Files
pdfcraft/crates/cos
Grzegorz MartensandClaude Opus 5.5 6771ddffe8 M2.6: inflate with zlib-rs everywhere, so reading Flate costs 2 steps a byte
The Resources review of the commits before this one found that pricing what Flate reads at 128
steps a byte refused real redactions. Redaction decodes a picture once for
each page that draws it, so 240 pages sharing a 1 MB scan-like Flate picture,
a mark on each, went from Ok (240 cleared, 20.3 s) to ImagesPastAllowance, its
allowance paying for 188 decodes: each paid about 27 times what it had.

The cost being priced is miniz_oxide's: about 1.7 us for every fixed-Huffman
block, empty or not. flate2's zlib-rs backend took about 11 to 13 ns a fixed block,
and the Windows build already used it (craft-update turns
on zip's deflate-flate2-zlib-rs, and with both backends on, flate2 inflates
with zlib-rs, measured). Every platform now builds flate2 with zlib-rs, which
it prefers to the miniz_oxide other dependencies still build.
In a release build (zlib-rs 0.6.8), per byte read: empty fixed blocks 9 ns,
fixed blocks of one literal 6, empty dynamic 21, empty stored 1.3, against
1,383, 768, 106 and 2 before. Each backend inflating what its own
encoder made, 16 MiB of text took 2.6 ms against 6.3 to 13.7, and a
scan-like picture 38 ms against 53.

FLATE_READ_WORK is now 2: per byte read, the worst shape comes to about 7 ns a
step, and a picture an encoder made pays twice its stored length more than unpriced
(an_image_an_encoder_made_pays_little_for_its_reading: at most half again).

The same review: export and compare's Office export read an indexed image's
lookup table outside the allowance, once per image, even with nothing left
(1.4 s each for 1 MiB of empty blocks under miniz_oxide). It is read through
ImageBudget::decode_filters now, and refused past it
(a_lookup_table_is_read_within_the_exports_allowance).

filters, cos, optimize, redact, create, edit, export and engine pass with the
new backend.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AmsdeUCdTaD72jkDw2pJKY
2026-11-03 00:14:52 +01:00
..

pdfcraft-cos

The PDF object layer (ISO 32000-2 §7): the COS object graph, parsed lazily and tolerantly, edited through a copy-on-write overlay, and written back incrementally or as a full rewrite.

  • Layer: L1, standalone. The layering table lets it use only pdfcraft-filters, pdfcraft-crypt and pdfcraft-geom (xtask/src/layers.rs::STANDALONE_DEPS); it currently uses the first two. That keeps the parser and the crypto independently auditable and reusable.
  • #![forbid(unsafe_code)], and the crate denies unwrap/expect/panic!/todo!. Builds for wasm32-unknown-unknown.

The object graph is the document model. Nothing is lifted into a tidier representation and written back out, because that is how data gets silently dropped. Typed views live above, in pdfcraft-model.

API

let doc = Document::open(bytes)?;                               // Arc<Vec<u8>>
let doc = Document::open_with_password(bytes, Some("owner"))?;  // encrypted files

let root = doc.root();                               // Option<ObjRef> — the catalog
let obj  = doc.get(ObjRef { num: 12, gen: 0 });      // load (and cache): Arc<Object>
let real = doc.resolve(&obj);                        // follow one indirect reference
let d    = doc.dict(&obj);                           // Option<Dict>, resolving first
let data = stream.decoded()?;                        // filters applied on demand
let data = doc.with_direct_parms(&stream).decoded()?; // /DecodeParms given by reference too
let px   = doc.decoded_image(&stream)?;              // an image's samples, a CCITT fax image's too
let fax  = doc.fax_params(&stream);                  // a CCITT fax image's /DecodeParms, Table 11 defaults

doc.set(r, Object::Dict(d));                         // edit: lands in the overlay
let r = doc.add(obj);                                // append a new object
doc.free(r);

let bytes = write_incremental(&doc, &SaveOptions::default())?;  // append a revision
let bytes = write_full(&doc, &SaveOptions::default())?;         // garbage-collected rewrite

Both writers return the bytes; the caller decides where they go, which is what lets the engine produce a working file in memory after every edit and keep saving atomic.

Useful alongside: doc.repair_log() (what was reconstructed), doc.revisions() and doc.revision_ends() (the incremental history), doc.is_edited(num) and doc.modified_objects(), doc.permissions() and doc.security(), doc.set_encryption() / doc.remove_encryption(), and doc.require_full_save() for edits that cannot be expressed incrementally.

Document is cheap to clone: the original bytes, the xref index and the parse caches are shared, so a clone is a snapshot. Undo/redo and background jobs hold clones while the UI edits another. Edits never touch the original bytes.

Reading

  • Every cross-reference form: classic tables with their quirky subsections, cross-reference streams, hybrid-reference files (/XRefStm), and object streams.
  • Repair. A file whose xref is wrong or missing is reconstructed by scanning for obj headers. Repairs are recorded, never silent, so the UI can say the file was damaged.
  • Tolerant, never fatal. The lexer works on a byte slice and a position; it allocates nothing for skipped data, bounds-checks every index, and caps nesting depth. Malformed input yields CosError, never a panic — this crate is the first thing a hostile file meets. Exhausted xref entries are skipped with a repair log; exhausted /Size hints allocate unused numbers without overflowing or replacing existing objects.
  • Lazy. Objects are parsed on first use and cached. (Whole-file buffering is still the case; ByteSource streaming is core.lazy-loading, not yet done.)
  • Encryption is handled on load and save through pdfcraft-crypt: every standard-security revision R2–R6 (RC4, AES-128/256), crypt filters, /EFF embedded-file encryption, and /EncryptMetadata. security.rs knows what the spec exempts — the /Encrypt dictionary, xref streams, the trailer /ID, and objects inside an object stream (the stream is encrypted whole).

Writing

write_incremental appends only the edited objects, a cross-reference section in the same style as the file's last one, and a trailer with /Prev. The original bytes are never modified, which is what keeps a signature over an earlier revision valid.

write_full writes only what the trailer can reach, renumbered densely — the Save As and garbage-collecting path. By default (SaveOptions::object_streams) it packs non-stream objects into compressed object streams with a cross-reference stream (PDF 1.5+); turning that off writes a classic table any reader accepts.

Dictionaries keep their key order and streams keep their encoded bytes, so rewritten objects stay recognisable in a diff and byte-stable when re-serialized.

Errors

CosError: NotPdf, Syntax { offset, detail }, MissingObject, NotADictionary, Filter, NeedsPassword, WrongPassword, Security, Poisoned.

Not done here

  • Streaming very large files within a memory budget (ByteSource, core.lazy-loading); files are read whole today.
  • Linearized output, Fast Web View (core.linearization, M11).
  • Structural validation against the Arlington model (core.arlington-validation), which will live in the reserved arlington crate.
  • Image codecs. pdfcraft-filters carries the non-image filters and CCITT fax and JBIG2 decoding (Document::decoded_image, a JBIG2 image's /JBIG2Globals within MAX_JBIG2_GLOBALS, 32 MiB, their filters undone once a document for the images that share them, up to 64 MiB kept, while their segments are decoded for every image, and ImageBudget for an operation over many images, which share one allowance of decoding, filters, decoders and pixels alike, and are refused past it: IMAGE_WORK, 2^34 steps, or 2 to 5 minutes of JBIG2 and about 1,000 pages scanned at 300 dpi, plus IMAGE_WORK_PER_BYTE, 2^13 steps for each byte of the file); DCT and JPX are still decoded through the bootstrap renderer (core.image-filters, M2.6).