arch/x86/platform/uv/tlb_uv.c

   1 /*
   2  *      SGI UltraViolet TLB flush routines.
   3  *
   4  *      (c) 2008-2012 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/debugfs.h>
  12 #include <linux/kernel.h>
  13 #include <linux/slab.h>
  14 #include <linux/delay.h>
  15
  16 #include <asm/mmu_context.h>
  17 #include <asm/uv/uv.h>
  18 #include <asm/uv/uv_mmrs.h>
  19 #include <asm/uv/uv_hub.h>
  20 #include <asm/uv/uv_bau.h>
  21 #include <asm/apic.h>
  22 #include <asm/idle.h>
  23 #include <asm/tsc.h>
  24 #include <asm/irq_vectors.h>
  25 #include <asm/timer.h>
  26
  27 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
  28 static int timeout_base_ns[] = {
  29                 20,
  30                 160,
  31                 1280,
  32                 10240,
  33                 81920,
  34                 655360,
  35                 5242880,
  36                 167772160
  37 };
  38
  39 static int timeout_us;
  40 static int nobau;
  41 static int nobau_perm;
  42 static cycles_t congested_cycles;
  43
  44 /* tunables: */
  45 static int max_concurr          = MAX_BAU_CONCURRENT;
  46 static int max_concurr_const    = MAX_BAU_CONCURRENT;
  47 static int plugged_delay        = PLUGGED_DELAY;
  48 static int plugsb4reset         = PLUGSB4RESET;
  49 static int giveup_limit         = GIVEUP_LIMIT;
  50 static int timeoutsb4reset      = TIMEOUTSB4RESET;
  51 static int ipi_reset_limit      = IPI_RESET_LIMIT;
  52 static int complete_threshold   = COMPLETE_THRESHOLD;
  53 static int congested_respns_us  = CONGESTED_RESPONSE_US;
  54 static int congested_reps       = CONGESTED_REPS;
  55 static int disabled_period      = DISABLED_PERIOD;
  56
  57 static struct tunables tunables[] = {
  58         {&max_concurr, MAX_BAU_CONCURRENT}, /* must be [0] */
  59         {&plugged_delay, PLUGGED_DELAY},
  60         {&plugsb4reset, PLUGSB4RESET},
  61         {&timeoutsb4reset, TIMEOUTSB4RESET},
  62         {&ipi_reset_limit, IPI_RESET_LIMIT},
  63         {&complete_threshold, COMPLETE_THRESHOLD},
  64         {&congested_respns_us, CONGESTED_RESPONSE_US},
  65         {&congested_reps, CONGESTED_REPS},
  66         {&disabled_period, DISABLED_PERIOD},
  67         {&giveup_limit, GIVEUP_LIMIT}
  68 };
  69
  70 static struct dentry *tunables_dir;
  71 static struct dentry *tunables_file;
  72
  73 /* these correspond to the statistics printed by ptc_seq_show() */
  74 static char *stat_description[] = {
  75         "sent:     number of shootdown messages sent",
  76         "stime:    time spent sending messages",
  77         "numuvhubs: number of hubs targeted with shootdown",
  78         "numuvhubs16: number times 16 or more hubs targeted",
  79         "numuvhubs8: number times 8 or more hubs targeted",
  80         "numuvhubs4: number times 4 or more hubs targeted",
  81         "numuvhubs2: number times 2 or more hubs targeted",
  82         "numuvhubs1: number times 1 hub targeted",
  83         "numcpus:  number of cpus targeted with shootdown",
  84         "dto:      number of destination timeouts",
  85         "retries:  destination timeout retries sent",
  86         "rok:   :  destination timeouts successfully retried",
  87         "resetp:   ipi-style resource resets for plugs",
  88         "resett:   ipi-style resource resets for timeouts",
  89         "giveup:   fall-backs to ipi-style shootdowns",
  90         "sto:      number of source timeouts",
  91         "bz:       number of stay-busy's",
  92         "throt:    number times spun in throttle",
  93         "swack:   image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE",
  94         "recv:     shootdown messages received",
  95         "rtime:    time spent processing messages",
  96         "all:      shootdown all-tlb messages",
  97         "one:      shootdown one-tlb messages",
  98         "mult:     interrupts that found multiple messages",
  99         "none:     interrupts that found no messages",
 100         "retry:    number of retry messages processed",
 101         "canc:     number messages canceled by retries",
 102         "nocan:    number retries that found nothing to cancel",
 103         "reset:    number of ipi-style reset requests processed",
 104         "rcan:     number messages canceled by reset requests",
 105         "disable:  number times use of the BAU was disabled",
 106         "enable:   number times use of the BAU was re-enabled"
 107 };
 108
 109 static int __init
 110 setup_nobau(char *arg)
 111 {
 112         nobau = 1;
 113         return 0;
 114 }
 115 early_param("nobau", setup_nobau);
 116
 117 /* base pnode in this partition */
 118 static int uv_base_pnode __read_mostly;
 119
 120 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
 121 static DEFINE_PER_CPU(struct bau_control, bau_control);
 122 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
 123
 124 static void
 125 set_bau_on(void)
 126 {
 127         int cpu;
 128         struct bau_control *bcp;
 129
 130         if (nobau_perm) {
 131                 pr_info("BAU not initialized; cannot be turned on\n");
 132                 return;
 133         }
 134         nobau = 0;
 135         for_each_present_cpu(cpu) {
 136                 bcp = &per_cpu(bau_control, cpu);
 137                 bcp->nobau = 0;
 138         }
 139         pr_info("BAU turned on\n");
 140         return;
 141 }
 142
 143 static void
 144 set_bau_off(void)
 145 {
 146         int cpu;
 147         struct bau_control *bcp;
 148
 149         nobau = 1;
 150         for_each_present_cpu(cpu) {
 151                 bcp = &per_cpu(bau_control, cpu);
 152                 bcp->nobau = 1;
 153         }
 154         pr_info("BAU turned off\n");
 155         return;
 156 }
 157
 158 /*
 159  * Determine the first node on a uvhub. 'Nodes' are used for kernel
 160  * memory allocation.
 161  */
 162 static int __init uvhub_to_first_node(int uvhub)
 163 {
 164         int node, b;
 165
 166         for_each_online_node(node) {
 167                 b = uv_node_to_blade_id(node);
 168                 if (uvhub == b)
 169                         return node;
 170         }
 171         return -1;
 172 }
 173
 174 /*
 175  * Determine the apicid of the first cpu on a uvhub.
 176  */
 177 static int __init uvhub_to_first_apicid(int uvhub)
 178 {
 179         int cpu;
 180
 181         for_each_present_cpu(cpu)
 182                 if (uvhub == uv_cpu_to_blade_id(cpu))
 183                         return per_cpu(x86_cpu_to_apicid, cpu);
 184         return -1;
 185 }
 186
 187 /*
 188  * Free a software acknowledge hardware resource by clearing its Pending
 189  * bit. This will return a reply to the sender.
 190  * If the message has timed out, a reply has already been sent by the
 191  * hardware but the resource has not been released. In that case our
 192  * clear of the Timeout bit (as well) will free the resource. No reply will
 193  * be sent (the hardware will only do one reply per message).
 194  */
 195 static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp,
 196                                                 int do_acknowledge)
 197 {
 198         unsigned long dw;
 199         struct bau_pq_entry *msg;
 200
 201         msg = mdp->msg;
 202         if (!msg->canceled && do_acknowledge) {
 203                 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
 204                 write_mmr_sw_ack(dw);
 205         }
 206         msg->replied_to = 1;
 207         msg->swack_vec = 0;
 208 }
 209
 210 /*
 211  * Process the receipt of a RETRY message
 212  */
 213 static void bau_process_retry_msg(struct msg_desc *mdp,
 214                                         struct bau_control *bcp)
 215 {
 216         int i;
 217         int cancel_count = 0;
 218         unsigned long msg_res;
 219         unsigned long mmr = 0;
 220         struct bau_pq_entry *msg = mdp->msg;
 221         struct bau_pq_entry *msg2;
 222         struct ptc_stats *stat = bcp->statp;
 223
 224         stat->d_retries++;
 225         /*
 226          * cancel any message from msg+1 to the retry itself
 227          */
 228         for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
 229                 if (msg2 > mdp->queue_last)
 230                         msg2 = mdp->queue_first;
 231                 if (msg2 == msg)
 232                         break;
 233
 234                 /* same conditions for cancellation as do_reset */
 235                 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
 236                     (msg2->swack_vec) && ((msg2->swack_vec &
 237                         msg->swack_vec) == 0) &&
 238                     (msg2->sending_cpu == msg->sending_cpu) &&
 239                     (msg2->msg_type != MSG_NOOP)) {
 240                         mmr = read_mmr_sw_ack();
 241                         msg_res = msg2->swack_vec;
 242                         /*
 243                          * This is a message retry; clear the resources held
 244                          * by the previous message only if they timed out.
 245                          * If it has not timed out we have an unexpected
 246                          * situation to report.
 247                          */
 248                         if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
 249                                 unsigned long mr;
 250                                 /*
 251                                  * Is the resource timed out?
 252                                  * Make everyone ignore the cancelled message.
