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/tsc.h>
  23 #include <asm/irq_vectors.h>
  24 #include <asm/timer.h>
  25
  26 static struct bau_operations ops;
  27
  28 static struct bau_operations uv123_bau_ops = {
  29         .bau_gpa_to_offset       = uv_gpa_to_offset,
  30         .read_l_sw_ack           = read_mmr_sw_ack,
  31         .read_g_sw_ack           = read_gmmr_sw_ack,
  32         .write_l_sw_ack          = write_mmr_sw_ack,
  33         .write_g_sw_ack          = write_gmmr_sw_ack,
  34         .write_payload_first     = write_mmr_payload_first,
  35         .write_payload_last      = write_mmr_payload_last,
  36 };
  37
  38 static struct bau_operations uv4_bau_ops = {
  39         .bau_gpa_to_offset       = uv_gpa_to_soc_phys_ram,
  40         .read_l_sw_ack           = read_mmr_proc_sw_ack,
  41         .read_g_sw_ack           = read_gmmr_proc_sw_ack,
  42         .write_l_sw_ack          = write_mmr_proc_sw_ack,
  43         .write_g_sw_ack          = write_gmmr_proc_sw_ack,
  44         .write_payload_first     = write_mmr_proc_payload_first,
  45         .write_payload_last      = write_mmr_proc_payload_last,
  46 };
  47
  48
  49 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
  50 static int timeout_base_ns[] = {
  51                 20,
  52                 160,
  53                 1280,
  54                 10240,
  55                 81920,
  56                 655360,
  57                 5242880,
  58                 167772160
  59 };
  60
  61 static int timeout_us;
  62 static bool nobau = true;
  63 static int nobau_perm;
  64 static cycles_t congested_cycles;
  65
  66 /* tunables: */
  67 static int max_concurr          = MAX_BAU_CONCURRENT;
  68 static int max_concurr_const    = MAX_BAU_CONCURRENT;
  69 static int plugged_delay        = PLUGGED_DELAY;
  70 static int plugsb4reset         = PLUGSB4RESET;
  71 static int giveup_limit         = GIVEUP_LIMIT;
  72 static int timeoutsb4reset      = TIMEOUTSB4RESET;
  73 static int ipi_reset_limit      = IPI_RESET_LIMIT;
  74 static int complete_threshold   = COMPLETE_THRESHOLD;
  75 static int congested_respns_us  = CONGESTED_RESPONSE_US;
  76 static int congested_reps       = CONGESTED_REPS;
  77 static int disabled_period      = DISABLED_PERIOD;
  78
  79 static struct tunables tunables[] = {
  80         {&max_concurr,           MAX_BAU_CONCURRENT}, /* must be [0] */
  81         {&plugged_delay,         PLUGGED_DELAY},
  82         {&plugsb4reset,          PLUGSB4RESET},
  83         {&timeoutsb4reset,       TIMEOUTSB4RESET},
  84         {&ipi_reset_limit,       IPI_RESET_LIMIT},
  85         {&complete_threshold,    COMPLETE_THRESHOLD},
  86         {&congested_respns_us,   CONGESTED_RESPONSE_US},
  87         {&congested_reps,        CONGESTED_REPS},
  88         {&disabled_period,       DISABLED_PERIOD},
  89         {&giveup_limit,          GIVEUP_LIMIT}
  90 };
  91
  92 static struct dentry *tunables_dir;
  93 static struct dentry *tunables_file;
  94
  95 /* these correspond to the statistics printed by ptc_seq_show() */
  96 static char *stat_description[] = {
  97         "sent:     number of shootdown messages sent",
  98         "stime:    time spent sending messages",
  99         "numuvhubs: number of hubs targeted with shootdown",
 100         "numuvhubs16: number times 16 or more hubs targeted",
 101         "numuvhubs8: number times 8 or more hubs targeted",
 102         "numuvhubs4: number times 4 or more hubs targeted",
 103         "numuvhubs2: number times 2 or more hubs targeted",
 104         "numuvhubs1: number times 1 hub targeted",
 105         "numcpus:  number of cpus targeted with shootdown",
 106         "dto:      number of destination timeouts",
 107         "retries:  destination timeout retries sent",
 108         "rok:   :  destination timeouts successfully retried",
 109         "resetp:   ipi-style resource resets for plugs",
 110         "resett:   ipi-style resource resets for timeouts",
 111         "giveup:   fall-backs to ipi-style shootdowns",
 112         "sto:      number of source timeouts",
 113         "bz:       number of stay-busy's",
 114         "throt:    number times spun in throttle",
 115         "swack:   image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE",
 116         "recv:     shootdown messages received",
 117         "rtime:    time spent processing messages",
 118         "all:      shootdown all-tlb messages",
 119         "one:      shootdown one-tlb messages",
 120         "mult:     interrupts that found multiple messages",
 121         "none:     interrupts that found no messages",
 122         "retry:    number of retry messages processed",
 123         "canc:     number messages canceled by retries",
 124         "nocan:    number retries that found nothing to cancel",
 125         "reset:    number of ipi-style reset requests processed",
 126         "rcan:     number messages canceled by reset requests",
 127         "disable:  number times use of the BAU was disabled",
 128         "enable:   number times use of the BAU was re-enabled"
 129 };
 130
 131 static int __init setup_bau(char *arg)
 132 {
 133         int result;
 134
 135         if (!arg)
 136                 return -EINVAL;
 137
 138         result = strtobool(arg, &nobau);
 139         if (result)
 140                 return result;
 141
 142         /* we need to flip the logic here, so that bau=y sets nobau to false */
 143         nobau = !nobau;
 144
 145         if (!nobau)
 146                 pr_info("UV BAU Enabled\n");
 147         else
 148                 pr_info("UV BAU Disabled\n");
 149
 150         return 0;
 151 }
 152 early_param("bau", setup_bau);
 153
 154 /* base pnode in this partition */
 155 static int uv_base_pnode __read_mostly;
 156
 157 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
 158 static DEFINE_PER_CPU(struct bau_control, bau_control);
 159 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
 160
 161 static void
 162 set_bau_on(void)
 163 {
 164         int cpu;
 165         struct bau_control *bcp;
 166
 167         if (nobau_perm) {
 168                 pr_info("BAU not initialized; cannot be turned on\n");
 169                 return;
 170         }
 171         nobau = false;
 172         for_each_present_cpu(cpu) {
 173                 bcp = &per_cpu(bau_control, cpu);
 174                 bcp->nobau = false;
 175         }
 176         pr_info("BAU turned on\n");
 177         return;
 178 }
 179
 180 static void
 181 set_bau_off(void)
 182 {
 183         int cpu;
 184         struct bau_control *bcp;
 185
 186         nobau = true;
 187         for_each_present_cpu(cpu) {
 188                 bcp = &per_cpu(bau_control, cpu);
 189                 bcp->nobau = true;
 190         }
 191         pr_info("BAU turned off\n");
 192         return;
 193 }
 194
 195 /*
 196  * Determine the first node on a uvhub. 'Nodes' are used for kernel
 197  * memory allocation.
 198  */
 199 static int __init uvhub_to_first_node(int uvhub)
 200 {
 201         int node, b;
 202
 203         for_each_online_node(node) {
 204                 b = uv_node_to_blade_id(node);
 205                 if (uvhub == b)
 206                         return node;
 207         }
 208         return -1;
 209 }
 210
 211 /*
 212  * Determine the apicid of the first cpu on a uvhub.
 213  */
 214 static int __init uvhub_to_first_apicid(int uvhub)
 215 {
 216         int cpu;
 217
 218         for_each_present_cpu(cpu)
 219                 if (uvhub == uv_cpu_to_blade_id(cpu))
 220                         return per_cpu(x86_cpu_to_apicid, cpu);
 221         return -1;
 222 }
 223
 224 /*
 225  * Free a software acknowledge hardware resource by clearing its Pending
 226  * bit. This will return a reply to the sender.
 227  * If the message has timed out, a reply has already been sent by the
 228  * hardware but the resource has not been released. In that case our
 229  * clear of the Timeout bit (as well) will free the resource. No reply will
 230  * be sent (the hardware will only do one reply per message).
 231  */
 232 static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp,
 233                                                 int do_acknowledge)
 234 {
 235         unsigned long dw;
 236         struct bau_pq_entry *msg;
 237
 238         msg = mdp->msg;
 239         if (!msg->canceled && do_acknowledge) {
 240                 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
 241                 ops.write_l_sw_ack(dw);
 242         }
 243         msg->replied_to = 1;
 244         msg->swack_vec = 0;
 245 }
 246
 247 /*
 248  * Process the receipt of a RETRY message
 249  */
 250 static void bau_process_retry_msg(struct msg_desc *mdp,
 251                                         struct bau_control *bcp)
 252 {
 253         int i;
 254         int cancel_count = 0;
 255         unsigned long msg_res;
 256         unsigned long mmr = 0;
 257         struct bau_pq_entry *msg = mdp->msg;
 258         struct bau_pq_entry *msg2;
 259         struct ptc_stats *stat = bcp->statp;
 260
 261         stat->d_retries++;
 262         /*
 263          * cancel any message from msg+1 to the retry itself
 264          */
 265         for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
 266                 if (msg2 > mdp->queue_last)
 267                         msg2 = mdp->queue_first;
 268                 if (msg2 == msg)
 269                         break;
 270
 271                 /* same conditions for cancellation as do_reset */
 272                 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
 273                     (msg2->swack_vec) && ((msg2->swack_vec &
 274                         msg->swack_vec) == 0) &&
 275                     (msg2->sending_cpu == msg->sending_cpu) &&
 276                     (msg2->msg_type != MSG_NOOP)) {
 277                         mmr = ops.read_l_sw_ack();
 278                         msg_res = msg2->swack_vec;
 279                         /*
 280                          * This is a message retry; clear the resources held
 281                          * by the previous message only if they timed out.
