arch/x86/platform/uv/tlb_uv.c

   1 /*
   2  *      SGI UltraViolet TLB flush routines.
   3  *
   4  *      (c) 2008-2010 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 static int timeout_us;
  39 static int nobau;
  40 static int baudisabled;
  41 static spinlock_t disable_lock;
  42 static cycles_t congested_cycles;
  43
  44 /* tunables: */
  45 static int max_bau_concurrent = MAX_BAU_CONCURRENT;
  46 static int max_bau_concurrent_constant = MAX_BAU_CONCURRENT;
  47 static int plugged_delay = PLUGGED_DELAY;
  48 static int plugsb4reset = PLUGSB4RESET;
  49 static int timeoutsb4reset = TIMEOUTSB4RESET;
  50 static int ipi_reset_limit = IPI_RESET_LIMIT;
  51 static int complete_threshold = COMPLETE_THRESHOLD;
  52 static int congested_response_us = CONGESTED_RESPONSE_US;
  53 static int congested_reps = CONGESTED_REPS;
  54 static int congested_period = CONGESTED_PERIOD;
  55 static struct dentry *tunables_dir;
  56 static struct dentry *tunables_file;
  57
  58 static int __init setup_nobau(char *arg)
  59 {
  60         nobau = 1;
  61         return 0;
  62 }
  63 early_param("nobau", setup_nobau);
  64
  65 /* base pnode in this partition */
  66 static int uv_partition_base_pnode __read_mostly;
  67 /* position of pnode (which is nasid>>1): */
  68 static int uv_nshift __read_mostly;
  69 static unsigned long uv_mmask __read_mostly;
  70
  71 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
  72 static DEFINE_PER_CPU(struct bau_control, bau_control);
  73 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
  74
  75 /*
  76  * Determine the first node on a uvhub. 'Nodes' are used for kernel
  77  * memory allocation.
  78  */
  79 static int __init uvhub_to_first_node(int uvhub)
  80 {
  81         int node, b;
  82
  83         for_each_online_node(node) {
  84                 b = uv_node_to_blade_id(node);
  85                 if (uvhub == b)
  86                         return node;
  87         }
  88         return -1;
  89 }
  90
  91 /*
  92  * Determine the apicid of the first cpu on a uvhub.
  93  */
  94 static int __init uvhub_to_first_apicid(int uvhub)
  95 {
  96         int cpu;
  97
  98         for_each_present_cpu(cpu)
  99                 if (uvhub == uv_cpu_to_blade_id(cpu))
 100                         return per_cpu(x86_cpu_to_apicid, cpu);
 101         return -1;
 102 }
 103
 104 /*
 105  * Free a software acknowledge hardware resource by clearing its Pending
 106  * bit. This will return a reply to the sender.
 107  * If the message has timed out, a reply has already been sent by the
 108  * hardware but the resource has not been released. In that case our
 109  * clear of the Timeout bit (as well) will free the resource. No reply will
 110  * be sent (the hardware will only do one reply per message).
 111  */
 112 static inline void uv_reply_to_message(struct msg_desc *mdp,
 113                                        struct bau_control *bcp)
 114 {
 115         unsigned long dw;
 116         struct bau_payload_queue_entry *msg;
 117
 118         msg = mdp->msg;
 119         if (!msg->canceled) {
 120                 dw = (msg->sw_ack_vector << UV_SW_ACK_NPENDING) |
 121                                                 msg->sw_ack_vector;
 122                 uv_write_local_mmr(
 123                                 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw);
 124         }
 125         msg->replied_to = 1;
 126         msg->sw_ack_vector = 0;
 127 }
 128
 129 /*
 130  * Process the receipt of a RETRY message
 131  */
 132 static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
 133                                             struct bau_control *bcp)
 134 {
 135         int i;
 136         int cancel_count = 0;
 137         int slot2;
 138         unsigned long msg_res;
 139         unsigned long mmr = 0;
 140         struct bau_payload_queue_entry *msg;
 141         struct bau_payload_queue_entry *msg2;
 142         struct ptc_stats *stat;
 143
 144         msg = mdp->msg;
 145         stat = bcp->statp;
 146         stat->d_retries++;
 147         /*
 148          * cancel any message from msg+1 to the retry itself
 149          */
 150         for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
 151                 if (msg2 > mdp->va_queue_last)
 152                         msg2 = mdp->va_queue_first;
 153                 if (msg2 == msg)
 154                         break;
 155
 156                 /* same conditions for cancellation as uv_do_reset */
 157                 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
 158                     (msg2->sw_ack_vector) && ((msg2->sw_ack_vector &
 159                         msg->sw_ack_vector) == 0) &&
 160                     (msg2->sending_cpu == msg->sending_cpu) &&
 161                     (msg2->msg_type != MSG_NOOP)) {
 162                         slot2 = msg2 - mdp->va_queue_first;
 163                         mmr = uv_read_local_mmr
 164                                 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
 165                         msg_res = msg2->sw_ack_vector;
 166                         /*
 167                          * This is a message retry; clear the resources held
 168                          * by the previous message only if they timed out.
 169                          * If it has not timed out we have an unexpected
 170                          * situation to report.
 171                          */
 172                         if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
 173                                 /*
 174                                  * is the resource timed out?
 175                                  * make everyone ignore the cancelled message.
 176                                  */
 177                                 msg2->canceled = 1;
 178                                 stat->d_canceled++;
 179                                 cancel_count++;
 180                                 uv_write_local_mmr(
 181                                     UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
 182                                         (msg_res << UV_SW_ACK_NPENDING) |
 183                                          msg_res);
 184                         }
 185                 }
 186         }
 187         if (!cancel_count)
 188                 stat->d_nocanceled++;
 189 }
 190
 191 /*
 192  * Do all the things a cpu should do for a TLB shootdown message.
 193  * Other cpu's may come here at the same time for this message.
 194  */
 195 static void uv_bau_process_message(struct msg_desc *mdp,
 196                                    struct bau_control *bcp)
 197 {
 198         int msg_ack_count;
 199         short socket_ack_count = 0;
 200         struct ptc_stats *stat;
 201         struct bau_payload_queue_entry *msg;
 202         struct bau_control *smaster = bcp->socket_master;
 203
 204         /*
 205          * This must be a normal message, or retry of a normal message
 206          */
 207         msg = mdp->msg;
 208         stat = bcp->statp;
 209         if (msg->address == TLB_FLUSH_ALL) {
 210                 local_flush_tlb();
 211                 stat->d_alltlb++;
 212         } else {
 213                 __flush_tlb_one(msg->address);
 214                 stat->d_onetlb++;
 215         }
 216         stat->d_requestee++;
 217
 218         /*
 219          * One cpu on each uvhub has the additional job on a RETRY
 220          * of releasing the resource held by the message that is
 221          * being retried.  That message is identified by sending
 222          * cpu number.
 223          */
 224         if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
 225                 uv_bau_process_retry_msg(mdp, bcp);
 226
 227         /*
 228          * This is a sw_ack message, so we have to reply to it.
 229          * Count each responding cpu on the socket. This avoids
 230          * pinging the count's cache line back and forth between
 231          * the sockets.
 232          */
 233         socket_ack_count = atomic_add_short_return(1, (struct atomic_short *)
 234                         &smaster->socket_acknowledge_count[mdp->msg_slot]);
 235         if (socket_ack_count == bcp->cpus_in_socket) {
 236                 /*
 237                  * Both sockets dump their completed count total into
 238                  * the message's count.
