Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Intel SMP support routines.
4 *
5 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
6 * (c) 1998-99, 2000, 2009 Ingo Molnar <mingo@redhat.com>
7 * (c) 2002,2003 Andi Kleen, SuSE Labs.
8 *
9 * i386 and x86_64 integration by Glauber Costa <gcosta@redhat.com>
10 */
11
12#include <linux/init.h>
13
14#include <linux/mm.h>
15#include <linux/delay.h>
16#include <linux/spinlock.h>
17#include <linux/export.h>
18#include <linux/kernel_stat.h>
19#include <linux/mc146818rtc.h>
20#include <linux/cache.h>
21#include <linux/interrupt.h>
22#include <linux/cpu.h>
23#include <linux/gfp.h>
24
25#include <asm/mtrr.h>
26#include <asm/tlbflush.h>
27#include <asm/mmu_context.h>
28#include <asm/proto.h>
29#include <asm/apic.h>
30#include <asm/idtentry.h>
31#include <asm/nmi.h>
32#include <asm/mce.h>
33#include <asm/trace/irq_vectors.h>
34#include <asm/kexec.h>
35#include <asm/virtext.h>
36
37/*
38 * Some notes on x86 processor bugs affecting SMP operation:
39 *
40 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
41 * The Linux implications for SMP are handled as follows:
42 *
43 * Pentium III / [Xeon]
44 * None of the E1AP-E3AP errata are visible to the user.
45 *
46 * E1AP. see PII A1AP
47 * E2AP. see PII A2AP
48 * E3AP. see PII A3AP
49 *
50 * Pentium II / [Xeon]
51 * None of the A1AP-A3AP errata are visible to the user.
52 *
53 * A1AP. see PPro 1AP
54 * A2AP. see PPro 2AP
55 * A3AP. see PPro 7AP
56 *
57 * Pentium Pro
58 * None of 1AP-9AP errata are visible to the normal user,
59 * except occasional delivery of 'spurious interrupt' as trap #15.
60 * This is very rare and a non-problem.
61 *
62 * 1AP. Linux maps APIC as non-cacheable
63 * 2AP. worked around in hardware
64 * 3AP. fixed in C0 and above steppings microcode update.
65 * Linux does not use excessive STARTUP_IPIs.
66 * 4AP. worked around in hardware
67 * 5AP. symmetric IO mode (normal Linux operation) not affected.
68 * 'noapic' mode has vector 0xf filled out properly.
69 * 6AP. 'noapic' mode might be affected - fixed in later steppings
70 * 7AP. We do not assume writes to the LVT deassering IRQs
71 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
72 * 9AP. We do not use mixed mode
73 *
74 * Pentium
75 * There is a marginal case where REP MOVS on 100MHz SMP
76 * machines with B stepping processors can fail. XXX should provide
77 * an L1cache=Writethrough or L1cache=off option.
78 *
79 * B stepping CPUs may hang. There are hardware work arounds
80 * for this. We warn about it in case your board doesn't have the work
81 * arounds. Basically that's so I can tell anyone with a B stepping
82 * CPU and SMP problems "tough".
83 *
84 * Specific items [From Pentium Processor Specification Update]
85 *
86 * 1AP. Linux doesn't use remote read
87 * 2AP. Linux doesn't trust APIC errors
88 * 3AP. We work around this
89 * 4AP. Linux never generated 3 interrupts of the same priority
90 * to cause a lost local interrupt.
91 * 5AP. Remote read is never used
92 * 6AP. not affected - worked around in hardware
93 * 7AP. not affected - worked around in hardware
94 * 8AP. worked around in hardware - we get explicit CS errors if not
95 * 9AP. only 'noapic' mode affected. Might generate spurious
96 * interrupts, we log only the first one and count the
97 * rest silently.
98 * 10AP. not affected - worked around in hardware
99 * 11AP. Linux reads the APIC between writes to avoid this, as per
100 * the documentation. Make sure you preserve this as it affects
101 * the C stepping chips too.
