2 * Intel SMP support routines.
4 * (c) 1995 Alan Cox, Building #3 <alan@redhat.com>
5 * (c) 1998-99, 2000 Ingo Molnar <mingo@redhat.com>
6 * (c) 2002,2003 Andi Kleen, SuSE Labs.
8 * This code is released under the GNU General Public License version 2 or
12 #include <linux/init.h>
15 #include <linux/delay.h>
16 #include <linux/spinlock.h>
17 #include <linux/smp_lock.h>
18 #include <linux/smp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mc146818rtc.h>
21 #include <linux/interrupt.h>
24 #include <asm/pgalloc.h>
25 #include <asm/tlbflush.h>
26 #include <asm/mach_apic.h>
27 #include <asm/mmu_context.h>
28 #include <asm/proto.h>
29 #include <asm/apicdef.h>
33 * Smarter SMP flushing macros.
36 * These mean you can really definitely utterly forget about
37 * writing to user space from interrupts. (Its not allowed anyway).
39 * Optimizations Manfred Spraul <manfred@colorfullife.com>
41 * More scalable flush, from Andi Kleen
43 * To avoid global state use 8 different call vectors.
44 * Each CPU uses a specific vector to trigger flushes on other
45 * CPUs. Depending on the received vector the target CPUs look into
46 * the right per cpu variable for the flush data.
48 * With more than 8 CPUs they are hashed to the 8 available
49 * vectors. The limited global vector space forces us to this right now.
50 * In future when interrupts are split into per CPU domains this could be
51 * fixed, at the cost of triggering multiple IPIs in some cases.
54 union smp_flush_state {
56 cpumask_t flush_cpumask;
57 struct mm_struct *flush_mm;
58 unsigned long flush_va;
59 #define FLUSH_ALL -1ULL
60 spinlock_t tlbstate_lock;
62 char pad[SMP_CACHE_BYTES];
63 } ____cacheline_aligned;
65 /* State is put into the per CPU data section, but padded
66 to a full cache line because other CPUs can access it and we don't
67 want false sharing in the per cpu data segment. */
68 static DEFINE_PER_CPU(union smp_flush_state, flush_state);
71 * We cannot call mmdrop() because we are in interrupt context,
72 * instead update mm->cpu_vm_mask.
74 static inline void leave_mm(int cpu)
76 if (read_pda(mmu_state) == TLBSTATE_OK)
78 cpu_clear(cpu, read_pda(active_mm)->cpu_vm_mask);
79 load_cr3(swapper_pg_dir);
84 * The flush IPI assumes that a thread switch happens in this order:
85 * [cpu0: the cpu that switches]
86 * 1) switch_mm() either 1a) or 1b)
87 * 1a) thread switch to a different mm
88 * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
89 * Stop ipi delivery for the old mm. This is not synchronized with
90 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
91 * for the wrong mm, and in the worst case we perform a superfluous
93 * 1a2) set cpu mmu_state to TLBSTATE_OK
94 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
95 * was in lazy tlb mode.
96 * 1a3) update cpu active_mm
97 * Now cpu0 accepts tlb flushes for the new mm.
98 * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
99 * Now the other cpus will send tlb flush ipis.
101 * 1b) thread switch without mm change
102 * cpu active_mm is correct, cpu0 already handles
104 * 1b1) set cpu mmu_state to TLBSTATE_OK
105 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
106 * Atomically set the bit [other cpus will start sending flush ipis],
108 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
109 * 2) switch %%esp, ie current
111 * The interrupt must handle 2 special cases:
112 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
113 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
114 * runs in kernel space, the cpu could load tlb entries for user space
117 * The good news is that cpu mmu_state is local to each cpu, no
118 * write/read ordering problems.
124 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
125 * 2) Leave the mm if we are in the lazy tlb mode.
127 * Interrupts are disabled.
130 asmlinkage void smp_invalidate_interrupt(struct pt_regs *regs)
134 union smp_flush_state *f;
136 cpu = smp_processor_id();
138 * orig_rax contains the negated interrupt vector.
139 * Use that to determine where the sender put the data.
