oweals/musl.git
13 years agorun pthread tsd destructors when a timer thread pretends to exit
Rich Felker [Sat, 9 Apr 2011 06:26:55 +0000 (02:26 -0400)]
run pthread tsd destructors when a timer thread pretends to exit

13 years agogreatly improve SIGEV_THREAD timers
Rich Felker [Sat, 9 Apr 2011 06:23:33 +0000 (02:23 -0400)]
greatly improve SIGEV_THREAD timers

calling pthread_exit from, or pthread_cancel on, the timer callback
thread will no longer destroy the timer.

13 years agoprepare notes for 0.7.8 release
Rich Felker [Sat, 9 Apr 2011 05:32:38 +0000 (01:32 -0400)]
prepare notes for 0.7.8 release

13 years agowork around a nasty bug in linux readv syscall
Rich Felker [Sat, 9 Apr 2011 05:17:55 +0000 (01:17 -0400)]
work around a nasty bug in linux readv syscall

according to posix, readv "shall be equivalent to read(), except..."
that it places the data into the buffers specified by the iov array.
however on linux, when reading from a terminal, each iov element
behaves almost like a separate read. this means that if the first iov
exactly satisfied the request (e.g. a length-one read of '\n') and the
second iov is nonzero length, the syscall will block again after
getting the blank line from the terminal until another line is read.
simply put, entering a single blank line becomes impossible.

the solution, fortunately, is simple. whenever the buffer size is
nonzero, reduce the length of the requested read by one byte and let
the last byte go through the buffer. this way, readv will already be
in the second (and last) iov, and won't re-block on the second iov.

13 years agobetter fix sysconf pthread stack min
Rich Felker [Fri, 8 Apr 2011 16:16:24 +0000 (12:16 -0400)]
better fix sysconf pthread stack min

13 years agoconsistency with pthread stack min in limits.h
Rich Felker [Fri, 8 Apr 2011 16:15:37 +0000 (12:15 -0400)]
consistency with pthread stack min in limits.h

13 years agofix broken sigsetjmp on x86_64
Rich Felker [Fri, 8 Apr 2011 15:56:52 +0000 (11:56 -0400)]
fix broken sigsetjmp on x86_64

13 years agoworkaround broken msghdr struct on 64bit linux
Rich Felker [Fri, 8 Apr 2011 13:24:19 +0000 (09:24 -0400)]
workaround broken msghdr struct on 64bit linux

POSIX clearly specifies the type of msg_iovlen and msg_controllen, and
Linux ignores it and makes them both size_t instead. to work around
this we add padding (instead of just using the wrong types like glibc
does), but we also need to patch-up the struct before passing it to
the kernel in case the caller did not zero-fill it.

if i could trust the kernel to just ignore the upper 32 bits, this
would not be necessary, but i don't think it will ignore them...

13 years agofix ipv6 address printing: 2001 appeared as 201, etc.
Rich Felker [Fri, 8 Apr 2011 13:21:51 +0000 (09:21 -0400)]
fix ipv6 address printing: 2001 appeared as 201, etc.

13 years agofix broken dns response parsing code that made most ipv6 lookups fail
Rich Felker [Fri, 8 Apr 2011 12:49:28 +0000 (08:49 -0400)]
fix broken dns response parsing code that made most ipv6 lookups fail

13 years agoreturn the requested string as the "canonical name" for numeric addresses
Rich Felker [Fri, 8 Apr 2011 12:14:28 +0000 (08:14 -0400)]
return the requested string as the "canonical name" for numeric addresses

previously NULL was returned in ai_canonname, resulting in crashes in
some callers. this behavior was incorrect. note however that the new
behavior differs from glibc, which performs reverse dns lookups. POSIX
is very clear that a reverse DNS lookup must not be performed for
numeric addresses.

13 years agofix uninitialized variables in dns lookup code
Rich Felker [Fri, 8 Apr 2011 03:18:12 +0000 (23:18 -0400)]
fix uninitialized variables in dns lookup code

13 years agofix bug in TRE found by clang (typo && instead of &)
Rich Felker [Fri, 8 Apr 2011 03:13:47 +0000 (23:13 -0400)]
fix bug in TRE found by clang (typo && instead of &)

13 years agofix misplaced *'s in string functions (harmless)
Rich Felker [Thu, 7 Apr 2011 20:19:30 +0000 (16:19 -0400)]
fix misplaced *'s in string functions (harmless)

13 years agofix broken unsigned comparison in wcstoumax
Rich Felker [Thu, 7 Apr 2011 20:13:47 +0000 (16:13 -0400)]
fix broken unsigned comparison in wcstoumax

13 years agofix breakage due to converting a return type to size_t in iconv...
Rich Felker [Thu, 7 Apr 2011 20:10:44 +0000 (16:10 -0400)]
fix breakage due to converting a return type to size_t in iconv...

