Merge branch 'for-linus' of master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband
[linux-2.6] / 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 #include <asm/irq_regs.h>
23
24 #include <asm/ptrace.h>
25 #include <asm/atomic.h>
26
27 #include <asm/delay.h>
28 #include <asm/irq.h>
29 #include <asm/page.h>
30 #include <asm/pgalloc.h>
31 #include <asm/pgtable.h>
32 #include <asm/oplib.h>
33 #include <asm/cpudata.h>
34
35 #define IRQ_RESCHEDULE          13
36 #define IRQ_STOP_CPU            14
37 #define IRQ_CROSS_CALL          15
38
39 extern ctxd_t *srmmu_ctx_table_phys;
40
41 extern void calibrate_delay(void);
42
43 extern volatile unsigned long cpu_callin_map[NR_CPUS];
44 extern unsigned char boot_cpu_id;
45
46 extern cpumask_t smp_commenced_mask;
47
48 extern int __smp4m_processor_id(void);
49
50 /*#define SMP_DEBUG*/
51
52 #ifdef SMP_DEBUG
53 #define SMP_PRINTK(x)   printk x
54 #else
55 #define SMP_PRINTK(x)
56 #endif
57
58 static inline unsigned long swap(volatile unsigned long *ptr, unsigned long val)
59 {
60         __asm__ __volatile__("swap [%1], %0\n\t" :
61                              "=&r" (val), "=&r" (ptr) :
62                              "0" (val), "1" (ptr));
63         return val;
64 }
65
66 static void smp_setup_percpu_timer(void);
67 extern void cpu_probe(void);
68
69 void __cpuinit smp4m_callin(void)
70 {
71         int cpuid = hard_smp_processor_id();
72
73         local_flush_cache_all();
74         local_flush_tlb_all();
75
76         /* Get our local ticker going. */
77         smp_setup_percpu_timer();
78
79         calibrate_delay();
80         smp_store_cpu_info(cpuid);
81
82         local_flush_cache_all();
83         local_flush_tlb_all();
84
85         /*
86          * Unblock the master CPU _only_ when the scheduler state
87          * of all secondary CPUs will be up-to-date, so after
88          * the SMP initialization the master will be just allowed
89          * to call the scheduler code.
90          */
91         /* Allow master to continue. */
92         swap(&cpu_callin_map[cpuid], 1);
93
94         /* XXX: What's up with all the flushes? */
95         local_flush_cache_all();
96         local_flush_tlb_all();
97         
98         cpu_probe();
99
100         /* Fix idle thread fields. */
101         __asm__ __volatile__("ld [%0], %%g6\n\t"
102                              : : "r" (&current_set[cpuid])
103                              : "memory" /* paranoid */);
104
105         /* Attach to the address space of init_task. */
106         atomic_inc(&init_mm.mm_count);
107         current->active_mm = &init_mm;
108
109         while (!cpu_isset(cpuid, smp_commenced_mask))
110                 mb();
111
112         local_irq_enable();
113
114         cpu_set(cpuid, cpu_online_map);
115 }
116
117 /*
118  *      Cycle through the processors asking the PROM to start each one.
119  */
120  
121 extern struct linux_prom_registers smp_penguin_ctable;
122 extern unsigned long trapbase_cpu1[];
123 extern unsigned long trapbase_cpu2[];
124 extern unsigned long trapbase_cpu3[];
125
126 void __init smp4m_boot_cpus(void)
127 {
128         smp_setup_percpu_timer();
129         local_flush_cache_all();
130 }
131
132 int __cpuinit smp4m_boot_one_cpu(int i)
133 {
134         extern unsigned long sun4m_cpu_startup;
135         unsigned long *entry = &sun4m_cpu_startup;
136         struct task_struct *p;
137         int timeout;
138         int cpu_node;
139
140         cpu_find_by_mid(i, &cpu_node);
141
142         /* Cook up an idler for this guy. */
143         p = fork_idle(i);
144         current_set[i] = task_thread_info(p);
145         /* See trampoline.S for details... */
146         entry += ((i-1) * 3);
147
148         /*
149          * Initialize the contexts table
150          * Since the call to prom_startcpu() trashes the structure,
151          * we need to re-initialize it for each cpu
152          */
153         smp_penguin_ctable.which_io = 0;
154         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
155         smp_penguin_ctable.reg_size = 0;
156
157         /* whirrr, whirrr, whirrrrrrrrr... */
158         printk("Starting CPU %d at %p\n", i, entry);
159         local_flush_cache_all();
160         prom_startcpu(cpu_node,
161                       &smp_penguin_ctable, 0, (char *)entry);
162
163         /* wheee... it's going... */
164         for(timeout = 0; timeout < 10000; timeout++) {
165                 if(cpu_callin_map[i])
166                         break;