141 sender = ~regs->orig_rax - INVALIDATE_TLB_VECTOR_START;
142 f = &per_cpu(flush_state, sender);
144 if (!cpu_isset(cpu, f->flush_cpumask))
147 * This was a BUG() but until someone can quote me the
148 * line from the intel manual that guarantees an IPI to
149 * multiple CPUs is retried _only_ on the erroring CPUs
150 * its staying as a return
155 if (f->flush_mm == read_pda(active_mm)) {
156 if (read_pda(mmu_state) == TLBSTATE_OK) {
157 if (f->flush_va == FLUSH_ALL)
160 __flush_tlb_one(f->flush_va);
166 cpu_clear(cpu, f->flush_cpumask);
169 static void flush_tlb_others(cpumask_t cpumask, struct mm_struct *mm,
173 union smp_flush_state *f;
175 /* Caller has disabled preemption */
176 sender = smp_processor_id() % NUM_INVALIDATE_TLB_VECTORS;
177 f = &per_cpu(flush_state, sender);
179 /* Could avoid this lock when
180 num_online_cpus() <= NUM_INVALIDATE_TLB_VECTORS, but it is
181 probably not worth checking this for a cache-hot lock. */
182 spin_lock(&f->tlbstate_lock);
186 cpus_or(f->flush_cpumask, cpumask, f->flush_cpumask);
189 * We have to send the IPI only to
192 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR_START + sender);
194 while (!cpus_empty(f->flush_cpumask))
199 spin_unlock(&f->tlbstate_lock);
202 int __cpuinit init_smp_flush(void)
205 for_each_cpu_mask(i, cpu_possible_map) {
206 spin_lock_init(&per_cpu(flush_state.tlbstate_lock, i));
211 core_initcall(init_smp_flush);
213 void flush_tlb_current_task(void)
215 struct mm_struct *mm = current->mm;
219 cpu_mask = mm->cpu_vm_mask;
220 cpu_clear(smp_processor_id(), cpu_mask);
223 if (!cpus_empty(cpu_mask))
224 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
227 EXPORT_SYMBOL(flush_tlb_current_task);
229 void flush_tlb_mm (struct mm_struct * mm)
234 cpu_mask = mm->cpu_vm_mask;
235 cpu_clear(smp_processor_id(), cpu_mask);
237 if (current->active_mm == mm) {
241 leave_mm(smp_processor_id());
243 if (!cpus_empty(cpu_mask))
244 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
248 EXPORT_SYMBOL(flush_tlb_mm);
250 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
252 struct mm_struct *mm = vma->vm_mm;
256 cpu_mask = mm->cpu_vm_mask;
257 cpu_clear(smp_processor_id(), cpu_mask);
259 if (current->active_mm == mm) {
263 leave_mm(smp_processor_id());
266 if (!cpus_empty(cpu_mask))
267 flush_tlb_others(cpu_mask, mm, va);
271 EXPORT_SYMBOL(flush_tlb_page);
273 static void do_flush_tlb_all(void* info)
275 unsigned long cpu = smp_processor_id();
278 if (read_pda(mmu_state) == TLBSTATE_LAZY)
282 void flush_tlb_all(void)
284 on_each_cpu(do_flush_tlb_all, NULL, 1, 1);
288 * this function sends a 'reschedule' IPI to another CPU.
289 * it goes straight through and wastes no time serializing
290 * anything. Worst case is that we lose a reschedule ...
293 void smp_send_reschedule(int cpu)
295 send_IPI_mask(cpumask_of_cpu(cpu), RESCHEDULE_VECTOR);
299 * Structure and data for smp_call_function(). This is designed to minimise
300 * static memory requirements. It also looks cleaner.
302 static DEFINE_SPINLOCK(call_lock);
304 struct call_data_struct {
305 void (*func) (void *info);
312 static struct call_data_struct * call_data;
314 void lock_ipi_call_lock(void)
316 spin_lock_irq(&call_lock);
319 void unlock_ipi_call_lock(void)
321 spin_unlock_irq(&call_lock);
325 * this function sends a 'generic call function' IPI to one other CPU
328 * cpu is a standard Linux logical CPU number.
