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 tcpu;
 703         int uvhub;
 704         int locals = 0;
 705         int remotes = 0;
 706         int hubs = 0;
 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
 713         /* kernel was booted 'nobau' */
 714         if (nobau)
 715                 return cpumask;
 716
 717         bcp = &per_cpu(bau_control, cpu);
 718         stat = bcp->statp;
 719
 720         /* bau was disabled due to slow response */
 721         if (bcp->baudisabled) {
 722                 /* the cpu that disabled it must re-enable it */
 723                 if (bcp->set_bau_off) {
 724                         if (get_cycles() >= bcp->set_bau_on_time) {
 725                                 stat->s_bau_reenabled++;
 726                                 baudisabled = 0;
 727                                 for_each_present_cpu(tcpu) {
 728                                         tbcp = &per_cpu(bau_control, tcpu);
 729                                         tbcp->baudisabled = 0;
 730                                         tbcp->period_requests = 0;
 731                                         tbcp->period_time = 0;
 732                                 }
 733                         }
 734                 }
 735                 return cpumask;
 736         }
 737
 738         /*
 739          * Each sending cpu has a per-cpu mask which it fills from the caller's
 740          * cpu mask.  All cpus are converted to uvhubs and copied to the
 741          * activation descriptor.
 742          */
 743         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
 744         /* don't actually do a shootdown of the local cpu */
 745         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
 746         if (cpu_isset(cpu, *cpumask))
 747                 stat->s_ntargself++;
 748
 749         bau_desc = bcp->descriptor_base;
 750         bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu;
 751         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
 752
 753         /* cpu statistics */
 754         for_each_cpu(tcpu, flush_mask) {
 755                 uvhub = uv_cpu_to_blade_id(tcpu);
 756                 bau_uvhub_set(uvhub, &bau_desc->distribution);
 757                 if (uvhub == bcp->uvhub)
 758                         locals++;
 759                 else
 760                         remotes++;
 761         }
 762         if ((locals + remotes) == 0)
 763                 return NULL;
 764         stat->s_requestor++;
 765         stat->s_ntargcpu += remotes + locals;
 766         stat->s_ntargremotes += remotes;
 767         stat->s_ntarglocals += locals;
 768         remotes = bau_uvhub_weight(&bau_desc->distribution);
 769
 770         /* uvhub statistics */
 771         hubs = bau_uvhub_weight(&bau_desc->distribution);
 772         if (locals) {
 773                 stat->s_ntarglocaluvhub++;
 774                 stat->s_ntargremoteuvhub += (hubs - 1);
 775         } else
 776                 stat->s_ntargremoteuvhub += hubs;
 777         stat->s_ntarguvhub += hubs;
 778         if (hubs >= 16)
 779                 stat->s_ntarguvhub16++;
 780         else if (hubs >= 8)
 781                 stat->s_ntarguvhub8++;
 782         else if (hubs >= 4)
 783                 stat->s_ntarguvhub4++;
 784         else if (hubs >= 2)
 785                 stat->s_ntarguvhub2++;
 786         else
 787                 stat->s_ntarguvhub1++;
 788
 789         bau_desc->payload.address = va;
 790         bau_desc->payload.sending_cpu = cpu;
 791
 792         /*
 793          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
 794          * or 1 if it gave up and the original cpumask should be returned.
 795          */
 796         if (!uv_flush_send_and_wait(bau_desc, flush_mask, bcp))
 797                 return NULL;
 798         else
 799                 return cpumask;
 800 }
 801
 802 /*
 803  * The BAU message interrupt comes here. (registered by set_intr_gate)
 804  * See entry_64.S
 805  *
 806  * We received a broadcast assist message.
 807  *
 808  * Interrupts are disabled; this interrupt could represent
 809  * the receipt of several messages.
 810  *
 811  * All cores/threads on this hub get this interrupt.
 812  * The last one to see it does the software ack.
