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/irq.h>
15 #include <linux/delay.h>
16 #include <linux/spinlock.h>
17 #include <linux/smp_lock.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/mc146818rtc.h>
20 #include <linux/cache.h>
21 #include <linux/interrupt.h>
22 #include <linux/dump.h>
25 #include <asm/tlbflush.h>
27 #include <mach_apic.h>
30 * Some notes on x86 processor bugs affecting SMP operation:
32 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
33 * The Linux implications for SMP are handled as follows:
35 * Pentium III / [Xeon]
36 * None of the E1AP-E3AP errata are visible to the user.
43 * None of the A1AP-A3AP errata are visible to the user.
50 * None of 1AP-9AP errata are visible to the normal user,
51 * except occasional delivery of 'spurious interrupt' as trap #15.
52 * This is very rare and a non-problem.
54 * 1AP. Linux maps APIC as non-cacheable
55 * 2AP. worked around in hardware
56 * 3AP. fixed in C0 and above steppings microcode update.
57 * Linux does not use excessive STARTUP_IPIs.
58 * 4AP. worked around in hardware
59 * 5AP. symmetric IO mode (normal Linux operation) not affected.
60 * 'noapic' mode has vector 0xf filled out properly.
61 * 6AP. 'noapic' mode might be affected - fixed in later steppings
62 * 7AP. We do not assume writes to the LVT deassering IRQs
63 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
64 * 9AP. We do not use mixed mode
67 * There is a marginal case where REP MOVS on 100MHz SMP
68 * machines with B stepping processors can fail. XXX should provide
69 * an L1cache=Writethrough or L1cache=off option.
71 * B stepping CPUs may hang. There are hardware work arounds
72 * for this. We warn about it in case your board doesn't have the work
73 * arounds. Basically thats so I can tell anyone with a B stepping
74 * CPU and SMP problems "tough".
76 * Specific items [From Pentium Processor Specification Update]
78 * 1AP. Linux doesn't use remote read
79 * 2AP. Linux doesn't trust APIC errors
80 * 3AP. We work around this
81 * 4AP. Linux never generated 3 interrupts of the same priority
82 * to cause a lost local interrupt.
83 * 5AP. Remote read is never used
84 * 6AP. not affected - worked around in hardware
85 * 7AP. not affected - worked around in hardware
86 * 8AP. worked around in hardware - we get explicit CS errors if not
87 * 9AP. only 'noapic' mode affected. Might generate spurious
88 * interrupts, we log only the first one and count the
90 * 10AP. not affected - worked around in hardware
91 * 11AP. Linux reads the APIC between writes to avoid this, as per
92 * the documentation. Make sure you preserve this as it affects
93 * the C stepping chips too.
94 * 12AP. not affected - worked around in hardware
95 * 13AP. not affected - worked around in hardware
96 * 14AP. we always deassert INIT during bootup
97 * 15AP. not affected - worked around in hardware
98 * 16AP. not affected - worked around in hardware
99 * 17AP. not affected - worked around in hardware
100 * 18AP. not affected - worked around in hardware
101 * 19AP. not affected - worked around in BIOS
103 * If this sounds worrying believe me these bugs are either ___RARE___,
104 * or are signal timing bugs worked around in hardware and there's
105 * about nothing of note with C stepping upwards.
108 DEFINE_PER_CPU(struct tlb_state, cpu_tlbstate) ____cacheline_aligned = { &init_mm, 0, };
111 * the following functions deal with sending IPIs between CPUs.
113 * We use 'broadcast', CPU->CPU IPIs and self-IPIs too.