 253                                  */
 254                                 msg2->canceled = 1;
 255                                 stat->d_canceled++;
 256                                 cancel_count++;
 257                                 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
 258                                 write_mmr_sw_ack(mr);
 259                         }
 260                 }
 261         }
 262         if (!cancel_count)
 263                 stat->d_nocanceled++;
 264 }
 265
 266 /*
 267  * Do all the things a cpu should do for a TLB shootdown message.
 268  * Other cpu's may come here at the same time for this message.
 269  */
 270 static void bau_process_message(struct msg_desc *mdp, struct bau_control *bcp,
 271                                                 int do_acknowledge)
 272 {
 273         short socket_ack_count = 0;
 274         short *sp;
 275         struct atomic_short *asp;
 276         struct ptc_stats *stat = bcp->statp;
 277         struct bau_pq_entry *msg = mdp->msg;
 278         struct bau_control *smaster = bcp->socket_master;
 279
 280         /*
 281          * This must be a normal message, or retry of a normal message
 282          */
 283         if (msg->address == TLB_FLUSH_ALL) {
 284                 local_flush_tlb();
 285                 stat->d_alltlb++;
 286         } else {
 287                 __flush_tlb_one(msg->address);
 288                 stat->d_onetlb++;
 289         }
 290         stat->d_requestee++;
 291
 292         /*
 293          * One cpu on each uvhub has the additional job on a RETRY
 294          * of releasing the resource held by the message that is
 295          * being retried.  That message is identified by sending
 296          * cpu number.
 297          */
 298         if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
 299                 bau_process_retry_msg(mdp, bcp);
 300
 301         /*
 302          * This is a swack message, so we have to reply to it.
 303          * Count each responding cpu on the socket. This avoids
 304          * pinging the count's cache line back and forth between
 305          * the sockets.
 306          */
 307         sp = &smaster->socket_acknowledge_count[mdp->msg_slot];
 308         asp = (struct atomic_short *)sp;
 309         socket_ack_count = atom_asr(1, asp);
 310         if (socket_ack_count == bcp->cpus_in_socket) {
 311                 int msg_ack_count;
 312                 /*
 313                  * Both sockets dump their completed count total into
 314                  * the message's count.
 315                  */
 316                 *sp = 0;
 317                 asp = (struct atomic_short *)&msg->acknowledge_count;
 318                 msg_ack_count = atom_asr(socket_ack_count, asp);
 319
 320                 if (msg_ack_count == bcp->cpus_in_uvhub) {
 321                         /*
 322                          * All cpus in uvhub saw it; reply
 323                          * (unless we are in the UV2 workaround)
 324                          */
 325                         reply_to_message(mdp, bcp, do_acknowledge);
 326                 }
 327         }
 328
 329         return;
 330 }
 331
 332 /*
 333  * Determine the first cpu on a pnode.
 334  */
 335 static int pnode_to_first_cpu(int pnode, struct bau_control *smaster)
 336 {
 337         int cpu;
 338         struct hub_and_pnode *hpp;
 339
 340         for_each_present_cpu(cpu) {
 341                 hpp = &smaster->thp[cpu];
 342                 if (pnode == hpp->pnode)
 343                         return cpu;
 344         }
 345         return -1;
 346 }
 347
 348 /*
 349  * Last resort when we get a large number of destination timeouts is
 350  * to clear resources held by a given cpu.
 351  * Do this with IPI so that all messages in the BAU message queue
 352  * can be identified by their nonzero swack_vec field.
 353  *
 354  * This is entered for a single cpu on the uvhub.
 355  * The sender want's this uvhub to free a specific message's
 356  * swack resources.
 357  */
 358 static void do_reset(void *ptr)
 359 {
 360         int i;
 361         struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id());
 362         struct reset_args *rap = (struct reset_args *)ptr;
 363         struct bau_pq_entry *msg;
 364         struct ptc_stats *stat = bcp->statp;
 365
 366         stat->d_resets++;
 367         /*
 368          * We're looking for the given sender, and
 369          * will free its swack resource.
 370          * If all cpu's finally responded after the timeout, its
 371          * message 'replied_to' was set.
 372          */
 373         for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
 374                 unsigned long msg_res;
 375                 /* do_reset: same conditions for cancellation as
 376                    bau_process_retry_msg() */
 377                 if ((msg->replied_to == 0) &&
 378                     (msg->canceled == 0) &&
 379                     (msg->sending_cpu == rap->sender) &&
 380                     (msg->swack_vec) &&
 381                     (msg->msg_type != MSG_NOOP)) {
 382                         unsigned long mmr;
 383                         unsigned long mr;
 384                         /*
 385                          * make everyone else ignore this message
 386                          */
 387                         msg->canceled = 1;
 388                         /*
 389                          * only reset the resource if it is still pending
 390                          */
 391                         mmr = read_mmr_sw_ack();
 392                         msg_res = msg->swack_vec;
 393                         mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
 394                         if (mmr & msg_res) {
 395                                 stat->d_rcanceled++;
 396                                 write_mmr_sw_ack(mr);
 397                         }
 398                 }
 399         }
 400         return;
 401 }
 402
 403 /*
 404  * Use IPI to get all target uvhubs to release resources held by
 405  * a given sending cpu number.
 406  */
 407 static void reset_with_ipi(struct pnmask *distribution, struct bau_control *bcp)
 408 {
 409         int pnode;
 410         int apnode;
 411         int maskbits;
 412         int sender = bcp->cpu;
 413         cpumask_t *mask = bcp->uvhub_master->cpumask;
 414         struct bau_control *smaster = bcp->socket_master;
 415         struct reset_args reset_args;
 416
 417         reset_args.sender = sender;
 418         cpus_clear(*mask);
 419         /* find a single cpu for each uvhub in this distribution mask */
 420         maskbits = sizeof(struct pnmask) * BITSPERBYTE;
 421         /* each bit is a pnode relative to the partition base pnode */
 422         for (pnode = 0; pnode < maskbits; pnode++) {
 423                 int cpu;
 424                 if (!bau_uvhub_isset(pnode, distribution))
 425                         continue;
 426                 apnode = pnode + bcp->partition_base_pnode;
 427                 cpu = pnode_to_first_cpu(apnode, smaster);
 428                 cpu_set(cpu, *mask);
 429         }
 430
 431         /* IPI all cpus; preemption is already disabled */
 432         smp_call_function_many(mask, do_reset, (void *)&reset_args, 1);
 433         return;
 434 }
 435
 436 static inline unsigned long cycles_2_us(unsigned long long cyc)
 437 {
 438         unsigned long long ns;
 439         unsigned long us;
 440         int cpu = smp_processor_id();
 441
 442         ns =  (cyc * per_cpu(cyc2ns, cpu)) >> CYC2NS_SCALE_FACTOR;
 443         us = ns / 1000;
 444         return us;
 445 }
 446
 447 /*
 448  * wait for all cpus on this hub to finish their sends and go quiet
 449  * leaves uvhub_quiesce set so that no new broadcasts are started by
 450  * bau_flush_send_and_wait()
 451  */
 452 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
 453 {
 454         atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 455 }
 456
 457 /*
 458  * mark this quiet-requestor as done
 459  */
 460 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
 461 {
 462         atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 463 }
 464
 465 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
 466 {
 467         unsigned long descriptor_status;
 468
 469         descriptor_status = uv_read_local_mmr(mmr_offset);
 470         descriptor_status >>= right_shift;
 471         descriptor_status &= UV_ACT_STATUS_MASK;
 472         return descriptor_status;
 473 }
 474
 475 /*
 476  * Wait for completion of a broadcast software ack message
 477  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
 478  */
 479 static int uv1_wait_completion(struct bau_desc *bau_desc,
 480                                 unsigned long mmr_offset, int right_shift,
 481                                 struct bau_control *bcp, long try)
 482 {
 483         unsigned long descriptor_status;
 484         cycles_t ttm;
 485         struct ptc_stats *stat = bcp->statp;
 486
 487         descriptor_status = uv1_read_status(mmr_offset, right_shift);
 488         /* spin on the status MMR, waiting for it to go idle */
 489         while ((descriptor_status != DS_IDLE)) {
 490                 /*
 491                  * Our software ack messages may be blocked because
 492                  * there are no swack resources available.  As long
 493                  * as none of them has timed out hardware will NACK
 494                  * our message and its state will stay IDLE.
 495                  */
 496                 if (descriptor_status == DS_SOURCE_TIMEOUT) {
 497                         stat->s_stimeout++;
 498                         return FLUSH_GIVEUP;
 499                 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
 500                         stat->s_dtimeout++;
 501                         ttm = get_cycles();
 502
 503                         /*
 504                          * Our retries may be blocked by all destination
 505                          * swack resources being consumed, and a timeout
 506                          * pending.  In that case hardware returns the
 507                          * ERROR that looks like a destination timeout.