 282                          * If it has not timed out we have an unexpected
 283                          * situation to report.
 284                          */
 285                         if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
 286                                 unsigned long mr;
 287                                 /*
 288                                  * Is the resource timed out?
 289                                  * Make everyone ignore the cancelled message.
 290                                  */
 291                                 msg2->canceled = 1;
 292                                 stat->d_canceled++;
 293                                 cancel_count++;
 294                                 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
 295                                 ops.write_l_sw_ack(mr);
 296                         }
 297                 }
 298         }
 299         if (!cancel_count)
 300                 stat->d_nocanceled++;
 301 }
 302
 303 /*
 304  * Do all the things a cpu should do for a TLB shootdown message.
 305  * Other cpu's may come here at the same time for this message.
 306  */
 307 static void bau_process_message(struct msg_desc *mdp, struct bau_control *bcp,
 308                                                 int do_acknowledge)
 309 {
 310         short socket_ack_count = 0;
 311         short *sp;
 312         struct atomic_short *asp;
 313         struct ptc_stats *stat = bcp->statp;
 314         struct bau_pq_entry *msg = mdp->msg;
 315         struct bau_control *smaster = bcp->socket_master;
 316
 317         /*
 318          * This must be a normal message, or retry of a normal message
 319          */
 320         if (msg->address == TLB_FLUSH_ALL) {
 321                 local_flush_tlb();
 322                 stat->d_alltlb++;
 323         } else {
 324                 __flush_tlb_one(msg->address);
 325                 stat->d_onetlb++;
 326         }
 327         stat->d_requestee++;
 328
 329         /*
 330          * One cpu on each uvhub has the additional job on a RETRY
 331          * of releasing the resource held by the message that is
 332          * being retried.  That message is identified by sending
 333          * cpu number.
 334          */
 335         if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
 336                 bau_process_retry_msg(mdp, bcp);
 337
 338         /*
 339          * This is a swack message, so we have to reply to it.
 340          * Count each responding cpu on the socket. This avoids
 341          * pinging the count's cache line back and forth between
 342          * the sockets.
 343          */
 344         sp = &smaster->socket_acknowledge_count[mdp->msg_slot];
 345         asp = (struct atomic_short *)sp;
 346         socket_ack_count = atom_asr(1, asp);
 347         if (socket_ack_count == bcp->cpus_in_socket) {
 348                 int msg_ack_count;
 349                 /*
 350                  * Both sockets dump their completed count total into
 351                  * the message's count.
 352                  */
 353                 *sp = 0;
 354                 asp = (struct atomic_short *)&msg->acknowledge_count;
 355                 msg_ack_count = atom_asr(socket_ack_count, asp);
 356
 357                 if (msg_ack_count == bcp->cpus_in_uvhub) {
 358                         /*
 359                          * All cpus in uvhub saw it; reply
 360                          * (unless we are in the UV2 workaround)
 361                          */
 362                         reply_to_message(mdp, bcp, do_acknowledge);
 363                 }
 364         }
 365
 366         return;
 367 }
 368
 369 /*
 370  * Determine the first cpu on a pnode.
 371  */
 372 static int pnode_to_first_cpu(int pnode, struct bau_control *smaster)
 373 {
 374         int cpu;
 375         struct hub_and_pnode *hpp;
 376
 377         for_each_present_cpu(cpu) {
 378                 hpp = &smaster->thp[cpu];
 379                 if (pnode == hpp->pnode)
 380                         return cpu;
 381         }
 382         return -1;
 383 }
 384
 385 /*
 386  * Last resort when we get a large number of destination timeouts is
 387  * to clear resources held by a given cpu.
 388  * Do this with IPI so that all messages in the BAU message queue
 389  * can be identified by their nonzero swack_vec field.
 390  *
 391  * This is entered for a single cpu on the uvhub.
 392  * The sender want's this uvhub to free a specific message's
 393  * swack resources.
 394  */
 395 static void do_reset(void *ptr)
 396 {
 397         int i;
 398         struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id());
 399         struct reset_args *rap = (struct reset_args *)ptr;
 400         struct bau_pq_entry *msg;
 401         struct ptc_stats *stat = bcp->statp;
 402
 403         stat->d_resets++;
 404         /*
 405          * We're looking for the given sender, and
 406          * will free its swack resource.
 407          * If all cpu's finally responded after the timeout, its
 408          * message 'replied_to' was set.
 409          */
 410         for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
 411                 unsigned long msg_res;
 412                 /* do_reset: same conditions for cancellation as
 413                    bau_process_retry_msg() */
 414                 if ((msg->replied_to == 0) &&
 415                     (msg->canceled == 0) &&
 416                     (msg->sending_cpu == rap->sender) &&
 417                     (msg->swack_vec) &&
 418                     (msg->msg_type != MSG_NOOP)) {
 419                         unsigned long mmr;
 420                         unsigned long mr;
 421                         /*
 422                          * make everyone else ignore this message
 423                          */
 424                         msg->canceled = 1;
 425                         /*
 426                          * only reset the resource if it is still pending
 427                          */
 428                         mmr = ops.read_l_sw_ack();
 429                         msg_res = msg->swack_vec;
 430                         mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
 431                         if (mmr & msg_res) {
 432                                 stat->d_rcanceled++;
 433                                 ops.write_l_sw_ack(mr);
 434                         }
 435                 }
 436         }
 437         return;
 438 }
 439
 440 /*
 441  * Use IPI to get all target uvhubs to release resources held by
 442  * a given sending cpu number.
 443  */
 444 static void reset_with_ipi(struct pnmask *distribution, struct bau_control *bcp)
 445 {
 446         int pnode;
 447         int apnode;
 448         int maskbits;
 449         int sender = bcp->cpu;
 450         cpumask_t *mask = bcp->uvhub_master->cpumask;
 451         struct bau_control *smaster = bcp->socket_master;
 452         struct reset_args reset_args;
 453
 454         reset_args.sender = sender;
 455         cpumask_clear(mask);
 456         /* find a single cpu for each uvhub in this distribution mask */
 457         maskbits = sizeof(struct pnmask) * BITSPERBYTE;
 458         /* each bit is a pnode relative to the partition base pnode */
 459         for (pnode = 0; pnode < maskbits; pnode++) {
 460                 int cpu;
 461                 if (!bau_uvhub_isset(pnode, distribution))
 462                         continue;
 463                 apnode = pnode + bcp->partition_base_pnode;
 464                 cpu = pnode_to_first_cpu(apnode, smaster);
 465                 cpumask_set_cpu(cpu, mask);
 466         }
 467
 468         /* IPI all cpus; preemption is already disabled */
 469         smp_call_function_many(mask, do_reset, (void *)&reset_args, 1);
 470         return;
 471 }
 472
 473 /*
 474  * Not to be confused with cycles_2_ns() from tsc.c; this gives a relative
 475  * number, not an absolute. It converts a duration in cycles to a duration in
 476  * ns.
 477  */
 478 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
 479 {
 480         struct cyc2ns_data *data = cyc2ns_read_begin();
 481         unsigned long long ns;
 482
 483         ns = mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift);
 484
 485         cyc2ns_read_end(data);
 486         return ns;
 487 }
 488
 489 /*
 490  * The reverse of the above; converts a duration in ns to a duration in cycles.
 491  */
 492 static inline unsigned long long ns_2_cycles(unsigned long long ns)
 493 {
 494         struct cyc2ns_data *data = cyc2ns_read_begin();
 495         unsigned long long cyc;
 496
 497         cyc = (ns << data->cyc2ns_shift) / data->cyc2ns_mul;
 498
 499         cyc2ns_read_end(data);
 500         return cyc;
 501 }
 502
 503 static inline unsigned long cycles_2_us(unsigned long long cyc)
 504 {
 505         return cycles_2_ns(cyc) / NSEC_PER_USEC;
 506 }
 507
 508 static inline cycles_t sec_2_cycles(unsigned long sec)
 509 {
 510         return ns_2_cycles(sec * NSEC_PER_SEC);
 511 }
 512
 513 static inline unsigned long long usec_2_cycles(unsigned long usec)
 514 {
 515         return ns_2_cycles(usec * NSEC_PER_USEC);
 516 }
 517
 518 /*
 519  * wait for all cpus on this hub to finish their sends and go quiet
 520  * leaves uvhub_quiesce set so that no new broadcasts are started by
 521  * bau_flush_send_and_wait()
 522  */
 523 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
 524 {
 525         atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 526 }
 527
 528 /*
 529  * mark this quiet-requestor as done
 530  */
 531 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
 532 {
 533         atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
 534 }
 535
 536 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
 537 {
 538         unsigned long descriptor_status;
 539
 540         descriptor_status = uv_read_local_mmr(mmr_offset);
 541         descriptor_status >>= right_shift;
 542         descriptor_status &= UV_ACT_STATUS_MASK;
 543         return descriptor_status;
 544 }
 545
 546 /*
 547  * Wait for completion of a broadcast software ack message
 548  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
 549  */
 550 static int uv1_wait_completion(struct bau_desc *bau_desc,
 551                                 unsigned long mmr_offset, int right_shift,
 552                                 struct bau_control *bcp, long try)
 553 {
 554         unsigned long descriptor_status;
 555         cycles_t ttm;
 556         struct ptc_stats *stat = bcp->statp;
 557
 558         descriptor_status = uv1_read_status(mmr_offset, right_shift);
 559         /* spin on the status MMR, waiting for it to go idle */
 560         while ((descriptor_status != DS_IDLE)) {
 561                 /*
 562                  * Our software ack messages may be blocked because
 563                  * there are no swack resources available.  As long
 564                  * as none of them has timed out hardware will NACK
 565                  * our message and its state will stay IDLE.
