Max Filippov [Wed, 30 Nov 2022 01:47:54 +0000 (17:47 -0800)]
xtensa: fix text relocations
The commit 5e08f1968316 ("Don't always update text in !pic_with_got case")
changed good_32bit_resolved_reloc to not do endianness swapping for
relocated entries in the text segment. This broke little-endian xtensa
FLAT images which after this change fail to start with the following
message:
binfmt_flat: reloc outside program 0x24c80100 (0 - 0x6e430/0x56a20)
Fix it by preserving 'update_text' when building for xtensa.
Fixes: 5e08f1968316 ("Don't always update text in !pic_with_got case") Reported-by: Niklas Cassel <[email protected]> Signed-off-by: Max Filippov <[email protected]>
Damien Le Moal [Wed, 9 Sep 2020 08:31:33 +0000 (17:31 +0900)]
elf2flt: add riscv 64-bits support
Add support for riscv 64bits ISA by defining the relocation types
R_RISCV_32_PCREL, R_RISCV_ADD32, R_RISCV_SUB32, R_RISCV_32 and
R_RISCV_64. riscv64 support also needs the __global_pointer$ symbol to
be defined right after the relocation tables in the data section. To
define this symbol, the "RISCV_GP" line prefix is added. The "RISCV_GP"
string is removed if the target CPU type is riscv64 and the definition
line is dropped for other CPU types.
With these changes, buildroot and busybox build and run on riscv NOMMU
systems with Linux kernel including patch 6045ab5fea4c
("binfmt_flat: do not stop relocating GOT entries prematurely on riscv")
fixing the binfmt_flat loader. Tested on QEMU and Canaan Kendryte K210
boards.
This patch is based on earlier work by Christoph Hellwig <[email protected]>.
elf2flt.ld: reinstate 32 byte alignment for .data section
Commit 8a3e74446fe7 ("allow to build arm flat binaries") moved the
following commands:
. = ALIGN(0x20) ;
@SYMBOL_PREFIX@_etext = . ;
from the .text section to the top level in the SECTIONS node.
The .text output section is being directed to a memory region using the
"> flatmem :text" output section attribute. Commands in the top level in
the SECTIONS node are not.
This means that the ALIGN() command is no longer being appended to the
flatmem memory region, it will simply update the Location Counter.
The heuristic for placing an output section is described here:
https://sourceware.org/binutils/docs-2.38/ld.html#Output-Section-Address
"If an output memory region is set for the section then it is added to this
region and its address will be the next free address in that region."
Since the .data section is being directed to the same memory region as the
.text section, this means that the Location Counter is not used when
assigning an address to the .data output section, it will simply use the
next free address.
No longer directing these commands to the flatmem memory region means that
the .data output section is no longer aligned to a 32 byte boundary.
Before commit 8a3e74446fe7 ("allow to build arm flat binaries"):
$ readelf -S busybox_unstripped.gdb | grep data
[ 3] .data PROGBITS 0000000000035ac000036ac0
$ readelf -s busybox_unstripped.gdb | grep _etext
19286: 0000000000035ac0 0 NOTYPE GLOBAL DEFAULT 1 _etext
After commit 8a3e74446fe7 ("allow to build arm flat binaries"):
$ readelf -S busybox_unstripped.gdb | grep data
[ 3] .data PROGBITS 0000000000035ab000036ab0
$ readelf -s busybox_unstripped.gdb | grep _etext
19287: 0000000000035ac0 0 NOTYPE GLOBAL DEFAULT 3 _etext
The .data output section has to be aligned to a 32 byte boundary, see the
FLAT_DATA_ALIGN 0x20 macro and its usage in fs/binfmt_flat.c:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/binfmt_flat.c?h=v5.17#n59
Readd an explicit ALIGN attribute on the .data section itself, since the
linker will obey this attribute regardless if being directed to a memory
region or not. Also remove the ALIGN() command before the .data section,
since this misleads the reader to think that the Location Counter is used
when assigning an address to the .data section, when it actually is not.
Fixes: 8a3e74446fe7 ("allow to build arm flat binaries") Signed-off-by: Niklas Cassel <[email protected]>
Binutils-2.40 introduced support for libzstd compressed bfd sections.
By default binutils configure will check for and use the system libzstd
library if it is present.
