Round out the facilty for a 'bound' thread to loan out its KSE

in specific situations. The owner thread must be blocked, and the
borrower can not proceed back to user space with the borrowed KSE.
The borrower will return the KSE on the next context switch where
teh owner wants it back. This removes a lot of possible
race conditions and deadlocks. It is consceivable that the
borrower should inherit the priority of the owner too.
that's another discussion and would be simple to do.

Also, as part of this, the "preallocatd spare thread" is attached to the
thread doing a syscall rather than the KSE. This removes the need to lock
the scheduler when we want to access it, as it's now "at hand".

DDB now shows a lot mor info for threaded proceses though it may need
some optimisation to squeeze it all back into 80 chars again.
(possible JKH project)

Upcalls are now "bound" threads, but "KSE Lending" now means that
other completing syscalls can be completed using that KSE before the upcall
finally makes it back to the UTS. (getting threads OUT OF THE KERNEL is
one of the highest priorities in the KSE system.) The upcall when it happens
will present all the completed syscalls to the KSE for selection.

15 files changed, 769 insertions, 513 deletions

diff --git a/sys/amd64/amd64/trap.c b/sys/amd64/amd64/trap.c
index 444bcf3..fdc0b49 100644
--- a/sys/amd64/amd64/trap.c
+++ b/sys/amd64/amd64/trap.c
@@ -274,7 +274,7 @@ trap(frame)
 		 * XXX p_singlethread not locked, but should be safe.
 		 */
 		if ((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) {
-			PROC_LOCK(p); /* check if thisis really needed */
+			PROC_LOCK(p);
 			mtx_lock_spin(&sched_lock);
 			thread_exit();
 			/* NOTREACHED */
@@ -965,18 +965,39 @@ syscall(frame)
 		cred_update_thread(td);
 	if (p->p_flag & P_KSES) {
 		/*
+		 * First check that we shouldn't just abort.
+		 * But check if we are the single thread first!
+		 * XXX p_singlethread not locked, but should be safe.
+		 */
+		if ((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) {
+			PROC_LOCK(p);
+			mtx_lock_spin(&sched_lock);
+			thread_exit();
+			/* NOTREACHED */
+		}
+
+		/*
 		 * If we are doing a syscall in a KSE environment,
 		 * note where our mailbox is. There is always the
 		 * possibility that we could do this lazily (in sleep()),
 		 * but for now do it every time.
 		 */
+#if 0
 		td->td_mailbox = (void *)fuword((caddr_t)td->td_kse->ke_mailbox
 		    + offsetof(struct kse_mailbox, km_curthread));
+#else /* if user pointer arithmetic is ok in the kernel */
+		td->td_mailbox =
+		    (void *)fuword(
+		    (void *)&td->td_kse->ke_mailbox->km_curthread);
+#endif
 		if ((td->td_mailbox == NULL) ||
 		(td->td_mailbox == (void *)-1)) {
 			td->td_mailbox = NULL;	/* single thread it.. */
 			td->td_flags &= ~TDF_UNBOUND;
 		} else {
+			if (td->td_standin == NULL) {
+				td->td_standin = thread_alloc();
+			}
 			td->td_flags |= TDF_UNBOUND;
 		}
 	}
diff --git a/sys/amd64/amd64/vm_machdep.c b/sys/amd64/amd64/vm_machdep.c
index 3e1329d..ff991ef 100644
--- a/sys/amd64/amd64/vm_machdep.c
+++ b/sys/amd64/amd64/vm_machdep.c
@@ -309,7 +309,10 @@ cpu_thread_setup(struct thread *td)
 
 /*
  * Initialize machine state (pcb and trap frame) for a new thread about to
- * upcall.
+ * upcall. Pu t enough state in the new thread's PCB to get it to go back 
+ * userret(), where we can intercept it again to set the return (upcall)
+ * Address and stack, along with those from upcals that are from other sources
+ * such as those generated in thread_userret() itself.
  */
 void
 cpu_set_upcall(struct thread *td, void *pcb)
@@ -369,8 +372,9 @@ cpu_set_upcall(struct thread *td, void *pcb)
 }
 
 /*
- * Set the machine state for performing an upcall that had to
- * wait until we selected a KSE to perform the upcall on.
+ * Set that machine state for performing an upcall that has to
+ * be done in thread_userret() so that those upcalls generated
+ * in thread_userret() itself can be done as well.
  */
 void
 cpu_set_upcall_kse(struct thread *td, struct kse *ke)
diff --git a/sys/ddb/db_ps.c b/sys/ddb/db_ps.c
index 26dc00b..8ddaaaf 100644
--- a/sys/ddb/db_ps.c
+++ b/sys/ddb/db_ps.c
@@ -127,7 +127,7 @@ db_ps(dummy1, dummy2, dummy3, dummy4)
 			db_printf("(threaded)  %s\n", p->p_comm);
 		FOREACH_THREAD_IN_PROC(p, td) {
 			if (p->p_flag & P_KSES) 
-				db_printf( "       thread %p ", td);
+				db_printf( "   thread %p ksegrp %p ", td, td->td_ksegrp);
 			if (TD_ON_SLEEPQ(td)) {
 				if (td->td_flags & TDF_CVWAITQ)
 					db_printf("[CVQ ");
@@ -155,6 +155,9 @@ db_ps(dummy1, dummy2, dummy3, dummy4)
 				if (TD_AWAITING_INTR(td)) {
 					db_printf("[IWAIT]");
 				}
+				if (TD_LENT(td)) {
+					db_printf("[LOAN]");
+				}
 				break;
 			case TDS_CAN_RUN:
 				db_printf("[Can run]");
@@ -168,9 +171,11 @@ db_ps(dummy1, dummy2, dummy3, dummy4)
 			default:
 				panic("unknown thread state");
 			}
-			if (p->p_flag & P_KSES)
+			if (p->p_flag & P_KSES) {
+				if (td->td_kse)
+					db_printf("[kse %p]", td->td_kse);
 				db_printf("\n");
-			else
+			} else
 				db_printf(" %s\n", p->p_comm);
 					
 		}
diff --git a/sys/i386/i386/trap.c b/sys/i386/i386/trap.c
index 444bcf3..fdc0b49 100644
--- a/sys/i386/i386/trap.c
+++ b/sys/i386/i386/trap.c
@@ -274,7 +274,7 @@ trap(frame)
 		 * XXX p_singlethread not locked, but should be safe.
 		 */
 		if ((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) {
-			PROC_LOCK(p); /* check if thisis really needed */
+			PROC_LOCK(p);
 			mtx_lock_spin(&sched_lock);
 			thread_exit();
 			/* NOTREACHED */
@@ -965,18 +965,39 @@ syscall(frame)
 		cred_update_thread(td);
 	if (p->p_flag & P_KSES) {
 		/*
+		 * First check that we shouldn't just abort.
+		 * But check if we are the single thread first!
+		 * XXX p_singlethread not locked, but should be safe.
+		 */
+		if ((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) {
+			PROC_LOCK(p);
+			mtx_lock_spin(&sched_lock);
+			thread_exit();
+			/* NOTREACHED */
+		}
+
+		/*
 		 * If we are doing a syscall in a KSE environment,
 		 * note where our mailbox is. There is always the
 		 * possibility that we could do this lazily (in sleep()),
 		 * but for now do it every time.
 		 */
+#if 0
 		td->td_mailbox = (void *)fuword((caddr_t)td->td_kse->ke_mailbox
 		    + offsetof(struct kse_mailbox, km_curthread));
+#else /* if user pointer arithmetic is ok in the kernel */
+		td->td_mailbox =
+		    (void *)fuword(
+		    (void *)&td->td_kse->ke_mailbox->km_curthread);
+#endif
 		if ((td->td_mailbox == NULL) ||
 		(td->td_mailbox == (void *)-1)) {
 			td->td_mailbox = NULL;	/* single thread it.. */
 			td->td_flags &= ~TDF_UNBOUND;
 		} else {
+			if (td->td_standin == NULL) {
+				td->td_standin = thread_alloc();
+			}
 			td->td_flags |= TDF_UNBOUND;
 		}
 	}
diff --git a/sys/i386/i386/vm_machdep.c b/sys/i386/i386/vm_machdep.c
index 3e1329d..ff991ef 100644
--- a/sys/i386/i386/vm_machdep.c
+++ b/sys/i386/i386/vm_machdep.c
@@ -309,7 +309,10 @@ cpu_thread_setup(struct thread *td)
 
 /*
  * Initialize machine state (pcb and trap frame) for a new thread about to
- * upcall.
+ * upcall. Pu t enough state in the new thread's PCB to get it to go back 
+ * userret(), where we can intercept it again to set the return (upcall)
+ * Address and stack, along with those from upcals that are from other sources
+ * such as those generated in thread_userret() itself.
  */
 void
 cpu_set_upcall(struct thread *td, void *pcb)
@@ -369,8 +372,9 @@ cpu_set_upcall(struct thread *td, void *pcb)
 }
 
