We ran into a problem with how DTrace processes relocations to determine the location of probe points in an ELF object. USDT probe points are just calls to (undefined) functions of a specific name, and when we process the ELF object we look for relocations to such symbols. We patch the text for the function at the call site with NOPs, and generate meta-data for this probe in a separate ELF object. We then change the type of the reloc to be R_X86_64_NONE, so that it will be ignored when the ELF object is linked into its target (so the linker won't overwrite our carefully placed NOPs). The reason why we don't just remove the reloc is that sometimes builds end up re-processing the ELF object and we want it to be able to regenerate the meta-data for the probes in that case. This does not work on x86 with PIC/PIE because the relocation validation in 382aae063227 ("x86: Only allow S + A relocations against absolute symbol") is not allowing R_X86_64_NONE relocations even though they have no functional purpose, i.e. they should be ignored. I realize that the use of the relocation type R_X86_64_NONE is perhaps a bit unusual but it does support the need here in a quite elegant manner. It would seem that allowing R_X86_64_NONE (and R_386_NONE) relocations in bfd_elf_x86_valid_reloc_p() would be the correct action?
Is it possible to create a testcase? A binary object is OK.
(In reply to H.J. Lu from comment #1) > Is it possible to create a testcase? A binary object is OK. If R_386_NONE is also impacted, please also provide such a testcase.
[hjl@gnu-cfl-3 pr31047]$ cat x.s .text .globl _start _start: .nop .reloc 0, BFD_RELOC_NONE, foo .section .note.GNU-stack,"",@progbits [hjl@gnu-cfl-3 pr31047]$ cat y.s .globl foo foo = 0x1000 .section .note.GNU-stack,"",@progbits [hjl@gnu-cfl-3 pr31047]$ make as -o y.o y.s as -o x.o x.s ./ld -pie -o x y.o x.o ./ld: x.o: relocation R_X86_64_NONE against absolute symbol `foo' in section `.text' is disallowed make: *** [Makefile:13: x] Error 1
The master branch has been updated by H.J. Lu <hjl@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=2e93abb858ae4ff2b8605b4a94988068869c0ff1 commit 2e93abb858ae4ff2b8605b4a94988068869c0ff1 Author: H.J. Lu <hjl.tools@gmail.com> Date: Tue Jan 9 06:30:28 2024 -0800 x86: Don't check R_386_NONE nor R_X86_64_NONE Update x86 ELF linker to skip R_386_NONE/R_X86_64_NONE when scanning relocations. bfd/ * PR ld/31047 * elf32-i386.c (elf_i386_scan_relocs): Don't check R_386_NONE. * elf64-x86-64.c (elf_x86_64_scan_relocs): Don't check R_X86_64_NONE. ld/ * PR ld/31047 * testsuite/ld-i386/i386.exp: Run PR ld/31047 test. * testsuite/ld-x86-64/x86-64.exp: Likewise. * testsuite/ld-i386/pr31047.d: New file. * testsuite/ld-x86-64/pr31047-x32.d: Likewise. * testsuite/ld-x86-64/pr31047.d: Likewise. * testsuite/ld-x86-64/pr31047a.s: Likewise. * testsuite/ld-x86-64/pr31047b.s: Likewise.
Fixed for 2.42.