xref: /openbmc/linux/arch/sparc/kernel/smp_64.c (revision c8dbaa22)
1 /* smp.c: Sparc64 SMP support.
2  *
3  * Copyright (C) 1997, 2007, 2008 David S. Miller (davem@davemloft.net)
4  */
5 
6 #include <linux/export.h>
7 #include <linux/kernel.h>
8 #include <linux/sched/mm.h>
9 #include <linux/sched/hotplug.h>
10 #include <linux/mm.h>
11 #include <linux/pagemap.h>
12 #include <linux/threads.h>
13 #include <linux/smp.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>
19 #include <linux/fs.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>
25 #include <linux/vmalloc.h>
26 #include <linux/ftrace.h>
27 #include <linux/cpu.h>
28 #include <linux/slab.h>
29 #include <linux/kgdb.h>
30 
31 #include <asm/head.h>
32 #include <asm/ptrace.h>
33 #include <linux/atomic.h>
34 #include <asm/tlbflush.h>
35 #include <asm/mmu_context.h>
36 #include <asm/cpudata.h>
37 #include <asm/hvtramp.h>
38 #include <asm/io.h>
39 #include <asm/timer.h>
40 #include <asm/setup.h>
41 
42 #include <asm/irq.h>
43 #include <asm/irq_regs.h>
44 #include <asm/page.h>
45 #include <asm/pgtable.h>
46 #include <asm/oplib.h>
47 #include <linux/uaccess.h>
48 #include <asm/starfire.h>
49 #include <asm/tlb.h>
50 #include <asm/sections.h>
51 #include <asm/prom.h>
52 #include <asm/mdesc.h>
53 #include <asm/ldc.h>
54 #include <asm/hypervisor.h>
55 #include <asm/pcr.h>
56 
57 #include "cpumap.h"
58 #include "kernel.h"
59 
60 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
61 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
62 	{ [0 ... NR_CPUS-1] = CPU_MASK_NONE };
63 
64 cpumask_t cpu_core_sib_map[NR_CPUS] __read_mostly = {
65 	[0 ... NR_CPUS-1] = CPU_MASK_NONE };
66 
67 cpumask_t cpu_core_sib_cache_map[NR_CPUS] __read_mostly = {
68 	[0 ... NR_CPUS - 1] = CPU_MASK_NONE };
69 
70 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
71 EXPORT_SYMBOL(cpu_core_map);
72 EXPORT_SYMBOL(cpu_core_sib_map);
73 EXPORT_SYMBOL(cpu_core_sib_cache_map);
74 
75 static cpumask_t smp_commenced_mask;
76 
77 void smp_info(struct seq_file *m)
78 {
79 	int i;
80 
81 	seq_printf(m, "State:\n");
82 	for_each_online_cpu(i)
83 		seq_printf(m, "CPU%d:\t\tonline\n", i);
84 }
85 
86 void smp_bogo(struct seq_file *m)
87 {
88 	int i;
89 
90 	for_each_online_cpu(i)
91 		seq_printf(m,
92 			   "Cpu%dClkTck\t: %016lx\n",
93 			   i, cpu_data(i).clock_tick);
94 }
95 
96 extern void setup_sparc64_timer(void);
97 
98 static volatile unsigned long callin_flag = 0;
99 
100 void smp_callin(void)
101 {
102 	int cpuid = hard_smp_processor_id();
103 
104 	__local_per_cpu_offset = __per_cpu_offset(cpuid);
105 
106 	if (tlb_type == hypervisor)
107 		sun4v_ktsb_register();
108 
109 	__flush_tlb_all();
110 
111 	setup_sparc64_timer();
112 
113 	if (cheetah_pcache_forced_on)
114 		cheetah_enable_pcache();
115 
116 	callin_flag = 1;
117 	__asm__ __volatile__("membar #Sync\n\t"
118 			     "flush  %%g6" : : : "memory");
119 
120 	/* Clear this or we will die instantly when we
121 	 * schedule back to this idler...
122 	 */
123 	current_thread_info()->new_child = 0;
124 
125 	/* Attach to the address space of init_task. */
126 	mmgrab(&init_mm);
127 	current->active_mm = &init_mm;
128 
129 	/* inform the notifiers about the new cpu */
130 	notify_cpu_starting(cpuid);
131 
132 	while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
133 		rmb();
134 
135 	set_cpu_online(cpuid, true);
136 
137 	/* idle thread is expected to have preempt disabled */
138 	preempt_disable();
139 
140 	local_irq_enable();
141 
142 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
143 }
144 
145 void cpu_panic(void)
146 {
147 	printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
148 	panic("SMP bolixed\n");
149 }
150 
151 /* This tick register synchronization scheme is taken entirely from
152  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
153  *
154  * The only change I've made is to rework it so that the master
155  * initiates the synchonization instead of the slave. -DaveM
156  */
157 
158 #define MASTER	0
159 #define SLAVE	(SMP_CACHE_BYTES/sizeof(unsigned long))
160 
161 #define NUM_ROUNDS	64	/* magic value */
162 #define NUM_ITERS	5	/* likewise */
163 
164 static DEFINE_RAW_SPINLOCK(itc_sync_lock);
165 static unsigned long go[SLAVE + 1];
166 
167 #define DEBUG_TICK_SYNC	0
168 
169 static inline long get_delta (long *rt, long *master)
170 {
171 	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
172 	unsigned long tcenter, t0, t1, tm;
173 	unsigned long i;
174 
175 	for (i = 0; i < NUM_ITERS; i++) {
176 		t0 = tick_ops->get_tick();
177 		go[MASTER] = 1;
178 		membar_safe("#StoreLoad");
179 		while (!(tm = go[SLAVE]))
180 			rmb();
181 		go[SLAVE] = 0;
182 		wmb();
183 		t1 = tick_ops->get_tick();
184 
185 		if (t1 - t0 < best_t1 - best_t0)
186 			best_t0 = t0, best_t1 = t1, best_tm = tm;
187 	}
188 
189 	*rt = best_t1 - best_t0;
190 	*master = best_tm - best_t0;
191 
192 	/* average best_t0 and best_t1 without overflow: */
193 	tcenter = (best_t0/2 + best_t1/2);
194 	if (best_t0 % 2 + best_t1 % 2 == 2)
195 		tcenter++;
196 	return tcenter - best_tm;
197 }
198 
199 void smp_synchronize_tick_client(void)
200 {
201 	long i, delta, adj, adjust_latency = 0, done = 0;
202 	unsigned long flags, rt, master_time_stamp;
203 #if DEBUG_TICK_SYNC
204 	struct {
205 		long rt;	/* roundtrip time */
206 		long master;	/* master's timestamp */
207 		long diff;	/* difference between midpoint and master's timestamp */
208 		long lat;	/* estimate of itc adjustment latency */
