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 errror 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,
 179                               u64 eqe,
 180                               enum ib_event_type event_type)
 181 {
 182         struct ib_event event;
 183         struct ehca_qp *qp;
 184         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
 185
 186         read_lock(&ehca_qp_idr_lock);
 187         qp = idr_find(&ehca_qp_idr, token);
 188         read_unlock(&ehca_qp_idr_lock);
 189
 190
 191         if (!qp)
 192                 return;
 193
 194         ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
 195
 196         if (!qp->ib_qp.event_handler)
 197                 return;
 198
 199         event.device     = &shca->ib_device;
 200         event.event      = event_type;
 201         event.element.qp = &qp->ib_qp;
 202
 203         qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 204
 205         return;
 206 }
 207
 208 static void cq_event_callback(struct ehca_shca *shca,
 209                               u64 eqe)
 210 {
 211         struct ehca_cq *cq;
 212         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
 213
 214         read_lock(&ehca_cq_idr_lock);
 215         cq = idr_find(&ehca_cq_idr, token);
 216         if (cq)
 217                 atomic_inc(&cq->nr_events);
 218         read_unlock(&ehca_cq_idr_lock);
 219
 220         if (!cq)
 221                 return;
 222
 223         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
 224
 225         if (atomic_dec_and_test(&cq->nr_events))
 226                 wake_up(&cq->wait_completion);
 227
 228         return;
 229 }
 230
 231 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
 232 {
 233         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
 234
 235         switch (identifier) {
 236         case 0x02: /* path migrated */
 237                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG);
 238                 break;
 239         case 0x03: /* communication established */
 240                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST);
 241                 break;
 242         case 0x04: /* send queue drained */
 243                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED);
 244                 break;
 245         case 0x05: /* QP error */
 246         case 0x06: /* QP error */
 247                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL);
 248                 break;
 249         case 0x07: /* CQ error */
 250         case 0x08: /* CQ error */
 251                 cq_event_callback(shca, eqe);
 252                 break;
 253         case 0x09: /* MRMWPTE error */
 254                 ehca_err(&shca->ib_device, "MRMWPTE error.");
 255                 break;
 256         case 0x0A: /* port event */
 257                 ehca_err(&shca->ib_device, "Port event.");
 258                 break;
 259         case 0x0B: /* MR access error */
 260                 ehca_err(&shca->ib_device, "MR access error.");
 261                 break;
 262         case 0x0C: /* EQ error */
 263                 ehca_err(&shca->ib_device, "EQ error.");
 264                 break;
 265         case 0x0D: /* P/Q_Key mismatch */
 266                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
 267                 break;
 268         case 0x10: /* sampling complete */
 269                 ehca_err(&shca->ib_device, "Sampling complete.");
 270                 break;
 271         case 0x11: /* unaffiliated access error */
 272                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
 273                 break;
 274         case 0x12: /* path migrating error */
 275                 ehca_err(&shca->ib_device, "Path migration error.");
 276                 break;
 277         case 0x13: /* interface trace stopped */
 278                 ehca_err(&shca->ib_device, "Interface trace stopped.");
 279                 break;
 280         case 0x14: /* first error capture info available */
 281         default:
 282                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
 283                          identifier, shca->ib_device.name);
 284                 break;
 285         }
 286
 287         return;
 288 }
 289
 290 static void dispatch_port_event(struct ehca_shca *shca, int port_num,
 291                                 enum ib_event_type type, const char *msg)
 292 {
 293         struct ib_event event;
 294
 295         ehca_info(&shca->ib_device, "port %d %s.", port_num, msg);
 296         event.device = &shca->ib_device;
 297         event.event = type;
 298         event.element.port_num = port_num;
 299         ib_dispatch_event(&event);
 300 }
 301
 302 static void notify_port_conf_change(struct ehca_shca *shca, int port_num)
 303 {
 304         struct ehca_sma_attr  new_attr;
 305         struct ehca_sma_attr *old_attr = &shca->sport[port_num - 1].saved_attr;
 306
 307         ehca_query_sma_attr(shca, port_num, &new_attr);
 308
 309         if (new_attr.sm_sl  != old_attr->sm_sl ||
 310             new_attr.sm_lid != old_attr->sm_lid)
 311                 dispatch_port_event(shca, port_num, IB_EVENT_SM_CHANGE,
 312                                     "SM changed");
 313
 314         if (new_attr.lid != old_attr->lid ||
