Under --two-process the compiled program is a child of the daemon, and the
daemon is the only thing that ever ends it: the program is a loop with no
reason of its own to stop, and lib/dev.ml kills it from a Fun.protect finally
block on the way out. That block does not run when the daemon is SIGKILLed —
by a harness tearing down a daemon it gave up on, by a watchdog, by anybody —
and what was left behind was a program with ppid 1, sleeping, holding a socket
nothing will ever connect to. Eight of those were on the machine this was
written on, the oldest six days old, and one had been minted by this suite.
The child now arms PR_SET_PDEATHSIG when it starts its agent, gated on a
FLAN_DEV_PARENT the daemon sets and nothing else does, and closes the race
where the daemon died before the arming by comparing getppid against that pid
rather than against 1 — under a subreaper an orphan is not reparented to init.
The signal is SIGPWR and not SIGTERM, because the daemon's ordinary teardown
already sends the child SIGTERM and the child's stderr is the daemon's own; a
handler there would print "the daemon is gone" into the daemon's output on
every clean close. The handler writes one sentence, unlinks the agent socket
it bound, and exits 0, and it does nothing else because it can run on the game
thread inside stdio's lock.
A merged build arms none of this and must not. There the daemon is the
process, so a dead daemon is a dead program; its parent is whoever typed
flan dev, which is a shell or an emacs and not the session's owner. Merged
daemons do leak on this machine — four of the eight — but for a different
reason, which is that accept_loop has no idea whether its editor still exists.
That wants a mechanism that tracks the client rather than the parent, and it
is not this change.
The daemon's read end of the program's stdout pipe is also cloexec now. The
child used to inherit it, which is both a leak and the reason waiting for EOF
on that pipe could never have been the mechanism: the pipe always had a
reader. A printing program now takes SIGPIPE once its daemon is gone, which is
a second net under chatty programs and not a substitute — the test uses
dev-watch, which is silent, so that only the thing under test can end it.
The test spawns the two-process shape, sends one request so that what it kills
is demonstrably serving, reads the program's pid out of the daemon's
/proc children before killing it — it is a grandchild, so waitpid is ECHILD —
and SIGKILLs the daemon. SIGKILL and not SIGTERM: SIGTERM lets the finally
block run and the assertion would be green with this change reverted. It fails
that way, and passes with it, in well under a second.
Six refusals in the runtime called _exit(134) where every other error had
learned to park: no restart by that name, a restart taken with the wrong
arguments or with none, a defer that invoked one, a null allocator, and
free-all on something with no region. Under a merged flan dev the compiler is
in that process, so a program that named a restart nobody established took the
session down with it, which is the one thing the break loop exists to prevent.
They park now. Not through flan_break_hook, which is what bounds and
arithmetic use: that hook may answer by aiming a transfer channel, and these
six are called by emitted code that falls off the end with no channel anywhere
in the call, so a restart chosen against one would be accepted and dropped.
flan_trap_hook says the other thing instead — stop here, let everything be
read, and refuse the resume with a reason.
All six park, for two reasons rather than one. Four are guards that fire
before the operation they guard, so nothing is half done and the frame reads
like any other. The other two fire mid-transfer, with the frame's defers
possibly half run, and they park only to be looked at: stopping on a torn
unwind is strictly more than exiting before anyone can ask what tore it.
The break loop grew a per-snapshot resumable flag for it. Restarts are still
listed and still numbered, the terminal marks them untakeable and the socket
reports the same positions as unreachable, and the listener refuses a choice
with the trap's own sentence rather than the thunk-boundary one.
Standalone builds die exactly as they did: nothing installs the hook in a
program that did not import the agent, and the acceptance case for free-all
still wants exit 134 and the same message.
The review entry that asked for this named flan_exit_hook, which is normal
termination and not this at all; it is struck out with the correction.
dev-chatty.flan outlives the surveys' twenty seconds by design, and unlike
dev-repl it prints while it does -- so the two backends stop at different
lines and the diff reports on scheduling rather than on lowering. It joins
dev-loop and dev-watch in the excluded-by-name list in both sweeps, with the
distinction written down.
rt_flush_out is guarded on __wasm__: the pipe it is careful about belongs to a
merged flan dev, which is only ever a native host, and wasm32 need not answer
for a descriptor mode its runtime may model differently.
