Merge branch 'x86-uv-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip

* 'x86-uv-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
  x86, UV: uv_irq.c: Fix all sparse warnings
  x86, UV: Improve BAU performance and error recovery
  x86, UV: Delete unneeded boot messages
  x86, UV: Clean up UV headers for MMR definitions

3 files changed, 912 insertions, 383 deletions

diff --git a/arch/x86/kernel/apic/x2apic_uv_x.c b/arch/x86/kernel/apic/x2apic_uv_x.c
index c085d52dbaf..e46f98f36e3 100644
--- a/arch/x86/kernel/apic/x2apic_uv_x.c
+++ b/arch/x86/kernel/apic/x2apic_uv_x.c
@@ -735,9 +735,6 @@ void __init uv_system_init(void)
 		uv_node_to_blade[nid] = blade;
 		uv_cpu_to_blade[cpu] = blade;
 		max_pnode = max(pnode, max_pnode);
-
-		printk(KERN_DEBUG "UV: cpu %d, apicid 0x%x, pnode %d, nid %d, lcpu %d, blade %d\n",
-			cpu, apicid, pnode, nid, lcpu, blade);
 	}
 
 	/* Add blade/pnode info for nodes without cpus */
diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c
index 17b03dd3a6b..7fea555929e 100644
--- a/arch/x86/kernel/tlb_uv.c
+++ b/arch/x86/kernel/tlb_uv.c
@@ -1,7 +1,7 @@
 /*
  *	SGI UltraViolet TLB flush routines.
  *
- *	(c) 2008 Cliff Wickman <cpw@sgi.com>, SGI.
+ *	(c) 2008-2010 Cliff Wickman <cpw@sgi.com>, SGI.
  *
  *	This code is released under the GNU General Public License version 2 or
  *	later.
@@ -20,42 +20,67 @@
 #include <asm/idle.h>
 #include <asm/tsc.h>
 #include <asm/irq_vectors.h>
+#include <asm/timer.h>
 
-static struct bau_control	**uv_bau_table_bases __read_mostly;
-static int			uv_bau_retry_limit __read_mostly;
+struct msg_desc {
+	struct bau_payload_queue_entry *msg;
+	int msg_slot;
+	int sw_ack_slot;
+	struct bau_payload_queue_entry *va_queue_first;
+	struct bau_payload_queue_entry *va_queue_last;
+};
 
-/* base pnode in this partition */
-static int			uv_partition_base_pnode __read_mostly;
+#define UV_INTD_SOFT_ACK_TIMEOUT_PERIOD	0x000000000bUL
+
+static int uv_bau_max_concurrent __read_mostly;
+
+static int nobau;
+static int __init setup_nobau(char *arg)
+{
+	nobau = 1;
+	return 0;
+}
+early_param("nobau", setup_nobau);
 
-static unsigned long		uv_mmask __read_mostly;
+/* base pnode in this partition */
+static int uv_partition_base_pnode __read_mostly;
+/* position of pnode (which is nasid>>1): */
+static int uv_nshift __read_mostly;
+static unsigned long uv_mmask __read_mostly;
 
 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
 static DEFINE_PER_CPU(struct bau_control, bau_control);
+static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
+
+struct reset_args {
+	int sender;
+};
 
 /*
- * Determine the first node on a blade.
+ * Determine the first node on a uvhub. 'Nodes' are used for kernel
+ * memory allocation.
  */
-static int __init blade_to_first_node(int blade)
+static int __init uvhub_to_first_node(int uvhub)
 {
 	int node, b;
 
 	for_each_online_node(node) {
 		b = uv_node_to_blade_id(node);
-		if (blade == b)
+		if (uvhub == b)
 			return node;
 	}
-	return -1; /* shouldn't happen */
+	return -1;
 }
 
 /*
- * Determine the apicid of the first cpu on a blade.
+ * Determine the apicid of the first cpu on a uvhub.
  */
-static int __init blade_to_first_apicid(int blade)
+static int __init uvhub_to_first_apicid(int uvhub)
 {
 	int cpu;
 
 	for_each_present_cpu(cpu)
-		if (blade == uv_cpu_to_blade_id(cpu))
+		if (uvhub == uv_cpu_to_blade_id(cpu))
 			return per_cpu(x86_cpu_to_apicid, cpu);
 	return -1;
 }
@@ -68,195 +93,459 @@ static int __init blade_to_first_apicid(int blade)
  * clear of the Timeout bit (as well) will free the resource. No reply will
  * be sent (the hardware will only do one reply per message).
  */
-static void uv_reply_to_message(int resource,
-				struct bau_payload_queue_entry *msg,
-				struct bau_msg_status *msp)
+static inline void uv_reply_to_message(struct msg_desc *mdp,
+				       struct bau_control *bcp)
 {
 	unsigned long dw;
+	struct bau_payload_queue_entry *msg;
 
-	dw = (1 << (resource + UV_SW_ACK_NPENDING)) | (1 << resource);
+	msg = mdp->msg;
+	if (!msg->canceled) {
+		dw = (msg->sw_ack_vector << UV_SW_ACK_NPENDING) |
+						msg->sw_ack_vector;
+		uv_write_local_mmr(
+				UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw);
+	}
 	msg->replied_to = 1;
 	msg->sw_ack_vector = 0;
-	if (msp)
-		msp->seen_by.bits = 0;
-	uv_write_local_mmr(UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw);
 }
 
 /*
- * Do all the things a cpu should do for a TLB shootdown message.
- * Other cpu's may come here at the same time for this message.
+ * Process the receipt of a RETRY message
  */
-static void uv_bau_process_message(struct bau_payload_queue_entry *msg,
-				   int msg_slot, int sw_ack_slot)
+static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
+					    struct bau_control *bcp)
 {
-	unsigned long this_cpu_mask;
-	struct bau_msg_status *msp;
-	int cpu;
+	int i;
+	int cancel_count = 0;
+	int slot2;
+	unsigned long msg_res;
+	unsigned long mmr = 0;
+	struct bau_payload_queue_entry *msg;
+	struct bau_payload_queue_entry *msg2;
+	struct ptc_stats *stat;
 
-	msp = __get_cpu_var(bau_control).msg_statuses + msg_slot;
-	cpu = uv_blade_processor_id();
-	msg->number_of_cpus =
-		uv_blade_nr_online_cpus(uv_node_to_blade_id(numa_node_id()));
-	this_cpu_mask = 1UL << cpu;
-	if (msp->seen_by.bits & this_cpu_mask)
-		return;
-	atomic_or_long(&msp->seen_by.bits, this_cpu_mask);
+	msg = mdp->msg;
+	stat = &per_cpu(ptcstats, bcp->cpu);
+	stat->d_retries++;
+	/*
+	 * cancel any message from msg+1 to the retry itself
+	 */
+	for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
+		if (msg2 > mdp->va_queue_last)
+			msg2 = mdp->va_queue_first;
+		if (msg2 == msg)
+			break;
+
+		/* same conditions for cancellation as uv_do_reset */
+		if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
+		    (msg2->sw_ack_vector) && ((msg2->sw_ack_vector &
+			msg->sw_ack_vector) == 0) &&
+		    (msg2->sending_cpu == msg->sending_cpu) &&
+		    (msg2->msg_type != MSG_NOOP)) {
+			slot2 = msg2 - mdp->va_queue_first;
+			mmr = uv_read_local_mmr
+				(UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
+			msg_res = ((msg2->sw_ack_vector << 8) |
+				   msg2->sw_ack_vector);
+			/*
+			 * This is a message retry; clear the resources held
+			 * by the previous message only if they timed out.
+			 * If it has not timed out we have an unexpected
+			 * situation to report.
+			 */
+			if (mmr & (msg_res << 8)) {
+				/*
+				 * is the resource timed out?
+				 * make everyone ignore the cancelled message.
+				 */
+				msg2->canceled = 1;
+				stat->d_canceled++;
+				cancel_count++;
+				uv_write_local_mmr(
+				    UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
+					(msg_res << 8) | msg_res);
+			} else
+				printk(KERN_INFO "note bau retry: no effect\n");
+		}
+	}
+	if (!cancel_count)
+		stat->d_nocanceled++;
+}
 
