Re: [pushed] [gdb/testsuite] Use maint ignore-probes in gdb.base/longjmp.exp

[Luis Machado <luis.machado@arm.com>]

On 2/8/23 20:36, Tom de Vries wrote:
> On 2/8/23 19:06, Luis Machado wrote:
>> On 2/8/23 15:38, Tom de Vries wrote:
>>> On 2/8/23 15:51, Luis Machado wrote:
>>>> On 2/8/23 14:48, Tom de Vries wrote:
>>>>> On 2/8/23 14:27, Luis Machado wrote:
>>>>>> Hi Tom,
>>>>>>
>>>>>> Is the entire test supposed to PASS? I'm seeing the following on my aarch64/Ubuntu 22.04 setup:
>>>>>>
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 2: next over call_longjmp (the program is no longer running)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 2: next over setjmp (the program is no longer running)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 2: setup: breakpoint at pattern start (got interactive prompt)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 2: setup: breakpoint at safety net (got interactive prompt)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 2: setup: continue to breakpoint at pattern start (the program exited)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 3: next over pattern (the program is no longer running)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 3: setup: breakpoint at pattern start (got interactive prompt)
>>>>>> FAIL: gdb.base/longjmp.exp: with_probes=0: pattern 3: setup: continue to breakpoint at pattern start (the program is no longer running)
>>>>>>
>>>>>> Maybe something is genuinely broken for aarch64 though, or I'm missing some packages/debuginfo.
>>>>>
>>>>> Hi,
>>>>>
>>>>> I just ran this test-case on openSUSE Leap 15.4 aarch64, no problems found.
>>>>>
>>>>
>>>> Alright. That's good to know.
>>>
>>> FWIW, I've tried this test-case also on various x86_64 distros other than the usual openSUSE Leap 15.4: ubuntu 20.04, fedora 37 and opensuse tumbleweed, again no problems found.
>>
>> I did a brief investigation on this one, and gdb seems to be doing something strange.
>>
>> For Ubuntu 20.04 we have the following, just after deleting the breakpoints leading into the pattern 2 check:
>>
>>
>> (gdb) info source
>> Current source file is longjmp.c
>> Compilation directory is /build/glibc-RIFKjK/glibc-2.31/setjmp
>> Located in /repos/binutils-gdb/gdb/testsuite/gdb.base/longjmp.c
>> Contains 82 lines.
>> Source language is c.
>> Producer is GNU C11 9.4.0 -moutline-atomics -mlittle-endian -mabi=lp64 -g -O2 -std=gnu11 -fgnu89-inline -fmerge-all-constants -frounding-math
>>   -fstack-protector-strong -fmath-errno -fPIC -ftls-model=initial-exec -fasynchronous-unwind-tables -fstack-protector-strong -fstack-clash-pro
>> tection.
>> Compiled with DWARF 4 debugging format.
>> Does not include preprocessor macro info.
>> (gdb) b 69
>> Breakpoint 4 at 0xaaaaaaaa08ec: file /builds/binutils-gdb-arm64-focal/gdb/testsuite/../../../../repos/binutils-gdb/gdb/tes
>> tsuite/gdb.base/longjmp.c, line 69.
>> (gdb)
>>
>> And for Ubuntu 22.04:
>>
>> (gdb) info source
>> Current source file is ./setjmp/longjmp.c
>> Compilation directory is ./setjmp
>> Located in /repos/binutils-gdb/gdb/testsuite/gdb.base/longjmp.c
>> Contains 82 lines.
>> Source language is c.
>> Producer is GNU C11 11.2.0 -mlittle-endian -mabi=lp64 -g -O2 -std=gnu11 -fgnu89-inline -fmerge-all-constants -frounding-math -fstack-protecto
>> r-strong -fno-common -fmath-errno -fPIC -ftls-model=initial-exec -fasynchronous-unwind-tables -fstack-protector-strong -fstack-clash-protecti
>> on.
>> Compiled with DWARF 5 debugging format.
>> Does not include preprocessor macro info.
>> (gdb) b 69
>> No line 69 in the current file.
>> Make breakpoint pending on future shared library load? (y or [n]) n
>> (gdb)
>>
>> There is a small difference in debug info (dwarf 4 for 20.04 and dwarf 5 for 22.04), source file name and compilation directory.
>>
>> What is strange is that gdb's 'info source' output seems to refer to the glibc longjmp source file as the current one. And the compilation directory
>> is also glibc's. The "Located in" field is from the testcase source, also named longjmp.c. The "Contains" line is also based on the testcase source file.
>>
>> Investigating further, if you "list", it will output the sources from the testcase file as well.
>>
>> Finally, for 20.04, the "break" command will use the testcase source file, but in 22.04 it will use the glibc source file. I'm guessing the fact that glibc's
>> source file in 20.04 is also called longjmp.c makes it work somehow. But in 22.04 the glibc source file is now ./setjmp/longjmp.c, and I guess gdb now
>> attempts to insert a breakpoint in the glibc source file, which doesn't have line 63. So it all goes downhill from there.
>>
>> I'm not sure if this is a long-standing bug or if it is a somewhat recent regression. But gdb seems to be genuinely confused about which source file is the current one
>> and which one to use for various commands.
>>
>> I'd expect gdb to pick one and stick with it, but it doesn't seem to be the case.
>>
>> Maybe we just uncovered a new bug with source handling.
> 
> I suspect the FAILs will disappear if we replace "break <n>" with "break $srcfile:<n>".  I'm not sure yet whether this is a fix or a workaround.

I suppose. But it seems there is a different underlying issue of gdb getting confused about what is the current source file.

> 
> Please file a PR and attach the entire gdb.log, I want to take a look at it.

Will do.

> 
> Thanks,
> - Tom