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         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
 184
 185         read_lock(&ehca_qp_idr_lock);
 186         qp = idr_find(&ehca_qp_idr, token);
 187         read_unlock(&ehca_qp_idr_lock);
 188
 189
 190         if (!qp)
 191                 return;
 192
 193         ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
 194
 195         if (!qp->ib_qp.event_handler)
 196                 return;
 197
 198         event.device     = &shca->ib_device;
 199         event.event      = event_type;
 200         event.element.qp = &qp->ib_qp;
 201
 202         qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 203
 204         return;
 205 }
 206
 207 static void cq_event_callback(struct ehca_shca *shca,
 208                               u64 eqe)
 209 {
 210         struct ehca_cq *cq;
 211         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
 212
 213         read_lock(&ehca_cq_idr_lock);
 214         cq = idr_find(&ehca_cq_idr, token);
 215         if (cq)
 216                 atomic_inc(&cq->nr_events);
 217         read_unlock(&ehca_cq_idr_lock);
 218
 219         if (!cq)
 220                 return;
 221
 222         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
 223
 224         if (atomic_dec_and_test(&cq->nr_events))
 225                 wake_up(&cq->wait_completion);
 226
 227         return;
 228 }
 229
 230 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
 231 {
 232         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
 233
 234         switch (identifier) {
 235         case 0x02: /* path migrated */
 236                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG);
 237                 break;
 238         case 0x03: /* communication established */
 239                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST);
 240                 break;
 241         case 0x04: /* send queue drained */
 242                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED);
 243                 break;
 244         case 0x05: /* QP error */
 245         case 0x06: /* QP error */
 246                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL);
 247                 break;
 248         case 0x07: /* CQ error */
 249         case 0x08: /* CQ error */
 250                 cq_event_callback(shca, eqe);
 251                 break;
 252         case 0x09: /* MRMWPTE error */
 253                 ehca_err(&shca->ib_device, "MRMWPTE error.");
 254                 break;
 255         case 0x0A: /* port event */
 256                 ehca_err(&shca->ib_device, "Port event.");
 257                 break;
 258         case 0x0B: /* MR access error */
 259                 ehca_err(&shca->ib_device, "MR access error.");
 260                 break;
 261         case 0x0C: /* EQ error */
 262                 ehca_err(&shca->ib_device, "EQ error.");
 263                 break;
 264         case 0x0D: /* P/Q_Key mismatch */
 265                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
 266                 break;
 267         case 0x10: /* sampling complete */
 268                 ehca_err(&shca->ib_device, "Sampling complete.");
 269                 break;
 270         case 0x11: /* unaffiliated access error */
 271                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
 272                 break;
 273         case 0x12: /* path migrating error */
 274                 ehca_err(&shca->ib_device, "Path migration error.");
 275                 break;
 276         case 0x13: /* interface trace stopped */
 277                 ehca_err(&shca->ib_device, "Interface trace stopped.");
 278                 break;
 279         case 0x14: /* first error capture info available */
 280         default:
 281                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
 282                          identifier, shca->ib_device.name);
 283                 break;
 284         }
 285
 286         return;
 287 }
 288
 289 static void parse_ec(struct ehca_shca *shca, u64 eqe)
 290 {
 291         struct ib_event event;
 292         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
 293         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
 294
 295         switch (ec) {
 296         case 0x30: /* port availability change */
 297                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
 298                         ehca_info(&shca->ib_device,
 299                                   "port %x is active.", port);
 300                         event.device = &shca->ib_device;
 301                         event.event = IB_EVENT_PORT_ACTIVE;
 302                         event.element.port_num = port;
 303                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 304                         ib_dispatch_event(&event);
 305                 } else {
 306                         ehca_info(&shca->ib_device,
 307                                   "port %x is inactive.", port);
 308                         event.device = &shca->ib_device;