209 	} t[NUM_ROUNDS];
210 #endif
211 
212 	go[MASTER] = 1;
213 
214 	while (go[MASTER])
215 		rmb();
216 
217 	local_irq_save(flags);
218 	{
219 		for (i = 0; i < NUM_ROUNDS; i++) {
220 			delta = get_delta(&rt, &master_time_stamp);
221 			if (delta == 0)
222 				done = 1;	/* let's lock on to this... */
223 
224 			if (!done) {
225 				if (i > 0) {
226 					adjust_latency += -delta;
227 					adj = -delta + adjust_latency/4;
228 				} else
229 					adj = -delta;
230 
231 				tick_ops->add_tick(adj);
232 			}
233 #if DEBUG_TICK_SYNC
234 			t[i].rt = rt;
235 			t[i].master = master_time_stamp;
236 			t[i].diff = delta;
237 			t[i].lat = adjust_latency/4;
238 #endif
239 		}
240 	}
241 	local_irq_restore(flags);
242 
243 #if DEBUG_TICK_SYNC
244 	for (i = 0; i < NUM_ROUNDS; i++)
245 		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
246 		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
247 #endif
248 
249 	printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
250 	       "(last diff %ld cycles, maxerr %lu cycles)\n",
251 	       smp_processor_id(), delta, rt);
252 }
253 
254 static void smp_start_sync_tick_client(int cpu);
255 
256 static void smp_synchronize_one_tick(int cpu)
257 {
258 	unsigned long flags, i;
259 
260 	go[MASTER] = 0;
261 
262 	smp_start_sync_tick_client(cpu);
263 
264 	/* wait for client to be ready */
265 	while (!go[MASTER])
266 		rmb();
267 
268 	/* now let the client proceed into his loop */
269 	go[MASTER] = 0;
270 	membar_safe("#StoreLoad");
271 
272 	raw_spin_lock_irqsave(&itc_sync_lock, flags);
273 	{
274 		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
275 			while (!go[MASTER])
276 				rmb();
277 			go[MASTER] = 0;
278 			wmb();
279 			go[SLAVE] = tick_ops->get_tick();
280 			membar_safe("#StoreLoad");
281 		}
282 	}
283 	raw_spin_unlock_irqrestore(&itc_sync_lock, flags);
284 }
285 
286 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
287 static void ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg,
288 				void **descrp)
289 {
290 	extern unsigned long sparc64_ttable_tl0;
291 	extern unsigned long kern_locked_tte_data;
292 	struct hvtramp_descr *hdesc;
293 	unsigned long trampoline_ra;
294 	struct trap_per_cpu *tb;
295 	u64 tte_vaddr, tte_data;
296 	unsigned long hv_err;
297 	int i;
298 
299 	hdesc = kzalloc(sizeof(*hdesc) +
300 			(sizeof(struct hvtramp_mapping) *
301 			 num_kernel_image_mappings - 1),
302 			GFP_KERNEL);
303 	if (!hdesc) {
304 		printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
305 		       "hvtramp_descr.\n");
306 		return;
307 	}
308 	*descrp = hdesc;
309 
310 	hdesc->cpu = cpu;
311 	hdesc->num_mappings = num_kernel_image_mappings;
312 
313 	tb = &trap_block[cpu];
314 
315 	hdesc->fault_info_va = (unsigned long) &tb->fault_info;
316 	hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
317 
318 	hdesc->thread_reg = thread_reg;
319 
320 	tte_vaddr = (unsigned long) KERNBASE;
321 	tte_data = kern_locked_tte_data;
322 
323 	for (i = 0; i < hdesc->num_mappings; i++) {
324 		hdesc->maps[i].vaddr = tte_vaddr;
325 		hdesc->maps[i].tte   = tte_data;
326 		tte_vaddr += 0x400000;
327 		tte_data  += 0x400000;
328 	}
329 
330 	trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
331 
332 	hv_err = sun4v_cpu_start(cpu, trampoline_ra,
333 				 kimage_addr_to_ra(&sparc64_ttable_tl0),
334 				 __pa(hdesc));
335 	if (hv_err)
336 		printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
337 		       "gives error %lu\n", hv_err);
338 }
339 #endif
340 
341 extern unsigned long sparc64_cpu_startup;
342 
343 /* The OBP cpu startup callback truncates the 3rd arg cookie to
344  * 32-bits (I think) so to be safe we have it read the pointer
345  * contained here so we work on >4GB machines. -DaveM
346  */
347 static struct thread_info *cpu_new_thread = NULL;
348 
349 static int smp_boot_one_cpu(unsigned int cpu, struct task_struct *idle)
350 {
351 	unsigned long entry =
352 		(unsigned long)(&sparc64_cpu_startup);
353 	unsigned long cookie =
354 		(unsigned long)(&cpu_new_thread);
355 	void *descr = NULL;
356 	int timeout, ret;
357 
358 	callin_flag = 0;
359 	cpu_new_thread = task_thread_info(idle);
360 
361 	if (tlb_type == hypervisor) {
362 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
363 		if (ldom_domaining_enabled)
364 			ldom_startcpu_cpuid(cpu,
365 					    (unsigned long) cpu_new_thread,
366 					    &descr);
367 		else
368 #endif
369 			prom_startcpu_cpuid(cpu, entry, cookie);
370 	} else {
371 		struct device_node *dp = of_find_node_by_cpuid(cpu);
372 
373 		prom_startcpu(dp->phandle, entry, cookie);
374 	}
375 
376 	for (timeout = 0; timeout < 50000; timeout++) {
377 		if (callin_flag)
378 			break;
379 		udelay(100);
380 	}
381 
382 	if (callin_flag) {
383 		ret = 0;
384 	} else {
385 		printk("Processor %d is stuck.\n", cpu);
386 		ret = -ENODEV;
387 	}
388 	cpu_new_thread = NULL;
389 
390 	kfree(descr);
391 
392 	return ret;
393 }
394 
395 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
396 {
397 	u64 result, target;
398 	int stuck, tmp;
399 
400 	if (this_is_starfire) {
401 		/* map to real upaid */
402 		cpu = (((cpu & 0x3c) << 1) |
403 			((cpu & 0x40) >> 4) |
404 			(cpu & 0x3));
405 	}
406 
407 	target = (cpu << 14) | 0x70;
408 again:
409 	/* Ok, this is the real Spitfire Errata #54.
410 	 * One must read back from a UDB internal register
411 	 * after writes to the UDB interrupt dispatch, but
412 	 * before the membar Sync for that write.
