[Tarantool-patches] [PATCH luajit] Check frame size limit before returning to a lower frame.

[Sergey Bronnikov]

Hi, Sergey


thanks for the patch!

On 3/12/24 08:26, Sergey Kaplun wrote:
> From: Mike Pall <mike>
>
> Thanks to Sergey Kaplun.
>
> (cherry picked from commit 302366a33853b730f1b7eb61d792abc4f84f0caa)
>
> When compiling a stitched (or side) trace, there is no check for the
> frame size of the current prototype during recording. Hence, when we
> return (for example, after stitching) to the lower frame with a maximum
> possible frame size (249), the 251 = `baseslot` (2) + `maxslot` (249)
> slot for GC64 mode may be used. This leads to the corresponding assertion
> failure in `rec_check_slots()`.
>
> This patch adds the corresponding check.
>
> Sergey Kaplun:
> * added the description and the test for the problem
>
> Part of tarantool/tarantool#9595
> ---
>
> Branch: https://github.com/tarantool/luajit/tree/skaplun/lj-1173-frame-limit-lower-frame
> Tarantool PR: https://github.com/tarantool/tarantool/pull/9791
> Related issues:
> * https://github.com/tarantool/tarantool/issues/9595
> * https://github.com/LuaJIT/LuaJIT/issues/1173
>
>   src/lj_record.c                               |  2 +
>   .../lj-1173-frame-limit-lower-frame.test.lua  | 83 +++++++++++++++++++
>   2 files changed, 85 insertions(+)
>   create mode 100644 test/tarantool-tests/lj-1173-frame-limit-lower-frame.test.lua
>
> diff --git a/src/lj_record.c b/src/lj_record.c
> index c01c1f0b..e3590b1a 100644
> --- a/src/lj_record.c
> +++ b/src/lj_record.c
> @@ -886,6 +886,8 @@ void lj_record_ret(jit_State *J, BCReg rbase, ptrdiff_t gotresults)
>         lj_trace_err(J, LJ_TRERR_LLEAVE);
>       } else if (J->needsnap) {  /* Tailcalled to ff with side-effects. */
>         lj_trace_err(J, LJ_TRERR_NYIRETL);  /* No way to insert snapshot here. */
> +    } else if (1 + pt->framesize >= LJ_MAX_JSLOTS) {
> +      lj_trace_err(J, LJ_TRERR_STACKOV);
>       } else {  /* Return to lower frame. Guard for the target we return to. */
>         TRef trpt = lj_ir_kgc(J, obj2gco(pt), IRT_PROTO);
>         TRef trpc = lj_ir_kptr(J, (void *)frame_pc(frame));
> diff --git a/test/tarantool-tests/lj-1173-frame-limit-lower-frame.test.lua b/test/tarantool-tests/lj-1173-frame-limit-lower-frame.test.lua
> new file mode 100644
> index 00000000..91e2c603
> --- /dev/null
> +++ b/test/tarantool-tests/lj-1173-frame-limit-lower-frame.test.lua
> @@ -0,0 +1,83 @@
> +local tap = require('tap')
> +
> +-- Test file to demonstrate LuaJIT assertion failure during
> +-- recording of side trace in GC64 mode with return to lower
> +-- frame, which has the maximum possible frame size.
> +-- See also: https://github.com/LuaJIT/LuaJIT/issues/1173.
> +
> +local test = tap.test('lj-1173-frame-limit-lower-frame'):skipcond({
> +  ['Test requires JIT enabled'] = not jit.status(),
> +})
> +
> +test:plan(1)
> +
> +-- Parent trace with stitching and returning to a lower frame.
> +local function retf()
> +  math.modf(1)
> +end
> +_G.retf = retf
> +
> +local LJ_MAX_JSLOTS = 250

I would say in a comment that constant is from <src/lj_def.h>.

Your test depends on this constant (if it will be changed the test will 
test nothing),

how to make sure that LJ_MAX_JSLOTS is equal to LJ_MAX_JSLOTS in 
<src/lj_def.h>?

> +
> +-- Generate the following function:
> +-- | local uv = {key = 1}
> +-- | return function()
> +-- |   local r = retf()
> +-- |   uv.key, uv.key, --[[124 times in total ...]] uv.key = nil
> +-- | end
> +-- It will have the following bytecode:
> +-- | 0001    GGET     0   0      ; "retf"
> +-- | 0002    CALL     0   2   1
> +-- | 0003    UGET     1   0      ; uv
> +-- | ...
> +-- | 0126    UGET   124   0      ; uv
> +-- | 0127    KNIL   125 248
> +-- | 0128    TSETS  248 124   1  ; "key"
> +-- | ...
> +-- | 0251    TSETS  125   1   1  ; "key"
> +-- | 0252    RET0     0   1
> +-- As you can see, the 249 slots (from 0 to 248) are occupied in
> +-- total.
> +-- When we return to the lower frame for the side trace, we may
> +-- hit the slot limit mentioned above: 2 slots are occupied
> +-- by the frame (`baseslot`) and `KNIL` bytecode recording sets
> +-- `maxslot` (the first free slot) to 249. Hence, the JIT slots
> +-- are overflowing.
> +
> +local chunk = 'local uv = {key = 1}\n'
> +chunk = chunk .. 'return function()\n'
> +chunk = chunk .. 'local r = retf()\n'
> +
> +-- Each `UGET` occupies 1 slot, `KNIL` occupies the same amount.
> +-- 1 slot is reserved (`r` variable), 1 pair is set outside the
> +-- cycle. 249 slots (the maximum available amount, see
> +-- <src/lj_parse.c>, `bcreg_bump()` for details) are occupied in
> +-- total.
> +for _ = 1, LJ_MAX_JSLOTS / 2 - 2 do
> +  chunk = chunk .. ('uv.key, ')
> +end
> +chunk = chunk .. 'uv.key = nil\n'
> +chunk = chunk .. 'end\n'
Why not to use multiline here and after the loop?
> +local get_func = assert(loadstring(chunk))
> +local function_max_framesize = get_func()
> +
> +jit.opt.start('hotloop=1', 'hotexit=1')
> +
> +-- Compile the parent trace first.
> +retf()
> +retf()
> +
> +-- Try to compile the side trace with a return to a lower frame
> +-- with a huge frame size.
> +function_max_framesize()
> +function_max_framesize()
> +
> +-- XXX: The limit check is OK with default defines for non-GC64
> +-- mode, the trace is compiled for it. The test fails only with
> +-- GC64 mode enabled. Still run the test for non-GC64 mode to
> +-- avoid regressions.
> +
> +test:ok(true, 'no assertion failure during recording')
> +
> +test:done(true)