 309                         event.event = IB_EVENT_PORT_ERR;
 310                         event.element.port_num = port;
 311                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 312                         ib_dispatch_event(&event);
 313                 }
 314                 break;
 315         case 0x31:
 316                 /* port configuration change
 317                  * disruptive change is caused by
 318                  * LID, PKEY or SM change
 319                  */
 320                 ehca_warn(&shca->ib_device,
 321                           "disruptive port %x configuration change", port);
 322
 323                 ehca_info(&shca->ib_device,
 324                           "port %x is inactive.", port);
 325                 event.device = &shca->ib_device;
 326                 event.event = IB_EVENT_PORT_ERR;
 327                 event.element.port_num = port;
 328                 shca->sport[port - 1].port_state = IB_PORT_DOWN;
 329                 ib_dispatch_event(&event);
 330
 331                 ehca_info(&shca->ib_device,
 332                           "port %x is active.", port);
 333                 event.device = &shca->ib_device;
 334                 event.event = IB_EVENT_PORT_ACTIVE;
 335                 event.element.port_num = port;
 336                 shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 337                 ib_dispatch_event(&event);
 338                 break;
 339         case 0x32: /* adapter malfunction */
 340                 ehca_err(&shca->ib_device, "Adapter malfunction.");
 341                 break;
 342         case 0x33:  /* trace stopped */
 343                 ehca_err(&shca->ib_device, "Traced stopped.");
 344                 break;
 345         default:
 346                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
 347                          ec, shca->ib_device.name);
 348                 break;
 349         }
 350
 351         return;
 352 }
 353
 354 static inline void reset_eq_pending(struct ehca_cq *cq)
 355 {
 356         u64 CQx_EP;
 357         struct h_galpa gal = cq->galpas.kernel;
 358
 359         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
 360         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
 361
 362         return;
 363 }
 364
 365 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
 366 {
 367         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 368
 369         tasklet_hi_schedule(&shca->neq.interrupt_task);
 370
 371         return IRQ_HANDLED;
 372 }
 373
 374 void ehca_tasklet_neq(unsigned long data)
 375 {
 376         struct ehca_shca *shca = (struct ehca_shca*)data;
 377         struct ehca_eqe *eqe;
 378         u64 ret;
 379
 380         eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 381
 382         while (eqe) {
 383                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
 384                         parse_ec(shca, eqe->entry);
 385
 386                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 387         }
 388
 389         ret = hipz_h_reset_event(shca->ipz_hca_handle,
 390                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
 391
 392         if (ret != H_SUCCESS)
 393                 ehca_err(&shca->ib_device, "Can't clear notification events.");
 394
 395         return;
 396 }
 397
 398 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 399 {
 400         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 401
 402         tasklet_hi_schedule(&shca->eq.interrupt_task);
 403
 404         return IRQ_HANDLED;
 405 }
 406
 407
 408 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 409 {
 410         u64 eqe_value;
 411         u32 token;
 412         struct ehca_cq *cq;
 413
 414         eqe_value = eqe->entry;
 415         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
 416         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 417                 ehca_dbg(&shca->ib_device, "Got completion event");
 418                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 419                 read_lock(&ehca_cq_idr_lock);
 420                 cq = idr_find(&ehca_cq_idr, token);
 421                 if (cq)
 422                         atomic_inc(&cq->nr_events);
 423                 read_unlock(&ehca_cq_idr_lock);
 424                 if (cq == NULL) {
 425                         ehca_err(&shca->ib_device,
 426                                  "Invalid eqe for non-existing cq token=%x",
 427                                  token);
 428                         return;
 429                 }
 430                 reset_eq_pending(cq);
 431                 if (ehca_scaling_code)
 432                         queue_comp_task(cq);
 433                 else {
 434                         comp_event_callback(cq);
 435                         if (atomic_dec_and_test(&cq->nr_events))
 436                                 wake_up(&cq->wait_completion);
 437                 }
 438         } else {
