x86, UV: Improve BAU performance and error recovery

- increase performance of the interrupt handler

- release timed-out software acknowledge resources

- recover from continuous-busy status due to a hardware issue

- add a 'throttle' to keep a uvhub from sending more than a
  specified number of broadcasts concurrently (work around the hardware issue)

- provide a 'nobau' boot command line option

- rename 'pnode' and 'node' to 'uvhub' (the 'node' terminology
  is ambiguous)

- add some new statistics about the scope of broadcasts, retries, the
  hardware issue and the 'throttle'

- split off new function uv_bau_retry_msg() from
  uv_bau_process_message() per community coding style feedback.

- simplify the argument list to uv_bau_process_message(), per
  community coding style feedback.

Signed-off-by: Cliff Wickman <cpw@sgi.com>
Cc: linux-mm@kvack.org
Cc: Jack Steiner <steiner@sgi.com>
Cc: Russ Anderson <rja@sgi.com>
Cc: Mike Travis <travis@sgi.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <E1O25Z4-0004Ur-PB@eag09.americas.sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

2 files changed, 1075 insertions, 442 deletions

diff --git a/arch/x86/include/asm/uv/uv_bau.h b/arch/x86/include/asm/uv/uv_bau.h
index b414d2b401f..aa558ac0306 100644
--- a/arch/x86/include/asm/uv/uv_bau.h
+++ b/arch/x86/include/asm/uv/uv_bau.h
@@ -27,13 +27,14 @@
  * set 2 is at BASE + 2*512, set 3 at BASE + 3*512, and so on.
  *
  * We will use 31 sets, one for sending BAU messages from each of the 32
- * cpu's on the node.
+ * cpu's on the uvhub.
  *
  * TLB shootdown will use the first of the 8 descriptors of each set.
  * Each of the descriptors is 64 bytes in size (8*64 = 512 bytes in a set).
  */
 
 #define UV_ITEMS_PER_DESCRIPTOR		8
+#define MAX_BAU_CONCURRENT		3
 #define UV_CPUS_PER_ACT_STATUS		32
 #define UV_ACT_STATUS_MASK		0x3
 #define UV_ACT_STATUS_SIZE		2
@@ -45,6 +46,9 @@
 #define UV_PAYLOADQ_PNODE_SHIFT		49
 #define UV_PTC_BASENAME			"sgi_uv/ptc_statistics"
 #define uv_physnodeaddr(x)		((__pa((unsigned long)(x)) & uv_mmask))
+#define UV_ENABLE_INTD_SOFT_ACK_MODE_SHIFT 15
+#define UV_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHIFT 16
+#define UV_INTD_SOFT_ACK_TIMEOUT_PERIOD 0x000000000bUL
 
 /*
  * bits in UVH_LB_BAU_SB_ACTIVATION_STATUS_0/1
@@ -55,15 +59,29 @@
 #define DESC_STATUS_SOURCE_TIMEOUT	3
 
 /*
- * source side thresholds at which message retries print a warning
+ * source side threshholds at which message retries print a warning
  */
 #define SOURCE_TIMEOUT_LIMIT		20
 #define DESTINATION_TIMEOUT_LIMIT	20
 
 /*
+ * misc. delays, in microseconds
+ */
+#define THROTTLE_DELAY			10
+#define TIMEOUT_DELAY			10
+#define BIOS_TO				1000
+/* BIOS is assumed to set the destination timeout to 1003520 nanoseconds */
+
+/*
+ * threshholds at which to use IPI to free resources
+ */
+#define PLUGSB4RESET 100
+#define TIMEOUTSB4RESET 100
+
+/*
  * number of entries in the destination side payload queue
  */
-#define DEST_Q_SIZE			17
+#define DEST_Q_SIZE			20
 /*
  * number of destination side software ack resources
  */
@@ -72,9 +90,10 @@
 /*
  * completion statuses for sending a TLB flush message
  */
-#define	FLUSH_RETRY			1
-#define	FLUSH_GIVEUP			2
-#define	FLUSH_COMPLETE			3
+#define FLUSH_RETRY_PLUGGED		1
+#define FLUSH_RETRY_TIMEOUT		2
+#define FLUSH_GIVEUP			3
+#define FLUSH_COMPLETE			4
 
 /*
  * Distribution: 32 bytes (256 bits) (bytes 0-0x1f of descriptor)
@@ -86,14 +105,14 @@
  * 'base_dest_nodeid' field of the header corresponds to the
  * destination nodeID associated with that specified bit.
  */
-struct bau_target_nodemask {
-	unsigned long bits[BITS_TO_LONGS(256)];
+struct bau_target_uvhubmask {
+	unsigned long bits[BITS_TO_LONGS(UV_DISTRIBUTION_SIZE)];
 };
 
