vserver 1.9.3
[linux-2.6.git] / arch / sparc / kernel / sun4m_smp.c
1 /* sun4m_smp.c: Sparc SUN4M SMP support.
2  *
3  * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
4  */
5
6 #include <asm/head.h>
7
8 #include <linux/kernel.h>
9 #include <linux/sched.h>
10 #include <linux/threads.h>
11 #include <linux/smp.h>
12 #include <linux/smp_lock.h>
13 #include <linux/interrupt.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/init.h>
16 #include <linux/spinlock.h>
17 #include <linux/mm.h>
18 #include <linux/swap.h>
19 #include <linux/profile.h>
20 #include <asm/cacheflush.h>
21 #include <asm/tlbflush.h>
22
23 #include <asm/ptrace.h>
24 #include <asm/atomic.h>
25
26 #include <asm/delay.h>
27 #include <asm/irq.h>
28 #include <asm/page.h>
29 #include <asm/pgalloc.h>
30 #include <asm/pgtable.h>
31 #include <asm/oplib.h>
32 #include <asm/cpudata.h>
33
34 #define IRQ_RESCHEDULE          13
35 #define IRQ_STOP_CPU            14
36 #define IRQ_CROSS_CALL          15
37
38 extern ctxd_t *srmmu_ctx_table_phys;
39
40 extern void calibrate_delay(void);
41
42 extern volatile int smp_processors_ready;
43 extern int smp_num_cpus;
44 extern int smp_threads_ready;
45 extern volatile unsigned long cpu_callin_map[NR_CPUS];
46 extern unsigned char boot_cpu_id;
47 extern int smp_activated;
48 extern volatile int __cpu_number_map[NR_CPUS];
49 extern volatile int __cpu_logical_map[NR_CPUS];
50 extern volatile unsigned long ipi_count;
51 extern volatile int smp_process_available;
52 extern volatile int smp_commenced;
53 extern int __smp4m_processor_id(void);
54
55 /*#define SMP_DEBUG*/
56
57 #ifdef SMP_DEBUG
58 #define SMP_PRINTK(x)   printk x
59 #else
60 #define SMP_PRINTK(x)
61 #endif
62
63 static inline unsigned long swap(volatile unsigned long *ptr, unsigned long val)
64 {
65         __asm__ __volatile__("swap [%1], %0\n\t" :
66                              "=&r" (val), "=&r" (ptr) :
67                              "0" (val), "1" (ptr));
68         return val;
69 }
70
71 static void smp_setup_percpu_timer(void);
72 extern void cpu_probe(void);
73
74 void __init smp4m_callin(void)
75 {
76         int cpuid = hard_smp_processor_id();
77
78         local_flush_cache_all();
79         local_flush_tlb_all();
80
81         set_irq_udt(boot_cpu_id);
82
83         /* Get our local ticker going. */
84         smp_setup_percpu_timer();
85
86         calibrate_delay();
87         smp_store_cpu_info(cpuid);
88
89         local_flush_cache_all();
90         local_flush_tlb_all();
91
92         /*
93          * Unblock the master CPU _only_ when the scheduler state
94          * of all secondary CPUs will be up-to-date, so after
95          * the SMP initialization the master will be just allowed
96          * to call the scheduler code.
97          */
98         /* Allow master to continue. */
99         swap((unsigned long *)&cpu_callin_map[cpuid], 1);
100
101         local_flush_cache_all();
102         local_flush_tlb_all();
103         
104         cpu_probe();
105
106         /* Fix idle thread fields. */
107         __asm__ __volatile__("ld [%0], %%g6\n\t"
108                              : : "r" (&current_set[cpuid])
109                              : "memory" /* paranoid */);
110
111         /* Attach to the address space of init_task. */
112         atomic_inc(&init_mm.mm_count);
113         current->active_mm = &init_mm;
114
115         while(!smp_commenced)
116                 barrier();
117
118         local_flush_cache_all();
119         local_flush_tlb_all();
120
121         local_irq_enable();
122 }
123
124 extern int cpu_idle(void *unused);
125 extern void init_IRQ(void);
126 extern void cpu_panic(void);
127
128 /*
129  *      Cycle through the processors asking the PROM to start each one.
