arch/x86/kernel/tlb_uv.c

   1 /*
   2  *      SGI UltraViolet TLB flush routines.
   3  *
   4  *      (c) 2008-2010 Cliff Wickman <cpw@sgi.com>, SGI.
   5  *
   6  *      This code is released under the GNU General Public License version 2 or
   7  *      later.
   8  */
   9 #include <linux/seq_file.h>
  10 #include <linux/proc_fs.h>
  11 #include <linux/kernel.h>
  12 #include <linux/slab.h>
  13
  14 #include <asm/mmu_context.h>
  15 #include <asm/uv/uv.h>
  16 #include <asm/uv/uv_mmrs.h>
  17 #include <asm/uv/uv_hub.h>
  18 #include <asm/uv/uv_bau.h>
  19 #include <asm/apic.h>
  20 #include <asm/idle.h>
  21 #include <asm/tsc.h>
  22 #include <asm/irq_vectors.h>
  23 #include <asm/timer.h>
  24
  25 struct msg_desc {
  26         struct bau_payload_queue_entry *msg;
  27         int msg_slot;
  28         int sw_ack_slot;
  29         struct bau_payload_queue_entry *va_queue_first;
  30         struct bau_payload_queue_entry *va_queue_last;
  31 };
  32
  33 #define UV_INTD_SOFT_ACK_TIMEOUT_PERIOD 0x000000000bUL
  34
  35 static int uv_bau_max_concurrent __read_mostly;
  36
  37 static int nobau;
  38 static int __init setup_nobau(char *arg)
  39 {
  40         nobau = 1;
  41         return 0;
  42 }
  43 early_param("nobau", setup_nobau);
  44
  45 /* base pnode in this partition */
  46 static int uv_partition_base_pnode __read_mostly;
  47 /* position of pnode (which is nasid>>1): */
  48 static int uv_nshift __read_mostly;
  49 static unsigned long uv_mmask __read_mostly;
  50
  51 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
  52 static DEFINE_PER_CPU(struct bau_control, bau_control);
  53 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
  54
  55 struct reset_args {
  56         int sender;
  57 };
  58
  59 /*
  60  * Determine the first node on a uvhub. 'Nodes' are used for kernel
  61  * memory allocation.
  62  */
  63 static int __init uvhub_to_first_node(int uvhub)
  64 {
  65         int node, b;
  66
  67         for_each_online_node(node) {
  68                 b = uv_node_to_blade_id(node);
  69                 if (uvhub == b)
  70                         return node;
  71         }
  72         return -1;
  73 }
  74
  75 /*
  76  * Determine the apicid of the first cpu on a uvhub.
  77  */
  78 static int __init uvhub_to_first_apicid(int uvhub)
  79 {
  80         int cpu;
  81
  82         for_each_present_cpu(cpu)
  83                 if (uvhub == uv_cpu_to_blade_id(cpu))
  84                         return per_cpu(x86_cpu_to_apicid, cpu);
  85         return -1;
  86 }
  87
  88 /*
  89  * Free a software acknowledge hardware resource by clearing its Pending
  90  * bit. This will return a reply to the sender.
  91  * If the message has timed out, a reply has already been sent by the
  92  * hardware but the resource has not been released. In that case our
  93  * clear of the Timeout bit (as well) will free the resource. No reply will
  94  * be sent (the hardware will only do one reply per message).
  95  */
  96 static inline void uv_reply_to_message(struct msg_desc *mdp,
  97                                        struct bau_control *bcp)
  98 {
  99         unsigned long dw;
 100         struct bau_payload_queue_entry *msg;
 101
 102         msg = mdp->msg;
 103         if (!msg->canceled) {
 104                 dw = (msg->sw_ack_vector << UV_SW_ACK_NPENDING) |
 105                                                 msg->sw_ack_vector;
 106                 uv_write_local_mmr(
 107                                 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw);
 108         }
 109         msg->replied_to = 1;
 110         msg->sw_ack_vector = 0;
 111 }
 112
 113 /*
 114  * Process the receipt of a RETRY message
 115  */
 116 static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
 117                                             struct bau_control *bcp)
 118 {
 119         int i;
 120         int cancel_count = 0;
 121         int slot2;
 122         unsigned long msg_res;
 123         unsigned long mmr = 0;
 124         struct bau_payload_queue_entry *msg;
 125         struct bau_payload_queue_entry *msg2;
 126         struct ptc_stats *stat;
 127
 128         msg = mdp->msg;
 129         stat = &per_cpu(ptcstats, bcp->cpu);
 130         stat->d_retries++;
 131         /*
 132          * cancel any message from msg+1 to the retry itself
 133          */
 134         for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
 135                 if (msg2 > mdp->va_queue_last)
 136                         msg2 = mdp->va_queue_first;
 137                 if (msg2 == msg)
 138                         break;
 139
 140                 /* same conditions for cancellation as uv_do_reset */
 141                 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
 142                     (msg2->sw_ack_vector) && ((msg2->sw_ack_vector &
 143                         msg->sw_ack_vector) == 0) &&
 144                     (msg2->sending_cpu == msg->sending_cpu) &&
 145                     (msg2->msg_type != MSG_NOOP)) {
 146                         slot2 = msg2 - mdp->va_queue_first;
 147                         mmr = uv_read_local_mmr
 148                                 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
 149                         msg_res = ((msg2->sw_ack_vector << 8) |
 150                                    msg2->sw_ack_vector);
 151                         /*
 152                          * This is a message retry; clear the resources held
 153                          * by the previous message only if they timed out.
 154                          * If it has not timed out we have an unexpected
 155                          * situation to report.
 156                          */
 157                         if (mmr & (msg_res << 8)) {
 158                                 /*
 159                                  * is the resource timed out?
 160                                  * make everyone ignore the cancelled message.
 161                                  */
 162                                 msg2->canceled = 1;
 163                                 stat->d_canceled++;
 164                                 cancel_count++;
 165                                 uv_write_local_mmr(
 166                                     UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
 167                                         (msg_res << 8) | msg_res);
 168                         } else
 169                                 printk(KERN_INFO "note bau retry: no effect\n");
 170                 }
 171         }
 172         if (!cancel_count)
 173                 stat->d_nocanceled++;
 174 }
 175
 176 /*
 177  * Do all the things a cpu should do for a TLB shootdown message.
 178  * Other cpu's may come here at the same time for this message.
