Grafts cause discrepancies in debug symbols file names (debug symbols missing in GDB).

  • Open
  • quality assurance status badge
Details
4 participants
  • Ludovic Courtès
  • Maxim Cournoyer
  • pelzflorian (Florian Pelz)
  • Richard Sent
Owner
unassigned
Submitted by
Maxim Cournoyer
Severity
important
M
M
Maxim Cournoyer wrote on 7 Jun 2021 20:19
Debug symbols file name discrepancies
(name . bug-guix)(address . bug-guix@gnu.org)
87r1hdtu4t.fsf@gmail.com
Hello,

While attempting to debug a crash in jami-qt, I've noticed that GDB
would fail to load the symbol tables of the shared libraries it uses,
for example qtdeclarative.

It seems that grafts are to blame. Let's start by looking at the debug
files found for the qtdeclarative libQt5Qml.so.5 shared library:

Toggle snippet (27 lines)
$ guix build qtdeclarative | xargs -I{} find -L {} -name *libQt5Qml.so.5*
substitute: updating substitutes from 'http://127.0.0.1:8080'... 100.0%
substitute: updating substitutes from 'https://ci.guix.gnu.org'... 100.0%
The following files will be downloaded:
/gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2
substituting /gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug...
downloading from https://ci.guix.gnu.org/nar/lzip/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug ...
qtdeclarative-5.15.2-debug 94.9MiB 1.2MiB/s 01:21 [##################] 100.0%

The following graft will be made:
/gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv
applying 2 grafts for /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv ...
grafting '/gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug' -> '/gnu/store/l3h4ka7v3j1yhik0f1phwch08a09p0bx-qtdeclarative-5.15.2-debug'...
grafting '/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2' -> '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2'...
updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/bin/qml'
updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/bin/q
[...]
updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/qt5/qml/QtQuick/Window.2/libwindowplugin.so'
updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/qt5/qml/QtTest/libqmltestplugin.so'
successfully built /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv
/gnu/store/l3h4ka7v3j1yhik0f1phwch08a09p0bx-qtdeclarative-5.15.2-debug/lib/debug/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2.debug
/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2
/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5
/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15

So far so good. The file hierarchy under the debug output matches the
actual shared library file name. Next, let's verify which qtdeclarative
shared libraries jami-qt is dynamically linked against:

Toggle snippet (7 lines)
$ guix build jami-qt | tail -1 | xargs -I{} ldd {}/bin/.jami-qt-real | grep qtdeclarative
libQt5QuickWidgets.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5QuickWidgets.so.5 (0x00007fb9e38a8000)
libQt5Quick.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5Quick.so.5 (0x00007fb9dba47000)
libQt5QmlModels.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5QmlModels.so.5 (0x00007fb9db9c3000)
libQt5Qml.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5Qml.so.5 (0x00007fb9dae4e000)

Oops! The actual store file name of the libQt5Qml.so.5 known to jami-qt
is *not* the same as the one obtained earlier, which explains why GDB
doesn't find its symbols. Without grafts, the first command gives:

Toggle snippet (7 lines)
$ guix build --no-grafts qtdeclarative | xargs -I{} find -L {} -name *libQt5Qml.so.5*
/gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug/lib/debug/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2.debug
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2

Which still doesn't match the libraries jami-qt is linked with.

I'm out of ideas for now. Would someone have a clue?

Thank you!

Maxim
M
M
Maxim Cournoyer wrote on 7 Jun 2021 21:26
(address . 48907@debbugs.gnu.org)
87mts1tr1u.fsf@gmail.com
Hello again,

Building everything without grafts does resolve the file name
discrepancy issue, e.g.:

guix build --no-grafts qtdeclarative | xargs -I{} find -L {} -name *libQt5Qml.so.5*
/gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug/lib/debug/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2.debug
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15
/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2

guix build --no-grafts jami-qt | tail -1 | xargs -I{} ldd {}/bin/.jami-qt-real | grep qtdeclarative
libQt5QuickWidgets.so.5 => /gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5QuickWidgets.so.5 (0x00007f42c79b2000)
libQt5Quick.so.5 => /gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Quick.so.5 (0x00007f42bfb52000)
libQt5QmlModels.so.5 => /gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5QmlModels.so.5 (0x00007f42bface000)
libQt5Qml.so.5 => /gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2/lib/libQt5Qml.so.5 (0x00007f42bef59000)

And the debug symbols are now discovered and usable by GDB. So this
issue is indeed caused by grafts.