 566                  */
 567                 if (descriptor_status == DS_SOURCE_TIMEOUT) {
 568                         stat->s_stimeout++;
 569                         return FLUSH_GIVEUP;
 570                 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
 571                         stat->s_dtimeout++;
 572                         ttm = get_cycles();
 573
 574                         /*
 575                          * Our retries may be blocked by all destination
 576                          * swack resources being consumed, and a timeout
 577                          * pending.  In that case hardware returns the
 578                          * ERROR that looks like a destination timeout.
 579                          */
 580                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 581                                 bcp->conseccompletes = 0;
 582                                 return FLUSH_RETRY_PLUGGED;
 583                         }
 584
 585                         bcp->conseccompletes = 0;
 586                         return FLUSH_RETRY_TIMEOUT;
 587                 } else {
 588                         /*
 589                          * descriptor_status is still BUSY
 590                          */
 591                         cpu_relax();
 592                 }
 593                 descriptor_status = uv1_read_status(mmr_offset, right_shift);
 594         }
 595         bcp->conseccompletes++;
 596         return FLUSH_COMPLETE;
 597 }
 598
 599 /*
 600  * UV2 could have an extra bit of status in the ACTIVATION_STATUS_2 register.
 601  * But not currently used.
 602  */
 603 static unsigned long uv2_3_read_status(unsigned long offset, int rshft, int desc)
 604 {
 605         return ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK) << 1;
 606 }
 607
 608 /*
 609  * Return whether the status of the descriptor that is normally used for this
 610  * cpu (the one indexed by its hub-relative cpu number) is busy.
 611  * The status of the original 32 descriptors is always reflected in the 64
 612  * bits of UVH_LB_BAU_SB_ACTIVATION_STATUS_0.
 613  * The bit provided by the activation_status_2 register is irrelevant to
 614  * the status if it is only being tested for busy or not busy.
 615  */
 616 int normal_busy(struct bau_control *bcp)
 617 {
 618         int cpu = bcp->uvhub_cpu;
 619         int mmr_offset;
 620         int right_shift;
 621
 622         mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 623         right_shift = cpu * UV_ACT_STATUS_SIZE;
 624         return (((((read_lmmr(mmr_offset) >> right_shift) &
 625                                 UV_ACT_STATUS_MASK)) << 1) == UV2H_DESC_BUSY);
 626 }
 627
 628 /*
 629  * Entered when a bau descriptor has gone into a permanent busy wait because
 630  * of a hardware bug.
 631  * Workaround the bug.
 632  */
 633 int handle_uv2_busy(struct bau_control *bcp)
 634 {
 635         struct ptc_stats *stat = bcp->statp;
 636
 637         stat->s_uv2_wars++;
 638         bcp->busy = 1;
 639         return FLUSH_GIVEUP;
 640 }
 641
 642 static int uv2_3_wait_completion(struct bau_desc *bau_desc,
 643                                 unsigned long mmr_offset, int right_shift,
 644                                 struct bau_control *bcp, long try)
 645 {
 646         unsigned long descriptor_stat;
 647         cycles_t ttm;
 648         int desc = bcp->uvhub_cpu;
 649         long busy_reps = 0;
 650         struct ptc_stats *stat = bcp->statp;
 651
 652         descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
 653
 654         /* spin on the status MMR, waiting for it to go idle */
 655         while (descriptor_stat != UV2H_DESC_IDLE) {
 656                 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT)) {
 657                         /*
 658                          * A h/w bug on the destination side may
 659                          * have prevented the message being marked
 660                          * pending, thus it doesn't get replied to
 661                          * and gets continually nacked until it times
 662                          * out with a SOURCE_TIMEOUT.
 663                          */
 664                         stat->s_stimeout++;
 665                         return FLUSH_GIVEUP;
 666                 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
 667                         ttm = get_cycles();
 668
 669                         /*
 670                          * Our retries may be blocked by all destination
 671                          * swack resources being consumed, and a timeout
 672                          * pending.  In that case hardware returns the
 673                          * ERROR that looks like a destination timeout.
 674                          * Without using the extended status we have to
 675                          * deduce from the short time that this was a
 676                          * strong nack.
 677                          */
 678                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
 679                                 bcp->conseccompletes = 0;
 680                                 stat->s_plugged++;
 681                                 /* FLUSH_RETRY_PLUGGED causes hang on boot */
 682                                 return FLUSH_GIVEUP;
 683                         }
 684                         stat->s_dtimeout++;
 685                         bcp->conseccompletes = 0;
 686                         /* FLUSH_RETRY_TIMEOUT causes hang on boot */
 687                         return FLUSH_GIVEUP;
 688                 } else {
 689                         busy_reps++;
 690                         if (busy_reps > 1000000) {
 691                                 /* not to hammer on the clock */
 692                                 busy_reps = 0;
 693                                 ttm = get_cycles();
 694                                 if ((ttm - bcp->send_message) > bcp->timeout_interval)
 695                                         return handle_uv2_busy(bcp);
 696                         }
 697                         /*
 698                          * descriptor_stat is still BUSY
 699                          */
 700                         cpu_relax();
 701                 }
 702                 descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
 703         }
 704         bcp->conseccompletes++;
 705         return FLUSH_COMPLETE;
 706 }
 707
 708 /*
 709  * There are 2 status registers; each and array[32] of 2 bits. Set up for
 710  * which register to read and position in that register based on cpu in
 711  * current hub.
 712  */
 713 static int wait_completion(struct bau_desc *bau_desc, struct bau_control *bcp, long try)
 714 {
 715         int right_shift;
 716         unsigned long mmr_offset;
 717         int desc = bcp->uvhub_cpu;
 718
 719         if (desc < UV_CPUS_PER_AS) {
 720                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 721                 right_shift = desc * UV_ACT_STATUS_SIZE;
 722         } else {
 723                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
 724                 right_shift = ((desc - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
 725         }
 726
 727         if (bcp->uvhub_version == 1)
 728                 return uv1_wait_completion(bau_desc, mmr_offset, right_shift, bcp, try);
 729         else
 730                 return uv2_3_wait_completion(bau_desc, mmr_offset, right_shift, bcp, try);
 731 }
 732
 733 /*
 734  * Our retries are blocked by all destination sw ack resources being
 735  * in use, and a timeout is pending. In that case hardware immediately
 736  * returns the ERROR that looks like a destination timeout.
 737  */
 738 static void destination_plugged(struct bau_desc *bau_desc,
 739                         struct bau_control *bcp,
 740                         struct bau_control *hmaster, struct ptc_stats *stat)
 741 {
 742         udelay(bcp->plugged_delay);
 743         bcp->plugged_tries++;
 744
 745         if (bcp->plugged_tries >= bcp->plugsb4reset) {
 746                 bcp->plugged_tries = 0;
 747
 748                 quiesce_local_uvhub(hmaster);
 749
 750                 spin_lock(&hmaster->queue_lock);
 751                 reset_with_ipi(&bau_desc->distribution, bcp);
 752                 spin_unlock(&hmaster->queue_lock);
 753
 754                 end_uvhub_quiesce(hmaster);
 755
 756                 bcp->ipi_attempts++;
 757                 stat->s_resets_plug++;
 758         }
 759 }
 760
 761 static void destination_timeout(struct bau_desc *bau_desc,
 762                         struct bau_control *bcp, struct bau_control *hmaster,
 763                         struct ptc_stats *stat)
 764 {
 765         hmaster->max_concurr = 1;
 766         bcp->timeout_tries++;
 767         if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
 768                 bcp->timeout_tries = 0;
 769
 770                 quiesce_local_uvhub(hmaster);
 771
 772                 spin_lock(&hmaster->queue_lock);
 773                 reset_with_ipi(&bau_desc->distribution, bcp);
 774                 spin_unlock(&hmaster->queue_lock);
 775
 776                 end_uvhub_quiesce(hmaster);
 777
 778                 bcp->ipi_attempts++;
 779                 stat->s_resets_timeout++;
 780         }
 781 }
 782
 783 /*
 784  * Stop all cpus on a uvhub from using the BAU for a period of time.
 785  * This is reversed by check_enable.