 508                          */
 509                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 510                                 bcp->conseccompletes = 0;
 511                                 return FLUSH_RETRY_PLUGGED;
 512                         }
 513
 514                         bcp->conseccompletes = 0;
 515                         return FLUSH_RETRY_TIMEOUT;
 516                 } else {
 517                         /*
 518                          * descriptor_status is still BUSY
 519                          */
 520                         cpu_relax();
 521                 }
 522                 descriptor_status = uv1_read_status(mmr_offset, right_shift);
 523         }
 524         bcp->conseccompletes++;
 525         return FLUSH_COMPLETE;
 526 }
 527
 528 /*
 529  * UV2 could have an extra bit of status in the ACTIVATION_STATUS_2 register.
 530  * But not currently used.
 531  */
 532 static unsigned long uv2_read_status(unsigned long offset, int rshft, int desc)
 533 {
 534         unsigned long descriptor_status;
 535
 536         descriptor_status =
 537                 ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK) << 1;
 538         return descriptor_status;
 539 }
 540
 541 /*
 542  * Return whether the status of the descriptor that is normally used for this
 543  * cpu (the one indexed by its hub-relative cpu number) is busy.
 544  * The status of the original 32 descriptors is always reflected in the 64
 545  * bits of UVH_LB_BAU_SB_ACTIVATION_STATUS_0.
 546  * The bit provided by the activation_status_2 register is irrelevant to
 547  * the status if it is only being tested for busy or not busy.
 548  */
 549 int normal_busy(struct bau_control *bcp)
 550 {
 551         int cpu = bcp->uvhub_cpu;
 552         int mmr_offset;
 553         int right_shift;
 554
 555         mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 556         right_shift = cpu * UV_ACT_STATUS_SIZE;
 557         return (((((read_lmmr(mmr_offset) >> right_shift) &
 558                                 UV_ACT_STATUS_MASK)) << 1) == UV2H_DESC_BUSY);
 559 }
 560
 561 /*
 562  * Entered when a bau descriptor has gone into a permanent busy wait because
 563  * of a hardware bug.
 564  * Workaround the bug.
 565  */
 566 int handle_uv2_busy(struct bau_control *bcp)
 567 {
 568         struct ptc_stats *stat = bcp->statp;
 569
 570         stat->s_uv2_wars++;
 571         bcp->busy = 1;
 572         return FLUSH_GIVEUP;
 573 }
 574
 575 static int uv2_wait_completion(struct bau_desc *bau_desc,
 576                                 unsigned long mmr_offset, int right_shift,
 577                                 struct bau_control *bcp, long try)
 578 {
 579         unsigned long descriptor_stat;
 580         cycles_t ttm;
 581         int desc = bcp->uvhub_cpu;
 582         long busy_reps = 0;
 583         struct ptc_stats *stat = bcp->statp;
 584
 585         descriptor_stat = uv2_read_status(mmr_offset, right_shift, desc);
 586
 587         /* spin on the status MMR, waiting for it to go idle */
 588         while (descriptor_stat != UV2H_DESC_IDLE) {
 589                 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT)) {
 590                         /*
 591                          * A h/w bug on the destination side may
 592                          * have prevented the message being marked
 593                          * pending, thus it doesn't get replied to
 594                          * and gets continually nacked until it times
 595                          * out with a SOURCE_TIMEOUT.
 596                          */
 597                         stat->s_stimeout++;
 598                         return FLUSH_GIVEUP;
 599                 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
 600                         ttm = get_cycles();
 601
 602                         /*
 603                          * Our retries may be blocked by all destination
 604                          * swack resources being consumed, and a timeout
 605                          * pending.  In that case hardware returns the
 606                          * ERROR that looks like a destination timeout.
 607                          * Without using the extended status we have to
 608                          * deduce from the short time that this was a
 609                          * strong nack.
 610                          */
 611                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 612                                 bcp->conseccompletes = 0;
 613                                 stat->s_plugged++;
 614                                 /* FLUSH_RETRY_PLUGGED causes hang on boot */
 615                                 return FLUSH_GIVEUP;
 616                         }
 617                         stat->s_dtimeout++;
 618                         bcp->conseccompletes = 0;
 619                         /* FLUSH_RETRY_TIMEOUT causes hang on boot */
 620                         return FLUSH_GIVEUP;
 621                 } else {
 622                         busy_reps++;
 623                         if (busy_reps > 1000000) {
 624                                 /* not to hammer on the clock */
 625                                 busy_reps = 0;
 626                                 ttm = get_cycles();
 627                                 if ((ttm - bcp->send_message) >
 628                                                 bcp->timeout_interval)
 629                                         return handle_uv2_busy(bcp);
 630                         }
 631                         /*
 632                          * descriptor_stat is still BUSY
 633                          */
 634                         cpu_relax();
 635                 }
 636                 descriptor_stat = uv2_read_status(mmr_offset, right_shift,
 637                                                                         desc);
 638         }
 639         bcp->conseccompletes++;
 640         return FLUSH_COMPLETE;
 641 }
 642
 643 /*
 644  * There are 2 status registers; each and array[32] of 2 bits. Set up for
 645  * which register to read and position in that register based on cpu in
 646  * current hub.
 647  */
 648 static int wait_completion(struct bau_desc *bau_desc,
 649                                 struct bau_control *bcp, long try)
 650 {
 651         int right_shift;
 652         unsigned long mmr_offset;
 653         int desc = bcp->uvhub_cpu;
 654
 655         if (desc < UV_CPUS_PER_AS) {
 656                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 657                 right_shift = desc * UV_ACT_STATUS_SIZE;
 658         } else {
 659                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
 660                 right_shift = ((desc - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
 661         }
 662
 663         if (bcp->uvhub_version == 1)
 664                 return uv1_wait_completion(bau_desc, mmr_offset, right_shift,
 665                                                                 bcp, try);
 666         else
 667                 return uv2_wait_completion(bau_desc, mmr_offset, right_shift,
 668                                                                 bcp, try);
 669 }
 670
 671 static inline cycles_t sec_2_cycles(unsigned long sec)
 672 {
 673         unsigned long ns;
 674         cycles_t cyc;
 675
 676         ns = sec * 1000000000;
 677         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
 678         return cyc;
 679 }
 680
 681 /*
 682  * Our retries are blocked by all destination sw ack resources being
 683  * in use, and a timeout is pending. In that case hardware immediately
 684  * returns the ERROR that looks like a destination timeout.
 685  */
 686 static void destination_plugged(struct bau_desc *bau_desc,
 687                         struct bau_control *bcp,
 688                         struct bau_control *hmaster, struct ptc_stats *stat)
 689 {
 690         udelay(bcp->plugged_delay);
 691         bcp->plugged_tries++;
 692
 693         if (bcp->plugged_tries >= bcp->plugsb4reset) {
 694                 bcp->plugged_tries = 0;
 695
 696                 quiesce_local_uvhub(hmaster);
 697
 698                 spin_lock(&hmaster->queue_lock);
 699                 reset_with_ipi(&bau_desc->distribution, bcp);
 700                 spin_unlock(&hmaster->queue_lock);
 701
 702                 end_uvhub_quiesce(hmaster);
 703
 704                 bcp->ipi_attempts++;
 705                 stat->s_resets_plug++;
 706         }
 707 }
 708
 709 static void destination_timeout(struct bau_desc *bau_desc,
 710                         struct bau_control *bcp, struct bau_control *hmaster,
 711                         struct ptc_stats *stat)
 712 {
 713         hmaster->max_concurr = 1;
 714         bcp->timeout_tries++;
 715         if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
 716                 bcp->timeout_tries = 0;
 717
 718                 quiesce_local_uvhub(hmaster);
 719
 720                 spin_lock(&hmaster->queue_lock);
 721                 reset_with_ipi(&bau_desc->distribution, bcp);
 722                 spin_unlock(&hmaster->queue_lock);
 723
 724                 end_uvhub_quiesce(hmaster);
 725
 726                 bcp->ipi_attempts++;
 727                 stat->s_resets_timeout++;
 728         }
 729 }
 730
 731 /*
 732  * Stop all cpus on a uvhub from using the BAU for a period of time.
 733  * This is reversed by check_enable.
