There are six (normally) asynchronous actions about which vi cares:
SIGHUP, SIGINT, SIGQUIT, SIGTERM, SIGTSTP and SIGWINCH.
The assumptions:
1: The DB routines are not reentrant.
2: The curses routines may not be reentrant.
3: Neither DB nor curses will restart system calls.
XXX
Note, most C library functions don't restart system calls. So, we should
*probably* start blocking around any imported function that we don't know
doesn't make a system call. This is going to be a genuine annoyance...
SIGHUP, SIGTERM
Used for file recovery. The DB routines can't be reentered, nor
can they handle interrupted system calls, so the vi routines that
call DB block signals. This means that DB routines could be
called at interrupt time, if necessary.
SIGQUIT
Disabled by the signal initialization routines. Historically, ^\
switched vi into ex mode, and we continue that practice.
SIGWINCH:
The interrupt routine sets a global bit which is checked by the
key-read routine, so there are no reentrancy issues. This means
that the screen will not resize until vi runs out of keys, but
that doesn't seem like a problem.
SIGINT and SIGTSTP are a much more difficult issue to resolve. Vi has
to permit the user to interrupt long-running operations. Generally, a
search, substitution or read/write is done on a large file, or, the user
creates a key mapping with an infinite loop. This problem will become
worse as more complex semantics are added to vi, especially things like
making it a pure text widget. There are four major solutions on the table,
each of which have minor permutations.
1: Run in raw mode.
The up side is that there's no asynchronous behavior to worry about,
and obviously no reentrancy problems. The down side is that it's easy
to misinterpret characters (e.g. :w big_file^Mi^V^C is going to look
like an interrupt) and it's easy to get into places where we won't see
interrupt characters (e.g. ":map a ixx^[hxxaXXX" infinitely loops in
historic implementations of vi). Periodically reading the terminal
input buffer might solve the latter problem, but it's not going to be
pretty.
Also, we're going to be checking for ^C's and ^Z's both, all over
the place -- I hate to litter the source code with that. For example,
the historic version of vi didn't permit you to suspend the screen if
you were on the colon command line. This isn't right. ^Z isn't a vi
command, it's a terminal event. (Dammit.)
2: Run in cbreak mode. There are two problems in this area. First, the
current curses implementations (both System V and Berkeley) don't give
you clean cbreak modes. For example, the IEXTEN bit is left on, turning
on DISCARD and LNEXT. To clarify, what vi WANTS is 8-bit clean, with
the exception that flow control and signals are turned on, and curses
cbreak mode doesn't give you this.
We can either set raw mode and twiddle the tty, or cbreak mode and
twiddle the tty. I chose to use raw mode, on the grounds that raw
mode is better defined and I'm less likely to be surprised by a curses
implementation down the road. The twiddling consists of setting ISIG,
IXON/IXOFF, and disabling some of the interrupt characters (see the
comments in cl_init.c). This is all found in historic System V (SVID
3) and POSIX 1003.1-1992, so it should be fairly portable.
The second problem is that vi permits you to enter literal signal
characters, e.g. ^V^C. There are two possible solutions. First, you
can turn off signals when you get a ^V, but that means that a network
packet containing ^V and ^C will lose, since the ^C may take effect
before vi reads the ^V. (This is particularly problematic if you're
talking over a protocol that recognizes signals locally and sends OOB
packets when it sees them.) Second, you can turn the ^C into a literal
character in vi, but that means that there's a race between entering
^V<character>^C, i.e. the sequence may end up being ^V^C<character>.
Also, the second solution doesn't work for flow control characters, as
they aren't delivered to the program as signals.
Generally, this is what historic vi did. (It didn't have the curses
problems because it didn't use curses.) It entered signals following
^V characters into the input stream, (which is why there's no way to
enter a literal flow control character).
3: Run in mostly raw mode; turn signals on when doing an operation the
user might want to interrupt, but leave them off most of the time.
