d2c7e2c4733b72130cbd90020b0ae61cc7ed4897
[linux-2.6.git] / arch / ppc64 / kernel / pSeries_smp.c
1 /*
2  * SMP support for pSeries and BPA machines.
3  *
4  * Dave Engebretsen, Peter Bergner, and
5  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
6  *
7  * Plus various changes from other IBM teams...
8  *
9  *      This program is free software; you can redistribute it and/or
10  *      modify it under the terms of the GNU General Public License
11  *      as published by the Free Software Foundation; either version
12  *      2 of the License, or (at your option) any later version.
13  */
14
15 #undef DEBUG
16
17 #include <linux/config.h>
18 #include <linux/kernel.h>
19 #include <linux/module.h>
20 #include <linux/sched.h>
21 #include <linux/smp.h>
22 #include <linux/interrupt.h>
23 #include <linux/delay.h>
24 #include <linux/init.h>
25 #include <linux/spinlock.h>
26 #include <linux/cache.h>
27 #include <linux/err.h>
28 #include <linux/sysdev.h>
29 #include <linux/cpu.h>
30
31 #include <asm/ptrace.h>
32 #include <asm/atomic.h>
33 #include <asm/irq.h>
34 #include <asm/page.h>
35 #include <asm/pgtable.h>
36 #include <asm/io.h>
37 #include <asm/prom.h>
38 #include <asm/smp.h>
39 #include <asm/paca.h>
40 #include <asm/time.h>
41 #include <asm/machdep.h>
42 #include <asm/xics.h>
43 #include <asm/cputable.h>
44 #include <asm/firmware.h>
45 #include <asm/system.h>
46 #include <asm/rtas.h>
47 #include <asm/plpar_wrappers.h>
48 #include <asm/pSeries_reconfig.h>
49
50 #include "mpic.h"
51 #include "bpa_iic.h"
52
53 #ifdef DEBUG
54 #define DBG(fmt...) udbg_printf(fmt)
55 #else
56 #define DBG(fmt...)
57 #endif
58
59 /*
60  * The primary thread of each non-boot processor is recorded here before
61  * smp init.
62  */
63 static cpumask_t of_spin_map;
64
65 extern void pSeries_secondary_smp_init(unsigned long);
66
67 #ifdef CONFIG_HOTPLUG_CPU
68
69 /* Get state of physical CPU.
70  * Return codes:
71  *      0       - The processor is in the RTAS stopped state
72  *      1       - stop-self is in progress
73  *      2       - The processor is not in the RTAS stopped state
74  *      -1      - Hardware Error
75  *      -2      - Hardware Busy, Try again later.
76  */
77 static int query_cpu_stopped(unsigned int pcpu)
78 {
79         int cpu_status;
80         int status, qcss_tok;
81
82         qcss_tok = rtas_token("query-cpu-stopped-state");
83         if (qcss_tok == RTAS_UNKNOWN_SERVICE)
84                 return -1;
85         status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
86         if (status != 0) {
87                 printk(KERN_ERR
88                        "RTAS query-cpu-stopped-state failed: %i\n", status);
89                 return status;
90         }
91
92         return cpu_status;
93 }
94
95 int pSeries_cpu_disable(void)
96 {
97         int cpu = smp_processor_id();
98
99         cpu_clear(cpu, cpu_online_map);
100         systemcfg->processorCount--;
101
102         /*fix boot_cpuid here*/
103         if (cpu == boot_cpuid)
104                 boot_cpuid = any_online_cpu(cpu_online_map);
105
106         /* FIXME: abstract this to not be platform specific later on */
107         xics_migrate_irqs_away();
108         return 0;
109 }
110
111 void pSeries_cpu_die(unsigned int cpu)
112 {
113         int tries;
114         int cpu_status;
115         unsigned int pcpu = get_hard_smp_processor_id(cpu);
116
117         for (tries = 0; tries < 25; tries++) {
118                 cpu_status = query_cpu_stopped(pcpu);
119                 if (cpu_status == 0 || cpu_status == -1)
120                         break;
121                 msleep(200);
122         }
123         if (cpu_status != 0) {
124                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
125                        cpu, pcpu, cpu_status);
126         }
127
128         /* Isolation and deallocation are definatly done by
129          * drslot_chrp_cpu.  If they were not they would be
130          * done here.  Change isolate state to Isolate and
131          * change allocation-state to Unusable.