 786  */
 787 static void disable_for_period(struct bau_control *bcp, struct ptc_stats *stat)
 788 {
 789         int tcpu;
 790         struct bau_control *tbcp;
 791         struct bau_control *hmaster;
 792         cycles_t tm1;
 793
 794         hmaster = bcp->uvhub_master;
 795         spin_lock(&hmaster->disable_lock);
 796         if (!bcp->baudisabled) {
 797                 stat->s_bau_disabled++;
 798                 tm1 = get_cycles();
 799                 for_each_present_cpu(tcpu) {
 800                         tbcp = &per_cpu(bau_control, tcpu);
 801                         if (tbcp->uvhub_master == hmaster) {
 802                                 tbcp->baudisabled = 1;
 803                                 tbcp->set_bau_on_time =
 804                                         tm1 + bcp->disabled_period;
 805                         }
 806                 }
 807         }
 808         spin_unlock(&hmaster->disable_lock);
 809 }
 810
 811 static void count_max_concurr(int stat, struct bau_control *bcp,
 812                                 struct bau_control *hmaster)
 813 {
 814         bcp->plugged_tries = 0;
 815         bcp->timeout_tries = 0;
 816         if (stat != FLUSH_COMPLETE)
 817                 return;
 818         if (bcp->conseccompletes <= bcp->complete_threshold)
 819                 return;
 820         if (hmaster->max_concurr >= hmaster->max_concurr_const)
 821                 return;
 822         hmaster->max_concurr++;
 823 }
 824
 825 static void record_send_stats(cycles_t time1, cycles_t time2,
 826                 struct bau_control *bcp, struct ptc_stats *stat,
 827                 int completion_status, int try)
 828 {
 829         cycles_t elapsed;
 830
 831         if (time2 > time1) {
 832                 elapsed = time2 - time1;
 833                 stat->s_time += elapsed;
 834
 835                 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
 836                         bcp->period_requests++;
 837                         bcp->period_time += elapsed;
 838                         if ((elapsed > congested_cycles) &&
 839                             (bcp->period_requests > bcp->cong_reps) &&
 840                             ((bcp->period_time / bcp->period_requests) >
 841                                                         congested_cycles)) {
 842                                 stat->s_congested++;
 843                                 disable_for_period(bcp, stat);
 844                         }
 845                 }
 846         } else
 847                 stat->s_requestor--;
 848
 849         if (completion_status == FLUSH_COMPLETE && try > 1)
 850                 stat->s_retriesok++;
 851         else if (completion_status == FLUSH_GIVEUP) {
 852                 stat->s_giveup++;
 853                 if (get_cycles() > bcp->period_end)
 854                         bcp->period_giveups = 0;
 855                 bcp->period_giveups++;
 856                 if (bcp->period_giveups == 1)
 857                         bcp->period_end = get_cycles() + bcp->disabled_period;
 858                 if (bcp->period_giveups > bcp->giveup_limit) {
 859                         disable_for_period(bcp, stat);
 860                         stat->s_giveuplimit++;
 861                 }
 862         }
 863 }
 864
 865 /*
 866  * Because of a uv1 hardware bug only a limited number of concurrent
 867  * requests can be made.
 868  */
 869 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
 870 {
 871         spinlock_t *lock = &hmaster->uvhub_lock;
 872         atomic_t *v;
 873
 874         v = &hmaster->active_descriptor_count;
 875         if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
 876                 stat->s_throttles++;
 877                 do {
 878                         cpu_relax();
 879                 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
 880         }
 881 }
 882
 883 /*
 884  * Handle the completion status of a message send.
 885  */
 886 static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
 887                         struct bau_control *bcp, struct bau_control *hmaster,
 888                         struct ptc_stats *stat)
 889 {
 890         if (completion_status == FLUSH_RETRY_PLUGGED)
 891                 destination_plugged(bau_desc, bcp, hmaster, stat);
 892         else if (completion_status == FLUSH_RETRY_TIMEOUT)
 893                 destination_timeout(bau_desc, bcp, hmaster, stat);
 894 }
 895
 896 /*
 897  * Send a broadcast and wait for it to complete.
 898  *
 899  * The flush_mask contains the cpus the broadcast is to be sent to including
 900  * cpus that are on the local uvhub.
 901  *
 902  * Returns 0 if all flushing represented in the mask was done.
 903  * Returns 1 if it gives up entirely and the original cpu mask is to be
 904  * returned to the kernel.
 905  */
 906 int uv_flush_send_and_wait(struct cpumask *flush_mask, struct bau_control *bcp,
 907         struct bau_desc *bau_desc)
 908 {
 909         int seq_number = 0;
 910         int completion_stat = 0;
 911         int uv1 = 0;
 912         long try = 0;
 913         unsigned long index;
 914         cycles_t time1;
 915         cycles_t time2;
 916         struct ptc_stats *stat = bcp->statp;
 917         struct bau_control *hmaster = bcp->uvhub_master;
 918         struct uv1_bau_msg_header *uv1_hdr = NULL;
 919         struct uv2_3_bau_msg_header *uv2_3_hdr = NULL;
 920
 921         if (bcp->uvhub_version == 1) {
 922                 uv1 = 1;
 923                 uv1_throttle(hmaster, stat);
 924         }
 925
 926         while (hmaster->uvhub_quiesce)
 927                 cpu_relax();
 928
 929         time1 = get_cycles();
 930         if (uv1)
 931                 uv1_hdr = &bau_desc->header.uv1_hdr;
 932         else
 933                 /* uv2 and uv3 */
 934                 uv2_3_hdr = &bau_desc->header.uv2_3_hdr;
 935
 936         do {
 937                 if (try == 0) {
 938                         if (uv1)
 939                                 uv1_hdr->msg_type = MSG_REGULAR;
 940                         else
 941                                 uv2_3_hdr->msg_type = MSG_REGULAR;
 942                         seq_number = bcp->message_number++;
 943                 } else {
 944                         if (uv1)
 945                                 uv1_hdr->msg_type = MSG_RETRY;
 946                         else
 947                                 uv2_3_hdr->msg_type = MSG_RETRY;
 948                         stat->s_retry_messages++;
 949                 }
 950
 951                 if (uv1)
 952                         uv1_hdr->sequence = seq_number;
 953                 else
 954                         uv2_3_hdr->sequence = seq_number;
 955                 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
 956                 bcp->send_message = get_cycles();
 957
 958                 write_mmr_activation(index);
 959
 960                 try++;
 961                 completion_stat = wait_completion(bau_desc, bcp, try);
 962
 963                 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
 964
 965                 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
 966                         bcp->ipi_attempts = 0;
 967                         stat->s_overipilimit++;
 968                         completion_stat = FLUSH_GIVEUP;
 969                         break;
 970                 }
 971                 cpu_relax();
 972         } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
 973                  (completion_stat == FLUSH_RETRY_TIMEOUT));
 974
 975         time2 = get_cycles();
 976
 977         count_max_concurr(completion_stat, bcp, hmaster);
 978
 979         while (hmaster->uvhub_quiesce)
 980                 cpu_relax();
 981
 982         atomic_dec(&hmaster->active_descriptor_count);
 983
 984         record_send_stats(time1, time2, bcp, stat, completion_stat, try);
 985
 986         if (completion_stat == FLUSH_GIVEUP)
 987                 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
 988                 return 1;
 989         return 0;
 990 }
 991
 992 /*
 993  * The BAU is disabled for this uvhub. When the disabled time period has
 994  * expired re-enable it.
 995  * Return 0 if it is re-enabled for all cpus on this uvhub.
 996  */
 997 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
 998 {
 999         int tcpu;
1000         struct bau_control *tbcp;
1001         struct bau_control *hmaster;
1002
1003         hmaster = bcp->uvhub_master;
1004         spin_lock(&hmaster->disable_lock);
1005         if (bcp->baudisabled && (get_cycles() >= bcp->set_bau_on_time)) {
1006                 stat->s_bau_reenabled++;
1007                 for_each_present_cpu(tcpu) {
1008                         tbcp = &per_cpu(bau_control, tcpu);
1009                         if (tbcp->uvhub_master == hmaster) {
1010                                 tbcp->baudisabled = 0;
1011                                 tbcp->period_requests = 0;
1012                                 tbcp->period_time = 0;
1013                                 tbcp->period_giveups = 0;
1014                         }
1015                 }
1016                 spin_unlock(&hmaster->disable_lock);
1017                 return 0;
1018         }
1019         spin_unlock(&hmaster->disable_lock);
1020         return -1;
1021 }
1022
1023 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
1024                                 int remotes, struct bau_desc *bau_desc)
1025 {
1026         stat->s_requestor++;
1027         stat->s_ntargcpu += remotes + locals;
1028         stat->s_ntargremotes += remotes;
1029         stat->s_ntarglocals += locals;
1030
1031         /* uvhub statistics */
1032         hubs = bau_uvhub_weight(&bau_desc->distribution);
1033         if (locals) {
1034                 stat->s_ntarglocaluvhub++;
1035                 stat->s_ntargremoteuvhub += (hubs - 1);
1036         } else
1037                 stat->s_ntargremoteuvhub += hubs;
1038
1039         stat->s_ntarguvhub += hubs;
1040
1041         if (hubs >= 16)
1042                 stat->s_ntarguvhub16++;
1043         else if (hubs >= 8)
1044                 stat->s_ntarguvhub8++;
1045         else if (hubs >= 4)
1046                 stat->s_ntarguvhub4++;
1047         else if (hubs >= 2)
1048                 stat->s_ntarguvhub2++;
1049         else
1050                 stat->s_ntarguvhub1++;
1051 }
1052
1053 /*
1054  * Translate a cpu mask to the uvhub distribution mask in the BAU
1055  * activation descriptor.
1056  */
1057 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
1058                         struct bau_desc *bau_desc, int *localsp, int *remotesp)
1059 {
1060         int cpu;
1061         int pnode;
1062         int cnt = 0;
1063         struct hub_and_pnode *hpp;
1064
1065         for_each_cpu(cpu, flush_mask) {
1066                 /*
1067                  * The distribution vector is a bit map of pnodes, relative
1068                  * to the partition base pnode (and the partition base nasid
1069                  * in the header).
1070                  * Translate cpu to pnode and hub using a local memory array.
