powerpc/cpumask: Update some comments
[linux-2.6.git] / arch / powerpc / platforms / pseries / hotplug-cpu.c
1 /*
2  * pseries CPU Hotplug infrastructure.
3  *
4  * Split out from arch/powerpc/platforms/pseries/setup.c
5  *  arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
6  *
7  * Peter Bergner, IBM   March 2001.
8  * Copyright (C) 2001 IBM.
9  * Dave Engebretsen, Peter Bergner, and
10  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11  * Plus various changes from other IBM teams...
12  *
13  * Copyright (C) 2006 Michael Ellerman, IBM Corporation
14  *
15  *      This program is free software; you can redistribute it and/or
16  *      modify it under the terms of the GNU General Public License
17  *      as published by the Free Software Foundation; either version
18  *      2 of the License, or (at your option) any later version.
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/delay.h>
23 #include <linux/cpu.h>
24 #include <asm/system.h>
25 #include <asm/prom.h>
26 #include <asm/rtas.h>
27 #include <asm/firmware.h>
28 #include <asm/machdep.h>
29 #include <asm/vdso_datapage.h>
30 #include <asm/pSeries_reconfig.h>
31 #include "xics.h"
32 #include "plpar_wrappers.h"
33 #include "offline_states.h"
34
35 /* This version can't take the spinlock, because it never returns */
36 static struct rtas_args rtas_stop_self_args = {
37         .token = RTAS_UNKNOWN_SERVICE,
38         .nargs = 0,
39         .nret = 1,
40         .rets = &rtas_stop_self_args.args[0],
41 };
42
43 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
44                                                         CPU_STATE_OFFLINE;
45 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
46
47 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
48
49 static int cede_offline_enabled __read_mostly = 1;
50
51 /*
52  * Enable/disable cede_offline when available.
53  */
54 static int __init setup_cede_offline(char *str)
55 {
56         if (!strcmp(str, "off"))
57                 cede_offline_enabled = 0;
58         else if (!strcmp(str, "on"))
59                 cede_offline_enabled = 1;
60         else
61                 return 0;
62         return 1;
63 }
64
65 __setup("cede_offline=", setup_cede_offline);
66
67 enum cpu_state_vals get_cpu_current_state(int cpu)
68 {
69         return per_cpu(current_state, cpu);
70 }
71
72 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
73 {
74         per_cpu(current_state, cpu) = state;
75 }
76
77 enum cpu_state_vals get_preferred_offline_state(int cpu)
78 {
79         return per_cpu(preferred_offline_state, cpu);
80 }
81
82 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
83 {
84         per_cpu(preferred_offline_state, cpu) = state;
85 }
86
87 void set_default_offline_state(int cpu)
88 {
89         per_cpu(preferred_offline_state, cpu) = default_offline_state;
90 }
91
92 static void rtas_stop_self(void)
93 {
94         struct rtas_args *args = &rtas_stop_self_args;
95
96         local_irq_disable();
97
98         BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
99
100         printk("cpu %u (hwid %u) Ready to die...\n",
101                smp_processor_id(), hard_smp_processor_id());
102         enter_rtas(__pa(args));
103
104         panic("Alas, I survived.\n");
105 }
106
107 static void pseries_mach_cpu_die(void)
108 {
109         unsigned int cpu = smp_processor_id();
110         unsigned int hwcpu = hard_smp_processor_id();
111         u8 cede_latency_hint = 0;
112
113         local_irq_disable();
114         idle_task_exit();
115         xics_teardown_cpu();
116
117         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
118                 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
119                 cede_latency_hint = 2;
120
121                 get_lppaca()->idle = 1;
122                 if (!get_lppaca()->shared_proc)
123                         get_lppaca()->donate_dedicated_cpu = 1;
124
125                 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
126                         extended_cede_processor(cede_latency_hint);
127                 }
128
129                 if (!get_lppaca()->shared_proc)
130                         get_lppaca()->donate_dedicated_cpu = 0;
131                 get_lppaca()->idle = 0;
132
133                 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
134                         unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
135
136                         /*
137                          * Call to start_secondary_resume() will not return.
138                          * Kernel stack will be reset and start_secondary()
139                          * will be called to continue the online operation.
