Skip to content

Fix 32-bit zstdmt max job size allocation failure - #4675

Open
danyalahmed1995 wants to merge 1 commit into
meta:devfrom
danyalahmed1995:fix/4628-zstdmt-32bit-job-size
Open

danyalahmed1995 wants to merge 1 commit into
meta:devfrom
danyalahmed1995:fix/4628-zstdmt-32bit-job-size

Conversation

@danyalahmed1995

Copy link
Copy Markdown

Fixes #4628.

Summary

Reduces the 32-bit zstdmt maximum job size from 512 MB / log 29 to 256 MB / log 28.

On 32-bit builds, the previous bound could allow a threaded compression configuration at max compression level to attempt an oversized allocation, matching the failure described in #4628.

This follows the issue reporter's diagnosis and proposed cap reduction.

Changes

  • Lower ZSTDMT_JOBLOG_MAX on 32-bit builds from 29 to 28.
  • Lower ZSTDMT_JOBSIZE_MAX on 32-bit builds from 512 MB to 256 MB.
  • Add a bounds check in the fuzzer unit tests so ZSTD_c_jobSize reports the expected architecture-specific upper bound.

Validation

  • git diff --check
  • Verified with a temporary GitHub Actions workflow on linux/386 using Debian testing:
    • make clean
    • make -j$(nproc)
    • make check

The temporary validation workflow was kept out of this PR branch.

@meta-cla meta-cla Bot added the CLA Signed label May 18, 2026
kevinjzhang pushed a commit to kevinjzhang/zstd that referenced this pull request Nov 12, 2026
On 32-bit builds, ZSTDMT_initCStream_internal() computed the round buffer size in size_t. With a 1 GiB job (windowLog 30, btultra2, LDM disabled) the 4 GiB total wraps to 0, init succeeds with a 0-byte buffer, and the first streaming copy writes past it. The same wrap happens with many workers at the maximum job size, which PR meta#4675's cap reduction alone does not prevent.

Compute the slack/section totals in 64 bits and reject the configuration with memory_allocation when the total exceeds SIZE_MAX, before allocating. Also adopt meta#4675's 32-bit cap reduction (job log 29->28, job size 512 MB->256 MB) so the reported 1-worker, level 22 case needs 1 GiB instead of 2 GiB.

Add a 32-bit-only fuzzer regression test that expects the oversized configuration to fail cleanly with memory_allocation, and a bounds test pinning the architecture-specific job size maximum.

Fixes meta#4628.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Compression fails on linux/i386 with ZSTD_maxCLevel + ZSTD_c_nbWorkers

1 participant