1use crash_handler::{CrashEventResult, CrashHandler};
2use log::info;
3use minidumper::{Client, LoopAction, MinidumpBinary};
4use release_channel::{RELEASE_CHANNEL, ReleaseChannel};
5use serde::{Deserialize, Serialize};
6
7#[cfg(not(target_os = "windows"))]
8use smol::process::Command;
9
10#[cfg(target_os = "macos")]
11use std::sync::atomic::AtomicU32;
12use std::{
13 env,
14 fs::{self, File},
15 io,
16 panic::{self, PanicHookInfo},
17 path::{Path, PathBuf},
18 process::{self},
19 sync::{
20 Arc, OnceLock,
21 atomic::{AtomicBool, Ordering},
22 },
23 thread,
24 time::Duration,
25};
26
27// set once the crash handler has initialized and the client has connected to it
28pub static CRASH_HANDLER: OnceLock<Arc<Client>> = OnceLock::new();
29// set when the first minidump request is made to avoid generating duplicate crash reports
30pub static REQUESTED_MINIDUMP: AtomicBool = AtomicBool::new(false);
31const CRASH_HANDLER_PING_TIMEOUT: Duration = Duration::from_secs(60);
32const CRASH_HANDLER_CONNECT_TIMEOUT: Duration = Duration::from_secs(10);
33
34#[cfg(target_os = "macos")]
35static PANIC_THREAD_ID: AtomicU32 = AtomicU32::new(0);
36
37pub async fn init(crash_init: InitCrashHandler) {
38 let gen_var = match env::var("ZED_GENERATE_MINIDUMPS") {
39 Ok(v) => {
40 if v == "false" || v == "0" {
41 Some(false)
42 } else {
43 Some(true)
44 }
45 }
46 Err(_) => None,
47 };
48
49 match (gen_var, *RELEASE_CHANNEL) {
50 (Some(false), _) | (None, ReleaseChannel::Dev) => {
51 let old_hook = panic::take_hook();
52 panic::set_hook(Box::new(move |info| {
53 unsafe { env::set_var("RUST_BACKTRACE", "1") };
54 old_hook(info);
55 // prevent the macOS crash dialog from popping up
56 if cfg!(target_os = "macos") {
57 std::process::exit(1);
58 }
59 }));
60 return;
61 }
62 _ => {
63 panic::set_hook(Box::new(panic_hook));
64 }
65 }
66
67 let exe = env::current_exe().expect("unable to find ourselves");
68 let zed_pid = process::id();
69 // TODO: we should be able to get away with using 1 crash-handler process per machine,
70 // but for now we append the PID of the current process which makes it unique per remote
71 // server or interactive zed instance. This solves an issue where occasionally the socket
72 // used by the crash handler isn't destroyed correctly which causes it to stay on the file
73 // system and block further attempts to initialize crash handlers with that socket path.
74 let socket_name = paths::temp_dir().join(format!("zed-crash-handler-{zed_pid}"));
75 #[cfg(not(target_os = "windows"))]
76 let _crash_handler = Command::new(exe)
77 .arg("--crash-handler")
78 .arg(&socket_name)
79 .spawn()
80 .expect("unable to spawn server process");
81
82 #[cfg(target_os = "windows")]
83 spawn_crash_handler_windows(&exe, &socket_name);
84
85 #[cfg(target_os = "linux")]
86 let server_pid = _crash_handler.id();
87 info!("spawning crash handler process");
88
89 let mut elapsed = Duration::ZERO;
90 let retry_frequency = Duration::from_millis(100);
91 let mut maybe_client = None;
92 while maybe_client.is_none() {
93 if let Ok(client) = Client::with_name(socket_name.as_path()) {
94 maybe_client = Some(client);
95 info!("connected to crash handler process after {elapsed:?}");
96 break;
97 }
98 elapsed += retry_frequency;
99 smol::Timer::after(retry_frequency).await;
100 }
101 let client = maybe_client.unwrap();
102 client
103 .send_message(1, serde_json::to_vec(&crash_init).unwrap())
104 .unwrap();
105
106 let client = Arc::new(client);
107 let handler = CrashHandler::attach(unsafe {
108 let client = client.clone();
109 crash_handler::make_crash_event(move |crash_context: &crash_handler::CrashContext| {
110 // only request a minidump once
111 let res = if REQUESTED_MINIDUMP
112 .compare_exchange(false, true, Ordering::Acquire, Ordering::Relaxed)
113 .is_ok()
114 {
115 #[cfg(target_os = "macos")]
116 suspend_all_other_threads();
117
118 // on macos this "ping" is needed to ensure that all our
119 // `client.send_message` calls have been processed before we trigger the
120 // minidump request.
