On Mon, Jul 29, 2024 at 3:03 PM Kienan Stewart <kstew...@efficios.com> wrote:
Hi Brian,
On 7/25/24 3:54 PM, Brian Hutchinson wrote:
Hi Kienan,
I'll answer your questions below, but I've got questions on what I saw
building and installing lttng-tools (2.13.13) and lttng-ust (2.13.8).
Based on the struggles I've had trying to get lttng to work with my
app over various Yocto versions (Dunfell & Kirkstone) and lttng
version, I think the problems I'm facing are mostly around C++ and
weak and hidden symbols in Yocto toolchain.
When I started my app with the options you mentioned previously a
while back, Id see things like:
# LTTNG_UST_DEBUG=1 LTTNG_UST_REGISTER_TIMEOUT=-1 /opt/tc/TrafficController
liblttng_ust_tracepoint[4012/4012]: Your compiler treats weak symbols
with hidden visibility for integer objects as SAME address between
compile units part of the same module. (in check_weak_hidden() at
tracepoint.c:1012)
liblttng_ust_tracepoint[4012/4012]: Your compiler treats weak symbols
with hidden visibility for pointer objects as SAME address between
compile units part of the same module. (in check_weak_hidden() at
tracepoint.c:1016)
liblttng_ust_tracepoint[4012/4012]: Your compiler treats weak symbols
with hidden visibility for 24-byte structure objects as SAME address
between compile units part of the same module. (in check_weak_hidden()
at tracepoint.c:1020)
These messages are extra information for debugging and not indicative of
a problem in of itself. C.f.
https://github.com/lttng/lttng-ust/blob/24f7193c9b918bf714a40e9fc908eeb4978ada1c/src/lib/lttng-ust-tracepoint/tracepoint.c#L1010
There is a unit test related to this:
https://github.com/lttng/lttng-ust/blob/24f7193c9b918bf714a40e9fc908eeb4978ada1c/tests/unit/gcc-weak-hidden/main.c#L76
I further researched this whole 'weak symbol' and 'hidden visibility'
topic in the lttng-dev archives and it smells a lot like what I've
been seeing. You should be able to mix both tracef and tracepoint
calls in souce code ... but I could not. I could get a tracef call to
work but if I put a tracepoint call in the same code then nothing
would work. This was with Dunfell 3.1.7 and earlier versions of
lttng.
At one point I could get a tracepoint call to work but I'd have to let
our cmake build system build and link the tpp.c file and then turn
around and use gcc to recompile it and copy it to where all the
objects were to create the huge .a library the app was built against.
That's when I first learned there are issues with C++. I think g++ is
used to build even .c files that aren't c++.
Then if I tried to put a tracepoint in another sub project, none of
the tracepoints would work and I'd get empty traces. This is a
symptom of the 'weak symbols with hidden visibility' issue ... and I
finally found others that were having same issue in the archives. I
don't fully understand the issues here, although I do understand some
of what's going on ... I just don't know what to do about it.
You said initially said that you're using `lttng_ust_tracepoint` exactly
as the hello world from the documentation; however, you have just
described several attempts at doing different things. Which case are we
trying to understand here?
lttng_ust_tracepoint. I only mentioned prior tests for context to
similar struggles from a year or more ago.
At this point I was being encouraged to keep upgrading to newer
versions of lttng. Our app never changed, gcc & lttng etc., kept
changing. Now with newer versions nothing runs, all I get is an
immediate segfault. Again, I'm building just like I did before a year
or so ago with older versions of Yocto and lttng. I say all of that
to give perspective and history of what I've seen and experienced.
Now this TLS thing has entered the picture too and so far I've only
changed lttng, I don't know if I should be applying patches to my gcc
for that issue. Like I said, I'm currently using Yocto Kirkstone
4.0.18 and 6.1.38 kernel.
Now I'll move into the area of things I've seen building/installing
lttng-tools and lttng-ust natively on the target environment I've
setup where I can run 'make check' etc. These are in the category of
"hey, is this ok, should I be worried about this":
While building lttng-tools I see things like:
*** Warning: Linking the executable userspace-probe-elf-binary against
the loadable module
*** libfoo.so is not portable!
The library is for a test program. My understanding is that the library
is compiled that way to force a stripped shared object to be produced in
order to validate that symbol lookups in libraries with no symtab
function as expected by using the dynsym table.
