CVS log for epic5/source/newio.c

Request diff between arbitrary revisions

* Roll in a bunch of changes from larne to make epic5 more vms-friendly
* Rename if.h -> ifcmd.h  (larne) -- Make sure you rerun configure!
* Run make depend to pick up name change of if.h
* Check for <sys/param.h> in configure (larne)
* Extern funcs are case insensitive in strict c90, fix (larne)
* Rename options -> options.h since vms requires a dot in filename (larne)
* Change "readonly" -> "rdonly" since "readonly" is a symbol in vms. (larne)
* Pull in "compat.h" for gailib.c and glob.c (larne)
* Set stdout to be explciitly line buffered at startup (larne)
* Returning a value from void functino forbidden by strict c90. fix (larne)
* Fix some compiler warnings that were lying around.
* There are more changes forthcoming, these were the easy ones.

Remove u_* bsd-isms so we can compile with -ansi on linux.

* Add new level SYSERR (see UPDATES)
* Add a server refnum to syserr() to send the syserr to the proper window.
* Add a server refnum to new_open() so we can pass it to syserr().
* Add new flag to /XECHO, /XECHO -AS (output to all windows on server)
* This should make syserrs go to the "correct window" most of the time (larne)

* Add arg to panic() controlling whether panic is sent via QUIT (rb black)

* Interrogate, and pass upwards, the socket error from getsockopt()
for nonblocking connects, and then output that error to the user.
Requested by twincest.

* Change startup code so it outputs the pid of the bot process for -b.
* Fix $info(o) bug.  rb zlonix
* Output the errno value ret by failed nonblocking server connects on error.
* Add auto-close boolean flag for server (see UPDATES)

* Fix from black for 'save' script, to save notify nicks.
* Fix from larne for configure checks for tcl for solaris (ick)
* Fix solaris ports support, inspired by patch from larne, rb zlonix

Fix some bugs with /window query that kreca found.

* Fix breakage for solaris found by zlonix
* Fix %{2}+ so it behaves like %+ and not like %{1}+, as per the docs.

Fix cut & paste error. doh.

* Add SEDSHA cipher type, which is better than SED, and always available.
* Add my_sleep|isreadable|iswritable() funcs to phase out rogue selects()h
* Add support for multiple concurrent servers to /encrypt. yay!
* Store the length of the Crypt key within the key itself.
* This moves us closer to being able to decrypt arbitrary data.
* Test all of the cipher types, incl. compat of CAST with ircII.  All set!

* Support cast5/blowfish encryption a la ircII.
* Fix build for non-ssl users.

* Support AI_ADDRCONFIG in our builtin AI_MASK/getaddrinfo().
* Make set[_non]_blocking extern.
* Support NEWIO_SSL_CONNECT as a new_open() type.
* Add a new ssl_connect() io callback function for NEWIO_SSL_CONNECT.
* New scripts from weirdo: autojoin, autoquery, floodprot, renumber
* New version of reconnect script from weirdo.
* GC the fixed-alist support (last used by /quote and protocol handler)
* Reorg server SSL negotation stuff to be nonblocking. yay!
* Comment all sorts of stuff in ssl.c.

* Add support for dgets buffer type -2 (only return full buffers)
* Use buffer type -2 for reading nonblocking dns addrs, fixes probs on solaris.

* Improve /dcc get
* Fix warnings

* Add marshalling functions for getaddrinfo()
* Use getaddrinfo marshalls + dgets to simulate nonblocking dns lookups.
* New newio io type, "NULL" which i'll probably whack later.

Add support from larne for solaris ports, which I modified up a bit, so if
there are any problems, blame me, and not him!

* Change /on leave to /on part.
* Only ever do one io event per call through select() or poll()
* Do put_echo() and not add_to_screen() in /window echo so it can be logged

* Take a new arg to new_open() to control whether INFO -- errors are output.
* Don't output the "INFO --" things for dcc or execs

Fix fd leak, observed by ce.

* Change "ERROR --" to "INFO --" to not alarm people as much.
* Change a few other warnings to not be so alarming.
* Change new_realloc() to not move ptr if it is big enough to hold new size.
* Fix all the status_* sub-formats to not have any hardcoded size limits.

