Files
lix/lix/nix/daemon.cc
T
eldritch horrorsandpiegames d0a4b55a0e treewide: pour the rpc foundations
pour the foundations for rpc, and let them set. this is very much
unstable and must be opted into with explicit store uris (e.g. by
setting `NIX_REMOTE=daemon?protocol=any`). the daemon sockets are
not enabled by default and must be enabled with the `rpc-sockets`
experimental feature. we will not advertise this just yet because
in the current state it one has to be *very* dedicated to the rpc
cause to deploy this, but once we have some more bits migrated we
may want to add release notes and officially as for beta testing.

Co-Authored-By: piegames <git@piegames.de>
Change-Id: I85a96ccb700b91190c1eb37154bcc6ae1c03401a
2026-05-06 10:55:08 +00:00

770 lines
25 KiB
C++

///@file
#include "lix/libcmd/command.hh"
#include "lix/libmain/shared.hh"
#include "lix/libstore/local-store.hh"
#include "lix/libstore/remote-store.hh"
#include "lix/libstore/remote-store-connection.hh"
#include "lix/libstore/store-api.hh"
#include "lix/libutil/async-collect.hh"
#include "lix/libutil/async-io.hh"
#include "lix/libutil/async.hh"
#include "lix/libutil/config.hh"
#include "lix/libutil/current-process.hh"
#include "lix/libutil/error.hh"
#include "lix/libutil/experimental-features.hh"
#include "lix/libutil/file-descriptor.hh"
#include "lix/libutil/logging.hh"
#include "lix/libutil/processes.hh"
#include "lix/libutil/result.hh"
#include "lix/libutil/serialise.hh"
#include "lix/libutil/archive.hh"
#include "lix/libutil/c-calls.hh"
#include "lix/libstore/globals.hh"
#include "lix/libstore/derivations.hh"
#include "lix/libutil/finally.hh"
#include "lix/libcmd/legacy.hh"
#include "lix/libutil/signals.hh"
#include "lix/libstore/daemon.hh"
#include "lix/libutil/strings.hh"
#include "lix/libutil/unix-domain-socket.hh"
#include "daemon.hh"
#include <algorithm>
#include <cerrno>
#include <climits>
#include <cstdint>
#include <cstring>
#include <exception>
#include <kj/async.h>
#include <kj/exception.h>
#include <string>
#include <unistd.h>
#include <signal.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <sys/stat.h>
#include <sys/socket.h>
#include <sys/un.h>
#include <sys/select.h>
#include <errno.h>
#include <pwd.h>
#include <grp.h>
#include <fcntl.h>
#if __APPLE__ || __FreeBSD__
#include <sys/ucred.h>
#endif
#if __APPLE__
#include <membership.h>
#endif
static constexpr int SUBDAEMON_CONNECTION_FD = 0;
namespace nix {
using namespace nix::daemon;
/**
* Settings related to authenticating clients for the Nix daemon.
*
* For pipes we have little good information about the client side, but
* for Unix domain sockets we do. So currently these options implemented
* mandatory access control based on user names and group names (looked
* up and translated to UID/GIDs in the CLI process that runs the code
* in this file).
*
* No code outside of this file knows about these settings (this is not
* exposed in a header); all authentication and authorization happens in
* `daemon.cc`.
*/
struct AuthorizationSettings : Config {
#include "daemon-settings.gen.inc"
};
AuthorizationSettings authorizationSettings;
Config & daemonAuthorizationSettings = authorizationSettings;
static void sigChldHandler(int sigNo)
{
// Ensure we don't modify errno of whatever we've interrupted
auto saved_errno = errno;
// Reap all dead children.
while (waitpid(-1, 0, WNOHANG) > 0) ;
errno = saved_errno;
}
static void setSigChldAction(bool autoReap)
{
struct sigaction act, oact;
act.sa_handler = autoReap ? sigChldHandler : SIG_DFL;
sigfillset(&act.sa_mask);
act.sa_flags = 0;
if (sigaction(SIGCHLD, &act, &oact))
throw SysError("setting SIGCHLD handler");
}
/**
* @return Is the given user a member of this group?
*
* @param user User specified by username.
*
* @param group Group the user might be a member of.