 734  */
 735 static void disable_for_period(struct bau_control *bcp, struct ptc_stats *stat)
 736 {
 737         int tcpu;
 738         struct bau_control *tbcp;
 739         struct bau_control *hmaster;
 740         cycles_t tm1;
 741
 742         hmaster = bcp->uvhub_master;
 743         spin_lock(&hmaster->disable_lock);
 744         if (!bcp->baudisabled) {
 745                 stat->s_bau_disabled++;
 746                 tm1 = get_cycles();
 747                 for_each_present_cpu(tcpu) {
 748                         tbcp = &per_cpu(bau_control, tcpu);
 749                         if (tbcp->uvhub_master == hmaster) {
 750                                 tbcp->baudisabled = 1;
 751                                 tbcp->set_bau_on_time =
 752                                         tm1 + bcp->disabled_period;
 753                         }
 754                 }
 755         }
 756         spin_unlock(&hmaster->disable_lock);
 757 }
 758
 759 static void count_max_concurr(int stat, struct bau_control *bcp,
 760                                 struct bau_control *hmaster)
 761 {
 762         bcp->plugged_tries = 0;
 763         bcp->timeout_tries = 0;
 764         if (stat != FLUSH_COMPLETE)
 765                 return;
 766         if (bcp->conseccompletes <= bcp->complete_threshold)
 767                 return;
 768         if (hmaster->max_concurr >= hmaster->max_concurr_const)
 769                 return;
 770         hmaster->max_concurr++;
 771 }
 772
 773 static void record_send_stats(cycles_t time1, cycles_t time2,
 774                 struct bau_control *bcp, struct ptc_stats *stat,
 775                 int completion_status, int try)
 776 {
 777         cycles_t elapsed;
 778
 779         if (time2 > time1) {
 780                 elapsed = time2 - time1;
 781                 stat->s_time += elapsed;
 782
 783                 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
 784                         bcp->period_requests++;
 785                         bcp->period_time += elapsed;
 786                         if ((elapsed > congested_cycles) &&
 787                             (bcp->period_requests > bcp->cong_reps) &&
 788                             ((bcp->period_time / bcp->period_requests) >
 789                                                         congested_cycles)) {
 790                                 stat->s_congested++;
 791                                 disable_for_period(bcp, stat);
 792                         }
 793                 }
 794         } else
 795                 stat->s_requestor--;
 796
 797         if (completion_status == FLUSH_COMPLETE && try > 1)
 798                 stat->s_retriesok++;
 799         else if (completion_status == FLUSH_GIVEUP) {
 800                 stat->s_giveup++;
 801                 if (get_cycles() > bcp->period_end)
 802                         bcp->period_giveups = 0;
 803                 bcp->period_giveups++;
 804                 if (bcp->period_giveups == 1)
 805                         bcp->period_end = get_cycles() + bcp->disabled_period;
 806                 if (bcp->period_giveups > bcp->giveup_limit) {
 807                         disable_for_period(bcp, stat);
 808                         stat->s_giveuplimit++;
 809                 }
 810         }
 811 }
 812
 813 /*
 814  * Because of a uv1 hardware bug only a limited number of concurrent
 815  * requests can be made.
 816  */
 817 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
 818 {
 819         spinlock_t *lock = &hmaster->uvhub_lock;
 820         atomic_t *v;
 821
 822         v = &hmaster->active_descriptor_count;
 823         if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
 824                 stat->s_throttles++;
 825                 do {
 826                         cpu_relax();
 827                 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
 828         }
 829 }
 830
 831 /*
 832  * Handle the completion status of a message send.
 833  */
 834 static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
 835                         struct bau_control *bcp, struct bau_control *hmaster,
 836                         struct ptc_stats *stat)
 837 {
 838         if (completion_status == FLUSH_RETRY_PLUGGED)
 839                 destination_plugged(bau_desc, bcp, hmaster, stat);
 840         else if (completion_status == FLUSH_RETRY_TIMEOUT)
 841                 destination_timeout(bau_desc, bcp, hmaster, stat);
 842 }
 843
 844 /*
 845  * Send a broadcast and wait for it to complete.
 846  *
 847  * The flush_mask contains the cpus the broadcast is to be sent to including
 848  * cpus that are on the local uvhub.
 849  *
 850  * Returns 0 if all flushing represented in the mask was done.
 851  * Returns 1 if it gives up entirely and the original cpu mask is to be
 852  * returned to the kernel.
 853  */
 854 int uv_flush_send_and_wait(struct cpumask *flush_mask, struct bau_control *bcp,
 855         struct bau_desc *bau_desc)
 856 {
 857         int seq_number = 0;
 858         int completion_stat = 0;
 859         int uv1 = 0;
 860         long try = 0;
 861         unsigned long index;
 862         cycles_t time1;
 863         cycles_t time2;
 864         struct ptc_stats *stat = bcp->statp;
 865         struct bau_control *hmaster = bcp->uvhub_master;
 866         struct uv1_bau_msg_header *uv1_hdr = NULL;
 867         struct uv2_bau_msg_header *uv2_hdr = NULL;
 868
 869         if (bcp->uvhub_version == 1) {
 870                 uv1 = 1;
 871                 uv1_throttle(hmaster, stat);
 872         }
 873
 874         while (hmaster->uvhub_quiesce)
 875                 cpu_relax();
 876
 877         time1 = get_cycles();
 878         if (uv1)
 879                 uv1_hdr = &bau_desc->header.uv1_hdr;
 880         else
 881                 uv2_hdr = &bau_desc->header.uv2_hdr;
 882
 883         do {
 884                 if (try == 0) {
 885                         if (uv1)
 886                                 uv1_hdr->msg_type = MSG_REGULAR;
 887                         else
 888                                 uv2_hdr->msg_type = MSG_REGULAR;
 889                         seq_number = bcp->message_number++;
 890                 } else {
 891                         if (uv1)
 892                                 uv1_hdr->msg_type = MSG_RETRY;
 893                         else
 894                                 uv2_hdr->msg_type = MSG_RETRY;
 895                         stat->s_retry_messages++;
 896                 }
 897
 898                 if (uv1)
 899                         uv1_hdr->sequence = seq_number;
 900                 else
 901                         uv2_hdr->sequence = seq_number;
 902                 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
 903                 bcp->send_message = get_cycles();
 904
 905                 write_mmr_activation(index);
 906
 907                 try++;
 908                 completion_stat = wait_completion(bau_desc, bcp, try);
 909
 910                 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
 911
 912                 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
 913                         bcp->ipi_attempts = 0;
 914                         stat->s_overipilimit++;
 915                         completion_stat = FLUSH_GIVEUP;
 916                         break;
 917                 }
 918                 cpu_relax();
 919         } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
 920                  (completion_stat == FLUSH_RETRY_TIMEOUT));
 921
 922         time2 = get_cycles();
 923
 924         count_max_concurr(completion_stat, bcp, hmaster);
 925
 926         while (hmaster->uvhub_quiesce)
 927                 cpu_relax();
 928
 929         atomic_dec(&hmaster->active_descriptor_count);
 930
 931         record_send_stats(time1, time2, bcp, stat, completion_stat, try);
 932
 933         if (completion_stat == FLUSH_GIVEUP)
 934                 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
 935                 return 1;
 936         return 0;
 937 }
 938
 939 /*
 940  * The BAU is disabled for this uvhub. When the disabled time period has
 941  * expired re-enable it.
 942  * Return 0 if it is re-enabled for all cpus on this uvhub.
 943  */
 944 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
 945 {
 946         int tcpu;
 947         struct bau_control *tbcp;
 948         struct bau_control *hmaster;
 949
 950         hmaster = bcp->uvhub_master;
 951         spin_lock(&hmaster->disable_lock);
 952         if (bcp->baudisabled && (get_cycles() >= bcp->set_bau_on_time)) {
 953                 stat->s_bau_reenabled++;
 954                 for_each_present_cpu(tcpu) {
 955                         tbcp = &per_cpu(bau_control, tcpu);
 956                         if (tbcp->uvhub_master == hmaster) {
 957                                 tbcp->baudisabled = 0;
 958                                 tbcp->period_requests = 0;
 959                                 tbcp->period_time = 0;
 960                                 tbcp->period_giveups = 0;
 961                         }
 962                 }
 963                 spin_unlock(&hmaster->disable_lock);
 964                 return 0;
 965         }
 966         spin_unlock(&hmaster->disable_lock);
 967         return -1;
 968 }
 969
 970 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
 971                                 int remotes, struct bau_desc *bau_desc)
 972 {
 973         stat->s_requestor++;
 974         stat->s_ntargcpu += remotes + locals;
 975         stat->s_ntargremotes += remotes;
 976         stat->s_ntarglocals += locals;
 977
 978         /* uvhub statistics */
 979         hubs = bau_uvhub_weight(&bau_desc->distribution);
 980         if (locals) {
 981                 stat->s_ntarglocaluvhub++;
 982                 stat->s_ntargremoteuvhub += (hubs - 1);
 983         } else
 984                 stat->s_ntargremoteuvhub += hubs;
 985
 986         stat->s_ntarguvhub += hubs;
 987
 988         if (hubs >= 16)
 989                 stat->s_ntarguvhub16++;
 990         else if (hubs >= 8)
 991                 stat->s_ntarguvhub8++;
 992         else if (hubs >= 4)
 993                 stat->s_ntarguvhub4++;
 994         else if (hubs >= 2)
 995                 stat->s_ntarguvhub2++;
 996         else
 997                 stat->s_ntarguvhub1++;
 998 }
 999
1000 /*
1001  * Translate a cpu mask to the uvhub distribution mask in the BAU
1002  * activation descriptor.
1003  */
1004 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
1005                         struct bau_desc *bau_desc, int *localsp, int *remotesp)
1006 {
1007         int cpu;
1008         int pnode;
1009         int cnt = 0;
1010         struct hub_and_pnode *hpp;
1011
1012         for_each_cpu(cpu, flush_mask) {
1013                 /*
1014                  * The distribution vector is a bit map of pnodes, relative
1015                  * to the partition base pnode (and the partition base nasid
1016                  * in the header).
1017                  * Translate cpu to pnode and hub using a local memory array.
