process_executer 4.0.3 → 4.1.0

data/lib/process_executer/commands/spawn_with_timeout.rb

@@ -48,7 +48,7 @@ module ProcessExecuter
       #
       def call
         begin
-          @pid = Process.spawn(*command, **options.spawn_options)
+          @pid = Process.spawn(*command, **spawn_options)
         rescue StandardError => e
           raise ProcessExecuter::SpawnError, "Failed to spawn process: #{e.message}"
         end
@@ -80,10 +80,14 @@ module ProcessExecuter
 
       # The status returned by Process.wait2
       #
+      # nil when the timeout was delivered after the wait had already reaped
+      # the subprocess: the status was lost to the raise and {#timed_out?} is
+      # true.
+      #
       # @example
       #   spawn.status #=> #<Process::Status: pid 12345 exit 0>
       #
-      # @return [Process::Status]
+      # @return [Process::Status, nil]
       #
       attr_reader :status
 
@@ -118,6 +122,47 @@ module ProcessExecuter
 
       private
 
+      # The options to pass to Process.spawn
+      #
+      # Subclasses may override this method to combine internal redirections
+      # with the user's options without modifying the options object the
+      # caller gave.
+      #
+      # @return [Hash]
+      #
+      def spawn_options = options.spawn_options.merge(process_group_options)
+
+      # Spawn options that place the subprocess into its own process group
+      #
+      # When `timeout_after` is set to a value that can fire (`nil` and `0`
+      # mean "no timeout"), the subprocess is made the leader of a new process
+      # group so that a timeout can kill the whole group -- including
+      # descendants that inherited the redirections -- instead of just the
+      # direct child. Empty when no timeout can fire or when the caller gave a
+      # `pgroup`/`new_pgroup` option themselves (their setting is honored).
+      #
+      # A new process group is a background group for any terminal the
+      # subprocess inherits, so an interactive subprocess that reads the
+      # terminal is stopped by `SIGTTIN` and then killed when the timeout
+      # fires -- which is the bound `timeout_after` promises. A caller who
+      # needs an interactive subprocess to stay in the foreground process
+      # group can pass their own `pgroup` option.
+      #
+      # @return [Hash]
+      #
+      def process_group_options
+        return {} unless options.timeout_after&.positive?
+        return {} unless options.pgroup == :not_set && options.new_pgroup == :not_set
+
+        windows? ? { new_pgroup: true } : { pgroup: true }
+      end
+
+      # Whether the current platform is Windows
+      #
+      # @return [Boolean]
+      #
+      def windows? = Gem.win_platform?
+
       # Wait for process to terminate
       #
       # If a `:timeout_after` is specified in options, terminate the process after the
@@ -142,21 +187,145 @@ module ProcessExecuter
 
