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 #define NEQE_SPECIFIC_EVENT    EHCA_BMASK_IBM(16, 23)
  66
  67 #define ERROR_DATA_LENGTH      EHCA_BMASK_IBM(52, 63)
  68 #define ERROR_DATA_TYPE        EHCA_BMASK_IBM( 0,  7)
  69
  70 static void queue_comp_task(struct ehca_cq *__cq);
  71
  72 static struct ehca_comp_pool *pool;
  73
  74 static inline void comp_event_callback(struct ehca_cq *cq)
  75 {
  76         if (!cq->ib_cq.comp_handler)
  77                 return;
  78
  79         spin_lock(&cq->cb_lock);
  80         cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
  81         spin_unlock(&cq->cb_lock);
  82
  83         return;
  84 }
  85
  86 static void print_error_data(struct ehca_shca *shca, void *data,
  87                              u64 *rblock, int length)
  88 {
  89         u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
  90         u64 resource = rblock[1];
  91
  92         switch (type) {
  93         case 0x1: /* Queue Pair */
  94         {
  95                 struct ehca_qp *qp = (struct ehca_qp *)data;
  96
  97                 /* only print error data if AER is set */
  98                 if (rblock[6] == 0)
  99                         return;
 100
 101                 ehca_err(&shca->ib_device,
 102                          "QP 0x%x (resource=%llx) has errors.",
 103                          qp->ib_qp.qp_num, resource);
 104                 break;
 105         }
 106         case 0x4: /* Completion Queue */
 107         {
 108                 struct ehca_cq *cq = (struct ehca_cq *)data;
 109
 110                 ehca_err(&shca->ib_device,
 111                          "CQ 0x%x (resource=%llx) has errors.",
 112                          cq->cq_number, resource);
 113                 break;
 114         }
 115         default:
 116                 ehca_err(&shca->ib_device,
 117                          "Unknown error type: %llx on %s.",
 118                          type, shca->ib_device.name);
 119                 break;
 120         }
 121
 122         ehca_err(&shca->ib_device, "Error data is available: %llx.", resource);
 123         ehca_err(&shca->ib_device, "EHCA ----- error data begin "
 124                  "---------------------------------------------------");
 125         ehca_dmp(rblock, length, "resource=%llx", resource);
 126         ehca_err(&shca->ib_device, "EHCA ----- error data end "
 127                  "----------------------------------------------------");
 128
 129         return;
 130 }
 131
 132 int ehca_error_data(struct ehca_shca *shca, void *data,
 133                     u64 resource)
 134 {
 135
 136         unsigned long ret;
 137         u64 *rblock;
 138         unsigned long block_count;
 139
 140         rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
 141         if (!rblock) {
 142                 ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
 143                 ret = -ENOMEM;
 144                 goto error_data1;
 145         }
 146
 147         /* rblock must be 4K aligned and should be 4K large */
 148         ret = hipz_h_error_data(shca->ipz_hca_handle,
 149                                 resource,
 150                                 rblock,
 151                                 &block_count);
 152
 153         if (ret == H_R_STATE)
 154                 ehca_err(&shca->ib_device,
 155                          "No error data is available: %llx.", resource);
 156         else if (ret == H_SUCCESS) {
 157                 int length;
 158
 159                 length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
 160
 161                 if (length > EHCA_PAGESIZE)
 162                         length = EHCA_PAGESIZE;
 163
 164                 print_error_data(shca, data, rblock, length);
 165         } else
 166                 ehca_err(&shca->ib_device,
 167                          "Error data could not be fetched: %llx", resource);
 168
 169         ehca_free_fw_ctrlblock(rblock);
 170
 171 error_data1:
 172         return ret;
 173
 174 }
 175
 176 static void dispatch_qp_event(struct ehca_shca *shca, struct ehca_qp *qp,
 177                               enum ib_event_type event_type)
 178 {
 179         struct ib_event event;
 180
 181         /* PATH_MIG without the QP ever having been armed is false alarm */
 182         if (event_type == IB_EVENT_PATH_MIG && !qp->mig_armed)
 183                 return;
 184
 185         event.device = &shca->ib_device;
 186         event.event = event_type;
 187
 188         if (qp->ext_type == EQPT_SRQ) {
 189                 if (!qp->ib_srq.event_handler)
 190                         return;
 191
 192                 event.element.srq = &qp->ib_srq;
 193                 qp->ib_srq.event_handler(&event, qp->ib_srq.srq_context);
 194         } else {
 195                 if (!qp->ib_qp.event_handler)
 196                         return;
 197
 198                 event.element.qp = &qp->ib_qp;
 199                 qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 200         }
 201 }
 202
 203 static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
 204                               enum ib_event_type event_type, int fatal)
 205 {
 206         struct ehca_qp *qp;
 207         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
 208
 209         read_lock(&ehca_qp_idr_lock);
 210         qp = idr_find(&ehca_qp_idr, token);
 211         if (qp)
 212                 atomic_inc(&qp->nr_events);
 213         read_unlock(&ehca_qp_idr_lock);
 214
 215         if (!qp)
 216                 return;
 217
 218         if (fatal)
 219                 ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
 220
 221         dispatch_qp_event(shca, qp, fatal && qp->ext_type == EQPT_SRQ ?
