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 <asm/plpar_wrappers.h>
34 
35 #include "offline_states.h"
36 
37 /* This version can't take the spinlock, because it never returns */
38 static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
39 
40 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
41 							CPU_STATE_OFFLINE;
42 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
43 
44 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
45 
46 static int cede_offline_enabled __read_mostly = 1;
47 
48 /*
49  * Enable/disable cede_offline when available.
50  */
51 static int __init setup_cede_offline(char *str)
52 {
53 	if (!strcmp(str, "off"))
54 		cede_offline_enabled = 0;
55 	else if (!strcmp(str, "on"))
56 		cede_offline_enabled = 1;
57 	else
58 		return 0;
59 	return 1;
60 }
61 
62 __setup("cede_offline=", setup_cede_offline);
63 
64 enum cpu_state_vals get_cpu_current_state(int cpu)
65 {
66 	return per_cpu(current_state, cpu);
67 }
68 
69 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
70 {
71 	per_cpu(current_state, cpu) = state;
72 }
73 
74 enum cpu_state_vals get_preferred_offline_state(int cpu)
75 {
76 	return per_cpu(preferred_offline_state, cpu);
77 }
78 
79 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
80 {
81 	per_cpu(preferred_offline_state, cpu) = state;
82 }
83 
84 void set_default_offline_state(int cpu)
85 {
86 	per_cpu(preferred_offline_state, cpu) = default_offline_state;
87 }
88 
89 static void rtas_stop_self(void)
90 {
91 	struct rtas_args args = {
92 		.token = cpu_to_be32(rtas_stop_self_token),
93 		.nargs = 0,
94 		.nret = 1,
95 		.rets = &args.args[0],
96 	};
97 
98 	local_irq_disable();
99 
100 	BUG_ON(rtas_stop_self_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 (!lppaca_shared_proc(get_lppaca()))
128 			get_lppaca()->donate_dedicated_cpu = 1;
129 
130 		while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
131 			while (!prep_irq_for_idle()) {
132 				local_irq_enable();
133 				local_irq_disable();
134 			}
135 
136 			extended_cede_processor(cede_latency_hint);
137 		}
138 
139 		local_irq_disable();
140 
141 		if (!lppaca_shared_proc(get_lppaca()))
142 			get_lppaca()->donate_dedicated_cpu = 0;
143 		get_lppaca()->idle = 0;
144 
145 		if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
146 			unregister_slb_shadow(hwcpu);
147 
148 			hard_irq_disable();
149 			/*
150 			 * Call to start_secondary_resume() will not return.
151 			 * Kernel stack will be reset and start_secondary()
152 			 * will be called to continue the online operation.
153 			 */
154 			start_secondary_resume();
155 		}
156 	}
157 
158 	/* Requested state is CPU_STATE_OFFLINE at this point */
159 	WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
160 
161 	set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
162 	unregister_slb_shadow(hwcpu);
163 	rtas_stop_self();
164 
165 	/* Should never get here... */
166 	BUG();
167 	for(;;);
168 }
169 
170 static int pseries_cpu_disable(void)
171 {
172 	int cpu = smp_processor_id();
173 
174 	set_cpu_online(cpu, false);
175 	vdso_data->processorCount--;
176 
177 	/*fix boot_cpuid here*/
178 	if (cpu == boot_cpuid)
179 		boot_cpuid = cpumask_any(cpu_online_mask);
180 
181 	/* FIXME: abstract this to not be platform specific later on */
182 	xics_migrate_irqs_away();
183 	return 0;
184 }
185 
186 /*
187  * pseries_cpu_die: Wait for the cpu to die.
188  * @cpu: logical processor id of the CPU whose death we're awaiting.
189  *
190  * This function is called from the context of the thread which is performing
191  * the cpu-offline. Here we wait for long enough to allow the cpu in question
192  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
193  * notifications.
194  *
195  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
196  * self-destruct.
