NOTES.UNIX: expand the description of RPATHs

Hopefully, this will make it more clear that it isn't only ELF
specific, even though there is a part that is (or even more
restrictively GNU ld.so specific).

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Andy Polyakov <appro@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6520)
This commit is contained in:
Richard Levitte 2018-06-19 13:03:55 +02:00
parent 8d983739bb
commit 445bc80861

View file

@ -5,38 +5,87 @@
For Unix/POSIX runtime systems on Windows, please see NOTES.WIN.
Shared libraries and installation in non-standard locations
-----------------------------------------------------------
OpenSSL uses the compiler to link programs and shared libraries
---------------------------------------------------------------
Binaries on Unix variants expect to find shared libraries in standard
locations, such as /usr/lib, /usr/local/lib and some other locations
configured in the system (for example /etc/ld.so.conf on some systems).
If the libraries are installed in non-standard locations, binaries
will not find them and therefore fail to run unless they get a bit of
help from a defined RPATH or RUNPATH. This can be applied by adding
the appropriate linker flags to the configuration command, such as
this (/usr/local/ssl was the default location for OpenSSL installation
in versions before 1.1.0):
OpenSSL's generated Makefile uses the C compiler command line to
link programs, shared libraries and dynamically loadable shared
objects. Because of this, any linking option that's given to the
configuration scripts MUST be in a form that the compiler can accept.
This varies between systems, where some have compilers that accept
linker flags directly, while others take them in '-Wl,' form. You need
to read your compiler documentation to figure out what is acceptable,
and ld(1) to figure out what linker options are available.
Shared libraries and installation in non-default locations
----------------------------------------------------------
Every Unix system has its own set of default locations for shared
libraries, such as /lib, /usr/lib or possibly /usr/local/lib. If
libraries are installed in non-default locations, dynamically linked
binaries will not find them and therefore fail to run unless they get a
bit of help from a defined runtime shared library search path.
For OpenSSL's application (the 'openssl' command), our configuration
scripts do NOT generally set the runtime shared library search path for
you. It's therefore advisable to set it explicitly when configuring
unless the libraries are to be installed in directories that you know
to be in the default list.
Runtime shared library search paths are specified with different
linking options depending on operating system and versions thereof, and
are talked about differently in their respective documentation;
variations of RPATH are the most usual (note: ELF systems have two such
tags, more on that below).
Possible options to set the runtime shared library search path include
the following:
-Wl,-rpath,/whatever/path
-R /whatever/path
-rpath /whatever/path
OpenSSL's configuration scripts recognise all these options and pass
them to the Makefile that they build. (In fact, it recognises anything
starting with '-Wl,' as a linker option, so for example, HP-UX'
'-Wl,+b,/whatever/path' would be used correctly)
Please do not use verbatim directories in your runtime shared library
search path! Some OpenSSL config targets add an extra directory level
for multilib installations. To help with that, the produced Makefile
includes the variable LIBRPATH, which is a convenience variable to be
used with the runtime shared library search path options, as shown in
this example:
$ ./config --prefix=/usr/local/ssl --openssldir=/usr/local/ssl \
-Wl,-rpath,/usr/local/ssl/lib
'-Wl,-rpath,$(LIBRPATH)'
Because the actual library location may vary further (for example on
multilib installations), there is a convenience variable in Makefile
that holds the exact installation directory and that can be used like
this:
On modern ELF based systems, there are two runtime search paths tags to
consider, DT_RPATH and DT_RUNPATH. Shared objects are searched for in
this order:
1. Using directories specified in DT_RPATH, unless DT_RUNPATH is
also set.
2. Using the environment variable LD_LIBRARY_PATH
3. Using directories specified in DT_RUNPATH.
4. Using system shared object caches and default directories.
This means that the values in the environment variable LD_LIBRARY_PATH
won't matter if the library is found in the paths given by DT_RPATH
(and DT_RUNPATH isn't set).
Exactly which of DT_RPATH or DT_RUNPATH is set by default appears to
depend on the system. For example, according to documentation,
DT_RPATH appears to be deprecated on Solaris in favor of DT_RUNPATH,
while on Debian GNU/Linux, either can be set, and DT_RPATH is the
default at the time of writing.
How to choose which runtime search path tag is to be set depends on
your system, please refer to ld(1) for the exact information on your
system. As an example, the way to ensure the DT_RUNPATH is set on
Debian GNU/Linux systems rather than DT_RPATH is to tell the linker to
set new dtags, like this:
$ ./config --prefix=/usr/local/ssl --openssldir=/usr/local/ssl \
-Wl,-rpath,'$(LIBRPATH)'
On modern systems using GNU ld.so, a better choice may be to use the
new dtags, like this:
$ ./config --prefix=/usr/local/ssl --openssldir=/usr/local/ssl \
-Wl,--enable-new-dtags,-rpath,'$(LIBRPATH)'
This sets DT_RUNPATH instead of DT_RPATH. DT_RUNPATH is considered after
the environment variable LD_LIBRARY_PATH, while DT_RPATH is considered
before that environment variable (which means that the values in that
environment variable won't matter if the library is found in the
paths given by DT_RPATH).
'-Wl,--enable-new-dtags,-rpath,$(LIBRPATH)'