]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - arch/x86/kernel/smp.c
x86, NMI: Add knob to disable using NMI IPIs to stop cpus
[mirror_ubuntu-artful-kernel.git] / arch / x86 / kernel / smp.c
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 /*
34 * Some notes on x86 processor bugs affecting SMP operation:
35 *
36 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
37 * The Linux implications for SMP are handled as follows:
38 *
39 * Pentium III / [Xeon]
40 * None of the E1AP-E3AP errata are visible to the user.
41 *
42 * E1AP. see PII A1AP
43 * E2AP. see PII A2AP
44 * E3AP. see PII A3AP
45 *
46 * Pentium II / [Xeon]
47 * None of the A1AP-A3AP errata are visible to the user.
48 *
49 * A1AP. see PPro 1AP
50 * A2AP. see PPro 2AP
51 * A3AP. see PPro 7AP
52 *
53 * Pentium Pro
54 * None of 1AP-9AP errata are visible to the normal user,
55 * except occasional delivery of 'spurious interrupt' as trap #15.
56 * This is very rare and a non-problem.
57 *
58 * 1AP. Linux maps APIC as non-cacheable
59 * 2AP. worked around in hardware
60 * 3AP. fixed in C0 and above steppings microcode update.
61 * Linux does not use excessive STARTUP_IPIs.
62 * 4AP. worked around in hardware
63 * 5AP. symmetric IO mode (normal Linux operation) not affected.
64 * 'noapic' mode has vector 0xf filled out properly.
65 * 6AP. 'noapic' mode might be affected - fixed in later steppings
66 * 7AP. We do not assume writes to the LVT deassering IRQs
67 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
68 * 9AP. We do not use mixed mode
69 *
70 * Pentium
71 * There is a marginal case where REP MOVS on 100MHz SMP
72 * machines with B stepping processors can fail. XXX should provide
73 * an L1cache=Writethrough or L1cache=off option.
74 *
75 * B stepping CPUs may hang. There are hardware work arounds
76 * for this. We warn about it in case your board doesn't have the work
77 * arounds. Basically that's so I can tell anyone with a B stepping
78 * CPU and SMP problems "tough".
79 *
80 * Specific items [From Pentium Processor Specification Update]
81 *
82 * 1AP. Linux doesn't use remote read
83 * 2AP. Linux doesn't trust APIC errors
84 * 3AP. We work around this
85 * 4AP. Linux never generated 3 interrupts of the same priority
86 * to cause a lost local interrupt.
87 * 5AP. Remote read is never used
88 * 6AP. not affected - worked around in hardware
89 * 7AP. not affected - worked around in hardware
90 * 8AP. worked around in hardware - we get explicit CS errors if not
91 * 9AP. only 'noapic' mode affected. Might generate spurious
92 * interrupts, we log only the first one and count the
93 * rest silently.
94 * 10AP. not affected - worked around in hardware
95 * 11AP. Linux reads the APIC between writes to avoid this, as per
96 * the documentation. Make sure you preserve this as it affects
97 * the C stepping chips too.
98 * 12AP. not affected - worked around in hardware
99 * 13AP. not affected - worked around in hardware
100 * 14AP. we always deassert INIT during bootup
101 * 15AP. not affected - worked around in hardware
102 * 16AP. not affected - worked around in hardware
103 * 17AP. not affected - worked around in hardware
104 * 18AP. not affected - worked around in hardware
105 * 19AP. not affected - worked around in BIOS
106 *
107 * If this sounds worrying believe me these bugs are either ___RARE___,
108 * or are signal timing bugs worked around in hardware and there's
109 * about nothing of note with C stepping upwards.
110 */
111
112 /*
113 * this function sends a 'reschedule' IPI to another CPU.
114 * it goes straight through and wastes no time serializing
115 * anything. Worst case is that we lose a reschedule ...
116 */
117 static void native_smp_send_reschedule(int cpu)
118 {
119 if (unlikely(cpu_is_offline(cpu))) {
120 WARN_ON(1);
121 return;
122 }
123 apic->send_IPI_mask(cpumask_of(cpu), RESCHEDULE_VECTOR);
124 }
125
126 void native_send_call_func_single_ipi(int cpu)
127 {
128 apic->send_IPI_mask(cpumask_of(cpu), CALL_FUNCTION_SINGLE_VECTOR);
129 }
130
131 void native_send_call_func_ipi(const struct cpumask *mask)
132 {
133 cpumask_var_t allbutself;
134
135 if (!alloc_cpumask_var(&allbutself, GFP_ATOMIC)) {
136 apic->send_IPI_mask(mask, CALL_FUNCTION_VECTOR);
137 return;
138 }
139
140 cpumask_copy(allbutself, cpu_online_mask);
141 cpumask_clear_cpu(smp_processor_id(), allbutself);
142
143 if (cpumask_equal(mask, allbutself) &&
144 cpumask_equal(cpu_online_mask, cpu_callout_mask))
145 apic->send_IPI_allbutself(CALL_FUNCTION_VECTOR);
146 else
147 apic->send_IPI_mask(mask, CALL_FUNCTION_VECTOR);
148
149 free_cpumask_var(allbutself);
150 }
151
152 static atomic_t stopping_cpu = ATOMIC_INIT(-1);
153
154 static int smp_stop_nmi_callback(unsigned int val, struct pt_regs *regs)
155 {
156 /* We are registered on stopping cpu too, avoid spurious NMI */
157 if (raw_smp_processor_id() == atomic_read(&stopping_cpu))
158 return NMI_HANDLED;
159
160 stop_this_cpu(NULL);
161
162 return NMI_HANDLED;
163 }
164
165 static void native_nmi_stop_other_cpus(int wait)
166 {
167 unsigned long flags;
168 unsigned long timeout;
169
170 if (reboot_force)
171 return;
172
173 /*
174 * Use an own vector here because smp_call_function
175 * does lots of things not suitable in a panic situation.
