1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * SMP support for pSeries machines.
4  *
5  * Dave Engebretsen, Peter Bergner, and
6  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
7  *
8  * Plus various changes from other IBM teams...
9  */
10 
11 
12 #include <linux/kernel.h>
13 #include <linux/sched.h>
14 #include <linux/smp.h>
15 #include <linux/interrupt.h>
16 #include <linux/delay.h>
17 #include <linux/init.h>
18 #include <linux/spinlock.h>
19 #include <linux/cache.h>
20 #include <linux/err.h>
21 #include <linux/device.h>
22 #include <linux/cpu.h>
23 #include <linux/pgtable.h>
24 
25 #include <asm/ptrace.h>
26 #include <linux/atomic.h>
27 #include <asm/irq.h>
28 #include <asm/page.h>
29 #include <asm/io.h>
30 #include <asm/prom.h>
31 #include <asm/smp.h>
32 #include <asm/paca.h>
33 #include <asm/machdep.h>
34 #include <asm/cputable.h>
35 #include <asm/firmware.h>
36 #include <asm/rtas.h>
37 #include <asm/vdso_datapage.h>
38 #include <asm/cputhreads.h>
39 #include <asm/xics.h>
40 #include <asm/xive.h>
41 #include <asm/dbell.h>
42 #include <asm/plpar_wrappers.h>
43 #include <asm/code-patching.h>
44 #include <asm/svm.h>
45 #include <asm/kvm_guest.h>
46 
47 #include "pseries.h"
48 
49 /*
50  * The Primary thread of each non-boot processor was started from the OF client
51  * interface by prom_hold_cpus and is spinning on secondary_hold_spinloop.
52  */
53 static cpumask_var_t of_spin_mask;
54 
55 /* Query where a cpu is now.  Return codes #defined in plpar_wrappers.h */
56 int smp_query_cpu_stopped(unsigned int pcpu)
57 {
58 	int cpu_status, status;
59 	int qcss_tok = rtas_token("query-cpu-stopped-state");
60 
61 	if (qcss_tok == RTAS_UNKNOWN_SERVICE) {
62 		printk_once(KERN_INFO
63 			"Firmware doesn't support query-cpu-stopped-state\n");
64 		return QCSS_HARDWARE_ERROR;
65 	}
66 
67 	status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
68 	if (status != 0) {
69 		printk(KERN_ERR
70 		       "RTAS query-cpu-stopped-state failed: %i\n", status);
71 		return status;
72 	}
73 
74 	return cpu_status;
75 }
76 
77 /**
78  * smp_startup_cpu() - start the given cpu
79  *
80  * At boot time, there is nothing to do for primary threads which were
81  * started from Open Firmware.  For anything else, call RTAS with the
82  * appropriate start location.
83  *
84  * Returns:
85  *	0	- failure
86  *	1	- success
87  */
88 static inline int smp_startup_cpu(unsigned int lcpu)
89 {
90 	int status;
91 	unsigned long start_here =
92 			__pa(ppc_function_entry(generic_secondary_smp_init));
93 	unsigned int pcpu;
94 	int start_cpu;
95 
96 	if (cpumask_test_cpu(lcpu, of_spin_mask))
97 		/* Already started by OF and sitting in spin loop */
98 		return 1;
99 
100 	pcpu = get_hard_smp_processor_id(lcpu);
101 
102 	/* Check to see if the CPU out of FW already for kexec */
103 	if (smp_query_cpu_stopped(pcpu) == QCSS_NOT_STOPPED){
104 		cpumask_set_cpu(lcpu, of_spin_mask);
105 		return 1;
106 	}
107 
108 	/*
109 	 * If the RTAS start-cpu token does not exist then presume the
110 	 * cpu is already spinning.
