]> pilppa.org Git - linux-2.6-omap-h63xx.git/blob - drivers/infiniband/hw/ehca/ehca_irq.c
Pull misc-for-upstream into release branch
[linux-2.6-omap-h63xx.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
408         eqe_value = eqe->entry;
409         ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
410         if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
411                 ehca_dbg(&shca->ib_device, "Got completion event");
412                 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
413                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
414                 cq = idr_find(&ehca_cq_idr, token);
415                 if (cq == NULL) {
416                         spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
417                         ehca_err(&shca->ib_device,
418                                  "Invalid eqe for non-existing cq token=%x",
419                                  token);
420                         return;
421                 }
422                 reset_eq_pending(cq);
423                 cq->nr_events++;
424                 spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
425                 if (ehca_scaling_code)
426                         queue_comp_task(cq);
427                 else {
428                         comp_event_callback(cq);
429                         spin_lock_irqsave(&ehca_cq_idr_lock, flags);
430                         cq->nr_events--;
431                         if (!cq->nr_events)
432                                 wake_up(&cq->wait_completion);
433                         spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
434                 }
435         } else {
436                 ehca_dbg(&shca->ib_device, "Got non completion event");
437                 parse_identifier(shca, eqe_value);
438         }
439 }
440
441 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
442 {
443         struct ehca_eq *eq = &shca->eq;
444         struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
445         u64 eqe_value;
446         unsigned long flags;
447         int eqe_cnt, i;
448         int eq_empty = 0;
449
450         spin_lock_irqsave(&eq->irq_spinlock, flags);
451         if (is_irq) {
452                 const int max_query_cnt = 100;
453                 int query_cnt = 0;
454                 int int_state = 1;
455                 do {
456                         int_state = hipz_h_query_int_state(
457                                 shca->ipz_hca_handle, eq->ist);
458                         query_cnt++;
459                         iosync();
460                 } while (int_state && query_cnt < max_query_cnt);
461                 if (unlikely((query_cnt == max_query_cnt)))
462                         ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
463                                  int_state, query_cnt);
464         }
465
466         /* read out all eqes */
467         eqe_cnt = 0;
468         do {
469                 u32 token;
470                 eqe_cache[eqe_cnt].eqe =
471                         (struct ehca_eqe *)ehca_poll_eq(shca, eq);
472                 if (!eqe_cache[eqe_cnt].eqe)
473                         break;
474                 eqe_value = eqe_cache[eqe_cnt].eqe->entry;
475                 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
476                         token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
477                         spin_lock(&ehca_cq_idr_lock);
478                         eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
479                         if (!eqe_cache[eqe_cnt].cq) {
480                                 spin_unlock(&ehca_cq_idr_lock);
481                                 ehca_err(&shca->ib_device,
482                                          "Invalid eqe for non-existing cq "
483                                          "token=%x", token);
484                                 continue;
485                         }
486                         eqe_cache[eqe_cnt].cq->nr_events++;
487                         spin_unlock(&ehca_cq_idr_lock);
488                 } else
489                         eqe_cache[eqe_cnt].cq = NULL;
490                 eqe_cnt++;
491         } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
492         if (!eqe_cnt) {
493                 if (is_irq)
494                         ehca_dbg(&shca->ib_device,
495                                  "No eqe found for irq event");
496                 goto unlock_irq_spinlock;
497         } else if (!is_irq)
498                 ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
499         if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
500                 ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
501         /* enable irq for new packets */
502         for (i = 0; i < eqe_cnt; i++) {
503                 if (eq->eqe_cache[i].cq)
504                         reset_eq_pending(eq->eqe_cache[i].cq);
505         }
506         /* check eq */
507         spin_lock(&eq->spinlock);
508         eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
509         spin_unlock(&eq->spinlock);
510         /* call completion handler for cached eqes */
511         for (i = 0; i < eqe_cnt; i++)
512                 if (eq->eqe_cache[i].cq) {
513                         if (ehca_scaling_code)
514                                 queue_comp_task(eq->eqe_cache[i].cq);
515                         else {
516                                 struct ehca_cq *cq = eq->eqe_cache[i].cq;
