When using chroots, use a private PID namespace

In a private PID namespace, processes have PIDs that are separate from
the rest of the system.  The initial child gets PID 1.  Processes in
the chroot cannot see processes outside of the chroot.  This improves
isolation between builds.  However, processes on the outside can see
processes in the chroot and send signals to them (if they have
appropriate rights).

Since the builder gets PID 1, it serves as the reaper for zombies in
the chroot.  This might turn out to be a problem.  In that case we'll
need to have a small PID 1 process that sits in a loop calling wait().

src/libstore/build.cc

@@ -793,6 +793,11 @@ private:
     typedef void (DerivationGoal::*GoalState)();
     GoalState state;
     
+    /* Stuff we need to pass to initChild(). */
+    PathSet dirsInChroot;
+    typedef map<string, string> Environment;
+    Environment env;
+
 public:
     DerivationGoal(const Path & drvPath, Worker & worker);
     ~DerivationGoal();
@@ -821,6 +826,11 @@ private:
     /* Start building a derivation. */
     void startBuilder();
 
+    /* Initialise the builder's process. */
+    void initChild();
+
+    friend int childEntry(void *);
+
     /* Must be called after the output paths have become valid (either
        due to a successful build or hook, or because they already
        were). */
@@ -1468,6 +1478,13 @@ void chmod(const Path & path, mode_t mode)
 }
 
 
+int childEntry(void * arg)
+{
+    ((DerivationGoal *) arg)->initChild();
+    return 1;
+}
+
+
 void DerivationGoal::startBuilder()
 {
     startNest(nest, lvlInfo,
@@ -1480,8 +1497,6 @@ void DerivationGoal::startBuilder()
             % drv.platform % thisSystem % drvPath);
 
     /* Construct the environment passed to the builder. */
-    typedef map<string, string> Environment;
-    Environment env; 
     
     /* Most shells initialise PATH to some default (/bin:/usr/bin:...) when
        PATH is not set.  We don't want this, so we fill it in with some dummy
@@ -1635,7 +1650,6 @@ void DerivationGoal::startBuilder()
        work properly.  Purity checking for fixed-output derivations
        is somewhat pointless anyway. */
     useChroot = queryBoolSetting("build-use-chroot", false);
-    PathSet dirsInChroot;
 
     if (fixedOutput) useChroot = false;
 
@@ -1691,7 +1705,6 @@ void DerivationGoal::startBuilder()
         Paths defaultDirs;
         defaultDirs.push_back("/dev");
         defaultDirs.push_back("/dev/pts");
-        defaultDirs.push_back("/proc");
 
         Paths dirsInChroot_ = querySetting("build-chroot-dirs", defaultDirs);
         dirsInChroot.insert(dirsInChroot_.begin(), dirsInChroot_.end());
@@ -1760,42 +1773,64 @@ void DerivationGoal::startBuilder()
     /* Fork a child to build the package.  Note that while we
        currently use forks to run and wait for the children, it
        shouldn't be hard to use threads for this on systems where
-       fork() is unavailable or inefficient. */
+       fork() is unavailable or inefficient.
+
+       If we're building in a chroot, then also set up private
+       namespaces for the build:
+
+       - The PID namespace causes the build to start as PID 1.
+         Processes outside of the chroot are not visible to those on
+         the inside, but processes inside the chroot are visible from
+         the outside (though with different PIDs).
+
+       - The private mount namespace ensures that all the bind mounts
+         we do will only show up in this process and its children, and
+         will disappear automatically when we're done.
+         
+       - The private network namespace ensures that the builder cannot
+         talk to the outside world (or vice versa).  It only has a
+         private loopback interface.
+
+       - The IPC namespace prevents the builder from communicating
+         with outside processes using SysV IPC mechanisms (shared
+         memory, message queues, semaphores).  It also ensures that
+         all IPC objects are destroyed when the builder exits.
+    */
+#if CHROOT_ENABLED
+    if (useChroot) {
+        char stack[32 * 1024];
+        pid = clone(childEntry, stack + sizeof(stack) - 8,
+            CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET | CLONE_NEWIPC | CLONE_NEWUTS | SIGCHLD, this);
+    } else
+#endif
+    {
         pid = fork();
-    switch (pid) {
+        if (pid == 0) initChild();
+        else if (pid == -1) throw SysError("unable to fork");
+    }
 
-    case -1:
+    /* parent */
-        throw SysError("unable to fork");
+    pid.setSeparatePG(true);
+    builderOut.writeSide.close();
+    worker.childStarted(shared_from_this(), pid,
+        singleton<set<int> >(builderOut.readSide), true, true);
 
-    case 0:
+    if (printBuildTrace) {
+        printMsg(lvlError, format("@ build-started %1% %2% %3% %4%")
+            % drvPath % drv.outputs["out"].path % drv.platform % logFile);
+    }
+}
 
