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