1 /* smp.c: Sparc64 SMP support.
3 * Copyright (C) 1997, 2007 David S. Miller (davem@davemloft.net)
6 #include <linux/module.h>
7 #include <linux/kernel.h>
8 #include <linux/sched.h>
10 #include <linux/pagemap.h>
11 #include <linux/threads.h>
12 #include <linux/smp.h>
13 #include <linux/interrupt.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/delay.h>
16 #include <linux/init.h>
17 #include <linux/spinlock.h>
19 #include <linux/seq_file.h>
20 #include <linux/cache.h>
21 #include <linux/jiffies.h>
22 #include <linux/profile.h>
23 #include <linux/bootmem.h>
26 #include <asm/ptrace.h>
27 #include <asm/atomic.h>
28 #include <asm/tlbflush.h>
29 #include <asm/mmu_context.h>
30 #include <asm/cpudata.h>
31 #include <asm/hvtramp.h>
35 #include <asm/irq_regs.h>
37 #include <asm/pgtable.h>
38 #include <asm/oplib.h>
39 #include <asm/uaccess.h>
40 #include <asm/timer.h>
41 #include <asm/starfire.h>
43 #include <asm/sections.h>
45 #include <asm/mdesc.h>
47 #include <asm/hypervisor.h>
49 extern void calibrate_delay(void);
51 int sparc64_multi_core __read_mostly
;
53 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_NONE
;
54 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_NONE
;
55 DEFINE_PER_CPU(cpumask_t
, cpu_sibling_map
) = CPU_MASK_NONE
;
56 cpumask_t cpu_core_map
[NR_CPUS
] __read_mostly
=
57 { [0 ... NR_CPUS
-1] = CPU_MASK_NONE
};
59 EXPORT_SYMBOL(cpu_possible_map
);
60 EXPORT_SYMBOL(cpu_online_map
);
61 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
62 EXPORT_SYMBOL(cpu_core_map
);
64 static cpumask_t smp_commenced_mask
;
66 void smp_info(struct seq_file
*m
)
70 seq_printf(m
, "State:\n");
71 for_each_online_cpu(i
)
72 seq_printf(m
, "CPU%d:\t\tonline\n", i
);
75 void smp_bogo(struct seq_file
*m
)
79 for_each_online_cpu(i
)
81 "Cpu%dClkTck\t: %016lx\n",
82 i
, cpu_data(i
).clock_tick
);
85 static __cacheline_aligned_in_smp
DEFINE_SPINLOCK(call_lock
);
87 extern void setup_sparc64_timer(void);
89 static volatile unsigned long callin_flag
= 0;
91 void __devinit
smp_callin(void)
93 int cpuid
= hard_smp_processor_id();
95 __local_per_cpu_offset
= __per_cpu_offset(cpuid
);
97 if (tlb_type
== hypervisor
)
98 sun4v_ktsb_register();
102 setup_sparc64_timer();
104 if (cheetah_pcache_forced_on
)
105 cheetah_enable_pcache();
110 __asm__
__volatile__("membar #Sync\n\t"
111 "flush %%g6" : : : "memory");
113 /* Clear this or we will die instantly when we
114 * schedule back to this idler...
116 current_thread_info()->new_child
= 0;
118 /* Attach to the address space of init_task. */
119 atomic_inc(&init_mm
.mm_count
);
120 current
->active_mm
= &init_mm
;
122 while (!cpu_isset(cpuid
, smp_commenced_mask
))
125 spin_lock(&call_lock
);
126 cpu_set(cpuid
, cpu_online_map
);
127 spin_unlock(&call_lock
);
129 /* idle thread is expected to have preempt disabled */
135 printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
136 panic("SMP bolixed\n");
139 /* This tick register synchronization scheme is taken entirely from
140 * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
142 * The only change I've made is to rework it so that the master
143 * initiates the synchonization instead of the slave. -DaveM
147 #define SLAVE (SMP_CACHE_BYTES/sizeof(unsigned long))
149 #define NUM_ROUNDS 64 /* magic value */
150 #define NUM_ITERS 5 /* likewise */
152 static DEFINE_SPINLOCK(itc_sync_lock
);
153 static unsigned long go
[SLAVE
+ 1];
155 #define DEBUG_TICK_SYNC 0
157 static inline long get_delta (long *rt
, long *master
)
159 unsigned long best_t0
= 0, best_t1
= ~0UL, best_tm
= 0;
160 unsigned long tcenter
, t0
, t1
, tm
;
163 for (i
= 0; i
< NUM_ITERS
; i
++) {
164 t0
= tick_ops
->get_tick();
167 while (!(tm
= go
[SLAVE
]))
171 t1
= tick_ops
->get_tick();
173 if (t1
- t0
< best_t1
- best_t0
)
174 best_t0
= t0
, best_t1
= t1
, best_tm
= tm
;
177 *rt
= best_t1
- best_t0
;
178 *master
= best_tm
- best_t0
;
180 /* average best_t0 and best_t1 without overflow: */
181 tcenter
= (best_t0
/2 + best_t1
/2);
182 if (best_t0
% 2 + best_t1
% 2 == 2)
184 return tcenter
- best_tm
;
187 void smp_synchronize_tick_client(void)
189 long i
, delta
, adj
, adjust_latency
= 0, done
= 0;
190 unsigned long flags
, rt
, master_time_stamp
, bound
;
193 long rt
