8c746ce1
feat(agent): gzipped images are decompressed in-process, sparsely
a73x 2026-08-06 09:12
Commit message
docs/assumptions.md
| Old | New | ||
|---|---|---|---|
| @@ -123,3 +123,23 @@ reaches it over the LAN rather than hairpinning out through the uplink, and | |||
| 123 | that hosts stay on that LAN. | 123 | that hosts stay on that LAN. |
| 124 | **Unverified**: one fleet host is a laptop that leaves the network, and a | 124 | **Unverified**: one fleet host is a laptop that leaves the network, and a |
| 125 | public hostname resolving to a public address would route the long way round. | 125 | public hostname resolving to a public address would route the long way round. |
| 126 | |||
| 127 | ### A gzipped image wraps a raw one | ||
| 128 | |||
| 129 | Gzip is transport, not a disk format, so a `.gz` download is decompressed and | ||
| 130 | the result treated as the image. Underpins publishing guest images gzipped—802 | ||
| 131 | MB on the wire against 3.5 GB raw, and smaller than the qcow2 it replaces. | ||
| 132 | **Proven** for the failure that matters: the decompressed head is sniffed | ||
| 133 | before a byte is written, so a gzipped qcow2 is rejected permanently instead of | ||
| 134 | landing in the cache under a `.raw` name and failing later at boot. Corrupt | ||
| 135 | gzip is permanent for the same reason a checksum mismatch is—the sha already | ||
| 136 | matched, so re-fetching returns the same bytes. | ||
| 137 | |||
| 138 | ### Decompression is cheaper than the download it saves | ||
| 139 | |||
| 140 | Measured on the real image: 802 MB gzipped, 6.8s to decompress with C gunzip | ||
| 141 | and 14.4s writing sparsely. Go's `compress/gzip` is slower, so budget 15-25s. | ||
| 142 | Underpins choosing gzipped raw over plain raw. | ||
| 143 | **Partly proven**: measured on one host (6 vCPU, nested). A slow host pays more | ||
| 144 | CPU for less transfer, which is the right trade on a LAN and the wrong one only | ||
| 145 | if a host is very slow and very well connected. | ||
internal/agent/imagecache/format.go
| Old | New | ||
|---|---|---|---|
| @@ -43,6 +43,20 @@ func sniff(head []byte) imageFormat { | |||
| 43 | } | 43 | } |
| 44 | } | 44 | } |
| 45 | 45 | ||
| 46 | // permanentError marks a failure no retry can fix. reconcile matches the | ||
| 47 | // Permanent() method structurally (errors.As against an anonymous interface), | ||
| 48 | // so this mirrors cloudhv's and inert's markers instead of sharing one: what a | ||
| 49 | // VMM driver and an image cache call permanent have nothing else in common. | ||
| 50 | type permanentError struct{ err error } | ||
| 51 | |||
| 52 | func (e permanentError) Error() string { return e.err.Error() } | ||
| 53 | func (e permanentError) Unwrap() error { return e.err } | ||
| 54 | func (e permanentError) Permanent() bool { return true } | ||
| 55 | |||
| 56 | func permanentf(format string, args ...any) error { | ||
| 57 | return permanentError{err: fmt.Errorf(format, args...)} | ||
| 58 | } | ||
| 59 | |||
| 46 | // sniffFile classifies the image file at path. A file shorter than magicLen is | 60 | // sniffFile classifies the image file at path. A file shorter than magicLen is |
