this is the first step in an overhaul aimed at greatly simplifying and
optimizing everything dealing with thread-local state.
previously, the thread pointer was initialized lazily on first access,
or at program startup if stack protector was in use, or at certain
random places where inconsistent state could be reached if it were not
initialized early. while believed to be fully correct, the logic was
fragile and non-obvious.
in the first phase of the thread pointer overhaul, support is retained
(and in some cases improved) for systems/situation where loading the
thread pointer fails, e.g. old kernels.
some notes on specific changes:
- the confusing use of libc.main_thread as an indicator that the
thread pointer is initialized is eliminated in favor of an explicit
has_thread_pointer predicate.
- sigaction no longer needs to ensure that the thread pointer is
initialized before installing a signal handler (this was needed to
prevent a situation where the signal handler caused the thread
pointer to be initialized and the subsequent sigreturn cleared it
again) but it still needs to ensure that implementation-internal
thread-related signals are not blocked.
- pthread tsd initialization for the main thread is deferred in a new
manner to minimize bloat in the static-linked __init_tp code.
- pthread_setcancelstate no longer needs special handling for the
situation before the thread pointer is initialized. it simply fails
on systems that cannot support a thread pointer, which are
non-conforming anyway.
- pthread_cleanup_push/pop now check for missing thread pointer and
nop themselves out in this case, so stdio no longer needs to avoid
the cancellable path when the thread pointer is not available.
a number of cases remain where certain interfaces may crash if the
system does not support a thread pointer. at this point, these should
be limited to pthread interfaces, and the number of such cases should
be fewer than before.
the external mmap function is heavy because it has to handle error
reporting that the kernel cannot do, and has to do some locking for
arcane race-condition-avoidance purposes. for allocating initial TLS,
we do not need any of that; the raw syscall suffices.
on i386, this change shaves off 13% of the size of .text for the empty
program.
in general, we aim to always include the header that's declaring a
function before defining it so that the compiler can check that
prototypes match.
additionally, the internal syscall.h declares __syscall_ret with a
visibility attribute to improve code generation for shared libc (to
prevent gratuitous GOT-register loads). this declaration should be
visible at the point where __syscall_ret is defined, too, or the
inconsistency could theoretically lead to problems at link-time.
in addition to the dbm functions (which we don't intent to implement
anyway), fmtmsg is still missing too. rather than adding exceptions I
think it's best just to avoid making the claim.
the text covering an ill-advised procedure for 'bootstrapping' a new
musl-based system in-place is removed. new information on targets and
compilers is added. formatting improved. the remaining text is
adjusted to cover both usage with musl-gcc on a non-musl-based system
and upgrading a musl-based system or toolchain.
the excess space was unused and unintentional. this change does not
affect the ABI between applications and libc. while it does
theoretically affect linkage between third-party translation units
using jmp_buf as part of a structure, we've already changed jmp_buf at
least once on all archs, and problems were never observed, likely
because such usage would be very unusual. in any case it's best to get
things right now rather than making changes sometime during the 1.0.x
series or later.
on x32, this change allows programs which use syscall() with pointers
or 64-bit values as arguments to work correctly, i.e. without
truncation or incorrect sign extension. on all other supported archs,
syscall_arg_t is defined as long, so this change is a no-op.
the previous pattern required "x32" to be used as the second field of
the gcc tuple, which is usually reserved for vendor use and not
appropriate as an ABI specifier. with this change, putting "x32" at
the end of the tuple, the way ABI specifiers are normally done, is
also permitted.
the incorrect error codes also made their way into errno when
__ptsname_r was called by plain ptsname, which reports errors via
errno rather than a return value.
these have been wrong for a long time and were never detected or
corrected. powerpc needs some gratuitous extra padding/reserved slots
in ipc_perm, big-endian ordering for the padding of time_t slots that
was intended by the kernel folks to allow a transition to 64-bit
time_t, and some minor gratuitous reordering of struct members.
the definition was found to be incorrect at least for powerpc, and
fixing this cleanly requires making the definition arch-specific. this
will allow cleaning up the definition for other archs to make it more
specific, and reversing some of the ugliness (time_t hacks) introduced
with the x32 port.
