glibc supports some alternate incorrect format specifiers, like %Ld as an alias for %lld. musl does not support these, and returns an error upon encountering them.
musl does not support glibc's custom format specifier registration system.
musl's floating point conversions honor rounding mode. glibc's do not.
musl's setbuf and setvbuf functions ignore the caller-provided buffer and only switch between fully-buffered, line-buffered, and non-buffered modes.
ISO C and POSIX require the end-of-file status for a FILE to be sticky, i.e. for it to preclude further reads once it's set, unless it's explicitly cleared. musl honors this requirement. glibc ignores it and always returns new input, if available, even after the EOF flag is set.
musl uses readv and writev for implementing stdio:
Normally this just yields good performance independent of heuristics on when to buffer and when to transfer. However, it does also impact how Linux interprets some writes to files in /proc.
By default, glibc provides the legacy BSD behavior for setjmp/longjmp, saving and restoring the signal mask as part of the jmp_buf. musl never includes signal masks unless you use sigsetjmp/siglongjmp and specifically request that the signal mask be saved/restored. Per POSIX, correct applications cannot rely on setjmp to save the signal mask, so musl's behavior gives much better performance for applications which do not care and which are using the interfaces correctly.
musl's regex implementation is based on TRE, with significant modifications. Some popular extensions are supported, but not all; in particular, it lacks some of the GNU extensions to POSIX BRE that add ERE-like capabilities to BRE.
The GNU regex implementation also has an alternate API which can be used instead of the POSIX API. This alternate API is not supported by musl at all.
Finally, glibc's regex uses a 32-bit regoff_t even on 64-bit archs. This is non-conforming to POSIX and precludes giving correct output for strings larger than 2GB. musl uses a correct type, but this renders the ABI of the regex functions incompatible on 64-bit archs. The ABI is compatible on 32-bit.
Calling exit more than once invokes undefined behavior. musl does not support recursive calls to exit (via atexit functions/destructors) or simultaneous calls to exit (from multiple threads). The error is not diagnosed and may give unexpected results, or may happen to do what the programmer intended. musl used to deadlock when exit was called recursively and gracefully handled exit from multiple threads, but the code that caused this behavior was removed in 1.1.2. glibc does not deadlock under recursive exit, but it's unclear whether such usage just "happens to work" or whether it's supported.
glibc's dynamic linker supports lazy binding: deferring resolution of symbol references to the first time the function is called. In theory this reduces startup latency if a large number of references are never used; in reality, it sometimes increases it by delocalizing the accesses. In any case, musl does not support lazy binding. This is both for robustness reasons (reporting failure of lazy binding is impossible) and because it greatly reduces the amount of fragile, arch-dependent code needed in the dynamic linker. Such code has been a perpetual source of bugs in the glibc implementation.
The one place this difference is visible to applications is when applications call dlopen on multiple libraries without proper dependency information, such that, once all libraries are loaded, they all satisfy each other's dependencies, but individually, the first library has unresolved symbols. This is erroneous usage and could also break on archs where full lazy binding is not possible or practical. It would be possible to add support for this dlopen usage case via emulation of lazy bindings, where binding is deferred until a subsequent dlopen call rather than the point of the function call; however, it's unclear whether the implementation cost and complexity are justified in order to support broken applications.
musl's dynamic loader loads libraries permanently for the lifetime of the process, until it exits or calls exec. dlclose is a no-op. This is very different from glibc's approach of reference counting libraries and unloading them when the count reaches zero. The differences in observable behavior to applications:
Either behavior conforms to POSIX, but only the musl behavior can satisfy the robustness conditions musl aims to provide. In particular:
POSIX 2008+TC1 allows dlerror to be thread-safe (with a thread-local error buffer) or non-thread-safe (with a global buffer). Prior versions required it to use a global buffer.
glibc has provided a thread-local dlerror result for a long time. Versions of musl up through 1.1.8 honored an old POSIX requirement and had global dlerror state. Starting with 1.1.9, musl's dlerror state is thread-local and now matches the glibc behavior.
glibc allows the usage of versioned symbols through version tables in gnu-specific elf sections. the musl dynlinker only supports the subset of symbol versioning that allows to pick the default symbol version, instead of the oldest version, which comes first in the regular symbol tables.
The default stack size for new threads on glibc is determined based on the resource limit governing the main thread's stack (RLIMIT_STACK). It generally ends up being 2-10 MB.
musl provides a default stack size of 80k. This does not include the guard page, nor does it include the space used for TLS unless total TLS size is very small. So the actual map size may appear closer to 90k, with around 80k usable by the application. This size was determined empirically with the goals of not gratuitously breaking applications but also not causing large amounts of memory and virtual address space to be committed in programs with large numbers of threads. Programs needing larger stacks, or which explicitly want a smaller stack, should make this explicit with pthread_attr_setstacksize. For largely unrestrained use of the standard library, a minimum of 12k is recommended, but stack sizes down to 2k are allowed.
musl provides strong guarantees of the POSIX-specified side effects being observed when a function which is a cancellation point is cancelled: either the side effects are as if the function returned failure with a status of EINTR, or the function returns success and does not act on cancellation. glibc has no such guarantees.
Cancellation cleanup handling in musl has no relationship to C++ exceptions and unwinding. Any destructors or exception handlers present when acting on cancellation will not be run. Per the standards, cancellation under such conditions yields undefined behavior, but glibc implements cancellation as an exception and thus both cancellation cleanup handlers and exception handlers/dtors get run.
glibc presently has a pseudo-ASCII, pseudo-8-bit C locale: it only defines characters 0 through 127, and the multibyte interfaces treat high bytes as EILSEQ, but on the other hand regex, fnmatch, glob, etc. accept high bytes.