| 47 | // read in full and classified on what it has, so a truncated or empty image is | 61 | // read in full and classified on what it has, so a truncated or empty image is |
| 48 | // raw rather than an error — the checksum has already vouched for the bytes by | 62 | // raw rather than an error — the checksum has already vouched for the bytes by |
internal/agent/imagecache/imagecache.go
| Old | New | ||
|---|---|---|---|
| @@ -15,9 +15,12 @@ | |||
| 15 | package imagecache | 15 | package imagecache |
| 16 | 16 | ||
| 17 | import ( | 17 | import ( |
| 18 | "bytes" | ||
| 19 | "compress/gzip" | ||
| 18 | "context" | 20 | "context" |
| 19 | "crypto/sha256" | 21 | "crypto/sha256" |
| 20 | "encoding/hex" | 22 | "encoding/hex" |
| 23 | "errors" | ||
| 21 | "fmt" | 24 | "fmt" |
| 22 | "io" | 25 | "io" |
| 23 | "log/slog" | 26 | "log/slog" |
| @@ -176,8 +179,8 @@ func (c *Cache) ensureOnce(ctx context.Context, url, sha string) (string, error) | |||
| 176 | return final, nil | 179 | return final, nil |
| 177 | } | 180 | } |
| 178 | 181 | ||
| 179 | // Convert to a UNIQUE temp file first, then rename onto final atomically so a | 182 | // Materialize into a UNIQUE temp file first, then rename onto final atomically |
| 180 | // crash mid-convert cannot leave a corrupt file at the final path. The temp | 183 | // so a crash mid-write cannot leave a corrupt file at the final path. The temp |
| 181 | // name is unique (not a fixed "<sha>.converting") so two concurrent Ensure | 184 | // name is unique (not a fixed "<sha>.converting") so two concurrent Ensure |
| 182 | // calls for the same sha can never converge on one file and corrupt it. The | 185 | // calls for the same sha can never converge on one file and corrupt it. The |
| 183 | // suffix deliberately does NOT end in ".raw", so evict()'s "*.raw" glob never | 186 | // suffix deliberately does NOT end in ".raw", so evict()'s "*.raw" glob never |
| @@ -188,10 +191,19 @@ func (c *Cache) ensureOnce(ctx context.Context, url, sha string) (string, error) | |||
| 188 | } | 191 | } |
| 189 | converting := cf.Name() | 192 | converting := cf.Name() |
| 190 | _ = cf.Close() | 193 | _ = cf.Close() |
| 191 | _ = os.Remove(converting) // reserve the unique name; let qemu-img create it fresh | 194 | _ = os.Remove(converting) // reserve the unique name; let the writer create it fresh |
| 192 | defer os.Remove(converting) | 195 | defer os.Remove(converting) |
| 193 | if _, err := c.run(ctx, "qemu-img", "convert", "-O", "raw", tmp, converting); err != nil { | 196 | |
| 194 | return "", fmt.Errorf("qemu-img convert: %w", err) | 197 | // sniff returns exactly three formats and raw returned above, so these two |
| 198 | // cases are the whole space; a fourth format means editing both. | ||
| 199 | switch format { | ||
| 200 | case formatGzip: | ||
| 201 | err = decompress(url, tmp, converting) | ||
| 202 | case formatQcow2: | ||
| 203 | err = c.convert(ctx, url, tmp, converting) | ||
| 204 | } | ||
| 205 | if err != nil { | ||
| 206 | return "", err | ||
| 195 | } | 207 | } |
| 196 | if err := os.Rename(converting, final); err != nil { | 208 | if err := os.Rename(converting, final); err != nil { |
| 197 | return "", fmt.Errorf("imagecache: rename to final: %w", err) | 209 | return "", fmt.Errorf("imagecache: rename to final: %w", err) |
