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 <linux/of.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/xics.h>
33 #include "plpar_wrappers.h"
34 #include "offline_states.h"
35 
36 /* This version can't take the spinlock, because it never returns */
37 static struct rtas_args rtas_stop_self_args = {
38 	.token = RTAS_UNKNOWN_SERVICE,
39 	.nargs = 0,
40 	.nret = 1,
41 	.rets = &rtas_stop_self_args.args[0],
42 };
43 
44 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
45 							CPU_STATE_OFFLINE;
46 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
47 
48 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
49 
50 static int cede_offline_enabled __read_mostly = 1;
51 
52 /*
53  * Enable/disable cede_offline when available.
54  */
55 static int __init setup_cede_offline(char *str)
56 {
57 	if (!strcmp(str, "off"))
58 		cede_offline_enabled = 0;
59 	else if (!strcmp(str, "on"))
60 		cede_offline_enabled = 1;
61 	else
62 		return 0;
63 	return 1;
64 }
65 
66 __setup("cede_offline=", setup_cede_offline);
67 
68 enum cpu_state_vals get_cpu_current_state(int cpu)
69 {
70 	return per_cpu(current_state, cpu);
71 }
72 
73 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
74 {
75 	per_cpu(current_state, cpu) = state;
76 }
77 
78 enum cpu_state_vals get_preferred_offline_state(int cpu)
79 {
80 	return per_cpu(preferred_offline_state, cpu);
81 }
82 
83 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
84 {
85 	per_cpu(preferred_offline_state, cpu) = state;
86 }
87 
88 void set_default_offline_state(int cpu)
89 {
90 	per_cpu(preferred_offline_state, cpu) = default_offline_state;
91 }
92 
93 static void rtas_stop_self(void)
94 {
95 	struct rtas_args *args = &rtas_stop_self_args;
96 
97 	local_irq_disable();
98 
99 	BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
100 
101 	printk("cpu %u (hwid %u) Ready to die...\n",
102 	       smp_processor_id(), hard_smp_processor_id());
103 	enter_rtas(__pa(args));
104 
105 	panic("Alas, I survived.\n");
106 }
107 
108 static void pseries_mach_cpu_die(void)
109 {
110 	unsigned int cpu = smp_processor_id();
111 	unsigned int hwcpu = hard_smp_processor_id();
112 	u8 cede_latency_hint = 0;
113 
114 	local_irq_disable();
115 	idle_task_exit();
116 	xics_teardown_cpu();
117 
118 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
119 		set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
120 		if (ppc_md.suspend_disable_cpu)
121 			ppc_md.suspend_disable_cpu();
122 
123 		cede_latency_hint = 2;
124 
125 		get_lppaca()->idle = 1;
126 		if (!get_lppaca()->shared_proc)
127 			get_lppaca()->donate_dedicated_cpu = 1;
128 
129 		while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
130 			while (!prep_irq_for_idle()) {
131 				local_irq_enable();
132 				local_irq_disable();
133 			}
134 
135 			extended_cede_processor(cede_latency_hint);
136 		}
137 
138 		local_irq_disable();
139 
140 		if (!get_lppaca()->shared_proc)
141 			get_lppaca()->donate_dedicated_cpu = 0;
142 		get_lppaca()->idle = 0;
143 
144 		if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
145 			unregister_slb_shadow(hwcpu);
146 
147 			hard_irq_disable();
148 			/*
149 			 * Call to start_secondary_resume() will not return.
150 			 * Kernel stack will be reset and start_secondary()
151 			 * will be called to continue the online operation.
152 			 */
153 			start_secondary_resume();
154 		}
155 	}
156 
157 	/* Requested state is CPU_STATE_OFFLINE at this point */
158 	WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
159 
160 	set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
161 	unregister_slb_shadow(hwcpu);
162 	rtas_stop_self();
163 
164 	/* Should never get here... */
165 	BUG();
166 	for(;;);
167 }
168 
169 static int pseries_cpu_disable(void)
170 {
171 	int cpu = smp_processor_id();
172 
173 	set_cpu_online(cpu, false);
174 	vdso_data->processorCount--;
175 
176 	/*fix boot_cpuid here*/
177 	if (cpu == boot_cpuid)
178 		boot_cpuid = cpumask_any(cpu_online_mask);
179 
180 	/* FIXME: abstract this to not be platform specific later on */
181 	xics_migrate_irqs_away();
182 	return 0;
183 }
184 
185 /*
186  * pseries_cpu_die: Wait for the cpu to die.
