git.oblomov.eu Git - linux-2.6/blob - 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 #define NEQE_DISRUPTIVE        EHCA_BMASK_IBM(16, 16)
  65
  66 #define ERROR_DATA_LENGTH      EHCA_BMASK_IBM(52, 63)
  67 #define ERROR_DATA_TYPE        EHCA_BMASK_IBM( 0,  7)
  68
  69 static void queue_comp_task(struct ehca_cq *__cq);
  70
  71 static struct ehca_comp_pool *pool;
  72 #ifdef CONFIG_HOTPLUG_CPU
  73 static struct notifier_block comp_pool_callback_nb;
  74 #endif
  75
  76 static inline void comp_event_callback(struct ehca_cq *cq)
  77 {
  78         if (!cq->ib_cq.comp_handler)
  79                 return;
  80
  81         spin_lock(&cq->cb_lock);
  82         cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
  83         spin_unlock(&cq->cb_lock);
  84
  85         return;
  86 }
  87
  88 static void print_error_data(struct ehca_shca *shca, void *data,
  89                              u64 *rblock, int length)
  90 {
  91         u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
  92         u64 resource = rblock[1];
  93
  94         switch (type) {
  95         case 0x1: /* Queue Pair */
  96         {
  97                 struct ehca_qp *qp = (struct ehca_qp *)data;
  98
  99                 /* only print error data if AER is set */
 100                 if (rblock[6] == 0)
 101                         return;
 102
 103                 ehca_err(&shca->ib_device,
 104                          "QP 0x%x (resource=%lx) has errors.",
 105                          qp->ib_qp.qp_num, resource);
 106                 break;
 107         }
 108         case 0x4: /* Completion Queue */
 109         {
 110                 struct ehca_cq *cq = (struct ehca_cq *)data;
 111
 112                 ehca_err(&shca->ib_device,
 113                          "CQ 0x%x (resource=%lx) has errors.",
 114                          cq->cq_number, resource);
 115                 break;
 116         }
 117         default:
 118                 ehca_err(&shca->ib_device,
 119                          "Unknown error type: %lx on %s.",
 120                          type, shca->ib_device.name);
 121                 break;
 122         }
 123
 124         ehca_err(&shca->ib_device, "Error data is available: %lx.", resource);
 125         ehca_err(&shca->ib_device, "EHCA ----- error data begin "
 126                  "---------------------------------------------------");
 127         ehca_dmp(rblock, length, "resource=%lx", resource);
 128         ehca_err(&shca->ib_device, "EHCA ----- error data end "
 129                  "----------------------------------------------------");
 130
 131         return;
 132 }
 133
 134 int ehca_error_data(struct ehca_shca *shca, void *data,
 135                     u64 resource)
 136 {
 137
 138         unsigned long ret;
 139         u64 *rblock;
 140         unsigned long block_count;
 141
 142         rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
 143         if (!rblock) {
 144                 ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
 145                 ret = -ENOMEM;
 146                 goto error_data1;
 147         }
 148
 149         /* rblock must be 4K aligned and should be 4K large */
 150         ret = hipz_h_error_data(shca->ipz_hca_handle,
 151                                 resource,
 152                                 rblock,
 153                                 &block_count);
 154
 155         if (ret == H_R_STATE)
 156                 ehca_err(&shca->ib_device,
 157                          "No error data is available: %lx.", resource);
 158         else if (ret == H_SUCCESS) {
 159                 int length;
 160
 161                 length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
 162
 163                 if (length > EHCA_PAGESIZE)
 164                         length = EHCA_PAGESIZE;
 165
 166                 print_error_data(shca, data, rblock, length);
 167         } else
 168                 ehca_err(&shca->ib_device,
 169                          "Error data could not be fetched: %lx", resource);
 170
 171         ehca_free_fw_ctrlblock(rblock);
 172
 173 error_data1:
 174         return ret;
 175
 176 }
 177
 178 static void dispatch_qp_event(struct ehca_shca *shca, struct ehca_qp *qp,
 179                               enum ib_event_type event_type)
 180 {
 181         struct ib_event event;
 182
 183         event.device = &shca->ib_device;
 184         event.event = event_type;
 185
 186         if (qp->ext_type == EQPT_SRQ) {
 187                 if (!qp->ib_srq.event_handler)