 315             new_attr.lmc != old_attr->lmc)
 316                 dispatch_port_event(shca, port_num, IB_EVENT_LID_CHANGE,
 317                                     "LID changed");
 318
 319         if (new_attr.pkey_tbl_len != old_attr->pkey_tbl_len ||
 320             memcmp(new_attr.pkeys, old_attr->pkeys,
 321                    sizeof(u16) * new_attr.pkey_tbl_len))
 322                 dispatch_port_event(shca, port_num, IB_EVENT_PKEY_CHANGE,
 323                                     "P_Key changed");
 324
 325         *old_attr = new_attr;
 326 }
 327
 328 static void parse_ec(struct ehca_shca *shca, u64 eqe)
 329 {
 330         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
 331         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
 332
 333         switch (ec) {
 334         case 0x30: /* port availability change */
 335                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
 336                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 337                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 338                                             "is active");
 339                         ehca_query_sma_attr(shca, port,
 340                                             &shca->sport[port - 1].saved_attr);
 341                 } else {
 342                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
 343                         dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
 344                                             "is inactive");
 345                 }
 346                 break;
 347         case 0x31:
 348                 /* port configuration change
 349                  * disruptive change is caused by
 350                  * LID, PKEY or SM change
 351                  */
 352                 if (EHCA_BMASK_GET(NEQE_DISRUPTIVE, eqe)) {
 353                         ehca_warn(&shca->ib_device, "disruptive port "
 354                                   "%d configuration change", port);
 355
 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                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
 361                         dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
 362                                             "is active");
 363                 } else
 364                         notify_port_conf_change(shca, port);
 365                 break;
 366         case 0x32: /* adapter malfunction */
 367                 ehca_err(&shca->ib_device, "Adapter malfunction.");
 368                 break;
 369         case 0x33:  /* trace stopped */
 370                 ehca_err(&shca->ib_device, "Traced stopped.");
 371                 break;
 372         default:
 373                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
 374                          ec, shca->ib_device.name);
 375                 break;
 376         }
 377
 378         return;
 379 }
 380
 381 static inline void reset_eq_pending(struct ehca_cq *cq)
 382 {
 383         u64 CQx_EP;
 384         struct h_galpa gal = cq->galpas.kernel;
 385
 386         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
 387         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
 388
 389         return;
 390 }
 391
 392 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
 393 {
 394         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 395
 396         tasklet_hi_schedule(&shca->neq.interrupt_task);
 397
 398         return IRQ_HANDLED;
 399 }
 400
 401 void ehca_tasklet_neq(unsigned long data)
 402 {
 403         struct ehca_shca *shca = (struct ehca_shca*)data;
 404         struct ehca_eqe *eqe;
 405         u64 ret;
 406
 407         eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 408
 409         while (eqe) {
 410                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
 411                         parse_ec(shca, eqe->entry);
 412
 413                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
 414         }
 415
 416         ret = hipz_h_reset_event(shca->ipz_hca_handle,
 417                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
 418
 419         if (ret != H_SUCCESS)
 420                 ehca_err(&shca->ib_device, "Can't clear notification events.");
 421
 422         return;
 423 }
 424
 425 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 426 {
 427         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
 428
 429         tasklet_hi_schedule(&shca->eq.interrupt_task);
 430
 431         return IRQ_HANDLED;
 432 }
 433
 434
 435 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 436 {
 437         u64 eqe_value;
 438         u32 token;
 439         struct ehca_cq *cq;
 440
 441         eqe_value = eqe->entry;
 442         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
 443         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 444                 ehca_dbg(&shca->ib_device, "Got completion event");
 445                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 446                 read_lock(&ehca_cq_idr_lock);
 447                 cq = idr_find(&ehca_cq_idr, token);
 448                 if (cq)
 449                         atomic_inc(&cq->nr_events);
 450                 read_unlock(&ehca_cq_idr_lock);