187  * @cpu: logical processor id of the CPU whose death we're awaiting.
188  *
189  * This function is called from the context of the thread which is performing
190  * the cpu-offline. Here we wait for long enough to allow the cpu in question
191  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
192  * notifications.
193  *
194  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
195  * self-destruct.
196  */
197 static void pseries_cpu_die(unsigned int cpu)
198 {
199 	int tries;
200 	int cpu_status = 1;
201 	unsigned int pcpu = get_hard_smp_processor_id(cpu);
202 
203 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
204 		cpu_status = 1;
205 		for (tries = 0; tries < 5000; tries++) {
206 			if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
207 				cpu_status = 0;
208 				break;
209 			}
210 			msleep(1);
211 		}
212 	} else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
213 
214 		for (tries = 0; tries < 25; tries++) {
215 			cpu_status = smp_query_cpu_stopped(pcpu);
216 			if (cpu_status == QCSS_STOPPED ||
217 			    cpu_status == QCSS_HARDWARE_ERROR)
218 				break;
219 			cpu_relax();
220 		}
221 	}
222 
223 	if (cpu_status != 0) {
224 		printk("Querying DEAD? cpu %i (%i) shows %i\n",
225 		       cpu, pcpu, cpu_status);
226 	}
227 
228 	/* Isolation and deallocation are definitely done by
229 	 * drslot_chrp_cpu.  If they were not they would be
230 	 * done here.  Change isolate state to Isolate and
231 	 * change allocation-state to Unusable.
232 	 */
233 	paca[cpu].cpu_start = 0;
234 }
235 
236 /*
237  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
238  * here is that a cpu device node may represent up to two logical cpus
239  * in the SMT case.  We must honor the assumption in other code that
240  * the logical ids for sibling SMT threads x and y are adjacent, such
241  * that x^1 == y and y^1 == x.
242  */
243 static int pseries_add_processor(struct device_node *np)
244 {
245 	unsigned int cpu;
246 	cpumask_var_t candidate_mask, tmp;
247 	int err = -ENOSPC, len, nthreads, i;
248 	const u32 *intserv;
249 
250 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
251 	if (!intserv)
252 		return 0;
253 
254 	zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
255 	zalloc_cpumask_var(&tmp, GFP_KERNEL);
256 
257 	nthreads = len / sizeof(u32);
258 	for (i = 0; i < nthreads; i++)
259 		cpumask_set_cpu(i, tmp);
260 
261 	cpu_maps_update_begin();
262 
263 	BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
264 
265 	/* Get a bitmap of unoccupied slots. */
266 	cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
267 	if (cpumask_empty(candidate_mask)) {
268 		/* If we get here, it most likely means that NR_CPUS is
269 		 * less than the partition's max processors setting.
270 		 */
271 		printk(KERN_ERR "Cannot add cpu %s; this system configuration"
272 		       " supports %d logical cpus.\n", np->full_name,
273 		       cpumask_weight(cpu_possible_mask));
274 		goto out_unlock;
275 	}
276 
277 	while (!cpumask_empty(tmp))
278 		if (cpumask_subset(tmp, candidate_mask))
279 			/* Found a range where we can insert the new cpu(s) */
280 			break;
281 		else
282 			cpumask_shift_left(tmp, tmp, nthreads);
283 
284 	if (cpumask_empty(tmp)) {
285 		printk(KERN_ERR "Unable to find space in cpu_present_mask for"
286 		       " processor %s with %d thread(s)\n", np->name,
287 		       nthreads);
288 		goto out_unlock;
289 	}
290 
291 	for_each_cpu(cpu, tmp) {
292 		BUG_ON(cpu_present(cpu));
293 		set_cpu_present(cpu, true);
294 		set_hard_smp_processor_id(cpu, *intserv++);
295 	}
296 	err = 0;
297 out_unlock:
298 	cpu_maps_update_done();
299 	free_cpumask_var(candidate_mask);
300 	free_cpumask_var(tmp);
301 	return err;
302 }
303 
304 /*
305  * Update the present map for a cpu node which is going away, and set
306  * the hard id in the paca(s) to -1 to be consistent with boot time
307  * convention for non-present cpus.