And three comments that went false with the _exit: the atexit registration in
the merged entry point is no longer there for rt_die, which unlinks the socket
for itself now, so both places that said so say what it is actually left
covering.
bounds.flan covers this already and the x86 survey cannot see it: the
program picks its case out of (at args 1) and survey.sh runs everything
with no arguments, so re-gating x86's lo <= hi would have failed
nothing. This probe reaches the reversed slice on its own, through
(len args) so that neither optimiser can fold the branch and the checker
has no literal to object to. It matches under the default sweep and
under SURVEY_FLAGS=--no-bounds-checks, which is the claim.
Also records what the same reading turned up and did not fix: x86 still
reports a negative slice-from-ptr promise through flan_slice_error, so
the two backends print different sentences for it, and no corpus program
reaches that case without arguments.
Retrying immediately looked like eight chances and was one: a walk that bails
at the epoch check costs almost nothing, so all eight fit inside the single
compaction they were all losing to, and the listing refused having waited for
nothing. A quarter of a millisecond between attempts -- the agent's break-loop
idiom, legal here because the waiter is the listener thread and never the game
loop -- bounds the whole refusal at two milliseconds. Measured with a writer
noting and freeing on top of three thousand live blocks: 8 right answers in 200
without the pause, 200 in 200 with it. It is not magic, and the comment says
so: a writer that spends most of its time rearranging the table still gets
refused, which is the honest answer and used to be a zero-row lie.
The two cases the last commit left unwired are wired now, and a third joins
them: a listing taken while the table really is being compacted, which nothing
covered -- the full-of-live case never compacts and the churn case is
single-threaded, so the retry itself was exercised by nothing. It asserts only
what a slower machine cannot change: never zero rows, never a count that is
neither right nor a refusal. How the rest divides is printed, not pinned.
It ended "send it again once the program is running", which reads as a
withdrawal of what eval promises a parked program on every other evaluation:
that a body redefined while parked installs when the program is run again.
What is true is narrower — this module was not taken, and the run that drains
the sixty-four already queued is what makes room for it.
The three named arms are not every way a step after the check can fail.
write_file makes the daemon's copy of the module's text before llc is called at
all, and a Sys_error from it went out through serve's guard with the session
already holding the declaration — the same stranded declaration under a
different exception. Dev.eval now restores on the way past anything that
escapes, and carries an accepted flag so that a module the agent has already
taken is never rolled back out from under the process.
That is also the seam the crash can be reproduced through. test_dev takes the
daemon's working directory away from it — its own, not the macro cache the two
failures above use, which cannot fail a build with no macro in it — evaluates a
defn, and then calls it. Without the restore the daemon lists probe-two in
describe and the call closes the socket: the null cell, the game thread, and
address 0.
Two seconds was measured on one warm machine and has no margin against the
6.8s cold build listening already records. The status and the absence of the
frame-boundary sentence are what discriminate; this bound exists only so a
timeout cannot be read as an answer, so it moves to 4.5 and says so.
The unreadable-frame case wrote both frames as one string, where `(:a14' reads
as a typo rather than as the end of one frame and the start of the next; they
are two inserts now, with a line saying which is which. And the check that a
request after a daemon restart does not wait out a reply the old connection
owed was asserting against the same number it had bound the timeout to -- the
one arrangement that cannot tell "fast" from "timed out a moment ago". The
timeout is ten seconds and the assertion is three, which is the gap the claim
is actually about.
(/ 0.0 0.0) printed nan through LLVM, which folds it at compile time to
the positive quiet NaN, and -nan through x86, where divsd computes the
negative one. Put the operands in globals so nothing folds and both say
-nan, so the divergence is the folding path and not the arithmetic.
