ld-specific flags. LDFLAGS is really for ld-related flags for cc,
not for ld, and some flags, e.g., -Bshareable, mean completely
different things to cc and ld. Having the wrong things in LDFLAGS
also broke the standard ${PROG} target. This was kludged around
by using a special rule that depended on LDFLAGS being bogus.
Fixing `make depend' broke the special rule but fixed the standard
rule (except in the DESTDIR case, which was handled more strictly
here than elsewhere).
dependency on `bar' is very unlikely to be correct.
This is a quick fix for broken dependencies in gdb and many other
places. The dependencies on internal libraries are now missing
instead of wrong when `make depend' is run before the libraries
are created.
library with a shared object dependency that contained alias symbols,
the linker incorrectly counted the number of symbols that would be
written, resulting in a fatal internal error. Since our libc now
contains some alias symbols (in "net/res_stubs.c"), this was
sufficient to tickle the bug: "ld -Bshareable foo.so -lc". To
fix it, I moved the accounting of alias symbols to a later point
in the processing, where it is possible to count only those symbols
that will actually be written to the output file.
This fix is well-confined to affect alias symbols only. I have
tested it with a full "make world". I am going to merge it into
-2.2 after a few more days of living with it in -current.
If it is set to a nonempty string, then simply skip any missing
shared libraries. This came up in a discussion long ago as a
potentially useful feature at sysinstall time. For example, an
X11 utility could be used without the X libraries being present,
provided the utility had a mode in which no X functions were actually
called.
This will make a number of things easier in the future, as well as (finally!)
avoiding the Id-smashing problem which has plagued developers for so long.
Boy, I'm glad we're not using sup anymore. This update would have been
insane otherwise.
nonempty string, then function calls are relocated at program start-up
rather than lazily. This variable is standard on Sun and SVR4 systems.
The dlopen() function now supports both lazy and immediate binding, as
determined by its "mode" argument, which can be either 1 (RTLD_LAZY) or
2 (RTLD_NOW). I will add defines of these symbols to <dlfcn.h> as soon
as I've done a little more checking to make sure they won't cause
collisions or bootstrapping problems that would break "make world".
The "LD_*" environment variables which alter dynamic linker behavior are
now treated as unset if they are set to the empty string. This agrees
with the standard SVR4 conventions for the dynamic linker.
Add a work-around for programs compiled with certain buggy versions of
crt0.o. The buggy versions failed to set the "crt_ldso" member of the
interface structure. This caused certain error messages from the
dynamic linker to begin with "(null)" instead of the pathname of the
dynamic linker.
nonempty string, then function calls are relocated at program start-up
rather than lazily. This variable is standard on Sun and SVR4 systems.
The dlopen() function now supports both lazy and immediate binding, as
determined by its "mode" argument, which can be either 1 (RTLD_LAZY) or
2 (RTLD_NOW). I will add defines of these symbols to <dlfcn.h> as soon
as I've done a little more checking to make sure they won't cause
collisions or bootstrapping problems that would break "make world".
that is stored in the hints file. If that search path contained
a non-existent directory (one, say, that had been removed), and
"ldconfig -m /a/perfectly/good/directory" was run, ldconfig returned
an error status without printing an error message. This caused
some confusing bombs when installing ports, in particular.
I changed it so that non-existent directories from the stored search
path are silently ignored. Only non-existent directories named
explicitly on the command line are treated as errors. Also, a
diagnostic is printed if and only if an error status is returned.
In an unrelated fix, ldconfig now silently ignores any directories
named on the command line when the "-r" option is given. Formerly,
these directories incorrectly made their way into the "search
directories" line of the listing. It really should be an error to
specify directories together with "-r", but I don't have time to
fix the manual page in that way right now.
2.2 Candidate.
the main program, report them directly from the dynamic linker and die
there, rather than returning an error message to crt0.o. This enables
the printing of error messages even for old executables, whose version
of crt0.o is not able to print them.
This fix closes PR bin/1869.
The code in crt0.o for printing error messages from the dynamic linker
is no longer used, because of this change. But it must remain, for
backward compatibility with older dynamic linkers.
opened. After that, the directories are already present, and there is
no point in adding them again. This doesn't fix any bugs; it's just for
efficiency.
since rt_readenv() already takes care of not setting unsafe variables.
