drivers/infiniband/hw/ehca/ehca_irq.c

   1 /*
   2  *  IBM eServer eHCA Infiniband device driver for Linux on POWER
   3  *
   4  *  Functions for EQs, NEQs and interrupts
   5  *
   6  *  Authors: Heiko J Schick <schickhj@de.ibm.com>
   7  *           Khadija Souissi <souissi@de.ibm.com>
   8  *           Hoang-Nam Nguyen <hnguyen@de.ibm.com>
   9  *           Joachim Fenkes <fenkes@de.ibm.com>
  10  *
  11  *  Copyright (c) 2005 IBM Corporation
  12  *
  13  *  All rights reserved.
  14  *
  15  *  This source code is distributed under a dual license of GPL v2.0 and OpenIB
  16  *  BSD.
  17  *
  18  * OpenIB BSD License
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions are met:
  22  *
  23  * Redistributions of source code must retain the above copyright notice, this
  24  * list of conditions and the following disclaimer.
  25  *
  26  * Redistributions in binary form must reproduce the above copyright notice,
  27  * this list of conditions and the following disclaimer in the documentation
  28  * and/or other materials
  29  * provided with the distribution.
  30  *
  31  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  32  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  34  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
  35  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  36  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  37  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  38  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER
  39  * IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  40  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  41  * POSSIBILITY OF SUCH DAMAGE.
  42  */
  43
  44 #include "ehca_classes.h"
  45 #include "ehca_irq.h"
  46 #include "ehca_iverbs.h"
  47 #include "ehca_tools.h"
  48 #include "hcp_if.h"
  49 #include "hipz_fns.h"
  50 #include "ipz_pt_fn.h"
  51
  52 #define EQE_COMPLETION_EVENT   EHCA_BMASK_IBM( 1,  1)
  53 #define EQE_CQ_QP_NUMBER       EHCA_BMASK_IBM( 8, 31)
  54 #define EQE_EE_IDENTIFIER      EHCA_BMASK_IBM( 2,  7)
  55 #define EQE_CQ_NUMBER          EHCA_BMASK_IBM( 8, 31)
  56 #define EQE_QP_NUMBER          EHCA_BMASK_IBM( 8, 31)
  57 #define EQE_QP_TOKEN           EHCA_BMASK_IBM(32, 63)
  58 #define EQE_CQ_TOKEN           EHCA_BMASK_IBM(32, 63)
  59
  60 #define NEQE_COMPLETION_EVENT  EHCA_BMASK_IBM( 1,  1)
  61 #define NEQE_EVENT_CODE        EHCA_BMASK_IBM( 2,  7)
  62 #define NEQE_PORT_NUMBER       EHCA_BMASK_IBM( 8, 15)
  63 #define NEQE_PORT_AVAILABILITY EHCA_BMASK_IBM(16, 16)
  64 #define NEQE_DISRUPTIVE        EHCA_BMASK_IBM(16, 16)
  65
  66 #define ERROR_DATA_LENGTH      EHCA_BMASK_IBM(52, 63)
  67 #define ERROR_DATA_TYPE        EHCA_BMASK_IBM( 0,  7)
  68
  69 static void queue_comp_task(struct ehca_cq *__cq);
  70
  71 static struct ehca_comp_pool *pool;
  72
  73 static inline void comp_event_callback(struct ehca_cq *cq)
  74 {
  75         if (!cq->ib_cq.comp_handler)
  76                 return;
  77
  78         spin_lock(&cq->cb_lock);
  79         cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
  80         spin_unlock(&cq->cb_lock);
  81
  82         return;
  83 }
  84
  85 static void print_error_data(struct ehca_shca *shca, void *data,
  86                              u64 *rblock, int length)
  87 {
  88         u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
  89         u64 resource = rblock[1];
  90
  91         switch (type) {
  92         case 0x1: /* Queue Pair */
  93         {
  94                 struct ehca_qp *qp = (struct ehca_qp *)data;
  95
  96                 /* only print error data if AER is set */
  97                 if (rblock[6] == 0)
  98                         return;
  99
 100                 ehca_err(&shca->ib_device,
 101                          "QP 0x%x (resource=%lx) has errors.",
 102                          qp->ib_qp.qp_num, resource);
 103                 break;
 104         }
 105         case 0x4: /* Completion Queue */
 106         {
 107                 struct ehca_cq *cq = (struct ehca_cq *)data;
 108
 109                 ehca_err(&shca->ib_device,
