perf(render): a picture that doesn't decode is decoded once, not at every draw (hayro-interpret patch 91)

The decoded-image cache stored a failed decode as an empty entry and read
an empty entry as a miss, so a picture or stencil mask that doesn't decode
(damaged data, a JPX or JBIG2 stream its decoder refuses) was decoded again
at every draw, and at every tile, preview and zoom of a render worker that
shares the cache.

The cache now keeps a failure for its key, with the warnings the decode
reported; a later draw reports the same warnings and gets nothing, as the
decode did. A failure counts 128 bytes towards the cache's 64 MiB budget.
A failure the page's JBIG2 allowance may have caused (ContentTruncated, or
the allowance spent) isn't kept, since the next page starts afresh.
Successful decodes are cached as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Pq7HyDBXygTf7jcqdsjcfn

crates/render/src/raster/tests.rs

@@ -8196,6 +8196,10 @@ fn decoded_images_serve_every_size_they_are_drawn_at() {
 // ICC profiles whose stream dictionaries are the same (vendored hayro-interpret patch 48).
 mod icc_profile_cache;
 
+// A picture that doesn't decode is decoded once however often it is drawn (vendored
+// hayro-interpret patch 91).
+mod failed_image_decodes;
+
 // A symbolic TrueType font draws through its (3, 0) cmap subtable first (#4156).
 mod symbolic_truetype_cmap;
 

