[clangd] Shutdown cleanly on signals.

[sammccall]

Summary

This avoids leaking PCH files if editors don't use the LSP shutdown protocol.

This is one fix for https://github.com/clangd/clangd/issues/209
(Though I think we should *also* be unlinking the files)

[merge_guards_bot]

Build result: fail - 60239 tests passed, 1 failed and 732 were skipped.

failed: Clangd.Clangd/exit-signal.test

Log files: console-log.txt, CMakeCache.txt

[sammccall]

Don't require temp file to already exist!

[kadircet]

Thanks, LGTM!

[merge_guards_bot]

Build result: fail - 60239 tests passed, 1 failed and 732 were skipped.

failed: Clangd.Clangd/exit-signal.test

Log files: console-log.txt, CMakeCache.txt

[sammccall]

Don't close the input stream early.

[merge_guards_bot]

Build result: pass - 60240 tests passed, 0 failed and 732 were skipped.

Log files: console-log.txt, CMakeCache.txt

clang-tools-extra/clangd/CMakeLists.txt

add_clang_library(clangDaemon
   Protocol.cpp
   Quality.cpp
   ParsedAST.cpp
   Preamble.cpp
   RIFF.cpp
   Selection.cpp
   SemanticHighlighting.cpp
   SemanticSelection.cpp
+  Shutdown.cpp
   SourceCode.cpp
   QueryDriverDatabase.cpp
   Threading.cpp
   Trace.cpp
   TUScheduler.cpp
   URI.cpp
   XRefs.cpp
 

clang-tools-extra/clangd/JSONTransport.cpp

 //===--- JSONTransport.cpp - sending and receiving LSP messages over JSON -===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 #include "Logger.h"
 #include "Protocol.h" // For LSPError
+#include "Shutdown.h"
 #include "Transport.h"
 #include "llvm/Support/Errno.h"
+#include "llvm/Support/Error.h"
 
 namespace clang {
 namespace clangd {
 namespace {
 
 llvm::json::Object encodeError(llvm::Error E) {
   std::string Message;
   ErrorCode Code = ErrorCode::UnknownErrorCode;
if (Result) {
           {"id", std::move(ID)},
           {"error", encodeError(Result.takeError())},
       });
     }
   }
 
   llvm::Error loop(MessageHandler &Handler) override {
     while (!feof(In)) {
+      if (shutdownRequested())
+        return llvm::createStringError(
+            std::make_error_code(std::errc::operation_canceled),
+            "Got signal, shutting down");
       if (ferror(In))
         return llvm::errorCodeToError(
             std::error_code(errno, std::system_category()));
       if (auto JSON = readRawMessage()) {
         if (auto Doc = llvm::json::parse(*JSON)) {
           vlog(Pretty ? "<<< {0:2}\n" : "<<< {0}\n", *Doc);
           if (!handleMessage(std::move(*Doc), Handler))
             return llvm::Error::success(); // we saw the "exit" notification.
bool JSONTransport::handleMessage(llvm::json::Value Message,
 
   if (ID)
     return Handler.onCall(*Method, std::move(Params), std::move(*ID));
   else
     return Handler.onNotify(*Method, std::move(Params));
 }
 
 // Tries to read a line up to and including \n.
-// If failing, feof() or ferror() will be set.
+// If failing, feof(), ferror(), or shutdownRequested() will be set.
 bool readLine(std::FILE *In, std::string &Out) {
   static constexpr int BufSize = 1024;
   size_t Size = 0;
   Out.clear();
   for (;;) {
     Out.resize(Size + BufSize);
     // Handle EINTR which is sent when a debugger attaches on some platforms.
-    if (!llvm::sys::RetryAfterSignal(nullptr, ::fgets, &Out[Size], BufSize, In))
+    if (!retryAfterSignalUnlessShutdown(
+            nullptr, [&] { return std::fgets(&Out[Size], BufSize, In); }))
       return false;
     clearerr(In);
     // If the line contained null bytes, anything after it (including \n) will
     // be ignored. Fortunately this is not a legal header or JSON.
     size_t Read = std::strlen(&Out[Size]);
     if (Read > 0 && Out[Size + Read - 1] == '\n') {
       Out.resize(Size + Read);
       return true;
     }
     Size += Read;
   }
 }
 