 179  */
 180 static void uv_bau_process_message(struct msg_desc *mdp,
 181                                    struct bau_control *bcp)
 182 {
 183         int msg_ack_count;
 184         short socket_ack_count = 0;
 185         struct ptc_stats *stat;
 186         struct bau_payload_queue_entry *msg;
 187         struct bau_control *smaster = bcp->socket_master;
 188
 189         /*
 190          * This must be a normal message, or retry of a normal message
 191          */
 192         msg = mdp->msg;
 193         stat = &per_cpu(ptcstats, bcp->cpu);
 194         if (msg->address == TLB_FLUSH_ALL) {
 195                 local_flush_tlb();
 196                 stat->d_alltlb++;
 197         } else {
 198                 __flush_tlb_one(msg->address);
 199                 stat->d_onetlb++;
 200         }
 201         stat->d_requestee++;
 202
 203         /*
 204          * One cpu on each uvhub has the additional job on a RETRY
 205          * of releasing the resource held by the message that is
 206          * being retried.  That message is identified by sending
 207          * cpu number.
 208          */
 209         if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
 210                 uv_bau_process_retry_msg(mdp, bcp);
 211
 212         /*
 213          * This is a sw_ack message, so we have to reply to it.
 214          * Count each responding cpu on the socket. This avoids
 215          * pinging the count's cache line back and forth between
 216          * the sockets.
 217          */
 218         socket_ack_count = atomic_add_short_return(1, (struct atomic_short *)
 219                         &smaster->socket_acknowledge_count[mdp->msg_slot]);
 220         if (socket_ack_count == bcp->cpus_in_socket) {
 221                 /*
 222                  * Both sockets dump their completed count total into
 223                  * the message's count.
 224                  */
 225                 smaster->socket_acknowledge_count[mdp->msg_slot] = 0;
 226                 msg_ack_count = atomic_add_short_return(socket_ack_count,
 227                                 (struct atomic_short *)&msg->acknowledge_count);
 228
 229                 if (msg_ack_count == bcp->cpus_in_uvhub) {
 230                         /*
 231                          * All cpus in uvhub saw it; reply
 232                          */
 233                         uv_reply_to_message(mdp, bcp);
 234                 }
 235         }
 236
 237         return;
 238 }
 239
 240 /*
 241  * Determine the first cpu on a uvhub.
 242  */
 243 static int uvhub_to_first_cpu(int uvhub)
 244 {
 245         int cpu;
 246         for_each_present_cpu(cpu)
 247                 if (uvhub == uv_cpu_to_blade_id(cpu))
 248                         return cpu;
 249         return -1;
 250 }
 251
 252 /*
 253  * Last resort when we get a large number of destination timeouts is
 254  * to clear resources held by a given cpu.
 255  * Do this with IPI so that all messages in the BAU message queue
 256  * can be identified by their nonzero sw_ack_vector field.
 257  *
 258  * This is entered for a single cpu on the uvhub.
 259  * The sender want's this uvhub to free a specific message's
 260  * sw_ack resources.
 261  */
 262 static void
 263 uv_do_reset(void *ptr)
 264 {
 265         int i;
 266         int slot;
 267         int count = 0;
 268         unsigned long mmr;
 269         unsigned long msg_res;
 270         struct bau_control *bcp;
 271         struct reset_args *rap;
 272         struct bau_payload_queue_entry *msg;
 273         struct ptc_stats *stat;
 274
 275         bcp = &per_cpu(bau_control, smp_processor_id());
 276         rap = (struct reset_args *)ptr;
 277         stat = &per_cpu(ptcstats, bcp->cpu);
 278         stat->d_resets++;
 279
 280         /*
 281          * We're looking for the given sender, and
 282          * will free its sw_ack resource.
 283          * If all cpu's finally responded after the timeout, its
 284          * message 'replied_to' was set.
 285          */
 286         for (msg = bcp->va_queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
 287                 /* uv_do_reset: same conditions for cancellation as
 288                    uv_bau_process_retry_msg() */
 289                 if ((msg->replied_to == 0) &&
 290                     (msg->canceled == 0) &&
 291                     (msg->sending_cpu == rap->sender) &&
 292                     (msg->sw_ack_vector) &&
 293                     (msg->msg_type != MSG_NOOP)) {
 294                         /*
 295                          * make everyone else ignore this message
 296                          */
 297                         msg->canceled = 1;
 298                         slot = msg - bcp->va_queue_first;
 299                         count++;
 300                         /*
 301                          * only reset the resource if it is still pending
 302                          */
 303                         mmr = uv_read_local_mmr
 304                                         (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
 305                         msg_res = ((msg->sw_ack_vector << 8) |
 306                                                    msg->sw_ack_vector);
 307                         if (mmr & msg_res) {
 308                                 stat->d_rcanceled++;
 309                                 uv_write_local_mmr(
 310                                     UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
 311                                                         msg_res);
 312                         }
 313                 }
 314         }
 315         return;
 316 }
 317
 318 /*
 319  * Use IPI to get all target uvhubs to release resources held by
 320  * a given sending cpu number.
 321  */
 322 static void uv_reset_with_ipi(struct bau_target_uvhubmask *distribution,
 323                               int sender)
 324 {
 325         int uvhub;
 326         int cpu;
 327         cpumask_t mask;
 328         struct reset_args reset_args;
 329
 330         reset_args.sender = sender;
 331
 332         cpus_clear(mask);
 333         /* find a single cpu for each uvhub in this distribution mask */
 334         for (uvhub = 0;
 335                     uvhub < sizeof(struct bau_target_uvhubmask) * BITSPERBYTE;
 336                     uvhub++) {
 337                 if (!bau_uvhub_isset(uvhub, distribution))
 338                         continue;
 339                 /* find a cpu for this uvhub */
 340                 cpu = uvhub_to_first_cpu(uvhub);
 341                 cpu_set(cpu, mask);
 342         }
 343         /* IPI all cpus; Preemption is already disabled */
 344         smp_call_function_many(&mask, uv_do_reset, (void *)&reset_args, 1);
 345         return;
 346 }
 347
 348 static inline unsigned long
 349 cycles_2_us(unsigned long long cyc)
 350 {
 351         unsigned long long ns;
 352         unsigned long us;
 353         ns =  (cyc * per_cpu(cyc2ns, smp_processor_id()))
 354                                                 >> CYC2NS_SCALE_FACTOR;
 355         us = ns / 1000;
 356         return us;
 357 }
 358
 359 /*
 360  * wait for all cpus on this hub to finish their sends and go quiet
 361  * leaves uvhub_quiesce set so that no new broadcasts are started by
 362  * bau_flush_send_and_wait()
 363  */
 364 static inline void
 365 quiesce_local_uvhub(struct bau_control *hmaster)
 366 {
 367         atomic_add_short_return(1, (struct atomic_short *)
 368                  &hmaster->uvhub_quiesce);
 369 }
 370
 371 /*
 372  * mark this quiet-requestor as done
 373  */
 374 static inline void
 375 end_uvhub_quiesce(struct bau_control *hmaster)
 376 {
 377         atomic_add_short_return(-1, (struct atomic_short *)
 378                 &hmaster->uvhub_quiesce);
 379 }
 380
 381 /*
 382  * Wait for completion of a broadcast software ack message
 383  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
 384  */
 385 static int uv_wait_completion(struct bau_desc *bau_desc,
 386         unsigned long mmr_offset, int right_shift, int this_cpu,
 387         struct bau_control *bcp, struct bau_control *smaster, long try)
 388 {
 389         int relaxes = 0;
 390         unsigned long descriptor_status;
 391         unsigned long mmr;
 392         unsigned long mask;
 393         cycles_t ttime;
 394         cycles_t timeout_time;
 395         struct ptc_stats *stat = &per_cpu(ptcstats, this_cpu);
 396         struct bau_control *hmaster;
 397
 398         hmaster = bcp->uvhub_master;
 399         timeout_time = get_cycles() + bcp->timeout_interval;
 400
 401         /* spin on the status MMR, waiting for it to go idle */
 402         while ((descriptor_status = (((unsigned long)
 403                 uv_read_local_mmr(mmr_offset) >>
 404                         right_shift) & UV_ACT_STATUS_MASK)) !=
 405                         DESC_STATUS_IDLE) {
 406                 /*
 407                  * Our software ack messages may be blocked because there are
 408                  * no swack resources available.  As long as none of them
 409                  * has timed out hardware will NACK our message and its
 410                  * state will stay IDLE.