Maxim
M
M
Maxim Cournoyer wrote on 7 Jun 2021 21:28
control message for bug #48907
(address . control@debbugs.gnu.org)
87lf7ltqyd.fsf@gmail.com
retitle 48907 Grafts cause discrepancies in debug symbols file names (debug symbols missing in GDB).
quit
L
L
Ludovic Courtès wrote on 18 Jun 2021 11:29
Re: bug#48907: Debug symbols file name discrepancies
(name . Maxim Cournoyer)(address . maxim.cournoyer@gmail.com)(address . 48907@debbugs.gnu.org)
87eeczikq5.fsf@gnu.org
Hi,

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

Toggle quote (37 lines)
> $ guix build qtdeclarative | xargs -I{} find -L {} -name *libQt5Qml.so.5*
> substitute: updating substitutes from 'http://127.0.0.1:8080'... 100.0%
> substitute: updating substitutes from 'https://ci.guix.gnu.org'... 100.0%
> The following files will be downloaded:
> /gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug
> /gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2
> substituting /gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug...
> downloading from https://ci.guix.gnu.org/nar/lzip/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug ...
> qtdeclarative-5.15.2-debug 94.9MiB 1.2MiB/s 01:21 [##################] 100.0%
>
> The following graft will be made:
> /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv
> applying 2 grafts for /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv ...
> grafting '/gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug' -> '/gnu/store/l3h4ka7v3j1yhik0f1phwch08a09p0bx-qtdeclarative-5.15.2-debug'...
> grafting '/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2' -> '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2'...
> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/bin/qml'
> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/bin/q
> [...]
> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/qt5/qml/QtQuick/Window.2/libwindowplugin.so'
> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/qt5/qml/QtTest/libqmltestplugin.so'
> successfully built /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv
> /gnu/store/l3h4ka7v3j1yhik0f1phwch08a09p0bx-qtdeclarative-5.15.2-debug/lib/debug/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2.debug
> /gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2
> /gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5
> /gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15
>
>
> So far so good. The file hierarchy under the debug output matches the
> actual shared library file name. Next, let's verify which qtdeclarative
> shared libraries jami-qt is dynamically linked against:
>
> $ guix build jami-qt | tail -1 | xargs -I{} ldd {}/bin/.jami-qt-real | grep qtdeclarative
> libQt5QuickWidgets.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5QuickWidgets.so.5 (0x00007fb9e38a8000)
> libQt5Quick.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5Quick.so.5 (0x00007fb9dba47000)
> libQt5QmlModels.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5QmlModels.so.5 (0x00007fb9db9c3000)
> libQt5Qml.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5Qml.so.5 (0x00007fb9dae4e000)

This is due to the fact that, when you run ‘guix build jami-qt’, the
grafting derivation dismisses the “debug” output of qtdeclarative, since
jami-qt does not depend on it. That way it doesn’t have to
build/download and graft qtdeclarative:debug.

Conversely, when you run ‘guix build qtdeclarative’, it grafts both
outputs of qtdeclarative. Thus, this grafting derivation is different
from the one jami-qt.drv depends on.

This behavior was implemented in commit
482fda2729c3e76999892cb8f9a0391a7bd37119. Not sure what a good solution
would be.

Thoughts?

Ludo’.
M
M
Maxim Cournoyer wrote on 24 Sep 2021 04:32
Re: bug#48907: Grafts cause discrepancies in debug symbols file names (debug symbols missing in GDB).
(name . Ludovic Courtès)(address . ludo@gnu.org)(address . 48907@debbugs.gnu.org)
87ilyqk8jg.fsf_-_@gmail.com
Hi!

