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 /*
 437  * Not to be confused with cycles_2_ns() from tsc.c; this gives a relative
 438  * number, not an absolute. It converts a duration in cycles to a duration in
 439  * ns.
 440  */
 441 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
 442 {
 443         struct cyc2ns_data *data = cyc2ns_read_begin();
 444         unsigned long long ns;
 445
 446         ns = mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift);
 447
 448         cyc2ns_read_end(data);
 449         return ns;
 450 }
 451
 452 /*
 453  * The reverse of the above; converts a duration in ns to a duration in cycles.
 454  */
 455 static inline unsigned long long ns_2_cycles(unsigned long long ns)
 456 {
 457         struct cyc2ns_data *data = cyc2ns_read_begin();
 458         unsigned long long cyc;
 459
 460         cyc = (ns << data->cyc2ns_shift) / data->cyc2ns_mul;
 461
 462         cyc2ns_read_end(data);
 463         return cyc;
 464 }
 465
 466 static inline unsigned long cycles_2_us(unsigned long long cyc)
 467 {
 468         return cycles_2_ns(cyc) / NSEC_PER_USEC;
 469 }
 470
 471 static inline cycles_t sec_2_cycles(unsigned long sec)
 472 {
 473         return ns_2_cycles(sec * NSEC_PER_SEC);
 474 }
 475
 476 static inline unsigned long long usec_2_cycles(unsigned long usec)
 477 {
 478         return ns_2_cycles(usec * NSEC_PER_USEC);
 479 }
 480
 481 /*
 482  * wait for all cpus on this hub to finish their sends and go quiet
 483  * leaves uvhub_quiesce set so that no new broadcasts are started by
 484  * bau_flush_send_and_wait()
 485  */
 486 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
 487 {
 488         atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 489 }
 490
 491 /*
 492  * mark this quiet-requestor as done
 493  */
 494 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
 495 {
 496         atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 497 }
 498
 499 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
 500 {
 501         unsigned long descriptor_status;
 502
 503         descriptor_status = uv_read_local_mmr(mmr_offset);
 504         descriptor_status >>= right_shift;
 505         descriptor_status &= UV_ACT_STATUS_MASK;
 506         return descriptor_status;
 507 }
 508
 509 /*
 510  * Wait for completion of a broadcast software ack message
 511  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
 512  */
 513 static int uv1_wait_completion(struct bau_desc *bau_desc,
 514                                 unsigned long mmr_offset, int right_shift,
 515                                 struct bau_control *bcp, long try)
 516 {
 517         unsigned long descriptor_status;
 518         cycles_t ttm;
 519         struct ptc_stats *stat = bcp->statp;
 520
 521         descriptor_status = uv1_read_status(mmr_offset, right_shift);
 522         /* spin on the status MMR, waiting for it to go idle */
 523         while ((descriptor_status != DS_IDLE)) {
 524                 /*
 525                  * Our software ack messages may be blocked because
 526                  * there are no swack resources available.  As long
 527                  * as none of them has timed out hardware will NACK
 528                  * our message and its state will stay IDLE.
 529                  */
 530                 if (descriptor_status == DS_SOURCE_TIMEOUT) {
 531                         stat->s_stimeout++;
 532                         return FLUSH_GIVEUP;
 533                 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
 534                         stat->s_dtimeout++;
 535                         ttm = get_cycles();
 536
 537                         /*
 538                          * Our retries may be blocked by all destination
 539                          * swack resources being consumed, and a timeout
 540                          * pending.  In that case hardware returns the
 541                          * ERROR that looks like a destination timeout.
 542                          */
 543                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 544                                 bcp->conseccompletes = 0;
 545                                 return FLUSH_RETRY_PLUGGED;
 546                         }
 547
 548                         bcp->conseccompletes = 0;
 549                         return FLUSH_RETRY_TIMEOUT;
 550                 } else {
 551                         /*
 552                          * descriptor_status is still BUSY
 553                          */
 554                         cpu_relax();
 555                 }
 556                 descriptor_status = uv1_read_status(mmr_offset, right_shift);
 557         }
 558         bcp->conseccompletes++;
 559         return FLUSH_COMPLETE;
 560 }
 561
 562 /*
 563  * UV2 could have an extra bit of status in the ACTIVATION_STATUS_2 register.
 564  * But not currently used.
 565  */
 566 static unsigned long uv2_read_status(unsigned long offset, int rshft, int desc)
 567 {
 568         unsigned long descriptor_status;
 569
 570         descriptor_status =
 571                 ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK) << 1;
 572         return descriptor_status;
 573 }
 574
 575 /*
 576  * Return whether the status of the descriptor that is normally used for this
 577  * cpu (the one indexed by its hub-relative cpu number) is busy.
 578  * The status of the original 32 descriptors is always reflected in the 64
 579  * bits of UVH_LB_BAU_SB_ACTIVATION_STATUS_0.
 580  * The bit provided by the activation_status_2 register is irrelevant to
 581  * the status if it is only being tested for busy or not busy.
 582  */
 583 int normal_busy(struct bau_control *bcp)
 584 {
 585         int cpu = bcp->uvhub_cpu;
 586         int mmr_offset;
 587         int right_shift;
 588
 589         mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 590         right_shift = cpu * UV_ACT_STATUS_SIZE;
 591         return (((((read_lmmr(mmr_offset) >> right_shift) &
 592                                 UV_ACT_STATUS_MASK)) << 1) == UV2H_DESC_BUSY);
 593 }
 594
 595 /*
 596  * Entered when a bau descriptor has gone into a permanent busy wait because
 597  * of a hardware bug.
 598  * Workaround the bug.
 599  */
 600 int handle_uv2_busy(struct bau_control *bcp)
 601 {
 602         struct ptc_stats *stat = bcp->statp;
 603
 604         stat->s_uv2_wars++;
 605         bcp->busy = 1;
 606         return FLUSH_GIVEUP;
 607 }
 608
 609 static int uv2_wait_completion(struct bau_desc *bau_desc,
 610                                 unsigned long mmr_offset, int right_shift,
 611                                 struct bau_control *bcp, long try)
 612 {
 613         unsigned long descriptor_stat;
 614         cycles_t ttm;
 615         int desc = bcp->uvhub_cpu;
 616         long busy_reps = 0;
 617         struct ptc_stats *stat = bcp->statp;
 618
 619         descriptor_stat = uv2_read_status(mmr_offset, right_shift, desc);
 620
 621         /* spin on the status MMR, waiting for it to go idle */
 622         while (descriptor_stat != UV2H_DESC_IDLE) {
 623                 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT)) {
 624                         /*
 625                          * A h/w bug on the destination side may
 626                          * have prevented the message being marked
 627                          * pending, thus it doesn't get replied to
 628                          * and gets continually nacked until it times
 629                          * out with a SOURCE_TIMEOUT.
 630                          */
 631                         stat->s_stimeout++;
 632                         return FLUSH_GIVEUP;
 633                 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
 634                         ttm = get_cycles();
 635
 636                         /*
 637                          * Our retries may be blocked by all destination
 638                          * swack resources being consumed, and a timeout
 639                          * pending.  In that case hardware returns the
 640                          * ERROR that looks like a destination timeout.
 641                          * Without using the extended status we have to
 642                          * deduce from the short time that this was a
 643                          * strong nack.
