1 /* smp.c: Sparc64 SMP support.
3 * Copyright (C) 1997 David S. Miller (davem@caip.rutgers.edu)
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/smp_lock.h>
14 #include <linux/interrupt.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/delay.h>
17 #include <linux/init.h>
18 #include <linux/spinlock.h>
20 #include <linux/seq_file.h>
21 #include <linux/cache.h>
22 #include <linux/jiffies.h>
23 #include <linux/profile.h>
24 #include <linux/bootmem.h>
27 #include <asm/ptrace.h>
28 #include <asm/atomic.h>
29 #include <asm/tlbflush.h>
30 #include <asm/mmu_context.h>
31 #include <asm/cpudata.h>
35 #include <asm/pgtable.h>
36 #include <asm/oplib.h>
37 #include <asm/uaccess.h>
38 #include <asm/timer.h>
39 #include <asm/starfire.h>
42 extern int linux_num_cpus
;
43 extern void calibrate_delay(void);
45 /* Please don't make this stuff initdata!!! --DaveM */
46 static unsigned char boot_cpu_id
;
48 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_NONE
;
49 cpumask_t phys_cpu_present_map __read_mostly
= CPU_MASK_NONE
;
50 static cpumask_t smp_commenced_mask
;
51 static cpumask_t cpu_callout_map
;
53 void smp_info(struct seq_file
*m
)
57 seq_printf(m
, "State:\n");
58 for (i
= 0; i
< NR_CPUS
; i
++) {
61 "CPU%d:\t\tonline\n", i
);
65 void smp_bogo(struct seq_file
*m
)
69 for (i
= 0; i
< NR_CPUS
; i
++)
72 "Cpu%dBogo\t: %lu.%02lu\n"
73 "Cpu%dClkTck\t: %016lx\n",
74 i
, cpu_data(i
).udelay_val
/ (500000/HZ
),
75 (cpu_data(i
).udelay_val
/ (5000/HZ
)) % 100,
76 i
, cpu_data(i
).clock_tick
);
79 void __init
smp_store_cpu_info(int id
)
83 /* multiplier and counter set by
84 smp_setup_percpu_timer() */
85 cpu_data(id
).udelay_val
= loops_per_jiffy
;
87 cpu_find_by_mid(id
, &cpu_node
);
88 cpu_data(id
).clock_tick
= prom_getintdefault(cpu_node
,
89 "clock-frequency", 0);
91 cpu_data(id
).pgcache_size
= 0;
92 cpu_data(id
).pte_cache
[0] = NULL
;
93 cpu_data(id
).pte_cache
[1] = NULL
;
94 cpu_data(id
).pgd_cache
= NULL
;
95 cpu_data(id
).idle_volume
= 1;
97 cpu_data(id
).dcache_size
= prom_getintdefault(cpu_node
, "dcache-size",
99 cpu_data(id
).dcache_line_size
=
100 prom_getintdefault(cpu_node
, "dcache-line-size", 32);
101 cpu_data(id
).icache_size
= prom_getintdefault(cpu_node
, "icache-size",
103 cpu_data(id
).icache_line_size
=
104 prom_getintdefault(cpu_node
, "icache-line-size", 32);
105 cpu_data(id
).ecache_size
= prom_getintdefault(cpu_node
, "ecache-size",
107 cpu_data(id
).ecache_line_size
=
108 prom_getintdefault(cpu_node
, "ecache-line-size", 64);
109 printk("CPU[%d]: Caches "
110 "D[sz(%d):line_sz(%d)] "
111 "I[sz(%d):line_sz(%d)] "
112 "E[sz(%d):line_sz(%d)]\n",
114 cpu_data(id
).dcache_size
, cpu_data(id
).dcache_line_size
,
115 cpu_data(id
).icache_size
, cpu_data(id
).icache_line_size
,
116 cpu_data(id
).ecache_size
, cpu_data(id
).ecache_line_size
);
119 static void smp_setup_percpu_timer(void);
121 static volatile unsigned long callin_flag
= 0;
123 extern void inherit_locked_prom_mappings(int save_p
);
125 static inline void cpu_setup_percpu_base(unsigned long cpu_id
)
127 __asm__
__volatile__("mov %0, %%g5\n\t"
128 "stxa %0, [%1] %2\n\t"
131 : "r" (__per_cpu_offset(cpu_id
)),
132 "r" (TSB_REG
), "i" (ASI_IMMU
));
135 void __init
smp_callin(void)
137 int cpuid
= hard_smp_processor_id();
139 inherit_locked_prom_mappings(0);
143 cpu_setup_percpu_base(cpuid
);
145 smp_setup_percpu_timer();
147 if (cheetah_pcache_forced_on
)
148 cheetah_enable_pcache();
153 smp_store_cpu_info(cpuid
);
155 __asm__
__volatile__("membar #Sync\n\t"
156 "flush %%g6" : : : "memory");
158 /* Clear this or we will die instantly when we
159 * schedule back to this idler...