* Prefer timers over i/o to prevent polling loop panics.

Don't emit spurious warnings for control-c

* Fix SSL with pthreads -- need to establish SSL before calling SSL_read!
* Fix compiling with kqueue, call it --with-multiplex=freebsd-kqueue
* Hide SSL capability behind "client_ssl_enabled()", nix all #ifdef HAVE_SSL

Fix SSL with pthreads.

* Add --with-multiplex=[select|poll|kqueue|pthread] to configure
* Check thread-safety of SSL in configure when using pthreads.
* Don't unconditionally skip leading :'s in symbol names.
* When calling functions, handle skipping leading :'s specially.
* Add support for pthread locking for ssl to ssl.c
* Add %{2}+ and %{3}+ status expandos that only show the mode string.
* Use AI_ADDRCONFIG flag to tell getaddrinfo() we only want addrs we can use.

Fix Dain Bramage.

Fix more compiler warnings.

* Refactor the error stuff by eliminating my_strerror() and dgets_errno.
* Error messages are now delivered on multiple levels as info is available.
* This may get annoying, but it's a lot more accurate now!
* Change inet_* functions to return int's since their retval is unused.
* Add a new internal "syserr()" function that generates formatted errors.
* Syserr's look like say()s but they act like yell()s. (can hook with /on yell)
* Keep an counter in the server stuff so we can count off as addrs are used.
* So the error message in /ON DCC_LOST is always just "ERROR" now.
* Fix various compiler warnings and stuff.
* Add $hookctl(GET HOOK <ref> STRING), return /load'able format of ON
* Fix some debugging outbound in switch_hostname(), doh!
* Roll set_socket_options() into a new Socket() function.
* Roll connections to AF_UNIX sockets into the mainline stuff.
* Roll nonblocking/connect/nonblocking into Connect().
* Add AF_UNIX support to Getnameinfo() to remove annoying errors.
* When doing recovery from EBADF in select, only check fd's we're using
* Just puke if configured for kqueue() and kqueue() fails.
* Don't puke on default 353 output if the userlist arg is missing. (for ratbox)
* Permit doing /window name to change case of window's name.

* Implement a proper built in /INVITE command.  See UPDATES
* Revert 2.8script change.
* Add "subpackage" command, which i'll document later.
* Package up 'global', and '2.8script' and 'basical' so they can be /unload'ed.
* Don't load "global" if user has their own startup script.
* Don't load "global" if the -q option is set, either.
* Thus, the "global" script acts as a default "epicrc" for those without one.
* THIS IS BACKWARDS INCOMPATABLE.  SEE UPDATES FOR IMPORTANT INFORMATION.
* Nix the "loading_global" variable, only used by hooks.
* Fix /on 005 so it shows the /on 005's and not all numeric hooks.
* If Accept() fails, tell the user straight-way so I can debug this.
* Fix /window name - so it redraws the status bar.

* Fix accept() with pthreads.  Nonblocking + pthreads == Bad.
* Throw away all unused addrs when we are registered, to ensure no reconnects.

Try to auto-repair select(2) failures.

Fix more bugs.

Remove a panic that broke wait().

Fix a minor brain-o.

* Refactor $sar() and $msar() so they share as much code as possible.
* Refactor newio, support select/kqueue/poll (#define's in newio.h)
* ... Getting closer to not needing select!
* Fix SSL.
* Fix a bunch of misc compiler warnings
* Remove restrictions on /QUOTE, so there!
* Add pthread support to newio for larne. HUZZAH!
* I'm sure a zillion things are broken, so it's time to test!