308  */
309 static void pseries_remove_processor(struct device_node *np)
310 {
311 	unsigned int cpu;
312 	int len, nthreads, i;
313 	const u32 *intserv;
314 
315 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
316 	if (!intserv)
317 		return;
318 
319 	nthreads = len / sizeof(u32);
320 
321 	cpu_maps_update_begin();
322 	for (i = 0; i < nthreads; i++) {
323 		for_each_present_cpu(cpu) {
324 			if (get_hard_smp_processor_id(cpu) != intserv[i])
325 				continue;
326 			BUG_ON(cpu_online(cpu));
327 			set_cpu_present(cpu, false);
328 			set_hard_smp_processor_id(cpu, -1);
329 			break;
330 		}
331 		if (cpu >= nr_cpu_ids)
332 			printk(KERN_WARNING "Could not find cpu to remove "
333 			       "with physical id 0x%x\n", intserv[i]);
334 	}
335 	cpu_maps_update_done();
336 }
337 
338 static int pseries_smp_notifier(struct notifier_block *nb,
339 				unsigned long action, void *node)
340 {
341 	int err = 0;
342 
343 	switch (action) {
344 	case OF_RECONFIG_ATTACH_NODE:
345 		err = pseries_add_processor(node);
346 		break;
347 	case OF_RECONFIG_DETACH_NODE:
348 		pseries_remove_processor(node);
349 		break;
350 	}
351 	return notifier_from_errno(err);
352 }
353 
354 static struct notifier_block pseries_smp_nb = {
355 	.notifier_call = pseries_smp_notifier,
356 };
357 
358 #define MAX_CEDE_LATENCY_LEVELS		4
359 #define	CEDE_LATENCY_PARAM_LENGTH	10
360 #define CEDE_LATENCY_PARAM_MAX_LENGTH	\
361 	(MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
362 #define CEDE_LATENCY_TOKEN		45
363 
364 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
365 
366 static int parse_cede_parameters(void)
367 {
368 	memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
369 	return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
370 			 NULL,
371 			 CEDE_LATENCY_TOKEN,
372 			 __pa(cede_parameters),
373 			 CEDE_LATENCY_PARAM_MAX_LENGTH);
374 }
375 
376 static int __init pseries_cpu_hotplug_init(void)
377 {
378 	struct device_node *np;
379 	const char *typep;
380 	int cpu;
381 	int qcss_tok;
382 
383 	for_each_node_by_name(np, "interrupt-controller") {
384 		typep = of_get_property(np, "compatible", NULL);
385 		if (strstr(typep, "open-pic")) {
386 			of_node_put(np);
387 
388 			printk(KERN_INFO "CPU Hotplug not supported on "
389 				"systems using MPIC\n");
390 			return 0;
391 		}
392 	}
393 
394 	rtas_stop_self_args.token = rtas_token("stop-self");
395 	qcss_tok = rtas_token("query-cpu-stopped-state");
396 
397 	if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
398 			qcss_tok == RTAS_UNKNOWN_SERVICE) {
399 		printk(KERN_INFO "CPU Hotplug not supported by firmware "
400 				"- disabling.\n");
401 		return 0;
402 	}
403 
404 	ppc_md.cpu_die = pseries_mach_cpu_die;
405 	smp_ops->cpu_disable = pseries_cpu_disable;
406 	smp_ops->cpu_die = pseries_cpu_die;
407 
408 	/* Processors can be added/removed only on LPAR */
409 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
410 		of_reconfig_notifier_register(&pseries_smp_nb);
411 		cpu_maps_update_begin();
412 		if (cede_offline_enabled && parse_cede_parameters() == 0) {
413 			default_offline_state = CPU_STATE_INACTIVE;
414 			for_each_online_cpu(cpu)
415 				set_default_offline_state(cpu);
416 		}
417 		cpu_maps_update_done();
418 	}
419 
420 	return 0;
421 }
422 arch_initcall(pseries_cpu_hotplug_init);
423