13 years agofixed crash in new rsyscall (failure to set sa_flags for signal handler)
Rich Felker [Thu, 7 Apr 2011 00:43:39 +0000 (20:43 -0400)]
fixed crash in new rsyscall (failure to set sa_flags for signal handler)

13 years agoconsistency: change all remaining syscalls to use SYS_ rather than __NR_ prefix
Rich Felker [Thu, 7 Apr 2011 00:32:53 +0000 (20:32 -0400)]
consistency: change all remaining syscalls to use SYS_ rather than __NR_ prefix

13 years agomove rsyscall out of pthread_create module
Rich Felker [Thu, 7 Apr 2011 00:27:07 +0000 (20:27 -0400)]
move rsyscall out of pthread_create module

this is something of a tradeoff, as now set*id() functions, rather
than pthread_create, are what pull in the code overhead for dealing
with linux's refusal to implement proper POSIX thread-vs-process
semantics. my motivations are:

1. it's cleaner this way, especially cleaner to optimize out the
rsyscall locking overhead from pthread_create when it's not needed.
2. it's expected that only a tiny number of core system programs will
ever use set*id() functions, whereas many programs may want to use
threads, and making thread overhead tiny is an incentive for "light"
programs to try threads.

13 years agopthread exit stuff: don't bother setting errno when we won't check it.
Rich Felker [Wed, 6 Apr 2011 23:47:50 +0000 (19:47 -0400)]
pthread exit stuff: don't bother setting errno when we won't check it.

13 years agofix rsyscall handler: must not clobber errno from signal context
Rich Felker [Wed, 6 Apr 2011 23:46:46 +0000 (19:46 -0400)]
fix rsyscall handler: must not clobber errno from signal context

13 years agofix typo in sys/msg.h
Rich Felker [Wed, 6 Apr 2011 21:50:38 +0000 (17:50 -0400)]
fix typo in sys/msg.h

13 years agoadd startup abi functions, dummy for now. eventually needed for c++ support.
Rich Felker [Wed, 6 Apr 2011 20:40:19 +0000 (16:40 -0400)]
add startup abi functions, dummy for now. eventually needed for c++ support.

13 years agoadd _res (__res_state()) dummy
Rich Felker [Wed, 6 Apr 2011 19:47:26 +0000 (15:47 -0400)]
add _res (__res_state()) dummy

13 years agoadd IN_LOOPBACKNET constant (nonstandard but in reserved namespace)
Rich Felker [Wed, 6 Apr 2011 19:44:39 +0000 (15:44 -0400)]
add IN_LOOPBACKNET constant (nonstandard but in reserved namespace)

13 years agodocument more changes for 0.7.7 v0.7.7
Rich Felker [Wed, 6 Apr 2011 18:46:37 +0000 (14:46 -0400)]
document more changes for 0.7.7

13 years agofix prototype for strsep
Rich Felker [Wed, 6 Apr 2011 18:28:29 +0000 (14:28 -0400)]
fix prototype for strsep

13 years agofix completely bogus loop condition in getmntent_r
Rich Felker [Wed, 6 Apr 2011 16:35:05 +0000 (12:35 -0400)]
fix completely bogus loop condition in getmntent_r

somehow this worked on my simple fstab, but horribly broke in general,
leading to use of uninitialized offset array and crashes.