102 * 12AP. not affected - worked around in hardware
103 * 13AP. not affected - worked around in hardware
104 * 14AP. we always deassert INIT during bootup
105 * 15AP. not affected - worked around in hardware
106 * 16AP. not affected - worked around in hardware
107 * 17AP. not affected - worked around in hardware
108 * 18AP. not affected - worked around in hardware
109 * 19AP. not affected - worked around in BIOS
110 *
111 * If this sounds worrying believe me these bugs are either ___RARE___,
112 * or are signal timing bugs worked around in hardware and there's
113 * about nothing of note with C stepping upwards.
114 */
115
116static atomic_t stopping_cpu = ATOMIC_INIT(-1);
117static bool smp_no_nmi_ipi = false;
118
119static int smp_stop_nmi_callback(unsigned int val, struct pt_regs *regs)
120{
121 /* We are registered on stopping cpu too, avoid spurious NMI */
122 if (raw_smp_processor_id() == atomic_read(&stopping_cpu))
123 return NMI_HANDLED;
124
125 cpu_emergency_vmxoff();
126 stop_this_cpu(NULL);
127
128 return NMI_HANDLED;
129}
130
131/*
132 * this function calls the 'stop' function on all other CPUs in the system.
133 */
134DEFINE_IDTENTRY_SYSVEC(sysvec_reboot)
135{
136 ack_APIC_irq();
137 cpu_emergency_vmxoff();
138 stop_this_cpu(NULL);
139}
140
141static int register_stop_handler(void)
142{
143 return register_nmi_handler(NMI_LOCAL, smp_stop_nmi_callback,
144 NMI_FLAG_FIRST, "smp_stop");
145}
146
147static void native_stop_other_cpus(int wait)
148{
149 unsigned long flags;
150 unsigned long timeout;
151
152 if (reboot_force)
153 return;
154
155 /*
156 * Use an own vector here because smp_call_function
157 * does lots of things not suitable in a panic situation.
158 */
159
160 /*
161 * We start by using the REBOOT_VECTOR irq.
162 * The irq is treated as a sync point to allow critical
163 * regions of code on other cpus to release their spin locks
164 * and re-enable irqs. Jumping straight to an NMI might
165 * accidentally cause deadlocks with further shutdown/panic
166 * code. By syncing, we give the cpus up to one second to
167 * finish their work before we force them off with the NMI.
168 */
169 if (num_online_cpus() > 1) {
170 /* did someone beat us here? */
171 if (atomic_cmpxchg(&stopping_cpu, -1, safe_smp_processor_id()) != -1)
172 return;
173
174 /* sync above data before sending IRQ */
175 wmb();
176
177 apic_send_IPI_allbutself(REBOOT_VECTOR);
178
179 /*
180 * Don't wait longer than a second for IPI completion. The
181 * wait request is not checked here because that would
182 * prevent an NMI shutdown attempt in case that not all
183 * CPUs reach shutdown state.
184 */
185 timeout = USEC_PER_SEC;
186 while (num_online_cpus() > 1 && timeout--)
187 udelay(1);
188 }
189
190 /* if the REBOOT_VECTOR didn't work, try with the NMI */
191 if (num_online_cpus() > 1) {
192 /*
193 * If NMI IPI is enabled, try to register the stop handler
194 * and send the IPI. In any case try to wait for the other
195 * CPUs to stop.
196 */
197 if (!smp_no_nmi_ipi && !register_stop_handler()) {
198 /* Sync above data before sending IRQ */
199 wmb();
200
201 pr_emerg("Shutting down cpus with NMI\n");
202
203 apic_send_IPI_allbutself(NMI_VECTOR);
204 }
205 /*
206 * Don't wait longer than 10 ms if the caller didn't
207 * reqeust it. If wait is true, the machine hangs here if
208 * one or more CPUs do not reach shutdown state.
