glibc.txt: Barret Rhoden Last updated: 2010-08-22 ---------------------------------------------- This document is an attempt to gather knowledge about how to work with the glibc port. This is based on my half-assed struggling with the port done by kevin and andrew. When to recompile (other than when changing libc) -------------------------- Whenever you change a kernel header (something in kern/include/ros), technically you should rebuild. It isn't critical in all places, such as for files that aren't included by glibc, or for files that haven't changed something that glibc relies on. After any recompile, don't forget to rebuild userspace apps (make userclean, make install-libs) and to refresh your FS. How to recompile glibc -------------------------- Normally, make x86 (or whatever) works fairly well. However, sometimes you need to rebuild all of glibc. This might be on a hunch, or to get rid of things that might have compiled, but are failing to link and don't seem to rebuild. This happens when you change parts of glibc. If you know what file(s) you changed and it doesn't seem like the changes are being applied, you can copy the file from the -ros directory to the glibc-XXX directory. The glibc make process doesn't apply diffs or track the changes from -ros to the glibc directory, so your recompile won't notice the changes. If things are stil messed up, the next step is to remove the build directories (i686-ros-glibc-*) and also the hidden files (.i686-ros-glibc*). If you get errors from make very early on about not finding some targets (such as install-headers), you forgot to delete the hidden files. Then next step up would be to remove the glibc-XXX folder (whatever the version is). Sometimes you aren't sure if old changes are sitting around in this folder that aren't getting overwritten by the real source of our changes, the -ros folder. A lovely example of this is an old Makefile for the nss subdir. It was altered to remove some things, notably the files service, which translates into libnss_files - which is needed by other things (nis, etc). To fix this, I had to notice an -ros version of the Makefile (after I realized it was the problem), and that the Makefile was being included - it was just faulty. Finally, you can always make clean, but the overall XCC Makefile will clean everything - including gcc and binutils. Note that if you are making a trivial addition to a kernel header, you can get away with just copying it to its location in the XCC install directory (sys-include). The -ros folder -------------------------- All changes to glibc must be made in the glibc-XXX-ros folder. Any changes here will be copied over to the real glibc folder when it is made. If you want to remove a file that was originally in the -ros folder, you need to manually remove it from the glibc-XXX folder as well. The updating scripts will not remove a file. The sysdeps folder -------------------------- Most of your changes will go here. Every system that has a glibc port should have one of these folders, with its system-specific ports. The files here are normally files expected elsewhere in libc, and the glibc build system knows to look in these places. Simply dropping a file in here will not get it built. It will for many files, but some others that weren't expected will need to be manually added to a Makefile. If you notice some files not getting compiled, (drop a #error in there), you'll need to add it to a Makefile. In the main sysdeps/Makefile, add the filename (minus .c) to the sysdeps var. Look at sa_len for an example. Sometimes you'll need to be careful about adding it for some subdirectories and not others (you probably only want to add a C file for one subdir). Check out Linux's port for help. Sometimes you have files that you want to change outside of the sysdeps folder. These still go in the glibc-XXX-ros folder, but not in sysdeps. The main example we have right now is features.h. At one point, I wanted to change sys/socket.h. Simply adding it to sysdeps/ros/sys/socket.h would not change the main sys/socket.h, nor would the sysdep version get included first. Putting it in the -ros/sys/ folder did it (though ultimately I didn't want the change). The point is, sysdeps doesn't mirror and override the main tree for all files - it is behind some others in the search/include path. Another situation requiring a change outside of the sysdeps directory was sunrpc/netname.c. I wanted to change the functions (stub out the ones that used NSS). Adding the sysdep worked, but it turns out that *both* the sysdep netname.c and the original sunrpc/netname.c were being compiled. The root cause seems to be compat-netname.os. There are make rules in sunrpc to generate some compatibility routines. Grep for rpc-compat-routines.os. The rule seems to ignore sysdeps and just use the normal C file - in this case netname.c. Subdirs -------------------------- As a note, these 'subdirectories' are the "primary folders" (i.e. addons), such as inet, ncsd, libio, whatever. These are the root folders in glibc, representing some major functionality. They can be built with the --enable-addons switch, which by default builds all of them. Sort of! To really get them to even be available for a port, you need to "include" them in a certain manner. There are two ways. One is the Subdirs file in the sysdeps/ros/ directory. Putting a subdir name in