111 	 */
112 	start_cpu = rtas_token("start-cpu");
113 	if (start_cpu == RTAS_UNKNOWN_SERVICE)
114 		return 1;
115 
116 	status = rtas_call(start_cpu, 3, 1, NULL, pcpu, start_here, pcpu);
117 	if (status != 0) {
118 		printk(KERN_ERR "start-cpu failed: %i\n", status);
119 		return 0;
120 	}
121 
122 	return 1;
123 }
124 
125 static void smp_setup_cpu(int cpu)
126 {
127 	if (xive_enabled())
128 		xive_smp_setup_cpu();
129 	else if (cpu != boot_cpuid)
130 		xics_setup_cpu();
131 
132 	if (firmware_has_feature(FW_FEATURE_SPLPAR))
133 		vpa_init(cpu);
134 
135 	cpumask_clear_cpu(cpu, of_spin_mask);
136 }
137 
138 static int smp_pSeries_kick_cpu(int nr)
139 {
140 	if (nr < 0 || nr >= nr_cpu_ids)
141 		return -EINVAL;
142 
143 	if (!smp_startup_cpu(nr))
144 		return -ENOENT;
145 
146 	/*
147 	 * The processor is currently spinning, waiting for the
148 	 * cpu_start field to become non-zero After we set cpu_start,
149 	 * the processor will continue on to secondary_start
150 	 */
151 	paca_ptrs[nr]->cpu_start = 1;
152 
153 	return 0;
154 }
155 
156 static int pseries_smp_prepare_cpu(int cpu)
157 {
158 	if (xive_enabled())
159 		return xive_smp_prepare_cpu(cpu);
160 	return 0;
161 }
162 
163 /* Cause IPI as setup by the interrupt controller (xics or xive) */
164 static void (*ic_cause_ipi)(int cpu) __ro_after_init;
165 
166 /* Use msgsndp doorbells target is a sibling, else use interrupt controller */
167 static void dbell_or_ic_cause_ipi(int cpu)
168 {
169 	if (doorbell_try_core_ipi(cpu))
170 		return;
171 
172 	ic_cause_ipi(cpu);
173 }
174 
175 static int pseries_cause_nmi_ipi(int cpu)
176 {
177 	int hwcpu;
178 
179 	if (cpu == NMI_IPI_ALL_OTHERS) {
180 		hwcpu = H_SIGNAL_SYS_RESET_ALL_OTHERS;
181 	} else {
182 		if (cpu < 0) {
183 			WARN_ONCE(true, "incorrect cpu parameter %d", cpu);
184 			return 0;
185 		}
186 
187 		hwcpu = get_hard_smp_processor_id(cpu);
188 	}
189 
190 	if (plpar_signal_sys_reset(hwcpu) == H_SUCCESS)
191 		return 1;
192 
193 	return 0;
194 }
195 
196 static __init void pSeries_smp_probe(void)
197 {
198 	if (xive_enabled())
199 		xive_smp_probe();
200 	else
201 		xics_smp_probe();
202 
203 	/* No doorbell facility, must use the interrupt controller for IPIs */
204 	if (!cpu_has_feature(CPU_FTR_DBELL))
205 		return;
206 
207 	/* Doorbells can only be used for IPIs between SMT siblings */
208 	if (!cpu_has_feature(CPU_FTR_SMT))
209 		return;
210 
211 	check_kvm_guest();
212 
213 	if (is_kvm_guest()) {
214 		/*
215 		 * KVM emulates doorbells by disabling FSCR[MSGP] so msgsndp
216 		 * faults to the hypervisor which then reads the instruction
217 		 * from guest memory, which tends to be slower than using XIVE.
218 		 */
219 		if (xive_enabled())
220 			return;
221 
222 		/*
223 		 * XICS hcalls aren't as fast, so we can use msgsndp (which
224 		 * also helps exercise KVM emulation), however KVM can't
225 		 * emulate secure guests because it can't read the instruction
226 		 * out of their memory.
227 		 */
228 		if (is_secure_guest())
229 			return;
230 	}
231 
232 	/*
233 	 * Under PowerVM, FSCR[MSGP] is enabled as guest vCPU siblings are
234 	 * gang scheduled on the same physical core, so doorbells are always
235 	 * faster than the interrupt controller, and they can be used by
236 	 * secure guests.
237 	 */
238 
239 	ic_cause_ipi = smp_ops->cause_ipi;
240 	smp_ops->cause_ipi = dbell_or_ic_cause_ipi;
241 }
242 
243 static struct smp_ops_t pseries_smp_ops = {
244 	.message_pass	= NULL,	/* Use smp_muxed_ipi_message_pass */
245 	.cause_ipi	= NULL,	/* Filled at runtime by pSeries_smp_probe() */
246 	.cause_nmi_ipi	= pseries_cause_nmi_ipi,
247 	.probe		= pSeries_smp_probe,
248 	.prepare_cpu	= pseries_smp_prepare_cpu,
249 	.kick_cpu	= smp_pSeries_kick_cpu,
250 	.setup_cpu	= smp_setup_cpu,
251 	.cpu_bootable	= smp_generic_cpu_bootable,
252 };
253 
254 /* This is called very early */
255 void __init smp_init_pseries(void)
256 {
257 	int i;
258 
259 	pr_debug(" -> smp_init_pSeries()\n");
260 	smp_ops = &pseries_smp_ops;
261 
262 	alloc_bootmem_cpumask_var(&of_spin_mask);
263 
264 	/*
265 	 * Mark threads which are still spinning in hold loops
266 	 *
267 	 * We know prom_init will not have started them if RTAS supports
268 	 * query-cpu-stopped-state.
269 	 */
270 	if (rtas_token("query-cpu-stopped-state") == RTAS_UNKNOWN_SERVICE) {
271 		if (cpu_has_feature(CPU_FTR_SMT)) {
272 			for_each_present_cpu(i) {
273 				if (cpu_thread_in_core(i) == 0)
274 					cpumask_set_cpu(i, of_spin_mask);
275 			}
276 		} else
277 			cpumask_copy(of_spin_mask, cpu_present_mask);
278 
279 		cpumask_clear_cpu(boot_cpuid, of_spin_mask);
280 	}
281 
282 	/* Non-lpar has additional take/give timebase */
283 	if (rtas_token("freeze-time-base") != RTAS_UNKNOWN_SERVICE) {
284 		smp_ops->give_timebase = rtas_give_timebase;
285 		smp_ops->take_timebase = rtas_take_timebase;
286 	}
287 
288 	pr_debug(" <- smp_init_pSeries()\n");
289 }
290