| @@ -200,6 +212,65 @@ func (c *Cache) ensureOnce(ctx context.Context, url, sha string) (string, error) | |||
| 200 | return final, nil | 212 | return final, nil |
| 201 | } | 213 | } |
| 202 | 214 | ||
| 215 | // decompress writes the gzipped raw image at src to dst, punching the long runs | ||
| 216 | // of zeros a cloud image is mostly made of into holes. | ||
| 217 | func decompress(url, src, dst string) error { | ||
| 218 | in, err := os.Open(src) | ||
| 219 | if err != nil { | ||
| 220 | return fmt.Errorf("imagecache: open download: %w", err) | ||
| 221 | } | ||
| 222 | defer in.Close() | ||
| 223 | zr, err := gzip.NewReader(in) | ||
| 224 | if err != nil { | ||
| 225 | // The download's sha already matched, so a broken gzip header is a | ||
| 226 | // property of the published artifact: re-fetching returns the same bytes | ||
| 227 | // and fails the same way. | ||
| 228 | return permanentf("imagecache: gzip header: %v", err) | ||
| 229 | } | ||
| 230 | defer zr.Close() | ||
| 231 | |||
| 232 | // Gzip wraps a RAW image. Sniff what is inside before writing a byte: a | ||
| 233 | // .qcow2.gz would otherwise decompress into a cache entry named <sha>.raw | ||
| 234 | // holding qcow2 bytes, and the guest would fail to boot for a reason nothing | ||
| 235 | // in the error names — exactly the illegible failure this package exists to | ||
| 236 | // prevent. The peeked bytes go back in front of the stream so the writer | ||
| 237 | // still sees the whole image. Failing before the output file is created is | ||
| 238 | // what leaves no temp behind. | ||
| 239 | head := make([]byte, magicLen) | ||
| 240 | n, err := io.ReadFull(zr, head) | ||
| 241 | if err != nil && !errors.Is(err, io.EOF) && !errors.Is(err, io.ErrUnexpectedEOF) { | ||
| 242 | return fmt.Errorf("imagecache: read decompressed header: %w", err) | ||
| 243 | } | ||
| 244 | head = head[:n] | ||
| 245 | if sniff(head) == formatQcow2 { | ||
| 246 | // Permanent for the same reason a broken header is: the sha matched, so | ||
| 247 | // these bytes are what the publisher published. | ||
| 248 | return permanentf("imagecache: %s decompresses to a qcow2 image; eitri guest images must be raw or gzipped raw (.raw.gz) — convert it with `qemu-img convert -O raw` where you publish it", url) | ||
| 249 | } | ||
| 250 | |||
| 251 | out, err := os.OpenFile(dst, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o644) | ||
| 252 | if err != nil { | ||
| 253 | return fmt.Errorf("imagecache: create image: %w", err) | ||
| 254 | } | ||
| 255 | if err := writeSparse(out, io.MultiReader(bytes.NewReader(head), zr)); err != nil { | ||
| 256 | out.Close() | ||
| 257 | return fmt.Errorf("imagecache: decompress: %w", err) | ||
| 258 | } | ||
| 259 | if err := out.Close(); err != nil { | ||
| 260 | return fmt.Errorf("imagecache: close image: %w", err) | ||
| 261 | } | ||
| 262 | return nil | ||
| 263 | } | ||
| 264 | |||
| 265 | // convert turns a qcow2 download into a raw image with qemu-img — the one | ||
| 266 | // format that still needs a tool, and the one a host without it cannot boot. | ||
| 267 | func (c *Cache) convert(ctx context.Context, url, src, dst string) error { | ||
| 268 | if _, err := c.run(ctx, "qemu-img", "convert", "-O", "raw", src, dst); err != nil { | ||