161 current_thread_info()->new_child
= 0;
163 /* Attach to the address space of init_task. */
164 atomic_inc(&init_mm
.mm_count
);
165 current
->active_mm
= &init_mm
;
167 while (!cpu_isset(cpuid
, smp_commenced_mask
))
170 cpu_set(cpuid
, cpu_online_map
);
175 printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
176 panic("SMP bolixed\n");
179 static unsigned long current_tick_offset __read_mostly
;
181 /* This tick register synchronization scheme is taken entirely from
182 * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
184 * The only change I've made is to rework it so that the master
185 * initiates the synchonization instead of the slave. -DaveM
189 #define SLAVE (SMP_CACHE_BYTES/sizeof(unsigned long))
191 #define NUM_ROUNDS 64 /* magic value */
192 #define NUM_ITERS 5 /* likewise */
194 static DEFINE_SPINLOCK(itc_sync_lock
);
195 static unsigned long go
[SLAVE
+ 1];
197 #define DEBUG_TICK_SYNC 0
199 static inline long get_delta (long *rt
, long *master
)
201 unsigned long best_t0
= 0, best_t1
= ~0UL, best_tm
= 0;
202 unsigned long tcenter
, t0
, t1
, tm
;
205 for (i
= 0; i
< NUM_ITERS
; i
++) {
206 t0
= tick_ops
->get_tick();
209 while (!(tm
= go
[SLAVE
]))
213 t1
= tick_ops
->get_tick();
215 if (t1
- t0
< best_t1
- best_t0
)
216 best_t0
= t0
, best_t1
= t1
, best_tm
= tm
;
219 *rt
= best_t1
- best_t0
;
220 *master
= best_tm
- best_t0
;
222 /* average best_t0 and best_t1 without overflow: */
223 tcenter
= (best_t0
/2 + best_t1
/2);
224 if (best_t0
% 2 + best_t1
% 2 == 2)
226 return tcenter
- best_tm
;
229 void smp_synchronize_tick_client(void)
231 long i
, delta
, adj
, adjust_latency
= 0, done
= 0;
232 unsigned long flags
, rt
, master_time_stamp
, bound
;
235 long rt
; /* roundtrip time */
236 long master
; /* master's timestamp */
237 long diff
; /* difference between midpoint and master's timestamp */
238 long lat
; /* estimate of itc adjustment latency */
247 local_irq_save(flags
);
249 for (i
= 0; i
< NUM_ROUNDS
; i
++) {
250 delta
= get_delta(&rt
, &master_time_stamp
);
252 done
= 1; /* let's lock on to this... */
258 adjust_latency
+= -delta
;
259 adj
= -delta
+ adjust_latency
/4;
263 tick_ops
->add_tick(adj
, current_tick_offset
);
267 t
[i
].master
= master_time_stamp
;
269 t
[i
].lat
= adjust_latency
/4;
273 local_irq_restore(flags
);
276 for (i
= 0; i
< NUM_ROUNDS
; i
++)
277 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
278 t
[i
].rt
, t
[i
].master
, t
[i
].diff
, t
[i
].lat
);
281 printk(KERN_INFO
"CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
282 "maxerr %lu cycles)\n", smp_processor_id(), delta
, rt
);
285 static void smp_start_sync_tick_client(int cpu
);
287 static void smp_synchronize_one_tick(int cpu
)
289 unsigned long flags
, i
;
293 smp_start_sync_tick_client(cpu
);
295 /* wait for client to be ready */
299 /* now let the client proceed into his loop */
303 spin_lock_irqsave(&itc_sync_lock
, flags
);
305 for (i
= 0; i
< NUM_ROUNDS
*NUM_ITERS
; i
++) {
310 go
[SLAVE
] = tick_ops
