[gold][patch] Control caching of memory mapped regions

[Cary Coutant <ccoutant at google dot com>]

When linking a huge binary, gold uses a lot of memory because it keeps
the input files mapped throughout the link. This can slow down the
link considerably and sometimes the linker runs out of memory. The
attached patch adds an option that can reduce gold's memory usage
considerably by letting it unmap the memory used by input files
between linker passes.

I had a bit of a dilemma with the option name. Ideally, I wanted to
use --[no-]keep-files-mapped, but that option already exists, to
control whether the linker maps files as a whole instead of piecemeal.
I chose to rename that option to --[no-]map-whole-files, which is
closer to its actual meaning, and adopt --keep-files-mapped for the
new control. The defaults are the same as before: --map-whole-files is
the default for 64-bit linkers, --no-map-whole-files is the default
for 32-bit linkers, and --keep-files-mapped is the default for either
case.

For mapping whole files, I changed it to create the initial whole file
mapping at the first request for any view, rather than as soon as the
file is opened. I also keep the whole file view on the list of views
for the file, making it eligible for removal.

I found a couple of gotchas in Add_symbols::run() and
Scan_relocs::run() -- gold was calling release() on the File object
before deleting the sd_ and rd_ objects, both of which contain several
File_view objects obtained from get_lasting_view(). With those
File_view objects still alive, release() was not actually clearing the
views.

Linking a 400MB binary, according to gold --stats output, with the
64-bit default --map-whole-files, the --no-keep-files-mapped option
reduced the total number of bytes mmaped at one time from 1011 MB to
only 32 MB; link time went down marginally from 19.6s to 19.3s (wall
time). According to ps, the memory high-water mark was reduced from
478,297 pages to 236,276 pages. Experiments with other combinations of
the two options showed that --no-keep-files-mapped
--no-map-whole-files increased link time slightly, while
--keep-files-mapped --no-map-whole-files increased link time quite a
bit (to 37s).

All tests pass on x86_64, debug and opt. OK to commit?

-cary

        * fileread.cc (File_read::~File_read): Don't delete whole_file_view_.
        (File_read::open[1]): Remove initial mapping of whole_file_view_.
        (File_read::open[2]): Add whole_file_view_ to list of views.
        (File_read::make_view): Remove test of whole_file_view_.
        (File_read::find_or_make_view): Create whole_file_view_ if
        necessary.
        (File_read::clear_views): Replace bool parameter with enum;
        adjust all callers.  Don't delete views with unowned data;
        do delete cached views and views from archives if
        --no-keep-files-mapped is set.  Set whole_file_view_ to NULL
        if clearing the corresponding view.
        * fileread.h (File_read::Clear_views_mode): New enum.
        (File_read::View::is_data_owned): New method.
        (File_read::clear_views): Replace bool parameter
        with enum; adjust all callers.
        * options.h (General_options): Change keep_files_mapped option;
        add map_whole_files.
        * readsyms.cc (Add_symbols::run): Delete sd_ object before
        releasing the file.
        * reloc.cc (Scan_relocs::run): Delete rd_ object before releasing
        the file.

Attachment: gold-no-keep-patch.txt
Description: Text document