 /*
- * mask of cpu's on a node
+ * mask of cpu's on a uvhub
  * (during initialization we need to check that unsigned long has
- *  enough bits for max. cpu's per node)
+ *  enough bits for max. cpu's per uvhub)
  */
 struct bau_local_cpumask {
 	unsigned long bits;
@@ -135,8 +154,8 @@ struct bau_msg_payload {
 struct bau_msg_header {
 	unsigned int dest_subnodeid:6;	/* must be 0x10, for the LB */
 	/* bits 5:0 */
-	unsigned int base_dest_nodeid:15; /* nasid>>1 (pnode) of */
-	/* bits 20:6 */			  /* first bit in node_map */
+	unsigned int base_dest_nodeid:15; /* nasid (pnode<<1) of */
+	/* bits 20:6 */			  /* first bit in uvhub map */
 	unsigned int command:8;	/* message type */
 	/* bits 28:21 */
 				/* 0x38: SN3net EndPoint Message */
@@ -146,26 +165,38 @@ struct bau_msg_header {
 	unsigned int rsvd_2:9;	/* must be zero */
 	/* bits 40:32 */
 				/* Suppl_A is 56-41 */
-	unsigned int payload_2a:8;/* becomes byte 16 of msg */
-	/* bits 48:41 */	/* not currently using */
-	unsigned int payload_2b:8;/* becomes byte 17 of msg */
-	/* bits 56:49 */	/* not currently using */
+	unsigned int sequence:16;/* message sequence number */
+	/* bits 56:41 */	/* becomes bytes 16-17 of msg */
 				/* Address field (96:57) is never used as an
 				   address (these are address bits 42:3) */
+
 	unsigned int rsvd_3:1;	/* must be zero */
 	/* bit 57 */
 				/* address bits 27:4 are payload */
-				/* these 24 bits become bytes 12-14 of msg */
+	/* these next 24  (58-81) bits become bytes 12-14 of msg */
+
+	/* bits 65:58 land in byte 12 */
 	unsigned int replied_to:1;/* sent as 0 by the source to byte 12 */
 	/* bit 58 */
-
-	unsigned int payload_1a:5;/* not currently used */
-	/* bits 63:59 */
-	unsigned int payload_1b:8;/* not currently used */
-	/* bits 71:64 */
-	unsigned int payload_1c:8;/* not currently used */
-	/* bits 79:72 */
-	unsigned int payload_1d:2;/* not currently used */
+	unsigned int msg_type:3; /* software type of the message*/
+	/* bits 61:59 */
+	unsigned int canceled:1; /* message canceled, resource to be freed*/
+	/* bit 62 */
+	unsigned int payload_1a:1;/* not currently used */
+	/* bit 63 */
+	unsigned int payload_1b:2;/* not currently used */
+	/* bits 65:64 */
+
+	/* bits 73:66 land in byte 13 */
+	unsigned int payload_1ca:6;/* not currently used */
+	/* bits 71:66 */
+	unsigned int payload_1c:2;/* not currently used */
+	/* bits 73:72 */
+
+	/* bits 81:74 land in byte 14 */
+	unsigned int payload_1d:6;/* not currently used */
+	/* bits 79:74 */
+	unsigned int payload_1e:2;/* not currently used */
 	/* bits 81:80 */
 
 	unsigned int rsvd_4:7;	/* must be zero */
@@ -178,7 +209,7 @@ struct bau_msg_header {
 	/* bits 95:90 */
 	unsigned int rsvd_6:5;	/* must be zero */
 	/* bits 100:96 */
-	unsigned int int_both:1;/* if 1, interrupt both sockets on the blade */
+	unsigned int int_both:1;/* if 1, interrupt both sockets on the uvhub */
 	/* bit 101*/
 	unsigned int fairness:3;/* usually zero */
 	/* bits 104:102 */
@@ -191,13 +222,18 @@ struct bau_msg_header {
 	/* bits 127:107 */
 };
 
+/* see msg_type: */
+#define MSG_NOOP 0
+#define MSG_REGULAR 1
+#define MSG_RETRY 2
+
 /*
  * The activation descriptor:
  * The format of the message to send, plus all accompanying control
  * Should be 64 bytes
  */
 struct bau_desc {
-	struct bau_target_nodemask distribution;
+	struct bau_target_uvhubmask distribution;
 	/*
 	 * message template, consisting of header and payload:
 	 */
@@ -237,19 +273,25 @@ struct bau_payload_queue_entry {
 	unsigned short acknowledge_count; /* filled in by destination */
 	/* 16 bits, bytes 10-11 */
 