; /* roundtrip time */
194 long master
; /* master's timestamp */
195 long diff
; /* difference between midpoint and master's timestamp */
196 long lat
; /* estimate of itc adjustment latency */
205 local_irq_save(flags
);
207 for (i
= 0; i
< NUM_ROUNDS
; i
++) {
208 delta
= get_delta(&rt
, &master_time_stamp
);
210 done
= 1; /* let's lock on to this... */
216 adjust_latency
+= -delta
;
217 adj
= -delta
+ adjust_latency
/4;
221 tick_ops
->add_tick(adj
);
225 t
[i
].master
= master_time_stamp
;
227 t
[i
].lat
= adjust_latency
/4;
231 local_irq_restore(flags
);
234 for (i
= 0; i
< NUM_ROUNDS
; i
++)
235 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
236 t
[i
].rt
, t
[i
].master
, t
[i
].diff
, t
[i
].lat
);
239 printk(KERN_INFO
"CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
240 "maxerr %lu cycles)\n", smp_processor_id(), delta
, rt
);
243 static void smp_start_sync_tick_client(int cpu
);
245 static void smp_synchronize_one_tick(int cpu
)
247 unsigned long flags
, i
;
251 smp_start_sync_tick_client(cpu
);
253 /* wait for client to be ready */
257 /* now let the client proceed into his loop */
261 spin_lock_irqsave(&itc_sync_lock
, flags
);
263 for (i
= 0; i
< NUM_ROUNDS
*NUM_ITERS
; i
++) {
268 go
[SLAVE
] = tick_ops
->get_tick();
272 spin_unlock_irqrestore(&itc_sync_lock
, flags
);
275 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
276 /* XXX Put this in some common place. XXX */
277 static unsigned long kimage_addr_to_ra(void *p
)
279 unsigned long val
= (unsigned long) p
;
281 return kern_base
+ (val
- KERNBASE
);
284 static void ldom_startcpu_cpuid(unsigned int cpu
, unsigned long thread_reg
)
286 extern unsigned long sparc64_ttable_tl0
;
287 extern unsigned long kern_locked_tte_data
;
288 extern int bigkernel
;
289 struct hvtramp_descr
*hdesc
;
290 unsigned long trampoline_ra
;
291 struct trap_per_cpu
*tb
;
292 u64 tte_vaddr
, tte_data
;
293 unsigned long hv_err
;
295 hdesc
= kzalloc(sizeof(*hdesc
), GFP_KERNEL
);
297 printk(KERN_ERR
"ldom_startcpu_cpuid: Cannot allocate "
303 hdesc
->num_mappings
= (bigkernel
? 2 : 1);
305 tb
= &trap_block
[cpu
];
308 hdesc
->fault_info_va
= (unsigned long) &tb
->fault_info
;
309 hdesc
->fault_info_pa
= kimage_addr_to_ra(&tb
->fault_info
);
311 hdesc
->thread_reg
= thread_reg
;
313 tte_vaddr
= (unsigned long) KERNBASE
;
314 tte_data
= kern_locked_tte_data
;
316 hdesc
->maps
[0].vaddr
= tte_vaddr
;
317 hdesc
->maps
[0].tte
= tte_data
;
319 tte_vaddr
+= 0x400000;
320 tte_data
+= 0x400000;
321 hdesc
->maps
[1].vaddr
= tte_vaddr
;
322 hdesc
->maps
[1].tte
= tte_data
;
325 trampoline_ra
= kimage_addr_to_ra(hv_cpu_startup
);
327 hv_err
= sun4v_cpu_start(cpu
, trampoline_ra
,
328 kimage_addr_to_ra(&sparc64_ttable_tl0
),
331 printk(KERN_ERR
"ldom_startcpu_cpuid: sun4v_cpu_start() "
332 "gives error %lu\n", hv_err
);
336 extern unsigned long sparc64_cpu_startup
;
338 /* The OBP cpu startup callback truncates the 3rd arg cookie to
339 * 32-bits (I think) so to be safe we have it read the pointer
340 * contained here so we work on >4GB machines. -DaveM
342 static struct thread_info
*cpu_new_thread
= NULL
;
344 static int __devinit
smp_boot_one_cpu(unsigned int cpu
)
346 struct trap_per_cpu
*tb
= &trap_block
[cpu
];
347 unsigned long entry
=
348 (unsigned long)(&sparc64_cpu_startup
);
349 unsigned long cookie
=
350 (unsigned long)(&cpu_new_thread
);
351 struct task_struct
*p
;
358 cpu_new_thread
= task_thread_info(p
);
360 if (tlb_type
== hypervisor
) {
361 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
362 if (ldom_domaining_enabled
)
363 ldom_startcpu_cpuid(cpu
,
364 (unsigned long) cpu_new_thread
);
367 prom_startcpu_cpuid(cpu
, entry
, cookie
);
369 struct device_node
*dp
= of_find_node_by_cpuid(cpu
);
371 prom_startcpu(dp
->node
, entry
, cookie
);
374 for (timeout
= 0; timeout
< 50000; timeout
++) {
383 printk("Processor %d is stuck.\n", cpu
);
386 cpu_new_thread
= NULL
;
396 static void spitfire_xcall_helper(u64 data0
, u64 data1
, u64 data2
, u64 pstate
, unsigned long cpu
)
401 if (this_is_starfire
) {
402 /* map to real upaid */
403 cpu
= (((cpu
& 0x3c) << 1) |
404 ((cpu
& 0x40) >> 4) |
408 target
= (cpu
<< 14) | 0x70;
410 /* Ok, this is the real Spitfire Errata #54.