* Add a dgets_buffer() to inject data into dgets() system.
* Fix up dgets() so it has read() like semantics.
* Fix up dgets() callers to use it like a read() type function.
* Turn off /xdebug server_connected hardcoded at startup for now.
* Fix up unix_* io functions to use dgets_buffer().
* Fix do_filedesc() to repeatedly call callback until buffer is clean
* Move set_socket_options() to network.c, make do_filedesc() global func

Fix minor bogon (maybe)

PHEAR the breakage!
* Major refactoring of newio system:
* Change main io() look into a "fill" cycle and a "drain" cycle, using buffers
* Change new_open() to take an io policy to perform when fd is "ready"
* Use token to determine if an fd is "dirty" or "clean" (is data ready?)
* Process accept()s through newio/dgets.
* Merge nonblocking connects as just another type of new_open() policy.
* Nonblocking connects emit sockaddr info when connected for dgets().
* Change dgets() to not perform io, but just drain from the buffer
* Now all data goes through dgets()! huzzah!
* Nuke kqueue support for the moment (will come back later)
* Hide the "holding" status of a fd inside the IO metadata.
* This allows me to support "holding" in other IO models.
* Major refactoring of ssl support:
* Move all meta-data about ssl in ssl.c in a linked list
* As much as possible, have server treat the ssl case as "normal"
* Implement all ssl interfaces even in non-ssl case, guard with panics
^^^^ None of the above should be user-visible changes.

* Fix /dmsg =<numeric> again, fix compiler warnings.
* Whack dgets() into submission regarding incomplete unbuffered lines.
* Revamp how partial lines from an /exec process are handled.

* Refactor window_check_servers() to work with server's current window only.
* Roll forward all the rest of changes from epic4 (i think)
* Add /ON NEW_NICKNAME, when epic needs a new nickname from you.
* Run 'make depend', fix compile errors, fix 'environment' script

Fix a defect where, if two fd's were ready at the same time (say, dcc chats)
and in the /on for the first one you new_close() the second one, epic would
panic trying to process the 2nd connection ("got a callback, not set up").
Now we only do that panic if the 2nd connection looks like it's still valid,
which it definitely won't be after a new_close().

This bug was discovered by ce.

Fix bugorz.

Ick.  Hack select__do_wait() to prefer timers over i/o.   This is a nasty
hack, but i think hsoc has suffered enough already.

Add some help for polling loop detections.

Expound a bit more in newio.c about the changes and purpose.

* Abstract the reading function out of dgets() via an argument
* Add unix_read(), the default reader for dgets()
* Add ssl_reader(), a dgets-like reader for SSL connections
* Add server_ssl_reader(), a dgets reader for servers using SSL.
* I probably broke SSL servers by doing this -- report bugs please!

* Abstract fd handling in newio.[ch] through #defines
* Add fd handling with freebsd's kqueue().  Huzzah! (not turned on by default)
* Refactor how dgets() works with incomplete SSL reads, slurp up everything
* Nix the bletcherous do_server(-1) hack in io().
* Fix a few more bogons in do_server().

* Hide new_select() behind wait_select(), which will hold the result fd_set's
* Don't make do_filedesc() take fd_set's, but rather get them from wait_select.
* Be cautious and when redrawing one screen, redraw all screens.
* Refactor io() some more; no explicit dependancy on select() now
* Refactor io(), optimize placement of some of the post-select operations.
* When wserv changes size, recalculate windows, and allow io() to redraw it.

* Make new_open[_for_write] take a callback function, stash it in io_rec.
* Add handle_filedescs() which takes select() results and calls callbacks
* Make io() call handle_filedescs() instead of the four callback functions.
* This is the first stage of an extendable callback event looper!  huzzah!

