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  * @start: start virtual address to be removed from TLB
1038  * @end: end virtual address to be remove from TLB
1039  * @cpu: the current cpu
1040  *
1041  * This is the entry point for initiating any UV global TLB shootdown.
1042  *
1043  * Purges the translation caches of all specified processors of the given
1044  * virtual address, or purges all TLB's on specified processors.
1045  *
1046  * The caller has derived the cpumask from the mm_struct.  This function
1047  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1048  *
1049  * The cpumask is converted into a uvhubmask of the uvhubs containing
1050  * those cpus.
1051  *
1052  * Note that this function should be called with preemption disabled.
1053  *
1054  * Returns NULL if all remote flushing was done.
1055  * Returns pointer to cpumask if some remote flushing remains to be
1056  * done.  The returned pointer is valid till preemption is re-enabled.
1057  */
1058 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1059                                 struct mm_struct *mm, unsigned long start,
1060                                 unsigned long end, unsigned int cpu)
1061 {
1062         int locals = 0;
1063         int remotes = 0;
1064         int hubs = 0;
1065         struct bau_desc *bau_desc;
1066         struct cpumask *flush_mask;
1067         struct ptc_stats *stat;
1068         struct bau_control *bcp;
1069         unsigned long descriptor_status;
1070         unsigned long status;
1071
1072         bcp = &per_cpu(bau_control, cpu);
1073
1074         if (bcp->nobau)
1075                 return cpumask;
1076
1077         stat = bcp->statp;
1078         stat->s_enters++;
1079
1080         if (bcp->busy) {
1081                 descriptor_status =
1082                         read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1083                 status = ((descriptor_status >> (bcp->uvhub_cpu *
1084                         UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1085                 if (status == UV2H_DESC_BUSY)
1086                         return cpumask;
1087                 bcp->busy = 0;
1088         }
1089
1090         /* bau was disabled due to slow response */
1091         if (bcp->baudisabled) {
1092                 if (check_enable(bcp, stat)) {
1093                         stat->s_ipifordisabled++;
1094                         return cpumask;
1095                 }
1096         }
1097
1098         /*
1099          * Each sending cpu has a per-cpu mask which it fills from the caller's
1100          * cpu mask.  All cpus are converted to uvhubs and copied to the
1101          * activation descriptor.
1102          */
1103         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1104         /* don't actually do a shootdown of the local cpu */
1105         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1106
1107         if (cpu_isset(cpu, *cpumask))
1108                 stat->s_ntargself++;
1109
1110         bau_desc = bcp->descriptor_base;
1111         bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1112         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1113         if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1114                 return NULL;
1115
1116         record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1117
1118         if (!end || (end - start) <= PAGE_SIZE)
1119                 bau_desc->payload.address = start;
1120         else
1121                 bau_desc->payload.address = TLB_FLUSH_ALL;
1122         bau_desc->payload.sending_cpu = cpu;
1123         /*
1124          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1125          * or 1 if it gave up and the original cpumask should be returned.
1126          */
1127         if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1128                 return NULL;
1129         else
1130                 return cpumask;
1131 }
1132
1133 /*
1134  * Search the message queue for any 'other' unprocessed message with the
1135  * same software acknowledge resource bit vector as the 'msg' message.
1136  */
1137 struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1138                                            struct bau_control *bcp)
1139 {
1140         struct bau_pq_entry *msg_next = msg + 1;
1141         unsigned char swack_vec = msg->swack_vec;
1142
1143         if (msg_next > bcp->queue_last)
1144                 msg_next = bcp->queue_first;
1145         while (msg_next != msg) {
1146                 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1147                                 (msg_next->swack_vec == swack_vec))
1148                         return msg_next;
1149                 msg_next++;
1150                 if (msg_next > bcp->queue_last)
1151                         msg_next = bcp->queue_first;
1152         }
1153         return NULL;
1154 }
1155
1156 /*
1157  * UV2 needs to work around a bug in which an arriving message has not
1158  * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1159  * Such a message must be ignored.
1160  */
1161 void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1162 {
1163         unsigned long mmr_image;
1164         unsigned char swack_vec;
1165         struct bau_pq_entry *msg = mdp->msg;
1166         struct bau_pq_entry *other_msg;
1167
1168         mmr_image = read_mmr_sw_ack();
1169         swack_vec = msg->swack_vec;
1170
1171         if ((swack_vec & mmr_image) == 0) {
1172                 /*
1173                  * This message was assigned a swack resource, but no
1174                  * reserved acknowlegment is pending.
1175                  * The bug has prevented this message from setting the MMR.