-	unsigned short replied_to:1;	/* sent as 0 by the source */
-	/* 1 bit */
-	unsigned short unused1:7;       /* not currently using */
-	/* 7 bits: byte 12) */
+	/* these next 3 bytes come from bits 58-81 of the message header */
+	unsigned short replied_to:1;    /* sent as 0 by the source */
+	unsigned short msg_type:3;      /* software message type */
+	unsigned short canceled:1;      /* sent as 0 by the source */
+	unsigned short unused1:3;       /* not currently using */
+	/* byte 12 */
 
-	unsigned char unused2[2];	/* not currently using */
-	/* bytes 13-14 */
+	unsigned char unused2a;		/* not currently using */
+	/* byte 13 */
+	unsigned char unused2;		/* not currently using */
+	/* byte 14 */
 
 	unsigned char sw_ack_vector;	/* filled in by the hardware */
 	/* byte 15 (bits 127:120) */
 
-	unsigned char unused4[3];	/* not currently using bytes 17-19 */
-	/* bytes 17-19 */
+	unsigned short sequence;	/* message sequence number */
+	/* bytes 16-17 */
+	unsigned char unused4[2];	/* not currently using bytes 18-19 */
+	/* bytes 18-19 */
 
 	int number_of_cpus;		/* filled in at destination */
 	/* 32 bits, bytes 20-23 (aligned) */
@@ -259,63 +301,93 @@ struct bau_payload_queue_entry {
 };
 
 /*
- * one for every slot in the destination payload queue
- */
-struct bau_msg_status {
-	struct bau_local_cpumask seen_by;	/* map of cpu's */
-};
-
-/*
- * one for every slot in the destination software ack resources
- */
-struct bau_sw_ack_status {
-	struct bau_payload_queue_entry *msg;	/* associated message */
-	int watcher;				/* cpu monitoring, or -1 */
-};
-
-/*
- * one on every node and per-cpu; to locate the software tables
+ * one per-cpu; to locate the software tables
  */
 struct bau_control {
 	struct bau_desc *descriptor_base;
-	struct bau_payload_queue_entry *bau_msg_head;
 	struct bau_payload_queue_entry *va_queue_first;
 	struct bau_payload_queue_entry *va_queue_last;
-	struct bau_msg_status *msg_statuses;
-	int *watching; /* pointer to array */
+	struct bau_payload_queue_entry *bau_msg_head;
+	struct bau_control *uvhub_master;
+	struct bau_control *socket_master;
+	unsigned long timeout_interval;
+	atomic_t active_descriptor_count;
+	int max_concurrent;
+	int max_concurrent_constant;
+	int retry_message_scans;
+	int plugged_tries;
+	int timeout_tries;
+	int ipi_attempts;
+	int conseccompletes;
+	short cpu;
+	short uvhub_cpu;
+	short uvhub;
+	short cpus_in_socket;
+	short cpus_in_uvhub;
+	unsigned short message_number;
+	unsigned short uvhub_quiesce;
+	short socket_acknowledge_count[DEST_Q_SIZE];
+	cycles_t send_message;
+	spinlock_t masks_lock;
+	spinlock_t uvhub_lock;
+	spinlock_t queue_lock;
 };
 
 /*
  * This structure is allocated per_cpu for UV TLB shootdown statistics.
  */
 struct ptc_stats {
-	unsigned long ptc_i;	/* number of IPI-style flushes */
-	unsigned long requestor;	/* number of nodes this cpu sent to */
-	unsigned long requestee;	/* times cpu was remotely requested */
-	unsigned long alltlb;	/* times all tlb's on this cpu were flushed */
-	unsigned long onetlb;	/* times just one tlb on this cpu was flushed */
-	unsigned long s_retry;	/* retries on source side timeouts */
-	unsigned long d_retry;	/* retries on destination side timeouts */
-	unsigned long sflush;	/* cycles spent in uv_flush_tlb_others */
-	unsigned long dflush;	/* cycles spent on destination side */
-	unsigned long retriesok; /* successes on retries */
-	unsigned long nomsg;	/* interrupts with no message */
-	unsigned long multmsg;	/* interrupts with multiple messages */
-	unsigned long ntargeted;/* nodes targeted */
+	/* sender statistics */
+	unsigned long s_giveup; /* number of fall backs to IPI-style flushes */
+	unsigned long s_requestor; /* number of shootdown requests */
+	unsigned long s_stimeout; /* source side timeouts */
+	unsigned long s_dtimeout; /* destination side timeouts */
+	unsigned long s_time; /* time spent in sending side */
+	unsigned long s_retriesok; /* successful retries */
+	unsigned long s_ntargcpu; /* number of cpus targeted */
+	unsigned long s_ntarguvhub; /* number of uvhubs targeted */
+	unsigned long s_ntarguvhub16; /* number of times >= 16 target hubs */
+	unsigned long s_ntarguvhub8; /* number of times >= 8 target hubs */
+	unsigned long s_ntarguvhub4; /* number of times >= 4 target hubs */
+	unsigned long s_ntarguvhub2; /* number of times >= 2 target hubs */
+	unsigned long s_ntarguvhub1; /* number of times == 1 target hub */
+	unsigned long s_resets_plug; /* ipi-style resets from plug state */
+	unsigned long s_resets_timeout; /* ipi-style resets from timeouts */
+	unsigned long s_busy; /* status stayed busy past s/w timer */
+	unsigned long s_throttles; /* waits in throttle */
+	unsigned long s_retry_messages; /* retry broadcasts */
+	/* destination statistics */
+	unsigned long d_alltlb; /* times all tlb's on this cpu were flushed */
+	unsigned long d_onetlb; /* times just one tlb on this cpu was flushed */
+	unsigned long d_multmsg; /* interrupts with multiple messages */
+	unsigned long d_nomsg; /* interrupts with no message */
+	unsigned long d_time; /* time spent on destination side */
+	unsigned long d_requestee; /* number of messages processed */
+	unsigned long d_retries; /* number of retry messages processed */
+	unsigned long d_canceled; /* number of messages canceled by retries */
+	unsigned long d_nocanceled; /* retries that found nothing to cancel */
+	unsigned long d_resets; /* number of ipi-style requests processed */
+	unsigned long d_rcanceled; /* number of messages canceled by resets */
 };
 