Ludovic Courtès <ludo@gnu.org> writes:

Toggle quote (56 lines)
> Hi,
>
> Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:
>
>> $ guix build qtdeclarative | xargs -I{} find -L {} -name *libQt5Qml.so.5*
>> substitute: updating substitutes from 'http://127.0.0.1:8080'... 100.0%
>> substitute: updating substitutes from 'https://ci.guix.gnu.org'... 100.0%
>> The following files will be downloaded:
>> /gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug
>> /gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2
>> substituting /gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug...
>> downloading from https://ci.guix.gnu.org/nar/lzip/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug ...
>> qtdeclarative-5.15.2-debug 94.9MiB 1.2MiB/s 01:21 [##################] 100.0%
>>
>> The following graft will be made:
>> /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv
>> applying 2 grafts for /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv ...
>> grafting '/gnu/store/g1gxfbkyxilnx7s6mjdlj697y5n5wazn-qtdeclarative-5.15.2-debug' -> '/gnu/store/l3h4ka7v3j1yhik0f1phwch08a09p0bx-qtdeclarative-5.15.2-debug'...
>> grafting '/gnu/store/nvzvrr137qfqsn2875yrs9ilfd12wi96-qtdeclarative-5.15.2' -> '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2'...
>> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/bin/qml'
>> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/bin/q
>> [...]
>> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/qt5/qml/QtQuick/Window.2/libwindowplugin.so'
>> updating '.gnu_debuglink' CRC in '/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/qt5/qml/QtTest/libqmltestplugin.so'
>> successfully built /gnu/store/djhcai9rixm2j3jlamwdhsgwgidg7w74-qtdeclarative-5.15.2.drv
>> /gnu/store/l3h4ka7v3j1yhik0f1phwch08a09p0bx-qtdeclarative-5.15.2-debug/lib/debug/gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2.debug
>> /gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15.2
>> /gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5
>> /gnu/store/pryhgzb6cwnzsskqwldwc6dxr6nwnywf-qtdeclarative-5.15.2/lib/libQt5Qml.so.5.15
>>
>>
>> So far so good. The file hierarchy under the debug output matches the
>> actual shared library file name. Next, let's verify which qtdeclarative
>> shared libraries jami-qt is dynamically linked against:
>>
>> $ guix build jami-qt | tail -1 | xargs -I{} ldd {}/bin/.jami-qt-real | grep qtdeclarative
>> libQt5QuickWidgets.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5QuickWidgets.so.5 (0x00007fb9e38a8000)
>> libQt5Quick.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5Quick.so.5 (0x00007fb9dba47000)
>> libQt5QmlModels.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5QmlModels.so.5 (0x00007fb9db9c3000)
>> libQt5Qml.so.5 => /gnu/store/mjl02yma4r5xjark6d8pp5h2j0a2vccs-qtdeclarative-5.15.2/lib/libQt5Qml.so.5 (0x00007fb9dae4e000)
>
> This is due to the fact that, when you run ‘guix build jami-qt’, the
> grafting derivation dismisses the “debug” output of qtdeclarative, since
> jami-qt does not depend on it. That way it doesn’t have to
> build/download and graft qtdeclarative:debug.
>
> Conversely, when you run ‘guix build qtdeclarative’, it grafts both
> outputs of qtdeclarative. Thus, this grafting derivation is different
> from the one jami-qt.drv depends on.
>
> This behavior was implemented in commit
> 482fda2729c3e76999892cb8f9a0391a7bd37119. Not sure what a good solution
> would be.
>
> Thoughts?

Yikes! This means that debugging with grafts (with the aid of debugging
symbols) is no longer possible, right?

I remember reading about a 2nd option to locate the separate debug
symbol files with GDB in info '(gdb) Separate Debug Files':


* The executable contains a "build ID", a unique bit string that is
also present in the corresponding debug info file. (This is
supported only on some operating systems, when using the ELF or PE
file formats for binary files and the GNU Binutils.) For more
details about this feature, see the description of the '--build-id'
command-line option in *note Command Line Options: (ld)Options.
The debug info file's name is not specified explicitly by the build
ID, but can be computed from the build ID, see below.