411 * One must read back from a UDB internal register
412 * after writes to the UDB interrupt dispatch, but
413 * before the membar Sync for that write.
414 * So we use the high UDB control register (ASI 0x7f,
415 * ADDR 0x20) for the dummy read. -DaveM
418 __asm__
__volatile__(
419 "wrpr %1, %2, %%pstate\n\t"
420 "stxa %4, [%0] %3\n\t"
421 "stxa %5, [%0+%8] %3\n\t"
423 "stxa %6, [%0+%8] %3\n\t"
425 "stxa %%g0, [%7] %3\n\t"
428 "ldxa [%%g1] 0x7f, %%g0\n\t"
431 : "r" (pstate
), "i" (PSTATE_IE
), "i" (ASI_INTR_W
),
432 "r" (data0
), "r" (data1
), "r" (data2
), "r" (target
),
433 "r" (0x10), "0" (tmp
)
436 /* NOTE: PSTATE_IE is still clear. */
439 __asm__
__volatile__("ldxa [%%g0] %1, %0"
441 : "i" (ASI_INTR_DISPATCH_STAT
));
443 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
450 } while (result
& 0x1);
451 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
454 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
455 smp_processor_id(), result
);
462 static __inline__
void spitfire_xcall_deliver(u64 data0
, u64 data1
, u64 data2
, cpumask_t mask
)
467 __asm__
__volatile__("rdpr %%pstate, %0" : "=r" (pstate
));
468 for_each_cpu_mask(i
, mask
)
469 spitfire_xcall_helper(data0
, data1
, data2
, pstate
, i
);
472 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
473 * packet, but we have no use for that. However we do take advantage of
474 * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
476 static void cheetah_xcall_deliver(u64 data0
, u64 data1
, u64 data2
, cpumask_t mask
)
479 int nack_busy_id
, is_jbus
, need_more
;
481 if (cpus_empty(mask
))
484 /* Unfortunately, someone at Sun had the brilliant idea to make the
485 * busy/nack fields hard-coded by ITID number for this Ultra-III
486 * derivative processor.
488 __asm__ ("rdpr %%ver, %0" : "=r" (ver
));
489 is_jbus
= ((ver
>> 32) == __JALAPENO_ID
||
490 (ver
>> 32) == __SERRANO_ID
);
492 __asm__
__volatile__("rdpr %%pstate, %0" : "=r" (pstate
));
496 __asm__
__volatile__("wrpr %0, %1, %%pstate\n\t"
497 : : "r" (pstate
), "i" (PSTATE_IE
));
499 /* Setup the dispatch data registers. */
500 __asm__
__volatile__("stxa %0, [%3] %6\n\t"
501 "stxa %1, [%4] %6\n\t"
502 "stxa %2, [%5] %6\n\t"
505 : "r" (data0
), "r" (data1
), "r" (data2
),
506 "r" (0x40), "r" (0x50), "r" (0x60),
513 for_each_cpu_mask(i
, mask
) {
514 u64 target
= (i
<< 14) | 0x70;
517 target
|= (nack_busy_id
<< 24);
518 __asm__
__volatile__(
519 "stxa %%g0, [%0] %1\n\t"
522 : "r" (target
), "i" (ASI_INTR_W
));
524 if (nack_busy_id
== 32) {
531 /* Now, poll for completion. */
536 stuck
= 100000 * nack_busy_id
;
538 __asm__
__volatile__("ldxa [%%g0] %1, %0"
539 : "=r" (dispatch_stat
)
540 : "i" (ASI_INTR_DISPATCH_STAT
));
541 if (dispatch_stat
== 0UL) {
542 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
544 if (unlikely(need_more
)) {
546 for_each_cpu_mask(i
, mask
) {
558 } while (dispatch_stat
& 0x5555555555555555UL
);
560 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
563 if ((dispatch_stat
& ~(0x5555555555555555UL
)) == 0) {
564 /* Busy bits will not clear, continue instead
565 * of freezing up on this cpu.
