Merge branch 'x86-uv-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
authorLinus Torvalds <torvalds@linux-foundation.org>
Tue, 18 May 2010 16:46:35 +0000 (09:46 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 18 May 2010 16:46:35 +0000 (09:46 -0700)
* '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

1  2 
arch/x86/kernel/apic/x2apic_uv_x.c
arch/x86/kernel/tlb_uv.c
arch/x86/kernel/uv_irq.c

index c085d52dbaf2f1305ea35c16e0877fb11abcb7e7,9a93b07fa14120c75b5d9202c717957723f60d54..e46f98f36e31f6e1b85211a87ea62de0c497cf6c
@@@ -17,7 -17,6 +17,7 @@@
  #include <linux/ctype.h>
  #include <linux/sched.h>
  #include <linux/timer.h>
 +#include <linux/slab.h>
  #include <linux/cpu.h>
  #include <linux/init.h>
  #include <linux/io.h>
@@@ -121,9 -120,11 +121,9 @@@ EXPORT_SYMBOL_GPL(uv_possible_blades)
  unsigned long sn_rtc_cycles_per_second;
  EXPORT_SYMBOL(sn_rtc_cycles_per_second);
  
 -/* Start with all IRQs pointing to boot CPU.  IRQ balancing will shift them. */
 -
  static const struct cpumask *uv_target_cpus(void)
  {
 -      return cpumask_of(0);
 +      return cpu_online_mask;
  }
  
  static void uv_vector_allocation_domain(int cpu, struct cpumask *retmask)
@@@ -735,9 -736,6 +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 --combined arch/x86/kernel/tlb_uv.c
index 17b03dd3a6b50f45cb87e7f4cfb1312414a35274,414f7c4fe76c9b9e762fb925cdc71154b992331f..7fea555929e24f2f411ea019b426e64c05ac2df0
@@@ -1,7 -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.
@@@ -9,7 -9,6 +9,7 @@@
  #include <linux/seq_file.h>
  #include <linux/proc_fs.h>
  #include <linux/kernel.h>
 +#include <linux/slab.h>
  
  #include <asm/mmu_context.h>
  #include <asm/uv/uv.h>
  #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;
  }
   * 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;
        }
        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;
        }
  
         * 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
   * 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 -706,82 +707,82 @@@ const struct cpumask *uv_flush_tlb_othe
                                          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);
  }
  
  /*
   *
   * 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);
                /*
                 * 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);
                /*
                 * 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 -906,20 +907,20 @@@ static void uv_ptc_seq_stop(struct seq_
  {
  }
  
+ 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)
  {
  
        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;
@@@ -649,80 -1121,31 +1122,31 @@@ static int __init uv_ptc_init(void
        return 0;
  }
  
- /*
-  * 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,
        /*
         * 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);
  }
  
  /*
   */
  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 --combined arch/x86/kernel/uv_irq.c
index 1d40336b030adc9206a48cc0a5e39ba0c22dcf7e,1a9f55a33489630848a9e7116275ec5fd75ab804..1132129db792b704fe14f4adaf02fda8ffafaa24
@@@ -10,7 -10,6 +10,7 @@@
  
  #include <linux/module.h>
  #include <linux/rbtree.h>
 +#include <linux/slab.h>
  #include <linux/irq.h>
  
  #include <asm/apic.h>
@@@ -44,7 -43,7 +44,7 @@@ static void uv_ack_apic(unsigned int ir
        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 -140,7 +141,7 @@@ int uv_irq_2_mmr_info(int irq, unsigne
   */
  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);
        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 -213,7 +214,7 @@@ static int uv_set_irq_affinity(unsigne
        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;
   * 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;
  
                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