-	if (msg->replied_to == 1)
-		return;
+/*
+ * Do all the things a cpu should do for a TLB shootdown message.
+ * Other cpu's may come here at the same time for this message.
+ */
+static void uv_bau_process_message(struct msg_desc *mdp,
+				   struct bau_control *bcp)
+{
+	int msg_ack_count;
+	short socket_ack_count = 0;
+	struct ptc_stats *stat;
+	struct bau_payload_queue_entry *msg;
+	struct bau_control *smaster = bcp->socket_master;
 
+	/*
+	 * This must be a normal message, or retry of a normal message
+	 */
+	msg = mdp->msg;
+	stat = &per_cpu(ptcstats, bcp->cpu);
 	if (msg->address == TLB_FLUSH_ALL) {
 		local_flush_tlb();
-		__get_cpu_var(ptcstats).alltlb++;
+		stat->d_alltlb++;
 	} else {
 		__flush_tlb_one(msg->address);
-		__get_cpu_var(ptcstats).onetlb++;
+		stat->d_onetlb++;
 	}
+	stat->d_requestee++;
+
+	/*
+	 * One cpu on each uvhub has the additional job on a RETRY
+	 * of releasing the resource held by the message that is
+	 * being retried.  That message is identified by sending
+	 * cpu number.
+	 */
+	if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
+		uv_bau_process_retry_msg(mdp, bcp);
 
-	__get_cpu_var(ptcstats).requestee++;
+	/*
+	 * This is a sw_ack message, so we have to reply to it.
+	 * Count each responding cpu on the socket. This avoids
+	 * pinging the count's cache line back and forth between
+	 * the sockets.
+	 */
+	socket_ack_count = atomic_add_short_return(1, (struct atomic_short *)
+			&smaster->socket_acknowledge_count[mdp->msg_slot]);
+	if (socket_ack_count == bcp->cpus_in_socket) {
+		/*
+		 * Both sockets dump their completed count total into
+		 * the message's count.
+		 */
+		smaster->socket_acknowledge_count[mdp->msg_slot] = 0;
+		msg_ack_count = atomic_add_short_return(socket_ack_count,
+				(struct atomic_short *)&msg->acknowledge_count);
+
+		if (msg_ack_count == bcp->cpus_in_uvhub) {
+			/*
+			 * All cpus in uvhub saw it; reply
+			 */
+			uv_reply_to_message(mdp, bcp);
+		}
+	}
 
-	atomic_inc_short(&msg->acknowledge_count);
-	if (msg->number_of_cpus == msg->acknowledge_count)
-		uv_reply_to_message(sw_ack_slot, msg, msp);
+	return;
 }
 
 /*
- * Examine the payload queue on one distribution node to see
- * which messages have not been seen, and which cpu(s) have not seen them.
+ * Determine the first cpu on a uvhub.
+ */
+static int uvhub_to_first_cpu(int uvhub)
+{
+	int cpu;
+	for_each_present_cpu(cpu)
+		if (uvhub == uv_cpu_to_blade_id(cpu))
+			return cpu;
+	return -1;
+}
+
+/*
+ * Last resort when we get a large number of destination timeouts is
+ * to clear resources held by a given cpu.
+ * Do this with IPI so that all messages in the BAU message queue
+ * can be identified by their nonzero sw_ack_vector field.
  *
- * Returns the number of cpu's that have not responded.
+ * This is entered for a single cpu on the uvhub.
+ * The sender want's this uvhub to free a specific message's
+ * sw_ack resources.
  */
-static int uv_examine_destination(struct bau_control *bau_tablesp, int sender)
+static void
+uv_do_reset(void *ptr)
 {
-	struct bau_payload_queue_entry *msg;
-	struct bau_msg_status *msp;
-	int count = 0;
 	int i;
-	int j;
+	int slot;
+	int count = 0;
+	unsigned long mmr;
+	unsigned long msg_res;
+	struct bau_control *bcp;
+	struct reset_args *rap;
+	struct bau_payload_queue_entry *msg;
+	struct ptc_stats *stat;
 
-	for (msg = bau_tablesp->va_queue_first, i = 0; i < DEST_Q_SIZE;
-	     msg++, i++) {
-		if ((msg->sending_cpu == sender) && (!msg->replied_to)) {
-			msp = bau_tablesp->msg_statuses + i;
-			printk(KERN_DEBUG
-			       "blade %d: address:%#lx %d of %d, not cpu(s): ",
-			       i, msg->address, msg->acknowledge_count,
-			       msg->number_of_cpus);
-			for (j = 0; j < msg->number_of_cpus; j++) {
-				if (!((1L << j) & msp->seen_by.bits)) {
-					count++;
-					printk("%d ", j);
-				}
+	bcp = &per_cpu(bau_control, smp_processor_id());
+	rap = (struct reset_args *)ptr;
+	stat = &per_cpu(ptcstats, bcp->cpu);
+	stat->d_resets++;
+
+	/*
+	 * We're looking for the given sender, and
+	 * will free its sw_ack resource.
+	 * If all cpu's finally responded after the timeout, its
+	 * message 'replied_to' was set.
+	 */
+	for (msg = bcp->va_queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
+		/* uv_do_reset: same conditions for cancellation as
+		   uv_bau_process_retry_msg() */
+		if ((msg->replied_to == 0) &&
+		    (msg->canceled == 0) &&
+		    (msg->sending_cpu == rap->sender) &&
+		    (msg->sw_ack_vector) &&
+		    (msg->msg_type != MSG_NOOP)) {
+			/*
+			 * make everyone else ignore this message
+			 */
+			msg->canceled = 1;
+			slot = msg - bcp->va_queue_first;
+			count++;
+			/*
+			 * only reset the resource if it is still pending
+			 */
+			mmr = uv_read_local_mmr
+					(UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
+			msg_res = ((msg->sw_ack_vector << 8) |
+						   msg->sw_ack_vector);
+			if (mmr & msg_res) {
+				stat->d_rcanceled++;
+				uv_write_local_mmr(
+				    UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
+							msg_res);
 			}
-			printk("\n");
 		}
 	}
-	return count;
+	return;
 }
 
 /*
- * Examine the payload queue on all the distribution nodes to see
- * which messages have not been seen, and which cpu(s) have not seen them.
- *
- * Returns the number of cpu's that have not responded.
+ * Use IPI to get all target uvhubs to release resources held by
+ * a given sending cpu number.
  */
-static int uv_examine_destinations(struct bau_target_nodemask *distribution)
+static void uv_reset_with_ipi(struct bau_target_uvhubmask *distribution,
+			      int sender)
 {
-	int sender;
-	int i;
-	int count = 0;
+	int uvhub;
+	int cpu;
+	cpumask_t mask;
+	struct reset_args reset_args;
+
+	reset_args.sender = sender;
 
-	sender = smp_processor_id();
-	for (i = 0; i < sizeof(struct bau_target_nodemask) * BITSPERBYTE; i++) {
-		if (!bau_node_isset(i, distribution))
+	cpus_clear(mask);
+	/* find a single cpu for each uvhub in this distribution mask */
+	for (uvhub = 0;
+		    uvhub < sizeof(struct bau_target_uvhubmask) * BITSPERBYTE;
+		    uvhub++) {
+		if (!bau_uvhub_isset(uvhub, distribution))
 			continue;
-		count += uv_examine_destination(uv_bau_table_bases[i], sender);
+		/* find a cpu for this uvhub */
+		cpu = uvhub_to_first_cpu(uvhub);
+		cpu_set(cpu, mask);
 	}
-	return count;
+	/* IPI all cpus; Preemption is already disabled */
+	smp_call_function_many(&mask, uv_do_reset, (void *)&reset_args, 1);
+	return;
+}
+
+static inline unsigned long
+cycles_2_us(unsigned long long cyc)
+{
+	unsigned long long ns;
+	unsigned long us;
+	ns =  (cyc * per_cpu(cyc2ns, smp_processor_id()))
+						>> CYC2NS_SCALE_FACTOR;
+	us = ns / 1000;
+	return us;
 }
 
