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>
7 * This code is released under the GNU General Public License version 2 or
11 #include <linux/init.h>
14 #include <linux/delay.h>
15 #include <linux/spinlock.h>
16 #include <linux/smp_lock.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/mc146818rtc.h>
19 #include <linux/cache.h>
20 #include <linux/interrupt.h>
21 #include <linux/cpu.h>
22 #include <linux/module.h>
25 #include <asm/tlbflush.h>
26 #include <mach_apic.h>
29 * Some notes on x86 processor bugs affecting SMP operation:
31 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
32 * The Linux implications for SMP are handled as follows:
34 * Pentium III / [Xeon]
35 * None of the E1AP-E3AP errata are visible to the user.
42 * None of the A1AP-A3AP errata are visible to the user.
49 * None of 1AP-9AP errata are visible to the normal user,
50 * except occasional delivery of 'spurious interrupt' as trap #15.
51 * This is very rare and a non-problem.
53 * 1AP. Linux maps APIC as non-cacheable
54 * 2AP. worked around in hardware
55 * 3AP. fixed in C0 and above steppings microcode update.
56 * Linux does not use excessive STARTUP_IPIs.
57 * 4AP. worked around in hardware
58 * 5AP. symmetric IO mode (normal Linux operation) not affected.
59 * 'noapic' mode has vector 0xf filled out properly.
60 * 6AP. 'noapic' mode might be affected - fixed in later steppings
61 * 7AP. We do not assume writes to the LVT deassering IRQs
62 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
63 * 9AP. We do not use mixed mode
66 * There is a marginal case where REP MOVS on 100MHz SMP
67 * machines with B stepping processors can fail. XXX should provide
68 * an L1cache=Writethrough or L1cache=off option.
70 * B stepping CPUs may hang. There are hardware work arounds
71 * for this. We warn about it in case your board doesn't have the work
72 * arounds. Basically thats so I can tell anyone with a B stepping
73 * CPU and SMP problems "tough".
75 * Specific items [From Pentium Processor Specification Update]
77 * 1AP. Linux doesn't use remote read
78 * 2AP. Linux doesn't trust APIC errors
79 * 3AP. We work around this
80 * 4AP. Linux never generated 3 interrupts of the same priority
81 * to cause a lost local interrupt.
82 * 5AP. Remote read is never used
83 * 6AP. not affected - worked around in hardware
84 * 7AP. not affected - worked around in hardware
85 * 8AP. worked around in hardware - we get explicit CS errors if not
86 * 9AP. only 'noapic' mode affected. Might generate spurious
87 * interrupts, we log only the first one and count the
89 * 10AP. not affected - worked around in hardware
90 * 11AP. Linux reads the APIC between writes to avoid this, as per
91 * the documentation. Make sure you preserve this as it affects
92 * the C stepping chips too.
93 * 12AP. not affected - worked around in hardware
94 * 13AP. not affected - worked around in hardware
95 * 14AP. we always deassert INIT during bootup
96 * 15AP. not affected - worked around in hardware
97 * 16AP. not affected - worked around in hardware
98 * 17AP. not affected - worked around in hardware
99 * 18AP. not affected - worked around in hardware
100 * 19AP. not affected - worked around in BIOS
102 * If this sounds worrying believe me these bugs are either ___RARE___,
103 * or are signal timing bugs worked around in hardware and there's
104 * about nothing of note with C stepping upwards.
107 DEFINE_PER_CPU(struct tlb_state, cpu_tlbstate) ____cacheline_aligned = { &init_mm, 0, };
110 * the following functions deal with sending IPIs between CPUs.
112 * We use 'broadcast', CPU->CPU IPIs and self-IPIs too.
115 static inline int __prepare_ICR (unsigned int shortcut, int vector)
117 unsigned int icr = shortcut | APIC_DEST_LOGICAL;
121 icr |= APIC_DM_FIXED | vector;
130 static inline int __prepare_ICR2 (unsigned int mask)
132 return SET_APIC_DEST_FIELD(mask);
135 void __send_IPI_shortcut(unsigned int shortcut, int vector)
138 * Subtle. In the case of the 'never do double writes' workaround
139 * we have to lock out interrupts to be safe. As we don't care
140 * of the value read we use an atomic rmw access to avoid costly
141 * cli/sti. Otherwise we use an even cheaper single atomic write
149 apic_wait_icr_idle();
152 * No need to touch the target chip field
154 cfg = __prepare_ICR(shortcut, vector);
157 * Send the IPI. The write to APIC_ICR fires this off.
159 apic_write_around(APIC_ICR, cfg);
162 void fastcall send_IPI_self(int vector)
164 __send_IPI_shortcut(APIC_DEST_SELF, vector);
168 * This is only used on smaller machines.
170 void send_IPI_mask_bitmask(cpumask_t cpumask, int vector)
172 unsigned long mask = cpus_addr(cpumask)[0];
176 local_irq_save(flags);
177 WARN_ON(mask & ~cpus_addr(cpu_online_map)[0]);
181 apic_wait_icr_idle();
184 * prepare target chip field
186 cfg = __prepare_ICR2(mask);
187 apic_write_around(APIC_ICR2, cfg);
192 cfg = __prepare_ICR(0, vector);
195 * Send the IPI. The write to APIC_ICR fires this off.