 813  * (the resource will not be freed until noninterruptable cpus see this
 814  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
 815  */
 816 void uv_bau_message_interrupt(struct pt_regs *regs)
 817 {
 818         int count = 0;
 819         cycles_t time_start;
 820         struct bau_payload_queue_entry *msg;
 821         struct bau_control *bcp;
 822         struct ptc_stats *stat;
 823         struct msg_desc msgdesc;
 824
 825         time_start = get_cycles();
 826         bcp = &per_cpu(bau_control, smp_processor_id());
 827         stat = bcp->statp;
 828         msgdesc.va_queue_first = bcp->va_queue_first;
 829         msgdesc.va_queue_last = bcp->va_queue_last;
 830         msg = bcp->bau_msg_head;
 831         while (msg->sw_ack_vector) {
 832                 count++;
 833                 msgdesc.msg_slot = msg - msgdesc.va_queue_first;
 834                 msgdesc.sw_ack_slot = ffs(msg->sw_ack_vector) - 1;
 835                 msgdesc.msg = msg;
 836                 uv_bau_process_message(&msgdesc, bcp);
 837                 msg++;
 838                 if (msg > msgdesc.va_queue_last)
 839                         msg = msgdesc.va_queue_first;
 840                 bcp->bau_msg_head = msg;
 841         }
 842         stat->d_time += (get_cycles() - time_start);
 843         if (!count)
 844                 stat->d_nomsg++;
 845         else if (count > 1)
 846                 stat->d_multmsg++;
 847         ack_APIC_irq();
 848 }
 849
 850 /*
 851  * uv_enable_timeouts
 852  *
 853  * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
 854  * shootdown message timeouts enabled.  The timeout does not cause
 855  * an interrupt, but causes an error message to be returned to
 856  * the sender.
 857  */
 858 static void uv_enable_timeouts(void)
 859 {
 860         int uvhub;
 861         int nuvhubs;
 862         int pnode;
 863         unsigned long mmr_image;
 864
 865         nuvhubs = uv_num_possible_blades();
 866
 867         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
 868                 if (!uv_blade_nr_possible_cpus(uvhub))
 869                         continue;
 870
 871                 pnode = uv_blade_to_pnode(uvhub);
 872                 mmr_image =
 873                     uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL);
 874                 /*
 875                  * Set the timeout period and then lock it in, in three
 876                  * steps; captures and locks in the period.
 877                  *
 878                  * To program the period, the SOFT_ACK_MODE must be off.
 879                  */
 880                 mmr_image &= ~((unsigned long)1 <<
 881                     UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 882                 uv_write_global_mmr64
 883                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 884                 /*
 885                  * Set the 4-bit period.
 886                  */
 887                 mmr_image &= ~((unsigned long)0xf <<
 888                      UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 889                 mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD <<
 890                      UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
 891                 uv_write_global_mmr64
 892                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 893                 /*
 894                  * Subsequent reversals of the timebase bit (3) cause an
 895                  * immediate timeout of one or all INTD resources as
 896                  * indicated in bits 2:0 (7 causes all of them to timeout).
 897                  */
 898                 mmr_image |= ((unsigned long)1 <<
 899                     UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
 900                 uv_write_global_mmr64
 901                     (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
 902         }
 903 }
 904
 905 static void *uv_ptc_seq_start(struct seq_file *file, loff_t *offset)
 906 {
 907         if (*offset < num_possible_cpus())
 908                 return offset;
 909         return NULL;
 910 }
 911
 912 static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
 913 {
 914         (*offset)++;
 915         if (*offset < num_possible_cpus())
 916                 return offset;
 917         return NULL;
 918 }
 919
 920 static void uv_ptc_seq_stop(struct seq_file *file, void *data)
 921 {
 922 }
 923
 924 static inline unsigned long long
 925 microsec_2_cycles(unsigned long microsec)
 926 {
 927         unsigned long ns;
 928         unsigned long long cyc;
 929
 930         ns = microsec * 1000;
 931         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
 932         return cyc;
 933 }
 934
 935 /*
 936  * Display the statistics thru /proc.
