* [Buildroot] [PATCHv2 1/3] package/gdb: fix gdb 9.2 segfault with Python 3.9 support
2021-06-08 15:18 [Buildroot] [PATCHv2 0/3] Fix gdb pretty printers for libstdcxx Thomas De Schampheleire
@ 2021-06-08 15:18 ` Thomas De Schampheleire
2021-07-25 21:21 ` Thomas Petazzoni
2021-08-05 11:10 ` Peter Korsgaard
2021-06-08 15:18 ` [Buildroot] [PATCHv2 2/3] toolchain/toolchain-external: fixup gdb pretty-printer loader for libstdcxx Thomas De Schampheleire
2021-06-08 15:18 ` [Buildroot] [PATCHv2 3/3] toolchain: mark sysroot as 'safe' path for gdb auto-load Thomas De Schampheleire
2 siblings, 2 replies; 8+ messages in thread
From: Thomas De Schampheleire @ 2021-06-08 15:18 UTC (permalink / raw)
To: buildroot
From: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
When enabling Python 3 support in gdb 9.2, gdb segfaults at startup.
The issue is tracked back to following upstream gdb commit, present since
gdb 10.1.
commit c47bae859a5af0d95224d90000df0e529f7c5aa0
Author: Kevin Buettner <kevinb@redhat.com>
Date: Wed May 27 20:05:40 2020 -0700
Fix Python3.9 related runtime problems
[...]
Backport this commit to 9.2 in Buildroot.
Signed-off-by: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
---
...x-Python3.9-related-runtime-problems.patch | 227 ++++++++++++++++++
1 file changed, 227 insertions(+)
create mode 100644 package/gdb/9.2/0007-Fix-Python3.9-related-runtime-problems.patch
diff --git a/package/gdb/9.2/0007-Fix-Python3.9-related-runtime-problems.patch b/package/gdb/9.2/0007-Fix-Python3.9-related-runtime-problems.patch
new file mode 100644
index 0000000000..89b0cdc4a0
--- /dev/null
+++ b/package/gdb/9.2/0007-Fix-Python3.9-related-runtime-problems.patch
@@ -0,0 +1,227 @@
+From c47bae859a5af0d95224d90000df0e529f7c5aa0 Mon Sep 17 00:00:00 2001
+From: Kevin Buettner <kevinb@redhat.com>
+Date: Wed, 27 May 2020 20:05:40 -0700
+Subject: [PATCH] Fix Python3.9 related runtime problems
+
+Python3.9b1 is now available on Rawhide. GDB w/ Python 3.9 support
+can be built using the configure switch -with-python=/usr/bin/python3.9.
+
+Attempting to run gdb/Python3.9 segfaults on startup:
+
+ #0 0x00007ffff7b0582c in PyEval_ReleaseLock () from /lib64/libpython3.9.so.1.0
+ #1 0x000000000069ccbf in do_start_initialization ()
+ at worktree-test1/gdb/python/python.c:1789
+ #2 _initialize_python ()
+ at worktree-test1/gdb/python/python.c:1877
+ #3 0x00000000007afb0a in initialize_all_files () at init.c:237
+ ...
+
+Consulting the the documentation...
+
+https://docs.python.org/3/c-api/init.html
+
+...we find that PyEval_ReleaseLock() has been deprecated since version
+3.2. It recommends using PyEval_SaveThread or PyEval_ReleaseThread()
+instead. In do_start_initialization, in gdb/python/python.c, we
+can replace the calls to PyThreadState_Swap() and PyEval_ReleaseLock()
+with a single call to PyEval_SaveThread. (Thanks to Keith Seitz
+for working this out.)
