Rob Swindell authored 2 months ago and Deucе committed 2 months ago

No functional change.

Rob Swindell authored 2 months ago and Deucе committed 2 months ago

Many sysops probably don't realize these keys do special stuff.

Rob Swindell authored 2 months ago and Deucе committed 2 months ago

so we can use it in nopen.c without requiring sbbsdefs.h ... smbdefs.h

This should fix the SVDM build.

Rob Swindell authored 2 months ago and Deucе committed 2 months ago

I don't know if these seek failures are actually happening or not, but
reading from or writing to the wrong offset in the node.dab file could
explain some of the node.dab corruption I'm seeing from macOS (over
SMB share).

For the SBBS binaries, set @executable_path and
@executable_path/../${LIBODIR} so they can all be in one dir, or
they can be in the build output dirs and still work.

For utilities, set to @executable_path and
@executable_path/../../${LIBODIR} for the same reason.

With this, we shouldn't need to fiddle with DYLD_LIBRARY_PATH

Remove the -E linker flag.
This should not be needed anymore, and isn't supported on macOS.

Have shared libraries include their full path.
This allows linked dylibs to work from where they were built, so
as long as you don't build the binaries on a CI machine, then try
to run them on a users machine (*cough*), it'll work out.

Use the correct rpath argument format on macOS.
It uses -rpath, not --rpath.

Rob Swindell authored 2 months ago and Deucе committed 2 months ago

Kind of a shot in the dark here: Max (WESTLINE) is reporting HEAP CORRUPTION
debug assertion in websrvr.dll. In the 2 instances reported, a long (336
char) JSON "query value" was logged by apparent spam-bot trying to create a
a new user account ("send-me-free-stuff" is one of the JSON properties).

JS_NewStringCopyZ() can return NULL in a low memory situation, though I don't
know that explains possible heap corruption.

Check and clear/invalidate node_socket while holding the node.dab record
lock.

This should fix the error reported by kk4qnb (KK4QBN)

Fix CID 516461

Fix CID 516462

Include O_CREAT access mode flag in opennodedat().

I experimented with excluding O_DENYNONE when the NM_CLOSENODEDAB flag is set
(to hopefully work-around MacOS Samba node.dab corruption issue), but that
didn't work with SBBSCTRL leaving the file open in SH_DENYNONE mode, so will
have revist that, but using a common open function helps.

Include a little more detail in 550 responses sent to clients too

This explains why Deon was still seeing time_t encoded time values.

This is a situation we can auto-correct and log a message when we do.

Could be helpful for debugging node status issues at some point

Bug

We already know such nodes are in-use, so no need to read their node.dab
record and put extra contention on the node.dab file. Hoepfully this
reduces or eliminates occurrences of the error:
    Node n status is WFC, but the node socket (s) and thread are still in use!
Though I kind of expect occurrences of "NODE STATUS FIXUP" errors to likely
return.

We could in theory just track status of nodes in memory (for those
nodes that this instance of sbbs controls), and not read the node.dab file at
all when checking those nodes' status, but:
- that would prevent out of process control of node status
  e.g. using the node utility to mark a node as offline
- we'd have to protect instance of in-memory node status checking/changing
  with a mutex

Hitting down-arrow key at a pause prompt normally displaye just one more line
of the display text/file, but if you received a node message/telegram/notice
after hitting down arrow, you'd get a screen full of text instead of just a
single (one more) line, as you wanted.

This looks to be because of the anti-recursive protection implemented in
pause() - when it calls nodesync() after the key press, that displays
node/user messages (if there are any) and if pause is called as a result
(e.g. because the line counter was already set to cause a pause after the
next line of output), it'd do nothing since that would be recursive. The fix
is to simply set the line counter as a result of the down-arrow key press
*after* the call to nodesync(), which might display multiple lines, but I
think that's fine.

Increment SMBLIB version to 3.10

Fix issue #845: Changing system/OS time zone, changes dates/times of posted
messages

Sysops and users shouldn't notice any change unless they change the time zone
of their system/OS (not accounting changes for daylight/standard time) and
the result will be that message dates appear the same after such a change.

For backward compatibily, any stored time_t's in msghdr_t.when_written.time
(i.e. all existing SMB messages) will still be decoded and displayed properly.
We detect a time_t value by the upper 6 bits being non-zero. When the upper
6 bits of a when_written.time value are zero, then we know the 'year' is
stored in the 16-bits before the when_written field (never used bits of the
netattr field, now part of the when_t structure definition) and the Month,
Day, Hour, Minute, and Second of the wallclock at the poster's site are
encoded in the low 26 bits of the time field.

