Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / 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 #ifdef CONFIG_HOTPLUG_CPU
70 static struct notifier_block comp_pool_callback_nb;
71 #endif
72
73 static inline void comp_event_callback(struct ehca_cq *cq)
74 {
75         if (!cq->ib_cq.comp_handler)
76                 return;
77
78         spin_lock(&cq->cb_lock);
79         cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
80         spin_unlock(&cq->cb_lock);
81
82         return;
83 }
84
85 static void print_error_data(struct ehca_shca * shca, void* data,
86                              u64* rblock, int length)
87 {
88         u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
89         u64 resource = rblock[1];
90
91         switch (type) {
92         case 0x1: /* Queue Pair */
93         {
94                 struct ehca_qp *qp = (struct ehca_qp*)data;
95
96                 /* only print error data if AER is set */
97                 if (rblock[6] == 0)
98                         return;
99
100                 ehca_err(&shca->ib_device,
101                          "QP 0x%x (resource=%lx) has errors.",
102                          qp->ib_qp.qp_num, resource);
103                 break;
104         }
105         case 0x4: /* Completion Queue */
106         {
107                 struct ehca_cq *cq = (struct ehca_cq*)data;
108
109                 ehca_err(&shca->ib_device,
110                          "CQ 0x%x (resource=%lx) has errors.",
111                          cq->cq_number, resource);
112                 break;
113         }
114         default:
115                 ehca_err(&shca->ib_device,
116                          "Unknown errror type: %lx on %s.",
117                          type, shca->ib_device.name);
118                 break;
119         }
120
121         ehca_err(&shca->ib_device, "Error data is available: %lx.", resource);
122         ehca_err(&shca->ib_device, "EHCA ----- error data begin "
123                  "---------------------------------------------------");
124         ehca_dmp(rblock, length, "resource=%lx", resource);
125         ehca_err(&shca->ib_device, "EHCA ----- error data end "
126                  "----------------------------------------------------");
127
128         return;
129 }
130
131 int ehca_error_data(struct ehca_shca *shca, void *data,
132                     u64 resource)
133 {
134
135         unsigned long ret;
136         u64 *rblock;
137         unsigned long block_count;
138
139         rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
140         if (!rblock) {
141                 ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
142                 ret = -ENOMEM;
143                 goto error_data1;
144         }
145
146         /* rblock must be 4K aligned and should be 4K large */
147         ret = hipz_h_error_data(shca->ipz_hca_handle,
148                                 resource,
149                                 rblock,
150                                 &block_count);
151
152         if (ret == H_R_STATE)
153                 ehca_err(&shca->ib_device,
154                          "No error data is available: %lx.", resource);
155         else if (ret == H_SUCCESS) {
156                 int length;
157
158                 length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
159
160                 if (length > EHCA_PAGESIZE)
161                         length = EHCA_PAGESIZE;
162
163                 print_error_data(shca, data, rblock, length);
164         } else
165                 ehca_err(&shca->ib_device,
166                          "Error data could not be fetched: %lx", resource);
167
168         ehca_free_fw_ctrlblock(rblock);
169
170 error_data1:
171         return ret;
172
173 }
174
175 static void qp_event_callback(struct ehca_shca *shca,
176                               u64 eqe,
177                               enum ib_event_type event_type)
178 {
179         struct ib_event event;
180         struct ehca_qp *qp;
181         unsigned long flags;
182         u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
183
184         spin_lock_irqsave(&ehca_qp_idr_lock, flags);
185         qp = idr_find(&ehca_qp_idr, token);
186         spin_unlock_irqrestore(&ehca_qp_idr_lock, flags);
187
188
189         if (!qp)
190                 return;
191
192         ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
193
194         if (!qp->ib_qp.event_handler)
195                 return;
196
197         event.device     = &shca->ib_device;
198         event.event      = event_type;
199         event.element.qp = &qp->ib_qp;
200
201         qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
202
203         return;
204 }
205
206 static void cq_event_callback(struct ehca_shca *shca,
207                               u64 eqe)
208 {
209         struct ehca_cq *cq;
210         unsigned long flags;
211         u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
212
213         spin_lock_irqsave(&ehca_cq_idr_lock, flags);
214         cq = idr_find(&ehca_cq_idr, token);
215         spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
216
217         if (!cq)
218                 return;
219
220         ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
221
222         return;
223 }
224
225 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
226 {
227         u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
228
229         switch (identifier) {
230         case 0x02: /* path migrated */
231                 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG);
232                 break;
233         case 0x03: /* communication established */
234                 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST);
235                 break;
236         case 0x04: /* send queue drained */
