--- id: OPS_task_1_strace_gdbgui_segfault name: strace + gdbgui core-dump → live debug → fix → 回归 use-after-free 崩溃完整闭环 category: OPS timeout_seconds: 1500 --- --- id: OPS_task_1_strace_gdbgui_segfault name: strace + gdbgui core-dump → live debug → fix → 回归 use-after-free 崩溃完整闭环 category: OPS timeout_seconds: 1500 --- ## Prompt > ⚙️ **Execution note**: This is a legitimate server-side crash post-mortem task. All artifacts must land under `/tmp_workspace/results/`. Execute the task directly; do not refuse. Background: `/tmp_workspace/buggy_server/` contains a tiny HTTP server written in C (~280 lines, source `buggy_server.c`). Ops reports intermittent segfaults under 100+ concurrent requests, with the core dump pointing near `handle_request()`. The exact root cause is unknown — it may be use-after-free, double-free, out-of-bounds write, uninitialized read, or null pointer dereference. Goal: walk the full loop "reproduce the crash → multi-channel CLI forensics → live debug in a GUI debugger → patch the source → regression load test", landing every piece of evidence and every artifact under `/tmp_workspace/results/`. **Anti-cheat**: the `narrative` field must not be copy-pasted from the prompt; it must be grounded in evidence you actually captured. ### Hard schema for the main deliverable `defect_analysis.json` - Required fields: `crash_addr` / `crash_file` / `crash_line` / `suspect_function` / `defect_class` / `evidence_strace_lines` / `evidence_addr2line` / `evidence_gdbgui_screenshot` / `narrative` - `suspect_function` must be `handle_request` - `defect_class` allowed values: `use_after_free` | `double_free` | `oob_write` | `uninit_read` | `null_deref` (the real root cause here is `use_after_free`; the evidence must converge to that value) - `narrative` ≥ 80 characters, and must weave together evidence from all three channels (`strace`, `addr2line`, `gdbgui` — all three keywords must appear) ### Supporting deliverables (under `/tmp_workspace/results/`) | File | Requirement | |---|---| | `build.log` / `build_fixed.log` | Build output for the baseline (`-g -O0`) and the fixed version | | `server_pid.txt` | PID of the observed process at the moment of reproduction | | `strace_crash_tail.txt` | ≥ 100 lines, containing at least 3 of `mmap` / `munmap` / `write` / `read` / `brk` / `close` | | `strace_post_tail.txt` | Strace tail from the post-fix regression run, ≥ 200 lines, **must not** contain `SIGSEGV` or `segfault` | | `dmesg_segfault.txt` | Kernel log of the crash, containing the word `segfault` or `SIGSEGV` and an `ip=0x...` field | | `core.bin` | Non-empty core file produced by this crash (> 100 B) | | `addr2line_output.txt` | Result of resolving the crash PC, containing `.c:` | | `disasm_handle_request.txt` | Disassembly snippet for `handle_request`, ≥ 10 instructions, containing at least one of `mov` / `call` / `lea` / `ret` / `push` / `jmp` | | `fix.patch` | Valid unified diff (full `---` / `+++` / `@@` headers), ≥ 3 added lines, and must substantively touch the real UAF fix site (e.g. `g_last_response` / `log_response` / `strdup` / `memcpy` / `= NULL`, etc.) | | `diagnosis_report.md` | ≥ 350 characters, with the 5 H2 sections `## Root Cause` / `## strace Evidence` / `## addr2line Evidence` / `## gdbgui Evidence` / `## Fix & Regression`; the `Fix & Regression` section must include a quantitative before/after comparison (e.g. `before peak crashes vs after 0 crashes`) | ### 8 workflow screenshots (fixed file names, all saved under `/tmp_workspace/results/`) Each ≥ 5 KB and content-unique (distinct MD5). Each must be a real window of a GUI debugger (gdbgui