 110                          "CQ 0x%x (resource=%lx) has errors.",
 111                          cq->cq_number, resource);
 112                 break;
 113         }
 114         default:
 115                 ehca_err(&shca->ib_device,
 116                          "Unknown error type: %lx on %s.",
 117                          type, shca->ib_device.name);
 118                 break;
 119         }
 120
 121         ehca_err(&shca->ib_device, "Error data is available: %lx.", resource);
 122         ehca_err(&shca->ib_device, "EHCA ----- error data begin "
 123                  "---------------------------------------------------");
 124         ehca_dmp(rblock, length, "resource=%lx", resource);
 125         ehca_err(&shca->ib_device, "EHCA ----- error data end "
 126                  "----------------------------------------------------");
 127
 128         return;
 129 }
 130
 131 int ehca_error_data(struct ehca_shca *shca, void *data,
 132                     u64 resource)
 133 {
 134
 135         unsigned long ret;
 136         u64 *rblock;
 137         unsigned long block_count;
 138
 139         rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
 140         if (!rblock) {
 141                 ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
 142                 ret = -ENOMEM;
 143                 goto error_data1;
 144         }
 145
 146         /* rblock must be 4K aligned and should be 4K large */
 147         ret = hipz_h_error_data(shca->ipz_hca_handle,
 148                                 resource,
 149                                 rblock,
 150                                 &block_count);
 151
 152         if (ret == H_R_STATE)
 153                 ehca_err(&shca->ib_device,
 154                          "No error data is available: %lx.", resource);
 155         else if (ret == H_SUCCESS) {
 156                 int length;
 157
 158                 length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
 159
 160                 if (length > EHCA_PAGESIZE)
 161                         length = EHCA_PAGESIZE;
 162
 163                 print_error_data(shca, data, rblock, length);
 164         } else
 165                 ehca_err(&shca->ib_device,
 166                          "Error data could not be fetched: %lx", resource);
 167
 168         ehca_free_fw_ctrlblock(rblock);
 169
 170 error_data1:
 171         return ret;
 172
 173 }
 174
 175 static void dispatch_qp_event(struct ehca_shca *shca, struct ehca_qp *qp,
 176                               enum ib_event_type event_type)
 177 {
 178         struct ib_event event;
 179
 180         event.device = &shca->ib_device;
 181         event.event = event_type;
 182
 183         if (qp->ext_type == EQPT_SRQ) {
 184                 if (!qp->ib_srq.event_handler)
 185                         return;
 186
 187                 event.element.srq = &qp->ib_srq;
 188                 qp->ib_srq.event_handler(&event, qp->ib_srq.srq_context);
 189         } else {
 190                 if (!qp->ib_qp.event_handler)
 191                         return;
 192
 193                 event.element.qp = &qp->ib_qp;
 194                 qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 195         }
 196 }
 197
 198 static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
 199                               enum ib_event_type event_type, int fatal)
 200 {
 201         struct ehca_qp *qp;
 202         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
 203
 204         read_lock(&ehca_qp_idr_lock);
 205         qp = idr_find(&ehca_qp_idr, token);
 206         read_unlock(&ehca_qp_idr_lock);
 207
 208         if (!qp)
 209                 return;
 210
 211         if (fatal)
 212                 ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
 213
 214         dispatch_qp_event(shca, qp, fatal && qp->ext_type == EQPT_SRQ ?
 215                           IB_EVENT_SRQ_ERR : event_type);
 216
 217         /*
 218          * eHCA only processes one WQE at a time for SRQ base QPs,
 219          * so the last WQE has been processed as soon as the QP enters
 220          * error state.
