Re: [PATCH v3] nptl_db: Support different libpthread/ld.so load orders (bug 27744)

[Florian Weimer <fweimer@redhat.com>]

* H. J. Lu:

> On Thu, Apr 22, 2021 at 10:16 AM Szabolcs Nagy via Gdb-patches
> <gdb-patches@sourceware.org> wrote:
>>
>> The 04/22/2021 16:25, Florian Weimer wrote:
>> > * Szabolcs Nagy:
>> >
>> > > if i rerun the link command of the test exe but with -no-pie
>> > > instead of -pie then the test passes with that binary.
>> > >
>> > > i suspect gdb places the breakpoint at the wrong place in pie
>> > > for some reason. can be ubuntu tooling specific. see the
>> > > breakpoint location (the exe base offset is missing):
>> >
>> > Thanks for investigating.
>> >
>> > > +attach 1254516
>> > > [New LWP 1254517]
>> > > Trying host libthread_db library: /home/szabolcs/try/build/nptl_db/libthread_db.so.1.
>> > > td_ta_new failed: application not linked with libthread
>> > > thread_db_load_search returning 0
>> > > Trying host libthread_db library: /home/szabolcs/try/build/nptl_db/libthread_db.so.1.
>> > > [Thread debugging using libthread_db enabled]
>> > > Using host libthread_db library "/home/szabolcs/try/build/nptl_db/libthread_db.so.1".
>> > > thread_db_load_search returning 1
>> > > 0x0000ffff9d3d89c4 in __futex_abstimed_wait_common64 (futex_word=0xffff9d363210, expected=1254517, clockid=<optimized out>, abstime=0x0, private=<optimized out>, cancel=cancel@entry=true) at futex-internal.c:74
>> > > 74          err = INTERNAL_SYSCALL_CANCEL (futex_time64, futex_word, op, expected,
>> > > +break debugger_inspection_point
>> > > Breakpoint 1 at 0x20c0: file tst-pthread-gdb-attach.c, line 123.
>> > > +continue
>> >
>> > Would you please check if the issue goes away if you replace
>> >
>> >            "add-symbol-file %1$s/nptl/tst-pthread-gdb-attach\n"
>> >
>> > with
>> >            "file %1$s/nptl/tst-pthread-gdb-attach\n"
>> >
>> > ?
>> >
>> > (I assume this happens without --enable-hardcoded-path-in-tests.)
>>
>> yes, it seems gdb does not work with
>>
>> ld.so ./exe
>>
>> if exe is pie. i could not get it to work wit file either.
>
> Just pass -no-pie to build exe.

Ah right, the penny finally dropped.  I've seen this failure before
during regular glibc debugging.  Fedora apparently has a downstream-only
patch. 8-/

This is what I came up with:

nptl: Do not build nptl/tst-pthread-gdb-attach as PIE

diff --git a/nptl/Makefile b/nptl/Makefile
index a3d1ef8d66..294bb2faa4 100644
--- a/nptl/Makefile
+++ b/nptl/Makefile
@@ -377,6 +377,9 @@ endif
 CFLAGS-tst-pthread-gdb-attach-static.c := $(CFLAGS-printers-tests)
 CPPFLAGS-tst-pthread-gdb-attach-static.c := \
   $(CFLAGS-printers-tests) -DDO_ADD_SYMBOL_FILE=0
+# As of version 9.2, GDB cannot attach properly to PIE programs that
+# were launched with an explicit ld.so invocation.
+tst-pthread-gdb-attach-no-pie = yes
 
 ifeq ($(build-shared),yes)
 tests-printers-libs := $(shared-thread-library)

It's not needed for the static test even with --enable-static-pie.

Thanks,
Florian