13 years agomajor semaphore improvements (performance and correctness)
Rich Felker [Wed, 6 Apr 2011 16:24:34 +0000 (12:24 -0400)]
major semaphore improvements (performance and correctness)

1. make sem_[timed]wait interruptible by signals, per POSIX
2. keep a waiter count in order to avoid unnecessary futex wake syscalls

13 years agofix signal-based timers with null sigevent argument
Rich Felker [Wed, 6 Apr 2011 13:26:41 +0000 (09:26 -0400)]
fix signal-based timers with null sigevent argument

since timer_create is no longer allocating a structure for the timer_t
and simply using the kernel timer id, it was impossible to specify the
timer_t as the argument to the signal handler. the solution is to pass
the null sigevent pointer on to the kernel, rather than filling it in
userspace, so that the kernel does the right thing. however, that
precludes the clever timerid-versus-threadid encoding we were doing.

instead, just assume timerids are below 1M and thread pointers are
above 1M. (in perspective: timerids are sequentially allocated and
seem limited to 32k, and thread pointers are at roughly 3G.)

13 years agofix incorrect (and conflicting on LP64 archs) types for sysv ipc msgq functions
Rich Felker [Wed, 6 Apr 2011 04:02:20 +0000 (00:02 -0400)]
fix incorrect (and conflicting on LP64 archs) types for sysv ipc msgq functions

13 years agofix (hopefully) statvfs breakage on x86_64 that resulted from fixing i386...
Rich Felker [Wed, 6 Apr 2011 03:58:36 +0000 (23:58 -0400)]
fix (hopefully) statvfs breakage on x86_64 that resulted from fixing i386...

13 years agodocument more changes
Rich Felker [Tue, 5 Apr 2011 22:58:47 +0000 (18:58 -0400)]
document more changes

13 years agonew framework to inhibit thread cancellation when needed
Rich Felker [Tue, 5 Apr 2011 22:00:28 +0000 (18:00 -0400)]
new framework to inhibit thread cancellation when needed

with these small changes, libc functions which need to call functions
which are cancellation points, but which themselves must not be
cancellation points, can use the CANCELPT_INHIBIT and CANCELPT_RESUME
macros to temporarily inhibit all cancellation.

13 years agoadd ip6 pktinfo stuff for x86_64
Rich Felker [Tue, 5 Apr 2011 21:27:28 +0000 (17:27 -0400)]
add ip6 pktinfo stuff for x86_64

these defs should probably all be moved out of bits and unified...

13 years agoadd sysv ipc message queues (completely untested)
Rich Felker [Tue, 5 Apr 2011 20:33:55 +0000 (16:33 -0400)]
add sysv ipc message queues (completely untested)

13 years agogetopt.h is a GNU-getopt-specific header. always expose GNU functions.
Rich Felker [Tue, 5 Apr 2011 19:45:24 +0000 (15:45 -0400)]
getopt.h is a GNU-getopt-specific header. always expose GNU functions.

13 years agoadd sysexits.h legacy header
Rich Felker [Tue, 5 Apr 2011 19:40:36 +0000 (15:40 -0400)]
add sysexits.h legacy header

13 years agoimplement the adjtime and adjtimex functions (nonstandard)
Rich Felker [Tue, 5 Apr 2011 19:38:20 +0000 (15:38 -0400)]
implement the adjtime and adjtimex functions (nonstandard)

13 years agoadd getmntent_r interface (all of mntent is nonstandard anyway)
Rich Felker [Tue, 5 Apr 2011 18:11:14 +0000 (14:11 -0400)]
add getmntent_r interface (all of mntent is nonstandard anyway)

13 years agoadd some missing ipv6 stuff
Rich Felker [Tue, 5 Apr 2011 18:05:11 +0000 (14:05 -0400)]
add some missing ipv6 stuff

13 years agoimplement if_indextoname and if_nametoindex functions
Rich Felker [Tue, 5 Apr 2011 17:20:08 +0000 (13:20 -0400)]
implement if_indextoname and if_nametoindex functions

13 years agoadd (nonstandard) cfmakeraw function
Rich Felker [Tue, 5 Apr 2011 16:35:09 +0000 (12:35 -0400)]
add (nonstandard) cfmakeraw function

13 years agoadd pivot_root syscall wrapper
Rich Felker [Tue, 5 Apr 2011 16:32:10 +0000 (12:32 -0400)]
add pivot_root syscall wrapper

13 years agoadd more legacy functions: setlinebuf and setbuffer
Rich Felker [Tue, 5 Apr 2011 16:25:31 +0000 (12:25 -0400)]
add more legacy functions: setlinebuf and setbuffer

13 years agoalternate name for syslog.h
Rich Felker [Tue, 5 Apr 2011 16:13:54 +0000 (12:13 -0400)]
alternate name for syslog.h

13 years agouncomment IP_PKTINFO
Rich Felker [Tue, 5 Apr 2011 16:07:42 +0000 (12:07 -0400)]
uncomment IP_PKTINFO

this was a hack leftover from testing before the initial
check-in to git.