 221          */
 222         if (fatal && qp->ext_type == EQPT_SRQBASE)
 223                 dispatch_qp_event(shca, qp, IB_EVENT_QP_LAST_WQE_REACHED);
 224
 225         return;
 226 }
 227
 228 static void cq_event_callback(struct ehca_shca *shca,
 229                               u64 eqe)
 230 {
 231         struct ehca_cq *cq;
 232         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
 233
 234         read_lock(&ehca_cq_idr_lock);
 235         cq = idr_find(&ehca_cq_idr, token);
 236         if (cq)
 237                 atomic_inc(&cq->nr_events);
 238         read_unlock(&ehca_cq_idr_lock);
 239
 240         if (!cq)
 241                 return;
 242
 243         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
 244
 245         if (atomic_dec_and_test(&cq->nr_events))
 246                 wake_up(&cq->wait_completion);
 247
 248         return;
 249 }
 250
 251 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
 252 {
 253         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
 254
 255         switch (identifier) {
 256         case 0x02: /* path migrated */
 257                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG, 0);
 258                 break;
 259         case 0x03: /* communication established */
 260                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST, 0);
 261                 break;
 262         case 0x04: /* send queue drained */
 263                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED, 0);
 264                 break;
 265         case 0x05: /* QP error */
 266         case 0x06: /* QP error */
 267                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL, 1);
 268                 break;
 269         case 0x07: /* CQ error */
 270         case 0x08: /* CQ error */
 271                 cq_event_callback(shca, eqe);
 272                 break;
 273         case 0x09: /* MRMWPTE error */
 274                 ehca_err(&shca->ib_device, "MRMWPTE error.");
 275                 break;
 276         case 0x0A: /* port event */
 277                 ehca_err(&shca->ib_device, "Port event.");
 278                 break;
 279         case 0x0B: /* MR access error */
 280                 ehca_err(&shca->ib_device, "MR access error.");
 281                 break;
 282         case 0x0C: /* EQ error */
 283                 ehca_err(&shca->ib_device, "EQ error.");
 284                 break;
 285         case 0x0D: /* P/Q_Key mismatch */
 286                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
 287                 break;
 288         case 0x10: /* sampling complete */
 289                 ehca_err(&shca->ib_device, "Sampling complete.");
 290                 break;
 291         case 0x11: /* unaffiliated access error */
 292                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
 293                 break;
 294         case 0x12: /* path migrating */
 295                 ehca_err(&shca->ib_device, "Path migrating.");
 296                 break;
 297         case 0x13: /* interface trace stopped */
 298                 ehca_err(&shca->ib_device, "Interface trace stopped.");
 299                 break;
 300         case 0x14: /* first error capture info available */
 301                 ehca_info(&shca->ib_device, "First error capture available");
 302                 break;
 303         case 0x15: /* SRQ limit reached */
 304                 qp_event_callback(shca, eqe, IB_EVENT_SRQ_LIMIT_REACHED, 0);
 305                 break;
 306         default:
 307                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
 308                          identifier, shca->ib_device.name);
 309                 break;
 310         }
 311
 312         return;
 313 }
 314
 315 static void dispatch_port_event(struct ehca_shca *shca, int port_num,
 316                                 enum ib_event_type type, const char *msg)
 317 {
 318         struct ib_event event;
 319
 320         ehca_info(&shca->ib_device, "port %d %s.", port_num, msg);
 321         event.device = &shca->ib_device;
 322         event.event = type;
 323         event.element.port_num = port_num;
 324         ib_dispatch_event(&event);
 325 }
 326
 327 static void notify_port_conf_change(struct ehca_shca *shca, int port_num)
 328 {
 329         struct ehca_sma_attr  new_attr;
 330         struct ehca_sma_attr *old_attr = &shca->sport[port_num - 1].saved_attr;
 331
 332         ehca_query_sma_attr(shca, port_num, &new_attr);
 333
 334         if (new_attr.sm_sl  != old_attr->sm_sl ||
 335             new_attr.sm_lid != old_attr->sm_lid)
 336                 dispatch_port_event(shca, port_num, IB_EVENT_SM_CHANGE,
 337                                     "SM changed");
 338
 339         if (new_attr.lid != old_attr->lid ||
 340             new_attr.lmc != old_attr->lmc)
 341                 dispatch_port_event(shca, port_num, IB_EVENT_LID_CHANGE,
 342                                     "LID changed");
 343
 344         if (new_attr.pkey_tbl_len != old_attr->pkey_tbl_len ||
 345             memcmp(new_attr.pkeys, old_attr->pkeys,
 346                    sizeof(u16) * new_attr.pkey_tbl_len))
