Re: [Tarantool-patches] [PATCH luajit] core: fix cdata decrementing

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

Sergey,

On 04.03.21, Sergey Kaplun wrote:
> Hi, Igor!
> 
> Thanks for the review!
> 
> Added your tag:
> | Reviewed-by: Igor Munkin <imun@tarantool.org>
> 
> On 27.02.21, Igor Munkin wrote:
> > Sergey,
> > 
> > Thanks for the patch, nice catch! I guess we missed this bug since the
> > tests do not cover the case for GCcdata object with a custom finalizer.
> > But these are frequently used in Tarantool though. The changes LGTM,
> > except several nits below.
> > 
> > On 16.02.21, Sergey Kaplun wrote:
> > > When cdata has custom finalizer (and so LJ_GC_CDATA_FIN flag) it is not
> > > collected immediately, when lj_cdata_free() is called. Instead, it is
> > > resurrected and marked finalized, so it is collected at the next GC
> > 
> > Minor: AFAIR, the payload is released (or finalized in LuaJIT terms) at
> > the current GC cycle. Hence, the GCobject is marked finalized at this
> > cycle too, but it is processed once more at the next one to be released.
> 
> Yes, as you said below: it is made white (is resurrected) in sweep phase
> and finalized additional payload in GCfinalize phase. The cdata
> object is collected at the next GC cycle.
> 
> I haven't change commit message since it means the same. Please
> correct me if I'm missing something. You may update this commit
> message in way what you want during cherry picking.

Sorry, you're right and v buhgalterii vse pereputali. I left the commit
message unchanged.

> 
> > 
> > > cycle. The reason of the bug is that gc_cdatanum is decremented when
> > > cdata is resurrected too (i.e. twice).
> > > 
> > > This patch excludes cdata decrementing from resurrection branch and
> > > adds corresponding tests.
> > > 
> > > Resolves tarantool/tarantool#5820
> > > Follows up tarantool/tarantool#5187
> > > ---
> > > 
> > > Branch: https://github.com/tarantool/luajit/tree/skaplun/gh-5820-improperly-cdata-counting
> > > Test Branch: https://github.com/tarantool/tarantool/tree/skaplun/gh-5820-improperly-cdata-counting
> > > Issue: https://github.com/tarantool/tarantool/issues/5820
> > > 
> > > ChangeLog entry for bumping LuaJIT:
> > > ===================================================================
> > > ## bugfix/LuaJIT
> > > 
> > > * Fix double cdata decrementing in platform metrics when finalizer is set (gh-5820).
> > > ===================================================================
> > > 
> > >  src/lj_cdata.c                                |  3 +-
> > >  test/misclib-getmetrics-capi.test.lua         | 15 +++++++++-
> > >  test/misclib-getmetrics-capi/testgetmetrics.c | 28 +++++++++++++++++++
> > >  test/misclib-getmetrics-lapi.test.lua         | 15 +++++++++-
> > >  4 files changed, 58 insertions(+), 3 deletions(-)

<snipped>

> > 
> > Side note: It's worth to check whether Tarantool internals do not spoil
> > the counters the same way they do after making LuaJIT testing machinery
> > self-sufficient. There is <collectgarbage> call at the beginning of the
> > test, but I haven't tested your patch against my series.
> 
> I've checked it at the new branch version:
> https://github.com/tarantool/tarantool/tree/skaplun/gh-5820-improperly-cdata-counting
> Looks OK.

Yes, CI is green and this is great. Thanks for checking!

> 

<snipped>

> 
> -- 
> Best regards,
> Sergey Kaplun

-- 
Best regards,
IM