+
+With that in place, GDB gets a little bit further. It still dies
+on startup, but the backtrace is different:
+
+ #0 0x00007ffff7b04306 in PyOS_InterruptOccurred ()
+ from /lib64/libpython3.9.so.1.0
+ #1 0x0000000000576e86 in check_quit_flag ()
+ at worktree-test1/gdb/extension.c:776
+ #2 0x0000000000576f8a in set_active_ext_lang (now_active=now_active at entry=0x983c00 <extension_language_python>)
+ at worktree-test1/gdb/extension.c:705
+ #3 0x000000000069d399 in gdbpy_enter::gdbpy_enter (this=0x7fffffffd2d0,
+ gdbarch=0x0, language=0x0)
+ at worktree-test1/gdb/python/python.c:211
+ #4 0x0000000000686e00 in python_new_inferior (inf=0xddeb10)
+ at worktree-test1/gdb/python/py-inferior.c:251
+ #5 0x00000000005d9fb9 in std::function<void (inferior*)>::operator()(inferior*) const (__args#0=<optimized out>, this=0xccad20)
+ at /usr/include/c++/10/bits/std_function.h:617
+ #6 gdb::observers::observable<inferior*>::notify (args#0=0xddeb10,
+ this=<optimized out>)
+ at worktree-test1/gdb/../gdbsupport/observable.h:106
+ #7 add_inferior_silent (pid=0)
+ at worktree-test1/gdb/inferior.c:113
+ #8 0x00000000005dbcb8 in initialize_inferiors ()
+ at worktree-test1/gdb/inferior.c:947
+ ...
+
+We checked with some Python Developers and were told that we should
+acquire the GIL prior to calling any Python C API function. We
+definitely don't have the GIL for calls of PyOS_InterruptOccurred().
+
+I moved class_gdbpy_gil earlier in the file and use it in
+gdbpy_check_quit_flag() to acquire (and automatically release) the
+GIL.
+
+With those changes in place, I was able to run to a GDB prompt. But,
+when trying to quit, it segfaulted again due to due to some other
+problems with gdbpy_check_quit_flag():
+
+ Thread 1 "gdb" received signal SIGSEGV, Segmentation fault.
+ 0x00007ffff7bbab0c in new_threadstate () from /lib64/libpython3.9.so.1.0
+ (top-gdb) bt 8
+ #0 0x00007ffff7bbab0c in new_threadstate () from /lib64/libpython3.9.so.1.0
+ #1 0x00007ffff7afa5ea in PyGILState_Ensure.cold ()
+ from /lib64/libpython3.9.so.1.0
+ #2 0x000000000069b58c in gdbpy_gil::gdbpy_gil (this=<synthetic pointer>)
+ at worktree-test1/gdb/python/python.c:278
+ #3 gdbpy_check_quit_flag (extlang=<optimized out>)
+ at worktree-test1/gdb/python/python.c:278
+ #4 0x0000000000576e96 in check_quit_flag ()
+ at worktree-test1/gdb/extension.c:776
+ #5 0x000000000057700c in restore_active_ext_lang (previous=0xe9c050)
+ at worktree-test1/gdb/extension.c:729
+ #6 0x000000000088913a in do_my_cleanups (
+ pmy_chain=0xc31870 <final_cleanup_chain>,
+ old_chain=0xae5720 <sentinel_cleanup>)
+ at worktree-test1/gdbsupport/cleanups.cc:131
+ #7 do_final_cleanups ()
+ at worktree-test1/gdbsupport/cleanups.cc:143
+
+In this case, we're trying to call a Python C API function after
+Py_Finalize() has been called from finalize_python(). I made
+finalize_python set gdb_python_initialized to false and then cause
+check_quit_flag() to return early when it's false.
+
+With these changes in place, GDB seems to be working again with
+Python3.9b1. I think it likely that there are other problems lurking.
+I wouldn't be surprised to find that there are other calls into Python
+where we don't first make sure that we have the GIL. Further changes
+may well be needed.
+
+I see no regressions testing on Rawhide using a GDB built with the
+default Python version (3.8.3) versus one built using Python 3.9b1.
+
+I've also tested on Fedora 28, 29, 30, 31, and 32 (all x86_64) using
+the default (though updated) system installed versions of Python on
+those OSes. This means that I've tested against Python versions
+2.7.15, 2.7.17, 2.7.18, 3.7.7, 3.8.2, and 3.8.3. In each case GDB
+still builds without problem and shows no regressions after applying
+this patch.
+
+gdb/ChangeLog:
+
+2020-MM-DD Kevin Buettner <kevinb@redhat.com>
+ Keith Seitz <keiths@redhat.com>
+
+ * python/python.c (do_start_initialization): For Python 3.9 and
+ later, call PyEval_SaveThread instead of PyEval_ReleaseLock.