The sign bit of a NaN is not a property of the number and IEEE 754 does
not specify it, so the print site is where this is answered.
flan_f64_to_bytes renders any NaN as nan, and the two dev emitters do
the same. That is not a new rule: format-f64 in the prelude has always
answered nan for this value, so a build where (print x) said -nan and
(show x 2) said nan was contradicting itself inside one backend. An
infinity still prints signed.
format.flan prints the three non-finite values through print as well as
through show. It is in the survey corpus, so the one program pins the
printed form under dune test and the agreement between backends under
the survey.
The registry compacted whenever the table was three quarters full, and a
compaction reclaims dead entries and nothing else. A program holding more than
three quarters of the table in live blocks therefore compacted on every
allocation for the rest of its life, reclaiming nothing each time and holding
the table-wide epoch odd while it did. A listing racing that loop lost all
eight of its attempts and answered with zero rows -- "nothing is held", about a
program holding three thousand blocks, from the verb that exists to find a
leak. Measured at 199 wrong answers in 200.
The trigger now also asks whether there is an eighth of a table's worth of dead
to reclaim, which is a count four places maintain: a death, an arena's
free-all, a note written over a dead slot, and the sweep itself. That bounds
the cost from the other side too, since a sweep that runs reclaims at least 512
slots and so cannot run twice in 512 allocations.
Separately, flan_dev_reg_by_type answered a walk it could not take with zero
rows, which is the same number a program that had freed everything gets, and
stepped past slots flan_reg_snap could not copy while still calling the walk
whole. It now counts those slots and returns -1 with the count, the agent
refuses in a sentence the daemon already renders, and the snap contract says
which caller keeps it and why reg_at is allowed not to.
A note that finds no slot is still dropped -- dying because a diagnostic ran out
of room would be the diagnostic shooting the patient -- and now says so on
stderr once, quoting how many entries were dead rather than claiming the table
is all live.
test/dev_limits.c gains three modes, driven from test_reload: 3100 live blocks
read under a writer thread (1 right in 200 before, 200 after), 3000 live with
600 churned on top of them to prove the sweep still runs, and a genuinely full
table that must say so exactly once.
The lo <= hi test in check_slice and slice-from-ptr's n >= 0 sat behind
--no-bounds-checks in both backends, while the comment beside each said
they could not be dropped. They are not bounds checks: hi <= len asks
whether a range fits inside a length, and lo <= hi asks whether the word
about to be written into a %slice's length field is a count at all. The
first stays behind the flag, the second is now emitted everywhere, the
way flan_vec_as_slice has always validated its own l > h in plain C.
emit.ml emits two signal blocks rather than one and i1, so an unchecked
build carries one compare. x86.ml keeps all three frame temporaries
stored outside the flag and gates only the second compare, because the
third is the length the message prints.
The IR assertion in test_acceptance now says the two slice calls are
present under --no-bounds-checks rather than absent, and the same build
is run: case 2 and case -2 of bounds.flan must still die.
rt_die began with fflush(stdout) and ended with exit(134), and under a merged
flan dev both could hang. The flush waits on a 64K pipe whose reader is busy;
exit runs the atexit chain and the ELF destructors, which want the loader lock
a dlopening listener thread may be holding, and in that build the chain also
holds OCaml's shutdown. A bounds failure that hangs is worse than one that
kills the process, because nothing is printed and nothing exits.
So the flush every trap does first goes through rt_flush_out, which puts fd 1
into non-blocking mode and accepts a truncated tail, and rt_die takes _exit
and unlinks FLAN_DEV_SOCK by hand -- the same shape as die_now in
flan_agent.c, which argued all of this first and is quoted rather than
diverged from. On a terminal, a file, or a pipe with room, O_NONBLOCK changes
nothing, and the acceptance corpus diffs that output.
One consequence worth knowing: _exit skips ASan's exit-time leak report, so a
trapping program under @sanitize no longer reports leaks.
`flan-dev-quit' promises to stop the daemon this Emacs started. It sent `close'
down the current connection as well, on the assumption that the two are the
same program -- and `flan-connect' is the command for attaching to one that is
not, so the assumption was one keystroke from being false. Connect to a daemon
running in a terminal, quit, and two programs ended: the one Emacs owned, and
somebody else's.