*/
static bool matchUser(const std::string & user, const struct group & gr)
{
for (char * * mem = gr.gr_mem; *mem; mem++)
if (user == *mem) return true;
#if __APPLE__
// FIXME: we should probably pipe the uid through these functions
// instead of converting the username back into the uid
if (auto pw = getpwnam(user.c_str())) {
uuid_t uuid, gruuid;
if (!mbr_uid_to_uuid(pw->pw_uid, uuid) && !mbr_gid_to_uuid(gr.gr_gid, gruuid)) {
int ismember = 0;
if (!mbr_check_membership(uuid, gruuid, &ismember)) {
return !!ismember;
}
}
}
#endif
return false;
}
/**
* Does the given user (specified by user name, primary group name and supplementary group names)
* match the given user/group whitelist?
*
* If the list allows all users: Yes.
*
* If the username is in the set: Yes.
*
* If the user is in another group which is in the set: yes.
* If the groups intersects in the set: Yes.
*
* Otherwise: No.
*/
static bool
matchUser(const std::string & user, const std::unordered_set<std::string> & groups, const Strings & users)
{
if (find(users.begin(), users.end(), "*") != users.end())
return true;
if (find(users.begin(), users.end(), user) != users.end())
return true;
for (auto & i : users)
if (i.substr(0, 1) == "@") {
auto rest = i.substr(1);
if (groups.contains(rest)) {
return true;
}
struct group * gr = sys::getgrnam(rest);
if (!gr) continue;
if (matchUser(user, *gr)) return true;
}
return false;
}
struct PeerInfo
{
std::optional<pid_t> pid;
uid_t uid;
gid_t gid;
std::vector<gid_t> supplementaryGids;
};
/**
* Get the identity of the caller, if possible.
*/
static PeerInfo getPeerInfo(int remote)
{
std::vector<gid_t> supplementaryGids;
#if defined(SO_PEERCRED)
ucred cred;
socklen_t credLen = sizeof(cred);
if (getsockopt(remote, SOL_SOCKET, SO_PEERCRED, &cred, &credLen) == -1)
throw SysError("getting peer credentials");
#if defined(SO_PEERGROUPS)
// NOTE(Raito):
// Linux can go up to NGROUPS_MAX supplementary groups (65K).
// It is safe to assume that most users will have a couple of supplementary
// groups by default (here, my Linux user has ~7).
// We allocate 128 because integers are tiny.
supplementaryGids.resize(128);
// Initially attempt to retrieve 128 groups.
socklen_t nrSupplementaryGids = supplementaryGids.size();
while (true) {
if (getsockopt(remote, SOL_SOCKET, SO_PEERGROUPS, supplementaryGids.data(), &nrSupplementaryGids)
== -1
&& errno != ERANGE)
{
throw SysError("getting peer groups");
}
// If the number of groups returned is less than the requested size, we are done.
if (nrSupplementaryGids <= supplementaryGids.size()) {
// We ensure the vector matches exactly the number of groups to avoid
// letting the rest of the vector imply that the vector is full of `root` groups.
supplementaryGids.resize(nrSupplementaryGids);
break;
}
// Otherwise, the vector is too small. Resize and try again.
nrSupplementaryGids *= 2;
// We ensure the vector is big enough in response to our latest known allocation requirement.
supplementaryGids.resize(nrSupplementaryGids);
}
#endif
PeerInfo peer = {cred.pid, cred.uid, cred.gid, supplementaryGids};
#elif defined(LOCAL_PEERCRED)
#if !defined(SOL_LOCAL)
#define SOL_LOCAL 0
#endif
xucred cred;
socklen_t credLen = sizeof(cred);
if (getsockopt(remote, SOL_LOCAL, LOCAL_PEERCRED, &cred, &credLen) == -1)
throw SysError("getting peer credentials");
size_t nrGroups = cred.cr_ngroups;
std::copy(cred.cr_groups, cred.cr_groups + nrGroups, std::back_inserter(supplementaryGids));
PeerInfo peer = {std::nullopt, cred.cr_uid, cred.cr_gid, supplementaryGids};
#if defined(LOCAL_PEERPID)
socklen_t pidLen = sizeof(peer.pid);
pid_t peerPid;
if (!getsockopt(remote, SOL_LOCAL, LOCAL_PEERPID, &peerPid, &pidLen)) {
peer.pid = peerPid;
}
#endif
#else
#error \
"Your platform does not provide a mean (SO_PEERCRED or LOCAL_PEERCRED) to receive the user credentials when a connection to a socket is made. Please provide one."