413 	 * So we use the high UDB control register (ASI 0x7f,
414 	 * ADDR 0x20) for the dummy read. -DaveM
415 	 */
416 	tmp = 0x40;
417 	__asm__ __volatile__(
418 	"wrpr	%1, %2, %%pstate\n\t"
419 	"stxa	%4, [%0] %3\n\t"
420 	"stxa	%5, [%0+%8] %3\n\t"
421 	"add	%0, %8, %0\n\t"
422 	"stxa	%6, [%0+%8] %3\n\t"
423 	"membar	#Sync\n\t"
424 	"stxa	%%g0, [%7] %3\n\t"
425 	"membar	#Sync\n\t"
426 	"mov	0x20, %%g1\n\t"
427 	"ldxa	[%%g1] 0x7f, %%g0\n\t"
428 	"membar	#Sync"
429 	: "=r" (tmp)
430 	: "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
431 	  "r" (data0), "r" (data1), "r" (data2), "r" (target),
432 	  "r" (0x10), "0" (tmp)
433         : "g1");
434 
435 	/* NOTE: PSTATE_IE is still clear. */
436 	stuck = 100000;
437 	do {
438 		__asm__ __volatile__("ldxa [%%g0] %1, %0"
439 			: "=r" (result)
440 			: "i" (ASI_INTR_DISPATCH_STAT));
441 		if (result == 0) {
442 			__asm__ __volatile__("wrpr %0, 0x0, %%pstate"
443 					     : : "r" (pstate));
444 			return;
445 		}
446 		stuck -= 1;
447 		if (stuck == 0)
448 			break;
449 	} while (result & 0x1);
450 	__asm__ __volatile__("wrpr %0, 0x0, %%pstate"
451 			     : : "r" (pstate));
452 	if (stuck == 0) {
453 		printk("CPU[%d]: mondo stuckage result[%016llx]\n",
454 		       smp_processor_id(), result);
455 	} else {
456 		udelay(2);
457 		goto again;
458 	}
459 }
460 
461 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
462 {
463 	u64 *mondo, data0, data1, data2;
464 	u16 *cpu_list;
465 	u64 pstate;
466 	int i;
467 
468 	__asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
469 	cpu_list = __va(tb->cpu_list_pa);
470 	mondo = __va(tb->cpu_mondo_block_pa);
471 	data0 = mondo[0];
472 	data1 = mondo[1];
473 	data2 = mondo[2];
474 	for (i = 0; i < cnt; i++)
475 		spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
476 }
477 
478 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
479  * packet, but we have no use for that.  However we do take advantage of
480  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
481  */
482 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
483 {
484 	int nack_busy_id, is_jbus, need_more;
485 	u64 *mondo, pstate, ver, busy_mask;
486 	u16 *cpu_list;
487 
488 	cpu_list = __va(tb->cpu_list_pa);
489 	mondo = __va(tb->cpu_mondo_block_pa);
490 
491 	/* Unfortunately, someone at Sun had the brilliant idea to make the
492 	 * busy/nack fields hard-coded by ITID number for this Ultra-III
493 	 * derivative processor.
494 	 */
495 	__asm__ ("rdpr %%ver, %0" : "=r" (ver));
496 	is_jbus = ((ver >> 32) == __JALAPENO_ID ||
497 		   (ver >> 32) == __SERRANO_ID);
498 
499 	__asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
500 
501 retry:
502 	need_more = 0;
503 	__asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
504 			     : : "r" (pstate), "i" (PSTATE_IE));
505 
506 	/* Setup the dispatch data registers. */
507 	__asm__ __volatile__("stxa	%0, [%3] %6\n\t"
508 			     "stxa	%1, [%4] %6\n\t"
509 			     "stxa	%2, [%5] %6\n\t"
510 			     "membar	#Sync\n\t"
511 			     : /* no outputs */
512 			     : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
513 			       "r" (0x40), "r" (0x50), "r" (0x60),
514 			       "i" (ASI_INTR_W));
515 
516 	nack_busy_id = 0;
517 	busy_mask = 0;
518 	{
519 		int i;
520 
521 		for (i = 0; i < cnt; i++) {
522 			u64 target, nr;
523 
524 			nr = cpu_list[i];
525 			if (nr == 0xffff)
526 				continue;
527 
528 			target = (nr << 14) | 0x70;
529 			if (is_jbus) {
530 				busy_mask |= (0x1UL << (nr * 2));
531 			} else {
532 				target |= (nack_busy_id << 24);
533 				busy_mask |= (0x1UL <<
534 					      (nack_busy_id * 2));
535 			}
536 			__asm__ __volatile__(
537 				"stxa	%%g0, [%0] %1\n\t"
538 				"membar	#Sync\n\t"
539 				: /* no outputs */
540 				: "r" (target), "i" (ASI_INTR_W));
541 			nack_busy_id++;
542 			if (nack_busy_id == 32) {
543 				need_more = 1;
544 				break;
545 			}
546 		}
547 	}
548 
549 	/* Now, poll for completion. */
550 	{
551 		u64 dispatch_stat, nack_mask;
552 		long stuck;
553 
554 		stuck = 100000 * nack_busy_id;
555 		nack_mask = busy_mask << 1;
556 		do {
557 			__asm__ __volatile__("ldxa	[%%g0] %1, %0"
558 					     : "=r" (dispatch_stat)
559 					     : "i" (ASI_INTR_DISPATCH_STAT));
560 			if (!(dispatch_stat & (busy_mask | nack_mask))) {
561 				__asm__ __volatile__("wrpr %0, 0x0, %%pstate"
562 						     : : "r" (pstate));
563 				if (unlikely(need_more)) {
564 					int i, this_cnt = 0;
565 					for (i = 0; i < cnt; i++) {
566 						if (cpu_list[i] == 0xffff)
567 							continue;
568 						cpu_list[i] = 0xffff;
569 						this_cnt++;
570 						if (this_cnt == 32)
571 							break;
572 					}
573 					goto retry;
574 				}
575 				return;
576 			}
577 			if (!--stuck)
578 				break;
579 		} while (dispatch_stat & busy_mask);
580 
581 		__asm__ __volatile__("wrpr %0, 0x0, %%pstate"
582 				     : : "r" (pstate));
583 
584 		if (dispatch_stat & busy_mask) {
585 			/* Busy bits will not clear, continue instead
586 			 * of freezing up on this cpu.
587 			 */
588 			printk("CPU[%d]: mondo stuckage result[%016llx]\n",
589 			       smp_processor_id(), dispatch_stat);
590 		} else {
591 			int i, this_busy_nack = 0;
592 
593 			/* Delay some random time with interrupts enabled
594 			 * to prevent deadlock.
595 			 */
596 			udelay(2 * nack_busy_id);
597 
598 			/* Clear out the mask bits for cpus which did not
599 			 * NACK us.
600 			 */
601 			for (i = 0; i < cnt; i++) {
602 				u64 check_mask, nr;
603 
604 				nr = cpu_list[i];
605 				if (nr == 0xffff)
606 					continue;
607 
608 				if (is_jbus)
609 					check_mask = (0x2UL << (2*nr));
610 				else
611 					check_mask = (0x2UL <<
612 						      this_busy_nack);
613 				if ((dispatch_stat & check_mask) == 0)
614 					cpu_list[i] = 0xffff;
615 				this_busy_nack += 2;
616 				if (this_busy_nack == 64)
617 					break;
618 			}
619 
620 			goto retry;
621 		}
622 	}
623 }
624 
625 #define	CPU_MONDO_COUNTER(cpuid)	(cpu_mondo_counter[cpuid])
626 #define	MONDO_USEC_WAIT_MIN		2
627 #define	MONDO_USEC_WAIT_MAX		100
628 #define	MONDO_RETRY_LIMIT		500000
629 
630 /* Multi-cpu list version.