 411                  */
 412                 if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) {
 413                         stat->s_stimeout++;
 414                         return FLUSH_GIVEUP;
 415                 } else if (descriptor_status ==
 416                                         DESC_STATUS_DESTINATION_TIMEOUT) {
 417                         stat->s_dtimeout++;
 418                         ttime = get_cycles();
 419
 420                         /*
 421                          * Our retries may be blocked by all destination
 422                          * swack resources being consumed, and a timeout
 423                          * pending.  In that case hardware returns the
 424                          * ERROR that looks like a destination timeout.
 425                          */
 426                         if (cycles_2_us(ttime - bcp->send_message) < BIOS_TO) {
 427                                 bcp->conseccompletes = 0;
 428                                 return FLUSH_RETRY_PLUGGED;
 429                         }
 430
 431                         bcp->conseccompletes = 0;
 432                         return FLUSH_RETRY_TIMEOUT;
 433                 } else {
 434                         /*
 435                          * descriptor_status is still BUSY
 436                          */
 437                         cpu_relax();
 438                         relaxes++;
 439                         if (relaxes >= 10000) {
 440                                 relaxes = 0;
 441                                 if (get_cycles() > timeout_time) {
 442                                         quiesce_local_uvhub(hmaster);
 443
 444                                         /* single-thread the register change */
 445                                         spin_lock(&hmaster->masks_lock);
 446                                         mmr = uv_read_local_mmr(mmr_offset);
 447                                         mask = 0UL;
 448                                         mask |= (3UL < right_shift);
 449                                         mask = ~mask;
 450                                         mmr &= mask;
 451                                         uv_write_local_mmr(mmr_offset, mmr);
 452                                         spin_unlock(&hmaster->masks_lock);
 453                                         end_uvhub_quiesce(hmaster);
 454                                         stat->s_busy++;
 455                                         return FLUSH_GIVEUP;
 456                                 }
 457                         }
 458                 }
 459         }
 460         bcp->conseccompletes++;
 461         return FLUSH_COMPLETE;
 462 }
 463
 464 static inline cycles_t
 465 sec_2_cycles(unsigned long sec)
 466 {
 467         unsigned long ns;
 468         cycles_t cyc;
 469
 470         ns = sec * 1000000000;
 471         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
 472         return cyc;
 473 }
 474
 475 /*
 476  * conditionally add 1 to *v, unless *v is >= u
 477  * return 0 if we cannot add 1 to *v because it is >= u
 478  * return 1 if we can add 1 to *v because it is < u
 479  * the add is atomic
 480  *
 481  * This is close to atomic_add_unless(), but this allows the 'u' value
 482  * to be lowered below the current 'v'.  atomic_add_unless can only stop
 483  * on equal.
 484  */
 485 static inline int atomic_inc_unless_ge(spinlock_t *lock, atomic_t *v, int u)
 486 {
 487         spin_lock(lock);
 488         if (atomic_read(v) >= u) {
 489                 spin_unlock(lock);
 490                 return 0;
 491         }
 492         atomic_inc(v);
 493         spin_unlock(lock);
 494         return 1;
 495 }
 496
 497 /**
 498  * uv_flush_send_and_wait
 499  *
 500  * Send a broadcast and wait for it to complete.
 501  *
 502  * The flush_mask contains the cpus the broadcast is to be sent to, plus
 503  * cpus that are on the local uvhub.
 504  *
 505  * Returns NULL if all flushing represented in the mask was done. The mask
 506  * is zeroed.
 507  * Returns @flush_mask if some remote flushing remains to be done. The
 508  * mask will have some bits still set, representing any cpus on the local
 509  * uvhub (not current cpu) and any on remote uvhubs if the broadcast failed.
 510  */
 511 const struct cpumask *uv_flush_send_and_wait(struct bau_desc *bau_desc,
 512                                              struct cpumask *flush_mask,
 513                                              struct bau_control *bcp)
 514 {
 515         int right_shift;
 516         int uvhub;
 517         int bit;
 518         int completion_status = 0;
 519         int seq_number = 0;
 520         long try = 0;
 521         int cpu = bcp->uvhub_cpu;
 522         int this_cpu = bcp->cpu;
 523         int this_uvhub = bcp->uvhub;
 524         unsigned long mmr_offset;
 525         unsigned long index;
 526         cycles_t time1;
 527         cycles_t time2;
 528         struct ptc_stats *stat = &per_cpu(ptcstats, bcp->cpu);
 529         struct bau_control *smaster = bcp->socket_master;
 530         struct bau_control *hmaster = bcp->uvhub_master;
 531
 532         /*
 533          * Spin here while there are hmaster->max_concurrent or more active
 534          * descriptors. This is the per-uvhub 'throttle'.