 188                         return;
 189
 190                 event.element.srq = &qp->ib_srq;
 191                 qp->ib_srq.event_handler(&event, qp->ib_srq.srq_context);
 192         } else {
 193                 if (!qp->ib_qp.event_handler)
 194                         return;
 195
 196                 event.element.qp = &qp->ib_qp;
 197                 qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 198         }
 199 }
 200
 201 static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
 202                               enum ib_event_type event_type, int fatal)
 203 {
 204         struct ehca_qp *qp;
 205         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
 206
 207         read_lock(&ehca_qp_idr_lock);
 208         qp = idr_find(&ehca_qp_idr, token);
 209         read_unlock(&ehca_qp_idr_lock);
 210
 211         if (!qp)
 212                 return;
 213
 214         if (fatal)
 215                 ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
 216
 217         dispatch_qp_event(shca, qp, fatal && qp->ext_type == EQPT_SRQ ?
 218                           IB_EVENT_SRQ_ERR : event_type);
 219
 220         /*
 221          * eHCA only processes one WQE at a time for SRQ base QPs,
 222          * so the last WQE has been processed as soon as the QP enters
 223          * error state.
 224          */
 225         if (fatal && qp->ext_type == EQPT_SRQBASE)
 226                 dispatch_qp_event(shca, qp, IB_EVENT_QP_LAST_WQE_REACHED);
 227
 228         return;
 229 }
 230
 231 static void cq_event_callback(struct ehca_shca *shca,
 232                               u64 eqe)
 233 {
 234         struct ehca_cq *cq;
 235         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
 236
 237         read_lock(&ehca_cq_idr_lock);
 238         cq = idr_find(&ehca_cq_idr, token);
 239         if (cq)
 240                 atomic_inc(&cq->nr_events);
 241         read_unlock(&ehca_cq_idr_lock);
 242
 243         if (!cq)
 244                 return;
 245
 246         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
 247
 248         if (atomic_dec_and_test(&cq->nr_events))
 249                 wake_up(&cq->wait_completion);
 250
 251         return;
 252 }
 253
 254 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
 255 {
 256         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
 257
 258         switch (identifier) {
 259         case 0x02: /* path migrated */
 260                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG, 0);
 261                 break;
 262         case 0x03: /* communication established */
 263                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST, 0);
 264                 break;
 265         case 0x04: /* send queue drained */
 266                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED, 0);
 267                 break;
 268         case 0x05: /* QP error */
 269         case 0x06: /* QP error */
 270                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL, 1);
 271                 break;
 272         case 0x07: /* CQ error */
 273         case 0x08: /* CQ error */
 274                 cq_event_callback(shca, eqe);
 275                 break;
 276         case 0x09: /* MRMWPTE error */
 277                 ehca_err(&shca->ib_device, "MRMWPTE error.");
 278                 break;
 279         case 0x0A: /* port event */
 280                 ehca_err(&shca->ib_device, "Port event.");
 281                 break;
 282         case 0x0B: /* MR access error */
 283                 ehca_err(&shca->ib_device, "MR access error.");
 284                 break;
 285         case 0x0C: /* EQ error */
 286                 ehca_err(&shca->ib_device, "EQ error.");
 287                 break;
 288         case 0x0D: /* P/Q_Key mismatch */
 289                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
 290                 break;
 291         case 0x10: /* sampling complete */
 292                 ehca_err(&shca->ib_device, "Sampling complete.");
 293                 break;
 294         case 0x11: /* unaffiliated access error */
 295                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
 296                 break;
 297         case 0x12: /* path migrating error */
 298                 ehca_err(&shca->ib_device, "Path migration error.");
 299                 break;
 300         case 0x13: /* interface trace stopped */
 301                 ehca_err(&shca->ib_device, "Interface trace stopped.");
 302                 break;
 303         case 0x14: /* first error capture info available */