#endif
return peer;
}
#define SD_LISTEN_FDS_START 3
/**
* Open a store without a path info cache.
*/
static kj::Promise<Result<ref<Store>>> openUncachedStore(AllowDaemon allowDaemon = AllowDaemon::Allow)
try {
StoreConfig::Params params; // FIXME: get params from somewhere
// Disable caching since the client already does that.
params["path-info-cache-size"] = "0";
co_return TRY_AWAIT(openStore(settings.storeUri, params, allowDaemon));
} catch (...) {
co_return result::current_exception();
}
/**
* Authenticate a potential client
*
* @param peer Information about other end of the connection, the client which
* wants to communicate with us.
*
* @return A pair of a `TrustedFlag`, whether the potential client is trusted,
* and the name of the user (useful for printing messages).
*
* If the potential client is not allowed to talk to us, we throw an `Error`.
*/
static std::pair<TrustedFlag, std::string> authPeer(const PeerInfo & peer)
{
TrustedFlag trusted = NotTrusted;
struct passwd * pw = getpwuid(peer.uid);
std::string user = pw ? pw->pw_name : std::to_string(peer.uid);
auto assertHealthyGroup = [&](const std::string & group) {
if (group == settings.buildUsersGroup) {
throw Error(
"the user '%1%' is not allowed to connect to the Nix daemon as its group is '%2%', "
"which is the group of users running the sandboxed builds.",
user,
group
);
}
};
std::unordered_set<std::string> groups;
auto insertGroup = [&](gid_t gid) {
auto gr = getgrgid(gid);
std::string group = gr ? gr->gr_name : std::to_string(gid);
assertHealthyGroup(group);
groups.insert(group);
};
// This ensures that the `groups` set is always non-empty with the primary group name.
insertGroup(peer.gid);
for (const gid_t suppGid : peer.supplementaryGids) {
insertGroup(suppGid);
}
const Strings & trustedUsers = authorizationSettings.trustedUsers;
const Strings & allowedUsers = authorizationSettings.allowedUsers;
if (matchUser(user, groups, trustedUsers)) {
trusted = Trusted;
}
if (!trusted && !matchUser(user, groups, allowedUsers)) {
throw Error("user '%1%' is not allowed to connect to the Nix daemon", user);
}
return { trusted, std::move(user) };
}
static kj::Promise<Result<void>> daemonLoopForSocket(
const Path & self,
const daemon::Protocol & socket,
AutoCloseFD & fdSocket,
std::optional<TrustedFlag> forceTrustClientOpt
)
try {
makeNonBlocking(fdSocket.get());
auto observer = kj::UnixEventPort::FdObserver{
AIO().unixEventPort, fdSocket.get(), kj::UnixEventPort::FdObserver::OBSERVE_READ
};
// Loop accepting connections.
while (1) {
try {
// Accept a connection.
struct sockaddr_un remoteAddr;
socklen_t remoteAddrLen = sizeof(remoteAddr);
AutoCloseFD remote{accept(
fdSocket.get(), reinterpret_cast<struct sockaddr *>(&remoteAddr), &remoteAddrLen
)};
if (!remote) {
if (errno == EAGAIN || errno == EWOULDBLOCK) {
co_await observer.whenBecomesReadable();
continue;
} else if (errno == EINTR) {
continue;
}
throw SysError("accepting connection");
}
// On macOS, accepted sockets inherit the non-blocking flag from the server socket, so
// explicitly make it blocking.
makeBlocking(remote.get());
closeOnExec(remote.get());
PeerInfo peer = getPeerInfo(remote.get());
printInfo("accepted connection from %1%", peer.pid ? fmt("pid %1%", *peer.pid) : "unknown peer");
// Fork a child to handle the connection. make sure it's called with
// argv0 `nix-daemon` so we don't try to run `nix --for` when called
// from more modern scripts that assume nix-command being available.