This also eliminates more uses of 32-bit time_t that'll likely start being
a problem 2038 and really fall over and die in 2106. At least messages'
posting dates won't have any issue now. The "when_imported" values could use
a similar treatment someday I suppose - and we could get rid of the
when_imported.zone value as its not really needed we could use those 16-bits
for the when_imported.year.

Didn't change anything with filebases (still using time_t's though the
when_written hdr field isn't used for much with regards to files).

Yes, we could have converted all imported "broken down" message dates to
UTC and continued to store them as a time_t using timegm() instead of mktime()
for conversion to time_t, and I considered that. But we would have needed to
create/use a flag in the message header to indicate such stored date/times
(since they'd have to go through different adjustment for original time zone
before display, basically reversing the logic of all the places we display the
message dates/times using localtime verus gmtime/UTC C RTL functions),
couldn't just initialize the time with a call to time() upon import of local
messages (unless the local timezone happened to be UTC). And in the end, we'd
still have a 32-bit time_t value. So this seemed the better path.

I would have liked to have stored the date fields in a more human readable
encoding (BCD or decimal, ala isoDate and isoTime_t), but I just didn't have
the spare bits in the fixed portion of message headers to be wasteful like
that.

Here's an example from smbutil v of a message header posted after this change:
  when_written     03292595 41E0 Fri Dec 20 18:22:21 2024 PST
  when_imported    6766265D 41E0 Fri Dec 20 18:22:21 2024 PST

Notice the difference in the hex encoding of the date/time between the 2
header fields: when_imported still uses time_t. The when_written.year value
isn't output here.

When the system time zone is not UTC, but the sysop wants to display some
dates and times in UTC, they can now use these @-codes to do that:
- TIME_UTC
- DATE_UTC
- UTC:fmt
- DATETIME_UTC
- MSG_DATE_UTC

This partially reverts commit 03b84df8.

I observed deadlocks on Linux attempting locks of user.tab on Samba share,
which also deadlocked my Windows nodes. Interestingly, the Windows nodes
never deadlocked on their own (after a week of testing) when using blocking
locks.

Double the frequency of lock retries - this has helped reduce the observed
user.tab lock failures on Vertrauen.

As Deon pointed out in DOVE-Net / Synchronet Discussion, having a local time
zone configured with a different UTC offset than your system time zone can
produce strange/unexpected results (e.g. displayed age of messages). Since
it's possible that not all sysops will complete the configuration wizard or
actually set their timezone to the correct value (and ignore the startup
warning message), we now make the default Local Time Zone to be
"automatic" - query the OS every time the local time zone is needed/used.

This has the downside of only storing (e.g. in message headers) the UTC
offset of the current time zone (not the time zone abbreviation/name as
encoded by SMB). I considered making an option to dynamically figure out
the actual time zone (not just the UTC offset) and while I think that's
doable, Deon just wanted his UTC offset (e.g. UTC+11:00) and not his
time zone name (e.g. AEDT) stored in message headers, so this setting
w...

... when current timezone is negative offset from UTC, but not a US time zone,
the default should be "No".

Looking into potential causes of issue #843, I found several instances where
we call getuserdat() without checking the return value and restoring the
useron.number to the current user number upon error: getuserdat() zeroes out
the user struct/number upon error, a bad API choice made 23 years ago.

Replace those instances with calls to sbbs_t::getuseron() which logs any
open/lock/read failures of the user.tab and does not modify/zero-out the
sbbs_t::useron struct upon error.

... since all (almost all) callerd do a config check first.

Also, some callers of openuserdat() were expecting -1 on failure (always).

Functions that returned the return value of openuserdat() upon failure, now
return USER_OPEN_ERROR instead, to be consistent.

A copy/paste from websrvr.c

This needed a custom solution (not errprintf) since the filename is passed-in
is likely from dynamically allocated memory, so a pointer comparison isn't
enough - and we don't get the function name.

This is a solution for issue #842, but only for messages posted/sent from the
terminal server using built-in functions and not via JS or other means. A more
universal/generic solution would be nice (beyond just always storing message
headers and bodies in UTF-8), but nothing has come to mind.