167                 udelay(200);
168         }
169
170         if (!(cpu_callin_map[i])) {
171                 printk("Processor %d is stuck.\n", i);
172                 return -ENODEV;
173         }
174
175         local_flush_cache_all();
176         return 0;
177 }
178
179 void __init smp4m_smp_done(void)
180 {
181         int i, first;
182         int *prev;
183
184         /* setup cpu list for irq rotation */
185         first = 0;
186         prev = &first;
187         for (i = 0; i < NR_CPUS; i++) {
188                 if (cpu_online(i)) {
189                         *prev = i;
190                         prev = &cpu_data(i).next;
191                 }
192         }
193         *prev = first;
194         local_flush_cache_all();
195
196         /* Free unneeded trap tables */
197         if (!cpu_isset(1, cpu_present_map)) {
198                 ClearPageReserved(virt_to_page(trapbase_cpu1));
199                 init_page_count(virt_to_page(trapbase_cpu1));
200                 free_page((unsigned long)trapbase_cpu1);
201                 totalram_pages++;
202                 num_physpages++;
203         }
204         if (!cpu_isset(2, cpu_present_map)) {
205                 ClearPageReserved(virt_to_page(trapbase_cpu2));
206                 init_page_count(virt_to_page(trapbase_cpu2));
207                 free_page((unsigned long)trapbase_cpu2);
208                 totalram_pages++;
209                 num_physpages++;
210         }
211         if (!cpu_isset(3, cpu_present_map)) {
212                 ClearPageReserved(virt_to_page(trapbase_cpu3));
213                 init_page_count(virt_to_page(trapbase_cpu3));
214                 free_page((unsigned long)trapbase_cpu3);
215                 totalram_pages++;
216                 num_physpages++;
217         }
218
219         /* Ok, they are spinning and ready to go. */
220 }
221
222 /* At each hardware IRQ, we get this called to forward IRQ reception
223  * to the next processor.  The caller must disable the IRQ level being
224  * serviced globally so that there are no double interrupts received.
225  *
226  * XXX See sparc64 irq.c.
227  */
228 void smp4m_irq_rotate(int cpu)
229 {
230         int next = cpu_data(cpu).next;
231         if (next != cpu)
232                 set_irq_udt(next);
233 }
234
235 /* Cross calls, in order to work efficiently and atomically do all
236  * the message passing work themselves, only stopcpu and reschedule
237  * messages come through here.
238  */
239 void smp4m_message_pass(int target, int msg, unsigned long data, int wait)
240 {
241         static unsigned long smp_cpu_in_msg[NR_CPUS];
242         cpumask_t mask;
243         int me = smp_processor_id();
244         int irq, i;
245
246         if(msg == MSG_RESCHEDULE) {
247                 irq = IRQ_RESCHEDULE;
248
249                 if(smp_cpu_in_msg[me])
250                         return;
251         } else if(msg == MSG_STOP_CPU) {
252                 irq = IRQ_STOP_CPU;
253         } else {
254                 goto barf;
255         }
256
257         smp_cpu_in_msg[me]++;
258         if(target == MSG_ALL_BUT_SELF || target == MSG_ALL) {
259                 mask = cpu_online_map;
260                 if(target == MSG_ALL_BUT_SELF)
261                         cpu_clear(me, mask);
262                 for(i = 0; i < 4; i++) {
263                         if (cpu_isset(i, mask))
264                                 set_cpu_int(i, irq);
265                 }
266         } else {
267                 set_cpu_int(target, irq);
268         }
269         smp_cpu_in_msg[me]--;
270
271         return;
272 barf:
273         printk("Yeeee, trying to send SMP msg(%d) on cpu %d\n", msg, me);
274         panic("Bogon SMP message pass.");
275 }
276
277 static struct smp_funcall {
278         smpfunc_t func;
279         unsigned long arg1;
280         unsigned long arg2;
281         unsigned long arg3;
282         unsigned long arg4;
283         unsigned long arg5;
284         unsigned long processors_in[SUN4M_NCPUS];  /* Set when ipi entered. */
285         unsigned long processors_out[SUN4M_NCPUS]; /* Set when ipi exited. */
286 } ccall_info;
287
288 static DEFINE_SPINLOCK(cross_call_lock);
289
290 /* Cross calls must be serialized, at least currently. */
291 void smp4m_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
292                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
293 {
294                 register int ncpus = SUN4M_NCPUS;
295                 unsigned long flags;
