1 | /* |
---|
2 | * Intel SMP support routines. |
---|
3 | * |
---|
4 | * (c) 1995 Alan Cox, Building #3 <alan@redhat.com> |
---|
5 | * (c) 1998-99, 2000 Ingo Molnar <mingo@redhat.com> |
---|
6 | * |
---|
7 | * This code is released under the GNU General Public License version 2 or |
---|
8 | * later. |
---|
9 | */ |
---|
10 | |
---|
11 | #include <xen/config.h> |
---|
12 | #include <xen/irq.h> |
---|
13 | #include <xen/sched.h> |
---|
14 | #include <xen/delay.h> |
---|
15 | #include <xen/perfc.h> |
---|
16 | #include <xen/spinlock.h> |
---|
17 | #include <asm/current.h> |
---|
18 | #include <asm/smp.h> |
---|
19 | #include <asm/mc146818rtc.h> |
---|
20 | #include <asm/flushtlb.h> |
---|
21 | #include <asm/smpboot.h> |
---|
22 | #include <asm/hardirq.h> |
---|
23 | #include <asm/ipi.h> |
---|
24 | #include <asm/hvm/support.h> |
---|
25 | #include <mach_apic.h> |
---|
26 | |
---|
27 | /* |
---|
28 | * Some notes on x86 processor bugs affecting SMP operation: |
---|
29 | * |
---|
30 | * Pentium, Pentium Pro, II, III (and all CPUs) have bugs. |
---|
31 | * The Linux implications for SMP are handled as follows: |
---|
32 | * |
---|
33 | * Pentium III / [Xeon] |
---|
34 | * None of the E1AP-E3AP errata are visible to the user. |
---|
35 | * |
---|
36 | * E1AP. see PII A1AP |
---|
37 | * E2AP. see PII A2AP |
---|
38 | * E3AP. see PII A3AP |
---|
39 | * |
---|
40 | * Pentium II / [Xeon] |
---|
41 | * None of the A1AP-A3AP errata are visible to the user. |
---|
42 | * |
---|
43 | * A1AP. see PPro 1AP |
---|
44 | * A2AP. see PPro 2AP |
---|
45 | * A3AP. see PPro 7AP |
---|
46 | * |
---|
47 | * Pentium Pro |
---|
48 | * None of 1AP-9AP errata are visible to the normal user, |
---|
49 | * except occasional delivery of 'spurious interrupt' as trap #15. |
---|
50 | * This is very rare and a non-problem. |
---|
51 | * |
---|
52 | * 1AP. Linux maps APIC as non-cacheable |
---|
53 | * 2AP. worked around in hardware |
---|
54 | * 3AP. fixed in C0 and above steppings microcode update. |
---|
55 | * Linux does not use excessive STARTUP_IPIs. |
---|
56 | * 4AP. worked around in hardware |
---|
57 | * 5AP. symmetric IO mode (normal Linux operation) not affected. |
---|
58 | * 'noapic' mode has vector 0xf filled out properly. |
---|
59 | * 6AP. 'noapic' mode might be affected - fixed in later steppings |
---|
60 | * 7AP. We do not assume writes to the LVT deassering IRQs |
---|
61 | * 8AP. We do not enable low power mode (deep sleep) during MP bootup |
---|
62 | * 9AP. We do not use mixed mode |
---|
63 | */ |
---|
64 | |
---|
65 | /* |
---|
66 | * The following functions deal with sending IPIs between CPUs. |
---|
67 | */ |
---|
68 | |
---|
69 | static inline int __prepare_ICR (unsigned int shortcut, int vector) |
---|
70 | { |
---|
71 | return APIC_DM_FIXED | shortcut | vector; |
---|
72 | } |
---|
73 | |
---|
74 | static inline int __prepare_ICR2 (unsigned int mask) |
---|
75 | { |
---|
76 | return SET_APIC_DEST_FIELD(mask); |