 /*
- * Set the machine state for performing an upcall that had to
- * wait until we selected a KSE to perform the upcall on.
+ * Set that machine state for performing an upcall that has to
+ * be done in thread_userret() so that those upcalls generated
+ * in thread_userret() itself can be done as well.
  */
 void
 cpu_set_upcall_kse(struct thread *td, struct kse *ke)
diff --git a/sys/kern/init_main.c b/sys/kern/init_main.c
index 66b58d2..3b97e60 100644
--- a/sys/kern/init_main.c
+++ b/sys/kern/init_main.c
@@ -374,12 +374,8 @@ proc0_init(void *dummy __unused)
 	ke->ke_oncpu = 0;
 	ke->ke_state = KES_THREAD;
 	ke->ke_thread = td;
-	/* proc_linkup puts it in the idle queue, that's not what we want. */
-	TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
-	kg->kg_idle_kses--;
 	p->p_peers = 0;
 	p->p_leader = p;
-KASSERT((ke->ke_kgrlist.tqe_next != ke), ("linked to self!"));
 
 
 	bcopy("swapper", p->p_comm, sizeof ("swapper"));
diff --git a/sys/kern/kern_condvar.c b/sys/kern/kern_condvar.c
index 26f5376..3ce9aff 100644
--- a/sys/kern/kern_condvar.c
+++ b/sys/kern/kern_condvar.c
@@ -130,16 +130,14 @@ cv_check_upcall(struct thread *td)
 	if ((td->td_proc->p_flag & P_KSES) && td->td_mailbox &&
 	    (td->td_flags & TDF_INMSLEEP) == 0) {
 		/*
-		 * If we have no queued work to do,
-		 * upcall to the UTS to see if it has more work.
 		 * We don't need to upcall now, just queue it.
+		 * The upcall will happen when other n-kernel work
+		 * in this SKEGRP has completed.
+		 * Don't recurse here!
 		 */
-		if (TAILQ_FIRST(&td->td_ksegrp->kg_runq) == NULL) {
-			/* Don't recurse here! */
-			td->td_flags |= TDF_INMSLEEP;
-			thread_schedule_upcall(td, td->td_kse);
-			td->td_flags &= ~TDF_INMSLEEP;
-		}
+		td->td_flags |= TDF_INMSLEEP;
+		thread_schedule_upcall(td, td->td_kse);
+		td->td_flags &= ~TDF_INMSLEEP;
 	}
 }
 
diff --git a/sys/kern/kern_exit.c b/sys/kern/kern_exit.c
index 83d5149..a586bef 100644
--- a/sys/kern/kern_exit.c
+++ b/sys/kern/kern_exit.c
@@ -561,9 +561,10 @@ wait1(td, uap, compat)
 	int compat;
 {
 	struct rusage ru;
-	register int nfound;
-	register struct proc *p, *q, *t;
+	int nfound;
+	struct proc *p, *q, *t;
 	int status, error;
+	struct thread *td2;
 	struct kse *ke;
 	struct ksegrp *kg;
 
@@ -718,8 +719,8 @@ loop:
 			}
 
 			/*
-			 * There should only be one KSE/KSEGRP but
-			 * do it right anyhow.
+			 * There should only be one 
+			 * but do it right anyhow.
 			 */
 			FOREACH_KSEGRP_IN_PROC(p, kg) {
 				FOREACH_KSE_IN_GROUP(kg, ke) {
@@ -730,6 +731,12 @@ loop:
 					}
 				}
 			}
+			FOREACH_THREAD_IN_PROC(p, td2) {
+				if (td2->td_standin != NULL) {
+					thread_free(td2->td_standin);
+					td2->td_standin = NULL;
+				}
+			}
 			thread_reap();	/* check for zombie threads */
 
 			/*
diff --git a/sys/kern/kern_fork.c b/sys/kern/kern_fork.c
index 9fbf602..0af883b 100644
--- a/sys/kern/kern_fork.c
+++ b/sys/kern/kern_fork.c
@@ -500,8 +500,6 @@ again:
 #undef RANGEOF
 
 	/* Set up the thread as an active thread (as if runnable). */
-	TAILQ_REMOVE(&kg2->kg_iq, ke2, ke_kgrlist);
-	kg2->kg_idle_kses--;
 	ke2->ke_state = KES_THREAD;
 	ke2->ke_thread = td2;
 	td2->td_kse = ke2;
diff --git a/sys/kern/kern_kse.c b/sys/kern/kern_kse.c
index 3326119..407b777 100644
--- a/sys/kern/kern_kse.c
+++ b/sys/kern/kern_kse.c
@@ -67,7 +67,7 @@ static int oiks_debug = 1;	/* 0 disable, 1 printf, 2 enter debugger */
 SYSCTL_INT(_kern_threads, OID_AUTO, oiks, CTLFLAG_RW,
 	&oiks_debug, 0, "OIKS thread debug");
 
-static int max_threads_per_proc = 6;
+static int max_threads_per_proc = 10;
 SYSCTL_INT(_kern_threads, OID_AUTO, max_per_proc, CTLFLAG_RW,
 	&max_threads_per_proc, 0, "Limit on threads per proc");
 
@@ -470,6 +470,11 @@ thread_exit(void)
 		thread_stash(ke->ke_tdspare);
 		ke->ke_tdspare = NULL;
 	}
+	if (td->td_standin != NULL) {
+		thread_stash(td->td_standin);
+		td->td_standin = NULL;
+	}
+
 	cpu_thread_exit(td);	/* XXXSMP */
 
 	/*
@@ -478,14 +483,6 @@ thread_exit(void)
 	 * all this stuff.
 	 */
 	if (p->p_numthreads > 1) {
-		/* Reassign this thread's KSE. */
-		ke->ke_thread = NULL;
-		td->td_kse = NULL;
-		ke->ke_state = KES_UNQUEUED;
-		if (ke->ke_bound == td)
-			ke->ke_bound = NULL;
-		kse_reassign(ke);
-
 		/* Unlink this thread from its proc. and the kseg */
 		TAILQ_REMOVE(&p->p_threads, td, td_plist);
 		p->p_numthreads--;
@@ -501,12 +498,41 @@ thread_exit(void)
 				thread_unsuspend_one(p->p_singlethread);
 			}
 		}
+
+		/* Reassign this thread's KSE. */
+		ke->ke_thread = NULL;
+		td->td_kse = NULL;
+		ke->ke_state = KES_UNQUEUED;
+		if (ke->ke_bound == td) {
+			printf("thread_exit: entered with ke_bound set\n");
+			ke->ke_bound = NULL; /* should never happen */
+		}
+
+		kse_reassign(ke);
 		PROC_UNLOCK(p);
 		td->td_state	= TDS_INACTIVE;
 		td->td_proc	= NULL;
 		td->td_ksegrp	= NULL;
 		td->td_last_kse	= NULL;
-		ke->ke_tdspare = td;
+		/* 
+		 * For now stash this here, however
+		 * it's not a permanent solution.
+		 *  When we want to make KSEs exit as well
+		 * we'll have to face this one again.
+		 * Where will we hide it then?
+		 *
+		 * In borrower threads, stash it in the lender
+		 * Where it won't be needed until
+		 * this thread is long gone.
+		 */
+		if (ke->ke_bound) {
+			if (ke->ke_bound->td_standin) {
+				thread_stash(ke->ke_bound->td_standin);
+			}
+			ke->ke_bound->td_standin = td;
+		} else {
+			ke->ke_tdspare = td;
+		}
 	} else {
 		PROC_UNLOCK(p);
 	}
@@ -555,40 +581,85 @@ struct thread *
 thread_schedule_upcall(struct thread *td, struct kse *ke)
 {
 	struct thread *td2;
+	int newkse;
 
 	mtx_assert(&sched_lock, MA_OWNED);
-	if (ke->ke_tdspare != NULL) {
-		td2 = ke->ke_tdspare;
-		ke->ke_tdspare = NULL;
+	newkse = (ke != td->td_kse);
+
+	/* 
+	 * If the kse is already owned by another thread then we can't
+	 * schedule an upcall because the other thread must be BOUND
+	 * which means it is not in a position to take an upcall.
+	 * We must be borrowing the KSE to allow us to complete some in-kernel
+	 * work. When we complete, the Bound thread will have teh chance to 
+	 * complete. This thread will sleep as planned. Hopefully there will
+	 * eventually be un unbound thread that can be converted to an
+	 * upcall to report the completion of this thread.
+	 */
+	if (ke->ke_bound && ((ke->ke_bound->td_flags & TDF_UNBOUND) == 0)) {
+		return (NULL);
+	}
+	KASSERT((ke->ke_bound == NULL), ("kse already bound"));
+
+	if ((td2 = td->td_standin) != NULL) {
+		td->td_standin = NULL;
 	} else {
-		mtx_unlock_spin(&sched_lock);
-		td2 = thread_alloc();
-		mtx_lock_spin(&sched_lock);
+		if (newkse)
+			panic("no reserve thread when called with a new kse");
+		/*
+		 * If called from (e.g.) sleep and we do not have
+		 * a reserve thread, then we've used it, so do not
+		 * create an upcall.
+		 */
+		return(NULL);
 	}
 	CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)",
-	     td, td->td_proc->p_pid, td->td_proc->p_comm);
+	     td2, td->td_proc->p_pid, td->td_proc->p_comm);
 	bzero(&td2->td_startzero,
 	    (unsigned)RANGEOF(struct thread, td_startzero, td_endzero));
 	bcopy(&td->td_startcopy, &td2->td_startcopy,
 	    (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy));
 	thread_link(td2, ke->ke_ksegrp);
 	cpu_set_upcall(td2, td->td_pcb);
+
+	/*
+	 * XXXKSE do we really need this? (default values for the
+	 * frame).
+	 */
 	bcopy(td->td_frame, td2->td_frame, sizeof(struct trapframe));
+
 	/*
-	 * The user context for this thread is selected when we choose
-	 * a KSE and return to userland on it. All we need do here is
-	 * note that the thread exists in order to perform an upcall.
-	 *
-	 * Since selecting a KSE to perform the upcall involves locking
-	 * that KSE's context to our upcall, its best to wait until the
-	 * last possible moment before grabbing a KSE. We do this in
-	 * userret().
+	 * Bind the new thread to the KSE,
+	 * and if it's our KSE, lend it back to ourself
+	 * so we can continue running.
 	 */
 	td2->td_ucred = crhold(td->td_ucred);
-	td2->td_flags = TDF_UNBOUND|TDF_UPCALLING;
-	TD_SET_CAN_RUN(td2);
-	setrunqueue(td2);
-	return (td2);
+	td2->td_flags = TDF_UPCALLING; /* note: BOUND */
+	td2->td_kse = ke;
+	td2->td_state = TDS_CAN_RUN;
+	td2->td_inhibitors = 0;
+	/*
+	 * If called from msleep(), we are working on the current
+	 * KSE so fake that we borrowed it. If called from
+	 * kse_create(), don't, as we have a new kse too.
+	 */
+	if (!newkse) {
+		/*
+		 * This thread will be scheduled when the current thread
+		 * blocks, exits or tries to enter userspace, (which ever
+		 * happens first). When that happens the KSe will "revert"
+		 * to this thread in a BOUND manner. Since we are called
+		 * from msleep() this is going to be "very soon" in nearly
+		 * all cases.
+		 */
+		ke->ke_bound = td2;
+		TD_SET_LOAN(td2);
+	} else {
+		ke->ke_bound = NULL;
+		ke->ke_thread = td2;
+		setrunqueue(td2);
+	}
+	return (td2);	/* bogus.. should be a void function */
 }
 