13 years agosupport the nonstandard err.h interfaces
Rich Felker [Tue, 5 Apr 2011 15:16:49 +0000 (11:16 -0400)]
support the nonstandard err.h interfaces

note that unlike the originals, these do not print the program
name/argv[0] because we have not saved it anywhere. this could be
changed in __libc_start_main if desired.

13 years agorelease notes for next bugfix release
Rich Felker [Tue, 5 Apr 2011 14:28:08 +0000 (10:28 -0400)]
release notes for next bugfix release

13 years agofix misaligned read on early string termination in strchr
Rich Felker [Tue, 5 Apr 2011 13:27:41 +0000 (09:27 -0400)]
fix misaligned read on early string termination in strchr

this could actually cause rare crashes in the case where a short
string is located at the end of a page and the following page is not
readable, and in fact this was seen in gcc compiling certain files.

13 years agofix overflow in printf %N$ argument handling
Rich Felker [Tue, 5 Apr 2011 13:24:03 +0000 (09:24 -0400)]
fix overflow in printf %N$ argument handling

13 years agofix various floating point rounding and formatting errors in *printf
Rich Felker [Tue, 5 Apr 2011 13:16:40 +0000 (09:16 -0400)]
fix various floating point rounding and formatting errors in *printf

13 years agorelease notes for 0.7.6 bugfix release v0.7.6
Rich Felker [Mon, 4 Apr 2011 22:00:02 +0000 (18:00 -0400)]
release notes for 0.7.6 bugfix release

13 years agofix rare but nasty under-allocation bug in malloc with large requests
Rich Felker [Mon, 4 Apr 2011 21:26:41 +0000 (17:26 -0400)]
fix rare but nasty under-allocation bug in malloc with large requests

the bug appeared only with requests roughly 2*sizeof(size_t) to
4*sizeof(size_t) bytes smaller than a multiple of the page size, and
only for requests large enough to be serviced by mmap instead of the
normal heap. it was only ever observed on 64-bit machines but
presumably could also affect 32-bit (albeit with a smaller window of
opportunity).

13 years agosimplify vdprintf implementation greatly based on recent vfprintf changes
Rich Felker [Mon, 4 Apr 2011 20:30:39 +0000 (16:30 -0400)]
simplify vdprintf implementation greatly based on recent vfprintf changes

since vfprintf will provide a temporary buffer in the case where the
target FILE has a zero buffer size, don't bother setting up a real
buffer for vdprintf. this also allows us to skip the call to fflush
since we know everything will be written out before vfprintf returns.

13 years agouse a local temp buffer for unbuffered streams in vfprintf
Rich Felker [Mon, 4 Apr 2011 20:24:49 +0000 (16:24 -0400)]
use a local temp buffer for unbuffered streams in vfprintf

this change makes it so most calls to fprintf(stderr, ...) will result
in a single writev syscall, as opposed to roughly 2*N syscalls (and
possibly more) where N is the number of format specifiers. in
principle we could use a much larger buffer, but it's best not to
increase the stack requirements too much. most messages are under 80
chars.

13 years agofix nl_langinfo to actually use the existing, correct internal version
Rich Felker [Sun, 3 Apr 2011 23:51:14 +0000 (19:51 -0400)]
fix nl_langinfo to actually use the existing, correct internal version

13 years agomake ualarm actually work (obsolete function removed from SUS)
Rich Felker [Sun, 3 Apr 2011 23:18:43 +0000 (19:18 -0400)]
make ualarm actually work (obsolete function removed from SUS)

13 years agofix various bugs in strtold:
Rich Felker [Sun, 3 Apr 2011 22:44:37 +0000 (18:44 -0400)]
fix various bugs in strtold:

0e10000000000000000000000000000000 was setting ERANGE

exponent char e/p was considered part of the match even if not
followed by a valid decimal value

"1e +10" was parsed as "1e+10"

hex digits were misinterpreted as 0..5 instead of 10..15

13 years agofix serious bug in strchr - char signedness
Rich Felker [Sun, 3 Apr 2011 22:16:11 +0000 (18:16 -0400)]
fix serious bug in strchr - char signedness

search for bytes with high bit set was giving (potentially dangerous)
wrong results. i've tested, cleaned up, and hopefully sped up this
function now.