 222                           IB_EVENT_SRQ_ERR : event_type);
 223
 224         /*
 225          * eHCA only processes one WQE at a time for SRQ base QPs,
 226          * so the last WQE has been processed as soon as the QP enters
 227          * error state.
 228          */
 229         if (fatal && qp->ext_type == EQPT_SRQBASE)
 230                 dispatch_qp_event(shca, qp, IB_EVENT_QP_LAST_WQE_REACHED);
 231
 232         if (atomic_dec_and_test(&qp->nr_events))
 233                 wake_up(&qp->wait_completion);
 234         return;
 235 }
 236
 237 static void cq_event_callback(struct ehca_shca *shca,
 238                               u64 eqe)
 239 {
 240         struct ehca_cq *cq;
 241         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
 242
 243         read_lock(&ehca_cq_idr_lock);
 244         cq = idr_find(&ehca_cq_idr, token);
 245         if (cq)
 246                 atomic_inc(&cq->nr_events);
 247         read_unlock(&ehca_cq_idr_lock);
 248
 249         if (!cq)
 250                 return;
 251
 252         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
 253
 254         if (atomic_dec_and_test(&cq->nr_events))
 255                 wake_up(&cq->wait_completion);
 256
 257         return;
 258 }
 259
 260 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
 261 {
 262         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
 263
 264         switch (identifier) {
 265         case 0x02: /* path migrated */
 266                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG, 0);
 267                 break;
 268         case 0x03: /* communication established */
 269                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST, 0);
 270                 break;
 271         case 0x04: /* send queue drained */
 272                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED, 0);
 273                 break;
 274         case 0x05: /* QP error */
 275         case 0x06: /* QP error */
 276                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL, 1);
 277                 break;
 278         case 0x07: /* CQ error */
 279         case 0x08: /* CQ error */
 280                 cq_event_callback(shca, eqe);
 281                 break;
 282         case 0x09: /* MRMWPTE error */
 283                 ehca_err(&shca->ib_device, "MRMWPTE error.");
 284                 break;
 285         case 0x0A: /* port event */
 286                 ehca_err(&shca->ib_device, "Port event.");
 287                 break;
 288         case 0x0B: /* MR access error */
 289                 ehca_err(&shca->ib_device, "MR access error.");
 290                 break;
 291         case 0x0C: /* EQ error */
 292                 ehca_err(&shca->ib_device, "EQ error.");
 293                 break;
 294         case 0x0D: /* P/Q_Key mismatch */
 295                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
 296                 break;
 297         case 0x10: /* sampling complete */
 298                 ehca_err(&shca->ib_device, "Sampling complete.");
 299                 break;
 300         case 0x11: /* unaffiliated access error */
 301                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
 302                 break;
 303         case 0x12: /* path migrating */
 304                 ehca_err(&shca->ib_device, "Path migrating.");
 305                 break;
 306         case 0x13: /* interface trace stopped */
 307                 ehca_err(&shca->ib_device, "Interface trace stopped.");
 308                 break;