 /*
@@ -605,6 +676,7 @@ signal_upcall(struct proc *p, int sig)
 	int error;
 
 	PROC_LOCK_ASSERT(p, MA_OWNED);
+return (NULL);
 
 	td = FIRST_THREAD_IN_PROC(p);
 	ke = td->td_kse;
@@ -619,94 +691,15 @@ signal_upcall(struct proc *p, int sig)
 	PROC_LOCK(p);
 	if (error)
 		return (NULL);
+	if (td->td_standin == NULL)
+		td->td_standin = thread_alloc();
 	mtx_lock_spin(&sched_lock);
-	td2 = thread_schedule_upcall(td, ke);
+	td2 = thread_schedule_upcall(td, ke); /* Bogus JRE */
 	mtx_unlock_spin(&sched_lock);
 	return (td2);
 }
 
 /*
- * Consider whether or not an upcall should be made, and update the
- * TDF_UPCALLING flag appropriately.
- *
- * This function is called when the current thread had been bound to a user
- * thread that performed a syscall that blocked, and is now returning.
- * Got that? syscall -> msleep -> wakeup -> syscall_return -> us.
- *
- * This thread will be returned to the UTS in its mailbox as a completed
- * thread.  We need to decide whether or not to perform an upcall now,
- * or simply queue the thread for later.
- *
- * XXXKSE Future enhancement: We could also return back to
- * the thread if we haven't had to do an upcall since then.
- * If the KSE's copy is == the thread's copy, and there are
- * no other completed threads.
- */
-static int
-thread_consider_upcalling(struct thread *td)
-{
-	struct proc *p;
-	struct ksegrp *kg;
-	int error;
-
-	/*
-	 * Save the thread's context, and link it
-	 * into the KSEGRP's list of completed threads.
-	 */
-	error = thread_export_context(td);
-	td->td_flags &= ~TDF_UNBOUND;
-	td->td_mailbox = NULL;
-	if (error)
-		/*
-		 * Failing to do the KSE operation just defaults
-		 * back to synchonous operation, so just return from
-		 * the syscall.
-		 */
-		return (error);
-
-	/*
-	 * Decide whether to perform an upcall now.
-	 */
-	/* Make sure there are no other threads waiting to run. */
-	p = td->td_proc;
-	kg = td->td_ksegrp;
-	PROC_LOCK(p);
-	mtx_lock_spin(&sched_lock);
-	/* bogus test, ok for testing though */
-	if (TAILQ_FIRST(&kg->kg_runq) && 
-	    (TAILQ_LAST(&kg->kg_runq, threadqueue) 
-		!= kg->kg_last_assigned)) {
-		/*
-		 * Another thread in this KSEG needs to run.
-		 * Switch to it instead of performing an upcall,
-		 * abondoning this thread.  Perform the upcall
-		 * later; discard this thread for now.
-		 *
-		 * XXXKSE - As for the other threads to run;
-		 * we COULD rush through all the threads
-		 * in this KSEG at this priority, or we
-		 * could throw the ball back into the court
-		 * and just run the highest prio kse available.
-		 * What is OUR priority?  The priority of the highest
-		 * sycall waiting to be returned?
-		 * For now, just let another KSE run (easiest).
-		 */
-		thread_exit(); /* Abandon current thread. */
-		/* NOTREACHED */
-	} 
-	/*
-	 * Perform an upcall now.
-	 *
-	 * XXXKSE - Assumes we are going to userland, and not
-	 * nested in the kernel.
-	 */
-	td->td_flags |= TDF_UPCALLING;
-	mtx_unlock_spin(&sched_lock);
-	PROC_UNLOCK(p);
-	return (0);
-}
-
-/*
  * The extra work we go through if we are a threaded process when we
  * return to userland.
  *
@@ -724,86 +717,188 @@ thread_userret(struct thread *td, struct trapframe *frame)
 	int error;
 	int unbound;
 	struct kse *ke;
+	struct ksegrp *kg;
+	struct thread *td2;
+	struct proc *p;
 
-	if (td->td_kse->ke_bound) {
-		thread_export_context(td);
-		PROC_LOCK(td->td_proc);
-		mtx_lock_spin(&sched_lock);
-		thread_exit();
-	}
+	error = 0;
 
-	/* Make the thread bound from now on, but remember what it was. */
 	unbound = td->td_flags & TDF_UNBOUND;
-	td->td_flags &= ~TDF_UNBOUND;
-	/*
-	 * Ensure that we have a spare thread available.
-	 */
-	ke = td->td_kse;
-	if (ke->ke_tdspare == NULL) {
-		mtx_lock(&Giant);
-		ke->ke_tdspare = thread_alloc();
-		mtx_unlock(&Giant);
-	}
-	/*
-	 * Originally bound threads need no additional work.
-	 */
-	if (unbound == 0)
-		return (0);
-	error = 0;
+
+	kg = td->td_ksegrp;
+	p = td->td_proc;
+
 	/*
-	 * Decide whether or not we should perform an upcall now.
+	 * Originally bound threads never upcall but they may 
+	 * loan out their KSE at this point.
+	 * Upcalls imply bound.. They also may want to do some Philantropy.
+	 * Unbound threads on the other hand either yield to other work
+	 * or transform into an upcall.
+	 * (having saved their context to user space in both cases)
 	 */
-	if (((td->td_flags & TDF_UPCALLING) == 0) && unbound) {
-		/* if we have other threads to run we will not return */
-		if ((error = thread_consider_upcalling(td)))
-			return (error); /* coundn't go async , just go sync. */
-	}
-	if (td->td_flags & TDF_UPCALLING) {
+	if (unbound ) {
 		/*
-		 * There is no more work to do and we are going to ride
-		 * this thead/KSE up to userland as an upcall.
+		 * We are an unbound thread, looking to return to 
+		 * user space.
+		 * THere are several possibilities:
+		 * 1) we are using a borrowed KSE. save state and exit.
+		 *    kse_reassign() will recycle the kse as needed,
+		 * 2) we are not.. save state, and then convert ourself
+		 *    to be an upcall, bound to the KSE.
+		 *    if there are others that need the kse,
+		 *    give them a chance by doing an mi_switch().
+		 *    Because we are bound, control will eventually return
+		 *    to us here.
+		 * ***
+		 * Save the thread's context, and link it
+		 * into the KSEGRP's list of completed threads.
 		 */
-		CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
-		    td, td->td_proc->p_pid, td->td_proc->p_comm);
+		error = thread_export_context(td);
+		td->td_mailbox = NULL;
+		if (error) {
+			/*
+			 * If we are not running on a borrowed KSE, then
+			 * failing to do the KSE operation just defaults
+			 * back to synchonous operation, so just return from
+			 * the syscall. If it IS borrowed, there is nothing
+			 * we can do. We just lose that context. We
+			 * probably should note this somewhere and send
+			 * the process a signal.
+			 */
+			PROC_LOCK(td->td_proc);
+			psignal(td->td_proc, SIGSEGV);
+			mtx_lock_spin(&sched_lock);
+			if (td->td_kse->ke_bound == NULL) {
+				td->td_flags &= ~TDF_UNBOUND;
+				PROC_UNLOCK(td->td_proc);
+				mtx_unlock_spin(&sched_lock);
+				return (error);	/* go sync */
+			}
+			thread_exit();
+		}
 
