arch/x86/platform/uv/tlb_uv.c

   1 /*
   2  *      SGI UltraViolet TLB flush routines.
   3  *
   4  *      (c) 2008-2014 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 bool nobau = true;
  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 setup_bau(char *arg)
 110 {
 111         int result;
 112
 113         if (!arg)
 114                 return -EINVAL;
 115
 116         result = strtobool(arg, &nobau);
 117         if (result)
 118                 return result;
 119
 120         /* we need to flip the logic here, so that bau=y sets nobau to false */
 121         nobau = !nobau;
 122
 123         if (!nobau)
 124                 pr_info("UV BAU Enabled\n");
 125         else
 126                 pr_info("UV BAU Disabled\n");
 127
 128         return 0;
 129 }
 130 early_param("bau", setup_bau);
 131
 132 /* base pnode in this partition */
 133 static int uv_base_pnode __read_mostly;
 134
 135 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
 136 static DEFINE_PER_CPU(struct bau_control, bau_control);
 137 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
 138
 139 static void
 140 set_bau_on(void)
 141 {
 142         int cpu;
 143         struct bau_control *bcp;
 144
 145         if (nobau_perm) {
 146                 pr_info("BAU not initialized; cannot be turned on\n");
 147                 return;
 148         }
 149         nobau = false;
 150         for_each_present_cpu(cpu) {
 151                 bcp = &per_cpu(bau_control, cpu);
 152                 bcp->nobau = false;
 153         }
 154         pr_info("BAU turned on\n");
 155         return;
 156 }
 157
 158 static void
 159 set_bau_off(void)
 160 {
 161         int cpu;
 162         struct bau_control *bcp;
 163
 164         nobau = true;
 165         for_each_present_cpu(cpu) {
 166                 bcp = &per_cpu(bau_control, cpu);
 167                 bcp->nobau = true;
 168         }
 169         pr_info("BAU turned off\n");
 170         return;
 171 }
 172
 173 /*
 174  * Determine the first node on a uvhub. 'Nodes' are used for kernel
 175  * memory allocation.
 176  */
 177 static int __init uvhub_to_first_node(int uvhub)
 178 {
 179         int node, b;
 180
 181         for_each_online_node(node) {
 182                 b = uv_node_to_blade_id(node);
 183                 if (uvhub == b)
 184                         return node;
 185         }
 186         return -1;
 187 }
 188
 189 /*
 190  * Determine the apicid of the first cpu on a uvhub.
 191  */
 192 static int __init uvhub_to_first_apicid(int uvhub)
 193 {
 194         int cpu;
 195
 196         for_each_present_cpu(cpu)
 197                 if (uvhub == uv_cpu_to_blade_id(cpu))
 198                         return per_cpu(x86_cpu_to_apicid, cpu);
 199         return -1;
 200 }
 201
 202 /*
 203  * Free a software acknowledge hardware resource by clearing its Pending
 204  * bit. This will return a reply to the sender.
 205  * If the message has timed out, a reply has already been sent by the
 206  * hardware but the resource has not been released. In that case our
 207  * clear of the Timeout bit (as well) will free the resource. No reply will
 208  * be sent (the hardware will only do one reply per message).
 209  */
 210 static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp,
 211                                                 int do_acknowledge)
 212 {
 213         unsigned long dw;
 214         struct bau_pq_entry *msg;
 215
 216         msg = mdp->msg;
 217         if (!msg->canceled && do_acknowledge) {
 218                 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
 219                 write_mmr_sw_ack(dw);
 220         }
 221         msg->replied_to = 1;
 222         msg->swack_vec = 0;
 223 }
 224
 225 /*
 226  * Process the receipt of a RETRY message
 227  */
 228 static void bau_process_retry_msg(struct msg_desc *mdp,
 229                                         struct bau_control *bcp)
 230 {
 231         int i;
 232         int cancel_count = 0;
 233         unsigned long msg_res;
 234         unsigned long mmr = 0;
 235         struct bau_pq_entry *msg = mdp->msg;
 236         struct bau_pq_entry *msg2;
 237         struct ptc_stats *stat = bcp->statp;
 238
 239         stat->d_retries++;
 240         /*
 241          * cancel any message from msg+1 to the retry itself
 242          */
 243         for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
 244                 if (msg2 > mdp->queue_last)
 245                         msg2 = mdp->queue_first;
 246                 if (msg2 == msg)
 247                         break;
 248
 249                 /* same conditions for cancellation as do_reset */
 250                 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
 251                     (msg2->swack_vec) && ((msg2->swack_vec &
 252                         msg->swack_vec) == 0) &&
 253                     (msg2->sending_cpu == msg->sending_cpu) &&
 254                     (msg2->msg_type != MSG_NOOP)) {
 255                         mmr = read_mmr_sw_ack();
 256                         msg_res = msg2->swack_vec;
 257                         /*
 258                          * This is a message retry; clear the resources held
 259                          * by the previous message only if they timed out.
 260                          * If it has not timed out we have an unexpected
 261                          * situation to report.
 262                          */
 263                         if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
 264                                 unsigned long mr;
 265                                 /*
 266                                  * Is the resource timed out?
 267                                  * Make everyone ignore the cancelled message.
 268                                  */
 269                                 msg2->canceled = 1;
 270                                 stat->d_canceled++;
 271                                 cancel_count++;
 272                                 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
 273                                 write_mmr_sw_ack(mr);
 274                         }
 275                 }
 276         }
 277         if (!cancel_count)
 278                 stat->d_nocanceled++;
 279 }
 280
 281 /*
 282  * Do all the things a cpu should do for a TLB shootdown message.
 283  * Other cpu's may come here at the same time for this message.
 284  */
 285 static void bau_process_message(struct msg_desc *mdp, struct bau_control *bcp,
 286                                                 int do_acknowledge)
 287 {
 288         short socket_ack_count = 0;
 289         short *sp;
 290         struct atomic_short *asp;
 291         struct ptc_stats *stat = bcp->statp;
 292         struct bau_pq_entry *msg = mdp->msg;
 293         struct bau_control *smaster = bcp->socket_master;
 294
 295         /*
 296          * This must be a normal message, or retry of a normal message
 297          */
 298         if (msg->address == TLB_FLUSH_ALL) {
 299                 local_flush_tlb();
 300                 stat->d_alltlb++;
 301         } else {
 302                 __flush_tlb_one(msg->address);
 303                 stat->d_onetlb++;
 304         }
 305         stat->d_requestee++;
 306
 307         /*
 308          * One cpu on each uvhub has the additional job on a RETRY
 309          * of releasing the resource held by the message that is
 310          * being retried.  That message is identified by sending
 311          * cpu number.
 312          */
 313         if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
 314                 bau_process_retry_msg(mdp, bcp);
 315
 316         /*
 317          * This is a swack message, so we have to reply to it.
 318          * Count each responding cpu on the socket. This avoids
 319          * pinging the count's cache line back and forth between
 320          * the sockets.
 321          */
 322         sp = &smaster->socket_acknowledge_count[mdp->msg_slot];
 323         asp = (struct atomic_short *)sp;
 324         socket_ack_count = atom_asr(1, asp);
 325         if (socket_ack_count == bcp->cpus_in_socket) {
 326                 int msg_ack_count;
 327                 /*
 328                  * Both sockets dump their completed count total into
 329                  * the message's count.
 330                  */
 331                 *sp = 0;
 332                 asp = (struct atomic_short *)&msg->acknowledge_count;
 333                 msg_ack_count = atom_asr(socket_ack_count, asp);
 334
 335                 if (msg_ack_count == bcp->cpus_in_uvhub) {
 336                         /*
 337                          * All cpus in uvhub saw it; reply
 338                          * (unless we are in the UV2 workaround)
 339                          */
 340                         reply_to_message(mdp, bcp, do_acknowledge);
 341                 }
 342         }
 343
 344         return;
 345 }
 346
 347 /*
 348  * Determine the first cpu on a pnode.
 349  */
 350 static int pnode_to_first_cpu(int pnode, struct bau_control *smaster)
 351 {
 352         int cpu;
 353         struct hub_and_pnode *hpp;
 354
 355         for_each_present_cpu(cpu) {
 356                 hpp = &smaster->thp[cpu];
 357                 if (pnode == hpp->pnode)
 358                         return cpu;
 359         }
 360         return -1;
 361 }
 362
 363 /*
 364  * Last resort when we get a large number of destination timeouts is
 365  * to clear resources held by a given cpu.
 366  * Do this with IPI so that all messages in the BAU message queue
 367  * can be identified by their nonzero swack_vec field.
 368  *
 369  * This is entered for a single cpu on the uvhub.
 370  * The sender want's this uvhub to free a specific message's
 371  * swack resources.
