A Hacker's Guide to NCURSES

Contents

Abstract

This document is a hacker's tour of the ncurses library and utilities. It discusses design philosophy, implementation methods, and the conventions used for coding and documentation. It is recommended reading for anyone who is interested in porting, extending or improving the package.

Objective of the Package

The objective of the ncurses package is to provide a freeware API for character-cell terminals and terminal emulators with the following characteristics:

These objectives are in priority order. So, for example, source compatibility with older version must trump featurefulness -- we cannot add features if it means breaking the portion of the API corresponding to historical curses versions.

Why System V Curses?

We used System V curses as a model, reverse-engineering their API, in order to fulfill the first two objectives.

System V curses implementations can support BSD curses programs with just a recompilation, so by capturing the System V API we also capture BSD's.

More importantly for the future, the XSI Curses standard issued by X/Open is explicitly and closely modeled on System V. So conformance with System V took us most of the way to base-level XSI conformance.

How to Design Extensions

The third objective (standards conformance) requires that it be easy to condition source code using ncurses so that the absence of nonstandard extensions does not break the code.

Accordingly, we have a policy of associating with each nonstandard extension a feature macro, so that ncurses client code can use this macro to condition in or out the code that requires the ncurses extension.

For example, there is a macro NCURSES_MOUSE_VERSION which XSI Curses does not define, but which is defined in the ncurses library header. You can use this to condition the calls to the mouse API calls.

Portability and Configuration

Code written for ncurses may assume an ANSI-standard C compiler and POSIX-compatible OS interface. It may also assume the presence of a System-V-compatible select(2) call.

We encourage (but do not require) developers to make the code friendly to less-capable UNIX environments wherever possible.

We encourage developers to support OS-specific optimizations and methods not available under POSIX/ANSI, provided only that:

We use GNU autoconf(1) as a tool to deal with portability issues. The right way to leverage an OS-specific feature is to modify the autoconf specification files (configure.in and aclocal.m4) to set up a new feature macro, which you then use to condition your code.

If autoconf Fails

The 'configure' script usually gets your system environment right automatically. Here are some -D options you might need to compile with if it fails:

-DHAVE_UNISTD_H
if <unistd.h> is present
-DHAVE_SIGACTION
if the sigaction function is present
-DHAVE_USLEEP
if the usleep function is present
-DSVR4_ACTION
if (e.g., svr4) you need _POSIX_SOURCE to have sigaction
-DHAVE_TERMIOS_H
if you have <termios.h>
-DHAVE_TERMIO_H
if you have <termio.h>; otherwise it uses <sgtty.h>
-DBROKEN_TIOCGETWINSZ
on SVR4 and HPUX, the get window size ioctl is broken.

Documentation Conventions

There are three kinds of documentation associated with this package. Each has a different preferred format:

Our conventions are simple:

  1. Maintain package-internal files in plain text. The expected viewer for them more(1) or an editor window; there's no point in elaborate mark-up.

  2. Mark up manual pages in the man macros. These have to be viewable through traditional man(1) programs.

  3. Write everything else in HTML.
When in doubt, HTMLize a master and use lynx(1) to generate plain ASCII (as we do for the announcement document).

The reason for choosing HTML is that it's (a) well-adapted for on-line browsing through viewers that are everywhere; (b) more easily readable as plain text than most other mark-ups, if you don't have a viewer; and (c) carries enough information that you can generate a nice-looking printed version from it. Also, of course, it make exporting things like the announcement document to WWW pretty trivial.

How to Report Bugs

The reporting address for bugs is bug-ncurses@gnu.org. This is a majordomo list; to join, write to ncurses-request@gnu.org with a message containing the line:
             subscribe <name>@<host.domain>
The ncurses code is maintained by a small group of volunteers. While we try our best to fix bugs promptly, we simply don't have a lot of hours to spend on elementary hand-holding. We rely on intelligent cooperation from our users. If you think you have found a bug in ncurses, there are some steps you can take before contacting us that will help get the bug fixed quickly.

In order to use our bug-fixing time efficiently, we put people who show us they've taken these steps at the head of our queue. This means that if you don't, you'll probably end up at the tail end and have to wait a while.