 		/*
-		 * Set user context to the UTS.
+		 * if the KSE is owned and we are borrowing it,
+		 * don't make an upcall, just exit so that the owner
+		 * can get its KSE if it wants it.
+		 * Our context is already safely stored for later
+		 * use by the UTS.
 		 */
-		cpu_set_upcall_kse(td, ke);
-
+		PROC_LOCK(p);
+		mtx_lock_spin(&sched_lock);
+		if (td->td_kse->ke_bound) {
+			thread_exit();
+		}
+		PROC_UNLOCK(p);
+				
 		/*
-		 * Put any completed mailboxes on this KSE's list.
+		 * Turn ourself into a bound upcall.
+		 * We will rely on kse_reassign()
+		 * to make us run at a later time.
+		 * We should look just like a sheduled upcall
+		 * from msleep() or cv_wait().
 		 */
-		error = thread_link_mboxes(td->td_ksegrp, ke);
-		if (error)
-			goto bad;
+		td->td_flags &= ~TDF_UNBOUND;
+		td->td_flags |= TDF_UPCALLING;
+		/* Only get here if we have become an upcall */
 
-		/*
-		 * Set state and mailbox.
+	} else {
+		mtx_lock_spin(&sched_lock);
+	}
+	/* 
+	 * We ARE going back to userland with this KSE.
+	 * Check for threads that need to borrow it.
+	 * Optimisation: don't call mi_switch if no-one wants the KSE.
+	 * Any other thread that comes ready after this missed the boat.
+	 */
+	ke = td->td_kse;
+	if ((td2 = kg->kg_last_assigned)) 
+		td2 = TAILQ_NEXT(td2, td_runq);
+	else
+		td2 = TAILQ_FIRST(&kg->kg_runq);
+	if (td2)  {
+		/* 
+		 * force a switch to more urgent 'in kernel'
+		 * work. Control will return to this thread
+		 * when there is no more work to do.
+		 * kse_reassign() will do tha for us.
 		 */
-		td->td_flags &= ~TDF_UPCALLING;
-#if 0
-		error = suword((caddr_t)ke->ke_mailbox +
-		    offsetof(struct kse_mailbox, km_curthread),
-		    0);
-#else	/* if user pointer arithmetic is ok in the kernel */
-		error = suword((caddr_t)&ke->ke_mailbox->km_curthread, 0);
-#endif
-		if (error)
-			goto bad;
+		TD_SET_LOAN(td);
+		ke->ke_bound = td;
+		ke->ke_thread = NULL;
+		mi_switch(); /* kse_reassign() will (re)find td2 */
+	}
+	mtx_unlock_spin(&sched_lock);
+
+	/*
+	 * Optimisation:
+	 * Ensure that we have a spare thread available,
+	 * for when we re-enter the kernel.
+	 */
+	if (td->td_standin == NULL) {
+		if (ke->ke_tdspare) {
+			td->td_standin = ke->ke_tdspare;
+			ke->ke_tdspare = NULL;
+		} else {
+			td->td_standin = thread_alloc();
+		}
 	}
+
+	/* 
+	 * To get here, we know there is no other need for our
+	 * KSE so we can proceed. If not upcalling, go back to 
+	 * userspace. If we are, get the upcall set up.
+	 */
+	if ((td->td_flags & TDF_UPCALLING) == 0)
+		return (0);
+
+	/* 
+	 * We must be an upcall to get this far.
+	 * There is no more work to do and we are going to ride
+	 * this thead/KSE up to userland as an upcall.
+	 * Do the last parts of the setup needed for the upcall.
+	 */
+	CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
+	    td, td->td_proc->p_pid, td->td_proc->p_comm);
+
+	/*
+	 * Set user context to the UTS.
+	 */
+	cpu_set_upcall_kse(td, ke);
+
 	/*
-	 * Stop any chance that we may be separated from
-	 * the KSE we are currently on. This is "biting the bullet",
-	 * we are committing to go to user space as as this KSE here.
+	 * Put any completed mailboxes on this KSE's list.
 	 */
-	return (error);
+	error = thread_link_mboxes(kg, ke);
+	if (error)
+		goto bad;
+
+	/*
+	 * Set state and mailbox.
+	 * From now on we are just a bound outgoing process.
+	 * **Problem** userret is often called several times.
+	 * it would be nice if this all happenned only on the first time 
+	 * through. (the scan for extra work etc.)
+	 */
+	td->td_flags &= ~TDF_UPCALLING;
+#if 0
+	error = suword((caddr_t)ke->ke_mailbox +
+	    offsetof(struct kse_mailbox, km_curthread), 0);
+#else	/* if user pointer arithmetic is ok in the kernel */
+	error = suword((caddr_t)&ke->ke_mailbox->km_curthread, 0);
+#endif
+	if (!error)
+		return (0);
+
 bad:
 	/*
 	 * Things are going to be so screwed we should just kill the process.
  	 * how do we do that?
 	 */
-	 panic ("thread_userret.. need to kill proc..... how?");
+	PROC_LOCK(td->td_proc);
+	psignal(td->td_proc, SIGSEGV);
+	PROC_UNLOCK(td->td_proc);
+	return (error);	/* go sync */
 }
 
 /*
diff --git a/sys/kern/kern_proc.c b/sys/kern/kern_proc.c
index bd613c45..82138a5 100644
--- a/sys/kern/kern_proc.c
+++ b/sys/kern/kern_proc.c
@@ -241,9 +241,7 @@ kse_link(struct kse *ke, struct ksegrp *kg)
 
 	TAILQ_INSERT_HEAD(&kg->kg_kseq, ke, ke_kglist);
 	kg->kg_kses++;
-	ke->ke_state = KES_IDLE;
-	TAILQ_INSERT_HEAD(&kg->kg_iq, ke, ke_kgrlist);
-	kg->kg_idle_kses++;
+	ke->ke_state = KES_UNQUEUED;
 	ke->ke_proc	= p;
 	ke->ke_ksegrp	= kg;
 	ke->ke_thread	= NULL;
@@ -310,24 +308,17 @@ kse_exit(struct thread *td, struct kse_exit_args *uap)
 int
 kse_release(struct thread *td, struct kse_release_args *uap)
 {
-	struct thread *td2;
+	struct proc *p;
 
+	p = td->td_proc;
 	/* KSE-enabled processes only, please. */
-	if ((td->td_proc->p_flag & P_KSES) == 0)
-		return (EINVAL);
-
-	/* Don't discard the last thread. */
-	td2 = FIRST_THREAD_IN_PROC(td->td_proc);
-	KASSERT(td2 != NULL, ("kse_release: no threads in our proc"));
-	if (TAILQ_NEXT(td, td_plist) == NULL)
-		return (EINVAL);
-
-	/* Abandon thread. */
-	PROC_LOCK(td->td_proc);
-	mtx_lock_spin(&sched_lock);
-	thread_exit();
-	/* NOTREACHED */
-	return (0);
+	if (p->p_flag & P_KSES) {
+		PROC_LOCK(p);
+		mtx_lock_spin(&sched_lock);
+		thread_exit();
+		/* NOTREACHED */
+	}
+	return (EINVAL);
 }
 
 /* struct kse_wakeup_args {
@@ -423,6 +414,10 @@ kse_create(struct thread *td, struct kse_create_args *uap)
 		if (SIGPENDING(p))
 			newke->ke_flags |= KEF_ASTPENDING;
 		PROC_UNLOCK(p);
+		/* For the first call this may not have been set */
+		if (td->td_standin == NULL) {
+			td->td_standin = thread_alloc();
+		}
 		mtx_lock_spin(&sched_lock);
 		if (newkg)
 			ksegrp_link(newkg, p);
diff --git a/sys/kern/kern_switch.c b/sys/kern/kern_switch.c
index b4b8879..37500a1 100644
--- a/sys/kern/kern_switch.c
+++ b/sys/kern/kern_switch.c
@@ -107,7 +107,11 @@ CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS);
 static struct runq runq;
 SYSINIT(runq, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, runq_init, &runq)
 
+void panc(char *string1, char *string2);
+
+#if 0
 static void runq_readjust(struct runq *rq, struct kse *ke);
+#endif
 /************************************************************************
  * Functions that manipulate runnability from a thread perspective.	*
  ************************************************************************/
@@ -169,9 +173,10 @@ retry:
 }
 
 /*
- * Given a KSE (now surplus), either assign a new runable thread to it
+ * Given a KSE (now surplus or at least loanable), either assign a new
+ * runable thread to it
  * (and put it in the run queue) or put it in the ksegrp's idle KSE list.
- * Assumes the kse is not linked to any threads any more. (has been cleaned).
+ * Or aybe give it back to its owner if it's been loaned.
  */
 void
 kse_reassign(struct kse *ke)
@@ -179,23 +184,15 @@ kse_reassign(struct kse *ke)
 	struct ksegrp *kg;
 	struct thread *td;
 	struct thread *owner;
+	struct thread *original;
 
 	mtx_assert(&sched_lock, MA_OWNED);
 	kg = ke->ke_ksegrp;
 	owner = ke->ke_bound;
+	original = ke->ke_thread;
 	KASSERT(!(owner && ((owner->td_kse != ke) || 
 		    (owner->td_flags & TDF_UNBOUND))), 
 		("kse_reassign: bad thread bound state"));
-	if (owner && (owner->td_inhibitors == TDI_LOAN)) {
-		TD_CLR_LOAN(owner);
-		ke->ke_bound = NULL;
-		ke->ke_thread = owner;
-		owner->td_kse = ke;
-		setrunqueue(owner);
-		CTR2(KTR_RUNQ, "kse_reassign: ke%p -> td%p (give back)",
-			 ke, owner);
-		return;
-	}
 