 347                 dispatch_port_event(shca, port_num, IB_EVENT_PKEY_CHANGE,
 348                                     "P_Key changed");
 349
 350         *old_attr = new_attr;
 351 }
 352
 353 static void parse_ec(struct ehca_shca *shca, u64 eqe)
 354 {
 355         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
 356         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
 357
 358         switch (ec) {
 359         case 0x30: /* port availability change */
 360                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
 361                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 362                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 363                                             "is active");
 364                         ehca_query_sma_attr(shca, port,
 365                                             &shca->sport[port - 1].saved_attr);
 366                 } else {
 367                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 368                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 369                                             "is inactive");
 370                 }
 371                 break;
 372         case 0x31:
 373                 /* port configuration change
 374                  * disruptive change is caused by
 375                  * LID, PKEY or SM change
 376                  */
 377                 if (EHCA_BMASK_GET(NEQE_DISRUPTIVE, eqe)) {
 378                         ehca_warn(&shca->ib_device, "disruptive port "
 379                                   "%d configuration change", port);
 380
 381                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 382                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 383                                             "is inactive");
 384
 385                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 386                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 387                                             "is active");
 388                 } else
 389                         notify_port_conf_change(shca, port);
 390                 break;
 391         case 0x32: /* adapter malfunction */
 392                 ehca_err(&shca->ib_device, "Adapter malfunction.");
 393                 break;
 394         case 0x33:  /* trace stopped */
 395                 ehca_err(&shca->ib_device, "Traced stopped.");
 396                 break;
 397         default:
 398                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
 399                          ec, shca->ib_device.name);
 400                 break;
 401         }
 402
 403         return;
 404 }
 405
 406 static inline void reset_eq_pending(struct ehca_cq *cq)
 407 {
 408         u64 CQx_EP;
 409         struct h_galpa gal = cq->galpas.kernel;
 410
 411         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
 412         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
 413
 414         return;
 415 }
 416
 417 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
 418 {
 419         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 420
 421         tasklet_hi_schedule(&shca->neq.interrupt_task);
 422
 423         return IRQ_HANDLED;
 424 }
 425
 426 void ehca_tasklet_neq(unsigned long data)
 427 {
 428         struct ehca_shca *shca = (struct ehca_shca*)data;
 429         struct ehca_eqe *eqe;
 430         u64 ret;
 431
 432         eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 433
 434         while (eqe) {
 435                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
 436                         parse_ec(shca, eqe->entry);
 437
 438                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 439         }
 440
 441         ret = hipz_h_reset_event(shca->ipz_hca_handle,
 442                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
 443
 444         if (ret != H_SUCCESS)
 445                 ehca_err(&shca->ib_device, "Can't clear notification events.");
 446
 447         return;
 448 }
 449
 450 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 451 {
 452         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 453
 454         tasklet_hi_schedule(&shca->eq.interrupt_task);
 455
 456         return IRQ_HANDLED;
 457 }
 458
 459
 460 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 461 {
 462         u64 eqe_value;
 463         u32 token;
 464         struct ehca_cq *cq;
 465
 466         eqe_value = eqe->entry;
 467         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
 468         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 469                 ehca_dbg(&shca->ib_device, "Got completion event");
 470                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 471                 read_lock(&ehca_cq_idr_lock);
 472                 cq = idr_find(&ehca_cq_idr, token);
 473                 if (cq)
 474                         atomic_inc(&cq->nr_events);
 475                 read_unlock(&ehca_cq_idr_lock);
 476                 if (cq == NULL) {
 477                         ehca_err(&shca->ib_device,
 478                                  "Invalid eqe for non-existing cq token=%x",
 479                                  token);