 309         case 0x14: /* first error capture info available */
 310                 ehca_info(&shca->ib_device, "First error capture available");
 311                 break;
 312         case 0x15: /* SRQ limit reached */
 313                 qp_event_callback(shca, eqe, IB_EVENT_SRQ_LIMIT_REACHED, 0);
 314                 break;
 315         default:
 316                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
 317                          identifier, shca->ib_device.name);
 318                 break;
 319         }
 320
 321         return;
 322 }
 323
 324 static void dispatch_port_event(struct ehca_shca *shca, int port_num,
 325                                 enum ib_event_type type, const char *msg)
 326 {
 327         struct ib_event event;
 328
 329         ehca_info(&shca->ib_device, "port %d %s.", port_num, msg);
 330         event.device = &shca->ib_device;
 331         event.event = type;
 332         event.element.port_num = port_num;
 333         ib_dispatch_event(&event);
 334 }
 335
 336 static void notify_port_conf_change(struct ehca_shca *shca, int port_num)
 337 {
 338         struct ehca_sma_attr  new_attr;
 339         struct ehca_sma_attr *old_attr = &shca->sport[port_num - 1].saved_attr;
 340
 341         ehca_query_sma_attr(shca, port_num, &new_attr);
 342
 343         if (new_attr.sm_sl  != old_attr->sm_sl ||
 344             new_attr.sm_lid != old_attr->sm_lid)
 345                 dispatch_port_event(shca, port_num, IB_EVENT_SM_CHANGE,
 346                                     "SM changed");
 347
 348         if (new_attr.lid != old_attr->lid ||
 349             new_attr.lmc != old_attr->lmc)
 350                 dispatch_port_event(shca, port_num, IB_EVENT_LID_CHANGE,
 351                                     "LID changed");
 352
 353         if (new_attr.pkey_tbl_len != old_attr->pkey_tbl_len ||
 354             memcmp(new_attr.pkeys, old_attr->pkeys,
 355                    sizeof(u16) * new_attr.pkey_tbl_len))
 356                 dispatch_port_event(shca, port_num, IB_EVENT_PKEY_CHANGE,
 357                                     "P_Key changed");
 358
 359         *old_attr = new_attr;
 360 }
 361
 362 /* replay modify_qp for sqps -- return 0 if all is well, 1 if AQP1 destroyed */
 363 static int replay_modify_qp(struct ehca_sport *sport)
 364 {
 365         int aqp1_destroyed;
 366         unsigned long flags;
 367
 368         spin_lock_irqsave(&sport->mod_sqp_lock, flags);
 369
 370         aqp1_destroyed = !sport->ibqp_sqp[IB_QPT_GSI];
 371
 372         if (sport->ibqp_sqp[IB_QPT_SMI])
 373                 ehca_recover_sqp(sport->ibqp_sqp[IB_QPT_SMI]);
 374         if (!aqp1_destroyed)
 375                 ehca_recover_sqp(sport->ibqp_sqp[IB_QPT_GSI]);
 376
 377         spin_unlock_irqrestore(&sport->mod_sqp_lock, flags);
 378
 379         return aqp1_destroyed;
 380 }
 381
 382 static void parse_ec(struct ehca_shca *shca, u64 eqe)
 383 {
 384         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
 385         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
 386         u8 spec_event;
 387         struct ehca_sport *sport = &shca->sport[port - 1];
 388
 389         switch (ec) {
 390         case 0x30: /* port availability change */
 391                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
 392                         /* only replay modify_qp calls in autodetect mode;
 393                          * if AQP1 was destroyed, the port is already down
 394                          * again and we can drop the event.