[...]
* For the "build ID" method, GDB looks in the '.build-id'
subdirectory of each one of the global debug directories for a file
named 'NN/NNNNNNNN.debug', where NN are the first 2 hex characters
of the build ID bit string, and NNNNNNNN are the rest of the bit
string. (Real build ID strings are 32 or more hex characters, not
10.)

What may help us here, compared to debug links, is that it seems to be
file name agnostic; the debug files would be matched by an internal ID
that they got at build time rather than from their file names (which
doesn't work with the current different derivations in the presence of
grafts).

Perhaps it'd also lift the need to recompute the CRC checksum of the
debug links produced when grafting!

HTH!

Maxim
L
L
Ludovic Courtès wrote on 24 Sep 2021 16:07
control message for bug #48907
(address . control@debbugs.gnu.org)
87sfxum5i2.fsf@gnu.org
severity 48907 important
quit
L
L
Ludovic Courtès wrote on 24 Sep 2021 16:14
Re: bug#48907: Grafts cause discrepancies in debug symbols file names (debug symbols missing in GDB).
(name . Maxim Cournoyer)(address . maxim.cournoyer@gmail.com)(address . 48907@debbugs.gnu.org)
87lf3mm560.fsf@gnu.org
Hi,

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

Toggle quote (2 lines)
> Ludovic Courtès <ludo@gnu.org> writes:

[...]

Toggle quote (5 lines)
>> This is due to the fact that, when you run ‘guix build jami-qt’, the
>> grafting derivation dismisses the “debug” output of qtdeclarative, since
>> jami-qt does not depend on it. That way it doesn’t have to
>> build/download and graft qtdeclarative:debug.

[...]

Toggle quote (3 lines)
> Yikes! This means that debugging with grafts (with the aid of debugging
> symbols) is no longer possible, right?

It depends on whether the separate “debug” output gets grafted or not,
but yeah, if a dependency tree has this shape (app -> lib + lib:debug),
running ‘guix install app’ alone will prevent you from getting debugging
symbols from ‘lib:debug’ I believe. That sucks.

I wonder if we should revert 482fda2729c3e76999892cb8f9a0391a7bd37119.
It’s often not very helpful anyway (we often find ourselves downloading
unnecessary package outputs because of grafting).

Toggle quote (6 lines)
> I remember reading about a 2nd option to locate the separate debug
> symbol files with GDB in info '(gdb) Separate Debug Files':
>
>
> * The executable contains a "build ID", a unique bit string that is

We’d have to check if this is applicable. Looking at the ld manual
(info "(ld) Options"), it seems that the UUID “style” is ruled out
because it’s non-deterministic, and the md5 and sha1 styles would
require us to rewrite build IDs IIUC, similar to how we rewrite CRCs.

Thanks,
Ludo’.
M
M
Maxim Cournoyer wrote on 28 Sep 2021 04:25
(name . Ludovic Courtès)(address . ludo@gnu.org)(address . 48907@debbugs.gnu.org)
87v92lh1wc.fsf@gmail.com
Heya,

Ludovic Courtès <ludo@gnu.org> writes:

[...]

Toggle quote (12 lines)
>> Yikes! This means that debugging with grafts (with the aid of debugging
>> symbols) is no longer possible, right?
>
> It depends on whether the separate “debug” output gets grafted or not,
> but yeah, if a dependency tree has this shape (app -> lib + lib:debug),
> running ‘guix install app’ alone will prevent you from getting debugging
> symbols from ‘lib:debug’ I believe. That sucks.
>
> I wonder if we should revert 482fda2729c3e76999892cb8f9a0391a7bd37119.
> It’s often not very helpful anyway (we often find ourselves downloading
> unnecessary package outputs because of grafting).

Hmm. Perhaps. But it'd also suck to have to download 1 GiB of unneeded
debugging symbols to just apply a graft to Qt, for example.

Toggle quote (11 lines)
>> I remember reading about a 2nd option to locate the separate debug
>> symbol files with GDB in info '(gdb) Separate Debug Files':
>>
>>
>> * The executable contains a "build ID", a unique bit string that is
>
> We’d have to check if this is applicable. Looking at the ld manual
> (info "(ld) Options"), it seems that the UUID “style” is ruled out
> because it’s non-deterministic, and the md5 and sha1 styles would
> require us to rewrite build IDs IIUC, similar to how we rewrite CRCs.