567 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
568 smp_processor_id(), dispatch_stat
);
570 int i
, this_busy_nack
= 0;
572 /* Delay some random time with interrupts enabled
573 * to prevent deadlock.
575 udelay(2 * nack_busy_id
);
577 /* Clear out the mask bits for cpus which did not
580 for_each_cpu_mask(i
, mask
) {
584 check_mask
= (0x2UL
<< (2*i
));
586 check_mask
= (0x2UL
<<
588 if ((dispatch_stat
& check_mask
) == 0)
591 if (this_busy_nack
== 64)
600 /* Multi-cpu list version. */
601 static void hypervisor_xcall_deliver(u64 data0
, u64 data1
, u64 data2
, cpumask_t mask
)
603 struct trap_per_cpu
*tb
;
606 cpumask_t error_mask
;
607 unsigned long flags
, status
;
608 int cnt
, retries
, this_cpu
, prev_sent
, i
;
610 if (cpus_empty(mask
))
613 /* We have to do this whole thing with interrupts fully disabled.
614 * Otherwise if we send an xcall from interrupt context it will
615 * corrupt both our mondo block and cpu list state.
617 * One consequence of this is that we cannot use timeout mechanisms
618 * that depend upon interrupts being delivered locally. So, for
619 * example, we cannot sample jiffies and expect it to advance.
621 * Fortunately, udelay() uses %stick/%tick so we can use that.
623 local_irq_save(flags
);
625 this_cpu
= smp_processor_id();
626 tb
= &trap_block
[this_cpu
];
628 mondo
= __va(tb
->cpu_mondo_block_pa
);
634 cpu_list
= __va(tb
->cpu_list_pa
);
636 /* Setup the initial cpu list. */
638 for_each_cpu_mask(i
, mask
)
641 cpus_clear(error_mask
);
645 int forward_progress
, n_sent
;
647 status
= sun4v_cpu_mondo_send(cnt
,
649 tb
->cpu_mondo_block_pa
);
651 /* HV_EOK means all cpus received the xcall, we're done. */
652 if (likely(status
== HV_EOK
))
655 /* First, see if we made any forward progress.
657 * The hypervisor indicates successful sends by setting
658 * cpu list entries to the value 0xffff.
661 for (i
= 0; i
< cnt
; i
++) {
662 if (likely(cpu_list
[i
] == 0xffff))
666 forward_progress
= 0;
667 if (n_sent
> prev_sent
)
668 forward_progress
= 1;
672 /* If we get a HV_ECPUERROR, then one or more of the cpus
673 * in the list are in error state. Use the cpu_state()
674 * hypervisor call to find out which cpus are in error state.
676 if (unlikely(status
== HV_ECPUERROR
)) {
677 for (i
= 0; i
< cnt
; i
++) {
685 err
= sun4v_cpu_state(cpu
);
687 err
== HV_CPU_STATE_ERROR
) {
688 cpu_list
[i
] = 0xffff;
689 cpu_set(cpu
, error_mask
);
692 } else if (unlikely(status
!= HV_EWOULDBLOCK
))
693 goto fatal_mondo_error
;
695 /* Don't bother rewriting the CPU list, just leave the
696 * 0xffff and non-0xffff entries in there and the
697 * hypervisor will do the right thing.
699 * Only advance timeout state if we didn't make any
702 if (unlikely(!forward_progress
)) {
703 if (unlikely(++retries
> 10000))
704 goto fatal_mondo_timeout
;
706 /* Delay a little bit to let other cpus catch up
707 * on their cpu mondo queue work.
713 local_irq_restore(flags
);
715 if (unlikely(!cpus_empty(error_mask
)))
716 goto fatal_mondo_cpu_error
;
720 fatal_mondo_cpu_error
:
721 printk(KERN_CRIT
"CPU[%d]: SUN4V mondo cpu error, some target cpus "
722 "were in error state\n",
724 printk(KERN_CRIT
"CPU[%d]: Error mask [ ", this_cpu
);
725 for_each_cpu_mask(i
, error_mask
)
731 local_irq_restore(flags
);
732 printk(KERN_CRIT
"CPU[%d]: SUN4V mondo timeout, no forward "
733 " progress after %d retries.\n",
735 goto dump_cpu_list_and_out
;
738 local_irq_restore(flags
);
739 printk(KERN_CRIT
"CPU[%d]: Unexpected SUN4V mondo error %lu\n",
741 printk(KERN_CRIT
"CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
742 "mondo_block_pa(%lx)\n",
743 this_cpu
, cnt
, tb
->cpu_list_pa
, tb
->cpu_mondo_block_pa
);
745 dump_cpu_list_and_out
:
746 printk(KERN_CRIT
"CPU[%d]: CPU list [ ", this_cpu
);
747 for (i
= 0; i
< cnt
; i
++)
748 printk("%u ", cpu_list
[i
]);
752 /* Send cross call to all processors mentioned in MASK
755 static void smp_cross_call_masked(unsigned long *func
, u32 ctx
, u64 data1
, u64 data2
, cpumask_t mask
)
757 u64 data0
= (((u64
)ctx
)<<32 | (((u64
)func
) & 0xffffffff));
758 int this_cpu
= get_cpu();
760 cpus_and(mask
, mask
, cpu_online_map
);
761 cpu_clear(this_cpu
, mask
);
763 if (tlb_type
== spitfire
)
764 spitfire_xcall_deliver(data0
, data1
, data2
, mask
);
765 else if (tlb_type
== cheetah
|| tlb_type
== cheetah_plus
)
766 cheetah_xcall_deliver(data0
, data1
, data2
, mask
);
768 hypervisor_xcall_deliver(data0
, data1
, data2
, mask
);
769 /* NOTE: Caller runs local copy on master. */
774 extern unsigned long xcall_sync_tick
;
776 static void smp_start_sync_tick_client(int cpu
)
778 cpumask_t mask
= cpumask_of_cpu(cpu
);
780 smp_cross_call_masked(&xcall_sync_tick
,
784 /* Send cross call to all processors except self. */
785 #define smp_cross_call(func, ctx, data1, data2) \
786 smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
788 struct call_data_struct
{
789 void (*func
) (void *info
);
795 static struct call_data_struct
*call_data
;
797 extern unsigned long xcall_call_function
;
800 * smp_call_function(): Run a function on all other CPUs.