 	/*
 	 * Find the first unassigned thread
@@ -212,29 +209,77 @@ kse_reassign(struct kse *ke)
 	 * If we found one assign it the kse, otherwise idle the kse.
 	 */
 	if (td) {
+		/*
+		 * If the original is bound to us we can only be lent out so
+		 * make a loan, otherwise we just drop the 
+		 * original thread.
+		 */
+		if (original) {
+			if (((original->td_flags & TDF_UNBOUND) == 0)) {
+				/*
+				 * Put the owner on the side
+				 */
+				ke->ke_bound = original;
+				TD_SET_LOAN(original);
+			} else {
+				original->td_kse = NULL;
+			}
+		}
 		kg->kg_last_assigned = td;
 		td->td_kse = ke;
 		ke->ke_thread = td;
 		runq_add(&runq, ke);
-		if (owner) 
-			TD_SET_LOAN(owner);
+		/*
+		 * if we have already borrowed this,
+		 * just pass it to the new thread,
+		 * otherwise, enact the loan.
+		 */
 		CTR2(KTR_RUNQ, "kse_reassign: ke%p -> td%p", ke, td);
-	} else if (!owner) {
-		ke->ke_state = KES_IDLE;
-		ke->ke_thread = NULL;
-		TAILQ_INSERT_HEAD(&kg->kg_iq, ke, ke_kgrlist);
-		kg->kg_idle_kses++;
-		CTR1(KTR_RUNQ, "kse_reassign: ke%p idled", ke);
-	} else {
+		return;
+	}
+	if (owner) { /* already loaned out */
+		/* effectivly unloan it */
 		TD_CLR_LOAN(owner);
-		ke->ke_state = KES_THREAD;
 		ke->ke_thread = owner;
-		owner->td_kse = ke;
-		ke->ke_flags |= KEF_ONLOANQ;
-		TAILQ_INSERT_HEAD(&kg->kg_lq, ke, ke_kgrlist);
-		kg->kg_loan_kses++;
-		CTR1(KTR_RUNQ, "kse_reassign: ke%p is on loan queue", ke);
+		ke->ke_bound = NULL;
+		if (original)
+			original->td_kse = NULL;
+		original = owner;
+
+		if (TD_CAN_RUN(owner)) {
+			/*
+			 * If the owner thread is now runnable,  run it..
+			 * Let it have its KSE back.
+			 */
+			setrunqueue(owner);
+			CTR2(KTR_RUNQ, "kse_reassign: ke%p -> td%p (give back)",
+			    ke, owner);
+			return;
+		}
+	}
+	/*
+	 * Presetly NOT loaned out.
+	 * If we are bound, we go on the loanable queue
+	 * otherwise onto the free queue.
+	 */
+	if (original) {
+		if (((original->td_flags & TDF_UNBOUND) == 0)) {
+			ke->ke_state = KES_THREAD;
+			ke->ke_flags |= KEF_ONLOANQ;
+			ke->ke_bound = NULL;
+			TAILQ_INSERT_HEAD(&kg->kg_lq, ke, ke_kgrlist);
+			kg->kg_loan_kses++;
+			CTR1(KTR_RUNQ, "kse_reassign: ke%p on loan queue", ke);
+			return;
+		} else {
+			original->td_kse = NULL;
+		}
 	}
+	ke->ke_state = KES_IDLE;
+	ke->ke_thread = NULL;
+	TAILQ_INSERT_HEAD(&kg->kg_iq, ke, ke_kgrlist);
+	kg->kg_idle_kses++;
+	CTR1(KTR_RUNQ, "kse_reassign: ke%p idled", ke);
 }
 
 int
@@ -252,7 +297,7 @@ kserunnable(void)
 void
 remrunqueue(struct thread *td)
 {
-	struct thread *td2, *td3, *owner;
+	struct thread *td2, *td3;
 	struct ksegrp *kg;
 	struct kse *ke;
 
@@ -273,6 +318,8 @@ remrunqueue(struct thread *td)
 		ke->ke_state = KES_THREAD; 
 		return;
 	}
+   	td3 = TAILQ_PREV(td, threadqueue, td_runq);
+	TAILQ_REMOVE(&kg->kg_runq, td, td_runq);
 	if (ke) {
 		/*
 		 * This thread has been assigned to a KSE.
@@ -282,62 +329,12 @@ remrunqueue(struct thread *td)
 		 */
 		td2 = kg->kg_last_assigned;
 		KASSERT((td2 != NULL), ("last assigned has wrong value "));
-		td->td_kse = NULL;
-		if ((td3 = TAILQ_NEXT(td2, td_runq))) {
-			KASSERT(td3 != td, ("td3 somehow matched td"));
-			/*
-			 * Give the next unassigned thread to the KSE
-			 * so the number of runnable KSEs remains
-			 * constant.
-			 */
-			td3->td_kse = ke;
-			ke->ke_thread = td3;
+		if (td2 == td) 
 			kg->kg_last_assigned = td3;
-			runq_readjust(&runq, ke);
-		} else {
-			/*
-			 * There is no unassigned thread.
-			 * If we were the last assigned one,
-			 * adjust the last assigned pointer back
-			 * one, which may result in NULL.
-			 */
-			if (td == td2) {
-				kg->kg_last_assigned =
-				    TAILQ_PREV(td, threadqueue, td_runq);
-			}
-			runq_remove(&runq, ke);
-			KASSERT((ke->ke_state != KES_IDLE),
-			    ("kse already idle"));
-			if (ke->ke_bound) {
-				owner = ke->ke_bound;
-				if (owner->td_inhibitors == TDI_LOAN) {
-					TD_CLR_LOAN(owner);
-					ke->ke_bound = NULL;
-					ke->ke_thread = owner;
-					owner->td_kse = ke;
-					setrunqueue(owner);
-					CTR2(KTR_RUNQ, 
-					"remrunqueue: ke%p -> td%p (give back)",
-			 			ke, owner);
-				} else {
-					TD_CLR_LOAN(owner);
-					ke->ke_state = KES_THREAD;
-					ke->ke_thread = owner;
-					owner->td_kse = ke;
-					ke->ke_flags |= KEF_ONLOANQ;
-					TAILQ_INSERT_HEAD(&kg->kg_lq, ke, 
-						ke_kgrlist);
-					kg->kg_loan_kses++;
-				}
-			} else {
-				ke->ke_state = KES_IDLE;
-				ke->ke_thread = NULL;
-				TAILQ_INSERT_HEAD(&kg->kg_iq, ke, ke_kgrlist);
-				kg->kg_idle_kses++;
-			}
-		}
+		td->td_kse = NULL;
+		ke->ke_thread = NULL;
+		kse_reassign(ke);
 	}
-	TAILQ_REMOVE(&kg->kg_runq, td, td_runq);
 }
 
 void
@@ -355,6 +352,15 @@ setrunqueue(struct thread *td)
 	TD_SET_RUNQ(td);
 	kg = td->td_ksegrp;
 	kg->kg_runnable++;
+	if ((td->td_proc->p_flag & P_KSES) == 0) {
+		/*
+		 * Common path optimisation: Only one of everything
+		 * and the KSE is always already attached.
+		 * Totally ignore the ksegrp run queue.
+		 */
+		runq_add(&runq, td->td_kse);
+		return;
+	}
 	if ((td->td_flags & TDF_UNBOUND) == 0) {
 		KASSERT((td->td_kse != NULL),
 		    ("queueing BAD thread to run queue"));
@@ -365,14 +371,10 @@ setrunqueue(struct thread *td)
 			TAILQ_REMOVE(&kg->kg_lq, ke, ke_kgrlist);
 			kg->kg_loan_kses--;
 		}
-		/*
-		 * Common path optimisation: Only one of everything
-		 * and the KSE is always already attached.
-		 * Totally ignore the ksegrp run queue.
-		 */
 		runq_add(&runq, td->td_kse);
 		return;
 	}
+
 	/* 
 	 * Ok, so we are threading with this thread.
 	 * We don't have a KSE, see if we can get one..
@@ -394,11 +396,16 @@ setrunqueue(struct thread *td)
 			ke->ke_state = KES_THREAD;
 			kg->kg_idle_kses--;
 		} else if (kg->kg_loan_kses) {
+			/*
+			 * Failing that see if we can borrow one.
+			 */
 			ke = TAILQ_FIRST(&kg->kg_lq);
 			TAILQ_REMOVE(&kg->kg_lq, ke, ke_kgrlist);
 			ke->ke_flags &= ~KEF_ONLOANQ;
 			ke->ke_state = KES_THREAD;
-			TD_SET_LOAN(ke->ke_bound);
+			TD_SET_LOAN(ke->ke_thread);
+			ke->ke_bound = ke->ke_thread;
+			ke->ke_thread  = NULL;
 			kg->kg_loan_kses--;
 		} else if (tda && (tda->td_priority > td->td_priority)) {
 			/*
@@ -697,6 +704,7 @@ runq_remove(struct runq *rq, struct kse *ke)
 	ke->ke_ksegrp->kg_runq_kses--;
 }
 
+#if 0
 static void 
 runq_readjust(struct runq *rq, struct kse *ke)
 {
@@ -706,19 +714,27 @@ runq_readjust(struct runq *rq, struct kse *ke)
 		runq_add(rq, ke);
 	}
 }
+#endif
 
