[Tarantool-patches] [PATCH luajit][v2] Prevent loop in snap_usedef().
Sergey Kaplun
skaplun at tarantool.org
Tue Jan 21 15:27:53 MSK 2025
Hi, Sergey!
Thanks for the patch and the fixes!
LGTM, after fixing my several suggestions about comments rephrasing and
a bunch of nits below.
On 16.01.25, Sergey Bronnikov wrote:
> Reported by XmiliaH.
>
> (cherry picked from commit 0e66fc96377853d898390f1a02723c54ec3a42f7)
>
> It is possible to get an infinite loop in a function `snap_usedef`
> when a `UCLO` makes a tight loop. This infinite loop could happen
> when `snap_usedef()` is called on trace exit processes UCLO
It is called during trace recording (more precisely on creation of the
snapshot for the guarded trace check).
> bytecode instruction and this instruction attempts a jump with
> negative value. The patch fixes the problem by checking a number
> of slots in a jump argument and replace this value my `maxslot` if
> a value is negative.
Please also mention that "This means that no values will be purged from
the snapshot".
>
> Sergey Bronnikov:
> * added the description and the test for the problem
>
> Part of tarantool/tarantool#10709
> ---
>
> Branch: https://github.com/tarantool/luajit/tree/ligurio/lj-736-prevent-loop-in-snap_usedef
>
> Related issues:
> * https://github.com/tarantool/tarantool/issues/10709
> * https://github.com/LuaJIT/LuaJIT/issues/736
>
> v2 changes:
>
> - Updated test, now it hangs without patch with fix.
> - Added more comments to the test with explanations.
>
> src/lj_snap.c | 7 +-
> .../lj-736-BC_UCLO-triggers-infinite-loop.lua | 82 +++++++++++++++++++
> 2 files changed, 88 insertions(+), 1 deletion(-)
> create mode 100644 test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
>
> diff --git a/src/lj_snap.c b/src/lj_snap.c
> index 8a33dc22..8d7bd868 100644
> --- a/src/lj_snap.c
> +++ b/src/lj_snap.c
<snipped>
> diff --git a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
> new file mode 100644
> index 00000000..fb053e9a
> --- /dev/null
> +++ b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
Filename suffix should be ".test.lua".
The ctest doesn't recognize this file.
> @@ -0,0 +1,82 @@
> +local tap = require('tap')
> +local test = tap.test('lj-736-BC_UCLO-triggers-infinite-loop'):skipcond({
> + ['Test requires JIT enabled'] = not jit.status(),
> +})
> +
> +test:plan(2)
> +
> +-- Test reproduces an issue when BC_UCLO triggers an infinite loop.
> +-- See details in https://github.com/LuaJIT/LuaJIT/issues/736.
Nit: Lets be consistent with most of our test files and put the short
description before the `tap.test()`:
===================================================================
diff --git a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
index fb053e9a..f38098a8 100644
--- a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
+++ b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
@@ -1,13 +1,13 @@
local tap = require('tap')
+-- Test file to demonstrate the infinite loop in LuaJIT during the
+-- use-def analysis for upvalues.
+-- See also https://github.com/LuaJIT/LuaJIT/issues/736.
local test = tap.test('lj-736-BC_UCLO-triggers-infinite-loop'):skipcond({
['Test requires JIT enabled'] = not jit.status(),
})
test:plan(2)
--- Test reproduces an issue when BC_UCLO triggers an infinite loop.
--- See details in https://github.com/LuaJIT/LuaJIT/issues/736.
