Re: [Tarantool-patches] [PATCH luajit] ARM64: Fix write barrier in BC_USETS.

[Igor Munkin via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Sergey,

Thanks for the fixes! See some new comments below.

On 01.08.21, Sergey Kaplun wrote:
> Igor,
> 
> Thanks for the review!
> Update commit message on the branch, considering you comments.

Got it, but I still have some more comments regarding it.

> 
> See answers to you questions below.
> 

<snipped>

> 
> > 
> > > | ccmp TMP0w, #0, #0, ne
> > > | beq <1 // branch out from barrier movement
> > > `TMP0w` contains `upvalue->closed` field. If it equals NULL (the first
> > > `#0`). The second zero is the value of NZCV condition flags set if the
> > > condition (`ne`) is FALSE [1][2]. If the set value is not white, then
> > > flags are set to zero and branch is not taken (no Zero flag). If it
> > > happens at propagate or atomic GC State and the `lj_gc_barrieruv()`
> > > function is called then the gray value to set is marked as white. That
> > > leads to the assertion failure in the `gc_mark()` function.
> > 
> > OK, I understand almost nothing from the part above. Here are the
> > comments:
> > 1. "If it equals NULL (the first `#0`)", then what?
> 
> My bad:
> I mean here:
> If it equals NULL (the first `#0`), then the upvalue is open.

So why do you use NULL instead of 0? The field is uint8_t type, so 0 is
much clearer.

> Added this.
> 
> > 2. Just to check we are on the same page: the second "immediate"
> > mentioned in docs[1] is NZCV?
> 
> Yes.
> 
> >                               Then beq <1 branch is not taken since
> > (TMP0w != 0) is FALSE (i.e. upvalue is not closed), but zero flag in
> > NZCV value is not set?
> 
> Yes.
> 
> >                        So how does the color of the value to be stored
> > relate to this control flow?
> 
> This NZCV value isn't set if the upvalue is white, because condition is
> of the following instruction
> 
> |    tst TMP1w, #LJ_GC_WHITES	// iswhite(str)
> 
> is TRUE. So the <1 branch is taken, because the upvalue is closed.

Well... I can't imagine how I needed to find this... This relates mostly
to ARM docs you've mentioned, but it would be nice to describe this
behaviour in the commit message (since you're writing a verbose one).

> 
> > 3. AFAICS, if the branch is not taken and <lj_gc_barrieruv> is called at
> > propagate or atomic phase, the value is colored either to gray or black.
> 
> Yes, that leads to the assertion failure mentioned in the ticket in the
> LuaJIT upstream.
> 
> > 
> > > 
> > > This patch changes yielded NZCV condition flag to 4 (Zero flag is up) to
> > > take the correct branch after `ccmp` instruction.
> > > 
> > > Sergey Kaplun:
> > > * added the description and the test for the problem
> > > 
> > > [1]: https://developer.arm.com/documentation/dui0801/g/pge1427897656225
> > > [2]: https://community.arm.com/developer/ip-products/processors/b/processors-ip-blog/posts/condition-codes-1-condition-flags-and-codes
> > 
> > Minor: Why #5629 is not mentioned?
> 
> Added.

Considering everything above, I propose the following wording:
| Contributed by Javier Guerra Giraldez.
|
| (cherry picked from commit c785131ca5a6d24adc519e5e0bf1b69b671d912f)
|
|
| Closed upvalues are never gray. Hence when closed upvalue is marked, it
| is marked as black. Black objects can't refer white objects, so for
| storing a white value in a closed upvalue, we need to move the barrier
| forward and color our value to gray by using `lj_gc_barrieruv()`. This
| function can't be called on closed upvalues with non-white values since
| there is no need to mark it again.
|
| USETS bytecode for arm64 architecture has the incorrect NZCV condition
| flag value in the instruction that checks the upvalue is closed:
| | tst TMP1w, #LJ_GC_WHITES
| | ccmp TMP0w, #0, #0, ne
| | beq <1 // branch out from barrier movement
| `TMP0w` contains `upvalue->closed` field, so the upvalue is open if this
| field equals to zero (the first one in `ccmp`). The second zero is the
| value of NZCV condition flags[1] yielded if the specified condition
| (`ne`) is met for the current values of the condition flags[2]. Hence,
| if the value to be stored is not white (`TMP1w` holds its color), then
| the condition is FALSE and all flags bits are set to zero so branch is
| not taken (Zero flag is not set). If this happens at propagate or atomic
| GC phase, the `lj_gc_barrieruv()` function is called and the gray value
| to be set is marked like if it is white. That leads to the assertion
| failure in the `gc_mark()` function.
|
| This patch changes NZCV condition flag to 4 (Zero flag is set) to take
| the correct branch after `ccmp` instruction.
|
| Sergey Kaplun:
| * added the description and the test for the problem
|
| [1]: https://community.arm.com/developer/ip-products/processors/b/processors-ip-blog/posts/condition-codes-1-condition-flags-and-codes
| [2]: https://developer.arm.com/documentation/dui0801/g/pge1427897656225
|
| Part of tarantool/tarantool#5629