C.f.
https://github.com/lttng/lttng-tools/commit/ef3dfe5d31c88fb548189a6441aaf8b2afc0bd4b
In file included from ../../../src/common/macros.h:15,
from ../../../include/lttng/health-internal.h:19,
from lttng-ctl-health.c:19:
In function 'lttng_strnlen',
inlined from 'lttng_strncpy' at ../../../src/common/macros.h:128:6,
inlined from 'set_health_socket_path' at lttng-ctl-health.c:146:9,
inlined from 'lttng_health_query' at lttng-ctl-health.c:264:8:
../../../src/common/compat/string.h:19:16: warning: 'strnlen'
specified bound 4096 may exceed source size 37 [-Wstringop-overread]
19 | return strnlen(str, max);
| ^~~~~~~~~~~~~~~~~
lttng-ctl-health.c: At top level:
cc1: note: unrecognized command-line option
'-Wno-incomplete-setjmp-declaration' may have been intended to silence
earlier diagnostics
This warning is addressed in
https://github.com/lttng/lttng-tools/commit/b25a59916106e5055be516f61f183a48f459b0b3
** Warning: Linking the shared library libbar.la against the loadable module
*** libzzz.so is not portable!
*** Warning: Linking the shared library libfoo.la against the loadable module
*** libbar.so is not portable!
While installing lttng-tools I see things like this:
make[4]: Entering directory '/opt/lttng/lttng-tools-2.13.13/src/lib/lttng-ctl'
CC lttng-ctl.lo
CC snapshot.lo
CC lttng-ctl-health.lo
In file included from ../../../src/common/macros.h:15,
from ../../../include/lttng/health-internal.h:19,
from lttng-ctl-health.c:19:
In function 'lttng_strnlen',
inlined from 'lttng_strncpy' at ../../../src/common/macros.h:128:6,
inlined from 'set_health_socket_path' at lttng-ctl-health.c:146:9,
inlined from 'lttng_health_query' at lttng-ctl-health.c:264:8:
../../../src/common/compat/string.h:19:16: warning: 'strnlen'
specified bound 4096 may exceed source size 37 [-Wstringop-overread]
19 | return strnlen(str, max);
| ^~~~~~~~~~~~~~~~~
lttng-ctl-health.c: At top level:
cc1: note: unrecognized command-line option
'-Wno-incomplete-setjmp-declaration' may have been intended to silence
earlier diagnostics
Making install in trigger-condition-event-matches
make[2]: Entering directory
'/opt/lttng/lttng-tools-2.13.13/doc/examples/trigger-condition-event-matches'
CC instrumented-app.o
CC tracepoint-trigger-example.o
AR libtracepoint-trigger-example.a
ar: `u' modifier ignored since `D' is the default (see `U')
While building lttng-ust I see things like:
Making all in utils
make[2]: Entering directory
'/home/iadmin/lttng-ust/lttng-ust-2.13.8/tests/utils'
CC tap.o
AR libtap.a
ar: `u' modifier ignored since `D' is the default (see `U')
While libtool now uses `cr` by default, automake still defines the
default to `cru` which is what ends up getting used in the example.
Since many distros have changed the configuration of ar such that 'D' is
the default rather than the previous behaviour 'U', 'u' is redundant.
The behaviour in automake has been changed in automake 1.16.90+.
C.f.
https://github.com/autotools-mirror/libtool/commit/418129bc63afc312701e84cb8afa5ca413df1ab5
C.f.
http://git.savannah.gnu.org/cgit/automake.git/commit/?id=8cdbdda5aec652c356fe6dbba96810202176ae75
