module Shell_sys

: sig

Calls external programs, creates pipelines, etc. (full interface)

This module is now thread-safe (as of May 2009), provided the threads do no share the same Shell or Shell_sys values. Problems reported earlier here have been resolved.

If you get errors like "Netsys_posix.watch_subprocess: uninitialized" you should call Shell_sys.install_job_handlers.

Common exceptions

exception Fatal_error of exn

An error is fatal if it is not possible to recover from it in a predictable manner. In this case, many function wrap such exceptions x into Fatal_error x.

Environments

type environment

The abstract type of a process environment

val create_env : unit -> environment

Creates an empty environment

val current_env : unit -> environment

Returns the environment of the current process as abstract environment value

val copy_env : environment -> environment

Copies an environment

val set_env : environment -> string array -> unit

Sets the contents of the environment to the passed string array

val get_env : environment -> string array

Gets the contents of the environment as string array

val iter_env : f:(string -> unit) -> environment -> unit

Iterates over the strings of the environment, and calls f s for every string s.

val set_env_var : environment -> string -> string -> unit

set_env_var env varname varval: Sets the value of the variable varname in the environment env to varval.

val get_env_var : environment -> string -> string

Returns the value of the variable in the environment

val iter_env_vars : f:(string -> string -> unit) -> environment -> unit

Iterates over the variables of the environment, and calls f name value for every variable with name and value.

Commands

type command

A command describes how to start a new process

val command : ?cmdname:string -> ?arguments:string array -> ?chdir:string -> ?environment:environment -> ?descriptors:Unix.file_descr list -> ?assignments:(Unix.file_descr * Unix.file_descr) list -> filename:string -> unit -> command

Creates a command from the passed arguments:

filename The name of the executable to start. The executable file is not searched, use [root:Shell_sys].lookup_executable for this purpose.

cmdname The name of the command passed in argv[0]. By default, this argument is derived from filename.

arguments The arguments of the command (starting with the first real argument, skipping cmdname). By default [].

chdir Before the command is executed it is changed to this directory.

environment The environment of the command. By default, the current environment

assignments A list of descriptor pairs (fd_from,fd_to). The descriptor fd_from in the current process will be assigned to fd_to in the subprocess started for the command. The list of assignments is executed sequentially, so later assignments must take the effect of previous assignments into account. For example, to make stderr of the subprocess write to stdout of the parent process, pass [(stdout; stderr)].

By default, there are no assignments.

descriptors The list of file descriptors to share with the current process. In the subprocess only those descriptors remain open that are either mentioned in descriptors, or that are the final target of assignments. By default, [stdin; stdout; stderr].
Note that only the final targets of assignments remain open in the subprocess (unless they are also listed in descriptors). If there are cascaded assignments like (fd1, fd2); (fd2, fd3) the intermediate descriptors like fd2 are not considered as final targets; only fd3 would be a final target in this example.

exception Executable_not_found of string

Raised when an executable file cannot be found; the argument is the search name

val lookup_executable : ?path:string list -> string -> string

Searches an executable file. If the passed search name contains a slash, it is expected that this name is already the path name of the executable. If the search name does not contain a slash character, it is searched in the directories enumerated by the search path.

path The search path. By default, the contents of the variable PATH of the current environment, split by ':', are used (Win32: SearchPath is used)

val get_cmdname : command -> string

Returns the name of the command

val get_arguments : command -> string array

Returns the argument array of the command (skipping the command name)

val get_chdir : command -> string option

Returns the chdir parameter of the command

val get_environment : command -> environment

Returns the designated environment of the command

val get_descriptors : command -> Unix.file_descr list

Returns the list of active descriptors

val get_assignments : command -> (Unix.file_descr * Unix.file_descr) list

Returns the list of assignments (fd_from,fd_to)

val get_filename : command -> string

Returns the file name of the executable

val set_cmdname : command -> string -> unit

Sets the command name

val set_arguments : command -> string array -> unit

Sets the argument array

val set_chdir : command -> string option -> unit

Sets the chdir parameter of the command

val set_environment : command -> environment -> unit

Sets the environment

val set_descriptors : command -> Unix.file_descr list -> unit

Sets the list of active descriptors

val set_assignments : command -> (Unix.file_descr * Unix.file_descr) list -> unit

Sets the list of assignments (fd_from,fd_to)

val set_filename : command -> string -> unit

Sets the file name of the executable to start

val copy_command : command -> command

Returns a duplicate of the command description

val is_executable : command -> bool

Returns true if there is an executable file for the command, and it is permitted to run this file (as stated by the file permissions).

false means that the command can definitely not be executed. However, even if the function returns true there may be still reasons that execution will fail.