crates/render/src/raster/tests/failed_image_decodes.rs

@@ -0,0 +1,216 @@
+//! A picture that doesn't decode is decoded once however often a page draws it (vendored
+//! hayro-interpret patch 91). The decoded-image cache kept only decodes that succeeded, so a
+//! damaged picture, or a JPX or JBIG2 stream its decoder refuses, was decoded again at every
+//! draw: a page drawing it a thousand times decoded it a thousand times. The failure is kept for
+//! the same key now, and every draw still reports it as it did.
+
+use super::image_cache_pdf;
+use hayro::hayro_interpret::font::Glyph;
+use hayro::hayro_interpret::hayro_syntax::Pdf;
+use hayro::hayro_interpret::{
+    BlendMode, ClipPath, Context, Device, GlyphDrawMode, Image, InterpreterCache, InterpreterSettings, InterpreterWarning, Paint, PathDrawMode,
+    SoftMask, TransformExt, interpret_page,
+};
+use kurbo::{Affine, BezPath, Rect};
+use std::io::Write;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+
+/// What a page handed its device: the raster images with their pixels, those of them with a
+/// mask, and the stencil masks with theirs. Every image is asked for at the size it is drawn
+/// at, as the renderer does.
+#[derive(Default)]
+struct Draws {
+    rasters: usize,
+    masked: usize,
+    stencils: usize,
+}
+
+impl<'a> Device<'a> for Draws {
+    fn set_soft_mask(&mut self, _: Option<SoftMask<'a>>) {}
+    fn set_blend_mode(&mut self, _: BlendMode) {}
+    fn draw_path(&mut self, _: &BezPath, _: Affine, _: &Paint<'a>, _: &PathDrawMode) {}
+    fn push_clip_path(&mut self, _: &ClipPath) {}
+    fn push_transparency_group(&mut self, _: f32, _: Option<SoftMask<'a>>, _: BlendMode) {}
+    fn draw_glyph(&mut self, _: &Glyph<'a>, _: Affine, _: Affine, _: &Paint<'a>, _: &GlyphDrawMode) {}
+    fn draw_image(&mut self, image: Image<'a, '_>, transform: Affine) {
+        let [a, b, c, d, ..] = transform.as_coeffs();
+        let target = ((a.hypot(b) * f64::from(image.width())).ceil() as u32, (c.hypot(d) * f64::from(image.height())).ceil() as u32);
+        match image {
+            Image::Raster(r) => r.with_rgba(
+                |_, alpha| {
+                    self.rasters += 1;
+                    self.masked += usize::from(alpha.is_some());
+                },
+                Some(target),
+            ),
+            Image::Stencil(s) => s.with_stencil(|_, _| self.stencils += 1, Some(target)),
+        }
+    }
+    fn pop_clip_path(&mut self) {}
+    fn pop_transparency_group(&mut self) {}
+}
+
+/// What drawing `/name` five times at one size, and then the picture that decodes (`/G`) once
+/// on the right, did: the decodes the cache ran, the image decode failures reported, and what
+/// the device was handed.
+struct Outcome {
+    decodes: usize,
+    failures: usize,
+    draws: Draws,
+}
+
+fn draw_five_times(name: &str) -> Outcome {
+    let flate = |data: &[u8]| {
+        let mut z = flate2::write::ZlibEncoder::new(Vec::new(), flate2::Compression::fast());
+        z.write_all(data).unwrap();
+        z.finish().unwrap()
+    };
+    let pixels = flate(&(0..64 * 64 * 3).map(|i| (i % 251) as u8).collect::<Vec<_>>());
+    let image = |filter: &str, extra: &str, len: usize| {
+        format!(
+            "<< /Type /XObject /Subtype /Image /Width 64 /Height 64 /ColorSpace /DeviceRGB /BitsPerComponent 8 /Filter /{filter} {extra}/Length {len} >>"
+        )
+    };
+    let good = image("FlateDecode", "", pixels.len());
+    // A JPEG 2000 picture whose codestream is four zero bytes.
+    let bad_jpx = image("JPXDecode", "", 4);
+    // A picture whose soft mask is such a JPEG 2000 stream.
+    let bad_mask = image("FlateDecode", "/SMask 9 0 R ", pixels.len());
+    // A stencil mask whose JBIG2 data is four zero bytes.
+    let bad_stencil = "<< /Type /XObject /Subtype /Image /Width 64 /Height 64 /ImageMask true /Filter /JBIG2Decode /Length 4 >>";
+    let jpx_mask =
+        "<< /Type /XObject /Subtype /Image /Width 64 /Height 64 /ColorSpace /DeviceGray /BitsPerComponent 8 /Filter /JPXDecode /Length 4 >>";
+    let content = [0, 24, 48, 72, 96].map(|x| format!("q 20 0 0 20 {x} 10 cm /{name} Do Q")).join(" ") + " q 10 0 0 10 118 10 cm /G Do Q";
+    let bytes = image_cache_pdf(
+        &content,
+        "<< /XObject << /G 5 0 R /J 6 0 R /S 7 0 R /M 8 0 R >> >>",
+        &[(&good, &pixels), (&bad_jpx, b"\0\0\0\0"), (&bad_mask, &pixels), (bad_stencil, b"\0\0\0\0"), (jpx_mask, b"\0\0\0\0")],
+    );
+    let parsed = Pdf::new(Arc::new(bytes)).expect("parses");
+    let page = &parsed.pages()[0];
+    let cache = InterpreterCache::new();
+    let failures = Arc::new(AtomicUsize::new(0));
+    let others = Arc::new(Mutex::new(Vec::new()));
+    let settings = InterpreterSettings {
+        warning_sink: {
+            let (failures, others) = (failures.clone(), others.clone());
+            Arc::new(move |warning| match warning {
+                InterpreterWarning::ImageDecodeFailure => {
+                    failures.fetch_add(1, Ordering::Relaxed);
+                }
+                other => others.lock().unwrap().push(format!("{other:?}")),
+            })
+        },
+        ..InterpreterSettings::default()
+    };
+    let initial = page.initial_transform(true).to_kurbo();
+    let mut ctx = Context::new(initial, Rect::new(0.0, 0.0, 130.0, 70.0), &cache, page.xref(), settings);
+    let mut draws = Draws::default();
+    interpret_page(page, &mut ctx, &mut draws);
+    assert!(others.lock().unwrap().is_empty(), "/{name} reported {:?}", others.lock().unwrap());
+    Outcome { decodes: cache.decoded_image_decodes(), failures: failures.load(Ordering::Relaxed), draws }