 535          */
 536         if (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
 537                         &hmaster->active_descriptor_count,
 538                         hmaster->max_concurrent)) {
 539                 stat->s_throttles++;
 540                 do {
 541                         cpu_relax();
 542                 } while (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
 543                         &hmaster->active_descriptor_count,
 544                         hmaster->max_concurrent));
 545         }
 546
 547         while (hmaster->uvhub_quiesce)
 548                 cpu_relax();
 549
 550         if (cpu < UV_CPUS_PER_ACT_STATUS) {
 551                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 552                 right_shift = cpu * UV_ACT_STATUS_SIZE;
 553         } else {
 554                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
 555                 right_shift =
 556                     ((cpu - UV_CPUS_PER_ACT_STATUS) * UV_ACT_STATUS_SIZE);
 557         }
 558         time1 = get_cycles();
 559         do {
 560                 /*
 561                  * Every message from any given cpu gets a unique message
 562                  * sequence number. But retries use that same number.
 563                  * Our message may have timed out at the destination because
 564                  * all sw-ack resources are in use and there is a timeout
 565                  * pending there.  In that case, our last send never got
 566                  * placed into the queue and we need to persist until it
 567                  * does.
 568                  *
 569                  * Make any retry a type MSG_RETRY so that the destination will
 570                  * free any resource held by a previous message from this cpu.
 571                  */
 572                 if (try == 0) {
 573                         /* use message type set by the caller the first time */
 574                         seq_number = bcp->message_number++;
 575                 } else {
 576                         /* use RETRY type on all the rest; same sequence */
 577                         bau_desc->header.msg_type = MSG_RETRY;
 578                         stat->s_retry_messages++;
 579                 }
 580                 bau_desc->header.sequence = seq_number;
 581                 index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) |
 582                         bcp->uvhub_cpu;
 583                 bcp->send_message = get_cycles();
 584
 585                 uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index);
 586
 587                 try++;
 588                 completion_status = uv_wait_completion(bau_desc, mmr_offset,
 589                         right_shift, this_cpu, bcp, smaster, try);
 590
 591                 if (completion_status == FLUSH_RETRY_PLUGGED) {
 592                         /*
 593                          * Our retries may be blocked by all destination swack
 594                          * resources being consumed, and a timeout pending. In
 595                          * that case hardware immediately returns the ERROR
 596                          * that looks like a destination timeout.
 597                          */
 598                         udelay(TIMEOUT_DELAY);
 599                         bcp->plugged_tries++;
 600                         if (bcp->plugged_tries >= PLUGSB4RESET) {
 601                                 bcp->plugged_tries = 0;
 602                                 quiesce_local_uvhub(hmaster);
 603                                 spin_lock(&hmaster->queue_lock);
 604                                 uv_reset_with_ipi(&bau_desc->distribution,
 605                                                         this_cpu);
 606                                 spin_unlock(&hmaster->queue_lock);
 607                                 end_uvhub_quiesce(hmaster);
 608                                 bcp->ipi_attempts++;
 609                                 stat->s_resets_plug++;
 610                         }
 611                 } else if (completion_status == FLUSH_RETRY_TIMEOUT) {
 612                         hmaster->max_concurrent = 1;
 613                         bcp->timeout_tries++;
 614                         udelay(TIMEOUT_DELAY);
 615                         if (bcp->timeout_tries >= TIMEOUTSB4RESET) {
 616                                 bcp->timeout_tries = 0;
 617                                 quiesce_local_uvhub(hmaster);
 618                                 spin_lock(&hmaster->queue_lock);
 619                                 uv_reset_with_ipi(&bau_desc->distribution,
 620                                                                 this_cpu);
 621                                 spin_unlock(&hmaster->queue_lock);
 622                                 end_uvhub_quiesce(hmaster);
 623                                 bcp->ipi_attempts++;
 624                                 stat->s_resets_timeout++;
 625                         }
 626                 }
 627                 if (bcp->ipi_attempts >= 3) {
 628                         bcp->ipi_attempts = 0;
 629                         completion_status = FLUSH_GIVEUP;
 630                         break;
 631                 }
 632                 cpu_relax();
 633         } while ((completion_status == FLUSH_RETRY_PLUGGED) ||
 634                  (completion_status == FLUSH_RETRY_TIMEOUT));
 635         time2 = get_cycles();
 636
 637         if ((completion_status == FLUSH_COMPLETE) && (bcp->conseccompletes > 5)
 638             && (hmaster->max_concurrent < hmaster->max_concurrent_constant))
 639                         hmaster->max_concurrent++;
 640
 641         /*
 642          * hold any cpu not timing out here; no other cpu currently held by
 643          * the 'throttle' should enter the activation code
 644          */
 645         while (hmaster->uvhub_quiesce)
 646                 cpu_relax();
 647         atomic_dec(&hmaster->active_descriptor_count);
 648
 649         /* guard against cycles wrap */
 650         if (time2 > time1)
 651                 stat->s_time += (time2 - time1);
 652         else
 653                 stat->s_requestor--; /* don't count this one */
 654         if (completion_status == FLUSH_COMPLETE && try > 1)
 655                 stat->s_retriesok++;
 656         else if (completion_status == FLUSH_GIVEUP) {
 657                 /*
 658                  * Cause the caller to do an IPI-style TLB shootdown on
 659                  * the target cpu's, all of which are still in the mask.
 660                  */
 661                 stat->s_giveup++;
 662                 return flush_mask;
 663         }
 664
 665         /*
 666          * Success, so clear the remote cpu's from the mask so we don't
 667          * use the IPI method of shootdown on them.
 668          */
 669         for_each_cpu(bit, flush_mask) {
 670                 uvhub = uv_cpu_to_blade_id(bit);
 671                 if (uvhub == this_uvhub)
 672                         continue;
 673                 cpumask_clear_cpu(bit, flush_mask);
 674         }
 675         if (!cpumask_empty(flush_mask))
 676                 return flush_mask;
 677
 678         return NULL;
 679 }
 680
 681 /**
 682  * uv_flush_tlb_others - globally purge translation cache of a virtual
 683  * address or all TLB's
 684  * @cpumask: mask of all cpu's in which the address is to be removed
 685  * @mm: mm_struct containing virtual address range
 686  * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
 687  * @cpu: the current cpu
 688  *
 689  * This is the entry point for initiating any UV global TLB shootdown.
 690  *
 691  * Purges the translation caches of all specified processors of the given
 692  * virtual address, or purges all TLB's on specified processors.
 693  *
 694  * The caller has derived the cpumask from the mm_struct.  This function
 695  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
 696  *
 697  * The cpumask is converted into a uvhubmask of the uvhubs containing
 698  * those cpus.
 699  *
 700  * Note that this function should be called with preemption disabled.
 701  *
 702  * Returns NULL if all remote flushing was done.
 703  * Returns pointer to cpumask if some remote flushing remains to be
 704  * done.  The returned pointer is valid till preemption is re-enabled.