1176                  */
1177                 /*
1178                  * Some message has set the MMR 'pending' bit; it might have
1179                  * been another message.  Look for that message.
1180                  */
1181                 other_msg = find_another_by_swack(msg, bcp);
1182                 if (other_msg) {
1183                         /*
1184                          * There is another. Process this one but do not
1185                          * ack it.
1186                          */
1187                         bau_process_message(mdp, bcp, 0);
1188                         /*
1189                          * Let the natural processing of that other message
1190                          * acknowledge it. Don't get the processing of sw_ack's
1191                          * out of order.
1192                          */
1193                         return;
1194                 }
1195         }
1196
1197         /*
1198          * Either the MMR shows this one pending a reply or there is no
1199          * other message using this sw_ack, so it is safe to acknowledge it.
1200          */
1201         bau_process_message(mdp, bcp, 1);
1202
1203         return;
1204 }
1205
1206 /*
1207  * The BAU message interrupt comes here. (registered by set_intr_gate)
1208  * See entry_64.S
1209  *
1210  * We received a broadcast assist message.
1211  *
1212  * Interrupts are disabled; this interrupt could represent
1213  * the receipt of several messages.
1214  *
1215  * All cores/threads on this hub get this interrupt.
1216  * The last one to see it does the software ack.
1217  * (the resource will not be freed until noninterruptable cpus see this
1218  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
1219  */
1220 void uv_bau_message_interrupt(struct pt_regs *regs)
1221 {
1222         int count = 0;
1223         cycles_t time_start;
1224         struct bau_pq_entry *msg;
1225         struct bau_control *bcp;
1226         struct ptc_stats *stat;
1227         struct msg_desc msgdesc;
1228
1229         ack_APIC_irq();
1230         time_start = get_cycles();
1231
1232         bcp = &per_cpu(bau_control, smp_processor_id());
1233         stat = bcp->statp;
1234
1235         msgdesc.queue_first = bcp->queue_first;
1236         msgdesc.queue_last = bcp->queue_last;
1237
1238         msg = bcp->bau_msg_head;
1239         while (msg->swack_vec) {
1240                 count++;
1241
1242                 msgdesc.msg_slot = msg - msgdesc.queue_first;
1243                 msgdesc.msg = msg;
1244                 if (bcp->uvhub_version == 2)
1245                         process_uv2_message(&msgdesc, bcp);
1246                 else
1247                         bau_process_message(&msgdesc, bcp, 1);
1248
1249                 msg++;
1250                 if (msg > msgdesc.queue_last)
1251                         msg = msgdesc.queue_first;
1252                 bcp->bau_msg_head = msg;
1253         }
1254         stat->d_time += (get_cycles() - time_start);
1255         if (!count)
1256                 stat->d_nomsg++;
1257         else if (count > 1)
1258                 stat->d_multmsg++;
1259 }
1260
1261 /*
1262  * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1263  * shootdown message timeouts enabled.  The timeout does not cause
1264  * an interrupt, but causes an error message to be returned to
1265  * the sender.
1266  */
1267 static void __init enable_timeouts(void)
1268 {
1269         int uvhub;
1270         int nuvhubs;
1271         int pnode;
1272         unsigned long mmr_image;
1273
1274         nuvhubs = uv_num_possible_blades();
1275
1276         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1277                 if (!uv_blade_nr_possible_cpus(uvhub))
1278                         continue;
1279
1280                 pnode = uv_blade_to_pnode(uvhub);
1281                 mmr_image = read_mmr_misc_control(pnode);
1282                 /*
1283                  * Set the timeout period and then lock it in, in three
1284                  * steps; captures and locks in the period.
1285                  *
1286                  * To program the period, the SOFT_ACK_MODE must be off.
1287                  */
1288                 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1289                 write_mmr_misc_control(pnode, mmr_image);
1290                 /*
1291                  * Set the 4-bit period.
1292                  */
1293                 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1294                 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1295                 write_mmr_misc_control(pnode, mmr_image);
1296                 /*
1297                  * UV1:
1298                  * Subsequent reversals of the timebase bit (3) cause an
1299                  * immediate timeout of one or all INTD resources as
1300                  * indicated in bits 2:0 (7 causes all of them to timeout).