This works well for things like file reads and writes. This doesn't
work well for trying to detect infinite maps. The problem is that
you can write the code so that you don't have to turn on interrupts
per keystroke, but the code isn't pretty and it's hard to make sure
that an optimization doesn't cover up an infinite loop. This also
requires interaction or state between the vi parser and the key
reading routines, as an infinite loop may still be returning keys
to the parser.
Also, if the user inserts an interrupt into the tty queue while the
interrupts are turned off, the key won't be treated as an interrupt,
and requiring the user to pound the keyboard to catch an interrupt
window is nasty.
4: Run in mostly raw mode, leaving signals on all of the time. Done
by setting raw mode, and twiddling the tty's termios ISIG bit.
This works well for the interrupt cases, because the code only has
to check to see if the interrupt flag has been set, and can otherwise
ignore signals. It's also less likely that we'll miss a case, and we
don't have to worry about synchronizing between the vi parser and the
key read routines.
The down side is that we have to turn signals off if the user wants
to enter a literal character (e.g. ^V^C). If the user enters the
combination fast enough, or as part of a single network packet,
the text input routines will treat it as a signal instead of as a
literal character. To some extent, we have this problem already,
since we turn off flow control so that the user can enter literal
XON/XOFF characters.
This is probably the easiest to code, and provides the smoothest
programming interface.
There are a couple of other problems to consider.
First, System V's curses doesn't handle SIGTSTP correctly. If you use the
newterm() interface, the TSTP signal will leave you in raw mode, and the
final endwin() will leave you in the correct shell mode. If you use the
initscr() interface, the TSTP signal will return you to the correct shell
mode, but the final endwin() will leave you in raw mode. There you have
it: proof that drug testing is not making any significant headway in the
computer industry. The 4BSD curses is deficient in that it does not have
an interface to the terminal keypad. So, regardless, we have to do our
own SIGTSTP handling.
The problem with this is that if we do our own SIGTSTP handling, in either
models #3 or #4, we're going to have to call curses routines at interrupt
time, which means that we might be reentering curses, which is something we
don't want to do.
Second, SIGTSTP has its own little problems. It's broadcast to the entire
process group, not sent to a single process. The scenario goes something
like this: the shell execs the mail program, which execs vi. The user hits
^Z, and all three programs get the signal, in some random order. The mail
program goes to sleep immediately (since it probably didn't have a SIGTSTP
handler in place). The shell gets a SIGCHLD, does a wait, and finds out
that the only child in its foreground process group (of which it's aware)
is asleep. It then optionally resets the terminal (because the modes aren't
how it left them), and starts prompting the user for input. The problem is
that somewhere in the middle of all of this, vi is resetting the terminal,
and getting ready to send a SIGTSTP to the process group in order to put
itself to sleep. There's a solution to all of this: when vi starts, it puts
itself into its own process group, and then only it (and possible child
processes) receive the SIGTSTP. This permits it to clean up the terminal
and switch back to the original process group, where it sends that process
group a SIGTSTP, putting everyone to sleep and waking the shell.
Third, handing SIGTSTP asynchronously is further complicated by the child
processes vi may fork off. If vi calls ex, ex resets the terminal and
starts running some filter, and SIGTSTP stops them both, vi has to know
when it restarts that it can't repaint the screen until ex's child has
finished running. This is solveable, but it's annoying.
Well, somebody had to make a decision, and this is the way it's going to be
(unless I get talked out of it). SIGINT is handled asynchronously, so
that we can pretty much guarantee that the user can interrupt any operation
at any time. SIGTSTP is handled synchronously, so that we don't have to
reenter curses and so that we don't have to play the process group games.
^Z is recognized in the standard text input and command modes. (^Z should
also be recognized during operations that may potentially take a long time.
The simplest solution is probably to twiddle the tty, install a handler for
SIGTSTP, and then restore normal tty modes when the operation is complete.)
110d525ec6