There are two underlying ideas here: (1) Keeping signals queued around
anything that's doing memory management (including push/pop of the heap)
has become crucial. (2) Anytime the shell is going to run a command, be
it buitin or external, it must be both safe and necessary to process any
queued signals, so that the apparent order of signal arrival and command
execution is preserved.
Garbage input (nul bytes, etc.) can cause the $(...) parser to become
confused during look-ahead and attempt to back up the input too far.
This commit catches the error but does not fix the underlying cause.
Combination of 12 commits from interrupt_abort branch.
Basic strategy is to introduce bits to errflag and to set and
reset them separately.
Remove interrupt status on return to main keymap.
Turn off ERRFLAG_INT for always block.
Restore bit thereafter: we probably need a new variable in order
to allow user interrupts to be reset in the always block.
Add TRY_BLOCK_INTERRUPT
This works the same as TRY_BLOCK_ERROR, but for a SIGINT, too.
Ensure propagation of SIGINT from exited job.
If received by foreground job, shell uses ERRFLAG_INT, not
ERRFLAG_ERROR, to set the new state.
Reset errflag before precmd()
Add always block in _main_completion to fix ZLS_COLORS
Ensures we get the right state of $ZLS_COLORS at the end of _main_complete
even if there's an interrupt. However, the "right state" is a bit messy
as it depends on styles.
Gets rid of all the remaining "ignoring return value" compiler
warnings, and makes some read/write operations safer by ensuring
that an EINTR is handled.