1301                  */
1302                 mmr_image |= (1L << SOFTACK_MSHIFT);
1303                 if (is_uv2_hub()) {
1304                         /* hw bug workaround; do not use extended status */
1305                         mmr_image &= ~(1L << UV2_EXT_SHFT);
1306                 }
1307                 write_mmr_misc_control(pnode, mmr_image);
1308         }
1309 }
1310
1311 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1312 {
1313         if (*offset < num_possible_cpus())
1314                 return offset;
1315         return NULL;
1316 }
1317
1318 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1319 {
1320         (*offset)++;
1321         if (*offset < num_possible_cpus())
1322                 return offset;
1323         return NULL;
1324 }
1325
1326 static void ptc_seq_stop(struct seq_file *file, void *data)
1327 {
1328 }
1329
1330 static inline unsigned long long usec_2_cycles(unsigned long microsec)
1331 {
1332         unsigned long ns;
1333         unsigned long long cyc;
1334
1335         ns = microsec * 1000;
1336         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
1337         return cyc;
1338 }
1339
1340 /*
1341  * Display the statistics thru /proc/sgi_uv/ptc_statistics
1342  * 'data' points to the cpu number
1343  * Note: see the descriptions in stat_description[].
1344  */
1345 static int ptc_seq_show(struct seq_file *file, void *data)
1346 {
1347         struct ptc_stats *stat;
1348         struct bau_control *bcp;
1349         int cpu;
1350
1351         cpu = *(loff_t *)data;
1352         if (!cpu) {
1353                 seq_printf(file,
1354                  "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1355                 seq_printf(file,
1356                         "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1357                 seq_printf(file,
1358                         "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1359                 seq_printf(file,
1360                         "rok resetp resett giveup sto bz throt disable ");
1361                 seq_printf(file,
1362                         "enable wars warshw warwaits enters ipidis plugged ");
1363                 seq_printf(file,
1364                         "ipiover glim cong swack recv rtime all one mult ");
1365                 seq_printf(file,
1366                         "none retry canc nocan reset rcan\n");
1367         }
1368         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1369                 bcp = &per_cpu(bau_control, cpu);
1370                 stat = bcp->statp;
1371                 /* source side statistics */
1372                 seq_printf(file,
1373                         "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1374                            cpu, bcp->nobau, stat->s_requestor,
1375                            cycles_2_us(stat->s_time),
1376                            stat->s_ntargself, stat->s_ntarglocals,
1377                            stat->s_ntargremotes, stat->s_ntargcpu,
1378                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1379                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
1380                 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1381                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1382                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1383                            stat->s_dtimeout, stat->s_strongnacks);
1384                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1385                            stat->s_retry_messages, stat->s_retriesok,
1386                            stat->s_resets_plug, stat->s_resets_timeout,
1387                            stat->s_giveup, stat->s_stimeout,
1388                            stat->s_busy, stat->s_throttles);
1389                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1390                            stat->s_bau_disabled, stat->s_bau_reenabled,
1391                            stat->s_uv2_wars, stat->s_uv2_wars_hw,
1392                            stat->s_uv2_war_waits, stat->s_enters,
1393                            stat->s_ipifordisabled, stat->s_plugged,
1394                            stat->s_overipilimit, stat->s_giveuplimit,
1395                            stat->s_congested);
1396
1397                 /* destination side statistics */
1398                 seq_printf(file,
1399                         "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1400                            read_gmmr_sw_ack(uv_cpu_to_pnode(cpu)),
1401                            stat->d_requestee, cycles_2_us(stat->d_time),
1402                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1403                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
1404                            stat->d_nocanceled, stat->d_resets,
1405                            stat->d_rcanceled);
1406         }
1407         return 0;
1408 }
1409
1410 /*
1411  * Display the tunables thru debugfs
1412  */
1413 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1414                                 size_t count, loff_t *ppos)
1415 {
1416         char *buf;
1417         int ret;
1418
1419         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1420                 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1421                 "ipi_reset_limit complete_threshold congested_response_us",
1422                 "congested_reps disabled_period giveup_limit",
1423                 max_concurr, plugged_delay, plugsb4reset,
1424                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1425                 congested_respns_us, congested_reps, disabled_period,
1426                 giveup_limit);
1427
1428         if (!buf)
1429                 return -ENOMEM;
1430
1431         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1432         kfree(buf);
1433         return ret;
1434 }
1435
1436 /*
1437  * handle a write to /proc/sgi_uv/ptc_statistics
1438  * -1: reset the statistics
1439  *  0: display meaning of the statistics
1440  */
1441 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1442                                 size_t count, loff_t *data)
1443 {
1444         int cpu;
1445         int i;
1446         int elements;
1447         long input_arg;
1448         char optstr[64];
1449         struct ptc_stats *stat;
1450
1451         if (count == 0 || count > sizeof(optstr))
1452                 return -EINVAL;
1453         if (copy_from_user(optstr, user, count))
1454                 return -EFAULT;
1455         optstr[count - 1] = '\0';
1456
1457         if (!strcmp(optstr, "on")) {
1458                 set_bau_on();
1459                 return count;
1460         } else if (!strcmp(optstr, "off")) {
1461                 set_bau_off();
1462                 return count;
1463         }
1464
1465         if (strict_strtol(optstr, 10, &input_arg) < 0) {
1466                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1467                 return -EINVAL;
1468         }
1469
1470         if (input_arg == 0) {
1471                 elements = ARRAY_SIZE(stat_description);
1472                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1473                 printk(KERN_DEBUG "Sender statistics:\n");
1474                 for (i = 0; i < elements; i++)
1475                         printk(KERN_DEBUG "%s\n", stat_description[i]);
1476         } else if (input_arg == -1) {
1477                 for_each_present_cpu(cpu) {
1478                         stat = &per_cpu(ptcstats, cpu);
1479                         memset(stat, 0, sizeof(struct ptc_stats));
1480                 }
1481         }
1482
1483         return count;
1484 }
1485
1486 static int local_atoi(const char *name)
1487 {
1488         int val = 0;
1489
1490         for (;; name++) {
1491                 switch (*name) {
1492                 case '0' ... '9':
1493                         val = 10*val+(*name-'0');
1494                         break;
1495                 default:
1496                         return val;
1497                 }
1498         }
1499 }
1500
1501 /*
1502  * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1503  * Zero values reset them to defaults.
