Files
syft/cmd/syft/internal/commands/convert.go
T
JuanandAlex Goodman e6d364437a feat(convert): passthrough-exact-format copies matching syft-json input unchanged (#5279)
* feat(convert): passthrough-exact-format copies matching syft-json input unchanged

When a syft-json input SBOM already exactly matches the requested output
format and schema version, `syft convert` can write the input bytes out
verbatim instead of decoding and re-encoding them. This avoids the lossy
round trip (and the associated cost) for callers that normalise SBOMs into
a target format without knowing up front whether the source already matches.

The behaviour is opt-in and configuration-only: it is enabled with
convert.passthrough-exact-format in the config file or the
SYFT_CONVERT_PASSTHROUGH_EXACT_FORMAT environment variable, and there is
deliberately no command-line flag, since it changes how every invocation in
a pipeline behaves rather than varying call by call. It defaults to off, so
existing behaviour is unchanged.

Pass-through applies only to syft-json input, and only when the document's
schema version is identical to the one this build of syft produces; any
other input is always converted. Outputs that do not match are still
converted, so a mixed set of `-o` values behaves as expected, and every
output is resolved and validated before anything is written. Pass-through
output is a byte-for-byte copy, so format options such as `format.pretty`
and `format.json.legacy` are not applied to it.

When the option is enabled a warning is logged stating so and noting the
syft-json-only scope, and a second warning is logged whenever an input is
found to match and the shortcut is actually taken.

The option lives under a `convert:` configuration section (options.Convert),
mirroring how attest scopes its own options. The description documents the
caveats above and notes that file-path input written with `-o <format>=<path>`
keeps memory minimal, whereas piped input or STDOUT output must hold the
whole document in memory.

Signed-off-by: juanjsebgarcia <github@juangarcia.co.uk>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>

* perf(convert): stream pass-through output instead of buffering the input

Identify, copy, and decode the input through a seekable reader rather than
loading the whole document into memory first. File inputs are read in place
and pass-through output is an io.Copy from the file. STDIN is used in place
when it is a redirected regular file. A pipe or terminal cannot seek, so its
contents are drained into an internal/spillbuf buffer: the first 1MB stays in
memory and the rest goes to a temp file under the user's TMPDIR (created
lazily, only when something actually spills), giving a seekable view without
ever holding the document in memory. The temp dir is rooted per invocation
via internal/tmpdir, honouring one already attached to the context, and is
removed when the command returns.

Writing pass-through output over the input file itself is now a no-op rather
than truncating the file before it is copied.

Measured on a 1GB syft-json document (Apple M1 Max), file in, file out:

  origin/main full re-encode      86.3s wall   99s CPU   12.0GB max RSS
  in-memory pass-through           7.4s wall  6.7s CPU    4.2GB max RSS
  streaming pass-through           7.0s wall  6.9s CPU    3.1GB max RSS

and the same 1GB document piped over STDIN with pass-through enabled:

  io.ReadAll into memory           1.3s wall  0.8s CPU    2.7GB max RSS
  spill buffer                     1.8s wall  1.0s CPU     66MB max RSS

The remaining cost of the streaming pass-through is entirely format.Identify,
which for syft-json decodes the whole document through a json.Decoder that
buffers the full value: measured alone it takes 7.1s and 3.1GB on the same
input. Reducing that further belongs to the format package.

Signed-off-by: juanjsebgarcia <github@juangarcia.co.uk>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>

* perf(syftjson): identify documents from their tail instead of a full parse

syft always writes the schema block as the last key of a syft-json document,
so the format and version can be read from the final bytes of the file. The
syft-json decoder's Identify now reads at most 2MB from the end of a seekable
input, finds the last "schema" key there, and verifies it sits directly within
the top-level object by scanning the remaining bytes for the single closing
brace of the document. Anything the tail cannot vouch for (non-seekable input,
a reordered document, a nested key) falls back to the existing full parse, so
behaviour is unchanged for documents the fast path declines.

The verification is position-independent: it also accepts documents whose
keys have been reordered so that other small keys follow schema. It rejects
nested schema objects, array-wrapped documents, truncated input, trailing
content, and schema text inside strings.

Supporting changes:
- the offset-aware reader the decoder collection hands to decoders now
  supports io.SeekEnd, which the tail read requires
- model.Document documents that Schema must remain its last field, and a
  test guards that the encoder writes it last in both compact and pretty
  output
- the decoder list documents that syft-json must stay first

Since Decode calls Identify before parsing, this also removes one of the two
full parses every syft-json conversion used to perform.

