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

source code of linux/arch/powerpc/platforms/pseries/smp.c