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>
23 #include <asm/pgalloc.h>
24 #include <asm/smpboot.h>
27 * Some notes on x86 processor bugs affecting SMP operation:
29 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
30 * The Linux implications for SMP are handled as follows:
32 * Pentium III / [Xeon]
33 * None of the E1AP-E3AP errata are visible to the user.
40 * None of the A1AP-A3AP errata are visible to the user.
47 * None of 1AP-9AP errata are visible to the normal user,
48 * except occasional delivery of 'spurious interrupt' as trap #15.
49 * This is very rare and a non-problem.
51 * 1AP. Linux maps APIC as non-cacheable
52 * 2AP. worked around in hardware
53 * 3AP. fixed in C0 and above steppings microcode update.
54 * Linux does not use excessive STARTUP_IPIs.
55 * 4AP. worked around in hardware
56 * 5AP. symmetric IO mode (normal Linux operation) not affected.
57 * 'noapic' mode has vector 0xf filled out properly.
58 * 6AP. 'noapic' mode might be affected - fixed in later steppings
59 * 7AP. We do not assume writes to the LVT deassering IRQs
60 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
61 * 9AP. We do not use mixed mode
64 * There is a marginal case where REP MOVS on 100MHz SMP
65 * machines with B stepping processors can fail. XXX should provide
66 * an L1cache=Writethrough or L1cache=off option.
68 * B stepping CPUs may hang. There are hardware work arounds
69 * for this. We warn about it in case your board doesn't have the work
70 * arounds. Basically thats so I can tell anyone with a B stepping
71 * CPU and SMP problems "tough".
73 * Specific items [From Pentium Processor Specification Update]
75 * 1AP. Linux doesn't use remote read
76 * 2AP. Linux doesn't trust APIC errors
77 * 3AP. We work around this
78 * 4AP. Linux never generated 3 interrupts of the same priority
79 * to cause a lost local interrupt.
80 * 5AP. Remote read is never used
81 * 6AP. not affected - worked around in hardware
82 * 7AP. not affected - worked around in hardware
83 * 8AP. worked around in hardware - we get explicit CS errors if not
84 * 9AP. only 'noapic' mode affected. Might generate spurious
85 * interrupts, we log only the first one and count the
87 * 10AP. not affected - worked around in hardware
88 * 11AP. Linux reads the APIC between writes to avoid this, as per
89 * the documentation. Make sure you preserve this as it affects
90 * the C stepping chips too.
91 * 12AP. not affected - worked around in hardware
92 * 13AP. not affected - worked around in hardware
93 * 14AP. we always deassert INIT during bootup
94 * 15AP. not affected - worked around in hardware
95 * 16AP. not affected - worked around in hardware
96 * 17AP. not affected - worked around in hardware
97 * 18AP. not affected - worked around in hardware
98 * 19AP. not affected - worked around in BIOS
100 * If this sounds worrying believe me these bugs are either ___RARE___,
101 * or are signal timing bugs worked around in hardware and there's
102 * about nothing of note with C stepping upwards.
105 /* The 'big kernel lock' */
106 spinlock_cacheline_t kernel_flag_cacheline = {SPIN_LOCK_UNLOCKED};
108 struct tlb_state cpu_tlbstate[NR_CPUS] __cacheline_aligned = {[0 ... NR_CPUS-1] = { &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 | INT_DEST_ADDR_MODE;
121 static inline int __prepare_ICR2 (unsigned int mask)
123 return SET_APIC_DEST_FIELD(mask);
126 static inline 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);
158 static inline void send_IPI_mask_bitmask(int mask, int vector)
170 apic_wait_icr_idle();
173 * prepare target chip field
175 cfg = __prepare_ICR2(mask);
176 apic_write_around(APIC_ICR2, cfg);
181 cfg = __prepare_ICR(0, vector);
184 * Send the IPI. The write to APIC_ICR fires this off.
186 apic_write_around(APIC_ICR, cfg);
188 __restore_flags(flags);
191 static inline void send_IPI_mask_sequence(int mask, int vector)
193 unsigned long cfg, flags;
194 unsigned int query_cpu, query_mask;
197 * Hack. The clustered APIC addressing mode doesn't allow us to send
198 * to an arbitrary mask, so I do a unicasts to each CPU instead. This
199 * should be modified to do 1 message per cluster ID - mbligh
205 for (query_cpu = 0; query_cpu < NR_CPUS; ++query_cpu) {
206 query_mask = 1 << query_cpu;
207 if (query_mask & mask) {
212 apic_wait_icr_idle();
215 * prepare target chip field
217 if(clustered_apic_mode == CLUSTERED_APIC_XAPIC)
218 cfg = __prepare_ICR2(cpu_to_physical_apicid(query_cpu));
220 cfg = __prepare_ICR2(cpu_to_logical_apicid(query_cpu));
221 apic_write_around(APIC_ICR2, cfg);
226 cfg = __prepare_ICR(0, vector);
229 * Send the IPI. The write to APIC_ICR fires this off.