This was part of the changes I submitted to Peter and John during the
review which must have gotten missed.
how I managed to get this out of sync, but I did. I guess that's what I
get for directly committing from different machines that I was testing on.
Pointed out by: Paul Traina <pst@freebsd.org>
configurable fallback search paths, as well as new crt interface version.
Also:
- even faster getenv(), get all environment variable settings in a single
pass.
- ldd printf-like format specifications
- minor code cleanups, one vsprintf -> vsnprintf (harmless)
The library search sequence is a little more complete now. Before,
it'd search $LD_LIBRARY_PATH (by opendir/readdir/closedir), then read
the hints file, then read /usr/lib (again by scanning thr directory). It
would then fail if there was no "found" library.
Now, it does LD_LIBRARY_PATH and the hints file the same, but then uses
a longer fallback path. The -R path is fetched from the executable if
specified at build time, the ldconfig path is appended, and /usr/lib is
appended to that. Duplicates are suppressed. This means that simply
placing a new library in /usr/local/lib will work (the same as it did in
/usr/lib) without needing ldconfig -m. It will find it quicker if the
ldconfig is run though.
Similar changes have been made to the NetBSD ld.so, but ours is rather
different now due to John Polstra's speedups and fixes from a while back.
The ldd printf-like format support came direct from NetBSD.
Reviewed by: nate, jdp
ldconfig path (from NetBSD). I added code to make sure there were no
duplicates in the path when multiple ldconfig -m's were used.
Reviewed by: nate, jdp
Obtained from: NetBSD (partly)
with the -R option and store the path in the dynamic header when specified.
The $LD_RUN_PATH environment variable is not checked yet.
While here, split up the code a bit more to enable more selective replacing
of GPL'ed components that are linked with ld.so with others.
Obtained from: NetBSD (mostly, the breakup is my fault)
as atomically as possible.
(Immutable targets can't be renamed without opening a window when
neither the source nor the target is immutable. Perhaps there
should be a rename_immutable syscall to do this if unsetting the
immutable flags would work.)
need this.
Consider the following code:
case 'O':
output_filename = malloc(strlen(arg)+4);
strcpy(output_filename, arg);
strcat(output_filename, ".tmp");
real_output_filename = arg;
return;
The idea here is to malloc() a buffer big enough to hold the name of
a supplied file name, plus ".tmp". So we malloc() 'size of filename'
bytes plus 4, right? Wrong! ".tmp" is _FIVE_ bytes long! There's a
traling '\0' which strcat() gleefully tacks on _outside_ the bounds
of the buffer. Result: program corrupts own memory. Program SEGVs at
seemingly random times. Bill not like random SEGVs. Bill smash.
Know how I found this? I've been trying to bootstrap -current on my
2.1.0-RELEASE machine at work and I couldn't seem to get libc.a built
because the linker would intermittently blow chunks while executing
things like 'ld -O foo.o -X -r foo.o'. Since this is an initial
bootstrap version of ld, it was linked against the 2.1.0 libc, who's
malloc() behaves differently than that in -current.
Presumeably ld -O doesn't blow up in -current, otherwise someone would
have spotted this already. I don't know if this is a bug or a feature.
Anyway. I'm changing the strlen(arg)+4 to strlen(arg)+5. Bah.
-O filename
Specifies the name of the output file. The file is created as
filename.tmp and when output is complete renamed to filename.
This allows us to:
ld -O ${.TARGET} -x -r ${.TARGET}
file. The field formerly contained random garbage, leading to spurious
differences between otherwise identical executables and libraries.
Submitted by: Bruce Evans <bde@freebsd.org>
been loaded, look for a match by device and inode number if the
traditional pathname comparisons don't find a match. This detects
the case in which a library is requested using two different names
which are really links to the same file, and avoids loading it
twice.
Requested by: peter@freebsd.org
Reviewed by: peter@freebsd.org
"-X" must survive, because of references from relocations, don't
qualify the symbol name with the name of the input file. This
saves some string space. It makes libc_pic.a about 2.4% smaller.
