Merge /pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / arch / powerpc / oprofile / op_model_cell.c
1 /*
2  * Cell Broadband Engine OProfile Support
3  *
4  * (C) Copyright IBM Corporation 2006
5  *
6  * Author: David Erb (djerb@us.ibm.com)
7  * Modifications:
8  *         Carl Love <carll@us.ibm.com>
9  *         Maynard Johnson <maynardj@us.ibm.com>
10  *
11  * This program is free software; you can redistribute it and/or
12  * modify it under the terms of the GNU General Public License
13  * as published by the Free Software Foundation; either version
14  * 2 of the License, or (at your option) any later version.
15  */
16
17 #include <linux/cpufreq.h>
18 #include <linux/delay.h>
19 #include <linux/init.h>
20 #include <linux/jiffies.h>
21 #include <linux/kthread.h>
22 #include <linux/oprofile.h>
23 #include <linux/percpu.h>
24 #include <linux/smp.h>
25 #include <linux/spinlock.h>
26 #include <linux/timer.h>
27 #include <asm/cell-pmu.h>
28 #include <asm/cputable.h>
29 #include <asm/firmware.h>
30 #include <asm/io.h>
31 #include <asm/oprofile_impl.h>
32 #include <asm/processor.h>
33 #include <asm/prom.h>
34 #include <asm/ptrace.h>
35 #include <asm/reg.h>
36 #include <asm/rtas.h>
37 #include <asm/system.h>
38
39 #include "../platforms/cell/interrupt.h"
40 #include "../platforms/cell/cbe_regs.h"
41
42 #define PPU_CYCLES_EVENT_NUM 1  /*  event number for CYCLES */
43 #define PPU_CYCLES_GRP_NUM   1  /* special group number for identifying
44                                  * PPU_CYCLES event
45                                  */
46 #define CBE_COUNT_ALL_CYCLES 0x42800000 /* PPU cycle event specifier */
47
48 #define NUM_THREADS 2         /* number of physical threads in
49                                * physical processor
50                                */
51 #define NUM_TRACE_BUS_WORDS 4
52 #define NUM_INPUT_BUS_WORDS 2
53
54
55 struct pmc_cntrl_data {
56         unsigned long vcntr;
57         unsigned long evnts;
58         unsigned long masks;
59         unsigned long enabled;
60 };
61
62 /*
63  * ibm,cbe-perftools rtas parameters
64  */
65
66 struct pm_signal {
67         u16 cpu;                /* Processor to modify */
68         u16 sub_unit;           /* hw subunit this applies to (if applicable) */
69         short int signal_group; /* Signal Group to Enable/Disable */
70         u8 bus_word;            /* Enable/Disable on this Trace/Trigger/Event
71                                  * Bus Word(s) (bitmask)
72                                  */
73         u8 bit;                 /* Trigger/Event bit (if applicable) */
74 };
75
76 /*
77  * rtas call arguments
78  */
79 enum {
80         SUBFUNC_RESET = 1,
81         SUBFUNC_ACTIVATE = 2,
82         SUBFUNC_DEACTIVATE = 3,
83
84         PASSTHRU_IGNORE = 0,
85         PASSTHRU_ENABLE = 1,
86         PASSTHRU_DISABLE = 2,
87 };
88
89 struct pm_cntrl {
90         u16 enable;
91         u16 stop_at_max;
92         u16 trace_mode;
93         u16 freeze;
94         u16 count_mode;
95 };
96
97 static struct {
98         u32 group_control;
99         u32 debug_bus_control;
100         struct pm_cntrl pm_cntrl;
101         u32 pm07_cntrl[NR_PHYS_CTRS];
102 } pm_regs;
103
104 #define GET_SUB_UNIT(x) ((x & 0x0000f000) >> 12)
105 #define GET_BUS_WORD(x) ((x & 0x000000f0) >> 4)
106 #define GET_BUS_TYPE(x) ((x & 0x00000300) >> 8)
107 #define GET_POLARITY(x) ((x & 0x00000002) >> 1)
108 #define GET_COUNT_CYCLES(x) (x & 0x00000001)
109 #define GET_INPUT_CONTROL(x) ((x & 0x00000004) >> 2)
110
111 static DEFINE_PER_CPU(unsigned long[NR_PHYS_CTRS], pmc_values);
112
113 static struct pmc_cntrl_data pmc_cntrl[NUM_THREADS][NR_PHYS_CTRS];
114
115 /* Interpetation of hdw_thread:
116  * 0 - even virtual cpus 0, 2, 4,...