197  */
198 static void pseries_cpu_die(unsigned int cpu)
199 {
200 	int tries;
201 	int cpu_status = 1;
202 	unsigned int pcpu = get_hard_smp_processor_id(cpu);
203 
204 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
205 		cpu_status = 1;
206 		for (tries = 0; tries < 5000; tries++) {
207 			if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
208 				cpu_status = 0;
209 				break;
210 			}
211 			msleep(1);
212 		}
213 	} else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
214 
215 		for (tries = 0; tries < 25; tries++) {
216 			cpu_status = smp_query_cpu_stopped(pcpu);
217 			if (cpu_status == QCSS_STOPPED ||
218 			    cpu_status == QCSS_HARDWARE_ERROR)
219 				break;
220 			cpu_relax();
221 		}
222 	}
223 
224 	if (cpu_status != 0) {
225 		printk("Querying DEAD? cpu %i (%i) shows %i\n",
226 		       cpu, pcpu, cpu_status);
227 	}
228 
229 	/* Isolation and deallocation are definitely done by
230 	 * drslot_chrp_cpu.  If they were not they would be
231 	 * done here.  Change isolate state to Isolate and
232 	 * change allocation-state to Unusable.
233 	 */
234 	paca[cpu].cpu_start = 0;
235 }
236 
237 /*
238  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
239  * here is that a cpu device node may represent up to two logical cpus
240  * in the SMT case.  We must honor the assumption in other code that
241  * the logical ids for sibling SMT threads x and y are adjacent, such
242  * that x^1 == y and y^1 == x.
243  */
244 static int pseries_add_processor(struct device_node *np)
245 {
246 	unsigned int cpu;
247 	cpumask_var_t candidate_mask, tmp;
248 	int err = -ENOSPC, len, nthreads, i;
249 	const u32 *intserv;
250 
251 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
252 	if (!intserv)
253 		return 0;
254 
255 	zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
256 	zalloc_cpumask_var(&tmp, GFP_KERNEL);
257 
258 	nthreads = len / sizeof(u32);
259 	for (i = 0; i < nthreads; i++)
260 		cpumask_set_cpu(i, tmp);
261 
262 	cpu_maps_update_begin();
263 
264 	BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
265 
266 	/* Get a bitmap of unoccupied slots. */
267 	cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
268 	if (cpumask_empty(candidate_mask)) {
269 		/* If we get here, it most likely means that NR_CPUS is
270 		 * less than the partition's max processors setting.
271 		 */
272 		printk(KERN_ERR "Cannot add cpu %s; this system configuration"
273 		       " supports %d logical cpus.\n", np->full_name,
274 		       cpumask_weight(cpu_possible_mask));
275 		goto out_unlock;
276 	}
277 
278 	while (!cpumask_empty(tmp))
279 		if (cpumask_subset(tmp, candidate_mask))
280 			/* Found a range where we can insert the new cpu(s) */
281 			break;
282 		else
283 			cpumask_shift_left(tmp, tmp, nthreads);
284 
285 	if (cpumask_empty(tmp)) {
286 		printk(KERN_ERR "Unable to find space in cpu_present_mask for"
287 		       " processor %s with %d thread(s)\n", np->name,
288 		       nthreads);
289 		goto out_unlock;
290 	}
291 
292 	for_each_cpu(cpu, tmp) {
293 		BUG_ON(cpu_present(cpu));
294 		set_cpu_present(cpu, true);
295 		set_hard_smp_processor_id(cpu, *intserv++);
296 	}
297 	err = 0;
298 out_unlock:
299 	cpu_maps_update_done();
300 	free_cpumask_var(candidate_mask);
301 	free_cpumask_var(tmp);
302 	return err;
303 }
304 
305 /*
306  * Update the present map for a cpu node which is going away, and set
307  * the hard id in the paca(s) to -1 to be consistent with boot time
308  * convention for non-present cpus.