Processes

type process

A process is the running instance of a command (a Unix process)

type group_action =

# \| New_bg_group	(*	Start process in new background process group	*)
# \| New_fg_group	(*	Start process in new foreground process group	*)
# \| Join_group of int	(*	Started process joins this existing process group	*)
# \| Current_group	(*	Started process remains in the current group	*)

Determines in which process group the new process will run

type fwd_mode =

# \| No_forward	(*	No forwarding of keyboard signals	*)
# \| Forward_to_process	(*	Forward signals directly to subprocess	*)
# \| Forward_to_group	(*	Forward signals to the process group of the subprocess	*)

Determines whether and how keyboard signals (SIGINT, SIGQUIT) are forwarded from the caller to the new child. There is no forwarding in Win32 - all console applications get the keyboard signals anyway.

val run : ?group:group_action -> ?forward_mode:fwd_mode -> ?pipe_assignments:(Unix.file_descr * Unix.file_descr) list -> command -> process

Executes the command concurrently with the current process. The function does not wait until the process terminates; it returns immediately after the exec system call has been successfully performed; errors that occur until exec are caught and reported as exception (even errors in the fresh subprocess).

On error, one can assume that the process state has been cleaned up: any forked child process has terminated; any modifications of the global process state has been restored.

File descriptor assignments: First, the assignments in pipe_assignments are performed, then the assignments contained in the command. The pipe_assignments are interpreted as parallel assignment, not as sequential assignment.

Note: For users without very special needs, it is recommended to run jobs instead of processes. See below for the job API.

group Determines in which process group the new process will run. By default Current_group.

forward_mode Whether and how to forward keyboard signals to the new child. By default No_forward. The Win32 implementation ignores this argument.

pipe_assignments A list of descriptor pairs (fd_from,fd_to). The descriptor fd_from in the current process will be assigned to fd_to in the started subprocess. In order to take effect, fd_to must also be passed in the descriptors property of the started command. Furthermore, fd_from may or may not be member of descriptors; in the first case it will remain open, in the latter case it will be closed. The list of assignments is executed in parallel. For example, to swap the roles of stdout and stderr, pass the list [(stdout,stderr); (stderr,stdout)].

val process_id : process -> int

Returns the process ID of the process

val status : process -> Unix.process_status

Reports the status so far known: If the process has terminated, the status of the process is returned. If the process is still running, Not_found will be raised.

val command_of_process : process -> command

Returns the command that is now running as the process

val call : command -> process

Executes the command and waits until the process terminates (synchronous execution a la system, but no intermediate shell). status is guaranteed to return WEXITED or WSIGNALED.

val kill : ?signal:int -> process -> unit

Sends a signal to the passed process.

signal The signal to send, by default SIGTERM

Jobs

A job is the description of how to run several commands which are linked by pipelines (or which are just a logical unit). A job_instance is the running instance of a job.

Jobs are implemented on a higher layer than commands; the following means of the operating system are used by job invocations:

Normally a job_instance corresponds to a Unix process group. In this case the last added command will result in the process group leader.
Controlling the execution of jobs requires that signal handlers are set in many cases (see install_job_handlers)
The processes of jobs are often interconnected by pipelines (see add_pipeline).
It is possible to handle pipelines between the current process and processes of the job (see add_producer and add_consumer)

Important:

In order to run jobs efficiently (without busy waiting) and properly it is strongly recommended to install the signal handlers using install_job_handlers

type job

type job_instance

val new_job : unit -> job

Creates a new job descriptor. Initially the job is empty, but you can fill it with commands (add_command), pipelines (add_pipeline), consumers (add_consumer) and producers (add_producer). When the job is set up, you can start it (run_job/finish_job or call_job).

val add_command : command -> job -> unit

Adds a command to a job.

