drivers/infiniband/hw/ehca/ehca_irq.c

   1 /*
   2  *  IBM eServer eHCA Infiniband device driver for Linux on POWER
   3  *
   4  *  Functions for EQs, NEQs and interrupts
   5  *
   6  *  Authors: Heiko J Schick <schickhj@de.ibm.com>
   7  *           Khadija Souissi <souissi@de.ibm.com>
   8  *           Hoang-Nam Nguyen <hnguyen@de.ibm.com>
   9  *           Joachim Fenkes <fenkes@de.ibm.com>
  10  *
  11  *  Copyright (c) 2005 IBM Corporation
  12  *
  13  *  All rights reserved.
  14  *
  15  *  This source code is distributed under a dual license of GPL v2.0 and OpenIB
  16  *  BSD.
  17  *
  18  * OpenIB BSD License
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions are met:
  22  *
  23  * Redistributions of source code must retain the above copyright notice, this
  24  * list of conditions and the following disclaimer.
  25  *
  26  * Redistributions in binary form must reproduce the above copyright notice,
  27  * this list of conditions and the following disclaimer in the documentation
  28  * and/or other materials
  29  * provided with the distribution.
  30  *
  31  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  32  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  34  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
  35  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  36  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  37  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  38  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER
  39  * IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  40  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  41  * POSSIBILITY OF SUCH DAMAGE.
  42  */
  43
  44 #include "ehca_classes.h"
  45 #include "ehca_irq.h"
  46 #include "ehca_iverbs.h"
  47 #include "ehca_tools.h"
  48 #include "hcp_if.h"
  49 #include "hipz_fns.h"
  50 #include "ipz_pt_fn.h"
  51
  52 #define EQE_COMPLETION_EVENT   EHCA_BMASK_IBM(1,1)
  53 #define EQE_CQ_QP_NUMBER       EHCA_BMASK_IBM(8,31)
  54 #define EQE_EE_IDENTIFIER      EHCA_BMASK_IBM(2,7)
  55 #define EQE_CQ_NUMBER          EHCA_BMASK_IBM(8,31)
  56 #define EQE_QP_NUMBER          EHCA_BMASK_IBM(8,31)
  57 #define EQE_QP_TOKEN           EHCA_BMASK_IBM(32,63)
  58 #define EQE_CQ_TOKEN           EHCA_BMASK_IBM(32,63)
  59
  60 #define NEQE_COMPLETION_EVENT  EHCA_BMASK_IBM(1,1)
  61 #define NEQE_EVENT_CODE        EHCA_BMASK_IBM(2,7)
  62 #define NEQE_PORT_NUMBER       EHCA_BMASK_IBM(8,15)
  63 #define NEQE_PORT_AVAILABILITY EHCA_BMASK_IBM(16,16)
  64
  65 #define ERROR_DATA_LENGTH      EHCA_BMASK_IBM(52,63)
  66 #define ERROR_DATA_TYPE        EHCA_BMASK_IBM(0,7)
  67
  68 static void queue_comp_task(struct ehca_cq *__cq);
  69
  70 static struct ehca_comp_pool* pool;
  71 #ifdef CONFIG_HOTPLUG_CPU
  72 static struct notifier_block comp_pool_callback_nb;
  73 #endif
  74
  75 static inline void comp_event_callback(struct ehca_cq *cq)
  76 {
  77         if (!cq->ib_cq.comp_handler)
  78                 return;
  79
  80         spin_lock(&cq->cb_lock);
  81         cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
  82         spin_unlock(&cq->cb_lock);
  83
  84         return;
  85 }
  86
  87 static void print_error_data(struct ehca_shca * shca, void* data,
  88                              u64* rblock, int length)
  89 {
  90         u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
  91         u64 resource = rblock[1];
  92
  93         switch (type) {
  94         case 0x1: /* Queue Pair */
  95         {
  96                 struct ehca_qp *qp = (struct ehca_qp*)data;
  97
  98                 /* only print error data if AER is set */
  99                 if (rblock[6] == 0)
 100                         return;
 101
 102                 ehca_err(&shca->ib_device,
 103                          "QP 0x%x (resource=%lx) has errors.",
 104                          qp->ib_qp.qp_num, resource);
 105                 break;
 106         }
 107         case 0x4: /* Completion Queue */
 108         {
 109                 struct ehca_cq *cq = (struct ehca_cq*)data;
 110
 111                 ehca_err(&shca->ib_device,
 112                          "CQ 0x%x (resource=%lx) has errors.",
 113                          cq->cq_number, resource);
 114                 break;
 115         }
 116         default:
 117                 ehca_err(&shca->ib_device,
 118                          "Unknown errror type: %lx on %s.",
 119                          type, shca->ib_device.name);
 120                 break;
 121         }
 122
 123         ehca_err(&shca->ib_device, "Error data is available: %lx.", resource);