517                                 comp_event_callback(cq);
518                                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
519                                 cq->nr_events--;
520                                 if (!cq->nr_events)
521                                         wake_up(&cq->wait_completion);
522                                 spin_unlock_irqrestore(&ehca_cq_idr_lock,
523                                                        flags);
524                         }
525                 } else {
526                         ehca_dbg(&shca->ib_device, "Got non completion event");
527                         parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
528                 }
529         /* poll eq if not empty */
530         if (eq_empty)
531                 goto unlock_irq_spinlock;
532         do {
533                 struct ehca_eqe *eqe;
534                 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
535                 if (!eqe)
536                         break;
537                 process_eqe(shca, eqe);
538         } while (1);
539
540 unlock_irq_spinlock:
541         spin_unlock_irqrestore(&eq->irq_spinlock, flags);
542 }
543
544 void ehca_tasklet_eq(unsigned long data)
545 {
546         ehca_process_eq((struct ehca_shca*)data, 1);
547 }
548
549 static inline int find_next_online_cpu(struct ehca_comp_pool* pool)
550 {
551         int cpu;
552         unsigned long flags;
553
554         WARN_ON_ONCE(!in_interrupt());
555         if (ehca_debug_level)
556                 ehca_dmp(&cpu_online_map, sizeof(cpumask_t), "");
557
558         spin_lock_irqsave(&pool->last_cpu_lock, flags);
559         cpu = next_cpu(pool->last_cpu, cpu_online_map);
560         if (cpu == NR_CPUS)
561                 cpu = first_cpu(cpu_online_map);
562         pool->last_cpu = cpu;
563         spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
564
565         return cpu;
566 }
567
568 static void __queue_comp_task(struct ehca_cq *__cq,
569                               struct ehca_cpu_comp_task *cct)
570 {
571         unsigned long flags;
572
573         spin_lock_irqsave(&cct->task_lock, flags);
574         spin_lock(&__cq->task_lock);
575
576         if (__cq->nr_callbacks == 0) {
577                 __cq->nr_callbacks++;
578                 list_add_tail(&__cq->entry, &cct->cq_list);
579                 cct->cq_jobs++;
580                 wake_up(&cct->wait_queue);
581         } else
582                 __cq->nr_callbacks++;
583
584         spin_unlock(&__cq->task_lock);
585         spin_unlock_irqrestore(&cct->task_lock, flags);
586 }
587
588 static void queue_comp_task(struct ehca_cq *__cq)
589 {
590         int cpu_id;
591         struct ehca_cpu_comp_task *cct;
592         int cq_jobs;
593         unsigned long flags;
594
595         cpu_id = find_next_online_cpu(pool);
596         BUG_ON(!cpu_online(cpu_id));
597
598         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
599         BUG_ON(!cct);
600
601         spin_lock_irqsave(&cct->task_lock, flags);
602         cq_jobs = cct->cq_jobs;
603         spin_unlock_irqrestore(&cct->task_lock, flags);
604         if (cq_jobs > 0) {
605                 cpu_id = find_next_online_cpu(pool);
606                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
607                 BUG_ON(!cct);
608         }
609
610         __queue_comp_task(__cq, cct);
611 }
612
613 static void run_comp_task(struct ehca_cpu_comp_task* cct)
614 {
615         struct ehca_cq *cq;
616         unsigned long flags;
617
618         spin_lock_irqsave(&cct->task_lock, flags);
619
620         while (!list_empty(&cct->cq_list)) {
621                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
622                 spin_unlock_irqrestore(&cct->task_lock, flags);
623                 comp_event_callback(cq);
624
625                 spin_lock_irqsave(&ehca_cq_idr_lock, flags);
626                 cq->nr_events--;
627                 if (!cq->nr_events)
628                         wake_up(&cq->wait_completion);
629                 spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
630
631                 spin_lock_irqsave(&cct->task_lock, flags);
632                 spin_lock(&cq->task_lock);
633                 cq->nr_callbacks--;
634                 if (!cq->nr_callbacks) {
635                         list_del_init(cct->cq_list.next);
636                         cct->cq_jobs--;
637                 }
638                 spin_unlock(&cq->task_lock);
639         }
640
641         spin_unlock_irqrestore(&cct->task_lock, flags);
642 }
643
644 static int comp_task(void *__cct)
645 {
646         struct ehca_cpu_comp_task* cct = __cct;
647         int cql_empty;
648         DECLARE_WAITQUEUE(wait, current);
649
650         set_current_state(TASK_INTERRUPTIBLE);
651         while(!kthread_should_stop()) {
652                 add_wait_queue(&cct->wait_queue, &wait);
653
654                 spin_lock_irq(&cct->task_lock);
655                 cql_empty = list_empty(&cct->cq_list);
656                 spin_unlock_irq(&cct->task_lock);
657                 if (cql_empty)
658                         schedule();
659                 else
660                         __set_current_state(TASK_RUNNING);
661
662                 remove_wait_queue(&cct->wait_queue, &wait);
663
664                 spin_lock_irq(&cct->task_lock);
665                 cql_empty = list_empty(&cct->cq_list);
666                 spin_unlock_irq(&cct->task_lock);
667                 if (!cql_empty)