 239                  */
 240                 smaster->socket_acknowledge_count[mdp->msg_slot] = 0;
 241                 msg_ack_count = atomic_add_short_return(socket_ack_count,
 242                                 (struct atomic_short *)&msg->acknowledge_count);
 243
 244                 if (msg_ack_count == bcp->cpus_in_uvhub) {
 245                         /*
 246                          * All cpus in uvhub saw it; reply
 247                          */
 248                         uv_reply_to_message(mdp, bcp);
 249                 }
 250         }
 251
 252         return;
 253 }
 254
 255 /*
 256  * Determine the first cpu on a uvhub.
 257  */
 258 static int uvhub_to_first_cpu(int uvhub)
 259 {
 260         int cpu;
 261         for_each_present_cpu(cpu)
 262                 if (uvhub == uv_cpu_to_blade_id(cpu))
 263                         return cpu;
 264         return -1;
 265 }
 266
 267 /*
 268  * Last resort when we get a large number of destination timeouts is
 269  * to clear resources held by a given cpu.
 270  * Do this with IPI so that all messages in the BAU message queue
 271  * can be identified by their nonzero sw_ack_vector field.
 272  *
 273  * This is entered for a single cpu on the uvhub.
 274  * The sender want's this uvhub to free a specific message's
 275  * sw_ack resources.
 276  */
 277 static void
 278 uv_do_reset(void *ptr)
 279 {
 280         int i;
 281         int slot;
 282         int count = 0;
 283         unsigned long mmr;
 284         unsigned long msg_res;
 285         struct bau_control *bcp;
 286         struct reset_args *rap;
 287         struct bau_payload_queue_entry *msg;
 288         struct ptc_stats *stat;
 289
 290         bcp = &per_cpu(bau_control, smp_processor_id());
 291         rap = (struct reset_args *)ptr;
 292         stat = bcp->statp;
 293         stat->d_resets++;
 294
 295         /*
 296          * We're looking for the given sender, and
 297          * will free its sw_ack resource.
 298          * If all cpu's finally responded after the timeout, its
 299          * message 'replied_to' was set.
 300          */
 301         for (msg = bcp->va_queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
 302                 /* uv_do_reset: same conditions for cancellation as
 303                    uv_bau_process_retry_msg() */
 304                 if ((msg->replied_to == 0) &&
 305                     (msg->canceled == 0) &&
 306                     (msg->sending_cpu == rap->sender) &&
 307                     (msg->sw_ack_vector) &&
 308                     (msg->msg_type != MSG_NOOP)) {
 309                         /*
 310                          * make everyone else ignore this message
 311                          */
 312                         msg->canceled = 1;
 313                         slot = msg - bcp->va_queue_first;
 314                         count++;
 315                         /*
 316                          * only reset the resource if it is still pending
 317                          */
 318                         mmr = uv_read_local_mmr
 319                                         (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
 320                         msg_res = msg->sw_ack_vector;
 321                         if (mmr & msg_res) {
 322                                 stat->d_rcanceled++;
 323                                 uv_write_local_mmr(
 324                                     UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
 325                                         (msg_res << UV_SW_ACK_NPENDING) |
 326                                          msg_res);
 327                         }
 328                 }
 329         }
 330         return;
 331 }
 332
 333 /*
 334  * Use IPI to get all target uvhubs to release resources held by
 335  * a given sending cpu number.
 336  */
 337 static void uv_reset_with_ipi(struct bau_target_uvhubmask *distribution,
 338                               int sender)
 339 {
 340         int uvhub;
 341         int cpu;
 342         cpumask_t mask;
 343         struct reset_args reset_args;
 344
 345         reset_args.sender = sender;
 346
 347         cpus_clear(mask);
 348         /* find a single cpu for each uvhub in this distribution mask */
 349         for (uvhub = 0;
 350                     uvhub < sizeof(struct bau_target_uvhubmask) * BITSPERBYTE;
 351                     uvhub++) {
 352                 if (!bau_uvhub_isset(uvhub, distribution))
 353                         continue;
 354                 /* find a cpu for this uvhub */
 355                 cpu = uvhub_to_first_cpu(uvhub);
 356                 cpu_set(cpu, mask);
 357         }
 358         /* IPI all cpus; Preemption is already disabled */
 359         smp_call_function_many(&mask, uv_do_reset, (void *)&reset_args, 1);
 360         return;
 361 }
 362
 363 static inline unsigned long
 364 cycles_2_us(unsigned long long cyc)
 365 {
 366         unsigned long long ns;
 367         unsigned long us;
 368         ns =  (cyc * per_cpu(cyc2ns, smp_processor_id()))
 369                                                 >> CYC2NS_SCALE_FACTOR;
 370         us = ns / 1000;
 371         return us;
 372 }
 373
 374 /*
 375  * wait for all cpus on this hub to finish their sends and go quiet
 376  * leaves uvhub_quiesce set so that no new broadcasts are started by
 377  * bau_flush_send_and_wait()
 378  */
 379 static inline void
 380 quiesce_local_uvhub(struct bau_control *hmaster)
 381 {
 382         atomic_add_short_return(1, (struct atomic_short *)
 383                  &hmaster->uvhub_quiesce);
 384 }
 385
 386 /*
 387  * mark this quiet-requestor as done
 388  */
 389 static inline void
 390 end_uvhub_quiesce(struct bau_control *hmaster)
 391 {
 392         atomic_add_short_return(-1, (struct atomic_short *)
 393                 &hmaster->uvhub_quiesce);
 394 }
 395
 396 /*
 397  * Wait for completion of a broadcast software ack message
 398  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
 399  */
 400 static int uv_wait_completion(struct bau_desc *bau_desc,
 401         unsigned long mmr_offset, int right_shift, int this_cpu,
 402         struct bau_control *bcp, struct bau_control *smaster, long try)
 403 {
 404         unsigned long descriptor_status;
 405         cycles_t ttime;
 406         struct ptc_stats *stat = bcp->statp;
 407         struct bau_control *hmaster;
 408
 409         hmaster = bcp->uvhub_master;
 410
 411         /* spin on the status MMR, waiting for it to go idle */
 412         while ((descriptor_status = (((unsigned long)
 413                 uv_read_local_mmr(mmr_offset) >>
 414                         right_shift) & UV_ACT_STATUS_MASK)) !=
 415                         DESC_STATUS_IDLE) {
 416                 /*
 417                  * Our software ack messages may be blocked because there are
 418                  * no swack resources available.  As long as none of them
 419                  * has timed out hardware will NACK our message and its
 420                  * state will stay IDLE.
 421                  */
 422                 if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) {
 423                         stat->s_stimeout++;
 424                         return FLUSH_GIVEUP;
 425                 } else if (descriptor_status ==
 426                                         DESC_STATUS_DESTINATION_TIMEOUT) {
 427                         stat->s_dtimeout++;
 428                         ttime = get_cycles();
 429
 430                         /*
 431                          * Our retries may be blocked by all destination
 432                          * swack resources being consumed, and a timeout
 433                          * pending.  In that case hardware returns the
 434                          * ERROR that looks like a destination timeout.
 435                          */
 436                         if (cycles_2_us(ttime - bcp->send_message) <
 437                                                         timeout_us) {
 438                                 bcp->conseccompletes = 0;
 439                                 return FLUSH_RETRY_PLUGGED;
 440                         }
 441
 442                         bcp->conseccompletes = 0;
 443                         return FLUSH_RETRY_TIMEOUT;
 444                 } else {
 445                         /*
 446                          * descriptor_status is still BUSY
 447                          */
 448                         cpu_relax();
 449                 }
 450         }
 451         bcp->conseccompletes++;
 452         return FLUSH_COMPLETE;
 453 }
 454
 455 static inline cycles_t
 456 sec_2_cycles(unsigned long sec)
 457 {
 458         unsigned long ns;
 459         cycles_t cyc;
 460
 461         ns = sec * 1000000000;
 462         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
 463         return cyc;
 464 }
 465
 466 /*
 467  * conditionally add 1 to *v, unless *v is >= u
 468  * return 0 if we cannot add 1 to *v because it is >= u
 469  * return 1 if we can add 1 to *v because it is < u
 470  * the add is atomic
 471  *
 472  * This is close to atomic_add_unless(), but this allows the 'u' value
 473  * to be lowered below the current 'v'.  atomic_add_unless can only stop
 474  * on equal.