 304                 ehca_info(&shca->ib_device, "First error capture available");
 305                 break;
 306         case 0x15: /* SRQ limit reached */
 307                 qp_event_callback(shca, eqe, IB_EVENT_SRQ_LIMIT_REACHED, 0);
 308                 break;
 309         default:
 310                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
 311                          identifier, shca->ib_device.name);
 312                 break;
 313         }
 314
 315         return;
 316 }
 317
 318 static void dispatch_port_event(struct ehca_shca *shca, int port_num,
 319                                 enum ib_event_type type, const char *msg)
 320 {
 321         struct ib_event event;
 322
 323         ehca_info(&shca->ib_device, "port %d %s.", port_num, msg);
 324         event.device = &shca->ib_device;
 325         event.event = type;
 326         event.element.port_num = port_num;
 327         ib_dispatch_event(&event);
 328 }
 329
 330 static void notify_port_conf_change(struct ehca_shca *shca, int port_num)
 331 {
 332         struct ehca_sma_attr  new_attr;
 333         struct ehca_sma_attr *old_attr = &shca->sport[port_num - 1].saved_attr;
 334
 335         ehca_query_sma_attr(shca, port_num, &new_attr);
 336
 337         if (new_attr.sm_sl  != old_attr->sm_sl ||
 338             new_attr.sm_lid != old_attr->sm_lid)
 339                 dispatch_port_event(shca, port_num, IB_EVENT_SM_CHANGE,
 340                                     "SM changed");
 341
 342         if (new_attr.lid != old_attr->lid ||
 343             new_attr.lmc != old_attr->lmc)
 344                 dispatch_port_event(shca, port_num, IB_EVENT_LID_CHANGE,
 345                                     "LID changed");
 346
 347         if (new_attr.pkey_tbl_len != old_attr->pkey_tbl_len ||
 348             memcmp(new_attr.pkeys, old_attr->pkeys,
 349                    sizeof(u16) * new_attr.pkey_tbl_len))
 350                 dispatch_port_event(shca, port_num, IB_EVENT_PKEY_CHANGE,
 351                                     "P_Key changed");
 352
 353         *old_attr = new_attr;
 354 }
 355
 356 static void parse_ec(struct ehca_shca *shca, u64 eqe)
 357 {
 358         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
 359         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
 360
 361         switch (ec) {
 362         case 0x30: /* port availability change */
 363                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
 364                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 365                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 366                                             "is active");
 367                         ehca_query_sma_attr(shca, port,
 368                                             &shca->sport[port - 1].saved_attr);
 369                 } else {
 370                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 371                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 372                                             "is inactive");
 373                 }
 374                 break;
 375         case 0x31:
 376                 /* port configuration change
 377                  * disruptive change is caused by
 378                  * LID, PKEY or SM change
 379                  */
 380                 if (EHCA_BMASK_GET(NEQE_DISRUPTIVE, eqe)) {
 381                         ehca_warn(&shca->ib_device, "disruptive port "
 382                                   "%d configuration change", port);
 383
 384                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 385                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 386                                             "is inactive");
 387
 388                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 389                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 390                                             "is active");
 391                 } else
 392                         notify_port_conf_change(shca, port);
 393                 break;
 394         case 0x32: /* adapter malfunction */
 395                 ehca_err(&shca->ib_device, "Adapter malfunction.");
 396                 break;
 397         case 0x33:  /* trace stopped */
 398                 ehca_err(&shca->ib_device, "Traced stopped.");
 399                 break;
 400         default:
 401                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
 402                          ec, shca->ib_device.name);
 403                 break;