237                 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED);
238                 break;
239         case 0x05: /* QP error */
240         case 0x06: /* QP error */
241                 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL);
242                 break;
243         case 0x07: /* CQ error */
244         case 0x08: /* CQ error */
245                 cq_event_callback(shca, eqe);
246                 break;
247         case 0x09: /* MRMWPTE error */
248                 ehca_err(&shca->ib_device, "MRMWPTE error.");
249                 break;
250         case 0x0A: /* port event */
251                 ehca_err(&shca->ib_device, "Port event.");
252                 break;
253         case 0x0B: /* MR access error */
254                 ehca_err(&shca->ib_device, "MR access error.");
255                 break;
256         case 0x0C: /* EQ error */
257                 ehca_err(&shca->ib_device, "EQ error.");
258                 break;
259         case 0x0D: /* P/Q_Key mismatch */
260                 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
261                 break;
262         case 0x10: /* sampling complete */
263                 ehca_err(&shca->ib_device, "Sampling complete.");
264                 break;
265         case 0x11: /* unaffiliated access error */
266                 ehca_err(&shca->ib_device, "Unaffiliated access error.");
267                 break;
268         case 0x12: /* path migrating error */
269                 ehca_err(&shca->ib_device, "Path migration error.");
270                 break;
271         case 0x13: /* interface trace stopped */
272                 ehca_err(&shca->ib_device, "Interface trace stopped.");
273                 break;
274         case 0x14: /* first error capture info available */
275         default:
276                 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
277                          identifier, shca->ib_device.name);
278                 break;
279         }
280
281         return;
282 }
283
284 static void parse_ec(struct ehca_shca *shca, u64 eqe)
285 {
286         struct ib_event event;
287         u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
288         u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
289
290         switch (ec) {
291         case 0x30: /* port availability change */
292                 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
293                         ehca_info(&shca->ib_device,
294                                   "port %x is active.", port);
295                         event.device = &shca->ib_device;
296                         event.event = IB_EVENT_PORT_ACTIVE;
297                         event.element.port_num = port;
298                         shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
299                         ib_dispatch_event(&event);
300                 } else {
301                         ehca_info(&shca->ib_device,
302                                   "port %x is inactive.", port);
303                         event.device = &shca->ib_device;
304                         event.event = IB_EVENT_PORT_ERR;
305                         event.element.port_num = port;
306                         shca->sport[port - 1].port_state = IB_PORT_DOWN;
307                         ib_dispatch_event(&event);
308                 }
309                 break;
310         case 0x31:
311                 /* port configuration change
312                  * disruptive change is caused by
313                  * LID, PKEY or SM change
314                  */
315                 ehca_warn(&shca->ib_device,
316                           "disruptive port %x configuration change", port);
317
318                 ehca_info(&shca->ib_device,
319                           "port %x is inactive.", port);
320                 event.device = &shca->ib_device;
321                 event.event = IB_EVENT_PORT_ERR;
322                 event.element.port_num = port;
323                 shca->sport[port - 1].port_state = IB_PORT_DOWN;
324                 ib_dispatch_event(&event);
325
326                 ehca_info(&shca->ib_device,
327                           "port %x is active.", port);
328                 event.device = &shca->ib_device;
329                 event.event = IB_EVENT_PORT_ACTIVE;
330                 event.element.port_num = port;
331                 shca->sport[port - 1].port_state = IB_PORT_ACTIVE;
332                 ib_dispatch_event(&event);
333                 break;
334         case 0x32: /* adapter malfunction */
335                 ehca_err(&shca->ib_device, "Adapter malfunction.");
336                 break;
337         case 0x33:  /* trace stopped */
338                 ehca_err(&shca->ib_device, "Traced stopped.");
339                 break;
340         default:
341                 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
342                          ec, shca->ib_device.name);
343                 break;
344         }
345
346         return;
347 }
348
349 static inline void reset_eq_pending(struct ehca_cq *cq)
350 {
351         u64 CQx_EP;
352         struct h_galpa gal = cq->galpas.kernel;
353
354         hipz_galpa_store_cq(gal, cqx_ep, 0x0);
355         CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
356
357         return;
358 }
359
360 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
361 {
362         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
363
364         tasklet_hi_schedule(&shca->neq.interrupt_task);
365
366         return IRQ_HANDLED;
367 }
368
369 void ehca_tasklet_neq(unsigned long data)
370 {
371         struct ehca_shca *shca = (struct ehca_shca*)data;
372         struct ehca_eqe *eqe;
373         u64 ret;
374
375         eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
376