 124         ehca_err(&shca->ib_device, "EHCA ----- error data begin "
 125                  "---------------------------------------------------");
 126         ehca_dmp(rblock, length, "resource=%lx", resource);
 127         ehca_err(&shca->ib_device, "EHCA ----- error data end "
 128                  "----------------------------------------------------");
 129
 130         return;
 131 }
 132
 133 int ehca_error_data(struct ehca_shca *shca, void *data,
 134                     u64 resource)
 135 {
 136
 137         unsigned long ret;
 138         u64 *rblock;
 139         unsigned long block_count;
 140
 141         rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
 142         if (!rblock) {
 143                 ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
 144                 ret = -ENOMEM;
 145                 goto error_data1;
 146         }
 147
 148         /* rblock must be 4K aligned and should be 4K large */
 149         ret = hipz_h_error_data(shca->ipz_hca_handle,
 150                                 resource,
 151                                 rblock,
 152                                 &block_count);
 153
 154         if (ret == H_R_STATE)
 155                 ehca_err(&shca->ib_device,
 156                          "No error data is available: %lx.", resource);
 157         else if (ret == H_SUCCESS) {
 158                 int length;
 159
 160                 length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
 161
 162                 if (length > EHCA_PAGESIZE)
 163                         length = EHCA_PAGESIZE;
 164
 165                 print_error_data(shca, data, rblock, length);
 166         } else
 167                 ehca_err(&shca->ib_device,
 168                          "Error data could not be fetched: %lx", resource);
 169
 170         ehca_free_fw_ctrlblock(rblock);
 171
 172 error_data1:
 173         return ret;
 174
 175 }
 176
 177 static void qp_event_callback(struct ehca_shca *shca,
 178                               u64 eqe,
 179                               enum ib_event_type event_type)
 180 {
 181         struct ib_event event;
 182         struct ehca_qp *qp;
 183         unsigned long flags;
 184         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
 185
 186         spin_lock_irqsave(&ehca_qp_idr_lock, flags);
 187         qp = idr_find(&ehca_qp_idr, token);
 188         spin_unlock_irqrestore(&ehca_qp_idr_lock, flags);
 189
 190
 191         if (!qp)
 192                 return;
 193
 194         ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
 195
 196         if (!qp->ib_qp.event_handler)
 197                 return;
 198
 199         event.device     = &shca->ib_device;
 200         event.event      = event_type;
 201         event.element.qp = &qp->ib_qp;
 202
 203         qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 204
 205         return;
 206 }
 207
 208 static void cq_event_callback(struct ehca_shca *shca,
 209                               u64 eqe)
 210 {
 211         struct ehca_cq *cq;
 212         unsigned long flags;
 213         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
 214
 215         spin_lock_irqsave(&ehca_cq_idr_lock, flags);
 216         cq = idr_find(&ehca_cq_idr, token);
 217         if (cq)
 218                 atomic_inc(&cq->nr_events);
 219         spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
 220
 221         if (!cq)
 222                 return;
 223
 224         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
 225
 226         if (atomic_dec_and_test(&cq->nr_events))
 227                 wake_up(&cq->wait_completion);
 228
 229         return;
 230 }
 231
 232 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
 233 {
 234         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
 235
 236         switch (identifier) {
 237         case 0x02: /* path migrated */
 238                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG);
 239                 break;
 240         case 0x03: /* communication established */
 241                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST);
 242                 break;
 243         case 0x04: /* send queue drained */
 244                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED);
 245                 break;
 246         case 0x05: /* QP error */
 247         case 0x06: /* QP error */
 248                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL);
 249                 break;
 250         case 0x07: /* CQ error */
 251         case 0x08: /* CQ error */