 644                          */
 645                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 646                                 bcp->conseccompletes = 0;
 647                                 stat->s_plugged++;
 648                                 /* FLUSH_RETRY_PLUGGED causes hang on boot */
 649                                 return FLUSH_GIVEUP;
 650                         }
 651                         stat->s_dtimeout++;
 652                         bcp->conseccompletes = 0;
 653                         /* FLUSH_RETRY_TIMEOUT causes hang on boot */
 654                         return FLUSH_GIVEUP;
 655                 } else {
 656                         busy_reps++;
 657                         if (busy_reps > 1000000) {
 658                                 /* not to hammer on the clock */
 659                                 busy_reps = 0;
 660                                 ttm = get_cycles();
 661                                 if ((ttm - bcp->send_message) >
 662                                                 bcp->timeout_interval)
 663                                         return handle_uv2_busy(bcp);
 664                         }
 665                         /*
 666                          * descriptor_stat is still BUSY
 667                          */
 668                         cpu_relax();
 669                 }
 670                 descriptor_stat = uv2_read_status(mmr_offset, right_shift,
 671                                                                         desc);
 672         }
 673         bcp->conseccompletes++;
 674         return FLUSH_COMPLETE;
 675 }
 676
 677 /*
 678  * There are 2 status registers; each and array[32] of 2 bits. Set up for
 679  * which register to read and position in that register based on cpu in
 680  * current hub.
 681  */
 682 static int wait_completion(struct bau_desc *bau_desc,
 683                                 struct bau_control *bcp, long try)
 684 {
 685         int right_shift;
 686         unsigned long mmr_offset;
 687         int desc = bcp->uvhub_cpu;
 688
 689         if (desc < UV_CPUS_PER_AS) {
 690                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 691                 right_shift = desc * UV_ACT_STATUS_SIZE;
 692         } else {
 693                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
 694                 right_shift = ((desc - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
 695         }
 696
 697         if (bcp->uvhub_version == 1)
 698                 return uv1_wait_completion(bau_desc, mmr_offset, right_shift,
 699                                                                 bcp, try);
 700         else
 701                 return uv2_wait_completion(bau_desc, mmr_offset, right_shift,
 702                                                                 bcp, try);
 703 }
 704
 705 /*
 706  * Our retries are blocked by all destination sw ack resources being
 707  * in use, and a timeout is pending. In that case hardware immediately
 708  * returns the ERROR that looks like a destination timeout.
 709  */
 710 static void destination_plugged(struct bau_desc *bau_desc,
 711                         struct bau_control *bcp,
 712                         struct bau_control *hmaster, struct ptc_stats *stat)
 713 {
 714         udelay(bcp->plugged_delay);
 715         bcp->plugged_tries++;
 716
 717         if (bcp->plugged_tries >= bcp->plugsb4reset) {
 718                 bcp->plugged_tries = 0;
 719
 720                 quiesce_local_uvhub(hmaster);
 721
 722                 spin_lock(&hmaster->queue_lock);
 723                 reset_with_ipi(&bau_desc->distribution, bcp);
 724                 spin_unlock(&hmaster->queue_lock);
 725
 726                 end_uvhub_quiesce(hmaster);
 727
 728                 bcp->ipi_attempts++;
 729                 stat->s_resets_plug++;
 730         }
 731 }
 732
 733 static void destination_timeout(struct bau_desc *bau_desc,
 734                         struct bau_control *bcp, struct bau_control *hmaster,
 735                         struct ptc_stats *stat)
 736 {
 737         hmaster->max_concurr = 1;
 738         bcp->timeout_tries++;
 739         if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
 740                 bcp->timeout_tries = 0;
 741
 742                 quiesce_local_uvhub(hmaster);
 743
 744                 spin_lock(&hmaster->queue_lock);
 745                 reset_with_ipi(&bau_desc->distribution, bcp);
 746                 spin_unlock(&hmaster->queue_lock);
 747
 748                 end_uvhub_quiesce(hmaster);
 749
 750                 bcp->ipi_attempts++;
 751                 stat->s_resets_timeout++;
 752         }
 753 }
 754
 755 /*
 756  * Stop all cpus on a uvhub from using the BAU for a period of time.
 757  * This is reversed by check_enable.
 758  */
 759 static void disable_for_period(struct bau_control *bcp, struct ptc_stats *stat)
 760 {
 761         int tcpu;
 762         struct bau_control *tbcp;
 763         struct bau_control *hmaster;
 764         cycles_t tm1;
 765
 766         hmaster = bcp->uvhub_master;
 767         spin_lock(&hmaster->disable_lock);
 768         if (!bcp->baudisabled) {
 769                 stat->s_bau_disabled++;
 770                 tm1 = get_cycles();
 771                 for_each_present_cpu(tcpu) {
 772                         tbcp = &per_cpu(bau_control, tcpu);
 773                         if (tbcp->uvhub_master == hmaster) {
 774                                 tbcp->baudisabled = 1;
 775                                 tbcp->set_bau_on_time =
 776                                         tm1 + bcp->disabled_period;
 777                         }
 778                 }
 779         }
 780         spin_unlock(&hmaster->disable_lock);
 781 }
 782
 783 static void count_max_concurr(int stat, struct bau_control *bcp,
 784                                 struct bau_control *hmaster)
 785 {
 786         bcp->plugged_tries = 0;
 787         bcp->timeout_tries = 0;
 788         if (stat != FLUSH_COMPLETE)
 789                 return;
 790         if (bcp->conseccompletes <= bcp->complete_threshold)
 791                 return;
 792         if (hmaster->max_concurr >= hmaster->max_concurr_const)
 793                 return;
 794         hmaster->max_concurr++;
 795 }
 796
 797 static void record_send_stats(cycles_t time1, cycles_t time2,
 798                 struct bau_control *bcp, struct ptc_stats *stat,
 799                 int completion_status, int try)
 800 {
 801         cycles_t elapsed;
 802
 803         if (time2 > time1) {
 804                 elapsed = time2 - time1;
 805                 stat->s_time += elapsed;
 806
 807                 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
 808                         bcp->period_requests++;
 809                         bcp->period_time += elapsed;
 810                         if ((elapsed > congested_cycles) &&
 811                             (bcp->period_requests > bcp->cong_reps) &&
 812                             ((bcp->period_time / bcp->period_requests) >
 813                                                         congested_cycles)) {
 814                                 stat->s_congested++;
 815                                 disable_for_period(bcp, stat);
 816                         }
 817                 }
 818         } else
 819                 stat->s_requestor--;
 820
 821         if (completion_status == FLUSH_COMPLETE && try > 1)
 822                 stat->s_retriesok++;
 823         else if (completion_status == FLUSH_GIVEUP) {
 824                 stat->s_giveup++;
 825                 if (get_cycles() > bcp->period_end)
 826                         bcp->period_giveups = 0;
 827                 bcp->period_giveups++;
 828                 if (bcp->period_giveups == 1)
 829                         bcp->period_end = get_cycles() + bcp->disabled_period;
 830                 if (bcp->period_giveups > bcp->giveup_limit) {
 831                         disable_for_period(bcp, stat);
 832                         stat->s_giveuplimit++;
 833                 }
 834         }
 835 }
 836
 837 /*
 838  * Because of a uv1 hardware bug only a limited number of concurrent
 839  * requests can be made.