Now when compiling elf2flt using the bfd library from a
binutils-2.40 and a system present libzstd you will get link errors:
/usr/bin/ld: ..//binutils-2.40/bfd/libbfd.a(compress.o): in function `decompress_contents':
../binutils-2.40/bfd/compress.c:517: undefined reference to `ZSTD_decompress'
/usr/bin/ld: ../binutils-2.40/bfd/compress.c:519: undefined reference to `ZSTD_isError'
/usr/bin/ld: ../binutils-2.40/bfd/libbfd.a(compress.o): in function `bfd_compress_section_contents':
../binutils-2.40/bfd/compress.c:656: undefined reference to `ZSTD_compress'
/usr/bin/ld: ../binutils-2.40/bfd/compress.c:661: undefined reference to `ZSTD_isError'
Add a library check to the elf2flt configure for libzstd presence.
When present link using it.
There is no real problem when linking with -lzstd if compiling with
bfd libraries from older binutils, it just won't be needed. So there
is no need to make this check selective on binutils version.
In binutils-2.40 the BFD_VMA_FMT definition used for printf style
formatting specifiers has been removed. For reference this was done in
commit b82817674f46 ("Don't use BFD_VMA_FMT in binutils") in the
binutils git development tree.
BFD_VMA_FMT is used in a number of places in the elf2flt code to output
bfd offsets, values and the like. So it is broken when using the bfd
code from binutils-2.40 and newer.
According to the binutils change PRIx64 (and friends) is used to replace
it, with appropriate casts to keep it clean for 32 and 64 bit platforms.
Change the elf2flt.c use of it in the same way to fix.
This does not change the output in any way in normal use. This fix can
be used on all versions of binutils (older and newer), there is no
need to only do this on 2.40 and newer.
Greg Ungerer [Wed, 15 Mar 2023 13:45:48 +0000 (23:45 +1000)]
elf2flt: force ARM.exidx section into text
Fix problem with el2flt core dumping on binutils version 2.33.1 and
newer. Previous attempts at fixing this have had a list of subtle
problems causing further breakage.
The problem stems from the .ARM.exidx section being marked as load
readonly data, but the elf2flt.ld linker script puts it into the
flat text section. The elf2flt processing code expects all separate
ELF sections marked as "data" to be in the flat data section and
calculates sizing and memory based on that. (Note that traditional
readonly data is treated as text in the linker script, so it is not
affected here and is merged into the flat text section with no
problems).
The retaining of separate ELF sections outside of the flat text and
flat data flatmem sections proper is fragile. Fundamentally the
flat format only has the two stored sections - text and data. By far
the best appraoch is to explicilty put everything into those named
sections in the elf2flt.ld.
The .ARM.exidx section must be part of the flat text section - due
to the relocations that it carries. So move it into the text section
proper in the elf2flt.ld linker script.
Although this fixed the simple problem case (elf2flt core dumping on
binutils newer than 2.33.1) it causes further breakage that required
further hacks to fix that resulted in more breakage.
Back all of these changes out and find a better way.
Note that this is actually a partial revert. It preserves the comment
fix in the original commit.
This change breaks at least the m68k architecture support. It causes
overlapping text and data sections resulting in errors like this:
ERROR: text=0x38f34 overlaps data=0x31a74 ?
The underlying cause here is that the .eh_frame section is marked
as readonly relocated data on m68k, resulting in it being treated
now as if it will be loaded into the flat text section. But it
is layed out in the linker script as part of the flat data section.
Note that this is actually a partial revert. It preserves the debug
printf fix in the original commit.
Romain Naour [Wed, 5 Feb 2020 09:31:32 +0000 (10:31 +0100)]
elf2flt: handle binutils >= 2.34
The latest Binutils release (2.34) is not compatible with elf2flt due
to a change in bfd_section_* macros [1]. The issue has been reported
to the Binutils mailing list but Alan Modra recommend to bundle
libbfd library sources into each projects using it [2]. That's
because the API is not stable over the time without any backward
compatibility guaranties.
On the other hand, the elf2flt tools needs to support modified
version of binutils for specific arch/target [3].
Add two tests in the configure script to detect this API change
in order to support binutils < 2.34 and binutils >= 2.34.