Seems like it could work? simark from #gdb says it should be
deterministic for reproducible builds. We'd need to fixup the grafted
debug output, but they could being done in a separate derivation would
no longer matter (as the debug symbols would be matched on a unique ID
that is not linked to that derivation, not on their file name, which
is).

Did I get the above right?

Thanks,

Maxim
L
L
Ludovic Courtès wrote on 28 Sep 2021 11:45
(name . Maxim Cournoyer)(address . maxim.cournoyer@gmail.com)(address . 48907@debbugs.gnu.org)
877df1f2zk.fsf@gnu.org
Hi,

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

Toggle quote (19 lines)
> Ludovic Courtès <ludo@gnu.org> writes:
>
> [...]
>
>>> Yikes! This means that debugging with grafts (with the aid of debugging
>>> symbols) is no longer possible, right?
>>
>> It depends on whether the separate “debug” output gets grafted or not,
>> but yeah, if a dependency tree has this shape (app -> lib + lib:debug),
>> running ‘guix install app’ alone will prevent you from getting debugging
>> symbols from ‘lib:debug’ I believe. That sucks.
>>
>> I wonder if we should revert 482fda2729c3e76999892cb8f9a0391a7bd37119.
>> It’s often not very helpful anyway (we often find ourselves downloading
>> unnecessary package outputs because of grafting).
>
> Hmm. Perhaps. But it'd also suck to have to download 1 GiB of unneeded
> debugging symbols to just apply a graft to Qt, for example.

Yeah. That’s already the case in some cases though, that’s what I
meant.

Toggle quote (20 lines)
>>> I remember reading about a 2nd option to locate the separate debug
>>> symbol files with GDB in info '(gdb) Separate Debug Files':
>>>
>>>
>>> * The executable contains a "build ID", a unique bit string that is
>>
>> We’d have to check if this is applicable. Looking at the ld manual
>> (info "(ld) Options"), it seems that the UUID “style” is ruled out
>> because it’s non-deterministic, and the md5 and sha1 styles would
>> require us to rewrite build IDs IIUC, similar to how we rewrite CRCs.
>
> Seems like it could work? simark from #gdb says it should be
> deterministic for reproducible builds. We'd need to fixup the grafted
> debug output, but they could being done in a separate derivation would
> no longer matter (as the debug symbols would be matched on a unique ID
> that is not linked to that derivation, not on their file name, which
> is).
>
> Did I get the above right?

To summarize, ‘.gnu_debuglink’ in executables/libraries contains the
CRC of the debug file.

Conversely, IIUC what the “normative parts of the output contents” (info
"(ld) Options") really are, build IDs are computed on the code, not on
debug info.

But the problems remains the same I think: if you have
/gnu/store/abc…/libfoo.so and /gnu/store/xyz…/libfoo.so, chances are
that they are different due to embedded store file names, and thus get a
different build ID.

Am I right?

(BTW, I just noticed build IDs were also discussed at

Ludo’.
L
L
Ludovic Courtès wrote on 28 Sep 2021 12:28
(name . Maxim Cournoyer)(address . maxim.cournoyer@gmail.com)(address . 48907@debbugs.gnu.org)
87lf3hdmeg.fsf@gnu.org
Ludovic Courtès <ludo@gnu.org> skribis:

Toggle quote (9 lines)
> Conversely, IIUC what the “normative parts of the output contents” (info
> "(ld) Options") really are, build IDs are computed on the code, not on
> debug info.
>
> But the problems remains the same I think: if you have
> /gnu/store/abc…/libfoo.so and /gnu/store/xyz…/libfoo.so, chances are
> that they are different due to embedded store file names, and thus get a
> different build ID.