209 */
210 timeout = USEC_PER_MSEC * 10;
211 while (num_online_cpus() > 1 && (wait || timeout--))
212 udelay(1);
213 }
214
215 local_irq_save(flags);
216 disable_local_APIC();
217 mcheck_cpu_clear(this_cpu_ptr(&cpu_info));
218 local_irq_restore(flags);
219}
220
221/*
222 * Reschedule call back. KVM uses this interrupt to force a cpu out of
223 * guest mode.
224 */
225DEFINE_IDTENTRY_SYSVEC_SIMPLE(sysvec_reschedule_ipi)
226{
227 ack_APIC_irq();
228 trace_reschedule_entry(RESCHEDULE_VECTOR);
229 inc_irq_stat(irq_resched_count);
230 scheduler_ipi();
231 trace_reschedule_exit(RESCHEDULE_VECTOR);
232}
233
234DEFINE_IDTENTRY_SYSVEC(sysvec_call_function)
235{
236 ack_APIC_irq();
237 trace_call_function_entry(CALL_FUNCTION_VECTOR);
238 inc_irq_stat(irq_call_count);
239 generic_smp_call_function_interrupt();
240 trace_call_function_exit(CALL_FUNCTION_VECTOR);
241}
242
243DEFINE_IDTENTRY_SYSVEC(sysvec_call_function_single)
244{
245 ack_APIC_irq();
246 trace_call_function_single_entry(CALL_FUNCTION_SINGLE_VECTOR);
247 inc_irq_stat(irq_call_count);
248 generic_smp_call_function_single_interrupt();
249 trace_call_function_single_exit(CALL_FUNCTION_SINGLE_VECTOR);
250}
251
252static int __init nonmi_ipi_setup(char *str)
253{
254 smp_no_nmi_ipi = true;
255 return 1;
256}
257
258__setup("nonmi_ipi", nonmi_ipi_setup);
259
260struct smp_ops smp_ops = {
261 .smp_prepare_boot_cpu = native_smp_prepare_boot_cpu,
262 .smp_prepare_cpus = native_smp_prepare_cpus,
263 .smp_cpus_done = native_smp_cpus_done,
264
265 .stop_other_cpus = native_stop_other_cpus,
266#if defined(CONFIG_KEXEC_CORE)
267 .crash_stop_other_cpus = kdump_nmi_shootdown_cpus,
268#endif
269 .smp_send_reschedule = native_smp_send_reschedule,
270
271 .cpu_up = native_cpu_up,
272 .cpu_die = native_cpu_die,
273 .cpu_disable = native_cpu_disable,
274 .play_dead = native_play_dead,
275
276 .send_call_func_ipi = native_send_call_func_ipi,
277 .send_call_func_single_ipi = native_send_call_func_single_ipi,
278};
279EXPORT_SYMBOL_GPL(smp_ops);
1/*
2 * Intel SMP support routines.
3 *
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998-99, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * (c) 2002,2003 Andi Kleen, SuSE Labs.
7 *
8 * i386 and x86_64 integration by Glauber Costa <gcosta@redhat.com>
9 *
10 * This code is released under the GNU General Public License version 2 or
11 * later.
12 */
13
14#include <linux/init.h>
15
16#include <linux/mm.h>
17#include <linux/delay.h>
18#include <linux/spinlock.h>
19#include <linux/export.h>
20#include <linux/kernel_stat.h>
21#include <linux/mc146818rtc.h>
22#include <linux/cache.h>
23#include <linux/interrupt.h>
24#include <linux/cpu.h>
25#include <linux/gfp.h>
26
27#include <asm/mtrr.h>
28#include <asm/tlbflush.h>
29#include <asm/mmu_context.h>
30#include <asm/proto.h>
31#include <asm/apic.h>
32#include <asm/nmi.h>
33#include <asm/mce.h>
34#include <asm/trace/irq_vectors.h>
35/*
36 * Some notes on x86 processor bugs affecting SMP operation:
37 *
38 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
39 * The Linux implications for SMP are handled as follows:
40 *
41 * Pentium III / [Xeon]
42 * None of the E1AP-E3AP errata are visible to the user.