RunOptions options{
.program = self,
.argv0 = "nix-daemon",
.args =
{
"--for",
peer.pid ? fmt("%1%", *peer.pid) : "unknown",
"--log-level",
fmt("%1%", int(verbosity)),
"--protocol",
std::string(socket.id()),
},
.redirections = {{.dup = SUBDAEMON_CONNECTION_FD, .from = remote.get()}}
};
if (forceTrustClientOpt) {
options.args.push_back(
*forceTrustClientOpt ? "--force-trusted" : "--force-untrusted"
);
}
auto [pid, _stdout] = runProgram2(options).release();
pid.release();
} catch (Error & error) {
auto ei = error.info();
// FIXME: add to trace?
ei.msg = HintFmt("error processing connection: %1%", ei.msg.str());
logError(ei);
}
}
co_return result::success();
} catch (...) {
co_return result::current_exception();
}
/**
* Run a server. The loop opens a socket and accepts new connections from that
* socket.
*
* @param forceTrustClientOpt If present, force trusting or not trusted
* the client. Otherwise, decide based on the authentication settings
* and user credentials (from the unix domain socket).
*/
static kj::Promise<Result<void>> daemonLoop(std::optional<TrustedFlag> forceTrustClientOpt)
try {
if (chdir("/") == -1) {
throw SysError("cannot change current directory");
}
const auto self = [] {
auto tmp = getSelfExe();
if (!tmp) {
throw Error("can't locate the daemon binary!");
}
return *tmp;
}();
std::list<std::pair<daemon::Protocol, AutoCloseFD>> sockets;
for (auto & socket : settings.nixDaemonSockets()) {
switch (socket.type) {
case daemon::Protocol::LEGACY_COMBINED:
case daemon::Protocol::LEGACY:
break;
case daemon::Protocol::RPC_V1:
if (!experimentalFeatureSettings.isEnabled(Xp::RpcSockets)) {
continue;
}
break;
}
createDirs(dirOf(socket.path));
sockets.emplace_back(socket, createUnixDomainSocket(socket.path, 0666));
}
// Get rid of children automatically; don't let them become zombies.
setSigChldAction(true);
TRY_AWAIT(asyncSpread(sockets, [&](auto & socket) {
return daemonLoopForSocket(self, socket.first, socket.second, forceTrustClientOpt);
}));
co_return result::success();
} catch (...) {
co_return result::current_exception();
}
static void daemonInstance(
daemon::Protocol protocol,
AsyncIoRoot & aio,
std::optional<TrustedFlag> forceTrustClientOpt,
char * peerPidArg
)
{
// Handle socket-based activation by systemd.
const auto [launchedByManager, connectionFd] = [&]() -> std::pair<bool, int> {
auto listenFds = getEnv("LISTEN_FDS");
if (listenFds) {
if (getEnv("LISTEN_PID") != std::to_string(getpid()) || listenFds != "1") {
throw Error("unexpected systemd environment variables");
}
if (auto fdnames = getEnv("LISTEN_FDNAMES")) {
protocol = daemon::getProtocol(*fdnames);
}
closeOnExec(SD_LISTEN_FDS_START);
// these unsets are not critical, we never did this for accept=no sockets either
(void) sys::unsetenv("LISTEN_FDS");
(void) sys::unsetenv("LISTEN_PID");
(void) sys::unsetenv("LISTEN_FDNAMES");
return {true, SD_LISTEN_FDS_START};
} else {
return {false, SUBDAEMON_CONNECTION_FD};
}
}();
PeerInfo peer = getPeerInfo(connectionFd);
TrustedFlag trusted;
std::string user;
// replace peerPidArg contents with the peer pid if possible. the forking daemon does
// this as a debugging aid and it is easy enough to do it here also, so we just do it
if (peerPidArg && peer.pid) {
auto pidForArgv = std::to_string(*peer.pid);
if (pidForArgv.size() < strlen(peerPidArg)) {
memset(peerPidArg, ' ', strlen(peerPidArg));
peerPidArg[0] = '\0';
memcpy(peerPidArg + 1, pidForArgv.c_str(), pidForArgv.size());
}
}
if (forceTrustClientOpt) {
trusted = *forceTrustClientOpt;
} else {
std::tie(trusted, user) = authPeer(peer);
}
printInfo(
"%1% is %2% (%3%%4%)",
Uncolored(peer.pid ? fmt("remote pid %s", *peer.pid) : "remote with unknown pid"),
peer.uid ? fmt("user %s", user) : "unknown user",
trusted ? "trusted" : "untrusted",
forceTrustClientOpt ? " by override" : ""
);
// Background the daemon.
