Lock Config and Preset Files Across Instances and Write Them Atomically (#15861)

* Lock Config and Preset Files Across Instances and Write Them Atomically

Every running instance shares one OrcaSlicer.conf and one user preset
tree, and nothing kept their writers apart. Two instances saving at the
same moment, or the cloud preset sync thread writing while the GUI thread
saved, could interleave, and a reader in another instance could open a
preset JSON or .info file between truncate and close and get a partial
file, dropping that preset for the session with a parse error.

Add InstanceLock, a scoped guard that serialises the threads of one
process through a recursive mutex and other processes through an advisory
OS file lock: flock on POSIX, held on the guard's own descriptor so no
other close in the process can drop it, and LockFileEx on Windows. The
outermost guard opens the lock file and closes it on release, so nothing
stays open between saves and a data dir can be removed once nothing is
saving into it; the file itself is kept, since deleting it would let a
third instance lock a fresh file while the second still holds the old
one. It is best effort: when the lock file cannot be opened or locked, or
another instance still holds it after a second, the guard logs once and
lets the write proceed, then leaves the file alone for ten seconds, so a
hung instance never blocks every other one and a holder stuck in a
debugger does not cost a stall per save. The guard sits at the leaf
readers and writers: set_sync_info_and_save() calls save_info() under the
preset collection mutex, so a batch lock around save_user_presets() would
invert the order against the sync thread. The user preset scan reads its
files on worker threads without the guard, since the mutex would
serialise them, and takes it per file in the serial commit step, so a
save never waits for the whole scan. Each read keeps the bytes of the
preset and its .info as they were before parsing; commit compares them
with the disk under the guard and reads a file that changed again, so it
never deletes or writes back over another instance's newer save, nor
installs a .json and .info from two different saves; a preset another
instance removed in the meantime is not installed. Without the guard, in
a cool-down, the scan still loads the presets but leaves their files
alone: an unreadable file stays for the next scan, and a derived
compatible printer is not written back. Read-only scans, which is what
the CLI does, take no lock and create no lock file.

AppConfig holds OrcaSlicer.conf.lock in load() and save(); load is
included because the Windows path restores from the .bak copy. Every
user preset writer and reader holds user.lock: Preset::save(), which
writes no .info when the preset itself could not be written, since an
.info without its preset reads as a cloud deletion request, save_info(),
reload() and remove_files(), each preset the scan commits, the
bundle metadata reads and write, the .info removal after a cloud-confirmed
delete, the orphaned-.info scan on the sync thread, the bundle folder
removal on unsubscribe and the physical printer writers and delete
paths. A bundle import extracts under cache/ into a folder per process
and per import, where no scan reads.

Preset JSON, .info, bundle metadata, physical printer and config files,
and the caches and state files that already used a temporary by hand,
now go through write_file_atomically(), which writes <file>.<pid>.<n>.tmp
beside the target and renames it over, so a reader that never waits sees
a complete old or new file. A symlink is followed; a target that is not
a regular file is written in place; and when no temporary can be created
beside an existing target, or the rename itself is refused, by a Windows
reader holding the file open or a mount that cannot replace in one step,
the helper writes in place as before, since losing the save is worse
than a torn read. On POSIX the rename replaces the
target atomically where the old code removed it first and left a window
with no file at all; only a mount that refuses a one-step replace gets
the old remove-then-rename. A crash between temporary and rename leaves
the temporary behind, which no scan reads. Preset::save() returns
whether it wrote the preset, so the scan counts a compatible printer it
could not write back as an error. A failed config write keeps
the config dirty, and the idle handler waits ten seconds before retrying
while an explicit save always tries.

* Run the Cross-Process Lock Test on Every Platform

The test that checks the guard yields to a lock held elsewhere forked a
child to hold it, so it was left out on Windows. The OS lock belongs to
the handle on Windows and to the open file description elsewhere, so a
second handle in the same process is refused like another instance
would be. The test now holds the lock that way and runs everywhere.