801 * @func: The function to run. This must be fast and non-blocking.
802 * @info: An arbitrary pointer to pass to the function.
803 * @nonatomic: currently unused.
804 * @wait: If true, wait (atomically) until function has completed on other CPUs.
806 * Returns 0 on success, else a negative status code. Does not return until
807 * remote CPUs are nearly ready to execute <<func>> or are or have executed.
809 * You must not call this function with disabled interrupts or from a
810 * hardware interrupt handler or from a bottom half handler.
812 static int smp_call_function_mask(void (*func
)(void *info
), void *info
,
813 int nonatomic
, int wait
, cpumask_t mask
)
815 struct call_data_struct data
;
818 /* Can deadlock when called with interrupts disabled */
819 WARN_ON(irqs_disabled());
823 atomic_set(&data
.finished
, 0);
826 spin_lock(&call_lock
);
828 cpu_clear(smp_processor_id(), mask
);
829 cpus
= cpus_weight(mask
);
836 smp_cross_call_masked(&xcall_call_function
, 0, 0, 0, mask
);
838 /* Wait for response */
839 while (atomic_read(&data
.finished
) != cpus
)
843 spin_unlock(&call_lock
);
848 int smp_call_function(void (*func
)(void *info
), void *info
,
849 int nonatomic
, int wait
)
851 return smp_call_function_mask(func
, info
, nonatomic
, wait
,
855 void smp_call_function_client(int irq
, struct pt_regs
*regs
)
857 void (*func
) (void *info
) = call_data
->func
;
858 void *info
= call_data
->info
;
860 clear_softint(1 << irq
);
861 if (call_data
->wait
) {
862 /* let initiator proceed only after completion */
864 atomic_inc(&call_data
->finished
);
866 /* let initiator proceed after getting data */
867 atomic_inc(&call_data
->finished
);
872 static void tsb_sync(void *info
)
874 struct trap_per_cpu
*tp
= &trap_block
[raw_smp_processor_id()];
875 struct mm_struct
*mm
= info
;
877 /* It is not valid to test "currrent->active_mm == mm" here.
879 * The value of "current" is not changed atomically with
880 * switch_mm(). But that's OK, we just need to check the
881 * current cpu's trap block PGD physical address.