631  *
632  * Deliver xcalls to 'cnt' number of cpus in 'cpu_list'.
633  * Sometimes not all cpus receive the mondo, requiring us to re-send
634  * the mondo until all cpus have received, or cpus are truly stuck
635  * unable to receive mondo, and we timeout.
636  * Occasionally a target cpu strand is borrowed briefly by hypervisor to
637  * perform guest service, such as PCIe error handling. Consider the
638  * service time, 1 second overall wait is reasonable for 1 cpu.
639  * Here two in-between mondo check wait time are defined: 2 usec for
640  * single cpu quick turn around and up to 100usec for large cpu count.
641  * Deliver mondo to large number of cpus could take longer, we adjusts
642  * the retry count as long as target cpus are making forward progress.
643  */
644 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
645 {
646 	int this_cpu, tot_cpus, prev_sent, i, rem;
647 	int usec_wait, retries, tot_retries;
648 	u16 first_cpu = 0xffff;
649 	unsigned long xc_rcvd = 0;
650 	unsigned long status;
651 	int ecpuerror_id = 0;
652 	int enocpu_id = 0;
653 	u16 *cpu_list;
654 	u16 cpu;
655 
656 	this_cpu = smp_processor_id();
657 	cpu_list = __va(tb->cpu_list_pa);
658 	usec_wait = cnt * MONDO_USEC_WAIT_MIN;
659 	if (usec_wait > MONDO_USEC_WAIT_MAX)
660 		usec_wait = MONDO_USEC_WAIT_MAX;
661 	retries = tot_retries = 0;
662 	tot_cpus = cnt;
663 	prev_sent = 0;
664 
665 	do {
666 		int n_sent, mondo_delivered, target_cpu_busy;
667 
668 		status = sun4v_cpu_mondo_send(cnt,
669 					      tb->cpu_list_pa,
670 					      tb->cpu_mondo_block_pa);
671 
672 		/* HV_EOK means all cpus received the xcall, we're done.  */
673 		if (likely(status == HV_EOK))
674 			goto xcall_done;
675 
676 		/* If not these non-fatal errors, panic */
677 		if (unlikely((status != HV_EWOULDBLOCK) &&
678 			(status != HV_ECPUERROR) &&
679 			(status != HV_ENOCPU)))
680 			goto fatal_errors;
681 
682 		/* First, see if we made any forward progress.
683 		 *
684 		 * Go through the cpu_list, count the target cpus that have
685 		 * received our mondo (n_sent), and those that did not (rem).
686 		 * Re-pack cpu_list with the cpus remain to be retried in the
687 		 * front - this simplifies tracking the truly stalled cpus.
688 		 *
689 		 * The hypervisor indicates successful sends by setting
690 		 * cpu list entries to the value 0xffff.
691 		 *
692 		 * EWOULDBLOCK means some target cpus did not receive the
693 		 * mondo and retry usually helps.
694 		 *
695 		 * ECPUERROR means at least one target cpu is in error state,
696 		 * it's usually safe to skip the faulty cpu and retry.
697 		 *
698 		 * ENOCPU means one of the target cpu doesn't belong to the
699 		 * domain, perhaps offlined which is unexpected, but not
700 		 * fatal and it's okay to skip the offlined cpu.
701 		 */
702 		rem = 0;
703 		n_sent = 0;
704 		for (i = 0; i < cnt; i++) {
705 			cpu = cpu_list[i];
706 			if (likely(cpu == 0xffff)) {
707 				n_sent++;
708 			} else if ((status == HV_ECPUERROR) &&
709 				(sun4v_cpu_state(cpu) == HV_CPU_STATE_ERROR)) {
710 				ecpuerror_id = cpu + 1;
711 			} else if (status == HV_ENOCPU && !cpu_online(cpu)) {
712 				enocpu_id = cpu + 1;
713 			} else {
714 				cpu_list[rem++] = cpu;
715 			}
716 		}
717 
718 		/* No cpu remained, we're done. */
719 		if (rem == 0)
720 			break;
721 
722 		/* Otherwise, update the cpu count for retry. */
723 		cnt = rem;
724 
725 		/* Record the overall number of mondos received by the
726 		 * first of the remaining cpus.
727 		 */
728 		if (first_cpu != cpu_list[0]) {
729 			first_cpu = cpu_list[0];
730 			xc_rcvd = CPU_MONDO_COUNTER(first_cpu);
731 		}
732 
733 		/* Was any mondo delivered successfully? */
734 		mondo_delivered = (n_sent > prev_sent);
735 		prev_sent = n_sent;
736 
737 		/* or, was any target cpu busy processing other mondos? */
738 		target_cpu_busy = (xc_rcvd < CPU_MONDO_COUNTER(first_cpu));
739 		xc_rcvd = CPU_MONDO_COUNTER(first_cpu);
740 
741 		/* Retry count is for no progress. If we're making progress,
742 		 * reset the retry count.
743 		 */
744 		if (likely(mondo_delivered || target_cpu_busy)) {
745 			tot_retries += retries;
746 			retries = 0;
747 		} else if (unlikely(retries > MONDO_RETRY_LIMIT)) {
748 			goto fatal_mondo_timeout;
749 		}
750 
751 		/* Delay a little bit to let other cpus catch up on
752 		 * their cpu mondo queue work.
753 		 */
754 		if (!mondo_delivered)
755 			udelay(usec_wait);
756 
757 		retries++;
758 	} while (1);
759 
760 xcall_done:
761 	if (unlikely(ecpuerror_id > 0)) {
762 		pr_crit("CPU[%d]: SUN4V mondo cpu error, target cpu(%d) was in error state\n",
763 		       this_cpu, ecpuerror_id - 1);
764 	} else if (unlikely(enocpu_id > 0)) {
765 		pr_crit("CPU[%d]: SUN4V mondo cpu error, target cpu(%d) does not belong to the domain\n",
766 		       this_cpu, enocpu_id - 1);
767 	}
768 	return;
769 
770 fatal_errors:
771 	/* fatal errors include bad alignment, etc */
772 	pr_crit("CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) mondo_block_pa(%lx)\n",
773 	       this_cpu, tot_cpus, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
774 	panic("Unexpected SUN4V mondo error %lu\n", status);
775 
776 fatal_mondo_timeout:
777 	/* some cpus being non-responsive to the cpu mondo */
778 	pr_crit("CPU[%d]: SUN4V mondo timeout, cpu(%d) made no forward progress after %d retries. Total target cpus(%d).\n",
779 	       this_cpu, first_cpu, (tot_retries + retries), tot_cpus);
780 	panic("SUN4V mondo timeout panic\n");
781 }
782 
783 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
784 
785 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
786 {
787 	struct trap_per_cpu *tb;
788 	int this_cpu, i, cnt;
789 	unsigned long flags;
790 	u16 *cpu_list;
791 	u64 *mondo;
792 
793 	/* We have to do this whole thing with interrupts fully disabled.
