H.J. Lu [Thu, 12 Mar 2015 21:51:42 +0000 (14:51 -0700)]
backends: Add x32_corenote.c
This patch adds x32_corenote.c to support x32 coredump. X32 coredump is
a hybrid between ia32 coredump and x86-64 coredump. The exact formats
are described in bfd/hosts/x86-64linux.h in GNU binutils source tree.
Petr Machata [Wed, 18 Mar 2015 18:42:02 +0000 (19:42 +0100)]
Publish known-dwarf.h
- Make the macro names properly scoped (DWARF_ALL_KNOWN_* and
DWARF_ONE_KNOWN_*).
- Simplify the calling convention: don't keep track of the comment
associated with the enumerator, always use the non-_DESC
DWARF_ONE_KNOWN_* callback.
- Install known-dwarf.h alongside libdw.h and others.
Mark Wielaard [Mon, 23 Mar 2015 21:19:12 +0000 (22:19 +0100)]
libelf: Don't extend ELF file size with SHT_NOBITS sh_offset.
Don't explicitly extend the file size for SHT_NOBITS sections. Since
that could cause a size beyond any actual file content it will cause
issues when the underlying ELF file has been mmapped or will extend
the file size to increase (writing fill bytes) when not mmapped. The
sh_offset value is essentially meaningless for SHT_NOBITS. gabi says
that a NOBITS section sh_offset member locates the "conceptual
placement" in the file. But it doesn't say this cannot be beyond the
enf of the file. When ELF_F_LAYOUT is set we should trust sh_offset
as given is what is wanted for an SHT_NOBITS section without extending
the file size.
Buggy binutils ld could generate files where SHT_NOBITS sections have
sh_offset outside the file.
https://sourceware.org/bugzilla/show_bug.cgi?id=12921
Mark Wielaard [Mon, 23 Mar 2015 14:08:27 +0000 (15:08 +0100)]
backends: aarch64 can have _GLOBAL_OFFSET_TABLE_ point anywhere in .got.
When building with relro aarch64 splits .got in a writable and read-only
part. _GLOBAL_OFFSET_TABLE_ points to the first writable symbol, not the
start of .got.
Mark Wielaard [Mon, 9 Mar 2015 15:41:48 +0000 (16:41 +0100)]
elflint: Correct gnu_symbias usage in compare_hash_gnu_hash.
commit 0a545e "elflint: Add various low-level checks." introduced a bug
in the handling of the gnu_symbias. The symbol in the bucket should first
be compared to STN_UNDEF before applying the bias. And the used symbol
index should have the bias added when compared against plain hash symbol.
Mark Wielaard [Wed, 18 Feb 2015 19:51:40 +0000 (20:51 +0100)]
libelf: Make __libelf_set_data_list_rdlock an internal_function.
__libelf_set_data_list_rdlock from elf_getdata.c is marked as an
internal_function in the implementation, but not in libelfP.h when it
is declared. Add internal_function to the declaration. This broke
the i686 build.
Reported-by: Alexander Cherepanov <ch3root@openwall.com> Signed-off-by: Mark Wielaard <mjw@redhat.com>
Josh Stone [Wed, 11 Feb 2015 21:53:14 +0000 (13:53 -0800)]
libdw: ensure read_encoded_value's value is set
With CFLAGS='-Og -g', F21 gcc hits -Werror=maybe-uninitialized in
read_encoded_value at "*result += value". It's fine with -O2/-O0.
In particular it seems to care about the __libdw_cfi_read_address_inc
calls. By my inspection, the only way those don't set value are for
error cases, which will then return immediately. This patch just sets
value = 0 to begin with, so gcc is always convinced it's fine.
Petr Machata [Wed, 11 Feb 2015 21:11:46 +0000 (22:11 +0100)]
dwarf_ranges: do not request base address attributes if not necessary
- Instead of bailing out at iteration start, we remember the
base-address-selecting attributes were not seen, and then bail out
later if no base address selection entry has been seen.
