1 /*
2  * SMP support for PowerNV machines.
3  *
4  * Copyright 2011 IBM Corp.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11 
12 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/sched.h>
15 #include <linux/sched/hotplug.h>
16 #include <linux/smp.h>
17 #include <linux/interrupt.h>
18 #include <linux/delay.h>
19 #include <linux/init.h>
20 #include <linux/spinlock.h>
21 #include <linux/cpu.h>
22 
23 #include <asm/irq.h>
24 #include <asm/smp.h>
25 #include <asm/paca.h>
26 #include <asm/machdep.h>
27 #include <asm/cputable.h>
28 #include <asm/firmware.h>
29 #include <asm/vdso_datapage.h>
30 #include <asm/cputhreads.h>
31 #include <asm/xics.h>
32 #include <asm/xive.h>
33 #include <asm/opal.h>
34 #include <asm/runlatch.h>
35 #include <asm/code-patching.h>
36 #include <asm/dbell.h>
37 #include <asm/kvm_ppc.h>
38 #include <asm/ppc-opcode.h>
39 #include <asm/cpuidle.h>
40 
41 #include "powernv.h"
42 
43 #ifdef DEBUG
44 #include <asm/udbg.h>
45 #define DBG(fmt...) udbg_printf(fmt)
46 #else
47 #define DBG(fmt...)
48 #endif
49 
50 static void pnv_smp_setup_cpu(int cpu)
51 {
52 	if (xive_enabled())
53 		xive_smp_setup_cpu();
54 	else if (cpu != boot_cpuid)
55 		xics_setup_cpu();
56 }
57 
58 static int pnv_smp_kick_cpu(int nr)
59 {
60 	unsigned int pcpu;
61 	unsigned long start_here =
62 			__pa(ppc_function_entry(generic_secondary_smp_init));
63 	long rc;
64 	uint8_t status;
65 
66 	if (nr < 0 || nr >= nr_cpu_ids)
67 		return -EINVAL;
68 
69 	pcpu = get_hard_smp_processor_id(nr);
70 	/*
71 	 * If we already started or OPAL is not supported, we just
72 	 * kick the CPU via the PACA
73 	 */
74 	if (paca[nr].cpu_start || !firmware_has_feature(FW_FEATURE_OPAL))
75 		goto kick;
76 
77 	/*
78 	 * At this point, the CPU can either be spinning on the way in
79 	 * from kexec or be inside OPAL waiting to be started for the
80 	 * first time. OPAL v3 allows us to query OPAL to know if it
81 	 * has the CPUs, so we do that
82 	 */
83 	rc = opal_query_cpu_status(pcpu, &status);
84 	if (rc != OPAL_SUCCESS) {
85 		pr_warn("OPAL Error %ld querying CPU %d state\n", rc, nr);
86 		return -ENODEV;
87 	}
88 
89 	/*
90 	 * Already started, just kick it, probably coming from
91 	 * kexec and spinning
92 	 */
93 	if (status == OPAL_THREAD_STARTED)
94 		goto kick;
95 
96 	/*
97 	 * Available/inactive, let's kick it
98 	 */
99 	if (status == OPAL_THREAD_INACTIVE) {
100 		pr_devel("OPAL: Starting CPU %d (HW 0x%x)...\n", nr, pcpu);
101 		rc = opal_start_cpu(pcpu, start_here);
102 		if (rc != OPAL_SUCCESS) {
103 			pr_warn("OPAL Error %ld starting CPU %d\n", rc, nr);
104 			return -ENODEV;
105 		}
106 	} else {
107 		/*
108 		 * An unavailable CPU (or any other unknown status)
109 		 * shouldn't be started. It should also
110 		 * not be in the possible map but currently it can
111 		 * happen
112 		 */
113 		pr_devel("OPAL: CPU %d (HW 0x%x) is unavailable"
114 			 " (status %d)...\n", nr, pcpu, status);
115 		return -ENODEV;
116 	}
117 
118 kick:
119 	return smp_generic_kick_cpu(nr);
120 }
121 
122 #ifdef CONFIG_HOTPLUG_CPU
123 
124 static int pnv_smp_cpu_disable(void)
125 {
126 	int cpu = smp_processor_id();
127 
128 	/* This is identical to pSeries... might consolidate by
129 	 * moving migrate_irqs_away to a ppc_md with default to
130 	 * the generic fixup_irqs. --BenH.