  1. Develop a recipe to reproduce the bug.

    Bugs we can reproduce are likely to be fixed very quickly, often within days. The most effective single thing you can do to get a quick fix is develop a way we can duplicate the bad behavior -- ideally, by giving us source for a small, portable test program that breaks the library. (Even better is a keystroke recipe using one of the test programs provided with the distribution.)

  2. Try to reproduce the bug on a different terminal type.

    In our experience, most of the behaviors people report as library bugs are actually due to subtle problems in terminal descriptions. This is especially likely to be true if you're using a traditional asynchronous terminal or PC-based terminal emulator, rather than xterm or a UNIX console entry.

    It's therefore extremely helpful if you can tell us whether or not your problem reproduces on other terminal types. Usually you'll have both a console type and xterm available; please tell us whether or not your bug reproduces on both.

    If you have xterm available, it is also good to collect xterm reports for different window sizes. This is especially true if you normally use an unusual xterm window size -- a surprising number of the bugs we've seen are either triggered or masked by these.

  3. Generate and examine a trace file for the broken behavior.

    Recompile your program with the debugging versions of the libraries. Insert a trace() call with the argument set to TRACE_UPDATE. (See "Writing Programs with NCURSES" for details on trace levels.) Reproduce your bug, then look at the trace file to see what the library was actually doing.

    Another frequent cause of apparent bugs is application coding errors that cause the wrong things to be put on the virtual screen. Looking at the virtual-screen dumps in the trace file will tell you immediately if this is happening, and save you from the possible embarrassment of being told that the bug is in your code and is your problem rather than ours.

    If the virtual-screen dumps look correct but the bug persists, it's possible to crank up the trace level to give more and more information about the library's update actions and the control sequences it issues to perform them. The test directory of the distribution contains a tool for digesting these logs to make them less tedious to wade through.

    Often you'll find terminfo problems at this stage by noticing that the escape sequences put out for various capabilities are wrong. If not, you're likely to learn enough to be able to characterize any bug in the screen-update logic quite exactly.

  4. Report details and symptoms, not just interpretations.

    If you do the preceding two steps, it is very likely that you'll discover the nature of the problem yourself and be able to send us a fix. This will create happy feelings all around and earn you good karma for the first time you run into a bug you really can't characterize and fix yourself.

    If you're still stuck, at least you'll know what to tell us. Remember, we need details. If you guess about what is safe to leave out, you are too likely to be wrong.

    If your bug produces a bad update, include a trace file. Try to make the trace at the least voluminous level that pins down the bug. Logs that have been through tracemunch are OK, it doesn't throw away any information (actually they're better than un-munched ones because they're easier to read).

    If your bug produces a core-dump, please include a symbolic stack trace generated by gdb(1) or your local equivalent.

    Tell us about every terminal on which you've reproduced the bug -- and every terminal on which you can't. Ideally, sent us terminfo sources for all of these (yours might differ from ours).

    Include your ncurses version and your OS/machine type, of course! You can find your ncurses version in the curses.h file.

If your problem smells like a logic error or in cursor movement or scrolling or a bad capability, there are a couple of tiny test frames for the library algorithms in the progs directory that may help you isolate it. These are not part of the normal build, but do have their own make productions.

The most important of these is mvcur, a test frame for the cursor-movement optimization code. With this program, you can see directly what control sequences will be emitted for any given cursor movement or scroll/insert/delete operations. If you think you've got a bad capability identified, you can disable it and test again. The program is command-driven and has on-line help.

If you think the vertical-scroll optimization is broken, or just want to understand how it works better, build hashmap and read the header comments of hardscroll.c and hashmap.c; then try it out. You can also test the hardware-scrolling optimization separately with hardscroll.

There's one other interactive tester, tctest, that exercises translation between termcap and terminfo formats. If you have a serious need to run this, you probably belong on our development team!