 937  * 'data' points to the cpu number
 938  */
 939 static int uv_ptc_seq_show(struct seq_file *file, void *data)
 940 {
 941         struct ptc_stats *stat;
 942         int cpu;
 943
 944         cpu = *(loff_t *)data;
 945
 946         if (!cpu) {
 947                 seq_printf(file,
 948                         "# cpu sent stime self locals remotes ncpus localhub ");
 949                 seq_printf(file,
 950                         "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
 951                 seq_printf(file,
 952                         "numuvhubs4 numuvhubs2 numuvhubs1 dto ");
 953                 seq_printf(file,
 954                         "retries rok resetp resett giveup sto bz throt ");
 955                 seq_printf(file,
 956                         "sw_ack recv rtime all ");
 957                 seq_printf(file,
 958                         "one mult none retry canc nocan reset rcan ");
 959                 seq_printf(file,
 960                         "disable enable\n");
 961         }
 962         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
 963                 stat = &per_cpu(ptcstats, cpu);
 964                 /* source side statistics */
 965                 seq_printf(file,
 966                         "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
 967                            cpu, stat->s_requestor, cycles_2_us(stat->s_time),
 968                            stat->s_ntargself, stat->s_ntarglocals,
 969                            stat->s_ntargremotes, stat->s_ntargcpu,
 970                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
 971                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
 972                 seq_printf(file, "%ld %ld %ld %ld %ld ",
 973                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
 974                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
 975                            stat->s_dtimeout);
 976                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
 977                            stat->s_retry_messages, stat->s_retriesok,
 978                            stat->s_resets_plug, stat->s_resets_timeout,
 979                            stat->s_giveup, stat->s_stimeout,
 980                            stat->s_busy, stat->s_throttles);
 981
 982                 /* destination side statistics */
 983                 seq_printf(file,
 984                            "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
 985                            uv_read_global_mmr64(uv_cpu_to_pnode(cpu),
 986                                         UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE),
 987                            stat->d_requestee, cycles_2_us(stat->d_time),
 988                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
 989                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
 990                            stat->d_nocanceled, stat->d_resets,
 991                            stat->d_rcanceled);
 992                 seq_printf(file, "%ld %ld\n",
 993                         stat->s_bau_disabled, stat->s_bau_reenabled);
 994         }
 995
 996         return 0;
 997 }
 998
 999 /*
1000  * Display the tunables thru debugfs
1001  */
1002 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1003                                                 size_t count, loff_t *ppos)
1004 {
1005         char *buf;
1006         int ret;
1007
1008         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1009                 "max_bau_concurrent plugged_delay plugsb4reset",
1010                 "timeoutsb4reset ipi_reset_limit complete_threshold",
1011                 "congested_response_us congested_reps congested_period",
1012                 max_bau_concurrent, plugged_delay, plugsb4reset,
1013                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1014                 congested_response_us, congested_reps, congested_period);
1015
1016         if (!buf)
1017                 return -ENOMEM;
1018
1019         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1020         kfree(buf);
1021         return ret;
1022 }
1023
1024 /*
1025  * -1: resetf the statistics
1026  *  0: display meaning of the statistics
1027  */
1028 static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user,
1029                                  size_t count, loff_t *data)
1030 {
1031         int cpu;
1032         long input_arg;
1033         char optstr[64];
1034         struct ptc_stats *stat;
1035
1036         if (count == 0 || count > sizeof(optstr))
1037                 return -EINVAL;
1038         if (copy_from_user(optstr, user, count))
1039                 return -EFAULT;
1040         optstr[count - 1] = '\0';
1041         if (strict_strtol(optstr, 10, &input_arg) < 0) {
1042                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1043                 return -EINVAL;
1044         }
1045
1046         if (input_arg == 0) {
1047                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1048                 printk(KERN_DEBUG "Sender statistics:\n");
1049                 printk(KERN_DEBUG
1050                 "sent:     number of shootdown messages sent\n");
1051                 printk(KERN_DEBUG
1052                 "stime:    time spent sending messages\n");
1053                 printk(KERN_DEBUG
1054                 "numuvhubs: number of hubs targeted with shootdown\n");
1055                 printk(KERN_DEBUG
1056                 "numuvhubs16: number times 16 or more hubs targeted\n");
1057                 printk(KERN_DEBUG
1058                 "numuvhubs8: number times 8 or more hubs targeted\n");
1059                 printk(KERN_DEBUG