131 	 */
132 	set_cpu_online(cpu, false);
133 	vdso_data->processorCount--;
134 	if (cpu == boot_cpuid)
135 		boot_cpuid = cpumask_any(cpu_online_mask);
136 	if (xive_enabled())
137 		xive_smp_disable_cpu();
138 	else
139 		xics_migrate_irqs_away();
140 	return 0;
141 }
142 
143 static void pnv_smp_cpu_kill_self(void)
144 {
145 	unsigned int cpu;
146 	unsigned long srr1, wmask;
147 
148 	/* Standard hot unplug procedure */
149 	/*
150 	 * This hard disables local interurpts, ensuring we have no lazy
151 	 * irqs pending.
152 	 */
153 	WARN_ON(irqs_disabled());
154 	hard_irq_disable();
155 	WARN_ON(lazy_irq_pending());
156 
157 	idle_task_exit();
158 	current->active_mm = NULL; /* for sanity */
159 	cpu = smp_processor_id();
160 	DBG("CPU%d offline\n", cpu);
161 	generic_set_cpu_dead(cpu);
162 	smp_wmb();
163 
164 	wmask = SRR1_WAKEMASK;
165 	if (cpu_has_feature(CPU_FTR_ARCH_207S))
166 		wmask = SRR1_WAKEMASK_P8;
167 
168 	while (!generic_check_cpu_restart(cpu)) {
169 		/*
170 		 * Clear IPI flag, since we don't handle IPIs while
171 		 * offline, except for those when changing micro-threading
172 		 * mode, which are handled explicitly below, and those
173 		 * for coming online, which are handled via
174 		 * generic_check_cpu_restart() calls.
175 		 */
176 		kvmppc_set_host_ipi(cpu, 0);
177 
178 		srr1 = pnv_cpu_offline(cpu);
179 
180 		WARN_ON(lazy_irq_pending());
181 
182 		/*
183 		 * If the SRR1 value indicates that we woke up due to
184 		 * an external interrupt, then clear the interrupt.
185 		 * We clear the interrupt before checking for the
186 		 * reason, so as to avoid a race where we wake up for
187 		 * some other reason, find nothing and clear the interrupt
188 		 * just as some other cpu is sending us an interrupt.
189 		 * If we returned from power7_nap as a result of
190 		 * having finished executing in a KVM guest, then srr1
191 		 * contains 0.
192 		 */
193 		if (((srr1 & wmask) == SRR1_WAKEEE) ||
194 		    ((srr1 & wmask) == SRR1_WAKEHVI)) {
195 			if (cpu_has_feature(CPU_FTR_ARCH_300)) {
196 				if (xive_enabled())
197 					xive_flush_interrupt();
198 				else
199 					icp_opal_flush_interrupt();
200 			} else
201 				icp_native_flush_interrupt();
202 		} else if ((srr1 & wmask) == SRR1_WAKEHDBELL) {
203 			unsigned long msg = PPC_DBELL_TYPE(PPC_DBELL_SERVER);
204 			asm volatile(PPC_MSGCLR(%0) : : "r" (msg));
205 		}
206 		smp_mb();
207 
208 		if (cpu_core_split_required())
209 			continue;
210 
211 		if (srr1 && !generic_check_cpu_restart(cpu))
212 			DBG("CPU%d Unexpected exit while offline srr1=%lx!\n",
213 					cpu, srr1);
214 
215 	}
216 
217 	DBG("CPU%d coming online...\n", cpu);
218 }
219 
220 #endif /* CONFIG_HOTPLUG_CPU */
221 
222 static int pnv_cpu_bootable(unsigned int nr)
223 {
224 	/*
225 	 * Starting with POWER8, the subcore logic relies on all threads of a
226 	 * core being booted so that they can participate in split mode
227 	 * switches. So on those machines we ignore the smt_enabled_at_boot
228 	 * setting (smt-enabled on the kernel command line).