794 	 * Otherwise if we send an xcall from interrupt context it will
795 	 * corrupt both our mondo block and cpu list state.
796 	 *
797 	 * One consequence of this is that we cannot use timeout mechanisms
798 	 * that depend upon interrupts being delivered locally.  So, for
799 	 * example, we cannot sample jiffies and expect it to advance.
800 	 *
801 	 * Fortunately, udelay() uses %stick/%tick so we can use that.
802 	 */
803 	local_irq_save(flags);
804 
805 	this_cpu = smp_processor_id();
806 	tb = &trap_block[this_cpu];
807 
808 	mondo = __va(tb->cpu_mondo_block_pa);
809 	mondo[0] = data0;
810 	mondo[1] = data1;
811 	mondo[2] = data2;
812 	wmb();
813 
814 	cpu_list = __va(tb->cpu_list_pa);
815 
816 	/* Setup the initial cpu list.  */
817 	cnt = 0;
818 	for_each_cpu(i, mask) {
819 		if (i == this_cpu || !cpu_online(i))
820 			continue;
821 		cpu_list[cnt++] = i;
822 	}
823 
824 	if (cnt)
825 		xcall_deliver_impl(tb, cnt);
826 
827 	local_irq_restore(flags);
828 }
829 
830 /* Send cross call to all processors mentioned in MASK_P
831  * except self.  Really, there are only two cases currently,
832  * "cpu_online_mask" and "mm_cpumask(mm)".
833  */
834 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
835 {
836 	u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
837 
838 	xcall_deliver(data0, data1, data2, mask);
839 }
840 
841 /* Send cross call to all processors except self. */
842 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
843 {
844 	smp_cross_call_masked(func, ctx, data1, data2, cpu_online_mask);
845 }
846 
847 extern unsigned long xcall_sync_tick;
848 
849 static void smp_start_sync_tick_client(int cpu)
850 {
851 	xcall_deliver((u64) &xcall_sync_tick, 0, 0,
852 		      cpumask_of(cpu));
853 }
854 
855 extern unsigned long xcall_call_function;
856 
857 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
858 {
859 	xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
860 }
861 
862 extern unsigned long xcall_call_function_single;
863 
864 void arch_send_call_function_single_ipi(int cpu)
865 {
866 	xcall_deliver((u64) &xcall_call_function_single, 0, 0,
867 		      cpumask_of(cpu));
868 }
869 
870 void __irq_entry smp_call_function_client(int irq, struct pt_regs *regs)
871 {
872 	clear_softint(1 << irq);
873 	irq_enter();
874 	generic_smp_call_function_interrupt();
875 	irq_exit();
876 }
877 
878 void __irq_entry smp_call_function_single_client(int irq, struct pt_regs *regs)
879 {
880 	clear_softint(1 << irq);
881 	irq_enter();
882 	generic_smp_call_function_single_interrupt();
883 	irq_exit();
884 }
885 
886 static void tsb_sync(void *info)
887 {
888 	struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
889 	struct mm_struct *mm = info;
890 
891 	/* It is not valid to test "current->active_mm == mm" here.
892 	 *
893 	 * The value of "current" is not changed atomically with
894 	 * switch_mm().  But that's OK, we just need to check the
895 	 * current cpu's trap block PGD physical address.
896 	 */
897 	if (tp->pgd_paddr == __pa(mm->pgd))
898 		tsb_context_switch(mm);
899 }
900 
901 void smp_tsb_sync(struct mm_struct *mm)
902 {
903 	smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
904 }
905 
906 extern unsigned long xcall_flush_tlb_mm;
907 extern unsigned long xcall_flush_tlb_page;
908 extern unsigned long xcall_flush_tlb_kernel_range;
909 extern unsigned long xcall_fetch_glob_regs;
910 extern unsigned long xcall_fetch_glob_pmu;
911 extern unsigned long xcall_fetch_glob_pmu_n4;
912 extern unsigned long xcall_receive_signal;
913 extern unsigned long xcall_new_mmu_context_version;
914 #ifdef CONFIG_KGDB
915 extern unsigned long xcall_kgdb_capture;
916 #endif
917 
918 #ifdef DCACHE_ALIASING_POSSIBLE
919 extern unsigned long xcall_flush_dcache_page_cheetah;
920 #endif
921 extern unsigned long xcall_flush_dcache_page_spitfire;
922 
923 static inline void __local_flush_dcache_page(struct page *page)
924 {
925 #ifdef DCACHE_ALIASING_POSSIBLE
926 	__flush_dcache_page(page_address(page),
927 			    ((tlb_type == spitfire) &&
928 			     page_mapping(page) != NULL));
929 #else
930 	if (page_mapping(page) != NULL &&
931 	    tlb_type == spitfire)
932 		__flush_icache_page(__pa(page_address(page)));
933 #endif
934 }
935 
936 void smp_flush_dcache_page_impl(struct page *page, int cpu)
937 {
938 	int this_cpu;
939 
940 	if (tlb_type == hypervisor)
941 		return;
942 
943 #ifdef CONFIG_DEBUG_DCFLUSH
944 	atomic_inc(&dcpage_flushes);
945 #endif
946 
947 	this_cpu = get_cpu();
948 
949 	if (cpu == this_cpu) {
950 		__local_flush_dcache_page(page);
951 	} else if (cpu_online(cpu)) {
952 		void *pg_addr = page_address(page);
953 		u64 data0 = 0;
954 
955 		if (tlb_type == spitfire) {
956 			data0 = ((u64)&xcall_flush_dcache_page_spitfire);
957 			if (page_mapping(page) != NULL)
958 				data0 |= ((u64)1 << 32);
959 		} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
960 #ifdef DCACHE_ALIASING_POSSIBLE
961 			data0 =	((u64)&xcall_flush_dcache_page_cheetah);
962 #endif
963 		}
964 		if (data0) {
965 			xcall_deliver(data0, __pa(pg_addr),
966 				      (u64) pg_addr, cpumask_of(cpu));
967 #ifdef CONFIG_DEBUG_DCFLUSH
968 			atomic_inc(&dcpage_flushes_xcall);
969 #endif
970 		}
971 	}
972 
973 	put_cpu();
974 }
975 
976 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
977 {
978 	void *pg_addr;
979 	u64 data0;
980 
981 	if (tlb_type == hypervisor)
982 		return;
983 
984 	preempt_disable();
985 
986 #ifdef CONFIG_DEBUG_DCFLUSH
987 	atomic_inc(&dcpage_flushes);
988 #endif
989 	data0 = 0;
990 	pg_addr = page_address(page);
991 	if (tlb_type == spitfire) {
992 		data0 = ((u64)&xcall_flush_dcache_page_spitfire);
993 		if (page_mapping(page) != NULL)
994 			data0 |= ((u64)1 << 32);
995 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
996 #ifdef DCACHE_ALIASING_POSSIBLE
997 		data0 = ((u64)&xcall_flush_dcache_page_cheetah);
998 #endif
999 	}
1000 	if (data0) {
1001 		xcall_deliver(data0, __pa(pg_addr),
1002 			      (u64) pg_addr, cpu_online_mask);
1003 #ifdef CONFIG_DEBUG_DCFLUSH
1004 		atomic_inc(&dcpage_flushes_xcall);
1005 #endif
1006 	}
1007 	__local_flush_dcache_page(page);
1008 
1009 	preempt_enable();
1010 }
1011 
1012 #ifdef CONFIG_KGDB
1013 void kgdb_roundup_cpus(unsigned long flags)
1014 {
1015 	smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1016 }
1017 #endif
1018 
1019 void smp_fetch_global_regs(void)
1020 {
1021 	smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1022 }
1023 
1024 void smp_fetch_global_pmu(void)
1025 {
1026 	if (tlb_type == hypervisor &&
1027 	    sun4v_chip_type >= SUN4V_CHIP_NIAGARA4)
1028 		smp_cross_call(&xcall_fetch_glob_pmu_n4, 0, 0, 0);
1029 	else
1030 		smp_cross_call(&xcall_fetch_glob_pmu, 0, 0, 0);
1031 }
1032 
1033 /* We know that the window frames of the user have been flushed
1034  * to the stack before we get here because all callers of us
1035  * are flush_tlb_*() routines, and these run after flush_cache_*()
1036  * which performs the flushw.