121 client.ping().ok();
122 client.request_dump(crash_context).is_ok()
123 } else {
124 true
125 };
126 CrashEventResult::Handled(res)
127 })
128 })
129 .expect("failed to attach signal handler");
130
131 #[cfg(target_os = "linux")]
132 {
133 handler.set_ptracer(Some(server_pid));
134 }
135 CRASH_HANDLER.set(client.clone()).ok();
136 std::mem::forget(handler);
137 info!("crash handler registered");
138
139 loop {
140 client.ping().ok();
141 smol::Timer::after(Duration::from_secs(10)).await;
142 }
143}
144
145#[cfg(target_os = "macos")]
146unsafe fn suspend_all_other_threads() {
147 let task = unsafe { mach2::traps::current_task() };
148 let mut threads: mach2::mach_types::thread_act_array_t = std::ptr::null_mut();
149 let mut count = 0;
150 unsafe {
151 mach2::task::task_threads(task, &raw mut threads, &raw mut count);
152 }
153 let current = unsafe { mach2::mach_init::mach_thread_self() };
154 let panic_thread = PANIC_THREAD_ID.load(Ordering::SeqCst);
155 for i in 0..count {
156 let t = unsafe { *threads.add(i as usize) };
157 if t != current && t != panic_thread {
158 unsafe { mach2::thread_act::thread_suspend(t) };
159 }
160 }
161}
162
163pub struct CrashServer {
164 initialization_params: OnceLock<InitCrashHandler>,
165 panic_info: OnceLock<CrashPanic>,
166 active_gpu: OnceLock<system_specs::GpuSpecs>,
167 has_connection: Arc<AtomicBool>,
168}
169
170#[derive(Debug, Deserialize, Serialize, Clone)]
171pub struct CrashInfo {
172 pub init: InitCrashHandler,
173 pub panic: Option<CrashPanic>,
174 pub minidump_error: Option<String>,
175 pub gpus: Vec<system_specs::GpuInfo>,
176 pub active_gpu: Option<system_specs::GpuSpecs>,
177}
178
179#[derive(Debug, Deserialize, Serialize, Clone)]
180pub struct InitCrashHandler {
181 pub session_id: String,
182 pub zed_version: String,
183 pub binary: String,
184 pub release_channel: String,
185 pub commit_sha: String,
186}
187
188#[derive(Deserialize, Serialize, Debug, Clone)]
189pub struct CrashPanic {
190 pub message: String,
191 pub span: String,
192}
193
194impl minidumper::ServerHandler for CrashServer {
195 fn create_minidump_file(&self) -> Result<(File, PathBuf), io::Error> {
196 let err_message = "Missing initialization data";
197 let dump_path = paths::logs_dir()
198 .join(
199 &self
200 .initialization_params
201 .get()
202 .expect(err_message)
203 .session_id,
204 )
205 .with_extension("dmp");
206 let file = File::create(&dump_path)?;
207 Ok((file, dump_path))
208 }
209
210 fn on_minidump_created(&self, result: Result<MinidumpBinary, minidumper::Error>) -> LoopAction {
211 let minidump_error = match result {
212 Ok(MinidumpBinary { mut file, path, .. }) => {