The daemon's socket is now kept apart from the connection's, which is what lets
one be told from the other. `flan-connect' names both programs before it
abandons a connection belonging to a daemon this Emacs started, and refuses
outright for a Lisp caller, as `flan-dev' already did for a running program.
`flan-dev-quit' on a connection that is not the daemon's closes that connection
and says what it left running. `flan-dev-restart-program' restarts on the
daemon's own socket, and refuses while the connection is elsewhere: there is no
reading of "build it again" that leaves one session where there was one.
Closing the window of a printing program left the pipe full, and the flushes
on the way out blocked on a compiler thread that was busy answering the very
request asking to run it again. program_state still said RUNNING, so rerun
refused a finished program as "the program is already running".
Two flushes moved. flan_exit no longer flushes before the hook -- exit(3)
flushes for itself, so that call only ever ran on the merged path, and there
it was the first thing to block. flan_merged_park flips the state under the
lock first and flushes after, because a state flip is two stores and cannot
wait on anything.
What widens is the window in which the program is parked and not yet on the
condvar. Nothing is lost there: program_asked is the fact and a signal
delivered to nobody is discarded. Two re-runs in that window are both
answered ok for one run, which is the trade against a refusal that was false.
`backward-sexp' does not signal when there is no form before point; it goes to
the beginning of the buffer and stays there. At point-min that is no movement,
so the form the dispatch then looked at was the one *after* point -- whose head
is very often a declaration -- and the region between start and end was empty.
The key installed the file's first declaration by name with no body, which the
daemon accepts, and the echo area reported it as a change. Both paths check for
movement now: the predicate refuses, and the expression arm refuses in words
rather than asking the daemon to evaluate the empty string.
The same predicate asked only for the depth at the open delimiter, and a form
written at column 1 inside a comment or a string is at depth 0 like any other.
A `defvar' in a comment is prose; it is not offered as a declaration any more.
Depth in a narrowed buffer still means depth within the narrowing, which is
noted where it is read rather than changed.
The merged build's stdout is a 64K pipe back into the daemon's own process,
and the accept loop is the only thing reading it -- which it is not doing
while serve is answering a request. The two five-second waits for a frame
boundary now drain the pipe on every tick, so a program stopped inside fwrite
is one the daemon lets go rather than one it waits out and then accuses of
not calling agent/poll.
drain and not take: the text stays in the buffer until with_output puts it on
the reply, which is where the output an evaluation caused belongs. And the
drain sits beside the sleep rather than inside the select, because a readable
pipe would make the tick free and count the timeout out in a fraction of it.
dev-chatty.flan prints 4K a frame, which is the only fixture here that fills
the pipe at all; without the drain it fails in 5.1s with the old sentence.
The members of a defenum are i32 at run time, but the reader hands the parser
an int64, so a value too large for the type arrived looking ordinary: truncated
by the x86 backend, malformed in the LLVM IR, and -- the reason this is a
correctness hole and not a nicety -- invisible to the duplicate-value rule
sitting right below it. That rule compares int64s, so (defenum E [A 0
B 4294967296]) passed it: the two differ as int64 and are both 0 as i32, and
the one check written to catch two names for one number waved through exactly
the case it exists for.
Each value is now checked where it is resolved, which is before the collision
scan runs, so the scan compares the numbers the program will actually have. A
value that does not fit is refused rather than quietly made to fit, naming the
member, its enum, and the value, with a different sentence for a value that was
written and one autoincrement walked into -- nothing in the source wrote
2147483648, so the refusal has to say where it came from before it can say it
is wrong.
The check is bound with a let rather than inlined into the cons, and that is
load-bearing: OCaml leaves :: operand order unspecified and takes the tail
first, so an inlined check would run after the recursive Int64.add and let
(defenum E [A 9223372036854775807 B]) wrap to min_int and refuse B for a number
in no one's source. Bound first, A is refused and the wrap is unreachable.
The parser is the only place this needs to happen: Parse.decl is the sole
constructor of Ast.Defenum's member values, and Load only re-qualifies the
enum's name.