117  * 1 - odd virtual cpus 1, 3, 5, ...
118  */
119 static u32 hdw_thread;
120
121 static u32 virt_cntr_inter_mask;
122 static struct timer_list timer_virt_cntr;
123
124 /* pm_signal needs to be global since it is initialized in
125  * cell_reg_setup at the time when the necessary information
126  * is available.
127  */
128 static struct pm_signal pm_signal[NR_PHYS_CTRS];
129 static int pm_rtas_token;
130
131 static u32 reset_value[NR_PHYS_CTRS];
132 static int num_counters;
133 static int oprofile_running;
134 static DEFINE_SPINLOCK(virt_cntr_lock);
135
136 static u32 ctr_enabled;
137
138 static unsigned char trace_bus[NUM_TRACE_BUS_WORDS];
139 static unsigned char input_bus[NUM_INPUT_BUS_WORDS];
140
141 /*
142  * Firmware interface functions
143  */
144 static int
145 rtas_ibm_cbe_perftools(int subfunc, int passthru,
146                        void *address, unsigned long length)
147 {
148         u64 paddr = __pa(address);
149
150         return rtas_call(pm_rtas_token, 5, 1, NULL, subfunc, passthru,
151                          paddr >> 32, paddr & 0xffffffff, length);
152 }
153
154 static void pm_rtas_reset_signals(u32 node)
155 {
156         int ret;
157         struct pm_signal pm_signal_local;
158
159         /*  The debug bus is being set to the passthru disable state.
160          *  However, the FW still expects atleast one legal signal routing
161          *  entry or it will return an error on the arguments.  If we don't
162          *  supply a valid entry, we must ignore all return values.  Ignoring
163          *  all return values means we might miss an error we should be
164          *  concerned about.
165          */
166
167         /*  fw expects physical cpu #. */
168         pm_signal_local.cpu = node;
169         pm_signal_local.signal_group = 21;
170         pm_signal_local.bus_word = 1;
171         pm_signal_local.sub_unit = 0;
172         pm_signal_local.bit = 0;
173
174         ret = rtas_ibm_cbe_perftools(SUBFUNC_RESET, PASSTHRU_DISABLE,
175                                      &pm_signal_local,
176                                      sizeof(struct pm_signal));
177
178         if (ret)
179                 printk(KERN_WARNING "%s: rtas returned: %d\n",
180                        __FUNCTION__, ret);
181 }
182
183 static void pm_rtas_activate_signals(u32 node, u32 count)
184 {
185         int ret;
186         int i, j;
187         struct pm_signal pm_signal_local[NR_PHYS_CTRS];
188
189         /* There is no debug setup required for the cycles event.
190          * Note that only events in the same group can be used.
191          * Otherwise, there will be conflicts in correctly routing
192          * the signals on the debug bus.  It is the responsiblity
193          * of the OProfile user tool to check the events are in
194          * the same group.