 /*
- * wait for completion of a broadcast message
- *
- * return COMPLETE, RETRY or GIVEUP
+ * wait for all cpus on this hub to finish their sends and go quiet
+ * leaves uvhub_quiesce set so that no new broadcasts are started by
+ * bau_flush_send_and_wait()
+ */
+static inline void
+quiesce_local_uvhub(struct bau_control *hmaster)
+{
+	atomic_add_short_return(1, (struct atomic_short *)
+		 &hmaster->uvhub_quiesce);
+}
+
+/*
+ * mark this quiet-requestor as done
+ */
+static inline void
+end_uvhub_quiesce(struct bau_control *hmaster)
+{
+	atomic_add_short_return(-1, (struct atomic_short *)
+		&hmaster->uvhub_quiesce);
+}
+
+/*
+ * Wait for completion of a broadcast software ack message
+ * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
  */
 static int uv_wait_completion(struct bau_desc *bau_desc,
-			      unsigned long mmr_offset, int right_shift)
+	unsigned long mmr_offset, int right_shift, int this_cpu,
+	struct bau_control *bcp, struct bau_control *smaster, long try)
 {
-	int exams = 0;
-	long destination_timeouts = 0;
-	long source_timeouts = 0;
+	int relaxes = 0;
 	unsigned long descriptor_status;
+	unsigned long mmr;
+	unsigned long mask;
+	cycles_t ttime;
+	cycles_t timeout_time;
+	struct ptc_stats *stat = &per_cpu(ptcstats, this_cpu);
+	struct bau_control *hmaster;
+
+	hmaster = bcp->uvhub_master;
+	timeout_time = get_cycles() + bcp->timeout_interval;
 
+	/* spin on the status MMR, waiting for it to go idle */
 	while ((descriptor_status = (((unsigned long)
 		uv_read_local_mmr(mmr_offset) >>
 			right_shift) & UV_ACT_STATUS_MASK)) !=
 			DESC_STATUS_IDLE) {
-		if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) {
-			source_timeouts++;
-			if (source_timeouts > SOURCE_TIMEOUT_LIMIT)
-				source_timeouts = 0;
-			__get_cpu_var(ptcstats).s_retry++;
-			return FLUSH_RETRY;
-		}
 		/*
-		 * spin here looking for progress at the destinations
+		 * Our software ack messages may be blocked because there are
+		 * no swack resources available.  As long as none of them
+		 * has timed out hardware will NACK our message and its
+		 * state will stay IDLE.
 		 */
-		if (descriptor_status == DESC_STATUS_DESTINATION_TIMEOUT) {
-			destination_timeouts++;
-			if (destination_timeouts > DESTINATION_TIMEOUT_LIMIT) {
-				/*
-				 * returns number of cpus not responding
-				 */
-				if (uv_examine_destinations
-				    (&bau_desc->distribution) == 0) {
-					__get_cpu_var(ptcstats).d_retry++;
-					return FLUSH_RETRY;
-				}
-				exams++;
-				if (exams >= uv_bau_retry_limit) {
-					printk(KERN_DEBUG
-					       "uv_flush_tlb_others");
-					printk("giving up on cpu %d\n",
-					       smp_processor_id());
+		if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) {
+			stat->s_stimeout++;
+			return FLUSH_GIVEUP;
+		} else if (descriptor_status ==
+					DESC_STATUS_DESTINATION_TIMEOUT) {
+			stat->s_dtimeout++;
+			ttime = get_cycles();
+
+			/*
+			 * Our retries may be blocked by all destination
+			 * swack resources being consumed, and a timeout
+			 * pending.  In that case hardware returns the
+			 * ERROR that looks like a destination timeout.
+			 */
+			if (cycles_2_us(ttime - bcp->send_message) < BIOS_TO) {
+				bcp->conseccompletes = 0;
+				return FLUSH_RETRY_PLUGGED;
+			}
+
+			bcp->conseccompletes = 0;
+			return FLUSH_RETRY_TIMEOUT;
+		} else {
+			/*
+			 * descriptor_status is still BUSY
+			 */
+			cpu_relax();
+			relaxes++;
+			if (relaxes >= 10000) {
+				relaxes = 0;
+				if (get_cycles() > timeout_time) {
+					quiesce_local_uvhub(hmaster);
+
+					/* single-thread the register change */
+					spin_lock(&hmaster->masks_lock);
+					mmr = uv_read_local_mmr(mmr_offset);
+					mask = 0UL;
+					mask |= (3UL < right_shift);
+					mask = ~mask;
+					mmr &= mask;
+					uv_write_local_mmr(mmr_offset, mmr);
+					spin_unlock(&hmaster->masks_lock);
+					end_uvhub_quiesce(hmaster);
+					stat->s_busy++;
 					return FLUSH_GIVEUP;
 				}
-				/*
-				 * delays can hang the simulator
-				   udelay(1000);
-				 */
-				destination_timeouts = 0;
 			}
 		}
-		cpu_relax();
 	}
+	bcp->conseccompletes++;
 	return FLUSH_COMPLETE;
 }
 
+static inline cycles_t
+sec_2_cycles(unsigned long sec)
+{
+	unsigned long ns;
+	cycles_t cyc;
+
+	ns = sec * 1000000000;
+	cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
+	return cyc;
+}
+
+/*
+ * conditionally add 1 to *v, unless *v is >= u
+ * return 0 if we cannot add 1 to *v because it is >= u
+ * return 1 if we can add 1 to *v because it is < u
+ * the add is atomic
+ *
+ * This is close to atomic_add_unless(), but this allows the 'u' value
+ * to be lowered below the current 'v'.  atomic_add_unless can only stop
+ * on equal.
+ */
+static inline int atomic_inc_unless_ge(spinlock_t *lock, atomic_t *v, int u)
+{
+	spin_lock(lock);
+	if (atomic_read(v) >= u) {
+		spin_unlock(lock);
+		return 0;
+	}
+	atomic_inc(v);
+	spin_unlock(lock);
+	return 1;
+}
+
 /**
  * uv_flush_send_and_wait
  *
- * Send a broadcast and wait for a broadcast message to complete.
+ * Send a broadcast and wait for it to complete.
  *
- * The flush_mask contains the cpus the broadcast was sent to.
+ * The flush_mask contains the cpus the broadcast is to be sent to, plus
+ * cpus that are on the local uvhub.
  *
- * Returns NULL if all remote flushing was done. The mask is zeroed.
+ * Returns NULL if all flushing represented in the mask was done. The mask
+ * is zeroed.
  * Returns @flush_mask if some remote flushing remains to be done. The
- * mask will have some bits still set.
+ * mask will have some bits still set, representing any cpus on the local
+ * uvhub (not current cpu) and any on remote uvhubs if the broadcast failed.
  */
-const struct cpumask *uv_flush_send_and_wait(int cpu, int this_pnode,
-					     struct bau_desc *bau_desc,
-					     struct cpumask *flush_mask)
+const struct cpumask *uv_flush_send_and_wait(struct bau_desc *bau_desc,
+					     struct cpumask *flush_mask,
+					     struct bau_control *bcp)
 {
-	int completion_status = 0;
 	int right_shift;
-	int tries = 0;
-	int pnode;
+	int uvhub;
 	int bit;
+	int completion_status = 0;
+	int seq_number = 0;
+	long try = 0;
+	int cpu = bcp->uvhub_cpu;
+	int this_cpu = bcp->cpu;
+	int this_uvhub = bcp->uvhub;
 	unsigned long mmr_offset;
 	unsigned long index;
 	cycles_t time1;
 	cycles_t time2;
+	struct ptc_stats *stat = &per_cpu(ptcstats, bcp->cpu);
+	struct bau_control *smaster = bcp->socket_master;
+	struct bau_control *hmaster = bcp->uvhub_master;
+
+	/*
+	 * Spin here while there are hmaster->max_concurrent or more active
+	 * descriptors. This is the per-uvhub 'throttle'.
+	 */
+	if (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
+			&hmaster->active_descriptor_count,
+			hmaster->max_concurrent)) {
+		stat->s_throttles++;
+		do {
+			cpu_relax();
+		} while (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
+			&hmaster->active_descriptor_count,
+			hmaster->max_concurrent));
+	}
+
+	while (hmaster->uvhub_quiesce)
+		cpu_relax();
 