->get_tick();
314 spin_unlock_irqrestore(&itc_sync_lock
, flags
);
317 extern unsigned long sparc64_cpu_startup
;
319 /* The OBP cpu startup callback truncates the 3rd arg cookie to
320 * 32-bits (I think) so to be safe we have it read the pointer
321 * contained here so we work on >4GB machines. -DaveM
323 static struct thread_info
*cpu_new_thread
= NULL
;
325 static int __devinit
smp_boot_one_cpu(unsigned int cpu
)
327 unsigned long entry
=
328 (unsigned long)(&sparc64_cpu_startup
);
329 unsigned long cookie
=
330 (unsigned long)(&cpu_new_thread
);
331 struct task_struct
*p
;
332 int timeout
, ret
, cpu_node
;
336 cpu_new_thread
= p
->thread_info
;
337 cpu_set(cpu
, cpu_callout_map
);
339 cpu_find_by_mid(cpu
, &cpu_node
);
340 prom_startcpu(cpu_node
, entry
, cookie
);
342 for (timeout
= 0; timeout
< 5000000; timeout
++) {
350 printk("Processor %d is stuck.\n", cpu
);
351 cpu_clear(cpu
, cpu_callout_map
);
354 cpu_new_thread
= NULL
;
359 static void spitfire_xcall_helper(u64 data0
, u64 data1
, u64 data2
, u64 pstate
, unsigned long cpu
)
364 if (this_is_starfire
) {
365 /* map to real upaid */
366 cpu
= (((cpu
& 0x3c) << 1) |
367 ((cpu
& 0x40) >> 4) |
371 target
= (cpu
<< 14) | 0x70;
373 /* Ok, this is the real Spitfire Errata #54.
374 * One must read back from a UDB internal register
375 * after writes to the UDB interrupt dispatch, but
376 * before the membar Sync for that write.
377 * So we use the high UDB control register (ASI 0x7f,
378 * ADDR 0x20) for the dummy read. -DaveM
381 __asm__
__volatile__(
382 "wrpr %1, %2, %%pstate\n\t"
383 "stxa %4, [%0] %3\n\t"
384 "stxa %5, [%0+%8] %3\n\t"
386 "stxa %6, [%0+%8] %3\n\t"
388 "stxa %%g0, [%7] %3\n\t"
391 "ldxa [%%g1] 0x7f, %%g0\n\t"
394 : "r" (pstate
), "i" (PSTATE_IE
), "i" (ASI_INTR_W
),
395 "r" (data0
), "r" (data1
), "r" (data2
), "r" (target
),
396 "r" (0x10), "0" (tmp
)
399 /* NOTE: PSTATE_IE is still clear. */
402 __asm__
__volatile__("ldxa [%%g0] %1, %0"
404 : "i" (ASI_INTR_DISPATCH_STAT
));
406 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
413 } while (result
& 0x1);
414 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
417 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
418 smp_processor_id(), result
);
425 static __inline__
void spitfire_xcall_deliver(u64 data0
, u64 data1
, u64 data2
, cpumask_t mask
)
430 __asm__
__volatile__("rdpr %%pstate, %0" : "=r" (pstate
));
431 for_each_cpu_mask(i
, mask
)
432 spitfire_xcall_helper(data0
, data1
, data2
, pstate
, i
);
435 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
436 * packet, but we have no use for that. However we do take advantage of
437 * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
439 static void cheetah_xcall_deliver(u64 data0
, u64 data1
, u64 data2
, cpumask_t mask
)
442 int nack_busy_id
, is_jalapeno
;
444 if (cpus_empty(mask
))
447 /* Unfortunately, someone at Sun had the brilliant idea to make the
448 * busy/nack fields hard-coded by ITID number for this Ultra-III
449 * derivative processor.