Up through version 1.1.10, musl provided a purely UTF-8 C locale. Despite being unusual, this had precedent (Plan 9) and is explicitly permitted by the ISO C standard (it's documented as an option in the MSE1 Rationale document). However, many scripts expect to be able to process data which is not valid UTF-8 using the standard utilities by setting LC_CTYPE=C, and as of Spring 2013, the Austin Group (responsible for maintaining POSIX) resolved to require in future versions of the standard that the C locale be "8 bit clear", i.e. treat each possible byte as an abstract character, to facilitate processing data in unknown or mixed encodings and other dubious uses. This rendered both glibc's and musl's behaviors non-conforming.
Starting with version 1.1.11, musl provides a special C locale where bytes 0x80-0xff are treated as abstract single-byte-character units with no actual character identity (they're mapped into wchar_t values that occupy the Unicode surrogates range). All other locales are still processed as multibyte UTF-8, and the intent is that the plain C locale's character set be thought of as "UTF-8, but processed byte-by-byte and without validation".
musl uses the Unicode and modern ISO 10646 definition of UTF-8, which is a one-to-one mapping between the Unicode Scalar Values 0-0xd7ff,0xe000-0x10ffff and valid 1-4 byte sequences. In particular, surrogates are rejected, "over-long sequences" are rejected, and pseudo-UTF-8 sequences of 5 or 6 bytes (which would correspond to non-Unicode-scalar-value numbers beyond 0x10ffff) are rejected. glibc on the other hand accepts 5 and 6 byte sequences.
Aside from being a matter of standards conformance, accepting sequences only up to 4 bytes allows for the possibility of generating UTF-8 output in a buffer intended for wchar_t, then converting it in-place, which is necessary for some interfaces that cannot rely on being able to allocate working space.
The iconv implementation musl is very small and oriented towards being unobtrusive to static link. Its character set/encoding coverage is very strong for its size, but not comprehensive like glibc's. In particular:
musl implements the library requirements for C99 Annex F. Unlike glibc, musl does not try to support non-conforming floating-point formats and arithmetics (eg. the powerpc abi with software long double using two doubles, the IBM long double format).
Glibc supports exception unmasking using fesetenv with non-standard fenv setting. Musl does not support exception unmasking and library functions assume exceptions are always masked (floating-point arithmetics never trap).
IEEE 754 specifies five exception signals and five exception status flags as well (inexact, invalid, divbyzero, underflow, overflow). The signals are all masked in the default environment so in case of an exceptional condition the system does not trap but delivers a default result and raises the appropriate status flag. (Note: exact underflow is a special case that signals the underflow exception but does not raise the underflow status flag). The C standard does not provide any way to unmask the exception signals (but the specification of feraiseexcept and fesetexceptflag takes unmasked exception signals into account). Thus a conformant implementation can behave as if floating-point computations never trap, the observable behaviour is conforming, unless the application circumvents the libc (eg. by using assembly) and unmasks the exception signals (or invokes other unspecified fenv operations, eg. changes the x87 FPU precision setting on i386) in which case the behaviour of library functions is undefined.
Glibc supports non-zero math_errhandling & MATH_ERRNO, ie. exceptional conditions of the mathematical library are reported in errno. Musl only supports MATH_ERREXCEPT.
At least one of MATH_ERRNO and MATH_ERREXCEPT must be supported and Annex F requires MATH_ERREXCEPT support. Old platforms and platforms without FPU may not provide the necessary IEEE 754 exception semantics for MATH_ERREXCEPT. On such platforms an implementation can only fall back to the classic errno based error reporting that was already present in C89. (Note: the errno based method provides less information, only EDOM and ERANGE is reported and the semantics is less precise).
On software FPU platforms without hardware register for floating-point status flags the software emulation library (provided by the compiler, eg. libgcc) can emulate the status flags with thread storage duration. Currently the status flags are not emulated, so musl is non-conforming on such platforms (neither errno nor fp exceptions are supported).
Traditional resolvers, including glibc's, make use of multiple nameserver lines in resolv.conf by trying each one in sequence and falling to the next after one times out. musl's resolver queries them all in parallel and accepts whichever response arrives first. This can increase network load (this is mitigated by only supporting up to three nameservers, and can be mitigated further at the configuration level by only configuring one nameserver) but drastically improves performance and reliability of DNS lookups, especially if diverse nameservers are used. An ideal configuration is:
musl's resolver previously did not support the "domain" and "search" keywords in resolv.conf. This feature was added in version 1.1.13, but its behavior differs slightly from glibc's: queries with fewer dots than the ndots configuration variable are processed with search first then tried literally (just like glibc), but those with at least as many dots as ndots are only tried in the global namespace (never falling back to search, which glibc would do if the name is not found in the global DNS namespace). This difference comes from a consistency requirement not to return different results subject to transient failures or to global DNS namespace changes outside of one's control (addition of new TLDs).
glibc supports IDN (non-ASCII name lookups via DNS) but requires the use of custom non-standard flags to getaddrinfo and getnameinfo to convert such names to/from the internal representation used in the DNS system (Punycode). musl does not yet support IDN conversions at all (but programs that know about IDN/Punycode and do the conversions themselves can of course lookup such names), but when support is added, it will always be enabled rather than requiring explicit options from the application.
You can contribute to this wiki! Submit pull-requests to somasis/musl-wiki.
This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported license, unless otherwise noted.
Last edited by Kylie McClain, 2016-07-01 16:55:02