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/cpu.h>
23 #include <linux/module.h>
26 #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);
148 * Send the IPI. The write to APIC_ICR fires this off.
150 apic_write_around(APIC_ICR, cfg);
153 void fastcall send_IPI_self(int vector)
155 __send_IPI_shortcut(APIC_DEST_SELF, vector);
159 * This is only used on smaller machines.
161 void send_IPI_mask_bitmask(cpumask_t cpumask, int vector)
163 unsigned long mask = cpus_addr(cpumask)[0];
167 local_irq_save(flags);
168 WARN_ON(mask & ~cpus_addr(cpu_online_map)[0]);
172 apic_wait_icr_idle();
175 * prepare target chip field
177 cfg = __prepare_ICR2(mask);
178 apic_write_around(APIC_ICR2, cfg);
183 cfg = __prepare_ICR(0, vector);
186 * Send the IPI. The write to APIC_ICR fires this off.
188 apic_write_around(APIC_ICR, cfg);
190 local_irq_restore(flags);
193 void send_IPI_mask_sequence(cpumask_t mask, int vector)
195 unsigned long cfg, flags;
196 unsigned int query_cpu;
199 * Hack. The clustered APIC addressing mode doesn't allow us to send
200 * to an arbitrary mask, so I do a unicasts to each CPU instead. This
201 * should be modified to do 1 message per cluster ID - mbligh
204 local_irq_save(flags);
206 for (query_cpu = 0; query_cpu < NR_CPUS; ++query_cpu) {
207 if (cpu_isset(query_cpu, mask)) {
212 apic_wait_icr_idle();
215 * prepare target chip field
217 cfg = __prepare_ICR2(cpu_to_logical_apicid(query_cpu));
218 apic_write_around(APIC_ICR2, cfg);
223 cfg = __prepare_ICR(0, vector);
226 * Send the IPI. The write to APIC_ICR fires this off.
228 apic_write_around(APIC_ICR, cfg);
231 local_irq_restore(flags);
234 #include <mach_ipi.h> /* must come after the send_IPI functions above for inlining */
237 * Smarter SMP flushing macros.
238 * c/o Linus Torvalds.
240 * These mean you can really definitely utterly forget about
241 * writing to user space from interrupts. (Its not allowed anyway).
243 * Optimizations Manfred Spraul <manfred@colorfullife.com>
246 static cpumask_t flush_cpumask;
247 static struct mm_struct * flush_mm;
248 static unsigned long flush_va;
249 static DEFINE_SPINLOCK(tlbstate_lock);
250 #define FLUSH_ALL 0xffffffff
253 * We cannot call mmdrop() because we are in interrupt context,
254 * instead update mm->cpu_vm_mask.
256 * We need to reload %cr3 since the page tables may be going
257 * away from under us..
259 static inline void leave_mm (unsigned long cpu)
261 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK)
263 cpu_clear(cpu, per_cpu(cpu_tlbstate, cpu).active_mm->cpu_vm_mask);
264 load_cr3(swapper_pg_dir);
269 * The flush IPI assumes that a thread switch happens in this order:
270 * [cpu0: the cpu that switches]
271 * 1) switch_mm() either 1a) or 1b)
272 * 1a) thread switch to a different mm
273 * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
274 * Stop ipi delivery for the old mm. This is not synchronized with
275 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
276 * for the wrong mm, and in the worst case we perform a superflous
278 * 1a2) set cpu_tlbstate to TLBSTATE_OK
279 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
280 * was in lazy tlb mode.
281 * 1a3) update cpu_tlbstate[].active_mm
282 * Now cpu0 accepts tlb flushes for the new mm.
283 * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
284 * Now the other cpus will send tlb flush ipis.
286 * 1b) thread switch without mm change
287 * cpu_tlbstate[].active_mm is correct, cpu0 already handles
289 * 1b1) set cpu_tlbstate to TLBSTATE_OK
290 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
291 * Atomically set the bit [other cpus will start sending flush ipis],
293 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
294 * 2) switch %%esp, ie current
296 * The interrupt must handle 2 special cases:
297 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
298 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
299 * runs in kernel space, the cpu could load tlb entries for user space
302 * The good news is that cpu_tlbstate is local to each cpu, no
303 * write/read ordering problems.
309 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
310 * 2) Leave the mm if we are in the lazy tlb mode.
313 fastcall void smp_invalidate_interrupt(struct pt_regs *regs)
319 if (!cpu_isset(cpu, flush_cpumask))
322 * This was a BUG() but until someone can quote me the
323 * line from the intel manual that guarantees an IPI to
324 * multiple CPUs is retried _only_ on the erroring CPUs
325 * its staying as a return
330 if (flush_mm == per_cpu(cpu_tlbstate, cpu).active_mm) {
331 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK) {
332 if (flush_va == FLUSH_ALL)
335 __flush_tlb_one(flush_va);
340 smp_mb__before_clear_bit();
341 cpu_clear(cpu, flush_cpumask);
342 smp_mb__after_clear_bit();
344 put_cpu_no_resched();
347 static void flush_tlb_others(cpumask_t cpumask, struct mm_struct *mm,
351 * A couple of (to be removed) sanity checks:
353 * - current CPU must not be in mask
354 * - mask must exist :)
356 BUG_ON(cpus_empty(cpumask));
357 BUG_ON(cpu_isset(smp_processor_id(), cpumask));
360 /* If a CPU which we ran on has gone down, OK. */
361 cpus_and(cpumask, cpumask, cpu_online_map);
362 if (cpus_empty(cpumask))
366 * i'm not happy about this global shared spinlock in the
367 * MM hot path, but we'll see how contended it is.