195          */
196         i = 0;
197         for (j = 0; j < count; j++) {
198                 if (pm_signal[j].signal_group != PPU_CYCLES_GRP_NUM) {
199
200                         /* fw expects physical cpu # */
201                         pm_signal_local[i].cpu = node;
202                         pm_signal_local[i].signal_group
203                                 = pm_signal[j].signal_group;
204                         pm_signal_local[i].bus_word = pm_signal[j].bus_word;
205                         pm_signal_local[i].sub_unit = pm_signal[j].sub_unit;
206                         pm_signal_local[i].bit = pm_signal[j].bit;
207                         i++;
208                 }
209         }
210
211         if (i != 0) {
212                 ret = rtas_ibm_cbe_perftools(SUBFUNC_ACTIVATE, PASSTHRU_ENABLE,
213                                              pm_signal_local,
214                                              i * sizeof(struct pm_signal));
215
216                 if (ret)
217                         printk(KERN_WARNING "%s: rtas returned: %d\n",
218                                __FUNCTION__, ret);
219         }
220 }
221
222 /*
223  * PM Signal functions
224  */
225 static void set_pm_event(u32 ctr, int event, u32 unit_mask)
226 {
227         struct pm_signal *p;
228         u32 signal_bit;
229         u32 bus_word, bus_type, count_cycles, polarity, input_control;
230         int j, i;
231
232         if (event == PPU_CYCLES_EVENT_NUM) {
233                 /* Special Event: Count all cpu cycles */
234                 pm_regs.pm07_cntrl[ctr] = CBE_COUNT_ALL_CYCLES;
235                 p = &(pm_signal[ctr]);
236                 p->signal_group = PPU_CYCLES_GRP_NUM;
237                 p->bus_word = 1;
238                 p->sub_unit = 0;
239                 p->bit = 0;
240                 goto out;
241         } else {
242                 pm_regs.pm07_cntrl[ctr] = 0;
243         }
244
245         bus_word = GET_BUS_WORD(unit_mask);
246         bus_type = GET_BUS_TYPE(unit_mask);
247         count_cycles = GET_COUNT_CYCLES(unit_mask);
248         polarity = GET_POLARITY(unit_mask);
249         input_control = GET_INPUT_CONTROL(unit_mask);
250         signal_bit = (event % 100);
251
252         p = &(pm_signal[ctr]);
253
254         p->signal_group = event / 100;
255         p->bus_word = bus_word;
256         p->sub_unit = (unit_mask & 0x0000f000) >> 12;
257
258         pm_regs.pm07_cntrl[ctr] = 0;
259         pm_regs.pm07_cntrl[ctr] |= PM07_CTR_COUNT_CYCLES(count_cycles);
260         pm_regs.pm07_cntrl[ctr] |= PM07_CTR_POLARITY(polarity);
261         pm_regs.pm07_cntrl[ctr] |= PM07_CTR_INPUT_CONTROL(input_control);
262
263         /* Some of the islands signal selection is based on 64 bit words.
264          * The debug bus words are 32 bits, the input words to the performance
265          * counters are defined as 32 bits.  Need to convert the 64 bit island
266          * specification to the appropriate 32 input bit and bus word for the
267          * performance counter event selection.  See the CELL Performance
268          * monitoring signals manual and the Perf cntr hardware descriptions
269          * for the details.
270          */
271         if (input_control == 0) {
272                 if (signal_bit > 31) {
273                         signal_bit -= 32;
274                         if (bus_word == 0x3)
275                                 bus_word = 0x2;
276                         else if (bus_word == 0xc)
277                                 bus_word = 0x8;
278                 }
279
280                 if ((bus_type == 0) && p->signal_group >= 60)
281                         bus_type = 2;
282                 if ((bus_type == 1) && p->signal_group >= 50)
283                         bus_type = 0;
284
285                 pm_regs.pm07_cntrl[ctr] |= PM07_CTR_INPUT_MUX(signal_bit);
286         } else {
287                 pm_regs.pm07_cntrl[ctr] = 0;
288                 p->bit = signal_bit;
289         }
290
291         for (i = 0; i < NUM_TRACE_BUS_WORDS; i++) {
292                 if (bus_word & (1 << i)) {
293                         pm_regs.debug_bus_control |=
294                             (bus_type << (31 - (2 * i) + 1));
295
296                         for (j = 0; j < NUM_INPUT_BUS_WORDS; j++) {
297                                 if (input_bus[j] == 0xff) {
298                                         input_bus[j] = i;
299                                         pm_regs.group_control |=
300                                             (i << (31 - i));
301                                         break;
302                                 }
303                         }
304                 }
305         }
306 out:
307         ;
308 }
309
310 static void write_pm_cntrl(int cpu)
311 {
312         /* Oprofile will use 32 bit counters, set bits 7:10 to 0
313          * pmregs.pm_cntrl is a global
314          */
315
316         u32 val = 0;
317         if (pm_regs.pm_cntrl.enable == 1)
318                 val |= CBE_PM_ENABLE_PERF_MON;
319
320         if (pm_regs.pm_cntrl.stop_at_max == 1)
321                 val |= CBE_PM_STOP_AT_MAX;
322
323         if (pm_regs.pm_cntrl.trace_mode == 1)
324                 val |= CBE_PM_TRACE_MODE_SET(pm_regs.pm_cntrl.trace_mode);
325
326         if (pm_regs.pm_cntrl.freeze == 1)
327                 val |= CBE_PM_FREEZE_ALL_CTRS;
328
329         /* Routine set_count_mode must be called previously to set
330          * the count mode based on the user selection of user and kernel.
