git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/ehca/ehca_irq.c

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