 705  */
 706 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
 707                                           struct mm_struct *mm,
 708                                           unsigned long va, unsigned int cpu)
 709 {
 710         int remotes;
 711         int tcpu;
 712         int uvhub;
 713         int locals = 0;
 714         struct bau_desc *bau_desc;
 715         struct cpumask *flush_mask;
 716         struct ptc_stats *stat;
 717         struct bau_control *bcp;
 718
 719         if (nobau)
 720                 return cpumask;
 721
 722         bcp = &per_cpu(bau_control, cpu);
 723         /*
 724          * Each sending cpu has a per-cpu mask which it fills from the caller's
 725          * cpu mask.  Only remote cpus are converted to uvhubs and copied.
 726          */
 727         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
 728         /*
 729          * copy cpumask to flush_mask, removing current cpu
 730          * (current cpu should already have been flushed by the caller and
 731          *  should never be returned if we return flush_mask)
 732          */
 733         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
 734         if (cpu_isset(cpu, *cpumask))
 735                 locals++;  /* current cpu was targeted */
 736
 737         bau_desc = bcp->descriptor_base;
 738         bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu;
 739
 740         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
 741         remotes = 0;
 742         for_each_cpu(tcpu, flush_mask) {
 743                 uvhub = uv_cpu_to_blade_id(tcpu);
 744                 if (uvhub == bcp->uvhub) {
 745                         locals++;
 746                         continue;
 747                 }
 748                 bau_uvhub_set(uvhub, &bau_desc->distribution);
 749                 remotes++;
 750         }
 751         if (remotes == 0) {
 752                 /*
 753                  * No off_hub flushing; return status for local hub.
 754                  * Return the caller's mask if all were local (the current
 755                  * cpu may be in that mask).
 756                  */
 757                 if (locals)
 758                         return cpumask;
 759                 else
 760                         return NULL;
 761         }
 762         stat = &per_cpu(ptcstats, cpu);
 763         stat->s_requestor++;
 764         stat->s_ntargcpu += remotes;
 765         remotes = bau_uvhub_weight(&bau_desc->distribution);
 766         stat->s_ntarguvhub += remotes;
 767         if (remotes >= 16)
 768                 stat->s_ntarguvhub16++;
 769         else if (remotes >= 8)
 770                 stat->s_ntarguvhub8++;
 771         else if (remotes >= 4)
 772                 stat->s_ntarguvhub4++;
 773         else if (remotes >= 2)
 774                 stat->s_ntarguvhub2++;
 775         else
 776                 stat->s_ntarguvhub1++;
 777
 778         bau_desc->payload.address = va;
 779         bau_desc->payload.sending_cpu = cpu;
 780
 781         /*
 782          * uv_flush_send_and_wait returns null if all cpu's were messaged, or
 783          * the adjusted flush_mask if any cpu's were not messaged.
 784          */
 785         return uv_flush_send_and_wait(bau_desc, flush_mask, bcp);
 786 }
 787
 788 /*
 789  * The BAU message interrupt comes here. (registered by set_intr_gate)
 790  * See entry_64.S
 791  *
 792  * We received a broadcast assist message.
 793  *
 794  * Interrupts are disabled; this interrupt could represent
 795  * the receipt of several messages.
 796  *
 797  * All cores/threads on this hub get this interrupt.
 798  * The last one to see it does the software ack.
 799  * (the resource will not be freed until noninterruptable cpus see this
 800  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
 801  */
 802 void uv_bau_message_interrupt(struct pt_regs *regs)
 803 {
 804         int count = 0;
 805         cycles_t time_start;
 806         struct bau_payload_queue_entry *msg;
 807         struct bau_control *bcp;
 808         struct ptc_stats *stat;
 809         struct msg_desc msgdesc;
 810
 811         time_start = get_cycles();
 812         bcp = &per_cpu(bau_control, smp_processor_id());
 813         stat = &per_cpu(ptcstats, smp_processor_id());
 814         msgdesc.va_queue_first = bcp->va_queue_first;
 815         msgdesc.va_queue_last = bcp->va_queue_last;
 816         msg = bcp->bau_msg_head;
 817         while (msg->sw_ack_vector) {
 818                 count++;
 819                 msgdesc.msg_slot = msg - msgdesc.va_queue_first;
 820                 msgdesc.sw_ack_slot = ffs(msg->sw_ack_vector) - 1;
 821                 msgdesc.msg = msg;
 822                 uv_bau_process_message(&msgdesc, bcp);
 823                 msg++;
 824                 if (msg > msgdesc.va_queue_last)
 825                         msg = msgdesc.va_queue_first;
 826                 bcp->bau_msg_head = msg;
 827         }
 828         stat->d_time += (get_cycles() - time_start);
 829         if (!count)
 830                 stat->d_nomsg++;
 831         else if (count > 1)
 832                 stat->d_multmsg++;
 833         ack_APIC_irq();
 834 }
 835
 836 /*
 837  * uv_enable_timeouts
 838  *
 839  * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
 840  * shootdown message timeouts enabled.  The timeout does not cause
 841  * an interrupt, but causes an error message to be returned to
 842  * the sender.
 843  */
 844 static void uv_enable_timeouts(void)
 845 {
 846         int uvhub;
 847         int nuvhubs;
 848         int pnode;
 849         unsigned long mmr_image;
 850
 851         nuvhubs = uv_num_possible_blades();
 852
 853         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
 854                 if (!uv_blade_nr_possible_cpus(uvhub))
 855                         continue;
 856
 857                 pnode = uv_blade_to_pnode(uvhub);
 858                 mmr_image =
 859                     uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL);
 860                 /*
 861                  * Set the timeout period and then lock it in, in three
 862                  * steps; captures and locks in the period.
 863                  *
 864                  * To program the period, the SOFT_ACK_MODE must be off.
 865                  */
 866                 mmr_image &= ~((unsigned long)1 <<
 867                     UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 868                 uv_write_global_mmr64
 869                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 870                 /*
 871                  * Set the 4-bit period.
 872                  */
 873                 mmr_image &= ~((unsigned long)0xf <<
 874                      UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 875                 mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD <<
 876                      UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 877                 uv_write_global_mmr64
 878                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 879                 /*
 880                  * Subsequent reversals of the timebase bit (3) cause an
 881                  * immediate timeout of one or all INTD resources as
 882                  * indicated in bits 2:0 (7 causes all of them to timeout).
