The exciting 'scrambled eggs' transformations performed on the
x86 register files as it passes from hw to perf to DWARF are
quite something and require some magic numbers (sources carefully cited)
to be made to work correctly. The previous code worked deceptively
well but missed various cases.
h/t to qt perfparser for showing the precedent that this can be done,
even if our magic numbers are slightly different (due to omitting
segment registers).
Serhei Makarov [Thu, 22 Aug 2024 20:24:44 +0000 (16:24 -0400)]
eu-stacktrace WIP: sketch how to handle PERF_SAMPLE_REGS_ABI_32
Needs some additional debugging. But we accept the abi constant
from Sysprof (added in the recent patch 49115e50
in https://git.sr.ht/~serhei/sysprof-experiments)
and check it at the correct locations. Now we just need the
register layout table to get rid of the possible-incorrect magic
number code....
Serhei Makarov [Mon, 24 Jun 2024 15:29:03 +0000 (11:29 -0400)]
eu-stacktrace WIP: match sysprof's handling of SIGINT
Without the signal handler, eu-stacktrace gets terminated early
by a SIGINT inherited from a parent sysprof process, and isn't
able to finish processing all the data packets.
Serhei Makarov [Thu, 16 Nov 2023 21:06:03 +0000 (16:06 -0500)]
eu-stacktrace WIP: introduce 'naive' unwinder mode, very initial code
Still a number of pieces missing:
- recycle Dwfl structs sensibly
- create and output callchain frame
- compare to baseline sysprof unwinding, improve data quality
eu-stacktrace WIP: include sysprof-capture-types.h
This header defines the Sysprof data format; we could make it optional
through configury, but for now the prototype does not support any other
data format.
eu-stacktrace is (will be) a utility to process a stream of stack
samples (such as those obtained from PERF_EVENT_SAMPLE) into a stream
of stack traces (such as those obtained from PERF_EVENT_CALLCHAIN),
freeing various profiling utilities from having to implement their own
backtracing logic.
Aaron Merey [Fri, 17 Mar 2023 15:20:50 +0000 (11:20 -0400)]
debuginfod-client.c: Skip empty file creation for cancelled queries
Empty files in the client cache are used to indicate that contacted
servers could not find a requested resource. Future queries for this
resource will not be attempted until the cache_miss_s duration has
passed.
Currently these empty files are also created when a query is cancelled
through the client's progressfn. This can occur, for example, when a
user cancels a download with ctrl-c.
This prevents user-cancelled queries from being retried promptly without
having to modify cache_miss_s. Fix this by skipping the creation of an
empty cache file when progressfn cancels a query.
The cause of the error is the use of a non-existent == operator in
expressions passed to the 'test' built-in command. POSIX shell
specifies only an = operator for testing string equality.
* configure.ac: Replace == by = in use_msan and enable_demangler
comparisons.
Mark Wielaard [Fri, 3 Mar 2023 16:10:30 +0000 (17:10 +0100)]
Fix static build when zstd is enabled for libelf
When doing a static build (as with --enable-gcov) the build fails
because the executables aren't linked with -lzstd. Also the
run-debuginfod-query-retry.sh test relies on libdebuginfod.so.1 which
isn't there.
Fix by adding $(zstd_LIBS) to the static libelf lines and using the
debuginfod-find executable itself in the testcase.
* src/Makefile.am (libelf): When BUILD_STATIC add $(zstd_LIBS)
* tests/Makefile.am (libelf): Likewise.
* tests/run-debuginfod-query-retry.sh: Use debuginfod-find instead
of libdebuginfod.so.1
Mark Wielaard [Fri, 3 Mar 2023 14:37:18 +0000 (15:37 +0100)]
libelf: Document and make ELFCOMPRESS_ZSTD usable with old system elf.h
Make sure that even if the system elf.h doesn't have ELF_COMPRESS_ZSTD
defined it can still be used as constant. Also update libelf.h
documentation and add new feature to NEWS.
Mark Wielaard [Fri, 3 Mar 2023 13:13:49 +0000 (14:13 +0100)]
readelf: Fix use-after-free ebl pointer issue
With -flto gcc 13 sees that we use the ebl pointer after closing and
freeing it.
In function ‘process_elf_file’,
inlined from ‘process_dwflmod’ at readelf.c:818:3:
readelf.c:1070:6: error: pointer ‘ebl_18’ used after ‘free’ [-Werror=use-after-free]
1070 | if (pure_ebl != ebl)
| ^
In function ‘ebl_closebackend’,
inlined from ‘process_elf_file’ at readelf.c:1068:3,
inlined from ‘process_dwflmod’ at readelf.c:818:3:
../libebl/eblclosebackend.c:47:7: note: call to ‘free’ here
47 | free (ebl);
| ^
Fix by only calling ebl_closebackend after using it in the comparison.