132          */
133         paca[cpu].cpu_start = 0;
134 }
135
136 /*
137  * Update cpu_present_map and paca(s) for a new cpu node.  The wrinkle
138  * here is that a cpu device node may represent up to two logical cpus
139  * in the SMT case.  We must honor the assumption in other code that
140  * the logical ids for sibling SMT threads x and y are adjacent, such
141  * that x^1 == y and y^1 == x.
142  */
143 static int pSeries_add_processor(struct device_node *np)
144 {
145         unsigned int cpu;
146         cpumask_t candidate_map, tmp = CPU_MASK_NONE;
147         int err = -ENOSPC, len, nthreads, i;
148         u32 *intserv;
149
150         intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
151         if (!intserv)
152                 return 0;
153
154         nthreads = len / sizeof(u32);
155         for (i = 0; i < nthreads; i++)
156                 cpu_set(i, tmp);
157
158         lock_cpu_hotplug();
159
160         BUG_ON(!cpus_subset(cpu_present_map, cpu_possible_map));
161
162         /* Get a bitmap of unoccupied slots. */
163         cpus_xor(candidate_map, cpu_possible_map, cpu_present_map);
164         if (cpus_empty(candidate_map)) {
165                 /* If we get here, it most likely means that NR_CPUS is
166                  * less than the partition's max processors setting.
167                  */
168                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
169                        " supports %d logical cpus.\n", np->full_name,
170                        cpus_weight(cpu_possible_map));
171                 goto out_unlock;
172         }
173
174         while (!cpus_empty(tmp))
175                 if (cpus_subset(tmp, candidate_map))
176                         /* Found a range where we can insert the new cpu(s) */
177                         break;
178                 else
179                         cpus_shift_left(tmp, tmp, nthreads);
180
181         if (cpus_empty(tmp)) {
182                 printk(KERN_ERR "Unable to find space in cpu_present_map for"
183                        " processor %s with %d thread(s)\n", np->name,
184                        nthreads);
185                 goto out_unlock;
186         }
187
188         for_each_cpu_mask(cpu, tmp) {
189                 BUG_ON(cpu_isset(cpu, cpu_present_map));
190                 cpu_set(cpu, cpu_present_map);
191                 set_hard_smp_processor_id(cpu, *intserv++);
192         }
193         err = 0;
194 out_unlock:
195         unlock_cpu_hotplug();
196         return err;
197 }
198
199 /*
200  * Update the present map for a cpu node which is going away, and set
201  * the hard id in the paca(s) to -1 to be consistent with boot time
202  * convention for non-present cpus.
203  */
204 static void pSeries_remove_processor(struct device_node *np)
205 {
206         unsigned int cpu;
207         int len, nthreads, i;
208         u32 *intserv;
209
210         intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
211         if (!intserv)
212                 return;
213
214         nthreads = len / sizeof(u32);
215
216         lock_cpu_hotplug();
217         for (i = 0; i < nthreads; i++) {
218                 for_each_present_cpu(cpu) {
219                         if (get_hard_smp_processor_id(cpu) != intserv[i])
220                                 continue;
221                         BUG_ON(cpu_online(cpu));
222                         cpu_clear(cpu, cpu_present_map);
223                         set_hard_smp_processor_id(cpu, -1);
224                         break;
225                 }
226                 if (cpu == NR_CPUS)
227                         printk(KERN_WARNING "Could not find cpu to remove "
228                                "with physical id 0x%x\n", intserv[i]);
229         }
230         unlock_cpu_hotplug();
231 }
232
233 static int pSeries_smp_notifier(struct notifier_block *nb, unsigned long action, void *node)
234 {
235         int err = NOTIFY_OK;
236
237         switch (action) {
238         case PSERIES_RECONFIG_ADD:
239                 if (pSeries_add_processor(node))
240                         err = NOTIFY_BAD;
241                 break;
242         case PSERIES_RECONFIG_REMOVE:
243                 pSeries_remove_processor(node);
244                 break;
245         default:
246                 err = NOTIFY_DONE;
247                 break;
248         }
249         return err;
250 }
251
252 static struct notifier_block pSeries_smp_nb = {
253         .notifier_call = pSeries_smp_notifier,
254 };
255
256 #endif /* CONFIG_HOTPLUG_CPU */
257
258 /**
259  * smp_startup_cpu() - start the given cpu
260  *
261  * At boot time, there is nothing to do for primary threads which were
262  * started from Open Firmware.  For anything else, call RTAS with the
263  * appropriate start location.