451 __asm__ ("rdpr %%ver, %0" : "=r" (ver
));
452 is_jalapeno
= ((ver
>> 32) == 0x003e0016);
454 __asm__
__volatile__("rdpr %%pstate, %0" : "=r" (pstate
));
457 __asm__
__volatile__("wrpr %0, %1, %%pstate\n\t"
458 : : "r" (pstate
), "i" (PSTATE_IE
));
460 /* Setup the dispatch data registers. */
461 __asm__
__volatile__("stxa %0, [%3] %6\n\t"
462 "stxa %1, [%4] %6\n\t"
463 "stxa %2, [%5] %6\n\t"
466 : "r" (data0
), "r" (data1
), "r" (data2
),
467 "r" (0x40), "r" (0x50), "r" (0x60),
474 for_each_cpu_mask(i
, mask
) {
475 u64 target
= (i
<< 14) | 0x70;
478 target
|= (nack_busy_id
<< 24);
479 __asm__
__volatile__(
480 "stxa %%g0, [%0] %1\n\t"
483 : "r" (target
), "i" (ASI_INTR_W
));
488 /* Now, poll for completion. */
493 stuck
= 100000 * nack_busy_id
;
495 __asm__
__volatile__("ldxa [%%g0] %1, %0"
496 : "=r" (dispatch_stat
)
497 : "i" (ASI_INTR_DISPATCH_STAT
));
498 if (dispatch_stat
== 0UL) {
499 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
505 } while (dispatch_stat
& 0x5555555555555555UL
);
507 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
510 if ((dispatch_stat
& ~(0x5555555555555555UL
)) == 0) {
511 /* Busy bits will not clear, continue instead
512 * of freezing up on this cpu.
514 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
515 smp_processor_id(), dispatch_stat
);
517 int i
, this_busy_nack
= 0;
519 /* Delay some random time with interrupts enabled
520 * to prevent deadlock.
522 udelay(2 * nack_busy_id
);
524 /* Clear out the mask bits for cpus which did not
527 for_each_cpu_mask(i
, mask
) {
531 check_mask
= (0x2UL
<< (2*i
));
533 check_mask
= (0x2UL
<<
535 if ((dispatch_stat
& check_mask
) == 0)
545 /* Send cross call to all processors mentioned in MASK
548 static void smp_cross_call_masked(unsigned long *func
, u32 ctx
, u64 data1
, u64 data2
, cpumask_t mask
)
550 u64 data0
= (((u64
)ctx
)<<32 | (((u64
)func
) & 0xffffffff));
551 int this_cpu
= get_cpu();
553 cpus_and(mask
, mask
, cpu_online_map
);
554 cpu_clear(this_cpu
, mask
);
556 if (tlb_type
== spitfire
)
557 spitfire_xcall_deliver(data0
, data1
, data2
, mask
);
559 cheetah_xcall_deliver(data0
, data1
, data2
, mask
);
560 /* NOTE: Caller runs local copy on master. */
565 extern unsigned long xcall_sync_tick
;
567 static void smp_start_sync_tick_client(int cpu
)
569 cpumask_t mask
= cpumask_of_cpu(cpu
);
571 smp_cross_call_masked(&xcall_sync_tick
,
575 /* Send cross call to all processors except self. */
576 #define smp_cross_call(func, ctx, data1, data2) \
577 smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
579 struct call_data_struct
{
580 void (*func
) (void *info
);
586 static DEFINE_SPINLOCK(call_lock
);
587 static struct call_data_struct
*call_data
;
589 extern unsigned long xcall_call_function
;
592 * You must not call this function with disabled interrupts or from a
593 * hardware interrupt handler or from a bottom half handler.
595 int smp_call_function(void (*func
)(void *info
), void *info
,
596 int nonatomic
, int wait
)
598 struct call_data_struct data
;
599 int cpus
= num_online_cpus() - 1;
605 /* Can deadlock when called with interrupts disabled */
606 WARN_ON(irqs_disabled());
610 atomic_set(&data
.finished
, 0);
613 spin_lock(&call_lock
);
617 smp_cross_call(&xcall_call_function
, 0, 0, 0);
620 * Wait for other cpus to complete function or at
621 * least snap the call data.