 883                  */
 884                 mmr_image |= ((unsigned long)1 <<
 885                     UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 886                 uv_write_global_mmr64
 887                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 888         }
 889 }
 890
 891 static void *uv_ptc_seq_start(struct seq_file *file, loff_t *offset)
 892 {
 893         if (*offset < num_possible_cpus())
 894                 return offset;
 895         return NULL;
 896 }
 897
 898 static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
 899 {
 900         (*offset)++;
 901         if (*offset < num_possible_cpus())
 902                 return offset;
 903         return NULL;
 904 }
 905
 906 static void uv_ptc_seq_stop(struct seq_file *file, void *data)
 907 {
 908 }
 909
 910 static inline unsigned long long
 911 millisec_2_cycles(unsigned long millisec)
 912 {
 913         unsigned long ns;
 914         unsigned long long cyc;
 915
 916         ns = millisec * 1000;
 917         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
 918         return cyc;
 919 }
 920
 921 /*
 922  * Display the statistics thru /proc.
 923  * 'data' points to the cpu number
 924  */
 925 static int uv_ptc_seq_show(struct seq_file *file, void *data)
 926 {
 927         struct ptc_stats *stat;
 928         int cpu;
 929
 930         cpu = *(loff_t *)data;
 931
 932         if (!cpu) {
 933                 seq_printf(file,
 934                         "# cpu sent stime numuvhubs numuvhubs16 numuvhubs8 ");
 935                 seq_printf(file,
 936                         "numuvhubs4 numuvhubs2 numuvhubs1 numcpus dto ");
 937                 seq_printf(file,
 938                         "retries rok resetp resett giveup sto bz throt ");
 939                 seq_printf(file,
 940                         "sw_ack recv rtime all ");
 941                 seq_printf(file,
 942                         "one mult none retry canc nocan reset rcan\n");
 943         }
 944         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
 945                 stat = &per_cpu(ptcstats, cpu);
 946                 /* source side statistics */
 947                 seq_printf(file,
 948                         "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
 949                            cpu, stat->s_requestor, cycles_2_us(stat->s_time),
 950                            stat->s_ntarguvhub, stat->s_ntarguvhub16,
 951                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
 952                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
 953                            stat->s_ntargcpu, stat->s_dtimeout);
 954                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
 955                            stat->s_retry_messages, stat->s_retriesok,
 956                            stat->s_resets_plug, stat->s_resets_timeout,
 957                            stat->s_giveup, stat->s_stimeout,
 958                            stat->s_busy, stat->s_throttles);
 959                 /* destination side statistics */
 960                 seq_printf(file,
 961                            "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
 962                            uv_read_global_mmr64(uv_cpu_to_pnode(cpu),
 963                                         UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE),
 964                            stat->d_requestee, cycles_2_us(stat->d_time),
 965                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
 966                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
 967                            stat->d_nocanceled, stat->d_resets,
 968                            stat->d_rcanceled);
 969         }
 970
 971         return 0;
 972 }
 973
 974 /*
 975  * -1: resetf the statistics
 976  *  0: display meaning of the statistics
 977  * >0: maximum concurrent active descriptors per uvhub (throttle)
 978  */
 979 static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user,
 980                                  size_t count, loff_t *data)
 981 {
 982         int cpu;
 983         long input_arg;
 984         char optstr[64];
 985         struct ptc_stats *stat;
 986         struct bau_control *bcp;
 987
 988         if (count == 0 || count > sizeof(optstr))
 989                 return -EINVAL;
 990         if (copy_from_user(optstr, user, count))
 991                 return -EFAULT;
 992         optstr[count - 1] = '\0';
 993         if (strict_strtol(optstr, 10, &input_arg) < 0) {
 994                 printk(KERN_DEBUG "%s is invalid\n", optstr);
 995                 return -EINVAL;
 996         }
 997
 998         if (input_arg == 0) {
 999                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1000                 printk(KERN_DEBUG "Sender statistics:\n");
1001                 printk(KERN_DEBUG
1002                 "sent:     number of shootdown messages sent\n");
1003                 printk(KERN_DEBUG
1004                 "stime:    time spent sending messages\n");
1005                 printk(KERN_DEBUG
1006                 "numuvhubs: number of hubs targeted with shootdown\n");
1007                 printk(KERN_DEBUG
1008                 "numuvhubs16: number times 16 or more hubs targeted\n");
1009                 printk(KERN_DEBUG
1010                 "numuvhubs8: number times 8 or more hubs targeted\n");
1011                 printk(KERN_DEBUG
1012                 "numuvhubs4: number times 4 or more hubs targeted\n");
1013                 printk(KERN_DEBUG
1014                 "numuvhubs2: number times 2 or more hubs targeted\n");
1015                 printk(KERN_DEBUG
1016                 "numuvhubs1: number times 1 hub targeted\n");
1017                 printk(KERN_DEBUG
1018                 "numcpus:  number of cpus targeted with shootdown\n");
1019                 printk(KERN_DEBUG
1020                 "dto:      number of destination timeouts\n");
1021                 printk(KERN_DEBUG
1022                 "retries:  destination timeout retries sent\n");
1023                 printk(KERN_DEBUG
1024                 "rok:   :  destination timeouts successfully retried\n");
1025                 printk(KERN_DEBUG
1026                 "resetp:   ipi-style resource resets for plugs\n");
1027                 printk(KERN_DEBUG
1028                 "resett:   ipi-style resource resets for timeouts\n");
1029                 printk(KERN_DEBUG
1030                 "giveup:   fall-backs to ipi-style shootdowns\n");
1031                 printk(KERN_DEBUG
1032                 "sto:      number of source timeouts\n");
1033                 printk(KERN_DEBUG
1034                 "bz:       number of stay-busy's\n");
1035                 printk(KERN_DEBUG
1036                 "throt:    number times spun in throttle\n");
1037                 printk(KERN_DEBUG "Destination side statistics:\n");
1038                 printk(KERN_DEBUG
1039                 "sw_ack:   image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
1040                 printk(KERN_DEBUG
1041                 "recv:     shootdown messages received\n");
1042                 printk(KERN_DEBUG
1043                 "rtime:    time spent processing messages\n");
1044                 printk(KERN_DEBUG
1045                 "all:      shootdown all-tlb messages\n");
1046                 printk(KERN_DEBUG