229 	 */
230 	if (cpu_has_feature(CPU_FTR_ARCH_207S))
231 		return 1;
232 
233 	return smp_generic_cpu_bootable(nr);
234 }
235 
236 static int pnv_smp_prepare_cpu(int cpu)
237 {
238 	if (xive_enabled())
239 		return xive_smp_prepare_cpu(cpu);
240 	return 0;
241 }
242 
243 /* Cause IPI as setup by the interrupt controller (xics or xive) */
244 static void (*ic_cause_ipi)(int cpu);
245 
246 static void pnv_cause_ipi(int cpu)
247 {
248 	if (doorbell_try_core_ipi(cpu))
249 		return;
250 
251 	ic_cause_ipi(cpu);
252 }
253 
254 static void pnv_p9_dd1_cause_ipi(int cpu)
255 {
256 	int this_cpu = get_cpu();
257 
258 	/*
259 	 * POWER9 DD1 has a global addressed msgsnd, but for now we restrict
260 	 * IPIs to same core, because it requires additional synchronization
261 	 * for inter-core doorbells which we do not implement.
262 	 */
263 	if (cpumask_test_cpu(cpu, cpu_sibling_mask(this_cpu)))
264 		doorbell_global_ipi(cpu);
265 	else
266 		ic_cause_ipi(cpu);
267 
268 	put_cpu();
269 }
270 
271 static void __init pnv_smp_probe(void)
272 {
273 	if (xive_enabled())
274 		xive_smp_probe();
275 	else
276 		xics_smp_probe();
277 
278 	if (cpu_has_feature(CPU_FTR_DBELL)) {
279 		ic_cause_ipi = smp_ops->cause_ipi;
280 		WARN_ON(!ic_cause_ipi);
281 
282 		if (cpu_has_feature(CPU_FTR_ARCH_300)) {
283 			if (cpu_has_feature(CPU_FTR_POWER9_DD1))
284 				smp_ops->cause_ipi = pnv_p9_dd1_cause_ipi;
285 			else
286 				smp_ops->cause_ipi = doorbell_global_ipi;
287 		} else {
288 			smp_ops->cause_ipi = pnv_cause_ipi;
289 		}
290 	}
291 }
292 
293 static struct smp_ops_t pnv_smp_ops = {
294 	.message_pass	= NULL, /* Use smp_muxed_ipi_message_pass */
295 	.cause_ipi	= NULL,	/* Filled at runtime by pnv_smp_probe() */
296 	.cause_nmi_ipi	= NULL,
297 	.probe		= pnv_smp_probe,
298 	.prepare_cpu	= pnv_smp_prepare_cpu,
299 	.kick_cpu	= pnv_smp_kick_cpu,
300 	.setup_cpu	= pnv_smp_setup_cpu,
301 	.cpu_bootable	= pnv_cpu_bootable,
302 #ifdef CONFIG_HOTPLUG_CPU
303 	.cpu_disable	= pnv_smp_cpu_disable,
304 	.cpu_die	= generic_cpu_die,
305 #endif /* CONFIG_HOTPLUG_CPU */
306 };
307 
308 /* This is called very early during platform setup_arch */
309 void __init pnv_smp_init(void)
310 {
311 	smp_ops = &pnv_smp_ops;
312 
313 #ifdef CONFIG_HOTPLUG_CPU
314 	ppc_md.cpu_die	= pnv_smp_cpu_kill_self;
315 #endif
316 }
317