368 * Temporarily this turns IRQs off, so that lockups are
369 * detected by the NMI watchdog.
371 spin_lock(&tlbstate_lock);
375 #if NR_CPUS <= BITS_PER_LONG
376 atomic_set_mask(cpumask, &flush_cpumask);
380 unsigned long *flush_mask = (unsigned long *)&flush_cpumask;
381 unsigned long *cpu_mask = (unsigned long *)&cpumask;
382 for (k = 0; k < BITS_TO_LONGS(NR_CPUS); ++k)
383 atomic_set_mask(cpu_mask[k], &flush_mask[k]);
387 * We have to send the IPI only to
390 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR);
392 while (!cpus_empty(flush_cpumask))
393 /* nothing. lockup detection does not belong here */
398 spin_unlock(&tlbstate_lock);
401 void flush_tlb_current_task(void)
403 struct mm_struct *mm = current->mm;
407 cpu_mask = mm->cpu_vm_mask;
408 cpu_clear(smp_processor_id(), cpu_mask);
411 if (!cpus_empty(cpu_mask))
412 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
416 void flush_tlb_mm (struct mm_struct * mm)
421 cpu_mask = mm->cpu_vm_mask;
422 cpu_clear(smp_processor_id(), cpu_mask);
424 if (current->active_mm == mm) {
428 leave_mm(smp_processor_id());
430 if (!cpus_empty(cpu_mask))
431 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
436 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
438 struct mm_struct *mm = vma->vm_mm;
442 cpu_mask = mm->cpu_vm_mask;
443 cpu_clear(smp_processor_id(), cpu_mask);
445 if (current->active_mm == mm) {
449 leave_mm(smp_processor_id());
452 if (!cpus_empty(cpu_mask))
453 flush_tlb_others(cpu_mask, mm, va);
457 EXPORT_SYMBOL(flush_tlb_page);
459 static void do_flush_tlb_all(void* info)
461 unsigned long cpu = smp_processor_id();
464 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_LAZY)
468 void flush_tlb_all(void)
470 on_each_cpu(do_flush_tlb_all, NULL, 1, 1);
474 * this function sends a 'reschedule' IPI to another CPU.
475 * it goes straight through and wastes no time serializing
476 * anything. Worst case is that we lose a reschedule ...
478 void smp_send_reschedule(int cpu)
480 WARN_ON(cpu_is_offline(cpu));
481 send_IPI_mask(cpumask_of_cpu(cpu), RESCHEDULE_VECTOR);
485 * Structure and data for smp_call_function(). This is designed to minimise
486 * static memory requirements. It also looks cleaner.
488 static DEFINE_SPINLOCK(call_lock);
490 struct call_data_struct {
491 void (*func) (void *info);
498 void lock_ipi_call_lock(void)
500 spin_lock_irq(&call_lock);
503 void unlock_ipi_call_lock(void)
505 spin_unlock_irq(&call_lock);
508 static struct call_data_struct * call_data;
511 * this function sends a 'generic call function' IPI to all other CPUs
515 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
518 * [SUMMARY] Run a function on all other CPUs.
519 * <func> The function to run. This must be fast and non-blocking.
520 * <info> An arbitrary pointer to pass to the function.
521 * <nonatomic> currently unused.
522 * <wait> If true, wait (atomically) until function has completed on other CPUs.
523 * [RETURNS] 0 on success, else a negative status code. Does not return until
524 * remote CPUs are nearly ready to execute <<func>> or are or have executed.
526 * You must not call this function with disabled interrupts or from a
527 * hardware interrupt handler or from a bottom half handler.
530 struct call_data_struct data;
533 /* Holding any lock stops cpus from going down. */
534 spin_lock(&call_lock);
535 cpus = num_online_cpus() - 1;
537 spin_unlock(&call_lock);
541 /* Can deadlock when called with interrupts disabled */
542 WARN_ON(irqs_disabled());
546 atomic_set(&data.started, 0);
549 atomic_set(&data.finished, 0);
554 /* Send a message to all other CPUs and wait for them to respond */
555 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
557 /* Wait for response */
558 while (atomic_read(&data.started) != cpus)
562 while (atomic_read(&data.finished) != cpus)
564 spin_unlock(&call_lock);
568 EXPORT_SYMBOL(smp_call_function);
570 static void stop_this_cpu (void * dummy)
575 cpu_clear(smp_processor_id(), cpu_online_map);
577 disable_local_APIC();
578 if (cpu_data[smp_processor_id()].hlt_works_ok)
584 * this function calls the 'stop' function on all other CPUs in the system.
587 void smp_send_stop(void)
589 smp_call_function(stop_this_cpu, NULL, 1, 0);
592 disable_local_APIC();
597 * Reschedule call back. Nothing to do,
598 * all the work is done automatically when
599 * we return from the interrupt.
601 fastcall void smp_reschedule_interrupt(struct pt_regs *regs)
606 fastcall void smp_call_function_interrupt(struct pt_regs *regs)
608 void (*func) (void *info) = call_data->func;
609 void *info = call_data->info;
610 int wait = call_data->wait;
614 * Notify initiating CPU that I've grabbed the data and am
615 * about to execute the function
618 atomic_inc(&call_data->started);
620 * At this point the info structure may be out of scope unless wait==1
628 atomic_inc(&call_data->finished);