* Allow client_connect() to be nonblocking
* Allow Getaddrinfo() and Freeaddrinfo() to be called extern.
* Track writable sockets in 'global_max_fd' too.
* Keep a list of AI's for server addresses in the server code
* When server connection lost before registration, use the next AI
* Implement set_server_status(), which controls where server is in lifetime
* Nix reconnect(), which is replaced by close_server + set_server_status().
* Nix server_reconnects_to(), we no longer reconnect to servers.
* Make /xdebug server_connect the default for this commit at least
* Call window_check_servers() every time through io(), because...
* Make window_check_servers() look for disconnected servers and restart them.
* Add add_mode_to_str() which adds a channel or usermode to a string (sorted)
* Make usermodes use add_mode_to_str()
* Make channel modes use add_mode_to_str()
* Now modes are not stored in bitmasks, not hard-coded at compile time. HUZZAH!
* Nix "inactive" field in channels, channels are thrown away at disconnect
* Don't allow channels to move between servers, natch.
* Nix all the code to handle disconnected channels and such.
* Use change_window_server() to move windows to new servers, because...
* Make window_check_servers() auto-close windowless servers.
* Make change_window_server(<server>, NOSERV) cause a disconnect from <server>
* Don't honor /set auto_reconnect, we don't ever reconnect.
* Fix printf() error in /on kick default output
* Add new server state, SERVER_RECONNECT which means "i can be connected"
* So now SERVER_CLOSED means "don't reconnect me"
* Allow /server, /window server, and /reconnect to change CLOSED -> RECONNECT
* Make server connections nonblocking.
* Server connections are now nonrecursive, nonblocking, multi-protocol. HUZZAH!
* GC a bunch of server code dealing with reconnection we don't use now.
* Don't call window_check_servers() other than io() for sanity reasons.
* Make /window server just change server, let w_c_s() bootstrap connection
* Remove many window->channel->server sanity checks.

* The last argument to open_file_for_write() needs to be lowercase.
* Add $numlines() from fudd.
EPIC4-1.2.3!

* Rename 'dcc()' to 'dcc_cmd()' so we can use dcc as variable
* Garbage collect unused 'irc_path' global variable.
* Add --with-warns to configure to do FreeBSD "WARNS" like compiler checks
* Many hundreds of changes to improve the "quality of code" in epic, including:
* - Treat all literal strings as (const char *) and fix const correctness.
* - Mopping up all remaining const-correctness issues
* - Do not "launder" a (const char *) to a (char *) in string searches, (ala
    strchr) but rather return a ssize_t offset value.
* - Particular, MatchingBracket() returns (ssize_t) and not (char *).
* - Eliminate all "shadow" variable names by changing them to something else.
* - Eliminate most function decls that do not include a prototype list.
* - Add prototypes for all extern functions that are missing them.
* - Ensure most function pointers include a prototype list.
* - In term.c, most of those (char *)'s are really (const char *)'s.
* - Explicitly specify 'static' functions as 'static'
* - Ensure every function is either 'static' or has a prototype.
* - Eliminate (unsigned) < (signed) type comparisons.
* Hopefully you should not notice any changes!

Nominally remove all uses of "unsafe" functions: (Commit 506)
* Convert all uses of strcpy() to strlcpy()
* Convert all uses of strcat() to strlcat()
* Convert all uses of strmcpy() to strlcpy()
* Convert all uses of strmcat() to strlcat()
* Convert all uses of sprintf() to snprintf()
* As much as possible, use 'sizeof var' in snprintf/strlcpy/strlcat.
* Qualify the size of some global vars so we can sizeof() them.
* Nuke strmcpy(), strmcat(), strmcat_ue(), strmccat(),
* Change strmopencat() to strlopencat().
* Change strmpcat() to strlpcat().
* Fix switch_hostname so it doesn't return a (const) string.
* Fix TimerTimeout so it doesn't return a (const) Timeval.
* Change things like (ret) to do ret while (0); for when "ret" is "{}".
* Fix initialization for load_level
* Stop passing in static strings into io(), so we can modify them.
* Make messages passed to io() more informative to the user.
* Fix CTCP FINGER handler not to assume (struct pw) strings can be overwritten.
* Use (socklen_t) and not an (int) for socket sizes.
* Don't return a value in a void function...
* Don't use return value of process_dcc_chat_ctcps() for assignment.
* Don't put semicolons after function bodies
* Fix $getopt() to save 'input_size' so we can use it with strl*().
* Convert all uses of stpcpy() to strlcat() and strlopencat().
* Nix checks for stpcpy() in configure -- regen configure.

Another merge.

* Reorganize param list to dgets() to include a length argument
* Add (set|get)_server_line_length(), the maximum size of an irc protocol line.
* Convert dgets() callers to the new calling style.

* $dccctl() changes as per UPDATES.
* Misc bug fixes as per KNOWNBUGS.