 439                 ehca_dbg(&shca->ib_device, "Got non completion event");
 440                 parse_identifier(shca, eqe_value);
 441         }
 442 }
 443
 444 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
 445 {
 446         struct ehca_eq *eq = &shca->eq;
 447         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
 448         u64 eqe_value;
 449         unsigned long flags;
 450         int eqe_cnt, i;
 451         int eq_empty = 0;
 452
 453         spin_lock_irqsave(&eq->irq_spinlock, flags);
 454         if (is_irq) {
 455                 const int max_query_cnt = 100;
 456                 int query_cnt = 0;
 457                 int int_state = 1;
 458                 do {
 459                         int_state = hipz_h_query_int_state(
 460                                 shca->ipz_hca_handle, eq->ist);
 461                         query_cnt++;
 462                         iosync();
 463                 } while (int_state && query_cnt < max_query_cnt);
 464                 if (unlikely((query_cnt == max_query_cnt)))
 465                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
 466                                  int_state, query_cnt);
 467         }
 468
 469         /* read out all eqes */
 470         eqe_cnt = 0;
 471         do {
 472                 u32 token;
 473                 eqe_cache[eqe_cnt].eqe =
 474                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
 475                 if (!eqe_cache[eqe_cnt].eqe)
 476                         break;
 477                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
 478                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 479                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 480                         read_lock(&ehca_cq_idr_lock);
 481                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
 482                         if (eqe_cache[eqe_cnt].cq)
 483                                 atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
 484                         read_unlock(&ehca_cq_idr_lock);
 485                         if (!eqe_cache[eqe_cnt].cq) {
 486                                 ehca_err(&shca->ib_device,
 487                                          "Invalid eqe for non-existing cq "
 488                                          "token=%x", token);
 489                                 continue;
 490                         }
 491                 } else
 492                         eqe_cache[eqe_cnt].cq = NULL;
 493                 eqe_cnt++;
 494         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
 495         if (!eqe_cnt) {
 496                 if (is_irq)
 497                         ehca_dbg(&shca->ib_device,
 498                                  "No eqe found for irq event");
 499                 goto unlock_irq_spinlock;
 500         } else if (!is_irq)
 501                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
 502         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
 503                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
 504         /* enable irq for new packets */
 505         for (i = 0; i < eqe_cnt; i++) {
 506                 if (eq->eqe_cache[i].cq)
 507                         reset_eq_pending(eq->eqe_cache[i].cq);
 508         }
 509         /* check eq */
 510         spin_lock(&eq->spinlock);
 511         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
 512         spin_unlock(&eq->spinlock);
 513         /* call completion handler for cached eqes */
 514         for (i = 0; i < eqe_cnt; i++)
 515                 if (eq->eqe_cache[i].cq) {
 516                         if (ehca_scaling_code)
 517                                 queue_comp_task(eq->eqe_cache[i].cq);
 518                         else {
 519                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
 520                                 comp_event_callback(cq);
 521                                 if (atomic_dec_and_test(&cq->nr_events))
 522                                         wake_up(&cq->wait_completion);
 523                         }
 524                 } else {
 525                         ehca_dbg(&shca->ib_device, "Got non completion event");
 526                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
 527                 }
 528         /* poll eq if not empty */
 529         if (eq_empty)
 530                 goto unlock_irq_spinlock;
 531         do {
 532                 struct ehca_eqe *eqe;
 533                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
 534                 if (!eqe)
 535                         break;
 536                 process_eqe(shca, eqe);
 537         } while (1);
 538
 539 unlock_irq_spinlock:
 540         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
 541 }
 542
 543 void ehca_tasklet_eq(unsigned long data)
 544 {
 545         ehca_process_eq((struct ehca_shca*)data, 1);
 546 }
 547
 548 static inline int find_next_online_cpu(struct ehca_comp_pool* pool)
 549 {
 550         int cpu;