 840  */
 841 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
 842 {
 843         spinlock_t *lock = &hmaster->uvhub_lock;
 844         atomic_t *v;
 845
 846         v = &hmaster->active_descriptor_count;
 847         if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
 848                 stat->s_throttles++;
 849                 do {
 850                         cpu_relax();
 851                 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
 852         }
 853 }
 854
 855 /*
 856  * Handle the completion status of a message send.
 857  */
 858 static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
 859                         struct bau_control *bcp, struct bau_control *hmaster,
 860                         struct ptc_stats *stat)
 861 {
 862         if (completion_status == FLUSH_RETRY_PLUGGED)
 863                 destination_plugged(bau_desc, bcp, hmaster, stat);
 864         else if (completion_status == FLUSH_RETRY_TIMEOUT)
 865                 destination_timeout(bau_desc, bcp, hmaster, stat);
 866 }
 867
 868 /*
 869  * Send a broadcast and wait for it to complete.
 870  *
 871  * The flush_mask contains the cpus the broadcast is to be sent to including
 872  * cpus that are on the local uvhub.
 873  *
 874  * Returns 0 if all flushing represented in the mask was done.
 875  * Returns 1 if it gives up entirely and the original cpu mask is to be
 876  * returned to the kernel.
 877  */
 878 int uv_flush_send_and_wait(struct cpumask *flush_mask, struct bau_control *bcp,
 879         struct bau_desc *bau_desc)
 880 {
 881         int seq_number = 0;
 882         int completion_stat = 0;
 883         int uv1 = 0;
 884         long try = 0;
 885         unsigned long index;
 886         cycles_t time1;
 887         cycles_t time2;
 888         struct ptc_stats *stat = bcp->statp;
 889         struct bau_control *hmaster = bcp->uvhub_master;
 890         struct uv1_bau_msg_header *uv1_hdr = NULL;
 891         struct uv2_bau_msg_header *uv2_hdr = NULL;
 892
 893         if (bcp->uvhub_version == 1) {
 894                 uv1 = 1;
 895                 uv1_throttle(hmaster, stat);
 896         }
 897
 898         while (hmaster->uvhub_quiesce)
 899                 cpu_relax();
 900
 901         time1 = get_cycles();
 902         if (uv1)
 903                 uv1_hdr = &bau_desc->header.uv1_hdr;
 904         else
 905                 uv2_hdr = &bau_desc->header.uv2_hdr;
 906
 907         do {
 908                 if (try == 0) {
 909                         if (uv1)
 910                                 uv1_hdr->msg_type = MSG_REGULAR;
 911                         else
 912                                 uv2_hdr->msg_type = MSG_REGULAR;
 913                         seq_number = bcp->message_number++;
 914                 } else {
 915                         if (uv1)
 916                                 uv1_hdr->msg_type = MSG_RETRY;
 917                         else
 918                                 uv2_hdr->msg_type = MSG_RETRY;
 919                         stat->s_retry_messages++;
 920                 }
 921
 922                 if (uv1)
 923                         uv1_hdr->sequence = seq_number;
 924                 else
 925                         uv2_hdr->sequence = seq_number;
 926                 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
 927                 bcp->send_message = get_cycles();
 928
 929                 write_mmr_activation(index);
 930
 931                 try++;
 932                 completion_stat = wait_completion(bau_desc, bcp, try);
 933
 934                 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
 935
 936                 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
 937                         bcp->ipi_attempts = 0;
 938                         stat->s_overipilimit++;
 939                         completion_stat = FLUSH_GIVEUP;
 940                         break;
 941                 }
 942                 cpu_relax();
 943         } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
 944                  (completion_stat == FLUSH_RETRY_TIMEOUT));
 945
 946         time2 = get_cycles();
 947
 948         count_max_concurr(completion_stat, bcp, hmaster);
 949
 950         while (hmaster->uvhub_quiesce)
 951                 cpu_relax();
 952
 953         atomic_dec(&hmaster->active_descriptor_count);
 954
 955         record_send_stats(time1, time2, bcp, stat, completion_stat, try);
 956
 957         if (completion_stat == FLUSH_GIVEUP)
 958                 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
 959                 return 1;
 960         return 0;
 961 }
 962
 963 /*
 964  * The BAU is disabled for this uvhub. When the disabled time period has
 965  * expired re-enable it.
 966  * Return 0 if it is re-enabled for all cpus on this uvhub.
 967  */
 968 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
 969 {
 970         int tcpu;
 971         struct bau_control *tbcp;
 972         struct bau_control *hmaster;
 973
 974         hmaster = bcp->uvhub_master;
 975         spin_lock(&hmaster->disable_lock);
 976         if (bcp->baudisabled && (get_cycles() >= bcp->set_bau_on_time)) {
 977                 stat->s_bau_reenabled++;
 978                 for_each_present_cpu(tcpu) {
 979                         tbcp = &per_cpu(bau_control, tcpu);
 980                         if (tbcp->uvhub_master == hmaster) {
 981                                 tbcp->baudisabled = 0;
 982                                 tbcp->period_requests = 0;
 983                                 tbcp->period_time = 0;
 984                                 tbcp->period_giveups = 0;
 985                         }
 986                 }
 987                 spin_unlock(&hmaster->disable_lock);
 988                 return 0;
 989         }
 990         spin_unlock(&hmaster->disable_lock);
 991         return -1;
 992 }
 993
 994 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
 995                                 int remotes, struct bau_desc *bau_desc)
 996 {
 997         stat->s_requestor++;
 998         stat->s_ntargcpu += remotes + locals;
 999         stat->s_ntargremotes += remotes;
1000         stat->s_ntarglocals += locals;
1001
1002         /* uvhub statistics */
1003         hubs = bau_uvhub_weight(&bau_desc->distribution);
1004         if (locals) {
1005                 stat->s_ntarglocaluvhub++;
1006                 stat->s_ntargremoteuvhub += (hubs - 1);
1007         } else
1008                 stat->s_ntargremoteuvhub += hubs;
1009
1010         stat->s_ntarguvhub += hubs;
1011
1012         if (hubs >= 16)
1013                 stat->s_ntarguvhub16++;
1014         else if (hubs >= 8)
1015                 stat->s_ntarguvhub8++;
1016         else if (hubs >= 4)
1017                 stat->s_ntarguvhub4++;
1018         else if (hubs >= 2)
1019                 stat->s_ntarguvhub2++;
1020         else
1021                 stat->s_ntarguvhub1++;
1022 }
1023
1024 /*
1025  * Translate a cpu mask to the uvhub distribution mask in the BAU
1026  * activation descriptor.
1027  */
1028 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
1029                         struct bau_desc *bau_desc, int *localsp, int *remotesp)
1030 {
1031         int cpu;
1032         int pnode;
1033         int cnt = 0;
1034         struct hub_and_pnode *hpp;
1035
1036         for_each_cpu(cpu, flush_mask) {
1037                 /*
1038                  * The distribution vector is a bit map of pnodes, relative
1039                  * to the partition base pnode (and the partition base nasid
1040                  * in the header).
1041                  * Translate cpu to pnode and hub using a local memory array.