 480                         return;
 481                 }
 482                 reset_eq_pending(cq);
 483                 if (ehca_scaling_code)
 484                         queue_comp_task(cq);
 485                 else {
 486                         comp_event_callback(cq);
 487                         if (atomic_dec_and_test(&cq->nr_events))
 488                                 wake_up(&cq->wait_completion);
 489                 }
 490         } else {
 491                 ehca_dbg(&shca->ib_device, "Got non completion event");
 492                 parse_identifier(shca, eqe_value);
 493         }
 494 }
 495
 496 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
 497 {
 498         struct ehca_eq *eq = &shca->eq;
 499         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
 500         u64 eqe_value;
 501         unsigned long flags;
 502         int eqe_cnt, i;
 503         int eq_empty = 0;
 504
 505         spin_lock_irqsave(&eq->irq_spinlock, flags);
 506         if (is_irq) {
 507                 const int max_query_cnt = 100;
 508                 int query_cnt = 0;
 509                 int int_state = 1;
 510                 do {
 511                         int_state = hipz_h_query_int_state(
 512                                 shca->ipz_hca_handle, eq->ist);
 513                         query_cnt++;
 514                         iosync();
 515                 } while (int_state && query_cnt < max_query_cnt);
 516                 if (unlikely((query_cnt == max_query_cnt)))
 517                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
 518                                  int_state, query_cnt);
 519         }
 520
 521         /* read out all eqes */
 522         eqe_cnt = 0;
 523         do {
 524                 u32 token;
 525                 eqe_cache[eqe_cnt].eqe =
 526                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
 527                 if (!eqe_cache[eqe_cnt].eqe)
 528                         break;
 529                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
 530                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 531                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 532                         read_lock(&ehca_cq_idr_lock);
 533                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
 534                         if (eqe_cache[eqe_cnt].cq)
 535                                 atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
 536                         read_unlock(&ehca_cq_idr_lock);
 537                         if (!eqe_cache[eqe_cnt].cq) {
 538                                 ehca_err(&shca->ib_device,
 539                                          "Invalid eqe for non-existing cq "
 540                                          "token=%x", token);
 541                                 continue;
 542                         }
 543                 } else
 544                         eqe_cache[eqe_cnt].cq = NULL;
 545                 eqe_cnt++;
 546         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
 547         if (!eqe_cnt) {
 548                 if (is_irq)
 549                         ehca_dbg(&shca->ib_device,
 550                                  "No eqe found for irq event");
 551                 goto unlock_irq_spinlock;
 552         } else if (!is_irq)
 553                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
 554         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
 555                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
 556         /* enable irq for new packets */
 557         for (i = 0; i < eqe_cnt; i++) {
 558                 if (eq->eqe_cache[i].cq)
 559                         reset_eq_pending(eq->eqe_cache[i].cq);
 560         }
 561         /* check eq */
 562         spin_lock(&eq->spinlock);
 563         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
 564         spin_unlock(&eq->spinlock);
 565         /* call completion handler for cached eqes */
 566         for (i = 0; i < eqe_cnt; i++)
 567                 if (eq->eqe_cache[i].cq) {
 568                         if (ehca_scaling_code)
 569                                 queue_comp_task(eq->eqe_cache[i].cq);
 570                         else {
 571                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
 572                                 comp_event_callback(cq);
 573                                 if (atomic_dec_and_test(&cq->nr_events))
 574                                         wake_up(&cq->wait_completion);
 575                         }
 576                 } else {
 577                         ehca_dbg(&shca->ib_device, "Got non completion event");
 578                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
 579                 }
 580         /* poll eq if not empty */
 581         if (eq_empty)
 582                 goto unlock_irq_spinlock;
 583         do {
 584                 struct ehca_eqe *eqe;
 585                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
 586                 if (!eqe)
 587                         break;
 588                 process_eqe(shca, eqe);