309  */
310 static void pseries_remove_processor(struct device_node *np)
311 {
312 	unsigned int cpu;
313 	int len, nthreads, i;
314 	const u32 *intserv;
315 
316 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
317 	if (!intserv)
318 		return;
319 
320 	nthreads = len / sizeof(u32);
321 
322 	cpu_maps_update_begin();
323 	for (i = 0; i < nthreads; i++) {
324 		for_each_present_cpu(cpu) {
325 			if (get_hard_smp_processor_id(cpu) != intserv[i])
326 				continue;
327 			BUG_ON(cpu_online(cpu));
328 			set_cpu_present(cpu, false);
329 			set_hard_smp_processor_id(cpu, -1);
330 			break;
331 		}
332 		if (cpu >= nr_cpu_ids)
333 			printk(KERN_WARNING "Could not find cpu to remove "
334 			       "with physical id 0x%x\n", intserv[i]);
335 	}
336 	cpu_maps_update_done();
337 }
338 
339 static int pseries_smp_notifier(struct notifier_block *nb,
340 				unsigned long action, void *node)
341 {
342 	int err = 0;
343 
344 	switch (action) {
345 	case OF_RECONFIG_ATTACH_NODE:
346 		err = pseries_add_processor(node);
347 		break;
348 	case OF_RECONFIG_DETACH_NODE:
349 		pseries_remove_processor(node);
350 		break;
351 	}
352 	return notifier_from_errno(err);
353 }
354 
355 static struct notifier_block pseries_smp_nb = {
356 	.notifier_call = pseries_smp_notifier,
357 };
358 
359 #define MAX_CEDE_LATENCY_LEVELS		4
360 #define	CEDE_LATENCY_PARAM_LENGTH	10
361 #define CEDE_LATENCY_PARAM_MAX_LENGTH	\
362 	(MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
363 #define CEDE_LATENCY_TOKEN		45
364 
365 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
366 
367 static int parse_cede_parameters(void)
368 {
369 	memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
370 	return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
371 			 NULL,
372 			 CEDE_LATENCY_TOKEN,
373 			 __pa(cede_parameters),
374 			 CEDE_LATENCY_PARAM_MAX_LENGTH);
375 }
376 
377 static int __init pseries_cpu_hotplug_init(void)
378 {
379 	struct device_node *np;
380 	const char *typep;
381 	int cpu;
382 	int qcss_tok;
383 
384 	for_each_node_by_name(np, "interrupt-controller") {
385 		typep = of_get_property(np, "compatible", NULL);
386 		if (strstr(typep, "open-pic")) {
387 			of_node_put(np);
388 
389 			printk(KERN_INFO "CPU Hotplug not supported on "
390 				"systems using MPIC\n");
391 			return 0;
392 		}
393 	}
394 
395 	rtas_stop_self_token = rtas_token("stop-self");
396 	qcss_tok = rtas_token("query-cpu-stopped-state");
397 
398 	if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
399 			qcss_tok == RTAS_UNKNOWN_SERVICE) {
400 		printk(KERN_INFO "CPU Hotplug not supported by firmware "
401 				"- disabling.\n");
402 		return 0;
403 	}
404 
405 	ppc_md.cpu_die = pseries_mach_cpu_die;
406 	smp_ops->cpu_disable = pseries_cpu_disable;
407 	smp_ops->cpu_die = pseries_cpu_die;
408 
409 	/* Processors can be added/removed only on LPAR */
410 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
411 		of_reconfig_notifier_register(&pseries_smp_nb);
412 		cpu_maps_update_begin();
413 		if (cede_offline_enabled && parse_cede_parameters() == 0) {
414 			default_offline_state = CPU_STATE_INACTIVE;
415 			for_each_online_cpu(cpu)
416 				set_default_offline_state(cpu);
417 		}
418 		cpu_maps_update_done();
419 	}
420 
421 	return 0;
422 }
423 machine_arch_initcall(pseries, pseries_cpu_hotplug_init);
424