 	if (cpu < UV_CPUS_PER_ACT_STATUS) {
 		mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
@@ -268,24 +557,108 @@ const struct cpumask *uv_flush_send_and_wait(int cpu, int this_pnode,
 	}
 	time1 = get_cycles();
 	do {
-		tries++;
+		/*
+		 * Every message from any given cpu gets a unique message
+		 * sequence number. But retries use that same number.
+		 * Our message may have timed out at the destination because
+		 * all sw-ack resources are in use and there is a timeout
+		 * pending there.  In that case, our last send never got
+		 * placed into the queue and we need to persist until it
+		 * does.
+		 *
+		 * Make any retry a type MSG_RETRY so that the destination will
+		 * free any resource held by a previous message from this cpu.
+		 */
+		if (try == 0) {
+			/* use message type set by the caller the first time */
+			seq_number = bcp->message_number++;
+		} else {
+			/* use RETRY type on all the rest; same sequence */
+			bau_desc->header.msg_type = MSG_RETRY;
+			stat->s_retry_messages++;
+		}
+		bau_desc->header.sequence = seq_number;
 		index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) |
-			cpu;
+			bcp->uvhub_cpu;
+		bcp->send_message = get_cycles();
+
 		uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index);
+
+		try++;
 		completion_status = uv_wait_completion(bau_desc, mmr_offset,
-					right_shift);
-	} while (completion_status == FLUSH_RETRY);
+			right_shift, this_cpu, bcp, smaster, try);
+
+		if (completion_status == FLUSH_RETRY_PLUGGED) {
+			/*
+			 * Our retries may be blocked by all destination swack
+			 * resources being consumed, and a timeout pending. In
+			 * that case hardware immediately returns the ERROR
+			 * that looks like a destination timeout.
+			 */
+			udelay(TIMEOUT_DELAY);
+			bcp->plugged_tries++;
+			if (bcp->plugged_tries >= PLUGSB4RESET) {
+				bcp->plugged_tries = 0;
+				quiesce_local_uvhub(hmaster);
+				spin_lock(&hmaster->queue_lock);
+				uv_reset_with_ipi(&bau_desc->distribution,
+							this_cpu);
+				spin_unlock(&hmaster->queue_lock);
+				end_uvhub_quiesce(hmaster);
+				bcp->ipi_attempts++;
+				stat->s_resets_plug++;
+			}
+		} else if (completion_status == FLUSH_RETRY_TIMEOUT) {
+			hmaster->max_concurrent = 1;
+			bcp->timeout_tries++;
+			udelay(TIMEOUT_DELAY);
+			if (bcp->timeout_tries >= TIMEOUTSB4RESET) {
+				bcp->timeout_tries = 0;
+				quiesce_local_uvhub(hmaster);
+				spin_lock(&hmaster->queue_lock);
+				uv_reset_with_ipi(&bau_desc->distribution,
+								this_cpu);
+				spin_unlock(&hmaster->queue_lock);
+				end_uvhub_quiesce(hmaster);
+				bcp->ipi_attempts++;
+				stat->s_resets_timeout++;
+			}
+		}
+		if (bcp->ipi_attempts >= 3) {
+			bcp->ipi_attempts = 0;
+			completion_status = FLUSH_GIVEUP;
+			break;
+		}
+		cpu_relax();
+	} while ((completion_status == FLUSH_RETRY_PLUGGED) ||
+		 (completion_status == FLUSH_RETRY_TIMEOUT));
 	time2 = get_cycles();
-	__get_cpu_var(ptcstats).sflush += (time2 - time1);
-	if (tries > 1)
-		__get_cpu_var(ptcstats).retriesok++;
 
-	if (completion_status == FLUSH_GIVEUP) {
+	if ((completion_status == FLUSH_COMPLETE) && (bcp->conseccompletes > 5)
+	    && (hmaster->max_concurrent < hmaster->max_concurrent_constant))
+			hmaster->max_concurrent++;
+
+	/*
+	 * hold any cpu not timing out here; no other cpu currently held by
+	 * the 'throttle' should enter the activation code
+	 */
+	while (hmaster->uvhub_quiesce)
+		cpu_relax();
+	atomic_dec(&hmaster->active_descriptor_count);
+
+	/* guard against cycles wrap */
+	if (time2 > time1)
+		stat->s_time += (time2 - time1);
+	else
+		stat->s_requestor--; /* don't count this one */
+	if (completion_status == FLUSH_COMPLETE && try > 1)
+		stat->s_retriesok++;
+	else if (completion_status == FLUSH_GIVEUP) {
 		/*
 		 * Cause the caller to do an IPI-style TLB shootdown on
-		 * the cpu's, all of which are still in the mask.
+		 * the target cpu's, all of which are still in the mask.
 		 */
-		__get_cpu_var(ptcstats).ptc_i++;
+		stat->s_giveup++;
 		return flush_mask;
 	}
 
@@ -294,18 +667,17 @@ const struct cpumask *uv_flush_send_and_wait(int cpu, int this_pnode,
 	 * use the IPI method of shootdown on them.
 	 */
 	for_each_cpu(bit, flush_mask) {
-		pnode = uv_cpu_to_pnode(bit);
-		if (pnode == this_pnode)
+		uvhub = uv_cpu_to_blade_id(bit);
+		if (uvhub == this_uvhub)
 			continue;
 		cpumask_clear_cpu(bit, flush_mask);
 	}
 	if (!cpumask_empty(flush_mask))
 		return flush_mask;
+
 	return NULL;
 }
 
-static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
-
 /**
  * uv_flush_tlb_others - globally purge translation cache of a virtual
  * address or all TLB's
@@ -322,8 +694,8 @@ static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
  * The caller has derived the cpumask from the mm_struct.  This function
  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
  *
- * The cpumask is converted into a nodemask of the nodes containing
- * the cpus.
+ * The cpumask is converted into a uvhubmask of the uvhubs containing
+ * those cpus.
  *
  * Note that this function should be called with preemption disabled.
  *
@@ -335,52 +707,82 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
 					  struct mm_struct *mm,
 					  unsigned long va, unsigned int cpu)
 {
-	struct cpumask *flush_mask = __get_cpu_var(uv_flush_tlb_mask);
-	int i;
-	int bit;
-	int pnode;
-	int uv_cpu;
-	int this_pnode;
+	int remotes;
+	int tcpu;
+	int uvhub;
 	int locals = 0;
 	struct bau_desc *bau_desc;
+	struct cpumask *flush_mask;
+	struct ptc_stats *stat;
+	struct bau_control *bcp;
 
-	cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
+	if (nobau)
+		return cpumask;
 
-	uv_cpu = uv_blade_processor_id();
-	this_pnode = uv_hub_info->pnode;
-	bau_desc = __get_cpu_var(bau_control).descriptor_base;
-	bau_desc += UV_ITEMS_PER_DESCRIPTOR * uv_cpu;
+	bcp = &per_cpu(bau_control, cpu);
+	/*
+	 * Each sending cpu has a per-cpu mask which it fills from the caller's
+	 * cpu mask.  Only remote cpus are converted to uvhubs and copied.
+	 */
+	flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
+	/*
+	 * copy cpumask to flush_mask, removing current cpu
+	 * (current cpu should already have been flushed by the caller and
+	 *  should never be returned if we return flush_mask)
+	 */
+	cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
+	if (cpu_isset(cpu, *cpumask))
+		locals++;  /* current cpu was targeted */
 
-	bau_nodes_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
+	bau_desc = bcp->descriptor_base;
+	bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu;
 
-	i = 0;
-	for_each_cpu(bit, flush_mask) {
-		pnode = uv_cpu_to_pnode(bit);
-		BUG_ON(pnode > (UV_DISTRIBUTION_SIZE - 1));
-		if (pnode == this_pnode) {
+	bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
+	remotes = 0;
+	for_each_cpu(tcpu, flush_mask) {
+		uvhub = uv_cpu_to_blade_id(tcpu);
+		if (uvhub == bcp->uvhub) {
 			locals++;
 			continue;
 		}
-		bau_node_set(pnode - uv_partition_base_pnode,
-				&bau_desc->distribution);
-		i++;
+		bau_uvhub_set(uvhub, &bau_desc->distribution);
+		remotes++;
 	}
-	if (i == 0) {
+	if (remotes == 0) {
 		/*
-		 * no off_node flushing; return status for local node
+		 * No off_hub flushing; return status for local hub.
+		 * Return the caller's mask if all were local (the current
+		 * cpu may be in that mask).
 		 */
 		if (locals)
-			return flush_mask;
+			return cpumask;
 		else
 			return NULL;
 	}
-	__get_cpu_var(ptcstats).requestor++;
-	__get_cpu_var(ptcstats).ntargeted += i;
+	stat = &per_cpu(ptcstats, cpu);
+	stat->s_requestor++;
+	stat->s_ntargcpu += remotes;
+	remotes = bau_uvhub_weight(&bau_desc->distribution);
+	stat->s_ntarguvhub += remotes;
+	if (remotes >= 16)
+		stat->s_ntarguvhub16++;
+	else if (remotes >= 8)
+		stat->s_ntarguvhub8++;
+	else if (remotes >= 4)
+		stat->s_ntarguvhub4++;
+	else if (remotes >= 2)
+		stat->s_ntarguvhub2++;
+	else
+		stat->s_ntarguvhub1++;
 
 	bau_desc->payload.address = va;
 	bau_desc->payload.sending_cpu = cpu;
 