---|
77 | } |
---|
78 | |
---|
79 | static inline void check_IPI_mask(cpumask_t cpumask) |
---|
80 | { |
---|
81 | /* |
---|
82 | * Sanity, and necessary. An IPI with no target generates a send accept |
---|
83 | * error with Pentium and P6 APICs. |
---|
84 | */ |
---|
85 | ASSERT(cpus_subset(cpumask, cpu_online_map)); |
---|
86 | ASSERT(!cpus_empty(cpumask)); |
---|
87 | } |
---|
88 | |
---|
89 | void send_IPI_mask_flat(cpumask_t cpumask, int vector) |
---|
90 | { |
---|
91 | unsigned long mask = cpus_addr(cpumask)[0]; |
---|
92 | unsigned long cfg; |
---|
93 | unsigned long flags; |
---|
94 | |
---|
95 | check_IPI_mask(cpumask); |
---|
96 | |
---|
97 | local_irq_save(flags); |
---|
98 | |
---|
99 | /* |
---|
100 | * Wait for idle. |
---|
101 | */ |
---|
102 | apic_wait_icr_idle(); |
---|
103 | |
---|
104 | /* |
---|
105 | * prepare target chip field |
---|
106 | */ |
---|
107 | cfg = __prepare_ICR2(mask); |
---|
108 | apic_write_around(APIC_ICR2, cfg); |
---|
109 | |
---|
110 | /* |
---|
111 | * program the ICR |
---|
112 | */ |
---|
113 | cfg = __prepare_ICR(0, vector) | APIC_DEST_LOGICAL; |
---|
114 | |
---|
115 | /* |
---|
116 | * Send the IPI. The write to APIC_ICR fires this off. |
---|
117 | */ |
---|
118 | apic_write_around(APIC_ICR, cfg); |
---|
119 | |
---|
120 | local_irq_restore(flags); |
---|
121 | } |
---|
122 | |
---|
123 | void send_IPI_mask_phys(cpumask_t mask, int vector) |
---|
124 | { |
---|
125 | unsigned long cfg, flags; |
---|
126 | unsigned int query_cpu; |
---|
127 | |
---|
128 | check_IPI_mask(mask); |
---|
129 | |
---|
130 | /* |
---|
131 | * Hack. The clustered APIC addressing mode doesn't allow us to send |
---|
132 | * to an arbitrary mask, so I do a unicasts to each CPU instead. This |
---|
133 | * should be modified to do 1 message per cluster ID - mbligh |
---|
134 | */ |
---|
135 | |
---|
136 | local_irq_save(flags); |
---|
137 | |
---|
138 | for_each_cpu_mask( query_cpu, mask ) |
---|
139 | { |
---|
140 | /* |
---|
141 | * Wait for idle. |
---|
142 | */ |
---|
143 | apic_wait_icr_idle(); |
---|
144 | |
---|
145 | /* |
---|
146 | * prepare target chip field |
---|
147 | */ |
---|
148 | cfg = __prepare_ICR2(cpu_physical_id(query_cpu)); |
---|
149 | apic_write_around(APIC_ICR2, cfg); |
---|
150 | |
---|
151 | /* |
---|
152 | * program the ICR |
---|
153 | */ |
---|
154 | cfg = __prepare_ICR(0, vector) | APIC_DEST_PHYSICAL; |
---|
155 | |
---|
156 | /* |
---|
157 | * Send the IPI. The write to APIC_ICR fires this off. |
---|
158 | */ |
---|
159 | apic_write_around(APIC_ICR, cfg); |
---|
160 | } |
---|
161 | |
---|
162 | local_irq_restore(flags); |
---|
163 | } |
---|
164 | |
---|
165 | static DEFINE_SPINLOCK(flush_lock); |
---|
166 | static cpumask_t flush_cpumask; |
---|
167 | static unsigned long flush_va; |
---|
168 | |
---|
169 | fastcall void smp_invalidate_interrupt(void) |
---|
170 | { |
---|
171 | ack_APIC_irq(); |
---|
172 | perfc_incr(ipis); |
---|