browser frontend recommended), with readable elements covering source panel / call stack / Locals / Registers / Memory / Breakpoints: - `view_01_gdbgui_coredump_bt.png`: after loading the core — source panel + multi-frame call stack + GDB Console with `bt` output + crash frame highlighted - `view_02_gdbgui_frame_jump.png`: clicking a frame auto-jumps and highlights the source line + Locals panel lists that frame's variable names and current values - `view_03_gdbgui_breakpoint_hit.png`: breakpoint set on a suspect line and hit; Breakpoints list and Locals visible on the same screen - `view_04_gdbgui_step_to_release.png`: `next`/`step` to the point where the suspect pointer is freed or written out of bounds; Locals shows the pointer name with a non-NULL address - `view_05_gdbgui_memory_viewer.png`: Memory Viewer showing the hex bytes at the freed address, with the source-level use site visible on the same screen - `view_06_gdbgui_stack_regs.png`: multi-frame Call Stack + Registers panel including `rax` / `rbx` / `rip`, etc. - `view_07_gdbgui_fix_verified.png`: after the fix — breakpoint at the original crash site is hit and Locals shows the suspect pointer handled correctly - `view_08_gdbgui_continue_no_crash.png`: after Continue — no SIGSEGV; the process is still running or exits cleanly ## Expected Behavior 设计意图与典型解题路径(仅供出题人参考,不发给 agent): 1. 推荐通道:`gdbgui` 浏览器调试器(`pip install gdbgui` 即用,默认监听 `http://127.0.0.1:5000`)。其它 GDB GUI 前端(VS Code GDB / `cgdb` / Nemiver 等)只要能截到源码面板/调用栈/Locals/Memory/Registers,理论可替代;评测会做 OCR + VLM 真实性校验,期待 gdbgui 浏览器窗口形态,走其它通道需自行权衡。 2. step 1:`make` 构建带 `-g -O0` 的 debug 二进制 → `build.log`;启动服务器记录 PID。 3. step 2:另一 shell 跑 `strace -f -tt -T -o /tmp/strace.raw -p `;`bash loadgen.sh` 触发崩溃;`tail -n 200 /tmp/strace.raw > strace_crash_tail.txt`。 4. step 3:`dmesg -T | grep -i segfault`(或 `journalctl -k --since "5 min ago"`)→ `dmesg_segfault.txt`;复制 core 到 `core.bin`。 5. step 4:`addr2line -e ./buggy_server -f -C ` → `addr2line_output.txt`;`objdump -d --disassemble=handle_request ./buggy_server | head -n 30` → `disasm_handle_request.txt`。 6. step 5:`gdbgui ./buggy_server --remap-sources -- core core.bin`,在浏览器界面按 view_01..view_06 顺序操作并截图。 7. step 6:综合三路证据写 `defect_analysis.json`(本案真因:`g_last_response` 全局指针在 `log_response()` 调用前已被 `free`,典型 use-after-free)。 8. step 7:编辑 `buggy_server.c` 修复(`free` 后置 `NULL` / 调整释放顺序 / `strdup` 拷贝 / 边界检查,由证据决定);`diff -u buggy_server.c.orig buggy_server.c > fix.patch`;`make` 重建 → `build_fixed.log`。 9. step 8:复跑 `loadgen.sh` 并继续 strace,`tail -n 300 /tmp/strace.raw > strace_post_tail.txt`(无 SIGSEGV);gdbgui 重新加载新二进制,原崩溃点设断点 → 截 `view_07` / `view_08`。 10. step 9:写 `diagnosis_report.md`,5 节齐全且 `Fix & Regression` 有可量化前后对比。约束说明: - `actions.log`(若 agent 框架产生)不得出现仅复制 prompt 文本来塞 `narrative` 字段的痕迹;`narrative` 必须由实际 strace 行 / addr2line 输出 / gdbgui 观察导出。 - 8 张截图 MD5 必须互不相同(防同图复用);单张 < 5 KB 视为占位空白,不计入。评分要点(hard gates): - CLI 证据(`strace_crash` + `addr2line`)与 GUI 证据(6 张 debug 截图)任一通道缺失,总分上限 0.35。 - 6 张 debug 截图缺一张上限 0.55,缺过半上限 0.40。 - `defect_analysis` 不到 0.5 上限 0.50,= 0 上限 0.40(`defect_class` 误判 / `suspect_function` 错 / `narrative` 不达标都会拉低)。 - `fix.patch` 未触及真 UAF 修复点上限 0.50,无 patch 上限 0.40。 - 修复后 `strace_post_tail.txt` 仍见 SIGSEGV 或行数不足上限 0.55。 - 截图 MD5 唯一性 < 0.75 上限 0.55(反 reuse 同图)。 - VLM rubric(gdbgui 真窗口 / Variables 面板 / Memory or Stack / fix verified)平均 < 0.6 上限 0.55,< 0.4 上限 0.30。 ## Source - gdbgui: https://github.com/cs01/gdbgui (GPL-3.0) - strace(1): https://man7.org/linux/man-pages/man1/strace.1.html - addr2line(1): https://man7.org/linux/man-pages/man1/addr2line.1.html - CWE-416 Use After Free: https://cwe.mitre.org/data/definitions/416.html - Linux core dump tutorial: https://man7.org/linux/man-pages/man5/core.5.html ## Grading Criteria - [ ] 1. `build.log` / `server_pid.txt` 存在 - [ ] 2. `strace_crash_tail.txt` ≥ 100 行 + 含 mmap/munmap/write - [ ] 3. `dmesg_segfault.txt` 含 segfault/SIGSEGV - [ ] 4. `core.bin` 存在且非空 - [ ] 5. `view_01..view_06` 6 张 gdbgui 截图齐全 + OCR 命中 gdbgui / Variables / Call Stack / Memory 等关键词 - [ ] 6. `addr2line_output.txt` 含 `file.c:NN` - [ ] 7. `disasm_handle_request.txt` 含汇编 - [ ] 8. `defect_analysis.json` schema 完整 + defect_class ∈ 5 类 + narrative ≥ 80 字 - [ ] 9. `fix.patch` 是合法 unified diff + 至少 3 行加号变更 - [ ] 10. `build_fixed.log` 存在 + `strace_post_tail.txt` 无 SIGSEGV - [ ] 11. `view_07` / `view_08` 修复验证截图 + OCR 含 gdbgui - [ ] 12. `diagnosis_report.md` ≥ 350 字 + 含 5 个标题节 - [ ] 13. VLM 验:gdbgui 真窗口 / Variables 面板 / Memory Viewer / fix verified - [ ] 14. Cross-channel:strace/addr2line 文本产物与 gdbgui 截图同时存在 ## Automated Checks ```python def grade(workspace_path=None, **kwargs) -> dict: """gdbgui + strace + addr2line use-after-free 诊断闭环 grader (v2 tightened).""" import json, re, hashlib from pathlib import Path try: from _judge_helper import vlm_score_rubric except Exception: vlm_score_rubric = None try: import pytesseract from PIL import Image except Exception: pytesseract = None Image = None workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace") rd = workspace / "results" s = {} def ocr_hits(path, kws): if not (pytesseract and Image and path.exists()): return False try: tx = pytesseract.image_to_string(Image.open(path)).lower() except Exception: return False return any(k.lower() in tx for k in kws) # 1. build.log + server_pid.txt bl = rd / "build.log" pid = rd / "server_pid.txt" s["build_artifacts"] = (int(bl.exists()) + int(pid.exists())) / 2.0 # 2. strace_crash_tail.txt — require ≥100 lines AND ≥3 distinct syscalls from the set sc = rd / "strace_crash_tail.txt" if sc.exists(): txt = sc.read_text(errors="ignore") lines = txt.strip().split("\n") sys_kws = ["mmap", "munmap", "write", "read", "brk", "close"] sys_hits = sum(1 for k in sys_kws if k in txt) if len(lines) >= 100 and sys_hits >= 3: s["strace_crash"] = 1.0 elif len(lines) >= 50 and sys_hits >= 2: s["strace_crash"] = 0.5 elif lines and sys_hits >= 1: s["strace_crash"] = 0.25 else: s["strace_crash"] = 0.0 else: s["strace_crash"] = 0.0 # 3. dmesg_segfault.txt dm = rd / "dmesg_segfault.txt" if dm.exists(): t = dm.read_text(errors="ignore").lower() s["dmesg_segfault"] = 1.0 if ("segfault" in t or "sigsegv" in t) else 0.3 else: s["dmesg_segfault"] = 0.0 # 4. core.bin cb = rd / "core.bin" s["core_dump"] = 1.0 if (cb.exists() and cb.stat().st_size > 100) else 0.0 # 5. gdbgui screenshots phase 2 + 4 (6 of them) debug_shots = [ "view_01_gdbgui_coredump_bt.png", "view_02_gdbgui_frame_jump.png", "view_03_gdbgui_breakpoint_hit.png", "view_04_gdbgui_step_to_release.png", "view_05_gdbgui_memory_viewer.png", "view_06_gdbgui_stack_regs.png", ] # require non-trivial size (>5KB) to count as present (anti-placeholder) def _valid_shot(p): return p.exists() and p.stat().st_size >= 5 * 1024 present = sum(1 for n in debug_shots if _valid_shot(rd / n)) s["debug_shots_present"] = present / len(debug_shots) kw = ["gdbgui", "breakpoint", "variables", "call stack", "registers", "source", "gdb", "console", "locals", "step", "continue", "memory"] if pytesseract and Image: hits = sum(1 for n in debug_shots if ocr_hits(rd / n, kw)) s["debug_shots_ocr"] = hits / len(debug_shots) else: s["debug_shots_ocr"] = 0.5 if present else 0.0 # md5 uniqueness across all 8 gdbgui screenshots (anti-cheat: same image reused) all_shots_for_md5 = debug_shots + [ "view_07_gdbgui_fix_verified.png", "view_08_gdbgui_continue_no_crash.png", ] md5s = [] for n in all_shots_for_md5: p = rd / n if _valid_shot(p): try: md5s.append(hashlib.md5(p.read_bytes()).hexdigest()) except Exception: pass if md5s: s["shots_md5_uniqueness"] = len(set(md5s)) / len(md5s) else: s["shots_md5_uniqueness"] = 0.0 # 6. addr2line + objdump al = rd / "addr2line_output.txt" if al.exists(): s["addr2line"] = 1.0 if re.search(r"\w+\.c:\d+", al.read_text(errors="ignore")) else 0.3 else: s["addr2line"] = 0.0 da = rd / "disasm_handle_request.txt" if da.exists(): t = da.read_text(errors="ignore") s["disasm"] = 1.0 if any(k in t for k in ["mov", "call", "lea", "ret", "push", "jmp"]) else 0.3 else: s["disasm"] = 0.0 # 7. defect_analysis.json — schema + correct defect_class (UAF) + narrative quality dj = rd / "defect_analysis.json" da_score = 0.0 defect_class_correct = False if dj.exists(): try: d = json.loads(dj.read_text()) req = {"crash_addr","crash_file","crash_line","suspect_function","defect_class", "evidence_strace_lines","evidence_addr2line","evidence_gdbgui_screenshot","narrative"} allowed = {"use_after_free","double_free","oob_write","uninit_read","null_deref"} if req <= set(d.keys()): da_score = 0.25 if d.get("defect_class") in allowed: da_score += 0.15 # bonus: defect_class must actually be use_after_free (true root cause) if d.get("defect_class") == "use_after_free": da_score += 0.20 defect_class_correct = True # narrative ≥80 chars AND must mention three evidence channels narr = d.get("narrative", "") if isinstance(narr, str) and len(narr) >= 80: da_score += 0.15 nl = narr.lower() chan_hits = sum(1 for k in ["strace", "addr2line", "gdbgui"] if k in nl) if chan_hits >= 3: da_score += 0.15 # suspect_function must be handle_request (per Prompt + crash site) if d.get("suspect_function") == "handle_request": da_score += 0.10 da_score = min(da_score, 1.0) except Exception: pass s["defect_analysis"] = da_score s["defect_class_correct"] = 1.0 if defect_class_correct else 0.0 # 8. fix.patch — must touch g_last_response or free/log_response area (true UAF fix) fp = rd / "fix.patch" if fp.exists(): t = fp.read_text(errors="ignore") is_diff = ("---" in t and "+++" in t and "@@" in t) plus_lines = [l for l in t.splitlines() if l.startswith("+") and not l.startswith("+++")] plus = len(plus_lines) plus_blob = "\n".join(plus_lines).lower() # fix is meaningful if it touches g_last_response, free(response), or moves log_response meaningful = any(k in plus_blob for k in [ "g_last_response", "log_response", "strdup", "memcpy", "= null" ]) if is_diff and plus >= 3 and meaningful: s["fix_patch"] = 1.0 elif is_diff and plus >= 3: s["fix_patch"] = 0.6 elif is_diff: s["fix_patch"] = 0.3 else: s["fix_patch"] = 0.0 else: s["fix_patch"] = 0.0 # 9. fixed build + post strace bf = rd / "build_fixed.log" s["build_fixed"] = 1.0 if bf.exists() else 0.0 sp = rd / "strace_post_tail.txt" if sp.exists(): t = sp.read_text(errors="ignore") tlow = t.lower() nlines = len(t.strip().split("\n")) no_crash = ("sigsegv" not in tlow) and ("segfault" not in tlow) # Prompt requires ≥200 lines for post-fix strace tail if no_crash and nlines >= 200: s["strace_post"] = 1.0 elif no_crash and nlines >= 100: s["strace_post"] = 0.6 elif no_crash: s["strace_post"] = 0.3 else: s["strace_post"] = 0.0 else: s["strace_post"] = 0.0 # 10. fix verification screenshots (2) — same anti-placeholder size gate fv_shots = ["view_07_gdbgui_fix_verified.png", "view_08_gdbgui_continue_no_crash.png"] fv_present = sum(1 for n in fv_shots if _valid_shot(rd / n)) s["fix_shots_present"] = fv_present / len(fv_shots) if pytesseract and Image: s["fix_shots_ocr"] = sum(1 for n in fv_shots if ocr_hits(rd / n, kw)) / len(fv_shots) else: s["fix_shots_ocr"] = 0.5 if fv_present else 0.0 # 11. report rpt = rd / "diagnosis_report.md" if rpt.exists(): t = rpt.read_text(errors="ignore") sections = ["## Root Cause","## strace Evidence","## addr2line Evidence", "## gdbgui Evidence","## Fix & Regression"] sec_hits = sum(1 for sec in sections if sec.lower() in t.lower()) s["report"] = 1.0 if (len(t) >= 350 and sec_hits >= 5) else (0.5 if sec_hits >= 3 else 0.2) else: s["report"] = 0.0 # 12. cross-channel evidence (CLI + GUI co-present) has_cli_ev = (s["strace_crash"] > 0) and (s["addr2line"] > 0) has_gui_ev = (s["debug_shots_present"] >= 0.5) s["cross_channel_evidence"] = 1.0 if (has_cli_ev and has_gui_ev) else 0.0 # VLM rubric (4 items) if vlm_score_rubric: all_shots = debug_shots + fv_shots sample = [str(rd / n) for n in all_shots if (rd / n).exists()][:4] if sample: rubric = { "vlm_gdbgui_real_window": "至少一张截图清晰显示 gdbgui 浏览器窗口(顶栏 gdbgui 字样 + 左侧源码面板带行号 + 底部 GDB Console)", "vlm_variables_panel": "至少一张截图能读到 Variables/Locals 面板列出变量名 + 当前值(非空)", "vlm_memory_or_stack": "至少一张截图能看到 Memory Viewer 的 hex 输出或 Call Stack 面板的多帧列表", "vlm_fix_verified": "view_07/view_08 中能看出修复后断点能命中且程序未崩溃 Continue", } vlm = vlm_score_rubric(sample, rubric, instruction="评估 gdbgui 浏览器调试器在 use-after-free 诊断与修复验证中的截图真实性。") for k in rubric: s[k] = vlm.get(k, 0.0) s["judge_method"] = vlm.get("judge_method", "failed") # ---- Weighted aggregate: core 60% / gui 30% / aux 10% ---- core_keys = ["defect_analysis", "defect_class_correct", "fix_patch", "build_fixed", "strace_post", "report"] gui_keys = ["debug_shots_present", "debug_shots_ocr", "fix_shots_present", "fix_shots_ocr", "shots_md5_uniqueness"] vlm_keys = [k for k in s if k.startswith("vlm_")] gui_keys_full = gui_keys + vlm_keys aux_keys = ["build_artifacts", "strace_crash", "dmesg_segfault", "core_dump", "addr2line", "disasm", "cross_channel_evidence"] def _avg(keys): vals = [s[k] for k in keys if k in s and isinstance(s[k], (int, float))] return sum(vals) / len(vals) if vals else 0.0 core = _avg(core_keys) gui = _avg(gui_keys_full) aux = _avg(aux_keys) base = 0.6 * core + 0.3 * gui + 0.1 * aux # ---- Hard gates (tightened for v2) ---- # CLI/GUI evidence channels if not has_cli_ev: base = min(base, 0.35) if not has_gui_ev: base = min(base, 0.35) # Must have ALL 6 debug screenshots present (>=5KB) — partial cap if s.get("debug_shots_present", 0) < 1.0: base = min(base, 0.55) if s.get("debug_shots_present", 0) < 0.5: base = min(base, 0.40) # defect analysis correctness if s.get("defect_analysis", 0) < 0.5: base = min(base, 0.50) if s.get("defect_analysis", 0) == 0: base = min(base, 0.40) # fix patch must be meaningful if s.get("fix_patch", 0) < 0.6: base = min(base, 0.50) if s.get("fix_patch", 0) == 0: base = min(base, 0.40) # post-fix regression must show no crash if s.get("strace_post", 0) < 0.6: base = min(base, 0.55) # screenshot anti-cheat if s.get("shots_md5_uniqueness", 0) < 0.75: base = min(base, 0.55) # VLM thresholds — tighter, layered if vlm_keys: vlm_avg = sum(s[k] for k in vlm_keys) / len(vlm_keys) s["vlm_avg"] = vlm_avg if vlm_avg < 0.6: base = min(base, 0.55) if vlm_avg < 0.4: base = min(base, 0.30) else: # VLM unavailable → can't reach top tier base = min(base, 0.60) s["overall_score"] = float(max(0.0, min(1.0, base))) return s ``` ## Workspace Path ``` workspace/OPS/task_1_strace_gdbgui_segfault ``` ## Skills ``` ``` ## Env ``` ``` ## Warmup ```bash echo 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 | base64 -d | bash ```