 589         } while (1);
 590
 591 unlock_irq_spinlock:
 592         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
 593 }
 594
 595 void ehca_tasklet_eq(unsigned long data)
 596 {
 597         ehca_process_eq((struct ehca_shca*)data, 1);
 598 }
 599
 600 static inline int find_next_online_cpu(struct ehca_comp_pool *pool)
 601 {
 602         int cpu;
 603         unsigned long flags;
 604
 605         WARN_ON_ONCE(!in_interrupt());
 606         if (ehca_debug_level)
 607                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
 608
 609         spin_lock_irqsave(&pool->last_cpu_lock, flags);
 610         cpu = next_cpu(pool->last_cpu, cpu_online_map);
 611         if (cpu == NR_CPUS)
 612                 cpu = first_cpu(cpu_online_map);
 613         pool->last_cpu = cpu;
 614         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
 615
 616         return cpu;
 617 }
 618
 619 static void __queue_comp_task(struct ehca_cq *__cq,
 620                               struct ehca_cpu_comp_task *cct)
 621 {
 622         unsigned long flags;
 623
 624         spin_lock_irqsave(&cct->task_lock, flags);
 625         spin_lock(&__cq->task_lock);
 626
 627         if (__cq->nr_callbacks == 0) {
 628                 __cq->nr_callbacks++;
 629                 list_add_tail(&__cq->entry, &cct->cq_list);
 630                 cct->cq_jobs++;
 631                 wake_up(&cct->wait_queue);
 632         } else
 633                 __cq->nr_callbacks++;
 634
 635         spin_unlock(&__cq->task_lock);
 636         spin_unlock_irqrestore(&cct->task_lock, flags);
 637 }
 638
 639 static void queue_comp_task(struct ehca_cq *__cq)
 640 {
 641         int cpu_id;
 642         struct ehca_cpu_comp_task *cct;
 643         int cq_jobs;
 644         unsigned long flags;
 645
 646         cpu_id = find_next_online_cpu(pool);
 647         BUG_ON(!cpu_online(cpu_id));
 648
 649         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 650         BUG_ON(!cct);
 651
 652         spin_lock_irqsave(&cct->task_lock, flags);
 653         cq_jobs = cct->cq_jobs;
 654         spin_unlock_irqrestore(&cct->task_lock, flags);
 655         if (cq_jobs > 0) {
 656                 cpu_id = find_next_online_cpu(pool);
 657                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 658                 BUG_ON(!cct);
 659         }
 660
 661         __queue_comp_task(__cq, cct);
 662 }
 663
 664 static void run_comp_task(struct ehca_cpu_comp_task *cct)
 665 {
 666         struct ehca_cq *cq;
 667         unsigned long flags;
 668
 669         spin_lock_irqsave(&cct->task_lock, flags);
 670
 671         while (!list_empty(&cct->cq_list)) {
 672                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 673                 spin_unlock_irqrestore(&cct->task_lock, flags);
 674
 675                 comp_event_callback(cq);
 676                 if (atomic_dec_and_test(&cq->nr_events))
 677                         wake_up(&cq->wait_completion);
 678
 679                 spin_lock_irqsave(&cct->task_lock, flags);
 680                 spin_lock(&cq->task_lock);
 681                 cq->nr_callbacks--;
 682                 if (!cq->nr_callbacks) {
 683                         list_del_init(cct->cq_list.next);
 684                         cct->cq_jobs--;
 685                 }
 686                 spin_unlock(&cq->task_lock);
 687         }
 688
 689         spin_unlock_irqrestore(&cct->task_lock, flags);
 690 }
 691
 692 static int comp_task(void *__cct)
 693 {
 694         struct ehca_cpu_comp_task *cct = __cct;
 695         int cql_empty;
 696         DECLARE_WAITQUEUE(wait, current);
 697
 698         set_current_state(TASK_INTERRUPTIBLE);
 699         while (!kthread_should_stop()) {
 700                 add_wait_queue(&cct->wait_queue, &wait);
 701
 702                 spin_lock_irq(&cct->task_lock);
 703                 cql_empty = list_empty(&cct->cq_list);
 704                 spin_unlock_irq(&cct->task_lock);
 705                 if (cql_empty)
 706                         schedule();
 707                 else
 708                         __set_current_state(TASK_RUNNING);
 709
 710                 remove_wait_queue(&cct->wait_queue, &wait);
 711
 712                 spin_lock_irq(&cct->task_lock);
 713                 cql_empty = list_empty(&cct->cq_list);
 714                 spin_unlock_irq(&cct->task_lock);
 715                 if (!cql_empty)
 716                         run_comp_task(__cct);
 717
 718                 set_current_state(TASK_INTERRUPTIBLE);
 719         }
 720         __set_current_state(TASK_RUNNING);
 721
 722         return 0;
 723 }
 724
 725 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
 726                                             int cpu)
 727 {
 728         struct ehca_cpu_comp_task *cct;
 729
 730         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 731         spin_lock_init(&cct->task_lock);
 732         INIT_LIST_HEAD(&cct->cq_list);
 733         init_waitqueue_head(&cct->wait_queue);
 734         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
 735
 736         return cct->task;
 737 }
 738
 739 static void destroy_comp_task(struct ehca_comp_pool *pool,