 372  */
 373 static void do_reset(void *ptr)
 374 {
 375         int i;
 376         struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id());
 377         struct reset_args *rap = (struct reset_args *)ptr;
 378         struct bau_pq_entry *msg;
 379         struct ptc_stats *stat = bcp->statp;
 380
 381         stat->d_resets++;
 382         /*
 383          * We're looking for the given sender, and
 384          * will free its swack resource.
 385          * If all cpu's finally responded after the timeout, its
 386          * message 'replied_to' was set.
 387          */
 388         for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
 389                 unsigned long msg_res;
 390                 /* do_reset: same conditions for cancellation as
 391                    bau_process_retry_msg() */
 392                 if ((msg->replied_to == 0) &&
 393                     (msg->canceled == 0) &&
 394                     (msg->sending_cpu == rap->sender) &&
 395                     (msg->swack_vec) &&
 396                     (msg->msg_type != MSG_NOOP)) {
 397                         unsigned long mmr;
 398                         unsigned long mr;
 399                         /*
 400                          * make everyone else ignore this message
 401                          */
 402                         msg->canceled = 1;
 403                         /*
 404                          * only reset the resource if it is still pending
 405                          */
 406                         mmr = read_mmr_sw_ack();
 407                         msg_res = msg->swack_vec;
 408                         mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
 409                         if (mmr & msg_res) {
 410                                 stat->d_rcanceled++;
 411                                 write_mmr_sw_ack(mr);
 412                         }
 413                 }
 414         }
 415         return;
 416 }
 417
 418 /*
 419  * Use IPI to get all target uvhubs to release resources held by
 420  * a given sending cpu number.
 421  */
 422 static void reset_with_ipi(struct pnmask *distribution, struct bau_control *bcp)
 423 {
 424         int pnode;
 425         int apnode;
 426         int maskbits;
 427         int sender = bcp->cpu;
 428         cpumask_t *mask = bcp->uvhub_master->cpumask;
 429         struct bau_control *smaster = bcp->socket_master;
 430         struct reset_args reset_args;
 431
 432         reset_args.sender = sender;
 433         cpumask_clear(mask);
 434         /* find a single cpu for each uvhub in this distribution mask */
 435         maskbits = sizeof(struct pnmask) * BITSPERBYTE;
 436         /* each bit is a pnode relative to the partition base pnode */
 437         for (pnode = 0; pnode < maskbits; pnode++) {
 438                 int cpu;
 439                 if (!bau_uvhub_isset(pnode, distribution))
 440                         continue;
 441                 apnode = pnode + bcp->partition_base_pnode;
 442                 cpu = pnode_to_first_cpu(apnode, smaster);
 443                 cpumask_set_cpu(cpu, mask);
 444         }
 445
 446         /* IPI all cpus; preemption is already disabled */
 447         smp_call_function_many(mask, do_reset, (void *)&reset_args, 1);
 448         return;
 449 }
 450
 451 /*
 452  * Not to be confused with cycles_2_ns() from tsc.c; this gives a relative
 453  * number, not an absolute. It converts a duration in cycles to a duration in
 454  * ns.
 455  */
 456 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
 457 {
 458         struct cyc2ns_data *data = cyc2ns_read_begin();
 459         unsigned long long ns;
 460
 461         ns = mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift);
 462
 463         cyc2ns_read_end(data);
 464         return ns;
 465 }
 466
 467 /*
 468  * The reverse of the above; converts a duration in ns to a duration in cycles.
 469  */
 470 static inline unsigned long long ns_2_cycles(unsigned long long ns)
 471 {
 472         struct cyc2ns_data *data = cyc2ns_read_begin();
 473         unsigned long long cyc;
 474
 475         cyc = (ns << data->cyc2ns_shift) / data->cyc2ns_mul;
 476
 477         cyc2ns_read_end(data);
 478         return cyc;
 479 }
 480
 481 static inline unsigned long cycles_2_us(unsigned long long cyc)
 482 {
 483         return cycles_2_ns(cyc) / NSEC_PER_USEC;
 484 }
 485
 486 static inline cycles_t sec_2_cycles(unsigned long sec)
 487 {
 488         return ns_2_cycles(sec * NSEC_PER_SEC);
 489 }
 490
 491 static inline unsigned long long usec_2_cycles(unsigned long usec)
 492 {
 493         return ns_2_cycles(usec * NSEC_PER_USEC);
 494 }
 495
 496 /*
 497  * wait for all cpus on this hub to finish their sends and go quiet
 498  * leaves uvhub_quiesce set so that no new broadcasts are started by
 499  * bau_flush_send_and_wait()
 500  */
 501 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
 502 {
 503         atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 504 }
 505
 506 /*
 507  * mark this quiet-requestor as done
 508  */
 509 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
 510 {
 511         atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 512 }
 513
 514 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
 515 {
 516         unsigned long descriptor_status;
 517
 518         descriptor_status = uv_read_local_mmr(mmr_offset);
 519         descriptor_status >>= right_shift;
 520         descriptor_status &= UV_ACT_STATUS_MASK;
 521         return descriptor_status;
 522 }
 523
 524 /*
 525  * Wait for completion of a broadcast software ack message
 526  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
 527  */
 528 static int uv1_wait_completion(struct bau_desc *bau_desc,
 529                                 unsigned long mmr_offset, int right_shift,
 530                                 struct bau_control *bcp, long try)
 531 {
 532         unsigned long descriptor_status;
 533         cycles_t ttm;
 534         struct ptc_stats *stat = bcp->statp;
 535
 536         descriptor_status = uv1_read_status(mmr_offset, right_shift);
 537         /* spin on the status MMR, waiting for it to go idle */
 538         while ((descriptor_status != DS_IDLE)) {
 539                 /*
 540                  * Our software ack messages may be blocked because
 541                  * there are no swack resources available.  As long
 542                  * as none of them has timed out hardware will NACK
 543                  * our message and its state will stay IDLE.
 544                  */
 545                 if (descriptor_status == DS_SOURCE_TIMEOUT) {
 546                         stat->s_stimeout++;
 547                         return FLUSH_GIVEUP;
 548                 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
 549                         stat->s_dtimeout++;
 550                         ttm = get_cycles();
 551
 552                         /*
 553                          * Our retries may be blocked by all destination
 554                          * swack resources being consumed, and a timeout
 555                          * pending.  In that case hardware returns the
 556                          * ERROR that looks like a destination timeout.
 557                          */
 558                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 559                                 bcp->conseccompletes = 0;
 560                                 return FLUSH_RETRY_PLUGGED;
 561                         }
 562
 563                         bcp->conseccompletes = 0;
 564                         return FLUSH_RETRY_TIMEOUT;
 565                 } else {
 566                         /*
 567                          * descriptor_status is still BUSY
 568                          */
 569                         cpu_relax();
 570                 }
 571                 descriptor_status = uv1_read_status(mmr_offset, right_shift);
 572         }
 573         bcp->conseccompletes++;
 574         return FLUSH_COMPLETE;
 575 }
 576
 577 /*
 578  * UV2 could have an extra bit of status in the ACTIVATION_STATUS_2 register.
 579  * But not currently used.
 580  */
 581 static unsigned long uv2_3_read_status(unsigned long offset, int rshft, int desc)
 582 {
 583         unsigned long descriptor_status;
 584
 585         descriptor_status =
 586                 ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK) << 1;
 587         return descriptor_status;
 588 }
 589
 590 /*
 591  * Return whether the status of the descriptor that is normally used for this
 592  * cpu (the one indexed by its hub-relative cpu number) is busy.
 593  * The status of the original 32 descriptors is always reflected in the 64
 594  * bits of UVH_LB_BAU_SB_ACTIVATION_STATUS_0.
 595  * The bit provided by the activation_status_2 register is irrelevant to
 596  * the status if it is only being tested for busy or not busy.
 597  */
 598 int normal_busy(struct bau_control *bcp)
 599 {
 600         int cpu = bcp->uvhub_cpu;
 601         int mmr_offset;
 602         int right_shift;
 603
 604         mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 605         right_shift = cpu * UV_ACT_STATUS_SIZE;
 606         return (((((read_lmmr(mmr_offset) >> right_shift) &
 607                                 UV_ACT_STATUS_MASK)) << 1) == UV2H_DESC_BUSY);
 608 }
 609
 610 /*
 611  * Entered when a bau descriptor has gone into a permanent busy wait because
 612  * of a hardware bug.
 613  * Workaround the bug.