231 apic_write_around(APIC_ICR, cfg);
234 __restore_flags(flags);
237 static inline void send_IPI_mask(int mask, int vector)
239 if (clustered_apic_mode)
240 send_IPI_mask_sequence(mask, vector);
242 send_IPI_mask_bitmask(mask, vector);
245 static inline void send_IPI_allbutself(int vector)
248 * if there are no other CPUs in the system then
249 * we get an APIC send error if we try to broadcast.
250 * thus we have to avoid sending IPIs in this case.
252 if (!(smp_num_cpus > 1))
255 if (clustered_apic_mode) {
256 // Pointless. Use send_IPI_mask to do this instead
259 if (smp_num_cpus > 1) {
260 for (cpu = 0; cpu < smp_num_cpus; ++cpu) {
261 if (cpu != smp_processor_id())
262 send_IPI_mask(1 << cpu, vector);
266 __send_IPI_shortcut(APIC_DEST_ALLBUT, vector);
271 static inline void send_IPI_all(int vector)
273 if (clustered_apic_mode) {
274 // Pointless. Use send_IPI_mask to do this instead
277 for (cpu = 0; cpu < smp_num_cpus; ++cpu) {
278 send_IPI_mask(1 << cpu, vector);
281 __send_IPI_shortcut(APIC_DEST_ALLINC, vector);
286 * Smarter SMP flushing macros.
287 * c/o Linus Torvalds.
289 * These mean you can really definitely utterly forget about
290 * writing to user space from interrupts. (Its not allowed anyway).
292 * Optimizations Manfred Spraul <manfred@colorfullife.com>
295 static volatile unsigned long flush_cpumask;
296 static struct mm_struct * flush_mm;
297 static unsigned long flush_va;
298 static spinlock_t tlbstate_lock = SPIN_LOCK_UNLOCKED;
299 #define FLUSH_ALL 0xffffffff
302 * We cannot call mmdrop() because we are in interrupt context,
303 * instead update mm->cpu_vm_mask.
305 * We need to reload %cr3 since the page tables may be going
306 * away frm under us...
308 static void inline leave_mm (unsigned long cpu)
310 BUG_ON(cpu_tlbstate[cpu].state == TLBSTATE_OK);
311 clear_bit(cpu, &cpu_tlbstate[cpu].active_mm->cpu_vm_mask);
312 load_cr3(swapper_pg_dir);
317 * The flush IPI assumes that a thread switch happens in this order:
318 * [cpu0: the cpu that switches]
319 * 1) switch_mm() either 1a) or 1b)
320 * 1a) thread switch to a different mm
321 * 1a1) clear_bit(cpu, &old_mm->cpu_vm_mask);
322 * Stop ipi delivery for the old mm. This is not synchronized with
323 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
324 * for the wrong mm, and in the worst case we perform a superflous
326 * 1a2) set cpu_tlbstate to TLBSTATE_OK
327 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
328 * was in lazy tlb mode.
329 * 1a3) update cpu_tlbstate[].active_mm
330 * Now cpu0 accepts tlb flushes for the new mm.
331 * 1a4) set_bit(cpu, &new_mm->cpu_vm_mask);
332 * Now the other cpus will send tlb flush ipis.
334 * 1b) thread switch without mm change
335 * cpu_tlbstate[].active_mm is correct, cpu0 already handles
337 * 1b1) set cpu_tlbstate to TLBSTATE_OK
338 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
339 * Atomically set the bit [other cpus will start sending flush ipis],
341 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
342 * 2) switch %%esp, ie current
344 * The interrupt must handle 2 special cases:
345 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
346 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
347 * runs in kernel space, the cpu could load tlb entries for user space
350 * The good news is that cpu_tlbstate is local to each cpu, no
351 * write/read ordering problems.
357 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
358 * 2) Leave the mm if we are in the lazy tlb mode.
361 asmlinkage void smp_invalidate_interrupt (void)
363 unsigned long cpu = smp_processor_id();
365 if (!test_bit(cpu, &flush_cpumask))
368 * This was a BUG() but until someone can quote me the
369 * line from the intel manual that guarantees an IPI to
370 * multiple CPUs is retried _only_ on the erroring CPUs
371 * its staying as a return
376 if (flush_mm == cpu_tlbstate[cpu].active_mm) {
377 if (cpu_tlbstate[cpu].state == TLBSTATE_OK) {
378 if (flush_va == FLUSH_ALL)
381 __flush_tlb_one(flush_va);
386 clear_bit(cpu, &flush_cpumask);
389 static void flush_tlb_others (unsigned long cpumask, struct mm_struct *mm,
393 * A couple of (to be removed) sanity checks:
395 * - we do not send IPIs to not-yet booted CPUs.