Measured on a 1GB syft-json document (Apple M1 Max), file in, file out:

  pass-through, previous commit     7.0s wall   6.9s CPU   3.1GB max RSS
  pass-through, this commit        0.85s wall  0.05s CPU    66MB max RSS
  full re-encode, previous commit  85.3s wall   99s CPU  11.5GB max RSS
  full re-encode, this commit      75.2s wall   91s CPU   9.9GB max RSS

Signed-off-by: juanjsebgarcia <github@juangarcia.co.uk>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>

* fix(syftjson): only identify from the tail when the document starts where the reader does

The tail scan vouched for the last 2MB, but nothing checked those bytes belonged to the
document the reader was pointed at. So `Identify` started claiming syft-json for files it
cannot decode: a document with a byte order mark or a stray log line in front of it, or a
syft document sitting after something else in the same stream, where `format.Decode` went
on to return an empty SBOM with a nil error.

Two O(1) checks:

- the first non-space byte at the reader's position has to be `{`
- `Decode` errors when the parsed document has no `schema.version`, instead of handing back
  the empty document it decoded into

Streams holding several concatenated documents are deliberately not detected. Proving there
is exactly one top-level value means reading all of it, which is the cost the tail path
exists to avoid, so `Identify` can still name a later document's version on such a file.
`Decode` reports what it actually parsed.

Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>

* fix(convert): recognize the input file by identity rather than by path

`convert - -o syft-json=sbom.json < sbom.json` truncated the file and exited 0. The
passthrough guard compared paths, and a shell redirect does not have one, so it was blind to
the one input mode where the truncation was silent.

`openConvertInput` now returns the input's `os.FileInfo`, which the STDIN branch already had
from its `Stat`, and `isSameFile` compares that instead.

Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>

---------

Signed-off-by: juanjsebgarcia <github@juangarcia.co.uk>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
2026-09-14 14:33:03 -04:00

