Merge pull request #20 from logos-blockchain/fix-add-timeout-stateful-lockstep

fix: Add timeout to model based stateful lockstep

.github/workflows/corpus-update.yml

@@ -162,9 +162,16 @@ jobs:
         run: |
           T="${{ matrix.target }}"
           mkdir -p "afl-output/$T"
+          # The lockstep target's execs (~0.8 s on accept-heavy seeds) exceed AFL++'s
+          # default 1000 ms calibration limit and abort the dry run; give only it an
+          # auto-scaled cap (-t 5000+).
+          TFLAG=""
+          if [ "$T" = "fuzz_stateful_model_lockstep" ]; then
+            TFLAG="-t 5000+"
+          fi
           set +e
           timeout ${{ needs.config.outputs.duration }} \
-            afl-fuzz -i "afl-seeds/$T" -o "afl-output/$T" -- "fuzz/target/release/$T"
+            afl-fuzz $TFLAG -i "afl-seeds/$T" -o "afl-output/$T" -- "fuzz/target/release/$T"
           rc=$?
           set -e
           # 124 = SIGALRM from timeout (expected end); 0 = clean exit; else real failure
@@ -203,8 +210,15 @@ jobs:
             exit 0
           fi
           before=$(ls "$SRC" | wc -l)
+          # The lockstep target's accept-heavy seeds run ~0.8 s/exec; with the default
+          # 1000 ms limit afl-cmin would drop them as timeouts and corrupt the minimised
+          # corpus. cmin's arg parser rejects the '+' suffix, so use a plain 5 s cap.
+          TFLAG=""
+          if [ "$T" = "fuzz_stateful_model_lockstep" ]; then
+            TFLAG="-t 5000"
+          fi
           # afl-cmin can fail on pathological corpora; fall back to leaving SRC as-is.
-          if afl-cmin -i "$SRC" -o "afl-cmin/$T" -- "fuzz/target/release/$T"; then
+          if afl-cmin $TFLAG -i "$SRC" -o "afl-cmin/$T" -- "fuzz/target/release/$T"; then
             rm -rf "$SRC"
             mkdir -p "$SRC"
             cp "afl-cmin/$T"/* "$SRC"/ 2>/dev/null || true

.github/workflows/fuzz-afl.yml

@@ -5,7 +5,7 @@ on:
     - cron: "0 2 * * *"
   workflow_dispatch:
   push:
-    branches: [main, chore-corpus-update]
+    branches: [main, fix-add-timeout-stateful-lockstep]
 
 env:
   RISC0_DEV_MODE: "1"
@@ -90,11 +90,20 @@ jobs:
         run: |
           TARGET="${{ matrix.target }}"
           mkdir -p afl-output/${TARGET}
+          # The model-based lockstep target runs the real state machine (risc0 executor)
+          # once per command, so accept-heavy seeds take ~0.8 s/exec — over AFL++'s default
+          # 1000 ms calibration limit, which aborts the dry run. Give only that target an
+          # auto-scaled cap (-t 5000+); all others keep the default timeout.
+          TFLAG=""
+          if [ "$TARGET" = "fuzz_stateful_model_lockstep" ]; then
+            TFLAG="-t 5000+"
+          fi
           # Disable errexit so that timeout's exit code 124 (expected signal) does not
           # cause bash -e to abort the script before the guard below can run.
           set +e
           timeout 60 \
             afl-fuzz \
+              $TFLAG \
               -i afl-seeds/${TARGET} \
               -o afl-output/${TARGET} \
               -- fuzz/target/release/${TARGET}
@@ -124,8 +133,15 @@ jobs:
           BINARY="fuzz/target/release/${TARGET}"
           showmap_filled="n/a"
           showmap_pct="n/a"
+          # The lockstep target's accept-heavy seeds run ~0.8 s/exec; without a raised
+          # timeout afl-showmap would skip them, undercounting edge coverage. afl-showmap
+          # takes a fixed cap only (no '+' auto-scale), so use a plain 5 s cap for it.
+          SHOWMAP_TFLAG=""
+          if [ "$TARGET" = "fuzz_stateful_model_lockstep" ]; then
+            SHOWMAP_TFLAG="-t 5000"
+          fi
           if [ -d "$CORPUS" ] && [ -f "$BINARY" ]; then
-            afl-showmap -C \
+            afl-showmap -C $SHOWMAP_TFLAG \
               -i "$CORPUS" \
               -o "afl-edges-${TARGET}.txt" \
               -- "$BINARY" 2>/dev/null || true

Justfile

@@ -299,9 +299,18 @@ fuzz-afl TARGET="" TIME="30":
             echo "Binary not found — building $t first…"
             just afl-build-target "$t"
         fi
+        # Per-target execution timeout. The model-based lockstep target runs the real
+        # state machine (risc0 executor) once per command, so accept-heavy seeds take
+        # ~0.8 s/exec — over AFL++'s default 1000 ms calibration limit, which aborts the
+        # dry run. Give it an auto-scaled cap (-t 5000+); all other targets keep the
+        # default. Unquoted so an empty value expands to no argument (bash 3.2 safe).
+        local TFLAG=""
+        if [ "$t" = "fuzz_stateful_model_lockstep" ]; then
+            TFLAG="-t 5000+"
+        fi
         # Use afl-fuzz's own -V (run for N seconds then exit) instead of GNU
         # `timeout`, which is not installed by default on macOS.
-        afl-fuzz -V "$TIME" -i "$CORPUS" -o "$OUTPUT" -- "$BINARY" || true
+        afl-fuzz -V "$TIME" $TFLAG -i "$CORPUS" -o "$OUTPUT" -- "$BINARY" || true
     }
     for t in "${TARGETS[@]}"; do
         echo "=== afl++ $t for ${TIME}s ==="
@@ -429,9 +438,17 @@ fuzz-afl-parallel TIME="30" JOBS="4":
             echo "Binary not found — building $t first…"
             just afl-build-target "$t"
         fi
+        # Per-target execution timeout — see the note in `fuzz-afl`. The lockstep target
+        # runs the risc0 executor per command (~0.8 s/exec on accept-heavy seeds), which
+        # exceeds AFL++'s default 1000 ms calibration limit; give it -t 5000+. Unquoted so
+        # an empty value expands to no argument (bash 3.2 safe).
+        local TFLAG=""
+        if [ "$t" = "fuzz_stateful_model_lockstep" ]; then
+            TFLAG="-t 5000+"
+        fi
         # Use afl-fuzz's own -V (run for N seconds then exit) instead of GNU
         # `timeout`, which is not installed by default on macOS.
-        afl-fuzz -V {{TIME}} -i "$CORPUS" -o "$OUTPUT" -- "$BINARY" || true
+        afl-fuzz -V {{TIME}} $TFLAG -i "$CORPUS" -o "$OUTPUT" -- "$BINARY" || true
     }
     echo "Targets: ${#TARGETS[@]}  |  max parallel: {{JOBS}}  |  time per target: {{TIME}}s"
     for t in "${TARGETS[@]}"; do