1018                  */
1019                 hpp = &bcp->socket_master->thp[cpu];
1020                 pnode = hpp->pnode - bcp->partition_base_pnode;
1021                 bau_uvhub_set(pnode, &bau_desc->distribution);
1022                 cnt++;
1023                 if (hpp->uvhub == bcp->uvhub)
1024                         (*localsp)++;
1025                 else
1026                         (*remotesp)++;
1027         }
1028         if (!cnt)
1029                 return 1;
1030         return 0;
1031 }
1032
1033 /*
1034  * globally purge translation cache of a virtual address or all TLB's
1035  * @cpumask: mask of all cpu's in which the address is to be removed
1036  * @mm: mm_struct containing virtual address range
1037  * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
1038  * @cpu: the current cpu
1039  *
1040  * This is the entry point for initiating any UV global TLB shootdown.
1041  *
1042  * Purges the translation caches of all specified processors of the given
1043  * virtual address, or purges all TLB's on specified processors.
1044  *
1045  * The caller has derived the cpumask from the mm_struct.  This function
1046  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1047  *
1048  * The cpumask is converted into a uvhubmask of the uvhubs containing
1049  * those cpus.
1050  *
1051  * Note that this function should be called with preemption disabled.
1052  *
1053  * Returns NULL if all remote flushing was done.
1054  * Returns pointer to cpumask if some remote flushing remains to be
1055  * done.  The returned pointer is valid till preemption is re-enabled.
1056  */
1057 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1058                                 struct mm_struct *mm, unsigned long start,
1059                                 unsigned end, unsigned int cpu)
1060 {
1061         int locals = 0;
1062         int remotes = 0;
1063         int hubs = 0;
1064         struct bau_desc *bau_desc;
1065         struct cpumask *flush_mask;
1066         struct ptc_stats *stat;
1067         struct bau_control *bcp;
1068         unsigned long descriptor_status;
1069         unsigned long status;
1070
1071         bcp = &per_cpu(bau_control, cpu);
1072         stat = bcp->statp;
1073         stat->s_enters++;
1074
1075         if (bcp->nobau)
1076                 return cpumask;
1077
1078         if (bcp->busy) {
1079                 descriptor_status =
1080                         read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1081                 status = ((descriptor_status >> (bcp->uvhub_cpu *
1082                         UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1083                 if (status == UV2H_DESC_BUSY)
1084                         return cpumask;
1085                 bcp->busy = 0;
1086         }
1087
1088         /* bau was disabled due to slow response */
1089         if (bcp->baudisabled) {
1090                 if (check_enable(bcp, stat)) {
1091                         stat->s_ipifordisabled++;
1092                         return cpumask;
1093                 }
1094         }
1095
1096         /*
1097          * Each sending cpu has a per-cpu mask which it fills from the caller's
1098          * cpu mask.  All cpus are converted to uvhubs and copied to the
1099          * activation descriptor.
1100          */
1101         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1102         /* don't actually do a shootdown of the local cpu */
1103         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1104
1105         if (cpu_isset(cpu, *cpumask))
1106                 stat->s_ntargself++;
1107
1108         bau_desc = bcp->descriptor_base;
1109         bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1110         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1111         if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1112                 return NULL;
1113
1114         record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1115
1116         bau_desc->payload.address = start;
1117         bau_desc->payload.sending_cpu = cpu;
1118         /*
1119          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1120          * or 1 if it gave up and the original cpumask should be returned.
1121          */
1122         if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1123                 return NULL;
1124         else
1125                 return cpumask;
1126 }
1127
1128 /*
1129  * Search the message queue for any 'other' unprocessed message with the
1130  * same software acknowledge resource bit vector as the 'msg' message.
1131  */
1132 struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1133                                            struct bau_control *bcp)
1134 {
1135         struct bau_pq_entry *msg_next = msg + 1;
1136         unsigned char swack_vec = msg->swack_vec;
1137
1138         if (msg_next > bcp->queue_last)
1139                 msg_next = bcp->queue_first;
1140         while (msg_next != msg) {
1141                 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1142                                 (msg_next->swack_vec == swack_vec))
1143                         return msg_next;
1144                 msg_next++;
1145                 if (msg_next > bcp->queue_last)
1146                         msg_next = bcp->queue_first;
1147         }
1148         return NULL;
1149 }
1150
1151 /*
1152  * UV2 needs to work around a bug in which an arriving message has not
1153  * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1154  * Such a message must be ignored.
1155  */
1156 void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1157 {
1158         unsigned long mmr_image;
1159         unsigned char swack_vec;
1160         struct bau_pq_entry *msg = mdp->msg;
1161         struct bau_pq_entry *other_msg;
1162
1163         mmr_image = read_mmr_sw_ack();
1164         swack_vec = msg->swack_vec;
1165
1166         if ((swack_vec & mmr_image) == 0) {
1167                 /*
1168                  * This message was assigned a swack resource, but no
1169                  * reserved acknowlegment is pending.
1170                  * The bug has prevented this message from setting the MMR.
1171                  */
1172                 /*
1173                  * Some message has set the MMR 'pending' bit; it might have
1174                  * been another message.  Look for that message.
1175                  */
1176                 other_msg = find_another_by_swack(msg, bcp);
1177                 if (other_msg) {
1178                         /*
1179                          * There is another. Process this one but do not
1180                          * ack it.
1181                          */
1182                         bau_process_message(mdp, bcp, 0);
1183                         /*
1184                          * Let the natural processing of that other message
1185                          * acknowledge it. Don't get the processing of sw_ack's
1186                          * out of order.
1187                          */
1188                         return;
1189                 }
1190         }
1191
1192         /*
1193          * Either the MMR shows this one pending a reply or there is no
1194          * other message using this sw_ack, so it is safe to acknowledge it.
1195          */
1196         bau_process_message(mdp, bcp, 1);
1197
1198         return;
1199 }
1200
1201 /*
1202  * The BAU message interrupt comes here. (registered by set_intr_gate)
1203  * See entry_64.S
1204  *
1205  * We received a broadcast assist message.
1206  *
1207  * Interrupts are disabled; this interrupt could represent
1208  * the receipt of several messages.
1209  *
1210  * All cores/threads on this hub get this interrupt.
1211  * The last one to see it does the software ack.
1212  * (the resource will not be freed until noninterruptable cpus see this
1213  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
1214  */
1215 void uv_bau_message_interrupt(struct pt_regs *regs)
1216 {
1217         int count = 0;
1218         cycles_t time_start;
1219         struct bau_pq_entry *msg;
1220         struct bau_control *bcp;
1221         struct ptc_stats *stat;
1222         struct msg_desc msgdesc;
1223
1224         ack_APIC_irq();
1225         time_start = get_cycles();
1226
1227         bcp = &per_cpu(bau_control, smp_processor_id());
1228         stat = bcp->statp;
1229
1230         msgdesc.queue_first = bcp->queue_first;
1231         msgdesc.queue_last = bcp->queue_last;
1232
1233         msg = bcp->bau_msg_head;
1234         while (msg->swack_vec) {
1235                 count++;
1236
1237                 msgdesc.msg_slot = msg - msgdesc.queue_first;
1238                 msgdesc.msg = msg;
1239                 if (bcp->uvhub_version == 2)
1240                         process_uv2_message(&msgdesc, bcp);
1241                 else
1242                         bau_process_message(&msgdesc, bcp, 1);
1243
1244                 msg++;
1245                 if (msg > msgdesc.queue_last)
1246                         msg = msgdesc.queue_first;
1247                 bcp->bau_msg_head = msg;
1248         }
1249         stat->d_time += (get_cycles() - time_start);
1250         if (!count)
1251                 stat->d_nomsg++;
1252         else if (count > 1)
1253                 stat->d_multmsg++;
1254 }
1255
1256 /*
1257  * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1258  * shootdown message timeouts enabled.  The timeout does not cause
1259  * an interrupt, but causes an error message to be returned to
1260  * the sender.
1261  */
1262 static void __init enable_timeouts(void)
1263 {
1264         int uvhub;
1265         int nuvhubs;
1266         int pnode;
1267         unsigned long mmr_image;
1268
1269         nuvhubs = uv_num_possible_blades();
1270
1271         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1272                 if (!uv_blade_nr_possible_cpus(uvhub))
1273                         continue;
1274
1275                 pnode = uv_blade_to_pnode(uvhub);
1276                 mmr_image = read_mmr_misc_control(pnode);
1277                 /*
1278                  * Set the timeout period and then lock it in, in three
1279                  * steps; captures and locks in the period.
1280                  *
1281                  * To program the period, the SOFT_ACK_MODE must be off.
1282                  */
1283                 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1284                 write_mmr_misc_control(pnode, mmr_image);
1285                 /*
1286                  * Set the 4-bit period.
1287                  */
1288                 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1289                 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1290                 write_mmr_misc_control(pnode, mmr_image);
1291                 /*
1292                  * UV1:
1293                  * Subsequent reversals of the timebase bit (3) cause an
1294                  * immediate timeout of one or all INTD resources as
1295                  * indicated in bits 2:0 (7 causes all of them to timeout).