Jan Kratochvil [Sat, 7 Feb 2015 14:08:20 +0000 (15:08 +0100)]
libelf: Consider sh_addralign 0 as 1
Currently the Koji build for arm32 fails with:
extracting debug info from /builddir/build/BUILDROOT/etcd-2.0.0-0.3.rc1.fc22.arm/usr/bin/etcd
Failed to write file: invalid section alignment
This is because the binary etcd
http://people.redhat.com/jkratoch/etcdctl.xz
contains:
Section Headers:
[Nr] Name Type Addr Off Size ES Flg Lk Inf Al
[11] .rel.plt REL 00459ee0 449ee0 000088 08 A 13 0 0
^
which corresponds to golang's code:
go/src/cmd/ld/elf.c
case EM_X86_64:
sh = elfshname(".rela.plt");
sh->addralign = RegSize;
default:
sh = elfshname(".rel.plt");
<nothing>
ELF spec says:
Values 0 and 1 mean the section has no alignment constraints.
and libelf/elf32_updatenull.c really parses it that way at line 204
ElfW2(LIBELFBITS,Word) sh_align = shdr->sh_addralign ?: 1;
but unfortunately the later line being patched no longer does.
libelf/
2015-02-07 Jan Kratochvil <jan.kratochvil@redhat.com>
* elf32_updatenull.c (__elfw2(LIBELFBITS,updatenull_wrlock)): Consider
sh_addralign 0 as 1.
Signed-off-by: Jan Kratochvil <jan.kratochvil@redhat.com>
Mark Wielaard [Mon, 26 Jan 2015 15:34:57 +0000 (16:34 +0100)]
libdwfl: Clear symtab result on error before using find_dynsym fallback.
find_dynsym could succeed after find_symtab failed but had already set up
symdata, syments and first_global. find_dynsym would not set or clear
first_global since all syms in dynsym are global. Causing lots of failing
lookups when calling __libdwfl_addrsym if first_global was some huge
bogus value.
Reported-by: Hanno Böck <hanno@hboeck.de> Signed-off-by: Mark Wielaard <mjw@redhat.com>
Mark Wielaard [Thu, 22 Jan 2015 11:49:29 +0000 (12:49 +0100)]
libelf: Make sure string returned by elf_strptr is NUL terminated.
The result of elf_strptr is often used directly to print or strcmp
the string. If the section data was truncated or corrupted that could
lead to invalid memory reads possibly crashing the application.
https://bugzilla.redhat.com/show_bug.cgi?id=1170810#c24
Reported-by: Alexander Cherepanov <cherepan@mccme.ru> Signed-off-by: Mark Wielaard <mjw@redhat.com>
Mark Wielaard [Tue, 20 Jan 2015 20:55:55 +0000 (21:55 +0100)]
libelf: Fix elf_newdata when raw ELF file/image data is available.
When ELF data for a section has been read by elf_rawdata, data_read
and rawdata_base are set, but data_list_rear will not be set until the
data will be converted (by elf_getdata). elf_newdata would overwrite
the existing data in that case. Both elf_getdata and elf_update rely
on the fact that when data_list_rear is set they don't have to look
at the raw data anymore. So make sure we update the data list properly
before adding any new data and raw data is available in elf_newdata.
Add newdata test that calls elf_newdata before and after elf_rawdata
and elf_getdata and checks the new size and contents of the section.
Mark Wielaard [Tue, 20 Jan 2015 13:52:41 +0000 (14:52 +0100)]
libelf: elf_strptr should fetch the shdr for the section if not yet known.
elf_strptr might be called before the shdrs are read in. In that case it
needs to explicitly call __elf[32|64]_getshdr_rdlock to check the section
type and size. The new strptr testcase triggers this corner case and crashes
before the fix.
Mark Wielaard [Thu, 15 Jan 2015 12:39:06 +0000 (13:39 +0100)]
tests: Make deleted and vdsosyms testcases work with "restricted ptrace".
Some systems might have "restricted ptrace" that doesn't allow process
inspection of arbitrary processes. Change the deleted testcase to
explicitly allow any other process to inspect it using the PR_SET_PTRACER
prctl set to PR_SET_PTRACER_ANY. Change the vdsosyms testcase to inspect
the process itself which should always be allowed.
Reported-by: Anatol Pomozov <anatol.pomozov@gmail.com> Signed-off-by: Mark Wielaard <mjw@redhat.com>
Mark Wielaard [Wed, 24 Dec 2014 12:17:23 +0000 (13:17 +0100)]
libdw: Search for the last matching address with dwarf_getsrc_die.
In commit 7d9b5a dwfl_module_getsrc was changed so that it returns the last
line record <= addr, rather than returning immediately on a match. This
changes dwarf_getsrc_die to do the same. And it adds a new test that checks
this by comparing against the same results from eu-addr2line (which uses
dwfl_module_getsrc) using dwarf_addrdie and dwarf_getsrc_die instead.
Mark Wielaard [Sat, 27 Dec 2014 15:16:29 +0000 (16:16 +0100)]
libdwfl: dwfl_module_getsrc should never match end_sequence line.
The line with end_sequence set has an address outside the current line
sequence. An end_sequence line has no other useful information except
marking the address as out of range.
Two tests, addrscopes and funcscopes, depended on matching the end_sequence
line. But that was because they included the high_pc address in the scope.