| 269 | return fmt.Errorf("qemu-img convert %s: %w", url, err) | ||
| 270 | } | ||
| 271 | return nil | ||
| 272 | } | ||
| 273 | |||
| 203 | // SweepTemps removes abandoned download and convert temporaries — the two temp | 274 | // SweepTemps removes abandoned download and convert temporaries — the two temp |
| 204 | // forms Ensure writes before its atomic rename. Neither matches evict's "*.raw" | 275 | // forms Ensure writes before its atomic rename. Neither matches evict's "*.raw" |
| 205 | // glob, so nothing else ever reclaims them, and an agent killed mid-fetch (the | 276 | // glob, so nothing else ever reclaims them, and an agent killed mid-fetch (the |
internal/agent/imagecache/imagecache_test.go
| Old | New | ||
|---|---|---|---|
| @@ -2,6 +2,7 @@ package imagecache | |||
| 2 | 2 | ||
| 3 | import ( | 3 | import ( |
| 4 | "bytes" | 4 | "bytes" |
| 5 | "compress/gzip" | ||
| 5 | "context" | 6 | "context" |
| 6 | "crypto/sha256" | 7 | "crypto/sha256" |
| 7 | "encoding/hex" | 8 | "encoding/hex" |
| @@ -342,3 +343,85 @@ func TestConcurrentEnsureFetchesOnce(t *testing.T) { | |||
| 342 | require.NoError(t, err) | 343 | require.NoError(t, err) |
| 343 | assert.Equal(t, body, data, "the shared result must be the real image") | 344 | assert.Equal(t, body, data, "the shared result must be the real image") |
| 344 | } | 345 | } |
| 346 | |||
| 347 | // gzipped returns body compressed, the shape a published .raw.gz has. | ||
| 348 | func gzipped(t *testing.T, body []byte) []byte { | ||
| 349 | t.Helper() | ||
| 350 | var buf bytes.Buffer | ||
| 351 | zw := gzip.NewWriter(&buf) | ||
| 352 | _, err := zw.Write(body) | ||
| 353 | require.NoError(t, err) | ||
| 354 | require.NoError(t, zw.Close()) | ||
| 355 | return buf.Bytes() | ||
| 356 | } | ||
| 357 | |||
| 358 | // qcow2 prefixes body with the qcow2 magic, so a test image is classified the | ||
| 359 | // way a real one would be. | ||
| 360 | func qcow2(body []byte) []byte { | ||
| 361 | return append([]byte{'Q', 'F', 'I', 0xfb}, body...) | ||
| 362 | } | ||
| 363 | |||
| 364 | // TestEnsureDecompressesGzippedRaw pins the format that keeps a published raw | ||
| 365 | // image smaller than the qcow2 it replaces: gzip is undone in-process, byte for | ||
| 366 | // byte, with no external tool. | ||
| 367 | func TestEnsureDecompressesGzippedRaw(t *testing.T) { | ||
| 368 | raw := concat( | ||
| 369 | bytes.Repeat([]byte{0xeb}, 4096), | ||
| 370 | make([]byte, 2<<20), | ||
| 371 | bytes.Repeat([]byte{0x77}, 4096), | ||
| 372 | ) | ||
| 373 | ts, sum := serve(t, gzipped(t, raw)) | ||
| 374 | dir := t.TempDir() | ||
| 375 | c := New(dir, noRunner(t)) | ||
| 376 | |||
| 377 | p, err := c.Ensure(context.Background(), ts.URL+"/img.raw.gz", sum) | ||
| 378 | require.NoError(t, err) | ||
| 379 | |||
| 380 | got, err := os.ReadFile(p) | ||
| 381 | require.NoError(t, err) | ||
| 382 | assert.True(t, bytes.Equal(raw, got), "decompressed image must be byte-identical") | ||
| 383 | |||
| 384 | fi, err := os.Stat(p) | ||
| 385 | require.NoError(t, err) | ||
| 386 | assert.Equal(t, int64(len(raw)), fi.Size()) | ||
| 387 | |||
| 388 | entries, err := os.ReadDir(dir) | ||
| 389 | require.NoError(t, err) | ||
| 390 | assert.Len(t, entries, 1, "no temp may survive a successful decompress") | ||