if (!launchedByManager && setsid() == -1) {
throw SysError("creating a new session");
}
auto store = aio.blockOn(openUncachedStore(AllowDaemon::Disallow));
if (auto local = dynamic_cast<LocalStore *>(&*store); local) {
local->associateWithCredentials(peer.uid, peer.gid);
}
// Handle the connection.
switch (protocol.type) {
case daemon::Protocol::LEGACY_COMBINED:
case daemon::Protocol::LEGACY: {
FdSource from(connectionFd);
FdSink to(connectionFd);
processLegacyConnection(aio, store, from, to, trusted);
break;
}
case daemon::Protocol::RPC_V1: {
auto stream = AIO().lowLevelProvider.wrapSocketFd(connectionFd);
aio.blockOn(processConnection(store, *stream, trusted));
break;
}
}
}
/**
* Forward a standard IO connection to the given remote store.
*
* We just act as a middleman blindly ferry output between the standard
* input/output and the remote store connection, not processing anything.
*
* Loops until standard input disconnects, or an error is encountered.
*/
static void forwardStdioConnection(AsyncIoRoot & aio, RemoteStore & store)
{
auto conn = aio.blockOn(store.openConnectionForDaemonForwarding());
auto connSocket = AIO().lowLevelProvider.wrapSocketFd(conn->getFD());
auto asyncStdin = AIO().lowLevelProvider.wrapInputFd(STDIN_FILENO);
auto asyncStdout = AIO().lowLevelProvider.wrapOutputFd(STDOUT_FILENO);
try {
connSocket->pumpTo(*asyncStdout)
.exclusiveJoin(asyncStdin->pumpTo(*connSocket))
.wait(aio.kj.waitScope);
} catch (kj::Exception & e) { // NOLINT(lix-foreign-exceptions): kj may throw immediate from pumpTo
if (e.getType() == kj::Exception::Type::DISCONNECTED) {
throw Error("peer disconnected unexpectedly");
} else {
throw Error("forwarding daemon connection: %s", e.getDescription().cStr());
}
}
}
/**
* Process a client connecting to us via standard input/output
*
* Unlike `forwardStdioConnection()` we do process commands ourselves in
* this case, not delegating to another daemon.
*
* @param trustClient Whether to trust the client. Forwarded directly to
* `processConnection()`.
*/
static void
processStdioConnection(AsyncIoRoot & aio, ref<Store> store, TrustedFlag trustClient)
{
FdSource from(STDIN_FILENO);
FdSink to(STDOUT_FILENO);
processLegacyConnection(aio, store, from, to, trustClient);
}
/**
* Entry point shared between the new CLI `nix daemon` and old CLI
* `nix-daemon`.
*
* @param forceTrustClientOpt See `daemonLoop()` and the parameter with
* the same name over there for details.
*/
static void
runDaemon(AsyncIoRoot & aio, bool stdio, std::optional<TrustedFlag> forceTrustClientOpt)
{
if (stdio) {
auto store = aio.blockOn(openUncachedStore());
// If --force-untrusted is passed, we cannot forward the connection and
// must process it ourselves (before delegating to the next store) to
// force untrusting the client.
if (auto remoteStore = store.try_cast_shared<RemoteStore>(); remoteStore && (!forceTrustClientOpt || *forceTrustClientOpt != NotTrusted))
forwardStdioConnection(aio, *remoteStore);
else
// `Trusted` is passed in the auto (no override case) because we
// cannot see who is on the other side of a plain pipe. Limiting
// access to those is explicitly not `nix-daemon`'s responsibility.
