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/irq.h>
16 #include <linux/delay.h>
17 #include <linux/spinlock.h>
18 #include <linux/smp_lock.h>
19 #include <linux/smp.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/mc146818rtc.h>
22 #include <linux/interrupt.h>
25 #include <asm/pgalloc.h>
26 #include <asm/tlbflush.h>
27 #include <asm/mach_apic.h>
30 * Smarter SMP flushing macros.
33 * These mean you can really definitely utterly forget about
34 * writing to user space from interrupts. (Its not allowed anyway).
36 * Optimizations Manfred Spraul <manfred@colorfullife.com>
39 static cpumask_t flush_cpumask;
40 static struct mm_struct * flush_mm;
41 static unsigned long flush_va;
42 static spinlock_t tlbstate_lock = SPIN_LOCK_UNLOCKED;
43 #define FLUSH_ALL 0xffffffff
46 * We cannot call mmdrop() because we are in interrupt context,
47 * instead update mm->cpu_vm_mask.
49 static inline void leave_mm (unsigned long cpu)
51 if (read_pda(mmu_state) == TLBSTATE_OK)
53 clear_bit(cpu, &read_pda(active_mm)->cpu_vm_mask);
59 * The flush IPI assumes that a thread switch happens in this order:
60 * [cpu0: the cpu that switches]
61 * 1) switch_mm() either 1a) or 1b)
62 * 1a) thread switch to a different mm
63 * 1a1) clear_bit(cpu, &old_mm->cpu_vm_mask);
64 * Stop ipi delivery for the old mm. This is not synchronized with
65 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
66 * for the wrong mm, and in the worst case we perform a superfluous
68 * 1a2) set cpu mmu_state to TLBSTATE_OK
69 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
70 * was in lazy tlb mode.
71 * 1a3) update cpu active_mm
72 * Now cpu0 accepts tlb flushes for the new mm.
73 * 1a4) set_bit(cpu, &new_mm->cpu_vm_mask);
74 * Now the other cpus will send tlb flush ipis.
76 * 1b) thread switch without mm change
77 * cpu active_mm is correct, cpu0 already handles
79 * 1b1) set cpu mmu_state to TLBSTATE_OK
80 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
81 * Atomically set the bit [other cpus will start sending flush ipis],
83 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
84 * 2) switch %%esp, ie current
86 * The interrupt must handle 2 special cases:
87 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
88 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
89 * runs in kernel space, the cpu could load tlb entries for user space
92 * The good news is that cpu mmu_state is local to each cpu, no
93 * write/read ordering problems.
99 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
100 * 2) Leave the mm if we are in the lazy tlb mode.
103 asmlinkage void smp_invalidate_interrupt (void)
109 if (!cpu_isset(cpu, flush_cpumask))
112 * This was a BUG() but until someone can quote me the
113 * line from the intel manual that guarantees an IPI to
114 * multiple CPUs is retried _only_ on the erroring CPUs
115 * its staying as a return
120 if (flush_mm == read_pda(active_mm)) {
121 if (read_pda(mmu_state) == TLBSTATE_OK) {
122 if (flush_va == FLUSH_ALL)
125 __flush_tlb_one(flush_va);
130 cpu_clear(cpu, flush_cpumask);
133 put_cpu_no_resched();
136 static void flush_tlb_others(cpumask_t cpumask, struct mm_struct *mm,
141 * A couple of (to be removed) sanity checks:
143 * - we do not send IPIs to not-yet booted CPUs.
144 * - current CPU must not be in mask
145 * - mask must exist :)
147 BUG_ON(cpus_empty(cpumask));
148 cpus_and(tmp, cpumask, cpu_online_map);
149 BUG_ON(!cpus_equal(tmp, cpumask));
150 BUG_ON(cpu_isset(smp_processor_id(), cpumask));
155 * I'm not happy about this global shared spinlock in the
156 * MM hot path, but we'll see how contended it is.
157 * Temporarily this turns IRQs off, so that lockups are
158 * detected by the NMI watchdog.
