vole
Clone
Read only, no credentials:
https://rodent.software/git/vole.git
Write over SSH with a registered key:
git clone ssh://git@0.0.0.0:2222/vole.git
A free and open-source forge that lives in your Git repo.
Bound and reclaim the runner's saved checkouts
f594f1a16f6f8a8a3ee2867a44249b72a2c76c96 · Keenan Nemetz · 5d ago · parent de84c32b
Bound and reclaim the runner's saved checkouts The runner already downloaded each repository once into --dir and cloned every job from that copy, but nothing kept the copies bounded: they grew with every tag, every branch, and every commit a job pinned by id, and a force-push left the old history reachable through the remote-tracking reflogs. A long-lived runner filled its disk and then could not check anything out at all. Repack a copy that is most of the cap, dropping the objects no branch or tag can reach and the temporary refs a killed runner leaves behind; then delete the least recently used copy when the cap is still reached. --git-cache-size sets the cap (10G by default, 0 keeps every copy). The repository this runner is about to build is never the one deleted, and a copy another process holds is passed over. Reading and writing a saved checkout now takes a per-repository lock, so a repack or a deletion in one runner process cannot run underneath another's clone, or underneath the site and release pushes that are written from the same cache. The first clone skips the working-tree checkout, which nothing ever read.
ci/agent/gitcache.go +347
| @@ -0,0 +1,347 @@ | |||
| 1 | package agent | ||
| 2 | |||
| 3 | // The runner keeps a trusted clone of every repository it builds at <dir>/<repo>, | ||
| 4 | // so a later job clones from disk and only the new commits cross the network. This | ||
| 5 | // file bounds that cache. | ||
| 6 | // | ||
| 7 | // Two operations, because they cost different things. Repacking one clone drops | ||
| 8 | // what no job can reach any more — the temporary refs a killed runner left behind, | ||
| 9 | // the reflogs that keep force-pushed history alive, and the objects only those | ||
| 10 | // held — and costs CPU. Deleting a clone costs the next job a full download, so it | ||
| 11 | // is the last resort, and whole clones are the only unit: half a clone is a job | ||
| 12 | // that fails in the middle of a fetch. | ||
| 13 | // | ||
| 14 | // Both take the repository's lock. A clone is a shared object store: a repack that | ||
| 15 | // prunes while another process is cloning from it, or a deletion that lands under a | ||
| 16 | // running job's feet, is data loss rather than a slow build. The lock is per | ||
| 17 | // repository, so two runners sharing a --dir still work in parallel on different | ||
| 18 | // repositories, and the only thing that waits is the same repository's next job. | ||
| 19 | |||
| 20 | import ( | ||
| 21 | "fmt" | ||
| 22 | "io/fs" | ||
| 23 | "os" | ||
| 24 | "os/exec" | ||
| 25 | "path/filepath" | ||
| 26 | "sort" | ||
| 27 | "strings" | ||
| 28 | "syscall" | ||
| 29 | "time" | ||
| 30 | |||
| 31 | "github.com/rodent-software/vole/internal/format" | ||
| 32 | ) | ||
| 33 | |||
| 34 | // DefaultGitCacheBytes caps the saved checkouts when a runner's operator says | ||
| 35 | // nothing. It is larger than the job cache because a clone only grows by history: | ||
| 36 | // a runner serving a handful of repositories should never reach it, and the cost | ||
| 37 | // of being wrong is one download. | ||