       # Wait for a process to terminate returning the status and timed out flag
       #
-      # @return [Array<Process::Status, Boolean>] an array containing the process status and a boolean
-      #   indicating whether the process timed out
+      # An exception other than the timeout (an `Interrupt` from Ctrl-C, for
+      # example) abandons the wait; {#kill_and_reap_abandoned_subprocess} then
+      # cleans up a subprocess this class isolated into its own process group
+      # before the exception propagates.
+      #
+      # @return [Array(Process::Status, Boolean), Array(nil, Boolean)] an array containing
+      #   the process status (nil when the timeout raced the wait and the status was lost,
+      #   see {#wait_with_timeout}) and a boolean indicating whether the process timed out
       def wait_for_process_raw
-        timed_out = false
-
-        process_status =
-          begin
-            Timeout.timeout(options.timeout_after) { Process.wait2(pid).last }
-          rescue Timeout::Error
-            Process.kill('KILL', pid)
-            timed_out = true
-            Process.wait2(pid).last
-          end
-
-        [process_status, timed_out]
+        wait_with_timeout
+      rescue Exception # rubocop:disable Lint/RescueException
+        kill_and_reap_abandoned_subprocess
+        raise
+      end
+
+      # Wait for the process, killing it when `timeout_after` expires first
+      #
+      # The timeout can be delivered after the timed wait has already reaped
+      # the subprocess but before it returns. In that race the subprocess's
+      # status was lost to the raise, so the status is nil and the timed out
+      # flag is still set.
+      #
+      # @return [Array(Process::Status, Boolean), Array(nil, Boolean)] an array containing
+      #   the process status (nil when the timeout raced the wait and the status was lost)
+      #   and a boolean indicating whether the process timed out
+      def wait_with_timeout
+        process_status = Timeout.timeout(options.timeout_after) { Process.wait2(pid).last }
+        [process_status, false]
+      rescue Timeout::Error
+        kill_subprocess
+        begin
+          [Process.wait2(pid).last, true]
+        rescue Errno::ECHILD
+          # the interrupted wait already reaped the subprocess; its status was
+          # lost to the raise
+          [nil, true]
+        end
+      end
+
+      # Kill and reap the subprocess when its wait was abandoned by an exception
+      #
+      # Only applies to a subprocess this class isolated into its own process
+      # group: such a subprocess no longer receives terminal-generated signals
+      # (Ctrl-C sends `SIGINT` to the caller's foreground group, not to the
+      # new group), so an exception that abandons the wait would otherwise
+      # leave it and its descendants running unsupervised and unreaped. A
+      # subprocess whose process group came from the caller's own options
+      # keeps its pre-existing signal semantics and is left alone.
+      #
+      # Rescues `Exception` (not just `StandardError`) so that a second async
+      # exception delivered during this best-effort cleanup cannot replace
+      # the exception already being re-raised by the caller.
+      #
+      # @return [void]
+      #
+      def kill_and_reap_abandoned_subprocess
+        return unless isolated_in_new_process_group?
+
+        kill_subprocess
+        Process.wait2(pid)
+      rescue Exception # rubocop:disable Lint/RescueException
+        # the subprocess may already be dead and reaped; the wait's exception
+        # is what must propagate
+      end
+
+      # Forcibly terminate the timed out subprocess and (if possible) its descendants
+      #
+      # When the subprocess was spawned as the leader of its own process
+      # group, the whole group is killed so that descendants that would
+      # otherwise survive the timeout (and keep any inherited redirection
+      # file descriptors open) are terminated too, falling back to killing
+      # the direct child if the group kill fails. A group signal reaches only
+      # the processes still in that group: a descendant that started its own
+      # session or joined another process group (a daemon, for example) is
+      # not killed. Otherwise the subprocess is in a process group this
+      # object did not create, so only the direct child is killed, matching
+      # the pre-process-group behavior.
+      #
+      # Killing a process group is only possible on POSIX platforms. On
+      # Windows, `Process.kill` cannot signal a process group (a negative pid
+      # raises an error), so the group kill always falls back to the direct
+      # child and descendants may survive the timeout; the bounded
+      # {MonitoredPipe#close} keeps such descendants from blocking
+      # {ProcessExecuter.run} indefinitely.
+      #
+      # A subprocess that already exited and was reaped before the signal is
+      # sent (the timeout racing the wait) leaves nothing to kill; that is not
+      # an error. In that same microsecond window the freed pid could in
+      # principle be recycled to an unrelated process, a hazard inherent to
+      # signaling by pid: Ruby exposes no race-free process handle (such as
+      # Linux's pidfd) that would eliminate it, and reuse would require the OS
+      # to cycle through its entire pid space within the window.
+      #
+      # @return [void]
+      #
+      def kill_subprocess
+        return if process_group_leader? && kill_process_group
+
+        Process.kill('KILL', pid)
+      rescue Errno::ESRCH
+        # the subprocess already exited and was reaped between the interrupted
+        # wait and the kill; there is nothing left to kill
+      end
+
+      # Whether the spawn options made the subprocess a new process group leader
+      #
+      # True when the process group option -- added by {#process_group_options}
+      # or given by the caller -- asks for a new process group with the
+      # subprocess as its leader (`pgroup: true`, `pgroup: 0`, or
+      # `new_pgroup: true`). False when there is no process group option or
+      # when `pgroup` places the subprocess in an existing process group.
+      #
+      # @return [Boolean]
+      #
+      def process_group_leader?
+        [true, 0].include?(spawn_options[:pgroup]) || spawn_options[:new_pgroup] == true
+      end
+
+      # Whether this class isolated the subprocess into its own process group
+      #
+      # True when the subprocess is a new process group leader and that came
+      # from {#process_group_options} rather than from a `pgroup`/`new_pgroup`
+      # option the caller supplied.
+      #
+      # @return [Boolean]
+      #
+      def isolated_in_new_process_group?
+        process_group_leader? && options.pgroup == :not_set && options.new_pgroup == :not_set
+      end
+
+      # Send SIGKILL to the subprocess's process group
+      #
+      # @return [Boolean] true if the signal was sent, false if doing so raised an error
+      #
+      def kill_process_group
+        Process.kill('KILL', -pid)
+        true
+      rescue StandardError
+        false
       end
     end
   end

data/lib/process_executer/destinations/monitored_pipe.rb

@@ -25,12 +25,15 @@ module ProcessExecuter
         destination.write data
       end
 
-      # Closes the pipe if it's open
+      # Does nothing: closing a caller-provided pipe is the caller's responsibility
+      #
+      # The library closes only pipes it creates. A {ProcessExecuter::MonitoredPipe}
+      # given as a redirection destination was created by the caller, and the
+      # {ProcessExecuter::MonitoredPipe} documentation makes closing it the
+      # caller's responsibility.
       #
       # @return [void]
-      def close
-        destination.close if destination.state == :open
-      end
+      def close; end
 
       # Determines if this class can handle the given destination
       #

data/lib/process_executer/destinations.rb

@@ -76,7 +76,7 @@ module ProcessExecuter
       destination_classes =
         ProcessExecuter::Destinations.constants
                                      .map { |const| ProcessExecuter::Destinations.const_get(const) }
-                                     .select { |const| const.is_a?(Class) }
+                                     .grep(Class)
                                      .reject { |klass| klass == ProcessExecuter::Destinations::DestinationBase }
 
       destination_classes.find { |klass| klass.handles?(destination) }

data/lib/process_executer/errors.rb

@@ -29,7 +29,7 @@ module ProcessExecuter
   # | `CommandError` | A subclass of this error is raised when there is a problem executing a command. |
   # | `FailedError` | Raised when the command exits with a non-zero exit status. |
   # | `SignaledError` | Raised when the command is terminated as a result of receiving a signal. This could happen if the process is forcibly terminated or if there is a serious system error. |
-  # | `TimeoutError` | This is a specific type of `SignaledError` that is raised when the command times out and is killed via the SIGKILL signal. |
+  # | `TimeoutError` | This is a specific type of `SignaledError` that is raised when the wait for the command is cut short by the configured `timeout_after` elapsing. The command is sent the SIGKILL signal, unless it had already exited in the moment between the timeout firing and the kill. |
   # | `ProcessIOError` | Raised when an error was encountered reading or writing to the command's subprocess. |
   # | `SpawnError` | Raised when the process could not execute. Check the `#cause` for the original exception from `Process.spawn`.  |
   #
@@ -139,7 +139,7 @@ module ProcessExecuter
   #
   class SignaledError < ProcessExecuter::CommandError; end
 
-  # Raised when the command takes longer than the configured timeout_after
+  # Raised when the wait for the command is cut short by `timeout_after` elapsing
   #
   # @example
   #   begin