diff --git a/io.h b/io.h
index 4736a32..8482282 100644
--- a/io.h
+++ b/io.h
@@ -186,8 +186,8 @@ unsigned int io_debugstring(int64 s,char* buf,unsigned int bufsize);
 enum { SLOTS=128 };
 typedef struct iomux {
   int ctx;
-  int working;	/* used to synchronize who is filling the queue */
-  unsigned int h,l; /* high, low */
+  volatile int working;		/* used to synchronize who is filling the queue */
+  volatile unsigned int h,l;	/* high, low */
   struct {
     int fd, events;
   } q[SLOTS];
diff --git a/io/iom_wait.c b/io/iom_wait.c
index ff59990..5510ada 100644
--- a/io/iom_wait.c
+++ b/io/iom_wait.c
@@ -27,16 +27,22 @@ int iom_wait(iomux_t* c,int64* s,unsigned int* revents,unsigned long timeout) {
     if (c->working==-2) return -2;	/* iomux was aborted */
 
     for (;;) {		// CAS-loop get the first element from the queue
-      unsigned int f=c->l;
+      unsigned int f=c->l;	// c is a ring buffer, c->l is low, c->h is high
+      // f is here to prevent double fetches from the volatile low water mark
       if (f == c->h)
 	break; /* no elements in queue */
-      int n=(f+1)%SLOTS;
+      // We want to grab the first element but other threads might be
+      // racing us. So first grab the event from the low water mark in
+      // the ring buffer, then increment the low water mark atomically,
+      // and if that worked we know we grabbed the right event.
+      int n=(f+1)%SLOTS;	// next value for c->l
 
-      *s=c->q[f].fd;
+      *s=c->q[f].fd;		// grab event we think is next in line
       *revents=c->q[f].events;
 
+      // now atomically increment low water mark
       if (__sync_bool_compare_and_swap(&c->l,f,n)) {
-	/* we got one */
+	/* Nobody snatched the event from us. Report success */
 	return 1;
       }
       /* collided with another thread, try again */
@@ -45,15 +51,17 @@ int iom_wait(iomux_t* c,int64* s,unsigned int* revents,unsigned long timeout) {
     /* The queue was empty. If someone else is already calling
     * epoll_wait/kevent, then use the semaphore */
     if (__sync_bool_compare_and_swap(&c->working,0,1)) {
-      /* race avoidance: another thread could have come in between and
-       * refilled the job queue */
+      /* If we get here, we got the lock and no other thread is
+       * attempting to fill the queue at the same time. However,
+       * another thread could have interrupted and refilled the job
+       * queue already, so check if that happened. */
       if (c->h != c->l) {
 	/* set working back to 0 unless someone set it to -2 in the mean time (iom_abort) */
 	if (__sync_val_compare_and_swap(&c->working,1,0)==-2) return -2;
 	continue;	// this is why we have an outer for loop, so we don't need goto here
       }
 
-      /* we have the job to fill the struct. */
+      /* We have the job to fill the struct. */
 
 #ifdef HAVE_EPOLL
       struct epoll_event ee[SLOTS];
@@ -88,9 +96,12 @@ int iom_wait(iomux_t* c,int64* s,unsigned int* revents,unsigned long timeout) {
 	   * still need to signal the semaphore below */
 	} else {
 	  /* The CAS loop on c->working above ensures we are the only one writing to c->h */
-	  c->q[c->h].fd=ee[i].data.fd;
-	  c->q[c->h].events=e;
-	  c->h = (c->h + 1) % SLOTS;
+	  size_t hcapture = __atomic_load_n(&c->h, __ATOMIC_ACQUIRE);	// c->h is volatile so make copy to perf-avoid double fetch
+	  c->q[hcapture].fd=ee[i].data.fd;
+	  c->q[hcapture].events=e;
+//	  c->h = (hcapture + 1) % SLOTS;
+	  // use __atomic_store so ARM hardware writes c->q before c->h
+	  __atomic_store_n(&c->h, (hcapture + 1) % SLOTS, __ATOMIC_RELEASE);
 	}
       }
 #elif defined(HAVE_KQUEUE)
@@ -122,9 +133,12 @@ int iom_wait(iomux_t* c,int64* s,unsigned int* revents,unsigned long timeout) {
 	   * still need to signal the semaphore below */
 	} else {
 	  /* The CAS loop on c->working above ensures we are the only one writing to c->h */
-	  c->q[c->h].fd=kev[i].ident;
-	  c->q[c->h].events=e;
-	  c->h = (c->h + 1) % SLOTS;
+	  size_t hcapture = __atomic_load_n(&c->h, __ATOMIC_ACQUIRE);	// c->h is volatile so make copy to perf-avoid double fetch
+	  c->q[hcapture].fd=kev[i].ident;
+	  c->q[hcapture].events=e;
+//	  c->h = (c->h + 1) % SLOTS;
+	  // use __atomic_store so ARM hardware writes c->q before c->h
+	  __atomic_store_n(&c->h, (hcapture + 1) % SLOTS, __ATOMIC_RELEASE);
 	}
       }
 #else