Mark Wielaard [Wed, 22 Feb 2023 22:34:00 +0000 (23:34 +0100)]
libdw: Fix dwarf_getscopes memory leak on error
When there is an error in dwarf_getscopes after the initial scopes
have been allocated, e.g. when looking for the inlined scopes, then
the scopes would leak. Fix this by explicitly free the scopes on error.
Mark Wielaard [Fri, 17 Feb 2023 13:52:04 +0000 (14:52 +0100)]
libasm: Fix use-after-free issue with circular single linked list cleanup
Pointed out by gcc 12 with -Wuse-after-free=3
In function ‘free_section’
asm_end.c:552:17: error: pointer ‘data’ used after ‘free’ [-Werror=use-after-free]
552 | while (oldp != scnp->content);
| ~~~~~^~~~~~~~~~~~~~~~
asm_end.c:550:9: note: call to ‘free’ here
550 | free (oldp);
| ^~~~~~~~~~~
Mark Wielaard [Wed, 15 Feb 2023 16:00:30 +0000 (17:00 +0100)]
configure: Error when demangler is enabled, but libstdc++ support isn't
There have been a couple of cases where demangler support has been
accidentially been disabled because libstdc++ wasn't there/installed
during build. Make it an error if --disable-demangler isn't explicitly
given.
Add sanity check making sure an leb128 isn't being read beyond the
end of the current data segment. Most code already had these guards,
but some were missing. This makes sure an appropriate error is
generated instead.
Aleksei Vetrov [Mon, 13 Feb 2023 20:10:05 +0000 (20:10 +0000)]
libdw: check memory access in get_(u|s)leb128
__libdw_get_uleb128 and __libdw_get_sleb128 should check if addrp has
already reached the end before unrolling the first step. It is done by
moving __libdw_max_len to the beginning of the function, which can
notice, that addrp is beyond the end. Then we just check the result of
this function.
Mark Wielaard [Thu, 26 Jan 2023 17:19:15 +0000 (18:19 +0100)]
backends: Handle DW_TAG_unspecified_type in dwarf_peeled_die_type
binutils 2.40 introduces DW_TAG_unspecified_type for assembly
functions with an unknown return type. This breaks the
run-funcretval.sh testcase because dwfl_module_return_value_location
returns an error for such functions because it cannot determine the
return value location. Fix that by treating DW_TAG_unspecified_type
as if the DIE doesn't have a DW_AT_type.
Also update the testcase to explicitly checking for
DW_TAG_unspecified_type and printing "returns unspecified type".
Ilya Leoshkevich [Mon, 13 Feb 2023 16:45:51 +0000 (17:45 +0100)]
configure: Add --enable-sanitize-memory
Add support for clang Memory Sanitizer [1], which detects the usage of
uninitialized values. While elfutils itself is already checked with
valgrind, checking code that depends on elfutils requires elfutils to
be built with MSan.
MSan is not linked into shared libraries, and is linked into
executables statically. Therefore, unlike the other sanitizers, MSan
needs to be configured fairly early, since we need to drop
-D_FORTIFY_SOURCE [2], -Wl,-z,defs and --no-undefined.
Disable a few tests that run for more than 5 minutes due to test files
being statically linked with MSan.
Ilya Leoshkevich [Mon, 13 Feb 2023 16:45:50 +0000 (17:45 +0100)]
backends: Support returning lvalue and rvalue references
On the low level, they are the same as pointers. The change needs to be
done for all backends, so define a function and a macro to avoid
repetition. Also add a native test, which has to be implemented in C++.
Add the configure check for it.
The default linkage for consts in C++ is internal, so declare them
extern. While at it, remove the debuginfod workaround for
argp_program_version_hook.
Co-developed-by: Mark Wielaard <mark@klomp.org> Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
By default, clang produces .llvm_addrsig sections [1]. The GNU
toolchain does not know how to handle them yet [2], so just ask clang
not to generate them for the time being.
readelf.c:12205:72: error: parameter 'desc' set but not used [-Werror,-Wunused-but-set-parameter]
handle_bit_registers (const Ebl_Register_Location *regloc, const void *desc,
^
Mark Wielaard says:
It is never really used since as far as I can see we don't have any
backend with a core register sets where a register doesn't have a
number of bits which isn't a multiple of 8 (only ia64 has some 1
bit registers, but those don't seem part of the core register set).
If we do accidentally try to handle such a register having an abort
is also not very nice. Lets just warn and return/continue.
asm_newscn.c:48:22: error: field 'pattern' with variable sized type 'struct FillPattern' not at the end of a struct or class is a GNU extension [-Werror,-Wgnu-variable-sized-type-not-at-end]
struct FillPattern pattern;
^
Fix by using a union instead. Define the second union member to be a
char array 1 byte larger than struct FillPattern. This should be legal
according to 6.7.9:
If an object that has static or thread storage duration is not
initialized explicitly, then ... if it is a union, the first named
member is initialized (recursively) according to these rules, and
any padding is initialized to zero bits.