 614  */
 615 int handle_uv2_busy(struct bau_control *bcp)
 616 {
 617         struct ptc_stats *stat = bcp->statp;
 618
 619         stat->s_uv2_wars++;
 620         bcp->busy = 1;
 621         return FLUSH_GIVEUP;
 622 }
 623
 624 static int uv2_3_wait_completion(struct bau_desc *bau_desc,
 625                                 unsigned long mmr_offset, int right_shift,
 626                                 struct bau_control *bcp, long try)
 627 {
 628         unsigned long descriptor_stat;
 629         cycles_t ttm;
 630         int desc = bcp->uvhub_cpu;
 631         long busy_reps = 0;
 632         struct ptc_stats *stat = bcp->statp;
 633
 634         descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
 635
 636         /* spin on the status MMR, waiting for it to go idle */
 637         while (descriptor_stat != UV2H_DESC_IDLE) {
 638                 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT)) {
 639                         /*
 640                          * A h/w bug on the destination side may
 641                          * have prevented the message being marked
 642                          * pending, thus it doesn't get replied to
 643                          * and gets continually nacked until it times
 644                          * out with a SOURCE_TIMEOUT.
 645                          */
 646                         stat->s_stimeout++;
 647                         return FLUSH_GIVEUP;
 648                 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
 649                         ttm = get_cycles();
 650
 651                         /*
 652                          * Our retries may be blocked by all destination
 653                          * swack resources being consumed, and a timeout
 654                          * pending.  In that case hardware returns the
 655                          * ERROR that looks like a destination timeout.
 656                          * Without using the extended status we have to
 657                          * deduce from the short time that this was a
 658                          * strong nack.
 659                          */
 660                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 661                                 bcp->conseccompletes = 0;
 662                                 stat->s_plugged++;
 663                                 /* FLUSH_RETRY_PLUGGED causes hang on boot */
 664                                 return FLUSH_GIVEUP;
 665                         }
 666                         stat->s_dtimeout++;
 667                         bcp->conseccompletes = 0;
 668                         /* FLUSH_RETRY_TIMEOUT causes hang on boot */
 669                         return FLUSH_GIVEUP;
 670                 } else {
 671                         busy_reps++;
 672                         if (busy_reps > 1000000) {
 673                                 /* not to hammer on the clock */
 674                                 busy_reps = 0;
 675                                 ttm = get_cycles();
 676                                 if ((ttm - bcp->send_message) > bcp->timeout_interval)
 677                                         return handle_uv2_busy(bcp);
 678                         }
 679                         /*
 680                          * descriptor_stat is still BUSY
 681                          */
 682                         cpu_relax();
 683                 }
 684                 descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
 685         }
 686         bcp->conseccompletes++;
 687         return FLUSH_COMPLETE;
 688 }
 689
 690 /*
 691  * There are 2 status registers; each and array[32] of 2 bits. Set up for
 692  * which register to read and position in that register based on cpu in
 693  * current hub.
 694  */
 695 static int wait_completion(struct bau_desc *bau_desc, struct bau_control *bcp, long try)
 696 {
 697         int right_shift;
 698         unsigned long mmr_offset;
 699         int desc = bcp->uvhub_cpu;
 700
 701         if (desc < UV_CPUS_PER_AS) {
 702                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 703                 right_shift = desc * UV_ACT_STATUS_SIZE;
 704         } else {
 705                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
 706                 right_shift = ((desc - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
 707         }
 708
 709         if (bcp->uvhub_version == 1)
 710                 return uv1_wait_completion(bau_desc, mmr_offset, right_shift, bcp, try);
 711         else
 712                 return uv2_3_wait_completion(bau_desc, mmr_offset, right_shift, bcp, try);
 713 }
 714
 715 /*
 716  * Our retries are blocked by all destination sw ack resources being
 717  * in use, and a timeout is pending. In that case hardware immediately
 718  * returns the ERROR that looks like a destination timeout.
 719  */
 720 static void destination_plugged(struct bau_desc *bau_desc,
 721                         struct bau_control *bcp,
 722                         struct bau_control *hmaster, struct ptc_stats *stat)
 723 {
 724         udelay(bcp->plugged_delay);
 725         bcp->plugged_tries++;
 726
 727         if (bcp->plugged_tries >= bcp->plugsb4reset) {
 728                 bcp->plugged_tries = 0;
 729
 730                 quiesce_local_uvhub(hmaster);
 731
 732                 spin_lock(&hmaster->queue_lock);
 733                 reset_with_ipi(&bau_desc->distribution, bcp);
 734                 spin_unlock(&hmaster->queue_lock);
 735
 736                 end_uvhub_quiesce(hmaster);
 737
 738                 bcp->ipi_attempts++;
 739                 stat->s_resets_plug++;
 740         }
 741 }
 742
 743 static void destination_timeout(struct bau_desc *bau_desc,
 744                         struct bau_control *bcp, struct bau_control *hmaster,
 745                         struct ptc_stats *stat)
 746 {
 747         hmaster->max_concurr = 1;
 748         bcp->timeout_tries++;
 749         if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
 750                 bcp->timeout_tries = 0;
 751
 752                 quiesce_local_uvhub(hmaster);
 753
 754                 spin_lock(&hmaster->queue_lock);
 755                 reset_with_ipi(&bau_desc->distribution, bcp);
 756                 spin_unlock(&hmaster->queue_lock);
 757
 758                 end_uvhub_quiesce(hmaster);
 759
 760                 bcp->ipi_attempts++;
 761                 stat->s_resets_timeout++;
 762         }
 763 }
 764
 765 /*
 766  * Stop all cpus on a uvhub from using the BAU for a period of time.
 767  * This is reversed by check_enable.
 768  */
 769 static void disable_for_period(struct bau_control *bcp, struct ptc_stats *stat)
 770 {
 771         int tcpu;
 772         struct bau_control *tbcp;
 773         struct bau_control *hmaster;
 774         cycles_t tm1;
 775
 776         hmaster = bcp->uvhub_master;
 777         spin_lock(&hmaster->disable_lock);
 778         if (!bcp->baudisabled) {
 779                 stat->s_bau_disabled++;
 780                 tm1 = get_cycles();
 781                 for_each_present_cpu(tcpu) {
 782                         tbcp = &per_cpu(bau_control, tcpu);
 783                         if (tbcp->uvhub_master == hmaster) {
 784                                 tbcp->baudisabled = 1;
 785                                 tbcp->set_bau_on_time =
 786                                         tm1 + bcp->disabled_period;
 787                         }
 788                 }
 789         }
 790         spin_unlock(&hmaster->disable_lock);
 791 }
 792
 793 static void count_max_concurr(int stat, struct bau_control *bcp,
 794                                 struct bau_control *hmaster)
 795 {
 796         bcp->plugged_tries = 0;
 797         bcp->timeout_tries = 0;
 798         if (stat != FLUSH_COMPLETE)
 799                 return;
 800         if (bcp->conseccompletes <= bcp->complete_threshold)
 801                 return;
 802         if (hmaster->max_concurr >= hmaster->max_concurr_const)
 803                 return;
 804         hmaster->max_concurr++;
 805 }
 806
 807 static void record_send_stats(cycles_t time1, cycles_t time2,
 808                 struct bau_control *bcp, struct ptc_stats *stat,
 809                 int completion_status, int try)
 810 {
 811         cycles_t elapsed;
 812
 813         if (time2 > time1) {
 814                 elapsed = time2 - time1;
 815                 stat->s_time += elapsed;
 816
 817                 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
 818                         bcp->period_requests++;
 819                         bcp->period_time += elapsed;
 820                         if ((elapsed > congested_cycles) &&
 821                             (bcp->period_requests > bcp->cong_reps) &&
 822                             ((bcp->period_time / bcp->period_requests) >
 823                                                         congested_cycles)) {
 824                                 stat->s_congested++;
 825                                 disable_for_period(bcp, stat);
 826                         }
 827                 }
 828         } else
 829                 stat->s_requestor--;
 830
 831         if (completion_status == FLUSH_COMPLETE && try > 1)
 832                 stat->s_retriesok++;
 833         else if (completion_status == FLUSH_GIVEUP) {
 834                 stat->s_giveup++;
 835                 if (get_cycles() > bcp->period_end)
 836                         bcp->period_giveups = 0;
 837                 bcp->period_giveups++;
 838                 if (bcp->period_giveups == 1)
 839                         bcp->period_end = get_cycles() + bcp->disabled_period;
 840                 if (bcp->period_giveups > bcp->giveup_limit) {
 841                         disable_for_period(bcp, stat);
 842                         stat->s_giveuplimit++;
 843                 }
 844         }
 845 }
 846
 847 /*
 848  * Because of a uv1 hardware bug only a limited number of concurrent
 849  * requests can be made.
 850  */
 851 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
 852 {
 853         spinlock_t *lock = &hmaster->uvhub_lock;
 854         atomic_t *v;
 855
 856         v = &hmaster->active_descriptor_count;
 857         if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
 858                 stat->s_throttles++;
 859                 do {
 860                         cpu_relax();
 861                 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
 862         }
 863 }
 864
 865 /*
 866  * Handle the completion status of a message send.