Signed-off-by: Romain Naour <[email protected]>
Two changes made to original patch:
. $binutils_include_dir used for binutils version check
. configure script regenerated (run autoconf)
Mike Frysinger [Fri, 20 Aug 2021 06:28:40 +0000 (02:28 -0400)]
elf2flt.ld: harmonize directive styles
We use "FOO:" style for "SINGLE_LINK:" and "TOR:" tags, but omit the
colon with "R_RODAT" and "W_RODAT" tags. There's no reason for this
difference, and can be confusing to the casual reader. Switch the
latter two over to use tags to be consistent.
Mike Pilawa [Thu, 19 Aug 2021 08:04:25 +0000 (18:04 +1000)]
elf2flt: fix for segfault on some ARM ELFs
I believe a bug was introduced in commit [1], which was done to move the .ARM.exidx input section from .data to .text output section. However, in doing so, the dynamic memory allocation for .text [elf2flt.c:~L1907: 'text = xmalloc(text_len);'] was not modified to allow for the additional .ARM.exidx section. The result was that most of the time malloc() allocated enough extra memory [due to page-size or similar allocation boundaries] such that a small additional section went unnoticed. However, in unlucky cases, the memory required for just the .text section fit almost perfectly within an allocation boundary, and the extra .ARM.exidx section exceeded the allocation, and thus produced a segfault when the memory was written.
The fix here modifies the calculation of 'text_len' in main() with logic similar to that of the original fix in [1] to output_relocs(), such that the correct amount of memory is allocated. The code is also modified such that 'data_len' is also calculated correctly, and does not over-allocate memory. This is necessary because the logic in main() is not a single grouping of if-elseif... priority logic as in output_relocs().
The change also attempts to be even more specific with input section selections for .text and .data output sections.
.text is only selected for input sections flagged as (SEC_CODE || (SEC_DATA && SEC_READONLY && SEC_RELOC))
.data is only selected for input sections flagged as (SEC_DATA && !(SEC_READONLY && SEC_RELOC))
The change appears to work correctly for previously segfault-causing ELF with these processed sections...
SEC_FLAGS:0x0000011f SEC_NAME:.text
SEC_FLAGS:0x0000012f SEC_NAME:.ARM.exidx
SEC_FLAGS:0x00000127 SEC_NAME:.data
SEC_FLAGS:0x00000123 SEC_NAME:.tm_clone_table
SEC_FLAGS:0x0000012b SEC_NAME:.eh_frame
SEC_FLAGS:0x00000001 SEC_NAME:.bss
SEC_FLAGS:0x00000100 SEC_NAME:.stack
SEC_FLAGS:0x01800108 SEC_NAME:.comment
SEC_FLAGS:0x00000108 SEC_NAME:.ARM.attributes
SEC_FLAGS:0x0000210c SEC_NAME:.debug_aranges
SEC_FLAGS:0x0000210c SEC_NAME:.debug_info
SEC_FLAGS:0x00002108 SEC_NAME:.debug_abbrev
SEC_FLAGS:0x0000210c SEC_NAME:.debug_line
SEC_FLAGS:0x0000210c SEC_NAME:.debug_frame
SEC_FLAGS:0x01802108 SEC_NAME:.debug_str
SEC_FLAGS:0x0000210c SEC_NAME:.debug_loc
SEC_FLAGS:0x00002108 SEC_NAME:.debug_ranges
It should also be pointed out that this commit may impact the regression noted in [2] and pull-request [3].
With this code change in place, elf2flt will put section .eh_frame with flags=0x12b in the .data output section. If the flags for an .eh_frame section were 0x12f (same as .ARM.exidx), then it would end up in .text output section.
Mike Frysinger [Wed, 18 Aug 2021 21:29:57 +0000 (17:29 -0400)]
ld-elf2flt: handle NULL pointers with printf(%s) ourselves
Since C & POSIX leave this as undefined behavior, gcc will warn about
it which breaks our -Werror builds. Glibc handles this just fine (and
will print "(null)"), but since we support more platforms than just
glibc, add a helper macro to avoid the edge case.
Mike Frysinger [Wed, 9 Sep 2020 09:21:51 +0000 (05:21 -0400)]
configure: deprecate --disable-ld-elf2flt-binary
We've been defaulting to the compiled linker for years now and no one
has complained. Lets deprecate the old shell script so we don't have
to maintain two large tools doing the same thing.