 // Returns None when:
-//  - ferror() or feof() are set.
+//  - ferror(), feof(), or shutdownRequested() are set.
 //  - Content-Length is missing or empty (protocol error)
 llvm::Optional<std::string> JSONTransport::readStandardMessage() {
   // A Language Server Protocol message starts with a set of HTTP headers,
   // delimited  by \r\n, and terminated by an empty line (\r\n).
   unsigned long long ContentLength = 0;
   std::string Line;
   while (true) {
     if (feof(In) || ferror(In) || !readLine(In, Line))
llvm::Optional<std::string> JSONTransport::readStandardMessage() {
   if (ContentLength == 0) {
     log("Warning: Missing Content-Length header, or zero-length message.");
     return llvm::None;
   }
 
   std::string JSON(ContentLength, '\0');
   for (size_t Pos = 0, Read; Pos < ContentLength; Pos += Read) {
     // Handle EINTR which is sent when a debugger attaches on some platforms.
-    Read = llvm::sys::RetryAfterSignal(0u, ::fread, &JSON[Pos], 1,
-                                       ContentLength - Pos, In);
+    Read = retryAfterSignalUnlessShutdown(0, [&]{
+      return std::fread(&JSON[Pos], 1, ContentLength - Pos, In);
+    });
     if (Read == 0) {
       elog("Input was aborted. Read only {0} bytes of expected {1}.", Pos,
            ContentLength);
       return llvm::None;
     }
     InMirror << llvm::StringRef(&JSON[Pos], Read);
     clearerr(In); // If we're done, the error was transient. If we're not done,
                   // either it was transient or we'll see it again on retry.
     Pos += Read;
   }
   return std::move(JSON);
 }
 
 // For lit tests we support a simplified syntax:
 // - messages are delimited by '---' on a line by itself
 // - lines starting with # are ignored.
 // This is a testing path, so favor simplicity over performance here.
-// When returning None, feof() or ferror() will be set.
+// When returning None, feof(), ferror(), or shutdownRequested() will be set.
 llvm::Optional<std::string> JSONTransport::readDelimitedMessage() {
   std::string JSON;
   std::string Line;
   while (readLine(In, Line)) {
     InMirror << Line;
     auto LineRef = llvm::StringRef(Line).trim();
     if (LineRef.startswith("#")) // comment
       continue;
 
     // found a delimiter
     if (LineRef.rtrim() == "---")
       break;
 
     JSON += Line;
   }
 
+  if (shutdownRequested())
+    return llvm::None;
   if (ferror(In)) {
     elog("Input error while reading message!");
     return llvm::None;
   }
   return std::move(JSON); // Including at EOF
 }
 