173 | irq_enter(); |
---|
174 | if ( !__sync_lazy_execstate() ) |
---|
175 | { |
---|
176 | if ( flush_va == FLUSHVA_ALL ) |
---|
177 | local_flush_tlb(); |
---|
178 | else |
---|
179 | local_flush_tlb_one(flush_va); |
---|
180 | } |
---|
181 | cpu_clear(smp_processor_id(), flush_cpumask); |
---|
182 | irq_exit(); |
---|
183 | } |
---|
184 | |
---|
185 | void __flush_tlb_mask(cpumask_t mask, unsigned long va) |
---|
186 | { |
---|
187 | ASSERT(local_irq_is_enabled()); |
---|
188 | |
---|
189 | if ( cpu_isset(smp_processor_id(), mask) ) |
---|
190 | { |
---|
191 | if ( va == FLUSHVA_ALL ) |
---|
192 | local_flush_tlb(); |
---|
193 | else |
---|
194 | local_flush_tlb_one(va); |
---|
195 | cpu_clear(smp_processor_id(), mask); |
---|
196 | } |
---|
197 | |
---|
198 | if ( !cpus_empty(mask) ) |
---|
199 | { |
---|
200 | spin_lock(&flush_lock); |
---|
201 | flush_cpumask = mask; |
---|
202 | flush_va = va; |
---|
203 | send_IPI_mask(mask, INVALIDATE_TLB_VECTOR); |
---|
204 | while ( !cpus_empty(flush_cpumask) ) |
---|
205 | cpu_relax(); |
---|
206 | spin_unlock(&flush_lock); |
---|
207 | } |
---|
208 | } |
---|
209 | |
---|
210 | /* Call with no locks held and interrupts enabled (e.g., softirq context). */ |
---|
211 | void new_tlbflush_clock_period(void) |
---|
212 | { |
---|
213 | cpumask_t allbutself; |
---|
214 | |
---|
215 | /* Flush everyone else. We definitely flushed just before entry. */ |
---|
216 | allbutself = cpu_online_map; |
---|
217 | cpu_clear(smp_processor_id(), allbutself); |
---|
218 | __flush_tlb_mask(allbutself, FLUSHVA_ALL); |
---|
219 | |
---|
220 | /* No need for atomicity: we are the only possible updater. */ |
---|
221 | ASSERT(tlbflush_clock == 0); |
---|
222 | tlbflush_clock++; |
---|
223 | } |
---|
224 | |
---|
225 | static void flush_tlb_all_pge_ipi(void *info) |
---|
226 | { |
---|
227 | local_flush_tlb_pge(); |
---|
228 | } |
---|
229 | |
---|
230 | void flush_tlb_all_pge(void) |
---|
231 | { |
---|
232 | smp_call_function(flush_tlb_all_pge_ipi, 0, 1, 1); |
---|
233 | local_flush_tlb_pge(); |
---|
234 | } |
---|
235 | |
---|
236 | void smp_send_event_check_mask(cpumask_t mask) |
---|
237 | { |
---|
238 | cpu_clear(smp_processor_id(), mask); |
---|
239 | if ( !cpus_empty(mask) ) |
---|
240 | send_IPI_mask(mask, EVENT_CHECK_VECTOR); |
---|
241 | } |
---|
242 | |
---|
243 | /* |
---|
244 | * Structure and data for smp_call_function()/on_selected_cpus(). |
---|
245 | */ |
---|
246 | |
---|
247 | struct call_data_struct { |
---|
248 | void (*func) (void *info); |
---|
249 | void *info; |
---|
250 | int wait; |
---|
251 | atomic_t started; |
---|
252 | atomic_t finished; |
---|
253 | cpumask_t selected; |
---|
254 | }; |
---|
255 | |
---|
256 | static DEFINE_SPINLOCK(call_lock); |
---|
257 | static struct call_data_struct *call_data; |
---|
258 | |
---|
259 | int smp_call_function( |
---|
260 | void (*func) (void *info), |
---|
261 | void *info, |
---|
262 | int retry, |
---|
