MustCompile(`^[a-z0-9](?:[a-z0-9._-]{0,98}[a-z0-9_])?$`) |
| 59 | |
| 60 | // validateName enforces the per-kind whitelist. Returns ErrInvalidPath |
| 61 | // wrapped with a precise reason on failure. |
| 62 | func validateName(kind, name string) error { |
| 63 | if name == "" { |
| 64 | return fmt.Errorf("%w: %s empty", ErrInvalidPath, kind) |
| 65 | } |
| 66 | maxLen, re, alphabet := 39, ownerNameRE, "[a-z0-9-]" |
| 67 | if kind == "repo" { |
| 68 | maxLen, re, alphabet = 100, repoNameRE, "[a-z0-9._-]" |
| 69 | } |
| 70 | if len(name) > maxLen { |
| 71 | return fmt.Errorf("%w: %s %q too long (max %d)", ErrInvalidPath, kind, name, maxLen) |
| 72 | } |
| 73 | if name != strings.ToLower(name) { |
| 74 | return fmt.Errorf("%w: %s %q must be lowercase", ErrInvalidPath, kind, name) |
| 75 | } |
| 76 | if strings.Contains(name, "..") { |
| 77 | return fmt.Errorf("%w: %s contains dot-dot", ErrInvalidPath, kind) |
| 78 | } |
| 79 | if strings.HasPrefix(name, ".") { |
| 80 | return fmt.Errorf("%w: %s starts with dot", ErrInvalidPath, kind) |
| 81 | } |
| 82 | if filepath.IsAbs(name) { |
| 83 | return fmt.Errorf("%w: %s is absolute", ErrInvalidPath, kind) |
| 84 | } |
| 85 | if !re.MatchString(name) { |
| 86 | return fmt.Errorf("%w: %s %q fails whitelist %s", ErrInvalidPath, kind, name, alphabet) |
| 87 | } |
| 88 | return nil |
| 89 | } |
| 90 | |
| 91 | // shardOf returns the two-character shard prefix for owner. When owner is |
| 92 | // shorter than two characters, pads with `_` so the path remains stable. |
| 93 | func shardOf(owner string) string { |
| 94 | switch len(owner) { |
| 95 | case 0: |
| 96 | return "__" |
| 97 | case 1: |
| 98 | return owner + "_" |
| 99 | default: |
| 100 | return owner[:2] |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | // RepoPath returns the absolute disk path for the bare repository at |
| 105 | // (owner, name). Validates inputs and guarantees the result is rooted at |
| 106 | // r.root. Both inputs are lowercased before path construction. |
| 107 | func (r *RepoFS) RepoPath(owner, name string) (string, error) { |
| 108 | owner = strings.ToLower(owner) |
| 109 | name = strings.ToLower(name) |
| 110 | if err := validateName("owner", owner); err != nil { |
| 111 | return "", err |
| 112 | } |
| 113 | if err := validateName("repo", name); err != nil { |
| 114 | return "", err |
| 115 | } |
| 116 | p := filepath.Join(r.root, shardOf(owner), owner, name+".git") |
| 117 | if err := r.containedInRoot(p); err != nil { |
| 118 | return "", err |
| 119 | } |
| 120 | return p, nil |
| 121 | } |
| 122 | |
| 123 | // DeletedRepoPath returns the internal tombstone path used while a |
| 124 | // soft-deleted repo is inside its restore grace window. Keeping |
| 125 | // tombstones outside the canonical <owner>/<name>.git path lets a new |
| 126 | // active repo reuse the name without losing the old row's restore data. |
| 127 | func (r *RepoFS) DeletedRepoPath(owner, name string, repoID int64) (string, error) { |
| 128 | if repoID <= 0 { |
| 129 | return "", fmt.Errorf("%w: repo id required", ErrInvalidPath) |
| 130 | } |
| 131 | canonical, err := r.RepoPath(owner, name) |
| 132 | if err != nil { |
| 133 | return "", err |
| 134 | } |
| 135 | p := filepath.Join(filepath.Dir(canonical), ".deleted", fmt.Sprintf("%d.git", repoID)) |
| 136 | if err := r.containedInRoot(p); err != nil { |
| 137 | return "", err |
| 138 | } |
| 139 | return p, nil |
| 140 | } |
| 141 | |
| 142 | // containedInRoot returns ErrEscapesRoot when p does not resolve under r.root. |
| 143 | // Defense-in-depth: validateName already rejects ".." and absolute paths, |
| 144 | // but a future caller might compose paths differently. |
| 145 | func (r *RepoFS) containedInRoot(p string) error { |
| 146 | clean := filepath.Clean(p) |
| 147 | if !strings.HasPrefix(clean, r.root+string(filepath.Separator)) && clean != r.root { |
| 148 | return fmt.Errorf("%w: %s not under %s", ErrEscapesRoot, clean, r.root) |
| 149 | } |
| 150 | return nil |