624 while (atomic_read(&data
.finished
) != cpus
) {
631 spin_unlock(&call_lock
);
636 spin_unlock(&call_lock
);
637 printk("XCALL: Remote cpus not responding, ncpus=%ld finished=%ld\n",
638 (long) num_online_cpus() - 1L,
639 (long) atomic_read(&data
.finished
));
643 void smp_call_function_client(int irq
, struct pt_regs
*regs
)
645 void (*func
) (void *info
) = call_data
->func
;
646 void *info
= call_data
->info
;
648 clear_softint(1 << irq
);
649 if (call_data
->wait
) {
650 /* let initiator proceed only after completion */
652 atomic_inc(&call_data
->finished
);
654 /* let initiator proceed after getting data */
655 atomic_inc(&call_data
->finished
);
660 extern unsigned long xcall_flush_tlb_mm
;
661 extern unsigned long xcall_flush_tlb_pending
;
662 extern unsigned long xcall_flush_tlb_kernel_range
;
663 extern unsigned long xcall_flush_tlb_all_spitfire
;
664 extern unsigned long xcall_flush_tlb_all_cheetah
;
665 extern unsigned long xcall_report_regs
;
666 extern unsigned long xcall_receive_signal
;
668 #ifdef DCACHE_ALIASING_POSSIBLE
669 extern unsigned long xcall_flush_dcache_page_cheetah
;
671 extern unsigned long xcall_flush_dcache_page_spitfire
;
673 #ifdef CONFIG_DEBUG_DCFLUSH
674 extern atomic_t dcpage_flushes
;
675 extern atomic_t dcpage_flushes_xcall
;
678 static __inline__
void __local_flush_dcache_page(struct page
*page
)
680 #ifdef DCACHE_ALIASING_POSSIBLE
681 __flush_dcache_page(page_address(page
),
682 ((tlb_type
== spitfire
) &&
683 page_mapping(page
) != NULL
));
685 if (page_mapping(page
) != NULL
&&
686 tlb_type
== spitfire
)
687 __flush_icache_page(__pa(page_address(page
)));
691 void smp_flush_dcache_page_impl(struct page
*page
, int cpu
)
693 cpumask_t mask
= cpumask_of_cpu(cpu
);
694 int this_cpu
= get_cpu();
696 #ifdef CONFIG_DEBUG_DCFLUSH
697 atomic_inc(&dcpage_flushes
);
699 if (cpu
== this_cpu
) {
700 __local_flush_dcache_page(page
);
701 } else if (cpu_online(cpu
)) {
702 void *pg_addr
= page_address(page
);
705 if (tlb_type
== spitfire
) {
707 ((u64
)&xcall_flush_dcache_page_spitfire
);
708 if (page_mapping(page
) != NULL
)
709 data0
|= ((u64
)1 << 32);
710 spitfire_xcall_deliver(data0
,
715 #ifdef DCACHE_ALIASING_POSSIBLE
717 ((u64
)&xcall_flush_dcache_page_cheetah
);
718 cheetah_xcall_deliver(data0
,
723 #ifdef CONFIG_DEBUG_DCFLUSH
724 atomic_inc(&dcpage_flushes_xcall
);
731 void flush_dcache_page_all(struct mm_struct
*mm
, struct page
*page
)
733 void *pg_addr
= page_address(page
);
734 cpumask_t mask
= cpu_online_map
;
736 int this_cpu
= get_cpu();
738 cpu_clear(this_cpu
, mask
);
740 #ifdef CONFIG_DEBUG_DCFLUSH
741 atomic_inc(&dcpage_flushes
);
743 if (cpus_empty(mask
))
745 if (tlb_type
== spitfire
) {
746 data0
= ((u64
)&xcall_flush_dcache_page_spitfire
);
747 if (page_mapping(page
) != NULL
)
748 data0
|= ((u64
)1 << 32);
749 spitfire_xcall_deliver(data0
,
754 #ifdef DCACHE_ALIASING_POSSIBLE
755 data0
= ((u64
)&xcall_flush_dcache_page_cheetah
);
756 cheetah_xcall_deliver(data0
,
761 #ifdef CONFIG_DEBUG_DCFLUSH
762 atomic_inc(&dcpage_flushes_xcall
);
765 __local_flush_dcache_page(page
);
770 void smp_receive_signal(int cpu
)
772 cpumask_t mask
= cpumask_of_cpu(cpu
);
774 if (cpu_online(cpu
)) {
775 u64 data0
= (((u64
)&xcall_receive_signal
) & 0xffffffff);
777 if (tlb_type
== spitfire
)
778 spitfire_xcall_deliver(data0
, 0, 0, mask
);
780 cheetah_xcall_deliver(data0
, 0, 0, mask
);
784 void smp_receive_signal_client(int irq
, struct pt_regs
*regs
)
786 /* Just return, rtrap takes care of the rest. */
787 clear_softint(1 << irq
);
790 void smp_report_regs(void)
792 smp_cross_call(&xcall_report_regs
, 0, 0, 0);
795 void smp_flush_tlb_all(void)
797 if (tlb_type
== spitfire
)
798 smp_cross_call(&xcall_flush_tlb_all_spitfire
, 0, 0, 0);
800 smp_cross_call(&xcall_flush_tlb_all_cheetah
, 0, 0, 0);
804 /* We know that the window frames of the user have been flushed
805 * to the stack before we get here because all callers of us
806 * are flush_tlb_*() routines, and these run after flush_cache_*()
807 * which performs the flushw.