1060                 "numuvhubs4: number times 4 or more hubs targeted\n");
1061                 printk(KERN_DEBUG
1062                 "numuvhubs2: number times 2 or more hubs targeted\n");
1063                 printk(KERN_DEBUG
1064                 "numuvhubs1: number times 1 hub targeted\n");
1065                 printk(KERN_DEBUG
1066                 "numcpus:  number of cpus targeted with shootdown\n");
1067                 printk(KERN_DEBUG
1068                 "dto:      number of destination timeouts\n");
1069                 printk(KERN_DEBUG
1070                 "retries:  destination timeout retries sent\n");
1071                 printk(KERN_DEBUG
1072                 "rok:   :  destination timeouts successfully retried\n");
1073                 printk(KERN_DEBUG
1074                 "resetp:   ipi-style resource resets for plugs\n");
1075                 printk(KERN_DEBUG
1076                 "resett:   ipi-style resource resets for timeouts\n");
1077                 printk(KERN_DEBUG
1078                 "giveup:   fall-backs to ipi-style shootdowns\n");
1079                 printk(KERN_DEBUG
1080                 "sto:      number of source timeouts\n");
1081                 printk(KERN_DEBUG
1082                 "bz:       number of stay-busy's\n");
1083                 printk(KERN_DEBUG
1084                 "throt:    number times spun in throttle\n");
1085                 printk(KERN_DEBUG "Destination side statistics:\n");
1086                 printk(KERN_DEBUG
1087                 "sw_ack:   image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
1088                 printk(KERN_DEBUG
1089                 "recv:     shootdown messages received\n");
1090                 printk(KERN_DEBUG
1091                 "rtime:    time spent processing messages\n");
1092                 printk(KERN_DEBUG
1093                 "all:      shootdown all-tlb messages\n");
1094                 printk(KERN_DEBUG
1095                 "one:      shootdown one-tlb messages\n");
1096                 printk(KERN_DEBUG
1097                 "mult:     interrupts that found multiple messages\n");
1098                 printk(KERN_DEBUG
1099                 "none:     interrupts that found no messages\n");
1100                 printk(KERN_DEBUG
1101                 "retry:    number of retry messages processed\n");
1102                 printk(KERN_DEBUG
1103                 "canc:     number messages canceled by retries\n");
1104                 printk(KERN_DEBUG
1105                 "nocan:    number retries that found nothing to cancel\n");
1106                 printk(KERN_DEBUG
1107                 "reset:    number of ipi-style reset requests processed\n");
1108                 printk(KERN_DEBUG
1109                 "rcan:     number messages canceled by reset requests\n");
1110                 printk(KERN_DEBUG
1111                 "disable:  number times use of the BAU was disabled\n");
1112                 printk(KERN_DEBUG
1113                 "enable:   number times use of the BAU was re-enabled\n");
1114         } else if (input_arg == -1) {
1115                 for_each_present_cpu(cpu) {
1116                         stat = &per_cpu(ptcstats, cpu);
1117                         memset(stat, 0, sizeof(struct ptc_stats));
1118                 }
1119         }
1120
1121         return count;
1122 }
1123
1124 static int local_atoi(const char *name)
1125 {
1126         int val = 0;
1127
1128         for (;; name++) {
1129                 switch (*name) {
1130                 case '0' ... '9':
1131                         val = 10*val+(*name-'0');
1132                         break;
1133                 default:
1134                         return val;
1135                 }
1136         }
1137 }
1138
1139 /*
1140  * set the tunables
1141  * 0 values reset them to defaults
1142  */
1143 static ssize_t tunables_write(struct file *file, const char __user *user,
1144                                  size_t count, loff_t *data)
1145 {
1146         int cpu;
1147         int cnt = 0;
1148         int val;
1149         char *p;
1150         char *q;
1151         char instr[64];
1152         struct bau_control *bcp;
1153
1154         if (count == 0 || count > sizeof(instr)-1)
1155                 return -EINVAL;
1156         if (copy_from_user(instr, user, count))
1157                 return -EFAULT;
1158
1159         instr[count] = '\0';
1160         /* count the fields */
1161         p = instr + strspn(instr, WHITESPACE);
1162         q = p;
1163         for (; *p; p = q + strspn(q, WHITESPACE)) {
1164                 q = p + strcspn(p, WHITESPACE);
1165                 cnt++;
1166                 if (q == p)
1167                         break;
1168         }
1169         if (cnt != 9) {
1170                 printk(KERN_INFO "bau tunable error: should be 9 numbers\n");
1171                 return -EINVAL;
1172         }
1173
1174         p = instr + strspn(instr, WHITESPACE);
1175         q = p;
1176         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1177                 q = p + strcspn(p, WHITESPACE);
1178                 val = local_atoi(p);
1179                 switch (cnt) {
1180                 case 0:
1181                         if (val == 0) {
1182                                 max_bau_concurrent = MAX_BAU_CONCURRENT;
1183                                 max_bau_concurrent_constant =
1184                                                         MAX_BAU_CONCURRENT;
1185                                 continue;
1186                         }
1187                         bcp = &per_cpu(bau_control, smp_processor_id());
1188                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1189                                 printk(KERN_DEBUG
