Out-of-bounds global write in medianfilter-cuda/ckMedian for small images

Aperta Adatta ai principianti
#475 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
2/5
Tempo stimato
1-3 ore
Idoneità per principianti
82/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
cpp

Direzione di ricerca

Leggi src/medianfilter-cuda/MedianFilter.cu intorno alla riga 252, dove il controllo dei limiti dell’output e la scrittura usano coordinate Y diverse. Inizia compilando ed eseguendo la riproduzione 1x1 PPM fornita con Compute Sanitizer. Il lavoro è terminato quando il kernel scrive solo all’interno dell’allocazione di output di larghezza per altezza e Compute Sanitizer non segnala errori.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Summary

The ckMedian kernel in src/medianfilter-cuda/MedianFilter.cu can write one row past the output allocation when the output buffer is allocated using the actual image dimensions (width * height * sizeof(unsigned int)).

A standard 1x1 binary P6 PPM image is sufficient to reproduce the issue. NVIDIA Compute Sanitizer reports an Invalid __global__ write of size 4 bytes at MedianFilter.cu:252.

The issue appears to be caused by an off-by-one error in the final output bounds check, rather than by the CUDA runtime or driver.

Confirmed versions

I originally reproduced the issue at:

  • af5d95689b3e5d83f6e50d1aae2787770e10569d

I also confirmed that it remains reproducible on the latest ORNL/HeCBench master available to me:

  • ffde18721e8c1090460e367c97b4af39240cc83e (2026-11-13)

The MedianFilter.cu file at both revisions has the same SHA-256:

38c557ed5810aaafc13f79c935c18e7746e3ec3fd91fac34a36f817b4615773e

Environment used for the latest-master confirmation

  • GPU: NVIDIA GeForce RTX 4080 (sm_89)
  • NVIDIA driver: 621.01
  • CUDA toolkit: 12.9, V12.9.41
  • NVIDIA Compute Sanitizer: 2025.2.0.0 (build 35692658)
  • Platform: WSL/Docker on Windows, x86_64
  • Build flags: -arch=sm_89 -O3 -lineinfo -std=c++17

Minimal reproduction

The attached archive contains the unmodified pinned MedianFilter.cu, the small fuzzing host harness, the PPM parser, crash inputs, and Compute Sanitizer logs.

The kernel itself is unchanged. The custom host harness allows arbitrary input dimensions and allocates the input/output buffers using the actual image size. This is important because the original benchmark host allocates a fixed 1920x1080 buffer, which can mask the out-of-bounds write for smaller images.

From the extracted archive, build the uninstrumented binary with:

nvcc -arch=sm_89 -O3 -lineinfo -std=c++17 \
  src/fuzz_main.cu src/MedianFilterHost.cpp -o medianfilter_repro

A minimal standard P6 PPM can also be generated independently:

python3 - <<'PY'
with open("minimal-1x1.ppm", "wb") as f:
    f.write(b"P6\n1 1\n255\n\x00\x00\x00")
PY

Run it with Compute Sanitizer:

compute-sanitizer --tool=memcheck --report-api-errors=no --error-exitcode=99 \
  ./medianfilter_repro minimal-1x1.ppm

Expected behavior

The kernel should only write within the width * height output elements. Compute Sanitizer should report no errors.

Actual behavior

The process exits with code 99 because of --error-exitcode=99, and Compute Sanitizer reports:

========= COMPUTE-SANITIZER
========= Invalid __global__ write of size 4 bytes
=========     at ckMedian(const uchar4 *, unsigned int *, int, int, int)+0x16a0 in MedianFilter.cu:252
=========     by thread (0,1,0) in block (0,0,0)
=========     Access at 0x711400204 is out of bounds
=========     and is 1 bytes after the nearest allocation at 0x711400200 of size 4 bytes
========= Target application returned an error
========= ERROR SUMMARY: 1 error

The device address varies between runs; the instruction, source location, allocation size, and error type are consistent.

Root-cause analysis

The kernel shifts the Y coordinate up by one row while loading the apron:

int iDevYPrime = iGroupIdY * iBlockY + iLocalIdY - 1;
int iDevGMEMOffset = iDevYPrime * iImageX + iImagePosX;

The final store restores that one-row shift by adding iImageX, but the bounds check still uses the un-restored iDevYPrime:

if ((iDevYPrime < iDevImageHeight) && (iImagePosX < iImageWidth)) {
    uiDest[iDevGMEMOffset + iImageX] = uiPackedPix;
}

The effective output row is therefore iDevYPrime + 1, not iDevYPrime.

When iDevYPrime == iDevImageHeight - 1, the current condition passes, but the store writes to output row iDevImageHeight, which is one row beyond the allocation. For a 1x1 image with a 4-byte output allocation, this becomes a 4-byte write immediately past that allocation.

Suggested fix

Check the restored output coordinates explicitly:

const int iDevYRestored = iDevYPrime + 1;
if (iDevYRestored >= 0 && iDevYRestored < iDevImageHeight &&
    iImagePosX < iImageWidth) {
    uiDest[iDevYRestored * iImageX + iImagePosX] = uiPackedPix;
}

This is a proposed source-level correction; the sanitizer results above were obtained with the unmodified kernel.

Additional evidence

The fuzzing campaign produced 15 distinct crash inputs:

  • 10 from the baseline configuration
  • 5 from the L2 configuration

All 15 files have distinct SHA-256 hashes. Every input was independently reproduced using an uninstrumented binary and standalone Compute Sanitizer, and all 15 returned rc=99 with the same MedianFilter.cu:252 invalid global write signature. They should therefore be treated as multiple triggers for one root-cause defect, not as separate vulnerabilities.

Observed triggering dimensions included:

1x1, 1x2, 1x6, 6x1, 6x2, 16x1, 26x1, and 96x1

The complete inputs and individual sanitizer logs are included in the attached archive. The separate minimal-1x1-memcheck.log file is the confirmation against ORNL/HeCBench master commit ffde18721e8c1090460e367c97b4af39240cc83e.

Lingua principale
C++
Stelle
310
Fork
125
Merge medio
5g 1h
PR unite (30g)
52

Preparare l'ambiente

Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di ORNL/HeCBench

Tutte le issue di ORNL/HeCBench

Issue simili

Altre issue su C++

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.