Adapted from a suggestion by Bruce Evans.
descriptions of LD_NO_INTERN_SEARCH and LD_NOSTD_PATH from the manual
page, since they are not supported.
Submitted by: Doug Ambrisko <ambrisko@ambrisko.roble.com>
man pages up to mdoc guidelines and fix some minor formatting glitches.
Also fixed a number of man pages to not abuse the .Xr macro to
display functions and path names and a lot of other junk.
This solves the problem of being unable to use shared libraries with dots
in their names before the ".so.<version>" code.
This should be brought into -stable.
There are more changes from Paul that look like they should be included,
but they change the format of the hints file, so I'm not going to bring them
in now (but we should in the future).
Obtained from: pk@netbsd.org
It addresses a number of problems that were present in earlier
versions.
The calls to the "init" and "fini" functions of shared libraries
have been reordered, so that they are called in a strictly nested
fashion, as is required for C++ constructors and destructors. In
addition, the "init" functions are called in better order relative
to each other. That makes the system more tolerant of C++ programs
which depend on a library's being initialized before its clients.
The dynamic linker is now more tolerant of shared libraries in
which dependencies on other shared libraries are incompletely
recorded.
Cleanup in the event of errors has been improved throughout the
dynamic linker. A number of memory leaks were eliminated.
The warning message for a shared library whose minor version number
is too old has been clarified.
The code dealing with the "ld.so.hints" file has been cleaned up.
A bug that caused the hints file to be unmapped incompletely has
been fixed. A different bug that could potentially cause the hints
file to be mapped on top of a loaded object has been fixed.
The code that searches for shared libraries has been cleaned up.
The searching is now more compatible with that done by SunOS and
SVR4. Also, some unnecessary and useless searches of both the
hints file and library directories have been eliminated.
Reviewed by: nate@freebsd.org
vector. Now it is called the "symbol caching" vector. This was made
possible and unconfusing by other changes that allowed me to localize
everything having to do with the caching vector in the function
reloc_map().
Switched to alloca() for allocating the caching vector, and eliminated
the special mmap-based allocation routines. Although this was motivated
by performance reasons, it led to significant simplification of the
code, and made it possible to confine the symbol caching code to the
single function reloc_map().
Got rid of the unnecessary and inefficient division loop at the
beginning of rtld().
Reduced the number of calls to getenv("LD_LIBRARY_PATH") to just 1, on
suggestion from <davidg@root.com>.
Added breaks out of the relocation loops when the relocation address is
found to be 0. A relocation address of 0 is caused by an unused
relocation entry. Unused relocation entries are caused by linking a
shared object with the "-Bsymbolic" switch. The runtime linker itself
is linked that way, and the last 40% of its relocation entries are
unused. Thus, breaking out of the loop on the first such entry is a
performance win when ld.so relocates itself. As a side benefit, it
permits removing a test from md_relocate_simple() in
../i386/md-static-funcs.c.
Unused relocation entries in other shared objects (linked with
"-Bsymbolic") caused even bigger problems in previous versions of the
runtime linker. The runtime linker interpreted the unused entries as if
they were valid. That caused it to perform repeated relocations of the
first byte of the shared object. In order to do that, it had to remap
the text segment writable. Breaking out of the loop on the first unused
relocation entry solves that.
Submitted by: John Polstra <jdp@polstra.com>
Implemented symbol memorizing to reduce the number of calls to lookup(),
making relocation go faster. While relocating a given shared object,
the dynamic linker maintains a memorizing vector that is directly
indexed by the symbol number in the relocation entry. The first time a
given symbol is looked up, the memorizing vector is filled in with a
pointer to the symbol table entry, and a pointer to the so_map of the
shared object in which the symbol was defined. On subsequent uses of
the same symbol, that information is retrieved directly from the
memorizing vector, without calling lookup() again.
A symbol that is referenced in a relocation entry is typically
referenced in many relocation entries, so this memorizing reduces the
number of calls to lookup() dramatically. The overall improvement in
the speed of dynamic linking is also dramatic -- as much as a factor of
three for programs that use many shared libaries.
