Complete shared repair journal ownership (c36e3cac) · Commits · hpc / gekkofs

README.md

+5 −0

Changes for README.md: 5 added lines, 0 removed lines.
Original line number	Diff line number	Diff line
		@@ -1046,6 +1046,11 @@ Client-metrics require the CMake argument `-DGKFS_ENABLE_CLIENT_METRICS=ON` (see
		environment.
		- `LIBGKFS_PROXY_PID_FILE` - Path to the proxy pid file (when using the GekkoFS proxy).
		- `LIBGKFS_NUM_REPL` - Number of replicas for data.
		- `LIBGKFS_REPAIR_STATUS_PATH` - Optional path for client-local repair counters and queue status.
		- `LIBGKFS_REPAIR_JOURNAL_PATH` - Optional path for client-local pending repair tasks. The journal is
		single-owner: do not share it between client processes. Its `.lock` sidecar records the owning host,
		PID, acquisition time, and a fingerprint of the configured hosts file; reuse it only with the same
		topology. A changed topology requires a new journal path or removal of the completed/stale sidecar.
		#### Optimization
		- `LIBGKFS_CREATE_WRITE_OPTIMIZATION` - Optimization for write operations (default: OFF).
		- `LIBGKFS_READ_INLINE_PREFETCH` - Prefetch inline data when opening files (default: OFF).

docs/sphinx/users/running.md

+6 −0

Changes for docs/sphinx/users/running.md: 6 added lines, 0 removed lines.
Original line number	Diff line number	Diff line
		@@ -367,3 +367,9 @@ The user can enable the data replication feature by setting the replication envi
		`LIBGKFS_NUM_REPL=<num repl>`.
		The number of replicas should go from `0` to the `number of servers - 1`. The replication environment variable can be
		set up for each client independently.

		For client-local repair recovery, set `LIBGKFS_REPAIR_JOURNAL_PATH` to a writable
		journal path. One client process owns a journal at a time. The adjacent `.lock`
		sidecar records the owner and a fingerprint of the configured hosts file; do not
		share a journal across client processes or reuse it after changing the topology.
		Use `LIBGKFS_REPAIR_STATUS_PATH` to publish a client-local repair queue snapshot.
		No newline at end of file

include/client/rpc/repair_journal_lock.hpp

+29 −0

Changes for include/client/rpc/repair_journal_lock.hpp: 29 added lines, 0 removed lines.
Original line number	Diff line number	Diff line
		@@ -21,6 +21,7 @@
		#include <cstring>
		#include <fstream>
		#include <iomanip>
		#include <optional>
		#include <sstream>
		#include <stdexcept>
		#include <string>
		@@ -57,6 +58,23 @@ repair_journal_topology_identity(const std::string& hostfile_path) {
		return result.str();
		}

		inline std::optional<std::string>
		read_repair_journal_topology_identity(const std::string& lock_path) {
		std::ifstream input(lock_path);
		if(!input.is_open()) {
		return std::nullopt;
		}

		std::string line;
		while(std::getline(input, line)) {
		constexpr auto prefix = "topology=";
		if(line.rfind(prefix, 0) == 0) {
		return line.substr(std::char_traits<char>::length(prefix));
		}
		}
		return std::string{};
		}

		/**
		* @brief Owns the advisory process lock for a client-local repair journal.
		*
		@@ -108,6 +126,17 @@ public:
		journal_path + "': " + ::strerror(error));
		}

		if(const auto previous_topology =
		read_repair_journal_topology_identity(lock_path_);
		previous_topology && !previous_topology->empty() &&
		topology_identity != "unknown" &&
		*previous_topology != topology_identity) {
		throw std::runtime_error(
		"Replica repair journal topology mismatch: expected '" +
		topology_identity + "', found '" + *previous_topology +
		"'");
		}

		repair_journal_owner owner;
		owner.hostname.resize(255);
		if(::gethostname(owner.hostname.data(), owner.hostname.size()) != 0) {

tests/unit/test_repair_journal_lock.cpp

+61 −0

Changes for tests/unit/test_repair_journal_lock.cpp: 61 added lines, 0 removed lines.
Original line number	Diff line number	Diff line
		@@ -14,6 +14,7 @@
		#include <filesystem>
		#include <fstream>
		#include <sstream>
		#include <sys/wait.h>
		#include <unistd.h>

		namespace {
		@@ -78,3 +79,63 @@ TEST_CASE("empty repair journal path does not create a lock",
		REQUIRE_FALSE(lock.owns_lock());
		REQUIRE(lock.owner().hostname.empty());
		}

		TEST_CASE("repair journal lock rejects a different topology",
		"[replication][repair][journal]") {
		const auto journal = test_journal_path();
		const auto lock_path = journal.string() + ".lock";
		std::filesystem::remove(lock_path);

		gkfs::rpc::repair_journal_lock first;
		first.acquire(journal.string(), "topology-a");
		first.release();

		gkfs::rpc::repair_journal_lock second;
		REQUIRE_THROWS_WITH(
		second.acquire(journal.string(), "topology-b"),
		"Replica repair journal topology mismatch: expected 'topology-b', found 'topology-a'");
		std::filesystem::remove(lock_path);
		}

		TEST_CASE("repair journal lock ownership is exclusive across processes",
		"[replication][repair][journal]") {
		const auto journal = test_journal_path();
		const auto lock_path = journal.string() + ".lock";
		std::filesystem::remove(lock_path);

		int ready[2]{};
		REQUIRE(::pipe(ready) == 0);
		const auto child = ::fork();
		REQUIRE(child >= 0);
		if(child == 0) {
		::close(ready[0]);
		gkfs::rpc::repair_journal_lock child_lock;
		try {
		child_lock.acquire(journal.string(), "topology-child");
		const char signal = '1';
		::write(ready[1], &signal, sizeof(signal));
		::sleep(1);
		_exit(0);
		} catch(...) {
		_exit(1);
		}
		}

		::close(ready[1]);
		char signal{};
		REQUIRE(::read(ready[0], &signal, sizeof(signal)) == sizeof(signal));
		::close(ready[0]);

		gkfs::rpc::repair_journal_lock parent_lock;
		REQUIRE_THROWS_WITH(
		parent_lock.acquire(journal.string(), "topology-child"),
		"Replica repair journal is already owned: " + journal.string());

		int status = 0;
		REQUIRE(::waitpid(child, &status, 0) == child);
		REQUIRE(WIFEXITED(status));
		REQUIRE(WEXITSTATUS(status) == 0);
		REQUIRE_NOTHROW(parent_lock.acquire(journal.string(), "topology-child"));
		parent_lock.release();
		std::filesystem::remove(lock_path);
		}