160 spin_lock(&tlbstate_lock);
164 cpus_or(flush_cpumask, cpumask, flush_cpumask);
167 * We have to send the IPI only to
170 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR);
172 while (!cpus_empty(flush_cpumask))
173 mb(); /* nothing. lockup detection does not belong here */;
177 spin_unlock(&tlbstate_lock);
180 void flush_tlb_current_task(void)
182 struct mm_struct *mm = current->mm;
186 cpu_mask = mm->cpu_vm_mask;
187 cpu_clear(smp_processor_id(), cpu_mask);
190 if (!cpus_empty(cpu_mask))
191 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
195 void flush_tlb_mm (struct mm_struct * mm)
200 cpu_mask = mm->cpu_vm_mask;
201 cpu_clear(smp_processor_id(), cpu_mask);
203 if (current->active_mm == mm) {
207 leave_mm(smp_processor_id());
209 if (!cpus_empty(cpu_mask))
210 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
215 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
217 struct mm_struct *mm = vma->vm_mm;
221 cpu_mask = mm->cpu_vm_mask;
222 cpu_clear(smp_processor_id(), cpu_mask);
224 if (current->active_mm == mm) {
228 leave_mm(smp_processor_id());
231 if (!cpus_empty(cpu_mask))
232 flush_tlb_others(cpu_mask, mm, va);
237 static void do_flush_tlb_all(void* info)
239 unsigned long cpu = smp_processor_id();
242 if (read_pda(mmu_state) == TLBSTATE_LAZY)
246 void flush_tlb_all(void)
248 on_each_cpu(do_flush_tlb_all, NULL, 1, 1);
251 void smp_kdb_stop(void)
253 send_IPI_allbutself(KDB_VECTOR);
257 * this function sends a 'reschedule' IPI to another CPU.
258 * it goes straight through and wastes no time serializing
259 * anything. Worst case is that we lose a reschedule ...
262 void smp_send_reschedule(int cpu)
264 send_IPI_mask(cpumask_of_cpu(cpu), RESCHEDULE_VECTOR);
268 * Structure and data for smp_call_function(). This is designed to minimise
269 * static memory requirements. It also looks cleaner.
271 static spinlock_t call_lock = SPIN_LOCK_UNLOCKED;
273 struct call_data_struct {
274 void (*func) (void *info);
281 static struct call_data_struct * call_data;
284 * this function sends a 'generic call function' IPI to all other CPUs
287 static void __smp_call_function (void (*func) (void *info), void *info,
288 int nonatomic, int wait)
290 static struct call_data_struct dumpdata;
291 struct call_data_struct normaldata;
292 struct call_data_struct *data;
293 int cpus = num_online_cpus()-1;
299 /* if another cpu beat us, they win! */
301 spin_unlock(&call_lock);
310 atomic_set(&data->started, 0);
311 data->wait = wait > 0 ? wait : 0;
313 atomic_set(&data->finished, 0);
317 /* Send a message to all other CPUs and wait for them to respond */
318 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
320 /* Wait for response */
322 while (atomic_read(&data->started) != cpus)
326 while (atomic_read(&data->finished) != cpus)
334 * smp_call_function - run a function on all other CPUs.
335 * @func: The function to run. This must be fast and non-blocking.
336 * @info: An arbitrary pointer to pass to the function.
337 * @nonatomic: currently unused.
338 * @wait: If 1, wait (atomically) until function has complete on other CPUs.
339 * If 0, wait for the IPI to be received by other CPUs, but do not wait
340 * for the completion of the IPI on each CPU. If -1, do not wait for
341 * other CPUs to receive IPI.
343 * Returns 0 on success, else a negative status code. Does not return until
344 * remote CPUs are nearly ready to execute func or are or have executed.
346 * You must not call this function with disabled interrupts or from a
347 * hardware interrupt handler or from a bottom half handler.
348 * Actually there are a few legal cases, like panic.
350 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
353 spin_lock(&call_lock);
354 __smp_call_function(func,info,nonatomic,wait);
355 spin_unlock(&call_lock);
359 void smp_stop_cpu(void)
364 cpu_clear(smp_processor_id(), cpu_online_map);
366 disable_local_APIC();
370 static void smp_really_stop_cpu(void *dummy)
377 void smp_send_stop(void)
380 /* Don't deadlock on the call lock in panic */
381 if (!spin_trylock(&call_lock)) {
382 /* ignore locking because we have paniced anyways */
385 __smp_call_function(smp_really_stop_cpu, NULL, 0, 0);
387 spin_unlock(&call_lock);
392 * Reschedule call back. Nothing to do,
393 * all the work is done automatically when
394 * we return from the interrupt.
396 asmlinkage void smp_reschedule_interrupt(void)
401 asmlinkage void smp_call_function_interrupt(void)
403 void (*func) (void *info) = call_data->func;
404 void *info = call_data->info;
405 int wait = call_data->wait;
409 * Notify initiating CPU that I've grabbed the data and am
410 * about to execute the function
413 atomic_inc(&call_data->started);
415 * At this point the info structure may be out of scope unless wait==1
422 atomic_inc(&call_data->finished);