-	return uv_flush_send_and_wait(uv_cpu, this_pnode, bau_desc, flush_mask);
+	/*
+	 * uv_flush_send_and_wait returns null if all cpu's were messaged, or
+	 * the adjusted flush_mask if any cpu's were not messaged.
+	 */
+	return uv_flush_send_and_wait(bau_desc, flush_mask, bcp);
 }
 
 /*
@@ -389,87 +791,70 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
  *
  * We received a broadcast assist message.
  *
- * Interrupts may have been disabled; this interrupt could represent
+ * Interrupts are disabled; this interrupt could represent
  * the receipt of several messages.
  *
- * All cores/threads on this node get this interrupt.
- * The last one to see it does the s/w ack.
+ * All cores/threads on this hub get this interrupt.
+ * The last one to see it does the software ack.
  * (the resource will not be freed until noninterruptable cpus see this
- *  interrupt; hardware will timeout the s/w ack and reply ERROR)
+ *  interrupt; hardware may timeout the s/w ack and reply ERROR)
  */
 void uv_bau_message_interrupt(struct pt_regs *regs)
 {
-	struct bau_payload_queue_entry *va_queue_first;
-	struct bau_payload_queue_entry *va_queue_last;
-	struct bau_payload_queue_entry *msg;
-	struct pt_regs *old_regs = set_irq_regs(regs);
-	cycles_t time1;
-	cycles_t time2;
-	int msg_slot;
-	int sw_ack_slot;
-	int fw;
 	int count = 0;
-	unsigned long local_pnode;
-
-	ack_APIC_irq();
-	exit_idle();
-	irq_enter();
-
-	time1 = get_cycles();
-
-	local_pnode = uv_blade_to_pnode(uv_numa_blade_id());
-
-	va_queue_first = __get_cpu_var(bau_control).va_queue_first;
-	va_queue_last = __get_cpu_var(bau_control).va_queue_last;
-
-	msg = __get_cpu_var(bau_control).bau_msg_head;
+	cycles_t time_start;
+	struct bau_payload_queue_entry *msg;
+	struct bau_control *bcp;
+	struct ptc_stats *stat;
+	struct msg_desc msgdesc;
+
+	time_start = get_cycles();
+	bcp = &per_cpu(bau_control, smp_processor_id());
+	stat = &per_cpu(ptcstats, smp_processor_id());
+	msgdesc.va_queue_first = bcp->va_queue_first;
+	msgdesc.va_queue_last = bcp->va_queue_last;
+	msg = bcp->bau_msg_head;
 	while (msg->sw_ack_vector) {
 		count++;
-		fw = msg->sw_ack_vector;
-		msg_slot = msg - va_queue_first;
-		sw_ack_slot = ffs(fw) - 1;
-
-		uv_bau_process_message(msg, msg_slot, sw_ack_slot);
-
+		msgdesc.msg_slot = msg - msgdesc.va_queue_first;
+		msgdesc.sw_ack_slot = ffs(msg->sw_ack_vector) - 1;
+		msgdesc.msg = msg;
+		uv_bau_process_message(&msgdesc, bcp);
 		msg++;
-		if (msg > va_queue_last)
-			msg = va_queue_first;
-		__get_cpu_var(bau_control).bau_msg_head = msg;
+		if (msg > msgdesc.va_queue_last)
+			msg = msgdesc.va_queue_first;
+		bcp->bau_msg_head = msg;
 	}
+	stat->d_time += (get_cycles() - time_start);
 	if (!count)
-		__get_cpu_var(ptcstats).nomsg++;
+		stat->d_nomsg++;
 	else if (count > 1)
-		__get_cpu_var(ptcstats).multmsg++;
-
-	time2 = get_cycles();
-	__get_cpu_var(ptcstats).dflush += (time2 - time1);
-
-	irq_exit();
-	set_irq_regs(old_regs);
+		stat->d_multmsg++;
+	ack_APIC_irq();
 }
 
 /*
  * uv_enable_timeouts
  *
- * Each target blade (i.e. blades that have cpu's) needs to have
+ * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
  * shootdown message timeouts enabled.  The timeout does not cause
  * an interrupt, but causes an error message to be returned to
  * the sender.
  */
 static void uv_enable_timeouts(void)
 {
-	int blade;
-	int nblades;
+	int uvhub;
+	int nuvhubs;
 	int pnode;
 	unsigned long mmr_image;
 
-	nblades = uv_num_possible_blades();
+	nuvhubs = uv_num_possible_blades();
 
-	for (blade = 0; blade < nblades; blade++) {
-		if (!uv_blade_nr_possible_cpus(blade))
+	for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
+		if (!uv_blade_nr_possible_cpus(uvhub))
 			continue;
 
-		pnode = uv_blade_to_pnode(blade);
+		pnode = uv_blade_to_pnode(uvhub);
 		mmr_image =
 		    uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL);
 		/*
@@ -479,16 +864,16 @@ static void uv_enable_timeouts(void)
 		 * To program the period, the SOFT_ACK_MODE must be off.
 		 */
 		mmr_image &= ~((unsigned long)1 <<
-			       UV_ENABLE_INTD_SOFT_ACK_MODE_SHIFT);
+		    UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 		uv_write_global_mmr64
 		    (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 		/*
 		 * Set the 4-bit period.
 		 */
 		mmr_image &= ~((unsigned long)0xf <<
-			UV_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHIFT);
+		     UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 		mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD <<
-			     UV_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHIFT);
+		     UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 		uv_write_global_mmr64
 		    (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 		/*
@@ -497,7 +882,7 @@ static void uv_enable_timeouts(void)
 		 * indicated in bits 2:0 (7 causes all of them to timeout).
 		 */
 		mmr_image |= ((unsigned long)1 <<
-			      UV_ENABLE_INTD_SOFT_ACK_MODE_SHIFT);
+		    UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 		uv_write_global_mmr64
 		    (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 	}
@@ -522,9 +907,20 @@ static void uv_ptc_seq_stop(struct seq_file *file, void *data)
 {
 }
 
+static inline unsigned long long
+millisec_2_cycles(unsigned long millisec)
+{
+	unsigned long ns;
+	unsigned long long cyc;
+
+	ns = millisec * 1000;
+	cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
+	return cyc;
+}
+
 /*
- * Display the statistics thru /proc
- * data points to the cpu number
+ * Display the statistics thru /proc.
+ * 'data' points to the cpu number
  */
 static int uv_ptc_seq_show(struct seq_file *file, void *data)
 {
@@ -535,78 +931,155 @@ static int uv_ptc_seq_show(struct seq_file *file, void *data)
 