Note that you cannot add the same command twice; however you can add a copy of a command already belonging to the job.

val add_pipeline : ?bidirectional:bool -> ?src_descr:Unix.file_descr -> ?dest_descr:Unix.file_descr -> src:command -> dest:command -> job -> unit

Adds a pipeline which redirects the output of the command src to the input of the command dest.

src_descr determines the file descriptor of the source command which is redirected. This is by default stdout.

dest_descr determines the file descriptor of the destination command to which the data stream is sent. This is by default stdin.

bidirectional if false (default), a classical pipe is created to connect the file descriptors. This normally restricts the data flow to one direction. If true, a socketpair is created which is roughly a bidirectional pipe. In this case, data flow in both directions is possible.

val add_producer : ?descr:Unix.file_descr -> producer:(Unix.file_descr -> bool) -> command -> job -> unit

Adds a producer to the job. A producer transfers data to the subprocess realizing the passed command. To do so, a pipe is created between the file descriptor descr of the subprocess and another descriptor descr' which is open in the current process. The function producer is called when data can be written into the pipe. The argument of producer is the writing end of the pipe descr'. This file descriptor is in non-blocking mode. The function producer must close descr' when all data are transferred. The return value of producer indicates whether the descriptor is still open.

descr The descriptor of the subprocess to which the reading end of the pipe is dup'ed. By default stdin.

val from_string : ?pos:int -> ?len:int -> ?epipe:(unit -> unit) -> string -> Unix.file_descr -> bool

from_string ?pos ?len ?epipe s returns a function which can be used as producer argument for add_producer. The data transferred to the subprocess is taken from the string s. After these data are sent, the pipeline is closed.

pos The position in s where the data slice to transfer begins. By default 0.

len The length of the data slice to transfer. By default, all bytes from the start position pos to the end of the string are taken.

epipe This function is called when the pipeline breaks (EPIPE). Default: the empty function. EPIPE exceptions are always caught, and implicitly handled by closing the pipeline.

val from_stream : ?epipe:(unit -> unit) -> string Stream.t -> Unix.file_descr -> bool

from_stream ?epipe s returns a function which can be used as producer argument for add_producer. The data transferred to the subprocess is taken from the string stream s. After these data are sent, the pipeline is closed.

epipe This function is called when the pipeline breaks (EPIPE). Default: the empty function. EPIPE exceptions are always caught, and implicitly handled by closing the pipeline.

val add_consumer : ?descr:Unix.file_descr -> consumer:(Unix.file_descr -> bool) -> command -> job -> unit

Adds a consumer to the job. A consumer transfers data from the subprocess realizing the passed command to the current process. To do so, a pipe is created between the file descriptor descr of the subprocess and another descriptor descr' which is open in the current process. The function consumer is called when data can be read from the pipe. The argument of consumer is reading end of the pipe descr'. This file descriptor is in non-blocking mode. The function consumer must close descr' after EOF is detected. The return value of consumer indicates whether the descriptor is still open.

descr The descriptor of the subprocess to which the writing end of the pipe is dup'ed. By default stdout.

val to_buffer : Buffer.t -> Unix.file_descr -> bool

to_buffer b returns a function which can be used as consumer argument for add_consumer. The data received from the subprocess is added to the buffer b.

type group_mode =

# \| Same_as_caller	(*	The job runs in the same process group as the current process	*)
# \| Foreground	(*	The job runs in a new foreground process group	*)
# \| Background	(*	The job runs in a new background process group	*)

Specifies how the job instance is related to process groups

val run_job : ?mode:group_mode -> ?forward_signals:bool -> job -> job_instance

Invokes the commands of the job such that they run concurrently with the main process.

The function returns a job_instance, i.e. a value recording which processes are started, and how they are related. Furthermore, the function has the side effect of adding the job to the global list of current jobs.

The mode argument specifies whether a new Unix process group is created for the job instance. A process group has the advantage that it is possible to send signals to all processes of the group at once. For example, one can terminate a group by sending SIGTERM to it: All member processes get the signal. Usually, these are not only the subprocesses initially created, but also further processes started by the initial members.

So if it is necessary to send signals to the processes of the job, it will be advantegous to run it in a new process group. However, this also means that signals sent to the current process group are not automatically forwarded to the created process group. For example, if the current process group is terminated, the job will continue running, because it is member of a different process group. One has to explicitly catch and forward signals to avoid wild-running jobs.

