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