+ (class gdbpy_gil): Move to earlier in file.
+ (finalize_python): Set gdb_python_initialized.
+ (gdbpy_check_quit_flag): Acquire GIL via gdbpy_gil. Return early
+ when not initialized.
+
+
+[import into Buildroot, removing ChangeLog change to avoid conflict]
+Signed-off-by: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
+
+---
+ gdb/ChangeLog | 10 ++++++++
+ gdb/python/python.c | 56 ++++++++++++++++++++++++---------------------
+ 2 files changed, 40 insertions(+), 26 deletions(-)
+
+diff --git a/gdb/python/python.c b/gdb/python/python.c
+index 67f362b852d..4bdd2201abc 100644
+--- a/gdb/python/python.c
++++ b/gdb/python/python.c
+@@ -238,6 +238,30 @@ gdbpy_enter::~gdbpy_enter ()
+ PyGILState_Release (m_state);
+ }
+
++/* A helper class to save and restore the GIL, but without touching
++ the other globals that are handled by gdbpy_enter. */
++
++class gdbpy_gil
++{
++public:
++
++ gdbpy_gil ()
++ : m_state (PyGILState_Ensure ())
++ {
++ }
++
++ ~gdbpy_gil ()
++ {
++ PyGILState_Release (m_state);
++ }
++
++ DISABLE_COPY_AND_ASSIGN (gdbpy_gil);
++
++private:
++
++ PyGILState_STATE m_state;
++};
++
+ /* Set the quit flag. */
+
+ static void
+@@ -251,6 +275,10 @@ gdbpy_set_quit_flag (const struct extension_language_defn *extlang)
+ static int
+ gdbpy_check_quit_flag (const struct extension_language_defn *extlang)
+ {
++ if (!gdb_python_initialized)
++ return 0;
++
++ gdbpy_gil gil;
+ return PyOS_InterruptOccurred ();
+ }
+
+@@ -943,30 +971,6 @@ gdbpy_source_script (const struct extension_language_defn *extlang,
+
+ /* Posting and handling events. */
+
+-/* A helper class to save and restore the GIL, but without touching
+- the other globals that are handled by gdbpy_enter. */
+-
+-class gdbpy_gil
+-{
+-public:
+-
+- gdbpy_gil ()
+- : m_state (PyGILState_Ensure ())
+- {
+- }
+-
+- ~gdbpy_gil ()
+- {
+- PyGILState_Release (m_state);
+- }
+-
+- DISABLE_COPY_AND_ASSIGN (gdbpy_gil);
+-
+-private:
+-
+- PyGILState_STATE m_state;
+-};
+-
+ /* A single event. */
+ struct gdbpy_event
+ {
+@@ -1616,6 +1620,7 @@ finalize_python (void *ignore)
+
+ Py_Finalize ();
+
++ gdb_python_initialized = false;
+ restore_active_ext_lang (previous_active);
+ }
+
+@@ -1785,8 +1790,7 @@ do_start_initialization ()
+ return false;
+
+ /* Release the GIL while gdb runs. */
+- PyThreadState_Swap (NULL);
+- PyEval_ReleaseLock ();
++ PyEval_SaveThread ();
+
+ make_final_cleanup (finalize_python, NULL);
+
+--
+2.26.3
+
--
2.26.3
^ permalink raw reply related [flat|nested] 8+ messages in thread* [Buildroot] [PATCHv2 2/3] toolchain/toolchain-external: fixup gdb pretty-printer loader for libstdcxx
2021-06-08 15:18 [Buildroot] [PATCHv2 0/3] Fix gdb pretty printers for libstdcxx Thomas De Schampheleire
2021-06-08 15:18 ` [Buildroot] [PATCHv2 1/3] package/gdb: fix gdb 9.2 segfault with Python 3.9 support Thomas De Schampheleire
@ 2021-06-08 15:18 ` Thomas De Schampheleire
2021-07-25 21:23 ` Thomas Petazzoni
2021-06-08 15:18 ` [Buildroot] [PATCHv2 3/3] toolchain: mark sysroot as 'safe' path for gdb auto-load Thomas De Schampheleire
2 siblings, 1 reply; 8+ messages in thread
From: Thomas De Schampheleire @ 2021-06-08 15:18 UTC (permalink / raw)
To: buildroot
From: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
gcc installs a libstdcxx-...so-gdb.py file that gdb will load automatically
when it loads libstdcxx.so, via the mechanism described at [1].