 867  */
 868 static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
 869                         struct bau_control *bcp, struct bau_control *hmaster,
 870                         struct ptc_stats *stat)
 871 {
 872         if (completion_status == FLUSH_RETRY_PLUGGED)
 873                 destination_plugged(bau_desc, bcp, hmaster, stat);
 874         else if (completion_status == FLUSH_RETRY_TIMEOUT)
 875                 destination_timeout(bau_desc, bcp, hmaster, stat);
 876 }
 877
 878 /*
 879  * Send a broadcast and wait for it to complete.
 880  *
 881  * The flush_mask contains the cpus the broadcast is to be sent to including
 882  * cpus that are on the local uvhub.
 883  *
 884  * Returns 0 if all flushing represented in the mask was done.
 885  * Returns 1 if it gives up entirely and the original cpu mask is to be
 886  * returned to the kernel.
 887  */
 888 int uv_flush_send_and_wait(struct cpumask *flush_mask, struct bau_control *bcp,
 889         struct bau_desc *bau_desc)
 890 {
 891         int seq_number = 0;
 892         int completion_stat = 0;
 893         int uv1 = 0;
 894         long try = 0;
 895         unsigned long index;
 896         cycles_t time1;
 897         cycles_t time2;
 898         struct ptc_stats *stat = bcp->statp;
 899         struct bau_control *hmaster = bcp->uvhub_master;
 900         struct uv1_bau_msg_header *uv1_hdr = NULL;
 901         struct uv2_3_bau_msg_header *uv2_3_hdr = NULL;
 902
 903         if (bcp->uvhub_version == 1) {
 904                 uv1 = 1;
 905                 uv1_throttle(hmaster, stat);
 906         }
 907
 908         while (hmaster->uvhub_quiesce)
 909                 cpu_relax();
 910
 911         time1 = get_cycles();
 912         if (uv1)
 913                 uv1_hdr = &bau_desc->header.uv1_hdr;
 914         else
 915                 /* uv2 and uv3 */
 916                 uv2_3_hdr = &bau_desc->header.uv2_3_hdr;
 917
 918         do {
 919                 if (try == 0) {
 920                         if (uv1)
 921                                 uv1_hdr->msg_type = MSG_REGULAR;
 922                         else
 923                                 uv2_3_hdr->msg_type = MSG_REGULAR;
 924                         seq_number = bcp->message_number++;
 925                 } else {
 926                         if (uv1)
 927                                 uv1_hdr->msg_type = MSG_RETRY;
 928                         else
 929                                 uv2_3_hdr->msg_type = MSG_RETRY;
 930                         stat->s_retry_messages++;
 931                 }
 932
 933                 if (uv1)
 934                         uv1_hdr->sequence = seq_number;
 935                 else
 936                         uv2_3_hdr->sequence = seq_number;
 937                 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
 938                 bcp->send_message = get_cycles();
 939
 940                 write_mmr_activation(index);
 941
 942                 try++;
 943                 completion_stat = wait_completion(bau_desc, bcp, try);
 944
 945                 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
 946
 947                 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
 948                         bcp->ipi_attempts = 0;
 949                         stat->s_overipilimit++;
 950                         completion_stat = FLUSH_GIVEUP;
 951                         break;
 952                 }
 953                 cpu_relax();
 954         } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
 955                  (completion_stat == FLUSH_RETRY_TIMEOUT));
 956
 957         time2 = get_cycles();
 958
 959         count_max_concurr(completion_stat, bcp, hmaster);
 960
 961         while (hmaster->uvhub_quiesce)
 962                 cpu_relax();
 963
 964         atomic_dec(&hmaster->active_descriptor_count);
 965
 966         record_send_stats(time1, time2, bcp, stat, completion_stat, try);
 967
 968         if (completion_stat == FLUSH_GIVEUP)
 969                 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
 970                 return 1;
 971         return 0;
 972 }
 973
 974 /*
 975  * The BAU is disabled for this uvhub. When the disabled time period has
 976  * expired re-enable it.
 977  * Return 0 if it is re-enabled for all cpus on this uvhub.
 978  */
 979 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
 980 {
 981         int tcpu;
 982         struct bau_control *tbcp;
 983         struct bau_control *hmaster;
 984
 985         hmaster = bcp->uvhub_master;
 986         spin_lock(&hmaster->disable_lock);
 987         if (bcp->baudisabled && (get_cycles() >= bcp->set_bau_on_time)) {
 988                 stat->s_bau_reenabled++;
 989                 for_each_present_cpu(tcpu) {
 990                         tbcp = &per_cpu(bau_control, tcpu);
 991                         if (tbcp->uvhub_master == hmaster) {
 992                                 tbcp->baudisabled = 0;
 993                                 tbcp->period_requests = 0;
 994                                 tbcp->period_time = 0;
 995                                 tbcp->period_giveups = 0;
 996                         }
 997                 }
 998                 spin_unlock(&hmaster->disable_lock);
 999                 return 0;
1000         }
1001         spin_unlock(&hmaster->disable_lock);
1002         return -1;
1003 }
1004
1005 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
1006                                 int remotes, struct bau_desc *bau_desc)
1007 {
1008         stat->s_requestor++;
1009         stat->s_ntargcpu += remotes + locals;
1010         stat->s_ntargremotes += remotes;
1011         stat->s_ntarglocals += locals;
1012
1013         /* uvhub statistics */
1014         hubs = bau_uvhub_weight(&bau_desc->distribution);
1015         if (locals) {
1016                 stat->s_ntarglocaluvhub++;
1017                 stat->s_ntargremoteuvhub += (hubs - 1);
1018         } else
1019                 stat->s_ntargremoteuvhub += hubs;
1020
1021         stat->s_ntarguvhub += hubs;
1022
1023         if (hubs >= 16)
1024                 stat->s_ntarguvhub16++;
1025         else if (hubs >= 8)
1026                 stat->s_ntarguvhub8++;
1027         else if (hubs >= 4)
1028                 stat->s_ntarguvhub4++;
1029         else if (hubs >= 2)
1030                 stat->s_ntarguvhub2++;
1031         else
1032                 stat->s_ntarguvhub1++;
1033 }
1034
1035 /*
1036  * Translate a cpu mask to the uvhub distribution mask in the BAU
1037  * activation descriptor.
1038  */
1039 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
1040                         struct bau_desc *bau_desc, int *localsp, int *remotesp)
1041 {
1042         int cpu;
1043         int pnode;
1044         int cnt = 0;
1045         struct hub_and_pnode *hpp;
1046
1047         for_each_cpu(cpu, flush_mask) {
1048                 /*
1049                  * The distribution vector is a bit map of pnodes, relative
1050                  * to the partition base pnode (and the partition base nasid
1051                  * in the header).
1052                  * Translate cpu to pnode and hub using a local memory array.
1053                  */
1054                 hpp = &bcp->socket_master->thp[cpu];
1055                 pnode = hpp->pnode - bcp->partition_base_pnode;
1056                 bau_uvhub_set(pnode, &bau_desc->distribution);
1057                 cnt++;
1058                 if (hpp->uvhub == bcp->uvhub)
1059                         (*localsp)++;
1060                 else
1061                         (*remotesp)++;
1062         }
1063         if (!cnt)
1064                 return 1;
1065         return 0;
1066 }
1067
1068 /*
1069  * globally purge translation cache of a virtual address or all TLB's
1070  * @cpumask: mask of all cpu's in which the address is to be removed
1071  * @mm: mm_struct containing virtual address range
1072  * @start: start virtual address to be removed from TLB
1073  * @end: end virtual address to be remove from TLB
1074  * @cpu: the current cpu
1075  *
1076  * This is the entry point for initiating any UV global TLB shootdown.
1077  *
1078  * Purges the translation caches of all specified processors of the given
1079  * virtual address, or purges all TLB's on specified processors.
1080  *
1081  * The caller has derived the cpumask from the mm_struct.  This function
1082  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1083  *
1084  * The cpumask is converted into a uvhubmask of the uvhubs containing
1085  * those cpus.
1086  *
1087  * Note that this function should be called with preemption disabled.
1088  *
1089  * Returns NULL if all remote flushing was done.
1090  * Returns pointer to cpumask if some remote flushing remains to be
1091  * done.  The returned pointer is valid till preemption is re-enabled.
1092  */
1093 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1094                                                 struct mm_struct *mm,
1095                                                 unsigned long start,
1096                                                 unsigned long end,
1097                                                 unsigned int cpu)
1098 {
1099         int locals = 0;
1100         int remotes = 0;
1101         int hubs = 0;
1102         struct bau_desc *bau_desc;
1103         struct cpumask *flush_mask;
1104         struct ptc_stats *stat;
1105         struct bau_control *bcp;
1106         unsigned long descriptor_status;
1107         unsigned long status;
1108
1109         bcp = &per_cpu(bau_control, cpu);
1110
1111         if (bcp->nobau)
1112                 return cpumask;
1113
1114         stat = bcp->statp;
1115         stat->s_enters++;
1116
1117         if (bcp->busy) {
1118                 descriptor_status =
1119                         read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1120                 status = ((descriptor_status >> (bcp->uvhub_cpu *
1121                         UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1122                 if (status == UV2H_DESC_BUSY)
1123                         return cpumask;
1124                 bcp->busy = 0;
1125         }
1126
1127         /* bau was disabled due to slow response */
1128         if (bcp->baudisabled) {
1129                 if (check_enable(bcp, stat)) {
1130                         stat->s_ipifordisabled++;
1131                         return cpumask;
1132                 }
1133         }
1134
1135         /*
1136          * Each sending cpu has a per-cpu mask which it fills from the caller's
1137          * cpu mask.  All cpus are converted to uvhubs and copied to the
1138          * activation descriptor.