1190                                 "Error: BAU max concurrent %d is invalid\n",
1191                                 val);
1192                                 return -EINVAL;
1193                         }
1194                         max_bau_concurrent = val;
1195                         max_bau_concurrent_constant = val;
1196                         continue;
1197                 case 1:
1198                         if (val == 0)
1199                                 plugged_delay = PLUGGED_DELAY;
1200                         else
1201                                 plugged_delay = val;
1202                         continue;
1203                 case 2:
1204                         if (val == 0)
1205                                 plugsb4reset = PLUGSB4RESET;
1206                         else
1207                                 plugsb4reset = val;
1208                         continue;
1209                 case 3:
1210                         if (val == 0)
1211                                 timeoutsb4reset = TIMEOUTSB4RESET;
1212                         else
1213                                 timeoutsb4reset = val;
1214                         continue;
1215                 case 4:
1216                         if (val == 0)
1217                                 ipi_reset_limit = IPI_RESET_LIMIT;
1218                         else
1219                                 ipi_reset_limit = val;
1220                         continue;
1221                 case 5:
1222                         if (val == 0)
1223                                 complete_threshold = COMPLETE_THRESHOLD;
1224                         else
1225                                 complete_threshold = val;
1226                         continue;
1227                 case 6:
1228                         if (val == 0)
1229                                 congested_response_us = CONGESTED_RESPONSE_US;
1230                         else
1231                                 congested_response_us = val;
1232                         continue;
1233                 case 7:
1234                         if (val == 0)
1235                                 congested_reps = CONGESTED_REPS;
1236                         else
1237                                 congested_reps = val;
1238                         continue;
1239                 case 8:
1240                         if (val == 0)
1241                                 congested_period = CONGESTED_PERIOD;
1242                         else
1243                                 congested_period = val;
1244                         continue;
1245                 }
1246                 if (q == p)
1247                         break;
1248         }
1249         for_each_present_cpu(cpu) {
1250                 bcp = &per_cpu(bau_control, cpu);
1251                 bcp->max_bau_concurrent = max_bau_concurrent;
1252                 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1253                 bcp->plugged_delay = plugged_delay;
1254                 bcp->plugsb4reset = plugsb4reset;
1255                 bcp->timeoutsb4reset = timeoutsb4reset;
1256                 bcp->ipi_reset_limit = ipi_reset_limit;
1257                 bcp->complete_threshold = complete_threshold;
1258                 bcp->congested_response_us = congested_response_us;
1259                 bcp->congested_reps = congested_reps;
1260                 bcp->congested_period = congested_period;
1261         }
1262         return count;
1263 }
1264
1265 static const struct seq_operations uv_ptc_seq_ops = {
1266         .start          = uv_ptc_seq_start,
1267         .next           = uv_ptc_seq_next,
1268         .stop           = uv_ptc_seq_stop,
1269         .show           = uv_ptc_seq_show
1270 };
1271
1272 static int uv_ptc_proc_open(struct inode *inode, struct file *file)
1273 {
1274         return seq_open(file, &uv_ptc_seq_ops);
1275 }
1276
1277 static int tunables_open(struct inode *inode, struct file *file)
1278 {
1279         return 0;
1280 }
1281
1282 static const struct file_operations proc_uv_ptc_operations = {
1283         .open           = uv_ptc_proc_open,
1284         .read           = seq_read,
1285         .write          = uv_ptc_proc_write,
1286         .llseek         = seq_lseek,
1287         .release        = seq_release,
1288 };
1289
1290 static const struct file_operations tunables_fops = {
1291         .open           = tunables_open,
1292         .read           = tunables_read,
1293         .write          = tunables_write,
1294         .llseek         = default_llseek,
1295 };
1296
1297 static int __init uv_ptc_init(void)
1298 {
1299         struct proc_dir_entry *proc_uv_ptc;
1300
1301         if (!is_uv_system())
1302                 return 0;
1303
1304         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1305                                   &proc_uv_ptc_operations);
1306         if (!proc_uv_ptc) {
1307                 printk(KERN_ERR "unable to create %s proc entry\n",
1308                        UV_PTC_BASENAME);
1309                 return -EINVAL;
1310         }
1311
1312         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1313         if (!tunables_dir) {
1314                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1315                        UV_BAU_TUNABLES_DIR);
1316                 return -EINVAL;
1317         }
1318         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1319                         tunables_dir, NULL, &tunables_fops);
1320         if (!tunables_file) {
1321                 printk(KERN_ERR "unable to create debugfs file %s\n",
1322                        UV_BAU_TUNABLES_FILE);
1323                 return -EINVAL;
1324         }
1325         return 0;
1326 }
1327