883 if (tp
->pgd_paddr
== __pa(mm
->pgd
))
884 tsb_context_switch(mm
);
887 void smp_tsb_sync(struct mm_struct
*mm
)
889 smp_call_function_mask(tsb_sync
, mm
, 0, 1, mm
->cpu_vm_mask
);
892 extern unsigned long xcall_flush_tlb_mm
;
893 extern unsigned long xcall_flush_tlb_pending
;
894 extern unsigned long xcall_flush_tlb_kernel_range
;
895 extern unsigned long xcall_report_regs
;
896 extern unsigned long xcall_receive_signal
;
897 extern unsigned long xcall_new_mmu_context_version
;
899 #ifdef DCACHE_ALIASING_POSSIBLE
900 extern unsigned long xcall_flush_dcache_page_cheetah
;
902 extern unsigned long xcall_flush_dcache_page_spitfire
;
904 #ifdef CONFIG_DEBUG_DCFLUSH
905 extern atomic_t dcpage_flushes
;
906 extern atomic_t dcpage_flushes_xcall
;
909 static __inline__
void __local_flush_dcache_page(struct page
*page
)
911 #ifdef DCACHE_ALIASING_POSSIBLE
912 __flush_dcache_page(page_address(page
),
913 ((tlb_type
== spitfire
) &&
914 page_mapping(page
) != NULL
));
916 if (page_mapping(page
) != NULL
&&
917 tlb_type
== spitfire
)
918 __flush_icache_page(__pa(page_address(page
)));
922 void smp_flush_dcache_page_impl(struct page
*page
, int cpu
)
924 cpumask_t mask
= cpumask_of_cpu(cpu
);
927 if (tlb_type
== hypervisor
)
930 #ifdef CONFIG_DEBUG_DCFLUSH
931 atomic_inc(&dcpage_flushes
);
934 this_cpu
= get_cpu();
936 if (cpu
== this_cpu
) {
937 __local_flush_dcache_page(page
);
938 } else if (cpu_online(cpu
)) {
939 void *pg_addr
= page_address(page
);
942 if (tlb_type
== spitfire
) {
944 ((u64
)&xcall_flush_dcache_page_spitfire
);
945 if (page_mapping(page
) != NULL
)
946 data0
|= ((u64
)1 << 32);
947 spitfire_xcall_deliver(data0
,
951 } else if (tlb_type
== cheetah
|| tlb_type
== cheetah_plus
) {
952 #ifdef DCACHE_ALIASING_POSSIBLE
954 ((u64
)&xcall_flush_dcache_page_cheetah
);
955 cheetah_xcall_deliver(data0
,
960 #ifdef CONFIG_DEBUG_DCFLUSH
961 atomic_inc(&dcpage_flushes_xcall
);
968 void flush_dcache_page_all(struct mm_struct
*mm
, struct page
*page
)
970 void *pg_addr
= page_address(page
);
971 cpumask_t mask
= cpu_online_map
;
975 if (tlb_type
== hypervisor
)
978 this_cpu
= get_cpu();
980 cpu_clear(this_cpu
, mask
);
982 #ifdef CONFIG_DEBUG_DCFLUSH
983 atomic_inc(&dcpage_flushes
);
985 if (cpus_empty(mask
))
987 if (tlb_type
== spitfire
) {
988 data0
= ((u64
)&xcall_flush_dcache_page_spitfire
);
989 if (page_mapping(page
) != NULL
)
990 data0
|= ((u64
)1 << 32);
991 spitfire_xcall_deliver(data0
,
995 } else if (tlb_type
== cheetah
|| tlb_type
== cheetah_plus
) {
996 #ifdef DCACHE_ALIASING_POSSIBLE
997 data0
= ((u64
)&xcall_flush_dcache_page_cheetah
);
998 cheetah_xcall_deliver(data0
,
1003 #ifdef CONFIG_DEBUG_DCFLUSH
1004 atomic_inc(&dcpage_flushes_xcall
);
1007 __local_flush_dcache_page(page
);
1012 static void __smp_receive_signal_mask(cpumask_t mask
)
1014 smp_cross_call_masked(&xcall_receive_signal
, 0, 0, 0, mask
);
1017 void smp_receive_signal(int cpu
)
1019 cpumask_t mask
= cpumask_of_cpu(cpu
);
1021 if (cpu_online(cpu
))
1022 __smp_receive_signal_mask(mask
);
1025 void smp_receive_signal_client(int irq
, struct pt_regs
*regs
)
1027 clear_softint(1 << irq
);
1030 void smp_new_mmu_context_version_client(int irq
, struct pt_regs
*regs
)
1032 struct mm_struct
*mm
;
1033 unsigned long flags
;
1035 clear_softint(1 << irq
);
1037 /* See if we need to allocate a new TLB context because
1038 * the version of the one we are using is now out of date.
1040 mm
= current
->active_mm
;
1041 if (unlikely(!mm
|| (mm
== &init_mm
)))
1044 spin_lock_irqsave(&mm
->context
.lock
, flags
);
1046 if (unlikely(!CTX_VALID(mm
->context
)))
1047 get_new_mmu_context(mm
);
1049 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
1051 load_secondary_context(mm
);
1052 __flush_tlb_mm(CTX_HWBITS(mm
->context
),
1056 void smp_new_mmu_context_version(void)
1058 smp_cross_call(&xcall_new_mmu_context_version
, 0, 0, 0);
1061 void smp_report_regs(void)
1063 smp_cross_call(&xcall_report_regs
, 0, 0, 0);
1066 /* We know that the window frames of the user have been flushed
1067 * to the stack before we get here because all callers of us
1068 * are flush_tlb_*() routines, and these run after flush_cache_*()
1069 * which performs the flushw.
1071 * The SMP TLB coherency scheme we use works as follows:
1073 * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1074 * space has (potentially) executed on, this is the heuristic
1075 * we use to avoid doing cross calls.
1077 * Also, for flushing from kswapd and also for clones, we
1078 * use cpu_vm_mask as the list of cpus to make run the TLB.
1080 * 2) TLB context numbers are shared globally across all processors
1081 * in the system, this allows us to play several games to avoid
1084 * One invariant is that when a cpu switches to a process, and
1085 * that processes tsk->active_mm->cpu_vm_mask does not have the
1086 * current cpu's bit set, that tlb context is flushed locally.