331          */
332         val |= CBE_PM_COUNT_MODE_SET(pm_regs.pm_cntrl.count_mode);
333         cbe_write_pm(cpu, pm_control, val);
334 }
335
336 static inline void
337 set_count_mode(u32 kernel, u32 user)
338 {
339         /* The user must specify user and kernel if they want them. If
340          *  neither is specified, OProfile will count in hypervisor mode.
341          *  pm_regs.pm_cntrl is a global
342          */
343         if (kernel) {
344                 if (user)
345                         pm_regs.pm_cntrl.count_mode = CBE_COUNT_ALL_MODES;
346                 else
347                         pm_regs.pm_cntrl.count_mode =
348                                 CBE_COUNT_SUPERVISOR_MODE;
349         } else {
350                 if (user)
351                         pm_regs.pm_cntrl.count_mode = CBE_COUNT_PROBLEM_MODE;
352                 else
353                         pm_regs.pm_cntrl.count_mode =
354                                 CBE_COUNT_HYPERVISOR_MODE;
355         }
356 }
357
358 static inline void enable_ctr(u32 cpu, u32 ctr, u32 * pm07_cntrl)
359 {
360
361         pm07_cntrl[ctr] |= CBE_PM_CTR_ENABLE;
362         cbe_write_pm07_control(cpu, ctr, pm07_cntrl[ctr]);
363 }
364
365 /*
366  * Oprofile is expected to collect data on all CPUs simultaneously.
367  * However, there is one set of performance counters per node.  There are
368  * two hardware threads or virtual CPUs on each node.  Hence, OProfile must
369  * multiplex in time the performance counter collection on the two virtual
370  * CPUs.  The multiplexing of the performance counters is done by this
371  * virtual counter routine.
372  *
373  * The pmc_values used below is defined as 'per-cpu' but its use is
374  * more akin to 'per-node'.  We need to store two sets of counter
375  * values per node -- one for the previous run and one for the next.
376  * The per-cpu[NR_PHYS_CTRS] gives us the storage we need.  Each odd/even
377  * pair of per-cpu arrays is used for storing the previous and next
378  * pmc values for a given node.
379  * NOTE: We use the per-cpu variable to improve cache performance.
380  */
381 static void cell_virtual_cntr(unsigned long data)
382 {
383         /* This routine will alternate loading the virtual counters for
384          * virtual CPUs
385          */
386         int i, prev_hdw_thread, next_hdw_thread;
387         u32 cpu;
388         unsigned long flags;
389
390         /* Make sure that the interrupt_hander and
391          * the virt counter are not both playing with
392          * the counters on the same node.
393          */
394
395         spin_lock_irqsave(&virt_cntr_lock, flags);
396
397         prev_hdw_thread = hdw_thread;
398
399         /* switch the cpu handling the interrupts */
400         hdw_thread = 1 ^ hdw_thread;
401         next_hdw_thread = hdw_thread;
402
403         for (i = 0; i < num_counters; i++)
404         /* There are some per thread events.  Must do the
405          * set event, for the thread that is being started
406          */
407                 set_pm_event(i,
408                         pmc_cntrl[next_hdw_thread][i].evnts,
409                         pmc_cntrl[next_hdw_thread][i].masks);
410
411         /* The following is done only once per each node, but
412          * we need cpu #, not node #, to pass to the cbe_xxx functions.
413          */
414         for_each_online_cpu(cpu) {
415                 if (cbe_get_hw_thread_id(cpu))
416                         continue;
417
418                 /* stop counters, save counter values, restore counts
419                  * for previous thread
420                  */
421                 cbe_disable_pm(cpu);
422                 cbe_disable_pm_interrupts(cpu);
423                 for (i = 0; i < num_counters; i++) {
424                         per_cpu(pmc_values, cpu + prev_hdw_thread)[i]
425                             = cbe_read_ctr(cpu, i);
426
427                         if (per_cpu(pmc_values, cpu + next_hdw_thread)[i]
428                             == 0xFFFFFFFF)
429                                 /* If the cntr value is 0xffffffff, we must
430                                  * reset that to 0xfffffff0 when the current
431                                  * thread is restarted.  This will generate a
432                                  * new interrupt and make sure that we never
433                                  * restore the counters to the max value.  If
434                                  * the counters were restored to the max value,
435                                  * they do not increment and no interrupts are
436                                  * generated.  Hence no more samples will be
437                                  * collected on that cpu.