 404         }
 405
 406         return;
 407 }
 408
 409 static inline void reset_eq_pending(struct ehca_cq *cq)
 410 {
 411         u64 CQx_EP;
 412         struct h_galpa gal = cq->galpas.kernel;
 413
 414         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
 415         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
 416
 417         return;
 418 }
 419
 420 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
 421 {
 422         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 423
 424         tasklet_hi_schedule(&shca->neq.interrupt_task);
 425
 426         return IRQ_HANDLED;
 427 }
 428
 429 void ehca_tasklet_neq(unsigned long data)
 430 {
 431         struct ehca_shca *shca = (struct ehca_shca*)data;
 432         struct ehca_eqe *eqe;
 433         u64 ret;
 434
 435         eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 436
 437         while (eqe) {
 438                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
 439                         parse_ec(shca, eqe->entry);
 440
 441                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 442         }
 443
 444         ret = hipz_h_reset_event(shca->ipz_hca_handle,
 445                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
 446
 447         if (ret != H_SUCCESS)
 448                 ehca_err(&shca->ib_device, "Can't clear notification events.");
 449
 450         return;
 451 }
 452
 453 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 454 {
 455         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 456
 457         tasklet_hi_schedule(&shca->eq.interrupt_task);
 458
 459         return IRQ_HANDLED;
 460 }
 461
 462
 463 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 464 {
 465         u64 eqe_value;
 466         u32 token;
 467         struct ehca_cq *cq;
 468
 469         eqe_value = eqe->entry;
 470         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
 471         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 472                 ehca_dbg(&shca->ib_device, "Got completion event");
 473                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 474                 read_lock(&ehca_cq_idr_lock);
 475                 cq = idr_find(&ehca_cq_idr, token);
 476                 if (cq)
 477                         atomic_inc(&cq->nr_events);
 478                 read_unlock(&ehca_cq_idr_lock);
 479                 if (cq == NULL) {
 480                         ehca_err(&shca->ib_device,
 481                                  "Invalid eqe for non-existing cq token=%x",
 482                                  token);
 483                         return;
 484                 }
 485                 reset_eq_pending(cq);
 486                 if (ehca_scaling_code)
 487                         queue_comp_task(cq);
 488                 else {
 489                         comp_event_callback(cq);
 490                         if (atomic_dec_and_test(&cq->nr_events))
 491                                 wake_up(&cq->wait_completion);
 492                 }
 493         } else {
 494                 ehca_dbg(&shca->ib_device, "Got non completion event");
 495                 parse_identifier(shca, eqe_value);
 496         }
 497 }
 498
 499 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
 500 {
 501         struct ehca_eq *eq = &shca->eq;
 502         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
 503         u64 eqe_value;
 504         unsigned long flags;
 505         int eqe_cnt, i;
 506         int eq_empty = 0;
 507
 508         spin_lock_irqsave(&eq->irq_spinlock, flags);
 509         if (is_irq) {
 510                 const int max_query_cnt = 100;
 511                 int query_cnt = 0;
 512                 int int_state = 1;
 513                 do {
 514                         int_state = hipz_h_query_int_state(
 515                                 shca->ipz_hca_handle, eq->ist);
 516                         query_cnt++;
 517                         iosync();
 518                 } while (int_state && query_cnt < max_query_cnt);
 519                 if (unlikely((query_cnt == max_query_cnt)))
 520                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
 521                                  int_state, query_cnt);
 522         }
 523
 524         /* read out all eqes */
 525         eqe_cnt = 0;
 526         do {
 527                 u32 token;
 528                 eqe_cache[eqe_cnt].eqe =
 529                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
 530                 if (!eqe_cache[eqe_cnt].eqe)