*** Warning: Linking the shared library libzero.la against the
loadable module
*** libfakeust0.so is not portable!
CCLD app_noust_indirect_abi0
*** Warning: Linking the executable app_noust_indirect_abi0 against
the loadable module
*** libzero.so is not portable!
CC app_noust_indirect_abi0_abi1-app_noust.o
CC libone.lo
CCLD libone.la
CCLD app_noust_indirect_abi0_abi1
*** Warning: Linking the executable app_noust_indirect_abi0_abi1
against the loadable module
*** libzero.so is not portable!
*** Warning: Linking the executable app_noust_indirect_abi0_abi1
against the loadable module
*** libone.so is not portable!
CC app_noust_indirect_abi1-app_noust.o
CCLD app_noust_indirect_abi1
*** Warning: Linking the executable app_noust_indirect_abi0_abi1
against the loadable module
*** libone.so is not portable!
CC app_noust_indirect_abi1-app_noust.o
CCLD app_noust_indirect_abi1
*** Warning: Linking the executable app_noust_indirect_abi1 against
the loadable module
*** libone.so is not portable!
CC app_ust.o
CC tp.o
CCLD app_ust
CC app_ust_dlopen.o
CCLD app_ust_dlopen
CC app_ust_indirect_abi0-app_ust.o
CC app_ust_indirect_abi0-tp.o
CCLD app_ust_indirect_abi0
*** Warning: Linking the executable app_ust_indirect_abi0 against the
loadable module
*** libzero.so is not portable!
CC app_ust_indirect_abi0_abi1-app_ust.o
CC app_ust_indirect_abi0_abi1-tp.o
CCLD app_ust_indirect_abi0_abi1
*** Warning: Linking the executable app_ust_indirect_abi0_abi1 against
the loadable module
*** libzero.so is not portable!
I don't know if these are ok or if I should be worried about any of that.
These are all for different tests.
... now on to your questions below.
On Wed, Jul 24, 2024 at 12:04 PM Kienan Stewart <kstew...@efficios.com> wrote:
Hi Brian,
On 7/22/24 6:00 PM, Brian Hutchinson wrote:
Hi Kienan,
Took a while to gather your grocery list but I think I have most of it
below ;)
thanks for all the extra info. Replies inline below, but I'll cut a lot
of the long output for readability.
tl;dr the environment continues to be weird, but my present suspicion is
that something in either compilation, the linking of your app (eg. with
ld when producing the executable), or some post linking stripping might
be causing issues.
I'm not aware of any stripping that's going on. In fact everything is
being built with debug symbols at the moment and I even turned off
optimization ... even used the debug friendly -O flag to see if that
made a difference.
I will stop digging into further hypotheticals on my side as there is no
reproducer for both the environment and the application. If you ever end
up with a minimal reproducer that you can share, I'd be more than happy
to examine it.
I'm planning on trying to make a small reproducer I can share but not there yet.
Great! I appreciate that you're taking the time to do so.
I may have not been clear. Most of the application components are
statically linked but I think there are some that are built as shared
objects (.so's) so that's what I was referring to. I know that
lttng-ust is dynamically linked ... I think the lttng-ust docs say this
is only option but also makes reference to the fact static linking was
once possible (in some versions of the documentation) but not supported
anymore (I probably have the docs memorized by now ha, ha ... I've
looked at many, many versions of them).
Just for full disclosure my ldd looks like:
linux-vdso.so.1 (0x0000ffffab196000)
libfcgi.so.0 => /usr/lib/libfcgi.so.0 (0x0000ffffa57f0000)
liblttng-ust.so.1 => /usr/lib/liblttng-ust.so.1
(0x0000ffffa5750000)
libxml2.so.2 => /usr/lib/libxml2.so.2 (0x0000ffffa55d0000)
librt.so.1 => /lib/librt.so.1 (0x0000ffffa55b0000)
libm.so.6 => /lib/libm.so.6 (0x0000ffffa5510000)
libstdc++.so.6 => /usr/lib/libstdc++.so.6 (0x0000ffffa52f0000)
libgcc_s.so.1 => /lib/libgcc_s.so.1 (0x0000ffffa52c0000)
libc.so.6 => /lib/libc.so.6 (0x0000ffffa5110000)
/lib/ld-linux-aarch64.so.1 (0x0000ffffab15d000)
liblttng-ust-common.so.1 =>
/usr/local/lib/liblttng-ust-common.so.1 (0x0000ffffa50e0000)
liblttng-ust-tracepoint.so.1 =>
/usr/local/lib/liblttng-ust-tracepoint.so.1 (0x0000ffffa50a0000)
libpthread.so.0 => /lib/libpthread.so.0 (0x0000ffffa5080000)
libz.so.1 => /lib/libz.so.1 (0x0000ffffa5050000)
I find it very suspicious that `liblttng-ust.so.1` is in `/usr/lib`,
while the other lttng-ust libraries are being loaded from `/usr/local/lib`.