src/libslic3r/utils.cpp

@@ -9,6 +9,8 @@
 #include <stdio.h>
 #include <filesystem>
 #include <sstream>
+#include <cerrno>
+#include <mutex>
 #include <iomanip>
 #include <algorithm>
 #include <cmath>
@@ -703,13 +705,99 @@ namespace WindowsSupport
 std::error_code rename_file(const std::string &from, const std::string &to)
 {
 #ifdef _WIN32
+	// Retries and moves an open destination aside itself.
 	return WindowsSupport::rename(from, to);
 #else
-	boost::nowide::remove(to.c_str());
-	return std::make_error_code(static_cast<std::errc>(boost::nowide::rename(from.c_str(), to.c_str())));
+	// rename(2) replaces an existing target atomically; removing it first would
+	// leave a window in which the file does not exist at all.
+	if (boost::nowide::rename(from.c_str(), to.c_str()) == 0)
+		return {};
+	const int err = errno;
+	// Some mounts (sshfs, gvfs, MTP and a few SMB setups) refuse to replace an
+	// existing target in one step, each with the error it sees fit; every error
+	// is worth the remove-then-rename this always did, except the ones no retry
+	// can help: nothing at the source, a different device, or a directory where
+	// a file was expected and the reverse.
+	const bool worth_retrying = err != ENOENT && err != EXDEV && err != ENOTDIR && err != EISDIR;
+	if (worth_retrying && boost::nowide::remove(to.c_str()) == 0 && boost::nowide::rename(from.c_str(), to.c_str()) == 0)
+		return {};
+	return std::make_error_code(static_cast<std::errc>(err));
 #endif
 }
 
+static std::error_code write_whole_file(const std::string &path, std::initializer_list<std::string_view> chunks, bool binary)
+{
+	errno = 0;
+	FILE *file = boost::nowide::fopen(path.c_str(), binary ? "wb" : "w");
+	if (file == nullptr)
+		return std::make_error_code(errno != 0 ? static_cast<std::errc>(errno) : std::errc::io_error);
+	bool ok = true;
+	for (const std::string_view chunk : chunks)
+		ok = ok && std::fwrite(chunk.data(), 1, chunk.size(), file) == chunk.size();
+	ok = ok && std::fflush(file) == 0;
+	const int err = ok ? 0 : errno;
+	ok = std::fclose(file) == 0 && ok;
+	if (ok)
+		return {};
+	return std::make_error_code(err != 0 ? static_cast<std::errc>(err) : std::errc::io_error);
+}
+
+// The in-place fallback truncates the target, so two threads of this process
+// on the same file must not both be in it. One mutex for all such writes: they
+// are the rare case. Never freed, like the InstanceLock registry, so a save
+// during static destruction still finds it.
+static std::error_code write_in_place(const std::string &path, std::initializer_list<std::string_view> chunks, bool binary)
+{
+	static auto *mutex = new std::mutex();
+	std::lock_guard<std::mutex> guard(*mutex);
+	return write_whole_file(path, chunks, binary);
+}
+
+std::error_code write_file_atomically(const std::string &path, std::initializer_list<std::string_view> chunks, bool binary)
+{
+	boost::system::error_code bec;
+	const boost::filesystem::file_status target = boost::filesystem::symlink_status(path, bec);
+	const bool target_exists = ! bec && boost::filesystem::exists(target);
+	if (target_exists && boost::filesystem::is_symlink(target)) {
+		// A config or preset kept in a dotfiles repository: the link stays,
+		// the file it points to is replaced like any other.
+		const boost::filesystem::path resolved = boost::filesystem::canonical(path, bec);
+		if (! bec && boost::filesystem::is_regular_file(resolved, bec))
+			return write_file_atomically(resolved.string(), chunks, binary);
+	}
+	if (target_exists && ! boost::filesystem::is_regular_file(target))
+		return write_in_place(path, chunks, binary);
+
+	// Unique per process and per call, so two threads writing one target
+	// without a lock never share a temporary.
+	static std::atomic<unsigned> counter{0};
+	const std::string tmp_path = path + "." + std::to_string(get_current_pid()) + "." + std::to_string(counter++) + ".tmp";
+	if (const std::error_code ec = write_whole_file(tmp_path, chunks, binary)) {
+		boost::nowide::remove(tmp_path.c_str());
+		if (! target_exists)
+			return ec;
+		// A directory that lets this process write its files but not create
+		// one: losing the save is worse than a reader seeing a partial file.
+		BOOST_LOG_TRIVIAL(warning) << "Cannot create a temporary beside " << path << " (" << ec.message() << "); writing in place";
+		return write_in_place(path, chunks, binary);
+	}
+#ifndef _WIN32
+	// Not on Windows, where a read-only bit on the temporary would stop the rename itself.
+	if (target_exists)
+		boost::filesystem::permissions(tmp_path, target.permissions(), bec);
+#endif
+	if (const std::error_code ec = rename_file(tmp_path, path)) {
+		boost::nowide::remove(tmp_path.c_str());
+		// A reader on Windows holding the target open without FILE_SHARE_DELETE,
+		// or a mount that cannot replace a file at all. Losing the save is worse
+		// than a reader seeing a partial file, so write in place the way this
+		// used to work before the atomic path existed.
+		BOOST_LOG_TRIVIAL(warning) << "Cannot replace " << path << " (" << ec.message() << "); writing in place";
+		return write_in_place(path, chunks, binary);
+	}
+	return {};
+}
+
 #ifdef __linux__
 // Copied from boost::filesystem.
 // Called by copy_file_linux() in case linux sendfile() API is not supported.