1328 /*
1329  * initialize the sending side's sending buffers
1330  */
1331 static void
1332 uv_activation_descriptor_init(int node, int pnode)
1333 {
1334         int i;
1335         int cpu;
1336         unsigned long pa;
1337         unsigned long m;
1338         unsigned long n;
1339         struct bau_desc *bau_desc;
1340         struct bau_desc *bd2;
1341         struct bau_control *bcp;
1342
1343         /*
1344          * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR)
1345          * per cpu; and one per cpu on the uvhub (UV_ADP_SIZE)
1346          */
1347         bau_desc = kmalloc_node(sizeof(struct bau_desc) * UV_ADP_SIZE
1348                                 * UV_ITEMS_PER_DESCRIPTOR, GFP_KERNEL, node);
1349         BUG_ON(!bau_desc);
1350
1351         pa = uv_gpa(bau_desc); /* need the real nasid*/
1352         n = pa >> uv_nshift;
1353         m = pa & uv_mmask;
1354
1355         uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE,
1356                               (n << UV_DESC_BASE_PNODE_SHIFT | m));
1357
1358         /*
1359          * initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each
1360          * cpu even though we only use the first one; one descriptor can
1361          * describe a broadcast to 256 uv hubs.
1362          */
1363         for (i = 0, bd2 = bau_desc; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR);
1364                 i++, bd2++) {
1365                 memset(bd2, 0, sizeof(struct bau_desc));
1366                 bd2->header.sw_ack_flag = 1;
1367                 /*
1368                  * base_dest_nodeid is the nasid of the first uvhub
1369                  * in the partition. The bit map will indicate uvhub numbers,
1370                  * which are 0-N in a partition. Pnodes are unique system-wide.
1371                  */
1372                 bd2->header.base_dest_nodeid = UV_PNODE_TO_NASID(uv_partition_base_pnode);
1373                 bd2->header.dest_subnodeid = 0x10; /* the LB */
1374                 bd2->header.command = UV_NET_ENDPOINT_INTD;
1375                 bd2->header.int_both = 1;
1376                 /*
1377                  * all others need to be set to zero:
1378                  *   fairness chaining multilevel count replied_to
1379                  */
1380         }
1381         for_each_present_cpu(cpu) {
1382                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1383                         continue;
1384                 bcp = &per_cpu(bau_control, cpu);
1385                 bcp->descriptor_base = bau_desc;
1386         }
1387 }
1388
1389 /*
1390  * initialize the destination side's receiving buffers
1391  * entered for each uvhub in the partition
1392  * - node is first node (kernel memory notion) on the uvhub
1393  * - pnode is the uvhub's physical identifier
1394  */
1395 static void
1396 uv_payload_queue_init(int node, int pnode)
1397 {
1398         int pn;
1399         int cpu;
1400         char *cp;
1401         unsigned long pa;
1402         struct bau_payload_queue_entry *pqp;
1403         struct bau_payload_queue_entry *pqp_malloc;
1404         struct bau_control *bcp;
1405
1406         pqp = kmalloc_node((DEST_Q_SIZE + 1)
1407                            * sizeof(struct bau_payload_queue_entry),
1408                            GFP_KERNEL, node);
1409         BUG_ON(!pqp);
1410         pqp_malloc = pqp;
1411
1412         cp = (char *)pqp + 31;
1413         pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5);
1414
1415         for_each_present_cpu(cpu) {
1416                 if (pnode != uv_cpu_to_pnode(cpu))
1417                         continue;
1418                 /* for every cpu on this pnode: */
1419                 bcp = &per_cpu(bau_control, cpu);
1420                 bcp->va_queue_first = pqp;
1421                 bcp->bau_msg_head = pqp;
1422                 bcp->va_queue_last = pqp + (DEST_Q_SIZE - 1);
1423         }
1424         /*
1425          * need the pnode of where the memory was really allocated
1426          */
1427         pa = uv_gpa(pqp);
1428         pn = pa >> uv_nshift;
1429         uv_write_global_mmr64(pnode,
1430                               UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST,
1431                               ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) |
1432                               uv_physnodeaddr(pqp));
1433         uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL,
1434                               uv_physnodeaddr(pqp));
1435         uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST,
1436                               (unsigned long)
1437                               uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1)));
1438         /* in effect, all msg_type's are set to MSG_NOOP */
1439         memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE);
1440 }
1441
1442 /*
1443  * Initialization of each UV hub's structures
1444  */
1445 static void __init uv_init_uvhub(int uvhub, int vector)
1446 {
1447         int node;
1448         int pnode;
1449         unsigned long apicid;
1450
1451         node = uvhub_to_first_node(uvhub);
1452         pnode = uv_blade_to_pnode(uvhub);
1453         uv_activation_descriptor_init(node, pnode);
1454         uv_payload_queue_init(node, pnode);
1455         /*
1456          * the below initialization can't be in firmware because the
1457          * messaging IRQ will be determined by the OS
1458          */
1459         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1460         uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG,
1461                                       ((apicid << 32) | vector));
1462 }
1463
1464 /*
1465  * We will set BAU_MISC_CONTROL with a timeout period.