140                          */
141                         start_secondary_resume();
142                 }
143         }
144
145         /* Requested state is CPU_STATE_OFFLINE at this point */
146         WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
147
148         set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
149         unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
150         rtas_stop_self();
151
152         /* Should never get here... */
153         BUG();
154         for(;;);
155 }
156
157 static int pseries_cpu_disable(void)
158 {
159         int cpu = smp_processor_id();
160
161         set_cpu_online(cpu, false);
162         vdso_data->processorCount--;
163
164         /*fix boot_cpuid here*/
165         if (cpu == boot_cpuid)
166                 boot_cpuid = cpumask_any(cpu_online_mask);
167
168         /* FIXME: abstract this to not be platform specific later on */
169         xics_migrate_irqs_away();
170         return 0;
171 }
172
173 /*
174  * pseries_cpu_die: Wait for the cpu to die.
175  * @cpu: logical processor id of the CPU whose death we're awaiting.
176  *
177  * This function is called from the context of the thread which is performing
178  * the cpu-offline. Here we wait for long enough to allow the cpu in question
179  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
180  * notifications.
181  *
182  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
183  * self-destruct.
184  */
185 static void pseries_cpu_die(unsigned int cpu)
186 {
187         int tries;
188         int cpu_status = 1;
189         unsigned int pcpu = get_hard_smp_processor_id(cpu);
190
191         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
192                 cpu_status = 1;
193                 for (tries = 0; tries < 1000; tries++) {
194                         if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
195                                 cpu_status = 0;
196                                 break;
197                         }
198                         cpu_relax();
199                 }
200         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
201
202                 for (tries = 0; tries < 25; tries++) {
203                         cpu_status = smp_query_cpu_stopped(pcpu);
204                         if (cpu_status == QCSS_STOPPED ||
205                             cpu_status == QCSS_HARDWARE_ERROR)
206                                 break;
207                         cpu_relax();
208                 }
209         }
210
211         if (cpu_status != 0) {
212                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
213                        cpu, pcpu, cpu_status);
214         }
215
216         /* Isolation and deallocation are definatly done by
217          * drslot_chrp_cpu.  If they were not they would be
218          * done here.  Change isolate state to Isolate and
219          * change allocation-state to Unusable.
220          */
221         paca[cpu].cpu_start = 0;
222 }
223
224 /*
225  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
226  * here is that a cpu device node may represent up to two logical cpus
227  * in the SMT case.  We must honor the assumption in other code that
228  * the logical ids for sibling SMT threads x and y are adjacent, such
229  * that x^1 == y and y^1 == x.
230  */
231 static int pseries_add_processor(struct device_node *np)
232 {
233         unsigned int cpu;
234         cpumask_var_t candidate_mask, tmp;
235         int err = -ENOSPC, len, nthreads, i;
236         const u32 *intserv;
237
238         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
239         if (!intserv)
240                 return 0;
241
242         zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
243         zalloc_cpumask_var(&tmp, GFP_KERNEL);
244
245         nthreads = len / sizeof(u32);
246         for (i = 0; i < nthreads; i++)
247                 cpumask_set_cpu(i, tmp);
248
249         cpu_maps_update_begin();
250
251         BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
252
253         /* Get a bitmap of unoccupied slots. */
254         cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
255         if (cpumask_empty(candidate_mask)) {
256                 /* If we get here, it most likely means that NR_CPUS is
257                  * less than the partition's max processors setting.
258                  */
259                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
260                        " supports %d logical cpus.\n", np->full_name,
261                        cpumask_weight(cpu_possible_mask));
262                 goto out_unlock;
263         }
264
265         while (!cpumask_empty(tmp))
266                 if (cpumask_subset(tmp, candidate_mask))
267                         /* Found a range where we can insert the new cpu(s) */
268                         break;
269                 else
270                         cpumask_shift_left(tmp, tmp, nthreads);
271
272         if (cpumask_empty(tmp)) {
273                 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
274                        " processor %s with %d thread(s)\n", np->name,
275                        nthreads);
276                 goto out_unlock;
277         }
278
279         for_each_cpu(cpu, tmp) {
280                 BUG_ON(cpumask_test_cpu(cpu, cpu_present_mask));
281                 set_cpu_present(cpu, true);
282                 set_hard_smp_processor_id(cpu, *intserv++);
283         }
284         err = 0;
285 out_unlock:
286         cpu_maps_update_done();
287         free_cpumask_var(candidate_mask);
288         free_cpumask_var(tmp);
289         return err;
290 }
291
292 /*
293  * Update the present map for a cpu node which is going away, and set
294  * the hard id in the paca(s) to -1 to be consistent with boot time
295  * convention for non-present cpus.