 #if 0
 void
-thread_sanity_check(struct thread *td)
+panc(char *string1, char *string2)
+{
+	printf("%s", string1);
+	Debugger(string2);
+}
+
+void
+thread_sanity_check(struct thread *td, char *string)
 {
 	struct proc *p;
 	struct ksegrp *kg;
 	struct kse *ke;
-	struct thread *td2;
+	struct thread *td2 = NULL;
 	unsigned int prevpri;
-	int	saw_lastassigned;
-	int unassigned;
-	int assigned;
+	int	saw_lastassigned = 0;
+	int unassigned = 0;
+	int assigned = 0;
 
 	p = td->td_proc;
 	kg = td->td_ksegrp;
@@ -727,16 +743,16 @@ thread_sanity_check(struct thread *td)
 
 	if (ke) {
 		if (p != ke->ke_proc) {
-			panic("wrong proc");
+			panc(string, "wrong proc");
 		}
 		if (ke->ke_thread != td) {
-			panic("wrong thread");
+			panc(string, "wrong thread");
 		}
 	}
 	
 	if ((p->p_flag & P_KSES) == 0) {
 		if (ke == NULL) {
-			panic("non KSE thread lost kse");
+			panc(string, "non KSE thread lost kse");
 		}
 	} else {
 		prevpri = 0;
@@ -745,22 +761,27 @@ thread_sanity_check(struct thread *td)
 		assigned = 0;
 		TAILQ_FOREACH(td2, &kg->kg_runq, td_runq) {
 			if (td2->td_priority < prevpri) {
-				panic("thread runqueue unosorted");
+				panc(string, "thread runqueue unosorted");
+			}
+			if ((td2->td_state == TDS_RUNQ) &&
+			    td2->td_kse &&
+			    (td2->td_kse->ke_state != KES_ONRUNQ)) {
+				panc(string, "KSE wrong state");
 			}
 			prevpri = td2->td_priority;
 			if (td2->td_kse) {
 				assigned++;
 				if (unassigned) {
-					panic("unassigned before assigned");
+					panc(string, "unassigned before assigned");
 				}
  				if  (kg->kg_last_assigned == NULL) {
-					panic("lastassigned corrupt");
+					panc(string, "lastassigned corrupt");
 				}
 				if (saw_lastassigned) {
-					panic("last assigned not last");
+					panc(string, "last assigned not last");
 				}
 				if (td2->td_kse->ke_thread != td2) {
-					panic("mismatched kse/thread");
+					panc(string, "mismatched kse/thread");
 				}
 			} else {
 				unassigned++;
@@ -768,28 +789,32 @@ thread_sanity_check(struct thread *td)
 			if (td2 == kg->kg_last_assigned) {
 				saw_lastassigned = 1;
 				if (td2->td_kse == NULL) {
-					panic("last assigned not assigned");
+					panc(string, "last assigned not assigned");
 				}
 			}
 		}
 		if (kg->kg_last_assigned && (saw_lastassigned == 0)) {
-			panic("where on earth does lastassigned point?");
+			panc(string, "where on earth does lastassigned point?");
 		}
 		FOREACH_THREAD_IN_GROUP(kg, td2) {
 			if (((td2->td_flags & TDF_UNBOUND) == 0) && 
 			    (TD_ON_RUNQ(td2))) {
 				assigned++;
 				if (td2->td_kse == NULL) {
-					panic ("BOUND thread with no KSE");
+					panc(string, "BOUND thread with no KSE");
 				}
 			}
 		}
 #if 0
 		if ((unassigned + assigned) != kg->kg_runnable) {
-			panic("wrong number in runnable");
+			panc(string, "wrong number in runnable");
 		}
 #endif
 	}
+	if (assigned == 12345) {
+		printf("%p %p %p %p %p %d, %d",
+		    td, td2, ke, kg, p, assigned, saw_lastassigned);
+	}
 }
 #endif
 
diff --git a/sys/kern/kern_synch.c b/sys/kern/kern_synch.c
index 42d23bd..b0f9d92 100644
--- a/sys/kern/kern_synch.c
+++ b/sys/kern/kern_synch.c
@@ -474,18 +474,16 @@ msleep(ident, mtx, priority, wmesg, timo)
 			return (EINTR);
 		if (td->td_mailbox && (!(td->td_flags & TDF_INMSLEEP))) {
 			/*
-			 * If we have no queued work to do, then
-			 * upcall to the UTS to see if it has more to do.
-			 * We don't need to upcall now, just make it and
-			 * queue it.
+			 * Arrange for an upcall to be readied.
+			 * it will not actually happen until all
+			 * pending in-kernel work for this KSEGRP
+			 * has been done.
 			 */
 			mtx_lock_spin(&sched_lock);
-			if (TAILQ_FIRST(&td->td_ksegrp->kg_runq) == NULL) {
-				/* Don't recurse here! */
-				td->td_flags |= TDF_INMSLEEP;
-				thread_schedule_upcall(td, td->td_kse);
-				td->td_flags &= ~TDF_INMSLEEP;
-			}
+			/* Don't recurse here! */
+			td->td_flags |= TDF_INMSLEEP;
+			thread_schedule_upcall(td, td->td_kse);
+			td->td_flags &= ~TDF_INMSLEEP;
 			mtx_unlock_spin(&sched_lock);
 		}
 	}
@@ -818,23 +816,15 @@ mi_switch(void)
 	 * or stopped or any thing else similar.
 	 */
 	if (TD_IS_RUNNING(td)) {
-		KASSERT(((ke->ke_flags & KEF_IDLEKSE) == 0),
-		    ("Idle thread in mi_switch with wrong state"));
 		/* Put us back on the run queue (kse and all). */
 		setrunqueue(td);
-	} else if (td->td_flags & TDF_UNBOUND) {
+	} else if (p->p_flag & P_KSES) {
 		/*
 		 * We will not be on the run queue. So we must be
-		 * sleeping or similar. If it's available,
+		 * sleeping or similar. As it's available,
 		 * someone else can use the KSE if they need it.
-		 * XXXKSE KSE loaning will change this.
+		 * (If bound LOANING can still occur).
 		 */
-		td->td_kse = NULL;
-		kse_reassign(ke);
-	} else if (p->p_flag & P_KSES) {
-		KASSERT(((ke->ke_bound == NULL) || (ke->ke_bound == td)),
-			("mi_switch: bad bound state"));
-		ke->ke_bound = td;
 		kse_reassign(ke);
 	}
 
diff --git a/sys/kern/kern_thread.c b/sys/kern/kern_thread.c
index 3326119..407b777 100644
--- a/sys/kern/kern_thread.c
+++ b/sys/kern/kern_thread.c
@@ -67,7 +67,7 @@ static int oiks_debug = 1;	/* 0 disable, 1 printf, 2 enter debugger */
 SYSCTL_INT(_kern_threads, OID_AUTO, oiks, CTLFLAG_RW,
 	&oiks_debug, 0, "OIKS thread debug");
 
-static int max_threads_per_proc = 6;
+static int max_threads_per_proc = 10;
 SYSCTL_INT(_kern_threads, OID_AUTO, max_per_proc, CTLFLAG_RW,
 	&max_threads_per_proc, 0, "Limit on threads per proc");
 
@@ -470,6 +470,11 @@ thread_exit(void)
 		thread_stash(ke->ke_tdspare);
 		ke->ke_tdspare = NULL;
 	}
+	if (td->td_standin != NULL) {
+		thread_stash(td->td_standin);
+		td->td_standin = NULL;
+	}
+
 	cpu_thread_exit(td);	/* XXXSMP */
 
 	/*
@@ -478,14 +483,6 @@ thread_exit(void)
 	 * all this stuff.
 	 */
 	if (p->p_numthreads > 1) {
-		/* Reassign this thread's KSE. */
-		ke->ke_thread = NULL;
-		td->td_kse = NULL;
-		ke->ke_state = KES_UNQUEUED;
-		if (ke->ke_bound == td)
-			ke->ke_bound = NULL;
-		kse_reassign(ke);
-
 		/* Unlink this thread from its proc. and the kseg */
 		TAILQ_REMOVE(&p->p_threads, td, td_plist);
 		p->p_numthreads--;
@@ -501,12 +498,41 @@ thread_exit(void)
 				thread_unsuspend_one(p->p_singlethread);
 			}
 		}
+
+		/* Reassign this thread's KSE. */
+		ke->ke_thread = NULL;
+		td->td_kse = NULL;
+		ke->ke_state = KES_UNQUEUED;
+		if (ke->ke_bound == td) {
+			printf("thread_exit: entered with ke_bound set\n");
+			ke->ke_bound = NULL; /* should never happen */
+		}
+
+		kse_reassign(ke);
 		PROC_UNLOCK(p);
 		td->td_state	= TDS_INACTIVE;
 		td->td_proc	= NULL;
 		td->td_ksegrp	= NULL;
 		td->td_last_kse	= NULL;
-		ke->ke_tdspare = td;
+		/* 
+		 * For now stash this here, however
+		 * it's not a permanent solution.
+		 *  When we want to make KSEs exit as well
+		 * we'll have to face this one again.
+		 * Where will we hide it then?
+		 *
+		 * In borrower threads, stash it in the lender
+		 * Where it won't be needed until
+		 * this thread is long gone.
+		 */
+		if (ke->ke_bound) {
+			if (ke->ke_bound->td_standin) {
+				thread_stash(ke->ke_bound->td_standin);
+			}
+			ke->ke_bound->td_standin = td;
+		} else {
+			ke->ke_tdspare = td;
+		}
 	} else {
 		PROC_UNLOCK(p);
 	}
@@ -555,40 +581,85 @@ struct thread *
 thread_schedule_upcall(struct thread *td, struct kse *ke)
 {
 	struct thread *td2;
+	int newkse;
 