 551         unsigned long flags;
 552
 553         WARN_ON_ONCE(!in_interrupt());
 554         if (ehca_debug_level)
 555                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
 556
 557         spin_lock_irqsave(&pool->last_cpu_lock, flags);
 558         cpu = next_cpu(pool->last_cpu, cpu_online_map);
 559         if (cpu == NR_CPUS)
 560                 cpu = first_cpu(cpu_online_map);
 561         pool->last_cpu = cpu;
 562         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
 563
 564         return cpu;
 565 }
 566
 567 static void __queue_comp_task(struct ehca_cq *__cq,
 568                               struct ehca_cpu_comp_task *cct)
 569 {
 570         unsigned long flags;
 571
 572         spin_lock_irqsave(&cct->task_lock, flags);
 573         spin_lock(&__cq->task_lock);
 574
 575         if (__cq->nr_callbacks == 0) {
 576                 __cq->nr_callbacks++;
 577                 list_add_tail(&__cq->entry, &cct->cq_list);
 578                 cct->cq_jobs++;
 579                 wake_up(&cct->wait_queue);
 580         } else
 581                 __cq->nr_callbacks++;
 582
 583         spin_unlock(&__cq->task_lock);
 584         spin_unlock_irqrestore(&cct->task_lock, flags);
 585 }
 586
 587 static void queue_comp_task(struct ehca_cq *__cq)
 588 {
 589         int cpu_id;
 590         struct ehca_cpu_comp_task *cct;
 591         int cq_jobs;
 592         unsigned long flags;
 593
 594         cpu_id = find_next_online_cpu(pool);
 595         BUG_ON(!cpu_online(cpu_id));
 596
 597         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 598         BUG_ON(!cct);
 599
 600         spin_lock_irqsave(&cct->task_lock, flags);
 601         cq_jobs = cct->cq_jobs;
 602         spin_unlock_irqrestore(&cct->task_lock, flags);
 603         if (cq_jobs > 0) {
 604                 cpu_id = find_next_online_cpu(pool);
 605                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 606                 BUG_ON(!cct);
 607         }
 608
 609         __queue_comp_task(__cq, cct);
 610 }
 611
 612 static void run_comp_task(struct ehca_cpu_comp_task* cct)
 613 {
 614         struct ehca_cq *cq;
 615         unsigned long flags;
 616
 617         spin_lock_irqsave(&cct->task_lock, flags);
 618
 619         while (!list_empty(&cct->cq_list)) {
 620                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 621                 spin_unlock_irqrestore(&cct->task_lock, flags);
 622
 623                 comp_event_callback(cq);
 624                 if (atomic_dec_and_test(&cq->nr_events))
 625                         wake_up(&cq->wait_completion);
 626
 627                 spin_lock_irqsave(&cct->task_lock, flags);
 628                 spin_lock(&cq->task_lock);
 629                 cq->nr_callbacks--;
 630                 if (!cq->nr_callbacks) {
 631                         list_del_init(cct->cq_list.next);
 632                         cct->cq_jobs--;
 633                 }
 634                 spin_unlock(&cq->task_lock);
 635         }
 636
 637         spin_unlock_irqrestore(&cct->task_lock, flags);
 638 }
 639
 640 static int comp_task(void *__cct)
 641 {
 642         struct ehca_cpu_comp_task* cct = __cct;
 643         int cql_empty;
 644         DECLARE_WAITQUEUE(wait, current);
 645
 646         set_current_state(TASK_INTERRUPTIBLE);
 647         while(!kthread_should_stop()) {
 648                 add_wait_queue(&cct->wait_queue, &wait);
 649
 650                 spin_lock_irq(&cct->task_lock);
 651                 cql_empty = list_empty(&cct->cq_list);
 652                 spin_unlock_irq(&cct->task_lock);
 653                 if (cql_empty)
 654                         schedule();
 655                 else
 656                         __set_current_state(TASK_RUNNING);
 657
 658                 remove_wait_queue(&cct->wait_queue, &wait);
 659
 660                 spin_lock_irq(&cct->task_lock);
 661                 cql_empty = list_empty(&cct->cq_list);
 662                 spin_unlock_irq(&cct->task_lock);
 663                 if (!cql_empty)
 664                         run_comp_task(__cct);
 665
 666                 set_current_state(TASK_INTERRUPTIBLE);
 667         }
 668         __set_current_state(TASK_RUNNING);
 669
 670         return 0;
 671 }
 672
 673 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
 674                                             int cpu)
 675 {
 676         struct ehca_cpu_comp_task *cct;
 677
 678         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 679         spin_lock_init(&cct->task_lock);
 680         INIT_LIST_HEAD(&cct->cq_list);
 681         init_waitqueue_head(&cct->wait_queue);
 682         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
 683
 684         return cct->task;
 685 }
 686
 687 static void destroy_comp_task(struct ehca_comp_pool *pool,
 688                               int cpu)
 689 {
 690         struct ehca_cpu_comp_task *cct;
 691         struct task_struct *task;
 692         unsigned long flags_cct;
 693
