xref: /linux/arch/powerpc/platforms/pseries/hotplug-cpu.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
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/interrupt.h>
23 #include <linux/delay.h>
24 #include <linux/sched.h>	/* for idle_task_exit */
25 #include <linux/cpu.h>
26 #include <asm/system.h>
27 #include <asm/prom.h>
28 #include <asm/rtas.h>
29 #include <asm/firmware.h>
30 #include <asm/machdep.h>
31 #include <asm/vdso_datapage.h>
32 #include <asm/pSeries_reconfig.h>
33 #include <asm/xics.h>
34 #include "plpar_wrappers.h"
35 #include "offline_states.h"
36 
37 /* This version can't take the spinlock, because it never returns */
38 static struct rtas_args rtas_stop_self_args = {
39 	.token = RTAS_UNKNOWN_SERVICE,
40 	.nargs = 0,
41 	.nret = 1,
42 	.rets = &rtas_stop_self_args.args[0],
43 };
44 
45 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
46 							CPU_STATE_OFFLINE;
47 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
48 
49 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
50 
51 static int cede_offline_enabled __read_mostly = 1;
52 
53 /*
54  * Enable/disable cede_offline when available.
55  */
56 static int __init setup_cede_offline(char *str)
57 {
58 	if (!strcmp(str, "off"))
59 		cede_offline_enabled = 0;
60 	else if (!strcmp(str, "on"))
61 		cede_offline_enabled = 1;
62 	else
63 		return 0;
64 	return 1;
65 }
66 
67 __setup("cede_offline=", setup_cede_offline);
68 
69 enum cpu_state_vals get_cpu_current_state(int cpu)
70 {
71 	return per_cpu(current_state, cpu);
72 }
73 
74 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
75 {
76 	per_cpu(current_state, cpu) = state;
77 }
78 
79 enum cpu_state_vals get_preferred_offline_state(int cpu)
80 {
81 	return per_cpu(preferred_offline_state, cpu);
82 }
83 
84 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
85 {
86 	per_cpu(preferred_offline_state, cpu) = state;
87 }
88 
89 void set_default_offline_state(int cpu)
90 {
91 	per_cpu(preferred_offline_state, cpu) = default_offline_state;
92 }
93 
94 static void rtas_stop_self(void)
95 {
96 	struct rtas_args *args = &rtas_stop_self_args;
97 
98 	local_irq_disable();
99 
100 	BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
101 
102 	printk("cpu %u (hwid %u) Ready to die...\n",
103 	       smp_processor_id(), hard_smp_processor_id());
104 	enter_rtas(__pa(args));
105 
106 	panic("Alas, I survived.\n");
107 }
108 
109 static void pseries_mach_cpu_die(void)
110 {
111 	unsigned int cpu = smp_processor_id();
112 	unsigned int hwcpu = hard_smp_processor_id();
113 	u8 cede_latency_hint = 0;
114 
115 	local_irq_disable();
116 	idle_task_exit();
117 	xics_teardown_cpu();
118 
119 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
120 		set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
121 		if (ppc_md.suspend_disable_cpu)
122 			ppc_md.suspend_disable_cpu();
123 
124 		cede_latency_hint = 2;
125 
126 		get_lppaca()->idle = 1;
127 		if (!get_lppaca()->shared_proc)
128 			get_lppaca()->donate_dedicated_cpu = 1;
129 
130 		while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
131 			extended_cede_processor(cede_latency_hint);
132 		}
133 
134 		if (!get_lppaca()->shared_proc)
135 			get_lppaca()->donate_dedicated_cpu = 0;
136 		get_lppaca()->idle = 0;
137 
138 		if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
139 			unregister_slb_shadow(hwcpu);
140 
141 			/*
142 			 * Call to start_secondary_resume() will not return.
143 			 * Kernel stack will be reset and start_secondary()
144 			 * will be called to continue the online operation.