1296                  */
1297                 mmr_image |= (1L << SOFTACK_MSHIFT);
1298                 if (is_uv2_hub()) {
1299                         /* hw bug workaround; do not use extended status */
1300                         mmr_image &= ~(1L << UV2_EXT_SHFT);
1301                 }
1302                 write_mmr_misc_control(pnode, mmr_image);
1303         }
1304 }
1305
1306 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1307 {
1308         if (*offset < num_possible_cpus())
1309                 return offset;
1310         return NULL;
1311 }
1312
1313 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1314 {
1315         (*offset)++;
1316         if (*offset < num_possible_cpus())
1317                 return offset;
1318         return NULL;
1319 }
1320
1321 static void ptc_seq_stop(struct seq_file *file, void *data)
1322 {
1323 }
1324
1325 static inline unsigned long long usec_2_cycles(unsigned long microsec)
1326 {
1327         unsigned long ns;
1328         unsigned long long cyc;
1329
1330         ns = microsec * 1000;
1331         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
1332         return cyc;
1333 }
1334
1335 /*
1336  * Display the statistics thru /proc/sgi_uv/ptc_statistics
1337  * 'data' points to the cpu number
1338  * Note: see the descriptions in stat_description[].
1339  */
1340 static int ptc_seq_show(struct seq_file *file, void *data)
1341 {
1342         struct ptc_stats *stat;
1343         struct bau_control *bcp;
1344         int cpu;
1345
1346         cpu = *(loff_t *)data;
1347         if (!cpu) {
1348                 seq_printf(file,
1349                  "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1350                 seq_printf(file,
1351                         "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1352                 seq_printf(file,
1353                         "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1354                 seq_printf(file,
1355                         "rok resetp resett giveup sto bz throt disable ");
1356                 seq_printf(file,
1357                         "enable wars warshw warwaits enters ipidis plugged ");
1358                 seq_printf(file,
1359                         "ipiover glim cong swack recv rtime all one mult ");
1360                 seq_printf(file,
1361                         "none retry canc nocan reset rcan\n");
1362         }
1363         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1364                 bcp = &per_cpu(bau_control, cpu);
1365                 stat = bcp->statp;
1366                 /* source side statistics */
1367                 seq_printf(file,
1368                         "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1369                            cpu, bcp->nobau, stat->s_requestor,
1370                            cycles_2_us(stat->s_time),
1371                            stat->s_ntargself, stat->s_ntarglocals,
1372                            stat->s_ntargremotes, stat->s_ntargcpu,
1373                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1374                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
1375                 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1376                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1377                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1378                            stat->s_dtimeout, stat->s_strongnacks);
1379                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1380                            stat->s_retry_messages, stat->s_retriesok,
1381                            stat->s_resets_plug, stat->s_resets_timeout,
1382                            stat->s_giveup, stat->s_stimeout,
1383                            stat->s_busy, stat->s_throttles);
1384                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1385                            stat->s_bau_disabled, stat->s_bau_reenabled,
1386                            stat->s_uv2_wars, stat->s_uv2_wars_hw,
1387                            stat->s_uv2_war_waits, stat->s_enters,
1388                            stat->s_ipifordisabled, stat->s_plugged,
1389                            stat->s_overipilimit, stat->s_giveuplimit,
1390                            stat->s_congested);
1391
1392                 /* destination side statistics */
1393                 seq_printf(file,
1394                         "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1395                            read_gmmr_sw_ack(uv_cpu_to_pnode(cpu)),
1396                            stat->d_requestee, cycles_2_us(stat->d_time),
1397                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1398                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
1399                            stat->d_nocanceled, stat->d_resets,
1400                            stat->d_rcanceled);
1401         }
1402         return 0;
1403 }
1404
1405 /*
1406  * Display the tunables thru debugfs
1407  */
1408 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1409                                 size_t count, loff_t *ppos)
1410 {
1411         char *buf;
1412         int ret;
1413
1414         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1415                 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1416                 "ipi_reset_limit complete_threshold congested_response_us",
1417                 "congested_reps disabled_period giveup_limit",
1418                 max_concurr, plugged_delay, plugsb4reset,
1419                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1420                 congested_respns_us, congested_reps, disabled_period,
1421                 giveup_limit);
1422
1423         if (!buf)
1424                 return -ENOMEM;
1425
1426         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1427         kfree(buf);
1428         return ret;
1429 }
1430
1431 /*
1432  * handle a write to /proc/sgi_uv/ptc_statistics
1433  * -1: reset the statistics
1434  *  0: display meaning of the statistics
1435  */
1436 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1437                                 size_t count, loff_t *data)
1438 {
1439         int cpu;
1440         int i;
1441         int elements;
1442         long input_arg;
1443         char optstr[64];
1444         struct ptc_stats *stat;
1445
1446         if (count == 0 || count > sizeof(optstr))
1447                 return -EINVAL;
1448         if (copy_from_user(optstr, user, count))
1449                 return -EFAULT;
1450         optstr[count - 1] = '\0';
1451
1452         if (!strcmp(optstr, "on")) {
1453                 set_bau_on();
1454                 return count;
1455         } else if (!strcmp(optstr, "off")) {
1456                 set_bau_off();
1457                 return count;
1458         }
1459
1460         if (strict_strtol(optstr, 10, &input_arg) < 0) {
1461                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1462                 return -EINVAL;
1463         }
1464
1465         if (input_arg == 0) {
1466                 elements = sizeof(stat_description)/sizeof(*stat_description);
1467                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1468                 printk(KERN_DEBUG "Sender statistics:\n");
1469                 for (i = 0; i < elements; i++)
1470                         printk(KERN_DEBUG "%s\n", stat_description[i]);
1471         } else if (input_arg == -1) {
1472                 for_each_present_cpu(cpu) {
1473                         stat = &per_cpu(ptcstats, cpu);
1474                         memset(stat, 0, sizeof(struct ptc_stats));
1475                 }
1476         }
1477
1478         return count;
1479 }
1480
1481 static int local_atoi(const char *name)
1482 {
1483         int val = 0;
1484
1485         for (;; name++) {
1486                 switch (*name) {
1487                 case '0' ... '9':
1488                         val = 10*val+(*name-'0');
1489                         break;
1490                 default:
1491                         return val;
1492                 }
1493         }
1494 }
1495
1496 /*
1497  * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1498  * Zero values reset them to defaults.
1499  */
1500 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1501                                 int count)
1502 {
1503         char *p;
1504         char *q;
1505         int cnt = 0;
1506         int val;
1507         int e = sizeof(tunables) / sizeof(*tunables);
1508
1509         p = instr + strspn(instr, WHITESPACE);
1510         q = p;
1511         for (; *p; p = q + strspn(q, WHITESPACE)) {
1512                 q = p + strcspn(p, WHITESPACE);
1513                 cnt++;
1514                 if (q == p)
1515                         break;
1516         }
1517         if (cnt != e) {
1518                 printk(KERN_INFO "bau tunable error: should be %d values\n", e);
1519                 return -EINVAL;
1520         }
1521
1522         p = instr + strspn(instr, WHITESPACE);
1523         q = p;
1524         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1525                 q = p + strcspn(p, WHITESPACE);
1526                 val = local_atoi(p);
1527                 switch (cnt) {
1528                 case 0:
1529                         if (val == 0) {
1530                                 max_concurr = MAX_BAU_CONCURRENT;
1531                                 max_concurr_const = MAX_BAU_CONCURRENT;
1532                                 continue;
1533                         }
1534                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1535                                 printk(KERN_DEBUG
1536                                 "Error: BAU max concurrent %d is invalid\n",
1537                                 val);
1538                                 return -EINVAL;
1539                         }
1540                         max_concurr = val;
1541                         max_concurr_const = val;
1542                         continue;
1543                 default:
1544                         if (val == 0)
1545                                 *tunables[cnt].tunp = tunables[cnt].deflt;
1546                         else
1547                                 *tunables[cnt].tunp = val;
1548                         continue;
1549                 }
1550                 if (q == p)
1551                         break;
1552         }
1553         return 0;
1554 }
1555
1556 /*
1557  * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1558  */
1559 static ssize_t tunables_write(struct file *file, const char __user *user,
1560                                 size_t count, loff_t *data)
1561 {
1562         int cpu;
1563         int ret;
1564         char instr[100];
1565         struct bau_control *bcp;
1566
1567         if (count == 0 || count > sizeof(instr)-1)
1568                 return -EINVAL;
1569         if (copy_from_user(instr, user, count))
1570                 return -EFAULT;
1571
1572         instr[count] = '\0';
1573
1574         cpu = get_cpu();
1575         bcp = &per_cpu(bau_control, cpu);
1576         ret = parse_tunables_write(bcp, instr, count);
1577         put_cpu();
1578         if (ret)
1579                 return ret;
1580
1581         for_each_present_cpu(cpu) {
1582                 bcp = &per_cpu(bau_control, cpu);
1583                 bcp->max_concurr =              max_concurr;
1584                 bcp->max_concurr_const =        max_concurr;
1585                 bcp->plugged_delay =            plugged_delay;
1586                 bcp->plugsb4reset =             plugsb4reset;
1587                 bcp->timeoutsb4reset =          timeoutsb4reset;
1588                 bcp->ipi_reset_limit =          ipi_reset_limit;
1589                 bcp->complete_threshold =       complete_threshold;
1590                 bcp->cong_response_us =         congested_respns_us;
1591                 bcp->cong_reps =                congested_reps;
1592                 bcp->disabled_period =          sec_2_cycles(disabled_period);
1593                 bcp->giveup_limit =             giveup_limit;
1594         }
1595         return count;
1596 }
1597
1598 static const struct seq_operations uv_ptc_seq_ops = {
1599         .start          = ptc_seq_start,
1600         .next           = ptc_seq_next,
1601         .stop           = ptc_seq_stop,
1602         .show           = ptc_seq_show
1603 };
1604
1605 static int ptc_proc_open(struct inode *inode, struct file *file)
1606 {
1607         return seq_open(file, &uv_ptc_seq_ops);
1608 }
1609
1610 static int tunables_open(struct inode *inode, struct file *file)
1611 {
1612         return 0;
1613 }
1614
1615 static const struct file_operations proc_uv_ptc_operations = {
1616         .open           = ptc_proc_open,
1617         .read           = seq_read,
1618         .write          = ptc_proc_write,
1619         .llseek         = seq_lseek,
1620         .release        = seq_release,
1621 };
1622
1623 static const struct file_operations tunables_fops = {
1624         .open           = tunables_open,
1625         .read           = tunables_read,
1626         .write          = tunables_write,
1627         .llseek         = default_llseek,
1628 };
1629
1630 static int __init uv_ptc_init(void)
1631 {
1632         struct proc_dir_entry *proc_uv_ptc;
1633
1634         if (!is_uv_system())
1635                 return 0;
1636
1637         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1638                                   &proc_uv_ptc_operations);
1639         if (!proc_uv_ptc) {
1640                 printk(KERN_ERR "unable to create %s proc entry\n",
1641                        UV_PTC_BASENAME);
1642                 return -EINVAL;
1643         }
1644
1645         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1646         if (!tunables_dir) {
1647                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1648                        UV_BAU_TUNABLES_DIR);
1649                 return -EINVAL;
1650         }
1651         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1652                                         tunables_dir, NULL, &tunables_fops);
1653         if (!tunables_file) {
1654                 printk(KERN_ERR "unable to create debugfs file %s\n",
1655                        UV_BAU_TUNABLES_FILE);
1656                 return -EINVAL;
1657         }
1658         return 0;
1659 }
1660
1661 /*
1662  * Initialize the sending side's sending buffers.