1139          */
1140         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1141         /* don't actually do a shootdown of the local cpu */
1142         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1143
1144         if (cpumask_test_cpu(cpu, cpumask))
1145                 stat->s_ntargself++;
1146
1147         bau_desc = bcp->descriptor_base;
1148         bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1149         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1150         if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1151                 return NULL;
1152
1153         record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1154
1155         if (!end || (end - start) <= PAGE_SIZE)
1156                 bau_desc->payload.address = start;
1157         else
1158                 bau_desc->payload.address = TLB_FLUSH_ALL;
1159         bau_desc->payload.sending_cpu = cpu;
1160         /*
1161          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1162          * or 1 if it gave up and the original cpumask should be returned.
1163          */
1164         if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1165                 return NULL;
1166         else
1167                 return cpumask;
1168 }
1169
1170 /*
1171  * Search the message queue for any 'other' unprocessed message with the
1172  * same software acknowledge resource bit vector as the 'msg' message.
1173  */
1174 struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1175                                            struct bau_control *bcp)
1176 {
1177         struct bau_pq_entry *msg_next = msg + 1;
1178         unsigned char swack_vec = msg->swack_vec;
1179
1180         if (msg_next > bcp->queue_last)
1181                 msg_next = bcp->queue_first;
1182         while (msg_next != msg) {
1183                 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1184                                 (msg_next->swack_vec == swack_vec))
1185                         return msg_next;
1186                 msg_next++;
1187                 if (msg_next > bcp->queue_last)
1188                         msg_next = bcp->queue_first;
1189         }
1190         return NULL;
1191 }
1192
1193 /*
1194  * UV2 needs to work around a bug in which an arriving message has not
1195  * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1196  * Such a message must be ignored.
1197  */
1198 void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1199 {
1200         unsigned long mmr_image;
1201         unsigned char swack_vec;
1202         struct bau_pq_entry *msg = mdp->msg;
1203         struct bau_pq_entry *other_msg;
1204
1205         mmr_image = read_mmr_sw_ack();
1206         swack_vec = msg->swack_vec;
1207
1208         if ((swack_vec & mmr_image) == 0) {
1209                 /*
1210                  * This message was assigned a swack resource, but no
1211                  * reserved acknowlegment is pending.
1212                  * The bug has prevented this message from setting the MMR.
1213                  */
1214                 /*
1215                  * Some message has set the MMR 'pending' bit; it might have
1216                  * been another message.  Look for that message.
1217                  */
1218                 other_msg = find_another_by_swack(msg, bcp);
1219                 if (other_msg) {
1220                         /*
1221                          * There is another. Process this one but do not
1222                          * ack it.
1223                          */
1224                         bau_process_message(mdp, bcp, 0);
1225                         /*
1226                          * Let the natural processing of that other message
1227                          * acknowledge it. Don't get the processing of sw_ack's
1228                          * out of order.
1229                          */
1230                         return;
1231                 }
1232         }
1233
1234         /*
1235          * Either the MMR shows this one pending a reply or there is no
1236          * other message using this sw_ack, so it is safe to acknowledge it.
1237          */
1238         bau_process_message(mdp, bcp, 1);
1239
1240         return;
1241 }
1242
1243 /*
1244  * The BAU message interrupt comes here. (registered by set_intr_gate)
1245  * See entry_64.S
1246  *
1247  * We received a broadcast assist message.
1248  *
1249  * Interrupts are disabled; this interrupt could represent
1250  * the receipt of several messages.
1251  *
1252  * All cores/threads on this hub get this interrupt.
1253  * The last one to see it does the software ack.
1254  * (the resource will not be freed until noninterruptable cpus see this
1255  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
1256  */
1257 void uv_bau_message_interrupt(struct pt_regs *regs)
1258 {
1259         int count = 0;
1260         cycles_t time_start;
1261         struct bau_pq_entry *msg;
1262         struct bau_control *bcp;
1263         struct ptc_stats *stat;
1264         struct msg_desc msgdesc;
1265
1266         ack_APIC_irq();
1267         time_start = get_cycles();
1268
1269         bcp = &per_cpu(bau_control, smp_processor_id());
1270         stat = bcp->statp;
1271
1272         msgdesc.queue_first = bcp->queue_first;
1273         msgdesc.queue_last = bcp->queue_last;
1274
1275         msg = bcp->bau_msg_head;
1276         while (msg->swack_vec) {
1277                 count++;
1278
1279                 msgdesc.msg_slot = msg - msgdesc.queue_first;
1280                 msgdesc.msg = msg;
1281                 if (bcp->uvhub_version == 2)
1282                         process_uv2_message(&msgdesc, bcp);
1283                 else
1284                         /* no error workaround for uv1 or uv3 */
1285                         bau_process_message(&msgdesc, bcp, 1);
1286
1287                 msg++;
1288                 if (msg > msgdesc.queue_last)
1289                         msg = msgdesc.queue_first;
1290                 bcp->bau_msg_head = msg;
1291         }
1292         stat->d_time += (get_cycles() - time_start);
1293         if (!count)
1294                 stat->d_nomsg++;
1295         else if (count > 1)
1296                 stat->d_multmsg++;
1297 }
1298
1299 /*
1300  * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1301  * shootdown message timeouts enabled.  The timeout does not cause
1302  * an interrupt, but causes an error message to be returned to
1303  * the sender.
1304  */
1305 static void __init enable_timeouts(void)
1306 {
1307         int uvhub;
1308         int nuvhubs;
1309         int pnode;
1310         unsigned long mmr_image;
1311
1312         nuvhubs = uv_num_possible_blades();
1313
1314         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1315                 if (!uv_blade_nr_possible_cpus(uvhub))
1316                         continue;
1317
1318                 pnode = uv_blade_to_pnode(uvhub);
1319                 mmr_image = read_mmr_misc_control(pnode);
1320                 /*
1321                  * Set the timeout period and then lock it in, in three
1322                  * steps; captures and locks in the period.
1323                  *
1324                  * To program the period, the SOFT_ACK_MODE must be off.
1325                  */
1326                 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1327                 write_mmr_misc_control(pnode, mmr_image);
1328                 /*
1329                  * Set the 4-bit period.
1330                  */
1331                 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1332                 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1333                 write_mmr_misc_control(pnode, mmr_image);
1334                 /*
1335                  * UV1:
1336                  * Subsequent reversals of the timebase bit (3) cause an
1337                  * immediate timeout of one or all INTD resources as
1338                  * indicated in bits 2:0 (7 causes all of them to timeout).
1339                  */
1340                 mmr_image |= (1L << SOFTACK_MSHIFT);
1341                 if (is_uv2_hub()) {
1342                         /* do not touch the legacy mode bit */
1343                         /* hw bug workaround; do not use extended status */
1344                         mmr_image &= ~(1L << UV2_EXT_SHFT);
1345                 } else if (is_uv3_hub()) {
1346                         mmr_image &= ~(1L << PREFETCH_HINT_SHFT);
1347                         mmr_image |= (1L << SB_STATUS_SHFT);
1348                 }
1349                 write_mmr_misc_control(pnode, mmr_image);
1350         }
1351 }
1352
1353 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1354 {
1355         if (*offset < num_possible_cpus())
1356                 return offset;
1357         return NULL;
1358 }
1359
1360 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1361 {
1362         (*offset)++;
1363         if (*offset < num_possible_cpus())
1364                 return offset;
1365         return NULL;
1366 }
1367
1368 static void ptc_seq_stop(struct seq_file *file, void *data)
1369 {
1370 }
1371
1372 /*
1373  * Display the statistics thru /proc/sgi_uv/ptc_statistics
1374  * 'data' points to the cpu number
1375  * Note: see the descriptions in stat_description[].