However the high_pc attributes has as address the first location past the
range associated with a given DIE. Adjust the tests to use high_pc - 1 as
end of the scope.
Mark Wielaard [Wed, 7 Jan 2015 22:39:28 +0000 (23:39 +0100)]
libdw: Robustify eh_frame_hdr and encoded-values reading.
Sanity check and keep track of binary_search_table data buffer length.
Add bounds check to encoded value reading. Also fix a bug when reading
the eh_frame header data from an other endian ELF image. Add a testcase
that would fail the new sanity checks because of the endian bug.
Mark Wielaard [Sun, 4 Jan 2015 23:12:53 +0000 (00:12 +0100)]
libdwfl: Handle seeing CU EOF marker without any other (real) CU.
intern_cu (and __libdwfl_nextcu) didn't correctly handle immediately seeing
the CU EOF marker. In that case the *nextp pointer (first_cu) returned by
intern_cu is -1. And less_layzy will immediately destroy the found CU node.
So don't use the *found result after calling less_lazy.
Mark Wielaard [Sun, 4 Jan 2015 23:03:03 +0000 (00:03 +0100)]
libdw: Check DW_AT_sibling attribute offset is after current DIE.
The sibling attribute should point after this DIE in the CU.
Otherwise various algorithms might loop or go into infinite recursion
walking the DIE tree.
Mark Wielaard [Sun, 4 Jan 2015 15:08:22 +0000 (16:08 +0100)]
libdw: Check register number in CFI isn't insanely large.
Some cfi.c array size allocation calculations might overflow when trying
to accommodate insanely large number of registers. Don't allow register
numbers larger than INT32_MAX / sizeof (dwarf_frame_register).
Mark Wielaard [Fri, 2 Jan 2015 23:27:58 +0000 (00:27 +0100)]
libdw: Add missing size check to parse_eh_frame_hdr.
afl-fuzz showed that when parse_eh_frame_hdr was called from
getcfi_scn_eh_frame invalid data could be read because of a missing size
check. That check was there when parse_eh_frame_hdr was called from
getcfi_gnu_eh_frame. Move the size check into parse_eh_frame. And add
an extra check to make sure d_buf is not NULL.
Mark Wielaard [Fri, 2 Jan 2015 23:09:45 +0000 (00:09 +0100)]
libdw: Handle NULL dbg for call_ref, GNU_implicit_pointer and addr DW_OPs.
afl-fuzz pointed out that __libdw_intern_expression didn't handle CFI
containing DW_OP_call_ref, DW_OP_implicit_pointer or DW_OP_addr. Because
in that case the Dwarf dbg is NULL. Both DW_OP_call_ref and
DW_OP_implicit_pointer cannot be used in CFI. That is just an error.
But DW_OP_addr can be. Without a Dwarf dbg we'll need to read the address
argument directly. Don't use __libdw_read_address_inc which might do a
relocation of the value read. But in practice the relocation hook isn't
implemented anyway.
Mark Wielaard [Tue, 30 Dec 2014 23:51:45 +0000 (00:51 +0100)]
libelf: gelf_getphdr should check phdr index is valid.
elf_getphdrnum does checks the phdrnum makes sense. But gelf_getphdr
checked the given index against the "raw" e_phnum or internal
__elf_getphdrnum_rdlock result without checking. Extract the checking
code into a new internal __elf_getphdrnum_chk_rdlock function and
use that.
Mark Wielaard [Thu, 25 Dec 2014 21:50:21 +0000 (22:50 +0100)]
libelf: ar_size cannot be negative. Fix max ar size.
Elf_Arhdr ar_size is loff_t, which is signed. Make sure it isn't negative.
When the parent start_offset is non-zero maxsize should include it to
compensate for ar offset.
Mark Wielaard [Sun, 11 Jan 2015 14:00:52 +0000 (15:00 +0100)]
libdw: Prevent infinite recursion when processing DW_TAG_imported_unit.
Invalid DWARF could create cycles with DW_TAG_imported_unit, which would
lead to infinite recursion and stack overflow in libdw_visit_scopes.
Keep track of imported units and error out when a cycle is detected.
read_long_names terminates names at the first '/' found but then skips
one character without checking (it's supposed to be '\n'). Hence the
next name could start with any character including '/'. This leads to
a directory traversal vulnerability at the time the contents of the
archive is extracted.
The danger is mitigated by the fact that only one '/' is possible in a
resulting filename and only in the leading position. Hence only files
in the root directory can be written via this vuln and only when ar is
executed as root.
The fix for the vuln is to not skip any characters while looking
for '/'.