1042                  */
1043                 hpp = &bcp->socket_master->thp[cpu];
1044                 pnode = hpp->pnode - bcp->partition_base_pnode;
1045                 bau_uvhub_set(pnode, &bau_desc->distribution);
1046                 cnt++;
1047                 if (hpp->uvhub == bcp->uvhub)
1048                         (*localsp)++;
1049                 else
1050                         (*remotesp)++;
1051         }
1052         if (!cnt)
1053                 return 1;
1054         return 0;
1055 }
1056
1057 /*
1058  * globally purge translation cache of a virtual address or all TLB's
1059  * @cpumask: mask of all cpu's in which the address is to be removed
1060  * @mm: mm_struct containing virtual address range
1061  * @start: start virtual address to be removed from TLB
1062  * @end: end virtual address to be remove from TLB
1063  * @cpu: the current cpu
1064  *
1065  * This is the entry point for initiating any UV global TLB shootdown.
1066  *
1067  * Purges the translation caches of all specified processors of the given
1068  * virtual address, or purges all TLB's on specified processors.
1069  *
1070  * The caller has derived the cpumask from the mm_struct.  This function
1071  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1072  *
1073  * The cpumask is converted into a uvhubmask of the uvhubs containing
1074  * those cpus.
1075  *
1076  * Note that this function should be called with preemption disabled.
1077  *
1078  * Returns NULL if all remote flushing was done.
1079  * Returns pointer to cpumask if some remote flushing remains to be
1080  * done.  The returned pointer is valid till preemption is re-enabled.
1081  */
1082 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1083                                 struct mm_struct *mm, unsigned long start,
1084                                 unsigned long end, unsigned int cpu)
1085 {
1086         int locals = 0;
1087         int remotes = 0;
1088         int hubs = 0;
1089         struct bau_desc *bau_desc;
1090         struct cpumask *flush_mask;
1091         struct ptc_stats *stat;
1092         struct bau_control *bcp;
1093         unsigned long descriptor_status;
1094         unsigned long status;
1095
1096         bcp = &per_cpu(bau_control, cpu);
1097
1098         if (bcp->nobau)
1099                 return cpumask;
1100
1101         stat = bcp->statp;
1102         stat->s_enters++;
1103
1104         if (bcp->busy) {
1105                 descriptor_status =
1106                         read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1107                 status = ((descriptor_status >> (bcp->uvhub_cpu *
1108                         UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1109                 if (status == UV2H_DESC_BUSY)
1110                         return cpumask;
1111                 bcp->busy = 0;
1112         }
1113
1114         /* bau was disabled due to slow response */
1115         if (bcp->baudisabled) {
1116                 if (check_enable(bcp, stat)) {
1117                         stat->s_ipifordisabled++;
1118                         return cpumask;
1119                 }
1120         }
1121
1122         /*
1123          * Each sending cpu has a per-cpu mask which it fills from the caller's
1124          * cpu mask.  All cpus are converted to uvhubs and copied to the
1125          * activation descriptor.
1126          */
1127         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1128         /* don't actually do a shootdown of the local cpu */
1129         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1130
1131         if (cpu_isset(cpu, *cpumask))
1132                 stat->s_ntargself++;
1133
1134         bau_desc = bcp->descriptor_base;
1135         bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1136         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1137         if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1138                 return NULL;
1139
1140         record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1141
1142         if (!end || (end - start) <= PAGE_SIZE)
1143                 bau_desc->payload.address = start;
1144         else
1145                 bau_desc->payload.address = TLB_FLUSH_ALL;
1146         bau_desc->payload.sending_cpu = cpu;
1147         /*
1148          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1149          * or 1 if it gave up and the original cpumask should be returned.
1150          */
1151         if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1152                 return NULL;
1153         else
1154                 return cpumask;
1155 }
1156
1157 /*
1158  * Search the message queue for any 'other' unprocessed message with the
1159  * same software acknowledge resource bit vector as the 'msg' message.
1160  */
1161 struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1162                                            struct bau_control *bcp)
1163 {
1164         struct bau_pq_entry *msg_next = msg + 1;
1165         unsigned char swack_vec = msg->swack_vec;
1166
1167         if (msg_next > bcp->queue_last)
1168                 msg_next = bcp->queue_first;
1169         while (msg_next != msg) {
1170                 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1171                                 (msg_next->swack_vec == swack_vec))
1172                         return msg_next;
1173                 msg_next++;
1174                 if (msg_next > bcp->queue_last)
1175                         msg_next = bcp->queue_first;
1176         }
1177         return NULL;
1178 }
1179
1180 /*
1181  * UV2 needs to work around a bug in which an arriving message has not
1182  * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1183  * Such a message must be ignored.
1184  */
1185 void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1186 {
1187         unsigned long mmr_image;
1188         unsigned char swack_vec;
1189         struct bau_pq_entry *msg = mdp->msg;
1190         struct bau_pq_entry *other_msg;
1191
1192         mmr_image = read_mmr_sw_ack();
1193         swack_vec = msg->swack_vec;
1194
1195         if ((swack_vec & mmr_image) == 0) {
1196                 /*
1197                  * This message was assigned a swack resource, but no
1198                  * reserved acknowlegment is pending.
1199                  * The bug has prevented this message from setting the MMR.
1200                  */
1201                 /*
1202                  * Some message has set the MMR 'pending' bit; it might have
1203                  * been another message.  Look for that message.
1204                  */
1205                 other_msg = find_another_by_swack(msg, bcp);
1206                 if (other_msg) {
1207                         /*
1208                          * There is another. Process this one but do not
1209                          * ack it.
1210                          */
1211                         bau_process_message(mdp, bcp, 0);
1212                         /*
1213                          * Let the natural processing of that other message
1214                          * acknowledge it. Don't get the processing of sw_ack's
1215                          * out of order.
1216                          */
1217                         return;
1218                 }
1219         }
1220
1221         /*
1222          * Either the MMR shows this one pending a reply or there is no
1223          * other message using this sw_ack, so it is safe to acknowledge it.
1224          */
1225         bau_process_message(mdp, bcp, 1);
1226
1227         return;
1228 }
1229
1230 /*
1231  * The BAU message interrupt comes here. (registered by set_intr_gate)
1232  * See entry_64.S
1233  *
1234  * We received a broadcast assist message.
1235  *
1236  * Interrupts are disabled; this interrupt could represent
1237  * the receipt of several messages.
1238  *
1239  * All cores/threads on this hub get this interrupt.
1240  * The last one to see it does the software ack.
1241  * (the resource will not be freed until noninterruptable cpus see this
1242  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
1243  */
1244 void uv_bau_message_interrupt(struct pt_regs *regs)
1245 {
1246         int count = 0;
1247         cycles_t time_start;
1248         struct bau_pq_entry *msg;
1249         struct bau_control *bcp;
1250         struct ptc_stats *stat;
1251         struct msg_desc msgdesc;
1252
1253         ack_APIC_irq();
1254         time_start = get_cycles();
1255
1256         bcp = &per_cpu(bau_control, smp_processor_id());
1257         stat = bcp->statp;
1258
1259         msgdesc.queue_first = bcp->queue_first;
1260         msgdesc.queue_last = bcp->queue_last;
1261
1262         msg = bcp->bau_msg_head;
1263         while (msg->swack_vec) {
1264                 count++;
1265
1266                 msgdesc.msg_slot = msg - msgdesc.queue_first;
1267                 msgdesc.msg = msg;
1268                 if (bcp->uvhub_version == 2)
1269                         process_uv2_message(&msgdesc, bcp);
1270                 else
1271                         bau_process_message(&msgdesc, bcp, 1);
1272
1273                 msg++;
1274                 if (msg > msgdesc.queue_last)
1275                         msg = msgdesc.queue_first;
1276                 bcp->bau_msg_head = msg;
1277         }
1278         stat->d_time += (get_cycles() - time_start);
1279         if (!count)
1280                 stat->d_nomsg++;
1281         else if (count > 1)
1282                 stat->d_multmsg++;
1283 }
1284
1285 /*
1286  * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1287  * shootdown message timeouts enabled.  The timeout does not cause
1288  * an interrupt, but causes an error message to be returned to
1289  * the sender.