296  */
297 static void pseries_remove_processor(struct device_node *np)
298 {
299         unsigned int cpu;
300         int len, nthreads, i;
301         const u32 *intserv;
302
303         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
304         if (!intserv)
305                 return;
306
307         nthreads = len / sizeof(u32);
308
309         cpu_maps_update_begin();
310         for (i = 0; i < nthreads; i++) {
311                 for_each_present_cpu(cpu) {
312                         if (get_hard_smp_processor_id(cpu) != intserv[i])
313                                 continue;
314                         BUG_ON(cpu_online(cpu));
315                         set_cpu_present(cpu, false);
316                         set_hard_smp_processor_id(cpu, -1);
317                         break;
318                 }
319                 if (cpu >= nr_cpu_ids)
320                         printk(KERN_WARNING "Could not find cpu to remove "
321                                "with physical id 0x%x\n", intserv[i]);
322         }
323         cpu_maps_update_done();
324 }
325
326 static int pseries_smp_notifier(struct notifier_block *nb,
327                                 unsigned long action, void *node)
328 {
329         int err = NOTIFY_OK;
330
331         switch (action) {
332         case PSERIES_RECONFIG_ADD:
333                 if (pseries_add_processor(node))
334                         err = NOTIFY_BAD;
335                 break;
336         case PSERIES_RECONFIG_REMOVE:
337                 pseries_remove_processor(node);
338                 break;
339         default:
340                 err = NOTIFY_DONE;
341                 break;
342         }
343         return err;
344 }
345
346 static struct notifier_block pseries_smp_nb = {
347         .notifier_call = pseries_smp_notifier,
348 };
349
350 #define MAX_CEDE_LATENCY_LEVELS         4
351 #define CEDE_LATENCY_PARAM_LENGTH       10
352 #define CEDE_LATENCY_PARAM_MAX_LENGTH   \
353         (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
354 #define CEDE_LATENCY_TOKEN              45
355
356 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
357
358 static int parse_cede_parameters(void)
359 {
360         memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
361         return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
362                          NULL,
363                          CEDE_LATENCY_TOKEN,
364                          __pa(cede_parameters),
365                          CEDE_LATENCY_PARAM_MAX_LENGTH);
366 }
367
368 static int __init pseries_cpu_hotplug_init(void)
369 {
370         struct device_node *np;
371         const char *typep;
372         int cpu;
373         int qcss_tok;
374
375         for_each_node_by_name(np, "interrupt-controller") {
376                 typep = of_get_property(np, "compatible", NULL);
377                 if (strstr(typep, "open-pic")) {
378                         of_node_put(np);
379
380                         printk(KERN_INFO "CPU Hotplug not supported on "
381                                 "systems using MPIC\n");
382                         return 0;
383                 }
384         }
385
386         rtas_stop_self_args.token = rtas_token("stop-self");
387         qcss_tok = rtas_token("query-cpu-stopped-state");
388
389         if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
390                         qcss_tok == RTAS_UNKNOWN_SERVICE) {
391                 printk(KERN_INFO "CPU Hotplug not supported by firmware "
392                                 "- disabling.\n");
393                 return 0;
394         }
395
396         ppc_md.cpu_die = pseries_mach_cpu_die;
397         smp_ops->cpu_disable = pseries_cpu_disable;
398         smp_ops->cpu_die = pseries_cpu_die;
399
400         /* Processors can be added/removed only on LPAR */
401         if (firmware_has_feature(FW_FEATURE_LPAR)) {
402                 pSeries_reconfig_notifier_register(&pseries_smp_nb);
403                 cpu_maps_update_begin();
404                 if (cede_offline_enabled && parse_cede_parameters() == 0) {
405                         default_offline_state = CPU_STATE_INACTIVE;
406                         for_each_online_cpu(cpu)
407                                 set_default_offline_state(cpu);
408                 }
409                 cpu_maps_update_done();
410         }
411
412         return 0;
413 }
414 arch_initcall(pseries_cpu_hotplug_init);