Merge HEAD from master.kernel.org:/pub/scm/linux/kernel/git/paulus/ppc64-2.6
[linux-2.6] / 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
276         if (cpu_isset(lcpu, of_spin_map))
277                 /* Already started by OF and sitting in spin loop */
278                 return 1;
279
280         pcpu = get_hard_smp_processor_id(lcpu);
281
282         /* Fixup atomic count: it exited inside IRQ handler. */
283         paca[lcpu].__current->thread_info->preempt_count        = 0;
284
285         status = rtas_call(rtas_token("start-cpu"), 3, 1, NULL,
286                            pcpu, start_here, lcpu);
287         if (status != 0) {
288                 printk(KERN_ERR "start-cpu failed: %i\n", status);
289                 return 0;
290         }
291         return 1;
292 }
293
294 #ifdef CONFIG_XICS
295 static inline void smp_xics_do_message(int cpu, int msg)
296 {
297         set_bit(msg, &xics_ipi_message[cpu].value);
298         mb();
299         xics_cause_IPI(cpu);
300 }
301
302 static void smp_xics_message_pass(int target, int msg)
303 {
304         unsigned int i;
305
306         if (target < NR_CPUS) {
307                 smp_xics_do_message(target, msg);
308         } else {
309                 for_each_online_cpu(i) {
310                         if (target == MSG_ALL_BUT_SELF
311                             && i == smp_processor_id())
312                                 continue;
313                         smp_xics_do_message(i, msg);
314                 }
315         }
316 }
317
318 static int __init smp_xics_probe(void)
319 {
320         xics_request_IPIs();
321
322         return cpus_weight(cpu_possible_map);
323 }
324
325 static void __devinit smp_xics_setup_cpu(int cpu)
326 {
327         if (cpu != boot_cpuid)
328                 xics_setup_cpu();
329
330         if (firmware_has_feature(FW_FEATURE_SPLPAR))
331                 vpa_init(cpu);
332
333         cpu_clear(cpu, of_spin_map);
334
335 }
336 #endif /* CONFIG_XICS */
337 #ifdef CONFIG_BPA_IIC
338 static void smp_iic_message_pass(int target, int msg)
339 {
340         unsigned int i;
341
342         if (target < NR_CPUS) {
343                 iic_cause_IPI(target, msg);
344         } else {
345                 for_each_online_cpu(i) {
346                         if (target == MSG_ALL_BUT_SELF
347                             && i == smp_processor_id())
348                                 continue;
349                         iic_cause_IPI(i, msg);
350                 }
351         }
352 }
353
354 static int __init smp_iic_probe(void)
355 {
356         iic_request_IPIs();
357
358         return cpus_weight(cpu_possible_map);
359 }
360
361 static void __devinit smp_iic_setup_cpu(int cpu)
362 {
363         if (cpu != boot_cpuid)
364                 iic_setup_cpu();
365 }
366 #endif /* CONFIG_BPA_IIC */
367
368 static DEFINE_SPINLOCK(timebase_lock);
369 static unsigned long timebase = 0;
370
371 static void __devinit pSeries_give_timebase(void)
372 {
373         spin_lock(&timebase_lock);
374         rtas_call(rtas_token("freeze-time-base"), 0, 1, NULL);
375         timebase = get_tb();
376         spin_unlock(&timebase_lock);
377
378         while (timebase)
379                 barrier();
380         rtas_call(rtas_token("thaw-time-base"), 0, 1, NULL);
381 }
382
383 static void __devinit pSeries_take_timebase(void)
384 {
385         while (!timebase)
386                 barrier();
387         spin_lock(&timebase_lock);
388         set_tb(timebase >> 32, timebase & 0xffffffff);
389         timebase = 0;
390         spin_unlock(&timebase_lock);
391 }
392
393 static void __devinit smp_pSeries_kick_cpu(int nr)
394 {
395         BUG_ON(nr < 0 || nr >= NR_CPUS);
396
397         if (!smp_startup_cpu(nr))
398                 return;
399
400         /*
401          * The processor is currently spinning, waiting for the
402          * cpu_start field to become non-zero After we set cpu_start,
403          * the processor will continue on to secondary_start
404          */
405         paca[nr].cpu_start = 1;
406 }
407
408 static int smp_pSeries_cpu_bootable(unsigned int nr)
409 {
410         /* Special case - we inhibit secondary thread startup
411          * during boot if the user requests it.  Odd-numbered
412          * cpus are assumed to be secondary threads.