1504  */
1505 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1506                                 int count)
1507 {
1508         char *p;
1509         char *q;
1510         int cnt = 0;
1511         int val;
1512         int e = ARRAY_SIZE(tunables);
1513
1514         p = instr + strspn(instr, WHITESPACE);
1515         q = p;
1516         for (; *p; p = q + strspn(q, WHITESPACE)) {
1517                 q = p + strcspn(p, WHITESPACE);
1518                 cnt++;
1519                 if (q == p)
1520                         break;
1521         }
1522         if (cnt != e) {
1523                 printk(KERN_INFO "bau tunable error: should be %d values\n", e);
1524                 return -EINVAL;
1525         }
1526
1527         p = instr + strspn(instr, WHITESPACE);
1528         q = p;
1529         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1530                 q = p + strcspn(p, WHITESPACE);
1531                 val = local_atoi(p);
1532                 switch (cnt) {
1533                 case 0:
1534                         if (val == 0) {
1535                                 max_concurr = MAX_BAU_CONCURRENT;
1536                                 max_concurr_const = MAX_BAU_CONCURRENT;
1537                                 continue;
1538                         }
1539                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1540                                 printk(KERN_DEBUG
1541                                 "Error: BAU max concurrent %d is invalid\n",
1542                                 val);
1543                                 return -EINVAL;
1544                         }
1545                         max_concurr = val;
1546                         max_concurr_const = val;
1547                         continue;
1548                 default:
1549                         if (val == 0)
1550                                 *tunables[cnt].tunp = tunables[cnt].deflt;
1551                         else
1552                                 *tunables[cnt].tunp = val;
1553                         continue;
1554                 }
1555                 if (q == p)
1556                         break;
1557         }
1558         return 0;
1559 }
1560
1561 /*
1562  * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1563  */
1564 static ssize_t tunables_write(struct file *file, const char __user *user,
1565                                 size_t count, loff_t *data)
1566 {
1567         int cpu;
1568         int ret;
1569         char instr[100];
1570         struct bau_control *bcp;
1571
1572         if (count == 0 || count > sizeof(instr)-1)
1573                 return -EINVAL;
1574         if (copy_from_user(instr, user, count))
1575                 return -EFAULT;
1576
1577         instr[count] = '\0';
1578
1579         cpu = get_cpu();
1580         bcp = &per_cpu(bau_control, cpu);
1581         ret = parse_tunables_write(bcp, instr, count);
1582         put_cpu();
1583         if (ret)
1584                 return ret;
1585
1586         for_each_present_cpu(cpu) {
1587                 bcp = &per_cpu(bau_control, cpu);
1588                 bcp->max_concurr =              max_concurr;
1589                 bcp->max_concurr_const =        max_concurr;
1590                 bcp->plugged_delay =            plugged_delay;
1591                 bcp->plugsb4reset =             plugsb4reset;
1592                 bcp->timeoutsb4reset =          timeoutsb4reset;
1593                 bcp->ipi_reset_limit =          ipi_reset_limit;
1594                 bcp->complete_threshold =       complete_threshold;
1595                 bcp->cong_response_us =         congested_respns_us;
1596                 bcp->cong_reps =                congested_reps;
1597                 bcp->disabled_period =          sec_2_cycles(disabled_period);
1598                 bcp->giveup_limit =             giveup_limit;
1599         }
1600         return count;
1601 }
1602
1603 static const struct seq_operations uv_ptc_seq_ops = {
1604         .start          = ptc_seq_start,
1605         .next           = ptc_seq_next,
1606         .stop           = ptc_seq_stop,
1607         .show           = ptc_seq_show
1608 };
1609
1610 static int ptc_proc_open(struct inode *inode, struct file *file)
1611 {
1612         return seq_open(file, &uv_ptc_seq_ops);
1613 }
1614
1615 static int tunables_open(struct inode *inode, struct file *file)
1616 {
1617         return 0;
1618 }
1619
1620 static const struct file_operations proc_uv_ptc_operations = {
1621         .open           = ptc_proc_open,