this first commit simply copies the existing definition to each arch
without any changes. this is intentional, to make it easier to review
changes made on a per-arch basis.
the printf floating point formatting code contains an optimization to
avoid computing digits that will be thrown away by rounding at the
specified (or default) precision. while it was correctly retaining all
places up to the last decimal place to be printed, it was not
retaining enough precision to see the next nonzero decimal place in
all cases. this could cause incorrect rounding down in round-to-even
(default) rounding mode, for example, when printing 0.5+DBL_EPSILON
with "%.0f".
in the fix, LDBL_MANT_DIG/3 is a lazy (non-sharp) upper bound on the
number of zeros between any two nonzero decimal digits.
empirically the overflow was an off-by-one, and it did not seem to be
overwriting meaningful data. rather than simply increasing the buffer
size by one, however, I have attempted to make the size obviously
correct in terms of bounds on the number of iterations for the loops
that fill the buffer. this still results in no more than a negligible
size increase of the buffer on the stack (6-7 32-bit slots) and is a
"safer" fix unless/until somebody wants to do the proof that a smaller
buffer would suffice.
this was problematic because several archs don't define __WORDSIZE. we
could add it, but I would rather phase this macro out in the long
term. in our version of the headers, UINTPTR_MAX is available here, so
just use it instead.
neither is correct; different commands take different argument types,
and some take no arguments at all. I have a much larger overhaul of
fcntl prepared to address this, but it's not appropriate to commit
during freeze.
the immediate problem being addressed affects forward-compatibility on
x32: if new commands are added and they take pointers, but the
libc-level fcntl function is not aware of them, using long would
sign-extend the pointer to 64 bits and give the kernel an invalid
pointer. on the kernel side, the argument to fcntl is always treated
as unsigned long, so no harm is done by treating possibly-signed
integer arguments as unsigned. for every command that takes an integer
argument except for F_SETOWN, large integer arguments and negative
arguments are handled identically anyway. in the case of F_SETOWN, the
kernel is responsible for converting the argument which it received as
unsigned long to int, so the sign of negative arguments is recovered.
the other problem that will be addressed later is that the type passed
to va_arg does not match the type in the caller of fcntl. an advanced
compiler doing cross-translation-unit analysis could potentially see
this mismatch and issue warnings or otherwise make trouble.
on i386, this patch was confirmed not to alter the code generated by
gcc 4.7.3. in principle the generated code should not be affected on
any arch except x32.
if we ever encounter other targets where error codes don't fit in the
8-bit range, the table should probably just be bumped to 16-bit, but
for now I don't want to increase the table size on all archs just
because of a bug in the mips abi.
most notably, it was failing to match sh4-*, etc., but in general the
explicit matching of hyphens for some archs was problematic because it
failed to accept simply the musl-style arch name (without a gcc-style
tuple) as an input. the original motivation of matching hyphens was to
prevent incorrectly identifying a 64-bit arch as the corresponding
32-bit arch (e.g. mips* matching mips64) but this is easily fixed by
simply checking (and for now, rejecting as unsupported) the relevant
64-bit archs.
linux, gcc, etc. all use "sh" as the name for the superh arch. there
was already some inconsistency internally in musl: the dynamic linker
was searching for "ld-musl-sh.path" as its path file despite its own
name being "ld-musl-superh.so.1". there was some sentiment in both
directions as to how to resolve the inconsistency, but overall "sh"
was favored.
per POSIX, ENOENT is reserved for invalid stream position; it is an
optional error and would only happen if the application performs
invalid seeks on the underlying file descriptor. however, linux's
getdents syscall also returns ENOENT if the directory was removed
between the time it was opened and the time of the read. we need to
catch this case and remap it to simple end-of-file condition (null
pointer return value like an error, but no change to errno). this
issue reportedly affects GNU make in certain corner cases.
rather than backing up and restoring errno, I've just changed the
syscall to be made in a way that doesn't affect errno (via an inline
syscall rather than a call to the __getdents function). the latter
still exists for the purpose of providing the public getdents alias
which sets errno.
