[PATCH v2] SHT_GROUP handling: Use _bfd_elf_link_m[un]map_section_contents

[H.J. Lu]

On Fri, Jun 28, 2024, 8:44 PM Alan Modra <amodra@gmail.com> wrote:

> On Thu, Jun 27, 2024 at 10:27:59PM -0700, H.J. Lu wrote:
> > Use _bfd_elf_link_m[un]map_section_content to mmap in SHT_GROUP section.
> > When mmap is used, data to link the 3.0GB clang executable in LLVM 19
> > debug build on Linux/x86-64 with 32GB RAM is:
>
> I don't believe any of the SHT_GROUP sections will use mmap.
> alan@squeak:~/src/llvm-project/build$ find . -name \*.o | xargs readelf
> -SW | egrep 'GROUP +[0-9a-f]+ [0-9a-f]+ 0{0,3}[^0]'
>   [200] .group            GROUP           0000000000000000 0008e4 0001ec
> 04     4986 3413  4
>   [236] .group            GROUP           0000000000000000 000a94 000274
> 04     5502 2821  4
>
> So my clang debug build only has one object with SHT_GROUP bigger
> than 512 bytes, far smaller than the 16k mmap threshold.  I'll only
> believe this is worth doing if you can show a group in real code that
> is large enough to actually use mmap.
>
> Likely you're using different compilers to me.  I just used the system
> compiler and linker on my Ubuntu-24.04 amd box.
> g++ (Ubuntu 13.2.0-23ubuntu4) 13.2.0
> GNU ld (GNU Binutils for Ubuntu) 2.42
>
> Here's the two biggest file in bin/
> -rwxrwxr-x 1 alan alan 4162474976 Jun 28 20:42 clang-19
> -rwxrwxr-x 1 alan alan 4256936064 Jun 28 20:55 clang-repl
>
> Please try with something like
>   if (ghdr->bfd_section->mmapped_p)
>     abort ();
> to prove the point.
>

It isn't about the SHT_GROUP section size.  Since kernel will allocate at
least
1 page for its I/O even if its size is 4 bytes, mmap is more efficient.  It
improves performance for parallel link.


> >                 malloc        mmap            improvement
> > user            42.15         42.18           -0.07%
> > system                  7.32          7.41            -1.2%
> > total           50.22         49.88           0.68%
> > context switches  4933                799             83.8%
> > major page faults 487         26              94.7%
> >
> >       * elf.c (process_sht_group_entries): Replace
> >       bfd_malloc_and_get_section and free with
> >       _bfd_elf_mmap_section_contents and
> >       _bfd_elf_munmap_section_contents.
> >
> > Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
> > ---
> >  bfd/elf.c | 8 +++++---
> >  1 file changed, 5 insertions(+), 3 deletions(-)
> >
> > diff --git a/bfd/elf.c b/bfd/elf.c
> > index 8bb296f9637..cefb92d2595 100644
> > --- a/bfd/elf.c
> > +++ b/bfd/elf.c
> > @@ -607,7 +607,8 @@ process_sht_group_entries (bfd *abfd,
> >    unsigned char *contents;
> >
> >    /* Read the raw contents.  */
> > -  if (!bfd_malloc_and_get_section (abfd, ghdr->bfd_section, &contents))
> > +  if (!_bfd_elf_mmap_section_contents (abfd, ghdr->bfd_section,
> > +                                    &contents))
> >      {
> >        _bfd_error_handler
> >       /* xgettext:c-format */
> > @@ -663,7 +664,8 @@ process_sht_group_entries (bfd *abfd,
> >                 gname = group_signature (abfd, ghdr);
> >                 if (gname == NULL)
> >                   {
> > -                   free (contents);
> > +                   _bfd_elf_munmap_section_contents
> > +                     (ghdr->bfd_section, contents);
> >                     return false;
> >                   }
> >               }
> > @@ -693,7 +695,7 @@ process_sht_group_entries (bfd *abfd,
> >           }
> >       }
> >      }
> > -  free (contents);
> > +  _bfd_elf_munmap_section_contents (ghdr->bfd_section, contents);
> >    return true;
> >  }
> >
> > --
> > 2.45.2
>
> --
> Alan Modra
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/binutils/attachments/20240629/496e5e29/attachment.htm>