 451                 if (cq == NULL) {
 452                         ehca_err(&shca->ib_device,
 453                                  "Invalid eqe for non-existing cq token=%x",
 454                                  token);
 455                         return;
 456                 }
 457                 reset_eq_pending(cq);
 458                 if (ehca_scaling_code)
 459                         queue_comp_task(cq);
 460                 else {
 461                         comp_event_callback(cq);
 462                         if (atomic_dec_and_test(&cq->nr_events))
 463                                 wake_up(&cq->wait_completion);
 464                 }
 465         } else {
 466                 ehca_dbg(&shca->ib_device, "Got non completion event");
 467                 parse_identifier(shca, eqe_value);
 468         }
 469 }
 470
 471 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
 472 {
 473         struct ehca_eq *eq = &shca->eq;
 474         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
 475         u64 eqe_value;
 476         unsigned long flags;
 477         int eqe_cnt, i;
 478         int eq_empty = 0;
 479
 480         spin_lock_irqsave(&eq->irq_spinlock, flags);
 481         if (is_irq) {
 482                 const int max_query_cnt = 100;
 483                 int query_cnt = 0;
 484                 int int_state = 1;
 485                 do {
 486                         int_state = hipz_h_query_int_state(
 487                                 shca->ipz_hca_handle, eq->ist);
 488                         query_cnt++;
 489                         iosync();
 490                 } while (int_state && query_cnt < max_query_cnt);
 491                 if (unlikely((query_cnt == max_query_cnt)))
 492                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
 493                                  int_state, query_cnt);
 494         }
 495
 496         /* read out all eqes */
 497         eqe_cnt = 0;
 498         do {
 499                 u32 token;
 500                 eqe_cache[eqe_cnt].eqe =
 501                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
 502                 if (!eqe_cache[eqe_cnt].eqe)
 503                         break;
 504                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
 505                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
 506                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
 507                         read_lock(&ehca_cq_idr_lock);
 508                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
 509                         if (eqe_cache[eqe_cnt].cq)
 510                                 atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
 511                         read_unlock(&ehca_cq_idr_lock);
 512                         if (!eqe_cache[eqe_cnt].cq) {
 513                                 ehca_err(&shca->ib_device,
 514                                          "Invalid eqe for non-existing cq "
 515                                          "token=%x", token);
 516                                 continue;
 517                         }
 518                 } else
 519                         eqe_cache[eqe_cnt].cq = NULL;
 520                 eqe_cnt++;
 521         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
 522         if (!eqe_cnt) {
 523                 if (is_irq)
 524                         ehca_dbg(&shca->ib_device,
 525                                  "No eqe found for irq event");
 526                 goto unlock_irq_spinlock;
 527         } else if (!is_irq)
 528                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
 529         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
 530                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
 531         /* enable irq for new packets */
 532         for (i = 0; i < eqe_cnt; i++) {
 533                 if (eq->eqe_cache[i].cq)
 534                         reset_eq_pending(eq->eqe_cache[i].cq);
 535         }
 536         /* check eq */
 537         spin_lock(&eq->spinlock);
 538         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
 539         spin_unlock(&eq->spinlock);
 540         /* call completion handler for cached eqes */
 541         for (i = 0; i < eqe_cnt; i++)
 542                 if (eq->eqe_cache[i].cq) {
 543                         if (ehca_scaling_code)
 544                                 queue_comp_task(eq->eqe_cache[i].cq);
 545                         else {
 546                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
 547                                 comp_event_callback(cq);
 548                                 if (atomic_dec_and_test(&cq->nr_events))
 549                                         wake_up(&cq->wait_completion);
 550                         }
 551                 } else {
 552                         ehca_dbg(&shca->ib_device, "Got non completion event");
 553                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
 554                 }
 555         /* poll eq if not empty */
 556         if (eq_empty)
 557                 goto unlock_irq_spinlock;
 558         do {
 559                 struct ehca_eqe *eqe;
 560                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