 252                 cq_event_callback(shca, eqe);
 253                 break;
 254         case 0x09: /* MRMWPTE error */
 255                 ehca_err(&shca->ib_device, "MRMWPTE error.");
 256                 break;
 257         case 0x0A: /* port event */
 258                 ehca_err(&shca->ib_device, "Port event.");
 259                 break;
 260         case 0x0B: /* MR access error */
 261                 ehca_err(&shca->ib_device, "MR access error.");
 262                 break;
 263         case 0x0C: /* EQ error */
 264                 ehca_err(&shca->ib_device, "EQ error.");
 265                 break;
 266         case 0x0D: /* P/Q_Key mismatch */
 267                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
 268                 break;
 269         case 0x10: /* sampling complete */
 270                 ehca_err(&shca->ib_device, "Sampling complete.");
 271                 break;
 272         case 0x11: /* unaffiliated access error */
 273                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
 274                 break;
 275         case 0x12: /* path migrating error */
 276                 ehca_err(&shca->ib_device, "Path migration error.");
 277                 break;
 278         case 0x13: /* interface trace stopped */
 279                 ehca_err(&shca->ib_device, "Interface trace stopped.");
 280                 break;
 281         case 0x14: /* first error capture info available */
 282         default:
 283                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
 284                          identifier, shca->ib_device.name);
 285                 break;
 286         }
 287
 288         return;
 289 }
 290
 291 static void parse_ec(struct ehca_shca *shca, u64 eqe)
 292 {
 293         struct ib_event event;
 294         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
 295         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
 296
 297         switch (ec) {
 298         case 0x30: /* port availability change */
 299                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
 300                         ehca_info(&shca->ib_device,
 301                                   "port %x is active.", port);
 302                         event.device = &shca->ib_device;
 303                         event.event = IB_EVENT_PORT_ACTIVE;
 304                         event.element.port_num = port;
 305                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 306                         ib_dispatch_event(&event);
 307                 } else {
 308                         ehca_info(&shca->ib_device,
 309                                   "port %x is inactive.", port);
 310                         event.device = &shca->ib_device;
 311                         event.event = IB_EVENT_PORT_ERR;
 312                         event.element.port_num = port;
 313                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 314                         ib_dispatch_event(&event);
 315                 }
 316                 break;
 317         case 0x31:
 318                 /* port configuration change
 319                  * disruptive change is caused by
 320                  * LID, PKEY or SM change
 321                  */
 322                 ehca_warn(&shca->ib_device,
 323                           "disruptive port %x configuration change", port);
 324
 325                 ehca_info(&shca->ib_device,
 326                           "port %x is inactive.", port);
 327                 event.device = &shca->ib_device;
 328                 event.event = IB_EVENT_PORT_ERR;
 329                 event.element.port_num = port;
 330                 shca->sport[port - 1].port_state = IB_PORT_DOWN;
 331                 ib_dispatch_event(&event);
 332
 333                 ehca_info(&shca->ib_device,
 334                           "port %x is active.", port);
 335                 event.device = &shca->ib_device;
 336                 event.event = IB_EVENT_PORT_ACTIVE;
 337                 event.element.port_num = port;
 338                 shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 339                 ib_dispatch_event(&event);
 340                 break;
 341         case 0x32: /* adapter malfunction */
 342                 ehca_err(&shca->ib_device, "Adapter malfunction.");
 343                 break;
 344         case 0x33:  /* trace stopped */
 345                 ehca_err(&shca->ib_device, "Traced stopped.");
 346                 break;
 347         default:
 348                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