130  */
131  
132 extern struct linux_prom_registers smp_penguin_ctable;
133 extern unsigned long trapbase_cpu1[];
134 extern unsigned long trapbase_cpu2[];
135 extern unsigned long trapbase_cpu3[];
136
137 void __init smp4m_boot_cpus(void)
138 {
139         int cpucount = 0;
140         int i, mid;
141
142         printk("Entering SMP Mode...\n");
143
144         local_irq_enable();
145         cpus_clear(cpu_present_map);
146
147         for (i = 0; !cpu_find_by_instance(i, NULL, &mid); i++)
148                 cpu_set(mid, cpu_present_map);
149
150         for(i=0; i < NR_CPUS; i++) {
151                 __cpu_number_map[i] = -1;
152                 __cpu_logical_map[i] = -1;
153         }
154
155         __cpu_number_map[boot_cpu_id] = 0;
156         __cpu_logical_map[0] = boot_cpu_id;
157         current_thread_info()->cpu = boot_cpu_id;
158
159         smp_store_cpu_info(boot_cpu_id);
160         set_irq_udt(boot_cpu_id);
161         smp_setup_percpu_timer();
162         local_flush_cache_all();
163         if(cpu_find_by_instance(1, NULL, NULL))
164                 return;  /* Not an MP box. */
165         for(i = 0; i < NR_CPUS; i++) {
166                 if(i == boot_cpu_id)
167                         continue;
168
169                 if (cpu_isset(i, cpu_present_map)) {
170                         extern unsigned long sun4m_cpu_startup;
171                         unsigned long *entry = &sun4m_cpu_startup;
172                         struct task_struct *p;
173                         int timeout;
174
175                         /* Cook up an idler for this guy. */
176                         p = fork_idle(i);
177                         cpucount++;
178                         current_set[i] = p->thread_info;
179                         /* See trampoline.S for details... */
180                         entry += ((i-1) * 3);
181
182                         /*
183                          * Initialize the contexts table
184                          * Since the call to prom_startcpu() trashes the structure,
185                          * we need to re-initialize it for each cpu
186                          */
187                         smp_penguin_ctable.which_io = 0;
188                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
189                         smp_penguin_ctable.reg_size = 0;
190
191                         /* whirrr, whirrr, whirrrrrrrrr... */
192                         printk("Starting CPU %d at %p\n", i, entry);
193                         local_flush_cache_all();
194                         prom_startcpu(cpu_data(i).prom_node,
195                                       &smp_penguin_ctable, 0, (char *)entry);
196
197                         /* wheee... it's going... */
198                         for(timeout = 0; timeout < 10000; timeout++) {
199                                 if(cpu_callin_map[i])
200                                         break;
201                                 udelay(200);
202                         }
203                         if(cpu_callin_map[i]) {
204                                 /* Another "Red Snapper". */
205                                 __cpu_number_map[i] = i;
206                                 __cpu_logical_map[i] = i;
207                         } else {
208                                 cpucount--;
209                                 printk("Processor %d is stuck.\n", i);
210                         }
211                 }
212                 if(!(cpu_callin_map[i])) {
213                         cpu_clear(i, cpu_present_map);
214                         __cpu_number_map[i] = -1;
215                 }
216         }
217         local_flush_cache_all();
218         if(cpucount == 0) {
219                 printk("Error: only one Processor found.\n");
220                 cpu_present_map = cpumask_of_cpu(smp_processor_id());
221         } else {
222                 unsigned long bogosum = 0;
223                 for(i = 0; i < NR_CPUS; i++) {
224                         if (cpu_isset(i, cpu_present_map))
225                                 bogosum += cpu_data(i).udelay_val;
226                 }
227                 printk("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n",