197 apic_write_around(APIC_ICR, cfg);
199 local_irq_restore(flags);
202 void send_IPI_mask_sequence(cpumask_t mask, int vector)
204 unsigned long cfg, flags;
205 unsigned int query_cpu;
208 * Hack. The clustered APIC addressing mode doesn't allow us to send
209 * to an arbitrary mask, so I do a unicasts to each CPU instead. This
210 * should be modified to do 1 message per cluster ID - mbligh
213 local_irq_save(flags);
215 for (query_cpu = 0; query_cpu < NR_CPUS; ++query_cpu) {
216 if (cpu_isset(query_cpu, mask)) {
221 apic_wait_icr_idle();
224 * prepare target chip field
226 cfg = __prepare_ICR2(cpu_to_logical_apicid(query_cpu));
227 apic_write_around(APIC_ICR2, cfg);
232 cfg = __prepare_ICR(0, vector);
235 * Send the IPI. The write to APIC_ICR fires this off.
237 apic_write_around(APIC_ICR, cfg);
240 local_irq_restore(flags);
243 #include <mach_ipi.h> /* must come after the send_IPI functions above for inlining */
246 * Smarter SMP flushing macros.
247 * c/o Linus Torvalds.
249 * These mean you can really definitely utterly forget about
250 * writing to user space from interrupts. (Its not allowed anyway).
252 * Optimizations Manfred Spraul <manfred@colorfullife.com>
255 static cpumask_t flush_cpumask;
256 static struct mm_struct * flush_mm;
257 static unsigned long flush_va;
258 static DEFINE_SPINLOCK(tlbstate_lock);
259 #define FLUSH_ALL 0xffffffff
262 * We cannot call mmdrop() because we are in interrupt context,
263 * instead update mm->cpu_vm_mask.
265 * We need to reload %cr3 since the page tables may be going
266 * away from under us..
268 static inline void leave_mm (unsigned long cpu)
270 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK)
272 cpu_clear(cpu, per_cpu(cpu_tlbstate, cpu).active_mm->cpu_vm_mask);
273 load_cr3(swapper_pg_dir);
278 * The flush IPI assumes that a thread switch happens in this order:
279 * [cpu0: the cpu that switches]
280 * 1) switch_mm() either 1a) or 1b)
281 * 1a) thread switch to a different mm
282 * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
283 * Stop ipi delivery for the old mm. This is not synchronized with
284 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
285 * for the wrong mm, and in the worst case we perform a superflous
287 * 1a2) set cpu_tlbstate to TLBSTATE_OK
288 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
289 * was in lazy tlb mode.
290 * 1a3) update cpu_tlbstate[].active_mm
291 * Now cpu0 accepts tlb flushes for the new mm.
292 * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
293 * Now the other cpus will send tlb flush ipis.
295 * 1b) thread switch without mm change
296 * cpu_tlbstate[].active_mm is correct, cpu0 already handles
298 * 1b1) set cpu_tlbstate to TLBSTATE_OK
299 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
300 * Atomically set the bit [other cpus will start sending flush ipis],
302 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
303 * 2) switch %%esp, ie current
305 * The interrupt must handle 2 special cases:
306 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
307 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
308 * runs in kernel space, the cpu could load tlb entries for user space
311 * The good news is that cpu_tlbstate is local to each cpu, no
312 * write/read ordering problems.
318 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
319 * 2) Leave the mm if we are in the lazy tlb mode.
322 fastcall void smp_invalidate_interrupt(struct pt_regs *regs)
328 if (!cpu_isset(cpu, flush_cpumask))
331 * This was a BUG() but until someone can quote me the
332 * line from the intel manual that guarantees an IPI to
333 * multiple CPUs is retried _only_ on the erroring CPUs
334 * its staying as a return
339 if (flush_mm == per_cpu(cpu_tlbstate, cpu).active_mm) {
340 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK) {
341 if (flush_va == FLUSH_ALL)
344 __flush_tlb_one(flush_va);
349 smp_mb__before_clear_bit();
350 cpu_clear(cpu, flush_cpumask);
351 smp_mb__after_clear_bit();
353 put_cpu_no_resched();
356 static void flush_tlb_others(cpumask_t cpumask, struct mm_struct *mm,
360 * A couple of (to be removed) sanity checks:
362 * - current CPU must not be in mask
363 * - mask must exist :)
365 BUG_ON(cpus_empty(cpumask));
366 BUG_ON(cpu_isset(smp_processor_id(), cpumask));
369 /* If a CPU which we ran on has gone down, OK. */
370 cpus_and(cpumask, cpumask, cpu_online_map);
371 if (cpus_empty(cpumask))
375 * i'm not happy about this global shared spinlock in the
376 * MM hot path, but we'll see how contended it is.
377 * Temporarily this turns IRQs off, so that lockups are
378 * detected by the NMI watchdog.