13 years agoprototype getdtablesize (nonstandard function)
Rich Felker [Sun, 3 Apr 2011 22:15:36 +0000 (18:15 -0400)]
prototype getdtablesize (nonstandard function)

13 years agoadd setresuid/setresgid functions (nonstandard)
Rich Felker [Sun, 3 Apr 2011 20:20:57 +0000 (16:20 -0400)]
add setresuid/setresgid functions (nonstandard)

13 years agopthread_create need not set errno
Rich Felker [Sun, 3 Apr 2011 20:15:15 +0000 (16:15 -0400)]
pthread_create need not set errno

13 years agofix statvfs syscalls (missing size argument)
Rich Felker [Sun, 3 Apr 2011 19:42:31 +0000 (15:42 -0400)]
fix statvfs syscalls (missing size argument)

13 years agoblock all signals during rsyscall
Rich Felker [Sun, 3 Apr 2011 17:15:42 +0000 (13:15 -0400)]
block all signals during rsyscall

otherwise a signal handler could see an inconsistent and nonconformant
program state where different threads have different uids/gids.

13 years agofix race condition in rsyscall handler
Rich Felker [Sun, 3 Apr 2011 17:03:18 +0000 (13:03 -0400)]
fix race condition in rsyscall handler

the problem: there is a (single-instruction) race condition window
between a thread flagging itself dead and decrementing itself from the
thread count. if it receives the rsyscall signal at this exact moment,
the rsyscall caller will never succeed in signalling enough flags to
succeed, and will deadlock forever. in previous versions of musl, the
about-to-terminate thread masked all signals prior to decrementing
the thread count, but this cost a whole syscall just to account for
extremely rare races.

the solution is a huge hack: rather than blocking in the signal
handler if the thread is dead, modify the signal mask of the saved
context and return in order to prevent further signal handling by the
dead thread. this allows the dead thread to continue decrementing the
thread count (if it had not yet done so) and exiting, even while the
live part of the program blocks for rsyscall.

13 years agodon't trust siginfo in rsyscall handler
Rich Felker [Sun, 3 Apr 2011 16:20:51 +0000 (12:20 -0400)]
don't trust siginfo in rsyscall handler

for some inexplicable reason, linux allows the sender of realtime
signals to spoof its identity. permission checks for sending signals
should limit the impact to same-user processes, but just to be safe,
we avoid trusting the siginfo structure and instead simply examine the
program state to see if we're in the middle of a legitimate rsyscall.

13 years agotimer threads should sleep and stay asleep... a long time
Rich Felker [Sun, 3 Apr 2011 16:10:24 +0000 (12:10 -0400)]
timer threads should sleep and stay asleep... a long time

13 years agorevert to deleting kernel-level timer from cancellation handler
Rich Felker [Sun, 3 Apr 2011 16:08:34 +0000 (12:08 -0400)]
revert to deleting kernel-level timer from cancellation handler

this is necessary in order to avoid breaking timer_getoverrun in the
last run of the timer event handler, if it has not yet finished.

13 years agosimplify calling of timer signal handler
Rich Felker [Sun, 3 Apr 2011 16:03:58 +0000 (12:03 -0400)]
simplify calling of timer signal handler

13 years agod_fileno alias for d_ino in dirent
Rich Felker [Sun, 3 Apr 2011 14:24:59 +0000 (10:24 -0400)]
d_fileno alias for d_ino in dirent

this is nonstandard but since POSIX reserved d_ prefix in dirent.h we
might as well define it unconditionally. some programs depend on it.

13 years agosimplify pthread tsd key handling
Rich Felker [Sun, 3 Apr 2011 06:40:18 +0000 (02:40 -0400)]
simplify pthread tsd key handling

13 years agoomit pthread tsd dtor code if tsd is not used
Rich Felker [Sun, 3 Apr 2011 06:33:50 +0000 (02:33 -0400)]
omit pthread tsd dtor code if tsd is not used

13 years agodon't disable seeking after first seek failure
Rich Felker [Sat, 2 Apr 2011 17:55:54 +0000 (13:55 -0400)]
don't disable seeking after first seek failure

this could cause problems if the application uses dup2(fd,fileno(f))
to redirect, and the old fd was not seekable but the new fd is.