 475  */
 476 static inline int atomic_inc_unless_ge(spinlock_t *lock, atomic_t *v, int u)
 477 {
 478         spin_lock(lock);
 479         if (atomic_read(v) >= u) {
 480                 spin_unlock(lock);
 481                 return 0;
 482         }
 483         atomic_inc(v);
 484         spin_unlock(lock);
 485         return 1;
 486 }
 487
 488 /*
 489  * Our retries are blocked by all destination swack resources being
 490  * in use, and a timeout is pending. In that case hardware immediately
 491  * returns the ERROR that looks like a destination timeout.
 492  */
 493 static void
 494 destination_plugged(struct bau_desc *bau_desc, struct bau_control *bcp,
 495                         struct bau_control *hmaster, struct ptc_stats *stat)
 496 {
 497         udelay(bcp->plugged_delay);
 498         bcp->plugged_tries++;
 499         if (bcp->plugged_tries >= bcp->plugsb4reset) {
 500                 bcp->plugged_tries = 0;
 501                 quiesce_local_uvhub(hmaster);
 502                 spin_lock(&hmaster->queue_lock);
 503                 uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu);
 504                 spin_unlock(&hmaster->queue_lock);
 505                 end_uvhub_quiesce(hmaster);
 506                 bcp->ipi_attempts++;
 507                 stat->s_resets_plug++;
 508         }
 509 }
 510
 511 static void
 512 destination_timeout(struct bau_desc *bau_desc, struct bau_control *bcp,
 513                         struct bau_control *hmaster, struct ptc_stats *stat)
 514 {
 515         hmaster->max_bau_concurrent = 1;
 516         bcp->timeout_tries++;
 517         if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
 518                 bcp->timeout_tries = 0;
 519                 quiesce_local_uvhub(hmaster);
 520                 spin_lock(&hmaster->queue_lock);
 521                 uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu);
 522                 spin_unlock(&hmaster->queue_lock);
 523                 end_uvhub_quiesce(hmaster);
 524                 bcp->ipi_attempts++;
 525                 stat->s_resets_timeout++;
 526         }
 527 }
 528
 529 /*
 530  * Completions are taking a very long time due to a congested numalink
 531  * network.
 532  */
 533 static void
 534 disable_for_congestion(struct bau_control *bcp, struct ptc_stats *stat)
 535 {
 536         int tcpu;
 537         struct bau_control *tbcp;
 538
 539         /* let only one cpu do this disabling */
 540         spin_lock(&disable_lock);
 541         if (!baudisabled && bcp->period_requests &&
 542             ((bcp->period_time / bcp->period_requests) > congested_cycles)) {
 543                 /* it becomes this cpu's job to turn on the use of the
 544                    BAU again */
 545                 baudisabled = 1;
 546                 bcp->set_bau_off = 1;
 547                 bcp->set_bau_on_time = get_cycles() +
 548                         sec_2_cycles(bcp->congested_period);
 549                 stat->s_bau_disabled++;
 550                 for_each_present_cpu(tcpu) {
 551                         tbcp = &per_cpu(bau_control, tcpu);
 552                                 tbcp->baudisabled = 1;
 553                 }
 554         }
 555         spin_unlock(&disable_lock);
 556 }
 557
 558 /**
 559  * uv_flush_send_and_wait
 560  *
 561  * Send a broadcast and wait for it to complete.
 562  *
 563  * The flush_mask contains the cpus the broadcast is to be sent to including
 564  * cpus that are on the local uvhub.
 565  *
 566  * Returns 0 if all flushing represented in the mask was done.
 567  * Returns 1 if it gives up entirely and the original cpu mask is to be
 568  * returned to the kernel.
 569  */
 570 int uv_flush_send_and_wait(struct bau_desc *bau_desc,
 571                            struct cpumask *flush_mask, struct bau_control *bcp)
 572 {
 573         int right_shift;
 574         int completion_status = 0;
 575         int seq_number = 0;
 576         long try = 0;
 577         int cpu = bcp->uvhub_cpu;
 578         int this_cpu = bcp->cpu;
 579         unsigned long mmr_offset;
 580         unsigned long index;
 581         cycles_t time1;
 582         cycles_t time2;
 583         cycles_t elapsed;
 584         struct ptc_stats *stat = bcp->statp;
 585         struct bau_control *smaster = bcp->socket_master;
 586         struct bau_control *hmaster = bcp->uvhub_master;
 587
 588         if (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
 589                         &hmaster->active_descriptor_count,
 590                         hmaster->max_bau_concurrent)) {
 591                 stat->s_throttles++;
 592                 do {
 593                         cpu_relax();
 594                 } while (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
 595                         &hmaster->active_descriptor_count,
 596                         hmaster->max_bau_concurrent));
 597         }
 598         while (hmaster->uvhub_quiesce)
 599                 cpu_relax();
 600
 601         if (cpu < UV_CPUS_PER_ACT_STATUS) {
 602                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
 603                 right_shift = cpu * UV_ACT_STATUS_SIZE;
 604         } else {
 605                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
 606                 right_shift =
 607                     ((cpu - UV_CPUS_PER_ACT_STATUS) * UV_ACT_STATUS_SIZE);
 608         }
 609         time1 = get_cycles();
 610         do {
 611                 if (try == 0) {
 612                         bau_desc->header.msg_type = MSG_REGULAR;
 613                         seq_number = bcp->message_number++;
 614                 } else {
 615                         bau_desc->header.msg_type = MSG_RETRY;
 616                         stat->s_retry_messages++;
 617                 }
 618                 bau_desc->header.sequence = seq_number;
 619                 index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) |
 620                         bcp->uvhub_cpu;
 621                 bcp->send_message = get_cycles();
 622                 uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index);
 623                 try++;
 624                 completion_status = uv_wait_completion(bau_desc, mmr_offset,
 625                         right_shift, this_cpu, bcp, smaster, try);
 626
 627                 if (completion_status == FLUSH_RETRY_PLUGGED) {
 628                         destination_plugged(bau_desc, bcp, hmaster, stat);
 629                 } else if (completion_status == FLUSH_RETRY_TIMEOUT) {
 630                         destination_timeout(bau_desc, bcp, hmaster, stat);
 631                 }
 632                 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
 633                         bcp->ipi_attempts = 0;
 634                         completion_status = FLUSH_GIVEUP;
 635                         break;
 636                 }
 637                 cpu_relax();
 638         } while ((completion_status == FLUSH_RETRY_PLUGGED) ||
 639                  (completion_status == FLUSH_RETRY_TIMEOUT));
 640         time2 = get_cycles();
 641         bcp->plugged_tries = 0;
 642         bcp->timeout_tries = 0;
 643         if ((completion_status == FLUSH_COMPLETE) &&
 644             (bcp->conseccompletes > bcp->complete_threshold) &&
 645             (hmaster->max_bau_concurrent <
 646                                         hmaster->max_bau_concurrent_constant))
 647                         hmaster->max_bau_concurrent++;
 648         while (hmaster->uvhub_quiesce)
 649                 cpu_relax();
 650         atomic_dec(&hmaster->active_descriptor_count);
 651         if (time2 > time1) {
 652                 elapsed = time2 - time1;
 653                 stat->s_time += elapsed;
 654                 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
 655                         bcp->period_requests++;
 656                         bcp->period_time += elapsed;
 657                         if ((elapsed > congested_cycles) &&
 658                             (bcp->period_requests > bcp->congested_reps)) {
 659                                 disable_for_congestion(bcp, stat);
 660                         }
 661                 }
 662         } else
 663                 stat->s_requestor--;
 664         if (completion_status == FLUSH_COMPLETE && try > 1)
 665                 stat->s_retriesok++;
 666         else if (completion_status == FLUSH_GIVEUP) {
 667                 stat->s_giveup++;
 668                 return 1;
 669         }
 670         return 0;
 671 }
 672
 673 /**
 674  * uv_flush_tlb_others - globally purge translation cache of a virtual
 675  * address or all TLB's
 676  * @cpumask: mask of all cpu's in which the address is to be removed
 677  * @mm: mm_struct containing virtual address range
 678  * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
 679  * @cpu: the current cpu
 680  *
 681  * This is the entry point for initiating any UV global TLB shootdown.
