While testing the coreutils programs the following utilities required getlogin()

• logname

Our pthread_mutex struct already has a member to store the level of nested locks but it is not used yet. Add a sys_futex_tid function to get a per-thread ID. When locking a mutex, store the sys_futex_tid in the mutex (instead of 0/1, as we currently do).

Prevents many programs from compiling, standarized by POSIX.

http://man7.org/linux/man-pages/man3/getline.3.html

Under Linux, libraries in subdirs of /usr/lib, for example, man-db in /usr/lib/man-db, are found and used as requested. When running managarm, the libraries in these directories are not found, and need to be copied over manually to /usr/lib.

Tested with man-db's man command

While testing the coreutils programs the following utilities required newlocale()

• sleep
• timeout

While testing on managarm the following program requires pthread_atfork()

• perl

While testing the coreutils programs the following utilities required linkat()

• ln

While testing the coreutils programs the following utilities required getpriority()

• nice

mlibc is failing in __cxa_atexit(), link. When memory is not zeroed before giving it to userland

When executing a symlink (in my test case, /usr/bin/cc pointing to /usr/bin/gcc) the following is printed

In function operator(), file ../../../src/mlibc/options/posix/generic/posix_stdlib.cpp:200
__ensure(!S_ISLNK(st.st_mode) && "TO

The full ensure line is as follows

__ensure(!S_ISLNK(st.st_mode) && "TODO: Use readlink in realpath()");

While testing the coreutils programs the following utilities required faccessat()

• sort

While testing git on managarm, the following bug came to my attention.
By doing the following operations in order, the program can be killed with an SIGSEGV
To reproduce:

• Have git installed and a git repo available, in this test case managarm
• Do the required git setup, for managarm, this is touch /dev/null, and setting the global username and email (git will complain about this)
• In the git repo, touch a file, and run git add .
• Run git commit -m <message>, where message is the commit message (irrelevant for triggering this bug) and watch it crash.

Running addr2line on libc.so gives the following line

dennis@DENNIS-MAIN-LAPTOP-SIDUCTION:~/managarm-clone/build$ addr2line --e system-root/usr/lib/libc.so 
0x987d0
/home/dennis/managarm-clone/build/pkg-builds/mlibc/../../../src/mlibc/options/internal/include/mlibc/lock.hpp:24

which translates to

if(__atomic_compare_exchange_n(&_futex, &expected, 1,
							false, __ATOMIC_ACQUIRE, __ATOMIC_ACQUIRE)) {

where the first part of the if statement is line 24 in lock.hpp.

While testing the coreutils programs the following utilities required an unimplemented printf() specifier ('76')

• seq
• dd

The following program required the unimplemented printf() specifier ('39'), which converts to an '

• xz

Furthermore, it seems we are lacking an implementation of %f

mlibc's ABI differs from Linux and the BSDs in the following aspects:

• File type constants returned by stat() have the same values on Linux and the BSDs but differ in mlibc.

Coreutils mv program calls renameat() to perform the move operation, this function is unimplemented. Implementing it would allow mv and programs relying on it to function.

While testing the coreutils programs on managarm the stat program crashed with the following error:

../../../src/mlibc/subprojects/frigg/include/frg/formatting.hpp:379: Assertion '!opts.left_justify' failed!

This can be reproduced by stating any regular file.

Running the following command gives the same error, on a different location, this time at line 459
readelf -I <executable>

As discussed on discord there are numerous improvements to be made to our current implementation of getaddrinfo()

• Handle DNS ipv6 records
• Parse /etc/resolv.conf for the nameserver
• Parse /etc/hosts and use it in name resolution
• Parse /etc/services to implement the services part of the function
• Handle CNAME records correctly
• Handle hints properly
• Handle all flags properly
• Sort output list according to RFC 3484

With the recent addition of tcp to managarm, it has become clear that we are lacking in name resolution. One major example is getaddrinfo(), required by at least git from my initial testing. We should strive to implement this as soon as possible

Currently, we do not use the dynamic linker in static builds. That is problematic for two reasons:

• dlopen() will not work
• ld.so is responsible for allocating thread-local storage (TLS)

Fix this by linking the dynamic linker into the static library. Use the (undocumented) __ehdr_start symbol to find the ELF EHDR of the executable. Read the PHDR offset from the EHDR and use that to find the executable's PHDRs. Then proceed to do dynamic linking as usual (naturally, there will be no relocations, so this will mostly be a no-op).

As an additional benefit, this change will unify the initialization paths (for .init_array etc.) in the static and dynamic cases.

The following program wants setgroups() to function correctly:

• locate from the package findutils

This allows us to call sysdeps as we like, without having to worry about corrupting errno.

