Re: [PATCH v2 12/14] i386: Remove lazy tlsdesc relocation related code

["H.J. Lu" <hjl.tools@gmail.com>]

On Tue, Apr 13, 2021 at 2:34 AM Szabolcs Nagy via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> Like in commit e75711ebfa976d5468ec292282566a18b07e4d67 for x86_64,
> remove unused lazy tlsdesc relocation processing code:
>
>   _dl_tlsdesc_resolve_abs_plus_addend
>   _dl_tlsdesc_resolve_rel
>   _dl_tlsdesc_resolve_rela
>   _dl_tlsdesc_resolve_hold
> ---
>  sysdeps/i386/dl-tlsdesc.S | 156 -------------------------
>  sysdeps/i386/dl-tlsdesc.h |   6 +-
>  sysdeps/i386/tlsdesc.c    | 231 +-------------------------------------
>  3 files changed, 2 insertions(+), 391 deletions(-)
>
> diff --git a/sysdeps/i386/dl-tlsdesc.S b/sysdeps/i386/dl-tlsdesc.S
> index e781d973b7..255fe88651 100644
> --- a/sysdeps/i386/dl-tlsdesc.S
> +++ b/sysdeps/i386/dl-tlsdesc.S
> @@ -134,159 +134,3 @@ _dl_tlsdesc_dynamic:
>         cfi_endproc
>         .size   _dl_tlsdesc_dynamic, .-_dl_tlsdesc_dynamic
>  #endif /* SHARED */
> -
> -     /* This function is a wrapper for a lazy resolver for TLS_DESC
> -       REL relocations that reference the *ABS* segment in their own
> -       link maps.  %ebx points to the caller's GOT.  %eax points to a
> -       TLS descriptor, such that 0(%eax) holds the address of the
> -       resolver wrapper itself (unless some other thread beat us to
> -       it) and 4(%eax) holds the addend in the relocation.
> -
> -       When the actual resolver returns, it will have adjusted the
> -       TLS descriptor such that we can tail-call it for it to return
> -       the TP offset of the symbol.  */
> -
> -       .hidden _dl_tlsdesc_resolve_abs_plus_addend
> -       .global _dl_tlsdesc_resolve_abs_plus_addend
> -       .type   _dl_tlsdesc_resolve_abs_plus_addend,@function
> -       cfi_startproc
> -       .align 16
> -_dl_tlsdesc_resolve_abs_plus_addend:
> -0:
> -       _CET_ENDBR
> -       pushl   %eax
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %ecx
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %edx
> -       cfi_adjust_cfa_offset (4)
> -       movl    $1f - 0b, %ecx
> -       movl    4(%ebx), %edx
> -       call    _dl_tlsdesc_resolve_abs_plus_addend_fixup
> -1:
> -       popl    %edx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %ecx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %eax
> -       cfi_adjust_cfa_offset (-4)
> -       jmp     *(%eax)
> -       cfi_endproc
> -       .size   _dl_tlsdesc_resolve_abs_plus_addend, .-_dl_tlsdesc_resolve_abs_plus_addend
> -
> -     /* This function is a wrapper for a lazy resolver for TLS_DESC
> -       REL relocations that had zero addends.  %ebx points to the
> -       caller's GOT.  %eax points to a TLS descriptor, such that
> -       0(%eax) holds the address of the resolver wrapper itself
> -       (unless some other thread beat us to it) and 4(%eax) holds a
> -       pointer to the relocation.
> -
> -       When the actual resolver returns, it will have adjusted the
> -       TLS descriptor such that we can tail-call it for it to return
> -       the TP offset of the symbol.  */
> -
> -       .hidden _dl_tlsdesc_resolve_rel
> -       .global _dl_tlsdesc_resolve_rel
> -       .type   _dl_tlsdesc_resolve_rel,@function
> -       cfi_startproc
> -       .align 16
> -_dl_tlsdesc_resolve_rel:
> -0:
> -       _CET_ENDBR
> -       pushl   %eax
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %ecx
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %edx
> -       cfi_adjust_cfa_offset (4)
> -       movl    $1f - 0b, %ecx
> -       movl    4(%ebx), %edx
> -       call    _dl_tlsdesc_resolve_rel_fixup
> -1:
> -       popl    %edx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %ecx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %eax
> -       cfi_adjust_cfa_offset (-4)
> -       jmp     *(%eax)
> -       cfi_endproc
> -       .size   _dl_tlsdesc_resolve_rel, .-_dl_tlsdesc_resolve_rel
> -
> -     /* This function is a wrapper for a lazy resolver for TLS_DESC
> -       RELA relocations.  %ebx points to the caller's GOT.  %eax
> -       points to a TLS descriptor, such that 0(%eax) holds the
> -       address of the resolver wrapper itself (unless some other
> -       thread beat us to it) and 4(%eax) holds a pointer to the
> -       relocation.
