[Tarantool-patches] [PATCH luajit 2/2] Different fix for partial snapshot restore due to stack overflow.

[Sergey Bronnikov]

Hello, Sergey!

Thanks for the patch! See a comment below.

Sergey

On 6/10/25 13:28, Sergey Kaplun wrote:
> From: Mike Pall <mike>
>
> Reported by Junlong Li. Fixed by Peter Cawley.
>
> (cherry picked from commit 86e7123bb1782a5f200ba5e83b8c4f3fbad4f7bc)
>
> This patch is a follow-up to the previous commit, which leads to a dirty
> read of the pseudo-valid PC set for the cframe on snapshot restoration.
> To avoid these dirty reads, this patch sets the PC to the outer frame
> as well before possible error throwing.
>
> Sergey Kaplun:
> * added the description and the test for the problem
>
> Part of tarantool/tarantool#11278
> ---
>   src/lj_snap.c                                 |  3 +-
>   src/lj_trace.c                                |  4 +-
>   ...-1196-stack-overflow-snap-restore.test.lua | 65 +++++++++++++++++++
>   3 files changed, 68 insertions(+), 4 deletions(-)
>   create mode 100644 test/tarantool-tests/lj-1196-stack-overflow-snap-restore.test.lua
>
> diff --git a/src/lj_snap.c b/src/lj_snap.c
> index 8d7bd868..4cfae579 100644
> --- a/src/lj_snap.c
> +++ b/src/lj_snap.c
> @@ -955,7 +955,8 @@ const BCIns *lj_snap_restore(jit_State *J, void *exptr)
>     lua_State *L = J->L;
>   
>     /* Set interpreter PC to the next PC to get correct error messages. */
> -  setcframe_pc(cframe_raw(L->cframe), pc+1);
> +  setcframe_pc(L->cframe, pc+1);
> +  setcframe_pc(cframe_raw(cframe_prev(L->cframe)), pc);
>   
>     /* Make sure the stack is big enough for the slots from the snapshot. */
>     if (LJ_UNLIKELY(L->base + snap->topslot >= tvref(L->maxstack))) {
> diff --git a/src/lj_trace.c b/src/lj_trace.c
> index 8a18d3cf..0d1d233a 100644
> --- a/src/lj_trace.c
> +++ b/src/lj_trace.c
> @@ -909,10 +909,8 @@ int LJ_FASTCALL lj_trace_exit(jit_State *J, void *exptr)
>     exd.J = J;
>     exd.exptr = exptr;
>     errcode = lj_vm_cpcall(L, NULL, &exd, trace_exit_cp);
> -  if (errcode) {
> -    setcframe_pc(cframe_raw(L->cframe), L);  /* Point to any valid memory. */
> +  if (errcode)
>       return -errcode;  /* Return negated error code. */
> -  }
>   
>     if (exitcode) copyTV(L, L->top++, &exiterr);  /* Anchor the error object. */
>   
> diff --git a/test/tarantool-tests/lj-1196-stack-overflow-snap-restore.test.lua b/test/tarantool-tests/lj-1196-stack-overflow-snap-restore.test.lua
> new file mode 100644
> index 00000000..942d1f82
> --- /dev/null
> +++ b/test/tarantool-tests/lj-1196-stack-overflow-snap-restore.test.lua
> @@ -0,0 +1,65 @@
> +local tap = require('tap')
> +
> +-- Test file to demonstrate LuaJIT dirty reads after stack
> +-- overflow during restoration from the snapshot.
> +-- The test fails before the patch under Valgrind.

Please specify valgrind option that is required for reproducing the bug.

Cannot reproduce with command line below with reverted patch:

VALGRIND_OPTS="--leak-check=no --malloc-fill=0x00 --free-fill=0x00" 
ctest -V -R test/tarantool-tests/lj-1196-partial-snap-restore.test.lua -V

> +--
> +-- luacheck: push no max_comment_line_length
> +--
> +-- See also:https://github.com/LuaJIT/LuaJIT/issues/1196,
> +--https://www.freelists.org/post/luajit/Invalid-read-found-by-valgrind.
> +--
> +-- luacheck: pop
> +
> +local test = tap.test('lj-1196-stack-overflow-snap-restore')
> +
> +test:plan(4)
> +
> +-- XXX: This file has the same tests as the
> +-- <test/LuaJIT-tests/lang/stackov.lua>, but without disabling the
> +-- compilation for the given functions. Hence, the check here is
> +-- less strict -- we just check that there are no dirty reads,
> +-- uninitialized memory access, etc.
> +
> +local function recursive_f_noarg()
> +  recursive_f_noarg()
> +end
> +
> +local function recursive_one_arg(argument)
> +  recursive_one_arg(argument)
> +end
> +
> +local function recursive_f_vararg(...)
> +  recursive_f_vararg(1, ...)
> +end
> +
> +local function recursive_f_vararg_tail(...)
> +  return recursive_f_vararg_tail(1, ...)
> +end
> +
> +-- Use `coroutine.wrap()`, for independent stack sizes.
> +-- The invalid read is done by the error handler
> +-- `debug.traceback()`, since it observes the pseudo PC (`L`) and
> +-- reads the memory by `L - 4` address before the patch.
> +
> +coroutine.wrap(function()
> +  local status = xpcall(recursive_f_noarg, debug.traceback)
> +test:ok(not status, 'correct status, recursive no arguments')
> +end)()
> +
> +coroutine.wrap(function()
> +  local status = xpcall(recursive_one_arg, debug.traceback, 1)
> +test:ok(not status, 'correct status, recursive one argument')
> +end)()
> +
> +coroutine.wrap(function()
> +  local status = xpcall(recursive_f_vararg, debug.traceback, 1)
> +test:ok(not status, 'correct status, recursive vararg')
> +end)()
> +
> +coroutine.wrap(function()
> +  local status = xpcall(recursive_f_vararg_tail, debug.traceback, 1)
> +test:ok(not status, 'correct status, recursive vararg tail')
> +end)()
> +
> +test:done(true)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.tarantool.org/pipermail/tarantool-patches/attachments/20250610/95df73e8/attachment.htm>