| 391 | } | ||
| 392 | |||
| 393 | // TestEnsureRejectsCorruptGzipPermanently pins the classification: the sha | ||
| 394 | // already matched, so the bytes are exactly what the publisher published and | ||
| 395 | // re-downloading them cannot produce a different outcome. | ||
| 396 | func TestEnsureRejectsCorruptGzipPermanently(t *testing.T) { | ||
| 397 | ts, sum := serve(t, []byte{0x1f, 0x8b, 'g', 'a', 'r', 'b', 'a', 'g', 'e', '!'}) | ||
| 398 | c := New(t.TempDir(), noRunner(t)) | ||
| 399 | |||
| 400 | _, err := c.Ensure(context.Background(), ts.URL+"/img.raw.gz", sum) | ||
| 401 | require.Error(t, err) | ||
| 402 | var p interface{ Permanent() bool } | ||
| 403 | require.ErrorAs(t, err, &p, "a broken gzip header must be marked permanent") | ||
| 404 | assert.True(t, p.Permanent()) | ||
| 405 | } | ||
| 406 | |||
| 407 | // TestEnsureRejectsGzippedQcow2Permanently pins the one trap gzip opens: gzip | ||
| 408 | // wraps a RAW image, and a .qcow2.gz would otherwise decompress into a cache | ||
| 409 | // entry named <sha>.raw holding qcow2 bytes — a guest that fails to boot for a | ||
| 410 | // reason no error names. Permanent because the sha already matched: the bytes | ||
| 411 | // are what the publisher published, and re-fetching cannot change them. | ||
| 412 | func TestEnsureRejectsGzippedQcow2Permanently(t *testing.T) { | ||
| 413 | ts, sum := serve(t, gzipped(t, qcow2(bytes.Repeat([]byte{0x42}, 4096)))) | ||
| 414 | dir := t.TempDir() | ||
| 415 | c := New(dir, noRunner(t)) | ||
| 416 | |||
| 417 | _, err := c.Ensure(context.Background(), ts.URL+"/img.qcow2.gz", sum) | ||
| 418 | require.Error(t, err) | ||
| 419 | var p interface{ Permanent() bool } | ||
| 420 | require.ErrorAs(t, err, &p, "a gzipped qcow2 can never decompress to a raw image") | ||
| 421 | assert.True(t, p.Permanent()) | ||
| 422 | assert.Contains(t, err.Error(), "qcow2", "the operator must learn what the image actually is") | ||
| 423 | |||
| 424 | entries, err := os.ReadDir(dir) | ||
| 425 | require.NoError(t, err) | ||
| 426 | assert.Empty(t, entries, "a rejected image must leave nothing in the cache") | ||
| 427 | } | ||
internal/agent/imagecache/sparse.go
| Old | New | ||
|---|---|---|---|
| @@ -0,0 +1,69 @@ | |||
| 1 | package imagecache | ||
| 2 | |||
| 3 | import ( | ||
| 4 | "bytes" | ||
| 5 | "errors" | ||
| 6 | "io" | ||
| 7 | "os" | ||
| 8 | ) | ||
| 9 | |||
| 10 | const ( | ||
| 11 | // blockSize is the granularity at which zeros are detected. A hole can only | ||
| 12 | // start on a filesystem block boundary, so 4 KiB — the block size on ext4, | ||
| 13 | // xfs, btrfs and APFS — is the finest granularity that buys anything. | ||
| 14 | blockSize = 4096 | ||
| 15 | // bufSize keeps reads chunky. It is a multiple of blockSize so every buffer | ||
| 16 | // boundary is also a block boundary, and a zero run crossing two buffers is | ||
| 17 | // still punched as one hole. | ||
| 18 | bufSize = 1 << 20 | ||
| 19 | ) | ||
| 20 | |||
| 21 | // zeroBlock is the comparand for hole detection: bytes.Equal against it is one | ||
| 22 | // memequal, which is why scanning a multi-gigabyte image block by block costs | ||
| 23 | // nothing next to the decompression feeding it. | ||
| 24 | var zeroBlock = make([]byte, blockSize) | ||