228                        cpucount + 1,
229                        bogosum/(500000/HZ),
230                        (bogosum/(5000/HZ))%100);
231                 smp_activated = 1;
232                 smp_num_cpus = cpucount + 1;
233         }
234
235         /* Free unneeded trap tables */
236         if (!cpu_isset(i, cpu_present_map)) {
237                 ClearPageReserved(virt_to_page(trapbase_cpu1));
238                 set_page_count(virt_to_page(trapbase_cpu1), 1);
239                 free_page((unsigned long)trapbase_cpu1);
240                 totalram_pages++;
241                 num_physpages++;
242         }
243         if (!cpu_isset(2, cpu_present_map)) {
244                 ClearPageReserved(virt_to_page(trapbase_cpu2));
245                 set_page_count(virt_to_page(trapbase_cpu2), 1);
246                 free_page((unsigned long)trapbase_cpu2);
247                 totalram_pages++;
248                 num_physpages++;
249         }
250         if (!cpu_isset(3, cpu_present_map)) {
251                 ClearPageReserved(virt_to_page(trapbase_cpu3));
252                 set_page_count(virt_to_page(trapbase_cpu3), 1);
253                 free_page((unsigned long)trapbase_cpu3);
254                 totalram_pages++;
255                 num_physpages++;
256         }
257
258         /* Ok, they are spinning and ready to go. */
259         smp_processors_ready = 1;
260 }
261
262 /* At each hardware IRQ, we get this called to forward IRQ reception
263  * to the next processor.  The caller must disable the IRQ level being
264  * serviced globally so that there are no double interrupts received.
265  *
266  * XXX See sparc64 irq.c.
267  */
268 void smp4m_irq_rotate(int cpu)
269 {
270 }
271
272 /* Cross calls, in order to work efficiently and atomically do all
273  * the message passing work themselves, only stopcpu and reschedule
274  * messages come through here.
275  */
276 void smp4m_message_pass(int target, int msg, unsigned long data, int wait)
277 {
278         static unsigned long smp_cpu_in_msg[NR_CPUS];
279         cpumask_t mask;
280         int me = smp_processor_id();
281         int irq, i;
282
283         if(msg == MSG_RESCHEDULE) {
284                 irq = IRQ_RESCHEDULE;
285
286                 if(smp_cpu_in_msg[me])
287                         return;
288         } else if(msg == MSG_STOP_CPU) {
289                 irq = IRQ_STOP_CPU;
290         } else {
291                 goto barf;
292         }
293
294         smp_cpu_in_msg[me]++;
295         if(target == MSG_ALL_BUT_SELF || target == MSG_ALL) {
296                 mask = cpu_present_map;
297                 if(target == MSG_ALL_BUT_SELF)
298                         cpu_clear(me, mask);
299                 for(i = 0; i < 4; i++) {
300                         if (cpu_isset(i, mask))
301                                 set_cpu_int(i, irq);
302                 }
303         } else {
304                 set_cpu_int(target, irq);
305         }
306         smp_cpu_in_msg[me]--;
307
308         return;
309 barf:
310         printk("Yeeee, trying to send SMP msg(%d) on cpu %d\n", msg, me);
311         panic("Bogon SMP message pass.");
312 }
313
314 static struct smp_funcall {
315         smpfunc_t func;
316         unsigned long arg1;
317         unsigned long arg2;
318         unsigned long arg3;
319         unsigned long arg4;
320         unsigned long arg5;
321         unsigned long processors_in[NR_CPUS];  /* Set when ipi entered. */
322         unsigned long processors_out[NR_CPUS]; /* Set when ipi exited. */
323 } ccall_info;
324
325 static spinlock_t cross_call_lock = SPIN_LOCK_UNLOCKED;
326
327 /* Cross calls must be serialized, at least currently. */
328 void smp4m_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
329                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
330 {
331         if(smp_processors_ready) {
332                 register int ncpus = smp_num_cpus;
333                 unsigned long flags;
334
335                 spin_lock_irqsave(&cross_call_lock, flags);
336
337                 /* Init function glue. */
338                 ccall_info.func = func;
339                 ccall_info.arg1 = arg1;