809 * The SMP TLB coherency scheme we use works as follows:
811 * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
812 * space has (potentially) executed on, this is the heuristic
813 * we use to avoid doing cross calls.
815 * Also, for flushing from kswapd and also for clones, we
816 * use cpu_vm_mask as the list of cpus to make run the TLB.
818 * 2) TLB context numbers are shared globally across all processors
819 * in the system, this allows us to play several games to avoid
822 * One invariant is that when a cpu switches to a process, and
823 * that processes tsk->active_mm->cpu_vm_mask does not have the
824 * current cpu's bit set, that tlb context is flushed locally.
826 * If the address space is non-shared (ie. mm->count == 1) we avoid
827 * cross calls when we want to flush the currently running process's
828 * tlb state. This is done by clearing all cpu bits except the current
829 * processor's in current->active_mm->cpu_vm_mask and performing the
830 * flush locally only. This will force any subsequent cpus which run
831 * this task to flush the context from the local tlb if the process
832 * migrates to another cpu (again).
834 * 3) For shared address spaces (threads) and swapping we bite the
835 * bullet for most cases and perform the cross call (but only to
836 * the cpus listed in cpu_vm_mask).
838 * The performance gain from "optimizing" away the cross call for threads is
839 * questionable (in theory the big win for threads is the massive sharing of
840 * address space state across processors).
843 /* This currently is only used by the hugetlb arch pre-fault
844 * hook on UltraSPARC-III+ and later when changing the pagesize
845 * bits of the context register for an address space.
847 void smp_flush_tlb_mm(struct mm_struct
*mm
)
849 u32 ctx
= CTX_HWBITS(mm
->context
);
852 if (atomic_read(&mm
->mm_users
) == 1) {
853 mm
->cpu_vm_mask
= cpumask_of_cpu(cpu
);
854 goto local_flush_and_out
;
857 smp_cross_call_masked(&xcall_flush_tlb_mm
,
862 __flush_tlb_mm(ctx
, SECONDARY_CONTEXT
);
867 void smp_flush_tlb_pending(struct mm_struct
*mm
, unsigned long nr
, unsigned long *vaddrs
)
869 u32 ctx
= CTX_HWBITS(mm
->context
);
872 if (mm
== current
->active_mm
&& atomic_read(&mm
->mm_users
) == 1)
873 mm
->cpu_vm_mask
= cpumask_of_cpu(cpu
);
875 smp_cross_call_masked(&xcall_flush_tlb_pending
,
876 ctx
, nr
, (unsigned long) vaddrs
,
879 __flush_tlb_pending(ctx
, nr
, vaddrs
);
884 void smp_flush_tlb_kernel_range(unsigned long start
, unsigned long end
)
887 end
= PAGE_ALIGN(end
);
889 smp_cross_call(&xcall_flush_tlb_kernel_range
,
892 __flush_tlb_kernel_range(start
, end
);
897 /* #define CAPTURE_DEBUG */
898 extern unsigned long xcall_capture
;
900 static atomic_t smp_capture_depth
= ATOMIC_INIT(0);
901 static atomic_t smp_capture_registry
= ATOMIC_INIT(0);
902 static unsigned long penguins_are_doing_time
;
904 void smp_capture(void)
906 int result
= atomic_add_ret(1, &smp_capture_depth
);
909 int ncpus
= num_online_cpus();
912 printk("CPU[%d]: Sending penguins to jail...",
915 penguins_are_doing_time
= 1;
916 membar_storestore_loadstore();
917 atomic_inc(&smp_capture_registry
);
918 smp_cross_call(&xcall_capture
, 0, 0, 0);
919 while (atomic_read(&smp_capture_registry
) != ncpus
)
927 void smp_release(void)
929 if (atomic_dec_and_test(&smp_capture_depth
)) {
931 printk("CPU[%d]: Giving pardon to "
932 "imprisoned penguins\n",
935 penguins_are_doing_time
= 0;
936 membar_storeload_storestore();
937 atomic_dec(&smp_capture_registry
);
941 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
942 * can service tlb flush xcalls...