Explicit-duplicate aliasing is untouched; that rule is deliberate.
`flan-dev--poll' guarded on `flan-dev--busy' and nothing else, and the watch
timer is precisely the sender that does not bind it: it sends without waiting,
so it has nothing to hold. A poll landing in that gap sent `describe' down a
connection that already owed a reply, read the watch's answer as its own, and
left its own for the watch to read a tick later -- after which the two
consumers stayed swapped for the session, each of them answering the other's
question. It runs `flan-dev-settle-hook' now, which is the invariant the hook's
docstring already claimed.
And the hook runs before `flan-dev--live-connection' rather than after. An
outstanding reply belongs to the connection it was asked on; when a daemon has
been restarted the old connection is gone and no reply is coming on the new
one, but the hook was being asked about the new one and blocked a full
`flan-dev-reply-timeout' -- a frozen Emacs on the first key pressed after a
restart, with the watch armed. `flan-dev--busy' now covers the reconnect too,
which asks the new daemon what it defines and so must not be interrupted by a
poll either.
Two orderings in the reply reader, and both of them were permanent. A frame
whose header arrived and whose body did not fell through the wait loop into
`flan-dev--extract-reply', where `byte-to-position' signalled a wrong-type
error on a position past the end of the buffer -- so the timeout message the
function goes to some trouble to word was never the one anybody read, and the
header stayed at the front of the buffer, where the next request took it as
its own and every request after that was answered by the one before it. The
body deadline is tested again now rather than trusted, and the dead frame is
erased: the timeout is said in the words meant for it, and the connection is
back in step. The header deadline still erases nothing, because a partial
header is a valid prefix of a reply that is merely slow.
The other is `flan-dev--extract-reply' reading the payload before deleting it,
so a payload that would not read was never consumed and the same bytes
signalled again on every later request. It is deleted first now. That makes
the frame gone whether or not the read succeeded, which `flan-watch--tick' has
to know: it cleared its pending flag only on a reply it got back, and would
otherwise wait for ever for one no longer in the buffer.
The condition was checked before in_loop, outside the diff that catches a
loop giving away what the next iteration needs. Emit puts the condition in
the loop header, so it runs again every trip: a condition that frees a
local freed it once per trip. It is now diffed against the same dead set,
with its own reason. A dotimes count and a loop's initial values stay
outside: those are evaluated exactly once.
The agent refuses a module it has no room to queue with "the program is not
calling agent/poll". For a running program that is the cause. For a parked one
it is confidently wrong: there is no game thread left to poll with, nothing
drains the ring until somebody runs the program again, and the ring is full
precisely because Dev.eval accepts evaluations while parked and promises they
install at the next run. At install sixty-five that promise breaks, and the
reply sent the reader to inspect a loop that is not running.
The agent cannot know this — parking is the merged shim's state — so the
daemon, which has just asked liveness, rewrites that one reply the way
abi_mismatch rewrites dlerror's. Every other refusal is still quoted in the
agent's own words, and a running program's is unchanged.
Pinned on Dev.refusal rather than end to end: filling the ring is sixty-four
real clang runs, and what the daemon does with the agent's words is the whole
of the change.
Session.eval committed the checked program the moment a form checked, and the
two steps that can still refuse it — the build, and the agent taking the
module — come after that, in the daemon. Either one left the editor holding an
error and the session holding a declaration the process has no body for. The
next module built for that session lists the name in its install prologue,
which interns a cell for it and never stores anything into it, and a dev
build's call through a cell has no null test in front of it: the game thread
jumps to address 0 at the next C-x C-e, locals render or globals refresh.
Session.held takes the four fields eval commits as one unit and Session.restore
puts them back. Dev.eval and Dev.eval_expr take one before checking and restore
it on every arm where nothing was accepted — a failed build, a refused
delivery, an unreachable agent. eval_expr needed it for the generic instances
it keeps, whose own comment already claimed the invariant: "the session must
not come to believe it holds a body that no module was ever written for."
A delivered module that then times out is not a refusal and does not roll back:
the agent has it and will install it at a frame boundary.