the other atomic FD_CLOEXEC interfaces (dup3, pipe2, socket) already
had such emulation in place. the justification for doing the emulation
here is the same as for the other functions: it allows applications to
simply use accept4 rather than having to have their own fallback code
for ENOSYS/EINVAL (which one you get is arch-specific!) and there is
no reasonable way an application could benefit from knowing the
operation is emulated/non-atomic since there is no workaround at the
application level for non-atomicity (that is the whole reason these
interfaces were added).
this was unlikely to lead to any crash or dangerous behavior, but
caused adjacent string constants to be treated as part of the
protocols table, possibly returning nonsensical results for unknown
protocol names/numbers or when getprotoent was called in a loop to
enumerate all protocols.
this saves a syscall in the case where the underlying open already
took place with O_APPEND, which is common because fopen with append
modes sets O_APPEND at the time of open before passing the file
descriptor to __fdopen.
when there is unflushed output, ftello (and ftell) compute the logical
stream position as the underlying file descriptor's offset plus an
adjustment for the amount of buffered data. however, this can give the
wrong result for append-mode streams where the unflushed writes should
adjust the logical position to be at the end of the file, as if a seek
to end-of-file takes place before the write.
the solution turns out to be a simple trick: when ftello (indirectly)
calls lseek to determine the current file offset, use SEEK_END instead
of SEEK_CUR if the stream is append-mode and there's unwritten
buffered data.
the ISO C rules regarding switching between reading and writing for a
stream opened in an update mode, along with the POSIX rules regarding
switching "active handles", conveniently leave undefined the
hypothetical usage cases where this fix might lead to observably
incorrect offsets.
the bug being fixed was discovered via the test case for glibc issue
the affected part of the header is responsible for providing both GNU
and BSD versions of the udphdr structure. previously, the
namespace-polluting GNU names were always used for the actual struct
members, and the BSD names, which are named in a manner resembling a
sane namespace, were always macros defined to expand to the GNU names.
now, unless _GNU_SOURCE is defined, the BSD names are used as the
actual structure members, and the macros and GNU names only come into
play when the application requests them.
there are two versions of this structure: the BSD version and the GNU
version. previously only the GNU version was supported. the only way
to support both simultaneously is with an anonymous union, which was a
nonstandard extension prior to C11, so some effort is made to avoid
breakage with compilers which do not support anonymous unions.
this commit is based on a patch by Timo Teräs, but with some changes.
in particular, the GNU version of the structure is not exposed unless
_GNU_SOURCE is defined; this both avoids namespace pollution and
dependency on anonymous unions in the default feature profile.
these are poorly designed (illogical argument order) and even poorly
implemented (brace issues) on glibc, but unfortunately some software
is using them. we could consider removing them again in the future at
some point if they're documented as deprecated, but for now the
simplest thing to do is just to provide them under _GNU_SOURCE.
STB_WEAK is only a weak reference for undefined symbols (those with a
section of SHN_UNDEF). otherwise, it's a weak definition. normally
this distinction would not matter, since a relocation referencing a
symbol that also provides a definition (not SHN_UNDEF) will always
succeed in finding the referenced symbol itself. however, in the case
of copy relocations, the referenced symbol itself is ignored in order
to search for another symbol to copy from, and thus it's possible that
no definition is found. in this case, if the symbol being resolved
happened to be a weak definition, it was misinterpreted as a weak
reference, suppressing the error path and causing a crash when the
copy relocation was performed with a null source pointer passed to
memcpy.
there are almost certainly still situations in which invalid
combinations of symbol and relocation types can cause the dynamic
linker to crash (this is pretty much inevitable), but the intent is
that crashes not be possible for symbol/relocation tables produced by
a valid linker.
install.sh was wrongly waiting until after atomically replacing the
old file to set the correct permissions on the new file. in the case
of the dynamic linker, this would cause a dynamic-linked chmod command
not to run (due to missing executable permissions on the dynamic
linker) and thus leave the system in an unusable state.
even if chmod is static-linked, the old behavior had a race window
where dynamic-linked programs could fail to run.
the operand size is unnecessary, since the assembler knows it from the
destination register size. removing the suffix makes it so the same
code should work for x32.
