Hi, Sergey,

thanks for the patch! LGTM

On 10.09.2024 17:05, Sergey Kaplun wrote:
> From: Mike Pall <mike>
>
> (cherry picked from commit 204cee2c917f55f288c0b166742e56c134fe578c)
>
> It is possible that a snapshot topslot is less than the possible topslot
> of the Lua stack. In that case, if the Lua stack overflows in
> `lj_vmevent_prepare()`, the error is raised inside
> `lj_vm_exit_handler()`, which has no corresponding DWARF eh_frame [1],
> so it leads to the crash.
>
> This patch fix-ups the topslot of the snapshot on trace exit to the
> maximum possible one.
>
> Sergey Kaplun:
> * added the description and the test for the problem
>
> [1]:https://refspecs.linuxfoundation.org/LSB_3.0.0/LSB-Core-generic/LSB-Core-generic/ehframechpt.html
>
> Part of tarantool/tarantool#10199
> ---
>
> Branch:https://github.com/tarantool/luajit/tree/skaplun/fix-stack-alloc-on-trace
> Issue:https://github.com/tarantool/tarantool/issues/10199
>
>   src/lj_trace.c                                |  6 ++-
>   .../fix-stack-alloc-on-trace-exit.test.lua    | 53 +++++++++++++++++++
>   2 files changed, 58 insertions(+), 1 deletion(-)
>   create mode 100644 test/tarantool-tests/fix-stack-alloc-on-trace-exit.test.lua
>
> diff --git a/src/lj_trace.c b/src/lj_trace.c
> index 20014ecb..94cb27e5 100644
> --- a/src/lj_trace.c
> +++ b/src/lj_trace.c
> @@ -522,7 +522,11 @@ static void trace_stop(jit_State *J)
>       lj_assertJ(J->parent != 0 && J->cur.root != 0, "not a side trace");
>       lj_asm_patchexit(J, traceref(J, J->parent), J->exitno, J->cur.mcode);
>       /* Avoid compiling a side trace twice (stack resizing uses parent exit). */
> -    traceref(J, J->parent)->snap[J->exitno].count = SNAPCOUNT_DONE;
> +    {
> +      SnapShot *snap = &traceref(J, J->parent)->snap[J->exitno];
> +      snap->count = SNAPCOUNT_DONE;
> +      if (J->cur.topslot > snap->topslot) snap->topslot = J->cur.topslot;
> +    }
>       /* Add to side trace chain in root trace. */
>       {
>         GCtrace *root = traceref(J, J->cur.root);
> diff --git a/test/tarantool-tests/fix-stack-alloc-on-trace-exit.test.lua b/test/tarantool-tests/fix-stack-alloc-on-trace-exit.test.lua
> new file mode 100644
> index 00000000..ca04e54e
> --- /dev/null
> +++ b/test/tarantool-tests/fix-stack-alloc-on-trace-exit.test.lua
> @@ -0,0 +1,53 @@
> +local tap = require('tap')
> +
> +-- Test file to demonstrate incorrect Lua stack restoration on
> +-- exit from trace by the stack overflow.
> +
> +local test = tap.test('fix-stack-alloc-on-trace-exit'):skipcond({
> +  ['Test requires JIT enabled'] = not jit.status(),
> +})
> +
> +local jit_dump = require('jit.dump')
> +
> +test:plan(2)
> +
> +-- Before the patch, it is possible that a snapshot topslot is
> +-- less than the possible topslot of the Lua stack. In that case,
> +-- if the Lua stack overflows in `lj_vmevent_prepare()`, the error
> +-- is raised inside `lj_vm_exit_handler()`, which has no
> +-- corresponding DWARF eh_frame, so it leads to the crash.
> +
> +-- Need for the stack growing in `lj_vmevent_prepare`.
> +jit_dump.start('x', '/dev/null')
> +
> +-- Create a coroutine with a fixed stack size.
> +local coro = coroutine.create(function()
> +  jit.opt.start('hotloop=1', 'hotexit=1', 'callunroll=1')
> +
> +  -- `math.modf` recording is NYI.
> +  -- Local `math_modf` simplifies `jit.dump()` output.
> +  local math_modf = math.modf
> +
> +  local function trace(n)
> +    n = n + 1
> +    -- luacheck: ignore
> +    -- Start a side trace here.
> +    if n % 2 == 0 then end
> +    -- Stop the recording of the side trace and a main trace,
> +    -- stitching.
> +    math_modf(1, 1)
> +    -- Grow stack, avoid tail calls.
> +    local unused = trace(n)
> +    return unused
> +  end
> +
> +  local n = 0
> +  trace(n)
> +end)
> +
> +local result, errmsg = coroutine.resume(coro)
> +
> +test:ok(not result, 'correct status and no crash')
> +test:like(errmsg, 'stack overflow', 'correct error message')
> +
> +test:done(true)