 682  *
 683  * Purges the translation caches of all specified processors of the given
 684  * virtual address, or purges all TLB's on specified processors.
 685  *
 686  * The caller has derived the cpumask from the mm_struct.  This function
 687  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
 688  *
 689  * The cpumask is converted into a uvhubmask of the uvhubs containing
 690  * those cpus.
 691  *
 692  * Note that this function should be called with preemption disabled.
 693  *
 694  * Returns NULL if all remote flushing was done.
 695  * Returns pointer to cpumask if some remote flushing remains to be
 696  * done.  The returned pointer is valid till preemption is re-enabled.
 697  */
 698 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
 699                                           struct mm_struct *mm,
 700                                           unsigned long va, unsigned int cpu)
 701 {
 702         int locals = 0;
 703         int remotes = 0;
 704         int hubs = 0;
 705         int tcpu;
 706         int tpnode;
 707         struct bau_desc *bau_desc;
 708         struct cpumask *flush_mask;
 709         struct ptc_stats *stat;
 710         struct bau_control *bcp;
 711         struct bau_control *tbcp;
 712         struct hub_and_pnode *hpp;
 713
 714         /* kernel was booted 'nobau' */
 715         if (nobau)
 716                 return cpumask;
 717
 718         bcp = &per_cpu(bau_control, cpu);
 719         stat = bcp->statp;
 720
 721         /* bau was disabled due to slow response */
 722         if (bcp->baudisabled) {
 723                 /* the cpu that disabled it must re-enable it */
 724                 if (bcp->set_bau_off) {
 725                         if (get_cycles() >= bcp->set_bau_on_time) {
 726                                 stat->s_bau_reenabled++;
 727                                 baudisabled = 0;
 728                                 for_each_present_cpu(tcpu) {
 729                                         tbcp = &per_cpu(bau_control, tcpu);
 730                                         tbcp->baudisabled = 0;
 731                                         tbcp->period_requests = 0;
 732                                         tbcp->period_time = 0;
 733                                 }
 734                         }
 735                 }
 736                 return cpumask;
 737         }
 738
 739         /*
 740          * Each sending cpu has a per-cpu mask which it fills from the caller's
 741          * cpu mask.  All cpus are converted to uvhubs and copied to the
 742          * activation descriptor.
 743          */
 744         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
 745         /* don't actually do a shootdown of the local cpu */
 746         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
 747         if (cpu_isset(cpu, *cpumask))
 748                 stat->s_ntargself++;
 749
 750         bau_desc = bcp->descriptor_base;
 751         bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu;
 752         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
 753
 754         for_each_cpu(tcpu, flush_mask) {
 755                 /*
 756                  * The distribution vector is a bit map of pnodes, relative
 757                  * to the partition base pnode (and the partition base nasid
 758                  * in the header).
 759                  * Translate cpu to pnode and hub using an array stored
 760                  * in local memory.
 761                  */
 762                 hpp = &bcp->socket_master->target_hub_and_pnode[tcpu];
 763                 tpnode = hpp->pnode - bcp->partition_base_pnode;
 764                 bau_uvhub_set(tpnode, &bau_desc->distribution);
 765                 if (hpp->uvhub == bcp->uvhub)
 766                         locals++;
 767                 else
 768                         remotes++;
 769         }
 770         if ((locals + remotes) == 0)
 771                 return NULL;
 772         stat->s_requestor++;
 773         stat->s_ntargcpu += remotes + locals;
 774         stat->s_ntargremotes += remotes;
 775         stat->s_ntarglocals += locals;
 776         remotes = bau_uvhub_weight(&bau_desc->distribution);
 777
 778         /* uvhub statistics */
 779         hubs = bau_uvhub_weight(&bau_desc->distribution);
 780         if (locals) {
 781                 stat->s_ntarglocaluvhub++;
 782                 stat->s_ntargremoteuvhub += (hubs - 1);
 783         } else
 784                 stat->s_ntargremoteuvhub += hubs;
 785         stat->s_ntarguvhub += hubs;
 786         if (hubs >= 16)
 787                 stat->s_ntarguvhub16++;
 788         else if (hubs >= 8)
 789                 stat->s_ntarguvhub8++;
 790         else if (hubs >= 4)
 791                 stat->s_ntarguvhub4++;
 792         else if (hubs >= 2)
 793                 stat->s_ntarguvhub2++;
 794         else
 795                 stat->s_ntarguvhub1++;
 796
 797         bau_desc->payload.address = va;
 798         bau_desc->payload.sending_cpu = cpu;
 799
 800         /*
 801          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
 802          * or 1 if it gave up and the original cpumask should be returned.
 803          */
 804         if (!uv_flush_send_and_wait(bau_desc, flush_mask, bcp))
 805                 return NULL;
 806         else
 807                 return cpumask;
 808 }
 809
 810 /*
 811  * The BAU message interrupt comes here. (registered by set_intr_gate)
 812  * See entry_64.S
 813  *
 814  * We received a broadcast assist message.
 815  *
 816  * Interrupts are disabled; this interrupt could represent
 817  * the receipt of several messages.
 818  *
 819  * All cores/threads on this hub get this interrupt.
 820  * The last one to see it does the software ack.
 821  * (the resource will not be freed until noninterruptable cpus see this
 822  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
 823  */
 824 void uv_bau_message_interrupt(struct pt_regs *regs)
 825 {
 826         int count = 0;
 827         cycles_t time_start;
 828         struct bau_payload_queue_entry *msg;
 829         struct bau_control *bcp;
 830         struct ptc_stats *stat;
 831         struct msg_desc msgdesc;
 832
 833         time_start = get_cycles();
 834         bcp = &per_cpu(bau_control, smp_processor_id());
 835         stat = bcp->statp;
 836         msgdesc.va_queue_first = bcp->va_queue_first;
 837         msgdesc.va_queue_last = bcp->va_queue_last;
 838         msg = bcp->bau_msg_head;
 839         while (msg->sw_ack_vector) {
 840                 count++;
 841                 msgdesc.msg_slot = msg - msgdesc.va_queue_first;
 842                 msgdesc.sw_ack_slot = ffs(msg->sw_ack_vector) - 1;
 843                 msgdesc.msg = msg;
 844                 uv_bau_process_message(&msgdesc, bcp);
 845                 msg++;
 846                 if (msg > msgdesc.va_queue_last)
 847                         msg = msgdesc.va_queue_first;
 848                 bcp->bau_msg_head = msg;
 849         }
 850         stat->d_time += (get_cycles() - time_start);
 851         if (!count)
 852                 stat->d_nomsg++;
 853         else if (count > 1)
 854                 stat->d_multmsg++;
 855         ack_APIC_irq();
 856 }
 857
 858 /*
 859  * uv_enable_timeouts
 860  *
 861  * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
 862  * shootdown message timeouts enabled.  The timeout does not cause
 863  * an interrupt, but causes an error message to be returned to
 864  * the sender.
