flan/test/test_dyn.ml
Joseph Ferano e6af2d3f77 The park kept the frames' roots off and the globals' on
flan_merged_park called flan_dyn_root_reset, which emptied the collector's
root stack. The frames' roots had to go — main is left by longjmp, so they
name stack the next run overwrites — but the dyn globals' roots are on that
same stack, pushed once by the emitted main and never popped, and the park
took them with the frames.

The park is not a quiet state. It services evaluated thunks, a thunk
allocates, and an allocation collects. So a program with (defvar config dyn)
answered (get config :s) with its string before any thunk ran and with nil
after one that allocated past the heap's floor — a read of memory the sweep
had freed, answering nil by luck of what the freed words decoded as.

The emitted main now brackets its global pushes: flan_dyn_root_globals_begin
empties the stack, the pushes go on, flan_dyn_root_globals_end records how
many of them there are, and the park resets to that line instead of to zero.
Nothing between the two allocates, which is what keeps the globals from being
swept in the window where they are unrooted — and [begin] emptying the stack
rather than adding to it is what makes a re-entered main re-root the same
globals rather than push a second copy of each, which also closes the other
half: a re-run used to re-push roots over slots left dangling by the park.

Both emitters, because the dev loop's default backend is x86 and a fix in one
lowering is not a fix. A program with no dyn globals emits neither call and
its root stack still resets to empty, which is what an empty push list should
leave behind.

flan_dyn_root_pop now clamps at the globals rather than at zero. An
over-popping frame eating the globals is the one way that clamp could turn a
miscount into this same use-after-free.

Covered twice. test/dyn_ops.c's park mode is the runtime's half — a run, a
park with a collecting thunk in it, and another run, three times over,
asserting both that the global survives and that the frame's five hundred
objects do not. Under ASan the old reset reports heap-use-after-free in
flan_dyn_tag with the free in gc_sweep; under memcheck it reports 24 errors
and still prints the right answer, which is the shape of the bug. test_dev.ml
drives the whole daemon over its socket on both backends against
programs/dev-dyn-global.flan.

Not touched, and it wants a decision rather than a patch: a re-run re-enters
flan_program_main, which re-runs the lifted startup function, so every global
with a computed initialiser is reset by a re-run. That contradicts dev.ml's
own note and FIX.org item 1. It is independent of this — the roots are right
whether or not the values are re-initialised.

Nor is this the reload path. A defvar added by an evaluation gets its storage
from flan_dev_global (emit.ml's new_globals, x86.ml's counterpart) and there
is no flan_dyn_root_push anywhere on that path in either backend, so a dyn
global added to a live session is unrooted. That is a separate defect with a
separate fix, and nothing here makes it better or worse.
2026-09-20 11:06:14 +07:00

