Hi everyone, Apologies if this has already been answered before. I am experimenting with musl. I see that musl produces a single DSO with all symbols, as opposed to glibc which produces multiple DSOs (libthread, libm, librt, etc.). I also notice that musl generates some empty archives for compatibility reasons, namely the lib{crypt,dl,m,pthread,resolv,rt,util,xnet}.a. These are already documented in the FAQ [1]. By looking at the code [2], I see that musl's dynamic linker ignores dynamic dependencies with names lib{c,pthread,rt,m,dl,util,xnet} and this makes sense based on the above. What doesn't make sense to me is that musl's dynamic linker does not ignore dynamic dependencies with names libresolv and libcrypt. Is there a reason for this? Thanks in advance, Nikos [1] https://wiki.musl-libc.org/faq.html#Q:-lib(m|pthread|crypt).a/so-are-empty? [2] https://git.musl-libc.org/cgit/musl/tree/ldso/dynlink.c#n991
On Wed, Sep 02, 2020 at 07:14:10PM +0300, Nikos Dragazis wrote: > Hi everyone, > > Apologies if this has already been answered before. > > I am experimenting with musl. I see that musl produces a single DSO with > all symbols, as opposed to glibc which produces multiple DSOs > (libthread, libm, librt, etc.). I also notice that musl generates some > empty archives for compatibility reasons, namely the > lib{crypt,dl,m,pthread,resolv,rt,util,xnet}.a. These are already > documented in the FAQ [1]. > > By looking at the code [2], I see that musl's dynamic linker ignores > dynamic dependencies with names lib{c,pthread,rt,m,dl,util,xnet} and > this makes sense based on the above. > > What doesn't make sense to me is that musl's dynamic linker does not > ignore dynamic dependencies with names libresolv and libcrypt. Is there > a reason for this? I believe the intent was only to reserve names that POSIX explicitly reserves: If a directory specified by a -L option contains files with names starting with any of the strings "libc.", "libl.", "libpthread.", "libm.", "librt.", "libtrace.", "libxnet.", or "liby.", the results are unspecified. https://pubs.opengroup.org/onlinepubs/9699919799/utilities/c99.html However it looks like the sets don't entirely match up. I'm not sure of the reason for the mismatch. The set of "builtin" library names should probably be broken into two parts: ones that will always be used (c, pthread, m, etc) and ones that will be used as fallbacks if no file is found (resolv, crypt, etc.). Note that the purpose of these built-in names is twofold: (1) it's part of glibc ABI-compat, for running glibc binaries with their names in DT_NEEDED, and this role could be moved out with the gcompat refactor if desired, and (2) supporting programs that (kinda dubiously) use dlopen with these names to access standard functionality. Rich
On 2/9/20 7:42 μ.μ., Rich Felker wrote: > On Wed, Sep 02, 2020 at 07:14:10PM +0300, Nikos Dragazis wrote: >> Hi everyone, >> >> Apologies if this has already been answered before. >> >> I am experimenting with musl. I see that musl produces a single DSO with >> all symbols, as opposed to glibc which produces multiple DSOs >> (libthread, libm, librt, etc.). I also notice that musl generates some >> empty archives for compatibility reasons, namely the >> lib{crypt,dl,m,pthread,resolv,rt,util,xnet}.a. These are already >> documented in the FAQ [1]. >> >> By looking at the code [2], I see that musl's dynamic linker ignores >> dynamic dependencies with names lib{c,pthread,rt,m,dl,util,xnet} and >> this makes sense based on the above. >> >> What doesn't make sense to me is that musl's dynamic linker does not >> ignore dynamic dependencies with names libresolv and libcrypt. Is there >> a reason for this? > I believe the intent was only to reserve names that POSIX explicitly > reserves: > > If a directory specified by a -L option contains files with names > starting with any of the strings "libc.", "libl.", "libpthread.", > "libm.", "librt.", "libtrace.", "libxnet.", or "liby.", the > results are unspecified. > > https://pubs.opengroup.org/onlinepubs/9699919799/utilities/c99.html > > However it looks like the sets don't entirely match up. I'm not sure > of the reason for the mismatch. > > The set of "builtin" library names should probably be broken into two > parts: ones that will always be used (c, pthread, m, etc) and ones > that will be used as fallbacks if no file is found (resolv, crypt, > etc.). It's not clear to me how the second part could be useful. Also, it would break the scenario where one wants to use musl to run a glibc-compat executable on a glibc-based system, because the linker would load glibc's libresolv.so. > > Note that the purpose of these built-in names is twofold: (1) it's > part of glibc ABI-compat, for running glibc binaries with their names > in DT_NEEDED Then definitely resolv and crypt should be part of it. > , and this role could be moved out with the gcompat > refactor if desired, and (2) supporting programs that (kinda > dubiously) use dlopen with these names to access standard > functionality. > > Rich