145 			 */
146 			start_secondary_resume();
147 		}
148 	}
149 
150 	/* Requested state is CPU_STATE_OFFLINE at this point */
151 	WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
152 
153 	set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
154 	unregister_slb_shadow(hwcpu);
155 	rtas_stop_self();
156 
157 	/* Should never get here... */
158 	BUG();
159 	for(;;);
160 }
161 
162 static int pseries_cpu_disable(void)
163 {
164 	int cpu = smp_processor_id();
165 
166 	set_cpu_online(cpu, false);
167 	vdso_data->processorCount--;
168 
169 	/*fix boot_cpuid here*/
170 	if (cpu == boot_cpuid)
171 		boot_cpuid = cpumask_any(cpu_online_mask);
172 
173 	/* FIXME: abstract this to not be platform specific later on */
174 	xics_migrate_irqs_away();
175 	return 0;
176 }
177 
178 /*
179  * pseries_cpu_die: Wait for the cpu to die.
180  * @cpu: logical processor id of the CPU whose death we're awaiting.
181  *
182  * This function is called from the context of the thread which is performing
183  * the cpu-offline. Here we wait for long enough to allow the cpu in question
184  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
185  * notifications.
186  *
187  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
188  * self-destruct.
189  */
190 static void pseries_cpu_die(unsigned int cpu)
191 {
192 	int tries;
193 	int cpu_status = 1;
194 	unsigned int pcpu = get_hard_smp_processor_id(cpu);
195 
196 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
197 		cpu_status = 1;
198 		for (tries = 0; tries < 5000; tries++) {
199 			if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
200 				cpu_status = 0;
201 				break;
202 			}
203 			msleep(1);
204 		}
205 	} else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
206 
207 		for (tries = 0; tries < 25; tries++) {
208 			cpu_status = smp_query_cpu_stopped(pcpu);
209 			if (cpu_status == QCSS_STOPPED ||
210 			    cpu_status == QCSS_HARDWARE_ERROR)
211 				break;
212 			cpu_relax();
213 		}
214 	}
215 
216 	if (cpu_status != 0) {
217 		printk("Querying DEAD? cpu %i (%i) shows %i\n",
218 		       cpu, pcpu, cpu_status);
219 	}
220 
221 	/* Isolation and deallocation are definitely done by
222 	 * drslot_chrp_cpu.  If they were not they would be
223 	 * done here.  Change isolate state to Isolate and
224 	 * change allocation-state to Unusable.
225 	 */
226 	paca[cpu].cpu_start = 0;
227 }
228 
229 /*
230  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
231  * here is that a cpu device node may represent up to two logical cpus
232  * in the SMT case.  We must honor the assumption in other code that
233  * the logical ids for sibling SMT threads x and y are adjacent, such
234  * that x^1 == y and y^1 == x.
235  */
236 static int pseries_add_processor(struct device_node *np)
237 {
238 	unsigned int cpu;
239 	cpumask_var_t candidate_mask, tmp;
240 	int err = -ENOSPC, len, nthreads, i;
241 	const u32 *intserv;
242 
243 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
244 	if (!intserv)
245 		return 0;
246 
247 	zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
248 	zalloc_cpumask_var(&tmp, GFP_KERNEL);
249 
250 	nthreads = len / sizeof(u32);
251 	for (i = 0; i < nthreads; i++)
252 		cpumask_set_cpu(i, tmp);
253 
254 	cpu_maps_update_begin();
255 
256 	BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
257 
258 	/* Get a bitmap of unoccupied slots. */
259 	cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
260 	if (cpumask_empty(candidate_mask)) {
261 		/* If we get here, it most likely means that NR_CPUS is
262 		 * less than the partition's max processors setting.
263 		 */
264 		printk(KERN_ERR "Cannot add cpu %s; this system configuration"
265 		       " supports %d logical cpus.\n", np->full_name,
266 		       cpumask_weight(cpu_possible_mask));
267 		goto out_unlock;
268 	}
269 
270 	while (!cpumask_empty(tmp))
271 		if (cpumask_subset(tmp, candidate_mask))
272 			/* Found a range where we can insert the new cpu(s) */
273 			break;
274 		else
275 			cpumask_shift_left(tmp, tmp, nthreads);
276 
277 	if (cpumask_empty(tmp)) {
278 		printk(KERN_ERR "Unable to find space in cpu_present_mask for"
279 		       " processor %s with %d thread(s)\n", np->name,
280 		       nthreads);
281 		goto out_unlock;
282 	}
283 
284 	for_each_cpu(cpu, tmp) {
285 		BUG_ON(cpu_present(cpu));
286 		set_cpu_present(cpu, true);
287 		set_hard_smp_processor_id(cpu, *intserv++);
288 	}
289 	err = 0;
290 out_unlock:
291 	cpu_maps_update_done();
292 	free_cpumask_var(candidate_mask);
293 	free_cpumask_var(tmp);
294 	return err;
295 }
296 
297 /*
298  * Update the present map for a cpu node which is going away, and set
299  * the hard id in the paca(s) to -1 to be consistent with boot time
300  * convention for non-present cpus.