264  *
265  * Returns:
266  *      0       - failure
267  *      1       - success
268  */
269 static inline int __devinit smp_startup_cpu(unsigned int lcpu)
270 {
271         int status;
272         unsigned long start_here = __pa((u32)*((unsigned long *)
273                                                pSeries_secondary_smp_init));
274         unsigned int pcpu;
275         int start_cpu;
276
277         if (cpu_isset(lcpu, of_spin_map))
278                 /* Already started by OF and sitting in spin loop */
279                 return 1;
280
281         pcpu = get_hard_smp_processor_id(lcpu);
282
283         /* Fixup atomic count: it exited inside IRQ handler. */
284         paca[lcpu].__current->thread_info->preempt_count        = 0;
285
286         /* 
287          * If the RTAS start-cpu token does not exist then presume the
288          * cpu is already spinning.
289          */
290         start_cpu = rtas_token("start-cpu");
291         if (start_cpu == RTAS_UNKNOWN_SERVICE)
292                 return 1;
293
294         status = rtas_call(start_cpu, 3, 1, NULL, pcpu, start_here, lcpu);
295         if (status != 0) {
296                 printk(KERN_ERR "start-cpu failed: %i\n", status);
297                 return 0;
298         }
299
300         return 1;
301 }
302
303 #ifdef CONFIG_XICS
304 static inline void smp_xics_do_message(int cpu, int msg)
305 {
306         set_bit(msg, &xics_ipi_message[cpu].value);
307         mb();
308         xics_cause_IPI(cpu);
309 }
310
311 static void smp_xics_message_pass(int target, int msg)
312 {
313         unsigned int i;
314
315         if (target < NR_CPUS) {
316                 smp_xics_do_message(target, msg);
317         } else {
318                 for_each_online_cpu(i) {
319                         if (target == MSG_ALL_BUT_SELF
320                             && i == smp_processor_id())
321                                 continue;
322                         smp_xics_do_message(i, msg);
323                 }
324         }
325 }
326
327 static int __init smp_xics_probe(void)
328 {
329         xics_request_IPIs();
330
331         return cpus_weight(cpu_possible_map);
332 }
333
334 static void __devinit smp_xics_setup_cpu(int cpu)
335 {
336         if (cpu != boot_cpuid)
337                 xics_setup_cpu();
338
339         if (firmware_has_feature(FW_FEATURE_SPLPAR))
340                 vpa_init(cpu);
341
342         cpu_clear(cpu, of_spin_map);
343
344 }
345 #endif /* CONFIG_XICS */
346 #ifdef CONFIG_BPA_IIC
347 static void smp_iic_message_pass(int target, int msg)
348 {
349         unsigned int i;
350
351         if (target < NR_CPUS) {
352                 iic_cause_IPI(target, msg);
353         } else {
354                 for_each_online_cpu(i) {
355                         if (target == MSG_ALL_BUT_SELF
356                             && i == smp_processor_id())
357                                 continue;
358                         iic_cause_IPI(i, msg);
359                 }
360         }
361 }
362
363 static int __init smp_iic_probe(void)
364 {
365         iic_request_IPIs();
366
367         return cpus_weight(cpu_possible_map);
368 }
369
370 static void __devinit smp_iic_setup_cpu(int cpu)
371 {
372         if (cpu != boot_cpuid)
373                 iic_setup_cpu();
374 }
375 #endif /* CONFIG_BPA_IIC */
376
377 static DEFINE_SPINLOCK(timebase_lock);
378 static unsigned long timebase = 0;
379
380 static void __devinit pSeries_give_timebase(void)
381 {
382         spin_lock(&timebase_lock);
383         rtas_call(rtas_token("freeze-time-base"), 0, 1, NULL);
384         timebase = get_tb();
385         spin_unlock(&timebase_lock);
386
387         while (timebase)
388                 barrier();
389         rtas_call(rtas_token("thaw-time-base"), 0, 1, NULL);
390 }
391
392 static void __devinit pSeries_take_timebase(void)
393 {
394         while (!timebase)
395                 barrier();
396         spin_lock(&timebase_lock);
397         set_tb(timebase >> 32, timebase & 0xffffffff);
398         timebase = 0;
399         spin_unlock(&timebase_lock);
400 }
401
402 static void __devinit smp_pSeries_kick_cpu(int nr)
403 {
404         BUG_ON(nr < 0 || nr >= NR_CPUS);
405
406         if (!smp_startup_cpu(nr))
407                 return;
408
409         /*
410          * The processor is currently spinning, waiting for the
411          * cpu_start field to become non-zero After we set cpu_start,
412          * the processor will continue on to secondary_start
413          */
414         paca[nr].cpu_start = 1;
415 }
416
417 static int smp_pSeries_cpu_bootable(unsigned int nr)
418 {
419         /* Special case - we inhibit secondary thread startup
420          * during boot if the user requests it.  Odd-numbered
421          * cpus are assumed to be secondary threads.