| 151 | } |
| 152 | |
| 153 | // Exists reports whether path exists. Validates that path is under root. |
| 154 | func (r *RepoFS) Exists(path string) (bool, error) { |
| 155 | if err := r.containedInRoot(path); err != nil { |
| 156 | return false, err |
| 157 | } |
| 158 | _, err := os.Stat(path) |
| 159 | if err == nil { |
| 160 | return true, nil |
| 161 | } |
| 162 | if errors.Is(err, os.ErrNotExist) { |
| 163 | return false, nil |
| 164 | } |
| 165 | return false, fmt.Errorf("storage: repofs: stat %s: %w", path, err) |
| 166 | } |
| 167 | |
| 168 | // DiskUsageBytes sums the byte size of every regular file under path. |
| 169 | // It is used for quota gates and repo:size_recalc, so it validates that |
| 170 | // the target stays inside the configured repo root and never follows |
| 171 | // directory symlinks. |
| 172 | func (r *RepoFS) DiskUsageBytes(ctx context.Context, path string) (int64, error) { |
| 173 | if err := r.containedInRoot(path); err != nil { |
| 174 | return 0, err |
| 175 | } |
| 176 | var total int64 |
| 177 | err := filepath.WalkDir(path, func(_ string, d fs.DirEntry, walkErr error) error { |
| 178 | if walkErr != nil { |
| 179 | return walkErr |
| 180 | } |
| 181 | if err := ctx.Err(); err != nil { |
| 182 | return err |
| 183 | } |
| 184 | if d.IsDir() { |
| 185 | return nil |
| 186 | } |
| 187 | info, err := d.Info() |
| 188 | if err != nil { |
| 189 | return err |
| 190 | } |
| 191 | if info.Mode().IsRegular() { |
| 192 | total += info.Size() |
| 193 | } |
| 194 | return nil |
| 195 | }) |
| 196 | return total, err |
| 197 | } |
| 198 | |
| 199 | // InitBare creates a bare git repository at path. Default branch is |
| 200 | // "trunk" — there is no path through this package that creates a bare |
| 201 | // repo with a different initial branch. |
| 202 | // |
| 203 | // The parent directory tree is created on demand. ErrAlreadyExists is |
| 204 | // returned if path is non-empty. |
| 205 | // |
| 206 | // The repo is initialized with `--shared=group`, which: |
| 207 | // |
| 208 | // - persists `core.sharedRepository=group` in config |
| 209 | // - sets the setgid bit on directories (2775) |
| 210 | // - keeps group-writable mode bits on files (0664) |
| 211 | // |
| 212 | // Both shithubd-web (web pushes via the HTTPS handler, runs as the |
| 213 | // `shithub` user) and the SSH `git` user (the AuthorizedKeysCommand |
| 214 | // dispatches into a process running as `git`, which is in the |
| 215 | // `shithub` group) write to the same bare repo on disk. Without |
| 216 | // `--shared=group`, git-receive-pack via SSH fails with |
| 217 | // "unable to create temporary object directory" because objects/ |
| 218 | // is 0755 and group write isn't set. |
| 219 | func (r *RepoFS) InitBare(ctx context.Context, path string) error { |
| 220 | if err := r.containedInRoot(path); err != nil { |
| 221 | return err |
| 222 | } |
| 223 | if entries, err := os.ReadDir(path); err == nil && len(entries) > 0 { |
| 224 | return fmt.Errorf("%w: %s", ErrAlreadyExists, path) |
| 225 | } |
| 226 | if err := os.MkdirAll(filepath.Dir(path), 0o2750); err != nil { |
| 227 | return fmt.Errorf("storage: repofs: mkdir parent: %w", err) |
| 228 | } |
| 229 | if err := os.MkdirAll(path, 0o2750); err != nil { |
| 230 | return fmt.Errorf("storage: repofs: mkdir target: %w", err) |
| 231 | } |
| 232 | // G204: path is constructed via RepoPath (strict whitelist) and verified |
| 233 | // to live under r.root. Caller cannot inject arbitrary args. |
| 234 | cmd := exec.CommandContext(ctx, "git", "init", "--bare", "--shared=group", "--initial-branch=trunk", path) //nolint:gosec |
| 235 | out, err := cmd.CombinedOutput() |
| 236 | if err != nil { |
| 237 | return fmt.Errorf("storage: repofs: git init --bare: %w (output: %s)", err, strings.TrimSpace(string(out))) |
| 238 | } |
| 239 | return nil |
| 240 | } |
| 241 | |
| 242 | // CloneBareShared clones src → dst as a bare repo with object |
| 243 | // alternates pointing back at src. Disk usage of the result is |