 561                 if (!eqe)
 562                         break;
 563                 process_eqe(shca, eqe);
 564         } while (1);
 565
 566 unlock_irq_spinlock:
 567         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
 568 }
 569
 570 void ehca_tasklet_eq(unsigned long data)
 571 {
 572         ehca_process_eq((struct ehca_shca*)data, 1);
 573 }
 574
 575 static inline int find_next_online_cpu(struct ehca_comp_pool* pool)
 576 {
 577         int cpu;
 578         unsigned long flags;
 579
 580         WARN_ON_ONCE(!in_interrupt());
 581         if (ehca_debug_level)
 582                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
 583
 584         spin_lock_irqsave(&pool->last_cpu_lock, flags);
 585         cpu = next_cpu(pool->last_cpu, cpu_online_map);
 586         if (cpu == NR_CPUS)
 587                 cpu = first_cpu(cpu_online_map);
 588         pool->last_cpu = cpu;
 589         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
 590
 591         return cpu;
 592 }
 593
 594 static void __queue_comp_task(struct ehca_cq *__cq,
 595                               struct ehca_cpu_comp_task *cct)
 596 {
 597         unsigned long flags;
 598
 599         spin_lock_irqsave(&cct->task_lock, flags);
 600         spin_lock(&__cq->task_lock);
 601
 602         if (__cq->nr_callbacks == 0) {
 603                 __cq->nr_callbacks++;
 604                 list_add_tail(&__cq->entry, &cct->cq_list);
 605                 cct->cq_jobs++;
 606                 wake_up(&cct->wait_queue);
 607         } else
 608                 __cq->nr_callbacks++;
 609
 610         spin_unlock(&__cq->task_lock);
 611         spin_unlock_irqrestore(&cct->task_lock, flags);
 612 }
 613
 614 static void queue_comp_task(struct ehca_cq *__cq)
 615 {
 616         int cpu_id;
 617         struct ehca_cpu_comp_task *cct;
 618         int cq_jobs;
 619         unsigned long flags;
 620
 621         cpu_id = find_next_online_cpu(pool);
 622         BUG_ON(!cpu_online(cpu_id));
 623
 624         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 625         BUG_ON(!cct);
 626
 627         spin_lock_irqsave(&cct->task_lock, flags);
 628         cq_jobs = cct->cq_jobs;
 629         spin_unlock_irqrestore(&cct->task_lock, flags);
 630         if (cq_jobs > 0) {
 631                 cpu_id = find_next_online_cpu(pool);
 632                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
 633                 BUG_ON(!cct);
 634         }
 635
 636         __queue_comp_task(__cq, cct);
 637 }
 638
 639 static void run_comp_task(struct ehca_cpu_comp_task* cct)
 640 {
 641         struct ehca_cq *cq;
 642         unsigned long flags;
 643
 644         spin_lock_irqsave(&cct->task_lock, flags);
 645
 646         while (!list_empty(&cct->cq_list)) {
 647                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 648                 spin_unlock_irqrestore(&cct->task_lock, flags);
 649
 650                 comp_event_callback(cq);
 651                 if (atomic_dec_and_test(&cq->nr_events))
 652                         wake_up(&cq->wait_completion);
 653
 654                 spin_lock_irqsave(&cct->task_lock, flags);
 655                 spin_lock(&cq->task_lock);
 656                 cq->nr_callbacks--;
 657                 if (!cq->nr_callbacks) {
 658                         list_del_init(cct->cq_list.next);
 659                         cct->cq_jobs--;
 660                 }
 661                 spin_unlock(&cq->task_lock);
 662         }
 663
 664         spin_unlock_irqrestore(&cct->task_lock, flags);
 665 }
 666
 667 static int comp_task(void *__cct)
 668 {
 669         struct ehca_cpu_comp_task* cct = __cct;
 670         int cql_empty;
 671         DECLARE_WAITQUEUE(wait, current);
 672
 673         set_current_state(TASK_INTERRUPTIBLE);
 674         while(!kthread_should_stop()) {
 675                 add_wait_queue(&cct->wait_queue, &wait);
 676
 677                 spin_lock_irq(&cct->task_lock);
 678                 cql_empty = list_empty(&cct->cq_list);
 679                 spin_unlock_irq(&cct->task_lock);
 680                 if (cql_empty)
 681                         schedule();
 682                 else
 683                         __set_current_state(TASK_RUNNING);
 684
 685                 remove_wait_queue(&cct->wait_queue, &wait);
 686
 687                 spin_lock_irq(&cct->task_lock);
 688                 cql_empty = list_empty(&cct->cq_list);
 689                 spin_unlock_irq(&cct->task_lock);
 690                 if (!cql_empty)
 691                         run_comp_task(__cct);
 692
 693                 set_current_state(TASK_INTERRUPTIBLE);
 694         }
 695         __set_current_state(TASK_RUNNING);
 696
 697         return 0;
 698 }
 699
 700 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
 701                                             int cpu)
 702 {
 703         struct ehca_cpu_comp_task *cct;
 704
 705         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 706         spin_lock_init(&cct->task_lock);
 707         INIT_LIST_HEAD(&cct->cq_list);
 708         init_waitqueue_head(&cct->wait_queue);
 709         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
 710
 711         return cct->task;
 712 }
 713
 714 static void destroy_comp_task(struct ehca_comp_pool *pool,