316 lines
11 KiB
Go

package commands
import (
"context"
"fmt"
"io"
"os"
"path/filepath"
"strings"
"github.com/spf13/cobra"
"github.com/anchore/clio"
"github.com/anchore/go-homedir"
"github.com/anchore/syft/cmd/syft/internal/options"
"github.com/anchore/syft/cmd/syft/internal/ui"
"github.com/anchore/syft/internal"
"github.com/anchore/syft/internal/bus"
"github.com/anchore/syft/internal/log"
"github.com/anchore/syft/internal/spillbuf"
"github.com/anchore/syft/internal/tmpdir"
"github.com/anchore/syft/syft/format"
"github.com/anchore/syft/syft/format/syftjson"
"github.com/anchore/syft/syft/sbom"
)
const (
convertExample = ` {{.appName}} {{.command}} img.syft.json -o spdx-json convert a syft SBOM to spdx-json, output goes to stdout
{{.appName}} {{.command}} img.syft.json -o cyclonedx-json=img.cdx.json convert a syft SBOM to CycloneDX, output is written to the file "img.cdx.json"
{{.appName}} {{.command}} - -o spdx-json convert an SBOM from STDIN to spdx-json
SYFT_CONVERT_PASSTHROUGH_EXACT_FORMAT=true {{.appName}} {{.command}} img.syft.json -o syft-json=out.syft.json copy the input to "out.syft.json" unchanged if it is already syft-json at the current schema version
`
)
type ConvertOptions struct {
options.Config `yaml:",inline" mapstructure:",squash"`
options.Output `yaml:",inline" mapstructure:",squash"`
options.UpdateCheck `yaml:",inline" mapstructure:",squash"`
Convert options.Convert `yaml:"convert" json:"convert" mapstructure:"convert"`
}
func Convert(app clio.Application) *cobra.Command {
id := app.ID()
opts := &ConvertOptions{
UpdateCheck: options.DefaultUpdateCheck(),
Output: options.DefaultOutput(),
Convert: options.DefaultConvert(),
}
return app.SetupCommand(&cobra.Command{
Use: "convert [SOURCE-SBOM] -o [FORMAT]",
Short: "Convert between SBOM formats",
Long: "[Experimental] Convert SBOM files to, and from, SPDX, CycloneDX and Syft's format. For more info about data loss between formats see https://github.com/anchore/syft/wiki/format-conversion",
Example: internal.Tprintf(convertExample, map[string]any{
"appName": id.Name,
"command": "convert",
}),
Args: validateConvertArgs,
PreRunE: applicationUpdateCheck(id, &opts.UpdateCheck),
RunE: func(cmd *cobra.Command, args []string) error {
restoreStdout := ui.CaptureStdoutToTraceLog()
defer restoreStdout()
return RunConvert(cmd.Context(), opts, args[0])
},
}, opts)
}
func validateConvertArgs(cmd *cobra.Command, args []string) error {
return validateArgs(cmd, args, "an SBOM argument is required")
}
func RunConvert(ctx context.Context, opts *ConvertOptions, userInput string) error {
log.Warn("convert is an experimental feature, run `syft convert -h` for help")
// a temp dir is only needed when STDIN is a pipe and has to be spilled to disk; it is created lazily on
// first use (under the user's TMPDIR), so a file input never touches it. Mirrors syft.CreateSBOM, which
// also honors a TempDir already attached to the context.
td := tmpdir.FromContext(ctx)
if td == nil {
_, td = tmpdir.Root(ctx, "syft-convert")
defer func() {
_ = td.Cleanup()
}()
}
reader, inputInfo, err := openConvertInput(userInput, td)
if err != nil {
return err
}
defer func() {
_ = reader.Close()
}()
outputs := opts.Outputs
var unchanged []string
if opts.Convert.PassthroughExactFormat {
log.Warn("convert.passthrough-exact-format is enabled: syft-json input that already exactly matches a requested output format and schema version will be copied to that output unchanged (this does not apply to any other input format)")
unchanged, outputs, err = partitionOutputsBySourceFormat(opts.Output, reader)
if err != nil {
return err
}
}
// resolve (and validate) every output that needs converting before writing anything, so that
// a bad -o value fails the command without leaving partial results behind
var writer sbom.Writer
if len(outputs) > 0 {
outputOpts := opts.Output
outputOpts.Outputs = outputs
writer, err = outputOpts.SBOMWriter()
if err != nil {
return err
}
}
for _, output := range unchanged {
if err := writeUnchanged(output, opts.LegacyFile, inputInfo, reader); err != nil {
return err
}
}
if writer == nil {
return nil
}
if _, err := reader.Seek(0, io.SeekStart); err != nil {
return fmt.Errorf("unable to seek to start of SBOM: %w", err)
}
s, _, _, err := format.Decode(reader)
if err != nil {
return fmt.Errorf("failed to decode SBOM: %w", err)
}
if s == nil {
return fmt.Errorf("no SBOM produced")
}
if err := writer.Write(*s); err != nil {
return fmt.Errorf("failed to write SBOM: %w", err)
}
return nil
}
// openConvertInput opens the SBOM document at the given path, or STDIN when "-". The returned reader is seekable so
// the document can be identified, copied, and decoded in turn without being loaded into memory. The returned
// os.FileInfo describes the file the document is being read from, so that writing back over it can be recognized;
// it is nil when the input is not a file (a pipe or a terminal on STDIN).
func openConvertInput(userInput string, td *tmpdir.TempDir) (io.ReadSeekCloser, os.FileInfo, error) {
if userInput == "-" {
return openStdin(td)
}
f, err := os.Open(userInput)
if err != nil {
return nil, nil, fmt.Errorf("failed to open SBOM file: %w", err)
}
info, err := f.Stat()
if err != nil {
// the document is readable, we just cannot recognize it as an output destination
