Merge pull request #7802 from edolstra/fix-7783

Fix PID namespace support check

src/libstore/build/local-derivation-goal.cc

@@ -209,7 +209,7 @@ void LocalDerivationGoal::tryLocalBuild()
 
     #if __linux__
     if (useChroot) {
-        if (!mountNamespacesSupported() || !pidNamespacesSupported()) {
+        if (!mountAndPidNamespacesSupported()) {
             if (!settings.sandboxFallback)
                 throw Error("this system does not support the kernel namespaces that are required for sandboxing; use '--no-sandbox' to disable sandboxing");
             debug("auto-disabling sandboxing because the prerequisite namespaces are not available");
@@ -385,12 +385,6 @@ void LocalDerivationGoal::cleanupPostOutputsRegisteredModeNonCheck()
 }
 
 
-int childEntry(void * arg)
-{
-    ((LocalDerivationGoal *) arg)->runChild();
-    return 1;
-}
-
 #if __linux__
 static void linkOrCopy(const Path & from, const Path & to)
 {
@@ -916,21 +910,15 @@ void LocalDerivationGoal::startBuilder()
             if (getuid() == 0 && setgroups(0, 0) == -1)
                 throw SysError("setgroups failed");
 
-            size_t stackSize = 1 * 1024 * 1024;
+            ProcessOptions options;
-            char * stack = (char *) mmap(0, stackSize,
+            options.cloneFlags = CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWIPC | CLONE_NEWUTS | CLONE_PARENT | SIGCHLD;
-                PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
-            if (stack == MAP_FAILED) throw SysError("allocating stack");
-
-            int flags = CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWIPC | CLONE_NEWUTS | CLONE_PARENT | SIGCHLD;
             if (privateNetwork)
-                flags |= CLONE_NEWNET;
+                options.cloneFlags |= CLONE_NEWNET;
             if (usingUserNamespace)
-                flags |= CLONE_NEWUSER;
+                options.cloneFlags |= CLONE_NEWUSER;
 
-            pid_t child = clone(childEntry, stack + stackSize, flags, this);
+            pid_t child = startProcess([&]() { runChild(); }, options);
 
-            if (child == -1)
-                throw SysError("creating sandboxed builder process using clone()");
             writeFull(builderOut.writeSide.get(),
                 fmt("%d %d\n", usingUserNamespace, child));
             _exit(0);

src/libutil/namespaces.cc

@@ -4,7 +4,7 @@
 #include "util.hh"
 #include "finally.hh"
 
-#include <mntent.h>
+#include <sys/mount.h>
 
 namespace nix {
 
@@ -33,60 +33,57 @@ bool userNamespacesSupported()
             return false;
         }
 
+        try {
             Pid pid = startProcess([&]()
             {
-            auto res = unshare(CLONE_NEWUSER);
+                _exit(0);
-            _exit(res ? 1 : 0);
+            }, {
+                .cloneFlags = CLONE_NEWUSER
             });
 
-        bool supported = pid.wait() == 0;
+            auto r = pid.wait();
+            assert(!r);
+        } catch (SysError & e) {
+            debug("user namespaces do not work on this system: %s", e.msg());
+            return false;
+        }
 
-        if (!supported)
+        return true;
-            debug("user namespaces do not work on this system");
-
-        return supported;
     }();
     return res;
 }
 
-bool mountNamespacesSupported()
+bool mountAndPidNamespacesSupported()
 {
     static auto res = [&]() -> bool
     {
-        bool useUserNamespace = userNamespacesSupported();
+        try {
 
             Pid pid = startProcess([&]()
             {
-            auto res = unshare(CLONE_NEWNS | (useUserNamespace ? CLONE_NEWUSER : 0));
+                /* Make sure we don't remount the parent's /proc. */
-            _exit(res ? 1 : 0);
+                if (mount(0, "/", 0, MS_PRIVATE | MS_REC, 0) == -1)
+                    _exit(1);
+
+                /* Test whether we can remount /proc. The kernel disallows
+                   this if /proc is not fully visible, i.e. if there are
+                   filesystems mounted on top of files inside /proc.  See
+                   https://lore.kernel.org/lkml/87tvsrjai0.fsf@xmission.com/T/. */
+                if (mount("none", "/proc", "proc", 0, 0) == -1)
+                    _exit(2);
+
+                _exit(0);
+            }, {
+                .cloneFlags = CLONE_NEWNS | CLONE_NEWPID | (userNamespacesSupported() ? CLONE_NEWUSER : 0)
             });
 