944 extern void prom_world(int);
945 extern void save_alternate_globals(unsigned long *);
946 extern void restore_alternate_globals(unsigned long *);
947 void smp_penguin_jailcell(int irq
, struct pt_regs
*regs
)
949 unsigned long global_save
[24];
951 clear_softint(1 << irq
);
955 __asm__
__volatile__("flushw");
956 save_alternate_globals(global_save
);
958 atomic_inc(&smp_capture_registry
);
959 membar_storeload_storestore();
960 while (penguins_are_doing_time
)
962 restore_alternate_globals(global_save
);
963 atomic_dec(&smp_capture_registry
);
969 #define prof_multiplier(__cpu) cpu_data(__cpu).multiplier
970 #define prof_counter(__cpu) cpu_data(__cpu).counter
972 void smp_percpu_timer_interrupt(struct pt_regs
*regs
)
974 unsigned long compare
, tick
, pstate
;
975 int cpu
= smp_processor_id();
976 int user
= user_mode(regs
);
979 * Check for level 14 softint.
982 unsigned long tick_mask
= tick_ops
->softint_mask
;
984 if (!(get_softint() & tick_mask
)) {
985 extern void handler_irq(int, struct pt_regs
*);
987 handler_irq(14, regs
);
990 clear_softint(tick_mask
);
994 profile_tick(CPU_PROFILING
, regs
);
995 if (!--prof_counter(cpu
)) {
998 if (cpu
== boot_cpu_id
) {
999 kstat_this_cpu
.irqs
[0]++;
1000 timer_tick_interrupt(regs
);
1003 update_process_times(user
);
1007 prof_counter(cpu
) = prof_multiplier(cpu
);
1010 /* Guarantee that the following sequences execute
1013 __asm__
__volatile__("rdpr %%pstate, %0\n\t"
1014 "wrpr %0, %1, %%pstate"
1018 compare
= tick_ops
->add_compare(current_tick_offset
);
1019 tick
= tick_ops
->get_tick();
1021 /* Restore PSTATE_IE. */
1022 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
1025 } while (time_after_eq(tick
, compare
));
1028 static void __init
smp_setup_percpu_timer(void)
1030 int cpu
= smp_processor_id();
1031 unsigned long pstate
;
1033 prof_counter(cpu
) = prof_multiplier(cpu
) = 1;
1035 /* Guarantee that the following sequences execute
1038 __asm__
__volatile__("rdpr %%pstate, %0\n\t"
1039 "wrpr %0, %1, %%pstate"
1043 tick_ops
->init_tick(current_tick_offset
);
1045 /* Restore PSTATE_IE. */
1046 __asm__
__volatile__("wrpr %0, 0x0, %%pstate"
1051 void __init
smp_tick_init(void)
1053 boot_cpu_id
= hard_smp_processor_id();
1054 current_tick_offset
= timer_tick_offset
;
1056 cpu_set(boot_cpu_id
, cpu_online_map
);
1057 prof_counter(boot_cpu_id
) = prof_multiplier(boot_cpu_id
) = 1;
1060 /* /proc/profile writes can call this, don't __init it please. */
1061 static DEFINE_SPINLOCK(prof_setup_lock
);
1063 int setup_profiling_timer(unsigned int multiplier
)
1065 unsigned long flags
;
1068 if ((!multiplier
) || (timer_tick_offset
/ multiplier
) < 1000)
1071 spin_lock_irqsave(&prof_setup_lock
, flags
);
1072 for (i
= 0; i
< NR_CPUS
; i
++)
1073 prof_multiplier(i
) = multiplier
;
1074 current_tick_offset
= (timer_tick_offset
/ multiplier
);
1075 spin_unlock_irqrestore(&prof_setup_lock
, flags
);
1080 void __init
smp_prepare_cpus(unsigned int max_cpus
)
1085 while (!cpu_find_by_instance(instance
, NULL
, &mid
)) {
1087 cpu_set(mid
, phys_cpu_present_map
);
1091 if (num_possible_cpus() > max_cpus
) {
1093 while (!cpu_find_by_instance(instance
, NULL
, &mid
)) {
1094 if (mid
!= boot_cpu_id
) {