1622         .read           = seq_read,
1623         .write          = ptc_proc_write,
1624         .llseek         = seq_lseek,
1625         .release        = seq_release,
1626 };
1627
1628 static const struct file_operations tunables_fops = {
1629         .open           = tunables_open,
1630         .read           = tunables_read,
1631         .write          = tunables_write,
1632         .llseek         = default_llseek,
1633 };
1634
1635 static int __init uv_ptc_init(void)
1636 {
1637         struct proc_dir_entry *proc_uv_ptc;
1638
1639         if (!is_uv_system())
1640                 return 0;
1641
1642         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1643                                   &proc_uv_ptc_operations);
1644         if (!proc_uv_ptc) {
1645                 printk(KERN_ERR "unable to create %s proc entry\n",
1646                        UV_PTC_BASENAME);
1647                 return -EINVAL;
1648         }
1649
1650         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1651         if (!tunables_dir) {
1652                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1653                        UV_BAU_TUNABLES_DIR);
1654                 return -EINVAL;
1655         }
1656         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1657                                         tunables_dir, NULL, &tunables_fops);
1658         if (!tunables_file) {
1659                 printk(KERN_ERR "unable to create debugfs file %s\n",
1660                        UV_BAU_TUNABLES_FILE);
1661                 return -EINVAL;
1662         }
1663         return 0;
1664 }
1665
1666 /*
1667  * Initialize the sending side's sending buffers.
1668  */
1669 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1670 {
1671         int i;
1672         int cpu;
1673         int uv1 = 0;
1674         unsigned long gpa;
1675         unsigned long m;
1676         unsigned long n;
1677         size_t dsize;
1678         struct bau_desc *bau_desc;
1679         struct bau_desc *bd2;
1680         struct uv1_bau_msg_header *uv1_hdr;
1681         struct uv2_bau_msg_header *uv2_hdr;
1682         struct bau_control *bcp;
1683
1684         /*
1685          * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1686          * per cpu; and one per cpu on the uvhub (ADP_SZ)
1687          */
1688         dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1689         bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1690         BUG_ON(!bau_desc);
1691
1692         gpa = uv_gpa(bau_desc);
1693         n = uv_gpa_to_gnode(gpa);
1694         m = uv_gpa_to_offset(gpa);
1695         if (is_uv1_hub())
1696                 uv1 = 1;
1697
1698         /* the 14-bit pnode */
1699         write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1700         /*
1701          * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1702          * cpu even though we only use the first one; one descriptor can
1703          * describe a broadcast to 256 uv hubs.
1704          */
1705         for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1706                 memset(bd2, 0, sizeof(struct bau_desc));
1707                 if (uv1) {
1708                         uv1_hdr = &bd2->header.uv1_hdr;
1709                         uv1_hdr->swack_flag =   1;
1710                         /*
1711                          * The base_dest_nasid set in the message header
1712                          * is the nasid of the first uvhub in the partition.
1713                          * The bit map will indicate destination pnode numbers
1714                          * relative to that base. They may not be consecutive
1715                          * if nasid striding is being used.
1716                          */
1717                         uv1_hdr->base_dest_nasid =
1718                                                 UV_PNODE_TO_NASID(base_pnode);
1719                         uv1_hdr->dest_subnodeid =       UV_LB_SUBNODEID;
1720                         uv1_hdr->command =              UV_NET_ENDPOINT_INTD;
1721                         uv1_hdr->int_both =             1;
1722                         /*
1723                          * all others need to be set to zero:
1724                          *   fairness chaining multilevel count replied_to
1725                          */
1726                 } else {
1727                         /*
1728                          * BIOS uses legacy mode, but UV2 hardware always
1729                          * uses native mode for selective broadcasts.