| 244 | // essentially refs + a small overhead; objects live in src's |
| 245 | // `objects/` until the fork is detached (S16 hard-delete cascade |
| 246 | // repacks each fork before removing the source). |
| 247 | // |
| 248 | // Both paths must be contained in r.root and on the same volume — |
| 249 | // the same-volume requirement is what makes alternates safe (S04). |
| 250 | // |
| 251 | // On success the dst directory exists with `git init --bare` shape |
| 252 | // plus an `objects/info/alternates` file pointing at src/objects. |
| 253 | // On failure the dst directory is removed so a retry sees a clean |
| 254 | // slate. |
| 255 | func (r *RepoFS) CloneBareShared(ctx context.Context, src, dst string) error { |
| 256 | if err := r.containedInRoot(src); err != nil { |
| 257 | return err |
| 258 | } |
| 259 | if err := r.containedInRoot(dst); err != nil { |
| 260 | return err |
| 261 | } |
| 262 | if entries, err := os.ReadDir(dst); err == nil && len(entries) > 0 { |
| 263 | return fmt.Errorf("%w: %s", ErrAlreadyExists, dst) |
| 264 | } |
| 265 | if err := os.MkdirAll(filepath.Dir(dst), 0o2750); err != nil { |
| 266 | return fmt.Errorf("storage: repofs: mkdir parent: %w", err) |
| 267 | } |
| 268 | // `git clone --shared` (here: object-alternates flag, NOT a perms |
| 269 | // flag — same name, different sense than init's --shared=group). |
| 270 | // To get group-writable perms we set core.sharedRepository=group |
| 271 | // via -c so the cloned config has it from byte zero. Without this, |
| 272 | // SSH-git push to a fork hits the same EACCES on objects/ that |
| 273 | // PR for SR2 #287 fixed for `git init --bare` (see InitBare). |
| 274 | // |
| 275 | // G204: src/dst are RepoPath-derived, both verified under r.root. |
| 276 | cmd := exec.CommandContext(ctx, "git", "-c", "core.sharedRepository=group", "clone", "--bare", "--shared", src, dst) //nolint:gosec |
| 277 | out, err := cmd.CombinedOutput() |
| 278 | if err != nil { |
| 279 | // Best-effort cleanup; if removal fails too, surface the |
| 280 | // original clone error since that's the actionable signal. |
| 281 | _ = os.RemoveAll(dst) |
| 282 | return fmt.Errorf("storage: repofs: git clone --bare --shared: %w (output: %s)", err, strings.TrimSpace(string(out))) |
| 283 | } |
| 284 | return nil |
| 285 | } |
| 286 | |
| 287 | // RepairSharedPerms brings an existing bare repo to the |
| 288 | // `--shared=group` contract InitBare now produces from byte zero |
| 289 | // (SR2 #287). Idempotent: a repo already at the contract is left |
| 290 | // alone except for explicitly setting the config (cheap). |
| 291 | // |
| 292 | // Steps: |
| 293 | // 1. `git config core.sharedRepository=group` |
| 294 | // 2. `chmod -R g+w` and `find -type d -exec chmod g+s` so future |
| 295 | // writes inherit the group on creation. |
| 296 | // |
| 297 | // Group ownership itself is NOT changed — the shipped invariant is |
| 298 | // that all repos are owned by the `shithub` group already (the |
| 299 | // shithub user creates them). If a repo's group is wrong, that's a |
| 300 | // separate provisioning bug; this method's job is only the bits. |
| 301 | // |
| 302 | // Runs git with safe.directory=* injected via env so the operator |
| 303 | // (typically root over SSH) can operate on repos owned by the |
| 304 | // `shithub` user. The same trick is used by the SSH dispatcher. |
| 305 | // Without it, git 2.35+ emits "fatal: not in a git directory" as |
| 306 | // part of the dubious-ownership early exit. |
| 307 | func (r *RepoFS) RepairSharedPerms(ctx context.Context, path string) error { |
| 308 | if err := r.containedInRoot(path); err != nil { |
| 309 | return err |
| 310 | } |
| 311 | if _, err := os.Stat(path); err != nil { |
| 312 | return fmt.Errorf("storage: repofs: stat %s: %w", path, err) |
| 313 | } |
| 314 | // Persist the contract in config. |
| 315 | cfg := exec.CommandContext(ctx, "git", "-C", path, "config", "core.sharedRepository", "group") //nolint:gosec |