We discussed this with Mark Wielaard on #guix¹, and one important
takeaway is:

Toggle snippet (17 lines)
<civodul> so gdb just checks that the separate debug file has the same
build-id as the code, right? [12:16]
<civodul> it doesn't matter whether it really is the sha1 of all those
sections, does it?
<mjw> that is kind of the whole point of the build-id, that it captures the
whole build environment, not just the generated code, but also how it
was generated (which is what the .debug sections kind of represent)
<civodul> ok [12:17]
<mjw> civodul, no, it doesn't need to be a hash over the actual bits
produced. It can be a completely different hash, it can be a different
number of bits (but not too short, they need to be globally unique).
<civodul> ok, so we could have our own way of caculating build IDs [12:18]
<mjw> civodul, all that really matters is that it uniquely identifies this
binary blob. If any input, source, compiler, flags, etc. changes, it
should be unique.

So I suspect that we would not need to rewrite build IDs upon grafting,
and we could use the ungrafted debug info with grafted code and vice
versa.

We should try it out to test the hypothesis, but if that works, that’d
be great.

Ludo’.

L
L
Ludovic Courtès wrote on 4 Oct 2021 15:14
(name . Maxim Cournoyer)(address . maxim.cournoyer@gmail.com)(address . 48907@debbugs.gnu.org)
8735phq6dy.fsf@gnu.org
Hi,

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

Toggle quote (19 lines)
> Ludovic Courtès <ludo@gnu.org> writes:
>
> [...]
>
>>> Yikes! This means that debugging with grafts (with the aid of debugging
>>> symbols) is no longer possible, right?
>>
>> It depends on whether the separate “debug” output gets grafted or not,
>> but yeah, if a dependency tree has this shape (app -> lib + lib:debug),
>> running ‘guix install app’ alone will prevent you from getting debugging
>> symbols from ‘lib:debug’ I believe. That sucks.
>>
>> I wonder if we should revert 482fda2729c3e76999892cb8f9a0391a7bd37119.
>> It’s often not very helpful anyway (we often find ourselves downloading
>> unnecessary package outputs because of grafting).
>
> Hmm. Perhaps. But it'd also suck to have to download 1 GiB of unneeded
> debugging symbols to just apply a graft to Qt, for example.

Yeah. That’s already the case in some cases though, that’s what I
meant.

Toggle quote (20 lines)
>>> I remember reading about a 2nd option to locate the separate debug
>>> symbol files with GDB in info '(gdb) Separate Debug Files':
>>>
>>>
>>> * The executable contains a "build ID", a unique bit string that is
>>
>> We’d have to check if this is applicable. Looking at the ld manual
>> (info "(ld) Options"), it seems that the UUID “style” is ruled out
>> because it’s non-deterministic, and the md5 and sha1 styles would
>> require us to rewrite build IDs IIUC, similar to how we rewrite CRCs.
>
> Seems like it could work? simark from #gdb says it should be
> deterministic for reproducible builds. We'd need to fixup the grafted
> debug output, but they could being done in a separate derivation would
> no longer matter (as the debug symbols would be matched on a unique ID
> that is not linked to that derivation, not on their file name, which
> is).
>
> Did I get the above right?

To summarize, ‘.gnu_debuglink’ in executables/libraries contains the
CRC of the debug file.

Conversely, IIUC what the “normative parts of the output contents” (info
"(ld) Options") really are, build IDs are computed on the code, not on
debug info.

But the problems remains the same I think: if you have
/gnu/store/abc…/libfoo.so and /gnu/store/xyz…/libfoo.so, chances are
that they are different due to embedded store file names, and thus get a
different build ID.

Am I right?

(BTW, I just noticed build IDs were also discussed at

Ludo’.
P
P
pelzflorian (Florian Pelz) wrote on 27 Apr 10:02 +0200
(address . 48907@debbugs.gnu.org)
871q6r9qsg.fsf@florianrock64.mail-host-address-is-not-set
Hello all.

Ludovic Courtès <ludo@gnu.org> writes:
Toggle quote (10 lines)
>>> I wonder if we should revert 482fda2729c3e76999892cb8f9a0391a7bd37119.
>>> It’s often not very helpful anyway (we often find ourselves downloading
>>> unnecessary package outputs because of grafting).
>>
>> Hmm. Perhaps. But it'd also suck to have to download 1 GiB of unneeded
>> debugging symbols to just apply a graft to Qt, for example.
>
> Yeah. That’s already the case in some cases though, that’s what I
> meant.