1730                          */
1731                         uv2_hdr = &bd2->header.uv2_hdr;
1732                         uv2_hdr->swack_flag =   1;
1733                         uv2_hdr->base_dest_nasid =
1734                                                 UV_PNODE_TO_NASID(base_pnode);
1735                         uv2_hdr->dest_subnodeid =       UV_LB_SUBNODEID;
1736                         uv2_hdr->command =              UV_NET_ENDPOINT_INTD;
1737                 }
1738         }
1739         for_each_present_cpu(cpu) {
1740                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1741                         continue;
1742                 bcp = &per_cpu(bau_control, cpu);
1743                 bcp->descriptor_base = bau_desc;
1744         }
1745 }
1746
1747 /*
1748  * initialize the destination side's receiving buffers
1749  * entered for each uvhub in the partition
1750  * - node is first node (kernel memory notion) on the uvhub
1751  * - pnode is the uvhub's physical identifier
1752  */
1753 static void pq_init(int node, int pnode)
1754 {
1755         int cpu;
1756         size_t plsize;
1757         char *cp;
1758         void *vp;
1759         unsigned long pn;
1760         unsigned long first;
1761         unsigned long pn_first;
1762         unsigned long last;
1763         struct bau_pq_entry *pqp;
1764         struct bau_control *bcp;
1765
1766         plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1767         vp = kmalloc_node(plsize, GFP_KERNEL, node);
1768         pqp = (struct bau_pq_entry *)vp;
1769         BUG_ON(!pqp);
1770
1771         cp = (char *)pqp + 31;
1772         pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1773
1774         for_each_present_cpu(cpu) {
1775                 if (pnode != uv_cpu_to_pnode(cpu))
1776                         continue;
1777                 /* for every cpu on this pnode: */
1778                 bcp = &per_cpu(bau_control, cpu);
1779                 bcp->queue_first        = pqp;
1780                 bcp->bau_msg_head       = pqp;
1781                 bcp->queue_last         = pqp + (DEST_Q_SIZE - 1);
1782         }
1783         /*
1784          * need the gnode of where the memory was really allocated
1785          */
1786         pn = uv_gpa_to_gnode(uv_gpa(pqp));
1787         first = uv_physnodeaddr(pqp);
1788         pn_first = ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | first;
1789         last = uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1));
1790         write_mmr_payload_first(pnode, pn_first);
1791         write_mmr_payload_tail(pnode, first);
1792         write_mmr_payload_last(pnode, last);
1793         write_gmmr_sw_ack(pnode, 0xffffUL);
1794
1795         /* in effect, all msg_type's are set to MSG_NOOP */
1796         memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1797 }
1798
1799 /*
1800  * Initialization of each UV hub's structures
1801  */
1802 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1803 {
1804         int node;
1805         int pnode;
1806         unsigned long apicid;
1807
1808         node = uvhub_to_first_node(uvhub);
1809         pnode = uv_blade_to_pnode(uvhub);
1810
1811         activation_descriptor_init(node, pnode, base_pnode);
1812
1813         pq_init(node, pnode);
1814         /*
1815          * The below initialization can't be in firmware because the
1816          * messaging IRQ will be determined by the OS.
1817          */
1818         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1819         write_mmr_data_config(pnode, ((apicid << 32) | vector));
1820 }
1821
1822 /*
1823  * We will set BAU_MISC_CONTROL with a timeout period.
1824  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1825  * So the destination timeout period has to be calculated from them.
1826  */
1827 static int calculate_destination_timeout(void)
1828 {
1829         unsigned long mmr_image;
1830         int mult1;
1831         int mult2;
1832         int index;
1833         int base;
1834         int ret;
1835         unsigned long ts_ns;
1836
1837         if (is_uv1_hub()) {
1838                 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1839                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1840                 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1841                 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1842                 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1843                 ts_ns = timeout_base_ns[index];
1844                 ts_ns *= (mult1 * mult2);
1845                 ret = ts_ns / 1000;
1846         } else {
1847                 /* 4 bits  0/1 for 10/80us base, 3 bits of multiplier */
1848                 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1849                 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1850                 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1851                         base = 80;
1852                 else
1853                         base = 10;
1854                 mult1 = mmr_image & UV2_ACK_MASK;
1855                 ret = mult1 * base;
1856         }
1857         return ret;
1858 }
1859
1860 static void __init init_per_cpu_tunables(void)
1861 {
1862         int cpu;
1863         struct bau_control *bcp;
1864
1865         for_each_present_cpu(cpu) {
1866                 bcp = &per_cpu(bau_control, cpu);
1867                 bcp->baudisabled                = 0;
1868                 if (nobau)
1869                         bcp->nobau              = 1;
1870                 bcp->statp                      = &per_cpu(ptcstats, cpu);
1871                 /* time interval to catch a hardware stay-busy bug */
1872                 bcp->timeout_interval           = usec_2_cycles(2*timeout_us);
1873                 bcp->max_concurr                = max_concurr;
1874                 bcp->max_concurr_const          = max_concurr;
1875                 bcp->plugged_delay              = plugged_delay;
1876                 bcp->plugsb4reset               = plugsb4reset;
1877                 bcp->timeoutsb4reset            = timeoutsb4reset;
1878                 bcp->ipi_reset_limit            = ipi_reset_limit;
1879                 bcp->complete_threshold         = complete_threshold;
1880                 bcp->cong_response_us           = congested_respns_us;
1881                 bcp->cong_reps                  = congested_reps;
1882                 bcp->disabled_period =          sec_2_cycles(disabled_period);
1883                 bcp->giveup_limit =             giveup_limit;
1884                 spin_lock_init(&bcp->queue_lock);
1885                 spin_lock_init(&bcp->uvhub_lock);
1886                 spin_lock_init(&bcp->disable_lock);
1887         }
1888 }
1889
1890 /*
1891  * Scan all cpus to collect blade and socket summaries.
