[PATCH v12 0/6] elf: Use mmap to map in section contents
H.J. Lu
hjl.tools@gmail.com
Thu Apr 4 13:53:43 GMT 2024
On Thu, Apr 4, 2024 at 6:12 AM Luis Machado <luis.machado@arm.com> wrote:
>
> Hi,
>
> On 3/17/24 12:19, H.J. Lu wrote:
> > Changes in v12:
> >
> > 1. Don't call _bfd_elf_link_keep_memory for x86 since it always returns
> > false.
> > 2. Make _bfd_elf_link_keep_memory static.
> >
> > Changes in v11:
> >
> > 1. Update _bfd_mmap_read_temporary to allocate buffer as needed.
> > 2. Don't allocate buffer when _bfd_mmap_read_temporary is called.
> >
> > Changes in v10:
> >
> > 1. Malloc a 4 * page size buffer for external relocations for the final
> > link to avoid mmap/munmap on small relocation sections.
> > 2. Malloc a buffer for input section contents which are smaller than
> > 4 * page size or can't be mapped for the final link.
> >
> > Changes in v9:
> >
> > 1. Use MAP_FAILED for mmap failure.
> >
> > Changes in v8:
> >
> > 1. Rebase against master branch.
> > 2. Add _bfd_elf_link_mmap_section_contents and
> > _bfd_elf_link_munmap_section_contents.
> >
> > Changes in v7:
> >
> > 1. Don't add the --keep-memory linker option.
> >
> > Changes in v6:
> >
> > 1. Add the --keep-memory linker option and always cache symbol and
> > relocation tables for --keep-memory.
> > 2. Always keep symbol table and relocation info for eh_frame to speedup
> > the Rust binary build by ~300x:
> >
> > https://sourceware.org/bugzilla/show_bug.cgi?id=31466
> >
> > Changes in v5:
> >
> > 1. Drop 2 patches which have been merged onto master branch.
> > 2. Rename _bfd_elf_mmap_section to _bfd_elf_mmap_section_contents.
> > 3. Rename _bfd_mmap_readonly_tracked, _bfd_mmap_readonly_untracked,
> > _bfd_munmap_readonly_untracked, _bfd_mmap_read_untracked to
> > _bfd_mmap_readonly_persistent, _bfd_mmap_readonly_temporary,
> > _bfd_munmap_readonly_temporary and _bfd_mmap_read_temporary.
> > 4. Drop the setup_group change.
> > 5. Fix a typo.
> > 6. Update comments.
> >
> > Changes in v4:
> >
> > 1. Change don't cache symbol nor relocation tables with mmap to opt-in.
> >
> > Changes in v3:
> >
> > 1. Fix non-mmap build.
> > 2. Change the argument name of bfd_mmap_local from flags to prot since
> > its values are PROT_XXX.
> >
> > Changes in v2:
> >
> > 1. Don't hard-code BFD_JUMP_TABLE_COPY in bfd so that elf-bfd.h can be
> > included in libbfd.c.
> > 2. Change the --with-mmap default to true.
> > 3. Check USE_MMAP instead of HAVE_MMAP.
> > 4. Remove the asize parameter to _bfd_mmap_readonly_tracked.
> > 5. Add contents_addr and contents_size to bfd_elf_section_data.
> > 6. Rename _bfd_link_keep_memory to _bfd_elf_link_keep_memory.
> >
> > ---
> > We can use mmap to map in ELF section contents, instead of copying them
> > into memory by hand. We don't need to cache symbol nor relocation tables
> > if they are mapped in. Data to link the 3.5GB clang executable in LLVM
> > 17 debug build on Linux/x86-64 with 32GB RAM is:
> >
> > stdio mmap improvement
> > user 86.73 87.02 -0.3%
> > system 9.55 9.21 3.6%
> > total 100.40 97.66 0.7%
> > maximum set(GB) 17.34 13.14 24%
> > page faults 4047667 3042877 25%
> >
> > and data to link the 275M cc1plus executable in GCC 14 stage 1 build is:
> >
> > user 5.41 5.44 -0.5%
> > system 0.80 0.76 5%
> > total 6.25 6.26 -0.2%
> > maximum set(MB) 1323 968 27%
> > page faults 323451 236371 27%
> >
> > Data shows that these won't improve the single copy linker performance.
> > But they improve the overall system performance when linker is used by
> > reducing linker memory usage and page faults. They allow more parallel
> > linker jobs on LLVM debug build.
> >
> > Here is a quote from Noah Goldstein: "on a large project they are an
> > extremely large speedup".
> >
> > H.J. Lu (6):
> > elf: Use mmap to map in read-only sections
> > elf: Add _bfd_elf_m[un]map_section_contents
> > elf: Use mmap to map in symbol and relocation tables
> > elf: Don't cache symbol nor relocation tables with mmap
> > elf: Always keep symbol table and relocation info for eh_frame
> > elf: Add _bfd_elf_link_m[un]map_section_contents
> >
> > bfd/bfd-in2.h | 24 ++++-
> > bfd/bfd.c | 17 +++
> > bfd/bfdwin.c | 8 +-
> > bfd/cache.c | 7 +-
> > bfd/compress.c | 2 +-
> > bfd/elf-bfd.h | 24 +++++
> > bfd/elf-eh-frame.c | 4 +-
> > bfd/elf-sframe.c | 4 +-
> > bfd/elf.c | 246 ++++++++++++++++++++++++++++++++++--------
> > bfd/elf32-i386.c | 8 +-
> > bfd/elf64-x86-64.c | 12 +--
> > bfd/elfcode.h | 7 +-
> > bfd/elflink.c | 213 ++++++++++++++++++++++++------------
> > bfd/elfxx-target.h | 6 +-
> > bfd/elfxx-x86.c | 8 +-
> > bfd/elfxx-x86.h | 1 +
> > bfd/libbfd-in.h | 33 +++++-
> > bfd/libbfd.c | 262 ++++++++++++++++++++++++++++++++++++++++++++-
> > bfd/libbfd.h | 33 +++++-
> > bfd/linker.c | 35 ------
> > bfd/lynx-core.c | 2 +-
> > bfd/opncls.c | 21 ++++
> > bfd/section.c | 9 +-
> > bfd/sysdep.h | 4 +
> > 24 files changed, 795 insertions(+), 195 deletions(-)
> >
>
> Probably known, judging from the e-mail reworking _bfd_mmap_read_temporary, but I thought I'd mention anyway.
>
> Looks like this might be causing gdb crashes of the following nature:
>
> Starting program: /home/builder/worker/gdb-fedora-arm64/gdb-build/gdb/testsuite/outputs/gdb.base/break-idempotent/break-idempotent-pie
> BFD: BFD (GNU Binutils) 2.42.50.20240404 internal error, aborting at ../../binutils-gdb/bfd/libbfd.c:1211 in _bfd_mmap_read_temporary
> Please report this bug.
The patch is at
https://sourceware.org/pipermail/binutils/2024-April/133374.html
--
H.J.
More information about the Binutils
mailing list