| 25 | |||
| 26 | // writeSparse copies src into dst, seeking past all-zero blocks instead of | ||
| 27 | // writing them, so a decompressed cloud image — which is mostly zeros — costs | ||
| 28 | // only the disk its real data occupies. | ||
| 29 | // | ||
| 30 | // dst ends at the stream's exact apparent size. That needs the closing | ||
| 31 | // Truncate: seeking past the end of a file does not extend it, so a stream | ||
| 32 | // ending in zeros would otherwise leave a file shorter than the image it | ||
| 33 | // claims to be, and PrepareRootDisk sizes a guest's disk against that number. | ||
| 34 | func writeSparse(dst *os.File, src io.Reader) error { | ||
| 35 | buf := make([]byte, bufSize) | ||
| 36 | var size int64 // logical bytes consumed, holes included | ||
| 37 | var hole int64 // consecutive zero bytes not yet seeked past | ||
| 38 | for { | ||
| 39 | n, readErr := io.ReadFull(src, buf) | ||
| 40 | for chunk := buf[:n]; len(chunk) > 0; { | ||
| 41 | b := chunk | ||
| 42 | if len(b) > blockSize { | ||
| 43 | b = b[:blockSize] | ||
| 44 | } | ||
| 45 | chunk = chunk[len(b):] | ||
| 46 | size += int64(len(b)) | ||
| 47 | if bytes.Equal(b, zeroBlock[:len(b)]) { | ||
| 48 | hole += int64(len(b)) | ||
| 49 | continue | ||
| 50 | } | ||
| 51 | if hole > 0 { | ||
| 52 | if _, err := dst.Seek(hole, io.SeekCurrent); err != nil { | ||
| 53 | return err | ||
| 54 | } | ||
| 55 | hole = 0 | ||
| 56 | } | ||
| 57 | if _, err := dst.Write(b); err != nil { | ||
| 58 | return err | ||
| 59 | } | ||
| 60 | } | ||
| 61 | if errors.Is(readErr, io.EOF) || errors.Is(readErr, io.ErrUnexpectedEOF) { | ||
| 62 | break | ||
| 63 | } | ||
| 64 | if readErr != nil { | ||
| 65 | return readErr | ||
| 66 | } | ||
| 67 | } | ||
| 68 | return dst.Truncate(size) | ||
| 69 | } | ||
internal/agent/imagecache/sparse_test.go
| Old | New | ||
|---|---|---|---|
| @@ -0,0 +1,127 @@ | |||
| 1 | package imagecache | ||
| 2 | |||
| 3 | import ( | ||
| 4 | "bytes" | ||
| 5 | "os" | ||
| 6 | "path/filepath" | ||
| 7 | "syscall" | ||
| 8 | "testing" | ||
| 9 | |||
| 10 | "github.com/stretchr/testify/assert" | ||
| 11 | "github.com/stretchr/testify/require" | ||
| 12 | ) | ||
| 13 | |||
| 14 | // allocatedBytes reports how much disk a file actually occupies, as opposed to | ||
| 15 | // the size it claims. st_blocks is in 512-byte units by POSIX definition, | ||
| 16 | // regardless of the filesystem's own block size. | ||
| 17 | func allocatedBytes(t *testing.T, path string) int64 { | ||
| 18 | t.Helper() | ||
| 19 | fi, err := os.Stat(path) | ||
| 20 | require.NoError(t, err) | ||
| 21 | st, ok := fi.Sys().(*syscall.Stat_t) | ||
| 22 | require.True(t, ok, "no syscall.Stat_t for %s", path) | ||
| 23 | return st.Blocks * 512 | ||
| 24 | } | ||
| 25 | |||
| 26 | // requireSparseFS skips the calling test when the temp filesystem cannot hold | ||
| 27 | // holes at all. Without this probe a sparseness assertion is really an | ||
| 28 | // assertion about the machine it runs on. | ||
| 29 | func requireSparseFS(t *testing.T, dir string) { | ||
| 30 | t.Helper() | ||
| 31 | p := filepath.Join(dir, "sparse-probe") | ||
| 32 | f, err := os.Create(p) | ||
| 33 | require.NoError(t, err) | ||
| 34 | require.NoError(t, f.Truncate(8<<20)) | ||