1466  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1467  * So the destination timeout period has be be calculated from them.
1468  */
1469 static int
1470 calculate_destination_timeout(void)
1471 {
1472         unsigned long mmr_image;
1473         int mult1;
1474         int mult2;
1475         int index;
1476         int base;
1477         int ret;
1478         unsigned long ts_ns;
1479
1480         mult1 = UV_INTD_SOFT_ACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1481         mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1482         index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1483         mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1484         mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1485         base = timeout_base_ns[index];
1486         ts_ns = base * mult1 * mult2;
1487         ret = ts_ns / 1000;
1488         return ret;
1489 }
1490
1491 /*
1492  * initialize the bau_control structure for each cpu
1493  */
1494 static int __init uv_init_per_cpu(int nuvhubs)
1495 {
1496         int i;
1497         int cpu;
1498         int pnode;
1499         int uvhub;
1500         int have_hmaster;
1501         short socket = 0;
1502         unsigned short socket_mask;
1503         unsigned char *uvhub_mask;
1504         struct bau_control *bcp;
1505         struct uvhub_desc *bdp;
1506         struct socket_desc *sdp;
1507         struct bau_control *hmaster = NULL;
1508         struct bau_control *smaster = NULL;
1509         struct socket_desc {
1510                 short num_cpus;
1511                 short cpu_number[MAX_CPUS_PER_SOCKET];
1512         };
1513         struct uvhub_desc {
1514                 unsigned short socket_mask;
1515                 short num_cpus;
1516                 short uvhub;
1517                 short pnode;
1518                 struct socket_desc socket[2];
1519         };
1520         struct uvhub_desc *uvhub_descs;
1521
1522         timeout_us = calculate_destination_timeout();
1523
1524         uvhub_descs = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1525         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1526         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
1527         for_each_present_cpu(cpu) {
1528                 bcp = &per_cpu(bau_control, cpu);
1529                 memset(bcp, 0, sizeof(struct bau_control));
1530                 pnode = uv_cpu_hub_info(cpu)->pnode;
1531                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1532                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1533                 bdp = &uvhub_descs[uvhub];
1534                 bdp->num_cpus++;
1535                 bdp->uvhub = uvhub;
1536                 bdp->pnode = pnode;
1537                 /* kludge: 'assuming' one node per socket, and assuming that
1538                    disabling a socket just leaves a gap in node numbers */
1539                 socket = (cpu_to_node(cpu) & 1);
1540                 bdp->socket_mask |= (1 << socket);
1541                 sdp = &bdp->socket[socket];
1542                 sdp->cpu_number[sdp->num_cpus] = cpu;
1543                 sdp->num_cpus++;
1544                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1545                         printk(KERN_EMERG "%d cpus per socket invalid\n", sdp->num_cpus);
1546                         return 1;
1547                 }
1548         }
1549         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1550                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
1551                         continue;
1552                 have_hmaster = 0;
1553                 bdp = &uvhub_descs[uvhub];
1554                 socket_mask = bdp->socket_mask;
1555                 socket = 0;
1556                 while (socket_mask) {
1557                         if (!(socket_mask & 1))
1558                                 goto nextsocket;
1559                         sdp = &bdp->socket[socket];
1560                         for (i = 0; i < sdp->num_cpus; i++) {
1561                                 cpu = sdp->cpu_number[i];
1562                                 bcp = &per_cpu(bau_control, cpu);
1563                                 bcp->cpu = cpu;
1564                                 if (i == 0) {
1565                                         smaster = bcp;
1566                                         if (!have_hmaster) {
1567                                                 have_hmaster++;
1568                                                 hmaster = bcp;
1569                                         }
1570                                 }