13 years agoapparently fseek should not set the error flag on failed seek
Rich Felker [Sat, 2 Apr 2011 17:54:55 +0000 (13:54 -0400)]
apparently fseek should not set the error flag on failed seek

13 years agoupdate release notes v0.7.5
Rich Felker [Sat, 2 Apr 2011 03:15:29 +0000 (23:15 -0400)]
update release notes

13 years agoavoid over-allocation of brk on first malloc
Rich Felker [Sat, 2 Apr 2011 03:07:03 +0000 (23:07 -0400)]
avoid over-allocation of brk on first malloc

if init_malloc returns positive (successful first init), malloc will
retry getting a chunk from the free bins rather than expanding the
heap again. also pass init_malloc a hint for the size of the initial
allocation.

13 years agoreorganize the __libc structure for threaded performance issues
Rich Felker [Sat, 2 Apr 2011 02:35:20 +0000 (22:35 -0400)]
reorganize the __libc structure for threaded performance issues

we want to keep atomically updated fields (locks and thread count) and
really anything writable far away from frequently-needed function
pointers. stuff some rarely-needed function pointers in between to
pad, hopefully up to a cache line boundary.

13 years agosimplify setting result on thread cancellation
Rich Felker [Sat, 2 Apr 2011 02:15:03 +0000 (22:15 -0400)]
simplify setting result on thread cancellation

13 years agouse bss instead of mmap for main thread's pthread thread-specific data
Rich Felker [Sat, 2 Apr 2011 02:07:59 +0000 (22:07 -0400)]
use bss instead of mmap for main thread's pthread thread-specific data

this simplifies code and removes a failure case

13 years agoremove obsolete and useless useconds_t type
Rich Felker [Sat, 2 Apr 2011 01:10:01 +0000 (21:10 -0400)]
remove obsolete and useless useconds_t type

13 years agosomehow timespec tv_nsec had the wrong type on x86_64... fixed
Rich Felker [Sat, 2 Apr 2011 00:58:40 +0000 (20:58 -0400)]
somehow timespec tv_nsec had the wrong type on x86_64... fixed

13 years agofix misspelled PTHREAD_CANCELED constant
Rich Felker [Sat, 2 Apr 2011 00:48:02 +0000 (20:48 -0400)]
fix misspelled PTHREAD_CANCELED constant

13 years agodocument more changes
Rich Felker [Sat, 2 Apr 2011 00:47:54 +0000 (20:47 -0400)]
document more changes

13 years agodocument changes for upcoming 0.7.5 release
Rich Felker [Sat, 2 Apr 2011 00:36:01 +0000 (20:36 -0400)]
document changes for upcoming 0.7.5 release

13 years agouse a_store to set cancel flag in pthread_cancel, to ensure a barrier
Rich Felker [Fri, 1 Apr 2011 23:53:16 +0000 (19:53 -0400)]
use a_store to set cancel flag in pthread_cancel, to ensure a barrier

13 years agosimplify pthread_key_delete
Rich Felker [Thu, 31 Mar 2011 23:06:22 +0000 (19:06 -0400)]
simplify pthread_key_delete

calling this function on an uninitialized key value is UB, so there is
no need to check that the table pointer was initialized.

13 years agogreatly simplify pthread_key_create (~20% size reduction)
Rich Felker [Thu, 31 Mar 2011 23:04:56 +0000 (19:04 -0400)]
greatly simplify pthread_key_create (~20% size reduction)

13 years agoadd some missing prototypes for nonstandard functions (strsep, clearenv)
Rich Felker [Wed, 30 Mar 2011 18:14:26 +0000 (14:14 -0400)]
add some missing prototypes for nonstandard functions (strsep, clearenv)

13 years agoavoid all malloc/free in timer creation/destruction
Rich Felker [Wed, 30 Mar 2011 17:04:55 +0000 (13:04 -0400)]
avoid all malloc/free in timer creation/destruction

instead of allocating a userspace structure for signal-based timers,
simply use the kernel timer id. we use the fact that thread pointers
will always be zero in the low bit (actually more) to encode integer
timerid values as pointers.

also, this change ensures that the timer_destroy syscall has completed
before the library timer_destroy function returns, in case it matters.