+}
+
+/// A JPEG 2000 picture that doesn't decode, drawn five times at one size, is decoded once, and
+/// each of the five draws still reports the failure.
+#[test]
+fn a_picture_that_does_not_decode_is_decoded_once_however_often_it_is_drawn() {
+    let out = draw_five_times("J");
+    assert_eq!(out.draws.rasters, 1, "only the picture after it drew");
+    assert_eq!(out.decodes, 2, "decoded again at every draw");
+    assert_eq!(out.failures, 5, "every draw reports the failure, as before");
+}
+
+/// A JBIG2 stencil mask that doesn't decode, drawn five times, is decoded once, and each draw
+/// still reports the failure: it isn't the page's JBIG2 allowance that refused it, which a
+/// later page would have afresh.
+#[test]
+fn a_stencil_mask_that_does_not_decode_is_decoded_once_however_often_it_is_drawn() {
+    let out = draw_five_times("M");
+    assert_eq!((out.draws.stencils, out.draws.rasters), (0, 1), "only the picture after it drew");
+    assert_eq!(out.decodes, 2, "decoded again at every draw");
+    assert_eq!(out.failures, 5, "every draw reports the failure, as before");
+}
+
+/// A picture whose soft mask doesn't decode draws without it at every draw, decoded once, and
+/// the mask's failure is reported once, as it was (the picture decodes; only its mask fails).
+#[test]
+fn a_picture_whose_soft_mask_does_not_decode_draws_unmasked_decoded_once() {
+    let out = draw_five_times("S");
+    assert_eq!((out.draws.rasters, out.draws.masked), (6, 0), "drawn five times without its mask, and the picture after it");
+    assert_eq!(out.decodes, 2);
+    assert_eq!(out.failures, 1);
+}
+
+/// A picture that decodes is still decoded once and handed to every draw.
+#[test]
+fn a_picture_that_decodes_is_still_decoded_once_and_drawn_every_time() {
+    let out = draw_five_times("G");
+    assert_eq!(out.draws.rasters, 6);
+    assert_eq!(out.decodes, 1);
+    assert_eq!(out.failures, 0);
+}
+
+/// A JBIG2 picture the page's allowance left out isn't kept as a failure: the next page drawing
+/// it starts with an allowance of its own and draws it, from the same cache.
+#[test]
+fn a_picture_left_out_for_the_pages_allowance_draws_on_the_next_page() {
+    use super::{RenderWarning, black_jbig2_page, jbig2_segment};
+    use crate::raster::{parse, render_page, take_warnings, warning_settings};
+    use crate::{RenderConfig, RenderRequest, RequestKind};
+    use hayro::RenderCache;
+
+    let image = |number: u8, id: u8| {
+        let mut stream = black_jbig2_page();
+        jbig2_segment(&mut stream, 1, 62, &[], &[0, 0, 0, 0, id]);
+        let mut object = format!(
+            "{number} 0 obj << /Type /XObject /Subtype /Image /Width 16 /Height 2 /BitsPerComponent 1 /ColorSpace /DeviceGray /Filter /JBIG2Decode /Length {} >> stream\n",
+            stream.len()
+        )
+        .into_bytes();
+        object.extend_from_slice(&stream);
+        object.extend_from_slice(b"\nendstream endobj\n");
+        object
+    };
+    let page = |number: u8, content: u8| {
+        format!("{number} 0 obj << /Type /Page /Parent 2 0 R /MediaBox [0 0 40 40] /Contents {content} 0 R /Resources << /XObject << /A 7 0 R /B 8 0 R >> >> >> endobj\n").into_bytes()
+    };
+    let content = |number: u8, text: &str| format!("{number} 0 obj << /Length {} >> stream\n{text}\nendstream endobj\n", text.len()).into_bytes();
+    let mut pdf =
+        b"%PDF-1.7\n1 0 obj << /Type /Catalog /Pages 2 0 R >> endobj\n2 0 obj << /Type /Pages /Kids [3 0 R 4 0 R] /Count 2 >> endobj\n".to_vec();
+    pdf.extend(page(3, 5));
+    pdf.extend(page(4, 6));
+    // The first page spends its allowance on A and has nothing left for B; the second draws B.
+    pdf.extend(content(5, "q 20 0 0 40 0 0 cm /A Do Q q 20 0 0 40 20 0 cm /B Do Q"));
+    pdf.extend(content(6, "q 20 0 0 40 20 0 cm /B Do Q"));
+    pdf.extend(image(7, 1));
+    pdf.extend(image(8, 2));
+    pdf.extend_from_slice(b"trailer << /Root 1 0 R >>\n%%EOF\n");
+    let pdf = parse(&Arc::new(pdf), None).expect("parses");
+    /// A page's pixels and the warnings it reported.
+    type Drawn = (Vec<u8>, Vec<RenderWarning>);
+    let black = |(rgba, _): &Drawn, x: usize| match &rgba[((20 * 40 + x) * 4)..][..4] {
+        [0, 0, 0, 255] => true,
+        [255, 255, 255, 255] => false,
+        other => panic!("neither an image nor the page: {other:?}"),
+    };
+    // Both pages through one cache, as a render worker keeps it.
+    let pages = |work: u64| {
+        let config = RenderConfig { jbig2_work: Some(work), ..RenderConfig::default() };
+        let cache = RenderCache::new();
+        [0, 1].map(|page| {
+            let warnings = Arc::new(Mutex::new(Vec::new()));
+            let settings = warning_settings(&config.settings(), warnings.clone());
+            let req = RenderRequest { page, kind: RequestKind::Pixels, tile: None, scale: 1.0, tag: 0, ..Default::default() };
+            let (_, _, rgba, _) = render_page(&pdf, &cache, &settings, &config, req).expect("renders");
+            (rgba[..].to_vec(), take_warnings(&warnings))
+        })
+    };
+    let one = (1..1_000).find(|work| black(&pages(*work)[0], 10)).expect("a small image is cheap");
+    let [first, second] = pages(one);
+    assert_eq!((black(&first, 10), black(&first, 30)), (true, false), "the first page had enough for A alone");
+    assert!(first.1.contains(&RenderWarning::ContentTruncated), "{:?}", first.1);
+    assert!(black(&second, 30), "B, left out of the first page, draws on the second");
+    assert!(second.1.is_empty(), "{:?}", second.1);
+}