 531                         break;
 532                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
 533                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 534                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 535                         read_lock(&ehca_cq_idr_lock);
 536                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
 537                         if (eqe_cache[eqe_cnt].cq)
 538                                 atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
 539                         read_unlock(&ehca_cq_idr_lock);
 540                         if (!eqe_cache[eqe_cnt].cq) {
 541                                 ehca_err(&shca->ib_device,
 542                                          "Invalid eqe for non-existing cq "
 543                                          "token=%x", token);
 544                                 continue;
 545                         }
 546                 } else
 547                         eqe_cache[eqe_cnt].cq = NULL;
 548                 eqe_cnt++;
 549         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
 550         if (!eqe_cnt) {
 551                 if (is_irq)
 552                         ehca_dbg(&shca->ib_device,
 553                                  "No eqe found for irq event");
 554                 goto unlock_irq_spinlock;
 555         } else if (!is_irq)
 556                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
 557         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
 558                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
 559         /* enable irq for new packets */
 560         for (i = 0; i < eqe_cnt; i++) {
 561                 if (eq->eqe_cache[i].cq)
 562                         reset_eq_pending(eq->eqe_cache[i].cq);
 563         }
 564         /* check eq */
 565         spin_lock(&eq->spinlock);
 566         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
 567         spin_unlock(&eq->spinlock);
 568         /* call completion handler for cached eqes */
 569         for (i = 0; i < eqe_cnt; i++)
 570                 if (eq->eqe_cache[i].cq) {
 571                         if (ehca_scaling_code)
 572                                 queue_comp_task(eq->eqe_cache[i].cq);
 573                         else {
 574                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
 575                                 comp_event_callback(cq);
 576                                 if (atomic_dec_and_test(&cq->nr_events))
 577                                         wake_up(&cq->wait_completion);
 578                         }
 579                 } else {
 580                         ehca_dbg(&shca->ib_device, "Got non completion event");
 581                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
 582                 }
 583         /* poll eq if not empty */
 584         if (eq_empty)
 585                 goto unlock_irq_spinlock;
 586         do {
 587                 struct ehca_eqe *eqe;
 588                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
 589                 if (!eqe)
 590                         break;
 591                 process_eqe(shca, eqe);
 592         } while (1);
 593
 594 unlock_irq_spinlock:
 595         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
 596 }
 597
 598 void ehca_tasklet_eq(unsigned long data)
 599 {
 600         ehca_process_eq((struct ehca_shca*)data, 1);
 601 }
 602
 603 static inline int find_next_online_cpu(struct ehca_comp_pool *pool)
 604 {
 605         int cpu;
 606         unsigned long flags;
 607
 608         WARN_ON_ONCE(!in_interrupt());
 609         if (ehca_debug_level)
 610                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
 611
 612         spin_lock_irqsave(&pool->last_cpu_lock, flags);
 613         cpu = next_cpu(pool->last_cpu, cpu_online_map);
 614         if (cpu == NR_CPUS)
 615                 cpu = first_cpu(cpu_online_map);
 616         pool->last_cpu = cpu;
 617         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
 618
 619         return cpu;
 620 }
 621
 622 static void __queue_comp_task(struct ehca_cq *__cq,
 623                               struct ehca_cpu_comp_task *cct)
 624 {
 625         unsigned long flags;
 626
 627         spin_lock_irqsave(&cct->task_lock, flags);
 628         spin_lock(&__cq->task_lock);
 629
 630         if (__cq->nr_callbacks == 0) {
 631                 __cq->nr_callbacks++;
 632                 list_add_tail(&__cq->entry, &cct->cq_list);