Steve Bennett [Tue, 28 Jan 2020 06:22:08 +0000 (16:22 +1000)]
microblaze: properly handle relocations in resolved symbol mode
The default is now to use -a -p, but microblaze currently has no support
for this.
Add support for the appropriate relocations and ignore
pc-relative and got(r20)-relative relocations.
Also:
- binfmt_flat can't support GOT data-relative relocations, so error out in that case
- Only output relocations debug if the relocation was actually created
- Tidy up the non-resolved symbol mode for microblaze
Mike Frysinger [Wed, 13 May 2020 11:23:08 +0000 (07:23 -0400)]
travis: disable PIE in tests
Since all current prebuilt binutils libs are built w/out PIC/PIE,
disable PIE usage in tests. Newer toolchains are defaulting this
to on which breaks us.
Max Filippov [Fri, 8 May 2020 04:11:43 +0000 (21:11 -0700)]
elf2flt.c: add new relocation types for xtensa
Xtensa have added new relocation types R_XTENSA_[NP]DIFF{8,16,32} with
the same properties as the existing types R_XTENSA_DIFF{8,16,32}.
Add them to the list of ignored relocation types.
This fixes the following error when invoking elf2flt on xtensa binaries
built with the recent binutils:
ERROR: reloc type R_XTENSA_PDIFF32 unsupported in this context
Thomas Petazzoni [Fri, 27 Dec 2019 15:40:42 +0000 (16:40 +0100)]
elf2flt.c: add support for SOURCE_DATE_EPOCH
The bFLT header has a "build_date" field which contains the date/time
at which the bFLT file was created. Unfortunately, this breaks
reproducible builds as two identical builds done at different times
will produce different results.
For example, on a bFLT binary, diffoscope reports the following
change:
In order to address this, this commit adds support for the
SOURCE_DATE_EPOCH environment variable, which is standardized by the
reproducible-builds.org group at
https://reproducible-builds.org/specs/source-date-epoch/.
We simply use the time from the SOURCE_DATE_EPOCH variable (which
contains the number of seconds since Epoch) when SOURCE_DATE_EPOCH is
available in the environment, and otherwise fallback to the existing
logic that takes the current time using time().
Thomas Petazzoni [Sun, 13 Aug 2017 14:03:20 +0000 (16:03 +0200)]
ld-elf2flt: behave properly when called with a name different from TARGET_ALIAS
ld-elf2flt currently handles two cases:
1 It is called as the wrapper for <TARGET_ALIAS>-ld, generally
installed in the bin/ directory of a toolchain.
2 It is called as the wrapper for "ld", generally installed in the
TARGET_ALIAS/bin/ directory of a toolchain.
Unfortunately, if for some reason it gets called using a FOOBAR-ld
name that is different from <TARGET_ALIAS>-ld, it assumes it is in
case (2), while it really is in case (1). Due to this, the path
mangling logic doesn't work, and it doesn't find ld.real.
This happens for example when the binary program in bin/ is named
arm-buildroot-uclinux-uclibcgnueabi-ld, but also has a simpler symlink
named arm-linux-ld. In this case,
arm-buildroot-uclinux-uclibcgnueabi-ld is recognized by ld-elf2flt as
containing TARGET_ALIAS, and therefore the proper logic to find
ld.real is applied. However, when arm-linux-ld is used, ld-elf2flt
doesn't find TARGET_ALIAS, and therefore assumes we're being called as
TARGET_ALIAS/bin/ld.. and searches for a program called ld.real in
bin/, which doesn't exist.
See:
$ ./output/host/bin/arm-buildroot-uclinux-uclibcgnueabi-ld
/home/thomas/buildroot/buildroot/output/host/bin/../arm-buildroot-uclinux-uclibcgnueabi/bin/ld.real: no input files
$ ./output/host/bin/arm-linux-ld
arm-linux-ld (ld-elf2flt): error trying to exec '/home/thomas/buildroot/buildroot/output/host/bin/ld.real': execvp: No such file or directory
$ ./output/host/arm-buildroot-uclinux-uclibcgnueabi/bin/ld
/home/thomas/buildroot/buildroot/output/host/arm-buildroot-uclinux-uclibcgnueabi/bin/ld.real: no input files
This commit fixes that by inverting the logic: if we're being called
as just "ld", then we assume it's the program in
TARGET_ALIAS/bin/. Otherwise, we're called through some variant of
TARGET-ld.