331 __smp_call_function_single(int cpu, void (*func) (void *info), void *info,
332 int nonatomic, int wait)
334 struct call_data_struct data;
339 atomic_set(&data.started, 0);
342 atomic_set(&data.finished, 0);
346 /* Send a message to all other CPUs and wait for them to respond */
347 send_IPI_mask(cpumask_of_cpu(cpu), CALL_FUNCTION_VECTOR);
349 /* Wait for response */
350 while (atomic_read(&data.started) != cpus)
356 while (atomic_read(&data.finished) != cpus)
361 * smp_call_function_single - Run a function on another CPU
362 * @func: The function to run. This must be fast and non-blocking.
363 * @info: An arbitrary pointer to pass to the function.
364 * @nonatomic: Currently unused.
365 * @wait: If true, wait until function has completed on other CPUs.
367 * Retrurns 0 on success, else a negative status code.
369 * Does not return until the remote CPU is nearly ready to execute <func>
370 * or is or has executed.
373 int smp_call_function_single (int cpu, void (*func) (void *info), void *info,
374 int nonatomic, int wait)
376 /* prevent preemption and reschedule on another processor */
383 spin_lock_bh(&call_lock);
384 __smp_call_function_single(cpu, func, info, nonatomic, wait);
385 spin_unlock_bh(&call_lock);
391 * this function sends a 'generic call function' IPI to all other CPUs
394 static void __smp_call_function (void (*func) (void *info), void *info,
395 int nonatomic, int wait)
397 struct call_data_struct data;
398 int cpus = num_online_cpus()-1;
405 atomic_set(&data.started, 0);
408 atomic_set(&data.finished, 0);
412 /* Send a message to all other CPUs and wait for them to respond */
413 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
415 /* Wait for response */
416 while (atomic_read(&data.started) != cpus)
422 while (atomic_read(&data.finished) != cpus)
427 * smp_call_function - run a function on all other CPUs.
428 * @func: The function to run. This must be fast and non-blocking.
429 * @info: An arbitrary pointer to pass to the function.
430 * @nonatomic: currently unused.
431 * @wait: If true, wait (atomically) until function has completed on other
434 * Returns 0 on success, else a negative status code. Does not return until
435 * remote CPUs are nearly ready to execute func or are or have executed.
437 * You must not call this function with disabled interrupts or from a
438 * hardware interrupt handler or from a bottom half handler.
439 * Actually there are a few legal cases, like panic.
441 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
444 spin_lock(&call_lock);
445 __smp_call_function(func,info,nonatomic,wait);
446 spin_unlock(&call_lock);
449 EXPORT_SYMBOL(smp_call_function);
451 void smp_stop_cpu(void)
457 cpu_clear(smp_processor_id(), cpu_online_map);
458 local_irq_save(flags);
459 disable_local_APIC();
460 local_irq_restore(flags);
463 static void smp_really_stop_cpu(void *dummy)
470 void smp_send_stop(void)
475 /* Don't deadlock on the call lock in panic */
476 if (!spin_trylock(&call_lock)) {
477 /* ignore locking because we have panicked anyways */
480 __smp_call_function(smp_really_stop_cpu, NULL, 0, 0);
482 spin_unlock(&call_lock);
485 disable_local_APIC();
490 * Reschedule call back. Nothing to do,
491 * all the work is done automatically when
492 * we return from the interrupt.
494 asmlinkage void smp_reschedule_interrupt(void)
499 asmlinkage void smp_call_function_interrupt(void)
501 void (*func) (void *info) = call_data->func;
502 void *info = call_data->info;
503 int wait = call_data->wait;
507 * Notify initiating CPU that I've grabbed the data and am
508 * about to execute the function
511 atomic_inc(&call_data->started);
513 * At this point the info structure may be out of scope unless wait==1
521 atomic_inc(&call_data->finished);
525 int safe_smp_processor_id(void)
532 apicid = hard_smp_processor_id();
533 if (apicid < NR_CPUS && x86_cpu_to_apicid[apicid] == apicid)
536 for (i = 0; i < NR_CPUS; ++i) {
537 if (x86_cpu_to_apicid[i] == apicid)
541 /* No entries in x86_cpu_to_apicid? Either no MPS|ACPI,
542 * or called too early. Either way, we must be CPU 0. */
543 if (x86_cpu_to_apicid[0] == BAD_APICID)
546 return 0; /* Should not happen */