Uninitialized value was created by an allocation of 'reglocs' in the stack frame
#0 0x562d35c686f0 in handle_core_note elfutils/src/readelf.c:12674:3
#const Ebl_Register_Location *reglocs;
==1006199==WARNING: MemorySanitizer: use-of-uninitialized-value
#0 0x562d35c68a2a in handle_core_note elfutils/src/readelf.c:12692:11
#colno = handle_core_registers (ebl, ebl->elf, desc + regs_offset,
# reglocs, nregloc);
Strictly speaking, this is not a problem, because nregloc == 0, but for
other note types we initialize it anyway, so do it here as well.
Aaron Merey [Wed, 8 Feb 2023 02:15:33 +0000 (21:15 -0500)]
debuginfod-client.c: Download section even if cached executable didn't contain it.
Before attempting to download a section, cache_find_section tries to
extract the section from existing files in the cache. If it's determined
that the section must not exist, cache_find_section returns -ENOENT to
indicate that the download should be skipped.
This patch fixes a bug where cache_find_section returns -ENOENT even
though the section exists. If the cache contains the executable but
not the debuginfo with the given build-id and the section only exists
in the debuginfo (such as any of the .debug_* sections), then
debuginfod_find_section returns -ENOENT even if the section could be
downloaded.
Fix this by having cache_find_section not return -ENOENT unless cached
debuginfo was able to be read.
Mark Wielaard [Mon, 6 Feb 2023 09:21:58 +0000 (10:21 +0100)]
debuginfod: Make sure there is only one typedef for debuginfod_client
Both debuginfod.h and libdwfl.h have a simple typedef for struct
debuginfod_client. Some compilers pedantically warn when including
both headers that such typedefs are only officially supported in
C11. So guard them with _ELFUTILS_DEBUGINFOD_CLIENT_TYPEDEF to
make them happy.
Mark Wielaard [Sat, 21 Jan 2023 23:31:57 +0000 (00:31 +0100)]
libdw: Search for abstract origin in the correct CU
With gcc -flto the abstract origin of an inlined subroutine
could be in a different CU. dwarf_getscopes might return an
empty scope if it cannot find the abstract origin scope. So
make sure to search in the
We also tried to add the origin match in pc_record directly
in the current inlined scope. This always failed, causing
to do a needless traversal, followed by the full CU scan in
dwarf_getscopes. Just always stop the pc_record search and
then do the CU origin_match in dwarf_getscopes.
Mark Wielaard [Sun, 22 Jan 2023 19:50:53 +0000 (20:50 +0100)]
addr2line: -C should not try to handle optional (ignored) argument
The --demangle option takes an option (ignored) argument. Since -C
is the short option of --demangle it also takes that optional
argument. But that means that something like -Cfi is like -C got
fi as argument, while the user expects -Cfi to be like -C -f -i.
Separate the --demangle and -C options. --demangle still takes
an optional (ignored) argument, but -C doesn't take any arguments
so -Cfi acts the same as -fiC.
Also fix --target, -b, so that it is properly ignored (and not
accidentially enables demangling).
Mark Wielaard [Thu, 19 Jan 2023 22:53:33 +0000 (23:53 +0100)]
tests: run-debuginfod-query-retry.sh /bin/sh might not have a build-id
If /bin/sh wasn't build with build-id the test fails because
debuginfod-find will not be able to even try to fetch something.
Use libdebuginfod.so.1 instead as test file because we just
build that and configure makes sure we always produce a build-id.
Mark Wielaard [Tue, 10 Jan 2023 17:18:05 +0000 (18:18 +0100)]
readelf: Check compression status of .debug section data
The various print_debug_*_section functions didn't get the section
data in the same way. Add a new get_debug_elf_data function that
gets the (possibly relocated) section data and that checks (and
warns) if the data might still be compressed in a way that we
cannot decompress.
Serhei Makarov [Tue, 10 Jan 2023 16:20:13 +0000 (11:20 -0500)]
debuginfod-find.1: note on how to find a comp_dir
doc/ChangeLog:
* debuginfod-find.1: add a note to DESCRIPTION section for the 'source'
subcommand, clarifying where to find the CU compilation-directory.
I'm looking at how to improve and document the workflow for using
debuginfod-find to study the behaviour of packaged software on a system, e.g.
in conjunction with a tracing tool like SystemTap. You can read the exact
compiled source code with debuginfod-find source and use that to figure out
which source code locations are interesting to trace. But that process has
several non-obvious or inconvenient steps.
This patch is a clarification to the debuginfod-find source man page pointing
out the eu-readelf command that can show the comp_dir in downloaded debuginfo.
Since debuginfod-find source could be picked up and used by a developer
unfamiliar with DWARF terminology, I think such a clarification saves time
for newbies figuring out what a CU compilation-directory is.