vendor/README.md

@@ -39,7 +39,7 @@ Other resource categories retain their existing parent-chain lookup.
 
 | Crate | Version | Licence | Patches | Test |
 |---|---|---|---|---|
-| hayro-interpret | 0.7.0 | Apache-2.0 OR MIT | (1) select `/AP /N` appearance states via `/AS` (checkboxes, radios, stateful icons) and honour the NoView flag; (2) Type3 Unicode fallback to `/Encoding` glyph names, then printable ASCII; (3) `Type3Glyph::advance_width`; (4) `InterpreterSettings::ocg_overrides` for viewer layer toggles; (5) soft masks whose group lacks `/CS` were dropped (Chrome gradient text); (6) paint nesting cap `MAX_PAINT_NESTING` for tiling patterns and Type 3 glyphs (self-referencing content overflowed the stack); (7) CID width ranges in `/W` and `/W2` clamped to `MAX_CID` (a `0 4294967295 w` range hung); (8) `InterpreterSettings::hide_comments` skips markup annotations but keeps widgets and links (View ▸ Hide all comments); (9) `image_size_ok`: images and masks over 2^28 pixels are not decoded (a fuzzed `/W 4294967295` stencil mask allocated 4 GiB); (10) `MAX_NESTED_PAINTS`: nested paints (forms, Type 3 glyphs and tiling patterns inside one another) share a per-page budget of 50,000 (a glyph showing eight glyphs of its own font, sixteen deep, hung); (11) depends on skrifa 0.47 instead of 0.42 (upstream read-fonts fixed a Type 1 tokenizer that recursed through runs of numbers: a fuzzed font hung or overflowed the stack), with the one API change (`Charset::glyph_id` returns `Option`); (12) annotation appearances whose `/BBox` or `/Matrix` collapses to a line or a point are skipped (mapping them onto `/Rect` divided by zero and handed the device infinite or NaN matrices); (13) a Highlight annotation without a usable normal appearance is drawn from its `/QuadPoints`, `/C` and `/CA` with Multiply blending, as pdfcraft-annot draws highlights (it was not drawn at all); (14) simple-font `/Widths` (`read_widths`): entries only for codes 0–255 (one entry per code below `/FirstChar` was allocated: `4294967295` asked for 34 GB), an entry that isn't a number leaves its code missing instead of ending the array, and a `/LastChar` below `/FirstChar` gives no widths instead of discarding the font for a standard one; (15) the appearance box is mapped onto `/Rect` corner to corner: the translation now accounts for the scale (an appearance whose box didn't start at the origin and differed in size from `/Rect` was drawn shifted); (16) a clip set by `W`/`W*` takes effect after whichever path-painting operator ends the path (`S`, `f`, `B`, … as well as `n`; ISO 32000-2 §8.5.4): it was applied only on `n`, so `re W* S` clipped nothing and the clip stayed pending for a later, unrelated `n`; (17) `InterpreterSettings::cancelled`: once set, interpretation stops before the next content operator (a render pool sets it for every worker when dropped, after an edit or on closing, and for one worker when the watchdog gives up on its render, so renders whose answers nobody can receive stop instead of running to the end); (18) `MAX_PAGE_CONTENT`: the content a page holds at once (its contents, and the forms, tiling patterns and Type 3 glyph procedures being painted) stays within 512 MiB (`decode_content`; once it runs out, `InterpreterWarning::ContentTruncated` is reported (once per page by pdfcraft-render) so a truncated page is not a clean render; each level gives its bytes back when it finishes): forms nest fifty deep and each level kept its content while the next ran, so a form painting itself around a stream inflating to 256 MiB would have held over 12 GB; and Type 0 sampled functions take at most `MAX_SAMPLE_ENTRIES` (2^21) table points and `MAX_SAMPLE_VALUES` (2^23) samples, read no further than `/Size` asks (256 MiB of 1-bit samples asked for over 8 GB); (19) mesh shadings (types 4–7) are sampled only inside the area being drawn (`ShadingPattern::encode_within`, which the vendored hayro renderer calls with that area plus a pixel), within 8 visits per pixel of it plus 4 per triangle (`TriangleSampler`), and patches are triangulated one at a time: points far off the page were sampled over their whole bounding box, up to 65536 × 65536 entries (a fuzzed 147-byte patch stream; a synthetic page passed 6 GB in six seconds), ten thousand overlapping triangles ran for minutes, and every patch became 722 triangles before any limit; meshes hold at most `MAX_MESH_PATCHES` (2^16) patches and `MAX_MESH_TRIANGLES` (2^20) triangles, a lattice needs two to 2^19 vertices per row (`/VerticesPerRow 0` pushed empty rows for ever; one huge row read the whole stream), and a page samples at most `MAX_PAGE_MESH_TRIANGLES` (2^25) mesh triangles in all (a mesh painted a thousand times was triangulated a thousand times); each cut logs a warning; (20) PDF function construction is limited to 64 levels and 10,000 total nodes with cycle detection; calculator parsing and evaluation share procedure depth/work budgets, integer division/bitshift/rotation/index are panic-free, and fixed-size stack errors propagate; (21) a Type 0 font drawn with a substitute that isn't CID-keyed (it isn't embedded, or its program doesn't load) draws `.nLine truncated
+| hayro-interpret | 0.7.0 | Apache-2.0 OR MIT | (1) select `/AP /N` appearance states via `/AS` (checkboxes, radios, stateful icons) and honour the NoView flag; (2) Type3 Unicode fallback to `/Encoding` glyph names, then printable ASCII; (3) `Type3Glyph::advance_width`; (4) `InterpreterSettings::ocg_overrides` for viewer layer toggles; (5) soft masks whose group lacks `/CS` were dropped (Chrome gradient text); (6) paint nesting cap `MAX_PAINT_NESTING` for tiling patterns and Type 3 glyphs (self-referencing content overflowed the stack); (7) CID width ranges in `/W` and `/W2` clamped to `MAX_CID` (a `0 4294967295 w` range hung); (8) `InterpreterSettings::hide_comments` skips markup annotations but keeps widgets and links (View ▸ Hide all comments); (9) `image_size_ok`: images and masks over 2^28 pixels are not decoded (a fuzzed `/W 4294967295` stencil mask allocated 4 GiB); (10) `MAX_NESTED_PAINTS`: nested paints (forms, Type 3 glyphs and tiling patterns inside one another) share a per-page budget of 50,000 (a glyph showing eight glyphs of its own font, sixteen deep, hung); (11) depends on skrifa 0.47 instead of 0.42 (upstream read-fonts fixed a Type 1 tokenizer that recursed through runs of numbers: a fuzzed font hung or overflowed the stack), with the one API change (`Charset::glyph_id` returns `Option`); (12) annotation appearances whose `/BBox` or `/Matrix` collapses to a line or a point are skipped (mapping them onto `/Rect` divided by zero and handed the device infinite or NaN matrices); (13) a Highlight annotation without a usable normal appearance is drawn from its `/QuadPoints`, `/C` and `/CA` with Multiply blending, as pdfcraft-annot draws highlights (it was not drawn at all); (14) simple-font `/Widths` (`read_widths`): entries only for codes 0–255 (one entry per code below `/FirstChar` was allocated: `4294967295` asked for 34 GB), an entry that isn't a number leaves its code missing instead of ending the array, and a `/LastChar` below `/FirstChar` gives no widths instead of discarding the font for a standard one; (15) the appearance box is mapped onto `/Rect` corner to corner: the translation now accounts for the scale (an appearance whose box didn't start at the origin and differed in size from `/Rect` was drawn shifted); (16) a clip set by `W`/`W*` takes effect after whichever path-painting operator ends the path (`S`, `f`, `B`, … as well as `n`; ISO 32000-2 §8.5.4): it was applied only on `n`, so `re W* S` clipped nothing and the clip