1663  */
1664 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1665 {
1666         int i;
1667         int cpu;
1668         int uv1 = 0;
1669         unsigned long gpa;
1670         unsigned long m;
1671         unsigned long n;
1672         size_t dsize;
1673         struct bau_desc *bau_desc;
1674         struct bau_desc *bd2;
1675         struct uv1_bau_msg_header *uv1_hdr;
1676         struct uv2_bau_msg_header *uv2_hdr;
1677         struct bau_control *bcp;
1678
1679         /*
1680          * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1681          * per cpu; and one per cpu on the uvhub (ADP_SZ)
1682          */
1683         dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1684         bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1685         BUG_ON(!bau_desc);
1686
1687         gpa = uv_gpa(bau_desc);
1688         n = uv_gpa_to_gnode(gpa);
1689         m = uv_gpa_to_offset(gpa);
1690         if (is_uv1_hub())
1691                 uv1 = 1;
1692
1693         /* the 14-bit pnode */
1694         write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1695         /*
1696          * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1697          * cpu even though we only use the first one; one descriptor can
1698          * describe a broadcast to 256 uv hubs.
1699          */
1700         for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1701                 memset(bd2, 0, sizeof(struct bau_desc));
1702                 if (uv1) {
1703                         uv1_hdr = &bd2->header.uv1_hdr;
1704                         uv1_hdr->swack_flag =   1;
1705                         /*
1706                          * The base_dest_nasid set in the message header
1707                          * is the nasid of the first uvhub in the partition.
1708                          * The bit map will indicate destination pnode numbers
1709                          * relative to that base. They may not be consecutive
1710                          * if nasid striding is being used.
1711                          */
1712                         uv1_hdr->base_dest_nasid =
1713                                                 UV_PNODE_TO_NASID(base_pnode);
1714                         uv1_hdr->dest_subnodeid =       UV_LB_SUBNODEID;
1715                         uv1_hdr->command =              UV_NET_ENDPOINT_INTD;
1716                         uv1_hdr->int_both =             1;
1717                         /*
1718                          * all others need to be set to zero:
1719                          *   fairness chaining multilevel count replied_to
1720                          */
1721                 } else {
1722                         /*
1723                          * BIOS uses legacy mode, but UV2 hardware always
1724                          * uses native mode for selective broadcasts.
1725                          */
1726                         uv2_hdr = &bd2->header.uv2_hdr;
1727                         uv2_hdr->swack_flag =   1;
1728                         uv2_hdr->base_dest_nasid =
1729                                                 UV_PNODE_TO_NASID(base_pnode);
1730                         uv2_hdr->dest_subnodeid =       UV_LB_SUBNODEID;
1731                         uv2_hdr->command =              UV_NET_ENDPOINT_INTD;
1732                 }
1733         }
1734         for_each_present_cpu(cpu) {
1735                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1736                         continue;
1737                 bcp = &per_cpu(bau_control, cpu);
1738                 bcp->descriptor_base = bau_desc;
1739         }
1740 }
1741
1742 /*
1743  * initialize the destination side's receiving buffers
1744  * entered for each uvhub in the partition
1745  * - node is first node (kernel memory notion) on the uvhub
1746  * - pnode is the uvhub's physical identifier
1747  */
1748 static void pq_init(int node, int pnode)
1749 {
1750         int cpu;
1751         size_t plsize;
1752         char *cp;
1753         void *vp;
1754         unsigned long pn;
1755         unsigned long first;
1756         unsigned long pn_first;
1757         unsigned long last;
1758         struct bau_pq_entry *pqp;
1759         struct bau_control *bcp;
1760
1761         plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1762         vp = kmalloc_node(plsize, GFP_KERNEL, node);
1763         pqp = (struct bau_pq_entry *)vp;
1764         BUG_ON(!pqp);
1765
1766         cp = (char *)pqp + 31;
1767         pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1768
1769         for_each_present_cpu(cpu) {
1770                 if (pnode != uv_cpu_to_pnode(cpu))
1771                         continue;
1772                 /* for every cpu on this pnode: */
1773                 bcp = &per_cpu(bau_control, cpu);
1774                 bcp->queue_first        = pqp;
1775                 bcp->bau_msg_head       = pqp;
1776                 bcp->queue_last         = pqp + (DEST_Q_SIZE - 1);
1777         }
1778         /*
1779          * need the gnode of where the memory was really allocated
1780          */
1781         pn = uv_gpa_to_gnode(uv_gpa(pqp));
1782         first = uv_physnodeaddr(pqp);
1783         pn_first = ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | first;
1784         last = uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1));
1785         write_mmr_payload_first(pnode, pn_first);
1786         write_mmr_payload_tail(pnode, first);
1787         write_mmr_payload_last(pnode, last);
1788         write_gmmr_sw_ack(pnode, 0xffffUL);
1789
1790         /* in effect, all msg_type's are set to MSG_NOOP */
1791         memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1792 }
1793
1794 /*
1795  * Initialization of each UV hub's structures
1796  */
1797 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1798 {
1799         int node;
1800         int pnode;
1801         unsigned long apicid;
1802
1803         node = uvhub_to_first_node(uvhub);
1804         pnode = uv_blade_to_pnode(uvhub);
1805
1806         activation_descriptor_init(node, pnode, base_pnode);
1807
1808         pq_init(node, pnode);
1809         /*
1810          * The below initialization can't be in firmware because the
1811          * messaging IRQ will be determined by the OS.
1812          */
1813         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1814         write_mmr_data_config(pnode, ((apicid << 32) | vector));
1815 }
1816
1817 /*
1818  * We will set BAU_MISC_CONTROL with a timeout period.
1819  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1820  * So the destination timeout period has to be calculated from them.
1821  */
1822 static int calculate_destination_timeout(void)
1823 {
1824         unsigned long mmr_image;
1825         int mult1;
1826         int mult2;
1827         int index;
1828         int base;
1829         int ret;
1830         unsigned long ts_ns;
1831
1832         if (is_uv1_hub()) {
1833                 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1834                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1835                 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1836                 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1837                 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1838                 ts_ns = timeout_base_ns[index];
1839                 ts_ns *= (mult1 * mult2);
1840                 ret = ts_ns / 1000;
1841         } else {
1842                 /* 4 bits  0/1 for 10/80us base, 3 bits of multiplier */
1843                 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1844                 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1845                 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1846                         base = 80;
1847                 else
1848                         base = 10;
1849                 mult1 = mmr_image & UV2_ACK_MASK;
1850                 ret = mult1 * base;
1851         }
1852         return ret;
1853 }
1854
1855 static void __init init_per_cpu_tunables(void)
1856 {
1857         int cpu;
1858         struct bau_control *bcp;
1859
1860         for_each_present_cpu(cpu) {
1861                 bcp = &per_cpu(bau_control, cpu);
1862                 bcp->baudisabled                = 0;
1863                 if (nobau)
1864                         bcp->nobau              = 1;
1865                 bcp->statp                      = &per_cpu(ptcstats, cpu);
1866                 /* time interval to catch a hardware stay-busy bug */
1867                 bcp->timeout_interval           = usec_2_cycles(2*timeout_us);
1868                 bcp->max_concurr                = max_concurr;
1869                 bcp->max_concurr_const          = max_concurr;
1870                 bcp->plugged_delay              = plugged_delay;
1871                 bcp->plugsb4reset               = plugsb4reset;
1872                 bcp->timeoutsb4reset            = timeoutsb4reset;
1873                 bcp->ipi_reset_limit            = ipi_reset_limit;
1874                 bcp->complete_threshold         = complete_threshold;
1875                 bcp->cong_response_us           = congested_respns_us;
1876                 bcp->cong_reps                  = congested_reps;
1877                 bcp->disabled_period =          sec_2_cycles(disabled_period);
1878                 bcp->giveup_limit =             giveup_limit;
1879                 spin_lock_init(&bcp->queue_lock);
1880                 spin_lock_init(&bcp->uvhub_lock);
1881                 spin_lock_init(&bcp->disable_lock);
1882         }
1883 }
1884
1885 /*
1886  * Scan all cpus to collect blade and socket summaries.