43 *
44 * E1AP. see PII A1AP
45 * E2AP. see PII A2AP
46 * E3AP. see PII A3AP
47 *
48 * Pentium II / [Xeon]
49 * None of the A1AP-A3AP errata are visible to the user.
50 *
51 * A1AP. see PPro 1AP
52 * A2AP. see PPro 2AP
53 * A3AP. see PPro 7AP
54 *
55 * Pentium Pro
56 * None of 1AP-9AP errata are visible to the normal user,
57 * except occasional delivery of 'spurious interrupt' as trap #15.
58 * This is very rare and a non-problem.
59 *
60 * 1AP. Linux maps APIC as non-cacheable
61 * 2AP. worked around in hardware
62 * 3AP. fixed in C0 and above steppings microcode update.
63 * Linux does not use excessive STARTUP_IPIs.
64 * 4AP. worked around in hardware
65 * 5AP. symmetric IO mode (normal Linux operation) not affected.
66 * 'noapic' mode has vector 0xf filled out properly.
67 * 6AP. 'noapic' mode might be affected - fixed in later steppings
68 * 7AP. We do not assume writes to the LVT deassering IRQs
69 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
70 * 9AP. We do not use mixed mode
71 *
72 * Pentium
73 * There is a marginal case where REP MOVS on 100MHz SMP
74 * machines with B stepping processors can fail. XXX should provide
75 * an L1cache=Writethrough or L1cache=off option.
76 *
77 * B stepping CPUs may hang. There are hardware work arounds
78 * for this. We warn about it in case your board doesn't have the work
79 * arounds. Basically that's so I can tell anyone with a B stepping
80 * CPU and SMP problems "tough".
81 *
82 * Specific items [From Pentium Processor Specification Update]
83 *
84 * 1AP. Linux doesn't use remote read
85 * 2AP. Linux doesn't trust APIC errors
86 * 3AP. We work around this
87 * 4AP. Linux never generated 3 interrupts of the same priority
88 * to cause a lost local interrupt.
89 * 5AP. Remote read is never used
90 * 6AP. not affected - worked around in hardware
91 * 7AP. not affected - worked around in hardware
92 * 8AP. worked around in hardware - we get explicit CS errors if not
93 * 9AP. only 'noapic' mode affected. Might generate spurious
94 * interrupts, we log only the first one and count the
95 * rest silently.
96 * 10AP. not affected - worked around in hardware
97 * 11AP. Linux reads the APIC between writes to avoid this, as per
98 * the documentation. Make sure you preserve this as it affects
99 * the C stepping chips too.
100 * 12AP. not affected - worked around in hardware
101 * 13AP. not affected - worked around in hardware
102 * 14AP. we always deassert INIT during bootup
103 * 15AP. not affected - worked around in hardware
104 * 16AP. not affected - worked around in hardware
105 * 17AP. not affected - worked around in hardware
106 * 18AP. not affected - worked around in hardware
107 * 19AP. not affected - worked around in BIOS
108 *
109 * If this sounds worrying believe me these bugs are either ___RARE___,
110 * or are signal timing bugs worked around in hardware and there's
111 * about nothing of note with C stepping upwards.
112 */
113
114static atomic_t stopping_cpu = ATOMIC_INIT(-1);
115static bool smp_no_nmi_ipi = false;
116
117/*
118 * this function sends a 'reschedule' IPI to another CPU.
119 * it goes straight through and wastes no time serializing
120 * anything. Worst case is that we lose a reschedule ...
