patch for multipage row

[pmattione-nvidia]

This adds a patch to release 26.06 for applying this fix for a parquet decoding bug in cuDF.

[pmattione-nvidia]

build

[greptile-apps]

Greptile Summary

This PR backports upstream cuDF fix rapidsai/cudf#23241 to the release/26.06 branch as a patch file, replacing the empty noop.patch placeholder. The fix addresses an out-of-bounds write during Parquet string-offset preprocessing caused by plain-encoded list<string> columns where a single list row spans multiple data pages, producing 0-row continuation pages that the host-side buffer sizing incorrectly skipped while GPU decode kernels still wrote offsets to them.

• Core fix (reader_impl_preprocess.cu, page_decode.cuh): Introduces page_has_rows_to_process() to correctly identify 0-row list continuation pages as needing processing, replacing a simple row-range exclusion that failed for these pages. Refactors is_bounds_page and is_page_contained to accept PageInfo const& page + chunk_start_row directly, enabling reuse from both GPU and the preprocess-functor context.
• Call-site updates (decode_preprocess.cu, page_delta_decode.cu, page_string_decode.cu): All existing call sites are updated to pass the explicit page and chunk_start_row arguments instead of the full page_state_s* struct.
• Regression test (parquet_chunked_reader_test.cu): Adds TestChunkedReadWithPlainListOfStringSpanningPages, covering full reads, chunked reads at multiple byte limits, and skip_rows/num_rows windows that land on 0-row pages — particularly effective under compute-sanitizer memcheck.

Confidence Score: 5/5

Safe to merge — this is a clean, well-scoped backport of an already-reviewed upstream fix with a comprehensive regression test.

The patch faithfully reproduces the upstream fix with no deviations: function signatures are correctly refactored, all call sites in the five changed CUDA files are updated consistently, and the new page_has_rows_to_process helper is logically equivalent to the old scattered checks while also handling the previously uncovered 0-row list-continuation-page case. The regression test covers the full read, chunked reads at multiple byte limits, and targeted skip_rows/num_rows windows that exercise exactly the boundary condition that triggered the OOB write.

No files require special attention.

Important Files Changed

Filename	Overview
patches/fix_parquet_preprocess_zero_row_string_pages.patch	New patch file backporting upstream fix for 0-row list page OOB write; logic is correct, signatures are properly refactored, and the helper page_has_rows_to_process correctly handles the list continuation-page edge case.
patches/noop.patch	Empty placeholder patch removed now that the real fix patch replaces it.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["compute_page_offset_count\n(host-side buffer sizing)"] --> B{page_has_rows_to_process?}
    B -->|"page.num_rows > 0\n& row range intersects"| C[Include page in buffer size]
    B -->|"page.num_rows == 0\n& !has_repetition"| D[Skip page → return 0]
    B -->|"page.num_rows == 0\n& has_repetition\n(0-row list continuation page)"| E{is_bounds_page OR\nis_page_contained?}
    E -->|Yes| C
    E -->|No| D
    C --> F["GPU decode kernels\nwrite offsets to allocated buffer ✓"]
    D --> G["GPU decode kernels\nstay within allocated buffer ✓"]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["compute_page_offset_count\n(host-side buffer sizing)"] --> B{page_has_rows_to_process?}
    B -->|"page.num_rows > 0\n& row range intersects"| C[Include page in buffer size]
    B -->|"page.num_rows == 0\n& !has_repetition"| D[Skip page → return 0]
    B -->|"page.num_rows == 0\n& has_repetition\n(0-row list continuation page)"| E{is_bounds_page OR\nis_page_contained?}
    E -->|Yes| C
    E -->|No| D
    C --> F["GPU decode kernels\nwrite offsets to allocated buffer ✓"]
    D --> G["GPU decode kernels\nstay within allocated buffer ✓"]

Loading

_{Reviews (2): Last reviewed commit: "delete noop patch" | Re-trigger Greptile}

[ttnghia]

Can we hot fix in cudf instead?

[pmattione-nvidia]

Can we hot fix in cudf instead?

No they aren't interested in it.

[ttnghia]

I don't think tests are needed here, as we never run them. However, it's good to keep the redundant tests here to reproduce if needed.

[ttnghia]

Please do not rebase this into main/26.08 release but fix in cudf instead.

[pmattione-nvidia]

Please do not rebase this into main/26.08 release but fix in cudf instead.

It's already fixed in cudf 26.08, as per the above PR link. this pr is just for a patch to 26.06 release only.

[NvTimLiu]

build

[pmattione-nvidia]

build