in fixing this, I've changed the logic from ugly #if/#else blocks
inside the struct shm_info definition to a fixed struct definition and
optional macros to rename the elements. this will be helpful if we
need to move shm_info to a bits header in the future, as it will keep
the feature test logic out of bits.
the fix should be complete on archs that use the generic definitions
(i386, arm, x86_64, microblaze), but mips and powerpc have not been
checked thoroughly and may need more fixes.
the imr_, imsf_, ip6_, ip6m_, ipi_, ipi6_, SCM_, and SOL_ prefixes are
not in the reserved namespace for this header. thus the constants and
structures using them need to be protected under appropriate feature
test macros.
this also affects some headers which are permitted to include
netinet/in.h, particularly netdb.h and arpa/inet.h.
the SOL_ macros are moved to sys/socket.h where they are in the
reserved namespace (SO*). they are still accessible via netinet/in.h
since it includes sys/socket.h implicitly (which is permitted).
the SCM_SRCRT macro is simply removed, since the definition used for
it, IPV6_RXSRCRT is not defined anywhere. it could be re-added, this
time in sys/socket.h, if the appropriate value can be determined;
however, given that the erroneous definition was not caught, it is
unlikely that any software actually attempts to use SCM_SRCRT.
per POSIX, the variadic argument has type union semun, which may
contain a pointer or int; the type read depends on the command being
issued. this allows the userspace part of the implementation to be
type-correct without requiring special-casing for different commands.
the kernel always expects to receive the argument interpreted as
unsigned long (or equivalently, a pointer), and does its own handling
of extracting the int portion from the representation, as needed.
this change fixes two possible issues: most immediately, reading the
argument as a (signed) long and passing it to the syscall would
perform incorrect sign-extension of pointers on the upcoming x32
target. the other possible issue is that some archs may use different
(user-space) argument-passing convention for unions, preventing va_arg
from correctly obtaining the argument when the type long (or even
unsigned long or void *) is passed to it.
really, fcntl should be changed to use the correct type corresponding
to cmd when calling va_arg, and to carry the correct type through
until making the syscall. however, this greatly increases binary size
and does not seem to offer any benefits except formal correctness, so
I'm holding off on that change for now.
the minimal changes made in this patch are in preparation for addition
of the x32 port, where the syscall macros need to know whether their
arguments are pointers or integers in order to properly pass them to
the 64-bit kernel.
it's unclear what the historical signature for this function was, but
semantically, the argument should be a pointer to const, and this is
what glibc uses. correct programs should not be using this function
anyway, so it's unlikely to matter.
this change is consistent with the corresponding glibc functions and
is semantically const-correct. the incorrect argument types without
const seem to have been taken from erroneous man pages.
this functionality has essentially always been deprecated in linux,
and was never supported by musl. the presence of the header was
reported to cause some software to attempt to use the nonexistant
function, so removing the header is the cleanest solution.
this was wrong since the original commit adding inotify, and I don't
see any explanation for it. not even the man pages have it wrong. it
was most likely a copy-and-paste error.
this practice came from very early, before internal/syscall.h defined
macros that could accept pointer arguments directly and handle them
correctly. aside from being ugly and unnecessary, it looks like it
will be problematic when we add support for 32-bit ABIs on archs where
registers (and syscall arguments) are 64-bit, e.g. x32 and mips n32.
this agrees with implementation practice on glibc and BSD systems, and
is the const-correct way to do things; it eliminates warnings from
passing pointers to const. the prototype without const came from
seemingly erroneous man pages.
the reasons are the same as for sbrk. unlike sbrk, there is no safe
usage because brk does not return any useful information, so it should
just fail unconditionally.
use of sbrk is never safe; it conflicts with malloc, and malloc may be
used internally by the implementation basically anywhere. prior to
this change, applications attempting to use sbrk to do their own heap
management simply caused untrackable memory corruption; now, they will
fail with ENOMEM allowing the errors to be fixed.
sbrk(0) is still permitted as a way to get the current brk; some
misguided applications use this as a measurement of their memory
usage or for other related purposes, and such usage is harmless.
eventually sbrk may be re-added if/when malloc is changed to avoid
using the brk by using mmap for all allocations.