1887  */
1888 static int __init get_cpu_topology(int base_pnode,
1889                                         struct uvhub_desc *uvhub_descs,
1890                                         unsigned char *uvhub_mask)
1891 {
1892         int cpu;
1893         int pnode;
1894         int uvhub;
1895         int socket;
1896         struct bau_control *bcp;
1897         struct uvhub_desc *bdp;
1898         struct socket_desc *sdp;
1899
1900         for_each_present_cpu(cpu) {
1901                 bcp = &per_cpu(bau_control, cpu);
1902
1903                 memset(bcp, 0, sizeof(struct bau_control));
1904
1905                 pnode = uv_cpu_hub_info(cpu)->pnode;
1906                 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1907                         printk(KERN_EMERG
1908                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1909                                 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1910                         return 1;
1911                 }
1912
1913                 bcp->osnode = cpu_to_node(cpu);
1914                 bcp->partition_base_pnode = base_pnode;
1915
1916                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1917                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1918                 bdp = &uvhub_descs[uvhub];
1919
1920                 bdp->num_cpus++;
1921                 bdp->uvhub = uvhub;
1922                 bdp->pnode = pnode;
1923
1924                 /* kludge: 'assuming' one node per socket, and assuming that
1925                    disabling a socket just leaves a gap in node numbers */
1926                 socket = bcp->osnode & 1;
1927                 bdp->socket_mask |= (1 << socket);
1928                 sdp = &bdp->socket[socket];
1929                 sdp->cpu_number[sdp->num_cpus] = cpu;
1930                 sdp->num_cpus++;
1931                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1932                         printk(KERN_EMERG "%d cpus per socket invalid\n",
1933                                 sdp->num_cpus);
1934                         return 1;
1935                 }
1936         }
1937         return 0;
1938 }
1939
1940 /*
1941  * Each socket is to get a local array of pnodes/hubs.
1942  */
1943 static void make_per_cpu_thp(struct bau_control *smaster)
1944 {
1945         int cpu;
1946         size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
1947
1948         smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
1949         memset(smaster->thp, 0, hpsz);
1950         for_each_present_cpu(cpu) {
1951                 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
1952                 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1953         }
1954 }
1955
1956 /*
1957  * Each uvhub is to get a local cpumask.
1958  */
1959 static void make_per_hub_cpumask(struct bau_control *hmaster)
1960 {
1961         int sz = sizeof(cpumask_t);
1962
1963         hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
1964 }
1965
1966 /*
1967  * Initialize all the per_cpu information for the cpu's on a given socket,
1968  * given what has been gathered into the socket_desc struct.
1969  * And reports the chosen hub and socket masters back to the caller.
1970  */
1971 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
1972                         struct bau_control **smasterp,
1973                         struct bau_control **hmasterp)
1974 {
1975         int i;
1976         int cpu;
1977         struct bau_control *bcp;
1978
1979         for (i = 0; i < sdp->num_cpus; i++) {
1980                 cpu = sdp->cpu_number[i];
1981                 bcp = &per_cpu(bau_control, cpu);
1982                 bcp->cpu = cpu;
1983                 if (i == 0) {
1984                         *smasterp = bcp;
1985                         if (!(*hmasterp))
1986                                 *hmasterp = bcp;
1987                 }
1988                 bcp->cpus_in_uvhub = bdp->num_cpus;
1989                 bcp->cpus_in_socket = sdp->num_cpus;
1990                 bcp->socket_master = *smasterp;
1991                 bcp->uvhub = bdp->uvhub;
1992                 if (is_uv1_hub())
1993                         bcp->uvhub_version = 1;
1994                 else if (is_uv2_hub())
1995                         bcp->uvhub_version = 2;
1996                 else {
1997                         printk(KERN_EMERG "uvhub version not 1 or 2\n");
1998                         return 1;
1999                 }
2000                 bcp->uvhub_master = *hmasterp;
2001                 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->blade_processor_id;
2002                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2003                         printk(KERN_EMERG "%d cpus per uvhub invalid\n",
2004                                 bcp->uvhub_cpu);
2005                         return 1;
2006                 }
2007         }
2008         return 0;
2009 }
2010
2011 /*
2012  * Summarize the blade and socket topology into the per_cpu structures.
2013  */
2014 static int __init summarize_uvhub_sockets(int nuvhubs,
2015                         struct uvhub_desc *uvhub_descs,
2016                         unsigned char *uvhub_mask)
2017 {
2018         int socket;
2019         int uvhub;
2020         unsigned short socket_mask;
2021
2022         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2023                 struct uvhub_desc *bdp;
2024                 struct bau_control *smaster = NULL;
2025                 struct bau_control *hmaster = NULL;
2026
2027                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2028                         continue;
2029
2030                 bdp = &uvhub_descs[uvhub];
2031                 socket_mask = bdp->socket_mask;
2032                 socket = 0;
2033                 while (socket_mask) {
2034                         struct socket_desc *sdp;
2035                         if ((socket_mask & 1)) {
2036                                 sdp = &bdp->socket[socket];
2037                                 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2038                                         return 1;
2039                                 make_per_cpu_thp(smaster);
2040                         }
2041                         socket++;
2042                         socket_mask = (socket_mask >> 1);
2043                 }
2044                 make_per_hub_cpumask(hmaster);
2045         }
2046         return 0;
2047 }
2048
2049 /*
2050  * initialize the bau_control structure for each cpu
2051  */
2052 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2053 {
2054         unsigned char *uvhub_mask;
2055         void *vp;
2056         struct uvhub_desc *uvhub_descs;
2057
2058         timeout_us = calculate_destination_timeout();
2059
2060         vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
2061         uvhub_descs = (struct uvhub_desc *)vp;
2062         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
2063         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2064
2065         if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2066                 goto fail;
2067
2068         if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2069                 goto fail;
2070
2071         kfree(uvhub_descs);
2072         kfree(uvhub_mask);
2073         init_per_cpu_tunables();
2074         return 0;
2075
2076 fail:
2077         kfree(uvhub_descs);
2078         kfree(uvhub_mask);
2079         return 1;
2080 }
2081
2082 /*
2083  * Initialization of BAU-related structures
2084  */
2085 static int __init uv_bau_init(void)
2086 {
2087         int uvhub;
2088         int pnode;
2089         int nuvhubs;
2090         int cur_cpu;
2091         int cpus;
2092         int vector;
2093         cpumask_var_t *mask;
2094
2095         if (!is_uv_system())
2096                 return 0;
2097
2098         for_each_possible_cpu(cur_cpu) {
2099                 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2100                 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2101         }
2102
2103         nuvhubs = uv_num_possible_blades();
2104         congested_cycles = usec_2_cycles(congested_respns_us);
2105
2106         uv_base_pnode = 0x7fffffff;
2107         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2108                 cpus = uv_blade_nr_possible_cpus(uvhub);
2109                 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2110                         uv_base_pnode = uv_blade_to_pnode(uvhub);
2111         }
2112
2113         enable_timeouts();
2114
2115         if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2116                 set_bau_off();
2117                 nobau_perm = 1;
2118                 return 0;
2119         }
2120
2121         vector = UV_BAU_MESSAGE;
2122         for_each_possible_blade(uvhub)
2123                 if (uv_blade_nr_possible_cpus(uvhub))
2124                         init_uvhub(uvhub, vector, uv_base_pnode);
2125
2126         alloc_intr_gate(vector, uv_bau_message_intr1);
2127
2128         for_each_possible_blade(uvhub) {
2129                 if (uv_blade_nr_possible_cpus(uvhub)) {
2130                         unsigned long val;
2131                         unsigned long mmr;
2132                         pnode = uv_blade_to_pnode(uvhub);
2133                         /* INIT the bau */
2134                         val = 1L << 63;
2135                         write_gmmr_activation(pnode, val);
2136                         mmr = 1; /* should be 1 to broadcast to both sockets */
2137                         if (!is_uv1_hub())
2138                                 write_mmr_data_broadcast(pnode, mmr);
2139                 }
2140         }
2141
2142         return 0;
2143 }
2144 core_initcall(uv_bau_init);
2145 fs_initcall(uv_ptc_init);