-        bool supported = pid.wait() == 0;
+            if (pid.wait()) {
-
+                debug("PID namespaces do not work on this system: cannot remount /proc");
-        if (!supported)
+                return false;
-            debug("mount namespaces do not work on this system");
-
-        return supported;
-    }();
-    return res;
             }
 
-bool pidNamespacesSupported()
+        } catch (SysError & e) {
-{
+            debug("mount namespaces do not work on this system: %s", e.msg());
-    static auto res = [&]() -> bool
-    {
-        /* Check whether /proc is fully visible, i.e. there are no
-           filesystems mounted on top of files inside /proc. If this
-           is not the case, then we cannot mount a new /proc inside
-           the sandbox that matches the sandbox's PID namespace.
-           See https://lore.kernel.org/lkml/87tvsrjai0.fsf@xmission.com/T/. */
-        auto fp = fopen("/proc/mounts", "r");
-        if (!fp) return false;
-        Finally delFP = [&]() { fclose(fp); };
-
-        while (auto ent = getmntent(fp))
-            if (hasPrefix(std::string_view(ent->mnt_dir), "/proc/")) {
-                debug("PID namespaces do not work because /proc is not fully visible; disabling sandboxing");
             return false;
         }
 

src/libutil/namespaces.hh

@@ -6,9 +6,7 @@ namespace nix {
 
 bool userNamespacesSupported();
 
-bool mountNamespacesSupported();
+bool mountAndPidNamespacesSupported();
-
-bool pidNamespacesSupported();
 
 #endif
 

src/libutil/util.cc

@@ -36,6 +36,7 @@
 #ifdef __linux__
 #include <sys/prctl.h>
 #include <sys/resource.h>
+#include <sys/mman.h>
 
 #include <cmath>
 #endif
@@ -1064,9 +1065,17 @@ static pid_t doFork(bool allowVfork, std::function<void()> fun)
 }
 
 
+static int childEntry(void * arg)
+{
+    auto main = (std::function<void()> *) arg;
+    (*main)();
+    return 1;
+}
+
+
 pid_t startProcess(std::function<void()> fun, const ProcessOptions & options)
 {
-    auto wrapper = [&]() {
+    std::function<void()> wrapper = [&]() {
         if (!options.allowVfork)
             logger = makeSimpleLogger();
         try {
@@ -1086,7 +1095,27 @@ pid_t startProcess(std::function<void()> fun, const ProcessOptions & options)
             _exit(1);
     };
 
-    pid_t pid = doFork(options.allowVfork, wrapper);
+    pid_t pid = -1;
+
+    if (options.cloneFlags) {
+        #ifdef __linux__
+        // Not supported, since then we don't know when to free the stack.
+        assert(!(options.cloneFlags & CLONE_VM));
+
+        size_t stackSize = 1 * 1024 * 1024;
+        auto stack = (char *) mmap(0, stackSize,
+            PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
+        if (stack == MAP_FAILED) throw SysError("allocating stack");
+
+        Finally freeStack([&]() { munmap(stack, stackSize); });
+
+        pid = clone(childEntry, stack + stackSize, options.cloneFlags | SIGCHLD, &wrapper);
+        #else
+        throw Error("clone flags are only supported on Linux");
+        #endif
+    } else
+        pid = doFork(options.allowVfork, wrapper);
+
     if (pid == -1) throw SysError("unable to fork");
 
     return pid;

src/libutil/util.hh

@@ -307,6 +307,7 @@ struct ProcessOptions
     bool dieWithParent = true;
     bool runExitHandlers = false;
     bool allowVfork = false;
+    int cloneFlags = 0; // use clone() with the specified flags (Linux only)
 };
 
 pid_t startProcess(std::function<void()> fun, const ProcessOptions & options = ProcessOptions());

tests/nixos/remote-builds.nix

@@ -11,6 +11,11 @@ let
     { services.openssh.enable = true;
       virtualisation.writableStore = true;
       nix.settings.sandbox = true;
+
+      # Regression test for use of PID namespaces when /proc has
+      # filesystems mounted on top of it
+      # (i.e. /proc/sys/fs/binfmt_misc).
+      boot.binfmt.emulatedSystems = [ "aarch64-linux" ];
     };
 
   # Trivial Nix expression to build remotely.