 349                          ec, shca->ib_device.name);
 350                 break;
 351         }
 352
 353         return;
 354 }
 355
 356 static inline void reset_eq_pending(struct ehca_cq *cq)
 357 {
 358         u64 CQx_EP;
 359         struct h_galpa gal = cq->galpas.kernel;
 360
 361         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
 362         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
 363
 364         return;
 365 }
 366
 367 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
 368 {
 369         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 370
 371         tasklet_hi_schedule(&shca->neq.interrupt_task);
 372
 373         return IRQ_HANDLED;
 374 }
 375
 376 void ehca_tasklet_neq(unsigned long data)
 377 {
 378         struct ehca_shca *shca = (struct ehca_shca*)data;
 379         struct ehca_eqe *eqe;
 380         u64 ret;
 381
 382         eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 383
 384         while (eqe) {
 385                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
 386                         parse_ec(shca, eqe->entry);
 387
 388                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 389         }
 390
 391         ret = hipz_h_reset_event(shca->ipz_hca_handle,
 392                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
 393
 394         if (ret != H_SUCCESS)
 395                 ehca_err(&shca->ib_device, "Can't clear notification events.");
 396
 397         return;
 398 }
 399
 400 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 401 {
 402         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 403
 404         tasklet_hi_schedule(&shca->eq.interrupt_task);
 405
 406         return IRQ_HANDLED;
 407 }
 408
 409
 410 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 411 {
 412         u64 eqe_value;
 413         u32 token;
 414         unsigned long flags;
 415         struct ehca_cq *cq;
 416
 417         eqe_value = eqe->entry;
 418         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
 419         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 420                 ehca_dbg(&shca->ib_device, "Got completion event");
 421                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 422                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
 423                 cq = idr_find(&ehca_cq_idr, token);
 424                 if (cq)
 425                         atomic_inc(&cq->nr_events);
 426                 spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
 427                 if (cq == NULL) {
 428                         ehca_err(&shca->ib_device,
 429                                  "Invalid eqe for non-existing cq token=%x",
 430                                  token);
 431                         return;
 432                 }
 433                 reset_eq_pending(cq);
 434                 if (ehca_scaling_code)
 435                         queue_comp_task(cq);
 436                 else {
 437                         comp_event_callback(cq);
 438                         if (atomic_dec_and_test(&cq->nr_events))
 439                                 wake_up(&cq->wait_completion);
 440                 }
 441         } else {
 442                 ehca_dbg(&shca->ib_device, "Got non completion event");
 443                 parse_identifier(shca, eqe_value);
 444         }
 445 }
 446
 447 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
 448 {
 449         struct ehca_eq *eq = &shca->eq;
 450         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
 451         u64 eqe_value;
 452         unsigned long flags;
 453         int eqe_cnt, i;
 454         int eq_empty = 0;
 455
 456         spin_lock_irqsave(&eq->irq_spinlock, flags);
 457         if (is_irq) {
 458                 const int max_query_cnt = 100;
 459                 int query_cnt = 0;
 460                 int int_state = 1;
 461                 do {
 462                         int_state = hipz_h_query_int_state(
 463                                 shca->ipz_hca_handle, eq->ist);
 464                         query_cnt++;
 465                         iosync();
 466                 } while (int_state && query_cnt < max_query_cnt);
 467                 if (unlikely((query_cnt == max_query_cnt)))
 468                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
 469                                  int_state, query_cnt);
 470         }
 471
 472         /* read out all eqes */
 473         eqe_cnt = 0;
 474         do {
 475                 u32 token;
 476                 eqe_cache[eqe_cnt].eqe =