| 316 | cfg.Env = append( |
| 317 | os.Environ(), |
| 318 | "GIT_CONFIG_COUNT=1", |
| 319 | "GIT_CONFIG_KEY_0=safe.directory", |
| 320 | "GIT_CONFIG_VALUE_0=*", |
| 321 | ) |
| 322 | if out, err := cfg.CombinedOutput(); err != nil { |
| 323 | return fmt.Errorf("storage: repofs: git config sharedRepository: %w (output: %s)", err, strings.TrimSpace(string(out))) |
| 324 | } |
| 325 | // Walk the tree once: directories get +g+s, files get +g+w. |
| 326 | // path is verified contained-in-root above; no symlinks span out |
| 327 | // of the repo (bare repos don't ship with symlinks under .git/). |
| 328 | // G122: filepath.Walk + os.Chmod is race-prone in adversarial |
| 329 | // trees, but our writer (this process running as root or shithub) |
| 330 | // is also the only writer for these paths, and the trees are not |
| 331 | // user-influenced beyond the validated owner/name slugs. Operator- |
| 332 | // only command, not user-triggered. |
| 333 | if err := filepath.Walk(path, func(p string, info os.FileInfo, err error) error { //nolint:gosec |
| 334 | if err != nil { |
| 335 | return err |
| 336 | } |
| 337 | mode := info.Mode() |
| 338 | newMode := mode | 0o060 // group rw |
| 339 | if info.IsDir() { |
| 340 | newMode |= os.ModeSetgid // g+s |
| 341 | } |
| 342 | if newMode == mode { |
| 343 | return nil |
| 344 | } |
| 345 | return os.Chmod(p, newMode) //nolint:gosec |
| 346 | }); err != nil { |
| 347 | return fmt.Errorf("storage: repofs: walk chmod: %w", err) |
| 348 | } |
| 349 | return nil |
| 350 | } |
| 351 | |
| 352 | // SetPreciousObjects marks a bare repo's objects as not-prunable. The |
| 353 | // canonical foot-gun for forks is source-repo `git gc` removing |
| 354 | // objects that forks reach via alternates; setting this on the source |
| 355 | // after a fork is created prevents that. Idempotent. |
| 356 | func (r *RepoFS) SetPreciousObjects(ctx context.Context, path string) error { |
| 357 | if err := r.containedInRoot(path); err != nil { |
| 358 | return err |
| 359 | } |
| 360 | cmd := exec.CommandContext(ctx, "git", "-C", path, "config", "extensions.preciousObjects", "true") //nolint:gosec |
| 361 | if out, err := cmd.CombinedOutput(); err != nil { |
| 362 | return fmt.Errorf("storage: repofs: set preciousObjects: %w (output: %s)", err, strings.TrimSpace(string(out))) |
| 363 | } |
| 364 | return nil |
| 365 | } |
| 366 | |
| 367 | // Move atomically renames oldPath to newPath. Both must be under root. |
| 368 | // If newPath already exists, returns ErrAlreadyExists rather than |
| 369 | // overwriting (avoids silent corruption on concurrent moves). |
| 370 | func (r *RepoFS) Move(oldPath, newPath string) error { |
| 371 | if err := r.containedInRoot(oldPath); err != nil { |
| 372 | return err |
| 373 | } |
| 374 | if err := r.containedInRoot(newPath); err != nil { |
| 375 | return err |
| 376 | } |
| 377 | if _, err := os.Stat(newPath); err == nil { |
| 378 | return fmt.Errorf("%w: %s", ErrAlreadyExists, newPath) |
| 379 | } else if !errors.Is(err, os.ErrNotExist) { |
| 380 | return fmt.Errorf("storage: repofs: stat dest: %w", err) |
| 381 | } |
| 382 | if err := os.MkdirAll(filepath.Dir(newPath), 0o750); err != nil { |
| 383 | return fmt.Errorf("storage: repofs: mkdir parent: %w", err) |
| 384 | } |
| 385 | if err := os.Rename(oldPath, newPath); err != nil { |
| 386 | return fmt.Errorf("storage: repofs: rename: %w", err) |
| 387 | } |
| 388 | return nil |
| 389 | } |
| 390 | |
| 391 | // Delete removes the bare repo at path. Refuses paths outside root. |
| 392 | func (r *RepoFS) Delete(path string) error { |
| 393 | if err := r.containedInRoot(path); err != nil { |
| 394 | return err |
| 395 | } |
| 396 | if path == r.root { |
| 397 | return fmt.Errorf("%w: refusing to delete root", ErrEscapesRoot) |
| 398 | } |
| 399 | if err := os.RemoveAll(path); err != nil { |
| 400 | return fmt.Errorf("storage: repofs: remove: %w", err) |
| 401 | } |
| 402 | return nil |
| 403 | } |
| 404 |