A Tour of the Ncurses Library

Library Overview

Most of the library is superstructure -- fairly trivial convenience interfaces to a small set of basic functions and data structures used to manipulate the virtual screen (in particular, none of this code does any I/O except through calls to more fundamental modules described below). The files lib_addch.c, lib_bkgnd.c, lib_box.c, lib_clear.c, lib_clrbot.c, lib_clreol.c, lib_data.c, lib_delch.c, lib_delwin.c, lib_erase.c, lib_getstr.c, lib_inchstr.c, lib_insch.c, lib_insdel.c, lib_insstr.c, lib_instr.c, lib_isendwin.c, lib_keyname.c, lib_move.c, lib_mvwin.c, lib_overlay.c, lib_pad.c, lib_printw.c, lib_scanw.c, lib_screen.c, lib_scroll.c, lib_scrreg.c, lib_set_term.c, lib_slk.c, lib_touch.c, lib_unctrl.c, and lib_window.c are all in this category. They are very unlikely to need change, barring bugs or some fundamental reorganization in the underlying data structures.

The lib_trace.c, lib_traceatr.c, and lib_tracechr.c file are used only for debugging support. It is rather unlikely you will ever need to change these, unless you want to introduce a new debug trace level for some reasoon.

There is another group of files that do direct I/O via tputs(), computations on the terminal capabilities, or queries to the OS environment, but nevertheless have only fairly low complexity. These include: lib_acs.c, lib_beep.c, lib_color.c, lib_endwin.c, lib_initscr.c, lib_longname.c, lib_newterm.c, lib_options.c, lib_termcap.c, lib_ti.c, lib_tparm.c, lib_tputs.c, lib_vidattr.c, and read_entry.c. These are likely to need revision only if ncurses is being ported to an environment without an underlying terminfo capability representation.

The files lib_kernel.c, lib_baudrate.c, lib_raw.c, lib_tstp.c, and lib_twait.c have serious hooks into the tty driver and signal facilities. If you run into porting snafus moving the package to another UNIX, the problem is likely to be in one of these files. The file lib_print.c uses sleep(2) and also falls in this category.

Almost all of the real work is done in the files hashmap.c, hardscroll.c, lib_addch.c, lib_doupdate.c, lib_mvcur.c, lib_getch.c, lib_mouse.c, lib_refresh.c, and lib_setup.c. Most of the algorithmic complexity in the library lives in these files. If there is a real bug in ncurses itself, it's probably here. We'll tour some of these files in detail below (see The Engine Room).

Finally, there is a group of files that is actually most of the terminfo compiler. The reason this code lives in the ncurses library is to support fallback to /etc/termcap. These files include alloc_entry.c, captoinfo.c, comp_captab.c, comp_error.c, comp_hash.c, comp_parse.c, comp_scan.c, and parse_entry.c, read_termcap.c, and write_entry.c. We'll discuss these in the compiler tour.

The Engine Room

Keyboard Input

All ncurses input funnels through the function wgetch(), defined in lib_getch.c. This function is tricky; it has to poll for keyboard and mouse events and do a running match of incoming input against the set of defined special keys.

The central data structure in this module is a FIFO queue, used to match multiple-character input sequences against special-key capabilities; also to implement pushback via ungetch().

The wgetch() code distinguishes between function key sequences and the same sequences typed manually by doing a timed wait after each input character that could lead a function key sequence. If the entire sequence takes less than 1 second, it is assumed to have been generated by a function key press.

Hackers bruised by previous encounters with variant select(2) calls may find the code in lib_twait.c interesting. It deals with the problem that some BSD selects don't return a reliable time-left value. The function timed_wait() effectively simulates a System V select.

Mouse Events

If the mouse interface is active, wgetch() polls for mouse events each call, before it goes to the keyboard for input. It is up to lib_mouse.c how the polling is accomplished; it may vary for different devices.

Under xterm, however, mouse event notifications come in via the keyboard input stream. They are recognized by having the kmous capability as a prefix. This is kind of klugey, but trying to wire in recognition of a mouse key prefix without going through the function-key machinery would be just too painful, and this turns out to imply having the prefix somewhere in the function-key capabilities at terminal-type initialization.

This kluge only works because kmous isn't actually used by any historic terminal type or curses implementation we know of. Best guess is it's a relic of some forgotten experiment in-house at Bell Labs that didn't leave any traces in the publicly-distributed System V terminfo files. If System V or XPG4 ever gets serious about using it again, this kluge may have to change.