340                 ccall_info.arg2 = arg2;
341                 ccall_info.arg3 = arg3;
342                 ccall_info.arg4 = arg4;
343                 ccall_info.arg5 = arg5;
344
345                 /* Init receive/complete mapping, plus fire the IPI's off. */
346                 {
347                         cpumask_t mask = cpu_present_map;
348                         register int i;
349
350                         cpu_clear(smp_processor_id(), mask);
351                         for(i = 0; i < ncpus; i++) {
352                                 if (cpu_isset(i, mask)) {
353                                         ccall_info.processors_in[i] = 0;
354                                         ccall_info.processors_out[i] = 0;
355                                         set_cpu_int(i, IRQ_CROSS_CALL);
356                                 } else {
357                                         ccall_info.processors_in[i] = 1;
358                                         ccall_info.processors_out[i] = 1;
359                                 }
360                         }
361                 }
362
363                 {
364                         register int i;
365
366                         i = 0;
367                         do {
368                                 while(!ccall_info.processors_in[i])
369                                         barrier();
370                         } while(++i < ncpus);
371
372                         i = 0;
373                         do {
374                                 while(!ccall_info.processors_out[i])
375                                         barrier();
376                         } while(++i < ncpus);
377                 }
378
379                 spin_unlock_irqrestore(&cross_call_lock, flags);
380         }
381 }
382
383 /* Running cross calls. */
384 void smp4m_cross_call_irq(void)
385 {
386         int i = smp_processor_id();
387
388         ccall_info.processors_in[i] = 1;
389         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
390                         ccall_info.arg4, ccall_info.arg5);
391         ccall_info.processors_out[i] = 1;
392 }
393
394 void smp4m_percpu_timer_interrupt(struct pt_regs *regs)
395 {
396         int cpu = smp_processor_id();
397
398         clear_profile_irq(cpu);
399
400         profile_tick(CPU_PROFILING, regs);
401
402         if(!--prof_counter(cpu)) {
403                 int user = user_mode(regs);
404
405                 irq_enter();
406                 update_process_times(user);
407                 irq_exit();
408
409                 prof_counter(cpu) = prof_multiplier(cpu);
410         }
411 }
412
413 extern unsigned int lvl14_resolution;
414
415 static void __init smp_setup_percpu_timer(void)
416 {
417         int cpu = smp_processor_id();
418
419         prof_counter(cpu) = prof_multiplier(cpu) = 1;
420         load_profile_irq(cpu, lvl14_resolution);
421
422         if(cpu == boot_cpu_id)
423                 enable_pil_irq(14);
424 }
425
426 void __init smp4m_blackbox_id(unsigned *addr)
427 {
428         int rd = *addr & 0x3e000000;
429         int rs1 = rd >> 11;
430         
431         addr[0] = 0x81580000 | rd;              /* rd %tbr, reg */
432         addr[1] = 0x8130200c | rd | rs1;        /* srl reg, 0xc, reg */
433         addr[2] = 0x80082003 | rd | rs1;        /* and reg, 3, reg */
434 }
435
436 void __init smp4m_blackbox_current(unsigned *addr)
437 {
438         int rd = *addr & 0x3e000000;
439         int rs1 = rd >> 11;
440         
441         addr[0] = 0x81580000 | rd;              /* rd %tbr, reg */
442         addr[2] = 0x8130200a | rd | rs1;        /* srl reg, 0xa, reg */
443         addr[4] = 0x8008200c | rd | rs1;        /* and reg, 3, reg */
444 }
445
446 void __init sun4m_init_smp(void)
447 {
448         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4m_blackbox_id);
449         BTFIXUPSET_BLACKBOX(load_current, smp4m_blackbox_current);
450         BTFIXUPSET_CALL(smp_cross_call, smp4m_cross_call, BTFIXUPCALL_NORM);
451         BTFIXUPSET_CALL(smp_message_pass, smp4m_message_pass, BTFIXUPCALL_NORM);
452         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4m_processor_id, BTFIXUPCALL_NORM);
453 }