| 38 | const DefaultGitCacheBytes int64 = 10 << 30 | ||
| 39 | |||
| 40 | const ( | ||
| 41 | // lastUsedStamp records when a job last built from a clone. The directory's | ||
| 42 | // mtime cannot stand in: git writes into .git, and a job that fetched nothing | ||
| 43 | // still used the cache. | ||
| 44 | lastUsedStamp = ".vole-last-used" | ||
| 45 | // repackStamp records when a clone was last repacked, so a repository under | ||
| 46 | // pressure is not repacked after every job. | ||
| 47 | repackStamp = ".vole-last-repack" | ||
| 48 | // repackInterval is the shortest gap between two repacks of the same clone. | ||
| 49 | repackInterval = 6 * time.Hour | ||
| 50 | // repackShare is the fraction of the cap one repository may hold before | ||
| 51 | // repacking it is worth the CPU: below it there is no pressure to relieve. | ||
| 52 | repackShare = 4 | ||
| 53 | ) | ||
| 54 | |||
| 55 | // lockCache takes the repository's exclusive lock and returns the release | ||
| 56 | // function. A process that dies holding it releases it; the file itself is never | ||
| 57 | // removed, because removing a lock file races with the next process opening it. | ||
| 58 | func lockCache(cache, repo string) (func(), error) { | ||
| 59 | f, err := openCacheLock(cache, repo) | ||
| 60 | if err != nil { | ||
| 61 | return nil, err | ||
| 62 | } | ||
| 63 | if err := syscall.Flock(int(f.Fd()), syscall.LOCK_EX); err != nil { | ||
| 64 | _ = f.Close() | ||
| 65 | return nil, fmt.Errorf("lock the saved checkout of %s: %w", repo, err) | ||
| 66 | } | ||
| 67 | return func() { | ||
| 68 | _ = syscall.Flock(int(f.Fd()), syscall.LOCK_UN) | ||
| 69 | _ = f.Close() | ||
| 70 | }, nil | ||
| 71 | } | ||
| 72 | |||
| 73 | // tryLockCache is lockCache for a caller that can leave the work for later: a | ||
| 74 | // clone another process is using must not be deleted or repacked underneath it, | ||
| 75 | // and that process's job is more important than this one's tidying. | ||
| 76 | func tryLockCache(cache, repo string) (func(), bool) { | ||
| 77 | f, err := openCacheLock(cache, repo) | ||
| 78 | if err != nil { | ||
| 79 | return nil, false | ||
| 80 | } | ||
| 81 | if err := syscall.Flock(int(f.Fd()), syscall.LOCK_EX|syscall.LOCK_NB); err != nil { | ||
| 82 | _ = f.Close() | ||
| 83 | return nil, false | ||
| 84 | } | ||
| 85 | return func() { | ||
| 86 | _ = syscall.Flock(int(f.Fd()), syscall.LOCK_UN) | ||
| 87 | _ = f.Close() | ||
| 88 | }, true | ||
| 89 | } | ||
| 90 | |||
| 91 | func openCacheLock(cache, repo string) (*os.File, error) { | ||
| 92 | dir := filepath.Join(cache, ".locks") | ||
| 93 | if err := os.MkdirAll(dir, 0o755); err != nil { | ||
| 94 | return nil, err | ||
| 95 | } | ||
| 96 | f, err := os.OpenFile(filepath.Join(dir, repo+".lock"), os.O_CREATE|os.O_RDWR, 0o644) | ||
| 97 | if err != nil { | ||
| 98 | return nil, fmt.Errorf("open the lock for the saved checkout of %s: %w", repo, err) | ||
| 99 | } | ||
| 100 | return f, nil | ||
| 101 | } | ||
| 102 | |||
| 103 | // trustCache marks a clone that prepareWorktree finished, and records the job that | ||
| 104 | // used it. The marker is what makes a clone trusted; the stamp is what makes it | ||
| 105 | // evictable in the right order. | ||
| 106 | func touchCache(cacheDir string) error { | ||
| 107 | if err := os.WriteFile(filepath.Join(cacheDir, cacheMarker), []byte("1\n"), 0o644); err != nil { | ||
| 108 | return fmt.Errorf("mark the trusted cache: %w", err) | ||