> 

<snipped>

> > 
> > > 
> > >  src/vm_arm64.dasc                             |  2 +-
> > >  ...6-arm64-incorrect-check-closed-uv.test.lua | 38 +++++++++++++++++++
> > >  2 files changed, 39 insertions(+), 1 deletion(-)
> > >  create mode 100644 test/tarantool-tests/lj-426-arm64-incorrect-check-closed-uv.test.lua
> > > 
> > 
> > <snipped>
> > 
> > > diff --git a/test/tarantool-tests/lj-426-arm64-incorrect-check-closed-uv.test.lua b/test/tarantool-tests/lj-426-arm64-incorrect-check-closed-uv.test.lua
> > > new file mode 100644
> > > index 00000000..b757133f
> > > --- /dev/null
> > > +++ b/test/tarantool-tests/lj-426-arm64-incorrect-check-closed-uv.test.lua
> > > @@ -0,0 +1,38 @@
> > > +local tap = require('tap')
> > > +
> > > +local test = tap.test('lj-426-arm64-incorrect-check-closed-uv')
> > > +test:plan(1)
> > > +
> > > +-- Test file to demonstrate LuaJIT USETS bytecode incorrect
> > > +-- behaviour on arm64 in case when non-white object is set to
> > > +-- closed upvalue.
> > > +-- See also, https://github.com/LuaJIT/LuaJIT/issues/426.
> > > +
> > > +-- First, create a closed upvalue.
> > > +do
> > 
> > Minor: I'm not sure, we need a separate lexical block here. Could you
> > please clarify the reason in the comment?
> 
> We need a closed upvalue. I suppose that it is the simpiest way to
> create one. Please, provide a simplier example if you know one.

My bad. Yes, the easiest way to emit UCLO bytecode is using a separate
lexical block.

> 
> > 
> > > +  local uv -- luacheck: no unused
> > > +  -- The function's prototype is created with the following
> > > +  -- constants at chunk parsing. After adding this constant to
> > > +  -- the function's prototype it will be marked as gray during
> > > +  -- propogate phase.
> > 
> > Then what does it test, if the constant is marked as gray? Will this
> > string be white later?
> 
> It shouldn't be white, it should be gray, otherwise the aforementioned
> condition is TRUE (remember, we need FALSE).

Again, PEBKAC, thanks for the explanation.

> 
> > 
> > > +  local function usets() uv = '' end
> > > +  _G.usets = usets
> > > +end
> > > +
> > > +-- Set GC state to GCpause.
> > > +collectgarbage()
> > > +-- Do GC step as often as possible.
> > > +collectgarbage('setstepmul', 100)
> > 
> > Minor: Don't get, why you need to make GC less aggressive for the test.
> > The test is run, until propagate phase is finished.
> 
> More likely, that it is run, until the upvalue is marked as black
> during traversing (with the bug). I can remove this line if you insist.

Drop it, please. I can't even *feel* its effect ;)

> 
> > 
> > > +
> > > +-- We don't know on what exactly step our upvalue is marked as
> > > +-- black and USETS become dangerous, so just check it at each
> > > +-- step.
> > > +-- Don't need to do the full GC cycle step by step.

Minor: It would be nice to drop a few words about string and upvalue
colours during this loop, but it's up to you.

> > > +local old_steps_atomic = misc.getmetrics().gc_steps_atomic
> > > +while (misc.getmetrics().gc_steps_atomic == old_steps_atomic) do
> > > +  collectgarbage('step')
> > > +  usets() -- luacheck: no global
> > > +end
> > > +
> > > +test:ok(true)
> > > +os.exit(test:check() and 0 or 1)
> > > -- 
> > > 2.31.0
> > > 
> > 
> > [1]: https://lists.tarantool.org/tarantool-patches/20210719073632.12008-1-skaplun@tarantool.org/T/#u
> > 
> > -- 
> > Best regards,
> > IM
> 
> -- 
> Best regards,
> Sergey Kaplun

-- 
Best regards,
IM