---
-- Listing below demonstrates a problem -
-- the bytecode UCLO on the line 13 makes a loop at 0013-0014:
--
===================================================================
> +--
> +-- Listing below demonstrates a problem -
> +-- the bytecode UCLO on the line 13 makes a loop at 0013-0014:
> +--
> +-- - BYTECODE -- bc_uclo.lua:0-20
> +-- 0001 KPRI 0 0
> +-- 0002 FNEW 1 0 ; bc_uclo.lua:5
> +-- 0003 KSHORT 2 1
> +-- 0004 KSHORT 3 4
> +-- 0005 KSHORT 4 1
> +-- 0006 FORI 2 => 0011
> +-- 0007 => ISNEN 5 0 ; 2
> +-- 0008 JMP 6 => 0010
> +-- 0009 UCLO 0 => 0012
> +-- 0010 => FORL 2 => 0007
> +-- 0011 => UCLO 0 => 0012
> +-- 0012 => KPRI 0 0
> +-- 0013 UCLO 0 => 0012
> +-- 0014 FNEW 1 1 ; bc_uclo.lua:18
> +-- 0015 UCLO 0 => 0016
> +-- 0016 => RET0 0 1
I suggest to list only the necessary part of the bytecodes, also
lets take the bytecode dump for this particular function:
===================================================================
diff --git a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
index f38098a8..8d0558e9 100644
--- a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
+++ b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
@@ -8,27 +8,18 @@ local test = tap.test('lj-736-BC_UCLO-triggers-infinite-loop'):skipcond({
test:plan(2)
--- Listing below demonstrates a problem -
--- the bytecode UCLO on the line 13 makes a loop at 0013-0014:
---
--- - BYTECODE -- bc_uclo.lua:0-20
--- 0001 KPRI 0 0
--- 0002 FNEW 1 0 ; bc_uclo.lua:5
--- 0003 KSHORT 2 1
--- 0004 KSHORT 3 4
--- 0005 KSHORT 4 1
--- 0006 FORI 2 => 0011
--- 0007 => ISNEN 5 0 ; 2
--- 0008 JMP 6 => 0010
--- 0009 UCLO 0 => 0012
--- 0010 => FORL 2 => 0007
--- 0011 => UCLO 0 => 0012
--- 0012 => KPRI 0 0
--- 0013 UCLO 0 => 0012
--- 0014 FNEW 1 1 ; bc_uclo.lua:18
--- 0015 UCLO 0 => 0016
--- 0016 => RET0 0 1
-
+-- Before the patch, the code flow like in the `testcase()` below
+-- may cause the problem -- use-def analysis for the 0019 UCLO
+-- creates an infinite loop in 0014 - 0019:
+-- | 0008 FORI base: 4 jump: => 0013
+-- | 0009 ISNEN var: 7 num: 0 ; number 2
+-- | 0010 JMP rbase: 8 jump: => 0012
+-- | 0011 UCLO rbase: 2 jump: => 0014
+-- | 0012 FORL base: 4 jump: => 0009
+-- | 0013 UCLO rbase: 2 jump: => 0014
+-- | 0014 KPRI dst: 2 pri: 0 ; Start of `assert()` line.
+-- | ...
+-- | 0019 UCLO rbase: 2 jump: => 0014
jit.opt.start('hotloop=1')
local assert_msg = 'Infinite loop is not reproduced.'
===================================================================
> +
> +jit.opt.start('hotloop=1')
> +
> +local assert_msg = 'Infinite loop is not reproduced.'
> +local assert = assert
> +
> +local function testcase()
> + -- The code in the first scope `do`/`end` is a prerequisite.
> + -- It is needed so that we have a trace at the exit from which
> + -- the creation of the snapshot will begin.
The needed part is not the creation of the snapshot but the use-def
analysis for the first (`uv1`) UCLO. I would rephrase it like the
following:
===================================================================
diff --git a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
index 8d0558e9..3e6508d1 100644
--- a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
+++ b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
@@ -27,11 +27,9 @@ local assert = assert
local function testcase()
-- The code in the first scope `do`/`end` is a prerequisite.
- -- It is needed so that we have a trace at the exit from which
- -- the creation of the snapshot will begin.
+ -- It contains the UCLO instruction for the `uv1`. The use-def
+ -- analysis for it escapes this `do`/`end` scope.
do
- -- Upvalue below is not used actually, but it is required
- -- for calling `snap_usedef()` on trace exit.
local uv1 -- luacheck: ignore
local _ = function() return uv1 end
===================================================================
> + do
> + -- Upvalue below is not used actually, but it is required
> + -- for calling `snap_usedef()` on trace exit.