> -
> -       When the actual resolver returns, it will have adjusted the
> -       TLS descriptor such that we can tail-call it for it to return
> -       the TP offset of the symbol.  */
> -
> -       .hidden _dl_tlsdesc_resolve_rela
> -       .global _dl_tlsdesc_resolve_rela
> -       .type   _dl_tlsdesc_resolve_rela,@function
> -       cfi_startproc
> -       .align 16
> -_dl_tlsdesc_resolve_rela:
> -0:
> -       _CET_ENDBR
> -       pushl   %eax
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %ecx
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %edx
> -       cfi_adjust_cfa_offset (4)
> -       movl    $1f - 0b, %ecx
> -       movl    4(%ebx), %edx
> -       call    _dl_tlsdesc_resolve_rela_fixup
> -1:
> -       popl    %edx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %ecx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %eax
> -       cfi_adjust_cfa_offset (-4)
> -       jmp     *(%eax)
> -       cfi_endproc
> -       .size   _dl_tlsdesc_resolve_rela, .-_dl_tlsdesc_resolve_rela
> -
> -     /* This function is a placeholder for lazy resolving of TLS
> -       relocations.  Once some thread starts resolving a TLS
> -       relocation, it sets up the TLS descriptor to use this
> -       resolver, such that other threads that would attempt to
> -       resolve it concurrently may skip the call to the original lazy
> -       resolver and go straight to a condition wait.
> -
> -       When the actual resolver returns, it will have adjusted the
> -       TLS descriptor such that we can tail-call it for it to return
> -       the TP offset of the symbol.  */
> -
> -       .hidden _dl_tlsdesc_resolve_hold
> -       .global _dl_tlsdesc_resolve_hold
> -       .type   _dl_tlsdesc_resolve_hold,@function
> -       cfi_startproc
> -       .align 16
> -_dl_tlsdesc_resolve_hold:
> -0:
> -       _CET_ENDBR
> -       pushl   %eax
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %ecx
> -       cfi_adjust_cfa_offset (4)
> -       pushl   %edx
> -       cfi_adjust_cfa_offset (4)
> -       movl    $1f - 0b, %ecx
> -       movl    4(%ebx), %edx
> -       call    _dl_tlsdesc_resolve_hold_fixup
> -1:
> -       popl    %edx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %ecx
> -       cfi_adjust_cfa_offset (-4)
> -       popl    %eax
> -       cfi_adjust_cfa_offset (-4)
> -       jmp     *(%eax)
> -       cfi_endproc
> -       .size   _dl_tlsdesc_resolve_hold, .-_dl_tlsdesc_resolve_hold
> diff --git a/sysdeps/i386/dl-tlsdesc.h b/sysdeps/i386/dl-tlsdesc.h
> index 753c03e79c..12e90da3a8 100644
> --- a/sysdeps/i386/dl-tlsdesc.h
> +++ b/sysdeps/i386/dl-tlsdesc.h
> @@ -43,11 +43,7 @@ struct tlsdesc_dynamic_arg
>
>  extern ptrdiff_t attribute_hidden __attribute__ ((regparm (1)))
>    _dl_tlsdesc_return (struct tlsdesc *),
> -  _dl_tlsdesc_undefweak (struct tlsdesc *),
> -  _dl_tlsdesc_resolve_abs_plus_addend (struct tlsdesc *),
> -  _dl_tlsdesc_resolve_rel (struct tlsdesc *),
> -  _dl_tlsdesc_resolve_rela (struct tlsdesc *),
> -  _dl_tlsdesc_resolve_hold (struct tlsdesc *);
> +  _dl_tlsdesc_undefweak (struct tlsdesc *);
>
>  # ifdef SHARED
>  extern void *_dl_make_tlsdesc_dynamic (struct link_map *map,
> diff --git a/sysdeps/i386/tlsdesc.c b/sysdeps/i386/tlsdesc.c
> index 0bc646541f..436a21f66b 100644
> --- a/sysdeps/i386/tlsdesc.c
> +++ b/sysdeps/i386/tlsdesc.c
> @@ -16,242 +16,13 @@
>     License along with the GNU C Library; if not, see
>     <https://www.gnu.org/licenses/>.  */
>
> -#include <link.h>
>  #include <ldsodefs.h>
> -#include <elf/dynamic-link.h>
>  #include <tls.h>
>  #include <dl-tlsdesc.h>
>  #include <dl-unmap-segments.h>
> +#define _dl_tlsdesc_resolve_hold 0
>  #include <tlsdeschtab.h>
>
> -/* The following 4 functions take an entry_check_offset argument.