380 spin_lock(&tlbstate_lock);
384 #if NR_CPUS <= BITS_PER_LONG
385 atomic_set_mask(cpumask, &flush_cpumask);
389 unsigned long *flush_mask = (unsigned long *)&flush_cpumask;
390 unsigned long *cpu_mask = (unsigned long *)&cpumask;
391 for (k = 0; k < BITS_TO_LONGS(NR_CPUS); ++k)
392 atomic_set_mask(cpu_mask[k], &flush_mask[k]);
396 * We have to send the IPI only to
399 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR);
401 while (!cpus_empty(flush_cpumask))
402 /* nothing. lockup detection does not belong here */
407 spin_unlock(&tlbstate_lock);
410 void flush_tlb_current_task(void)
412 struct mm_struct *mm = current->mm;
416 cpu_mask = mm->cpu_vm_mask;
417 cpu_clear(smp_processor_id(), cpu_mask);
420 if (!cpus_empty(cpu_mask))
421 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
425 void flush_tlb_mm (struct mm_struct * mm)
430 cpu_mask = mm->cpu_vm_mask;
431 cpu_clear(smp_processor_id(), cpu_mask);
433 if (current->active_mm == mm) {
437 leave_mm(smp_processor_id());
439 if (!cpus_empty(cpu_mask))
440 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
445 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
447 struct mm_struct *mm = vma->vm_mm;
451 cpu_mask = mm->cpu_vm_mask;
452 cpu_clear(smp_processor_id(), cpu_mask);
454 if (current->active_mm == mm) {
458 leave_mm(smp_processor_id());
461 if (!cpus_empty(cpu_mask))
462 flush_tlb_others(cpu_mask, mm, va);
466 EXPORT_SYMBOL(flush_tlb_page);
468 static void do_flush_tlb_all(void* info)
470 unsigned long cpu = smp_processor_id();
473 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_LAZY)
477 void flush_tlb_all(void)
479 on_each_cpu(do_flush_tlb_all, NULL, 1, 1);
483 * this function sends a 'reschedule' IPI to another CPU.
484 * it goes straight through and wastes no time serializing
485 * anything. Worst case is that we lose a reschedule ...
487 void smp_send_reschedule(int cpu)
489 WARN_ON(cpu_is_offline(cpu));
490 send_IPI_mask(cpumask_of_cpu(cpu), RESCHEDULE_VECTOR);
494 * Structure and data for smp_call_function(). This is designed to minimise
495 * static memory requirements. It also looks cleaner.
497 static DEFINE_SPINLOCK(call_lock);
499 struct call_data_struct {
500 void (*func) (void *info);
507 void lock_ipi_call_lock(void)
509 spin_lock_irq(&call_lock);
512 void unlock_ipi_call_lock(void)
514 spin_unlock_irq(&call_lock);
517 static struct call_data_struct *call_data;
520 * smp_call_function(): Run a function on all other CPUs.
521 * @func: The function to run. This must be fast and non-blocking.
522 * @info: An arbitrary pointer to pass to the function.
523 * @nonatomic: currently unused.
524 * @wait: If true, wait (atomically) until function has completed on other CPUs.
526 * Returns 0 on success, else a negative status code. Does not return until
527 * remote CPUs are nearly ready to execute <<func>> or are or have executed.
529 * You must not call this function with disabled interrupts or from a
530 * hardware interrupt handler or from a bottom half handler.
532 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
535 struct call_data_struct data;
538 /* Holding any lock stops cpus from going down. */
539 spin_lock(&call_lock);
540 cpus = num_online_cpus() - 1;
542 spin_unlock(&call_lock);
546 /* Can deadlock when called with interrupts disabled */
547 WARN_ON(irqs_disabled());
551 atomic_set(&data.started, 0);
554 atomic_set(&data.finished, 0);
559 /* Send a message to all other CPUs and wait for them to respond */
560 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
562 /* Wait for response */
563 while (atomic_read(&data.started) != cpus)
567 while (atomic_read(&data.finished) != cpus)
569 spin_unlock(&call_lock);
573 EXPORT_SYMBOL(smp_call_function);
575 static void stop_this_cpu (void * dummy)
580 cpu_clear(smp_processor_id(), cpu_online_map);
582 disable_local_APIC();
583 if (cpu_data[smp_processor_id()].hlt_works_ok)
589 * this function calls the 'stop' function on all other CPUs in the system.
592 void smp_send_stop(void)
594 smp_call_function(stop_this_cpu, NULL, 1, 0);
597 disable_local_APIC();
602 * Reschedule call back. Nothing to do,
603 * all the work is done automatically when
604 * we return from the interrupt.
606 fastcall void smp_reschedule_interrupt(struct pt_regs *regs)
611 fastcall void smp_call_function_interrupt(struct pt_regs *regs)
613 void (*func) (void *info) = call_data->func;
614 void *info = call_data->info;
615 int wait = call_data->wait;
619 * Notify initiating CPU that I've grabbed the data and am
620 * about to execute the function
623 atomic_inc(&call_data->started);
625 * At this point the info structure may be out of scope unless wait==1
633 atomic_inc(&call_data->finished);