116 static inline int __prepare_ICR (unsigned int shortcut, int vector)
118 return APIC_DM_FIXED | shortcut | vector | APIC_DEST_LOGICAL;
121 static inline int __prepare_ICR2 (unsigned int mask)
123 return SET_APIC_DEST_FIELD(mask);
126 void __send_IPI_shortcut(unsigned int shortcut, int vector)
129 * Subtle. In the case of the 'never do double writes' workaround
130 * we have to lock out interrupts to be safe. As we don't care
131 * of the value read we use an atomic rmw access to avoid costly
132 * cli/sti. Otherwise we use an even cheaper single atomic write
140 apic_wait_icr_idle();
143 * No need to touch the target chip field
145 cfg = __prepare_ICR(shortcut, vector);
147 if (vector == DUMP_VECTOR) {
149 * Setup DUMP IPI to be delivered as an NMI
151 cfg = (cfg&~APIC_VECTOR_MASK)|APIC_DM_NMI;
155 * Send the IPI. The write to APIC_ICR fires this off.
157 apic_write_around(APIC_ICR, cfg);
160 void fastcall send_IPI_self(int vector)
162 __send_IPI_shortcut(APIC_DEST_SELF, vector);
166 * This is only used on smaller machines.
168 void send_IPI_mask_bitmask(cpumask_t cpumask, int vector)
170 unsigned long mask = cpus_addr(cpumask)[0];
174 local_irq_save(flags);
179 apic_wait_icr_idle();
182 * prepare target chip field
184 cfg = __prepare_ICR2(mask);
185 apic_write_around(APIC_ICR2, cfg);
190 cfg = __prepare_ICR(0, vector);
193 * Send the IPI. The write to APIC_ICR fires this off.
195 apic_write_around(APIC_ICR, cfg);
197 local_irq_restore(flags);
200 inline void send_IPI_mask_sequence(cpumask_t mask, int vector)
202 unsigned long cfg, flags;
203 unsigned int query_cpu;
206 * Hack. The clustered APIC addressing mode doesn't allow us to send
207 * to an arbitrary mask, so I do a unicasts to each CPU instead. This
208 * should be modified to do 1 message per cluster ID - mbligh
211 local_irq_save(flags);
213 for (query_cpu = 0; query_cpu < NR_CPUS; ++query_cpu) {
214 if (cpu_isset(query_cpu, mask)) {
219 apic_wait_icr_idle();
222 * prepare target chip field
224 cfg = __prepare_ICR2(cpu_to_logical_apicid(query_cpu));
225 apic_write_around(APIC_ICR2, cfg);
230 cfg = __prepare_ICR(0, vector);
232 if (vector == DUMP_VECTOR) {
234 * Setup DUMP IPI to be delivered as an NMI
236 cfg = (cfg&~APIC_VECTOR_MASK)|APIC_DM_NMI;
239 * Send the IPI. The write to APIC_ICR fires this off.
241 apic_write_around(APIC_ICR, cfg);
244 local_irq_restore(flags);
247 #include <mach_ipi.h> /* must come after the send_IPI functions above for inlining */
250 * Smarter SMP flushing macros.
251 * c/o Linus Torvalds.
253 * These mean you can really definitely utterly forget about
254 * writing to user space from interrupts. (Its not allowed anyway).
256 * Optimizations Manfred Spraul <manfred@colorfullife.com>
259 static cpumask_t flush_cpumask;
260 static struct mm_struct * flush_mm;
261 static unsigned long flush_va;
262 static spinlock_t tlbstate_lock = SPIN_LOCK_UNLOCKED;
263 #define FLUSH_ALL 0xffffffff
266 * We cannot call mmdrop() because we are in interrupt context,
267 * instead update mm->cpu_vm_mask.
269 * We need to reload %cr3 since the page tables may be going
270 * away from under us..
272 static inline void leave_mm (unsigned long cpu)
274 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK)
276 cpu_clear(cpu, per_cpu(cpu_tlbstate, cpu).active_mm->cpu_vm_mask);
277 load_cr3(swapper_pg_dir);
282 * The flush IPI assumes that a thread switch happens in this order:
283 * [cpu0: the cpu that switches]
284 * 1) switch_mm() either 1a) or 1b)
285 * 1a) thread switch to a different mm
286 * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
287 * Stop ipi delivery for the old mm. This is not synchronized with
288 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
289 * for the wrong mm, and in the worst case we perform a superflous
291 * 1a2) set cpu_tlbstate to TLBSTATE_OK
292 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
293 * was in lazy tlb mode.