So Yocto puts all of the lttng libs into /usr/lib. When I sent the
previous info I was using lttng-tools and modules built by Yocto/OE
and I setup a native build environment on the target so I could run
'make check' etc., and that's why there were things in /usr/local/lib
because that's where you guys want stuff to be. So I actually left
the lttng-ust installables in /usr/local/build but also copied them to
/usr/lib to overwrite old Yocto versions there.
It's not so much that it's "where we want it to be". The documentation
uses `/usr/local/lib` because `/usr/local` is meant for software
installed by the sysadmin administrator, as is the case when building a
custom version. `/usr/lib` should be used by packages shipped with the
system.
C.f. https://refspecs.linuxfoundation.org/FHS_3.0/fhs/ch04s09.html
You're free to do as you see fit, but when you start mixing and matching
libraries and some are put in /usr/lib by your system packages and some
you move there manually I find it more difficult to follow what is going on.
This information also matches the statedump and the LD_DEBUG info from
later on.
Could you verify some of the following information:
1. In your build root for lttng-ust, enumerate all the liblttng*so
files. For each shared object, run `file $libname` and record the value
of the BuildID hash.R5jow
Sorry, I'm not following you here. The only buildID hash I can think
of is with 'eu-unstrip -n' but that's on core files, not individual
libs. And looking at the options I have for 'file' on my target, I
don't see anything that looks like what you are asking.
Perhaps I wasn't clear, the command to run is really just `file`. As a
fuller example:
```
$ file ./src/lib/lttng-ust-fork/.libs/liblttng-ust-fork.so.1.0.0
./src/lib/lttng-ust-fork/.libs/liblttng-ust-fork.so.1.0.0: ELF 64-bit
LSB shared object, x86-64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=b2b4a0fc449cf317e32c23e0bb57ea1ad702b702, with debug_info,
not stripped
Ok, feel stupid now. When I ran the command before, I used short name
and didn't do it on the long name and just got back:
# file /usr/lib/liblttng-ust.so
/usr/lib/liblttng-ust.so: symbolic link to liblttng-ust.so.1.0.0
... and immediately looked at man page to try to figure out what
switch showed BuildID etc., ha, ha.
When I do it on long name here is what I see:
# file /usr/lib/liblttng-ust-common.so.1.0.0
/usr/lib/liblttng-ust-common.so.1.0.0: ELF 64-bit LSB shared object,
ARM aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=a50c9a77163b6b91e1f84e57d167c7b77ae707a3, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-ctl.so.5.0.0
/usr/lib/liblttng-ust-ctl.so.5.0.0: ELF 64-bit LSB shared object, ARM
aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=547cccac08721ed1c9a7f3c7ebf1de84ddba7fba, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-cyg-profile-fast.so.1.0.0
/usr/lib/liblttng-ust-cyg-profile-fast.so.1.0.0: ELF 64-bit LSB shared
object, ARM aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=ad5f71ef5e83ab9a972488976c47db265d3360b9, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-cyg-profile.so.1.0.0
/usr/lib/liblttng-ust-cyg-profile.so.1.0.0: ELF 64-bit LSB shared
object, ARM aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=c2c246c1973bd3241aa4f7229fcbcb27ebe08e82, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-dl.so.1.0.0
/usr/lib/liblttng-ust-dl.so.1.0.0: ELF 64-bit LSB shared object, ARM
aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=6082ee88c9394319bc3adff16b0b3ea9f8d549ec, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-fd.so.1.0.0
/usr/lib/liblttng-ust-fd.so.1.0.0: ELF 64-bit LSB shared object, ARM
aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=a2813245af91abe98615771dfd7d5f19b033a410, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-fork.so.1.0.0
/usr/lib/liblttng-ust-fork.so.1.0.0: ELF 64-bit LSB shared object, ARM
aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=86afa53808502873830c02290f477d4ff8013afb, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-libc-wrapper.so.1.0.0
/usr/lib/liblttng-ust-libc-wrapper.so.1.0.0: ELF 64-bit LSB shared
object, ARM aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=52f391875a378b5f2c46747a58020b86cb7c9a83, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-pthread-wrapper.so.1.0.0
/usr/lib/liblttng-ust-pthread-wrapper.so.1.0.0: ELF 64-bit LSB shared
object, ARM aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=7127223a7e8ed67c6697b95ae1f8ac107df7e47e, with
debug_info, not stripped
# file /usr/lib/liblttng-ust-tracepoint.so.1.0.0
/usr/lib/liblttng-ust-tracepoint.so.1.0.0: ELF 64-bit LSB shared
object, ARM aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=5971b4d84ec1efe61c6d47c38e92de20569f0f49, with
debug_info, not stripped
# file /usr/lib/liblttng-ust.so.1.0.0
/usr/lib/liblttng-ust.so.1.0.0: ELF 64-bit LSB shared object, ARM
aarch64, version 1 (SYSV), dynamically linked,
BuildID[sha1]=ce7097ae9bbf42a02dccd386fdfbd37e3224858b, with
debug_info, not stripped
cutting some stuff out cause it's getting long again.