 477                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
 478                 if (!eqe_cache[eqe_cnt].eqe)
 479                         break;
 480                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
 481                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 482                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 483                         spin_lock(&ehca_cq_idr_lock);
 484                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
 485                         if (eqe_cache[eqe_cnt].cq)
 486                                 atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
 487                         spin_unlock(&ehca_cq_idr_lock);
 488                         if (!eqe_cache[eqe_cnt].cq) {
 489                                 ehca_err(&shca->ib_device,
 490                                          "Invalid eqe for non-existing cq "
 491                                          "token=%x", token);
 492                                 continue;
 493                         }
 494                 } else
 495                         eqe_cache[eqe_cnt].cq = NULL;
 496                 eqe_cnt++;
 497         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
 498         if (!eqe_cnt) {
 499                 if (is_irq)
 500                         ehca_dbg(&shca->ib_device,
 501                                  "No eqe found for irq event");
 502                 goto unlock_irq_spinlock;
 503         } else if (!is_irq)
 504                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
 505         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
 506                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
 507         /* enable irq for new packets */
 508         for (i = 0; i < eqe_cnt; i++) {
 509                 if (eq->eqe_cache[i].cq)
 510                         reset_eq_pending(eq->eqe_cache[i].cq);
 511         }
 512         /* check eq */
 513         spin_lock(&eq->spinlock);
 514         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
 515         spin_unlock(&eq->spinlock);
 516         /* call completion handler for cached eqes */
 517         for (i = 0; i < eqe_cnt; i++)
 518                 if (eq->eqe_cache[i].cq) {
 519                         if (ehca_scaling_code)
 520                                 queue_comp_task(eq->eqe_cache[i].cq);
 521                         else {
 522                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
 523                                 comp_event_callback(cq);
 524                                 if (atomic_dec_and_test(&cq->nr_events))
 525                                         wake_up(&cq->wait_completion);
 526                         }
 527                 } else {
 528                         ehca_dbg(&shca->ib_device, "Got non completion event");
 529                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
 530                 }
 531         /* poll eq if not empty */
 532         if (eq_empty)
 533                 goto unlock_irq_spinlock;
 534         do {
 535                 struct ehca_eqe *eqe;
 536                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
 537                 if (!eqe)
 538                         break;
 539                 process_eqe(shca, eqe);
 540         } while (1);
 541
 542 unlock_irq_spinlock:
 543         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
 544 }
 545
 546 void ehca_tasklet_eq(unsigned long data)
 547 {
 548         ehca_process_eq((struct ehca_shca*)data, 1);
 549 }
 550
 551 static inline int find_next_online_cpu(struct ehca_comp_pool* pool)
 552 {
 553         int cpu;
 554         unsigned long flags;
 555
 556         WARN_ON_ONCE(!in_interrupt());
 557         if (ehca_debug_level)
 558                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
 559
 560         spin_lock_irqsave(&pool->last_cpu_lock, flags);
 561         cpu = next_cpu(pool->last_cpu, cpu_online_map);
 562         if (cpu == NR_CPUS)
 563                 cpu = first_cpu(cpu_online_map);
 564         pool->last_cpu = cpu;
 565         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
 566
 567         return cpu;
 568 }
 569
 570 static void __queue_comp_task(struct ehca_cq *__cq,
 571                               struct ehca_cpu_comp_task *cct)
 572 {
 573         unsigned long flags;
 574
 575         spin_lock_irqsave(&cct->task_lock, flags);
 576         spin_lock(&__cq->task_lock);
 577
 578         if (__cq->nr_callbacks == 0) {
 579                 __cq->nr_callbacks++;