294 * 1a3) update cpu_tlbstate[].active_mm
295 * Now cpu0 accepts tlb flushes for the new mm.
296 * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
297 * Now the other cpus will send tlb flush ipis.
299 * 1b) thread switch without mm change
300 * cpu_tlbstate[].active_mm is correct, cpu0 already handles
302 * 1b1) set cpu_tlbstate to TLBSTATE_OK
303 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
304 * Atomically set the bit [other cpus will start sending flush ipis],
306 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
307 * 2) switch %%esp, ie current
309 * The interrupt must handle 2 special cases:
310 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
311 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
312 * runs in kernel space, the cpu could load tlb entries for user space
315 * The good news is that cpu_tlbstate is local to each cpu, no
316 * write/read ordering problems.
322 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
323 * 2) Leave the mm if we are in the lazy tlb mode.
326 fastcall void smp_invalidate_interrupt(struct pt_regs *regs)
331 if (current->active_mm)
332 load_user_cs_desc(cpu, current->active_mm);
334 if (!cpu_isset(cpu, flush_cpumask))
337 * This was a BUG() but until someone can quote me the
338 * line from the intel manual that guarantees an IPI to
339 * multiple CPUs is retried _only_ on the erroring CPUs
340 * its staying as a return
345 if (flush_mm == per_cpu(cpu_tlbstate, cpu).active_mm) {
346 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK) {
347 if (flush_va == FLUSH_ALL)
350 __flush_tlb_one(flush_va);
355 smp_mb__before_clear_bit();
356 cpu_clear(cpu, flush_cpumask);
357 smp_mb__after_clear_bit();
359 put_cpu_no_resched();
362 static void flush_tlb_others(cpumask_t cpumask, struct mm_struct *mm,
367 * A couple of (to be removed) sanity checks:
369 * - we do not send IPIs to not-yet booted CPUs.
370 * - current CPU must not be in mask
371 * - mask must exist :)
373 BUG_ON(cpus_empty(cpumask));
375 cpus_and(tmp, cpumask, cpu_online_map);
376 BUG_ON(!cpus_equal(cpumask, tmp));
377 BUG_ON(cpu_isset(smp_processor_id(), cpumask));
381 * i'm not happy about this global shared spinlock in the
382 * MM hot path, but we'll see how contended it is.
383 * Temporarily this turns IRQs off, so that lockups are
384 * detected by the NMI watchdog.
386 spin_lock(&tlbstate_lock);
390 #if NR_CPUS <= BITS_PER_LONG
391 atomic_set_mask(cpumask, &flush_cpumask);
395 unsigned long *flush_mask = (unsigned long *)&flush_cpumask;
396 unsigned long *cpu_mask = (unsigned long *)&cpumask;
397 for (k = 0; k < BITS_TO_LONGS(NR_CPUS); ++k)
398 atomic_set_mask(cpu_mask[k], &flush_mask[k]);
402 * We have to send the IPI only to
405 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR);
407 while (!cpus_empty(flush_cpumask))
408 /* nothing. lockup detection does not belong here */
413 spin_unlock(&tlbstate_lock);
416 void flush_tlb_current_task(void)
418 struct mm_struct *mm = current->mm;
422 cpu_mask = mm->cpu_vm_mask;
423 cpu_clear(smp_processor_id(), cpu_mask);
426 if (!cpus_empty(cpu_mask))
427 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
431 void flush_tlb_mm (struct mm_struct * mm)
436 cpu_mask = mm->cpu_vm_mask;
437 cpu_clear(smp_processor_id(), cpu_mask);
439 if (current->active_mm == mm) {
443 leave_mm(smp_processor_id());
445 if (!cpus_empty(cpu_mask))
446 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
451 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
453 struct mm_struct *mm = vma->vm_mm;
457 cpu_mask = mm->cpu_vm_mask;
458 cpu_clear(smp_processor_id(), cpu_mask);
460 if (current->active_mm == mm) {
464 leave_mm(smp_processor_id());
467 if (!cpus_empty(cpu_mask))
468 flush_tlb_others(cpu_mask, mm, va);
473 static void do_flush_tlb_all(void* info)
475 unsigned long cpu = smp_processor_id();
478 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_LAZY)
482 void flush_tlb_all(void)
484 on_each_cpu(do_flush_tlb_all, NULL, 1, 1);
487 void dump_send_ipi(void)
489 send_IPI_allbutself(DUMP_VECTOR);
493 * this function sends a 'reschedule' IPI to another CPU.