processStdioConnection(aio, store, forceTrustClientOpt.value_or(Trusted));
} else {
try {
aio.blockOn(makeInterruptible(daemonLoop(forceTrustClientOpt)));
} catch (Interrupted &) {
}
}
}
static int
main_nix_daemon(AsyncIoRoot & aio, std::string programName, Strings argv, std::span<char *> rawArgv)
{
{
auto stdio = false;
std::optional<TrustedFlag> isTrustedOpt = std::nullopt;
bool isInstance = false;
char * peerPidArg = nullptr;
Verbosity subdaemonLogLevel = lvlInfo;
std::string protocol = "legacy-combined";
LegacyArgs(aio, programName, [&](Strings::iterator & arg, const Strings::iterator & end) {
if (*arg == "--daemon")
; // ignored for backwards compatibility
else if (*arg == "--help")
showManPage("nix-daemon");
else if (*arg == "--version")
printVersion("nix-daemon");
else if (*arg == "--stdio")
stdio = true;
else if (*arg == "--force-trusted") {
experimentalFeatureSettings.require(Xp::DaemonTrustOverride);
isTrustedOpt = Trusted;
} else if (*arg == "--force-untrusted") {
experimentalFeatureSettings.require(Xp::DaemonTrustOverride);
isTrustedOpt = NotTrusted;
} else if (*arg == "--default-trust") {
experimentalFeatureSettings.require(Xp::DaemonTrustOverride);
isTrustedOpt = std::nullopt;
} else if (*arg == "--for") {
isInstance = true;
getArg(*arg, arg, end);
} else if (*arg == "--protocol") {
protocol = getArg(*arg, arg, end);
} else if (*arg == "--for-socket-activation") {
isInstance = true;
// HACK: too many copies and rewrites happen by the time we get here to
// be able to calculate a rawArgv offset. instead we will search for an
// exact match and blindly assume that it's the one we want to rewrite.
for (auto [i, rawArg] : enumerate(rawArgv)) {
if (rawArg == *arg) {
peerPidArg = rawArg + strlen("--for");
break;
}
}
} else if (*arg == "--log-level") {
if (auto level = string2Int<int>(getArg(*arg, arg, end)); level) {
subdaemonLogLevel = static_cast<Verbosity>(std::min<int>(lvlVomit, *level));
} else {
throw UsageError("--log-level expects an integer in the range [0..7]");
}
} else {
return false;
}
return true;
}).parseCmdline(argv);
if (isInstance) {
verbosity = Verbosity(std::min<uint64_t>(subdaemonLogLevel, lvlVomit));
daemonInstance(daemon::getProtocol(protocol), aio, isTrustedOpt, peerPidArg);
} else {
runDaemon(aio, stdio, isTrustedOpt);
}
return 0;
}
}
void registerLegacyNixDaemon() {
LegacyCommandRegistry::addWithRaw("nix-daemon", main_nix_daemon);
}
struct CmdDaemon : StoreCommand
{
bool stdio = false;
std::optional<TrustedFlag> isTrustedOpt = std::nullopt;
CmdDaemon()
{
addFlag({
.longName = "stdio",
.description = "Attach to standard I/O, instead of trying to bind to a UNIX socket.",
.handler = {&stdio, true},
});
addFlag({
.longName = "force-trusted",
.description = "Force the daemon to trust connecting clients.",
.handler = {[&]() {
isTrustedOpt = Trusted;
}},
.experimentalFeature = Xp::DaemonTrustOverride,
});
addFlag({
.longName = "force-untrusted",
.description = "Force the daemon to not trust connecting clients. The connection will be processed by the receiving daemon before forwarding commands.",
.handler = {[&]() {
isTrustedOpt = NotTrusted;
}},
.experimentalFeature = Xp::DaemonTrustOverride,
});
addFlag({
.longName = "default-trust",
.description = "Use Nix's default trust.",
.handler = {[&]() {
isTrustedOpt = std::nullopt;
}},
.experimentalFeature = Xp::DaemonTrustOverride,
});
}
std::string description() override
{
return "daemon to perform store operations on behalf of non-root clients";
}
Category category() override { return catUtility; }
std::string doc() override
{
return
#include "daemon.md"
;
}
void run(ref<Store> store) override
{
runDaemon(aio(), stdio, isTrustedOpt);
}
};
void registerNixDaemon()
{
registerCommand2<CmdDaemon>({"daemon"});
}
}