| 109 | } | ||
| 110 | if err := os.WriteFile(filepath.Join(cacheDir, lastUsedStamp), []byte(time.Now().UTC().Format(time.RFC3339)+"\n"), 0o644); err != nil { | ||
| 111 | return fmt.Errorf("record the use of the trusted cache: %w", err) | ||
| 112 | } | ||
| 113 | return nil | ||
| 114 | } | ||
| 115 | |||
| 116 | // savedCheckout is one repository's clone under --dir. | ||
| 117 | type savedCheckout struct { | ||
| 118 | repo string | ||
| 119 | path string | ||
| 120 | size int64 | ||
| 121 | used time.Time | ||
| 122 | } | ||
| 123 | |||
| 124 | // savedCheckouts lists the clones a runner may reclaim. A directory without the | ||
| 125 | // marker is not trusted and not this file's business: prepareWorktree rebuilds it | ||
| 126 | // on the next job. | ||
| 127 | func savedCheckouts(cache string) []savedCheckout { | ||
| 128 | entries, err := os.ReadDir(cache) | ||
| 129 | if err != nil { | ||
| 130 | return nil | ||
| 131 | } | ||
| 132 | var out []savedCheckout | ||
| 133 | for _, e := range entries { | ||
| 134 | if !e.IsDir() { | ||
| 135 | continue | ||
| 136 | } | ||
| 137 | // The job cache, the job trees and the locks live in the same directory and | ||
| 138 | // are bounded or removed elsewhere. | ||
| 139 | switch e.Name() { | ||
| 140 | case "cache", "jobs", ".locks": | ||
| 141 | continue | ||
| 142 | } | ||
| 143 | dir := filepath.Join(cache, e.Name()) | ||
| 144 | if !exists(filepath.Join(dir, ".git")) || !exists(filepath.Join(dir, cacheMarker)) { | ||
| 145 | continue | ||
| 146 | } | ||
| 147 | out = append(out, savedCheckout{ | ||
| 148 | repo: e.Name(), | ||
| 149 | path: dir, | ||
| 150 | size: dirSize(dir), | ||
| 151 | used: stampTime(dir, lastUsedStamp), | ||
| 152 | }) | ||
| 153 | } | ||
| 154 | return out | ||
| 155 | } | ||
| 156 | |||
| 157 | // maintainTrustedCaches keeps the saved checkouts under cache within maxBytes. The | ||
| 158 | // active repository is the one the next job will build from, so it is never | ||
| 159 | // deleted. A maxBytes of zero keeps every clone: the operator asked for disk to be | ||
| 160 | // spent on downloads saved. | ||
| 161 | // | ||
| 162 | // It runs between jobs, and a failure is reported and ignored: a runner that | ||
| 163 | // cannot tidy its disk is still a runner that can build. | ||
| 164 | func maintainTrustedCaches(cache string, maxBytes int64, active string) { | ||
| 165 | if maxBytes <= 0 { | ||
| 166 | return | ||
| 167 | } | ||
| 168 | all := savedCheckouts(cache) | ||
| 169 | if len(all) == 0 { | ||
| 170 | return | ||
| 171 | } | ||
| 172 | var total int64 | ||
| 173 | for _, c := range all { | ||
| 174 | total += c.size | ||
| 175 | } | ||
| 176 | share := maxBytes / repackShare | ||
| 177 | if share < 1 { | ||
| 178 | share = 1 | ||
| 179 | } | ||
| 180 | |||
| 181 | // Repack under pressure first: it is the only way to free space without | ||
| 182 | // costing the next job a download. Biggest first, so the clone that is most of | ||
| 183 | // the cache is dealt with before smaller ones are deleted to make room for it. | ||
| 184 | bySize := append([]savedCheckout(nil), all...) | ||
| 185 | sort.Slice(bySize, func(i, j int) bool { return bySize[i].size > bySize[j].size }) | ||
| 186 | for _, c := range bySize { | ||
| 187 | overCap := total > maxBytes | ||
| 188 | if !overCap && c.size <= share { | ||
| 189 | continue | ||
| 190 | } | ||
| 191 | // A repository over its share is repacked at most every repackInterval. When | ||