 } // namespace

clang-tools-extra/clangd/Shutdown.h

+//===--- Shutdown.h - Unclean exit scenarios --------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// LSP specifies a protocol for shutting down: a `shutdown` request followed
+// by an `exit` notification. If this protocol is followed, clangd should
+// finish outstanding work and exit with code 0.
+//
+// The way this works in the happy case:
+//  - when ClangdLSPServer gets `shutdown`, it sets a flag
+//  - when ClangdLSPServer gets `exit`, it returns false to indicate end-of-LSP
+//  - Transport::loop() returns with no error
+//  - ClangdServer::run() checks the shutdown flag and returns with no error.
+//  - we `return 0` from main()
+//  - destructor of ClangdServer and other main()-locals runs.
+//    This blocks until outstanding requests complete (results are ignored)
+//  - global destructors run, such as fallback deletion of temporary files
+//
+// There are a number of things that can go wrong. Some are handled here, and
+// some elsewhere.
+//  - `exit` notification with no `shutdown`:
+//    ClangdServer::run() sees this and returns false, main() returns nonzero.
+//  - stdin/stdout are closed
+//    The Transport detects this while doing IO and returns an error from loop()
+//    ClangdServer::run() logs a message and then returns false, etc
+//  - a request thread gets stuck, so the ClangdServer destructor hangs.
+//    Before returning from main(), we start a watchdog thread to abort() the
+//    process if it takes too long to exit. See abortAfterTimeout().
+//  - clangd crashes (e.g. segfault or assertion)
+//    A fatal signal is sent (SEGV, ABRT, etc)
+//    The installed signal handler prints a stack trace and exits.
+//  - parent process goes away or tells us to shut down
+//    A "graceful shutdown" signal is sent (TERM, HUP, etc).
+//    The installed signal handler calls requestShutdown() which sets a flag.
+//    The Transport IO is interrupted, and Transport::loop() checks the flag and
+//    returns an error, etc.
+//
+//===----------------------------------------------------------------------===//
+#ifndef LLVM_CLANG_TOOLS_EXTRA_CLANGD_SHUTDOWN_H
+#define LLVM_CLANG_TOOLS_EXTRA_CLANGD_SHUTDOWN_H
+
+#include <cerrno>
+#include <chrono>
+
+namespace clang {
+namespace clangd {
+
+/// Causes this process to crash if still running after Timeout.
+void abortAfterTimeout(std::chrono::seconds Timeout);
+
+/// Sets a flag to indicate that clangd was sent a shutdown signal, and the
+/// transport loop should exit at the next opportunity.
+/// If shutdown was already requested, aborts the process.
+/// This function is threadsafe and signal-safe.
+void requestShutdown();
+/// Checks whether requestShutdown() was called.
+/// This function is threadsafe and signal-safe.
+bool shutdownRequested();
+
+/// Retry an operation if it gets interrupted by a signal.
+/// This is like llvm::sys::RetryAfterSignal, except that if shutdown was
+/// requested (which interrupts IO), we'll fail rather than retry.
+template <typename Fun, typename Ret = decltype(std::declval<Fun>()())>
+Ret retryAfterSignalUnlessShutdown(
+    const typename std::enable_if<true, Ret>::type &Fail, // Suppress deduction.
+    const Fun &F) {
+  Ret Res;
+  do {
+    if (shutdownRequested())
+      return Fail;
+    errno = 0;
+    Res = F();
+  } while (Res == Fail && errno == EINTR);
+  return Res;
+}
+
+} // namespace clangd
+} // namespace clang
+
+#endif

clang-tools-extra/clangd/Shutdown.cpp

+//===--- Shutdown.cpp - Unclean exit scenarios ----------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "Shutdown.h"
+
+#include <atomic>
+#include <thread>
+
+namespace clang {
+namespace clangd {
+
+void abortAfterTimeout(std::chrono::seconds Timeout) {
+  // This is more portable than sys::WatchDog, and yields a stack trace.
+  std::thread([Timeout] {
+    std::this_thread::sleep_for(Timeout);
+    std::abort();
+  }).detach();
+}
+
+static std::atomic<bool> ShutdownRequested = {false};
+
+void requestShutdown() {
+  if (ShutdownRequested.exchange(true))
+    // This is the second shutdown request. Exit hard.
+    std::abort();
+}
+
+bool shutdownRequested() {
+  return ShutdownRequested;
+}
+
+} // namespace clangd
+} // namespace clang
+

clang-tools-extra/clangd/test/exit-eof.test

+# RUN: not clangd -sync < %s 2> %t.err
+# RUN: FileCheck %s < %t.err
+#
+# No LSP messages here, just let clangd see the end-of-file
+# CHECK: Transport error:
+# (Typically "Transport error: Input/output error" but platform-dependent).
+

clang-tools-extra/clangd/test/exit-signal.test

+# This is a fiddly signal test, we need POSIX and a real shell.
+UNSUPPORTED: win32
+REQUIRES: shell
+
+# Our goal is:
+#  1) spawn clangd
+#  2) wait for it to start up (install signal handlers)
+#  3) send SIGTERM
+#  4) wait for clangd to shut down (nonzero exit for a signal)
+#  4) verify the shutdown was clean
+
+RUN: rm %t.err
+     # To keep clangd running, we need to hold its input stream open.
+     # We redirect its input to a subshell that waits for it to start up.
+RUN: not clangd 2> %t.err < <( \
+       # Loop waiting for clangd to start up, so signal handlers are installed.
+       # Reading the PID line ensures this, and lets us send a signal.
+RUN:   while true; do \
+         # Relevant log line is I[timestamp] PID: <NNN>
+RUN:     CLANGD_PID=$(grep -a -m 1 "PID:" %t.err | cut -d' ' -f 3); \
+RUN:     [ ! -z "$CLANGD_PID" ] && break; \
+RUN:   done; \
+RUN:   kill $CLANGD_PID \
+     # Now wait for clangd to exit.
+RUN: )
+
+# Check that clangd caught the signal and shut down cleanly.
+RUN: FileCheck %s < %t.err
+CHECK: Transport error: Got signal
+CHECK: LSP finished
+