 633                 cct->cq_jobs++;
 634                 wake_up(&cct->wait_queue);
 635         } else
 636                 __cq->nr_callbacks++;
 637
 638         spin_unlock(&__cq->task_lock);
 639         spin_unlock_irqrestore(&cct->task_lock, flags);
 640 }
 641
 642 static void queue_comp_task(struct ehca_cq *__cq)
 643 {
 644         int cpu_id;
 645         struct ehca_cpu_comp_task *cct;
 646         int cq_jobs;
 647         unsigned long flags;
 648
 649         cpu_id = find_next_online_cpu(pool);
 650         BUG_ON(!cpu_online(cpu_id));
 651
 652         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 653         BUG_ON(!cct);
 654
 655         spin_lock_irqsave(&cct->task_lock, flags);
 656         cq_jobs = cct->cq_jobs;
 657         spin_unlock_irqrestore(&cct->task_lock, flags);
 658         if (cq_jobs > 0) {
 659                 cpu_id = find_next_online_cpu(pool);
 660                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 661                 BUG_ON(!cct);
 662         }
 663
 664         __queue_comp_task(__cq, cct);
 665 }
 666
 667 static void run_comp_task(struct ehca_cpu_comp_task *cct)
 668 {
 669         struct ehca_cq *cq;
 670         unsigned long flags;
 671
 672         spin_lock_irqsave(&cct->task_lock, flags);
 673
 674         while (!list_empty(&cct->cq_list)) {
 675                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 676                 spin_unlock_irqrestore(&cct->task_lock, flags);
 677
 678                 comp_event_callback(cq);
 679                 if (atomic_dec_and_test(&cq->nr_events))
 680                         wake_up(&cq->wait_completion);
 681
 682                 spin_lock_irqsave(&cct->task_lock, flags);
 683                 spin_lock(&cq->task_lock);
 684                 cq->nr_callbacks--;
 685                 if (!cq->nr_callbacks) {
 686                         list_del_init(cct->cq_list.next);
 687                         cct->cq_jobs--;
 688                 }
 689                 spin_unlock(&cq->task_lock);
 690         }
 691
 692         spin_unlock_irqrestore(&cct->task_lock, flags);
 693 }
 694
 695 static int comp_task(void *__cct)
 696 {
 697         struct ehca_cpu_comp_task *cct = __cct;
 698         int cql_empty;
 699         DECLARE_WAITQUEUE(wait, current);
 700
 701         set_current_state(TASK_INTERRUPTIBLE);
 702         while (!kthread_should_stop()) {
 703                 add_wait_queue(&cct->wait_queue, &wait);
 704
 705                 spin_lock_irq(&cct->task_lock);
 706                 cql_empty = list_empty(&cct->cq_list);
 707                 spin_unlock_irq(&cct->task_lock);
 708                 if (cql_empty)
 709                         schedule();
 710                 else
 711                         __set_current_state(TASK_RUNNING);
 712
 713                 remove_wait_queue(&cct->wait_queue, &wait);
 714
 715                 spin_lock_irq(&cct->task_lock);
 716                 cql_empty = list_empty(&cct->cq_list);
 717                 spin_unlock_irq(&cct->task_lock);
 718                 if (!cql_empty)
 719                         run_comp_task(__cct);
 720
 721                 set_current_state(TASK_INTERRUPTIBLE);
 722         }
 723         __set_current_state(TASK_RUNNING);
 724
 725         return 0;
 726 }
 727
 728 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
 729                                             int cpu)
 730 {
 731         struct ehca_cpu_comp_task *cct;
 732
 733         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 734         spin_lock_init(&cct->task_lock);
 735         INIT_LIST_HEAD(&cct->cq_list);
 736         init_waitqueue_head(&cct->wait_queue);
 737         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
 738
 739         return cct->task;
 740 }
 741
 742 static void destroy_comp_task(struct ehca_comp_pool *pool,
 743                               int cpu)
 744 {
 745         struct ehca_cpu_comp_task *cct;
 746         struct task_struct *task;
 747         unsigned long flags_cct;
 748
 749         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 750
 751         spin_lock_irqsave(&cct->task_lock, flags_cct);
 752
 753         task = cct->task;
 754         cct->task = NULL;
 755         cct->cq_jobs = 0;
 756
 757         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 758
 759         if (task)
 760                 kthread_stop(task);
 761 }
 762
 763 #ifdef CONFIG_HOTPLUG_CPU
 764 static void take_over_work(struct ehca_comp_pool *pool,
 765                            int cpu)