1571                                 bcp->cpus_in_uvhub = bdp->num_cpus;
1572                                 bcp->cpus_in_socket = sdp->num_cpus;
1573                                 bcp->socket_master = smaster;
1574                                 bcp->uvhub = bdp->uvhub;
1575                                 bcp->uvhub_master = hmaster;
1576                                 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->
1577                                                 blade_processor_id;
1578                                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
1579                                         printk(KERN_EMERG
1580                                                 "%d cpus per uvhub invalid\n",
1581                                                 bcp->uvhub_cpu);
1582                                         return 1;
1583                                 }
1584                         }
1585 nextsocket:
1586                         socket++;
1587                         socket_mask = (socket_mask >> 1);
1588                 }
1589         }
1590         kfree(uvhub_descs);
1591         kfree(uvhub_mask);
1592         for_each_present_cpu(cpu) {
1593                 bcp = &per_cpu(bau_control, cpu);
1594                 bcp->baudisabled = 0;
1595                 bcp->statp = &per_cpu(ptcstats, cpu);
1596                 /* time interval to catch a hardware stay-busy bug */
1597                 bcp->timeout_interval = microsec_2_cycles(2*timeout_us);
1598                 bcp->max_bau_concurrent = max_bau_concurrent;
1599                 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1600                 bcp->plugged_delay = plugged_delay;
1601                 bcp->plugsb4reset = plugsb4reset;
1602                 bcp->timeoutsb4reset = timeoutsb4reset;
1603                 bcp->ipi_reset_limit = ipi_reset_limit;
1604                 bcp->complete_threshold = complete_threshold;
1605                 bcp->congested_response_us = congested_response_us;
1606                 bcp->congested_reps = congested_reps;
1607                 bcp->congested_period = congested_period;
1608         }
1609         return 0;
1610 }
1611
1612 /*
1613  * Initialization of BAU-related structures
1614  */
1615 static int __init uv_bau_init(void)
1616 {
1617         int uvhub;
1618         int pnode;
1619         int nuvhubs;
1620         int cur_cpu;
1621         int vector;
1622         unsigned long mmr;
1623
1624         if (!is_uv_system())
1625                 return 0;
1626
1627         if (nobau)
1628                 return 0;
1629
1630         for_each_possible_cpu(cur_cpu)
1631                 zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu),
1632                                        GFP_KERNEL, cpu_to_node(cur_cpu));
1633
1634         uv_nshift = uv_hub_info->m_val;
1635         uv_mmask = (1UL << uv_hub_info->m_val) - 1;
1636         nuvhubs = uv_num_possible_blades();
1637         spin_lock_init(&disable_lock);
1638         congested_cycles = microsec_2_cycles(congested_response_us);
1639
1640         if (uv_init_per_cpu(nuvhubs)) {
1641                 nobau = 1;
1642                 return 0;
1643         }
1644
1645         uv_partition_base_pnode = 0x7fffffff;
1646         for (uvhub = 0; uvhub < nuvhubs; uvhub++)
1647                 if (uv_blade_nr_possible_cpus(uvhub) &&
1648                         (uv_blade_to_pnode(uvhub) < uv_partition_base_pnode))
1649                         uv_partition_base_pnode = uv_blade_to_pnode(uvhub);
1650
1651         vector = UV_BAU_MESSAGE;
1652         for_each_possible_blade(uvhub)
1653                 if (uv_blade_nr_possible_cpus(uvhub))
1654                         uv_init_uvhub(uvhub, vector);
1655
1656         uv_enable_timeouts();
1657         alloc_intr_gate(vector, uv_bau_message_intr1);
1658
1659         for_each_possible_blade(uvhub) {
1660                 if (uv_blade_nr_possible_cpus(uvhub)) {
1661                         pnode = uv_blade_to_pnode(uvhub);
1662                         /* INIT the bau */
1663                         uv_write_global_mmr64(pnode,
1664                                         UVH_LB_BAU_SB_ACTIVATION_CONTROL,
1665                                         ((unsigned long)1 << 63));
1666                         mmr = 1; /* should be 1 to broadcast to both sockets */
1667                         uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST,
1668                                                 mmr);
1669                 }
1670         }
1671
1672         return 0;
1673 }
1674 core_initcall(uv_bau_init);
1675 fs_initcall(uv_ptc_init);