> -   It's computed by the caller as an offset between its entry point
> -   and the call site, such that by adding the built-in return address
> -   that is implicitly passed to the function with this offset, we can
> -   easily obtain the caller's entry point to compare with the entry
> -   point given in the TLS descriptor.  If it's changed, we want to
> -   return immediately.  */
> -
> -/* This function is used to lazily resolve TLS_DESC REL relocations
> -   that reference the *ABS* segment in their own link maps.  The
> -   argument is the addend originally stored there.  */
> -
> -void
> -__attribute__ ((regparm (3))) attribute_hidden
> -_dl_tlsdesc_resolve_abs_plus_addend_fixup (struct tlsdesc volatile *td,
> -                                          struct link_map *l,
> -                                          ptrdiff_t entry_check_offset)
> -{
> -  ptrdiff_t addend = (ptrdiff_t) td->arg;
> -
> -  if (_dl_tlsdesc_resolve_early_return_p (td, __builtin_return_address (0)
> -                                         - entry_check_offset))
> -    return;
> -
> -#ifndef SHARED
> -  CHECK_STATIC_TLS (l, l);
> -#else
> -  if (!TRY_STATIC_TLS (l, l))
> -    {
> -      td->arg = _dl_make_tlsdesc_dynamic (l, addend);
> -      td->entry = _dl_tlsdesc_dynamic;
> -    }
> -  else
> -#endif
> -    {
> -      td->arg = (void*) (addend - l->l_tls_offset);
> -      td->entry = _dl_tlsdesc_return;
> -    }
> -
> -  _dl_tlsdesc_wake_up_held_fixups ();
> -}
> -
> -/* This function is used to lazily resolve TLS_DESC REL relocations
> -   that originally had zero addends.  The argument location, that
> -   originally held the addend, is used to hold a pointer to the
> -   relocation, but it has to be restored before we call the function
> -   that applies relocations.  */
> -
> -void
> -__attribute__ ((regparm (3))) attribute_hidden
> -_dl_tlsdesc_resolve_rel_fixup (struct tlsdesc volatile *td,
> -                              struct link_map *l,
> -                              ptrdiff_t entry_check_offset)
> -{
> -  const ElfW(Rel) *reloc = td->arg;
> -
> -  if (_dl_tlsdesc_resolve_early_return_p (td, __builtin_return_address (0)
> -                                         - entry_check_offset))
> -    return;
> -
> -  /* The code below was borrowed from _dl_fixup(),
> -     except for checking for STB_LOCAL.  */
> -  const ElfW(Sym) *const symtab
> -    = (const void *) D_PTR (l, l_info[DT_SYMTAB]);
> -  const char *strtab = (const void *) D_PTR (l, l_info[DT_STRTAB]);
> -  const ElfW(Sym) *sym = &symtab[ELFW(R_SYM) (reloc->r_info)];
> -  lookup_t result;
> -
> -   /* Look up the target symbol.  If the normal lookup rules are not
> -      used don't look in the global scope.  */
> -  if (ELFW(ST_BIND) (sym->st_info) != STB_LOCAL
> -      && __builtin_expect (ELFW(ST_VISIBILITY) (sym->st_other), 0) == 0)
> -    {
> -      const struct r_found_version *version = NULL;
> -
> -      if (l->l_info[VERSYMIDX (DT_VERSYM)] != NULL)
> -       {
> -         const ElfW(Half) *vernum =
> -           (const void *) D_PTR (l, l_info[VERSYMIDX (DT_VERSYM)]);
> -         ElfW(Half) ndx = vernum[ELFW(R_SYM) (reloc->r_info)] & 0x7fff;
> -         version = &l->l_versions[ndx];
> -         if (version->hash == 0)
> -           version = NULL;
> -       }
> -
> -      result = _dl_lookup_symbol_x (strtab + sym->st_name, l, &sym,
> -                                   l->l_scope, version, ELF_RTYPE_CLASS_PLT,
> -                                   DL_LOOKUP_ADD_DEPENDENCY, NULL);
> -    }
> -  else
> -    {
> -      /* We already found the symbol.  The module (and therefore its load
> -        address) is also known.  */
> -      result = l;
> -    }
> -
> -  if (!sym)
> -    {
> -      td->arg = 0;
> -      td->entry = _dl_tlsdesc_undefweak;
> -    }
> -  else
> -    {
> -#  ifndef SHARED
> -      CHECK_STATIC_TLS (l, result);
> -#  else
> -      if (!TRY_STATIC_TLS (l, result))
> -       {
> -         td->arg = _dl_make_tlsdesc_dynamic (result, sym->st_value);
> -         td->entry = _dl_tlsdesc_dynamic;
> -       }
> -      else
> -#  endif
> -       {
> -         td->arg = (void*)(sym->st_value - result->l_tls_offset);
> -         td->entry = _dl_tlsdesc_return;
> -       }
> -    }
> -
> -  _dl_tlsdesc_wake_up_held_fixups ();
> -}
> -
> -/* This function is used to lazily resolve TLS_DESC RELA relocations.