396 * - current CPU must not be in mask
397 * - mask must exist :)
401 if ((cpumask & cpu_online_map) != cpumask)
403 if (cpumask & (1 << smp_processor_id()))
409 * i'm not happy about this global shared spinlock in the
410 * MM hot path, but we'll see how contended it is.
411 * Temporarily this turns IRQs off, so that lockups are
412 * detected by the NMI watchdog.
414 spin_lock(&tlbstate_lock);
418 atomic_set_mask(cpumask, &flush_cpumask);
420 * We have to send the IPI only to
423 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR);
425 while (flush_cpumask)
426 /* nothing. lockup detection does not belong here */;
430 spin_unlock(&tlbstate_lock);
433 void flush_tlb_current_task(void)
435 struct mm_struct *mm = current->mm;
436 unsigned long cpu_mask = mm->cpu_vm_mask & ~(1 << smp_processor_id());
440 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
443 void flush_tlb_mm (struct mm_struct * mm)
445 unsigned long cpu_mask = mm->cpu_vm_mask & ~(1 << smp_processor_id());
447 if (current->active_mm == mm) {
451 leave_mm(smp_processor_id());
454 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
457 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
459 struct mm_struct *mm = vma->vm_mm;
460 unsigned long cpu_mask = mm->cpu_vm_mask & ~(1 << smp_processor_id());
462 if (current->active_mm == mm) {
466 leave_mm(smp_processor_id());
470 flush_tlb_others(cpu_mask, mm, va);
473 static inline void do_flush_tlb_all_local(void)
475 unsigned long cpu = smp_processor_id();
478 if (cpu_tlbstate[cpu].state == TLBSTATE_LAZY)
482 static void flush_tlb_all_ipi(void* info)
484 do_flush_tlb_all_local();
487 void flush_tlb_all(void)
489 smp_call_function (flush_tlb_all_ipi,0,1,1);
491 do_flush_tlb_all_local();
495 * this function sends a 'reschedule' IPI to another CPU.
496 * it goes straight through and wastes no time serializing
497 * anything. Worst case is that we lose a reschedule ...
500 void fastcall smp_send_reschedule(int cpu)
502 send_IPI_mask(1 << cpu, RESCHEDULE_VECTOR);
506 * Structure and data for smp_call_function(). This is designed to minimise
507 * static memory requirements. It also looks cleaner.
509 static spinlock_t call_lock = SPIN_LOCK_UNLOCKED;
511 struct call_data_struct {
512 void (*func) (void *info);
519 static struct call_data_struct * call_data;
522 * this function sends a 'generic call function' IPI to all other CPUs
526 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
529 * [SUMMARY] Run a function on all other CPUs.
530 * <func> The function to run. This must be fast and non-blocking.
531 * <info> An arbitrary pointer to pass to the function.
532 * <nonatomic> currently unused.
533 * <wait> If true, wait (atomically) until function has completed on other CPUs.
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 struct call_data_struct data;
542 int cpus = smp_num_cpus-1;
549 atomic_set(&data.started, 0);
552 atomic_set(&data.finished, 0);
554 spin_lock(&call_lock);
557 /* Send a message to all other CPUs and wait for them to respond */
558 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
560 /* Wait for response */
561 while (atomic_read(&data.started) != cpus)
565 while (atomic_read(&data.finished) != cpus)
567 spin_unlock(&call_lock);
572 static void stop_this_cpu (void * dummy)
577 clear_bit(smp_processor_id(), &cpu_online_map);
579 disable_local_APIC();
580 if (cpu_data[smp_processor_id()].hlt_works_ok)
581 for(;;) __asm__("hlt");
586 * this function calls the 'stop' function on all other CPUs in the system.
589 void smp_send_stop(void)
591 smp_call_function(stop_this_cpu, NULL, 1, 0);
595 disable_local_APIC();
600 * Reschedule call back. Nothing to do,
601 * all the work is done automatically when
602 * we return from the interrupt.
604 asmlinkage void smp_reschedule_interrupt(void)
609 asmlinkage void smp_call_function_interrupt(void)
611 void (*func) (void *info) = call_data->func;
612 void *info = call_data->info;
613 int wait = call_data->wait;
617 * Notify initiating CPU that I've grabbed the data and am
618 * about to execute the function
621 atomic_inc(&call_data->started);
623 * At this point the info structure may be out of scope unless wait==1
628 atomic_inc(&call_data->finished);