| 35 | require.NoError(t, f.Close()) | ||
| 36 | alloc := allocatedBytes(t, p) | ||
| 37 | require.NoError(t, os.Remove(p)) | ||
| 38 | if alloc > 1<<20 { | ||
| 39 | t.Skipf("filesystem under %s does not support sparse files (8 MiB hole allocated %d bytes)", dir, alloc) | ||
| 40 | } | ||
| 41 | } | ||
| 42 | |||
| 43 | // TestWriteSparse pins both halves of the contract at once: the file that comes | ||
| 44 | // out is byte-identical to the stream that went in AND its apparent size is | ||
| 45 | // exact, while the runs of zeros cost no disk. | ||
| 46 | func TestWriteSparse(t *testing.T) { | ||
| 47 | data := bytes.Repeat([]byte{0xAB}, blockSize) | ||
| 48 | |||
| 49 | tests := []struct { | ||
| 50 | name string | ||
| 51 | body []byte | ||
| 52 | sparse bool // must occupy far less disk than it claims | ||
| 53 | }{ | ||
| 54 | { | ||
| 55 | name: "solid data only", | ||
| 56 | body: bytes.Repeat([]byte{0x5A}, 3*blockSize), | ||
| 57 | }, | ||
| 58 | { | ||
| 59 | name: "hole between two data blocks", | ||
| 60 | body: concat(data, make([]byte, 8<<20), data), | ||
| 61 | sparse: true, | ||
| 62 | }, | ||
| 63 | { | ||
| 64 | name: "leading hole", | ||
| 65 | body: concat(make([]byte, 8<<20), data), | ||
| 66 | sparse: true, | ||
| 67 | }, | ||
| 68 | { | ||
| 69 | name: "trailing hole: apparent size must survive it", | ||
| 70 | body: concat(data, make([]byte, 8<<20)), | ||
| 71 | sparse: true, | ||
| 72 | }, | ||
| 73 | { | ||
| 74 | name: "zero run spanning several read buffers", | ||
| 75 | body: concat(data, make([]byte, 3*bufSize), data), | ||
| 76 | sparse: true, | ||
| 77 | }, | ||
| 78 | { | ||
| 79 | name: "sub-block tail", | ||
| 80 | body: concat(data, []byte("tail")), | ||
| 81 | }, | ||
| 82 | { | ||
| 83 | name: "unaligned data after a hole", | ||
| 84 | body: concat(make([]byte, 8<<20), []byte("x")), | ||
| 85 | }, | ||
| 86 | { | ||
| 87 | name: "empty stream", | ||
| 88 | body: nil, | ||
| 89 | }, | ||
| 90 | } | ||
| 91 | |||
| 92 | for _, tc := range tests { | ||
| 93 | t.Run(tc.name, func(t *testing.T) { | ||
| 94 | dir := t.TempDir() | ||
| 95 | if tc.sparse { | ||
| 96 | requireSparseFS(t, dir) | ||
| 97 | } | ||
| 98 | p := filepath.Join(dir, "out.raw") | ||
| 99 | f, err := os.Create(p) | ||
| 100 | require.NoError(t, err) | ||
| 101 | require.NoError(t, writeSparse(f, bytes.NewReader(tc.body))) | ||
| 102 | require.NoError(t, f.Close()) | ||
| 103 | |||
| 104 | got, err := os.ReadFile(p) | ||
| 105 | require.NoError(t, err) | ||
| 106 | assert.True(t, bytes.Equal(tc.body, got), "decompressed bytes must be identical") | ||
| 107 | |||
| 108 | fi, err := os.Stat(p) | ||
| 109 | require.NoError(t, err) | ||
| 110 | assert.Equal(t, int64(len(tc.body)), fi.Size(), "apparent size must be exact") | ||
| 111 | |||
| 112 | if tc.sparse { | ||
| 113 | alloc := allocatedBytes(t, p) | ||
| 114 | assert.Less(t, alloc, fi.Size()/2, | ||
| 115 | "zero runs must be holes: %d bytes allocated for an apparent %d", alloc, fi.Size()) | ||
| 116 | } | ||
| 117 | }) | ||
| 118 | } | ||
| 119 | } | ||
| 120 | |||
| 121 | func concat(parts ...[]byte) []byte { | ||
| 122 | var out []byte | ||
| 123 | for _, p := range parts { | ||
| 124 | out = append(out, p...) | ||
| 125 | } | ||
| 126 | return out | ||
| 127 | } | ||