438                                  */
439                                 cbe_write_ctr(cpu, i, 0xFFFFFFF0);
440                         else
441                                 cbe_write_ctr(cpu, i,
442                                               per_cpu(pmc_values,
443                                                       cpu +
444                                                       next_hdw_thread)[i]);
445                 }
446
447                 /* Switch to the other thread. Change the interrupt
448                  * and control regs to be scheduled on the CPU
449                  * corresponding to the thread to execute.
450                  */
451                 for (i = 0; i < num_counters; i++) {
452                         if (pmc_cntrl[next_hdw_thread][i].enabled) {
453                                 /* There are some per thread events.
454                                  * Must do the set event, enable_cntr
455                                  * for each cpu.
456                                  */
457                                 enable_ctr(cpu, i,
458                                            pm_regs.pm07_cntrl);
459                         } else {
460                                 cbe_write_pm07_control(cpu, i, 0);
461                         }
462                 }
463
464                 /* Enable interrupts on the CPU thread that is starting */
465                 cbe_enable_pm_interrupts(cpu, next_hdw_thread,
466                                          virt_cntr_inter_mask);
467                 cbe_enable_pm(cpu);
468         }
469
470         spin_unlock_irqrestore(&virt_cntr_lock, flags);
471
472         mod_timer(&timer_virt_cntr, jiffies + HZ / 10);
473 }
474
475 static void start_virt_cntrs(void)
476 {
477         init_timer(&timer_virt_cntr);
478         timer_virt_cntr.function = cell_virtual_cntr;
479         timer_virt_cntr.data = 0UL;
480         timer_virt_cntr.expires = jiffies + HZ / 10;
481         add_timer(&timer_virt_cntr);
482 }
483
484 /* This function is called once for all cpus combined */
485 static void
486 cell_reg_setup(struct op_counter_config *ctr,
487                struct op_system_config *sys, int num_ctrs)
488 {
489         int i, j, cpu;
490
491         pm_rtas_token = rtas_token("ibm,cbe-perftools");
492         if (pm_rtas_token == RTAS_UNKNOWN_SERVICE) {
493                 printk(KERN_WARNING "%s: RTAS_UNKNOWN_SERVICE\n",
494                        __FUNCTION__);
495                 goto out;
496         }
497
498         num_counters = num_ctrs;
499
500         pm_regs.group_control = 0;
501         pm_regs.debug_bus_control = 0;
502
503         /* setup the pm_control register */
504         memset(&pm_regs.pm_cntrl, 0, sizeof(struct pm_cntrl));
505         pm_regs.pm_cntrl.stop_at_max = 1;
506         pm_regs.pm_cntrl.trace_mode = 0;
507         pm_regs.pm_cntrl.freeze = 1;
508
509         set_count_mode(sys->enable_kernel, sys->enable_user);
510
511         /* Setup the thread 0 events */
512         for (i = 0; i < num_ctrs; ++i) {
513
514                 pmc_cntrl[0][i].evnts = ctr[i].event;
515                 pmc_cntrl[0][i].masks = ctr[i].unit_mask;
516                 pmc_cntrl[0][i].enabled = ctr[i].enabled;
517                 pmc_cntrl[0][i].vcntr = i;
518
519                 for_each_possible_cpu(j)
520                         per_cpu(pmc_values, j)[i] = 0;
521         }
522
523         /* Setup the thread 1 events, map the thread 0 event to the
524          * equivalent thread 1 event.