301  */
302 static void pseries_remove_processor(struct device_node *np)
303 {
304 	unsigned int cpu;
305 	int len, nthreads, i;
306 	const u32 *intserv;
307 
308 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
309 	if (!intserv)
310 		return;
311 
312 	nthreads = len / sizeof(u32);
313 
314 	cpu_maps_update_begin();
315 	for (i = 0; i < nthreads; i++) {
316 		for_each_present_cpu(cpu) {
317 			if (get_hard_smp_processor_id(cpu) != intserv[i])
318 				continue;
319 			BUG_ON(cpu_online(cpu));
320 			set_cpu_present(cpu, false);
321 			set_hard_smp_processor_id(cpu, -1);
322 			break;
323 		}
324 		if (cpu >= nr_cpu_ids)
325 			printk(KERN_WARNING "Could not find cpu to remove "
326 			       "with physical id 0x%x\n", intserv[i]);
327 	}
328 	cpu_maps_update_done();
329 }
330 
331 static int pseries_smp_notifier(struct notifier_block *nb,
332 				unsigned long action, void *node)
333 {
334 	int err = 0;
335 
336 	switch (action) {
337 	case PSERIES_RECONFIG_ADD:
338 		err = pseries_add_processor(node);
339 		break;
340 	case PSERIES_RECONFIG_REMOVE:
341 		pseries_remove_processor(node);
342 		break;
343 	}
344 	return notifier_from_errno(err);
345 }
346 
347 static struct notifier_block pseries_smp_nb = {
348 	.notifier_call = pseries_smp_notifier,
349 };
350 
351 #define MAX_CEDE_LATENCY_LEVELS		4
352 #define	CEDE_LATENCY_PARAM_LENGTH	10
353 #define CEDE_LATENCY_PARAM_MAX_LENGTH	\
354 	(MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
355 #define CEDE_LATENCY_TOKEN		45
356 
357 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
358 
359 static int parse_cede_parameters(void)
360 {
361 	memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
362 	return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
363 			 NULL,
364 			 CEDE_LATENCY_TOKEN,
365 			 __pa(cede_parameters),
366 			 CEDE_LATENCY_PARAM_MAX_LENGTH);
367 }
368 
369 static int __init pseries_cpu_hotplug_init(void)
370 {
371 	struct device_node *np;
372 	const char *typep;
373 	int cpu;
374 	int qcss_tok;
375 
376 	for_each_node_by_name(np, "interrupt-controller") {
377 		typep = of_get_property(np, "compatible", NULL);
378 		if (strstr(typep, "open-pic")) {
379 			of_node_put(np);
380 
381 			printk(KERN_INFO "CPU Hotplug not supported on "
382 				"systems using MPIC\n");
383 			return 0;
384 		}
385 	}
386 
387 	rtas_stop_self_args.token = rtas_token("stop-self");
388 	qcss_tok = rtas_token("query-cpu-stopped-state");
389 
390 	if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
391 			qcss_tok == RTAS_UNKNOWN_SERVICE) {
392 		printk(KERN_INFO "CPU Hotplug not supported by firmware "
393 				"- disabling.\n");
394 		return 0;
395 	}
396 
397 	ppc_md.cpu_die = pseries_mach_cpu_die;
398 	smp_ops->cpu_disable = pseries_cpu_disable;
399 	smp_ops->cpu_die = pseries_cpu_die;
400 
401 	/* Processors can be added/removed only on LPAR */
402 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
403 		pSeries_reconfig_notifier_register(&pseries_smp_nb);
404 		cpu_maps_update_begin();
405 		if (cede_offline_enabled && parse_cede_parameters() == 0) {
406 			default_offline_state = CPU_STATE_INACTIVE;
407 			for_each_online_cpu(cpu)
408 				set_default_offline_state(cpu);
409 		}
410 		cpu_maps_update_done();
411 	}
412 
413 	return 0;
414 }
415 arch_initcall(pseries_cpu_hotplug_init);
416