 	if (!cpu) {
 		seq_printf(file,
-		"# cpu requestor requestee one all sretry dretry ptc_i ");
+			"# cpu sent stime numuvhubs numuvhubs16 numuvhubs8 ");
 		seq_printf(file,
-		"sw_ack sflush dflush sok dnomsg dmult starget\n");
+			"numuvhubs4 numuvhubs2 numuvhubs1 numcpus dto ");
+		seq_printf(file,
+			"retries rok resetp resett giveup sto bz throt ");
+		seq_printf(file,
+			"sw_ack recv rtime all ");
+		seq_printf(file,
+			"one mult none retry canc nocan reset rcan\n");
 	}
 	if (cpu < num_possible_cpus() && cpu_online(cpu)) {
 		stat = &per_cpu(ptcstats, cpu);
-		seq_printf(file, "cpu %d %ld %ld %ld %ld %ld %ld %ld ",
-			   cpu, stat->requestor,
-			   stat->requestee, stat->onetlb, stat->alltlb,
-			   stat->s_retry, stat->d_retry, stat->ptc_i);
-		seq_printf(file, "%lx %ld %ld %ld %ld %ld %ld\n",
+		/* source side statistics */
+		seq_printf(file,
+			"cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
+			   cpu, stat->s_requestor, cycles_2_us(stat->s_time),
+			   stat->s_ntarguvhub, stat->s_ntarguvhub16,
+			   stat->s_ntarguvhub8, stat->s_ntarguvhub4,
+			   stat->s_ntarguvhub2, stat->s_ntarguvhub1,
+			   stat->s_ntargcpu, stat->s_dtimeout);
+		seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
+			   stat->s_retry_messages, stat->s_retriesok,
+			   stat->s_resets_plug, stat->s_resets_timeout,
+			   stat->s_giveup, stat->s_stimeout,
+			   stat->s_busy, stat->s_throttles);
+		/* destination side statistics */
+		seq_printf(file,
+			   "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
 			   uv_read_global_mmr64(uv_cpu_to_pnode(cpu),
 					UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE),
-			   stat->sflush, stat->dflush,
-			   stat->retriesok, stat->nomsg,
-			   stat->multmsg, stat->ntargeted);
+			   stat->d_requestee, cycles_2_us(stat->d_time),
+			   stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
+			   stat->d_nomsg, stat->d_retries, stat->d_canceled,
+			   stat->d_nocanceled, stat->d_resets,
+			   stat->d_rcanceled);
 	}
 
 	return 0;
 }
 
 /*
+ * -1: resetf the statistics
  *  0: display meaning of the statistics
- * >0: retry limit
+ * >0: maximum concurrent active descriptors per uvhub (throttle)
  */
 static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user,
 				 size_t count, loff_t *data)
 {
-	long newmode;
+	int cpu;
+	long input_arg;
 	char optstr[64];
+	struct ptc_stats *stat;
+	struct bau_control *bcp;
 
 	if (count == 0 || count > sizeof(optstr))
 		return -EINVAL;
 	if (copy_from_user(optstr, user, count))
 		return -EFAULT;
 	optstr[count - 1] = '\0';
-	if (strict_strtoul(optstr, 10, &newmode) < 0) {
+	if (strict_strtol(optstr, 10, &input_arg) < 0) {
 		printk(KERN_DEBUG "%s is invalid\n", optstr);
 		return -EINVAL;
 	}
 
-	if (newmode == 0) {
+	if (input_arg == 0) {
 		printk(KERN_DEBUG "# cpu:      cpu number\n");
+		printk(KERN_DEBUG "Sender statistics:\n");
+		printk(KERN_DEBUG
+		"sent:     number of shootdown messages sent\n");
+		printk(KERN_DEBUG
+		"stime:    time spent sending messages\n");
+		printk(KERN_DEBUG
+		"numuvhubs: number of hubs targeted with shootdown\n");
+		printk(KERN_DEBUG
+		"numuvhubs16: number times 16 or more hubs targeted\n");
+		printk(KERN_DEBUG
+		"numuvhubs8: number times 8 or more hubs targeted\n");
+		printk(KERN_DEBUG
+		"numuvhubs4: number times 4 or more hubs targeted\n");
+		printk(KERN_DEBUG
+		"numuvhubs2: number times 2 or more hubs targeted\n");
+		printk(KERN_DEBUG
+		"numuvhubs1: number times 1 hub targeted\n");
+		printk(KERN_DEBUG
+		"numcpus:  number of cpus targeted with shootdown\n");
+		printk(KERN_DEBUG
+		"dto:      number of destination timeouts\n");
+		printk(KERN_DEBUG
+		"retries:  destination timeout retries sent\n");
+		printk(KERN_DEBUG
+		"rok:   :  destination timeouts successfully retried\n");
+		printk(KERN_DEBUG
+		"resetp:   ipi-style resource resets for plugs\n");
+		printk(KERN_DEBUG
+		"resett:   ipi-style resource resets for timeouts\n");
+		printk(KERN_DEBUG
+		"giveup:   fall-backs to ipi-style shootdowns\n");
+		printk(KERN_DEBUG
+		"sto:      number of source timeouts\n");
+		printk(KERN_DEBUG
+		"bz:       number of stay-busy's\n");
+		printk(KERN_DEBUG
+		"throt:    number times spun in throttle\n");
+		printk(KERN_DEBUG "Destination side statistics:\n");
 		printk(KERN_DEBUG
-		"requestor:  times this cpu was the flush requestor\n");
+		"sw_ack:   image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
 		printk(KERN_DEBUG
-		"requestee:  times this cpu was requested to flush its TLBs\n");
+		"recv:     shootdown messages received\n");
 		printk(KERN_DEBUG
-		"one:        times requested to flush a single address\n");
+		"rtime:    time spent processing messages\n");
 		printk(KERN_DEBUG
-		"all:        times requested to flush all TLB's\n");
+		"all:      shootdown all-tlb messages\n");
 		printk(KERN_DEBUG
-		"sretry:     number of retries of source-side timeouts\n");
+		"one:      shootdown one-tlb messages\n");
 		printk(KERN_DEBUG
-		"dretry:     number of retries of destination-side timeouts\n");
+		"mult:     interrupts that found multiple messages\n");
 		printk(KERN_DEBUG
-		"ptc_i:      times UV fell through to IPI-style flushes\n");
+		"none:     interrupts that found no messages\n");
 		printk(KERN_DEBUG
-		"sw_ack:     image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
+		"retry:    number of retry messages processed\n");
 		printk(KERN_DEBUG
-		"sflush_us:  cycles spent in uv_flush_tlb_others()\n");
+		"canc:     number messages canceled by retries\n");
 		printk(KERN_DEBUG
-		"dflush_us:  cycles spent in handling flush requests\n");
-		printk(KERN_DEBUG "sok:        successes on retry\n");
-		printk(KERN_DEBUG "dnomsg:     interrupts with no message\n");
+		"nocan:    number retries that found nothing to cancel\n");
 		printk(KERN_DEBUG
-		"dmult:      interrupts with multiple messages\n");
-		printk(KERN_DEBUG "starget:    nodes targeted\n");
+		"reset:    number of ipi-style reset requests processed\n");
+		printk(KERN_DEBUG
+		"rcan:     number messages canceled by reset requests\n");
+	} else if (input_arg == -1) {
+		for_each_present_cpu(cpu) {
+			stat = &per_cpu(ptcstats, cpu);
+			memset(stat, 0, sizeof(struct ptc_stats));
+		}
 	} else {
-		uv_bau_retry_limit = newmode;
-		printk(KERN_DEBUG "timeout retry limit:%d\n",
-		       uv_bau_retry_limit);
+		uv_bau_max_concurrent = input_arg;
+		bcp = &per_cpu(bau_control, smp_processor_id());
+		if (uv_bau_max_concurrent < 1 ||
+		    uv_bau_max_concurrent > bcp->cpus_in_uvhub) {
+			printk(KERN_DEBUG
+				"Error: BAU max concurrent %d; %d is invalid\n",
+				bcp->max_concurrent, uv_bau_max_concurrent);
+			return -EINVAL;
+		}
+		printk(KERN_DEBUG "Set BAU max concurrent:%d\n",
+		       uv_bau_max_concurrent);
+		for_each_present_cpu(cpu) {
+			bcp = &per_cpu(bau_control, cpu);
+			bcp->max_concurrent = uv_bau_max_concurrent;
+		}
 	}
 