Submitted by: jdp@polstra.com "John Polstra"
shared library. Formerly, the message looked like this:
ld.so: run: libjdp1.so.1.0: Undefined error: 0
The new message looks like this:
ld.so: run: Can't find shared library "libjdp1.so.1.0"
(Where "run" is the name of the program being executed.)
Submitted by: jdp@polstra.com (John Polstra)
symbols.
An easy example to see this is to develop an X program which links
against Xt, but doesn't add -lX11 to the link line. It will link fine,
but cause run-time errors by ld.so because of missing symbols used by Xt
defined in X11. This patch makes the errors more readable.
Submitted by: jdp@polstra.com (John Polstra)
on dlclose. Also correctly call constructors and destructors for libraries
linked with /usr/lib/c++rt0.o.
Change interpretation of dlopen manpage to call _init() rather than init()
for dlopened objects.
Change c++rt0.o to avoid using atexit to call destructors, allowing dlclose to
call destructors when an object is unloaded.
Change interface between crt0 and ld.so to allow crt0 to call a function on
exit to call destructors for shared libraries explicitly.
These changes are backwards compatible. Old binaries will work with the new
ld.so and new binaries will work with the old ld.so. A version number has
been introduced in the crt0-ld.so interface to allow for future changes.
Reviewed by: GAWollman, Craig Struble <cstruble@singularity.bevc.blacksburg.va.us>
now safely add a line like
ldconfig -m ${PREFIX}/lib
in ports' Makefiles and packing lists without throwing away some
directories the user may have added.
Submitted by: Mostly by Paul Kranenburg <pk@cs.few.eur.nl>
underlying bugs which are caused by mixing static/shared libraries with
this change in place.
The shlib code is not capable of supporting this feature in it's present
state and will need significant modifications in order to do so.
members over shared library members. This modification causes the linker
to use the first definition it sees for a symbol instead of having
priorities based on the library type. This modification should allow
gdb to compile again.
Obtained from:
Email conversation with Paul Kranenbury, but implemented completely by
me. If it doesn't work, it's my fault not his.
first place and we were too long in finding out. Now we know, and the
damage is hard to fix. This is part one: ld will not link gcc dynamic,
if specified as "-lgcc".
Suggested by: dyson & davidg
NetBSD ld code except for local changes for dlopen() and friends and
the hashing on the minor value of the shlibs. We should be binary
compatible now with all their libraries.
Obtained from: NetBSD
misfeature caused troubles when a program attempted to access a shlib
where one with a higher minor number has been hashed. Ldconfig does
only include the highest-numbered shlib anyway, so this is in no way a
limitation of generality.
Caution: after installing the new programs, your /var/run/ld.so.hints
needs to be rebuiult; run ldconfig again as it's done from /etc/rc.
warning handling and allows for link-time warnings with a modified
version of gas.
Note: Not all of the newer bits were updated such as some of the non-x86
machine-dependant code is relevant to FreeBSD right now.
Obtained from: NetBSD
completely when ldconfig unlinks it. If init is shared, then the
referenced unlinked copy of the hints file created by running
ldconfig in /etc/rc caused the file system to be unclean after
every reboot.
automagically. -lfoo has to be right to work, but ${LIBFO0} is too
easy to forget or misspell; nothing checks it and it should be
different for shared libraries.
handling of errors through the standard err() and warn()
more fixes for Geoff Rehmet's NULL pointer bug.
fixes NULL pointer bugs when linking mono and nested X servers.
supports a `-nostdlib' option.
accept object files without a symbol table
don't attempt dynamic linking when `-A' is given
a few variable names have chaged (desc -> fd), and the formatting has
changed which should make it much easier to track his sources.
I tested 'make world' for /usr/src and X twice with these changes.
Paul Kranenburg's description:
ld is in error here, assuming that symbols with N_EXT set always have an
entry in the (global) symbol table: this is not the case for C++ generated
constructor/destructor symbols. I can reproduce your failure by fudging
a "multiply defined" constructor symbol by hand. Checking for `g == NULL'
seems to be a ok as a fence for now.
So:
for now, in do_file_warnings() we check if g == NULL, before trying to generate
any warning messages. This prevents a NULL pointer dereference.
to public. These functions are also used in /usr/include/link.h,
so it looks, like they shouldn't be private.
I will ask Paul about that, if this is correct.