1376  */
1377 static int ptc_seq_show(struct seq_file *file, void *data)
1378 {
1379         struct ptc_stats *stat;
1380         struct bau_control *bcp;
1381         int cpu;
1382
1383         cpu = *(loff_t *)data;
1384         if (!cpu) {
1385                 seq_puts(file,
1386                          "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1387                 seq_puts(file, "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1388                 seq_puts(file,
1389                          "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1390                 seq_puts(file,
1391                          "rok resetp resett giveup sto bz throt disable ");
1392                 seq_puts(file,
1393                          "enable wars warshw warwaits enters ipidis plugged ");
1394                 seq_puts(file,
1395                          "ipiover glim cong swack recv rtime all one mult ");
1396                 seq_puts(file, "none retry canc nocan reset rcan\n");
1397         }
1398         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1399                 bcp = &per_cpu(bau_control, cpu);
1400                 if (bcp->nobau) {
1401                         seq_printf(file, "cpu %d bau disabled\n", cpu);
1402                         return 0;
1403                 }
1404                 stat = bcp->statp;
1405                 /* source side statistics */
1406                 seq_printf(file,
1407                         "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1408                            cpu, bcp->nobau, stat->s_requestor,
1409                            cycles_2_us(stat->s_time),
1410                            stat->s_ntargself, stat->s_ntarglocals,
1411                            stat->s_ntargremotes, stat->s_ntargcpu,
1412                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1413                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
1414                 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1415                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1416                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1417                            stat->s_dtimeout, stat->s_strongnacks);
1418                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1419                            stat->s_retry_messages, stat->s_retriesok,
1420                            stat->s_resets_plug, stat->s_resets_timeout,
1421                            stat->s_giveup, stat->s_stimeout,
1422                            stat->s_busy, stat->s_throttles);
1423                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1424                            stat->s_bau_disabled, stat->s_bau_reenabled,
1425                            stat->s_uv2_wars, stat->s_uv2_wars_hw,
1426                            stat->s_uv2_war_waits, stat->s_enters,
1427                            stat->s_ipifordisabled, stat->s_plugged,
1428                            stat->s_overipilimit, stat->s_giveuplimit,
1429                            stat->s_congested);
1430
1431                 /* destination side statistics */
1432                 seq_printf(file,
1433                         "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1434                            read_gmmr_sw_ack(uv_cpu_to_pnode(cpu)),
1435                            stat->d_requestee, cycles_2_us(stat->d_time),
1436                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1437                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
1438                            stat->d_nocanceled, stat->d_resets,
1439                            stat->d_rcanceled);
1440         }
1441         return 0;
1442 }
1443
1444 /*
1445  * Display the tunables thru debugfs
1446  */
1447 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1448                                 size_t count, loff_t *ppos)
1449 {
1450         char *buf;
1451         int ret;
1452
1453         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1454                 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1455                 "ipi_reset_limit complete_threshold congested_response_us",
1456                 "congested_reps disabled_period giveup_limit",
1457                 max_concurr, plugged_delay, plugsb4reset,
1458                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1459                 congested_respns_us, congested_reps, disabled_period,
1460                 giveup_limit);
1461
1462         if (!buf)
1463                 return -ENOMEM;
1464
1465         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1466         kfree(buf);
1467         return ret;
1468 }
1469
1470 /*
1471  * handle a write to /proc/sgi_uv/ptc_statistics
1472  * -1: reset the statistics
1473  *  0: display meaning of the statistics
1474  */
1475 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1476                                 size_t count, loff_t *data)
1477 {
1478         int cpu;
1479         int i;
1480         int elements;
1481         long input_arg;
1482         char optstr[64];
1483         struct ptc_stats *stat;
1484
1485         if (count == 0 || count > sizeof(optstr))
1486                 return -EINVAL;
1487         if (copy_from_user(optstr, user, count))
1488                 return -EFAULT;
1489         optstr[count - 1] = '\0';
1490
1491         if (!strcmp(optstr, "on")) {
1492                 set_bau_on();
1493                 return count;
1494         } else if (!strcmp(optstr, "off")) {
1495                 set_bau_off();
1496                 return count;
1497         }
1498
1499         if (kstrtol(optstr, 10, &input_arg) < 0) {
1500                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1501                 return -EINVAL;
1502         }
1503
1504         if (input_arg == 0) {
1505                 elements = ARRAY_SIZE(stat_description);
1506                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1507                 printk(KERN_DEBUG "Sender statistics:\n");
1508                 for (i = 0; i < elements; i++)
1509                         printk(KERN_DEBUG "%s\n", stat_description[i]);
1510         } else if (input_arg == -1) {
1511                 for_each_present_cpu(cpu) {
1512                         stat = &per_cpu(ptcstats, cpu);
1513                         memset(stat, 0, sizeof(struct ptc_stats));
1514                 }
1515         }
1516
1517         return count;
1518 }
1519
1520 static int local_atoi(const char *name)
1521 {
1522         int val = 0;
1523
1524         for (;; name++) {
1525                 switch (*name) {
1526                 case '0' ... '9':
1527                         val = 10*val+(*name-'0');
1528                         break;
1529                 default:
1530                         return val;
1531                 }
1532         }
1533 }
1534
1535 /*
1536  * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1537  * Zero values reset them to defaults.
1538  */
1539 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1540                                 int count)
1541 {
1542         char *p;
1543         char *q;
1544         int cnt = 0;
1545         int val;
1546         int e = ARRAY_SIZE(tunables);
1547
1548         p = instr + strspn(instr, WHITESPACE);
1549         q = p;
1550         for (; *p; p = q + strspn(q, WHITESPACE)) {
1551                 q = p + strcspn(p, WHITESPACE);
1552                 cnt++;
1553                 if (q == p)
1554                         break;
1555         }
1556         if (cnt != e) {
1557                 printk(KERN_INFO "bau tunable error: should be %d values\n", e);
1558                 return -EINVAL;
1559         }
1560
1561         p = instr + strspn(instr, WHITESPACE);
1562         q = p;
1563         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1564                 q = p + strcspn(p, WHITESPACE);
1565                 val = local_atoi(p);
1566                 switch (cnt) {
1567                 case 0:
1568                         if (val == 0) {
1569                                 max_concurr = MAX_BAU_CONCURRENT;
1570                                 max_concurr_const = MAX_BAU_CONCURRENT;
1571                                 continue;
1572                         }
1573                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1574                                 printk(KERN_DEBUG
1575                                 "Error: BAU max concurrent %d is invalid\n",
1576                                 val);
1577                                 return -EINVAL;
1578                         }
1579                         max_concurr = val;
1580                         max_concurr_const = val;
1581                         continue;
1582                 default:
1583                         if (val == 0)
1584                                 *tunables[cnt].tunp = tunables[cnt].deflt;
1585                         else
1586                                 *tunables[cnt].tunp = val;
1587                         continue;
1588                 }
1589                 if (q == p)
1590                         break;
1591         }
1592         return 0;
1593 }
1594
1595 /*
1596  * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1597  */
1598 static ssize_t tunables_write(struct file *file, const char __user *user,
1599                                 size_t count, loff_t *data)
1600 {
1601         int cpu;
1602         int ret;
1603         char instr[100];
1604         struct bau_control *bcp;
1605
1606         if (count == 0 || count > sizeof(instr)-1)
1607                 return -EINVAL;
1608         if (copy_from_user(instr, user, count))
1609                 return -EFAULT;
1610
1611         instr[count] = '\0';
1612
1613         cpu = get_cpu();
1614         bcp = &per_cpu(bau_control, cpu);
1615         ret = parse_tunables_write(bcp, instr, count);
1616         put_cpu();
1617         if (ret)
1618                 return ret;
1619
1620         for_each_present_cpu(cpu) {
1621                 bcp = &per_cpu(bau_control, cpu);
1622                 bcp->max_concurr         = max_concurr;
1623                 bcp->max_concurr_const   = max_concurr;
1624                 bcp->plugged_delay       = plugged_delay;
1625                 bcp->plugsb4reset        = plugsb4reset;
1626                 bcp->timeoutsb4reset     = timeoutsb4reset;
1627                 bcp->ipi_reset_limit     = ipi_reset_limit;
1628                 bcp->complete_threshold  = complete_threshold;
1629                 bcp->cong_response_us    = congested_respns_us;
1630                 bcp->cong_reps           = congested_reps;
1631                 bcp->disabled_period     = sec_2_cycles(disabled_period);
1632                 bcp->giveup_limit        = giveup_limit;
1633         }
1634         return count;
1635 }
1636
1637 static const struct seq_operations uv_ptc_seq_ops = {
1638         .start          = ptc_seq_start,
1639         .next           = ptc_seq_next,
1640         .stop           = ptc_seq_stop,
1641         .show           = ptc_seq_show
1642 };
1643
1644 static int ptc_proc_open(struct inode *inode, struct file *file)
1645 {
1646         return seq_open(file, &uv_ptc_seq_ops);
1647 }
1648
1649 static int tunables_open(struct inode *inode, struct file *file)
1650 {
1651         return 0;
1652 }
1653
1654 static const struct file_operations proc_uv_ptc_operations = {
1655         .open           = ptc_proc_open,
1656         .read           = seq_read,
1657         .write          = ptc_proc_write,
1658         .llseek         = seq_lseek,
1659         .release        = seq_release,
1660 };
1661
1662 static const struct file_operations tunables_fops = {
1663         .open           = tunables_open,
1664         .read           = tunables_read,
1665         .write          = tunables_write,
1666         .llseek         = default_llseek,
1667 };
1668
1669 static int __init uv_ptc_init(void)
1670 {
1671         struct proc_dir_entry *proc_uv_ptc;
1672
1673         if (!is_uv_system())
1674                 return 0;
1675
1676         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1677                                   &proc_uv_ptc_operations);
1678         if (!proc_uv_ptc) {
1679                 printk(KERN_ERR "unable to create %s proc entry\n",
1680                        UV_PTC_BASENAME);
1681                 return -EINVAL;
1682         }
1683
1684         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1685         if (!tunables_dir) {
1686                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1687                        UV_BAU_TUNABLES_DIR);
1688                 return -EINVAL;
1689         }
1690         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1691                                         tunables_dir, NULL, &tunables_fops);
1692         if (!tunables_file) {
1693                 printk(KERN_ERR "unable to create debugfs file %s\n",
1694                        UV_BAU_TUNABLES_FILE);
1695                 return -EINVAL;
1696         }
1697         return 0;
1698 }
1699
1700 /*
1701  * Initialize the sending side's sending buffers.
