| Message ID | 20140603175042.GH15355@redacted.bos.redhat.com |
|---|---|
| State | New |
| Headers | show |
Hi, Kyle, On Jun 3, 2014, Kyle McMartin <kmcmarti@redhat.com> wrote: > AArch64 presently defaults to use TLS descriptors, as opposed to > traditional dynamic TLS relocations, and has hit a bug in the copy-pasta > used in elf_machine_rel{,a} for all TLS descriptor implementations. When > a program calls dlopen() with R_AARCH64_TLSDESC relocs, we're consuming > static TLS until it's exhausted (presumably as an optimisation.) > However, this means that once our static TLS space is consumed, > attempting to dlopen() any shared object which contain an > R_AARCH64_TPREL reloc will fail, since we've exhausted all the space. > Instead, always use dynamic TLS for TLS descriptors, which lets us load > arbitrarily many proper dynamic TLS using shared objects, and allow the > static TLS fallback only in the non-SHARED case. This approach severely worsens performance of TLS Descriptors. If what you're running into is the DTV limit for Static TLS, rather than Static TLS space proper, please consider using in its stead the final version (still under review) of the patch I posted this week, with Subject: “[BZ#17090/17620/17621]: fix DTV race, assert, and DTV_SURPLUS Static TLS limit”. It should allow TPREL relocs regardless of how many TLS modules you've loaded before, static or dynamic, except in the unlikely case of actually filling up the Static TLS segment. Please let me know if you have any problems with it.
--- a/sysdeps/aarch64/dl-machine.h +++ b/sysdeps/aarch64/dl-machine.h @@ -295,7 +295,7 @@ elf_machine_rela (struct link_map *map, const ElfW(Rela) *reloc, # ifndef SHARED CHECK_STATIC_TLS (map, sym_map); # else - if (!TRY_STATIC_TLS (map, sym_map)) + if (1) { td->arg = _dl_make_tlsdesc_dynamic (sym_map, sym->st_value + reloc->r_addend); diff --git a/sysdeps/arm/dl-machine.h b/sysdeps/arm/dl-machine.h index 899b256..f55a991 100644 --- a/sysdeps/arm/dl-machine.h +++ b/sysdeps/arm/dl-machine.h @@ -458,7 +458,7 @@ elf_machine_rel (struct link_map *map, const Elf32_Rel *reloc, # ifndef SHARED CHECK_STATIC_TLS (map, sym_map); # else - if (!TRY_STATIC_TLS (map, sym_map)) + if (1) { td->argument.pointer = _dl_make_tlsdesc_dynamic (sym_map, value); diff --git a/sysdeps/i386/dl-machine.h b/sysdeps/i386/dl-machine.h index 368bee2..b6b5802 100644 --- a/sysdeps/i386/dl-machine.h +++ b/sysdeps/i386/dl-machine.h @@ -394,7 +394,7 @@ elf_machine_rel (struct link_map *map, const Elf32_Rel *reloc, # ifndef SHARED CHECK_STATIC_TLS (map, sym_map); # else - if (!TRY_STATIC_TLS (map, sym_map)) + if (1) { td->arg = _dl_make_tlsdesc_dynamic (sym_map, sym->st_value + (ElfW(Word))td->arg); diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h index 8df04a9..4ec4340 100644 --- a/sysdeps/x86_64/dl-machine.h +++ b/sysdeps/x86_64/dl-machine.h @@ -359,7 +359,7 @@ elf_machine_rela (struct link_map *map, const ElfW(Rela) *reloc, # ifndef SHARED CHECK_STATIC_TLS (map, sym_map); # else - if (!TRY_STATIC_TLS (map, sym_map)) + if (1) { td->arg = _dl_make_tlsdesc_dynamic (sym_map, sym->st_value + reloc->r_addend);
AArch64 presently defaults to use TLS descriptors, as opposed to traditional dynamic TLS relocations, and has hit a bug in the copy-pasta used in elf_machine_rel{,a} for all TLS descriptor implementations. When a program calls dlopen() with R_AARCH64_TLSDESC relocs, we're consuming static TLS until it's exhausted (presumably as an optimisation.) However, this means that once our static TLS space is consumed, attempting to dlopen() any shared object which contain an R_AARCH64_TPREL reloc will fail, since we've exhausted all the space. Instead, always use dynamic TLS for TLS descriptors, which lets us load arbitrarily many proper dynamic TLS using shared objects, and allow the static TLS fallback only in the non-SHARED case. This issue is reproducible on x86_64, by specifying -mtls-dialect=gnu2 when building with gcc. I've written a test for this: https://github.com/jkkm/test-tls_desc.git Tweak DIALECT to be appropriate, and run ./test-tls_desc.sh, ./main and it should fail with: % ./test-tls_desc.sh; ./main ./tmp16.so: cannot allocate memory in static TLS block ./test-tls_desc.sh: line 18: 10141 Aborted (core dumped) ./main 2014-06-03 Kyle McMartin <kyle@redhat.com> * sysdeps/aarch64/dl-machine.h (elf_machine_rela): Always allocate dynamic TLS space for TLS descriptors. * sysdeps/x86_64/dl-machine.h (elf_machine_rela): Ditto. * sysdeps/arm/dl-machine.h (elf_machine_rel): Ditto. * sysdeps/i386/dl-machine.h (elf_machine_rel): Ditto.