stayed pending for a later, unrelated `n`; (17) `InterpreterSettings::cancelled`: once set, interpretation stops before the next content operator (a render pool sets it for every worker when dropped, after an edit or on closing, and for one worker when the watchdog gives up on its render, so renders whose answers nobody can receive stop instead of running to the end); (18) `MAX_PAGE_CONTENT`: the content a page holds at once (its contents, and the forms, tiling patterns and Type 3 glyph procedures being painted) stays within 512 MiB (`decode_content`; once it runs out, `InterpreterWarning::ContentTruncated` is reported (once per page by pdfcraft-render) so a truncated page is not a clean render; each level gives its bytes back when it finishes): forms nest fifty deep and each level kept its content while the next ran, so a form painting itself around a stream inflating to 256 MiB would have held over 12 GB; and Type 0 sampled functions take at most `MAX_SAMPLE_ENTRIES` (2^21) table points and `MAX_SAMPLE_VALUES` (2^23) samples, read no further than `/Size` asks (256 MiB of 1-bit samples asked for over 8 GB); (19) mesh shadings (types 4–7) are sampled only inside the area being drawn (`ShadingPattern::encode_within`, which the vendored hayro renderer calls with that area plus a pixel), within 8 visits per pixel of it plus 4 per triangle (`TriangleSampler`), and patches are triangulated one at a time: points far off the page were sampled over their whole bounding box, up to 65536 × 65536 entries (a fuzzed 147-byte patch stream; a synthetic page passed 6 GB in six seconds), ten thousand overlapping triangles ran for minutes, and every patch became 722 triangles before any limit; meshes hold at most `MAX_MESH_PATCHES` (2^16) patches and `MAX_MESH_TRIANGLES` (2^20) triangles, a lattice needs two to 2^19 vertices per row (`/VerticesPerRow 0` pushed empty rows for ever; one huge row read the whole stream), and a page samples at most `MAX_PAGE_MESH_TRIANGLES` (2^25) mesh triangles in all (a mesh painted a thousand times was triangulated a thousand times); each cut logs a warning; (20) PDF function construction is limited to 64 levels and 10,000 total nodes with cycle detection; calculator parsing and evaluation share procedure depth/work budgets, integer division/bitshift/rotation/index are panic-free, and fixed-size stack errors propagate; (21) a Type 0 font drawn with a substitute that isn't CID-keyed (it isn't embedded, or its program doesn't load) draws `.nLine truncated
 | hayro | 0.7.1 | Apache-2.0 OR MIT | (1) `RenderSettings::{x_offset, y_offset}` to render one tile of a large page; (2) images over `MAX_IMAGE_PIXELS` (2^28) are skipped (a fuzzed `/W 4294967295` hung in resampling); (3) `tiling_cell_scale` keeps a tiling pattern's cell pixmap within 3000 px a side whatever its `/XStep`/`/YStep` (a fuzzed step far beyond `/BBox` allocated 2 GB); (4) stroke widths clamped to 4 × (canvas width + height) device pixels (a fuzzed `/LW 9223372036854775807` allocated 10 GB in stroke expansion); (5) `usable_dash_pattern`: a stroke whose dash pattern could cut it into more than `MAX_DASHES_PER_STROKE` (250,000) pieces, counting kurbo's restart of the pattern at every subpath and `ClosePath`, is drawn solid (a fuzzed `/D [[11] 0]` on a line from x = 92234775807 allocated over 5 GB, `[0.000001] 0 d` on a 20 pt line over 10 GB); (6) a dash pattern with a negative or non-finite entry, or a period that isn't positive, is drawn solid (`[-1 -1] 0 d` looped for ever in kurbo); (7) `render_size` and `render_into`: render into a buffer the caller allocates (PdfCraft renders into pixel-aligned memory that the GUI takes over as texture data); `render` is `render_into` on a new pixmap; (8) checked image source/target/RGBA, alpha-premultiplication scratch and intermediate sizes before resampling, finite transforms and exact source/mask buffers; sources wider than u16 may shrink before backend limits, with a 2^20 source-axis cap and conservative 128 MiB filter-planning estimate; an over-limit or failed resize keeps the original image resolution and geometry when the source fits the backend, otherwise skips it, with a logged warning; staging, resampling output/scratch, RGBA conversion and padding buffers use fallible reserves; (9) `may_paint_canvas`: a fill or stroke whose control-point box, grown by how far its paint reaches past the path (half the stroke width times the larger of the miter limit and √2) plus 0.5 user units and 2 pixels, misses the canvas is not drawn, so a tile no longer strokes and fills every path of its page (item 5 of #623); `set_skip_offscreen_paths` turns this off for the regression test; (10) `/Interpolate` governs the pre-resize too (#624): a minified image that does not request interpolation is pre-resized with nearest-neighbour (since (21), only inside Type 3 glyphs), planned without the Catmull-Rom filter limits (`image_resampling_size` takes the flag); a Type 3 glyph image is sampled bicubically only when interpolated; a minified image with a matching soft mask is nearest-sampled straight into premultiplied RGBA; (11) `may_paint_rect`: an image whose stream-size box, mapped to the device, misses the canvas is skipped before decoding (decoded data can only shrink against that box), so a tile of an image-heavy page no longer decodes, resamples and rasterizes every image on it; (12) preview renders (`InterpreterSettings::preview`: backdrops, thumbnails, grid cells, signature-drag backgrounds) skip images under `PREVIEW_MIN_IMAGE_AREA` device pixels, while final renders always draw them; (13) reading colours (#897): `RenderSettings::color_replacement` takes a `ColorReplacement` (text colour, background, `only_black`, `line_art`) that remaps solid text and line-art paint while rendering, glyph outlines counting as text and stencil-mask ink as line art; images, image alpha masks, soft masks and pattern cells keep their colours; `render_pdf` gains a `color_replacement: Option<ColorReplacement>` argument (pass `None` for true colours); the upstream-only `[dev-dependencies.log]` entry is removed from `Cargo.toml` (log is a normal dependency since the logging patch), so `cargo test -p hayro` runs this crate's unit tests; (14) a shrunk image that is interpolated (or, since (21), area-averaged) and whose axes lie along the device's (upright, mirrored or a quarter turn, as on a page under `/Rotate`) is snapped to whole device pixels: shrunk once to its rounded device rectangle (halves round up, so a tile snaps an edge where the whole page does) and drawn 1:1 from the rounded corner, instead of a ceil-sized raster resampled again at a fractional scale and offset (`snap_minified_image`; scans looked soft next to Acrobat, and until #5530 a turned one still did, up to 200 grey levels off the same page drawn upright). An image counts as axis-aligned when the terms the snap drops move its far edges by at most a hundredth of a device pixel (`SNAP_SKEW`), not by a tolerance on the coefficients, which let a slight skew of a very wide image through; and the snapped 1:1 transform is used only when the shrink produced the snapped size, so a failed resize keeps the full transform instead of drawing the source raster at its source size; (15) a soft mask of another size than its image is filtered when shrunk, whatever its `/Interpolate` (#1649): such a mask carries the shape on its own (Word saves text as a tiny black image with a glyph-bitmap mask), and nearest-neighbour dropped every stroLine truncated