263 | int wait) |
---|
264 | { |
---|
265 | cpumask_t allbutself = cpu_online_map; |
---|
266 | cpu_clear(smp_processor_id(), allbutself); |
---|
267 | return on_selected_cpus(allbutself, func, info, retry, wait); |
---|
268 | } |
---|
269 | |
---|
270 | int on_selected_cpus( |
---|
271 | cpumask_t selected, |
---|
272 | void (*func) (void *info), |
---|
273 | void *info, |
---|
274 | int retry, |
---|
275 | int wait) |
---|
276 | { |
---|
277 | struct call_data_struct data; |
---|
278 | unsigned int nr_cpus = cpus_weight(selected); |
---|
279 | |
---|
280 | ASSERT(local_irq_is_enabled()); |
---|
281 | |
---|
282 | if ( nr_cpus == 0 ) |
---|
283 | return 0; |
---|
284 | |
---|
285 | data.func = func; |
---|
286 | data.info = info; |
---|
287 | data.wait = wait; |
---|
288 | atomic_set(&data.started, 0); |
---|
289 | atomic_set(&data.finished, 0); |
---|
290 | data.selected = selected; |
---|
291 | |
---|
292 | spin_lock(&call_lock); |
---|
293 | |
---|
294 | call_data = &data; |
---|
295 | wmb(); |
---|
296 | |
---|
297 | send_IPI_mask(selected, CALL_FUNCTION_VECTOR); |
---|
298 | |
---|
299 | while ( atomic_read(wait ? &data.finished : &data.started) != nr_cpus ) |
---|
300 | cpu_relax(); |
---|
301 | |
---|
302 | spin_unlock(&call_lock); |
---|
303 | |
---|
304 | return 0; |
---|
305 | } |
---|
306 | |
---|
307 | static void stop_this_cpu (void *dummy) |
---|
308 | { |
---|
309 | cpu_clear(smp_processor_id(), cpu_online_map); |
---|
310 | |
---|
311 | local_irq_disable(); |
---|
312 | disable_local_APIC(); |
---|
313 | hvm_disable(); |
---|
314 | |
---|
315 | for ( ; ; ) |
---|
316 | __asm__ __volatile__ ( "hlt" ); |
---|
317 | } |
---|
318 | |
---|
319 | void smp_send_stop(void) |
---|
320 | { |
---|
321 | /* Stop all other CPUs in the system. */ |
---|
322 | smp_call_function(stop_this_cpu, NULL, 1, 0); |
---|
323 | |
---|
324 | local_irq_disable(); |
---|
325 | disable_local_APIC(); |
---|
326 | local_irq_enable(); |
---|
327 | } |
---|
328 | |
---|
329 | fastcall void smp_event_check_interrupt(struct cpu_user_regs *regs) |
---|
330 | { |
---|
331 | ack_APIC_irq(); |
---|
332 | perfc_incr(ipis); |
---|
333 | } |
---|
334 | |
---|
335 | fastcall void smp_call_function_interrupt(struct cpu_user_regs *regs) |
---|
336 | { |
---|
337 | void (*func)(void *info) = call_data->func; |
---|
338 | void *info = call_data->info; |
---|
339 | |
---|
340 | ack_APIC_irq(); |
---|
341 | perfc_incr(ipis); |
---|
342 | |
---|
343 | if ( !cpu_isset(smp_processor_id(), call_data->selected) ) |
---|
344 | return; |
---|
345 | |
---|
346 | irq_enter(); |
---|
347 | |
---|
348 | if ( call_data->wait ) |
---|
349 | { |
---|
350 | (*func)(info); |
---|
351 | mb(); |
---|
352 | atomic_inc(&call_data->finished); |
---|
353 | } |
---|
354 | else |
---|
355 | { |
---|
356 | mb(); |
---|
357 | atomic_inc(&call_data->started); |
---|
358 | (*func)(info); |
---|
359 | } |
---|
360 | |
---|
361 | irq_exit(); |
---|
362 | } |
---|