1047                 "one:      shootdown one-tlb messages\n");
1048                 printk(KERN_DEBUG
1049                 "mult:     interrupts that found multiple messages\n");
1050                 printk(KERN_DEBUG
1051                 "none:     interrupts that found no messages\n");
1052                 printk(KERN_DEBUG
1053                 "retry:    number of retry messages processed\n");
1054                 printk(KERN_DEBUG
1055                 "canc:     number messages canceled by retries\n");
1056                 printk(KERN_DEBUG
1057                 "nocan:    number retries that found nothing to cancel\n");
1058                 printk(KERN_DEBUG
1059                 "reset:    number of ipi-style reset requests processed\n");
1060                 printk(KERN_DEBUG
1061                 "rcan:     number messages canceled by reset requests\n");
1062         } else if (input_arg == -1) {
1063                 for_each_present_cpu(cpu) {
1064                         stat = &per_cpu(ptcstats, cpu);
1065                         memset(stat, 0, sizeof(struct ptc_stats));
1066                 }
1067         } else {
1068                 uv_bau_max_concurrent = input_arg;
1069                 bcp = &per_cpu(bau_control, smp_processor_id());
1070                 if (uv_bau_max_concurrent < 1 ||
1071                     uv_bau_max_concurrent > bcp->cpus_in_uvhub) {
1072                         printk(KERN_DEBUG
1073                                 "Error: BAU max concurrent %d; %d is invalid\n",
1074                                 bcp->max_concurrent, uv_bau_max_concurrent);
1075                         return -EINVAL;
1076                 }
1077                 printk(KERN_DEBUG "Set BAU max concurrent:%d\n",
1078                        uv_bau_max_concurrent);
1079                 for_each_present_cpu(cpu) {
1080                         bcp = &per_cpu(bau_control, cpu);
1081                         bcp->max_concurrent = uv_bau_max_concurrent;
1082                 }
1083         }
1084
1085         return count;
1086 }
1087
1088 static const struct seq_operations uv_ptc_seq_ops = {
1089         .start          = uv_ptc_seq_start,
1090         .next           = uv_ptc_seq_next,
1091         .stop           = uv_ptc_seq_stop,
1092         .show           = uv_ptc_seq_show
1093 };
1094
1095 static int uv_ptc_proc_open(struct inode *inode, struct file *file)
1096 {
1097         return seq_open(file, &uv_ptc_seq_ops);
1098 }
1099
1100 static const struct file_operations proc_uv_ptc_operations = {
1101         .open           = uv_ptc_proc_open,
1102         .read           = seq_read,
1103         .write          = uv_ptc_proc_write,
1104         .llseek         = seq_lseek,
1105         .release        = seq_release,
1106 };
1107
1108 static int __init uv_ptc_init(void)
1109 {
1110         struct proc_dir_entry *proc_uv_ptc;
1111
1112         if (!is_uv_system())
1113                 return 0;
1114
1115         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1116                                   &proc_uv_ptc_operations);
1117         if (!proc_uv_ptc) {
1118                 printk(KERN_ERR "unable to create %s proc entry\n",
1119                        UV_PTC_BASENAME);
1120                 return -EINVAL;
1121         }
1122         return 0;
1123 }
1124
1125 /*
1126  * initialize the sending side's sending buffers
1127  */
1128 static void
1129 uv_activation_descriptor_init(int node, int pnode)
1130 {
1131         int i;
1132         int cpu;
1133         unsigned long pa;
1134         unsigned long m;
1135         unsigned long n;
1136         struct bau_desc *bau_desc;
1137         struct bau_desc *bd2;
1138         struct bau_control *bcp;
1139
1140         /*
1141          * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR)
1142          * per cpu; and up to 32 (UV_ADP_SIZE) cpu's per uvhub
1143          */
1144         bau_desc = (struct bau_desc *)kmalloc_node(sizeof(struct bau_desc)*
1145                 UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR, GFP_KERNEL, node);
1146         BUG_ON(!bau_desc);
1147
1148         pa = uv_gpa(bau_desc); /* need the real nasid*/
1149         n = pa >> uv_nshift;
1150         m = pa & uv_mmask;
1151
1152         uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE,
1153                               (n << UV_DESC_BASE_PNODE_SHIFT | m));
1154
1155         /*
1156          * initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each
1157          * cpu even though we only use the first one; one descriptor can
1158          * describe a broadcast to 256 uv hubs.
1159          */
1160         for (i = 0, bd2 = bau_desc; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR);
1161                 i++, bd2++) {
1162                 memset(bd2, 0, sizeof(struct bau_desc));
1163                 bd2->header.sw_ack_flag = 1;
1164                 /*
1165                  * base_dest_nodeid is the nasid (pnode<<1) of the first uvhub
1166                  * in the partition. The bit map will indicate uvhub numbers,
1167                  * which are 0-N in a partition. Pnodes are unique system-wide.
1168                  */
1169                 bd2->header.base_dest_nodeid = uv_partition_base_pnode << 1;
1170                 bd2->header.dest_subnodeid = 0x10; /* the LB */
1171                 bd2->header.command = UV_NET_ENDPOINT_INTD;
1172                 bd2->header.int_both = 1;
1173                 /*
1174                  * all others need to be set to zero:
1175                  *   fairness chaining multilevel count replied_to
1176                  */
1177         }
1178         for_each_present_cpu(cpu) {
1179                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1180                         continue;
1181                 bcp = &per_cpu(bau_control, cpu);
1182                 bcp->descriptor_base = bau_desc;
1183         }
1184 }
1185
1186 /*
1187  * initialize the destination side's receiving buffers
1188  * entered for each uvhub in the partition
1189  * - node is first node (kernel memory notion) on the uvhub
1190  * - pnode is the uvhub's physical identifier
1191  */
1192 static void
1193 uv_payload_queue_init(int node, int pnode)
1194 {
1195         int pn;
1196         int cpu;
1197         char *cp;
1198         unsigned long pa;
1199         struct bau_payload_queue_entry *pqp;
1200         struct bau_payload_queue_entry *pqp_malloc;
1201         struct bau_control *bcp;
1202
1203         pqp = (struct bau_payload_queue_entry *) kmalloc_node(
1204                 (DEST_Q_SIZE + 1) * sizeof(struct bau_payload_queue_entry),
1205                 GFP_KERNEL, node);
1206         BUG_ON(!pqp);
1207         pqp_malloc = pqp;
1208
1209         cp = (char *)pqp + 31;
1210         pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5);
1211
1212         for_each_present_cpu(cpu) {
1213                 if (pnode != uv_cpu_to_pnode(cpu))
1214                         continue;
1215                 /* for every cpu on this pnode: */
1216                 bcp = &per_cpu(bau_control, cpu);
1217                 bcp->va_queue_first = pqp;
1218                 bcp->bau_msg_head = pqp;
1219                 bcp->va_queue_last = pqp + (DEST_Q_SIZE - 1);
1220         }
1221         /*
1222          * need the pnode of where the memory was really allocated
1223          */
1224         pa = uv_gpa(pqp);