296
297                 spin_lock_irqsave(&cross_call_lock, flags);
298
299                 /* Init function glue. */
300                 ccall_info.func = func;
301                 ccall_info.arg1 = arg1;
302                 ccall_info.arg2 = arg2;
303                 ccall_info.arg3 = arg3;
304                 ccall_info.arg4 = arg4;
305                 ccall_info.arg5 = arg5;
306
307                 /* Init receive/complete mapping, plus fire the IPI's off. */
308                 {
309                         cpumask_t mask = cpu_online_map;
310                         register int i;
311
312                         cpu_clear(smp_processor_id(), mask);
313                         for(i = 0; i < ncpus; i++) {
314                                 if (cpu_isset(i, mask)) {
315                                         ccall_info.processors_in[i] = 0;
316                                         ccall_info.processors_out[i] = 0;
317                                         set_cpu_int(i, IRQ_CROSS_CALL);
318                                 } else {
319                                         ccall_info.processors_in[i] = 1;
320                                         ccall_info.processors_out[i] = 1;
321                                 }
322                         }
323                 }
324
325                 {
326                         register int i;
327
328                         i = 0;
329                         do {
330                                 while(!ccall_info.processors_in[i])
331                                         barrier();
332                         } while(++i < ncpus);
333
334                         i = 0;
335                         do {
336                                 while(!ccall_info.processors_out[i])
337                                         barrier();
338                         } while(++i < ncpus);
339                 }
340
341                 spin_unlock_irqrestore(&cross_call_lock, flags);
342 }
343
344 /* Running cross calls. */
345 void smp4m_cross_call_irq(void)
346 {
347         int i = smp_processor_id();
348
349         ccall_info.processors_in[i] = 1;
350         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
351                         ccall_info.arg4, ccall_info.arg5);
352         ccall_info.processors_out[i] = 1;
353 }
354
355 void smp4m_percpu_timer_interrupt(struct pt_regs *regs)
356 {
357         struct pt_regs *old_regs;
358         int cpu = smp_processor_id();
359
360         old_regs = set_irq_regs(regs);
361
362         clear_profile_irq(cpu);
363
364         profile_tick(CPU_PROFILING);
365
366         if(!--prof_counter(cpu)) {
367                 int user = user_mode(regs);
368
369                 irq_enter();
370                 update_process_times(user);
371                 irq_exit();
372
373                 prof_counter(cpu) = prof_multiplier(cpu);
374         }
375         set_irq_regs(old_regs);
376 }
377
378 extern unsigned int lvl14_resolution;
379
380 static void __init smp_setup_percpu_timer(void)
381 {
382         int cpu = smp_processor_id();
383
384         prof_counter(cpu) = prof_multiplier(cpu) = 1;
385         load_profile_irq(cpu, lvl14_resolution);
386
387         if(cpu == boot_cpu_id)
388                 enable_pil_irq(14);
389 }
390
391 void __init smp4m_blackbox_id(unsigned *addr)
392 {
393         int rd = *addr & 0x3e000000;
394         int rs1 = rd >> 11;
395         
396         addr[0] = 0x81580000 | rd;              /* rd %tbr, reg */
397         addr[1] = 0x8130200c | rd | rs1;        /* srl reg, 0xc, reg */
398         addr[2] = 0x80082003 | rd | rs1;        /* and reg, 3, reg */
399 }
400
401 void __init smp4m_blackbox_current(unsigned *addr)
402 {
403         int rd = *addr & 0x3e000000;
404         int rs1 = rd >> 11;
405         
406         addr[0] = 0x81580000 | rd;              /* rd %tbr, reg */
407         addr[2] = 0x8130200a | rd | rs1;        /* srl reg, 0xa, reg */
408         addr[4] = 0x8008200c | rd | rs1;        /* and reg, 3, reg */
409 }
410
411 void __init sun4m_init_smp(void)
412 {
413         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4m_blackbox_id);
414         BTFIXUPSET_BLACKBOX(load_current, smp4m_blackbox_current);
415         BTFIXUPSET_CALL(smp_cross_call, smp4m_cross_call, BTFIXUPCALL_NORM);
416         BTFIXUPSET_CALL(smp_message_pass, smp4m_message_pass, BTFIXUPCALL_NORM);
417         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4m_processor_id, BTFIXUPCALL_NORM);
418 }