Sounds like `make check` for lttng-tools passed then?
At first no. But I think this is because I built the new lttng-tools
in my on target native environment and ran make check and forgot to do
make install first, so it was using the older version of lttng-tools.
So then I ran make install of lttng-tools and even did a make clean
and rebuild of lttng-ust and re-installed it and ran make check on
both (and things looked a lot better) ... that's where those warnings
etc., I asked you about came from.
My understanding at this point is the unit tests are passing for
LTTng-UST on your system, as are the unit and regression tests for
LTTng-tools. The example programs shipped with LTTng-UST work on your
system, as does the example from the documentation. The statedump
tracepoints loaded from LTTng-UST are also working fine, as evinced by
the program logs and the LTTng trace you shared.
Despite my confusion about how exactly you're using the `hello world`
tracepoint in your application (as you've now described several
variations), the direction this points to for me are details related to
how you're using LTTng-UST and/or how are your building and linking your
application. To be clear, I don't mean to say that there is or is not an
issue in LTTng-UST, but to point at where to examine next in detail
including analysis of the produced object files.
I compared the doc/examples/hello-static-lib to what I picked out of
the LTTng documentation on the web site "Quick start" section and the
traceprovider headder file is including stddef.h and mine isn't and
the doc/examples/hello-static-lib/hello.c code is doing a sighandler
and mine isn't doing any of that either. I think I've probably posted
it before but will do it again. Here is what I'm calling my "hello".
It's from the lttng documentation but I cut it down even further just
to make sure I didn't fat finger something. Like I said before, the
full hello example from the documentation works. But when I call
pretty much the same code from my app it seg faults.
I don't know if the differences I see between my "hello" and the
"hello-static-lib" matter.
hello-tp.h:
#undef LTTNG_UST_TRACEPOINT_PROVIDER
#define LTTNG_UST_TRACEPOINT_PROVIDER hello_world
#undef LTTNG_UST_TRACEPOINT_INCLUDE
#define LTTNG_UST_TRACEPOINT_INCLUDE "./hello-tp.h"
#if !defined(_HELLO_TP_H) || defined(LTTNG_UST_TRACEPOINT_HEADER_MULTI_READ)
#define _HELLO_TP_H
#include <lttng/tracepoint.h>
LTTNG_UST_TRACEPOINT_EVENT(
hello_world,
my_first_tracepoint,
LTTNG_UST_TP_ARGS(
int, my_integer_arg
),
LTTNG_UST_TP_FIELDS(
lttng_ust_field_integer(int, my_integer_field, my_integer_arg)
)
)
#endif /* _HELLO_TP_H */
#include <lttng/tracepoint-event.h>
hello-tp.c:
#define LTTNG_UST_TRACEPOINT_CREATE_PROBES
#include "hello-tp.h"
From my_app:
#define LTTNG_UST_TRACEPOINT_DEFINE
//#define LTTNG_UST_TRACEPOINT_PROBE_DYNAMIC_LINKAGE
#include "hello-tp.h"
.
.
.
lttng_ust_tracepoint(hello_world, my_first_tracepoint, 23, "hi there!");
In the above case the tpp is static but I've tried to make it a shared
object too (thus the commented out DYNAMIC_LINKAGE above) but get the
same result.
Again, I think the issue is probably g++ and weak/hidden symbol
related and or TLS but that's based on the totality of what I've
experienced over the past year or so and seeing the
experiences/problems of others in the lttng-dev archives.
Regards,
Brian
