Re: [Tarantool-patches] [PATCH luajit] test: harden skipcond for libc in LuaJIT-tests

[Sergey Bronnikov via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

[-- Attachment #1: Type: text/plain, Size: 4313 bytes --]

Hi, Sergey,


On 20.12.2024 15:15, Sergey Kaplun wrote:
> Hi, Sergey!
> Thanks for the review!
> See my answers below.
>
> On 20.12.24, Sergey Bronnikov wrote:
>> Hello, Sergey,
>>
>> thanks for the patch! See comments below.
>>
>>
>> On 19.12.2024 12:30, Sergey Kaplun wrote:
>>> The `strtod parsing` subtest in the <lib/base/tonumber_scan.lua> checks
>>> the results yielded by the `strtod()` via FFI call. In glibc versions
>>> before 2.28 it returns an incorrect result (NaN instead of inf) for
>>> "0x3p1023" [1]. This patch hardens the skipcond for this test for a
>>> smaller version of the libc installed.
>>>
>>> [1]:https://sourceware.org/bugzilla/show_bug.cgi?id=23279
>>> ---
>>>
>>> Branch:https://github.com/tarantool/luajit/tree/skaplun/gh-noticket-fix-glibc-versions
>>>
>>> The CI failure for libc 2.23 (ubuntu 16):
>>> https://github.com/tarantool/tarantool/actions/runs/12376159316/job/34542626889?pr=10949#step:5:13916
>> Ubuntu 16.04 EOL and EOS are reached [1]
>>> The CI failure for libc 2.24 (debian 9):
>>> https://github.com/tarantool/tarantool/actions/runs/12376159307/job/34542633454?pr=10949#step:5:13859
>> Debian 9 EOL is reached 4 years ago [2]
>>> The CI failure for libc 2.27 (ubuntu 18):
>>> https://github.com/tarantool/tarantool/actions/runs/12376159278/job/34542625496?pr=10949#step:5:11743
>>> The CI pass for libc 2.31 (ubuntu 20):
>>> https://github.com/tarantool/tarantool/actions/runs/12376159289/job/34542630993?pr=10949
>> Ubuntu 20.04 support ends in 3 months and 2 week [1].
> Yes, unfortunately we should keep (AFAICS) these runners to check the
> installations for our customers.
>
I believe we will clarify a list of Linux distros in a separate activity.


>>
>>>    test/LuaJIT-tests/lib/base/tonumber_scan.lua | 8 +++++---
>>>    1 file changed, 5 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/test/LuaJIT-tests/lib/base/tonumber_scan.lua b/test/LuaJIT-tests/lib/base/tonumber_scan.lua
>>> index ac7d68a4..800c57b0 100644
>>> --- a/test/LuaJIT-tests/lib/base/tonumber_scan.lua
>>> +++ b/test/LuaJIT-tests/lib/base/tonumber_scan.lua
>>> @@ -186,9 +186,11 @@ do --- tonumber parsing
>>>      test_conv(tonumber)
>>>    end
>>>    
>>> --- Skip for the old libc version with the bug in the `strtod()`.
>>> --- See alsohttps://sourceware.org/bugzilla/show_bug.cgi?id=16151.
>>> -do --- strtod parsing -libc<2.19
>>> +-- Skip for the old libc version with the bugs in the `strtod()`.
>>> +-- Affected versions:
>>> +-- * <2.19:https://sourceware.org/bugzilla/show_bug.cgi?id=16151.
>>> +-- * <2.28:https://sourceware.org/bugzilla/show_bug.cgi?id=23279.
>> I propose to skip strtod test when Glibc is older than say 2.30 [3] and
>> avoid testing
>>
>> with ancient Glibc versions because Glibc implementation is buggy [4] and
>>
>> a there are lot of bugs in old and more or less new versions:
>>
>> v2.41 [5], v2.22 [6], v2.27 [7], v2.15 [8], v2.33 [9] etc [10].
>>
>>
>> I don't think we should waste our time with issues like this.
>>
>>
>> 1.https://endoflife.date/ubuntu
>>
>> 2.https://endoflife.date/debian
>>
>> 3.https://sourceware.org/glibc/wiki/Release
>>
>> 4.
>> https://github.com/tarantool/luajit/blob/cd5ab5da9b56bf560407554c0377cfa8567d4430/src/lj_strscan.c#L20-L27
>>
>> 5.https://sourceware.org/bugzilla/show_bug.cgi?id=30220
>>
>> 6.https://sourceware.org/bugzilla/show_bug.cgi?id=19380
>>
>> 7.https://sourceware.org/bugzilla/show_bug.cgi?id=15744
>>
>> 8.https://sourceware.org/bugzilla/show_bug.cgi?id=14049
>>
>> 9.https://sourceware.org/bugzilla/show_bug.cgi?id=26137
>>
>> 10.
>> https://sourceware.org/bugzilla/buglist.cgi?bug_status=__all__&content=strtod&no_redirect=1&order=bug_id%20DESC&product=glibc&query_format=specific
> Yes, I agree that this should be the last red line for us. If the test
> will fail again due to glibc, it will be easier just to drop it. For
> now, I see no failures for versions 2.31 -- 2.40, and find no bugs
> related to our tests that affected versions 2.28 -- 2.30, so let's skip
> it for versions less than 2.28 for now.

Okay, LGTM then.

Also discussed it and decided to remove the test if the same issue will 
happen next time.

>
>>> +do --- strtod parsing -libc<2.28
>>>      test_conv(function(s)
>>>        local d = ffi.C.strtod(s, e)
>>>        return (e[0][0] == 0 and #s ~= 0) and d or nil

[-- Attachment #2: Type: text/html, Size: 8039 bytes --]