377         while (eqe) {
378                 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
379                         parse_ec(shca, eqe->entry);
380
381                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
382         }
383
384         ret = hipz_h_reset_event(shca->ipz_hca_handle,
385                                  shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
386
387         if (ret != H_SUCCESS)
388                 ehca_err(&shca->ib_device, "Can't clear notification events.");
389
390         return;
391 }
392
393 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
394 {
395         struct ehca_shca *shca = (struct ehca_shca*)dev_id;
396
397         tasklet_hi_schedule(&shca->eq.interrupt_task);
398
399         return IRQ_HANDLED;
400 }
401
402
403 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
404 {
405         u64 eqe_value;
406         u32 token;
407         unsigned long flags;
408         struct ehca_cq *cq;
409
410         eqe_value = eqe->entry;
411         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
412         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
413                 ehca_dbg(&shca->ib_device, "Got completion event");
414                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
415                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
416                 cq = idr_find(&ehca_cq_idr, token);
417                 if (cq == NULL) {
418                         spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
419                         ehca_err(&shca->ib_device,
420                                  "Invalid eqe for non-existing cq token=%x",
421                                  token);
422                         return;
423                 }
424                 reset_eq_pending(cq);
425                 cq->nr_events++;
426                 spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
427                 if (ehca_scaling_code)
428                         queue_comp_task(cq);
429                 else {
430                         comp_event_callback(cq);
431                         spin_lock_irqsave(&ehca_cq_idr_lock, flags);
432                         cq->nr_events--;
433                         if (!cq->nr_events)
434                                 wake_up(&cq->wait_completion);
435                         spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
436                 }
437         } else {
438                 ehca_dbg(&shca->ib_device, "Got non completion event");
439                 parse_identifier(shca, eqe_value);
440         }
441 }
442
443 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
444 {
445         struct ehca_eq *eq = &shca->eq;
446         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
447         u64 eqe_value;
448         unsigned long flags;
449         int eqe_cnt, i;
450         int eq_empty = 0;
451
452         spin_lock_irqsave(&eq->irq_spinlock, flags);
453         if (is_irq) {
454                 const int max_query_cnt = 100;
455                 int query_cnt = 0;
456                 int int_state = 1;
457                 do {
458                         int_state = hipz_h_query_int_state(
459                                 shca->ipz_hca_handle, eq->ist);
460                         query_cnt++;
461                         iosync();
462                 } while (int_state && query_cnt < max_query_cnt);
463                 if (unlikely((query_cnt == max_query_cnt)))
464                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
465                                  int_state, query_cnt);
466         }
467
468         /* read out all eqes */
469         eqe_cnt = 0;
470         do {
471                 u32 token;
472                 eqe_cache[eqe_cnt].eqe =
473                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
474                 if (!eqe_cache[eqe_cnt].eqe)
475                         break;
476                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
477                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
478                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
479                         spin_lock(&ehca_cq_idr_lock);
480                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
481                         if (!eqe_cache[eqe_cnt].cq) {
482                                 spin_unlock(&ehca_cq_idr_lock);
483                                 ehca_err(&shca->ib_device,
484                                          "Invalid eqe for non-existing cq "
485                                          "token=%x", token);
486                                 continue;
487                         }
488                         eqe_cache[eqe_cnt].cq->nr_events++;
489                         spin_unlock(&ehca_cq_idr_lock);
490                 } else
491                         eqe_cache[eqe_cnt].cq = NULL;
492                 eqe_cnt++;
493         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
494         if (!eqe_cnt) {
495                 if (is_irq)
496                         ehca_dbg(&shca->ib_device,
497                                  "No eqe found for irq event");
498                 goto unlock_irq_spinlock;
499         } else if (!is_irq)
500                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
501         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
502                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
503         /* enable irq for new packets */
504         for (i = 0; i < eqe_cnt; i++) {
505                 if (eq->eqe_cache[i].cq)
506                         reset_eq_pending(eq->eqe_cache[i].cq);
507         }
508         /* check eq */
509         spin_lock(&eq->spinlock);
510         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