 395                          */
 396                         if (ehca_nr_ports < 0)
 397                                 if (replay_modify_qp(sport))
 398                                         break;
 399
 400                         sport->port_state = IB_PORT_ACTIVE;
 401                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 402                                             "is active");
 403                         ehca_query_sma_attr(shca, port, &sport->saved_attr);
 404                 } else {
 405                         sport->port_state = IB_PORT_DOWN;
 406                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 407                                             "is inactive");
 408                 }
 409                 break;
 410         case 0x31:
 411                 /* port configuration change
 412                  * disruptive change is caused by
 413                  * LID, PKEY or SM change
 414                  */
 415                 if (EHCA_BMASK_GET(NEQE_DISRUPTIVE, eqe)) {
 416                         ehca_warn(&shca->ib_device, "disruptive port "
 417                                   "%d configuration change", port);
 418
 419                         sport->port_state = IB_PORT_DOWN;
 420                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 421                                             "is inactive");
 422
 423                         sport->port_state = IB_PORT_ACTIVE;
 424                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 425                                             "is active");
 426                         ehca_query_sma_attr(shca, port,
 427                                             &sport->saved_attr);
 428                 } else
 429                         notify_port_conf_change(shca, port);
 430                 break;
 431         case 0x32: /* adapter malfunction */
 432                 ehca_err(&shca->ib_device, "Adapter malfunction.");
 433                 break;
 434         case 0x33:  /* trace stopped */
 435                 ehca_err(&shca->ib_device, "Traced stopped.");
 436                 break;
 437         case 0x34: /* util async event */
 438                 spec_event = EHCA_BMASK_GET(NEQE_SPECIFIC_EVENT, eqe);
 439                 if (spec_event == 0x80) /* client reregister required */
 440                         dispatch_port_event(shca, port,
 441                                             IB_EVENT_CLIENT_REREGISTER,
 442                                             "client reregister req.");
 443                 else
 444                         ehca_warn(&shca->ib_device, "Unknown util async "
 445                                   "event %x on port %x", spec_event, port);
 446                 break;
 447         default:
 448                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
 449                          ec, shca->ib_device.name);
 450                 break;
 451         }
 452
 453         return;
 454 }
 455
 456 static inline void reset_eq_pending(struct ehca_cq *cq)
 457 {
 458         u64 CQx_EP;
 459         struct h_galpa gal = cq->galpas.kernel;
 460
 461         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
 462         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
 463
 464         return;
 465 }
 466
 467 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
 468 {
 469         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 470
 471         tasklet_hi_schedule(&shca->neq.interrupt_task);
 472
 473         return IRQ_HANDLED;
 474 }
 475
 476 void ehca_tasklet_neq(unsigned long data)
 477 {
 478         struct ehca_shca *shca = (struct ehca_shca*)data;
 479         struct ehca_eqe *eqe;
 480         u64 ret;
 481
 482         eqe = ehca_poll_eq(shca, &shca->neq);
 483
 484         while (eqe) {
 485                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
 486                         parse_ec(shca, eqe->entry);
 487
 488                 eqe = ehca_poll_eq(shca, &shca->neq);
 489         }
 490
 491         ret = hipz_h_reset_event(shca->ipz_hca_handle,
 492                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
 493
 494         if (ret != H_SUCCESS)
 495                 ehca_err(&shca->ib_device, "Can't clear notification events.");
 496
 497         return;
 498 }
 499
 500 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 501 {
 502         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 503
 504         tasklet_hi_schedule(&shca->eq.interrupt_task);
 505
 506         return IRQ_HANDLED;
 507 }
 508
 509
 510 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 511 {
 512         u64 eqe_value;
 513         u32 token;
 514         struct ehca_cq *cq;
 515
 516         eqe_value = eqe->entry;
 517         ehca_dbg(&shca->ib_device, "eqe_value=%llx", eqe_value);
 518         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 519                 ehca_dbg(&shca->ib_device, "Got completion event");
 520                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 521                 read_lock(&ehca_cq_idr_lock);
 522                 cq = idr_find(&ehca_cq_idr, token);
 523                 if (cq)
 524                         atomic_inc(&cq->nr_events);