1088 * If the address space is non-shared (ie. mm->count == 1) we avoid
1089 * cross calls when we want to flush the currently running process's
1090 * tlb state. This is done by clearing all cpu bits except the current
1091 * processor's in current->active_mm->cpu_vm_mask and performing the
1092 * flush locally only. This will force any subsequent cpus which run
1093 * this task to flush the context from the local tlb if the process
1094 * migrates to another cpu (again).
1096 * 3) For shared address spaces (threads) and swapping we bite the
1097 * bullet for most cases and perform the cross call (but only to
1098 * the cpus listed in cpu_vm_mask).
1100 * The performance gain from "optimizing" away the cross call for threads is
1101 * questionable (in theory the big win for threads is the massive sharing of
1102 * address space state across processors).
1105 /* This currently is only used by the hugetlb arch pre-fault
1106 * hook on UltraSPARC-III+ and later when changing the pagesize
1107 * bits of the context register for an address space.
1109 void smp_flush_tlb_mm(struct mm_struct
*mm
)
1111 u32 ctx
= CTX_HWBITS(mm
->context
);
1112 int cpu
= get_cpu();
1114 if (atomic_read(&mm
->mm_users
) == 1) {
1115 mm
->cpu_vm_mask
= cpumask_of_cpu(cpu
);
1116 goto local_flush_and_out
;
1119 smp_cross_call_masked(&xcall_flush_tlb_mm
,
1123 local_flush_and_out
:
1124 __flush_tlb_mm(ctx
, SECONDARY_CONTEXT
);
1129 void smp_flush_tlb_pending(struct mm_struct
*mm
, unsigned long nr
, unsigned long *vaddrs
)
1131 u32 ctx
= CTX_HWBITS(mm
->context
);
1132 int cpu
= get_cpu();
1134 if (mm
== current
->active_mm
&& atomic_read(&mm
->mm_users
) == 1)
1135 mm
->cpu_vm_mask
= cpumask_of_cpu(cpu
);
1137 smp_cross_call_masked(&xcall_flush_tlb_pending
,
1138 ctx
, nr
, (unsigned long) vaddrs
,
1141 __flush_tlb_pending(ctx
, nr
, vaddrs
);
1146 void smp_flush_tlb_kernel_range(unsigned long start
, unsigned long end
)
1149 end
= PAGE_ALIGN(end
);
1151 smp_cross_call(&xcall_flush_tlb_kernel_range
,
1154 __flush_tlb_kernel_range(start
, end
);
1159 /* #define CAPTURE_DEBUG */
1160 extern unsigned long xcall_capture
;
1162 static atomic_t smp_capture_depth
= ATOMIC_INIT(0);
1163 static atomic_t smp_capture_registry
= ATOMIC_INIT(0);
1164 static unsigned long penguins_are_doing_time
;
1166 void smp_capture(void)
1168 int result
= atomic_add_ret(1, &smp_capture_depth
);
1171 int ncpus
= num_online_cpus();
1173 #ifdef CAPTURE_DEBUG
1174 printk("CPU[%d]: Sending penguins to jail...",
1175 smp_processor_id());
1177 penguins_are_doing_time
= 1;
1178 membar_storestore_loadstore();
1179 atomic_inc(&smp_capture_registry
);
1180 smp_cross_call(&xcall_capture
, 0, 0, 0);
1181 while (atomic_read(&smp_capture_registry
) != ncpus
)
1183 #ifdef CAPTURE_DEBUG
1189 void smp_release(void)
1191 if (atomic_dec_and_test(&smp_capture_depth
)) {
1192 #ifdef CAPTURE_DEBUG
1193 printk("CPU[%d]: Giving pardon to "
1194 "imprisoned penguins\n",
1195 smp_processor_id());
1197 penguins_are_doing_time
= 0;
1198 membar_storeload_storestore();
1199 atomic_dec(&smp_capture_registry
);
1203 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
1204 * can service tlb flush xcalls...