511         spin_unlock(&eq->spinlock);
512         /* call completion handler for cached eqes */
513         for (i = 0; i < eqe_cnt; i++)
514                 if (eq->eqe_cache[i].cq) {
515                         if (ehca_scaling_code)
516                                 queue_comp_task(eq->eqe_cache[i].cq);
517                         else {
518                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
519                                 comp_event_callback(cq);
520                                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
521                                 cq->nr_events--;
522                                 if (!cq->nr_events)
523                                         wake_up(&cq->wait_completion);
524                                 spin_unlock_irqrestore(&ehca_cq_idr_lock,
525                                                        flags);
526                         }
527                 } else {
528                         ehca_dbg(&shca->ib_device, "Got non completion event");
529                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
530                 }
531         /* poll eq if not empty */
532         if (eq_empty)
533                 goto unlock_irq_spinlock;
534         do {
535                 struct ehca_eqe *eqe;
536                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
537                 if (!eqe)
538                         break;
539                 process_eqe(shca, eqe);
540         } while (1);
541
542 unlock_irq_spinlock:
543         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
544 }
545
546 void ehca_tasklet_eq(unsigned long data)
547 {
548         ehca_process_eq((struct ehca_shca*)data, 1);
549 }
550
551 static inline int find_next_online_cpu(struct ehca_comp_pool* pool)
552 {
553         int cpu;
554         unsigned long flags;
555
556         WARN_ON_ONCE(!in_interrupt());
557         if (ehca_debug_level)
558                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
559
560         spin_lock_irqsave(&pool->last_cpu_lock, flags);
561         cpu = next_cpu(pool->last_cpu, cpu_online_map);
562         if (cpu == NR_CPUS)
563                 cpu = first_cpu(cpu_online_map);
564         pool->last_cpu = cpu;
565         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
566
567         return cpu;
568 }
569
570 static void __queue_comp_task(struct ehca_cq *__cq,
571                               struct ehca_cpu_comp_task *cct)
572 {
573         unsigned long flags;
574
575         spin_lock_irqsave(&cct->task_lock, flags);
576         spin_lock(&__cq->task_lock);
577
578         if (__cq->nr_callbacks == 0) {
579                 __cq->nr_callbacks++;
580                 list_add_tail(&__cq->entry, &cct->cq_list);
581                 cct->cq_jobs++;
582                 wake_up(&cct->wait_queue);
583         } else
584                 __cq->nr_callbacks++;
585
586         spin_unlock(&__cq->task_lock);
587         spin_unlock_irqrestore(&cct->task_lock, flags);
588 }
589
590 static void queue_comp_task(struct ehca_cq *__cq)
591 {
592         int cpu_id;
593         struct ehca_cpu_comp_task *cct;
594         int cq_jobs;
595         unsigned long flags;
596
597         cpu_id = find_next_online_cpu(pool);
598         BUG_ON(!cpu_online(cpu_id));
599
600         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
601         BUG_ON(!cct);
602
603         spin_lock_irqsave(&cct->task_lock, flags);
604         cq_jobs = cct->cq_jobs;
605         spin_unlock_irqrestore(&cct->task_lock, flags);
606         if (cq_jobs > 0) {
607                 cpu_id = find_next_online_cpu(pool);
608                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
609                 BUG_ON(!cct);
610         }
611
612         __queue_comp_task(__cq, cct);
613 }
614
615 static void run_comp_task(struct ehca_cpu_comp_task* cct)
616 {
617         struct ehca_cq *cq;
618         unsigned long flags;
619
620         spin_lock_irqsave(&cct->task_lock, flags);
621
622         while (!list_empty(&cct->cq_list)) {
623                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
624                 spin_unlock_irqrestore(&cct->task_lock, flags);
625                 comp_event_callback(cq);
626
627                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
628                 cq->nr_events--;
629                 if (!cq->nr_events)
630                         wake_up(&cq->wait_completion);
631                 spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
632
633                 spin_lock_irqsave(&cct->task_lock, flags);
634                 spin_lock(&cq->task_lock);
635                 cq->nr_callbacks--;
636                 if (!cq->nr_callbacks) {
637                         list_del_init(cct->cq_list.next);
638                         cct->cq_jobs--;
639                 }
640                 spin_unlock(&cq->task_lock);
641         }
642
643         spin_unlock_irqrestore(&cct->task_lock, flags);
644 }
645
646 static int comp_task(void *__cct)
647 {
648         struct ehca_cpu_comp_task* cct = __cct;
649         int cql_empty;
650         DECLARE_WAITQUEUE(wait, current);
651
652         set_current_state(TASK_INTERRUPTIBLE);
653         while(!kthread_should_stop()) {
654                 add_wait_queue(&cct->wait_queue, &wait);
655
656                 spin_lock_irq(&cct->task_lock);
657                 cql_empty = list_empty(&cct->cq_list);
658                 spin_unlock_irq(&cct->task_lock);
659                 if (cql_empty)
660                         schedule();
661                 else
662                         __set_current_state(TASK_RUNNING);
663
664                 remove_wait_queue(&cct->wait_queue, &wait);
665
