Hi, Sergey, thanks for tha patch! LGTM with a minor comment below. On 26.08.2024 15:37, Sergey Kaplun wrote: > From: Mike Pall > > Thanks to Sergey Kaplun. > > (cherry picked from commit e45fd4cb713b610506213692f3b55a1869febb03) > > `narrow_conv_backprop()` misses the stack pointer (`nc->sp`) limit check > after a bunch of recursive calls that may change its value. As a result, > it leads to stack-buffer-overflow during the instruction narrowing. This > patch adds a missing check. > > Sergey Kaplun: > * added the description and the test for the problem > > Part of tarantool/tarantool#10199 > --- > > Branch:https://github.com/tarantool/luajit/tree/skaplun/lj-1262-fix-limit-narrow-conv-backprop > Related issues: > *https://github.com/tarantool/tarantool/issues/10199 > *https://github.com/LuaJIT/LuaJIT/issues/1262 > > src/lj_opt_narrow.c | 3 +- > ...62-fix-limit-narrow-conv-backprop.test.lua | 61 +++++++++++++++++++ > 2 files changed, 63 insertions(+), 1 deletion(-) > create mode 100644 test/tarantool-tests/lj-1262-fix-limit-narrow-conv-backprop.test.lua > > diff --git a/src/lj_opt_narrow.c b/src/lj_opt_narrow.c > index db0da10f..6b6f20d3 100644 > --- a/src/lj_opt_narrow.c > +++ b/src/lj_opt_narrow.c > @@ -341,7 +341,8 @@ static int narrow_conv_backprop(NarrowConv *nc, IRRef ref, int depth) > NarrowIns *savesp = nc->sp; > int count = narrow_conv_backprop(nc, ir->op1, depth); > count += narrow_conv_backprop(nc, ir->op2, depth); > - if (count <= 1) { /* Limit total number of conversions. */ > + /* Limit total number of conversions. */ > + if (count <= 1 && nc->sp < nc->maxsp) { > *nc->sp++ = NARROWINS(IRT(ir->o, nc->t), ref); > return count; > } > diff --git a/test/tarantool-tests/lj-1262-fix-limit-narrow-conv-backprop.test.lua b/test/tarantool-tests/lj-1262-fix-limit-narrow-conv-backprop.test.lua > new file mode 100644 > index 00000000..6bb4025d > --- /dev/null > +++ b/test/tarantool-tests/lj-1262-fix-limit-narrow-conv-backprop.test.lua > @@ -0,0 +1,61 @@ > +local tap = require('tap') > + > +-- Test file to demonstrate stack-buffer-overflow during the > +-- narrowing optimization. > +-- See also:https://github.com/LuaJIT/LuaJIT/issues/1262. > + > +local test = tap.test('lj-1262-fix-limit-narrow-conv-backprop'):skipcond({ > + ['Test requires JIT enabled'] = not jit.status(), > +}) > + > +test:plan(1) > + > +-- XXX: Test fails only under ASAN. > +-- XXX: The original reproducer was found by fuzzer: > +--https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=70779. > +-- It creates a long side trace with a huge amount of ADD IRs, > +-- which are recursively used in the `narrow_conv_backprop()` many > +-- enough times to catch the stack-buffer-overflow. I can't > +-- simplify the reproducer any more (or write it from scratch), so > +-- I leave it in that state. > + I would remove a sentence "I can't simplify the reproducer ..." from the test, because it is useful information on review and I think it is useless for those who will use the test. Feel free to ignore and leave as is. s/any more/anymore/ > +local DEFAULT_NUMBER = 1 > +local tonumber = tonumber > + > +local always_number = function(val) > + return tonumber(val) or DEFAULT_NUMBER > +end > + > +local add = function(v1, v2) > + return always_number(v1) + always_number(v2) > +end > + > +jit.opt.start('hotloop=1', 'hotexit=1') > + > +local counter_0 = 0 > +local counter_1 = 0 > +local counter_2 = 0 > +local tmp = add(nil, 'Name') > +local Name0 = add(tmp, 'Name') > +-- Start a long side trace here. > +for _ = 0, 0, 0 do > + if counter_0 > 5 then break end > + counter_0 = counter_0 + 1 > + > + for _ = always_number(false), 1, always_number(Name0) do > + if counter_1 > 5 then break end > + counter_1 = counter_1 + 1 > + > + repeat > + if counter_2 > 5 then break end > + counter_2 = counter_2 + 1 > + > + Name0 = Name0 + Name0 + Name0 > + Name0 = add(Name0, nil) + Name0 > + until nil > + end > +end > + > +test:ok(true, 'no stack-buffer-overflow during narrowing') > + > +test:done(true)