-falign-functions=32, since these two really get hammered on. To make them
faster needs a threadreg or TLS, unless there is a way to tell gcc that a
library-local (pthread__threadmask) variable does not need to be PIC.
preparing to sleep on a mutex are slow in relative terms, so this allows
us to recover from short lock holds without blocking, while not wasting
too much time on longer holds.
architecture to provide asm versions of the RAS operations.
We do this because relying on the compiler to get the RAS right is not
sensible. (It gets alpha wrong and hppa is suboptimal)
Provide asm RAS ops for hppa.
(A slightly different version) reviewed by Andrew Doran.
- Allow callers to try and release an unheld rwlock. Just return EPERM
as mandated by IEEE Std 1003.1.
- Use pthread__atomic_swap_ptr() to set in the new lock value. At this
point the lock word can't have changed.
pthread__rwlock_wrlock, pthread__rwlock_rdlock:
- Mask out the waiter bits in the lock word before checking to see if
the current thread is about to lock against itself.
the following do not wake other threads early:
pthread_mutex_lock(&mutex);
pthread_cond_broadcast(&cond);
foo = malloc(100); /* takes libc mutexes */
pthread_mutex_unlock(&mutex);
- Eliminate mutexattr_private and just set a bit in ptm_owner if the mutex
is recursive. This forces the slow path to be taken for recursive mutexes.
Overload an unused field in pthread_mutex_t to record whether or not it's
an errorcheck mutex.
- Streamline pthread_mutex_lock / pthread_mutex_unlock a bit more. As a
side effect makes it possible to have assembly stubs for them.
- Update some comments and fix minor bugs. Minor cosmetic changes.
- Replace some spinlocks with mutexes and rwlocks.
- Change the process private semaphores to use mutexes and condition
variables instead of doing the synchronization directly. Spinlocks
are no longer used by the semaphore code.
Instead, make the deferred wakeup list a per-thread array and pass down
the lwpid_t's that way.
- In pthread_cond_wait(), take the mutex before dealing with early wakeup.
In this way there should never be contention on the CV's spinlock if
the app follows POSIX rules (there should only be contention on the
user-provided mutex).
- Add a port of the kernel's rwlocks. The rwlock's spinlock is only taken if
there is contention. This is enabled where atomic ops are available. Right
now that is only i386 and amd64 because I don't have other hardware to
test with. It's trivial to add stubs for other architectures as long as
they have compare-and-swap. When we have proper atomic ops the old rwlock
code can be removed.
- Add a new mutex implementation that's similar to the kernel's mutexes, but
uses compare-and-swap to maintain the waiters list, so no spinlocks are
involved. Same caveats apply as for the rwlocks.
yielding. This is a nasty band-aid but with many threads, looping over
sched_yield() wastes a huge amount of CPU time. It would be nice to have a
way to temporarily disable preemption, but it turns out that's yet another
no-brain concept that has been patented and the patent holder seems to be
suing people lately. Another alternative is probably to have kernel-assisted
spinlocks.
Chops another ~10% off create/join in a loop on i386.
- Disable low level debugging as this is stable. Improves benchmarks
across the board by a small percentage. Uncontested mutex acquire
and release in a loop becomes about 8% quicker.
- Minor cleanup.
hint pointer, but do so in a way that remains compatible with older
pthread libraries. This can be used to wake another thread before the
calling thread goes asleep, saving at least one syscall + involuntary
context switch. This turns out to be a fairly large win on the condvar
benchmarks that I have tried.
detach/join.
- Make mutex acquire spin for a short time, as done with spinlocks.
- Make the number of spins controllable with the env var PTHREAD_NSPINS.
- Reduce the amount of time that libpthread internal spinlocks are held.
- Rely more on the barrier effects of park/unpark to avoid taking spinlocks.
- Simplify the locking around pthreads and the global queues.
- Align per-thread sync data on a 128 byte boundary.
- Offset thread stacks by a small amount to try and reduce cache thrash.
ad