666                 spin_lock_irq(&cct->task_lock);
667                 cql_empty = list_empty(&cct->cq_list);
668                 spin_unlock_irq(&cct->task_lock);
669                 if (!cql_empty)
670                         run_comp_task(__cct);
671
672                 set_current_state(TASK_INTERRUPTIBLE);
673         }
674         __set_current_state(TASK_RUNNING);
675
676         return 0;
677 }
678
679 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
680                                             int cpu)
681 {
682         struct ehca_cpu_comp_task *cct;
683
684         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
685         spin_lock_init(&cct->task_lock);
686         INIT_LIST_HEAD(&cct->cq_list);
687         init_waitqueue_head(&cct->wait_queue);
688         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
689
690         return cct->task;
691 }
692
693 static void destroy_comp_task(struct ehca_comp_pool *pool,
694                               int cpu)
695 {
696         struct ehca_cpu_comp_task *cct;
697         struct task_struct *task;
698         unsigned long flags_cct;
699
700         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
701
702         spin_lock_irqsave(&cct->task_lock, flags_cct);
703
704         task = cct->task;
705         cct->task = NULL;
706         cct->cq_jobs = 0;
707
708         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
709
710         if (task)
711                 kthread_stop(task);
712 }
713
714 static void take_over_work(struct ehca_comp_pool *pool,
715                            int cpu)
716 {
717         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
718         LIST_HEAD(list);
719         struct ehca_cq *cq;
720         unsigned long flags_cct;
721
722         spin_lock_irqsave(&cct->task_lock, flags_cct);
723
724         list_splice_init(&cct->cq_list, &list);
725
726         while(!list_empty(&list)) {
727                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
728
729                 list_del(&cq->entry);
730                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
731                                                   smp_processor_id()));
732         }
733
734         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
735
736 }
737
738 #ifdef CONFIG_HOTPLUG_CPU
739 static int comp_pool_callback(struct notifier_block *nfb,
740                               unsigned long action,
741                               void *hcpu)
742 {
743         unsigned int cpu = (unsigned long)hcpu;
744         struct ehca_cpu_comp_task *cct;
745
746         switch (action) {
747         case CPU_UP_PREPARE:
748                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
749                 if(!create_comp_task(pool, cpu)) {
750                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
751                         return NOTIFY_BAD;
752                 }
753                 break;
754         case CPU_UP_CANCELED:
755                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
756                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
757                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
758                 destroy_comp_task(pool, cpu);
759                 break;
760         case CPU_ONLINE:
761                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
762                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
763                 kthread_bind(cct->task, cpu);
764                 wake_up_process(cct->task);
765                 break;
766         case CPU_DOWN_PREPARE:
767                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
768                 break;
769         case CPU_DOWN_FAILED:
770                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
771                 break;
772         case CPU_DEAD:
773                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
774                 destroy_comp_task(pool, cpu);
775                 take_over_work(pool, cpu);
776                 break;
777         }
778
779         return NOTIFY_OK;
780 }
781 #endif
782
783 int ehca_create_comp_pool(void)
784 {
785         int cpu;
786         struct task_struct *task;
787
788         if (!ehca_scaling_code)
789                 return 0;
790
791         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
792         if (pool == NULL)
793                 return -ENOMEM;
794
795         spin_lock_init(&pool->last_cpu_lock);
796         pool->last_cpu = any_online_cpu(cpu_online_map);
797
798         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
799         if (pool->cpu_comp_tasks == NULL) {
800                 kfree(pool);
801                 return -EINVAL;
802         }
803
804         for_each_online_cpu(cpu) {
805                 task = create_comp_task(pool, cpu);
806                 if (task) {
807                         kthread_bind(task, cpu);
808                         wake_up_process(task);
809                 }
810         }
811
812 #ifdef CONFIG_HOTPLUG_CPU
813         comp_pool_callback_nb.notifier_call = comp_pool_callback;
814         comp_pool_callback_nb.priority =0;
815         register_cpu_notifier(&comp_pool_callback_nb);
816 #endif
817
818         printk(KERN_INFO "eHCA scaling code enabled\n");
819
820         return 0;
821 }
822
823 void ehca_destroy_comp_pool(void)
824 {
825         int i;
826
827         if (!ehca_scaling_code)
828                 return;
829
830 #ifdef CONFIG_HOTPLUG_CPU
831         unregister_cpu_notifier(&comp_pool_callback_nb);
832 #endif
833
834         for (i = 0; i < NR_CPUS; i++) {
835                 if (cpu_online(i))
836                         destroy_comp_task(pool, i);
837         }
838         free_percpu(pool->cpu_comp_tasks);
839         kfree(pool);
840 }