1290  */
1291 static void __init enable_timeouts(void)
1292 {
1293         int uvhub;
1294         int nuvhubs;
1295         int pnode;
1296         unsigned long mmr_image;
1297
1298         nuvhubs = uv_num_possible_blades();
1299
1300         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1301                 if (!uv_blade_nr_possible_cpus(uvhub))
1302                         continue;
1303
1304                 pnode = uv_blade_to_pnode(uvhub);
1305                 mmr_image = read_mmr_misc_control(pnode);
1306                 /*
1307                  * Set the timeout period and then lock it in, in three
1308                  * steps; captures and locks in the period.
1309                  *
1310                  * To program the period, the SOFT_ACK_MODE must be off.
1311                  */
1312                 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1313                 write_mmr_misc_control(pnode, mmr_image);
1314                 /*
1315                  * Set the 4-bit period.
1316                  */
1317                 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1318                 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1319                 write_mmr_misc_control(pnode, mmr_image);
1320                 /*
1321                  * UV1:
1322                  * Subsequent reversals of the timebase bit (3) cause an
1323                  * immediate timeout of one or all INTD resources as
1324                  * indicated in bits 2:0 (7 causes all of them to timeout).
1325                  */
1326                 mmr_image |= (1L << SOFTACK_MSHIFT);
1327                 if (is_uv2_hub()) {
1328                         /* hw bug workaround; do not use extended status */
1329                         mmr_image &= ~(1L << UV2_EXT_SHFT);
1330                 }
1331                 write_mmr_misc_control(pnode, mmr_image);
1332         }
1333 }
1334
1335 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1336 {
1337         if (*offset < num_possible_cpus())
1338                 return offset;
1339         return NULL;
1340 }
1341
1342 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1343 {
1344         (*offset)++;
1345         if (*offset < num_possible_cpus())
1346                 return offset;
1347         return NULL;
1348 }
1349
1350 static void ptc_seq_stop(struct seq_file *file, void *data)
1351 {
1352 }
1353
1354 /*
1355  * Display the statistics thru /proc/sgi_uv/ptc_statistics
1356  * 'data' points to the cpu number
1357  * Note: see the descriptions in stat_description[].
1358  */
1359 static int ptc_seq_show(struct seq_file *file, void *data)
1360 {
1361         struct ptc_stats *stat;
1362         struct bau_control *bcp;
1363         int cpu;
1364
1365         cpu = *(loff_t *)data;
1366         if (!cpu) {
1367                 seq_printf(file,
1368                  "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1369                 seq_printf(file,
1370                         "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1371                 seq_printf(file,
1372                         "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1373                 seq_printf(file,
1374                         "rok resetp resett giveup sto bz throt disable ");
1375                 seq_printf(file,
1376                         "enable wars warshw warwaits enters ipidis plugged ");
1377                 seq_printf(file,
1378                         "ipiover glim cong swack recv rtime all one mult ");
1379                 seq_printf(file,
1380                         "none retry canc nocan reset rcan\n");
1381         }
1382         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1383                 bcp = &per_cpu(bau_control, cpu);
1384                 stat = bcp->statp;
1385                 /* source side statistics */
1386                 seq_printf(file,
1387                         "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1388                            cpu, bcp->nobau, stat->s_requestor,
1389                            cycles_2_us(stat->s_time),
1390                            stat->s_ntargself, stat->s_ntarglocals,
1391                            stat->s_ntargremotes, stat->s_ntargcpu,
1392                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1393                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
1394                 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1395                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1396                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1397                            stat->s_dtimeout, stat->s_strongnacks);
1398                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1399                            stat->s_retry_messages, stat->s_retriesok,
1400                            stat->s_resets_plug, stat->s_resets_timeout,
1401                            stat->s_giveup, stat->s_stimeout,
1402                            stat->s_busy, stat->s_throttles);
1403                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1404                            stat->s_bau_disabled, stat->s_bau_reenabled,
1405                            stat->s_uv2_wars, stat->s_uv2_wars_hw,
1406                            stat->s_uv2_war_waits, stat->s_enters,
1407                            stat->s_ipifordisabled, stat->s_plugged,
1408                            stat->s_overipilimit, stat->s_giveuplimit,
1409                            stat->s_congested);
1410
1411                 /* destination side statistics */
1412                 seq_printf(file,
1413                         "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1414                            read_gmmr_sw_ack(uv_cpu_to_pnode(cpu)),
1415                            stat->d_requestee, cycles_2_us(stat->d_time),
1416                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1417                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
1418                            stat->d_nocanceled, stat->d_resets,
1419                            stat->d_rcanceled);
1420         }
1421         return 0;
1422 }
1423
1424 /*
1425  * Display the tunables thru debugfs
1426  */
1427 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1428                                 size_t count, loff_t *ppos)
1429 {
1430         char *buf;
1431         int ret;
1432
1433         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1434                 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1435                 "ipi_reset_limit complete_threshold congested_response_us",
1436                 "congested_reps disabled_period giveup_limit",
1437                 max_concurr, plugged_delay, plugsb4reset,
1438                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1439                 congested_respns_us, congested_reps, disabled_period,
1440                 giveup_limit);
1441
1442         if (!buf)
1443                 return -ENOMEM;
1444
1445         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1446         kfree(buf);
1447         return ret;
1448 }
1449
1450 /*
1451  * handle a write to /proc/sgi_uv/ptc_statistics
1452  * -1: reset the statistics
1453  *  0: display meaning of the statistics
1454  */
1455 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1456                                 size_t count, loff_t *data)
1457 {
1458         int cpu;
1459         int i;
1460         int elements;
1461         long input_arg;
1462         char optstr[64];
1463         struct ptc_stats *stat;
1464
1465         if (count == 0 || count > sizeof(optstr))
1466                 return -EINVAL;
1467         if (copy_from_user(optstr, user, count))
1468                 return -EFAULT;
1469         optstr[count - 1] = '\0';
1470
1471         if (!strcmp(optstr, "on")) {
1472                 set_bau_on();
1473                 return count;
1474         } else if (!strcmp(optstr, "off")) {
1475                 set_bau_off();
1476                 return count;
1477         }
1478
1479         if (strict_strtol(optstr, 10, &input_arg) < 0) {
1480                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1481                 return -EINVAL;
1482         }
1483
1484         if (input_arg == 0) {
1485                 elements = ARRAY_SIZE(stat_description);
1486                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1487                 printk(KERN_DEBUG "Sender statistics:\n");
1488                 for (i = 0; i < elements; i++)
1489                         printk(KERN_DEBUG "%s\n", stat_description[i]);
1490         } else if (input_arg == -1) {
1491                 for_each_present_cpu(cpu) {
1492                         stat = &per_cpu(ptcstats, cpu);
1493                         memset(stat, 0, sizeof(struct ptc_stats));
1494                 }
1495         }
1496
1497         return count;
1498 }
1499
1500 static int local_atoi(const char *name)
1501 {
1502         int val = 0;
1503
1504         for (;; name++) {
1505                 switch (*name) {
1506                 case '0' ... '9':
1507                         val = 10*val+(*name-'0');
1508                         break;
1509                 default:
1510                         return val;
1511                 }
1512         }
1513 }
1514
1515 /*
1516  * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1517  * Zero values reset them to defaults.