As of e70ccf0, the following files are affected:

• options/rtdl/generic/linker.cpp
• options/glibc/generic/sys-ioctl.cpp
• options/posix/generic/dirent-stubs.cpp
• options/posix/generic/termios-stubs.cpp
• options/posix/generic/fcntl-stubs.cpp
• options/posix/generic/sys-select-stubs.cpp
• options/posix/generic/sys-mman-stubs.cpp
• options/posix/generic/posix_stdlib.cpp
• options/posix/generic/pthread-stubs.cpp
• options/posix/generic/sys-time-stubs.cpp
• options/posix/generic/sched-stubs.cpp
• options/posix/generic/sys-stat-stubs.cpp
• options/posix/generic/unistd-stubs.cpp
• options/posix/generic/sys-wait-stubs.cpp
• options/posix/generic/posix_signal.cpp
• options/posix/generic/sys-socket-stubs.cpp
• options/internal/include/mlibc/sysdeps.hpp
• options/internal/generic/debug.cpp
• options/internal/generic/frigg.cpp
• options/internal/generic/allocator.cpp
• options/internal/gcc/guard-abi.cpp
• options/internal/gcc/initfini.cpp
• options/ansi/generic/stdio-stubs.cpp
• options/ansi/generic/signal-stubs.cpp
• options/ansi/generic/stdlib-stubs.cpp
• options/ansi/generic/time-stubs.cpp
• options/ansi/generic/file-io.cpp
• options/linux/generic/poll-stubs.cpp
• options/linux/generic/pty-stubs.cpp
• options/linux/generic/sys-signalfd.cpp
• options/linux/generic/sys-inotify-stubs.cpp
• options/linux/generic/sys-epoll.cpp
• options/linux/generic/sys-timerfd.cpp
• options/linux/generic/sys-mount.cpp
• sysdeps/qword/generic/file.cpp
• sysdeps/managarm/rtdl-generic/support.cpp
• sysdeps/managarm/generic/memory.cpp
• sysdeps/managarm/generic/signals.cpp
• sysdeps/managarm/generic/file.cpp
• sysdeps/managarm/generic/fork-exec.cpp
• sysdeps/managarm/generic/socket.cpp
• sysdeps/managarm/generic/time.cpp
• sysdeps/managarm/generic/mount.cpp
• sysdeps/managarm/generic/ensure.cpp

GNU coreutils cp calls dirfd() when copying a directory.

The following programs, tested on managarm, require inet_pton() to properly function

• curl, running curl <ip> and curl <url> crashes due to the inet_pton() function being unimplemented

While testing the coreutils programs the following utilities required mkfifo()

• mkfifo

While testing the coreutils programs the following utilities required sync()

• sync

While testing on qword, dpkg failed to initialize when reading it's configuration directory.

While testing on managarm I found that the following programs required sigsetjmp() and siglongjmp()

• gdb
• perl

While testing the coreutils programs the following utilities required link()

• link

While testing git on managarm, the following unimplemented functions are observed

• pthread_setcancelstate(), required for most of the operations, including git commit and git log
• rmdir(), required for git checkout, possibly only when directories need to be removed (see issue #50).
• utime(), required for git add, this is implemented for managarm, but calls it with unsupported modes (meaning one of the fields is not UTIME_NOW).
• with git add working(ish), git commit now throws an SIGSEGV in an internal mlibc header, before triggering the pthread_setcancelstate() function. This problem is tracked in a seperate issue, issue #111
• fetching repo's with either git clone, git pull or git fetch (assuming git:// repo's) require handling of the /etc/services file to work (part of issue #99)
• fetching, cloning or pulling repo's (http(s):// repo's) require curl (tracked in managarm/bootstrap-managarm#33)

This is a tracking issue for all git related problems in mlibc

For example, SIGIO and SIGBUS have the same value

Currently, decoding unicode data can produce bad results, e.g., if cat is used to display a UTF-8 file.

The following programs require textdomain() to be implemented:

• gapplication from the package glib
• gdbus from the package glib
• gresource from the package glib
• gsettings from the package glib

While testing the coreutils programs the following utilities required _freadahead()

• cp
• install
• rm

The coreutilities program hostid wants gethostid() to function

On managarm, when running any command with nohup, for example, nohup true, the following line is returned

(null): failed to render standard input unusable: No such file or directory (ENOENT)

This can be reproduced by running nohup true, although nohup <command> should work as well.

The following programs, tested on managarm, require regcomp() to properly function

• file, file works for executables, but for other files, regcomp() is called, thus severely limiting it's use.
• flex, all programs from flex want regcomp(). Without it, even flex --help fails

This requires a PIC-enabled crt1.

While testing the coreutils programs the following utilities required freopen()

• shuf
• tsort
• uniq

Add an option to write some non-zero byte pattern to malloc()ed memory and/or memory obtained from sys_anon_allocate to simplify memory testing.

While testing managarm, the following program from coreutils crashed on the mknod() function being unimplemented

• mknod

mlibc always closes stdin, stdout and stderr on exit. This can fail (and, currently, crash mlibc) if the program decided to close those streams itself.

While testing the coreutils programs the following utilities required fdatasync()

• shred

The following programs also requires fdatasync()

• update-mime-info from the package shared-mime-info
• xbps-pkgdb from the package xbps

With the recent addition of tcp to managarm, it has become clear that we are lacking in name resolution. One major example is gethostbyname(), required by at least lynx and syslogd from my initial testing. We should strive to implement this as soon as possible

During testing on managarm, the following program required getentropy()

• wget, when fetching https:// links

The following programs want setmntent() to function correctly:

• df from the package coreutils
• updatedb from the package findutils

A lot of programs get their own name by looking at program_invocation_name or program_invocation_short_name. Right now, those are NULL. Setting these correctly would fix this.

When running /usr/bin/mknod <filename> c 1 3 on managarm, which should be a valid command according to LFS, mknod crashes with the following error

In function strtoxmax, file ../../../src/mlibc/options/ansi/generic/inttypes-stubs.cpp:25 __ensure(base) failed

This can be reliably be reproduced by running the command specified above

While testing the coreutils programs the following utilities required rmdir()

• rmdir

GNU-style hashes are enabled by default on modern Linux distributions. Supporting them would simplify linking against mlibc on Linux.