1037  *
1038  * The SMP TLB coherency scheme we use works as follows:
1039  *
1040  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1041  *    space has (potentially) executed on, this is the heuristic
1042  *    we use to avoid doing cross calls.
1043  *
1044  *    Also, for flushing from kswapd and also for clones, we
1045  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1046  *
1047  * 2) TLB context numbers are shared globally across all processors
1048  *    in the system, this allows us to play several games to avoid
1049  *    cross calls.
1050  *
1051  *    One invariant is that when a cpu switches to a process, and
1052  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1053  *    current cpu's bit set, that tlb context is flushed locally.
1054  *
1055  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1056  *    cross calls when we want to flush the currently running process's
1057  *    tlb state.  This is done by clearing all cpu bits except the current
1058  *    processor's in current->mm->cpu_vm_mask and performing the
1059  *    flush locally only.  This will force any subsequent cpus which run
1060  *    this task to flush the context from the local tlb if the process
1061  *    migrates to another cpu (again).
1062  *
1063  * 3) For shared address spaces (threads) and swapping we bite the
1064  *    bullet for most cases and perform the cross call (but only to
1065  *    the cpus listed in cpu_vm_mask).
1066  *
1067  *    The performance gain from "optimizing" away the cross call for threads is
1068  *    questionable (in theory the big win for threads is the massive sharing of
1069  *    address space state across processors).
1070  */
1071 
1072 /* This currently is only used by the hugetlb arch pre-fault
1073  * hook on UltraSPARC-III+ and later when changing the pagesize
1074  * bits of the context register for an address space.
1075  */
1076 void smp_flush_tlb_mm(struct mm_struct *mm)
1077 {
1078 	u32 ctx = CTX_HWBITS(mm->context);
1079 	int cpu = get_cpu();
1080 
1081 	if (atomic_read(&mm->mm_users) == 1) {
1082 		cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1083 		goto local_flush_and_out;
1084 	}
1085 
1086 	smp_cross_call_masked(&xcall_flush_tlb_mm,
1087 			      ctx, 0, 0,
1088 			      mm_cpumask(mm));
1089 
1090 local_flush_and_out:
1091 	__flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1092 
1093 	put_cpu();
1094 }
1095 
1096 struct tlb_pending_info {
1097 	unsigned long ctx;
1098 	unsigned long nr;
1099 	unsigned long *vaddrs;
1100 };
1101 
1102 static void tlb_pending_func(void *info)
1103 {
1104 	struct tlb_pending_info *t = info;
1105 
1106 	__flush_tlb_pending(t->ctx, t->nr, t->vaddrs);
1107 }
1108 
1109 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1110 {
1111 	u32 ctx = CTX_HWBITS(mm->context);
1112 	struct tlb_pending_info info;
1113 	int cpu = get_cpu();
1114 
1115 	info.ctx = ctx;
1116 	info.nr = nr;
1117 	info.vaddrs = vaddrs;
1118 
1119 	if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1120 		cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1121 	else
1122 		smp_call_function_many(mm_cpumask(mm), tlb_pending_func,
1123 				       &info, 1);
1124 
1125 	__flush_tlb_pending(ctx, nr, vaddrs);
1126 
1127 	put_cpu();
1128 }
1129 
1130 void smp_flush_tlb_page(struct mm_struct *mm, unsigned long vaddr)
1131 {
1132 	unsigned long context = CTX_HWBITS(mm->context);
1133 	int cpu = get_cpu();
1134 
1135 	if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1136 		cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1137 	else
1138 		smp_cross_call_masked(&xcall_flush_tlb_page,
1139 				      context, vaddr, 0,
1140 				      mm_cpumask(mm));
1141 	__flush_tlb_page(context, vaddr);
1142 
1143 	put_cpu();
1144 }
1145 
1146 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1147 {
1148 	start &= PAGE_MASK;
1149 	end    = PAGE_ALIGN(end);
1150 	if (start != end) {
1151 		smp_cross_call(&xcall_flush_tlb_kernel_range,
1152 			       0, start, end);
1153 
1154 		__flush_tlb_kernel_range(start, end);
1155 	}
1156 }
1157 
1158 /* CPU capture. */
1159 /* #define CAPTURE_DEBUG */
1160 extern unsigned long xcall_capture;
1161 
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;
1165 
1166 void smp_capture(void)
1167 {
1168 	int result = atomic_add_return(1, &smp_capture_depth);
1169 
1170 	if (result == 1) {
1171 		int ncpus = num_online_cpus();
1172 
1173 #ifdef CAPTURE_DEBUG
1174 		printk("CPU[%d]: Sending penguins to jail...",
1175 		       smp_processor_id());
1176 #endif
1177 		penguins_are_doing_time = 1;
1178 		atomic_inc(&smp_capture_registry);
1179 		smp_cross_call(&xcall_capture, 0, 0, 0);
1180 		while (atomic_read(&smp_capture_registry) != ncpus)
1181 			rmb();
1182 #ifdef CAPTURE_DEBUG
1183 		printk("done\n");
1184 #endif
1185 	}
1186 }
1187 
1188 void smp_release(void)
1189 {
1190 	if (atomic_dec_and_test(&smp_capture_depth)) {
1191 #ifdef CAPTURE_DEBUG
1192 		printk("CPU[%d]: Giving pardon to "
1193 		       "imprisoned penguins\n",
1194 		       smp_processor_id());
1195 #endif
1196 		penguins_are_doing_time = 0;
1197 		membar_safe("#StoreLoad");
1198 		atomic_dec(&smp_capture_registry);
1199 	}
1200 }
1201 
1202 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1203  * set, so they can service tlb flush xcalls...
