elf2flt: fix relocations for read-only .eh_frame section

[RomainNaour]

The commit [1] moved readonly data sections to "text" section. This was needed to fix a elf2flt segfault on ARM architectures with Binutils >= 2.33.1.

After this patch was applied to Buildroot's elf2flt package, new segfault appear while building several packages (acpica, augeas, binutils, cairo, fontconfig, gptfdisk, libopenssl, mimic...) [2].

We can reproduce the issue manually from the binutils build directory:

'output/host/m68k-buildroot-uclinux-uclibc/bin/elf2flt' '-a' '-o' 'readelf' '-r' 'readelf.gdb'

While debuging, we can notice the flags value (0x12f) for .eh_frame just before the crash.

RELOCS: .eh_frame [0x2185c20]: flags=0x12f vma=0x84384

/* bug case: flags = 0x12f (m68k)
 * SEC_HAS_CONTENTS 0x100
 * SEC_DATA         0x020
 * SEC_READONLY     0x008
 * SEC_RELOC        0x004
 * SEC_LOAD         0x002
 * SEC_ALLOC        0x001
 */

On ARM cortex-m4, we have the same flags: RELOCS: .ARM.exidx [0x9ac5b0]: flags=0x12f vma=0x4b4ec

So due to the new condition introduced by [1] the .eh_frame section located in a readonly data section will be moved to the "text" section.

Looking at the gcc code for m68k [3]:

"Because .eh_frame refers to both code and data, it follows that .eh_frame must be in the data segment itself. [...] In theory, we could create a read-only .eh_frame [...]. However, gcc currently handles indirect references using a per-TU constant pool. This means that if a function and its eh_frame are removed by the linker, the eh_frame's indirect references to the removed function will not be removed, leading to an unresolved symbol error."

Fix this crash by checking the section name and move .eh_frame section even if it is located in readonly data.

[1] 73325b7f209e0f68887333385184af275531427d [2] http://lists.busybox.net/pipermail/buildroot/2020-February/274593.html [3] https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=gcc/config/m68k/m68k.h;h=fc65e524b139a6d43e528956a788b9110aebaf2e;hb=a0c06cc27d2146b7d86758ffa236516c6143d62c#l785 [4] https://github.com/uclinux-dev/elf2flt/issues/12

Signed-off-by: Romain Naour [email protected]

[tpetazzoni]

I am not an expert if elf2flt, but I believe the above change will fix things for eh_frame, but the issue can happen for other sections as well. With commit 73325b7f209e0f68887333385184af275531427d applied, elf2flt now assumes that if the ELF section being processed is read-only, then this section will be "mapped" into the text segment of the bFLT binary. However, the ld-elf2flt linker script doesn't do this for eh_frame and eh_frame_hdr: they are mapped into the data segment.

So, I think a better change is to move eh_frame and eh_frame_hdr from the data segment to the text segment. But we would need to do that for all ELF sections that are marked read-only.

Or alternatively, we change 73325b7f209e0f68887333385184af275531427d and instead use an explicit list of ELF sections that the ld-elf2flt linker script places into the text segment.

@vapier @gregungerer any suggestion ?

[gregungerer]

Elf2flt pull request #27 (https://github.com/uclinux-dev/elf2flt/pull/27) would mean this change is no longer required. It reverts the 2 previous changes that led to this being a problem and instead fixes the ARM.exidx problem with hopefully no new problems created.

[vapier]

should be resolved in latest version with Greg's fixes