[gold][patch] Fix file descriptor leak

[Cary Coutant <ccoutant at google dot com>]

PR 10708 is about gold running out of file descriptors. I've found
several places where gold can re-open a file descriptor for reading
without holding a lock on the file; when that happens, the file
descriptor never gets marked as available for reuse. This can cause fd
exhaustion with --no-keep-files-mapped and a low limit on open file
descriptors (ulimit -n).

This patch fixes most of the file descriptor leakages by holding a
Task_lock_obj on the object while we're reading from it. There are
three more cases in the ARM backend -- two are easily fixable, but
Doug wants to restructure the third so that reading the file is
unnecessary, so I'll leave the arm.cc patches to him.

I originally added an assert in fileread.cc,
File_read::find_or_make_view(), to give an internal error if we ever
try to get a view on a file while the file is unlocked (that's how I
ended up finding most of these), but I'm leaving that out at least
until we fix the remaining known problems in arm.cc.

-cary

        PR gold/10708
        * copy-relocs.cc (Copy_relocs::emit_copy_reloc): Hold a lock on the
        object when reading from the file.
        * gold.cc (queue_middle_tasks): Hold a lock on the object when doing
        second layout pass.
        * icf.cc (preprocess_for_unique_sections): Hold a lock on the object
        when reading section contents.
        (get_section_contents): Likewise.
        (icf::find_identical_sections): Likewise.
        * mapfile.cc (Mapfile::print_discarded_sections): Hold a lock on the
        object when reading from the file.
        * plugin.cc (Plugin_manager::layout_deferred_objects): Hold a lock on
        the object when doing deferred section layout.

Attachment: gold-fd-leak-patch-1.txt
Description: Text document