1204  */
1205 extern void prom_world(int);
1206 
1207 void __irq_entry smp_penguin_jailcell(int irq, struct pt_regs *regs)
1208 {
1209 	clear_softint(1 << irq);
1210 
1211 	preempt_disable();
1212 
1213 	__asm__ __volatile__("flushw");
1214 	prom_world(1);
1215 	atomic_inc(&smp_capture_registry);
1216 	membar_safe("#StoreLoad");
1217 	while (penguins_are_doing_time)
1218 		rmb();
1219 	atomic_dec(&smp_capture_registry);
1220 	prom_world(0);
1221 
1222 	preempt_enable();
1223 }
1224 
1225 /* /proc/profile writes can call this, don't __init it please. */
1226 int setup_profiling_timer(unsigned int multiplier)
1227 {
1228 	return -EINVAL;
1229 }
1230 
1231 void __init smp_prepare_cpus(unsigned int max_cpus)
1232 {
1233 }
1234 
1235 void smp_prepare_boot_cpu(void)
1236 {
1237 }
1238 
1239 void __init smp_setup_processor_id(void)
1240 {
1241 	if (tlb_type == spitfire)
1242 		xcall_deliver_impl = spitfire_xcall_deliver;
1243 	else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1244 		xcall_deliver_impl = cheetah_xcall_deliver;
1245 	else
1246 		xcall_deliver_impl = hypervisor_xcall_deliver;
1247 }
1248 
1249 void __init smp_fill_in_cpu_possible_map(void)
1250 {
1251 	int possible_cpus = num_possible_cpus();
1252 	int i;
1253 
1254 	if (possible_cpus > nr_cpu_ids)
1255 		possible_cpus = nr_cpu_ids;
1256 
1257 	for (i = 0; i < possible_cpus; i++)
1258 		set_cpu_possible(i, true);
1259 	for (; i < NR_CPUS; i++)
1260 		set_cpu_possible(i, false);
1261 }
1262 
1263 void smp_fill_in_sib_core_maps(void)
1264 {
1265 	unsigned int i;
1266 
1267 	for_each_present_cpu(i) {
1268 		unsigned int j;
1269 
1270 		cpumask_clear(&cpu_core_map[i]);
1271 		if (cpu_data(i).core_id == 0) {
1272 			cpumask_set_cpu(i, &cpu_core_map[i]);
1273 			continue;
1274 		}
1275 
1276 		for_each_present_cpu(j) {
1277 			if (cpu_data(i).core_id ==
1278 			    cpu_data(j).core_id)
1279 				cpumask_set_cpu(j, &cpu_core_map[i]);
1280 		}
1281 	}
1282 
1283 	for_each_present_cpu(i)  {
1284 		unsigned int j;
1285 
1286 		for_each_present_cpu(j)  {
1287 			if (cpu_data(i).max_cache_id ==
1288 			    cpu_data(j).max_cache_id)
1289 				cpumask_set_cpu(j, &cpu_core_sib_cache_map[i]);
1290 
1291 			if (cpu_data(i).sock_id == cpu_data(j).sock_id)
1292 				cpumask_set_cpu(j, &cpu_core_sib_map[i]);
1293 		}
1294 	}
1295 
1296 	for_each_present_cpu(i) {
1297 		unsigned int j;
1298 
1299 		cpumask_clear(&per_cpu(cpu_sibling_map, i));
1300 		if (cpu_data(i).proc_id == -1) {
1301 			cpumask_set_cpu(i, &per_cpu(cpu_sibling_map, i));
1302 			continue;
1303 		}
1304 
1305 		for_each_present_cpu(j) {
1306 			if (cpu_data(i).proc_id ==
1307 			    cpu_data(j).proc_id)
1308 				cpumask_set_cpu(j, &per_cpu(cpu_sibling_map, i));
1309 		}
1310 	}
1311 }
1312 
1313 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1314 {
1315 	int ret = smp_boot_one_cpu(cpu, tidle);
1316 
1317 	if (!ret) {
1318 		cpumask_set_cpu(cpu, &smp_commenced_mask);
1319 		while (!cpu_online(cpu))
1320 			mb();
1321 		if (!cpu_online(cpu)) {
1322 			ret = -ENODEV;
1323 		} else {
1324 			/* On SUN4V, writes to %tick and %stick are
1325 			 * not allowed.
1326 			 */
1327 			if (tlb_type != hypervisor)
1328 				smp_synchronize_one_tick(cpu);
1329 		}
1330 	}
1331 	return ret;
1332 }
1333 
1334 #ifdef CONFIG_HOTPLUG_CPU
1335 void cpu_play_dead(void)
1336 {
1337 	int cpu = smp_processor_id();
1338 	unsigned long pstate;
1339 
1340 	idle_task_exit();
1341 
1342 	if (tlb_type == hypervisor) {
1343 		struct trap_per_cpu *tb = &trap_block[cpu];
1344 
1345 		sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1346 				tb->cpu_mondo_pa, 0);
1347 		sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1348 				tb->dev_mondo_pa, 0);
1349 		sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1350 				tb->resum_mondo_pa, 0);
1351 		sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1352 				tb->nonresum_mondo_pa, 0);
1353 	}
1354 
1355 	cpumask_clear_cpu(cpu, &smp_commenced_mask);
1356 	membar_safe("#Sync");
1357 
1358 	local_irq_disable();
1359 
1360 	__asm__ __volatile__(
1361 		"rdpr	%%pstate, %0\n\t"
1362 		"wrpr	%0, %1, %%pstate"
1363 		: "=r" (pstate)
1364 		: "i" (PSTATE_IE));
1365 
1366 	while (1)
1367 		barrier();
1368 }
1369 
1370 int __cpu_disable(void)
1371 {
1372 	int cpu = smp_processor_id();
1373 	cpuinfo_sparc *c;
1374 	int i;
1375 
1376 	for_each_cpu(i, &cpu_core_map[cpu])
1377 		cpumask_clear_cpu(cpu, &cpu_core_map[i]);
1378 	cpumask_clear(&cpu_core_map[cpu]);
1379 
1380 	for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
1381 		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
1382 	cpumask_clear(&per_cpu(cpu_sibling_map, cpu));
1383 
1384 	c = &cpu_data(cpu);
1385 
1386 	c->core_id = 0;
1387 	c->proc_id = -1;
1388 
1389 	smp_wmb();
1390 
1391 	/* Make sure no interrupts point to this cpu.  */
1392 	fixup_irqs();
1393 
1394 	local_irq_enable();
1395 	mdelay(1);
1396 	local_irq_disable();
1397 
1398 	set_cpu_online(cpu, false);
1399 
1400 	cpu_map_rebuild();
1401 
1402 	return 0;
1403 }
1404 
1405 void __cpu_die(unsigned int cpu)
1406 {
1407 	int i;
1408 
1409 	for (i = 0; i < 100; i++) {
1410 		smp_rmb();
1411 		if (!cpumask_test_cpu(cpu, &smp_commenced_mask))
1412 			break;
1413 		msleep(100);
1414 	}
1415 	if (cpumask_test_cpu(cpu, &smp_commenced_mask)) {
1416 		printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1417 	} else {
