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