 | hayro-jbig2 | 0.3.0 | Apache-2.0 OR MIT | (1) at most 2^28 pixels per bitmap and page bitmap, and 2^25 cells per halftone grid and grayscale image (each side could be 65535: 4.3 billion pixels decoded one by one, and a 4-byte value per cell); (2) at most 2^26 symbol pixels per symbol dictionary (a fuzzed dictionary decoded for minutes); (3) a symbol dictionary's height classes are capped at one per new symbol and its export runs at `2 × symbols + 1`, each plus `EXTRA_EMPTY_STEPS` (1024): an empty height class or a zero-length export run doesn't advance, and past the end of its data the arithmetic decoder can decode them for ever (one changed byte hung the renderer); (4) a segment header's long-form referred-to segment count is capped at `u16::MAX` and a larger count is refused before its retention or referred-to bytes are read: the 29-bit count went straight to `Vec::with_capacity`, a reservation of about 2 GiB whose failure aborts the process (from storytold/craft-libs#176); (5) what a page's segments hold for later ones (intermediate regions, pattern and symbol dictionaries with their headers and retained contexts, and Huffman tables) is at most 64 MiB, two square bitmaps of 16,384 a side (`MAX_RETAINED_WORDS`), and a dictionary's new and exported symbols are counted against it as they are made: each was bounded on its own, but a few dozen bytes an intermediate region held 33 MB each, and 8 MiB symbols exported again and again killed the process, so a few kilobytes could exhaust memory; (6) a segment may refer to at most 2^20 symbols, checked before any are collected: references may repeat, and 65,535 of them to one dictionary, in a 78 KB stream, asked for 17 GB; (7) a pattern dictionary keeps at most 16 MiB of shifted copies, which only speed up drawing, and none that wouldn't fit what the page has left (`MAX_SHIFTED_WORDS`); (8) a custom Huffman table has at most 2^16 lines, and a stream with its globals at most 2^20 segments and 2^22 references, checked as they are read: both grew before anything counted them, and PDFs of a kilobyte or so reached 5 GB and more behind nested Flate; (9) a page past the bounds of (1) is refused when the stream is opened: a caller sizes its output from the header before the page bitmap is made, and an Indexed image in a 743-byte PDF asked for 137 GB, which aborts; (10) a stream may ask for at most 2^30 units of decoding work, four of the largest pages, about 6 s to 17 s (`MAX_WORK`): decoding a bitmap costs a pixel each and 4 a row, copying one a word each and 1 a row, and each integer an item reads costs 38, what one takes at worst, so one with no pixels pays too, before it is decoded: each bitmap was bounded, but nothing bounded how many there were, and a 795-byte PDF took 265 s; (11) `Image::decode_within` spends from what a run of images may still do, starting with its page's pixels, and `charge_reading` charges reading a stream and its globals before anything is parsed, so an operation over a document can bound what `MAX_WORK` bounds for one image, and images that fail or decode nothing add up too: a file of thousands of images made to reach that limit took as many times as long; (12) adding to a location, a reference offset, a symbol's corner or a Huffman range line near the end of an `i32`, and negating a decoded integer of 2^31, overflowed: a panic wherever overflow is checked. The refinement context, a text region's corners, the integer decoder and the Huffman table wrap, as a release build did, and `Bitmap::combine` saturates, which comes to the same, as a location that overflows is past the end of any page. What is drawn doesn't change; (13) a bitmap filled with a default pixel of 1 (a page, or an intermediate text or halftone region) had 1s past its width in the last word of each row, and a refinement, which reads its reference a word at a time, read them as pixels where the standard has 0 (6.3.5.2): refining such a reference not a multiple of 32 wide decoded to other pixels than were coded, release builds too. `Bitmap::clear_padding` clears those bits when a bitmap is filled with 1s, as cutting symbols and patterns from a collective bitmap already did. Ghostscript reads the test's page and text streams as they decode now; (14) a region's location is unsigned (5.4.4), so one past `i32::MAX` is off the page, where the decoder took it for a signed one: a region there was drawn back onto the page from the left or the top (at 0xFFFFFFFC, four pixels left of it), and a refinement there, of the page or of a region at the same place (7.4.7.5), was refused as an overflow, and the page with it. `Bitmap::combine_region` keeps such a location past the page. A refinement of another region refines it where it is, as Table 35 has it (GRREFERENCEDX and GRREFERENCEDY 0), whatever the two locations say; a refinement of the page, see (16); (15) a refinement whose reference is offset right of it (GRREFERENCEDX above 0, which a refined text instance has from a wider bitmap, TaLine truncated
 | lopdf | 0.45.0 | MIT | (1) PNG predictor rows longer than the data are refused before their buffers are allocated (a fuzzed `/Columns 4294967295` allocated two 4 GiB rows), and empty frames return without row buffers; (2) predictor row widths use saturating arithmetic or reject overflow; (3) sub-byte TIFF Predictor 2 scratch is bounded by the samples present in the input instead of the declared `/Colors`, aligned with [lopdf#589](https://github.com/J-F-Liu/lopdf/pull/589); (4) the object lookup keeps its cycle history through compressed objects, requires an ordinary in-file container, and caps the chain of indirect lookups at 16 (a chain of new objects, each a `/Length` compressed in a container whose own `/Length` names the next, or a stream whose `/Length` names the next stream, overflowed the stack: a dev build's 2 MiB thread within 800 hops); (5) object-stream members are parsed within their own extent (up to the next distinct offset), an entry whose offset repeats an earlier entry's is skipped (so a repeated object number can resolve to its earlier entry), and offsets are processed in index order; (6) `find_latest_trailer`: a rebuild takes a `/Type /XRef` stream dictionary as the trailer when the file has no `trailer` keyword (the latest in the file wins, as a later `trailer` would), so a concatenated or truncated file keeps its /Info and /ID instead of failing to load; a candidate whose /Root is the catalog beats one whose /Root is another dictionary, and an in-file /Root must match its generation (#2365) | (1) `inspect::tests::png_predictor_rows_longer_than_the_data_are_refused_before_allocating`, `inspect::tests::empty_png_predictor_frames_do_not_allocate_rows`; (2) `inspect::tests::tiff_subbyte_predictor_row_width_overflow_is_refused`; (3) `inspect::tests::tiff_subbyte_predictor_scratch_is_bounded_by_available_samples`, `inspect::tests::tiff_subbyte_predictors_keep_components_rows_and_padding`; (4) `inspect::tests::compressed_object_self_container_does_not_recurse`, `inspect::tests::compressed_object_mutual_containers_do_not_recurse`, `inspect::tests::object_stream_length_cycle_does_not_recurse`, `inspect::tests::compressed_object_container_chain_does_not_recurse`, `inspect::tests::compressed_length_chain_does_not_recurse`, `inspect::tests::plain_stream_length_chain_does_not_recurse`, `inspect::tests::compressed_objects_and_indirect_stream_length_still_load`, `inspect::tests::compressed_length_in_a_container_with_an_indirect_length_still_loads`; (5) `inspect::tests::object_stream_member_extents_bound_owned_strings`, `inspect::tests::object_stream_member_extents_preserve_index_order_and_whitespace`; (6) `inspect::tests::rebuilt_file_keeps_the_info_of_its_cross_reference_stream`, `inspect::tests::rebuilt_file_prefers_a_trailer_whose_root_is_the_catalog` |