 	return count;
@@ -650,79 +1123,30 @@ static int __init uv_ptc_init(void)
 }
 
 /*
- * begin the initialization of the per-blade control structures
- */
-static struct bau_control * __init uv_table_bases_init(int blade, int node)
-{
-	int i;
-	struct bau_msg_status *msp;
-	struct bau_control *bau_tabp;
-
-	bau_tabp =
-	    kmalloc_node(sizeof(struct bau_control), GFP_KERNEL, node);
-	BUG_ON(!bau_tabp);
-
-	bau_tabp->msg_statuses =
-	    kmalloc_node(sizeof(struct bau_msg_status) *
-			 DEST_Q_SIZE, GFP_KERNEL, node);
-	BUG_ON(!bau_tabp->msg_statuses);
-
-	for (i = 0, msp = bau_tabp->msg_statuses; i < DEST_Q_SIZE; i++, msp++)
-		bau_cpubits_clear(&msp->seen_by, (int)
-				  uv_blade_nr_possible_cpus(blade));
-
-	uv_bau_table_bases[blade] = bau_tabp;
-
-	return bau_tabp;
-}
-
-/*
- * finish the initialization of the per-blade control structures
- */
-static void __init
-uv_table_bases_finish(int blade,
-		      struct bau_control *bau_tablesp,
-		      struct bau_desc *adp)
-{
-	struct bau_control *bcp;
-	int cpu;
-
-	for_each_present_cpu(cpu) {
-		if (blade != uv_cpu_to_blade_id(cpu))
-			continue;
-
-		bcp = (struct bau_control *)&per_cpu(bau_control, cpu);
-		bcp->bau_msg_head	= bau_tablesp->va_queue_first;
-		bcp->va_queue_first	= bau_tablesp->va_queue_first;
-		bcp->va_queue_last	= bau_tablesp->va_queue_last;
-		bcp->msg_statuses	= bau_tablesp->msg_statuses;
-		bcp->descriptor_base	= adp;
-	}
-}
-
-/*
  * initialize the sending side's sending buffers
  */
-static struct bau_desc * __init
+static void
 uv_activation_descriptor_init(int node, int pnode)
 {
 	int i;
+	int cpu;
 	unsigned long pa;
 	unsigned long m;
 	unsigned long n;
-	struct bau_desc *adp;
-	struct bau_desc *ad2;
+	struct bau_desc *bau_desc;
+	struct bau_desc *bd2;
+	struct bau_control *bcp;
 
 	/*
 	 * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR)
-	 * per cpu; and up to 32 (UV_ADP_SIZE) cpu's per blade
+	 * per cpu; and up to 32 (UV_ADP_SIZE) cpu's per uvhub
 	 */
-	adp = (struct bau_desc *)kmalloc_node(sizeof(struct bau_desc)*
+	bau_desc = (struct bau_desc *)kmalloc_node(sizeof(struct bau_desc)*
 		UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR, GFP_KERNEL, node);
-	BUG_ON(!adp);
+	BUG_ON(!bau_desc);
 
-	pa = uv_gpa(adp); /* need the real nasid*/
-	n = uv_gpa_to_pnode(pa);
+	pa = uv_gpa(bau_desc); /* need the real nasid*/
+	n = pa >> uv_nshift;
 	m = pa & uv_mmask;
 
 	uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE,
@@ -731,96 +1155,188 @@ uv_activation_descriptor_init(int node, int pnode)
 	/*
 	 * initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each
 	 * cpu even though we only use the first one; one descriptor can
-	 * describe a broadcast to 256 nodes.
+	 * describe a broadcast to 256 uv hubs.
 	 */
-	for (i = 0, ad2 = adp; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR);
-		i++, ad2++) {
-		memset(ad2, 0, sizeof(struct bau_desc));
-		ad2->header.sw_ack_flag = 1;
+	for (i = 0, bd2 = bau_desc; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR);
+		i++, bd2++) {
+		memset(bd2, 0, sizeof(struct bau_desc));
+		bd2->header.sw_ack_flag = 1;
 		/*
-		 * base_dest_nodeid is the first node in the partition, so
-		 * the bit map will indicate partition-relative node numbers.
-		 * note that base_dest_nodeid is actually a nasid.
+		 * base_dest_nodeid is the nasid (pnode<<1) of the first uvhub
+		 * in the partition. The bit map will indicate uvhub numbers,
+		 * which are 0-N in a partition. Pnodes are unique system-wide.
 		 */
-		ad2->header.base_dest_nodeid = uv_partition_base_pnode << 1;
-		ad2->header.dest_subnodeid = 0x10; /* the LB */
-		ad2->header.command = UV_NET_ENDPOINT_INTD;
-		ad2->header.int_both = 1;
+		bd2->header.base_dest_nodeid = uv_partition_base_pnode << 1;
+		bd2->header.dest_subnodeid = 0x10; /* the LB */
+		bd2->header.command = UV_NET_ENDPOINT_INTD;
+		bd2->header.int_both = 1;
 		/*
 		 * all others need to be set to zero:
 		 *   fairness chaining multilevel count replied_to
 		 */
 	}
-	return adp;
+	for_each_present_cpu(cpu) {
+		if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
+			continue;
+		bcp = &per_cpu(bau_control, cpu);
+		bcp->descriptor_base = bau_desc;
+	}
 }
 
 /*
  * initialize the destination side's receiving buffers
+ * entered for each uvhub in the partition
+ * - node is first node (kernel memory notion) on the uvhub
+ * - pnode is the uvhub's physical identifier
  */
-static struct bau_payload_queue_entry * __init
-uv_payload_queue_init(int node, int pnode, struct bau_control *bau_tablesp)
+static void
+uv_payload_queue_init(int node, int pnode)
 {
-	struct bau_payload_queue_entry *pqp;
-	unsigned long pa;
 	int pn;
+	int cpu;
 	char *cp;
+	unsigned long pa;
+	struct bau_payload_queue_entry *pqp;
+	struct bau_payload_queue_entry *pqp_malloc;
+	struct bau_control *bcp;
 
 	pqp = (struct bau_payload_queue_entry *) kmalloc_node(
 		(DEST_Q_SIZE + 1) * sizeof(struct bau_payload_queue_entry),
 		GFP_KERNEL, node);
 	BUG_ON(!pqp);
+	pqp_malloc = pqp;
 
 	cp = (char *)pqp + 31;
 	pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5);
-	bau_tablesp->va_queue_first = pqp;
+
+	for_each_present_cpu(cpu) {
+		if (pnode != uv_cpu_to_pnode(cpu))
+			continue;
+		/* for every cpu on this pnode: */
+		bcp = &per_cpu(bau_control, cpu);
+		bcp->va_queue_first = pqp;
+		bcp->bau_msg_head = pqp;
+		bcp->va_queue_last = pqp + (DEST_Q_SIZE - 1);
+	}
 	/*
 	 * need the pnode of where the memory was really allocated
 	 */
 	pa = uv_gpa(pqp);
-	pn = uv_gpa_to_pnode(pa);
+	pn = pa >> uv_nshift;
 	uv_write_global_mmr64(pnode,
 			      UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST,
 			      ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) |
 			      uv_physnodeaddr(pqp));
 	uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL,
 			      uv_physnodeaddr(pqp));
-	bau_tablesp->va_queue_last = pqp + (DEST_Q_SIZE - 1);
 	uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST,
 			      (unsigned long)
-			      uv_physnodeaddr(bau_tablesp->va_queue_last));
+			      uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1)));
+	/* in effect, all msg_type's are set to MSG_NOOP */
 	memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE);
-
-	return pqp;
 }
 
