Tarantool development patches archive
 help / color / mirror / Atom feed
From: Sergey Ostanevich via Tarantool-patches <tarantool-patches@dev.tarantool.org>
To: Sergey Kaplun <skaplun@tarantool.org>
Cc: tarantool-patches@dev.tarantool.org
Subject: Re: [Tarantool-patches] [PATCH luajit] ARM64: Fix write barrier in BC_USETS.
Date: Tue, 10 Aug 2021 19:40:45 +0300
Message-ID: <4BF6CB05-7904-4918-81BA-73149399FD04@tarantool.org> (raw)
In-Reply-To: <YRFRcjqezSJHtwyf@root>

[-- Attachment #1: Type: text/plain, Size: 10635 bytes --]

Hi!

Thanks for the patch, LGTM.

I have no other explanation than that ‘missed flag as a result of compare’.
But you did it better - took time for me to get to the point. 
BTW, will make easier to grasp with a simple C equivalent, like 

-	if (iswite(str) || closed != 0)
+	if (iswite(str) || closed == 0)

Sergos.



> On 9 Aug 2021, at 19:01, Sergey Kaplun <skaplun@tarantool.org> wrote:
> 
> Igor, thanks for the feedback!
> 
> On 08.08.21, Igor Munkin wrote:
>> 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.
> 
> Changed.
> 
>> 
>>> 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
> 
> Updated, as you've suggested.
> 
>> 
>>> 
>> 
>> <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 ;)
> 
> Done.
> 
>> 
>>> 
>>>> 
>>>>> +
>>>>> +-- 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.
> 
> Added.
> 
> The iterative patch is the following:
> 
> ===================================================================
> 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
> index b757133f..4cdf1211 100644
> --- 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
> @@ -21,9 +21,10 @@ end
> 
> -- Set GC state to GCpause.
> collectgarbage()
> --- Do GC step as often as possible.
> -collectgarbage('setstepmul', 100)
> 
> +-- We want to wait for the situation, when upvalue is black,
> +-- the string is gray. Both conditions are satisfied, when the
> +-- corresponding `usets()` function is marked, for example.
> -- 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.
> ===================================================================
> 
>> 
>>>>> +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
> 
> -- 
> Best regards,
> Sergey Kaplun


[-- Attachment #2: Type: text/html, Size: 41907 bytes --]

  parent reply	other threads:[~2021-08-10 16:40 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-07-07 14:36 Sergey Kaplun via Tarantool-patches
2021-08-01 10:39 ` Igor Munkin via Tarantool-patches
2021-08-01 17:00   ` Sergey Kaplun via Tarantool-patches
2021-08-08 19:28     ` Igor Munkin via Tarantool-patches
2021-08-09 16:01       ` Sergey Kaplun via Tarantool-patches
2021-08-09 19:46         ` Igor Munkin via Tarantool-patches
2021-08-10 16:40         ` Sergey Ostanevich via Tarantool-patches [this message]
2021-08-11  5:57           ` Vitaliia Ioffe via Tarantool-patches
2021-08-11  7:22 ` Igor Munkin via Tarantool-patches

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=4BF6CB05-7904-4918-81BA-73149399FD04@tarantool.org \
    --to=tarantool-patches@dev.tarantool.org \
    --cc=sergos@tarantool.org \
    --cc=skaplun@tarantool.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Tarantool development patches archive

This inbox may be cloned and mirrored by anyone:

	git clone --mirror https://lists.tarantool.org/tarantool-patches/0 tarantool-patches/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 tarantool-patches tarantool-patches/ https://lists.tarantool.org/tarantool-patches \
		tarantool-patches@dev.tarantool.org.
	public-inbox-index tarantool-patches

Example config snippet for mirrors.


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git