However, the auto-load file installed by gcc contains hardcoded paths
referring to the location where the (external) toolchain was built, which
are normally not available.
Fix up the paths in the load file so that the pretty printers can be loaded
automatically.
Note that gdb will only auto-load the file if its location is marked as
'safe'. A subsequent commit will take care of that.
Technically, there could be more than one load file, e.g. in lib and
usr/lib, so fix them all. This was for example observed in
BR2_TOOLCHAIN_EXTERNAL_ARM_AARCH64.
In a very specific case with a local custom toolchain, there were actually
two 'python' directories, which would break the sed command, so arbitrarily
limit to the first one encountered.
[1] https://sourceware.org/gdb/onlinedocs/gdb/objfile_002dgdbdotext-file.html
Signed-off-by: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
---
.../pkg-toolchain-external.mk | 21 +++++++++++++++++++
1 file changed, 21 insertions(+)
diff --git a/toolchain/toolchain-external/pkg-toolchain-external.mk b/toolchain/toolchain-external/pkg-toolchain-external.mk
index 6d91cb5d1e..d7fb93ee20 100644
--- a/toolchain/toolchain-external/pkg-toolchain-external.mk
+++ b/toolchain/toolchain-external/pkg-toolchain-external.mk
@@ -485,6 +485,26 @@ define TOOLCHAIN_EXTERNAL_INSTALL_GDBINIT
fi
endef
+# GCC installs a libstdcxx-...so-gdb.py file that gdb will load automatically,
+# but it contains hardcoded paths referring to the location where the (external)
+# toolchain was built. Fix up these paths so that the pretty printers can be
+# loaded automatically.
+# By default, the pretty printers are installed in
+# $(datadir)/gcc-$(gcc_version)/python but this could have been overwritten with
+# the gcc configure option: --with-python-dir. We thus have to search the
+# correct path first.
+define TOOLCHAIN_EXTERNAL_FIXUP_PRETTY_PRINTER_LOADER
+ $(Q)loadfiles=$$(find $(STAGING_DIR) -name 'libstdc++.so*-gdb.py' 2>/dev/null); \
+ pythondir=$$(find $(TOOLCHAIN_EXTERNAL_DOWNLOAD_INSTALL_DIR) -path '*/libstdcxx/__init__.py' 2>/dev/null | xargs dirname | xargs dirname | head -n1); \
+ if [ -n "$$loadfiles" ] && [ -n "$$pythondir" ]; then \
+ echo "Fixing up hardcoded paths in GDB pretty-printer auto-load file(s) for libstdcxx: $$loadfiles"; \
+ sed -ri \
+ -e 's%^libdir\s*=.*%libdir = "$(STAGING_DIR)/lib"%' \
+ -e "s%^pythondir\s*=.*%pythondir = '$$pythondir'%" \
+ $$loadfiles; \
+ fi
+endef
+
# uClibc-ng dynamic loader is called ld-uClibc.so.1, but gcc is not
# patched specifically for uClibc-ng, so it continues to generate
# binaries that expect the dynamic loader to be named ld-uClibc.so.0,
@@ -589,6 +609,7 @@ define $(2)_INSTALL_STAGING_CMDS
$$(TOOLCHAIN_EXTERNAL_INSTALL_SYSROOT_LIBS)
$$(TOOLCHAIN_EXTERNAL_INSTALL_WRAPPER)
$$(TOOLCHAIN_EXTERNAL_INSTALL_GDBINIT)
+ $$(TOOLCHAIN_EXTERNAL_FIXUP_PRETTY_PRINTER_LOADER)
endef
# Even though we're installing things in both the staging, the host
--
2.26.3
^ permalink raw reply related [flat|nested] 8+ messages in thread