1892  */
1893 static int __init get_cpu_topology(int base_pnode,
1894                                         struct uvhub_desc *uvhub_descs,
1895                                         unsigned char *uvhub_mask)
1896 {
1897         int cpu;
1898         int pnode;
1899         int uvhub;
1900         int socket;
1901         struct bau_control *bcp;
1902         struct uvhub_desc *bdp;
1903         struct socket_desc *sdp;
1904
1905         for_each_present_cpu(cpu) {
1906                 bcp = &per_cpu(bau_control, cpu);
1907
1908                 memset(bcp, 0, sizeof(struct bau_control));
1909
1910                 pnode = uv_cpu_hub_info(cpu)->pnode;
1911                 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1912                         printk(KERN_EMERG
1913                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1914                                 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1915                         return 1;
1916                 }
1917
1918                 bcp->osnode = cpu_to_node(cpu);
1919                 bcp->partition_base_pnode = base_pnode;
1920
1921                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1922                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1923                 bdp = &uvhub_descs[uvhub];
1924
1925                 bdp->num_cpus++;
1926                 bdp->uvhub = uvhub;
1927                 bdp->pnode = pnode;
1928
1929                 /* kludge: 'assuming' one node per socket, and assuming that
1930                    disabling a socket just leaves a gap in node numbers */
1931                 socket = bcp->osnode & 1;
1932                 bdp->socket_mask |= (1 << socket);
1933                 sdp = &bdp->socket[socket];
1934                 sdp->cpu_number[sdp->num_cpus] = cpu;
1935                 sdp->num_cpus++;
1936                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1937                         printk(KERN_EMERG "%d cpus per socket invalid\n",
1938                                 sdp->num_cpus);
1939                         return 1;
1940                 }
1941         }
1942         return 0;
1943 }
1944
1945 /*
1946  * Each socket is to get a local array of pnodes/hubs.
1947  */
1948 static void make_per_cpu_thp(struct bau_control *smaster)
1949 {
1950         int cpu;
1951         size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
1952
1953         smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
1954         memset(smaster->thp, 0, hpsz);
1955         for_each_present_cpu(cpu) {
1956                 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
1957                 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1958         }
1959 }
1960
1961 /*
1962  * Each uvhub is to get a local cpumask.
1963  */
1964 static void make_per_hub_cpumask(struct bau_control *hmaster)
1965 {
1966         int sz = sizeof(cpumask_t);
1967
1968         hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
1969 }
1970
1971 /*
1972  * Initialize all the per_cpu information for the cpu's on a given socket,
1973  * given what has been gathered into the socket_desc struct.
1974  * And reports the chosen hub and socket masters back to the caller.
1975  */
1976 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
1977                         struct bau_control **smasterp,
1978                         struct bau_control **hmasterp)
1979 {
1980         int i;
1981         int cpu;
1982         struct bau_control *bcp;
1983
1984         for (i = 0; i < sdp->num_cpus; i++) {
1985                 cpu = sdp->cpu_number[i];
1986                 bcp = &per_cpu(bau_control, cpu);
1987                 bcp->cpu = cpu;
1988                 if (i == 0) {
1989                         *smasterp = bcp;
1990                         if (!(*hmasterp))
1991                                 *hmasterp = bcp;
1992                 }
1993                 bcp->cpus_in_uvhub = bdp->num_cpus;
1994                 bcp->cpus_in_socket = sdp->num_cpus;
1995                 bcp->socket_master = *smasterp;
1996                 bcp->uvhub = bdp->uvhub;
1997                 if (is_uv1_hub())
1998                         bcp->uvhub_version = 1;
1999                 else if (is_uv2_hub())
2000                         bcp->uvhub_version = 2;
2001                 else {
2002                         printk(KERN_EMERG "uvhub version not 1 or 2\n");
2003                         return 1;
2004                 }
2005                 bcp->uvhub_master = *hmasterp;
2006                 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->blade_processor_id;
2007                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2008                         printk(KERN_EMERG "%d cpus per uvhub invalid\n",
2009                                 bcp->uvhub_cpu);
2010                         return 1;
2011                 }
2012         }
2013         return 0;
2014 }
2015
2016 /*
2017  * Summarize the blade and socket topology into the per_cpu structures.
