Re: [PATCH v2] libctf: ctf_member_next needs to return (ssize_t)-1 on error

[Torbjorn SVENSSON <torbjorn.svensson@foss.st.com>]

On 2023-09-12 16:23, Nick Alcock wrote:
> On 8 Sep 2023, Torbjorn SVENSSON told this:
> 
>> On 2023-09-07 14:10, Nick Alcock wrote:
>>> On 30 Aug 2023, Alan Modra via Binutils outgrape:
>>>
>>>> On Wed, Aug 30, 2023 at 10:34:05AM +0200, Torbjorn SVENSSON wrote:
>>>>> @Alan, any additional comments on this updated patch (except the missing
>>>>> semicolon that I mention below)?
>>>>
>>>> I'm leaving it to Nick Alcock to decide what to do here.
>>> I agree that we should indeed be returning -1 from all functions that
>>> return an int (it used to, but ctf_id_t has to be an unsigned long). But
>>> I think it might be less disruptive to do so via a new
>>> ctf_set_errno_int() which is just like ctf_set_errno but returns an int
>>> rather than an unsigned long. That eliminates a lot of {}ery and makes
>>> each individual hunk smaller.
>>
>> Ok, I can do that instead if that's considered the proper way.
> 
> I think it might be a good bit neater and a good bit less work :)

Ok, I'll send a revised patch in the next few days.

>>> My concern is that it's really hard to validate all this -- can anyone
>>> think of a trick that would emit *consistent* warnings if we called
>>> return (ctf_set_errno()) from a function returning int? I mean this
>>> returning-int thing is a change I made ages ago, and despite making it
>>> *and* attempting to validate on 64-bit Windows I have clearly not got it
>>> right because it's drifted right out of correctness again.
>>>
>>> (Similarly, does anyone have a build/target triplet on which this goes
>>> wrong? because it's not going wrong on any of my mingw64 or cygwin tests
>>> as far as I can tell.)
>>
>> I discovered the issue using the GCC12 package for arm-none-eabi that Arm released
>> (https://developer.arm.com/downloads/-/arm-gnu-toolchain-downloads), but built using the x86_64-w64-mingw GCC compiler on Linux.
> 
> I'm wondering what the configure options were for binutils itself --
> anything? was it a cross at all? (I'll admit this description has
> confused me a bit: is this a cross from mingw64 to aarch64 or what?)

Sorry for not being more clear on this point.
This is the configure line used to reproduce the issue outside a full 
build of an arm-none-eabi toolchain. I'm sure that a few of the 
arguments to configure can be dropped without impact on the problem, but 
I leave that to the reader to decide.

INSTALL=/tmp/pr30836-libctf-inifinit-loop
./configure \
   --build=x86_64-linux-gnu \
   --host=x86_64-w64-mingw32 \
   --target=arm-none-eabi \
   --prefix=$INSTALL \
   --infodir=$INSTALL/share/doc/gcc-arm-none-eabi/info \
   --mandir=$INSTALL/share/doc/gcc-arm-none-eabi/man \
   --htmldir=$INSTALL/share/doc/gcc-arm-none-eabi/html \
   --pdfdir=$INSTALL/share/doc/gcc-arm-none-eabi/pdf \
   --disable-nls \
   --disable-werror \
   --disable-sim \
   --disable-gdb \
   --enable-interwork \
   --enable-plugins \
   --with-sysroot=$INSTALL/arm-none-eabi \
   --with-pkgversion=foo
make -j8 V=1 CFLAGS="-O0 -g"
make install


>> I've opened a ticket for the issue where I've attached 2 object files that you can use to reproduce the issue without needing to
>> rebuild GCC + multlibs to verify the problem.
>> https://sourceware.org/bugzilla/show_bug.cgi?id=30836
> 
> Thanks! I hope I won't need them, but they might well come in handy...

Using the above built binutils (ld.exe in this case) with the 
pr41893-1.o object file is enough to trigger the loop.

>>> I kinda wish we could rely on having C11 -- type-generic macros are made
>>> for cases like this :(
>>
>> Looks like it would be a nice fix indeed, but is there anything else
>> that could be done to improve the situation without needing to go to
>> C11?
> 
> I've been trying to think of something other than adding whatever the
> build is you saw this on to my regular test matrix.

You know both the code and the CTF format better than I do, so maybe you 
can take a look at the .ctf section in pr41893-1.o and see why it is 
wrongly handled. It might well be that the error that I stumbled on (not 
returning -1 on failure) is not the real problem... Maybe the real 
problem is something less obvious in the parsing of the object file itself?