1071                  */
1072                 hpp = &bcp->socket_master->thp[cpu];
1073                 pnode = hpp->pnode - bcp->partition_base_pnode;
1074                 bau_uvhub_set(pnode, &bau_desc->distribution);
1075                 cnt++;
1076                 if (hpp->uvhub == bcp->uvhub)
1077                         (*localsp)++;
1078                 else
1079                         (*remotesp)++;
1080         }
1081         if (!cnt)
1082                 return 1;
1083         return 0;
1084 }
1085
1086 /*
1087  * globally purge translation cache of a virtual address or all TLB's
1088  * @cpumask: mask of all cpu's in which the address is to be removed
1089  * @mm: mm_struct containing virtual address range
1090  * @start: start virtual address to be removed from TLB
1091  * @end: end virtual address to be remove from TLB
1092  * @cpu: the current cpu
1093  *
1094  * This is the entry point for initiating any UV global TLB shootdown.
1095  *
1096  * Purges the translation caches of all specified processors of the given
1097  * virtual address, or purges all TLB's on specified processors.
1098  *
1099  * The caller has derived the cpumask from the mm_struct.  This function
1100  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1101  *
1102  * The cpumask is converted into a uvhubmask of the uvhubs containing
1103  * those cpus.
1104  *
1105  * Note that this function should be called with preemption disabled.
1106  *
1107  * Returns NULL if all remote flushing was done.
1108  * Returns pointer to cpumask if some remote flushing remains to be
1109  * done.  The returned pointer is valid till preemption is re-enabled.
1110  */
1111 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1112                                                 struct mm_struct *mm,
1113                                                 unsigned long start,
1114                                                 unsigned long end,
1115                                                 unsigned int cpu)
1116 {
1117         int locals = 0;
1118         int remotes = 0;
1119         int hubs = 0;
1120         struct bau_desc *bau_desc;
1121         struct cpumask *flush_mask;
1122         struct ptc_stats *stat;
1123         struct bau_control *bcp;
1124         unsigned long descriptor_status;
1125         unsigned long status;
1126
1127         bcp = &per_cpu(bau_control, cpu);
1128
1129         if (bcp->nobau)
1130                 return cpumask;
1131
1132         stat = bcp->statp;
1133         stat->s_enters++;
1134
1135         if (bcp->busy) {
1136                 descriptor_status =
1137                         read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1138                 status = ((descriptor_status >> (bcp->uvhub_cpu *
1139                         UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1140                 if (status == UV2H_DESC_BUSY)
1141                         return cpumask;
1142                 bcp->busy = 0;
1143         }
1144
1145         /* bau was disabled due to slow response */
1146         if (bcp->baudisabled) {
1147                 if (check_enable(bcp, stat)) {
1148                         stat->s_ipifordisabled++;
1149                         return cpumask;
1150                 }
1151         }
1152
1153         /*
1154          * Each sending cpu has a per-cpu mask which it fills from the caller's
1155          * cpu mask.  All cpus are converted to uvhubs and copied to the
1156          * activation descriptor.
1157          */
1158         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1159         /* don't actually do a shootdown of the local cpu */
1160         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1161
1162         if (cpumask_test_cpu(cpu, cpumask))
1163                 stat->s_ntargself++;
1164
1165         bau_desc = bcp->descriptor_base;
1166         bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1167         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1168         if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1169                 return NULL;
1170
1171         record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1172
1173         if (!end || (end - start) <= PAGE_SIZE)
1174                 bau_desc->payload.address = start;
1175         else
1176                 bau_desc->payload.address = TLB_FLUSH_ALL;
1177         bau_desc->payload.sending_cpu = cpu;
1178         /*
1179          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1180          * or 1 if it gave up and the original cpumask should be returned.
1181          */
1182         if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1183                 return NULL;
1184         else
1185                 return cpumask;
1186 }
1187
1188 /*
1189  * Search the message queue for any 'other' unprocessed message with the
1190  * same software acknowledge resource bit vector as the 'msg' message.
1191  */
1192 struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1193                                            struct bau_control *bcp)
1194 {
1195         struct bau_pq_entry *msg_next = msg + 1;
1196         unsigned char swack_vec = msg->swack_vec;
1197
1198         if (msg_next > bcp->queue_last)
1199                 msg_next = bcp->queue_first;
1200         while (msg_next != msg) {
1201                 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1202                                 (msg_next->swack_vec == swack_vec))
1203                         return msg_next;
1204                 msg_next++;
1205                 if (msg_next > bcp->queue_last)
1206                         msg_next = bcp->queue_first;
1207         }
1208         return NULL;
1209 }
1210
1211 /*
1212  * UV2 needs to work around a bug in which an arriving message has not
1213  * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1214  * Such a message must be ignored.
1215  */
1216 void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1217 {
1218         unsigned long mmr_image;
1219         unsigned char swack_vec;
1220         struct bau_pq_entry *msg = mdp->msg;
1221         struct bau_pq_entry *other_msg;
1222
1223         mmr_image = ops.read_l_sw_ack();
1224         swack_vec = msg->swack_vec;
1225
1226         if ((swack_vec & mmr_image) == 0) {
1227                 /*
1228                  * This message was assigned a swack resource, but no
1229                  * reserved acknowlegment is pending.
1230                  * The bug has prevented this message from setting the MMR.
1231                  */
1232                 /*
1233                  * Some message has set the MMR 'pending' bit; it might have
1234                  * been another message.  Look for that message.
1235                  */
1236                 other_msg = find_another_by_swack(msg, bcp);
1237                 if (other_msg) {
1238                         /*
1239                          * There is another. Process this one but do not
1240                          * ack it.
1241                          */
1242                         bau_process_message(mdp, bcp, 0);
1243                         /*
1244                          * Let the natural processing of that other message
1245                          * acknowledge it. Don't get the processing of sw_ack's
1246                          * out of order.
1247                          */
1248                         return;
1249                 }
1250         }
1251
1252         /*
1253          * Either the MMR shows this one pending a reply or there is no
1254          * other message using this sw_ack, so it is safe to acknowledge it.
1255          */
1256         bau_process_message(mdp, bcp, 1);
1257
1258         return;
1259 }
1260
1261 /*
1262  * The BAU message interrupt comes here. (registered by set_intr_gate)
1263  * See entry_64.S
1264  *
1265  * We received a broadcast assist message.
1266  *
1267  * Interrupts are disabled; this interrupt could represent
1268  * the receipt of several messages.
1269  *
1270  * All cores/threads on this hub get this interrupt.
1271  * The last one to see it does the software ack.
1272  * (the resource will not be freed until noninterruptable cpus see this
1273  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
1274  */
1275 void uv_bau_message_interrupt(struct pt_regs *regs)
1276 {
1277         int count = 0;
1278         cycles_t time_start;
1279         struct bau_pq_entry *msg;
1280         struct bau_control *bcp;
1281         struct ptc_stats *stat;
1282         struct msg_desc msgdesc;
1283
1284         ack_APIC_irq();
1285         time_start = get_cycles();
1286
1287         bcp = &per_cpu(bau_control, smp_processor_id());
1288         stat = bcp->statp;
1289
1290         msgdesc.queue_first = bcp->queue_first;
1291         msgdesc.queue_last = bcp->queue_last;
1292
1293         msg = bcp->bau_msg_head;
1294         while (msg->swack_vec) {
1295                 count++;
1296
1297                 msgdesc.msg_slot = msg - msgdesc.queue_first;
1298                 msgdesc.msg = msg;
1299                 if (bcp->uvhub_version == 2)
1300                         process_uv2_message(&msgdesc, bcp);
1301                 else
1302                         /* no error workaround for uv1 or uv3 */
1303                         bau_process_message(&msgdesc, bcp, 1);
1304
1305                 msg++;
1306                 if (msg > msgdesc.queue_last)
1307                         msg = msgdesc.queue_first;
1308                 bcp->bau_msg_head = msg;
1309         }
1310         stat->d_time += (get_cycles() - time_start);
1311         if (!count)
1312                 stat->d_nomsg++;
1313         else if (count > 1)
1314                 stat->d_multmsg++;
1315 }
1316
1317 /*
1318  * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1319  * shootdown message timeouts enabled.  The timeout does not cause
1320  * an interrupt, but causes an error message to be returned to
1321  * the sender.
1322  */
1323 static void __init enable_timeouts(void)
1324 {
1325         int uvhub;
1326         int nuvhubs;
1327         int pnode;
1328         unsigned long mmr_image;
1329
1330         nuvhubs = uv_num_possible_blades();
1331
1332         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1333                 if (!uv_blade_nr_possible_cpus(uvhub))
1334                         continue;
1335
1336                 pnode = uv_blade_to_pnode(uvhub);
1337                 mmr_image = read_mmr_misc_control(pnode);
1338                 /*
1339                  * Set the timeout period and then lock it in, in three
1340                  * steps; captures and locks in the period.
1341                  *
1342                  * To program the period, the SOFT_ACK_MODE must be off.
1343                  */
1344                 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1345                 write_mmr_misc_control(pnode, mmr_image);
1346                 /*
1347                  * Set the 4-bit period.
1348                  */
1349                 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1350                 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1351                 write_mmr_misc_control(pnode, mmr_image);
1352                 /*
1353                  * UV1:
1354                  * Subsequent reversals of the timebase bit (3) cause an
1355                  * immediate timeout of one or all INTD resources as
1356                  * indicated in bits 2:0 (7 causes all of them to timeout).