176 */
177 if (num_online_cpus() > 1) {
178 /* did someone beat us here? */
179 if (atomic_cmpxchg(&stopping_cpu, -1, safe_smp_processor_id() != -1))
180 return;
181
182 if (register_nmi_handler(NMI_LOCAL, smp_stop_nmi_callback,
183 NMI_FLAG_FIRST, "smp_stop"))
184 /* Note: we ignore failures here */
185 return;
186
187 /* sync above data before sending NMI */
188 wmb();
189
190 apic->send_IPI_allbutself(NMI_VECTOR);
191
192 /*
193 * Don't wait longer than a second if the caller
194 * didn't ask us to wait.
195 */
196 timeout = USEC_PER_SEC;
197 while (num_online_cpus() > 1 && (wait || timeout--))
198 udelay(1);
199 }
200
201 local_irq_save(flags);
202 disable_local_APIC();
203 local_irq_restore(flags);
204 }
205
206 /*
207 * this function calls the 'stop' function on all other CPUs in the system.
208 */
209
210 asmlinkage void smp_reboot_interrupt(void)
211 {
212 ack_APIC_irq();
213 irq_enter();
214 stop_this_cpu(NULL);
215 irq_exit();
216 }
217
218 static void native_irq_stop_other_cpus(int wait)
219 {
220 unsigned long flags;
221 unsigned long timeout;
222
223 if (reboot_force)
224 return;
225
226 /*
227 * Use an own vector here because smp_call_function
228 * does lots of things not suitable in a panic situation.
229 * On most systems we could also use an NMI here,
230 * but there are a few systems around where NMI
231 * is problematic so stay with an non NMI for now
232 * (this implies we cannot stop CPUs spinning with irq off
233 * currently)
234 */
235 if (num_online_cpus() > 1) {
236 apic->send_IPI_allbutself(REBOOT_VECTOR);
237
238 /*
239 * Don't wait longer than a second if the caller
240 * didn't ask us to wait.
241 */
242 timeout = USEC_PER_SEC;
243 while (num_online_cpus() > 1 && (wait || timeout--))
244 udelay(1);
245 }
246
247 local_irq_save(flags);
248 disable_local_APIC();
249 local_irq_restore(flags);
250 }
251
252 static void native_smp_disable_nmi_ipi(void)
253 {
254 smp_ops.stop_other_cpus = native_irq_stop_other_cpus;
255 }
256
257 /*
258 * Reschedule call back.
259 */
260 void smp_reschedule_interrupt(struct pt_regs *regs)
261 {
262 ack_APIC_irq();
263 inc_irq_stat(irq_resched_count);
264 scheduler_ipi();
265 /*
266 * KVM uses this interrupt to force a cpu out of guest mode
267 */
268 }
269
270 void smp_call_function_interrupt(struct pt_regs *regs)
271 {
272 ack_APIC_irq();
273 irq_enter();
274 generic_smp_call_function_interrupt();
275 inc_irq_stat(irq_call_count);
276 irq_exit();
277 }
278
279 void smp_call_function_single_interrupt(struct pt_regs *regs)
280 {
281 ack_APIC_irq();
282 irq_enter();
283 generic_smp_call_function_single_interrupt();
284 inc_irq_stat(irq_call_count);
285 irq_exit();
286 }
287
288 static int __init nonmi_ipi_setup(char *str)
289 {
290 native_smp_disable_nmi_ipi();
291 return 1;
292 }
293
294 __setup("nonmi_ipi", nonmi_ipi_setup);
295
296 struct smp_ops smp_ops = {
297 .smp_prepare_boot_cpu = native_smp_prepare_boot_cpu,
298 .smp_prepare_cpus = native_smp_prepare_cpus,
299 .smp_cpus_done = native_smp_cpus_done,
300
301 .stop_other_cpus = native_nmi_stop_other_cpus,
302 .smp_send_reschedule = native_smp_send_reschedule,
303
304 .cpu_up = native_cpu_up,
305 .cpu_die = native_cpu_die,
306 .cpu_disable = native_cpu_disable,
307 .play_dead = native_play_dead,
308
309 .send_call_func_ipi = native_send_call_func_ipi,
310 .send_call_func_single_ipi = native_send_call_func_single_ipi,
311 };
312 EXPORT_SYMBOL_GPL(smp_ops);