Here are some more details about mouse event handling:

The lib_mouse()code is logically split into a lower level that accepts event reports in a device-dependent format and an upper level that parses mouse gestures and filters events. The mediating data structure is a circular queue of event structures.

Functionally, the lower level's job is to pick up primitive events and put them on the circular queue. This can happen in one of two ways: either (a) _nc_mouse_event() detects a series of incoming mouse reports and queues them, or (b) code in lib_getch.c detects the kmous prefix in the keyboard input stream and calls _nc_mouse_inline to queue up a series of adjacent mouse reports.

In either case, _nc_mouse_parse() should be called after the series is accepted to parse the digested mouse reports (low-level events) into a gesture (a high-level or composite event).

Output and Screen Updating

With the single exception of character echoes during a wgetnstr() call (which simulates cooked-mode line editing in an ncurses window), the library normally does all its output at refresh time.

The main job is to go from the current state of the screen (as represented in the curscr window structure) to the desired new state (as represented in the newscr window structure), while doing as little I/O as possible.

The brains of this operation are the modules hashmap.c, hardscroll.c and lib_doupdate.c; the latter two use lib_mvcur.c. Essentially, what happens looks like this:

The hashmap.c module tries to detect vertical motion changes between the real and virtual screens. This information is represented by the oldindex members in the newscr structure. These are modified by vertical-motion and clear operations, and both are re-initialized after each update. To this change-journalling information, the hashmap code adds deductions made using a modified Heckel algorithm on hash values generated from the line contents.

The hardscroll.c module computes an optimum set of scroll, insertion, and deletion operations to make the indices match. It calls _nc_mvcur_scrolln() in lib_mvcur.c to do those motions.

Then lib_doupdate.c goes to work. Its job is to do line-by-line transformations of curscr lines to newscr lines. Its main tool is the routine mvcur() in lib_mvcur.c. This routine does cursor-movement optimization, attempting to get from given screen location A to given location B in the fewest output characters posible.

If you want to work on screen optimizations, you should use the fact that (in the trace-enabled version of the library) enabling the TRACE_TIMES trace level causes a report to be emitted after each screen update giving the elapsed time and a count of characters emitted during the update. You can use this to tell when an update optimization improves efficiency.

In the trace-enabled version of the library, it is also possible to disable and re-enable various optimizations at runtime by tweaking the variable _nc_optimize_enable. See the file include/curses.h.in for mask values, near the end.

The Forms and Menu Libraries

The forms and menu libraries should work reliably in any environment you can port ncurses to. The only portability issue anywhere in them is what flavor of regular expressions the built-in form field type TYPE_REGEXP will recognize.

The configuration code prefers the POSIX regex facility, modeled on System V's, but will settle for BSD regexps if the former isn't available.

Historical note: the panels code was written primarily to assist in porting u386mon 2.0 (comp.sources.misc v14i001-4) to systems lacking panels support; u386mon 2.10 and beyond use it. This version has been slightly cleaned up for ncurses.

A Tour of the Terminfo Compiler

The ncurses implementation of tic is rather complex internally; it has to do a trying combination of missions. This starts with the fact that, in addition to its normal duty of compiling terminfo sources into loadable terminfo binaries, it has to be able to handle termcap syntax and compile that too into terminfo entries.

The implementation therefore starts with a table-driven, dual-mode lexical analyzer (in comp_scan.c). The lexer chooses its mode (termcap or terminfo) based on the first `,' or `:' it finds in each entry. The lexer does all the work of recognizing capability names and values; the grammar above it is trivial, just "parse entries till you run out of file".

Translation of Non-use Capabilities

Translation of most things besides use capabilities is pretty straightforward. The lexical analyzer's tokenizer hands each capability name to a hash function, which drives a table lookup. The table entry yields an index which is used to look up the token type in another table, and controls interpretation of the value.

One possibly interesting aspect of the implementation is the way the compiler tables are initialized. All the tables are generated by various awk/sed/sh scripts from a master table include/Caps; these scripts actually write C initializers which are linked to the compiler. Furthermore, the hash table is generated in the same way, so it doesn't have to be generated at compiler startup time (another benefit of this organization is that the hash table can be in shareable text space).