here means glibc will try to build it; it is available to be an addon. Careful with these, since a lot of the folders tend to need each other (like most all of the ones in unix/inet). If you want a massive subsystem, such as "unix/inet" or "posix", you can add it to the sysdeps/ros/Implies file. You will get a bunch of these folders at once, plus some other unspecified stuff (necessary for the overall system, perhaps?). If you add "unix/inet", you get more than just its Subdirs. Also note that these subdirs can depend on each other, noted in the "Depends" file. Presumably this will cause them to get made... Errno & TLS -------------------------- errno is a macro that may point to different locations depending on where you are in Glibc. errno_location is simply one of the options. During dynamic linking, the linker points errno to a different location, so it is usable and can be referenced before TLS is fully set up in TLS_INIT_TP. Because errno is valid when the linker runs, regular syscalls can be made. However for statically linked apps, several syscalls cannot use the ros_syscall macro, because there is no valid errno set up. Unimplemented Stubs -------------------------- There are a lot of things we haven't ported, but we have the stub functions so that we can at least compile things against it. When you compile (including regular programs), you'll get warnings about this. Linux ASM bits -------------------------- We've included some header files from Linux's kernel. For now, we just need something to keep glibc happy, even though ROS doesn't have any networking syscalls. Don't rely on this stuff too much. These files, and other future glibc compatibility kernel headers, are in kern/include/ros/glibc-asm. Weak Symbols, start.c, and ros_syscall_blockon -------------------------- For a long time, __ros_syscall_blockon was not getting overridden by the strong symbol in parlib. This means that glibc's syscalls were not blocking uthreads properly. Why did the weak symbols work for vcore_entry() and vcore_event_init(), but not blockon? Side note: we needed to force the linker to find out about vcore_entry and vcore_event_init, via the -u tricks of commit f188983. The linker is not obligated to look in libraries to override a weak symbol. The crux of the matter is that start.c is not linked with applications in the same manner as the rest of glibc (notably all the I/O syscalls). start.c will get linked with the program at compile time, while libc can be linked dynamically. Because of this, start.c's weak symbols would get (correctly) overridden by the strong symbols of libparlib.a. But glibc would build libc.so, and that would not get a chance to link against the binary (and libparlib.a) until load time. The weak symbols in libc get promoted to strong symbols when it is done linking libc.so, and when it later is linked against the program, there is no longer an opportunity to override the weak symbol. Also note that rtld will never link with parlib, so any weak symbol in there will never get overriden. I briefly considered linking rtld with -lparlib, but it blows up in a nasty way: parlib needs lots of libc, which rtld is not built with (memset for example). Anyway, the moral of the story is to be careful with your weak symbols/aliases. Only put them in start.c, or similar files. Adding a Global Variable to the ABI -------------------------- It's not enough to simply 'extern' a variable (or declare a function, which is extern by default). At some point, glibc will change those symbols from GLOBAL to LOCAL. You need to add an entry to the Versions file (sysdeps/ros/Versions). Be sure to indent with spaces, not tabs, or else glibc's scripts will not handle your symbol. Putting them in the libc, glibc 2.0 section seems to work fine. Tips, Questions, and Misc Notes -------------------------- - Grep and find are your friend. - Watch what Linux does. - The kernel headers end up getting installed to the sys-include/ directory, while the regular glib headers go to the include directory. - atomic_swap might have issues (warnings about signedness differences) - It's not always clear which files in the -ros folder actually need to be there (presumably they all do), and in what manner they were modified and from which original copy. Ideally, the ROS-specific changes to large files will be commented. - the SHARED flag is a bit of a mess - things get compiled differently for shared vs static, and it can get complicated with start.c, tls.c, etc. - What things in one file rely heavily on another file? Are there non-obvious gotchas? (yes, and no one documented them). - Is the build failing without any clear error messages? Scroll up a lot, and there may be messages farther up (like a hundred+ lines up). I've had some gcc stage2 builds that fail with no obvious issue in the short term console output, but the real error is much higher. Some aspect of the build system will continue on failures and only fail much later, after building other packages. - Note that libstdc++ is a subpart of gcc, built during stage2, and has its own configure script and settings. Ghetto Things (Feel free to fix them): -------------------------- - ptsname: we needed to have the sysdep (was being looked for by the login subdir make process), but having the actual functions in it caused a link error (multiple declarations). So we have an empty file...