 766 {
 767         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 768         LIST_HEAD(list);
 769         struct ehca_cq *cq;
 770         unsigned long flags_cct;
 771
 772         spin_lock_irqsave(&cct->task_lock, flags_cct);
 773
 774         list_splice_init(&cct->cq_list, &list);
 775
 776         while (!list_empty(&list)) {
 777                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 778
 779                 list_del(&cq->entry);
 780                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
 781                                                   smp_processor_id()));
 782         }
 783
 784         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 785
 786 }
 787
 788 static int comp_pool_callback(struct notifier_block *nfb,
 789                               unsigned long action,
 790                               void *hcpu)
 791 {
 792         unsigned int cpu = (unsigned long)hcpu;
 793         struct ehca_cpu_comp_task *cct;
 794
 795         switch (action) {
 796         case CPU_UP_PREPARE:
 797         case CPU_UP_PREPARE_FROZEN:
 798                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
 799                 if (!create_comp_task(pool, cpu)) {
 800                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
 801                         return NOTIFY_BAD;
 802                 }
 803                 break;
 804         case CPU_UP_CANCELED:
 805         case CPU_UP_CANCELED_FROZEN:
 806                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
 807                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 808                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
 809                 destroy_comp_task(pool, cpu);
 810                 break;
 811         case CPU_ONLINE:
 812         case CPU_ONLINE_FROZEN:
 813                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
 814                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 815                 kthread_bind(cct->task, cpu);
 816                 wake_up_process(cct->task);
 817                 break;
 818         case CPU_DOWN_PREPARE:
 819         case CPU_DOWN_PREPARE_FROZEN:
 820                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
 821                 break;
 822         case CPU_DOWN_FAILED:
 823         case CPU_DOWN_FAILED_FROZEN:
 824                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
 825                 break;
 826         case CPU_DEAD:
 827         case CPU_DEAD_FROZEN:
 828                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
 829                 destroy_comp_task(pool, cpu);
 830                 take_over_work(pool, cpu);
 831                 break;
 832         }
 833
 834         return NOTIFY_OK;
 835 }
 836 #endif
 837
 838 int ehca_create_comp_pool(void)
 839 {
 840         int cpu;
 841         struct task_struct *task;
 842
 843         if (!ehca_scaling_code)
 844                 return 0;
 845
 846         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
 847         if (pool == NULL)
 848                 return -ENOMEM;
 849
 850         spin_lock_init(&pool->last_cpu_lock);
 851         pool->last_cpu = any_online_cpu(cpu_online_map);
 852
 853         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
 854         if (pool->cpu_comp_tasks == NULL) {
 855                 kfree(pool);
 856                 return -EINVAL;
 857         }
 858
 859         for_each_online_cpu(cpu) {
 860                 task = create_comp_task(pool, cpu);
 861                 if (task) {
 862                         kthread_bind(task, cpu);
 863                         wake_up_process(task);
 864                 }
 865         }
 866
 867 #ifdef CONFIG_HOTPLUG_CPU
 868         comp_pool_callback_nb.notifier_call = comp_pool_callback;
 869         comp_pool_callback_nb.priority = 0;
 870         register_cpu_notifier(&comp_pool_callback_nb);
 871 #endif
 872
 873         printk(KERN_INFO "eHCA scaling code enabled\n");
 874
 875         return 0;
 876 }
 877
 878 void ehca_destroy_comp_pool(void)
 879 {
 880         int i;
 881
 882         if (!ehca_scaling_code)
 883                 return;
 884
 885 #ifdef CONFIG_HOTPLUG_CPU
 886         unregister_cpu_notifier(&comp_pool_callback_nb);
 887 #endif
 888
 889         for (i = 0; i < NR_CPUS; i++) {
 890                 if (cpu_online(i))
 891                         destroy_comp_task(pool, i);
 892         }
 893         free_percpu(pool->cpu_comp_tasks);
 894         kfree(pool);
 895 }