-        /* Warning: in the child we should absolutely not make any
+
-           SQLite calls! */
+void DerivationGoal::initChild()
+{
+    /* Warning: in the child we should absolutely not make any SQLite
+       calls! */
 
     try { /* child */
 
 #if CHROOT_ENABLED
         if (useChroot) {
-                /* Set up private namespaces for the build:
-
-                   - The private mount namespace ensures that all the
-                     bind mounts we do will only show up in this
-                     process and its children, and will disappear
-                     automatically when we're done.
-
-                   - The private network namespace ensures that the
-                     builder cannot talk to the outside world (or vice
-                     versa).  It only has a private loopback
-                     interface.
-
-                   - The IPC namespace prevents the builder from
-                     communicating with outside processes using SysV
-                     IPC mechanisms (shared memory, message queues,
-                     semaphores).  It also ensures that all IPC
-                     objects are destroyed when the builder exits. */
-                if (unshare(CLONE_NEWNS | CLONE_NEWNET | CLONE_NEWIPC | CLONE_NEWUTS) == -1)
-                    throw SysError("cannot set up private namespaces");
-
             /* Initialise the loopback interface. */
             AutoCloseFD fd(socket(PF_INET, SOCK_DGRAM, IPPROTO_IP));
             if (fd == -1) throw SysError("cannot open IP socket");
@@ -1828,11 +1863,16 @@ void DerivationGoal::startBuilder()
                     throw SysError(format("bind mount from `%1%' to `%2%' failed") % source % target);
             }
 
+            /* Bind a new instance of procfs on /proc to reflect our
+               private PID namespace. */
+            if (mount("none", (chrootRootDir + "/proc").c_str(), "proc", 0, 0) == -1)
+                throw SysError("mounting /proc");
+                    
             /* Do the chroot().  Below we do a chdir() to the
                temporary build directory to make sure the current
                directory is in the chroot.  (Actually the order
-                   doesn't matter, since due to the bind mount tmpDir
+               doesn't matter, since due to the bind mount tmpDir and
-                   and tmpRootDit/tmpDir are the same directories.) */
+               tmpRootDit/tmpDir are the same directories.) */
             if (chroot(chrootRootDir.c_str()) == -1)
                 throw SysError(format("cannot change root directory to `%1%'") % chrootRootDir);
         }
@@ -1854,8 +1894,8 @@ void DerivationGoal::startBuilder()
                 throw SysError("cannot set i686-linux personality");
         }
 
-            /* Impersonate a Linux 2.6 machine to get some determinism
+        /* Impersonate a Linux 2.6 machine to get some determinism in
-               in builds that depend on the kernel version. */
+           builds that depend on the kernel version. */
         if ((drv.platform == "i686-linux" || drv.platform == "x86_64-linux") &&
             queryBoolSetting("build-impersonate-linux-26", true))
         {
@@ -1874,12 +1914,12 @@ void DerivationGoal::startBuilder()
         std::vector<const char *> args; /* careful with c_str()! */
         string user; /* must be here for its c_str()! */
             
-            /* If we are running in `build-users' mode, then switch to
+        /* If we are running in `build-users' mode, then switch to the
-               the user we allocated above.  Make sure that we drop
+           user we allocated above.  Make sure that we drop all root
-               all root privileges.  Note that above we have closed
+           privileges.  Note that above we have closed all file
-               all file descriptors except std*, so that's safe.  Also
+           descriptors except std*, so that's safe.  Also note that
-               note that setuid() when run as root sets the real,
+           setuid() when run as root sets the real, effective and
-               effective and saved UIDs. */
+           saved UIDs. */
         if (buildUser.enabled()) {
             printMsg(lvlChatty, format("switching to user `%1%'") % buildUser.getUser());
 
@@ -1931,19 +1971,6 @@ void DerivationGoal::startBuilder()
 }
 
 
-    /* parent */
-    pid.setSeparatePG(true);
-    builderOut.writeSide.close();
-    worker.childStarted(shared_from_this(), pid,
-        singleton<set<int> >(builderOut.readSide), true, true);
-
-    if (printBuildTrace) {
-        printMsg(lvlError, format("@ build-started %1% %2% %3% %4%")
-            % drvPath % drv.outputs["out"].path % drv.platform % logFile);
-    }
-}
-
-
 /* Parse a list of reference specifiers.  Each element must either be
    a store path, or the symbolic name of the output of the derivation
    (such as `out'). */

src/libutil/util.cc

@@ -779,8 +779,11 @@ void Pid::kill()
     int status;
     while (waitpid(pid, &status, 0) == -1) {
         checkInterrupt();
-        if (errno != EINTR) printMsg(lvlError,
+        if (errno != EINTR) {
+            printMsg(lvlError,
                 (SysError(format("waiting for process %1%") % pid).msg()));
+            break;
+        }
     }
 
     pid = -1;