Thus, adding a new capability is usually pretty trivial, just a matter of adding one line to the include/Caps file. We'll have more to say about this in the section on Source-Form Translation.

Use Capability Resolution

The background problem that makes tic tricky isn't the capability translation itself, it's the resolution of use capabilities. Older versions would not handle forward use references for this reason (that is, a using terminal always had to follow its use target in the source file). By doing this, they got away with a simple implementation tactic; compile everything as it blows by, then resolve uses from compiled entries.

This won't do for ncurses. The problem is that that the whole compilation process has to be embeddable in the ncurses library so that it can be called by the startup code to translate termcap entries on the fly. The embedded version can't go promiscuously writing everything it translates out to disk -- for one thing, it will typically be running with non-root permissions.

So our tic is designed to parse an entire terminfo file into a doubly-linked circular list of entry structures in-core, and then do use resolution in-memory before writing everything out. This design has other advantages: it makes forward and back use-references equally easy (so we get the latter for free), and it makes checking for name collisions before they're written out easy to do.

And this is exactly how the embedded version works. But the stand-alone user-accessible version of tic partly reverts to the historical strategy; it writes to disk (not keeping in core) any entry with no use references.

This is strictly a core-economy kluge, implemented because the terminfo master file is large enough that some core-poor systems swap like crazy when you compile it all in memory...there have been reports of this process taking three hours, rather than the twenty seconds or less typical on the author's development box.

So. The executable tic passes the entry-parser a hook that immediately writes out the referenced entry if it has no use capabilities. The compiler main loop refrains from adding the entry to the in-core list when this hook fires. If some other entry later needs to reference an entry that got written immediately, that's OK; the resolution code will fetch it off disk when it can't find it in core.

Name collisions will still be detected, just not as cleanly. The write_entry() code complains before overwriting an entry that postdates the time of tic's first call to write_entry(), Thus it will complain about overwriting entries newly made during the tic run, but not about overwriting ones that predate it.

Source-Form Translation

Another use of tic is to do source translation between various termcap and terminfo formats. There are more variants out there than you might think; the ones we know about are described in the captoinfo(1) manual page.

The translation output code (dump_entry() in ncurses/dump_entry.c) is shared with the infocmp(1) utility. It takes the same internal representation used to generate the binary form and dumps it to standard output in a specified format.

The include/Caps file has a header comment describing ways you can specify source translations for nonstandard capabilities just by altering the master table. It's possible to set up capability aliasing or tell the compiler to plain ignore a given capability without writing any C code at all.

For circumstances where you need to do algorithmic translation, there are functions in parse_entry.c called after the parse of each entry that are specifically intended to encapsulate such translations. This, for example, is where the AIX box1 capability get translated to an acsc string.

Other Utilities

The infocmp utility is just a wrapper around the same entry-dumping code used by tic for source translation. Perhaps the one interesting aspect of the code is the use of a predicate function passed in to dump_entry() to control which capabilities are dumped. This is necessary in order to handle both the ordinary De-compilation case and entry difference reporting.

The tput and clear utilities just do an entry load followed by a tputs() of a selected capability.

Style Tips for Developers

See the TO-DO file in the top-level directory of the source distribution for additions that would be particularly useful.

The prefix _nc_ should be used on library public functions that are not part of the curses API in order to prevent pollution of the application namespace. If you have to add to or modify the function prototypes in curses.h.in, read ncurses/MKlib_gen.sh first so you can avoid breaking XSI conformance. Please join the ncurses mailing list. See the INSTALL file in the top level of the distribution for details on the list.

Look for the string FIXME in source files to tag minor bugs and potential problems that could use fixing.

Don't try to auto-detect OS features in the main body of the C code. That's the job of the configuration system.

To hold down complexity, do make your code data-driven. Especially, if you can drive logic from a table filtered out of include/Caps, do it. If you find you need to augment the data in that file in order to generate the proper table, that's still preferable to ad-hoc code -- that's why the fifth field (flags) is there.

Have fun!

Porting Hints

The following notes are intended to be a first step towards DOS and Macintosh ports of the ncurses libraries.