| 192 | // the whole cache is over the cap it is worth doing now: the alternative is | ||
| 193 | // deleting a clone a repack might have kept. | ||
| 194 | if !overCap && !dueForRepack(c.path) { | ||
| 195 | continue | ||
| 196 | } | ||
| 197 | freed, err := repackSavedCheckout(cache, c.repo) | ||
| 198 | if err != nil { | ||
| 199 | fmt.Fprintf(os.Stderr, "vole runner: could not repack the saved checkout of %s: %v\n", c.repo, err) | ||
| 200 | continue | ||
| 201 | } | ||
| 202 | total -= freed | ||
| 203 | if freed > 0 { | ||
| 204 | fmt.Fprintf(os.Stderr, "vole runner: repacked the saved checkout of %s, freeing %s\n", | ||
| 205 | c.repo, format.Bytes(freed)) | ||
| 206 | } | ||
| 207 | } | ||
| 208 | if total <= maxBytes { | ||
| 209 | return | ||
| 210 | } | ||
| 211 | |||
| 212 | // Still over the cap. Delete the clones nobody has built from in the longest | ||
| 213 | // time, oldest first, and pass over one another process is using: its job is | ||
| 214 | // the one that would suffer. | ||
| 215 | sort.Slice(all, func(i, j int) bool { return all[i].used.Before(all[j].used) }) | ||
| 216 | for _, c := range all { | ||
| 217 | if total <= maxBytes { | ||
| 218 | return | ||
| 219 | } | ||
| 220 | if c.repo == active { | ||
| 221 | continue | ||
| 222 | } | ||
| 223 | freed, ok := deleteSavedCheckout(cache, c.repo) | ||
| 224 | if !ok { | ||
| 225 | continue | ||
| 226 | } | ||
| 227 | total -= freed | ||
| 228 | fmt.Fprintf(os.Stderr, "vole runner: deleted the saved checkout of %s to stay within the cap, freeing %s\n", | ||
| 229 | c.repo, format.Bytes(freed)) | ||
| 230 | } | ||
| 231 | } | ||
| 232 | |||
| 233 | // repackSavedCheckout drops everything in one clone that no branch or tag can | ||
| 234 | // reach. Branch tips and tags stay, because they are what turns the next job into | ||
| 235 | // a delta fetch. | ||
| 236 | func repackSavedCheckout(cache, repo string) (int64, error) { | ||
| 237 | unlock, ok := tryLockCache(cache, repo) | ||
| 238 | if !ok { | ||
| 239 | // Another process is using the clone right now. The space can wait for the | ||
| 240 | // next job rather than stalling this one. | ||
| 241 | return 0, nil | ||
| 242 | } | ||
| 243 | defer unlock() | ||
| 244 | |||
| 245 | dir := filepath.Join(cache, repo) | ||
| 246 | before := dirSize(dir) | ||
| 247 | if err := clearJobRefs(dir); err != nil { | ||
| 248 | return 0, err | ||
| 249 | } | ||
| 250 | // The cache is not a safety net: nobody recovers work from it, and a reflog for | ||
| 251 | // refs/remotes/origin/* is exactly what keeps a force-pushed-away history | ||
| 252 | // alive. Expire both kinds at once and let gc prune what that leaves. | ||
| 253 | for _, c := range [][2]string{ | ||
| 254 | {"gc.reflogExpire", "now"}, | ||
| 255 | {"gc.reflogExpireUnreachable", "now"}, | ||
| 256 | {"gc.pruneExpire", "now"}, | ||
| 257 | } { | ||
| 258 | if out, err := gitCache(dir, "config", c[0], c[1]); err != nil { | ||
| 259 | return 0, fmt.Errorf("git config %s: %v: %s", c[0], err, strings.TrimSpace(out)) | ||
| 260 | } | ||
| 261 | } | ||
| 262 | if out, err := gitCache(dir, "gc", "--quiet", "--prune=now"); err != nil { | ||
| 263 | return 0, fmt.Errorf("git gc: %v: %s", err, strings.TrimSpace(out)) | ||
| 264 | } | ||
| 265 | if err := os.WriteFile(filepath.Join(dir, repackStamp), []byte(time.Now().UTC().Format(time.RFC3339)+"\n"), 0o644); err != nil { | ||
