[Tarantool-patches] [PATCH luajit 3/6] FFI: Ensure returned string is alive in ffi.typeinfo().
Sergey Kaplun
skaplun at tarantool.org
Mon Oct 23 12:22:03 MSK 2023
From: Mike Pall <mike>
Reported by zhuizhuhaomeng.
(cherry-picked from commit 94a40bb238092e73f3dc0c3626911a7efa997c22)
`ffi.typeinfo()` doesn't check that the string containing the name type
is alive when it sets this string to the returned table. This leads to
the corresponding assertion failure in `checklivetv()`. This patch flips
the white colour of the stored string to resurrect it.
Sergey Kaplun:
* added the description and the test for the problem
Part of tarantool/tarantool#9145
---
src/lib_ffi.c | 1 +
.../lj-745-ffi-typeinfo-dead-names.test.lua | 28 +++++++++++++++++++
2 files changed, 29 insertions(+)
create mode 100644 test/tarantool-tests/lj-745-ffi-typeinfo-dead-names.test.lua
diff --git a/src/lib_ffi.c b/src/lib_ffi.c
index 62af54c1..e60e7b19 100644
--- a/src/lib_ffi.c
+++ b/src/lib_ffi.c
@@ -573,6 +573,7 @@ LJLIB_CF(ffi_typeinfo)
setintV(lj_tab_setstr(L, t, lj_str_newlit(L, "sib")), (int32_t)ct->sib);
if (gcref(ct->name)) {
GCstr *s = gco2str(gcref(ct->name));
+ if (isdead(G(L), obj2gco(s))) flipwhite(obj2gco(s));
setstrV(L, lj_tab_setstr(L, t, lj_str_newlit(L, "name")), s);
}
lj_gc_check(L);
diff --git a/test/tarantool-tests/lj-745-ffi-typeinfo-dead-names.test.lua b/test/tarantool-tests/lj-745-ffi-typeinfo-dead-names.test.lua
new file mode 100644
index 00000000..02e13759
--- /dev/null
+++ b/test/tarantool-tests/lj-745-ffi-typeinfo-dead-names.test.lua
@@ -0,0 +1,28 @@
+local tap = require('tap')
+local ffi = require('ffi')
+local test = tap.test('lj-745-ffi-typeinfo-dead-names')
+
+test:plan(1)
+
+-- Start from the beginning of the GC cycle.
+collectgarbage()
+
+local function ctypes_iteration()
+ local i = 1
+ -- Test `checklivetv()` assertion in `setstrV()` inside
+ -- `ffi.typeinfo()`.
+ while ffi.typeinfo(i) do i = i + 1 end
+end
+
+-- Call `ffi.typeinfo()` much enough to be sure that strings with
+-- names of types become dead. The number of iterations is big
+-- enough (more than x2 of the required value) to see assertion
+-- failure guaranteed under Tarantool too (it has much more alive
+-- objects on start).
+for _ = 1, 100 do
+ ctypes_iteration()
+end
+
+test:ok(true, 'no assertion failure inside ffi.typeinfo()')
+
+test:done(true)
--
2.42.0
More information about the Tarantool-patches
mailing list