 740                               int cpu)
 741 {
 742         struct ehca_cpu_comp_task *cct;
 743         struct task_struct *task;
 744         unsigned long flags_cct;
 745
 746         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 747
 748         spin_lock_irqsave(&cct->task_lock, flags_cct);
 749
 750         task = cct->task;
 751         cct->task = NULL;
 752         cct->cq_jobs = 0;
 753
 754         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 755
 756         if (task)
 757                 kthread_stop(task);
 758 }
 759
 760 static void __cpuinit take_over_work(struct ehca_comp_pool *pool, int cpu)
 761 {
 762         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 763         LIST_HEAD(list);
 764         struct ehca_cq *cq;
 765         unsigned long flags_cct;
 766
 767         spin_lock_irqsave(&cct->task_lock, flags_cct);
 768
 769         list_splice_init(&cct->cq_list, &list);
 770
 771         while (!list_empty(&list)) {
 772                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 773
 774                 list_del(&cq->entry);
 775                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
 776                                                   smp_processor_id()));
 777         }
 778
 779         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 780
 781 }
 782
 783 static int __cpuinit comp_pool_callback(struct notifier_block *nfb,
 784                                         unsigned long action,
 785                                         void *hcpu)
 786 {
 787         unsigned int cpu = (unsigned long)hcpu;
 788         struct ehca_cpu_comp_task *cct;
 789
 790         switch (action) {
 791         case CPU_UP_PREPARE:
 792         case CPU_UP_PREPARE_FROZEN:
 793                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
 794                 if (!create_comp_task(pool, cpu)) {
 795                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
 796                         return NOTIFY_BAD;
 797                 }
 798                 break;
 799         case CPU_UP_CANCELED:
 800         case CPU_UP_CANCELED_FROZEN:
 801                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
 802                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 803                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
 804                 destroy_comp_task(pool, cpu);
 805                 break;
 806         case CPU_ONLINE:
 807         case CPU_ONLINE_FROZEN:
 808                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
 809                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 810                 kthread_bind(cct->task, cpu);
 811                 wake_up_process(cct->task);
 812                 break;
 813         case CPU_DOWN_PREPARE:
 814         case CPU_DOWN_PREPARE_FROZEN:
 815                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
 816                 break;
 817         case CPU_DOWN_FAILED:
 818         case CPU_DOWN_FAILED_FROZEN:
 819                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
 820                 break;
 821         case CPU_DEAD:
 822         case CPU_DEAD_FROZEN:
 823                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
 824                 destroy_comp_task(pool, cpu);
 825                 take_over_work(pool, cpu);
 826                 break;
 827         }
 828
 829         return NOTIFY_OK;
 830 }
 831
 832 static struct notifier_block comp_pool_callback_nb __cpuinitdata = {
 833         .notifier_call  = comp_pool_callback,
 834         .priority       = 0,
 835 };
 836
 837 int ehca_create_comp_pool(void)
 838 {
 839         int cpu;
 840         struct task_struct *task;
 841
 842         if (!ehca_scaling_code)
 843                 return 0;
 844
 845         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
 846         if (pool == NULL)
 847                 return -ENOMEM;
 848
 849         spin_lock_init(&pool->last_cpu_lock);
 850         pool->last_cpu = any_online_cpu(cpu_online_map);
 851
 852         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
 853         if (pool->cpu_comp_tasks == NULL) {
 854                 kfree(pool);
 855                 return -EINVAL;
 856         }
 857
 858         for_each_online_cpu(cpu) {
 859                 task = create_comp_task(pool, cpu);
 860                 if (task) {
 861                         kthread_bind(task, cpu);
 862                         wake_up_process(task);
 863                 }
 864         }
 865
 866         register_hotcpu_notifier(&comp_pool_callback_nb);
 867
 868         printk(KERN_INFO "eHCA scaling code enabled\n");
 869
 870         return 0;
 871 }
 872
 873 void ehca_destroy_comp_pool(void)
 874 {
 875         int i;
 876
 877         if (!ehca_scaling_code)
 878                 return;
 879
 880         unregister_hotcpu_notifier(&comp_pool_callback_nb);
 881
 882         for (i = 0; i < NR_CPUS; i++) {
 883                 if (cpu_online(i))
 884                         destroy_comp_task(pool, i);
 885         }
 886         free_percpu(pool->cpu_comp_tasks);
 887         kfree(pool);
 888 }