1357                  */
1358                 mmr_image |= (1L << SOFTACK_MSHIFT);
1359                 if (is_uv2_hub()) {
1360                         /* do not touch the legacy mode bit */
1361                         /* hw bug workaround; do not use extended status */
1362                         mmr_image &= ~(1L << UV2_EXT_SHFT);
1363                 } else if (is_uv3_hub()) {
1364                         mmr_image &= ~(1L << PREFETCH_HINT_SHFT);
1365                         mmr_image |= (1L << SB_STATUS_SHFT);
1366                 }
1367                 write_mmr_misc_control(pnode, mmr_image);
1368         }
1369 }
1370
1371 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1372 {
1373         if (*offset < num_possible_cpus())
1374                 return offset;
1375         return NULL;
1376 }
1377
1378 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1379 {
1380         (*offset)++;
1381         if (*offset < num_possible_cpus())
1382                 return offset;
1383         return NULL;
1384 }
1385
1386 static void ptc_seq_stop(struct seq_file *file, void *data)
1387 {
1388 }
1389
1390 /*
1391  * Display the statistics thru /proc/sgi_uv/ptc_statistics
1392  * 'data' points to the cpu number
1393  * Note: see the descriptions in stat_description[].
1394  */
1395 static int ptc_seq_show(struct seq_file *file, void *data)
1396 {
1397         struct ptc_stats *stat;
1398         struct bau_control *bcp;
1399         int cpu;
1400
1401         cpu = *(loff_t *)data;
1402         if (!cpu) {
1403                 seq_puts(file,
1404                          "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1405                 seq_puts(file, "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1406                 seq_puts(file,
1407                          "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1408                 seq_puts(file,
1409                          "rok resetp resett giveup sto bz throt disable ");
1410                 seq_puts(file,
1411                          "enable wars warshw warwaits enters ipidis plugged ");
1412                 seq_puts(file,
1413                          "ipiover glim cong swack recv rtime all one mult ");
1414                 seq_puts(file, "none retry canc nocan reset rcan\n");
1415         }
1416         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1417                 bcp = &per_cpu(bau_control, cpu);
1418                 if (bcp->nobau) {
1419                         seq_printf(file, "cpu %d bau disabled\n", cpu);
1420                         return 0;
1421                 }
1422                 stat = bcp->statp;
1423                 /* source side statistics */
1424                 seq_printf(file,
1425                         "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1426                            cpu, bcp->nobau, stat->s_requestor,
1427                            cycles_2_us(stat->s_time),
1428                            stat->s_ntargself, stat->s_ntarglocals,
1429                            stat->s_ntargremotes, stat->s_ntargcpu,
1430                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1431                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
1432                 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1433                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1434                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1435                            stat->s_dtimeout, stat->s_strongnacks);
1436                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1437                            stat->s_retry_messages, stat->s_retriesok,
1438                            stat->s_resets_plug, stat->s_resets_timeout,
1439                            stat->s_giveup, stat->s_stimeout,
1440                            stat->s_busy, stat->s_throttles);
1441                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1442                            stat->s_bau_disabled, stat->s_bau_reenabled,
1443                            stat->s_uv2_wars, stat->s_uv2_wars_hw,
1444                            stat->s_uv2_war_waits, stat->s_enters,
1445                            stat->s_ipifordisabled, stat->s_plugged,
1446                            stat->s_overipilimit, stat->s_giveuplimit,
1447                            stat->s_congested);
1448
1449                 /* destination side statistics */
1450                 seq_printf(file,
1451                         "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1452                            ops.read_g_sw_ack(uv_cpu_to_pnode(cpu)),
1453                            stat->d_requestee, cycles_2_us(stat->d_time),
1454                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1455                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
1456                            stat->d_nocanceled, stat->d_resets,
1457                            stat->d_rcanceled);
1458         }
1459         return 0;
1460 }
1461
1462 /*
1463  * Display the tunables thru debugfs
1464  */
1465 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1466                                 size_t count, loff_t *ppos)
1467 {
1468         char *buf;
1469         int ret;
1470
1471         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1472                 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1473                 "ipi_reset_limit complete_threshold congested_response_us",
1474                 "congested_reps disabled_period giveup_limit",
1475                 max_concurr, plugged_delay, plugsb4reset,
1476                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1477                 congested_respns_us, congested_reps, disabled_period,
1478                 giveup_limit);
1479
1480         if (!buf)
1481                 return -ENOMEM;
1482
1483         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1484         kfree(buf);
1485         return ret;
1486 }
1487
1488 /*
1489  * handle a write to /proc/sgi_uv/ptc_statistics
1490  * -1: reset the statistics
1491  *  0: display meaning of the statistics
1492  */
1493 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1494                                 size_t count, loff_t *data)
1495 {
1496         int cpu;
1497         int i;
1498         int elements;
1499         long input_arg;
1500         char optstr[64];
1501         struct ptc_stats *stat;
1502
1503         if (count == 0 || count > sizeof(optstr))
1504                 return -EINVAL;
1505         if (copy_from_user(optstr, user, count))
1506                 return -EFAULT;
1507         optstr[count - 1] = '\0';
1508
1509         if (!strcmp(optstr, "on")) {
1510                 set_bau_on();
1511                 return count;
1512         } else if (!strcmp(optstr, "off")) {
1513                 set_bau_off();
1514                 return count;
1515         }
1516
1517         if (kstrtol(optstr, 10, &input_arg) < 0) {
1518                 pr_debug("%s is invalid\n", optstr);
1519                 return -EINVAL;
1520         }
1521
1522         if (input_arg == 0) {
1523                 elements = ARRAY_SIZE(stat_description);
1524                 pr_debug("# cpu:      cpu number\n");
1525                 pr_debug("Sender statistics:\n");
1526                 for (i = 0; i < elements; i++)
1527                         pr_debug("%s\n", stat_description[i]);
1528         } else if (input_arg == -1) {
1529                 for_each_present_cpu(cpu) {
1530                         stat = &per_cpu(ptcstats, cpu);
1531                         memset(stat, 0, sizeof(struct ptc_stats));
1532                 }
1533         }
1534
1535         return count;
1536 }
1537
1538 static int local_atoi(const char *name)
1539 {
1540         int val = 0;
1541
1542         for (;; name++) {
1543                 switch (*name) {
1544                 case '0' ... '9':
1545                         val = 10*val+(*name-'0');
1546                         break;
1547                 default:
1548                         return val;
1549                 }
1550         }
1551 }
1552
1553 /*
1554  * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1555  * Zero values reset them to defaults.
1556  */
1557 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1558                                 int count)
1559 {
1560         char *p;
1561         char *q;
1562         int cnt = 0;
1563         int val;
1564         int e = ARRAY_SIZE(tunables);
1565
1566         p = instr + strspn(instr, WHITESPACE);
1567         q = p;
1568         for (; *p; p = q + strspn(q, WHITESPACE)) {
1569                 q = p + strcspn(p, WHITESPACE);
1570                 cnt++;
1571                 if (q == p)
1572                         break;
1573         }
1574         if (cnt != e) {
1575                 pr_info("bau tunable error: should be %d values\n", e);
1576                 return -EINVAL;
1577         }
1578
1579         p = instr + strspn(instr, WHITESPACE);
1580         q = p;
1581         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1582                 q = p + strcspn(p, WHITESPACE);
1583                 val = local_atoi(p);
1584                 switch (cnt) {
1585                 case 0:
1586                         if (val == 0) {
1587                                 max_concurr = MAX_BAU_CONCURRENT;
1588                                 max_concurr_const = MAX_BAU_CONCURRENT;
1589                                 continue;
1590                         }
1591                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1592                                 pr_debug(
1593                                 "Error: BAU max concurrent %d is invalid\n",
1594                                 val);
1595                                 return -EINVAL;
1596                         }
1597                         max_concurr = val;
1598                         max_concurr_const = val;
1599                         continue;
1600                 default:
1601                         if (val == 0)
1602                                 *tunables[cnt].tunp = tunables[cnt].deflt;
1603                         else
1604                                 *tunables[cnt].tunp = val;
1605                         continue;
1606                 }
1607                 if (q == p)
1608                         break;
1609         }
1610         return 0;
1611 }
1612
1613 /*
1614  * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1615  */
1616 static ssize_t tunables_write(struct file *file, const char __user *user,
1617                                 size_t count, loff_t *data)
1618 {
1619         int cpu;
1620         int ret;
1621         char instr[100];
1622         struct bau_control *bcp;
1623
1624         if (count == 0 || count > sizeof(instr)-1)
1625                 return -EINVAL;
1626         if (copy_from_user(instr, user, count))
1627                 return -EFAULT;
1628
1629         instr[count] = '\0';
1630
1631         cpu = get_cpu();
1632         bcp = &per_cpu(bau_control, cpu);
1633         ret = parse_tunables_write(bcp, instr, count);
1634         put_cpu();
1635         if (ret)
1636                 return ret;
1637
1638         for_each_present_cpu(cpu) {
1639                 bcp = &per_cpu(bau_control, cpu);
1640                 bcp->max_concurr         = max_concurr;
1641                 bcp->max_concurr_const   = max_concurr;
1642                 bcp->plugged_delay       = plugged_delay;
1643                 bcp->plugsb4reset        = plugsb4reset;
1644                 bcp->timeoutsb4reset     = timeoutsb4reset;
1645                 bcp->ipi_reset_limit     = ipi_reset_limit;
1646                 bcp->complete_threshold  = complete_threshold;
1647                 bcp->cong_response_us    = congested_respns_us;
1648                 bcp->cong_reps           = congested_reps;
1649                 bcp->disabled_period     = sec_2_cycles(disabled_period);
1650                 bcp->giveup_limit        = giveup_limit;
1651         }
1652         return count;
1653 }
1654
1655 static const struct seq_operations uv_ptc_seq_ops = {
1656         .start          = ptc_seq_start,
1657         .next           = ptc_seq_next,
1658         .stop           = ptc_seq_stop,
1659         .show           = ptc_seq_show
1660 };
1661
1662 static int ptc_proc_open(struct inode *inode, struct file *file)
1663 {
1664         return seq_open(file, &uv_ptc_seq_ops);
1665 }
1666
1667 static int tunables_open(struct inode *inode, struct file *file)
1668 {
1669         return 0;
1670 }
1671
1672 static const struct file_operations proc_uv_ptc_operations = {
1673         .open           = ptc_proc_open,
1674         .read           = seq_read,
1675         .write          = ptc_proc_write,
1676         .llseek         = seq_lseek,
1677         .release        = seq_release,
1678 };
1679
1680 static const struct file_operations tunables_fops = {
1681         .open           = tunables_open,
1682         .read           = tunables_read,
1683         .write          = tunables_write,
1684         .llseek         = default_llseek,
1685 };
1686
1687 static int __init uv_ptc_init(void)
1688 {
1689         struct proc_dir_entry *proc_uv_ptc;
1690
1691         if (!is_uv_system())
1692                 return 0;
1693
1694         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1695                                   &proc_uv_ptc_operations);
1696         if (!proc_uv_ptc) {
1697                 pr_err("unable to create %s proc entry\n",
1698                        UV_PTC_BASENAME);
1699                 return -EINVAL;
1700         }
1701
1702         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1703         if (!tunables_dir) {
1704                 pr_err("unable to create debugfs directory %s\n",
1705                        UV_BAU_TUNABLES_DIR);
1706                 return -EINVAL;
1707         }
1708         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1709                                         tunables_dir, NULL, &tunables_fops);
1710         if (!tunables_file) {
1711                 pr_err("unable to create debugfs file %s\n",
1712                        UV_BAU_TUNABLES_FILE);
1713                 return -EINVAL;
1714         }
1715         return 0;
1716 }
1717
1718 /*
1719  * Initialize the sending side's sending buffers.