log.WithFields("error", err, "path", userInput).Debug("unable to stat SBOM file")
}
return f, info, nil
}
// openStdin returns a seekable view of STDIN. When STDIN is a regular file (such as a shell redirect) it is read in
// place. A pipe or terminal cannot seek (you will get errors such as "seek /dev/stdin: illegal seek"), so its
// contents are drained into a spill buffer: a small head stays in memory and everything beyond it goes to a temp
// file under td, so an arbitrarily large piped SBOM never has to be held in memory to be made seekable.
func openStdin(td *tmpdir.TempDir) (io.ReadSeekCloser, os.FileInfo, error) {
if info, err := os.Stdin.Stat(); err == nil && info.Mode().IsRegular() {
if start, err := os.Stdin.Seek(0, io.SeekCurrent); err == nil {
// a redirect reads the file in place, so it is just as much an input file as a path argument is
return readSeekNopCloser{io.NewSectionReader(os.Stdin, start, info.Size()-start)}, info, nil
}
}
buf := spillbuf.New(td)
if _, err := io.Copy(io.NewOffsetWriter(buf, 0), os.Stdin); err != nil {
_ = buf.Close()
return nil, nil, fmt.Errorf("failed to read SBOM from STDIN: %w", err)
}
return &spilledReader{
SectionReader: io.NewSectionReader(buf, 0, buf.Size()),
buf: buf,
}, nil, nil
}
type readSeekNopCloser struct {
io.ReadSeeker
}
func (readSeekNopCloser) Close() error {
return nil
}
// spilledReader is a seekable view over a spill buffer; closing it releases the buffer's memory and spill file.
type spilledReader struct {
*io.SectionReader
buf *spillbuf.Buffer
}
func (r *spilledReader) Close() error {
return r.buf.Close()
}
// partitionOutputsBySourceFormat splits the requested outputs (in "<format>[@<version>][=<path>]" form) into those
// that exactly match the format and schema version of a syft-json input document (which can be written out
// unchanged) and those that still require conversion. Input in any other format is always converted. Outputs that
// cannot be resolved to an encoder are left for the writer to report on.
func partitionOutputsBySourceFormat(output options.Output, reader io.ReadSeeker) (unchanged []string, toConvert []string, err error) {
if _, err := reader.Seek(0, io.SeekStart); err != nil {
return nil, nil, fmt.Errorf("unable to seek to start of SBOM: %w", err)
}
id, version := format.Identify(reader)
if id == "" || version == "" {
// let decoding surface the appropriate error
return nil, output.Outputs, nil
}
if id != syftjson.ID {
log.WithFields("format", id, "version", version).Debug("input is not syft-json, passthrough does not apply")
return nil, output.Outputs, nil
}
encoders, err := output.Encoders()
if err != nil {
return nil, nil, err
}
collection := format.NewEncoderCollection(encoders...)
for _, out := range output.Outputs {
name, _, _ := strings.Cut(strings.TrimSpace(out), "=")
enc := collection.GetByString(name)
if enc != nil && enc.ID() == id && enc.Version() == version {
log.WithFields("format", id, "version", version, "output", out).Warn("input already exactly matches the requested output format and schema version, copying it unchanged instead of converting it")
unchanged = append(unchanged, out)
continue
}
toConvert = append(toConvert, out)
}
return unchanged, toConvert, nil
}
// writeUnchanged copies the SBOM document as-is to the destination described by the given "<format>[=<path>]"
// output value, falling back to the (deprecated) --file path, and finally to the report bus (STDOUT). When the
// destination is the input file itself there is nothing to do.
func writeUnchanged(output string, defaultFile string, inputInfo os.FileInfo, reader io.ReadSeeker) error {
if _, err := reader.Seek(0, io.SeekStart); err != nil {
return fmt.Errorf("unable to seek to start of SBOM: %w", err)
}
_, path, _ := strings.Cut(strings.TrimSpace(output), "=")
if path == "" {
path = defaultFile
}
if path == "" {
// the report bus only carries strings, so STDOUT output must be materialized in memory
content, err := io.ReadAll(reader) //nolint:gocritic // the whole SBOM is the report; there is no meaningful bound
if err != nil {
return fmt.Errorf("unable to read SBOM: %w", err)
}
bus.Report(string(content))
return nil
}
expandedPath, err := homedir.Expand(path)
if err != nil {
log.Warnf("could not expand given writer output path=%q: %w", path, err)
expandedPath = path
}
if isSameFile(inputInfo, expandedPath) {
log.WithFields("path", expandedPath).Info("output is the input file, leaving it as-is")
return nil
}
if err := os.MkdirAll(filepath.Dir(expandedPath), 0755); err != nil {
return fmt.Errorf("unable to create report directory: %w", err)
}
f, err := os.OpenFile(expandedPath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0644)
if err != nil {
return fmt.Errorf("unable to create report file: %w", err)
}
if _, err := io.Copy(f, reader); err != nil {
_ = f.Close()
return fmt.Errorf("unable to write report file: %w", err)
}
return f.Close()
}
// isSameFile reports whether the given path names the file the input is being read from. Comparing the input's
// os.FileInfo rather than its path covers a shell redirect onto STDIN, which has no path to compare.
func isSameFile(inputInfo os.FileInfo, path string) bool {
if inputInfo == nil {
return false
}
pathInfo, err := os.Stat(path)
if err != nil {
return false
}
return os.SameFile(inputInfo, pathInfo)
}