213 use io::Write;
214 file.flush().ok();
215 // TODO: clean this up once https://github.com/EmbarkStudios/crash-handling/issues/101 is addressed
216 drop(file);
217 let original_file = File::open(&path).unwrap();
218 let compressed_path = path.with_extension("zstd");
219 let compressed_file = File::create(&compressed_path).unwrap();
220 zstd::stream::copy_encode(original_file, compressed_file, 0).ok();
221 fs::rename(&compressed_path, path).unwrap();
222 None
223 }
224 Err(e) => Some(format!("{e:?}")),
225 };
226
227 #[cfg(not(any(target_os = "linux", target_os = "freebsd")))]
228 let gpus = vec![];
229
230 #[cfg(any(target_os = "linux", target_os = "freebsd"))]
231 let gpus = match system_specs::read_gpu_info_from_sys_class_drm() {
232 Ok(gpus) => gpus,
233 Err(err) => {
234 log::warn!("Failed to collect GPU information for crash report: {err}");
235 vec![]
236 }
237 };
238
239 let crash_info = CrashInfo {
240 init: self
241 .initialization_params
242 .get()
243 .expect("not initialized")
244 .clone(),
245 panic: self.panic_info.get().cloned(),
246 minidump_error,
247 active_gpu: self.active_gpu.get().cloned(),
248 gpus,
249 };
250
251 let crash_data_path = paths::logs_dir()
252 .join(&crash_info.init.session_id)
253 .with_extension("json");
254
255 fs::write(crash_data_path, serde_json::to_vec(&crash_info).unwrap()).ok();
256
257 LoopAction::Exit
258 }
259
260 fn on_message(&self, kind: u32, buffer: Vec<u8>) {
261 match kind {
262 1 => {
263 let init_data =
264 serde_json::from_slice::<InitCrashHandler>(&buffer).expect("invalid init data");
265 self.initialization_params
266 .set(init_data)
267 .expect("already initialized");
268 }
269 2 => {
270 let panic_data =
271 serde_json::from_slice::<CrashPanic>(&buffer).expect("invalid panic data");
272 self.panic_info.set(panic_data).expect("already panicked");
273 }
274 3 => {
275 let gpu_specs: system_specs::GpuSpecs =
276 bincode::deserialize(&buffer).expect("gpu specs");
277 // we ignore the case where it was already set because this message is sent
278 // on each new window. in theory all zed windows should be using the same
279 // GPU so this is fine.
280 self.active_gpu.set(gpu_specs).ok();
281 }
282 _ => {
283 panic!("invalid message kind");
284 }
285 }
286 }
287
288 fn on_client_disconnected(&self, _clients: usize) -> LoopAction {
289 LoopAction::Exit
290 }
291
292 fn on_client_connected(&self, _clients: usize) -> LoopAction {
293 self.has_connection.store(true, Ordering::SeqCst);
294 LoopAction::Continue
295 }
296}
297
298pub fn panic_hook(info: &PanicHookInfo) {
299 // Don't handle a panic on threads that are not relevant to the main execution.