1720  */
1721 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1722 {
1723         int i;
1724         int cpu;
1725         int uv1 = 0;
1726         unsigned long gpa;
1727         unsigned long m;
1728         unsigned long n;
1729         size_t dsize;
1730         struct bau_desc *bau_desc;
1731         struct bau_desc *bd2;
1732         struct uv1_bau_msg_header *uv1_hdr;
1733         struct uv2_3_bau_msg_header *uv2_3_hdr;
1734         struct bau_control *bcp;
1735
1736         /*
1737          * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1738          * per cpu; and one per cpu on the uvhub (ADP_SZ)
1739          */
1740         dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1741         bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1742         BUG_ON(!bau_desc);
1743
1744         gpa = uv_gpa(bau_desc);
1745         n = uv_gpa_to_gnode(gpa);
1746         m = ops.bau_gpa_to_offset(gpa);
1747         if (is_uv1_hub())
1748                 uv1 = 1;
1749
1750         /* the 14-bit pnode */
1751         write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1752         /*
1753          * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1754          * cpu even though we only use the first one; one descriptor can
1755          * describe a broadcast to 256 uv hubs.
1756          */
1757         for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1758                 memset(bd2, 0, sizeof(struct bau_desc));
1759                 if (uv1) {
1760                         uv1_hdr = &bd2->header.uv1_hdr;
1761                         uv1_hdr->swack_flag = 1;
1762                         /*
1763                          * The base_dest_nasid set in the message header
1764                          * is the nasid of the first uvhub in the partition.
1765                          * The bit map will indicate destination pnode numbers
1766                          * relative to that base. They may not be consecutive
1767                          * if nasid striding is being used.
1768                          */
1769                         uv1_hdr->base_dest_nasid =
1770                                                   UV_PNODE_TO_NASID(base_pnode);
1771                         uv1_hdr->dest_subnodeid  = UV_LB_SUBNODEID;
1772                         uv1_hdr->command         = UV_NET_ENDPOINT_INTD;
1773                         uv1_hdr->int_both        = 1;
1774                         /*
1775                          * all others need to be set to zero:
1776                          *   fairness chaining multilevel count replied_to
1777                          */
1778                 } else {
1779                         /*
1780                          * BIOS uses legacy mode, but uv2 and uv3 hardware always
1781                          * uses native mode for selective broadcasts.
1782                          */
1783                         uv2_3_hdr = &bd2->header.uv2_3_hdr;
1784                         uv2_3_hdr->swack_flag      = 1;
1785                         uv2_3_hdr->base_dest_nasid =
1786                                                   UV_PNODE_TO_NASID(base_pnode);
1787                         uv2_3_hdr->dest_subnodeid  = UV_LB_SUBNODEID;
1788                         uv2_3_hdr->command         = UV_NET_ENDPOINT_INTD;
1789                 }
1790         }
1791         for_each_present_cpu(cpu) {
1792                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1793                         continue;
1794                 bcp = &per_cpu(bau_control, cpu);
1795                 bcp->descriptor_base = bau_desc;
1796         }
1797 }
1798
1799 /*
1800  * initialize the destination side's receiving buffers
1801  * entered for each uvhub in the partition
1802  * - node is first node (kernel memory notion) on the uvhub
1803  * - pnode is the uvhub's physical identifier
1804  */
1805 static void pq_init(int node, int pnode)
1806 {
1807         int cpu;
1808         size_t plsize;
1809         char *cp;
1810         void *vp;
1811         unsigned long gnode, first, last, tail;
1812         struct bau_pq_entry *pqp;
1813         struct bau_control *bcp;
1814
1815         plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1816         vp = kmalloc_node(plsize, GFP_KERNEL, node);
1817         pqp = (struct bau_pq_entry *)vp;
1818         BUG_ON(!pqp);
1819
1820         cp = (char *)pqp + 31;
1821         pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1822
1823         for_each_present_cpu(cpu) {
1824                 if (pnode != uv_cpu_to_pnode(cpu))
1825                         continue;
1826                 /* for every cpu on this pnode: */
1827                 bcp = &per_cpu(bau_control, cpu);
1828                 bcp->queue_first        = pqp;
1829                 bcp->bau_msg_head       = pqp;
1830                 bcp->queue_last         = pqp + (DEST_Q_SIZE - 1);
1831         }
1832
1833         first = ops.bau_gpa_to_offset(uv_gpa(pqp));
1834         last = ops.bau_gpa_to_offset(uv_gpa(pqp + (DEST_Q_SIZE - 1)));
1835
1836         /*
1837          * Pre UV4, the gnode is required to locate the payload queue
1838          * and the payload queue tail must be maintained by the kernel.
1839          */
1840         bcp = &per_cpu(bau_control, smp_processor_id());
1841         if (bcp->uvhub_version <= 3) {
1842                 tail = first;
1843                 gnode = uv_gpa_to_gnode(uv_gpa(pqp));
1844                 first = (gnode << UV_PAYLOADQ_GNODE_SHIFT) | tail;
1845                 write_mmr_payload_tail(pnode, tail);
1846         }
1847
1848         ops.write_payload_first(pnode, first);
1849         ops.write_payload_last(pnode, last);
1850
1851         /* in effect, all msg_type's are set to MSG_NOOP */
1852         memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1853 }
1854
1855 /*
1856  * Initialization of each UV hub's structures
1857  */
1858 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1859 {
1860         int node;
1861         int pnode;
1862         unsigned long apicid;
1863
1864         node = uvhub_to_first_node(uvhub);
1865         pnode = uv_blade_to_pnode(uvhub);
1866
1867         activation_descriptor_init(node, pnode, base_pnode);
1868
1869         pq_init(node, pnode);
1870         /*
1871          * The below initialization can't be in firmware because the
1872          * messaging IRQ will be determined by the OS.
1873          */
1874         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1875         write_mmr_data_config(pnode, ((apicid << 32) | vector));
1876 }
1877
1878 /*
1879  * We will set BAU_MISC_CONTROL with a timeout period.
1880  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1881  * So the destination timeout period has to be calculated from them.
1882  */
1883 static int calculate_destination_timeout(void)
1884 {
1885         unsigned long mmr_image;
1886         int mult1;
1887         int mult2;
1888         int index;
1889         int base;
1890         int ret;
1891         unsigned long ts_ns;
1892
1893         if (is_uv1_hub()) {
1894                 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1895                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1896                 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1897                 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1898                 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1899                 ts_ns = timeout_base_ns[index];
1900                 ts_ns *= (mult1 * mult2);
1901                 ret = ts_ns / 1000;
1902         } else {
1903                 /* same destination timeout for uv2 and uv3 */
1904                 /* 4 bits  0/1 for 10/80us base, 3 bits of multiplier */
1905                 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1906                 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1907                 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1908                         base = 80;
1909                 else
1910                         base = 10;
1911                 mult1 = mmr_image & UV2_ACK_MASK;
1912                 ret = mult1 * base;
1913         }
1914         return ret;
1915 }
1916
1917 static void __init init_per_cpu_tunables(void)
1918 {
1919         int cpu;
1920         struct bau_control *bcp;
1921
1922         for_each_present_cpu(cpu) {
1923                 bcp = &per_cpu(bau_control, cpu);
1924                 bcp->baudisabled                = 0;
1925                 if (nobau)
1926                         bcp->nobau              = true;
1927                 bcp->statp                      = &per_cpu(ptcstats, cpu);
1928                 /* time interval to catch a hardware stay-busy bug */
1929                 bcp->timeout_interval           = usec_2_cycles(2*timeout_us);
1930                 bcp->max_concurr                = max_concurr;
1931                 bcp->max_concurr_const          = max_concurr;
1932                 bcp->plugged_delay              = plugged_delay;
1933                 bcp->plugsb4reset               = plugsb4reset;
1934                 bcp->timeoutsb4reset            = timeoutsb4reset;
1935                 bcp->ipi_reset_limit            = ipi_reset_limit;
1936                 bcp->complete_threshold         = complete_threshold;
1937                 bcp->cong_response_us           = congested_respns_us;
1938                 bcp->cong_reps                  = congested_reps;
1939                 bcp->disabled_period            = sec_2_cycles(disabled_period);
1940                 bcp->giveup_limit               = giveup_limit;
1941                 spin_lock_init(&bcp->queue_lock);
1942                 spin_lock_init(&bcp->uvhub_lock);
1943                 spin_lock_init(&bcp->disable_lock);
1944         }
1945 }
1946
1947 /*
1948  * Scan all cpus to collect blade and socket summaries.