1225         pn = pa >> uv_nshift;
1226         uv_write_global_mmr64(pnode,
1227                               UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST,
1228                               ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) |
1229                               uv_physnodeaddr(pqp));
1230         uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL,
1231                               uv_physnodeaddr(pqp));
1232         uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST,
1233                               (unsigned long)
1234                               uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1)));
1235         /* in effect, all msg_type's are set to MSG_NOOP */
1236         memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE);
1237 }
1238
1239 /*
1240  * Initialization of each UV hub's structures
1241  */
1242 static void __init uv_init_uvhub(int uvhub, int vector)
1243 {
1244         int node;
1245         int pnode;
1246         unsigned long apicid;
1247
1248         node = uvhub_to_first_node(uvhub);
1249         pnode = uv_blade_to_pnode(uvhub);
1250         uv_activation_descriptor_init(node, pnode);
1251         uv_payload_queue_init(node, pnode);
1252         /*
1253          * the below initialization can't be in firmware because the
1254          * messaging IRQ will be determined by the OS
1255          */
1256         apicid = uvhub_to_first_apicid(uvhub);
1257         uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG,
1258                                       ((apicid << 32) | vector));
1259 }
1260
1261 /*
1262  * initialize the bau_control structure for each cpu
1263  */
1264 static void uv_init_per_cpu(int nuvhubs)
1265 {
1266         int i, j, k;
1267         int cpu;
1268         int pnode;
1269         int uvhub;
1270         short socket = 0;
1271         struct bau_control *bcp;
1272         struct uvhub_desc *bdp;
1273         struct socket_desc *sdp;
1274         struct bau_control *hmaster = NULL;
1275         struct bau_control *smaster = NULL;
1276         struct socket_desc {
1277                 short num_cpus;
1278                 short cpu_number[16];
1279         };
1280         struct uvhub_desc {
1281                 short num_sockets;
1282                 short num_cpus;
1283                 short uvhub;
1284                 short pnode;
1285                 struct socket_desc socket[2];
1286         };
1287         struct uvhub_desc *uvhub_descs;
1288
1289         uvhub_descs = (struct uvhub_desc *)
1290                 kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1291         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1292         for_each_present_cpu(cpu) {
1293                 bcp = &per_cpu(bau_control, cpu);
1294                 memset(bcp, 0, sizeof(struct bau_control));
1295                 spin_lock_init(&bcp->masks_lock);
1296                 bcp->max_concurrent = uv_bau_max_concurrent;
1297                 pnode = uv_cpu_hub_info(cpu)->pnode;
1298                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1299                 bdp = &uvhub_descs[uvhub];
1300                 bdp->num_cpus++;
1301                 bdp->uvhub = uvhub;
1302                 bdp->pnode = pnode;
1303                 /* time interval to catch a hardware stay-busy bug */
1304                 bcp->timeout_interval = millisec_2_cycles(3);
1305                 /* kludge: assume uv_hub.h is constant */
1306                 socket = (cpu_physical_id(cpu)>>5)&1;
1307                 if (socket >= bdp->num_sockets)
1308                         bdp->num_sockets = socket+1;
1309                 sdp = &bdp->socket[socket];
1310                 sdp->cpu_number[sdp->num_cpus] = cpu;
1311                 sdp->num_cpus++;
1312         }
1313         socket = 0;
1314         for_each_possible_blade(uvhub) {
1315                 bdp = &uvhub_descs[uvhub];
1316                 for (i = 0; i < bdp->num_sockets; i++) {
1317                         sdp = &bdp->socket[i];
1318                         for (j = 0; j < sdp->num_cpus; j++) {
1319                                 cpu = sdp->cpu_number[j];
1320                                 bcp = &per_cpu(bau_control, cpu);
1321                                 bcp->cpu = cpu;
1322                                 if (j == 0) {
1323                                         smaster = bcp;
1324                                         if (i == 0)
1325                                                 hmaster = bcp;
1326                                 }
1327                                 bcp->cpus_in_uvhub = bdp->num_cpus;
1328                                 bcp->cpus_in_socket = sdp->num_cpus;
1329                                 bcp->socket_master = smaster;
1330                                 bcp->uvhub_master = hmaster;
1331                                 for (k = 0; k < DEST_Q_SIZE; k++)
1332                                         bcp->socket_acknowledge_count[k] = 0;
1333                                 bcp->uvhub_cpu =
1334                                   uv_cpu_hub_info(cpu)->blade_processor_id;
1335                         }
1336                         socket++;
1337                 }
1338         }
1339         kfree(uvhub_descs);
1340 }
1341
1342 /*
1343  * Initialization of BAU-related structures
1344  */
1345 static int __init uv_bau_init(void)
1346 {
1347         int uvhub;
1348         int pnode;
1349         int nuvhubs;
1350         int cur_cpu;
1351         int vector;
1352         unsigned long mmr;
1353
1354         if (!is_uv_system())
1355                 return 0;
1356
1357         if (nobau)
1358                 return 0;
1359
1360         for_each_possible_cpu(cur_cpu)
1361                 zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu),
1362                                        GFP_KERNEL, cpu_to_node(cur_cpu));
1363
1364         uv_bau_max_concurrent = MAX_BAU_CONCURRENT;
1365         uv_nshift = uv_hub_info->m_val;
1366         uv_mmask = (1UL << uv_hub_info->m_val) - 1;
1367         nuvhubs = uv_num_possible_blades();
1368
1369         uv_init_per_cpu(nuvhubs);
1370
1371         uv_partition_base_pnode = 0x7fffffff;
1372         for (uvhub = 0; uvhub < nuvhubs; uvhub++)
1373                 if (uv_blade_nr_possible_cpus(uvhub) &&
1374                         (uv_blade_to_pnode(uvhub) < uv_partition_base_pnode))
1375                         uv_partition_base_pnode = uv_blade_to_pnode(uvhub);
1376
1377         vector = UV_BAU_MESSAGE;
1378         for_each_possible_blade(uvhub)
1379                 if (uv_blade_nr_possible_cpus(uvhub))
1380                         uv_init_uvhub(uvhub, vector);
1381
1382         uv_enable_timeouts();
1383         alloc_intr_gate(vector, uv_bau_message_intr1);
1384
1385         for_each_possible_blade(uvhub) {
1386                 pnode = uv_blade_to_pnode(uvhub);
1387                 /* INIT the bau */
1388                 uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_ACTIVATION_CONTROL,
1389                                       ((unsigned long)1 << 63));
1390                 mmr = 1; /* should be 1 to broadcast to both sockets */
1391                 uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST, mmr);
1392         }
1393
1394         return 0;
1395 }
1396 core_initcall(uv_bau_init);
1397 core_initcall(uv_ptc_init);