2018  */
2019 static int __init summarize_uvhub_sockets(int nuvhubs,
2020                         struct uvhub_desc *uvhub_descs,
2021                         unsigned char *uvhub_mask)
2022 {
2023         int socket;
2024         int uvhub;
2025         unsigned short socket_mask;
2026
2027         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2028                 struct uvhub_desc *bdp;
2029                 struct bau_control *smaster = NULL;
2030                 struct bau_control *hmaster = NULL;
2031
2032                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2033                         continue;
2034
2035                 bdp = &uvhub_descs[uvhub];
2036                 socket_mask = bdp->socket_mask;
2037                 socket = 0;
2038                 while (socket_mask) {
2039                         struct socket_desc *sdp;
2040                         if ((socket_mask & 1)) {
2041                                 sdp = &bdp->socket[socket];
2042                                 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2043                                         return 1;
2044                                 make_per_cpu_thp(smaster);
2045                         }
2046                         socket++;
2047                         socket_mask = (socket_mask >> 1);
2048                 }
2049                 make_per_hub_cpumask(hmaster);
2050         }
2051         return 0;
2052 }
2053
2054 /*
2055  * initialize the bau_control structure for each cpu
2056  */
2057 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2058 {
2059         unsigned char *uvhub_mask;
2060         void *vp;
2061         struct uvhub_desc *uvhub_descs;
2062
2063         timeout_us = calculate_destination_timeout();
2064
2065         vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
2066         uvhub_descs = (struct uvhub_desc *)vp;
2067         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
2068         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2069
2070         if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2071                 goto fail;
2072
2073         if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2074                 goto fail;
2075
2076         kfree(uvhub_descs);
2077         kfree(uvhub_mask);
2078         init_per_cpu_tunables();
2079         return 0;
2080
2081 fail:
2082         kfree(uvhub_descs);
2083         kfree(uvhub_mask);
2084         return 1;
2085 }
2086
2087 /*
2088  * Initialization of BAU-related structures
2089  */
2090 static int __init uv_bau_init(void)
2091 {
2092         int uvhub;
2093         int pnode;
2094         int nuvhubs;
2095         int cur_cpu;
2096         int cpus;
2097         int vector;
2098         cpumask_var_t *mask;
2099
2100         if (!is_uv_system())
2101                 return 0;
2102
2103         for_each_possible_cpu(cur_cpu) {
2104                 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2105                 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2106         }
2107
2108         nuvhubs = uv_num_possible_blades();
2109         congested_cycles = usec_2_cycles(congested_respns_us);
2110
2111         uv_base_pnode = 0x7fffffff;
2112         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2113                 cpus = uv_blade_nr_possible_cpus(uvhub);
2114                 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2115                         uv_base_pnode = uv_blade_to_pnode(uvhub);
2116         }
2117
2118         enable_timeouts();
2119
2120         if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2121                 set_bau_off();
2122                 nobau_perm = 1;
2123                 return 0;
2124         }
2125
2126         vector = UV_BAU_MESSAGE;
2127         for_each_possible_blade(uvhub)
2128                 if (uv_blade_nr_possible_cpus(uvhub))
2129                         init_uvhub(uvhub, vector, uv_base_pnode);
2130
2131         alloc_intr_gate(vector, uv_bau_message_intr1);
2132
2133         for_each_possible_blade(uvhub) {
2134                 if (uv_blade_nr_possible_cpus(uvhub)) {
2135                         unsigned long val;
2136                         unsigned long mmr;
2137                         pnode = uv_blade_to_pnode(uvhub);
2138                         /* INIT the bau */
2139                         val = 1L << 63;
2140                         write_gmmr_activation(pnode, val);
2141                         mmr = 1; /* should be 1 to broadcast to both sockets */
2142                         if (!is_uv1_hub())
2143                                 write_mmr_data_broadcast(pnode, mmr);
2144                 }
2145         }
2146
2147         return 0;
2148 }
2149 core_initcall(uv_bau_init);
2150 fs_initcall(uv_ptc_init);