422          */
423         if (system_state < SYSTEM_RUNNING &&
424             cpu_has_feature(CPU_FTR_SMT) &&
425             !smt_enabled_at_boot && nr % 2 != 0)
426                 return 0;
427
428         return 1;
429 }
430 #ifdef CONFIG_MPIC
431 static struct smp_ops_t pSeries_mpic_smp_ops = {
432         .message_pass   = smp_mpic_message_pass,
433         .probe          = smp_mpic_probe,
434         .kick_cpu       = smp_pSeries_kick_cpu,
435         .setup_cpu      = smp_mpic_setup_cpu,
436 };
437 #endif
438 #ifdef CONFIG_XICS
439 static struct smp_ops_t pSeries_xics_smp_ops = {
440         .message_pass   = smp_xics_message_pass,
441         .probe          = smp_xics_probe,
442         .kick_cpu       = smp_pSeries_kick_cpu,
443         .setup_cpu      = smp_xics_setup_cpu,
444         .cpu_bootable   = smp_pSeries_cpu_bootable,
445 };
446 #endif
447 #ifdef CONFIG_BPA_IIC
448 static struct smp_ops_t bpa_iic_smp_ops = {
449         .message_pass   = smp_iic_message_pass,
450         .probe          = smp_iic_probe,
451         .kick_cpu       = smp_pSeries_kick_cpu,
452         .setup_cpu      = smp_iic_setup_cpu,
453         .cpu_bootable   = smp_pSeries_cpu_bootable,
454 };
455 #endif
456
457 /* This is called very early */
458 void __init smp_init_pSeries(void)
459 {
460         int i;
461
462         DBG(" -> smp_init_pSeries()\n");
463
464         switch (ppc64_interrupt_controller) {
465 #ifdef CONFIG_MPIC
466         case IC_OPEN_PIC:
467                 smp_ops = &pSeries_mpic_smp_ops;
468                 break;
469 #endif
470 #ifdef CONFIG_XICS
471         case IC_PPC_XIC:
472                 smp_ops = &pSeries_xics_smp_ops;
473                 break;
474 #endif
475 #ifdef CONFIG_BPA_IIC
476         case IC_BPA_IIC:
477                 smp_ops = &bpa_iic_smp_ops;
478                 break;
479 #endif
480         default:
481                 panic("Invalid interrupt controller");
482         }
483
484 #ifdef CONFIG_HOTPLUG_CPU
485         smp_ops->cpu_disable = pSeries_cpu_disable;
486         smp_ops->cpu_die = pSeries_cpu_die;
487
488         /* Processors can be added/removed only on LPAR */
489         if (systemcfg->platform == PLATFORM_PSERIES_LPAR)
490                 pSeries_reconfig_notifier_register(&pSeries_smp_nb);
491 #endif
492
493         /* Mark threads which are still spinning in hold loops. */
494         if (cpu_has_feature(CPU_FTR_SMT)) {
495                 for_each_present_cpu(i) { 
496                         if (i % 2 == 0)
497                                 /*
498                                  * Even-numbered logical cpus correspond to
499                                  * primary threads.
500                                  */
501                                 cpu_set(i, of_spin_map);
502                 }
503         } else {
504                 of_spin_map = cpu_present_map;
505         }
506
507         cpu_clear(boot_cpuid, of_spin_map);
508
509         /* Non-lpar has additional take/give timebase */
510         if (rtas_token("freeze-time-base") != RTAS_UNKNOWN_SERVICE) {
511                 smp_ops->give_timebase = pSeries_give_timebase;
512                 smp_ops->take_timebase = pSeries_take_timebase;
513         }
514
515         DBG(" <- smp_init_pSeries()\n");
516 }
517