clang-tools-extra/clangd/tool/ClangdMain.cpp

 //===--- ClangdMain.cpp - clangd server loop ------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "ClangdLSPServer.h"
 #include "CodeComplete.h"
 #include "Features.inc"
 #include "Path.h"
 #include "Protocol.h"
+#include "Shutdown.h"
 #include "Trace.h"
 #include "Transport.h"
 #include "index/Background.h"
 #include "index/Serialization.h"
 #include "clang/Basic/Version.h"
 #include "clang/Format/Format.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/FileSystem.h"
 #include "llvm/Support/Path.h"
 #include "llvm/Support/Process.h"
 #include "llvm/Support/Program.h"
 #include "llvm/Support/Signals.h"
 #include "llvm/Support/TargetSelect.h"
 #include "llvm/Support/raw_ostream.h"
 #include <chrono>
 #include <cstdlib>
 #include <iostream>
 #include <memory>
 #include <mutex>
 #include <string>
 #include <thread>
 
+#ifndef _WIN32
+#include <unistd.h>
+#endif
+
 namespace clang {
 namespace clangd {
 namespace {
 
 using llvm::cl::cat;
 using llvm::cl::CommaSeparated;
 using llvm::cl::desc;
 using llvm::cl::Hidden;

 };
 
 int main(int argc, char *argv[]) {
   using namespace clang;
   using namespace clang::clangd;
 
   llvm::InitializeAllTargetInfos();
   llvm::sys::PrintStackTraceOnErrorSignal(argv[0]);
+  llvm::sys::SetInterruptFunction(&requestShutdown);
   llvm::cl::SetVersionPrinter([](llvm::raw_ostream &OS) {
     OS << clang::getClangToolFullVersion("clangd") << "\n";
   });
   const char *FlagsEnvVar = "CLANGD_FLAGS";
   const char *Overview =
       R"(clangd is a language server that provides IDE-like features to editors.
 
 It should be used via an editor plugin rather than invoked directly. For more information, see:
if (llvm::outs().is_displayed() && llvm::errs().is_displayed())
     llvm::errs() << Overview << "\n";
   // Use buffered stream to stderr (we still flush each log message). Unbuffered
   // stream can cause significant (non-deterministic) latency for the logger.
   llvm::errs().SetBuffered();
   StreamLogger Logger(llvm::errs(), LogLevel);
   LoggingSession LoggingSession(Logger);
   // Write some initial logs before we start doing any real work.
   log("{0}", clang::getClangToolFullVersion("clangd"));
+// FIXME: abstract this better, and print PID on windows too.
+#ifndef _WIN32
+  log("PID: {0}", getpid());
+#endif
   {
     SmallString<128> CWD;
     if (auto Err = llvm::sys::fs::current_path(CWD))
       log("Working directory unknown: {0}", Err.message());
     else
       log("Working directory: {0}", CWD);
   }
   for (int I = 0; I < argc; ++I)
#endif
 
   // There may still be lingering background threads (e.g. slow requests
   // whose results will be dropped, background index shutting down).
   //
   // These should terminate quickly, and ~ClangdLSPServer blocks on them.
   // However if a bug causes them to run forever, we want to ensure the process
   // eventually exits. As clangd isn't directly user-facing, an editor can
   // "leak" clangd processes. Crashing in this case contains the damage.
-  //
-  // This is more portable than sys::WatchDog, and yields a stack trace.
-  std::thread([] {
-    std::this_thread::sleep_for(std::chrono::minutes(5));
-    std::abort();
-  }).detach();
+  abortAfterTimeout(std::chrono::minutes(5));
 
   return ExitCode;
 }