> -   The argument location is used to hold a pointer to the relocation.  */
> -
> -void
> -__attribute__ ((regparm (3))) attribute_hidden
> -_dl_tlsdesc_resolve_rela_fixup (struct tlsdesc volatile *td,
> -                               struct link_map *l,
> -                               ptrdiff_t entry_check_offset)
> -{
> -  const ElfW(Rela) *reloc = td->arg;
> -
> -  if (_dl_tlsdesc_resolve_early_return_p (td, __builtin_return_address (0)
> -                                         - entry_check_offset))
> -    return;
> -
> -  /* The code below was borrowed from _dl_fixup(),
> -     except for checking for STB_LOCAL.  */
> -  const ElfW(Sym) *const symtab
> -    = (const void *) D_PTR (l, l_info[DT_SYMTAB]);
> -  const char *strtab = (const void *) D_PTR (l, l_info[DT_STRTAB]);
> -  const ElfW(Sym) *sym = &symtab[ELFW(R_SYM) (reloc->r_info)];
> -  lookup_t result;
> -
> -   /* Look up the target symbol.  If the normal lookup rules are not
> -      used don't look in the global scope.  */
> -  if (ELFW(ST_BIND) (sym->st_info) != STB_LOCAL
> -      && __builtin_expect (ELFW(ST_VISIBILITY) (sym->st_other), 0) == 0)
> -    {
> -      const struct r_found_version *version = NULL;
> -
> -      if (l->l_info[VERSYMIDX (DT_VERSYM)] != NULL)
> -       {
> -         const ElfW(Half) *vernum =
> -           (const void *) D_PTR (l, l_info[VERSYMIDX (DT_VERSYM)]);
> -         ElfW(Half) ndx = vernum[ELFW(R_SYM) (reloc->r_info)] & 0x7fff;
> -         version = &l->l_versions[ndx];
> -         if (version->hash == 0)
> -           version = NULL;
> -       }
> -
> -      result = _dl_lookup_symbol_x (strtab + sym->st_name, l, &sym,
> -                                   l->l_scope, version, ELF_RTYPE_CLASS_PLT,
> -                                   DL_LOOKUP_ADD_DEPENDENCY, NULL);
> -    }
> -  else
> -    {
> -      /* We already found the symbol.  The module (and therefore its load
> -        address) is also known.  */
> -      result = l;
> -    }
> -
> -  if (!sym)
> -    {
> -      td->arg = (void*) reloc->r_addend;
> -      td->entry = _dl_tlsdesc_undefweak;
> -    }
> -  else
> -    {
> -#  ifndef SHARED
> -      CHECK_STATIC_TLS (l, result);
> -#  else
> -      if (!TRY_STATIC_TLS (l, result))
> -       {
> -         td->arg = _dl_make_tlsdesc_dynamic (result, sym->st_value
> -                                             + reloc->r_addend);
> -         td->entry = _dl_tlsdesc_dynamic;
> -       }
> -      else
> -#  endif
> -       {
> -         td->arg = (void*) (sym->st_value - result->l_tls_offset
> -                            + reloc->r_addend);
> -         td->entry = _dl_tlsdesc_return;
> -       }
> -    }
> -
> -  _dl_tlsdesc_wake_up_held_fixups ();
> -}
> -
> -/* This function is used to avoid busy waiting for other threads to
> -   complete the lazy relocation.  Once another thread wins the race to
> -   relocate a TLS descriptor, it sets the descriptor up such that this
> -   function is called to wait until the resolver releases the
> -   lock.  */
> -
> -void
> -__attribute__ ((regparm (3))) attribute_hidden
> -_dl_tlsdesc_resolve_hold_fixup (struct tlsdesc volatile *td,
> -                               struct link_map *l __attribute__((__unused__)),
> -                               ptrdiff_t entry_check_offset)
> -{
> -  /* Maybe we're lucky and can return early.  */
> -  if (__builtin_return_address (0) - entry_check_offset != td->entry)
> -    return;
> -
> -  /* Locking here will stop execution until the running resolver runs
> -     _dl_tlsdesc_wake_up_held_fixups(), releasing the lock.
> -
> -     FIXME: We'd be better off waiting on a condition variable, such
> -     that we didn't have to hold the lock throughout the relocation
> -     processing.  */
> -  __rtld_lock_lock_recursive (GL(dl_load_lock));
> -  __rtld_lock_unlock_recursive (GL(dl_load_lock));
> -}
> -
> -
>  /* Unmap the dynamic object, but also release its TLS descriptor table
>     if there is one.  */
>
> --
> 2.17.1
>

LGTM.

Thanks.

-- 
H.J.