The following library modules are `pure curses'; they operate only on the curses internal structures, do all output through other curses calls (not including tputs() and putp()) and do not call any other UNIX routines such as signal(2) or the stdio library. Thus, they should not need to be modified for single-terminal ports.

lib_addch.c lib_addstr.c lib_bkgd.c lib_box.c lib_clear.c lib_clrbot.c lib_clreol.c lib_delch.c lib_delwin.c lib_erase.c lib_inchstr.c lib_insch.c lib_insdel.c lib_insstr.c lib_keyname.c lib_move.c lib_mvwin.c lib_newwin.c lib_overlay.c lib_pad.c lib_printw.c lib_refresh.c lib_scanw.c lib_scroll.c lib_scrreg.c lib_set_term.c lib_touch.c lib_tparm.c lib_tputs.c lib_unctrl.c lib_window.c panel.c

This module is pure curses, but calls outstr():

lib_getstr.c

These modules are pure curses, except that they use tputs() and putp():

lib_beep.c lib_endwin.c lib_color.c lib_options.c lib_slk.c lib_vidattr.c

This modules assist in POSIX emulation on non-POSIX systems:

sigaction.c
signal calls
The following source files will not be needed for a single-terminal-type port.

captoinfo.c clear.c comp_captab.c comp_error.c comp_hash.c comp_main.c comp_parse.c comp_scan.c alloc_entry.c dump_entry.c parse_entry.c read_entry.c write_entry.c infocmp.c tput.c

The following modules will use open()/read()/write()/close()/lseek() on files, but no other OS calls.

lib_screen.c
used to read/write screen dumps
lib_trace.c
used to write trace data to the logfile
Modules that would have to be modified for a port start here:

The following modules are `pure curses' but contain assumptions inappropriate for a memory-mapped port.

lib_longname.c	-- assumes there may be multiple terminals
	longname()		-- return long name of terminal
lib_acs.c	-- assumes acs_map as a double indirection
	init_acs()		-- initialize acs map
lib_mvcur.c	-- assumes cursor moves have variable cost
	mvcur_init()		-- initialize
	mvcur()			-- do physical cursor move
	mvcur_wrap()		-- wrap
	scrolln()		-- do physical scrolling
lib_termcap.c	-- assumes there may be multiple terminals
	tgetent()		-- load entry
	tgetflag()		-- get boolean capability
	tgetnum()		-- get numeric capability
	tgetstr()		-- get string capability
lib_ti.c	-- assumes there may be multiple terminals
	tigetent()		-- load entry
	tigetflag()		-- get boolean capability
	tigetnum()		-- get numeric capability
	tigetstr()		-- get string capability

The following modules use UNIX-specific calls:

lib_doupdate.c	-- input checking
	doupdate()		-- repaint real screen to match virtual
	_nc_outch()		-- put out a single character
lib_getch.c	-- read()
	wgetch()		-- get single character
	wungetch()		-- push back single character
lib_initscr.c	-- getenv()
	initscr()		-- initialize curses functions
lib_newterm.c
	newterm()		-- set up new terminal screen
lib_baudrate.c
	baudrate()		-- return the baudrate
lib_kernel.c	-- various tty-manipulation and system calls
	reset_prog_mode()	-- reset ccurses-raw mode
	reset_shell_mode()	-- reset cooked mode
	erasechar()		-- return the erase char
	killchar()		-- return the kill character
	flushinp()		-- flush pending input
	savetty()		-- save tty state
	resetty()		-- reset tty to state at last savetty() 
lib_raw.c	-- various tty-manipulation calls
	raw()
	echo()
	nl()
	qiflush()
	cbreak()
	noraw()
	noecho()
	nonl()
	noqiflush()
	nocbreak()
lib_setup.c	-- various tty-manipulation calls
	use_env()
	setupterm()
lib_restart.c	-- various tty-manipulation calls
	def_shell_mode()
	def_prog_mode()
	set_curterm()
	del_curterm()
lib_tstp.c	-- signal-manipulation calls
	_nc_signal_handler()	-- enable/disable window-mode signal catching
lib_twait.c	-- gettimeofday(), select().
	usleep()		-- microsecond sleep
	_nc_timed_wait()	-- timed wait for input
The package kernel could be made smaller.


Eric S. Raymond <esr@snark.thyrsus.com>
(Note: This is not the bug address!)