Archive element IO is performed on the file of the containing archive,
which leads to the BFD "where" field of archives and their elements
being out of sync with the real file position. (We're talking
traditional archives here, not thin archives.) The old bfd_seek code
recognized this by not attempting to optimize away seeks for
archives. However, there was other code that could return bogus
results. For example, cache.c limits the number of open files by
closing a file and remembering its state once the limit is reached.
If bfd_tell is called on an archive element when the containing
archive is closed, it will return an invalid file pointer.
It's possible to have a valid "where" field for archives by always
using and updating the containing archive BFD. That's what this patch
does. Note that cache.c used to find the containing archive BFD
anyway for the iostream, so we're not really doing extra work, just
transferring it up to the correct abstraction level.
The patch also gets rid of some hacks. bfd_tell was called when
bfd_seek failed, in an attempt to correct "where". That's got to be
papering over another problem, so that code has been removed.
bfd_read also had an "optimiziation" to return early when the number
of bytes was zero, and bfd_seek optimized calls that didn't move the
file pointer. This was covering for a coff_slurp_line_table bug where
IO was attempted on a pe-dll BFD without an iovec.
* bfd.c (struct bfd): Update comment on "where" usage.
* bfdio.c (bfd_bwrite, bfd_stat): Use and update "iovec",
"iostream", and "where" from containing archive file. Return
error on NULL iovec.
(bfd_bread): Similarly, and return error attempted out of
bounds archive element access.
(bfd_tell, bfd_flush): Use and update "iovec", "iostream", and
"where" from containing archive file.
(bfd_seek): Likewise. Return error on NULL iovec. Don't
attempt to optimize away seeks. Don't paper over errors by
calling bfd_tell.
(bfd_get_mtime): Call bfd_stat rather than iovec->bstat.
(bfd_get_size): Likewise.
(bfd_mmap): Operate on and use iovec of containing archive
file. Return error on NULL iovec.
* cache.c (bfd_cache_lookup_worker): Abort if working on
archive element bfd.
(cache_bread_1): Delete bfd parameter, add FILE* parameter.
Don't ignore zero byte reads.
(cache_bread): Look up FILE* in cache here. Error on NULL
lookup.
(cache_bwrite): Rename "where" to "from".
(cache_bmmap): Don't handle archive elements.
* coffcode.h (coff_slurp_line_table): Exit early on zero
lineno count.
* bfd-in2.h: Regenerate.
Give these bfd-internal symbols with external linkage a _bfd_ prefix
to avoid collisions in the global symbol namespace.
bfd/ChangeLog:
2017-02-17 Pedro Alves <palves@redhat.com>
* bfdio.c (real_ftell): Rename to ...
(_bfd_real_ftell): ... this.
(real_fseek): Rename to ...
(_bfd_real_fseek): ... this.
(real_fopen): Rename to ...
(_bfd_real_fopen): ... this.
* libbfd-in.h (real_ftell): Rename to ...
(_bfd_real_ftell): ... this.
(real_fseek): Rename to ...
(_bfd_real_fseek): ... this.
(real_fopen): Rename to ...
(_bfd_real_fopen): ... this.
* cache.c, dwarf2.c, opncls.c: Adjust all callers.
* libbfd.h: Regenerate.
LTO plugin support in plugin_maybe_claim wants to close the IR bfd
after replacing it with the recompiled object, but can't do so for
archive elements due to various pointers that access the archive bfd.
Thin archives have the same problem. They too cannot have their
element bfds closed.
PR ld/20241
bfd/
* archive.c (open_nested_file): Set my_archive.
* bfd.c (_bfd_default_error_handler <%B>): Exclude archive file name
for thin archives.
* bfdio.c (bfd_tell): Don't adjust origin for thin archives.
(bfd_seek): Likewise.
* bfdwin.c (bfd_get_file_window): Likewise.
* cache.c (cache_bmmap): Likewise.
(bfd_cache_lookup_worker): Don't look in my_archive for thin archives.
* mach-o.c (bfd_mach_o_follow_dsym): Don't open my_archive for
thin archives.
* plugin.c (try_claim): Likewise.
* xcofflink.c (xcoff_link_add_dynamic_symbols): Use import path of
file within thin archive, not the archive.
binutils/
* bucomm.c (bfd_get_archive_filename): Return file name within thin
archive.
ld/
* ldmain.c (add_archive_element): Just print file name of file within
thin archives.
* ldmisc.c (vfinfo): Likewise.
* plugin.c (plugin_object_p): Open file within thin archives.
(plugin_maybe_claim): Expand comment.
There is no need to check and initialize pagesize_m1 in cache_bmmap.
* cache.c (cache_bmmap): Move pagesize_m1 ... to
(pagesize_m1): Here.
(bfd_cache_init): Initialize pagesize_m1.
The current hard coded limit of open files in bfd/cache.c is 10. This
is pretty low these days. Binaries are often linked against much more
than 10 files (and sometimes against more than 100 shared libraries).
When debugging with GDB some files are opened and closed multiple
times because of this low limit. If possible make the BFD cache file
limit depend on the actual open file limit of the process so more BFD
files can be open at the same time.
* cache.c (BFD_CACHE_MAX_OPEN): Remove define.
(max_open_files): New static int initialized to zero.
(bfd_cache_max_open): New static function to set and return
max_open_files.
(bfd_cache_init): Use bfd_cache_max_open.
(bfd_open_file): Likewise.
* configure.in (AC_CHECK_HEADERS): Add sys/resource.h.
(AC_CHECK_FUNCS): Add getrlimit.
* configure: Regenerated.
* config.in: Likewise.
* sysdep.h: Check and include sys/resource.h for getrlimit.
* corefile.c (reopen_exec_file): Use exec_bfd_mtime.
* exec.c (exec_bfd_mtime): Define.
(exec_close): Clear it.
(exec_file_attach): Set it.
* gdbcore.h (exec_bfd_mtime): Declare.
* source.c (find_source_lines): Do not use bfd_get_mtime.
* cache.c (enum cache_flag): New.
(close_one): Save file mtime.
(bfd_cache_lookup): Add flag arg, adjust all users.
(bfd_cache_lookup_worker): Likewise.
(cache_btell): Use CACHE_NO_OPEN and return abfd->where if file
not open.
(cache_bflush): Similarly, and return success of file not open.
(cache_bseek): Use CACHE_NO_SEEK if SEEK_SET or SEEK_END.
(cache_bstat): Use CACHE_NO_SEEK_ERROR.
