save_currentstate() increments program_state.in_checkpoint and then
returns early without decrementing it when currentlevel_rewrite()
fails (full disk, quota, unwritable directory). The guard at the top
of the function suppresses every later checkpoint for the rest of the
game, leaving recover with nothing newer than the last checkpoint that
did get written, and nothing says so.
savestateinlock() returns early unless
program_state.something_worth_saving is set, but both call sites that
are meant to lay down the initial checkpoint run before that flag is
set: newgame() calls save_currentstate() two lines early, and
dorecover() calls savestateinlock() 73 lines before it, ahead of the
pass that writes out the level files.
Both calls are therefore no-ops, and the <uid><plname>.0 lock file
holds nothing but the pid written by getlock() until the hero first
changes dungeon level. A game that dies without a chance to save
during that window (SIGKILL, OOM killer, watchdog, host reboot)
cannot be rebuilt: recover has no save file name, no current level
number and no game state, and reports "Checkpointing was not in
effect", which is true of the outcome but misleading about the cause.
The window covers the whole of a long stay on one level, and in
particular the entire period right after a restore.
Set the flag before the call in newgame(). In dorecover(), drop the
ineffective call and checkpoint once the restore is complete instead;
at the original spot no level file for this session has been written
yet, so a checkpoint there would name a current level and save file
that are not on disk. The new call goes after program_state.restoring
is cleared, so that stairs and traps are written with the same
relative dlevel encoding a normal save uses, and it is
save_currentstate() rather than savestateinlock() so that
in_checkpoint is set while savegamestate() writes u.ustuck_mid and
u.usteed_mid.
This arrived with the something_worth_saving guard; 3.4.3, which has
no guard, is unaffected. Looks like this issue has been around since
version 3.6.0.
Only quantum mechanics are supposed to have a chance of death-dropping the
Schroedinger's cat box.
Slash'Em already had this but it was missed when Genetic engineers were ported
over.
k21971