The moral of the story is that one should only create new process groups when it is necessary (e.g. the user must be able to stop an action at any time). Furthermore, signal forwarding must be configured.

The Unix shell also allows the programmer to specify process group handling to a certain extent. Normally, commands are executed in the same process group as the caller. The syntax "command &" forces that the command is run in a new background process group. There is another situation when new process groups are created: when a new interactive shell is started the commands are run in new foreground process groups (so the keyboard signals like CTRL-C work).

mode Specifies the process group handling. By default, the job is executed in the same process group as the current process (Same_as_caller). The value Background causes that a new background process group is started. The value Foreground causes that a new foreground process group is started. For the latter, it is required that there is a controlling terminal (i.e. it does not work for daemons). Any existing foreground process group (there is at most one) is put into the background, but this is not restored when the job is over (the caller must do this). Foreground process groups should be avoided unless you are writing an interactive shell interpreter.

forward_signals If true, the default, keyboard signals (SIGINT, SIGQUIT) delivered to the current process are forwarded to the job. This has only a meaning if the job is running as background process group. Furthermore, it is required that install_job_handlers has been called to enable signal forwarding.
The function returns normally if at least one process could be started. If no process was startable (i.e. the first command was not startable), an exception is raised. If one or more processes could be started but not all, job_status will return Job_partially_running. The caller should then discard the job and any intermediate result that might already have been produced by the partial job.
When all processes could be started and no other exceptional condition happened, the function sets job_status to Job_running.

class type ['t] job_handler_engine_type =

inherit ['t] Uq_engines.engine

method job : job

Returns the called job

method job_instance : job_instance

Returns the job instance

This type of engine also returns the job and the job_instance.

class job_engine : Unixqueue.event_system -> job_instance -> [unit] job_handler_engine_type

The job_engine watches the job, and looks whether the processes are finished, and if so, it records the process statuses. Also, the engine takes care of pumping producer data into the job, and of collecting consumer data.

val finish_job : job_instance -> unit

This creates a job_engine internally and runs until it is finished, i.e. until the job has been executed.

In previous version of Ocamlnet there was an optional sys argument. This is gone now. Also, the error handling is different. It is no longer possible to restart finish_job when an error happens.

val call_job : ?mode:group_mode -> ?forward_signals:bool -> ?onerror:(job_instance -> unit) -> job -> job_instance

Starts the job (see run_job) and waits until it finishes (see finish_job); i.e. call_job = run_job + finish_job. The function returns normally if all processes can be started; you can examine job_status of the result to get the information whether all processes returned the exit code 0.

onerror If not all of the processes can be started, the function passed by onerror is invoked. By default, this function calls abandon_job to stop the already running processes. After the onerror function has returned, the original exception is raised again. Fatal error conditions are not caught.

mode See run_job

forward_signals See run_job

val processes : job_instance -> process list

Returns the processes that have actually been started for this job by run_job; note that the corresponding Unix process group may have additional processes (e.g. indirectly started processes).

exception No_Unix_process_group

Raised by functions referring to Unix process groups when the job has not been started in its own process group.

val process_group_leader : job_instance -> process

Returns the process group leader process. This function is not available for jobs in the mode Same_as_caller.

val process_group_id : job_instance -> int

Returns the Unix ID of the process group as number > 1. This function is not available for jobs in the mode Same_as_caller.

type job_status =

# \| Job_running	(*	All commands could be started, and at least one process is still running	*)
# \| Job_partially_running	(*	Not all commands could be started, and at least one process is still running	*)
# \| Job_ok	(*	all processes terminated with exit code 0	*)
# \| Job_error	(*	all processes terminated but some abnormally	*)
# \| Job_abandoned	(*	the job has been abandoned (see `abandon_job`)	*)

Indicates the status of the job

val job_status : job_instance -> job_status

Returns the status. The status may only change after finish_job has been called:

after run_job: status is Job_running or Job_partially_running
after finish_job: if returning normally: status is Job_ok or Job_error. After an exception happened the other states are possible, too

val kill_process_group : ?signal:int -> job_instance -> unit

Kills the process group if it is still (at least partially) running. This operation is not available if the mode is Same_as_caller (exception No_Unix_process_group).

Note 1: In the Unix terminology, "killing a job" only means to send a signal to the job. So the job may continue running, or it may terminate; in general we do not know this. Because of this, the job will still have an entry in the job list.