204 lines
9.0 KiB
OCaml

(* The dynamic-value runtime, driven from C.
runtime/flan_dyn.c is a C ABI with no Flan spelling yet — the compiler lane
is what gives it one — so the only way to reach every operation, and every
way each one refuses, is a C main. test/dyn_ops.c is that main and
programs/dyn-host.flan is the program it is linked against, which has no
[main] of its own for the same reason programs/reload.flan does not.
What is asserted here, and why each is its own thing:
ops every operation's happy path, the tags, the printed form of
each, text identity against text equality, and a vec that
contains itself
gc a million allocations against a hundred live, and the heap's
high-water mark bounded
unrooted the positive control: an object nothing points at is reclaimed.
park the root stack cut back to its globals with no frame having
returned, which is what the dev daemon's park does between one
run and the next: the globals stay rooted, the frames' go
desc an aggregate root: a struct whose dyn fields are named by a
descriptor rather than pushed one at a time.
Without it a collector that never freed would pass everything
nested a chain of vecs sixty-four deep, traced through one root
sharing one object held three times — written through one path and read
through another, and swept once when the last goes
refuse:* twenty-four refusals, one process each, asserted on the sentence
as well as on the status: a process that died some other way is
not the guard firing, and the exit code cannot tell them apart
One binary, built once, run thirty times. The build is the expensive
part and the runs are milliseconds, which is what keeps this inside
`dune test` rather than behind an alias. *)
open Flan
(* The watchdog first: a hang is the one failure mode that reports nothing at
all. See watchdog.ml. *)
let () = Watchdog.arm ~seconds:600 "test_dyn"
let failures = ref 0
let fail fmt =
Printf.ksprintf (fun s -> incr failures; print_endline ("FAIL " ^ s)) fmt
let scratch = Filename.get_temp_dir_name ()
let tmp name = Filename.concat scratch ("flan-dyn-" ^ name)
let has hay needle =
let n = String.length needle and h = String.length hay in
let rec go i = i + n <= h && (String.sub hay i n = needle || go (i + 1)) in
go 0
let () =
match Sys.command "command -v clang > /dev/null 2>&1" with
| 0 ->
let p =
Check.program
(Load.program ~file:"programs/dyn-host.flan"
(Reader.read_file "programs/dyn-host.flan")).Load.decls
in
let exe = tmp "ops" in
ignore (Build.executable ~csrcs:[ "dyn_ops.c" ] p ~out:exe);
let run mode =
let o = tmp (mode ^ ".out") and e = tmp (mode ^ ".err") in
let code =
Sys.command
(Printf.sprintf "%s %s > %s 2> %s" (Filename.quote exe)
(Filename.quote mode) (Filename.quote o) (Filename.quote e))
in
let out = In_channel.with_open_bin o In_channel.input_all in
let err = In_channel.with_open_bin e In_channel.input_all in
List.iter (fun x -> try Sys.remove x with Sys_error _ -> ()) [ o; e ];
(code, out, err)
in
(* The happy paths. Each assertion inside prints its own FAIL line, so the
output is the report and this only has to notice that there was one. *)
let code, out, err = run "ops" in
if code <> 0 || out <> "ops ok\n" then
fail "the operations\n got: %S (exit %d, err %S)" out code err;
(* The collector. Four sentences, each a yes: the high-water mark stayed
under half a megabyte across forty megabytes of allocation, the heap
settled small, the object count stayed bounded, and the hundred rooted
values were all still what they were set to. *)
let code, out, err = run "gc" in
let want_gc =
"peak under 512K: yes\nsettled under 32K: yes\n\
live objects bounded: yes\nlive set intact: yes\n"
in
if code <> 0 || out <> want_gc then
fail "a million allocations against a hundred live\n\
\ got: %S (exit %d, err %S)\n wanted: %S"
out code err want_gc;
(* And the control. A collector that never freed anything would pass every
other case in this file; this is the one it cannot. *)
let code, out, _ = run "unrooted" in
let want_un = "allocated: yes\nreclaimed all but the ring: yes\n" in
if code <> 0 || out <> want_un then
fail "an unrooted object\n got: %S (exit %d)\n wanted: %S"
out code want_un;
(* The park: a root stack cut back to its globals with no frame having
returned, which is what the merged dev build does between a run and the
next. Both halves are asserted, because each on its own is met by doing
nothing — a reset that dropped everything keeps no global, and a reset
that dropped nothing keeps every frame. *)
let code, out, _ = run "park" in
let want_park =
"a dyn global survives a parked thunk: yes\n\
a finished run's frame roots go: yes\n"
in
if code <> 0 || out <> want_park then
fail "the park's root reset\n got: %S (exit %d)\n wanted: %S"
out code want_park;
(* The aggregate roots the per-type descriptors added: a struct with dyn
fields at three offsets, one of them inside a nested struct, rooted by
address and descriptor rather than word by word.
The second line is the one that discriminates, and the first on its own
does not: a marker that walked every word of the struct would keep the
named fields too. So the struct also holds five hundred objects behind a
dyn word at an offset the descriptor leaves out, and the claim is that
they are *not* kept. *)
let code, out, _ = run "desc" in
let want_desc =
"aggregate root survives collection: yes\n\
a dyn at an offset the descriptor omits is not marked: yes\n\
and is reclaimed once dropped: yes\n"
in
if code <> 0 || out <> want_desc then
fail "an aggregate root\n got: %S (exit %d)\n wanted: %S"
out code want_desc;
let code, out, _ = run "nested" in
if code <> 0 || out <> "chain of 64 intact: yes\n" then
fail "a chain of nested vecs\n got: %S (exit %d)" out code;
let code, out, _ = run "sharing" in
let want_sh =
"three slots hold one object: yes\n\
write through one path is seen through another: yes\n\
shared object survives on the holder alone: yes\n\
still reachable: yes\n\
live bytes after dropping everything: ok\n"
in
if code <> 0 || out <> want_sh then
fail "interior sharing\n got: %S (exit %d)\n wanted: %S"
out code want_sh;
(* Every refusal. The pair is (mode, a phrase the sentence must contain);
the phrase is chosen to be the part that says *which* mistake it was,
so a message that named the wrong operation or the wrong tag would not
pass by accident.
The tag names are asserted here too, as words: "int" and "text" and not
2 and 4. A message with a number in it is a puzzle, and the rule is
written down in flan_dyn.c beside the table the words come from. *)
let refusals =
[ ("add", "dyn +: int and text");
("sub", "dyn -: nil and int");
("mul", "dyn *: bool and int");
("div", "dyn /: vec and int");
("rem", "dyn %: int and nil");
("divzero", "does not divide by zero");
("remzero", "does not divide by zero");
("divover", "one past the largest i64");
("lt", "dyn <: int and text");
("le", "dyn <=: text and nil");
("gt", "dyn >: vec and vec");
("ge", "dyn >=: bool and bool");
("len", "dyn len: int");
("at", "dyn at: int and int");
("atindex", "an index must be an int");
("atrange", "index 9 is out of bounds for text of length 2");
("atnegative", "index -1 is out of bounds");
("setattext", "a text is immutable");
("setatnotvec", "only a vec is assigned into");
("setatrange", "index 0 is out of bounds for vec of length 0");
("push", "only a vec is pushed to");
("needi64", "dyn i64: text");
("needf64", "dyn f64: int");
("needbool", "dyn bool: nil") ]
in
List.iter
(fun (mode, phrase) ->
let code, out, err = run ("refuse:" ^ mode) in
if code = 0 then
fail "%s returned rather than trapping: %S" mode out
else if not (has err phrase) then
fail "%s did not say %S; it said %S" mode phrase err)
refusals;
(try Sys.remove exe with Sys_error _ -> ());
(* A line on the way out, because a test that says nothing when it passes
is a test nobody can tell from a test that did not run. *)
if !failures = 0 then
Printf.printf " ok the dyn runtime: %d refusals and seven runs\n"
(List.length refusals)
else exit 1
| _ -> print_endline "SKIP test_dyn: no clang"