121 */
122static void native_smp_send_reschedule(int cpu)
123{
124 if (unlikely(cpu_is_offline(cpu))) {
125 WARN_ON(1);
126 return;
127 }
128 apic->send_IPI(cpu, RESCHEDULE_VECTOR);
129}
130
131void native_send_call_func_single_ipi(int cpu)
132{
133 apic->send_IPI(cpu, CALL_FUNCTION_SINGLE_VECTOR);
134}
135
136void native_send_call_func_ipi(const struct cpumask *mask)
137{
138 cpumask_var_t allbutself;
139
140 if (!alloc_cpumask_var(&allbutself, GFP_ATOMIC)) {
141 apic->send_IPI_mask(mask, CALL_FUNCTION_VECTOR);
142 return;
143 }
144
145 cpumask_copy(allbutself, cpu_online_mask);
146 cpumask_clear_cpu(smp_processor_id(), allbutself);
147
148 if (cpumask_equal(mask, allbutself) &&
149 cpumask_equal(cpu_online_mask, cpu_callout_mask))
150 apic->send_IPI_allbutself(CALL_FUNCTION_VECTOR);
151 else
152 apic->send_IPI_mask(mask, CALL_FUNCTION_VECTOR);
153
154 free_cpumask_var(allbutself);
155}
156
157static int smp_stop_nmi_callback(unsigned int val, struct pt_regs *regs)
158{
159 /* We are registered on stopping cpu too, avoid spurious NMI */
160 if (raw_smp_processor_id() == atomic_read(&stopping_cpu))
161 return NMI_HANDLED;
162
163 stop_this_cpu(NULL);
164
165 return NMI_HANDLED;
166}
167
168/*
169 * this function calls the 'stop' function on all other CPUs in the system.
170 */
171
172asmlinkage __visible void smp_reboot_interrupt(void)
173{
174 ipi_entering_ack_irq();
175 stop_this_cpu(NULL);
176 irq_exit();
177}
178
179static void native_stop_other_cpus(int wait)
180{
181 unsigned long flags;
182 unsigned long timeout;
183
184 if (reboot_force)
185 return;
186
187 /*
188 * Use an own vector here because smp_call_function
189 * does lots of things not suitable in a panic situation.
190 */
191
192 /*
193 * We start by using the REBOOT_VECTOR irq.
194 * The irq is treated as a sync point to allow critical
195 * regions of code on other cpus to release their spin locks
196 * and re-enable irqs. Jumping straight to an NMI might
197 * accidentally cause deadlocks with further shutdown/panic
198 * code. By syncing, we give the cpus up to one second to
199 * finish their work before we force them off with the NMI.
200 */
201 if (num_online_cpus() > 1) {
202 /* did someone beat us here? */
203 if (atomic_cmpxchg(&stopping_cpu, -1, safe_smp_processor_id()) != -1)
204 return;
205
206 /* sync above data before sending IRQ */
207 wmb();
208
209 apic->send_IPI_allbutself(REBOOT_VECTOR);
210
211 /*
212 * Don't wait longer than a second if the caller
213 * didn't ask us to wait.
214 */
215 timeout = USEC_PER_SEC;
216 while (num_online_cpus() > 1 && (wait || timeout--))
217 udelay(1);
218 }
219
220 /* if the REBOOT_VECTOR didn't work, try with the NMI */
221 if ((num_online_cpus() > 1) && (!smp_no_nmi_ipi)) {
222 if (register_nmi_handler(NMI_LOCAL, smp_stop_nmi_callback,
223 NMI_FLAG_FIRST, "smp_stop"))
224 /* Note: we ignore failures here */
225 /* Hope the REBOOT_IRQ is good enough */
226 goto finish;
227
228 /* sync above data before sending IRQ */
229 wmb();
230
231 pr_emerg("Shutting down cpus with NMI\n");
232
233 apic->send_IPI_allbutself(NMI_VECTOR);
234
235 /*
236 * Don't wait longer than a 10 ms if the caller
237 * didn't ask us to wait.