 865  */
 866 static void __init uv_enable_timeouts(void)
 867 {
 868         int uvhub;
 869         int nuvhubs;
 870         int pnode;
 871         unsigned long mmr_image;
 872
 873         nuvhubs = uv_num_possible_blades();
 874
 875         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
 876                 if (!uv_blade_nr_possible_cpus(uvhub))
 877                         continue;
 878
 879                 pnode = uv_blade_to_pnode(uvhub);
 880                 mmr_image =
 881                     uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL);
 882                 /*
 883                  * Set the timeout period and then lock it in, in three
 884                  * steps; captures and locks in the period.
 885                  *
 886                  * To program the period, the SOFT_ACK_MODE must be off.
 887                  */
 888                 mmr_image &= ~((unsigned long)1 <<
 889                     UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 890                 uv_write_global_mmr64
 891                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 892                 /*
 893                  * Set the 4-bit period.
 894                  */
 895                 mmr_image &= ~((unsigned long)0xf <<
 896                      UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 897                 mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD <<
 898                      UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 899                 uv_write_global_mmr64
 900                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 901                 /*
 902                  * Subsequent reversals of the timebase bit (3) cause an
 903                  * immediate timeout of one or all INTD resources as
 904                  * indicated in bits 2:0 (7 causes all of them to timeout).
 905                  */
 906                 mmr_image |= ((unsigned long)1 <<
 907                     UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 908                 uv_write_global_mmr64
 909                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 910         }
 911 }
 912
 913 static void *uv_ptc_seq_start(struct seq_file *file, loff_t *offset)
 914 {
 915         if (*offset < num_possible_cpus())
 916                 return offset;
 917         return NULL;
 918 }
 919
 920 static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
 921 {
 922         (*offset)++;
 923         if (*offset < num_possible_cpus())
 924                 return offset;
 925         return NULL;
 926 }
 927
 928 static void uv_ptc_seq_stop(struct seq_file *file, void *data)
 929 {
 930 }
 931
 932 static inline unsigned long long
 933 microsec_2_cycles(unsigned long microsec)
 934 {
 935         unsigned long ns;
 936         unsigned long long cyc;
 937
 938         ns = microsec * 1000;
 939         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
 940         return cyc;
 941 }
 942
 943 /*
 944  * Display the statistics thru /proc.
 945  * 'data' points to the cpu number
 946  */
 947 static int uv_ptc_seq_show(struct seq_file *file, void *data)
 948 {
 949         struct ptc_stats *stat;
 950         int cpu;
 951
 952         cpu = *(loff_t *)data;
 953
 954         if (!cpu) {
 955                 seq_printf(file,
 956                         "# cpu sent stime self locals remotes ncpus localhub ");
 957                 seq_printf(file,
 958                         "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
 959                 seq_printf(file,
 960                         "numuvhubs4 numuvhubs2 numuvhubs1 dto ");
 961                 seq_printf(file,
 962                         "retries rok resetp resett giveup sto bz throt ");
 963                 seq_printf(file,
 964                         "sw_ack recv rtime all ");
 965                 seq_printf(file,
 966                         "one mult none retry canc nocan reset rcan ");
 967                 seq_printf(file,
 968                         "disable enable\n");
 969         }
 970         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
 971                 stat = &per_cpu(ptcstats, cpu);
 972                 /* source side statistics */
 973                 seq_printf(file,
 974                         "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
 975                            cpu, stat->s_requestor, cycles_2_us(stat->s_time),
 976                            stat->s_ntargself, stat->s_ntarglocals,
 977                            stat->s_ntargremotes, stat->s_ntargcpu,
 978                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
 979                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
 980                 seq_printf(file, "%ld %ld %ld %ld %ld ",
 981                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
 982                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
 983                            stat->s_dtimeout);
 984                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
 985                            stat->s_retry_messages, stat->s_retriesok,
 986                            stat->s_resets_plug, stat->s_resets_timeout,
 987                            stat->s_giveup, stat->s_stimeout,
 988                            stat->s_busy, stat->s_throttles);
 989
 990                 /* destination side statistics */
 991                 seq_printf(file,
 992                            "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
 993                            uv_read_global_mmr64(uv_cpu_to_pnode(cpu),
 994                                         UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE),
 995                            stat->d_requestee, cycles_2_us(stat->d_time),
 996                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
 997                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
 998                            stat->d_nocanceled, stat->d_resets,
 999                            stat->d_rcanceled);
1000                 seq_printf(file, "%ld %ld\n",
1001                         stat->s_bau_disabled, stat->s_bau_reenabled);
1002         }
1003
1004         return 0;
1005 }
1006
1007 /*
1008  * Display the tunables thru debugfs
1009  */
1010 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1011                                                 size_t count, loff_t *ppos)
1012 {
1013         char *buf;
1014         int ret;
1015
1016         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1017                 "max_bau_concurrent plugged_delay plugsb4reset",
1018                 "timeoutsb4reset ipi_reset_limit complete_threshold",
1019                 "congested_response_us congested_reps congested_period",
1020                 max_bau_concurrent, plugged_delay, plugsb4reset,
1021                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1022                 congested_response_us, congested_reps, congested_period);
1023
1024         if (!buf)
1025                 return -ENOMEM;
1026
1027         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1028         kfree(buf);
1029         return ret;
1030 }
1031
1032 /*
1033  * -1: resetf the statistics
1034  *  0: display meaning of the statistics
1035  */
1036 static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user,
1037                                  size_t count, loff_t *data)
1038 {
1039         int cpu;
1040         long input_arg;
1041         char optstr[64];
1042         struct ptc_stats *stat;
1043
1044         if (count == 0 || count > sizeof(optstr))
1045                 return -EINVAL;
1046         if (copy_from_user(optstr, user, count))
1047                 return -EFAULT;
1048         optstr[count - 1] = '\0';
1049         if (strict_strtol(optstr, 10, &input_arg) < 0) {
1050                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1051                 return -EINVAL;
1052         }
1053
1054         if (input_arg == 0) {
1055                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1056                 printk(KERN_DEBUG "Sender statistics:\n");
1057                 printk(KERN_DEBUG
1058                 "sent:     number of shootdown messages sent\n");
1059                 printk(KERN_DEBUG
1060                 "stime:    time spent sending messages\n");
1061                 printk(KERN_DEBUG
1062                 "numuvhubs: number of hubs targeted with shootdown\n");
1063                 printk(KERN_DEBUG
1064                 "numuvhubs16: number times 16 or more hubs targeted\n");
1065                 printk(KERN_DEBUG
1066                 "numuvhubs8: number times 8 or more hubs targeted\n");
1067                 printk(KERN_DEBUG
1068                 "numuvhubs4: number times 4 or more hubs targeted\n");
1069                 printk(KERN_DEBUG
1070                 "numuvhubs2: number times 2 or more hubs targeted\n");
1071                 printk(KERN_DEBUG
1072                 "numuvhubs1: number times 1 hub targeted\n");
1073                 printk(KERN_DEBUG
1074                 "numcpus:  number of cpus targeted with shootdown\n");
1075                 printk(KERN_DEBUG
1076                 "dto:      number of destination timeouts\n");
1077                 printk(KERN_DEBUG
1078                 "retries:  destination timeout retries sent\n");
1079                 printk(KERN_DEBUG
1080                 "rok:   :  destination timeouts successfully retried\n");
1081                 printk(KERN_DEBUG
1082                 "resetp:   ipi-style resource resets for plugs\n");
1083                 printk(KERN_DEBUG
1084                 "resett:   ipi-style resource resets for timeouts\n");
1085                 printk(KERN_DEBUG
1086                 "giveup:   fall-backs to ipi-style shootdowns\n");
1087                 printk(KERN_DEBUG
1088                 "sto:      number of source timeouts\n");
1089                 printk(KERN_DEBUG
1090                 "bz:       number of stay-busy's\n");
1091                 printk(KERN_DEBUG
1092                 "throt:    number times spun in throttle\n");