 580                 list_add_tail(&__cq->entry, &cct->cq_list);
 581                 cct->cq_jobs++;
 582                 wake_up(&cct->wait_queue);
 583         } else
 584                 __cq->nr_callbacks++;
 585
 586         spin_unlock(&__cq->task_lock);
 587         spin_unlock_irqrestore(&cct->task_lock, flags);
 588 }
 589
 590 static void queue_comp_task(struct ehca_cq *__cq)
 591 {
 592         int cpu_id;
 593         struct ehca_cpu_comp_task *cct;
 594         int cq_jobs;
 595         unsigned long flags;
 596
 597         cpu_id = find_next_online_cpu(pool);
 598         BUG_ON(!cpu_online(cpu_id));
 599
 600         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 601         BUG_ON(!cct);
 602
 603         spin_lock_irqsave(&cct->task_lock, flags);
 604         cq_jobs = cct->cq_jobs;
 605         spin_unlock_irqrestore(&cct->task_lock, flags);
 606         if (cq_jobs > 0) {
 607                 cpu_id = find_next_online_cpu(pool);
 608                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 609                 BUG_ON(!cct);
 610         }
 611
 612         __queue_comp_task(__cq, cct);
 613 }
 614
 615 static void run_comp_task(struct ehca_cpu_comp_task* cct)
 616 {
 617         struct ehca_cq *cq;
 618         unsigned long flags;
 619
 620         spin_lock_irqsave(&cct->task_lock, flags);
 621
 622         while (!list_empty(&cct->cq_list)) {
 623                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 624                 spin_unlock_irqrestore(&cct->task_lock, flags);
 625
 626                 comp_event_callback(cq);
 627                 if (atomic_dec_and_test(&cq->nr_events))
 628                         wake_up(&cq->wait_completion);
 629
 630                 spin_lock_irqsave(&cct->task_lock, flags);
 631                 spin_lock(&cq->task_lock);
 632                 cq->nr_callbacks--;
 633                 if (!cq->nr_callbacks) {
 634                         list_del_init(cct->cq_list.next);
 635                         cct->cq_jobs--;
 636                 }
 637                 spin_unlock(&cq->task_lock);
 638         }
 639
 640         spin_unlock_irqrestore(&cct->task_lock, flags);
 641 }
 642
 643 static int comp_task(void *__cct)
 644 {
 645         struct ehca_cpu_comp_task* cct = __cct;
 646         int cql_empty;
 647         DECLARE_WAITQUEUE(wait, current);
 648
 649         set_current_state(TASK_INTERRUPTIBLE);
 650         while(!kthread_should_stop()) {
 651                 add_wait_queue(&cct->wait_queue, &wait);
 652
 653                 spin_lock_irq(&cct->task_lock);
 654                 cql_empty = list_empty(&cct->cq_list);
 655                 spin_unlock_irq(&cct->task_lock);
 656                 if (cql_empty)
 657                         schedule();
 658                 else
 659                         __set_current_state(TASK_RUNNING);
 660
 661                 remove_wait_queue(&cct->wait_queue, &wait);
 662
 663                 spin_lock_irq(&cct->task_lock);
 664                 cql_empty = list_empty(&cct->cq_list);
 665                 spin_unlock_irq(&cct->task_lock);
 666                 if (!cql_empty)
 667                         run_comp_task(__cct);
 668
 669                 set_current_state(TASK_INTERRUPTIBLE);
 670         }
 671         __set_current_state(TASK_RUNNING);
 672
 673         return 0;
 674 }
 675
 676 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
 677                                             int cpu)
 678 {
 679         struct ehca_cpu_comp_task *cct;
 680
 681         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 682         spin_lock_init(&cct->task_lock);
 683         INIT_LIST_HEAD(&cct->cq_list);
 684         init_waitqueue_head(&cct->wait_queue);
 685         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
 686
 687         return cct->task;
 688 }
 689
 690 static void destroy_comp_task(struct ehca_comp_pool *pool,
 691                               int cpu)
 692 {
 693         struct ehca_cpu_comp_task *cct;
 694         struct task_struct *task;
 695         unsigned long flags_cct;
 696
 697         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 698
 699         spin_lock_irqsave(&cct->task_lock, flags_cct);
 700
 701         task = cct->task;
 702         cct->task = NULL;
 703         cct->cq_jobs = 0;
 704
 705         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 706
 707         if (task)
 708                 kthread_stop(task);
 709 }
 710
 711 #ifdef CONFIG_HOTPLUG_CPU
 712 static void take_over_work(struct ehca_comp_pool *pool,