1518  */
1519 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1520                                 int count)
1521 {
1522         char *p;
1523         char *q;
1524         int cnt = 0;
1525         int val;
1526         int e = ARRAY_SIZE(tunables);
1527
1528         p = instr + strspn(instr, WHITESPACE);
1529         q = p;
1530         for (; *p; p = q + strspn(q, WHITESPACE)) {
1531                 q = p + strcspn(p, WHITESPACE);
1532                 cnt++;
1533                 if (q == p)
1534                         break;
1535         }
1536         if (cnt != e) {
1537                 printk(KERN_INFO "bau tunable error: should be %d values\n", e);
1538                 return -EINVAL;
1539         }
1540
1541         p = instr + strspn(instr, WHITESPACE);
1542         q = p;
1543         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1544                 q = p + strcspn(p, WHITESPACE);
1545                 val = local_atoi(p);
1546                 switch (cnt) {
1547                 case 0:
1548                         if (val == 0) {
1549                                 max_concurr = MAX_BAU_CONCURRENT;
1550                                 max_concurr_const = MAX_BAU_CONCURRENT;
1551                                 continue;
1552                         }
1553                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1554                                 printk(KERN_DEBUG
1555                                 "Error: BAU max concurrent %d is invalid\n",
1556                                 val);
1557                                 return -EINVAL;
1558                         }
1559                         max_concurr = val;
1560                         max_concurr_const = val;
1561                         continue;
1562                 default:
1563                         if (val == 0)
1564                                 *tunables[cnt].tunp = tunables[cnt].deflt;
1565                         else
1566                                 *tunables[cnt].tunp = val;
1567                         continue;
1568                 }
1569                 if (q == p)
1570                         break;
1571         }
1572         return 0;
1573 }
1574
1575 /*
1576  * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1577  */
1578 static ssize_t tunables_write(struct file *file, const char __user *user,
1579                                 size_t count, loff_t *data)
1580 {
1581         int cpu;
1582         int ret;
1583         char instr[100];
1584         struct bau_control *bcp;
1585
1586         if (count == 0 || count > sizeof(instr)-1)
1587                 return -EINVAL;
1588         if (copy_from_user(instr, user, count))
1589                 return -EFAULT;
1590
1591         instr[count] = '\0';
1592
1593         cpu = get_cpu();
1594         bcp = &per_cpu(bau_control, cpu);
1595         ret = parse_tunables_write(bcp, instr, count);
1596         put_cpu();
1597         if (ret)
1598                 return ret;
1599
1600         for_each_present_cpu(cpu) {
1601                 bcp = &per_cpu(bau_control, cpu);
1602                 bcp->max_concurr =              max_concurr;
1603                 bcp->max_concurr_const =        max_concurr;
1604                 bcp->plugged_delay =            plugged_delay;
1605                 bcp->plugsb4reset =             plugsb4reset;
1606                 bcp->timeoutsb4reset =          timeoutsb4reset;
1607                 bcp->ipi_reset_limit =          ipi_reset_limit;
1608                 bcp->complete_threshold =       complete_threshold;
1609                 bcp->cong_response_us =         congested_respns_us;
1610                 bcp->cong_reps =                congested_reps;
1611                 bcp->disabled_period =          sec_2_cycles(disabled_period);
1612                 bcp->giveup_limit =             giveup_limit;
1613         }
1614         return count;
1615 }
1616
1617 static const struct seq_operations uv_ptc_seq_ops = {
1618         .start          = ptc_seq_start,
1619         .next           = ptc_seq_next,
1620         .stop           = ptc_seq_stop,
1621         .show           = ptc_seq_show
1622 };
1623
1624 static int ptc_proc_open(struct inode *inode, struct file *file)
1625 {
1626         return seq_open(file, &uv_ptc_seq_ops);
1627 }
1628
1629 static int tunables_open(struct inode *inode, struct file *file)
1630 {
1631         return 0;
1632 }
1633
1634 static const struct file_operations proc_uv_ptc_operations = {
1635         .open           = ptc_proc_open,
1636         .read           = seq_read,
1637         .write          = ptc_proc_write,
1638         .llseek         = seq_lseek,
1639         .release        = seq_release,
1640 };
1641
1642 static const struct file_operations tunables_fops = {
1643         .open           = tunables_open,
1644         .read           = tunables_read,
1645         .write          = tunables_write,
1646         .llseek         = default_llseek,
1647 };
1648
1649 static int __init uv_ptc_init(void)
1650 {
1651         struct proc_dir_entry *proc_uv_ptc;
1652
1653         if (!is_uv_system())
1654                 return 0;
1655
1656         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1657                                   &proc_uv_ptc_operations);
1658         if (!proc_uv_ptc) {
1659                 printk(KERN_ERR "unable to create %s proc entry\n",
1660                        UV_PTC_BASENAME);
1661                 return -EINVAL;
1662         }
1663
1664         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1665         if (!tunables_dir) {
1666                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1667                        UV_BAU_TUNABLES_DIR);
1668                 return -EINVAL;
1669         }
1670         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1671                                         tunables_dir, NULL, &tunables_fops);
1672         if (!tunables_file) {
1673                 printk(KERN_ERR "unable to create debugfs file %s\n",
1674                        UV_BAU_TUNABLES_FILE);
1675                 return -EINVAL;
1676         }
1677         return 0;
1678 }
1679
1680 /*
1681  * Initialize the sending side's sending buffers.
1682  */
1683 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1684 {
1685         int i;
1686         int cpu;
1687         int uv1 = 0;
1688         unsigned long gpa;
1689         unsigned long m;
1690         unsigned long n;
1691         size_t dsize;
1692         struct bau_desc *bau_desc;
1693         struct bau_desc *bd2;
1694         struct uv1_bau_msg_header *uv1_hdr;
1695         struct uv2_bau_msg_header *uv2_hdr;
1696         struct bau_control *bcp;
1697
1698         /*
1699          * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1700          * per cpu; and one per cpu on the uvhub (ADP_SZ)
1701          */
1702         dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1703         bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1704         BUG_ON(!bau_desc);
1705
1706         gpa = uv_gpa(bau_desc);
1707         n = uv_gpa_to_gnode(gpa);
1708         m = uv_gpa_to_offset(gpa);
1709         if (is_uv1_hub())
1710                 uv1 = 1;
1711
1712         /* the 14-bit pnode */
1713         write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1714         /*
1715          * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1716          * cpu even though we only use the first one; one descriptor can
1717          * describe a broadcast to 256 uv hubs.
1718          */
1719         for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1720                 memset(bd2, 0, sizeof(struct bau_desc));
1721                 if (uv1) {
1722                         uv1_hdr = &bd2->header.uv1_hdr;
1723                         uv1_hdr->swack_flag =   1;
1724                         /*
1725                          * The base_dest_nasid set in the message header
1726                          * is the nasid of the first uvhub in the partition.
1727                          * The bit map will indicate destination pnode numbers
1728                          * relative to that base. They may not be consecutive
1729                          * if nasid striding is being used.
