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