 525                 read_unlock(&ehca_cq_idr_lock);
 526                 if (cq == NULL) {
 527                         ehca_err(&shca->ib_device,
 528                                  "Invalid eqe for non-existing cq token=%x",
 529                                  token);
 530                         return;
 531                 }
 532                 reset_eq_pending(cq);
 533                 if (ehca_scaling_code)
 534                         queue_comp_task(cq);
 535                 else {
 536                         comp_event_callback(cq);
 537                         if (atomic_dec_and_test(&cq->nr_events))
 538                                 wake_up(&cq->wait_completion);
 539                 }
 540         } else {
 541                 ehca_dbg(&shca->ib_device, "Got non completion event");
 542                 parse_identifier(shca, eqe_value);
 543         }
 544 }
 545
 546 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
 547 {
 548         struct ehca_eq *eq = &shca->eq;
 549         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
 550         u64 eqe_value, ret;
 551         unsigned long flags;
 552         int eqe_cnt, i;
 553         int eq_empty = 0;
 554
 555         spin_lock_irqsave(&eq->irq_spinlock, flags);
 556         if (is_irq) {
 557                 const int max_query_cnt = 100;
 558                 int query_cnt = 0;
 559                 int int_state = 1;
 560                 do {
 561                         int_state = hipz_h_query_int_state(
 562                                 shca->ipz_hca_handle, eq->ist);
 563                         query_cnt++;
 564                         iosync();
 565                 } while (int_state && query_cnt < max_query_cnt);
 566                 if (unlikely((query_cnt == max_query_cnt)))
 567                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
 568                                  int_state, query_cnt);
 569         }
 570
 571         /* read out all eqes */
 572         eqe_cnt = 0;
 573         do {
 574                 u32 token;
 575                 eqe_cache[eqe_cnt].eqe = ehca_poll_eq(shca, eq);
 576                 if (!eqe_cache[eqe_cnt].eqe)
 577                         break;
 578                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
 579                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 580                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 581                         read_lock(&ehca_cq_idr_lock);
 582                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
 583                         if (eqe_cache[eqe_cnt].cq)
 584                                 atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
 585                         read_unlock(&ehca_cq_idr_lock);
 586                         if (!eqe_cache[eqe_cnt].cq) {
 587                                 ehca_err(&shca->ib_device,
 588                                          "Invalid eqe for non-existing cq "
 589                                          "token=%x", token);
 590                                 continue;
 591                         }
 592                 } else
 593                         eqe_cache[eqe_cnt].cq = NULL;
 594                 eqe_cnt++;
 595         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
 596         if (!eqe_cnt) {
 597                 if (is_irq)
 598                         ehca_dbg(&shca->ib_device,
 599                                  "No eqe found for irq event");
 600                 goto unlock_irq_spinlock;
 601         } else if (!is_irq) {
 602                 ret = hipz_h_eoi(eq->ist);
 603                 if (ret != H_SUCCESS)
 604                         ehca_err(&shca->ib_device,
 605                                  "bad return code EOI -rc = %lld\n", ret);
 606                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
 607         }
 608         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
 609                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
 610         /* enable irq for new packets */
 611         for (i = 0; i < eqe_cnt; i++) {
 612                 if (eq->eqe_cache[i].cq)
 613                         reset_eq_pending(eq->eqe_cache[i].cq);
 614         }
 615         /* check eq */
 616         spin_lock(&eq->spinlock);
 617         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
 618         spin_unlock(&eq->spinlock);
 619         /* call completion handler for cached eqes */
 620         for (i = 0; i < eqe_cnt; i++)
 621                 if (eq->eqe_cache[i].cq) {
 622                         if (ehca_scaling_code)
 623                                 queue_comp_task(eq->eqe_cache[i].cq);
 624                         else {
 625                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
 626                                 comp_event_callback(cq);
 627                                 if (atomic_dec_and_test(&cq->nr_events))
 628                                         wake_up(&cq->wait_completion);
 629                         }
 630                 } else {
 631                         ehca_dbg(&shca->ib_device, "Got non completion event");
 632                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
 633                 }
 634         /* poll eq if not empty */