1093                 printk(KERN_DEBUG "Destination side statistics:\n");
1094                 printk(KERN_DEBUG
1095                 "sw_ack:   image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
1096                 printk(KERN_DEBUG
1097                 "recv:     shootdown messages received\n");
1098                 printk(KERN_DEBUG
1099                 "rtime:    time spent processing messages\n");
1100                 printk(KERN_DEBUG
1101                 "all:      shootdown all-tlb messages\n");
1102                 printk(KERN_DEBUG
1103                 "one:      shootdown one-tlb messages\n");
1104                 printk(KERN_DEBUG
1105                 "mult:     interrupts that found multiple messages\n");
1106                 printk(KERN_DEBUG
1107                 "none:     interrupts that found no messages\n");
1108                 printk(KERN_DEBUG
1109                 "retry:    number of retry messages processed\n");
1110                 printk(KERN_DEBUG
1111                 "canc:     number messages canceled by retries\n");
1112                 printk(KERN_DEBUG
1113                 "nocan:    number retries that found nothing to cancel\n");
1114                 printk(KERN_DEBUG
1115                 "reset:    number of ipi-style reset requests processed\n");
1116                 printk(KERN_DEBUG
1117                 "rcan:     number messages canceled by reset requests\n");
1118                 printk(KERN_DEBUG
1119                 "disable:  number times use of the BAU was disabled\n");
1120                 printk(KERN_DEBUG
1121                 "enable:   number times use of the BAU was re-enabled\n");
1122         } else if (input_arg == -1) {
1123                 for_each_present_cpu(cpu) {
1124                         stat = &per_cpu(ptcstats, cpu);
1125                         memset(stat, 0, sizeof(struct ptc_stats));
1126                 }
1127         }
1128
1129         return count;
1130 }
1131
1132 static int local_atoi(const char *name)
1133 {
1134         int val = 0;
1135
1136         for (;; name++) {
1137                 switch (*name) {
1138                 case '0' ... '9':
1139                         val = 10*val+(*name-'0');
1140                         break;
1141                 default:
1142                         return val;
1143                 }
1144         }
1145 }
1146
1147 /*
1148  * set the tunables
1149  * 0 values reset them to defaults
1150  */
1151 static ssize_t tunables_write(struct file *file, const char __user *user,
1152                                  size_t count, loff_t *data)
1153 {
1154         int cpu;
1155         int cnt = 0;
1156         int val;
1157         char *p;
1158         char *q;
1159         char instr[64];
1160         struct bau_control *bcp;
1161
1162         if (count == 0 || count > sizeof(instr)-1)
1163                 return -EINVAL;
1164         if (copy_from_user(instr, user, count))
1165                 return -EFAULT;
1166
1167         instr[count] = '\0';
1168         /* count the fields */
1169         p = instr + strspn(instr, WHITESPACE);
1170         q = p;
1171         for (; *p; p = q + strspn(q, WHITESPACE)) {
1172                 q = p + strcspn(p, WHITESPACE);
1173                 cnt++;
1174                 if (q == p)
1175                         break;
1176         }
1177         if (cnt != 9) {
1178                 printk(KERN_INFO "bau tunable error: should be 9 numbers\n");
1179                 return -EINVAL;
1180         }
1181
1182         p = instr + strspn(instr, WHITESPACE);
1183         q = p;
1184         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1185                 q = p + strcspn(p, WHITESPACE);
1186                 val = local_atoi(p);
1187                 switch (cnt) {
1188                 case 0:
1189                         if (val == 0) {
1190                                 max_bau_concurrent = MAX_BAU_CONCURRENT;
1191                                 max_bau_concurrent_constant =
1192                                                         MAX_BAU_CONCURRENT;
1193                                 continue;
1194                         }
1195                         bcp = &per_cpu(bau_control, smp_processor_id());
1196                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1197                                 printk(KERN_DEBUG
1198                                 "Error: BAU max concurrent %d is invalid\n",
1199                                 val);
1200                                 return -EINVAL;
1201                         }
1202                         max_bau_concurrent = val;
1203                         max_bau_concurrent_constant = val;
1204                         continue;
1205                 case 1:
1206                         if (val == 0)
1207                                 plugged_delay = PLUGGED_DELAY;
1208                         else
1209                                 plugged_delay = val;
1210                         continue;
1211                 case 2:
1212                         if (val == 0)
1213                                 plugsb4reset = PLUGSB4RESET;
1214                         else
1215                                 plugsb4reset = val;
1216                         continue;
1217                 case 3:
1218                         if (val == 0)
1219                                 timeoutsb4reset = TIMEOUTSB4RESET;
1220                         else
1221                                 timeoutsb4reset = val;
1222                         continue;
1223                 case 4:
1224                         if (val == 0)
1225                                 ipi_reset_limit = IPI_RESET_LIMIT;
1226                         else
1227                                 ipi_reset_limit = val;
1228                         continue;
1229                 case 5:
1230                         if (val == 0)
1231                                 complete_threshold = COMPLETE_THRESHOLD;
1232                         else
1233                                 complete_threshold = val;
1234                         continue;
1235                 case 6:
1236                         if (val == 0)
1237                                 congested_response_us = CONGESTED_RESPONSE_US;
1238                         else
1239                                 congested_response_us = val;
1240                         continue;
1241                 case 7:
1242                         if (val == 0)
1243                                 congested_reps = CONGESTED_REPS;
1244                         else
1245                                 congested_reps = val;
1246                         continue;
1247                 case 8:
1248                         if (val == 0)
1249                                 congested_period = CONGESTED_PERIOD;
1250                         else
1251                                 congested_period = val;
1252                         continue;
1253                 }
1254                 if (q == p)
1255                         break;
1256         }
1257         for_each_present_cpu(cpu) {
1258                 bcp = &per_cpu(bau_control, cpu);
1259                 bcp->max_bau_concurrent = max_bau_concurrent;
1260                 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1261                 bcp->plugged_delay = plugged_delay;
1262                 bcp->plugsb4reset = plugsb4reset;
1263                 bcp->timeoutsb4reset = timeoutsb4reset;
1264                 bcp->ipi_reset_limit = ipi_reset_limit;
1265                 bcp->complete_threshold = complete_threshold;
1266                 bcp->congested_response_us = congested_response_us;
1267                 bcp->congested_reps = congested_reps;
1268                 bcp->congested_period = congested_period;
1269         }
1270         return count;
1271 }
1272
1273 static const struct seq_operations uv_ptc_seq_ops = {
1274         .start          = uv_ptc_seq_start,
1275         .next           = uv_ptc_seq_next,
1276         .stop           = uv_ptc_seq_stop,
1277         .show           = uv_ptc_seq_show
1278 };
1279
1280 static int uv_ptc_proc_open(struct inode *inode, struct file *file)
1281 {
1282         return seq_open(file, &uv_ptc_seq_ops);
1283 }
1284
1285 static int tunables_open(struct inode *inode, struct file *file)
1286 {
1287         return 0;
1288 }
1289
1290 static const struct file_operations proc_uv_ptc_operations = {
1291         .open           = uv_ptc_proc_open,
1292         .read           = seq_read,
1293         .write          = uv_ptc_proc_write,
1294         .llseek         = seq_lseek,
1295         .release        = seq_release,
1296 };
1297
1298 static const struct file_operations tunables_fops = {
1299         .open           = tunables_open,
1300         .read           = tunables_read,
1301         .write          = tunables_write,
1302         .llseek         = default_llseek,
1303 };
1304
1305 static int __init uv_ptc_init(void)
1306 {
1307         struct proc_dir_entry *proc_uv_ptc;
1308
1309         if (!is_uv_system())
1310                 return 0;
1311
1312         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1313                                   &proc_uv_ptc_operations);
1314         if (!proc_uv_ptc) {
1315                 printk(KERN_ERR "unable to create %s proc entry\n",
1316                        UV_PTC_BASENAME);
1317                 return -EINVAL;
1318         }
1319
1320         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1321         if (!tunables_dir) {
1322                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1323                        UV_BAU_TUNABLES_DIR);
1324                 return -EINVAL;
1325         }
1326         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1327                         tunables_dir, NULL, &tunables_fops);
1328         if (!tunables_file) {
1329                 printk(KERN_ERR "unable to create debugfs file %s\n",
1330                        UV_BAU_TUNABLES_FILE);
1331                 return -EINVAL;
1332         }
1333         return 0;
1334 }
1335
1336 /*
1337  * Initialize the sending side's sending buffers.