494 * it goes straight through and wastes no time serializing
495 * anything. Worst case is that we lose a reschedule ...
497 void smp_send_reschedule(int cpu)
499 send_IPI_mask(cpumask_of_cpu(cpu), RESCHEDULE_VECTOR);
503 * Structure and data for smp_call_function(). This is designed to minimise
504 * static memory requirements. It also looks cleaner.
506 static spinlock_t call_lock = SPIN_LOCK_UNLOCKED;
508 struct call_data_struct {
509 void (*func) (void *info);
516 static struct call_data_struct * call_data;
519 * this function sends a 'generic call function' IPI to all other CPUs
523 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
526 * [SUMMARY] Run a function on all other CPUs.
527 * <func> The function to run. This must be fast and non-blocking.
528 * <info> An arbitrary pointer to pass to the function.
529 * <nonatomic> currently unused.
530 * <wait> If 1, wait (atomically) until function has completed on other CPUs.
531 * If 0, wait for the IPI to be received by other CPUs, but do not wait
532 * for the completion of the function on each CPU.
533 * If -1, do not wait for other CPUs to receive IPI.
534 * [RETURNS] 0 on success, else a negative status code. Does not return until
535 * remote CPUs are nearly ready to execute <<func>> or are or have executed.
537 * You must not call this function with disabled interrupts or from a
538 * hardware interrupt handler or from a bottom half handler.
541 static struct call_data_struct dumpdata;
542 struct call_data_struct normaldata;
543 struct call_data_struct *data;
544 int cpus = num_online_cpus()-1;
549 /* Can deadlock when called with interrupts disabled */
550 /* Only if we are waiting for other CPU to ack */
551 WARN_ON(irqs_disabled() && wait >= 0);
553 spin_lock(&call_lock);
555 /* if another cpu beat us, they win! */
557 spin_unlock(&call_lock);
566 atomic_set(&data->started, 0);
567 data->wait = wait > 0 ? wait : 0;
569 atomic_set(&data->finished, 0);
574 /* Send a message to all other CPUs and wait for them to respond */
575 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
577 /* Wait for response */
579 while (atomic_read(&data->started) != cpus)
583 while (atomic_read(&data->finished) != cpus)
585 spin_unlock(&call_lock);
590 void stop_this_cpu (void * dummy)
595 cpu_clear(smp_processor_id(), cpu_online_map);
597 disable_local_APIC();
598 if (cpu_data[smp_processor_id()].hlt_works_ok)
599 for(;;) __asm__("hlt");
604 * this function calls the 'stop' function on all other CPUs in the system.
607 void smp_send_stop(void)
609 smp_call_function(stop_this_cpu, NULL, 1, 0);
612 disable_local_APIC();
616 EXPORT_SYMBOL(smp_send_stop);
619 * Reschedule call back. Nothing to do,
620 * all the work is done automatically when
621 * we return from the interrupt.
623 fastcall void smp_reschedule_interrupt(struct pt_regs *regs)
628 fastcall void smp_call_function_interrupt(struct pt_regs *regs)
630 void (*func) (void *info) = call_data->func;
631 void *info = call_data->info;
632 int wait = call_data->wait;
636 * Notify initiating CPU that I've grabbed the data and am
637 * about to execute the function
640 atomic_inc(&call_data->started);
642 * At this point the info structure may be out of scope unless wait==1
650 atomic_inc(&call_data->finished);