525          */
526         for (i = 0; i < num_ctrs; ++i) {
527                 if ((ctr[i].event >= 2100) && (ctr[i].event <= 2111))
528                         pmc_cntrl[1][i].evnts = ctr[i].event + 19;
529                 else if (ctr[i].event == 2203)
530                         pmc_cntrl[1][i].evnts = ctr[i].event;
531                 else if ((ctr[i].event >= 2200) && (ctr[i].event <= 2215))
532                         pmc_cntrl[1][i].evnts = ctr[i].event + 16;
533                 else
534                         pmc_cntrl[1][i].evnts = ctr[i].event;
535
536                 pmc_cntrl[1][i].masks = ctr[i].unit_mask;
537                 pmc_cntrl[1][i].enabled = ctr[i].enabled;
538                 pmc_cntrl[1][i].vcntr = i;
539         }
540
541         for (i = 0; i < NUM_TRACE_BUS_WORDS; i++)
542                 trace_bus[i] = 0xff;
543
544         for (i = 0; i < NUM_INPUT_BUS_WORDS; i++)
545                 input_bus[i] = 0xff;
546
547         /* Our counters count up, and "count" refers to
548          * how much before the next interrupt, and we interrupt
549          * on overflow.  So we calculate the starting value
550          * which will give us "count" until overflow.
551          * Then we set the events on the enabled counters.
552          */
553         for (i = 0; i < num_counters; ++i) {
554                 /* start with virtual counter set 0 */
555                 if (pmc_cntrl[0][i].enabled) {
556                         /* Using 32bit counters, reset max - count */
557                         reset_value[i] = 0xFFFFFFFF - ctr[i].count;
558                         set_pm_event(i,
559                                      pmc_cntrl[0][i].evnts,
560                                      pmc_cntrl[0][i].masks);
561
562                         /* global, used by cell_cpu_setup */
563                         ctr_enabled |= (1 << i);
564                 }
565         }
566
567         /* initialize the previous counts for the virtual cntrs */
568         for_each_online_cpu(cpu)
569                 for (i = 0; i < num_counters; ++i) {
570                         per_cpu(pmc_values, cpu)[i] = reset_value[i];
571                 }
572 out:
573         ;
574 }
575
576 /* This function is called once for each cpu */
577 static void cell_cpu_setup(struct op_counter_config *cntr)
578 {
579         u32 cpu = smp_processor_id();
580         u32 num_enabled = 0;
581         int i;
582
583         /* There is one performance monitor per processor chip (i.e. node),
584          * so we only need to perform this function once per node.
585          */
586         if (cbe_get_hw_thread_id(cpu))
587                 goto out;
588
589         if (pm_rtas_token == RTAS_UNKNOWN_SERVICE) {
590                 printk(KERN_WARNING "%s: RTAS_UNKNOWN_SERVICE\n",
591                        __FUNCTION__);
592                 goto out;
593         }
594
595         /* Stop all counters */
596         cbe_disable_pm(cpu);
597         cbe_disable_pm_interrupts(cpu);
598
599         cbe_write_pm(cpu, pm_interval, 0);
600         cbe_write_pm(cpu, pm_start_stop, 0);
601         cbe_write_pm(cpu, group_control, pm_regs.group_control);
602         cbe_write_pm(cpu, debug_bus_control, pm_regs.debug_bus_control);
603         write_pm_cntrl(cpu);
604
605         for (i = 0; i < num_counters; ++i) {
606                 if (ctr_enabled & (1 << i)) {
607                         pm_signal[num_enabled].cpu = cbe_cpu_to_node(cpu);
608                         num_enabled++;
609                 }
610         }
611
612         pm_rtas_activate_signals(cbe_cpu_to_node(cpu), num_enabled);
613 out:
614         ;
615 }
616
617 static void cell_global_start(struct op_counter_config *ctr)
618 {
619         u32 cpu;
620         u32 interrupt_mask = 0;
621         u32 i;
622
623         /* This routine gets called once for the system.
624          * There is one performance monitor per node, so we
625          * only need to perform this function once per node.
626          */
627         for_each_online_cpu(cpu) {
628                 if (cbe_get_hw_thread_id(cpu))
629                         continue;
630
631                 interrupt_mask = 0;
632
633                 for (i = 0; i < num_counters; ++i) {
634                         if (ctr_enabled & (1 << i)) {
635                                 cbe_write_ctr(cpu, i, reset_value[i]);
636                                 enable_ctr(cpu, i, pm_regs.pm07_cntrl);
637                                 interrupt_mask |=
638                                     CBE_PM_CTR_OVERFLOW_INTR(i);
639                         } else {
640                                 /* Disable counter */
641                                 cbe_write_pm07_control(cpu, i, 0);
642                         }
643                 }
644
645                 cbe_get_and_clear_pm_interrupts(cpu);
646                 cbe_enable_pm_interrupts(cpu, hdw_thread, interrupt_mask);
647                 cbe_enable_pm(cpu);
648         }
649
650         virt_cntr_inter_mask = interrupt_mask;
651         oprofile_running = 1;
652         smp_wmb();
653
654         /* NOTE: start_virt_cntrs will result in cell_virtual_cntr() being
655          * executed which manipulates the PMU.  We start the "virtual counter"
656          * here so that we do not need to synchronize access to the PMU in
657          * the above for-loop.