300 if extension_host::wasm_host::IS_WASM_THREAD.with(|v| v.load(Ordering::Acquire)) {
301 log::error!("wasm thread panicked!");
302 return;
303 }
304
305 let message = info.payload_as_str().unwrap_or("Box<Any>").to_owned();
306
307 let span = info
308 .location()
309 .map(|loc| format!("{}:{}", loc.file(), loc.line()))
310 .unwrap_or_default();
311
312 let current_thread = std::thread::current();
313 let thread_name = current_thread.name().unwrap_or("<unnamed>");
314
315 // wait 500ms for the crash handler process to start up
316 // if it's still not there just write panic info and no minidump
317 let retry_frequency = Duration::from_millis(100);
318 for _ in 0..5 {
319 if let Some(client) = CRASH_HANDLER.get() {
320 let location = info
321 .location()
322 .map_or_else(|| "<unknown>".to_owned(), |location| location.to_string());
323 log::error!("thread '{thread_name}' panicked at {location}:\n{message}...");
324 client
325 .send_message(
326 2,
327 serde_json::to_vec(&CrashPanic { message, span }).unwrap(),
328 )
329 .ok();
330 log::error!("triggering a crash to generate a minidump...");
331
332 #[cfg(target_os = "macos")]
333 PANIC_THREAD_ID.store(
334 unsafe { mach2::mach_init::mach_thread_self() },
335 Ordering::SeqCst,
336 );
337
338 cfg_if::cfg_if! {
339 if #[cfg(target_os = "windows")] {
340 // https://learn.microsoft.com/en-us/windows/win32/debug/system-error-codes--0-499-
341 CrashHandler.simulate_exception(Some(234)); // (MORE_DATA_AVAILABLE)
342 break;
343 } else {
344 std::process::abort();
345 }
346 }
347 }
348 thread::sleep(retry_frequency);
349 }
350}
351
352#[cfg(target_os = "windows")]
353fn spawn_crash_handler_windows(exe: &Path, socket_name: &Path) {
354 use std::ffi::OsStr;
355 use std::iter::once;
356 use std::os::windows::ffi::OsStrExt;
357 use windows::Win32::System::Threading::{
358 CreateProcessW, PROCESS_CREATION_FLAGS, PROCESS_INFORMATION, STARTF_FORCEOFFFEEDBACK,
359 STARTUPINFOW,
360 };
361 use windows::core::PWSTR;
362
363 let mut command_line: Vec<u16> = OsStr::new(&format!(
364 "\"{}\" --crash-handler \"{}\"",
365 exe.display(),
366 socket_name.display()
367 ))
368 .encode_wide()
369 .chain(once(0))
370 .collect();
371
372 let mut startup_info = STARTUPINFOW::default();
373 startup_info.cb = std::mem::size_of::<STARTUPINFOW>() as u32;
374
375 // By default, Windows enables a "busy" cursor when a GUI application is launched.
376 // This cursor is disabled once the application starts processing window messages.
377 // Since the crash handler process doesn't process messages, this "busy" cursor stays enabled for a long time.
378 // Disable the cursor feedback to prevent this from happening.
379 startup_info.dwFlags = STARTF_FORCEOFFFEEDBACK;
380
381 let mut process_info = PROCESS_INFORMATION::default();
382
383 unsafe {
384 CreateProcessW(
385 None,
386 Some(PWSTR(command_line.as_mut_ptr())),
387 None,
388 None,
389 false,
390 PROCESS_CREATION_FLAGS(0),
391 None,
392 None,
393 &startup_info,
394 &mut process_info,
395 )
396 .expect("unable to spawn server process");
397
398 windows::Win32::Foundation::CloseHandle(process_info.hProcess).ok();
399 windows::Win32::Foundation::CloseHandle(process_info.hThread).ok();
400 }
401}
402
403pub fn crash_server(socket: &Path) {
404 let Ok(mut server) = minidumper::Server::with_name(socket) else {
405 log::info!("Couldn't create socket, there may already be a running crash server");
406 return;
407 };
408
409 let shutdown = Arc::new(AtomicBool::new(false));
410 let has_connection = Arc::new(AtomicBool::new(false));
411
412 thread::Builder::new()
413 .name("CrashServerTimeout".to_owned())
414 .spawn({
415 let shutdown = shutdown.clone();
416 let has_connection = has_connection.clone();
417 move || {
418 std::thread::sleep(CRASH_HANDLER_CONNECT_TIMEOUT);
419 if !has_connection.load(Ordering::SeqCst) {
420 shutdown.store(true, Ordering::SeqCst);
421 }
422 }
423 })
424 .unwrap();
425
426 server
427 .run(
428 Box::new(CrashServer {
429 initialization_params: OnceLock::new(),
430 panic_info: OnceLock::new(),
431 has_connection,
432 active_gpu: OnceLock::new(),
433 }),
434 &shutdown,
435 Some(CRASH_HANDLER_PING_TIMEOUT),
436 )
437 .expect("failed to run server");
438}