1338  */
1339 static void
1340 uv_activation_descriptor_init(int node, int pnode, int base_pnode)
1341 {
1342         int i;
1343         int cpu;
1344         unsigned long pa;
1345         unsigned long m;
1346         unsigned long n;
1347         struct bau_desc *bau_desc;
1348         struct bau_desc *bd2;
1349         struct bau_control *bcp;
1350
1351         /*
1352          * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR)
1353          * per cpu; and one per cpu on the uvhub (UV_ADP_SIZE)
1354          */
1355         bau_desc = kmalloc_node(sizeof(struct bau_desc) * UV_ADP_SIZE
1356                                 * UV_ITEMS_PER_DESCRIPTOR, GFP_KERNEL, node);
1357         BUG_ON(!bau_desc);
1358
1359         pa = uv_gpa(bau_desc); /* need the real nasid*/
1360         n = pa >> uv_nshift;
1361         m = pa & uv_mmask;
1362
1363         /* the 14-bit pnode */
1364         uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE,
1365                               (n << UV_DESC_BASE_PNODE_SHIFT | m));
1366         /*
1367          * Initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each
1368          * cpu even though we only use the first one; one descriptor can
1369          * describe a broadcast to 256 uv hubs.
1370          */
1371         for (i = 0, bd2 = bau_desc; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR);
1372                 i++, bd2++) {
1373                 memset(bd2, 0, sizeof(struct bau_desc));
1374                 bd2->header.sw_ack_flag = 1;
1375                 /*
1376                  * The base_dest_nasid set in the message header is the nasid
1377                  * of the first uvhub in the partition. The bit map will
1378                  * indicate destination pnode numbers relative to that base.
1379                  * They may not be consecutive if nasid striding is being used.
1380                  */
1381                 bd2->header.base_dest_nasid = UV_PNODE_TO_NASID(base_pnode);
1382                 bd2->header.dest_subnodeid = UV_LB_SUBNODEID;
1383                 bd2->header.command = UV_NET_ENDPOINT_INTD;
1384                 bd2->header.int_both = 1;
1385                 /*
1386                  * all others need to be set to zero:
1387                  *   fairness chaining multilevel count replied_to
1388                  */
1389         }
1390         for_each_present_cpu(cpu) {
1391                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1392                         continue;
1393                 bcp = &per_cpu(bau_control, cpu);
1394                 bcp->descriptor_base = bau_desc;
1395         }
1396 }
1397
1398 /*
1399  * initialize the destination side's receiving buffers
1400  * entered for each uvhub in the partition
1401  * - node is first node (kernel memory notion) on the uvhub
1402  * - pnode is the uvhub's physical identifier
1403  */
1404 static void
1405 uv_payload_queue_init(int node, int pnode)
1406 {
1407         int pn;
1408         int cpu;
1409         char *cp;
1410         unsigned long pa;
1411         struct bau_payload_queue_entry *pqp;
1412         struct bau_payload_queue_entry *pqp_malloc;
1413         struct bau_control *bcp;
1414
1415         pqp = kmalloc_node((DEST_Q_SIZE + 1)
1416                            * sizeof(struct bau_payload_queue_entry),
1417                            GFP_KERNEL, node);
1418         BUG_ON(!pqp);
1419         pqp_malloc = pqp;
1420
1421         cp = (char *)pqp + 31;
1422         pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5);
1423
1424         for_each_present_cpu(cpu) {
1425                 if (pnode != uv_cpu_to_pnode(cpu))
1426                         continue;
1427                 /* for every cpu on this pnode: */
1428                 bcp = &per_cpu(bau_control, cpu);
1429                 bcp->va_queue_first = pqp;
1430                 bcp->bau_msg_head = pqp;
1431                 bcp->va_queue_last = pqp + (DEST_Q_SIZE - 1);
1432         }
1433         /*
1434          * need the pnode of where the memory was really allocated
1435          */
1436         pa = uv_gpa(pqp);
1437         pn = pa >> uv_nshift;
1438         uv_write_global_mmr64(pnode,
1439                               UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST,
1440                               ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) |
1441                               uv_physnodeaddr(pqp));
1442         uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL,
1443                               uv_physnodeaddr(pqp));
1444         uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST,
1445                               (unsigned long)
1446                               uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1)));
1447         /* in effect, all msg_type's are set to MSG_NOOP */
1448         memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE);
1449 }
1450
1451 /*
1452  * Initialization of each UV hub's structures
1453  */
1454 static void __init uv_init_uvhub(int uvhub, int vector, int base_pnode)
1455 {
1456         int node;
1457         int pnode;
1458         unsigned long apicid;
1459
1460         node = uvhub_to_first_node(uvhub);
1461         pnode = uv_blade_to_pnode(uvhub);
1462         uv_activation_descriptor_init(node, pnode, base_pnode);
1463         uv_payload_queue_init(node, pnode);
1464         /*
1465          * The below initialization can't be in firmware because the
1466          * messaging IRQ will be determined by the OS.
1467          */
1468         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1469         uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG,
1470                                       ((apicid << 32) | vector));
1471 }
1472
1473 /*
1474  * We will set BAU_MISC_CONTROL with a timeout period.
1475  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1476  * So the destination timeout period has be be calculated from them.