1206 extern void prom_world(int);
1208 void smp_penguin_jailcell(int irq
, struct pt_regs
*regs
)
1210 clear_softint(1 << irq
);
1214 __asm__
__volatile__("flushw");
1216 atomic_inc(&smp_capture_registry
);
1217 membar_storeload_storestore();
1218 while (penguins_are_doing_time
)
1220 atomic_dec(&smp_capture_registry
);
1226 /* /proc/profile writes can call this, don't __init it please. */
1227 int setup_profiling_timer(unsigned int multiplier
)
1232 void __init
smp_prepare_cpus(unsigned int max_cpus
)
1236 void __devinit
smp_prepare_boot_cpu(void)
1240 void __devinit
smp_fill_in_sib_core_maps(void)
1244 for_each_present_cpu(i
) {
1247 cpus_clear(cpu_core_map
[i
]);
1248 if (cpu_data(i
).core_id
== 0) {
1249 cpu_set(i
, cpu_core_map
[i
]);
1253 for_each_present_cpu(j
) {
1254 if (cpu_data(i
).core_id
==
1255 cpu_data(j
).core_id
)
1256 cpu_set(j
, cpu_core_map
[i
]);
1260 for_each_present_cpu(i
) {
1263 cpus_clear(per_cpu(cpu_sibling_map
, i
));
1264 if (cpu_data(i
).proc_id
== -1) {
1265 cpu_set(i
, per_cpu(cpu_sibling_map
, i
));
1269 for_each_present_cpu(j
) {
1270 if (cpu_data(i
).proc_id
==
1271 cpu_data(j
).proc_id
)
1272 cpu_set(j
, per_cpu(cpu_sibling_map
, i
));
1277 int __cpuinit
__cpu_up(unsigned int cpu
)
1279 int ret
= smp_boot_one_cpu(cpu
);
1282 cpu_set(cpu
, smp_commenced_mask
);
1283 while (!cpu_isset(cpu
, cpu_online_map
))
1285 if (!cpu_isset(cpu
, cpu_online_map
)) {
1288 /* On SUN4V, writes to %tick and %stick are
1291 if (tlb_type
!= hypervisor
)
1292 smp_synchronize_one_tick(cpu
);
1298 #ifdef CONFIG_HOTPLUG_CPU
1299 void cpu_play_dead(void)
1301 int cpu
= smp_processor_id();
1302 unsigned long pstate
;
1306 if (tlb_type
== hypervisor
) {
1307 struct trap_per_cpu
*tb
= &trap_block
[cpu
];
1309 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO
,
1310 tb
->cpu_mondo_pa
, 0);
1311 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO
,
1312 tb
->dev_mondo_pa
, 0);
1313 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR
,
1314 tb
->resum_mondo_pa
, 0);
1315 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR
,
1316 tb
->nonresum_mondo_pa
, 0);
1319 cpu_clear(cpu
, smp_commenced_mask
);
1320 membar_safe("#Sync");
1322 local_irq_disable();
1324 __asm__
__volatile__(
1325 "rdpr %%pstate, %0\n\t"
1326 "wrpr %0, %1, %%pstate"
1334 int __cpu_disable(void)
1336 int cpu
= smp_processor_id();
1340 for_each_cpu_mask(i
, cpu_core_map
[cpu
])
1341 cpu_clear(cpu
, cpu_core_map
[i
]);
1342 cpus_clear(cpu_core_map
[cpu
]);
1344 for_each_cpu_mask(i
, per_cpu(cpu_sibling_map
, cpu
))
1345 cpu_clear(cpu
, per_cpu(cpu_sibling_map
, i
));
1346 cpus_clear(per_cpu(cpu_sibling_map
, cpu
));
1353 spin_lock(&call_lock
);
1354 cpu_clear(cpu
, cpu_online_map
);
1355 spin_unlock(&call_lock
);
1359 /* Make sure no interrupts point to this cpu. */
1364 local_irq_disable();
1369 void __cpu_die(unsigned int cpu
)
1373 for (i
= 0; i
< 100; i
++) {
1375 if (!cpu_isset(cpu
, smp_commenced_mask
))
1379 if (cpu_isset(cpu
, smp_commenced_mask
)) {
1380 printk(KERN_ERR
"CPU %u didn't die...\n", cpu
);
1382 #if defined(CONFIG_SUN_LDOMS)
1383 unsigned long hv_err
;
1387 hv_err
= sun4v_cpu_stop(cpu
);
1388 if (hv_err
== HV_EOK
) {
1389 cpu_clear(cpu
, cpu_present_map
);
1392 } while (--limit
> 0);
1394 printk(KERN_ERR
"sun4v_cpu_stop() fails err=%lu\n",
1402 void __init
smp_cpus_done(unsigned int max_cpus
)
1406 void smp_send_reschedule(int cpu
)
1408 smp_receive_signal(cpu
);
1411 /* This is a nop because we capture all other cpus
1412 * anyways when making the PROM active.
1414 void smp_send_stop(void)
1418 unsigned long __per_cpu_base __read_mostly
;
1419 unsigned long __per_cpu_shift __read_mostly
;
1421 EXPORT_SYMBOL(__per_cpu_base
);
1422 EXPORT_SYMBOL(__per_cpu_shift
);
1424 void __init
real_setup_per_cpu_areas(void)
1426 unsigned long goal
, size
, i
;
1429 /* Copy section for each CPU (we discard the original) */
1430 goal
= PERCPU_ENOUGH_ROOM
;
1432 __per_cpu_shift
= PAGE_SHIFT
;
1433 for (size
= PAGE_SIZE
; size
< goal
; size
<<= 1UL)
1436 ptr
= alloc_bootmem_pages(size
* NR_CPUS
);
1438 __per_cpu_base
= ptr
- __per_cpu_start
;
1440 for (i
= 0; i
< NR_CPUS
; i
++, ptr
+= size
)
1441 memcpy(ptr
, __per_cpu_start
, __per_cpu_end
- __per_cpu_start
);
1443 /* Setup %g5 for the boot cpu. */
1444 __local_per_cpu_offset
= __per_cpu_offset(smp_processor_id());