 635         if (eq_empty)
 636                 goto unlock_irq_spinlock;
 637         do {
 638                 struct ehca_eqe *eqe;
 639                 eqe = ehca_poll_eq(shca, &shca->eq);
 640                 if (!eqe)
 641                         break;
 642                 process_eqe(shca, eqe);
 643         } while (1);
 644
 645 unlock_irq_spinlock:
 646         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
 647 }
 648
 649 void ehca_tasklet_eq(unsigned long data)
 650 {
 651         ehca_process_eq((struct ehca_shca*)data, 1);
 652 }
 653
 654 static inline int find_next_online_cpu(struct ehca_comp_pool *pool)
 655 {
 656         int cpu;
 657         unsigned long flags;
 658
 659         WARN_ON_ONCE(!in_interrupt());
 660         if (ehca_debug_level >= 3)
 661                 ehca_dmp(cpu_online_mask, cpumask_size(), "");
 662
 663         spin_lock_irqsave(&pool->last_cpu_lock, flags);
 664         cpu = cpumask_next(pool->last_cpu, cpu_online_mask);
 665         if (cpu >= nr_cpu_ids)
 666                 cpu = cpumask_first(cpu_online_mask);
 667         pool->last_cpu = cpu;
 668         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
 669
 670         return cpu;
 671 }
 672
 673 static void __queue_comp_task(struct ehca_cq *__cq,
 674                               struct ehca_cpu_comp_task *cct)
 675 {
 676         unsigned long flags;
 677
 678         spin_lock_irqsave(&cct->task_lock, flags);
 679         spin_lock(&__cq->task_lock);
 680
 681         if (__cq->nr_callbacks == 0) {
 682                 __cq->nr_callbacks++;
 683                 list_add_tail(&__cq->entry, &cct->cq_list);
 684                 cct->cq_jobs++;
 685                 wake_up(&cct->wait_queue);
 686         } else
 687                 __cq->nr_callbacks++;
 688
 689         spin_unlock(&__cq->task_lock);
 690         spin_unlock_irqrestore(&cct->task_lock, flags);
 691 }
 692
 693 static void queue_comp_task(struct ehca_cq *__cq)
 694 {
 695         int cpu_id;
 696         struct ehca_cpu_comp_task *cct;
 697         int cq_jobs;
 698         unsigned long flags;
 699
 700         cpu_id = find_next_online_cpu(pool);
 701         BUG_ON(!cpu_online(cpu_id));
 702
 703         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 704         BUG_ON(!cct);
 705
 706         spin_lock_irqsave(&cct->task_lock, flags);
 707         cq_jobs = cct->cq_jobs;
 708         spin_unlock_irqrestore(&cct->task_lock, flags);
 709         if (cq_jobs > 0) {
 710                 cpu_id = find_next_online_cpu(pool);
 711                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 712                 BUG_ON(!cct);
 713         }
 714
 715         __queue_comp_task(__cq, cct);
 716 }
 717
 718 static void run_comp_task(struct ehca_cpu_comp_task *cct)
 719 {
 720         struct ehca_cq *cq;
 721         unsigned long flags;
 722
 723         spin_lock_irqsave(&cct->task_lock, flags);
 724
 725         while (!list_empty(&cct->cq_list)) {
 726                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 727                 spin_unlock_irqrestore(&cct->task_lock, flags);
 728
 729                 comp_event_callback(cq);
 730                 if (atomic_dec_and_test(&cq->nr_events))
 731                         wake_up(&cq->wait_completion);
 732
 733                 spin_lock_irqsave(&cct->task_lock, flags);
 734                 spin_lock(&cq->task_lock);
 735                 cq->nr_callbacks--;
 736                 if (!cq->nr_callbacks) {
 737                         list_del_init(cct->cq_list.next);
 738                         cct->cq_jobs--;
 739                 }
 740                 spin_unlock(&cq->task_lock);
 741         }
 742
 743         spin_unlock_irqrestore(&cct->task_lock, flags);
 744 }
 745
 746 static int comp_task(void *__cct)
 747 {
 748         struct ehca_cpu_comp_task *cct = __cct;
 749         int cql_empty;
 750         DECLARE_WAITQUEUE(wait, current);
 751
 752         set_current_state(TASK_INTERRUPTIBLE);
 753         while (!kthread_should_stop()) {
 754                 add_wait_queue(&cct->wait_queue, &wait);
 755
 756                 spin_lock_irq(&cct->task_lock);
 757                 cql_empty = list_empty(&cct->cq_list);
 758                 spin_unlock_irq(&cct->task_lock);
 759                 if (cql_empty)
 760                         schedule();
 761                 else
 762                         __set_current_state(TASK_RUNNING);
 763
 764                 remove_wait_queue(&cct->wait_queue, &wait);
 765
 766                 spin_lock_irq(&cct->task_lock);
 767                 cql_empty = list_empty(&cct->cq_list);
 768                 spin_unlock_irq(&cct->task_lock);
 769                 if (!cql_empty)
 770                         run_comp_task(__cct);
 771
 772                 set_current_state(TASK_INTERRUPTIBLE);
 773         }
 774         __set_current_state(TASK_RUNNING);
 775
 776         return 0;
 777 }
 778
 779 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
 780                                             int cpu)
 781 {
 782         struct ehca_cpu_comp_task *cct;
 783
 784         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 785         spin_lock_init(&cct->task_lock);
 786         INIT_LIST_HEAD(&cct->cq_list);