-static inline int bau_node_isset(int node, struct bau_target_nodemask *dstp)
+static inline int bau_uvhub_isset(int uvhub, struct bau_target_uvhubmask *dstp)
 {
-	return constant_test_bit(node, &dstp->bits[0]);
+	return constant_test_bit(uvhub, &dstp->bits[0]);
 }
-static inline void bau_node_set(int node, struct bau_target_nodemask *dstp)
+static inline void bau_uvhub_set(int uvhub, struct bau_target_uvhubmask *dstp)
 {
-	__set_bit(node, &dstp->bits[0]);
+	__set_bit(uvhub, &dstp->bits[0]);
 }
-static inline void bau_nodes_clear(struct bau_target_nodemask *dstp, int nbits)
+static inline void bau_uvhubs_clear(struct bau_target_uvhubmask *dstp,
+				    int nbits)
 {
 	bitmap_zero(&dstp->bits[0], nbits);
 }
+static inline int bau_uvhub_weight(struct bau_target_uvhubmask *dstp)
+{
+	return bitmap_weight((unsigned long *)&dstp->bits[0],
+				UV_DISTRIBUTION_SIZE);
+}
 
 static inline void bau_cpubits_clear(struct bau_local_cpumask *dstp, int nbits)
 {
@@ -328,4 +400,35 @@ static inline void bau_cpubits_clear(struct bau_local_cpumask *dstp, int nbits)
 extern void uv_bau_message_intr1(void);
 extern void uv_bau_timeout_intr1(void);
 
+struct atomic_short {
+	short counter;
+};
+
+/**
+ * atomic_read_short - read a short atomic variable
+ * @v: pointer of type atomic_short
+ *
+ * Atomically reads the value of @v.
+ */
+static inline int atomic_read_short(const struct atomic_short *v)
+{
+	return v->counter;
+}
+
+/**
+ * atomic_add_short_return - add and return a short int
+ * @i: short value to add
+ * @v: pointer of type atomic_short
+ *
+ * Atomically adds @i to @v and returns @i + @v
+ */
+static inline int atomic_add_short_return(short i, struct atomic_short *v)
+{
+	short __i = i;
+	asm volatile(LOCK_PREFIX "xaddw %0, %1"
+			: "+r" (i), "+m" (v->counter)
+			: : "memory");
+	return i + __i;
+}
+
 #endif /* _ASM_X86_UV_UV_BAU_H */
diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c
index ef68ba48564..414f7c4fe76 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.
@@ -19,44 +19,67 @@
 #include <asm/idle.h>
 #include <asm/tsc.h>
 #include <asm/irq_vectors.h>
+#include <asm/timer.h>
+
+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;
+};
 
 #define UV_INTD_SOFT_ACK_TIMEOUT_PERIOD	0x000000000bUL
 
-static struct bau_control	**uv_bau_table_bases __read_mostly;
-static int			uv_bau_retry_limit __read_mostly;
+static int uv_bau_max_concurrent __read_mostly;
 
-/* base pnode in this partition */
-static int			uv_partition_base_pnode __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;
 }
@@ -69,195 +92,459 @@ 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;
 
-	sender = smp_processor_id();
-	for (i = 0; i < sizeof(struct bau_target_nodemask) * BITSPERBYTE; i++) {
-		if (!bau_node_isset(i, distribution))
+	reset_args.sender = sender;
+
+	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;
@@ -269,24 +556,108 @@ const struct cpumask *uv_flush_send_and_wait(int cpu, int this_pnode,
 	}
 	time