 713                            int cpu)
 714 {
 715         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 716         LIST_HEAD(list);
 717         struct ehca_cq *cq;
 718         unsigned long flags_cct;
 719
 720         spin_lock_irqsave(&cct->task_lock, flags_cct);
 721
 722         list_splice_init(&cct->cq_list, &list);
 723
 724         while(!list_empty(&list)) {
 725                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 726
 727                 list_del(&cq->entry);
 728                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
 729                                                   smp_processor_id()));
 730         }
 731
 732         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 733
 734 }
 735
 736 static int comp_pool_callback(struct notifier_block *nfb,
 737                               unsigned long action,
 738                               void *hcpu)
 739 {
 740         unsigned int cpu = (unsigned long)hcpu;
 741         struct ehca_cpu_comp_task *cct;
 742
 743         switch (action) {
 744         case CPU_UP_PREPARE:
 745         case CPU_UP_PREPARE_FROZEN:
 746                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
 747                 if(!create_comp_task(pool, cpu)) {
 748                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
 749                         return NOTIFY_BAD;
 750                 }
 751                 break;
 752         case CPU_UP_CANCELED:
 753         case CPU_UP_CANCELED_FROZEN:
 754                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
 755                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 756                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
 757                 destroy_comp_task(pool, cpu);
 758                 break;
 759         case CPU_ONLINE:
 760         case CPU_ONLINE_FROZEN:
 761                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
 762                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 763                 kthread_bind(cct->task, cpu);
 764                 wake_up_process(cct->task);
 765                 break;
 766         case CPU_DOWN_PREPARE:
 767         case CPU_DOWN_PREPARE_FROZEN:
 768                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
 769                 break;
 770         case CPU_DOWN_FAILED:
 771         case CPU_DOWN_FAILED_FROZEN:
 772                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
 773                 break;
 774         case CPU_DEAD:
 775         case CPU_DEAD_FROZEN:
 776                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
 777                 destroy_comp_task(pool, cpu);
 778                 take_over_work(pool, cpu);
 779                 break;
 780         }
 781
 782         return NOTIFY_OK;
 783 }
 784 #endif
 785
 786 int ehca_create_comp_pool(void)
 787 {
 788         int cpu;
 789         struct task_struct *task;
 790
 791         if (!ehca_scaling_code)
 792                 return 0;
 793
 794         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
 795         if (pool == NULL)
 796                 return -ENOMEM;
 797
 798         spin_lock_init(&pool->last_cpu_lock);
 799         pool->last_cpu = any_online_cpu(cpu_online_map);
 800
 801         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
 802         if (pool->cpu_comp_tasks == NULL) {
 803                 kfree(pool);
 804                 return -EINVAL;
 805         }
 806
 807         for_each_online_cpu(cpu) {
 808                 task = create_comp_task(pool, cpu);
 809                 if (task) {
 810                         kthread_bind(task, cpu);
 811                         wake_up_process(task);
 812                 }
 813         }
 814
 815 #ifdef CONFIG_HOTPLUG_CPU
 816         comp_pool_callback_nb.notifier_call = comp_pool_callback;
 817         comp_pool_callback_nb.priority =0;
 818         register_cpu_notifier(&comp_pool_callback_nb);
 819 #endif
 820
 821         printk(KERN_INFO "eHCA scaling code enabled\n");
 822
 823         return 0;
 824 }
 825
 826 void ehca_destroy_comp_pool(void)
 827 {
 828         int i;
 829
 830         if (!ehca_scaling_code)
 831                 return;
 832
 833 #ifdef CONFIG_HOTPLUG_CPU
 834         unregister_cpu_notifier(&comp_pool_callback_nb);
 835 #endif
 836
 837         for (i = 0; i < NR_CPUS; i++) {
 838                 if (cpu_online(i))
 839                         destroy_comp_task(pool, i);
 840         }
 841         free_percpu(pool->cpu_comp_tasks);
 842         kfree(pool);
 843 }