vendor/hayro-interpret/src/cache.rs

@@ -1,10 +1,11 @@
 use crate::util::hash128;
+use crate::{InterpreterWarning, WarningSinkFn};
 use hayro_syntax::object::{Array, Dict, MaybeRef, Name, Null, ObjRef, Object, Stream};
 use kurbo::Affine;
 use std::any::Any;
 use std::collections::HashMap;
 use std::collections::hash_map::Entry;
-use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::atomic::{AtomicU8, AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
 
 type CacheMap = HashMap<u128, Option<Box<dyn Any + Send + Sync>>>;
@@ -69,6 +70,14 @@ impl Cache {
 /// them; when an insert would pass the budget the whole map is dropped (a fresh epoch), which
 /// only costs re-decoding, and an image larger than the whole budget is never kept. The budget
 /// caps memory per interpreter cache. A poisoned lock decodes without the cache.
+///
+/// PdfCraft patch (91): a decode that fails is kept too, as the warnings it reported, so an
+/// image drawn again and again that doesn't decode (damaged data, a JPX or JBIG2 stream its
+/// decoder refuses) isn't decoded again at every draw: a later draw reports the same warnings
+/// and gets nothing, as the decode did. A failure counts [`FAILED_DECODE_BYTES`] towards the
+/// budget, so failures can't grow the map without bound either. A failure that may come from
+/// what the page had left (its JBIG2 allowance, `ContentTruncated`) isn't kept: another page,
+/// or the same page drawn again, starts with a fresh allowance.
 #[derive(Clone)]
 pub(crate) struct DecodedImageCache {
     inner: Arc<Mutex<DecodedImageCacheInner>>,
@@ -110,9 +119,43 @@ pub(crate) struct DecodedImageKey {
     pub(crate) preview: bool,
 }
 
+/// PdfCraft patch (91): what the budget charges for a failed decode, about what its key and
+/// entry take in the map.
+const FAILED_DECODE_BYTES: usize = 128;
+
+/// PdfCraft patch (91): what a [`DecodedImageCache`] keeps for a key.
+enum DecodedEntry {
+    /// The decoded image.
+    Decoded(Box<dyn Any + Send + Sync>),
+    /// The decode failed, reporting these warnings ([`warning_bit`] flags).
+    Failed(u8),
+}
+
+/// PdfCraft patch (91): `warning` as a flag in a set of warnings.
+fn warning_bit(warning: InterpreterWarning) -> u8 {
+    match warning {
+        InterpreterWarning::UnsupportedFont => 1,
+        InterpreterWarning::ImageDecodeFailure => 2,
+        InterpreterWarning::ContentTruncated => 4,
+        InterpreterWarning::ColorFallback => 8,
+    }
+}
+
+/// PdfCraft patch (91): the warnings in a set of [`warning_bit`] flags.
+fn warnings_in(bits: u8) -> impl Iterator<Item = InterpreterWarning> {
+    [
+        InterpreterWarning::UnsupportedFont,
+        InterpreterWarning::ImageDecodeFailure,
+        InterpreterWarning::ContentTruncated,
+        InterpreterWarning::ColorFallback,
+    ]
+    .into_iter()
+    .filter(move |warning| bits & warning_bit(*warning) != 0)
+}
+
 #[derive(Default)]
 struct DecodedImageCacheInner {
-    map: HashMap<DecodedImageKey, Option<Box<dyn Any + Send + Sync>>>,
+    map: HashMap<DecodedImageKey, DecodedEntry>,
     bytes: usize,
     /// PdfCraft patch (32): how many keyed decodes ran (cache misses).
     decodes: usize,
@@ -140,44 +183,73 @@ impl DecodedImageCache {
         }
     }
 