1477  */
1478 static int
1479 calculate_destination_timeout(void)
1480 {
1481         unsigned long mmr_image;
1482         int mult1;
1483         int mult2;
1484         int index;
1485         int base;
1486         int ret;
1487         unsigned long ts_ns;
1488
1489         mult1 = UV_INTD_SOFT_ACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1490         mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1491         index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1492         mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1493         mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1494         base = timeout_base_ns[index];
1495         ts_ns = base * mult1 * mult2;
1496         ret = ts_ns / 1000;
1497         return ret;
1498 }
1499
1500 /*
1501  * initialize the bau_control structure for each cpu
1502  */
1503 static int __init uv_init_per_cpu(int nuvhubs, int base_part_pnode)
1504 {
1505         int i;
1506         int cpu;
1507         int tcpu;
1508         int pnode;
1509         int uvhub;
1510         int have_hmaster;
1511         short socket = 0;
1512         unsigned short socket_mask;
1513         unsigned char *uvhub_mask;
1514         struct bau_control *bcp;
1515         struct uvhub_desc *bdp;
1516         struct socket_desc *sdp;
1517         struct bau_control *hmaster = NULL;
1518         struct bau_control *smaster = NULL;
1519         struct socket_desc {
1520                 short num_cpus;
1521                 short cpu_number[MAX_CPUS_PER_SOCKET];
1522         };
1523         struct uvhub_desc {
1524                 unsigned short socket_mask;
1525                 short num_cpus;
1526                 short uvhub;
1527                 short pnode;
1528                 struct socket_desc socket[2];
1529         };
1530         struct uvhub_desc *uvhub_descs;
1531
1532         timeout_us = calculate_destination_timeout();
1533
1534         uvhub_descs = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1535         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1536         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
1537         for_each_present_cpu(cpu) {
1538                 bcp = &per_cpu(bau_control, cpu);
1539                 memset(bcp, 0, sizeof(struct bau_control));
1540                 pnode = uv_cpu_hub_info(cpu)->pnode;
1541                 if ((pnode - base_part_pnode) >= UV_DISTRIBUTION_SIZE) {
1542                         printk(KERN_EMERG
1543                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1544                                 cpu, pnode, base_part_pnode,
1545                                 UV_DISTRIBUTION_SIZE);
1546                         return 1;
1547                 }
1548                 bcp->osnode = cpu_to_node(cpu);
1549                 bcp->partition_base_pnode = uv_partition_base_pnode;
1550                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1551                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1552                 bdp = &uvhub_descs[uvhub];
1553                 bdp->num_cpus++;
1554                 bdp->uvhub = uvhub;
1555                 bdp->pnode = pnode;
1556                 /* kludge: 'assuming' one node per socket, and assuming that
1557                    disabling a socket just leaves a gap in node numbers */
1558                 socket = bcp->osnode & 1;
1559                 bdp->socket_mask |= (1 << socket);
1560                 sdp = &bdp->socket[socket];
1561                 sdp->cpu_number[sdp->num_cpus] = cpu;
1562                 sdp->num_cpus++;
1563                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1564                         printk(KERN_EMERG "%d cpus per socket invalid\n", sdp->num_cpus);
1565                         return 1;
1566                 }
1567         }
1568         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1569                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
1570                         continue;
1571                 have_hmaster = 0;
1572                 bdp = &uvhub_descs[uvhub];
1573                 socket_mask = bdp->socket_mask;
1574                 socket = 0;
1575                 while (socket_mask) {
1576                         if (!(socket_mask & 1))
1577                                 goto nextsocket;
1578                         sdp = &bdp->socket[socket];
1579                         for (i = 0; i < sdp->num_cpus; i++) {
1580                                 cpu = sdp->cpu_number[i];
1581                                 bcp = &per_cpu(bau_control, cpu);
1582                                 bcp->cpu = cpu;
1583                                 if (i == 0) {
1584                                         smaster = bcp;
1585                                         if (!have_hmaster) {
1586                                                 have_hmaster++;
1587                                                 hmaster = bcp;
1588                                         }
1589                                 }
1590                                 bcp->cpus_in_uvhub = bdp->num_cpus;
1591                                 bcp->cpus_in_socket = sdp->num_cpus;
1592                                 bcp->socket_master = smaster;
1593                                 bcp->uvhub = bdp->uvhub;
1594                                 bcp->uvhub_master = hmaster;
1595                                 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->
1596                                                 blade_processor_id;
1597                                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
1598                                         printk(KERN_EMERG
1599                                                 "%d cpus per uvhub invalid\n",
1600                                                 bcp->uvhub_cpu);
1601                                         return 1;
1602                                 }
1603                         }
1604 nextsocket:
1605                         socket++;
1606                         socket_mask = (socket_mask >> 1);
1607                         /* each socket gets a local array of pnodes/hubs */
1608                         bcp = smaster;
1609                         bcp->target_hub_and_pnode = kmalloc_node(
1610                                 sizeof(struct hub_and_pnode) *
1611                                 num_possible_cpus(), GFP_KERNEL, bcp->osnode);
1612                         memset(bcp->target_hub_and_pnode, 0,
1613                                 sizeof(struct hub_and_pnode) *
1614                                 num_possible_cpus());
1615                         for_each_present_cpu(tcpu) {
1616                                 bcp->target_hub_and_pnode[tcpu].pnode =
1617                                         uv_cpu_hub_info(tcpu)->pnode;
1618                                 bcp->target_hub_and_pnode[tcpu].uvhub =
1619                                         uv_cpu_hub_info(tcpu)->numa_blade_id;
1620                         }
1621                 }
1622         }
1623         kfree(uvhub_descs);
1624         kfree(uvhub_mask);
1625         for_each_present_cpu(cpu) {
1626                 bcp = &per_cpu(bau_control, cpu);
1627                 bcp->baudisabled = 0;
1628                 bcp->statp = &per_cpu(ptcstats, cpu);
1629                 /* time interval to catch a hardware stay-busy bug */
1630                 bcp->timeout_interval = microsec_2_cycles(2*timeout_us);
1631                 bcp->max_bau_concurrent = max_bau_concurrent;
1632                 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1633                 bcp->plugged_delay = plugged_delay;
1634                 bcp->plugsb4reset = plugsb4reset;
1635                 bcp->timeoutsb4reset = timeoutsb4reset;
1636                 bcp->ipi_reset_limit = ipi_reset_limit;
1637                 bcp->complete_threshold = complete_threshold;
1638                 bcp->congested_response_us = congested_response_us;
1639                 bcp->congested_reps = congested_reps;
1640                 bcp->congested_period = congested_period;
1641         }
1642         return 0;
1643 }
1644
1645 /*
1646  * Initialization of BAU-related structures
1647  */
1648 static int __init uv_bau_init(void)
1649 {
1650         int uvhub;
1651         int pnode;
1652         int nuvhubs;
1653         int cur_cpu;
1654         int vector;
1655         unsigned long mmr;
1656
1657         if (!is_uv_system())
1658                 return 0;
1659
1660         if (nobau)
1661                 return 0;
1662
1663         for_each_possible_cpu(cur_cpu)
1664                 zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu),
1665                                        GFP_KERNEL, cpu_to_node(cur_cpu));
1666
1667         uv_nshift = uv_hub_info->m_val;
1668         uv_mmask = (1UL << uv_hub_info->m_val) - 1;
1669         nuvhubs = uv_num_possible_blades();
1670         spin_lock_init(&disable_lock);
1671         congested_cycles = microsec_2_cycles(congested_response_us);
1672
1673         uv_partition_base_pnode = 0x7fffffff;
1674         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1675                 if (uv_blade_nr_possible_cpus(uvhub) &&
1676                         (uv_blade_to_pnode(uvhub) < uv_partition_base_pnode))
1677                         uv_partition_base_pnode = uv_blade_to_pnode(uvhub);
1678         }
1679
1680         if (uv_init_per_cpu(nuvhubs, uv_partition_base_pnode)) {
1681                 nobau = 1;
1682                 return 0;
1683         }
1684
1685         vector = UV_BAU_MESSAGE;
1686         for_each_possible_blade(uvhub)
1687                 if (uv_blade_nr_possible_cpus(uvhub))
1688                         uv_init_uvhub(uvhub, vector, uv_partition_base_pnode);
1689
1690         uv_enable_timeouts();
1691         alloc_intr_gate(vector, uv_bau_message_intr1);
1692
1693         for_each_possible_blade(uvhub) {
1694                 if (uv_blade_nr_possible_cpus(uvhub)) {
1695                         pnode = uv_blade_to_pnode(uvhub);
1696                         /* INIT the bau */
1697                         uv_write_global_mmr64(pnode,
1698                                         UVH_LB_BAU_SB_ACTIVATION_CONTROL,
1699                                         ((unsigned long)1 << 63));
1700                         mmr = 1; /* should be 1 to broadcast to both sockets */
1701                         uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST,
1702                                                 mmr);
1703                 }
1704         }
1705
1706         return 0;
1707 }
1708 core_initcall(uv_bau_init);
1709 fs_initcall(uv_ptc_init);