1730                          */
1731                         uv1_hdr->base_dest_nasid =
1732                                                 UV_PNODE_TO_NASID(base_pnode);
1733                         uv1_hdr->dest_subnodeid =       UV_LB_SUBNODEID;
1734                         uv1_hdr->command =              UV_NET_ENDPOINT_INTD;
1735                         uv1_hdr->int_both =             1;
1736                         /*
1737                          * all others need to be set to zero:
1738                          *   fairness chaining multilevel count replied_to
1739                          */
1740                 } else {
1741                         /*
1742                          * BIOS uses legacy mode, but UV2 hardware always
1743                          * uses native mode for selective broadcasts.
1744                          */
1745                         uv2_hdr = &bd2->header.uv2_hdr;
1746                         uv2_hdr->swack_flag =   1;
1747                         uv2_hdr->base_dest_nasid =
1748                                                 UV_PNODE_TO_NASID(base_pnode);
1749                         uv2_hdr->dest_subnodeid =       UV_LB_SUBNODEID;
1750                         uv2_hdr->command =              UV_NET_ENDPOINT_INTD;
1751                 }
1752         }
1753         for_each_present_cpu(cpu) {
1754                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1755                         continue;
1756                 bcp = &per_cpu(bau_control, cpu);
1757                 bcp->descriptor_base = bau_desc;
1758         }
1759 }
1760
1761 /*
1762  * initialize the destination side's receiving buffers
1763  * entered for each uvhub in the partition
1764  * - node is first node (kernel memory notion) on the uvhub
1765  * - pnode is the uvhub's physical identifier
1766  */
1767 static void pq_init(int node, int pnode)
1768 {
1769         int cpu;
1770         size_t plsize;
1771         char *cp;
1772         void *vp;
1773         unsigned long pn;
1774         unsigned long first;
1775         unsigned long pn_first;
1776         unsigned long last;
1777         struct bau_pq_entry *pqp;
1778         struct bau_control *bcp;
1779
1780         plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1781         vp = kmalloc_node(plsize, GFP_KERNEL, node);
1782         pqp = (struct bau_pq_entry *)vp;
1783         BUG_ON(!pqp);
1784
1785         cp = (char *)pqp + 31;
1786         pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1787
1788         for_each_present_cpu(cpu) {
1789                 if (pnode != uv_cpu_to_pnode(cpu))
1790                         continue;
1791                 /* for every cpu on this pnode: */
1792                 bcp = &per_cpu(bau_control, cpu);
1793                 bcp->queue_first        = pqp;
1794                 bcp->bau_msg_head       = pqp;
1795                 bcp->queue_last         = pqp + (DEST_Q_SIZE - 1);
1796         }
1797         /*
1798          * need the gnode of where the memory was really allocated
1799          */
1800         pn = uv_gpa_to_gnode(uv_gpa(pqp));
1801         first = uv_physnodeaddr(pqp);
1802         pn_first = ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | first;
1803         last = uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1));
1804         write_mmr_payload_first(pnode, pn_first);
1805         write_mmr_payload_tail(pnode, first);
1806         write_mmr_payload_last(pnode, last);
1807         write_gmmr_sw_ack(pnode, 0xffffUL);
1808
1809         /* in effect, all msg_type's are set to MSG_NOOP */
1810         memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1811 }
1812
1813 /*
1814  * Initialization of each UV hub's structures
1815  */
1816 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1817 {
1818         int node;
1819         int pnode;
1820         unsigned long apicid;
1821
1822         node = uvhub_to_first_node(uvhub);
1823         pnode = uv_blade_to_pnode(uvhub);
1824
1825         activation_descriptor_init(node, pnode, base_pnode);
1826
1827         pq_init(node, pnode);
1828         /*
1829          * The below initialization can't be in firmware because the
1830          * messaging IRQ will be determined by the OS.
1831          */
1832         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1833         write_mmr_data_config(pnode, ((apicid << 32) | vector));
1834 }
1835
1836 /*
1837  * We will set BAU_MISC_CONTROL with a timeout period.
1838  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1839  * So the destination timeout period has to be calculated from them.
1840  */
1841 static int calculate_destination_timeout(void)
1842 {
1843         unsigned long mmr_image;
1844         int mult1;
1845         int mult2;
1846         int index;
1847         int base;
1848         int ret;
1849         unsigned long ts_ns;
1850
1851         if (is_uv1_hub()) {
1852                 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1853                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1854                 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1855                 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1856                 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1857                 ts_ns = timeout_base_ns[index];
1858                 ts_ns *= (mult1 * mult2);
1859                 ret = ts_ns / 1000;
1860         } else {
1861                 /* 4 bits  0/1 for 10/80us base, 3 bits of multiplier */
1862                 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1863                 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1864                 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1865                         base = 80;
1866                 else
1867                         base = 10;
1868                 mult1 = mmr_image & UV2_ACK_MASK;
1869                 ret = mult1 * base;
1870         }
1871         return ret;
1872 }
1873
1874 static void __init init_per_cpu_tunables(void)
1875 {
1876         int cpu;
1877         struct bau_control *bcp;
1878
1879         for_each_present_cpu(cpu) {
1880                 bcp = &per_cpu(bau_control, cpu);
1881                 bcp->baudisabled                = 0;
1882                 if (nobau)
1883                         bcp->nobau              = 1;
1884                 bcp->statp                      = &per_cpu(ptcstats, cpu);
1885                 /* time interval to catch a hardware stay-busy bug */
1886                 bcp->timeout_interval           = usec_2_cycles(2*timeout_us);
1887                 bcp->max_concurr                = max_concurr;
1888                 bcp->max_concurr_const          = max_concurr;
1889                 bcp->plugged_delay              = plugged_delay;
1890                 bcp->plugsb4reset               = plugsb4reset;
1891                 bcp->timeoutsb4reset            = timeoutsb4reset;
1892                 bcp->ipi_reset_limit            = ipi_reset_limit;
1893                 bcp->complete_threshold         = complete_threshold;
1894                 bcp->cong_response_us           = congested_respns_us;
1895                 bcp->cong_reps                  = congested_reps;
1896                 bcp->disabled_period =          sec_2_cycles(disabled_period);
1897                 bcp->giveup_limit =             giveup_limit;
1898                 spin_lock_init(&bcp->queue_lock);
1899                 spin_lock_init(&bcp->uvhub_lock);
1900                 spin_lock_init(&bcp->disable_lock);
1901         }
1902 }
1903
1904 /*
1905  * Scan all cpus to collect blade and socket summaries.