 694         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 695
 696         spin_lock_irqsave(&cct->task_lock, flags_cct);
 697
 698         task = cct->task;
 699         cct->task = NULL;
 700         cct->cq_jobs = 0;
 701
 702         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 703
 704         if (task)
 705                 kthread_stop(task);
 706 }
 707
 708 #ifdef CONFIG_HOTPLUG_CPU
 709 static void take_over_work(struct ehca_comp_pool *pool,
 710                            int cpu)
 711 {
 712         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 713         LIST_HEAD(list);
 714         struct ehca_cq *cq;
 715         unsigned long flags_cct;
 716
 717         spin_lock_irqsave(&cct->task_lock, flags_cct);
 718
 719         list_splice_init(&cct->cq_list, &list);
 720
 721         while(!list_empty(&list)) {
 722                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 723
 724                 list_del(&cq->entry);
 725                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
 726                                                   smp_processor_id()));
 727         }
 728
 729         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 730
 731 }
 732
 733 static int comp_pool_callback(struct notifier_block *nfb,
 734                               unsigned long action,
 735                               void *hcpu)
 736 {
 737         unsigned int cpu = (unsigned long)hcpu;
 738         struct ehca_cpu_comp_task *cct;
 739
 740         switch (action) {
 741         case CPU_UP_PREPARE:
 742         case CPU_UP_PREPARE_FROZEN:
 743                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
 744                 if(!create_comp_task(pool, cpu)) {
 745                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
 746                         return NOTIFY_BAD;
 747                 }
 748                 break;
 749         case CPU_UP_CANCELED:
 750         case CPU_UP_CANCELED_FROZEN:
 751                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
 752                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 753                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
 754                 destroy_comp_task(pool, cpu);
 755                 break;
 756         case CPU_ONLINE:
 757         case CPU_ONLINE_FROZEN:
 758                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
 759                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 760                 kthread_bind(cct->task, cpu);
 761                 wake_up_process(cct->task);
 762                 break;
 763         case CPU_DOWN_PREPARE:
 764         case CPU_DOWN_PREPARE_FROZEN:
 765                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
 766                 break;
 767         case CPU_DOWN_FAILED:
 768         case CPU_DOWN_FAILED_FROZEN:
 769                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
 770                 break;
 771         case CPU_DEAD:
 772         case CPU_DEAD_FROZEN:
 773                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
 774                 destroy_comp_task(pool, cpu);
 775                 take_over_work(pool, cpu);
 776                 break;
 777         }
 778
 779         return NOTIFY_OK;
 780 }
 781 #endif
 782
 783 int ehca_create_comp_pool(void)
 784 {
 785         int cpu;
 786         struct task_struct *task;
 787
 788         if (!ehca_scaling_code)
 789                 return 0;
 790
 791         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
 792         if (pool == NULL)
 793                 return -ENOMEM;
 794
 795         spin_lock_init(&pool->last_cpu_lock);
 796         pool->last_cpu = any_online_cpu(cpu_online_map);
 797
 798         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
 799         if (pool->cpu_comp_tasks == NULL) {
 800                 kfree(pool);
 801                 return -EINVAL;
 802         }
 803
 804         for_each_online_cpu(cpu) {
 805                 task = create_comp_task(pool, cpu);
 806                 if (task) {
 807                         kthread_bind(task, cpu);
 808                         wake_up_process(task);
 809                 }
 810         }
 811
 812 #ifdef CONFIG_HOTPLUG_CPU
 813         comp_pool_callback_nb.notifier_call = comp_pool_callback;
 814         comp_pool_callback_nb.priority =0;
 815         register_cpu_notifier(&comp_pool_callback_nb);
 816 #endif
 817
 818         printk(KERN_INFO "eHCA scaling code enabled\n");
 819
 820         return 0;
 821 }
 822
 823 void ehca_destroy_comp_pool(void)
 824 {
 825         int i;
 826
 827         if (!ehca_scaling_code)
 828                 return;
 829
 830 #ifdef CONFIG_HOTPLUG_CPU
 831         unregister_cpu_notifier(&comp_pool_callback_nb);
 832 #endif
 833
 834         for (i = 0; i < NR_CPUS; i++) {
 835                 if (cpu_online(i))
 836                         destroy_comp_task(pool, i);
 837         }
 838         free_percpu(pool->cpu_comp_tasks);
 839         kfree(pool);
 840 }