1702  */
1703 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1704 {
1705         int i;
1706         int cpu;
1707         int uv1 = 0;
1708         unsigned long gpa;
1709         unsigned long m;
1710         unsigned long n;
1711         size_t dsize;
1712         struct bau_desc *bau_desc;
1713         struct bau_desc *bd2;
1714         struct uv1_bau_msg_header *uv1_hdr;
1715         struct uv2_3_bau_msg_header *uv2_3_hdr;
1716         struct bau_control *bcp;
1717
1718         /*
1719          * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1720          * per cpu; and one per cpu on the uvhub (ADP_SZ)
1721          */
1722         dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1723         bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1724         BUG_ON(!bau_desc);
1725
1726         gpa = uv_gpa(bau_desc);
1727         n = uv_gpa_to_gnode(gpa);
1728         m = uv_gpa_to_offset(gpa);
1729         if (is_uv1_hub())
1730                 uv1 = 1;
1731
1732         /* the 14-bit pnode */
1733         write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1734         /*
1735          * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1736          * cpu even though we only use the first one; one descriptor can
1737          * describe a broadcast to 256 uv hubs.
1738          */
1739         for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1740                 memset(bd2, 0, sizeof(struct bau_desc));
1741                 if (uv1) {
1742                         uv1_hdr = &bd2->header.uv1_hdr;
1743                         uv1_hdr->swack_flag = 1;
1744                         /*
1745                          * The base_dest_nasid set in the message header
1746                          * is the nasid of the first uvhub in the partition.
1747                          * The bit map will indicate destination pnode numbers
1748                          * relative to that base. They may not be consecutive
1749                          * if nasid striding is being used.
1750                          */
1751                         uv1_hdr->base_dest_nasid =
1752                                                   UV_PNODE_TO_NASID(base_pnode);
1753                         uv1_hdr->dest_subnodeid  = UV_LB_SUBNODEID;
1754                         uv1_hdr->command         = UV_NET_ENDPOINT_INTD;
1755                         uv1_hdr->int_both        = 1;
1756                         /*
1757                          * all others need to be set to zero:
1758                          *   fairness chaining multilevel count replied_to
1759                          */
1760                 } else {
1761                         /*
1762                          * BIOS uses legacy mode, but uv2 and uv3 hardware always
1763                          * uses native mode for selective broadcasts.
1764                          */
1765                         uv2_3_hdr = &bd2->header.uv2_3_hdr;
1766                         uv2_3_hdr->swack_flag      = 1;
1767                         uv2_3_hdr->base_dest_nasid =
1768                                                   UV_PNODE_TO_NASID(base_pnode);
1769                         uv2_3_hdr->dest_subnodeid  = UV_LB_SUBNODEID;
1770                         uv2_3_hdr->command         = UV_NET_ENDPOINT_INTD;
1771                 }
1772         }
1773         for_each_present_cpu(cpu) {
1774                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1775                         continue;
1776                 bcp = &per_cpu(bau_control, cpu);
1777                 bcp->descriptor_base = bau_desc;
1778         }
1779 }
1780
1781 /*
1782  * initialize the destination side's receiving buffers
1783  * entered for each uvhub in the partition
1784  * - node is first node (kernel memory notion) on the uvhub
1785  * - pnode is the uvhub's physical identifier
1786  */
1787 static void pq_init(int node, int pnode)
1788 {
1789         int cpu;
1790         size_t plsize;
1791         char *cp;
1792         void *vp;
1793         unsigned long pn;
1794         unsigned long first;
1795         unsigned long pn_first;
1796         unsigned long last;
1797         struct bau_pq_entry *pqp;
1798         struct bau_control *bcp;
1799
1800         plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1801         vp = kmalloc_node(plsize, GFP_KERNEL, node);
1802         pqp = (struct bau_pq_entry *)vp;
1803         BUG_ON(!pqp);
1804
1805         cp = (char *)pqp + 31;
1806         pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1807
1808         for_each_present_cpu(cpu) {
1809                 if (pnode != uv_cpu_to_pnode(cpu))
1810                         continue;
1811                 /* for every cpu on this pnode: */
1812                 bcp = &per_cpu(bau_control, cpu);
1813                 bcp->queue_first        = pqp;
1814                 bcp->bau_msg_head       = pqp;
1815                 bcp->queue_last         = pqp + (DEST_Q_SIZE - 1);
1816         }
1817         /*
1818          * need the gnode of where the memory was really allocated
1819          */
1820         pn = uv_gpa_to_gnode(uv_gpa(pqp));
1821         first = uv_physnodeaddr(pqp);
1822         pn_first = ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | first;
1823         last = uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1));
1824         write_mmr_payload_first(pnode, pn_first);
1825         write_mmr_payload_tail(pnode, first);
1826         write_mmr_payload_last(pnode, last);
1827         write_gmmr_sw_ack(pnode, 0xffffUL);
1828
1829         /* in effect, all msg_type's are set to MSG_NOOP */
1830         memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1831 }
1832
1833 /*
1834  * Initialization of each UV hub's structures
1835  */
1836 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1837 {
1838         int node;
1839         int pnode;
1840         unsigned long apicid;
1841
1842         node = uvhub_to_first_node(uvhub);
1843         pnode = uv_blade_to_pnode(uvhub);
1844
1845         activation_descriptor_init(node, pnode, base_pnode);
1846
1847         pq_init(node, pnode);
1848         /*
1849          * The below initialization can't be in firmware because the
1850          * messaging IRQ will be determined by the OS.
1851          */
1852         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1853         write_mmr_data_config(pnode, ((apicid << 32) | vector));
1854 }
1855
1856 /*
1857  * We will set BAU_MISC_CONTROL with a timeout period.
1858  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1859  * So the destination timeout period has to be calculated from them.
1860  */
1861 static int calculate_destination_timeout(void)
1862 {
1863         unsigned long mmr_image;
1864         int mult1;
1865         int mult2;
1866         int index;
1867         int base;
1868         int ret;
1869         unsigned long ts_ns;
1870
1871         if (is_uv1_hub()) {
1872                 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1873                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1874                 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1875                 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1876                 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1877                 ts_ns = timeout_base_ns[index];
1878                 ts_ns *= (mult1 * mult2);
1879                 ret = ts_ns / 1000;
1880         } else {
1881                 /* same destination timeout for uv2 and uv3 */
1882                 /* 4 bits  0/1 for 10/80us base, 3 bits of multiplier */
1883                 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1884                 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1885                 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1886                         base = 80;
1887                 else
1888                         base = 10;
1889                 mult1 = mmr_image & UV2_ACK_MASK;
1890                 ret = mult1 * base;
1891         }
1892         return ret;
1893 }
1894
1895 static void __init init_per_cpu_tunables(void)
1896 {
1897         int cpu;
1898         struct bau_control *bcp;
1899
1900         for_each_present_cpu(cpu) {
1901                 bcp = &per_cpu(bau_control, cpu);
1902                 bcp->baudisabled                = 0;
1903                 if (nobau)
1904                         bcp->nobau              = true;
1905                 bcp->statp                      = &per_cpu(ptcstats, cpu);
1906                 /* time interval to catch a hardware stay-busy bug */
1907                 bcp->timeout_interval           = usec_2_cycles(2*timeout_us);
1908                 bcp->max_concurr                = max_concurr;
1909                 bcp->max_concurr_const          = max_concurr;
1910                 bcp->plugged_delay              = plugged_delay;
1911                 bcp->plugsb4reset               = plugsb4reset;
1912                 bcp->timeoutsb4reset            = timeoutsb4reset;
1913                 bcp->ipi_reset_limit            = ipi_reset_limit;
1914                 bcp->complete_threshold         = complete_threshold;
1915                 bcp->cong_response_us           = congested_respns_us;
1916                 bcp->cong_reps                  = congested_reps;
1917                 bcp->disabled_period            = sec_2_cycles(disabled_period);
1918                 bcp->giveup_limit               = giveup_limit;
1919                 spin_lock_init(&bcp->queue_lock);
1920                 spin_lock_init(&bcp->uvhub_lock);
1921                 spin_lock_init(&bcp->disable_lock);
1922         }
1923 }
1924
1925 /*
1926  * Scan all cpus to collect blade and socket summaries.