1906  */
1907 static int __init get_cpu_topology(int base_pnode,
1908                                         struct uvhub_desc *uvhub_descs,
1909                                         unsigned char *uvhub_mask)
1910 {
1911         int cpu;
1912         int pnode;
1913         int uvhub;
1914         int socket;
1915         struct bau_control *bcp;
1916         struct uvhub_desc *bdp;
1917         struct socket_desc *sdp;
1918
1919         for_each_present_cpu(cpu) {
1920                 bcp = &per_cpu(bau_control, cpu);
1921
1922                 memset(bcp, 0, sizeof(struct bau_control));
1923
1924                 pnode = uv_cpu_hub_info(cpu)->pnode;
1925                 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1926                         printk(KERN_EMERG
1927                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1928                                 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1929                         return 1;
1930                 }
1931
1932                 bcp->osnode = cpu_to_node(cpu);
1933                 bcp->partition_base_pnode = base_pnode;
1934
1935                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1936                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1937                 bdp = &uvhub_descs[uvhub];
1938
1939                 bdp->num_cpus++;
1940                 bdp->uvhub = uvhub;
1941                 bdp->pnode = pnode;
1942
1943                 /* kludge: 'assuming' one node per socket, and assuming that
1944                    disabling a socket just leaves a gap in node numbers */
1945                 socket = bcp->osnode & 1;
1946                 bdp->socket_mask |= (1 << socket);
1947                 sdp = &bdp->socket[socket];
1948                 sdp->cpu_number[sdp->num_cpus] = cpu;
1949                 sdp->num_cpus++;
1950                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1951                         printk(KERN_EMERG "%d cpus per socket invalid\n",
1952                                 sdp->num_cpus);
1953                         return 1;
1954                 }
1955         }
1956         return 0;
1957 }
1958
1959 /*
1960  * Each socket is to get a local array of pnodes/hubs.
1961  */
1962 static void make_per_cpu_thp(struct bau_control *smaster)
1963 {
1964         int cpu;
1965         size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
1966
1967         smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
1968         memset(smaster->thp, 0, hpsz);
1969         for_each_present_cpu(cpu) {
1970                 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
1971                 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1972         }
1973 }
1974
1975 /*
1976  * Each uvhub is to get a local cpumask.
1977  */
1978 static void make_per_hub_cpumask(struct bau_control *hmaster)
1979 {
1980         int sz = sizeof(cpumask_t);
1981
1982         hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
1983 }
1984
1985 /*
1986  * Initialize all the per_cpu information for the cpu's on a given socket,
1987  * given what has been gathered into the socket_desc struct.
1988  * And reports the chosen hub and socket masters back to the caller.
1989  */
1990 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
1991                         struct bau_control **smasterp,
1992                         struct bau_control **hmasterp)
1993 {
1994         int i;
1995         int cpu;
1996         struct bau_control *bcp;
1997
1998         for (i = 0; i < sdp->num_cpus; i++) {
1999                 cpu = sdp->cpu_number[i];
2000                 bcp = &per_cpu(bau_control, cpu);
2001                 bcp->cpu = cpu;
2002                 if (i == 0) {
2003                         *smasterp = bcp;
2004                         if (!(*hmasterp))
2005                                 *hmasterp = bcp;
2006                 }
2007                 bcp->cpus_in_uvhub = bdp->num_cpus;
2008                 bcp->cpus_in_socket = sdp->num_cpus;
2009                 bcp->socket_master = *smasterp;
2010                 bcp->uvhub = bdp->uvhub;
2011                 if (is_uv1_hub())
2012                         bcp->uvhub_version = 1;
2013                 else if (is_uv2_hub())
2014                         bcp->uvhub_version = 2;
2015                 else {
2016                         printk(KERN_EMERG "uvhub version not 1 or 2\n");
2017                         return 1;
2018                 }
2019                 bcp->uvhub_master = *hmasterp;
2020                 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->blade_processor_id;
2021                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2022                         printk(KERN_EMERG "%d cpus per uvhub invalid\n",
2023                                 bcp->uvhub_cpu);
2024                         return 1;
2025                 }
2026         }
2027         return 0;
2028 }
2029
2030 /*
2031  * Summarize the blade and socket topology into the per_cpu structures.
2032  */
2033 static int __init summarize_uvhub_sockets(int nuvhubs,
2034                         struct uvhub_desc *uvhub_descs,
2035                         unsigned char *uvhub_mask)
2036 {
2037         int socket;
2038         int uvhub;
2039         unsigned short socket_mask;
2040
2041         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2042                 struct uvhub_desc *bdp;
2043                 struct bau_control *smaster = NULL;
2044                 struct bau_control *hmaster = NULL;
2045
2046                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2047                         continue;
2048
2049                 bdp = &uvhub_descs[uvhub];
2050                 socket_mask = bdp->socket_mask;
2051                 socket = 0;
2052                 while (socket_mask) {
2053                         struct socket_desc *sdp;
2054                         if ((socket_mask & 1)) {
2055                                 sdp = &bdp->socket[socket];
2056                                 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2057                                         return 1;
2058                                 make_per_cpu_thp(smaster);
2059                         }
2060                         socket++;
2061                         socket_mask = (socket_mask >> 1);
2062                 }
2063                 make_per_hub_cpumask(hmaster);
2064         }
2065         return 0;
2066 }
2067
2068 /*
2069  * initialize the bau_control structure for each cpu
2070  */
2071 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2072 {
2073         unsigned char *uvhub_mask;
2074         void *vp;
2075         struct uvhub_desc *uvhub_descs;
2076
2077         timeout_us = calculate_destination_timeout();
2078
2079         vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
2080         uvhub_descs = (struct uvhub_desc *)vp;
2081         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
2082         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2083
2084         if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2085                 goto fail;
2086
2087         if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2088                 goto fail;
2089
2090         kfree(uvhub_descs);
2091         kfree(uvhub_mask);
2092         init_per_cpu_tunables();
2093         return 0;
2094
2095 fail:
2096         kfree(uvhub_descs);
2097         kfree(uvhub_mask);
2098         return 1;
2099 }
2100
2101 /*
2102  * Initialization of BAU-related structures
2103  */
2104 static int __init uv_bau_init(void)
2105 {
2106         int uvhub;
2107         int pnode;
2108         int nuvhubs;
2109         int cur_cpu;
2110         int cpus;
2111         int vector;
2112         cpumask_var_t *mask;
2113
2114         if (!is_uv_system())
2115                 return 0;
2116
2117         for_each_possible_cpu(cur_cpu) {
2118                 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2119                 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2120         }
2121
2122         nuvhubs = uv_num_possible_blades();
2123         congested_cycles = usec_2_cycles(congested_respns_us);
2124
2125         uv_base_pnode = 0x7fffffff;
2126         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2127                 cpus = uv_blade_nr_possible_cpus(uvhub);
2128                 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2129                         uv_base_pnode = uv_blade_to_pnode(uvhub);
2130         }
2131
2132         enable_timeouts();
2133
2134         if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2135                 set_bau_off();
2136                 nobau_perm = 1;
2137                 return 0;
2138         }
2139
2140         vector = UV_BAU_MESSAGE;
2141         for_each_possible_blade(uvhub)
2142                 if (uv_blade_nr_possible_cpus(uvhub))
2143                         init_uvhub(uvhub, vector, uv_base_pnode);
2144
2145         alloc_intr_gate(vector, uv_bau_message_intr1);
2146
2147         for_each_possible_blade(uvhub) {
2148                 if (uv_blade_nr_possible_cpus(uvhub)) {
2149                         unsigned long val;
2150                         unsigned long mmr;
2151                         pnode = uv_blade_to_pnode(uvhub);
2152                         /* INIT the bau */
2153                         val = 1L << 63;
2154                         write_gmmr_activation(pnode, val);
2155                         mmr = 1; /* should be 1 to broadcast to both sockets */
2156                         if (!is_uv1_hub())
2157                                 write_mmr_data_broadcast(pnode, mmr);
2158                 }
2159         }
2160
2161         return 0;
2162 }
2163 core_initcall(uv_bau_init);
2164 fs_initcall(uv_ptc_init);