238 */
239 timeout = USEC_PER_MSEC * 10;
240 while (num_online_cpus() > 1 && (wait || timeout--))
241 udelay(1);
242 }
243
244finish:
245 local_irq_save(flags);
246 disable_local_APIC();
247 mcheck_cpu_clear(this_cpu_ptr(&cpu_info));
248 local_irq_restore(flags);
249}
250
251/*
252 * Reschedule call back.
253 */
254static inline void __smp_reschedule_interrupt(void)
255{
256 inc_irq_stat(irq_resched_count);
257 scheduler_ipi();
258}
259
260__visible void smp_reschedule_interrupt(struct pt_regs *regs)
261{
262 ack_APIC_irq();
263 __smp_reschedule_interrupt();
264 /*
265 * KVM uses this interrupt to force a cpu out of guest mode
266 */
267}
268
269__visible void smp_trace_reschedule_interrupt(struct pt_regs *regs)
270{
271 /*
272 * Need to call irq_enter() before calling the trace point.
273 * __smp_reschedule_interrupt() calls irq_enter/exit() too (in
274 * scheduler_ipi(). This is OK, since those functions are allowed
275 * to nest.
276 */
277 ipi_entering_ack_irq();
278 trace_reschedule_entry(RESCHEDULE_VECTOR);
279 __smp_reschedule_interrupt();
280 trace_reschedule_exit(RESCHEDULE_VECTOR);
281 exiting_irq();
282 /*
283 * KVM uses this interrupt to force a cpu out of guest mode
284 */
285}
286
287static inline void __smp_call_function_interrupt(void)
288{
289 generic_smp_call_function_interrupt();
290 inc_irq_stat(irq_call_count);
291}
292
293__visible void smp_call_function_interrupt(struct pt_regs *regs)
294{
295 ipi_entering_ack_irq();
296 __smp_call_function_interrupt();
297 exiting_irq();
298}
299
300__visible void smp_trace_call_function_interrupt(struct pt_regs *regs)
301{
302 ipi_entering_ack_irq();
303 trace_call_function_entry(CALL_FUNCTION_VECTOR);
304 __smp_call_function_interrupt();
305 trace_call_function_exit(CALL_FUNCTION_VECTOR);
306 exiting_irq();
307}
308
309static inline void __smp_call_function_single_interrupt(void)
310{
311 generic_smp_call_function_single_interrupt();
312 inc_irq_stat(irq_call_count);
313}
314
315__visible void smp_call_function_single_interrupt(struct pt_regs *regs)
316{
317 ipi_entering_ack_irq();
318 __smp_call_function_single_interrupt();
319 exiting_irq();
320}
321
322__visible void smp_trace_call_function_single_interrupt(struct pt_regs *regs)
323{
324 ipi_entering_ack_irq();
325 trace_call_function_single_entry(CALL_FUNCTION_SINGLE_VECTOR);
326 __smp_call_function_single_interrupt();
327 trace_call_function_single_exit(CALL_FUNCTION_SINGLE_VECTOR);
328 exiting_irq();
329}
330
331static int __init nonmi_ipi_setup(char *str)
332{
333 smp_no_nmi_ipi = true;
334 return 1;
335}
336
337__setup("nonmi_ipi", nonmi_ipi_setup);
338
339struct smp_ops smp_ops = {
340 .smp_prepare_boot_cpu = native_smp_prepare_boot_cpu,
341 .smp_prepare_cpus = native_smp_prepare_cpus,
342 .smp_cpus_done = native_smp_cpus_done,
343
344 .stop_other_cpus = native_stop_other_cpus,
345 .smp_send_reschedule = native_smp_send_reschedule,
346
347 .cpu_up = native_cpu_up,
348 .cpu_die = native_cpu_die,
349 .cpu_disable = native_cpu_disable,
350 .play_dead = native_play_dead,
351
352 .send_call_func_ipi = native_send_call_func_ipi,
353 .send_call_func_single_ipi = native_send_call_func_single_ipi,
354};
355EXPORT_SYMBOL_GPL(smp_ops);