 715                               int cpu)
 716 {
 717         struct ehca_cpu_comp_task *cct;
 718         struct task_struct *task;
 719         unsigned long flags_cct;
 720
 721         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 722
 723         spin_lock_irqsave(&cct->task_lock, flags_cct);
 724
 725         task = cct->task;
 726         cct->task = NULL;
 727         cct->cq_jobs = 0;
 728
 729         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 730
 731         if (task)
 732                 kthread_stop(task);
 733 }
 734
 735 #ifdef CONFIG_HOTPLUG_CPU
 736 static void take_over_work(struct ehca_comp_pool *pool,
 737                            int cpu)
 738 {
 739         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 740         LIST_HEAD(list);
 741         struct ehca_cq *cq;
 742         unsigned long flags_cct;
 743
 744         spin_lock_irqsave(&cct->task_lock, flags_cct);
 745
 746         list_splice_init(&cct->cq_list, &list);
 747
 748         while(!list_empty(&list)) {
 749                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
 750
 751                 list_del(&cq->entry);
 752                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
 753                                                   smp_processor_id()));
 754         }
 755
 756         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
 757
 758 }
 759
 760 static int comp_pool_callback(struct notifier_block *nfb,
 761                               unsigned long action,
 762                               void *hcpu)
 763 {
 764         unsigned int cpu = (unsigned long)hcpu;
 765         struct ehca_cpu_comp_task *cct;
 766
 767         switch (action) {
 768         case CPU_UP_PREPARE:
 769         case CPU_UP_PREPARE_FROZEN:
 770                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
 771                 if(!create_comp_task(pool, cpu)) {
 772                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
 773                         return NOTIFY_BAD;
 774                 }
 775                 break;
 776         case CPU_UP_CANCELED:
 777         case CPU_UP_CANCELED_FROZEN:
 778                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
 779                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 780                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
 781                 destroy_comp_task(pool, cpu);
 782                 break;
 783         case CPU_ONLINE:
 784         case CPU_ONLINE_FROZEN:
 785                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
 786                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
 787                 kthread_bind(cct->task, cpu);
 788                 wake_up_process(cct->task);
 789                 break;
 790         case CPU_DOWN_PREPARE:
 791         case CPU_DOWN_PREPARE_FROZEN:
 792                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
 793                 break;
 794         case CPU_DOWN_FAILED:
 795         case CPU_DOWN_FAILED_FROZEN:
 796                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
 797                 break;
 798         case CPU_DEAD:
 799         case CPU_DEAD_FROZEN:
 800                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
 801                 destroy_comp_task(pool, cpu);
 802                 take_over_work(pool, cpu);
 803                 break;
 804         }
 805
 806         return NOTIFY_OK;
 807 }
 808 #endif
 809
 810 int ehca_create_comp_pool(void)
 811 {
 812         int cpu;
 813         struct task_struct *task;
 814
 815         if (!ehca_scaling_code)
 816                 return 0;
 817
 818         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
 819         if (pool == NULL)
 820                 return -ENOMEM;
 821
 822         spin_lock_init(&pool->last_cpu_lock);
 823         pool->last_cpu = any_online_cpu(cpu_online_map);
 824
 825         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
 826         if (pool->cpu_comp_tasks == NULL) {
 827                 kfree(pool);
 828                 return -EINVAL;
 829         }
 830
 831         for_each_online_cpu(cpu) {
 832                 task = create_comp_task(pool, cpu);
 833                 if (task) {
 834                         kthread_bind(task, cpu);
 835                         wake_up_process(task);
 836                 }
 837         }
 838
 839 #ifdef CONFIG_HOTPLUG_CPU
 840         comp_pool_callback_nb.notifier_call = comp_pool_callback;
 841         comp_pool_callback_nb.priority =0;
 842         register_cpu_notifier(&comp_pool_callback_nb);
 843 #endif
 844
 845         printk(KERN_INFO "eHCA scaling code enabled\n");
 846
 847         return 0;
 848 }
 849
 850 void ehca_destroy_comp_pool(void)
 851 {
 852         int i;
 853
 854         if (!ehca_scaling_code)
 855                 return;
 856
 857 #ifdef CONFIG_HOTPLUG_CPU
 858         unregister_cpu_notifier(&comp_pool_callback_nb);
 859 #endif
 860
 861         for (i = 0; i < NR_CPUS; i++) {
 862                 if (cpu_online(i))
 863                         destroy_comp_task(pool, i);
 864         }
 865         free_percpu(pool->cpu_comp_tasks);
 866         kfree(pool);
 867 }