Note 2: Because sub-sub-processes are also killed, this function may send the signal to more processes than kill_processes (below). On the other hand, it is possible that sub-processes change their group ID such that it is also possible that this function sends the signal to fewer processes than kill_processes.

signal The signal number to send (O'Caml signal numbers as used by the Sys module). Default is Sys.sigterm.

val kill_processes : ?signal:int -> job_instance -> unit

Kills the individual processes of the job which are still running.

signal The signal number to send (O'Caml signal numbers as used by the Sys module). Default is Sys.sigterm.

val cancel_job : ?signal:int -> job_instance -> unit

Tries to get rid of a running job. If the mode is Same_as_caller, the signal is sent to the processes individually. If the mode is Foreground or Background, the signal is sent to the process group corresponding to the job.

This function removes the job from the job list; i.e. it is no longer watched. Because of some magic spells it is guaranteed that the job dies immediately without becoming a zombie (provided you have a SIGCHLD handler).

signal The signal number to send (O'Caml signal numbers as used by the Sys module). Default is Sys.sigterm.

val abandon_job : ?signal:int -> job_instance -> unit

Deprecated name for cancel_job

exception Already_installed

Raised when the job handlers are already installed

val configure_job_handlers : ?catch_sigint:bool -> ?catch_sigquit:bool -> ?catch_sigterm:bool -> ?catch_sighup:bool -> ?at_exit:bool -> unit -> unit

Configures signal and at_exit handlers for jobs:

The keyboard signals SIGINT and SIGQUIT are forwarded to all jobs which are running in the background (and thus are not automatically notified) and want to get such signals (forward_signals).
The signals SIGTERM and SIGHUP are (if the handler is installed) forwarded to all dependent processes (regardless whether they are running in their own Unix process group or not, and regardless of forward_signals).
The at_exit handler sends a SIGTERM to all dependent processes, too.

In previous versions of Ocamlnet it was also possible to configure catch_sigchld to set whether a SIGCHLD handler is installed. This is now always done.

In previous versions of Ocamlnet there was also a set_sigpipe flag. This flag is gone as a SIGPIPE handler is now always installed.

The handlers are now managed by [root:Netsys_signal]. The handlers of this module set the keep_default flag for SIGINT, SIGQUIT, SIGTERM, and SIGHUP, so that the default action for these signals is executed after the forwarding to the child processes is done. By setting another handler in [root:Netsys_signal] without that flag this behavior can be overridden.

Note that if an uncaught exception leads to program termination, this situation will not be detected; any running jobs will not be terminated (sorry, this cannot be fixed).

This function sets only which handlers will be installed when install_job_handlers (below) is invoked. The function fails if the handlers are already installed.

Win32: No handlers are installed. It would be desirable to some extent that at least at_exit is honoured, however, this is not yet done.

catch_sigint whether to install a SIGINT handler (default: true)

catch_sigquit whether to install a SIGQUIT handler (default: true)

catch_sigterm whether to install a SIGTERM handler (default: true)

catch_sighup whether to install a SIGHUP handler (default: true)

at_exit whether to set the at_exit handler (default: true)

val install_job_handlers : unit -> unit

Installs handlers as configured before. Raises Already_installed if the handlers are already installed.

Removed functions

The functions add_rd_polling and add_wr_polling have been removed. They were added prior to the merge with the equeue library. Use a Unixqueue now, which is much more powerful.

Also no longer supported because the type system_handler is gone:

type system_handler

type process_event

val wait : 
      ?wnohang:bool ->                     (* default: false *)
      ?wuntraced:bool ->                   (* default: false *)
      ?restart:bool ->                     (* default: false *)
      ?check_interval:float ->             (* default: 0.1 *)
      ?read:(Unix.file_descr list) ->      (* default: [] *)
      ?write:(Unix.file_descr list) ->     (* default: [] *)
      ?except:(Unix.file_descr list) ->    (* default: [] *)
      process list ->
        process_event list

val register_job : system_handler -> job_instance -> unit

val iter_job_instances : f:(job_instance -> unit) ->  unit

val watch_for_zombies : unit -> unit

val process_group_expects_signals : job_instance -> bool

Debugging

module Debug : sig

val enable : bool Pervasives.ref

Enables [root:Netlog]-style debugging of this module

end