1418 #if defined(CONFIG_SUN_LDOMS)
1419 		unsigned long hv_err;
1420 		int limit = 100;
1421 
1422 		do {
1423 			hv_err = sun4v_cpu_stop(cpu);
1424 			if (hv_err == HV_EOK) {
1425 				set_cpu_present(cpu, false);
1426 				break;
1427 			}
1428 		} while (--limit > 0);
1429 		if (limit <= 0) {
1430 			printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1431 			       hv_err);
1432 		}
1433 #endif
1434 	}
1435 }
1436 #endif
1437 
1438 void __init smp_cpus_done(unsigned int max_cpus)
1439 {
1440 }
1441 
1442 void smp_send_reschedule(int cpu)
1443 {
1444 	if (cpu == smp_processor_id()) {
1445 		WARN_ON_ONCE(preemptible());
1446 		set_softint(1 << PIL_SMP_RECEIVE_SIGNAL);
1447 	} else {
1448 		xcall_deliver((u64) &xcall_receive_signal,
1449 			      0, 0, cpumask_of(cpu));
1450 	}
1451 }
1452 
1453 void __irq_entry smp_receive_signal_client(int irq, struct pt_regs *regs)
1454 {
1455 	clear_softint(1 << irq);
1456 	scheduler_ipi();
1457 }
1458 
1459 static void stop_this_cpu(void *dummy)
1460 {
1461 	set_cpu_online(smp_processor_id(), false);
1462 	prom_stopself();
1463 }
1464 
1465 void smp_send_stop(void)
1466 {
1467 	int cpu;
1468 
1469 	if (tlb_type == hypervisor) {
1470 		int this_cpu = smp_processor_id();
1471 #ifdef CONFIG_SERIAL_SUNHV
1472 		sunhv_migrate_hvcons_irq(this_cpu);
1473 #endif
1474 		for_each_online_cpu(cpu) {
1475 			if (cpu == this_cpu)
1476 				continue;
1477 
1478 			set_cpu_online(cpu, false);
1479 #ifdef CONFIG_SUN_LDOMS
1480 			if (ldom_domaining_enabled) {
1481 				unsigned long hv_err;
1482 				hv_err = sun4v_cpu_stop(cpu);
1483 				if (hv_err)
1484 					printk(KERN_ERR "sun4v_cpu_stop() "
1485 					       "failed err=%lu\n", hv_err);
1486 			} else
1487 #endif
1488 				prom_stopcpu_cpuid(cpu);
1489 		}
1490 	} else
1491 		smp_call_function(stop_this_cpu, NULL, 0);
1492 }
1493 
1494 /**
1495  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1496  * @cpu: cpu to allocate for
1497  * @size: size allocation in bytes
1498  * @align: alignment
1499  *
1500  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1501  * does the right thing for NUMA regardless of the current
1502  * configuration.
1503  *
1504  * RETURNS:
1505  * Pointer to the allocated area on success, NULL on failure.
1506  */
1507 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
1508 					size_t align)
1509 {
1510 	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1511 #ifdef CONFIG_NEED_MULTIPLE_NODES
1512 	int node = cpu_to_node(cpu);
1513 	void *ptr;
1514 
1515 	if (!node_online(node) || !NODE_DATA(node)) {
1516 		ptr = __alloc_bootmem(size, align, goal);
1517 		pr_info("cpu %d has no node %d or node-local memory\n",
1518 			cpu, node);
1519 		pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1520 			 cpu, size, __pa(ptr));
1521 	} else {
1522 		ptr = __alloc_bootmem_node(NODE_DATA(node),
1523 					   size, align, goal);
1524 		pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1525 			 "%016lx\n", cpu, size, node, __pa(ptr));
1526 	}
1527 	return ptr;
1528 #else
1529 	return __alloc_bootmem(size, align, goal);
1530 #endif
1531 }
1532 
1533 static void __init pcpu_free_bootmem(void *ptr, size_t size)
1534 {
1535 	free_bootmem(__pa(ptr), size);
1536 }
1537 
1538 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
1539 {
1540 	if (cpu_to_node(from) == cpu_to_node(to))
1541 		return LOCAL_DISTANCE;
1542 	else
1543 		return REMOTE_DISTANCE;
1544 }
1545 
1546 static void __init pcpu_populate_pte(unsigned long addr)
1547 {
1548 	pgd_t *pgd = pgd_offset_k(addr);
1549 	pud_t *pud;
1550 	pmd_t *pmd;
1551 
1552 	if (pgd_none(*pgd)) {
1553 		pud_t *new;
1554 
1555 		new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1556 		pgd_populate(&init_mm, pgd, new);
1557 	}
1558 
1559 	pud = pud_offset(pgd, addr);
1560 	if (pud_none(*pud)) {
1561 		pmd_t *new;
1562 
1563 		new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1564 		pud_populate(&init_mm, pud, new);
1565 	}
1566 
1567 	pmd = pmd_offset(pud, addr);
1568 	if (!pmd_present(*pmd)) {
1569 		pte_t *new;
1570 
1571 		new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1572 		pmd_populate_kernel(&init_mm, pmd, new);
1573 	}
1574 }
1575 
1576 void __init setup_per_cpu_areas(void)
1577 {
1578 	unsigned long delta;
1579 	unsigned int cpu;
1580 	int rc = -EINVAL;
1581 
1582 	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
1583 		rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1584 					    PERCPU_DYNAMIC_RESERVE, 4 << 20,
1585 					    pcpu_cpu_distance,
1586 					    pcpu_alloc_bootmem,
1587 					    pcpu_free_bootmem);
1588 		if (rc)
1589 			pr_warning("PERCPU: %s allocator failed (%d), "
1590 				   "falling back to page size\n",
1591 				   pcpu_fc_names[pcpu_chosen_fc], rc);
1592 	}
1593 	if (rc < 0)
1594 		rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
1595 					   pcpu_alloc_bootmem,
1596 					   pcpu_free_bootmem,
1597 					   pcpu_populate_pte);
1598 	if (rc < 0)
1599 		panic("cannot initialize percpu area (err=%d)", rc);
1600 
1601 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1602 	for_each_possible_cpu(cpu)
1603 		__per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
1604 
1605 	/* Setup %g5 for the boot cpu.  */
1606 	__local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1607 
1608 	of_fill_in_cpu_data();
1609 	if (tlb_type == hypervisor)
1610 		mdesc_fill_in_cpu_data(cpu_all_mask);
1611 }
1612