 	mtx_assert(&sched_lock, MA_OWNED);
-	if (ke->ke_tdspare != NULL) {
-		td2 = ke->ke_tdspare;
-		ke->ke_tdspare = NULL;
+	newkse = (ke != td->td_kse);
+
+	/* 
+	 * If the kse is already owned by another thread then we can't
+	 * schedule an upcall because the other thread must be BOUND
+	 * which means it is not in a position to take an upcall.
+	 * We must be borrowing the KSE to allow us to complete some in-kernel
+	 * work. When we complete, the Bound thread will have teh chance to 
+	 * complete. This thread will sleep as planned. Hopefully there will
+	 * eventually be un unbound thread that can be converted to an
+	 * upcall to report the completion of this thread.
+	 */
+	if (ke->ke_bound && ((ke->ke_bound->td_flags & TDF_UNBOUND) == 0)) {
+		return (NULL);
+	}
+	KASSERT((ke->ke_bound == NULL), ("kse already bound"));
+
+	if ((td2 = td->td_standin) != NULL) {
+		td->td_standin = NULL;
 	} else {
-		mtx_unlock_spin(&sched_lock);
-		td2 = thread_alloc();
-		mtx_lock_spin(&sched_lock);
+		if (newkse)
+			panic("no reserve thread when called with a new kse");
+		/*
+		 * If called from (e.g.) sleep and we do not have
+		 * a reserve thread, then we've used it, so do not
+		 * create an upcall.
+		 */
+		return(NULL);
 	}
 	CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)",
-	     td, td->td_proc->p_pid, td->td_proc->p_comm);
+	     td2, td->td_proc->p_pid, td->td_proc->p_comm);
 	bzero(&td2->td_startzero,
 	    (unsigned)RANGEOF(struct thread, td_startzero, td_endzero));
 	bcopy(&td->td_startcopy, &td2->td_startcopy,
 	    (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy));
 	thread_link(td2, ke->ke_ksegrp);
 	cpu_set_upcall(td2, td->td_pcb);
+
+	/*
+	 * XXXKSE do we really need this? (default values for the
+	 * frame).
+	 */
 	bcopy(td->td_frame, td2->td_frame, sizeof(struct trapframe));
+
 	/*
-	 * The user context for this thread is selected when we choose
-	 * a KSE and return to userland on it. All we need do here is
-	 * note that the thread exists in order to perform an upcall.
-	 *
-	 * Since selecting a KSE to perform the upcall involves locking
-	 * that KSE's context to our upcall, its best to wait until the
-	 * last possible moment before grabbing a KSE. We do this in
-	 * userret().
+	 * Bind the new thread to the KSE,
+	 * and if it's our KSE, lend it back to ourself
+	 * so we can continue running.
 	 */
 	td2->td_ucred = crhold(td->td_ucred);
-	td2->td_flags = TDF_UNBOUND|TDF_UPCALLING;
-	TD_SET_CAN_RUN(td2);
-	setrunqueue(td2);
-	return (td2);
+	td2->td_flags = TDF_UPCALLING; /* note: BOUND */
+	td2->td_kse = ke;
+	td2->td_state = TDS_CAN_RUN;
+	td2->td_inhibitors = 0;
+	/*
+	 * If called from msleep(), we are working on the current
+	 * KSE so fake that we borrowed it. If called from
+	 * kse_create(), don't, as we have a new kse too.
+	 */
+	if (!newkse) {
+		/*
+		 * This thread will be scheduled when the current thread
+		 * blocks, exits or tries to enter userspace, (which ever
+		 * happens first). When that happens the KSe will "revert"
+		 * to this thread in a BOUND manner. Since we are called
+		 * from msleep() this is going to be "very soon" in nearly
+		 * all cases.
+		 */
+		ke->ke_bound = td2;
+		TD_SET_LOAN(td2);
+	} else {
+		ke->ke_bound = NULL;
+		ke->ke_thread = td2;
+		setrunqueue(td2);
+	}
+	return (td2);	/* bogus.. should be a void function */
 }
 
 /*
@@ -605,6 +676,7 @@ signal_upcall(struct proc *p, int sig)
 	int error;
 
 	PROC_LOCK_ASSERT(p, MA_OWNED);
+return (NULL);
 
 	td = FIRST_THREAD_IN_PROC(p);
 	ke = td->td_kse;
@@ -619,94 +691,15 @@ signal_upcall(struct proc *p, int sig)
 	PROC_LOCK(p);
 	if (error)
 		return (NULL);
+	if (td->td_standin == NULL)
+		td->td_standin = thread_alloc();
 	mtx_lock_spin(&sched_lock);
-	td2 = thread_schedule_upcall(td, ke);
+	td2 = thread_schedule_upcall(td, ke); /* Bogus JRE */
 	mtx_unlock_spin(&sched_lock);
 	return (td2);
 }
 
 /*
- * Consider whether or not an upcall should be made, and update the
- * TDF_UPCALLING flag appropriately.
- *
- * This function is called when the current thread had been bound to a user
- * thread that performed a syscall that blocked, and is now returning.
- * Got that? syscall -> msleep -> wakeup -> syscall_return -> us.
- *
- * This thread will be returned to the UTS in its mailbox as a completed
- * thread.  We need to decide whether or not to perform an upcall now,
- * or simply queue the thread for later.
- *
- * XXXKSE Future enhancement: We could also return back to
- * the thread if we haven't had to do an upcall since then.
- * If the KSE's copy is == the thread's copy, and there are
- * no other completed threads.
- */
-static int
-thread_consider_upcalling(struct thread *td)
-{
-	struct proc *p;
-	struct ksegrp *kg;
-	int error;
-
-	/*
-	 * Save the thread's context, and link it
-	 * into the KSEGRP's list of completed threads.
-	 */
-	error = thread_export_context(td);
-	td->td_flags &= ~TDF_UNBOUND;
-	td->td_mailbox = NULL;
-	if (error)
-		/*
-		 * Failing to do the KSE operation just defaults
-		 * back to synchonous operation, so just return from
-		 * the syscall.
-		 */
-		return (error);
-
-	/*
-	 * Decide whether to perform an upcall now.
-	 */
-	/* Make sure there are no other threads waiting to run. */
-	p = td->td_proc;
-	kg = td->td_ksegrp;
-	PROC_LOCK(p);
-	mtx_lock_spin(&sched_lock);
-	/* bogus test, ok for testing though */
-	if (TAILQ_FIRST(&kg->kg_runq) && 
-	    (TAILQ_LAST(&kg->kg_runq, threadqueue) 
-		!= kg->kg_last_assigned)) {
-		/*
-		 * Another thread in this KSEG needs to run.
-		 * Switch to it instead of performing an upcall,
-		 * abondoning this thread.  Perform the upcall
-		 * later; discard this thread for now.
-		 *
-		 * XXXKSE - As for the other threads to run;
-		 * we COULD rush through all the threads
-		 * in this KSEG at this priority, or we
-		 * could throw the ball back into the court
-		 * and just run the highest prio kse available.
-		 * What is OUR priority?  The priority of the highest
-		 * sycall waiting to be returned?
-		 * For now, just let another KSE run (easiest).
-		 */
-		thread_exit(); /* Abandon current thread. */
-		/* NOTREACHED */
-	} 
-	/*
-	 * Perform an upcall now.
-	 *
-	 * XXXKSE - Assumes we are going to userland, and not
-	 * nested in the kernel.
-	 */
-	td->td_flags |= TDF_UPCALLING;
-	mtx_unlock_spin(&sched_lock);
-	PROC_UNLOCK(p);
-	return (0);
-}
-
-/*
  * The extra work we go through if we are a threaded process when we
  * return to userland.
  *
@@ -724,86 +717,188 @@ thread_userret(struct thread *td, struct trapframe *frame)
 	int error;
 	int unbound;
 	struct kse *ke;
+	struct ksegrp *kg;
+	struct thread *td2;
+	struct proc *p;
 
-	if (td->td_kse->ke_bound) {
-		thread_export_context(td);
-		PROC_LOCK(td->td_proc);
-		mtx_lock_spin(&sched_lock);
-		thread_exit();
-	}
+	error = 0;
 