 /*
- * Initialization of each UV blade's structures
+ * Initialization of each UV hub's structures
  */
-static int __init uv_init_blade(int blade)
+static void __init uv_init_uvhub(int uvhub, int vector)
 {
 	int node;
 	int pnode;
-	unsigned long pa;
 	unsigned long apicid;
-	struct bau_desc *adp;
-	struct bau_payload_queue_entry *pqp;
-	struct bau_control *bau_tablesp;
-
-	node = blade_to_first_node(blade);
-	bau_tablesp = uv_table_bases_init(blade, node);
-	pnode = uv_blade_to_pnode(blade);
-	adp = uv_activation_descriptor_init(node, pnode);
-	pqp = uv_payload_queue_init(node, pnode, bau_tablesp);
-	uv_table_bases_finish(blade, bau_tablesp, adp);
+
+	node = uvhub_to_first_node(uvhub);
+	pnode = uv_blade_to_pnode(uvhub);
+	uv_activation_descriptor_init(node, pnode);
+	uv_payload_queue_init(node, pnode);
 	/*
 	 * the below initialization can't be in firmware because the
 	 * messaging IRQ will be determined by the OS
 	 */
-	apicid = blade_to_first_apicid(blade);
-	pa = uv_read_global_mmr64(pnode, UVH_BAU_DATA_CONFIG);
+	apicid = uvhub_to_first_apicid(uvhub);
 	uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG,
-				      ((apicid << 32) | UV_BAU_MESSAGE));
-	return 0;
+				      ((apicid << 32) | vector));
+}
+
+/*
+ * initialize the bau_control structure for each cpu
+ */
+static void uv_init_per_cpu(int nuvhubs)
+{
+	int i, j, k;
+	int cpu;
+	int pnode;
+	int uvhub;
+	short socket = 0;
+	struct bau_control *bcp;
+	struct uvhub_desc *bdp;
+	struct socket_desc *sdp;
+	struct bau_control *hmaster = NULL;
+	struct bau_control *smaster = NULL;
+	struct socket_desc {
+		short num_cpus;
+		short cpu_number[16];
+	};
+	struct uvhub_desc {
+		short num_sockets;
+		short num_cpus;
+		short uvhub;
+		short pnode;
+		struct socket_desc socket[2];
+	};
+	struct uvhub_desc *uvhub_descs;
+
+	uvhub_descs = (struct uvhub_desc *)
+		kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
+	memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
+	for_each_present_cpu(cpu) {
+		bcp = &per_cpu(bau_control, cpu);
+		memset(bcp, 0, sizeof(struct bau_control));
+		spin_lock_init(&bcp->masks_lock);
+		bcp->max_concurrent = uv_bau_max_concurrent;
+		pnode = uv_cpu_hub_info(cpu)->pnode;
+		uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
+		bdp = &uvhub_descs[uvhub];
+		bdp->num_cpus++;
+		bdp->uvhub = uvhub;
+		bdp->pnode = pnode;
+		/* time interval to catch a hardware stay-busy bug */
+		bcp->timeout_interval = millisec_2_cycles(3);
+		/* kludge: assume uv_hub.h is constant */
+		socket = (cpu_physical_id(cpu)>>5)&1;
+		if (socket >= bdp->num_sockets)
+			bdp->num_sockets = socket+1;
+		sdp = &bdp->socket[socket];
+		sdp->cpu_number[sdp->num_cpus] = cpu;
+		sdp->num_cpus++;
+	}
+	socket = 0;
+	for_each_possible_blade(uvhub) {
+		bdp = &uvhub_descs[uvhub];
+		for (i = 0; i < bdp->num_sockets; i++) {
+			sdp = &bdp->socket[i];
+			for (j = 0; j < sdp->num_cpus; j++) {
+				cpu = sdp->cpu_number[j];
+				bcp = &per_cpu(bau_control, cpu);
+				bcp->cpu = cpu;
+				if (j == 0) {
+					smaster = bcp;
+					if (i == 0)
+						hmaster = bcp;
+				}
+				bcp->cpus_in_uvhub = bdp->num_cpus;
+				bcp->cpus_in_socket = sdp->num_cpus;
+				bcp->socket_master = smaster;
+				bcp->uvhub_master = hmaster;
+				for (k = 0; k < DEST_Q_SIZE; k++)
+					bcp->socket_acknowledge_count[k] = 0;
+				bcp->uvhub_cpu =
+				  uv_cpu_hub_info(cpu)->blade_processor_id;
+			}
+			socket++;
+		}
+	}
+	kfree(uvhub_descs);
 }
 
 /*
@@ -828,38 +1344,54 @@ static int __init uv_init_blade(int blade)
  */
 static int __init uv_bau_init(void)
 {
-	int blade;
-	int nblades;
+	int uvhub;
+	int pnode;
+	int nuvhubs;
 	int cur_cpu;
+	int vector;
+	unsigned long mmr;
 
 	if (!is_uv_system())
 		return 0;
 
+	if (nobau)
+		return 0;
+
 	for_each_possible_cpu(cur_cpu)
 		zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu),
 				       GFP_KERNEL, cpu_to_node(cur_cpu));
 
-	uv_bau_retry_limit = 1;
+	uv_bau_max_concurrent = MAX_BAU_CONCURRENT;
+	uv_nshift = uv_hub_info->m_val;
 	uv_mmask = (1UL << uv_hub_info->m_val) - 1;
-	nblades = uv_num_possible_blades();
+	nuvhubs = uv_num_possible_blades();
 
-	uv_bau_table_bases = (struct bau_control **)
-	    kmalloc(nblades * sizeof(struct bau_control *), GFP_KERNEL);
-	BUG_ON(!uv_bau_table_bases);
+	uv_init_per_cpu(nuvhubs);
 
 	uv_partition_base_pnode = 0x7fffffff;
-	for (blade = 0; blade < nblades; blade++)
-		if (uv_blade_nr_possible_cpus(blade) &&
-			(uv_blade_to_pnode(blade) < uv_partition_base_pnode))
-			uv_partition_base_pnode = uv_blade_to_pnode(blade);
-	for (blade = 0; blade < nblades; blade++)
-		if (uv_blade_nr_possible_cpus(blade))
-			uv_init_blade(blade);
-
-	alloc_intr_gate(UV_BAU_MESSAGE, uv_bau_message_intr1);
+	for (uvhub = 0; uvhub < nuvhubs; uvhub++)
+		if (uv_blade_nr_possible_cpus(uvhub) &&
+			(uv_blade_to_pnode(uvhub) < uv_partition_base_pnode))
+			uv_partition_base_pnode = uv_blade_to_pnode(uvhub);
+
+	vector = UV_BAU_MESSAGE;
+	for_each_possible_blade(uvhub)
+		if (uv_blade_nr_possible_cpus(uvhub))
+			uv_init_uvhub(uvhub, vector);
+
 	uv_enable_timeouts();
+	alloc_intr_gate(vector, uv_bau_message_intr1);
+
+	for_each_possible_blade(uvhub) {
+		pnode = uv_blade_to_pnode(uvhub);
+		/* INIT the bau */
+		uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_ACTIVATION_CONTROL,
+				      ((unsigned long)1 << 63));
+		mmr = 1; /* should be 1 to broadcast to both sockets */
+		uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST, mmr);
+	}
 
 	return 0;
 }
-__initcall(uv_bau_init);
-__initcall(uv_ptc_init);
+core_initcall(uv_bau_init);
+core_initcall(uv_ptc_init);
diff --git a/arch/x86/kernel/uv_irq.c b/arch/x86/kernel/uv_irq.c
index 1d40336b030..1132129db79 100644
--- a/arch/x86/kernel/uv_irq.c
+++ b/arch/x86/kernel/uv_irq.c
@@ -44,7 +44,7 @@ static void uv_ack_apic(unsigned int irq)
 	ack_APIC_irq();
 }
 
-struct irq_chip uv_irq_chip = {
+static struct irq_chip uv_irq_chip = {
 	.name		= "UV-CORE",
 	.startup	= uv_noop_ret,
 	.shutdown	= uv_noop,
@@ -141,7 +141,7 @@ int uv_irq_2_mmr_info(int irq, unsigned long *offset, int *pnode)
  */
 static int
 arch_enable_uv_irq(char *irq_name, unsigned int irq, int cpu, int mmr_blade,
-		       unsigned long mmr_offset, int restrict)
+		       unsigned long mmr_offset, int limit)
 {
 	const struct cpumask *eligible_cpu = cpumask_of(cpu);
 	struct irq_desc *desc = irq_to_desc(irq);
@@ -160,7 +160,7 @@ arch_enable_uv_irq(char *irq_name, unsigned int irq, int cpu, int mmr_blade,
 	if (err != 0)
 		return err;
 
-	if (restrict == UV_AFFINITY_CPU)
+	if (limit == UV_AFFINITY_CPU)
 		desc->status |= IRQ_NO_BALANCING;
 	else
 		desc->status |= IRQ_MOVE_PCNTXT;
@@ -214,7 +214,7 @@ static int uv_set_irq_affinity(unsigned int irq, const struct cpumask *mask)
 	unsigned long mmr_value;
 	struct uv_IO_APIC_route_entry *entry;
 	unsigned long mmr_offset;
-	unsigned mmr_pnode;
+	int mmr_pnode;
 
 	if (set_desc_affinity(desc, mask, &dest))
 		return -1;
@@ -248,7 +248,7 @@ static int uv_set_irq_affinity(unsigned int irq, const struct cpumask *mask)
  * interrupt is raised.
  */
 int uv_setup_irq(char *irq_name, int cpu, int mmr_blade,
-		 unsigned long mmr_offset, int restrict)
+		 unsigned long mmr_offset, int limit)
 {
 	int irq, ret;
 
@@ -258,7 +258,7 @@ int uv_setup_irq(char *irq_name, int cpu, int mmr_blade,
 		return -EBUSY;
 
 	ret = arch_enable_uv_irq(irq_name, irq, cpu, mmr_blade, mmr_offset,
-		restrict);
+		limit);
 	if (ret == irq)
 		uv_set_irq_2_mmr_info(irq, mmr_offset, mmr_blade);
 	else