I believe I got bitten by this 482fda2729c3e76999892cb8f9a0391a7bd37119
too, not only for debug symbols, which annoys me, and not like the Go
users, but for me the issue surfaced when trying to use guile-g-golf,
whose hello-world.scm errors out with “cannot register existing type
'GdkPixbuf'” and GDB debugging shows (I think) it is because a variant
of gdk-pixbuf without a debug output is used by Gtk while guile-g-golf
loads gdk-pixbuf with debug output before. `guix shell --no-grafts` is
a workaround.

Perhaps guile-gi have the same issue. Quoting
Toggle quote (4 lines)
> It is pretty clear to me, that the main culprit here is a different
> version of GLib being linked to Guile-GI than the one that should be
> loaded through Guile-GI.

In essence it seems to me 482fda2729c3e76999892cb8f9a0391a7bd37119 is
implemented at the wrong level. It changed the grafting derivation and
thus the hash, when it could probably be done in guix-daemon directly.

Is it better to wait with fixing this until the guix-daemon is rewritten
in Guile? And use `--no-grafts' in the meantime?

Regards,
Florian
M
M
Maxim Cournoyer wrote on 30 May 03:11 +0200
(name . pelzflorian (Florian Pelz))(address . pelzflorian@pelzflorian.de)
87plt4f6j3.fsf@gmail.com
Hi Florian,

"pelzflorian (Florian Pelz)" <pelzflorian@pelzflorian.de> writes:

Toggle quote (22 lines)
> Hello all.
>
> Ludovic Courtès <ludo@gnu.org> writes:
>>>> I wonder if we should revert 482fda2729c3e76999892cb8f9a0391a7bd37119.
>>>> It’s often not very helpful anyway (we often find ourselves downloading
>>>> unnecessary package outputs because of grafting).
>>>
>>> Hmm. Perhaps. But it'd also suck to have to download 1 GiB of unneeded
>>> debugging symbols to just apply a graft to Qt, for example.
>>
>> Yeah. That’s already the case in some cases though, that’s what I
>> meant.
>
> I believe I got bitten by this 482fda2729c3e76999892cb8f9a0391a7bd37119
> too, not only for debug symbols, which annoys me, and not like the Go
> users, but for me the issue surfaced when trying to use guile-g-golf,
> whose hello-world.scm errors out with “cannot register existing type
> 'GdkPixbuf'” and GDB debugging shows (I think) it is because a variant
> of gdk-pixbuf without a debug output is used by Gtk while guile-g-golf
> loads gdk-pixbuf with debug output before. `guix shell --no-grafts` is
> a workaround.

I'm guessing bug #64836 ("pygobject GTK modules lookup fails following
CUPS graft") may be another manifestation of this problem.

--
Thanks,
Maxim
R
R
Richard Sent wrote on 30 May 04:38 +0200
Other possibly related issues
(address . 48907@debbugs.gnu.org)
87sey081n3.fsf@freakingpenguin.com
Hi all,

I believe https://issues.guix.gnu.org/62890is another example of this
problem. Possibly https://issues.guix.gnu.org/47115#23as well.

One idea I had was to somehow graft each output as a unique derivation
via repeated calls to graft-derivation. From what I've seen so far
though that looks much easier said than done. It hasn't progressed
beyond the "huh, I wonder" phase.

Funny, I just started investigating the grafting code myself to see if I
could understand what's going on and I stumble upon this 3 year old
issue just an hour after Maxim comments.

--
Take it easy,
Richard Sent
Making my computer weirder one commit at a time.
?
Your comment

Commenting via the web interface is currently disabled.

To comment on this conversation send an email to 48907@debbugs.gnu.org

To respond to this issue using the mumi CLI, first switch to it
mumi current 48907
Then, you may apply the latest patchset in this issue (with sign off)
mumi am -- -s
Or, compose a reply to this issue
mumi compose
Or, send patches to this issue
mumi send-email *.patch