1095 cpu_clear(mid
, phys_cpu_present_map
);
1096 if (num_possible_cpus() <= max_cpus
)
1103 smp_store_cpu_info(boot_cpu_id
);
1106 void __devinit
smp_prepare_boot_cpu(void)
1108 if (hard_smp_processor_id() >= NR_CPUS
) {
1109 prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
1113 current_thread_info()->cpu
= hard_smp_processor_id();
1115 cpu_set(smp_processor_id(), cpu_online_map
);
1116 cpu_set(smp_processor_id(), phys_cpu_present_map
);
1119 int __devinit
__cpu_up(unsigned int cpu
)
1121 int ret
= smp_boot_one_cpu(cpu
);
1124 cpu_set(cpu
, smp_commenced_mask
);
1125 while (!cpu_isset(cpu
, cpu_online_map
))
1127 if (!cpu_isset(cpu
, cpu_online_map
)) {
1130 smp_synchronize_one_tick(cpu
);
1136 void __init
smp_cpus_done(unsigned int max_cpus
)
1138 unsigned long bogosum
= 0;
1141 for (i
= 0; i
< NR_CPUS
; i
++) {
1143 bogosum
+= cpu_data(i
).udelay_val
;
1145 printk("Total of %ld processors activated "
1146 "(%lu.%02lu BogoMIPS).\n",
1147 (long) num_online_cpus(),
1148 bogosum
/(500000/HZ
),
1149 (bogosum
/(5000/HZ
))%100);
1152 /* This needn't do anything as we do not sleep the cpu
1153 * inside of the idler task, so an interrupt is not needed
1154 * to get a clean fast response.
1156 * XXX Reverify this assumption... -DaveM
1158 * Addendum: We do want it to do something for the signal
1159 * delivery case, we detect that by just seeing
1160 * if we are trying to send this to an idler or not.
1162 void smp_send_reschedule(int cpu
)
1164 if (cpu_data(cpu
).idle_volume
== 0)
1165 smp_receive_signal(cpu
);
1168 /* This is a nop because we capture all other cpus
1169 * anyways when making the PROM active.
1171 void smp_send_stop(void)
1175 unsigned long __per_cpu_base __read_mostly
;
1176 unsigned long __per_cpu_shift __read_mostly
;
1178 EXPORT_SYMBOL(__per_cpu_base
);
1179 EXPORT_SYMBOL(__per_cpu_shift
);
1181 void __init
setup_per_cpu_areas(void)
1183 unsigned long goal
, size
, i
;
1185 /* Created by linker magic */
1186 extern char __per_cpu_start
[], __per_cpu_end
[];
1188 /* Copy section for each CPU (we discard the original) */
1189 goal
= ALIGN(__per_cpu_end
- __per_cpu_start
, PAGE_SIZE
);
1191 #ifdef CONFIG_MODULES
1192 if (goal
< PERCPU_ENOUGH_ROOM
)
1193 goal
= PERCPU_ENOUGH_ROOM
;
1195 __per_cpu_shift
= 0;
1196 for (size
= 1UL; size
< goal
; size
<<= 1UL)
1199 /* Make sure the resulting __per_cpu_base value
1200 * will fit in the 43-bit sign extended IMMU
1203 ptr
= __alloc_bootmem(size
* NR_CPUS
, PAGE_SIZE
,
1204 (unsigned long) __per_cpu_start
);
1206 __per_cpu_base
= ptr
- __per_cpu_start
;
1208 if ((__per_cpu_shift
< PAGE_SHIFT
) ||
1209 (__per_cpu_base
& ~PAGE_MASK
) ||
1210 (__per_cpu_base
!= (((long) __per_cpu_base
<< 20) >> 20))) {
1211 prom_printf("PER_CPU: Invalid layout, "
1212 "ptr[%p] shift[%lx] base[%lx]\n",
1213 ptr
, __per_cpu_shift
, __per_cpu_base
);
1217 for (i
= 0; i
< NR_CPUS
; i
++, ptr
+= size
)
1218 memcpy(ptr
, __per_cpu_start
, __per_cpu_end
- __per_cpu_start
);
1220 /* Finally, load in the boot cpu's base value.
1221 * We abuse the IMMU TSB register for trap handler
1222 * entry and exit loading of %g5. That is why it
1223 * has to be page aligned.
1225 cpu_setup_percpu_base(hard_smp_processor_id());