-    /// The decoded image for `key`, from the cache or from `f`. `None` as the key decodes
-    /// without the cache.
+    /// The decoded image for `key`, from the cache or from `f`, which decodes reporting its
+    /// warnings to the sink it is handed. `None` as the key decodes without the cache.
+    /// PdfCraft patch (91): `sink` gets the warnings of the decode, or those of the failed
+    /// decode kept for `key`.
     pub(crate) fn get_or_insert_with<T: Clone + Send + Sync + 'static>(
         &self,
         key: Option<DecodedImageKey>,
         budget: usize,
         size_of: impl FnOnce(&T) -> usize,
-        f: impl FnOnce() -> Option<T>,
+        sink: &WarningSinkFn,
+        f: impl FnOnce(&WarningSinkFn) -> Option<T>,
     ) -> Option<T> {
-        let Some(key) = key else { return f() };
-        // Fast path: a hit clones the pixels without decoding.
+        let Some(key) = key else { return f(sink) };
+        // Fast path: a hit clones the pixels without decoding, and a kept failure reports its
+        // warnings again without decoding.
         let hit = match self.inner.lock() {
-            Ok(locked) => locked
-                .map
-                .get(&key)
-                .and_then(|val| val.as_ref())
-                .and_then(|val| val.downcast_ref::<T>().cloned()),
+            Ok(locked) => match locked.map.get(&key) {
+                Some(DecodedEntry::Decoded(val)) => val.downcast_ref::<T>().cloned().map(Ok),
+                Some(DecodedEntry::Failed(bits)) => Some(Err(*bits)),
+                None => None,
+            },
             // A thread panicked while holding the lock: decode without the cache.
-            Err(_) => return f(),
+            Err(_) => return f(sink),
         };
-        if let Some(hit) = hit {
-            return Some(hit);
+        match hit {
+            Some(Ok(hit)) => return Some(hit),
+            Some(Err(bits)) => {
+                warnings_in(bits).for_each(|warning| sink(warning));
+                return None;
+            }
+            None => {}
         }
-        // Miss: decode without holding the lock (decoding must never deadlock the cache).
-        let val = f();
+        // Miss: decode without holding the lock (decoding must never deadlock the cache),
+        // noting the warnings it reports.
+        let reported = Arc::new(AtomicU8::new(0));
+        let recording: WarningSinkFn = {
+            let (reported, sink) = (reported.clone(), sink.clone());
+            Arc::new(move |warning| {
+                reported.fetch_or(warning_bit(warning), Ordering::Relaxed);
+                sink(warning);
+            })
+        };
+        let val = f(&recording);
+        let reported = reported.load(Ordering::Relaxed);
         let Ok(mut locked) = self.inner.lock() else {
             return val;
         };
         locked.decodes = locked.decodes.saturating_add(1);
         // Another thread may have decoded it meanwhile: keep the first copy.
-        if let Some(existing) = locked.map.get(&key)
-            && let Some(existing) = existing.as_ref().and_then(|val| val.downcast_ref::<T>())
-        {
-            return Some(existing.clone());
+        match locked.map.get(&key) {
+            Some(DecodedEntry::Decoded(existing)) => {
+                if let Some(existing) = existing.downcast_ref::<T>() {
+                    return Some(existing.clone());
+                }
+            }
+            Some(DecodedEntry::Failed(_)) if val.is_none() => return None,
+            _ => {}
         }
         let size = match &val {
             Some(v) => size_of(v),
-            None => 0,
+            // A failure the page's JBIG2 allowance may have caused isn't the image's to keep.
+            None if reported & warning_bit(InterpreterWarning::ContentTruncated) != 0
+                || hayro_syntax::page_jbig2_work_spent() =>
+            {
+                return None;
+            }
+            None => FAILED_DECODE_BYTES,
         };
         if size > budget {
             return val;
@@ -186,12 +258,14 @@ impl DecodedImageCache {
             locked.map.clear();
             locked.bytes = 0;
         }
+        let entry = match &val {
+            Some(val) => DecodedEntry::Decoded(Box::new(val.clone())),
+            None => DecodedEntry::Failed(reported),
+        };
+        // An entry replaced (a failure another thread kept) isn't given back: the budget
+        // errs towards a fresh epoch.
         locked.bytes = locked.bytes.saturating_add(size);
-        locked.map.insert(
-            key,
-            val.clone()
-                .map(|val| Box::new(val) as Box<dyn Any + Send + Sync>),
-        );
+        locked.map.insert(key, entry);
         val
     }
 }

vendor/hayro-interpret/src/x_object.rs

@@ -578,7 +578,8 @@ impl<'a> ImageXObject<'a> {
             self.decode_key(cache, true, target_dimension),
             MAX_DECODED_IMAGE_BYTES,
             decoded_mask_size,
-            || decode_mask(self, target_dimension),
+            &self.warning_sink,
+            |sink| decode_mask(&self.reporting_to(sink), target_dimension),
         )
     }
 
@@ -598,7 +599,8 @@ impl<'a> ImageXObject<'a> {
             self.decode_key(cache, false, target_dimension),
             MAX_DECODED_IMAGE_BYTES,
             decoded_raster_size,
-            || decode_raster(self, target_dimension),
+            &self.warning_sink,
+            |sink| decode_raster(&self.reporting_to(sink), target_dimension),
         )
     }
 
@@ -649,6 +651,14 @@ impl<'a> ImageXObject<'a> {
             || dict.get::<Stream<'_>>(MASK).is_some_and(|s| uses_jpx(&s))
     }
 
+    /// PdfCraft patch (91): this image, reporting its warnings to `sink`.
+    fn reporting_to(&self, sink: &WarningSinkFn) -> Self {
+        Self {
+            warning_sink: sink.clone(),
+            ..self.clone()
+        }
+    }
+
     pub(crate) fn width(&self) -> u32 {
         self.width
     }

vendor/hayro-syntax/src/filter/jbig2.rs

@@ -28,7 +28,9 @@ pub fn set_page_jbig2_work(work: u64) {
 }
 
 /// Pays for reading `bytes` of a stream and its globals out of what the page has left, or `None`.
-/// An image that doesn't decode isn't kept, so a page drawing it again and again reads it again.
+/// An image that doesn't decode isn't always kept (not past the page's allowance, nor drawn under a
+/// transfer function: hayro-interpret patch 91), so a page drawing it again and again can read it
+/// again.
 #[cfg(feature = "std")]
 fn charge_page_reading(bytes: usize) -> Option<()> {
     let mut left = PAGE_WORK_LEFT.with(core::cell::Cell::get);