Greg Ungerer [Wed, 30 Oct 2019 06:08:19 +0000 (16:08 +1000)]
elf2flt: fix relocations for read-only data
Readonly data sections are mapped into the "text" section in the
elf2flt.ld linker script. The relocation generation code is not handling
that case properly though, and is actually mapping any data section type
into the "data" section of the target binary.
This problem case has been detected with elf2flt core dumping when used
with binutils-2.33.1 (on ARM architecture targets). See thread at:
Mike Frysinger [Sun, 17 Feb 2019 21:14:33 +0000 (16:14 -0500)]
elf2flt: drop v850 reloc ifdefs
The binutils elf/v850.h header has used the elf/reloc-macros.h helpers
to create R_V850_xxx enums. They haven't been defines for a long time.
Trying to use #ifdef checks on them doesn't work.
This means we require binutils-2.24+ now which was released in 2013.
We can see if anyone notices if we need to support older versions.
Alexey Neyman [Mon, 27 Feb 2017 09:20:10 +0000 (01:20 -0800)]
Symlink required binutils/BFD headers to local dir
... to minimize the chance of clashes with system headers.
Also, remove the -lcygwin from Makefile.in: this breaks canadian
build on cygwin, as it tries to pass -lcygwin into non-Cygwin host
CC. This chunk pre-dates the addition of -lc into configure.ac and
passing -lc should be sufficient (it works for me).
Mike Frysinger [Wed, 1 Mar 2017 06:56:29 +0000 (23:56 -0700)]
elf2flt: clean up elf headers dependencies
The only reason we included cygwin-elf.h & elf.h was for the various
target reloc defines. However, since these are all provided by the
bfd library which we already include, we can switch to using that for
every target.
Now we don't have any dep on the host's ELF headers (either existing,
or being up-to-date).
Mike Frysinger [Mon, 12 Dec 2016 05:35:04 +0000 (00:35 -0500)]
elf2flt: fix unused warning for e1/bfin targets
The one call site for this func is inside of an e1/bfin ifdef check,
so add the same logic to the definition to fix a build error:
elf2flt.c:212:1: error: ‘get_symbol_offset’ defined but not used [-Werror=unused-function]
Mike Frysinger [Tue, 13 Sep 2016 06:37:14 +0000 (02:37 -0400)]
Merge pull request #3 from KirillSmirnov/uninitialized-var
My gcc is 5.4.0, target is m68k, binutils 2.26.
elf2flt.c: In function 'output_relocs':
elf2flt.c:1604:5: error: 'sym_reloc_size' may be used uninitialized in this function [-Werror=maybe-uninitialized]
printf(" RELOC[%d]: offset=0x%"BFD_VMA_FMT"x symbol=%s%s "
Greg Ungerer [Fri, 19 Aug 2016 13:49:51 +0000 (23:49 +1000)]
elf2flt: fix relocation support for R_ARM_TARGET types
R_ARM_TARGET1 (and I think R_ARM_TARGET2) relocation types should be
treated in the same way as R_ARM_ABS32. Fix them to write out the addend
to the flat binary in network byte order.
In output_relocs, we do "sprintf(&addstr[0], "+0x%lx", ...)", with addrstr
being a 16 bytes array.
On 64bits hosts, in the unlikely case the value overflows 32bits, the buffer
may overflow.
Indeed, the maximum theorical size is 20 bytes (16 bytes for the value + 3
bytes for "+0x" + the end of string marker).
The reason the value overflows 32bits is yet to be understood, as the ARMV7-M
is 32bits architecture, but this patch first ensure the sprintf call is robust
enough.
The _stack_start symbol needs to be in the same flatmem memory region
as text/data/bss, otherwise it will not end up with the correct address.
Direct the section into the flatmem region.
The GNU linker uses -v as a shortcut to --version, not --verbose. So atm,
if you run `ld -v` to get the linker version, ld-elf2flt throws out a lot
of verbose debugging information. So drop the -v checking in ld-elf2flt
to keep from breaking systems that parse the linker version.