658          */
659         start_virt_cntrs();
660 }
661
662 static void cell_global_stop(void)
663 {
664         int cpu;
665
666         /* This routine will be called once for the system.
667          * There is one performance monitor per node, so we
668          * only need to perform this function once per node.
669          */
670         del_timer_sync(&timer_virt_cntr);
671         oprofile_running = 0;
672         smp_wmb();
673
674         for_each_online_cpu(cpu) {
675                 if (cbe_get_hw_thread_id(cpu))
676                         continue;
677
678                 cbe_sync_irq(cbe_cpu_to_node(cpu));
679                 /* Stop the counters */
680                 cbe_disable_pm(cpu);
681
682                 /* Deactivate the signals */
683                 pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
684
685                 /* Deactivate interrupts */
686                 cbe_disable_pm_interrupts(cpu);
687         }
688 }
689
690 static void
691 cell_handle_interrupt(struct pt_regs *regs, struct op_counter_config *ctr)
692 {
693         u32 cpu;
694         u64 pc;
695         int is_kernel;
696         unsigned long flags = 0;
697         u32 interrupt_mask;
698         int i;
699
700         cpu = smp_processor_id();
701
702         /* Need to make sure the interrupt handler and the virt counter
703          * routine are not running at the same time. See the
704          * cell_virtual_cntr() routine for additional comments.
705          */
706         spin_lock_irqsave(&virt_cntr_lock, flags);
707
708         /* Need to disable and reenable the performance counters
709          * to get the desired behavior from the hardware.  This
710          * is hardware specific.
711          */
712
713         cbe_disable_pm(cpu);
714
715         interrupt_mask = cbe_get_and_clear_pm_interrupts(cpu);
716
717         /* If the interrupt mask has been cleared, then the virt cntr
718          * has cleared the interrupt.  When the thread that generated
719          * the interrupt is restored, the data count will be restored to
720          * 0xffffff0 to cause the interrupt to be regenerated.
721          */
722
723         if ((oprofile_running == 1) && (interrupt_mask != 0)) {
724                 pc = regs->nip;
725                 is_kernel = is_kernel_addr(pc);
726
727                 for (i = 0; i < num_counters; ++i) {
728                         if ((interrupt_mask & CBE_PM_CTR_OVERFLOW_INTR(i))
729                             && ctr[i].enabled) {
730                                 oprofile_add_pc(pc, is_kernel, i);
731                                 cbe_write_ctr(cpu, i, reset_value[i]);
732                         }
733                 }
734
735                 /* The counters were frozen by the interrupt.
736                  * Reenable the interrupt and restart the counters.
737                  * If there was a race between the interrupt handler and
738                  * the virtual counter routine.  The virutal counter
739                  * routine may have cleared the interrupts.  Hence must
740                  * use the virt_cntr_inter_mask to re-enable the interrupts.
741                  */
742                 cbe_enable_pm_interrupts(cpu, hdw_thread,
743                                          virt_cntr_inter_mask);
744
745                 /* The writes to the various performance counters only writes
746                  * to a latch.  The new values (interrupt setting bits, reset
747                  * counter value etc.) are not copied to the actual registers
748                  * until the performance monitor is enabled.  In order to get
749                  * this to work as desired, the permormance monitor needs to
750                  * be disabled while writting to the latches.  This is a
751                  * HW design issue.
752                  */
753                 cbe_enable_pm(cpu);
754         }
755         spin_unlock_irqrestore(&virt_cntr_lock, flags);
756 }
757
758 struct op_powerpc_model op_model_cell = {
759         .reg_setup = cell_reg_setup,
760         .cpu_setup = cell_cpu_setup,
761         .global_start = cell_global_start,
762         .global_stop = cell_global_stop,
763         .handle_interrupt = cell_handle_interrupt,
764 };