13 years agooptimize timer creation and possibly protect against some minor races
Rich Felker [Wed, 30 Mar 2011 16:06:39 +0000 (12:06 -0400)]
optimize timer creation and possibly protect against some minor races

the major idea of this patch is not to depend on having the timer
pointer delivered to the signal handler, and instead use the thread
pointer to get the callback function address and argument. this way,
the parent thread can make the timer_create syscall while the child
thread is starting, and it should never have to block waiting for the
barrier.

13 years agoavoid crash on stupid but allowable usage of pthread_mutex_unlock
Rich Felker [Wed, 30 Mar 2011 14:32:45 +0000 (10:32 -0400)]
avoid crash on stupid but allowable usage of pthread_mutex_unlock

unlocking an unlocked mutex is not UB for robust or error-checking
mutexes, so we must avoid calling __pthread_self (which might crash
due to lack of thread-register initialization) until after checking
that the mutex is locked.

13 years agorename __simple_malloc.c to lite_malloc.c - yes this affects behavior!
Rich Felker [Wed, 30 Mar 2011 13:29:49 +0000 (09:29 -0400)]
rename __simple_malloc.c to lite_malloc.c - yes this affects behavior!

why does this affect behavior? well, the linker seems to traverse
archive files starting from its current position when resolving
symbols. since calloc.c comes alphabetically (and thus in sequence in
the archive file) between __simple_malloc.c and malloc.c, attempts to
resolve the "malloc" symbol for use by calloc.c were pulling in the
full malloc.c implementation rather than the __simple_malloc.c
implementation.

as of now, lite_malloc.c and malloc.c are adjacent in the archive and
in the correct order, so malloc.c should never be used to resolve
"malloc" unless it's already needed to resolve another symbol ("free"
or "realloc").

13 years agostreamline mutex unlock to remove a useless branch, use a_store to unlock
Rich Felker [Wed, 30 Mar 2011 13:06:00 +0000 (09:06 -0400)]
streamline mutex unlock to remove a useless branch, use a_store to unlock

this roughly halves the cost of pthread_mutex_unlock, at least for
non-robust, normal-type mutexes.

the a_store change is in preparation for future support of archs which
require a memory barrier or special atomic store operation, and also
should prevent the possibility of the compiler misordering writes.

13 years agocheap special-case optimization for normal mutexes
Rich Felker [Wed, 30 Mar 2011 12:58:25 +0000 (08:58 -0400)]
cheap special-case optimization for normal mutexes

cycle-level benchmark on atom cpu showed typical pthread_mutex_lock
call dropping from ~120 cycles to ~90 cycles with this change. benefit
may vary with compiler options and version, but this optimization is
very cheap to make and should always help some.

13 years agoreorder timer initialization so that timer_create does not depend on free
Rich Felker [Wed, 30 Mar 2011 02:43:13 +0000 (22:43 -0400)]
reorder timer initialization so that timer_create does not depend on free

this allows small programs which only create times, but never delete
them, to use simple_malloc instead of the full malloc.

13 years agomissing prototype for wcscoll (stub)
Rich Felker [Tue, 29 Mar 2011 22:30:27 +0000 (18:30 -0400)]
missing prototype for wcscoll (stub)

13 years agorevert mutex "optimization" that turned out to be worse
Rich Felker [Tue, 29 Mar 2011 19:11:25 +0000 (15:11 -0400)]
revert mutex "optimization" that turned out to be worse

13 years agoimplement POSIX timers
Rich Felker [Tue, 29 Mar 2011 17:01:25 +0000 (13:01 -0400)]
implement POSIX timers

this implementation is superior to the glibc/nptl implementation, in
that it gives true realtime behavior. there is no risk of timer
expiration events being lost due to failed thread creation or failed
malloc, because the thread is created as time creation time, and
reused until the timer is deleted.

13 years agomajor improvements to cancellation handling
Rich Felker [Tue, 29 Mar 2011 16:58:22 +0000 (12:58 -0400)]
major improvements to cancellation handling

- there is no longer any risk of spoofing cancellation requests, since
  the cancel flag is set in pthread_cancel rather than in the signal
  handler.

- cancellation signal is no longer unblocked when running the
  cancellation handlers. instead, pthread_create will cause any new
  threads created from a cancellation handler to unblock their own
  cancellation signal.

- various tweaks in preparation for POSIX timer support.