summaryrefslogtreecommitdiff
path: root/src/thread/pthread_barrier_wait.c
diff options
context:
space:
mode:
authorRich Felker <dalias@aerifal.cx>2011-09-27 13:50:29 -0400
committerRich Felker <dalias@aerifal.cx>2011-09-27 13:50:29 -0400
commit60164570111873175111cf8a5b973375e492eee9 (patch)
tree1557931c4e0bacd2be217965be08031d5e3a4e51 /src/thread/pthread_barrier_wait.c
parent3f39c9b3130cd6c142d358159879b799370a6663 (diff)
downloadmusl-60164570111873175111cf8a5b973375e492eee9.tar.gz
musl-60164570111873175111cf8a5b973375e492eee9.tar.bz2
musl-60164570111873175111cf8a5b973375e492eee9.tar.xz
musl-60164570111873175111cf8a5b973375e492eee9.zip
process-shared barrier support, based on discussion with bdonlan
this implementation is rather heavy-weight, but it's the first solution i've found that's actually correct. all waiters actually wait twice at the barrier so that they can synchronize exit, and they hold a "vm lock" that prevents changes to virtual memory mappings (and blocks pthread_barrier_destroy) until all waiters are finished inspecting the barrier. thus, it is safe for any thread to destroy and/or unmap the barrier's memory as soon as pthread_barrier_wait returns, without further synchronization.
Diffstat (limited to 'src/thread/pthread_barrier_wait.c')
-rw-r--r--src/thread/pthread_barrier_wait.c74
1 files changed, 67 insertions, 7 deletions
diff --git a/src/thread/pthread_barrier_wait.c b/src/thread/pthread_barrier_wait.c
index aed1adc8..db432ba3 100644
--- a/src/thread/pthread_barrier_wait.c
+++ b/src/thread/pthread_barrier_wait.c
@@ -1,5 +1,62 @@
#include "pthread_impl.h"
+static int vmlock[2];
+
+void __vm_lock(int inc)
+{
+ for (;;) {
+ int v = vmlock[0];
+ if (inc*v < 0) __wait(vmlock, vmlock+1, v, 1);
+ else if (a_cas(vmlock, v, v+inc)==v) break;
+ }
+}
+
+void __vm_unlock(void)
+{
+ if (vmlock[0]>0) a_dec(vmlock);
+ else a_inc(vmlock);
+ if (vmlock[1]) __wake(vmlock, 1, 1);
+}
+
+static int pshared_barrier_wait(pthread_barrier_t *b)
+{
+ int limit = (b->_b_limit & INT_MAX) + 1;
+ int seq;
+ int ret = 0;
+
+ if (!limit) return PTHREAD_BARRIER_SERIAL_THREAD;
+
+ while (a_swap(&b->_b_lock, 1))
+ __wait(&b->_b_lock, &b->_b_waiters, 1, 0);
+
+ seq = b->_b_seq;
+
+ if (++b->_b_count == limit) {
+ ret = PTHREAD_BARRIER_SERIAL_THREAD;
+ b->_b_seq++;
+ __wake(&b->_b_seq, -1, 0);
+ } else {
+ a_store(&b->_b_lock, 0);
+ if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+ __wait(&b->_b_seq, 0, seq, 0);
+ }
+
+ __vm_lock(+1);
+
+ if (a_fetch_add(&b->_b_count, -1)==1) {
+ b->_b_seq++;
+ __wake(&b->_b_seq, -1, 0);
+ a_store(&b->_b_lock, 0);
+ if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+ } else {
+ __wait(&b->_b_seq, 0, seq+1, 0);
+ }
+
+ __vm_unlock();
+
+ return 0;
+}
+
struct instance
{
int count;
@@ -16,9 +73,12 @@ int pthread_barrier_wait(pthread_barrier_t *b)
/* Trivial case: count was set at 1 */
if (!limit) return PTHREAD_BARRIER_SERIAL_THREAD;
+ /* Process-shared barriers require a separate, inefficient wait */
+ if (limit < 0) return pshared_barrier_wait(b);
+
/* Otherwise we need a lock on the barrier object */
while (a_swap(&b->_b_lock, 1))
- __wait(&b->_b_lock, &b->_b_waiters, 1, 0);
+ __wait(&b->_b_lock, &b->_b_waiters, 1, 1);
inst = b->_b_inst;
/* First thread to enter the barrier becomes the "instance owner" */
@@ -27,7 +87,7 @@ int pthread_barrier_wait(pthread_barrier_t *b)
int spins = 10000;
b->_b_inst = inst = &new_inst;
a_store(&b->_b_lock, 0);
- if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+ if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
while (spins-- && !inst->finished)
a_spin();
a_inc(&inst->finished);
@@ -40,19 +100,19 @@ int pthread_barrier_wait(pthread_barrier_t *b)
if (++inst->count == limit) {
b->_b_inst = 0;
a_store(&b->_b_lock, 0);
- if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+ if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
a_store(&inst->last, 1);
if (inst->waiters)
- __wake(&inst->last, -1, 0);
+ __wake(&inst->last, -1, 1);
} else {
a_store(&b->_b_lock, 0);
- if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
- __wait(&inst->last, &inst->waiters, 0, 0);
+ if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
+ __wait(&inst->last, &inst->waiters, 0, 1);
}
/* Last thread to exit the barrier wakes the instance owner */
if (a_fetch_add(&inst->count,-1)==1 && a_fetch_add(&inst->finished,1))
- __wake(&inst->finished, 1, 0);
+ __wake(&inst->finished, 1, 1);
return 0;
}