1927  */
1928 static int __init get_cpu_topology(int base_pnode,
1929                                         struct uvhub_desc *uvhub_descs,
1930                                         unsigned char *uvhub_mask)
1931 {
1932         int cpu;
1933         int pnode;
1934         int uvhub;
1935         int socket;
1936         struct bau_control *bcp;
1937         struct uvhub_desc *bdp;
1938         struct socket_desc *sdp;
1939
1940         for_each_present_cpu(cpu) {
1941                 bcp = &per_cpu(bau_control, cpu);
1942
1943                 memset(bcp, 0, sizeof(struct bau_control));
1944
1945                 pnode = uv_cpu_hub_info(cpu)->pnode;
1946                 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1947                         printk(KERN_EMERG
1948                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1949                                 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1950                         return 1;
1951                 }
1952
1953                 bcp->osnode = cpu_to_node(cpu);
1954                 bcp->partition_base_pnode = base_pnode;
1955
1956                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1957                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1958                 bdp = &uvhub_descs[uvhub];
1959
1960                 bdp->num_cpus++;
1961                 bdp->uvhub = uvhub;
1962                 bdp->pnode = pnode;
1963
1964                 /* kludge: 'assuming' one node per socket, and assuming that
1965                    disabling a socket just leaves a gap in node numbers */
1966                 socket = bcp->osnode & 1;
1967                 bdp->socket_mask |= (1 << socket);
1968                 sdp = &bdp->socket[socket];
1969                 sdp->cpu_number[sdp->num_cpus] = cpu;
1970                 sdp->num_cpus++;
1971                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1972                         printk(KERN_EMERG "%d cpus per socket invalid\n",
1973                                 sdp->num_cpus);
1974                         return 1;
1975                 }
1976         }
1977         return 0;
1978 }
1979
1980 /*
1981  * Each socket is to get a local array of pnodes/hubs.
1982  */
1983 static void make_per_cpu_thp(struct bau_control *smaster)
1984 {
1985         int cpu;
1986         size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
1987
1988         smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
1989         memset(smaster->thp, 0, hpsz);
1990         for_each_present_cpu(cpu) {
1991                 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
1992                 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1993         }
1994 }
1995
1996 /*
1997  * Each uvhub is to get a local cpumask.
1998  */
1999 static void make_per_hub_cpumask(struct bau_control *hmaster)
2000 {
2001         int sz = sizeof(cpumask_t);
2002
2003         hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
2004 }
2005
2006 /*
2007  * Initialize all the per_cpu information for the cpu's on a given socket,
2008  * given what has been gathered into the socket_desc struct.
2009  * And reports the chosen hub and socket masters back to the caller.
2010  */
2011 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
2012                         struct bau_control **smasterp,
2013                         struct bau_control **hmasterp)
2014 {
2015         int i;
2016         int cpu;
2017         struct bau_control *bcp;
2018
2019         for (i = 0; i < sdp->num_cpus; i++) {
2020                 cpu = sdp->cpu_number[i];
2021                 bcp = &per_cpu(bau_control, cpu);
2022                 bcp->cpu = cpu;
2023                 if (i == 0) {
2024                         *smasterp = bcp;
2025                         if (!(*hmasterp))
2026                                 *hmasterp = bcp;
2027                 }
2028                 bcp->cpus_in_uvhub = bdp->num_cpus;
2029                 bcp->cpus_in_socket = sdp->num_cpus;
2030                 bcp->socket_master = *smasterp;
2031                 bcp->uvhub = bdp->uvhub;
2032                 if (is_uv1_hub())
2033                         bcp->uvhub_version = 1;
2034                 else if (is_uv2_hub())
2035                         bcp->uvhub_version = 2;
2036                 else if (is_uv3_hub())
2037                         bcp->uvhub_version = 3;
2038                 else {
2039                         printk(KERN_EMERG "uvhub version not 1, 2 or 3\n");
2040                         return 1;
2041                 }
2042                 bcp->uvhub_master = *hmasterp;
2043                 bcp->uvhub_cpu = uv_cpu_blade_processor_id(cpu);
2044
2045                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2046                         printk(KERN_EMERG "%d cpus per uvhub invalid\n",
2047                                 bcp->uvhub_cpu);
2048                         return 1;
2049                 }
2050         }
2051         return 0;
2052 }
2053
2054 /*
2055  * Summarize the blade and socket topology into the per_cpu structures.
2056  */
2057 static int __init summarize_uvhub_sockets(int nuvhubs,
2058                         struct uvhub_desc *uvhub_descs,
2059                         unsigned char *uvhub_mask)
2060 {
2061         int socket;
2062         int uvhub;
2063         unsigned short socket_mask;
2064
2065         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2066                 struct uvhub_desc *bdp;
2067                 struct bau_control *smaster = NULL;
2068                 struct bau_control *hmaster = NULL;
2069
2070                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2071                         continue;
2072
2073                 bdp = &uvhub_descs[uvhub];
2074                 socket_mask = bdp->socket_mask;
2075                 socket = 0;
2076                 while (socket_mask) {
2077                         struct socket_desc *sdp;
2078                         if ((socket_mask & 1)) {
2079                                 sdp = &bdp->socket[socket];
2080                                 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2081                                         return 1;
2082                                 make_per_cpu_thp(smaster);
2083                         }
2084                         socket++;
2085                         socket_mask = (socket_mask >> 1);
2086                 }
2087                 make_per_hub_cpumask(hmaster);
2088         }
2089         return 0;
2090 }
2091
2092 /*
2093  * initialize the bau_control structure for each cpu
2094  */
2095 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2096 {
2097         unsigned char *uvhub_mask;
2098         void *vp;
2099         struct uvhub_desc *uvhub_descs;
2100
2101         timeout_us = calculate_destination_timeout();
2102
2103         vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
2104         uvhub_descs = (struct uvhub_desc *)vp;
2105         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
2106         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2107
2108         if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2109                 goto fail;
2110
2111         if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2112                 goto fail;
2113
2114         kfree(uvhub_descs);
2115         kfree(uvhub_mask);
2116         init_per_cpu_tunables();
2117         return 0;
2118
2119 fail:
2120         kfree(uvhub_descs);
2121         kfree(uvhub_mask);
2122         return 1;
2123 }
2124
2125 /*
2126  * Initialization of BAU-related structures
2127  */
2128 static int __init uv_bau_init(void)
2129 {
2130         int uvhub;
2131         int pnode;
2132         int nuvhubs;
2133         int cur_cpu;
2134         int cpus;
2135         int vector;
2136         cpumask_var_t *mask;
2137
2138         if (!is_uv_system())
2139                 return 0;
2140
2141         for_each_possible_cpu(cur_cpu) {
2142                 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2143                 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2144         }
2145
2146         nuvhubs = uv_num_possible_blades();
2147         congested_cycles = usec_2_cycles(congested_respns_us);
2148
2149         uv_base_pnode = 0x7fffffff;
2150         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2151                 cpus = uv_blade_nr_possible_cpus(uvhub);
2152                 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2153                         uv_base_pnode = uv_blade_to_pnode(uvhub);
2154         }
2155
2156         enable_timeouts();
2157
2158         if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2159                 set_bau_off();
2160                 nobau_perm = 1;
2161                 return 0;
2162         }
2163
2164         vector = UV_BAU_MESSAGE;
2165         for_each_possible_blade(uvhub) {
2166                 if (uv_blade_nr_possible_cpus(uvhub))
2167                         init_uvhub(uvhub, vector, uv_base_pnode);
2168         }
2169
2170         alloc_intr_gate(vector, uv_bau_message_intr1);
2171
2172         for_each_possible_blade(uvhub) {
2173                 if (uv_blade_nr_possible_cpus(uvhub)) {
2174                         unsigned long val;
2175                         unsigned long mmr;
2176                         pnode = uv_blade_to_pnode(uvhub);
2177                         /* INIT the bau */
2178                         val = 1L << 63;
2179                         write_gmmr_activation(pnode, val);
2180                         mmr = 1; /* should be 1 to broadcast to both sockets */
2181                         if (!is_uv1_hub())
2182                                 write_mmr_data_broadcast(pnode, mmr);
2183                 }
2184         }
2185
2186         return 0;
2187 }
2188 core_initcall(uv_bau_init);
2189 fs_initcall(uv_ptc_init);