| 266 | return 0, fmt.Errorf("record the repack: %w", err) | ||
| 267 | } | ||
| 268 | // A repack packs what is left, and on a small clone the new pack is bigger than | ||
| 269 | // the loose objects it removed. Report only what was actually freed, so the | ||
| 270 | // caller does not subtract a saving that did not happen. | ||
| 271 | after := dirSize(dir) | ||
| 272 | if after > before { | ||
| 273 | return 0, nil | ||
| 274 | } | ||
| 275 | return before - after, nil | ||
| 276 | } | ||
| 277 | |||
| 278 | // clearJobRefs removes the temporary refs prepareWorktree adds for a commit that is | ||
| 279 | // not on a branch. A killed runner leaves them, and they pin that commit forever. | ||
| 280 | // The lock is what makes this safe: a live job's ref is added under the same lock, | ||
| 281 | // so a running job's commit is never the one deleted here. | ||
| 282 | func clearJobRefs(dir string) error { | ||
| 283 | out, err := gitCache(dir, "for-each-ref", "--format=%(refname)", "refs/vole/jobs") | ||
| 284 | if err != nil { | ||
| 285 | return fmt.Errorf("git for-each-ref: %v: %s", err, strings.TrimSpace(out)) | ||
| 286 | } | ||
| 287 | for _, ref := range strings.Fields(out) { | ||
| 288 | if out, err := gitCache(dir, "update-ref", "-d", ref); err != nil { | ||
| 289 | return fmt.Errorf("git update-ref -d %s: %v: %s", ref, err, strings.TrimSpace(out)) | ||
| 290 | } | ||
| 291 | } | ||
| 292 | return nil | ||
| 293 | } | ||
| 294 | |||
| 295 | // deleteSavedCheckout removes one clone and reports what it freed. It declines when | ||
| 296 | // another process holds the lock, because that process is building from it. | ||
| 297 | func deleteSavedCheckout(cache, repo string) (int64, bool) { | ||
| 298 | unlock, ok := tryLockCache(cache, repo) | ||
| 299 | if !ok { | ||
| 300 | return 0, false | ||
| 301 | } | ||
| 302 | defer unlock() | ||
| 303 | dir := filepath.Join(cache, repo) | ||
| 304 | freed := dirSize(dir) | ||
| 305 | if err := os.RemoveAll(dir); err != nil { | ||
| 306 | fmt.Fprintf(os.Stderr, "vole runner: could not delete the saved checkout of %s: %v\n", repo, err) | ||
| 307 | return 0, false | ||
| 308 | } | ||
| 309 | return freed, true | ||
| 310 | } | ||
| 311 | |||
| 312 | func dueForRepack(dir string) bool { | ||
| 313 | return time.Since(stampTime(dir, repackStamp)) >= repackInterval | ||
| 314 | } | ||
| 315 | |||
| 316 | func stampTime(dir, name string) time.Time { | ||
| 317 | info, err := os.Stat(filepath.Join(dir, name)) | ||
| 318 | if err != nil { | ||
| 319 | return time.Time{} | ||
| 320 | } | ||
| 321 | return info.ModTime() | ||
| 322 | } | ||
| 323 | |||
| 324 | func exists(path string) bool { | ||
| 325 | _, err := os.Stat(path) | ||
| 326 | return err == nil | ||
| 327 | } | ||
| 328 | |||
| 329 | func dirSize(dir string) int64 { | ||
| 330 | var total int64 | ||
| 331 | _ = filepath.WalkDir(dir, func(_ string, d fs.DirEntry, err error) error { | ||
| 332 | if err != nil || d.IsDir() { | ||
| 333 | return nil | ||
| 334 | } | ||
| 335 | if info, err := d.Info(); err == nil { | ||
| 336 | total += info.Size() | ||
| 337 | } | ||
| 338 | return nil | ||
| 339 | }) | ||
| 340 | return total | ||
| 341 | } | ||
| 342 | |||
| 343 | func gitCache(dir string, args ...string) (string, error) { | ||
| 344 | cmd := exec.Command("git", append([]string{"-C", dir}, args...)...) | ||
| 345 | out, err := cmd.CombinedOutput() | ||
| 346 | return string(out), err | ||
| 347 | } | ||