David McCullough [Thu, 16 Dec 2010 01:37:41 +0000 (01:37 +0000)]
The .note.ABI-tag section exists to indicate to other projects (like gdb
or library loaders) information about the target OS. It doesn't actually
contain anything that is used at runtime. So while the current linker
script gathers this into the .data section, the final FLAT doesn't include
anything from it. But tools expect to find a dedicated section in ELFs
which the current section merge prevents.^M
So give .note.ABI-tag its own output section so gdb can locate and use it.
This shouldn't change the FLAT files produced in any way.
David McCullough [Tue, 17 Aug 2010 04:25:26 +0000 (04:25 +0000)]
When we converted ld-elf2flt from the shell script to C, one small nuance
was missed: argv[0] contains the full path only when invoked with the full
path. This is not the same behavior for shell scripts as $0 is always the
full path to the script in question. Most of the time this isn't an issue
as gcc will invoke all of its tools (like the linker) with a full relative
path to itself. However, if we attempt to invoke the linker directly, we
can see misbehavior such as:
bfin-uclinux-ld.real: cannot open linker script file ./../lib/elf2flt.ld:
No such file or directory
So, to fix this, we lean on more libiberty functions. Specifically, the
make_relative_prefix() function. This function locates a full argv[0] by
scanning $PATH to see where it was invoked. This might sound a little
dodgy, but this is fundamental to how gcc and binutils implement support
for their runtime relocation, so it can't break ld-elf2flt without first
breaking every one else ;).
In the fall out of this fix, we can cull a bunch of local code that does
custom path parsing. So not only do we get to fix an annoying bug, we get
to shrink code in the process.
David McCullough [Tue, 22 Jun 2010 06:12:47 +0000 (06:12 +0000)]
The current code misses checking a few args in order to determine the
default "print" mode (ktrace/l1stack/...). Rather than update a list
that people easily forget, rework the code to generically detect that
no arguments have been specified.
The bflt loader in the kernel will, however, add a small extra data table
just before .data's content (cf. handling of MAX_SHARED_LIBS in
binfmt_flat.c:load_flat_file).
Now, if .text and .data are in the same segment, directly following each
other in the binary file, but have that extra data table added in the
run-time memory layout, GDB will get very confused when trying to access
items in the now-moved .data section. Without any kernel (loader) / GDB
changes, the solution is to tell the linker to always put .text and .data
into separate segments, which GDB will handle gracefully then.
Poor getopt() implementations as found in many BSD/Darwin systems will
stop processing options after a non-option is encountered. That means
ld-elf2flt has to be careful to not stick options after non-options when
executing sub children. In a default setup, it will invoke `elf2flt` with
the output followed by the -a option which subsequently fails:
elf2flt: Can't open '-a': No such file or directory
David McCullough [Tue, 14 Jul 2009 23:00:33 +0000 (23:00 +0000)]
the "all" target should not be depending on "ld-elf2flt"
anymore as this is handled through the PROG vars. it isnt a problem
for Linux systems, but when EXEEXT is set, things go boom.
David McCullough [Sun, 12 Jul 2009 23:28:58 +0000 (23:28 +0000)]
Due to shell portability issues (which is to say shell scripts are not
portable -- think Windows), convert elf2flt to C code.
I've updated this code base to the latest elf2flt tree and actually done
some basic tests -- building the three Blackfin tuples (ELF, FLAT, FDPIC)
and running programs on my Blackfin boards. This process found errors in
the original implementation as well as some of the cleanups I did.
Unify the duplicated windows and other system fallback logic in stubs.h
and add some fatal() helper functions to standardize the error output when
falling over. This way we don't end up with obscure error messages with
no idea what util they are coming from.
This cleans up the Makefile handling of the different compiler flags such
that it uses standard names across the board as well as unifies the link
method.
Rather than putting the `rm` at the end of the script before the normal
exit point, create a trap to automatically delete the script when exiting.
This way the linker script gets cleaned up whenever there is an error as
well. Otherwise every link invocation that ends in a failure could leave
behind crap. On my system, i found almost 2 million of these suckers in
my /tmp dir.
David McCullough [Sun, 24 May 2009 23:33:48 +0000 (23:33 +0000)]
When the relocs are larger than 16bits, incorrect values are written when
the .H/.L loading are reversed. Normally this wouldn't happen because the
gcc compiler always outputs in the same order (first hi, then lo).