the workaround/fallback code for supporting O_PATH file descriptors
when the kernel lacks support for performing these operations on them
caused EBADF to get replaced by ENOENT (due to missing entry in
/proc/self/fd). this is unlikely to affect real-world code (calls that
might yield EBADF are generally unsafe, especially in library code)
but it was breaking some test cases.
the fix I've applied is something of a tradeoff: it adds one syscall
to these operations on kernels where the workaround is needed. the
alternative would be to catch ENOENT from the /proc lookup and
translate it to EBADF, but I want to avoid doing that in the interest
of not touching/depending on /proc at all in these functions as long
as the kernel correctly supports the operations. this is following the
general principle of isolating hacks to code paths that are taken on
broken systems, and keeping the code for correct systems completely
hack-free.
the ABI allows the callee to clobber stack slots that correspond to
arguments passed in registers, so the caller must adjust the stack
pointer to reserve space appropriately. prior to this fix, the argv
array was possibly clobbered by dynamic linker code before passing
control to the main program.
our getcwd already (as an extension) supports allocation of a buffer
when the buffer argument is a null pointer, so there's no need to
duplicate the allocation logic in this wrapper function. duplicating
it is actually harmful in that it doubles the stack usage from
PATH_MAX to 2*PATH_MAX.
the wildcard function in GNU make includes dangling symlinks; if any
exist under the .git directory, they would get added as dependencies,
causing make to exit with an error due to lacking a rule to build the
missing file.
as far as I can tell, git operations which should force version.h to
be rebuilt must all touch the mtime of the top-level .git directory.
previously this flag was defined and accepted as a no-op, possibly
breaking some software that uses it. given the choice to remove the
definition and possibly break applications that were already working,
or simply implement the feature, the latter turned out to be easy
enough to make the decision easy.
in the case where the FNM_PATHNAME flag is also set, this
implementation is clean and essentially optimal. otherwise, it's an
inefficient "brute force" implementation. at some point, when cleaning
up and refactoring this code, I may add a more direct code path for
handling FNM_LEADING_DIR in the non-FNM_PATHNAME case, but at this
point my main interest is avoiding introducing new bugs in the code
that implements the standard fnmatch features specified by POSIX.
this is still experimental and subject to change. for git checkouts,
an attempt is made to record the exact revision to aid in bug reports
and debugging. no version information is recorded in the static libc.a
or binaries it's linked into.
the FNM_PATHNAME logic for advancing by /-delimited components was
incorrect when the / character was escaped (i.e. \/), and a final \ at
the end of pattern was not handled correctly.
a v6 socket will only be used if there is at least one v6 nameserver
address. if the kernel lacks v6 support, the code will fall back to
using a v4 socket and requests to v6 servers will silently fail. when
using a v6 socket, v4 addresses are converted to v4-mapped form and
setsockopt is used to ensure that the v6 socket can accept both v4 and
v6 traffic (this is on-by-default on Linux but the default is
configurable in /proc and so it needs to be set explicitly on the
socket level). this scheme avoids increasing resource usage during
lookups and allows the existing network io loop to be used without
modification.
previously, nameservers whose address family did not match the address
family of the first-listed nameserver were simply ignored. prior to
recent __ipparse fixes, they were not ignored but erroneously parsed.
the old value of 20 was reported by Laurent Bercot as being
insufficient for a reasonable real-world usage case. actual problem
was the internal buffer used by ttyname(), but the implementation of
ttyname uses TTY_NAME_MAX, and for consistency it's best to increase
both. the new value is aligned with glibc.
subsequent code assumes the address family requested is either
unspecified or one of IPv4/IPv6, and could malfunction if this
constraint is not met, so other address families should be explicitly
rejected.
on archs with excess precision, the floating point constant 1e40f may
be evaluated such that it does not actually produce an infinity.
1e5000f is sufficiently large to produce an infinity for all supported
floating point formats. note that this definition of INFINITY is only
used for old or non-GNUC compilers anyway; despite being a portable,
conforming definition, it leads to erroneous warnings on many
compilers and thus using the builtin is preferred.
Rich Felker