413          */
414         if (system_state < SYSTEM_RUNNING &&
415             cpu_has_feature(CPU_FTR_SMT) &&
416             !smt_enabled_at_boot && nr % 2 != 0)
417                 return 0;
418
419         return 1;
420 }
421 #ifdef CONFIG_MPIC
422 static struct smp_ops_t pSeries_mpic_smp_ops = {
423         .message_pass   = smp_mpic_message_pass,
424         .probe          = smp_mpic_probe,
425         .kick_cpu       = smp_pSeries_kick_cpu,
426         .setup_cpu      = smp_mpic_setup_cpu,
427 };
428 #endif
429 #ifdef CONFIG_XICS
430 static struct smp_ops_t pSeries_xics_smp_ops = {
431         .message_pass   = smp_xics_message_pass,
432         .probe          = smp_xics_probe,
433         .kick_cpu       = smp_pSeries_kick_cpu,
434         .setup_cpu      = smp_xics_setup_cpu,
435         .cpu_bootable   = smp_pSeries_cpu_bootable,
436 };
437 #endif
438 #ifdef CONFIG_BPA_IIC
439 static struct smp_ops_t bpa_iic_smp_ops = {
440         .message_pass   = smp_iic_message_pass,
441         .probe          = smp_iic_probe,
442         .kick_cpu       = smp_pSeries_kick_cpu,
443         .setup_cpu      = smp_iic_setup_cpu,
444         .cpu_bootable   = smp_pSeries_cpu_bootable,
445 };
446 #endif
447
448 /* This is called very early */
449 void __init smp_init_pSeries(void)
450 {
451         int i;
452
453         DBG(" -> smp_init_pSeries()\n");
454
455         switch (ppc64_interrupt_controller) {
456 #ifdef CONFIG_MPIC
457         case IC_OPEN_PIC:
458                 smp_ops = &pSeries_mpic_smp_ops;
459                 break;
460 #endif
461 #ifdef CONFIG_XICS
462         case IC_PPC_XIC:
463                 smp_ops = &pSeries_xics_smp_ops;
464                 break;
465 #endif
466 #ifdef CONFIG_BPA_IIC
467         case IC_BPA_IIC:
468                 smp_ops = &bpa_iic_smp_ops;
469                 break;
470 #endif
471         default:
472                 panic("Invalid interrupt controller");
473         }
474
475 #ifdef CONFIG_HOTPLUG_CPU
476         smp_ops->cpu_disable = pSeries_cpu_disable;
477         smp_ops->cpu_die = pSeries_cpu_die;
478
479         /* Processors can be added/removed only on LPAR */
480         if (systemcfg->platform == PLATFORM_PSERIES_LPAR)
481                 pSeries_reconfig_notifier_register(&pSeries_smp_nb);
482 #endif
483
484         /* Mark threads which are still spinning in hold loops. */
485         if (cpu_has_feature(CPU_FTR_SMT)) {
486                 for_each_present_cpu(i) { 
487                         if (i % 2 == 0)
488                                 /*
489                                  * Even-numbered logical cpus correspond to
490                                  * primary threads.
491                                  */
492                                 cpu_set(i, of_spin_map);
493                 }
494         } else {
495                 of_spin_map = cpu_present_map;
496         }
497
498         cpu_clear(boot_cpuid, of_spin_map);
499
500         /* Non-lpar has additional take/give timebase */
501         if (rtas_token("freeze-time-base") != RTAS_UNKNOWN_SERVICE) {
502                 smp_ops->give_timebase = pSeries_give_timebase;
503                 smp_ops->take_timebase = pSeries_take_timebase;
504         }
505
506         DBG(" <- smp_init_pSeries()\n");
507 }
508