 787         init_waitqueue_head(&cct->wait_queue);
 788         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
 789
 790         return cct->task;
 791 }
 792
 793 static void destroy_comp_task(struct ehca_comp_pool *pool,
 794                               int cpu)
 795 {
 796         struct ehca_cpu_comp_task *cct;
 797         struct task_struct *task;
 798         unsigned long flags_cct;
 799
 800         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 801
 802         spin_lock_irqsave(&cct->task_lock, flags_cct);
 803
 804         task = cct->task;
 805         cct->task = NULL;
 806         cct->cq_jobs = 0;
 807
 808         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 809
 810         if (task)
 811                 kthread_stop(task);
 812 }
 813
 814 static void __cpuinit take_over_work(struct ehca_comp_pool *pool, int cpu)
 815 {
 816         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 817         LIST_HEAD(list);
 818         struct ehca_cq *cq;
 819         unsigned long flags_cct;
 820
 821         spin_lock_irqsave(&cct->task_lock, flags_cct);
 822
 823         list_splice_init(&cct->cq_list, &list);
 824
 825         while (!list_empty(&list)) {
 826                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 827
 828                 list_del(&cq->entry);
 829                 __queue_comp_task(cq, this_cpu_ptr(pool->cpu_comp_tasks));
 830         }
 831
 832         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 833
 834 }
 835
 836 static int __cpuinit comp_pool_callback(struct notifier_block *nfb,
 837                                         unsigned long action,
 838                                         void *hcpu)
 839 {
 840         unsigned int cpu = (unsigned long)hcpu;
 841         struct ehca_cpu_comp_task *cct;
 842
 843         switch (action) {
 844         case CPU_UP_PREPARE:
 845         case CPU_UP_PREPARE_FROZEN:
 846                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
 847                 if (!create_comp_task(pool, cpu)) {
 848                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
 849                         return NOTIFY_BAD;
 850                 }
 851                 break;
 852         case CPU_UP_CANCELED:
 853         case CPU_UP_CANCELED_FROZEN:
 854                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
 855                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 856                 kthread_bind(cct->task, cpumask_any(cpu_online_mask));
 857                 destroy_comp_task(pool, cpu);
 858                 break;
 859         case CPU_ONLINE:
 860         case CPU_ONLINE_FROZEN:
 861                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
 862                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 863                 kthread_bind(cct->task, cpu);
 864                 wake_up_process(cct->task);
 865                 break;
 866         case CPU_DOWN_PREPARE:
 867         case CPU_DOWN_PREPARE_FROZEN:
 868                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
 869                 break;
 870         case CPU_DOWN_FAILED:
 871         case CPU_DOWN_FAILED_FROZEN:
 872                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
 873                 break;
 874         case CPU_DEAD:
 875         case CPU_DEAD_FROZEN:
 876                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
 877                 destroy_comp_task(pool, cpu);
 878                 take_over_work(pool, cpu);
 879                 break;
 880         }
 881
 882         return NOTIFY_OK;
 883 }
 884
 885 static struct notifier_block comp_pool_callback_nb __cpuinitdata = {
 886         .notifier_call  = comp_pool_callback,
 887         .priority       = 0,
 888 };
 889
 890 int ehca_create_comp_pool(void)
 891 {
 892         int cpu;
 893         struct task_struct *task;
 894
 895         if (!ehca_scaling_code)
 896                 return 0;
 897
 898         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
 899         if (pool == NULL)
 900                 return -ENOMEM;
 901
 902         spin_lock_init(&pool->last_cpu_lock);
 903         pool->last_cpu = cpumask_any(cpu_online_mask);
 904
 905         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
 906         if (pool->cpu_comp_tasks == NULL) {
 907                 kfree(pool);
 908                 return -EINVAL;
 909         }
 910
 911         for_each_online_cpu(cpu) {
 912                 task = create_comp_task(pool, cpu);
 913                 if (task) {
 914                         kthread_bind(task, cpu);
 915                         wake_up_process(task);
 916                 }
 917         }
 918
 919         register_hotcpu_notifier(&comp_pool_callback_nb);
 920
 921         printk(KERN_INFO "eHCA scaling code enabled\n");
 922
 923         return 0;
 924 }
 925
 926 void ehca_destroy_comp_pool(void)
 927 {
 928         int i;
 929
 930         if (!ehca_scaling_code)
 931                 return;
 932
 933         unregister_hotcpu_notifier(&comp_pool_callback_nb);
 934
 935         for_each_online_cpu(i)
 936                 destroy_comp_task(pool, i);
 937
 938         free_percpu(pool->cpu_comp_tasks);
 939         kfree(pool);
 940 }