Side note: The use-def analysis is called on trace recording.
I would rather merge this comment with the comment above, as you can see in
the diff.
> + local uv1 -- luacheck: ignore
Nit: It is better to use "no unused" here to be more specific.
> + local _ = function() return uv1 end
> +
> + -- The loop below is required for recording a trace.
> + -- The condition inside a loop executes `goto` to a label
> + -- outside of the loop when the code executed by JIT and
> + -- this triggers snapshotting.
> + for i = 1, 2 do
> + -- Exit to interpreter once trace is compiled.
I would rephrase this 2 comments, since they a little bit misleading:
===================================================================
diff --git a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
index 3e6508d1..a95651fd 100644
--- a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
+++ b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
@@ -33,12 +33,11 @@ local function testcase()
local uv1 -- luacheck: ignore
local _ = function() return uv1 end
- -- The loop below is required for recording a trace.
- -- The condition inside a loop executes `goto` to a label
- -- outside of the loop when the code executed by JIT and
- -- this triggers snapshotting.
+ -- Records the trace for which use-def analysis is applied.
for i = 1, 2 do
- -- Exit to interpreter once trace is compiled.
+ -- This condition triggers snapshoting and use-def analysis.
+ -- Before the patch this triggers the infinite loop in the
+ -- `snap_usedef()`, so the `goto` is never taken.
if i == 2 then
goto x
end
===================================================================
> + if i == 2 then
> + goto x
> + end
> + end
> + end
> +
> +::x::
> + do
> + local uv2 -- luacheck: no unused
> +
> + -- `goto` if not executed without a patch and generates an
> + -- UCLO bytecode that makes an infinite loop in a function
> + -- `snap_usedef` when patch is not applied. `goto` must point
> + -- to the label on one of the previous lines. `assert()` is
> + -- executed when patch is applied.
> + assert(nil, assert_msg)
> + goto x
> +
> + -- Line below is required, it makes `uv` upvalue, and must be
> + -- placed after `goto`, otherwise reproducer become broken.
> + local _ = function() return uv2 end -- luacheck: ignore
I would also rephrase this 2 comments like the following since they
are repeated what we say before:
===================================================================
diff --git a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
index a95651fd..8111d752 100644
--- a/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
+++ b/test/tarantool-tests/lj-736-BC_UCLO-triggers-infinite-loop.lua
@@ -47,17 +47,11 @@ local function testcase()
::x::
do
local uv2 -- luacheck: no unused
-
- -- `goto` if not executed without a patch and generates an
- -- UCLO bytecode that makes an infinite loop in a function
- -- `snap_usedef` when patch is not applied. `goto` must point
- -- to the label on one of the previous lines. `assert()` is
- -- executed when patch is applied.
assert(nil, assert_msg)
+ -- Create a tight loop for the one more upvalue (`uv2`).
+ -- Before the patch, use-def analysis gets stuck in this code
+ -- flow.
goto x
-
- -- Line below is required, it makes `uv` upvalue, and must be
- -- placed after `goto`, otherwise reproducer become broken.
local _ = function() return uv2 end -- luacheck: ignore
end
end
===================================================================
Also I would add the following comment:
| -- This code is unreachable by design. Prevent luacheck warning.
> + end
> +end
> +
> +local ok, err = pcall(testcase)
> +
> +test:is(ok, false, 'assertion is triggered in a function with testcase')
> +test:ok(err:match(assert_msg), 'BC_UCLO does not trigger an infinite loop')
I would rather say here 'correct error message'. Or, as an alternative,
we may just use 1 test check:
| pcall(testcase)
|
| test:ok(true, 'no infinite loop in the use-def analysis')
instead of these 2 checks.
> +
> +os.exit(test:check() and 0 or 1)
Minor: Please use `test:done(true)` like in all other tests.
> --
> 2.34.1
>
--
Best regards,
Sergey Kaplun
More information about the Tarantool-patches
mailing list