-	/* Make the thread bound from now on, but remember what it was. */
 	unbound = td->td_flags & TDF_UNBOUND;
-	td->td_flags &= ~TDF_UNBOUND;
-	/*
-	 * Ensure that we have a spare thread available.
-	 */
-	ke = td->td_kse;
-	if (ke->ke_tdspare == NULL) {
-		mtx_lock(&Giant);
-		ke->ke_tdspare = thread_alloc();
-		mtx_unlock(&Giant);
-	}
-	/*
-	 * Originally bound threads need no additional work.
-	 */
-	if (unbound == 0)
-		return (0);
-	error = 0;
+
+	kg = td->td_ksegrp;
+	p = td->td_proc;
+
 	/*
-	 * Decide whether or not we should perform an upcall now.
+	 * Originally bound threads never upcall but they may 
+	 * loan out their KSE at this point.
+	 * Upcalls imply bound.. They also may want to do some Philantropy.
+	 * Unbound threads on the other hand either yield to other work
+	 * or transform into an upcall.
+	 * (having saved their context to user space in both cases)
 	 */
-	if (((td->td_flags & TDF_UPCALLING) == 0) && unbound) {
-		/* if we have other threads to run we will not return */
-		if ((error = thread_consider_upcalling(td)))
-			return (error); /* coundn't go async , just go sync. */
-	}
-	if (td->td_flags & TDF_UPCALLING) {
+	if (unbound ) {
 		/*
-		 * There is no more work to do and we are going to ride
-		 * this thead/KSE up to userland as an upcall.
+		 * We are an unbound thread, looking to return to 
+		 * user space.
+		 * THere are several possibilities:
+		 * 1) we are using a borrowed KSE. save state and exit.
+		 *    kse_reassign() will recycle the kse as needed,
+		 * 2) we are not.. save state, and then convert ourself
+		 *    to be an upcall, bound to the KSE.
+		 *    if there are others that need the kse,
+		 *    give them a chance by doing an mi_switch().
+		 *    Because we are bound, control will eventually return
+		 *    to us here.
+		 * ***
+		 * Save the thread's context, and link it
+		 * into the KSEGRP's list of completed threads.
 		 */
-		CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
-		    td, td->td_proc->p_pid, td->td_proc->p_comm);
+		error = thread_export_context(td);
+		td->td_mailbox = NULL;
+		if (error) {
+			/*
+			 * If we are not running on a borrowed KSE, then
+			 * failing to do the KSE operation just defaults
+			 * back to synchonous operation, so just return from
+			 * the syscall. If it IS borrowed, there is nothing
+			 * we can do. We just lose that context. We
+			 * probably should note this somewhere and send
+			 * the process a signal.
+			 */
+			PROC_LOCK(td->td_proc);
+			psignal(td->td_proc, SIGSEGV);
+			mtx_lock_spin(&sched_lock);
+			if (td->td_kse->ke_bound == NULL) {
+				td->td_flags &= ~TDF_UNBOUND;
+				PROC_UNLOCK(td->td_proc);
+				mtx_unlock_spin(&sched_lock);
+				return (error);	/* go sync */
+			}
+			thread_exit();
+		}
 
 		/*
-		 * Set user context to the UTS.
+		 * if the KSE is owned and we are borrowing it,
+		 * don't make an upcall, just exit so that the owner
+		 * can get its KSE if it wants it.
+		 * Our context is already safely stored for later
+		 * use by the UTS.
 		 */
-		cpu_set_upcall_kse(td, ke);
-
+		PROC_LOCK(p);
+		mtx_lock_spin(&sched_lock);
+		if (td->td_kse->ke_bound) {
+			thread_exit();
+		}
+		PROC_UNLOCK(p);
+				
 		/*
-		 * Put any completed mailboxes on this KSE's list.
+		 * Turn ourself into a bound upcall.
+		 * We will rely on kse_reassign()
+		 * to make us run at a later time.
+		 * We should look just like a sheduled upcall
+		 * from msleep() or cv_wait().
 		 */
-		error = thread_link_mboxes(td->td_ksegrp, ke);
-		if (error)
-			goto bad;
+		td->td_flags &= ~TDF_UNBOUND;
+		td->td_flags |= TDF_UPCALLING;
+		/* Only get here if we have become an upcall */
 
-		/*
-		 * Set state and mailbox.
+	} else {
+		mtx_lock_spin(&sched_lock);
+	}
+	/* 
+	 * We ARE going back to userland with this KSE.
+	 * Check for threads that need to borrow it.
+	 * Optimisation: don't call mi_switch if no-one wants the KSE.
+	 * Any other thread that comes ready after this missed the boat.
+	 */
+	ke = td->td_kse;
+	if ((td2 = kg->kg_last_assigned)) 
+		td2 = TAILQ_NEXT(td2, td_runq);
+	else
+		td2 = TAILQ_FIRST(&kg->kg_runq);
+	if (td2)  {
+		/* 
+		 * force a switch to more urgent 'in kernel'
+		 * work. Control will return to this thread
+		 * when there is no more work to do.
+		 * kse_reassign() will do tha for us.
 		 */
-		td->td_flags &= ~TDF_UPCALLING;
-#if 0
-		error = suword((caddr_t)ke->ke_mailbox +
-		    offsetof(struct kse_mailbox, km_curthread),
-		    0);
-#else	/* if user pointer arithmetic is ok in the kernel */
-		error = suword((caddr_t)&ke->ke_mailbox->km_curthread, 0);
-#endif
-		if (error)
-			goto bad;
+		TD_SET_LOAN(td);
+		ke->ke_bound = td;
+		ke->ke_thread = NULL;
+		mi_switch(); /* kse_reassign() will (re)find td2 */
+	}
+	mtx_unlock_spin(&sched_lock);
+
+	/*
+	 * Optimisation:
+	 * Ensure that we have a spare thread available,
+	 * for when we re-enter the kernel.
+	 */
+	if (td->td_standin == NULL) {
+		if (ke->ke_tdspare) {
+			td->td_standin = ke->ke_tdspare;
+			ke->ke_tdspare = NULL;
+		} else {
+			td->td_standin = thread_alloc();
+		}
 	}
+
+	/* 
+	 * To get here, we know there is no other need for our
+	 * KSE so we can proceed. If not upcalling, go back to 
+	 * userspace. If we are, get the upcall set up.
+	 */
+	if ((td->td_flags & TDF_UPCALLING) == 0)
+		return (0);
+
+	/* 
+	 * We must be an upcall to get this far.
+	 * There is no more work to do and we are going to ride
+	 * this thead/KSE up to userland as an upcall.
+	 * Do the last parts of the setup needed for the upcall.
+	 */
+	CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
+	    td, td->td_proc->p_pid, td->td_proc->p_comm);
+
+	/*
+	 * Set user context to the UTS.
+	 */
+	cpu_set_upcall_kse(td, ke);
+
 	/*
-	 * Stop any chance that we may be separated from
-	 * the KSE we are currently on. This is "biting the bullet",
-	 * we are committing to go to user space as as this KSE here.
+	 * Put any completed mailboxes on this KSE's list.
 	 */
-	return (error);
+	error = thread_link_mboxes(kg, ke);
+	if (error)
+		goto bad;
+
+	/*
+	 * Set state and mailbox.
+	 * From now on we are just a bound outgoing process.
+	 * **Problem** userret is often called several times.
+	 * it would be nice if this all happenned only on the first time 
+	 * through. (the scan for extra work etc.)
+	 */
+	td->td_flags &= ~TDF_UPCALLING;
+#if 0
+	error = suword((caddr_t)ke->ke_mailbox +
+	    offsetof(struct kse_mailbox, km_curthread), 0);
+#else	/* if user pointer arithmetic is ok in the kernel */
+	error = suword((caddr_t)&ke->ke_mailbox->km_curthread, 0);
+#endif
+	if (!error)
+		return (0);
+
 bad:
 	/*
 	 * Things are going to be so screwed we should just kill the process.
  	 * how do we do that?
 	 */
-	 panic ("thread_userret.. need to kill proc..... how?");
+	PROC_LOCK(td->td_proc);
+	psignal(td->td_proc, SIGSEGV);
+	PROC_UNLOCK(td->td_proc);
+	return (error);	/* go sync */
 }
 
 /*
diff --git a/sys/sys/proc.h b/sys/sys/proc.h
index 41a775f..33279db 100644
--- a/sys/sys/proc.h
+++ b/sys/sys/proc.h
@@ -286,6 +286,7 @@ struct thread {
 	struct kse_thr_mailbox *td_mailbox; /* the userland mailbox address */
 	struct ucred	*td_ucred;	/* (k) Reference to credentials. */
 	void		(*td_switchin)(void); /* (k) Switchin special func. */
+	struct thread	*td_standin;	/* (?) use this for an upcall */
 	u_int		td_critnest;	/* (k) Critical section nest level. */
 #define	td_endzero td_md
 
@@ -344,6 +345,7 @@ struct thread {
 #define	TD_IS_SUSPENDED(td)	((td)->td_inhibitors & TDI_SUSPENDED)
 #define	TD_IS_SWAPPED(td)	((td)->td_inhibitors & TDI_SWAPPED)
 #define	TD_ON_LOCK(td)		((td)->td_inhibitors & TDI_LOCK)
+#define	TD_LENT(td)		((td)->td_inhibitors & TDI_LOAN)
 #define	TD_AWAITING_INTR(td)	((td)->td_inhibitors & TDI_IWAIT)
 #define	TD_IS_RUNNING(td)	((td)->td_state == TDS_RUNNING)
 #define	TD_ON_RUNQ(td)		((td)->td_state == TDS_RUNQ)
@@ -929,7 +931,7 @@ void	thread_suspend_one(struct thread *td);
 void	thread_unsuspend_one(struct thread *td);
 int	thread_userret(struct thread *td, struct trapframe *frame);
 
-void	thread_sanity_check(struct thread *td);
+void	thread_sanity_check(struct thread *td, char *);
 #endif	/* _KERNEL */
 
 #endif	/* !_SYS_PROC_H_ */