668                         run_comp_task(__cct);
669
670                 set_current_state(TASK_INTERRUPTIBLE);
671         }
672         __set_current_state(TASK_RUNNING);
673
674         return 0;
675 }
676
677 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
678                                             int cpu)
679 {
680         struct ehca_cpu_comp_task *cct;
681
682         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
683         spin_lock_init(&cct->task_lock);
684         INIT_LIST_HEAD(&cct->cq_list);
685         init_waitqueue_head(&cct->wait_queue);
686         cct->task = kthread_create(comp_task, cct, "ehca_comp/%d", cpu);
687
688         return cct->task;
689 }
690
691 static void destroy_comp_task(struct ehca_comp_pool *pool,
692                               int cpu)
693 {
694         struct ehca_cpu_comp_task *cct;
695         struct task_struct *task;
696         unsigned long flags_cct;
697
698         cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
699
700         spin_lock_irqsave(&cct->task_lock, flags_cct);
701
702         task = cct->task;
703         cct->task = NULL;
704         cct->cq_jobs = 0;
705
706         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
707
708         if (task)
709                 kthread_stop(task);
710 }
711
712 static void take_over_work(struct ehca_comp_pool *pool,
713                            int cpu)
714 {
715         struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
716         LIST_HEAD(list);
717         struct ehca_cq *cq;
718         unsigned long flags_cct;
719
720         spin_lock_irqsave(&cct->task_lock, flags_cct);
721
722         list_splice_init(&cct->cq_list, &list);
723
724         while(!list_empty(&list)) {
725                 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
726
727                 list_del(&cq->entry);
728                 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
729                                                   smp_processor_id()));
730         }
731
732         spin_unlock_irqrestore(&cct->task_lock, flags_cct);
733
734 }
735
736 static int comp_pool_callback(struct notifier_block *nfb,
737                               unsigned long action,
738                               void *hcpu)
739 {
740         unsigned int cpu = (unsigned long)hcpu;
741         struct ehca_cpu_comp_task *cct;
742
743         switch (action) {
744         case CPU_UP_PREPARE:
745                 ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
746                 if(!create_comp_task(pool, cpu)) {
747                         ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
748                         return NOTIFY_BAD;
749                 }
750                 break;
751         case CPU_UP_CANCELED:
752                 ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
753                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
754                 kthread_bind(cct->task, any_online_cpu(cpu_online_map));
755                 destroy_comp_task(pool, cpu);
756                 break;
757         case CPU_ONLINE:
758                 ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
759                 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
760                 kthread_bind(cct->task, cpu);
761                 wake_up_process(cct->task);
762                 break;
763         case CPU_DOWN_PREPARE:
764                 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
765                 break;
766         case CPU_DOWN_FAILED:
767                 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
768                 break;
769         case CPU_DEAD:
770                 ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
771                 destroy_comp_task(pool, cpu);
772                 take_over_work(pool, cpu);
773                 break;
774         }
775
776         return NOTIFY_OK;
777 }
778
779 int ehca_create_comp_pool(void)
780 {
781         int cpu;
782         struct task_struct *task;
783
784         if (!ehca_scaling_code)
785                 return 0;
786
787         pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
788         if (pool == NULL)
789                 return -ENOMEM;
790
791         spin_lock_init(&pool->last_cpu_lock);
792         pool->last_cpu = any_online_cpu(cpu_online_map);
793
794         pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
795         if (pool->cpu_comp_tasks == NULL) {
796                 kfree(pool);
797                 return -EINVAL;
798         }
799
800         for_each_online_cpu(cpu) {
801                 task = create_comp_task(pool, cpu);
802                 if (task) {
803                         kthread_bind(task, cpu);
804                         wake_up_process(task);
805                 }
806         }
807
808         comp_pool_callback_nb.notifier_call = comp_pool_callback;
809         comp_pool_callback_nb.priority =0;
810         register_cpu_notifier(&comp_pool_callback_nb);
811
812         printk(KERN_INFO "eHCA scaling code enabled\n");
813
814         return 0;
815 }
816
817 void ehca_destroy_comp_pool(void)
818 {
819         int i;
820
821         if (!ehca_scaling_code)
822                 return;
823
824         unregister_cpu_notifier(&comp_pool_callback_nb);
825
826         for (i = 0; i < NR_CPUS; i++) {
827                 if (cpu_online(i))
828                         destroy_comp_task(pool, i);
829         }
830         free_percpu(pool->cpu_comp_tasks);
831         kfree(pool);
832 }