Yikes.  Merged to become consistent with HEAD.  Sorry for all the commits,
I'll be quiet again soon. :)
-wd

* CTCP-over-DCC fix.  There are some things that confuse me about the inner
  workings here, so I'm not sure if this fix was done properly, but I did
  a few tests including a DCC send over a SED encrypted DCC and everything
  seems to tick over well.
* Unix socket configure tests as per KNOWNBUGS.
* /set flood_ignore defaulted to off.
* Tuned commandqueues.
* The 324 hook is now called for joins as well.
* %{1}D status setting as per UPDATES.
* DCC hold mode as per UPDATES.

Fix all the rest of the compile issues. bleh.

Fix build error on non-ssl systems.

* Remove colon at end of RETURN_MSTR so it is like others.
* Unify dgets() and SSL_dgets()
* Add a 4th argument to dgets(), an (SSL *) for ssl reads, NULL for normals
* Integrate SSL stuff more tightly by removing unecessary #ifdef's
* Change 'enable_ssl' to 'try_ssl' to avoid confusion with 'ssl_enabled'
* Remove uneccesary members of (Window)
* #include "functions.h" in any file implementing built in functions.
* panic if ssl op attempted on non-ssl connection.
* Don't call functions in RETURN_* macros. (they're evaled twice)
* Nuke #ifdef 0'd uses of 'ceiling_of_display'

* Use server_list_size() and not number_of_servers outside of server.c
* Use NOSERV and not -1 for "no server"
* Use (Timeval) and not (struct timeval)
* Fix (get|set)_server_sent_*() et al to take server refnum.
* Use parse_server_index() and not my_atol() to parse server refnums
* re-indent quotecmd
* Rename is_server_connected() to is_server_registered
* Make /quote work as long as server is open (not registered)
* Use send_to_aserver() when it can avoid setting from_server unecessarily
* Change sent_to_server and clear_sent_to_server to (set|get)server_sent().
* Change doing_* to get_server_doing_*.
* Use (unsigned short) and not (u_short)
* Use (unsigned long) and not (u_long)
* Make output_reject_ctcp() take a refnum
* Lock a dcc while we are connecting it, just in case. ;-)
* Fix /dcc send -p <port>
* Store protocol enforcements as (get|set)_server_protocol_state().
* Use (Stat) and not (struct stat)
* Rewrite $timerctl() so it works like $logctl()
* Rename is_server_connected() to is_server_registered()
* Fix an off-by-one error with /FOR ... IN ...
* Keep track of both readable and writable fd's with newio stuff.
* Convert server_list to an array of pointers
* When user /server -delete's, leave a gap, don't change refnums around.
* Nix code that handles deleting server refnums.
* New function, new_open_for_writing(), for when we do nonblocking connects.
* Change new_open() to deregister fd as "writable"
* Renovate the notify stuff to take refnums and not just use from_server.
* Renovate /WHO, /ISON, /USERHOST to take refnums and not just use from_server.
* Make everything that uses server_list[] use get_server() instead.
* Make everything capable of dealing with "gaps" in server_list[].
* Use a bunch of macros, to handle with getting and setting server attrs.
* Make as much as possible use refnums as params and not from_server.
* And a lot of other stuff...

* Change do_input_timeouts() to a timer callback.
* Nuke the now unused global /wait stuff
* Revamp timer.c so you can add/change/delete/update timers during a timer
* Change add_timer() so it takes a winref not a window pointer.
* Make the "key timeout" thing happen on demand instead of a recurring timer
* Remove the recurring "key timeout" timer.
* Remove the 'key timeout' thing from io(), also 'first time' stuff.
* Call get_time() after select() in io().
* Fix two compile bugs found by larne.
* Fix bug when you did /flush on a /clear window.

Update copyright notices and copyright dates.

No functional changes, just removal of unecessary whitespace at the end
of lines.  From adm.

Fix new_select() to not expose a file descriptor with ready (cached) data
unless the caller said they were interested in that file descriptor.

Applied the SSL patch from ay-ar.

Details are in README.SSL.  This works well at this point, but there may
be teething problems.

Actually test for polls correctly in the sub-second-resolution world;
second == 0 and usecond == 0.

Initial import into CVS

Initial revision