Signed-off-by: Alexander Cherepanov <cherepan@mccme.ru>
Mark Wielaard [Fri, 19 Dec 2014 19:53:22 +0000 (20:53 +0100)]
tests: Don't match libfunc symbol on ppc64 in run-deleted.sh testcase.
For PPC64 we need access to the OPD table which we get through the shdrs
but for the deleted-lib we only have phdrs. So we don't have the name of
the function. But since we should find the EH_FRAME through phdrs just
fine, we can unwind into main. Skip the libfunc name check for now till
we find a nice way to enable the lookup without opd:
https://bugzilla.redhat.com/show_bug.cgi?id=1176238
Mark Wielaard [Fri, 19 Dec 2014 17:19:29 +0000 (18:19 +0100)]
libdwfl: Always try find_dynsym last, also on find_symtab ELF errors.
An ELF error during find_symtab might be incomplete shdrs or string table
from ELF images read through elf_from_memory. The phdrs and dynsym should
always be complete.
Mark Wielaard [Fri, 19 Dec 2014 14:40:16 +0000 (15:40 +0100)]
libdwfl: Remove p_align sanity check from elf_from_memory.
In commit f15bcd "elf_from_remote_memory should use pagesize, not p_align"
we already relaxed the p_align sanity check to allow alignment of the
segment against the pagesize since that is what the glibc dynamic linker
actually does. But on some architectures the kernel inserts a vdso with
a completely bogus p_align for some PT_LOAD segments. So just drop the
whole sanity check and allow anything since we won't use p_align, but
always already use pagesize anyway.
Ulrich Drepper [Thu, 18 Dec 2014 17:32:39 +0000 (12:32 -0500)]
avoid scary command output
The commands to check for invalid text relocations in the generated DSOs
shouldn't be displayed. They contain an echo which prints the text.
This patch suppresses the commands from being printed.
Mark Wielaard [Mon, 15 Dec 2014 18:43:02 +0000 (19:43 +0100)]
readelf: Warn, don't assert, if loclist or rangelist offset is too big.
We use a couple of bits to keep track of the addr size, dwarf size and
warning given for a loclist or rangelist offset. Normally offset won't
be that big and will fit in 61-bits easily. But if not, don't assert,
but just warn we don't handle such large offsets.
Mark Wielaard [Sun, 14 Dec 2014 20:48:23 +0000 (21:48 +0100)]
libdw: Add get_uleb128 and get_sleb128 bounds checking.
Both get_uleb128 and get_sleb128 now take an end pointer to prevent
reading too much data. Adjust all callers to provide the end pointer.
There are still two exceptions. "Raw" dwarf_getabbrevattr and
read_encoded_valued don't have a end pointer associated yet.
They will have to be provided in the future.
Mark Wielaard [Fri, 12 Dec 2014 15:43:04 +0000 (16:43 +0100)]
libdw: Make sure all attributes come with a (fake) CU for bound checks.
All attributes now have a reference to a (fake) CU that has startp and
endp set to the data section where the form data comes from. Use that
for bounds checking in __libdw_form_val_len and dwarf_formblock to make
sure data read doesn't overflow any data section. Remove libdwP.h cu_data
and use cu startp and endp directly where appropriate.
Mark Wielaard [Sat, 13 Dec 2014 23:09:29 +0000 (00:09 +0100)]
Guard against 64bit unsigned wrap around in (int) compare functions.
Dwarf_Adrr and Dwarf_Off are 64-bit unsigned, and comparison functions
used in qsort or tfind return int, it is possible for the difference to
be so large that it wraps around. Make sure to just return -1, 0 or 1
in compare_aranges and compare_cukey.
Mark Wielaard [Fri, 12 Dec 2014 20:51:27 +0000 (21:51 +0100)]
tests: Add an explicit -fasynchronous-unwind-tables for deleted-lib.so.
Some systems don't add a PT_GNU_EH_FRAME segment by default, but the
deleted test depends on it so that it can get at the .eh_frame_hdr table
through the phdrs from the remote memory ELF image. Request one explicitly
with -fasynchronous-unwind-tables.
Josh Stone [Fri, 12 Dec 2014 01:19:23 +0000 (17:19 -0800)]
tests: Test addr2line -i with lexical blocks present
This test is for the fix in commit aecdf2670c02, but it was hampered by
unlucky sorting of line records. With the new stable sort, the test now
gets the desired line numbers.
Josh Stone [Fri, 12 Dec 2014 00:23:46 +0000 (16:23 -0800)]
libdwfl: Search for the last matching address in lines
Now that libdw's srclines use a stable sort, we can reliably choose the
*last* matching line record for a given address, which should be the
innermost where inlines are concerned.
projects / thirdparty / elfutils.git / log