1949  */
1950 static int __init get_cpu_topology(int base_pnode,
1951                                         struct uvhub_desc *uvhub_descs,
1952                                         unsigned char *uvhub_mask)
1953 {
1954         int cpu;
1955         int pnode;
1956         int uvhub;
1957         int socket;
1958         struct bau_control *bcp;
1959         struct uvhub_desc *bdp;
1960         struct socket_desc *sdp;
1961
1962         for_each_present_cpu(cpu) {
1963                 bcp = &per_cpu(bau_control, cpu);
1964
1965                 memset(bcp, 0, sizeof(struct bau_control));
1966
1967                 pnode = uv_cpu_hub_info(cpu)->pnode;
1968                 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1969                         pr_emerg(
1970                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1971                                 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1972                         return 1;
1973                 }
1974
1975                 bcp->osnode = cpu_to_node(cpu);
1976                 bcp->partition_base_pnode = base_pnode;
1977
1978                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1979                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1980                 bdp = &uvhub_descs[uvhub];
1981
1982                 bdp->num_cpus++;
1983                 bdp->uvhub = uvhub;
1984                 bdp->pnode = pnode;
1985
1986                 /* kludge: 'assuming' one node per socket, and assuming that
1987                    disabling a socket just leaves a gap in node numbers */
1988                 socket = bcp->osnode & 1;
1989                 bdp->socket_mask |= (1 << socket);
1990                 sdp = &bdp->socket[socket];
1991                 sdp->cpu_number[sdp->num_cpus] = cpu;
1992                 sdp->num_cpus++;
1993                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1994                         pr_emerg("%d cpus per socket invalid\n",
1995                                 sdp->num_cpus);
1996                         return 1;
1997                 }
1998         }
1999         return 0;
2000 }
2001
2002 /*
2003  * Each socket is to get a local array of pnodes/hubs.
2004  */
2005 static void make_per_cpu_thp(struct bau_control *smaster)
2006 {
2007         int cpu;
2008         size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
2009
2010         smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
2011         memset(smaster->thp, 0, hpsz);
2012         for_each_present_cpu(cpu) {
2013                 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
2014                 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
2015         }
2016 }
2017
2018 /*
2019  * Each uvhub is to get a local cpumask.
2020  */
2021 static void make_per_hub_cpumask(struct bau_control *hmaster)
2022 {
2023         int sz = sizeof(cpumask_t);
2024
2025         hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
2026 }
2027
2028 /*
2029  * Initialize all the per_cpu information for the cpu's on a given socket,
2030  * given what has been gathered into the socket_desc struct.
2031  * And reports the chosen hub and socket masters back to the caller.
2032  */
2033 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
2034                         struct bau_control **smasterp,
2035                         struct bau_control **hmasterp)
2036 {
2037         int i;
2038         int cpu;
2039         struct bau_control *bcp;
2040
2041         for (i = 0; i < sdp->num_cpus; i++) {
2042                 cpu = sdp->cpu_number[i];
2043                 bcp = &per_cpu(bau_control, cpu);
2044                 bcp->cpu = cpu;
2045                 if (i == 0) {
2046                         *smasterp = bcp;
2047                         if (!(*hmasterp))
2048                                 *hmasterp = bcp;
2049                 }
2050                 bcp->cpus_in_uvhub = bdp->num_cpus;
2051                 bcp->cpus_in_socket = sdp->num_cpus;
2052                 bcp->socket_master = *smasterp;
2053                 bcp->uvhub = bdp->uvhub;
2054                 if (is_uv1_hub())
2055                         bcp->uvhub_version = 1;
2056                 else if (is_uv2_hub())
2057                         bcp->uvhub_version = 2;
2058                 else if (is_uv3_hub())
2059                         bcp->uvhub_version = 3;
2060                 else if (is_uv4_hub())
2061                         bcp->uvhub_version = 4;
2062                 else {
2063                         pr_emerg("uvhub version not 1, 2, 3, or 4\n");
2064                         return 1;
2065                 }
2066                 bcp->uvhub_master = *hmasterp;
2067                 bcp->uvhub_cpu = uv_cpu_blade_processor_id(cpu);
2068
2069                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2070                         pr_emerg("%d cpus per uvhub invalid\n",
2071                                 bcp->uvhub_cpu);
2072                         return 1;
2073                 }
2074         }
2075         return 0;
2076 }
2077
2078 /*
2079  * Summarize the blade and socket topology into the per_cpu structures.
2080  */
2081 static int __init summarize_uvhub_sockets(int nuvhubs,
2082                         struct uvhub_desc *uvhub_descs,
2083                         unsigned char *uvhub_mask)
2084 {
2085         int socket;
2086         int uvhub;
2087         unsigned short socket_mask;
2088
2089         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2090                 struct uvhub_desc *bdp;
2091                 struct bau_control *smaster = NULL;
2092                 struct bau_control *hmaster = NULL;
2093
2094                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2095                         continue;
2096
2097                 bdp = &uvhub_descs[uvhub];
2098                 socket_mask = bdp->socket_mask;
2099                 socket = 0;
2100                 while (socket_mask) {
2101                         struct socket_desc *sdp;
2102                         if ((socket_mask & 1)) {
2103                                 sdp = &bdp->socket[socket];
2104                                 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2105                                         return 1;
2106                                 make_per_cpu_thp(smaster);
2107                         }
2108                         socket++;
2109                         socket_mask = (socket_mask >> 1);
2110                 }
2111                 make_per_hub_cpumask(hmaster);
2112         }
2113         return 0;
2114 }
2115
2116 /*
2117  * initialize the bau_control structure for each cpu
2118  */
2119 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2120 {
2121         unsigned char *uvhub_mask;
2122         void *vp;
2123         struct uvhub_desc *uvhub_descs;
2124
2125         if (is_uv3_hub() || is_uv2_hub() || is_uv1_hub())
2126                 timeout_us = calculate_destination_timeout();
2127
2128         vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
2129         uvhub_descs = (struct uvhub_desc *)vp;
2130         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
2131         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2132
2133         if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2134                 goto fail;
2135
2136         if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2137                 goto fail;
2138
2139         kfree(uvhub_descs);
2140         kfree(uvhub_mask);
2141         init_per_cpu_tunables();
2142         return 0;
2143
2144 fail:
2145         kfree(uvhub_descs);
2146         kfree(uvhub_mask);
2147         return 1;
2148 }
2149
2150 /*
2151  * Initialization of BAU-related structures
2152  */
2153 static int __init uv_bau_init(void)
2154 {
2155         int uvhub;
2156         int pnode;
2157         int nuvhubs;
2158         int cur_cpu;
2159         int cpus;
2160         int vector;
2161         cpumask_var_t *mask;
2162
2163         if (!is_uv_system())
2164                 return 0;
2165
2166         if (is_uv4_hub())
2167                 ops = uv4_bau_ops;
2168         else if (is_uv3_hub())
2169                 ops = uv123_bau_ops;
2170         else if (is_uv2_hub())
2171                 ops = uv123_bau_ops;
2172         else if (is_uv1_hub())
2173                 ops = uv123_bau_ops;
2174
2175         for_each_possible_cpu(cur_cpu) {
2176                 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2177                 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2178         }
2179
2180         nuvhubs = uv_num_possible_blades();
2181         congested_cycles = usec_2_cycles(congested_respns_us);
2182
2183         uv_base_pnode = 0x7fffffff;
2184         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2185                 cpus = uv_blade_nr_possible_cpus(uvhub);
2186                 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2187                         uv_base_pnode = uv_blade_to_pnode(uvhub);
2188         }
2189
2190         /* software timeouts are not supported on UV4 */
2191         if (is_uv3_hub() || is_uv2_hub() || is_uv1_hub())
2192                 enable_timeouts();
2193
2194         if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2195                 set_bau_off();
2196                 nobau_perm = 1;
2197                 return 0;
2198         }
2199
2200         vector = UV_BAU_MESSAGE;
2201         for_each_possible_blade(uvhub) {
2202                 if (uv_blade_nr_possible_cpus(uvhub))
2203                         init_uvhub(uvhub, vector, uv_base_pnode);
2204         }
2205
2206         alloc_intr_gate(vector, uv_bau_message_intr1);
2207
2208         for_each_possible_blade(uvhub) {
2209                 if (uv_blade_nr_possible_cpus(uvhub)) {
2210                         unsigned long val;
2211                         unsigned long mmr;
2212                         pnode = uv_blade_to_pnode(uvhub);
2213                         /* INIT the bau */
2214                         val = 1L << 63;
2215                         write_gmmr_activation(pnode, val);
2216                         mmr = 1; /* should be 1 to broadcast to both sockets */
2217                         if (!is_uv1_hub())
2218                                 write_mmr_data_broadcast(pnode, mmr);
2219                 }
2220         }
2221
2222         return 0;
2223 }
2224 core_initcall(uv_bau_init);
2225 fs_initcall(uv_ptc_init);