xref: /linux/arch/powerpc/platforms/powermac/smp.c (revision 858259cf7d1c443c836a2022b78cb281f0a9b95e)
1 /*
2  * SMP support for power macintosh.
3  *
4  * We support both the old "powersurge" SMP architecture
5  * and the current Core99 (G4 PowerMac) machines.
6  *
7  * Note that we don't support the very first rev. of
8  * Apple/DayStar 2 CPUs board, the one with the funky
9  * watchdog. Hopefully, none of these should be there except
10  * maybe internally to Apple. I should probably still add some
11  * code to detect this card though and disable SMP. --BenH.
12  *
13  * Support Macintosh G4 SMP by Troy Benjegerdes (hozer@drgw.net)
14  * and Ben Herrenschmidt <benh@kernel.crashing.org>.
15  *
16  * Support for DayStar quad CPU cards
17  * Copyright (C) XLR8, Inc. 1994-2000
18  *
19  *  This program is free software; you can redistribute it and/or
20  *  modify it under the terms of the GNU General Public License
21  *  as published by the Free Software Foundation; either version
22  *  2 of the License, or (at your option) any later version.
23  */
24 #include <linux/config.h>
25 #include <linux/kernel.h>
26 #include <linux/sched.h>
27 #include <linux/smp.h>
28 #include <linux/smp_lock.h>
29 #include <linux/interrupt.h>
30 #include <linux/kernel_stat.h>
31 #include <linux/delay.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/errno.h>
35 #include <linux/hardirq.h>
36 #include <linux/cpu.h>
37 
38 #include <asm/ptrace.h>
39 #include <asm/atomic.h>
40 #include <asm/irq.h>
41 #include <asm/page.h>
42 #include <asm/pgtable.h>
43 #include <asm/sections.h>
44 #include <asm/io.h>
45 #include <asm/prom.h>
46 #include <asm/smp.h>
47 #include <asm/machdep.h>
48 #include <asm/pmac_feature.h>
49 #include <asm/time.h>
50 #include <asm/mpic.h>
51 #include <asm/cacheflush.h>
52 #include <asm/keylargo.h>
53 #include <asm/pmac_low_i2c.h>
54 
55 #undef DEBUG
56 
57 #ifdef DEBUG
58 #define DBG(fmt...) udbg_printf(fmt)
59 #else
60 #define DBG(fmt...)
61 #endif
62 
63 extern void __secondary_start_pmac_0(void);
64 
65 #ifdef CONFIG_PPC32
66 
67 /* Sync flag for HW tb sync */
68 static volatile int sec_tb_reset = 0;
69 
70 /*
71  * Powersurge (old powermac SMP) support.
72  */
73 
74 /* Addresses for powersurge registers */
75 #define HAMMERHEAD_BASE		0xf8000000
76 #define HHEAD_CONFIG		0x90
77 #define HHEAD_SEC_INTR		0xc0
78 
79 /* register for interrupting the primary processor on the powersurge */
80 /* N.B. this is actually the ethernet ROM! */
81 #define PSURGE_PRI_INTR		0xf3019000
82 
83 /* register for storing the start address for the secondary processor */
84 /* N.B. this is the PCI config space address register for the 1st bridge */
85 #define PSURGE_START		0xf2800000
86 
87 /* Daystar/XLR8 4-CPU card */
88 #define PSURGE_QUAD_REG_ADDR	0xf8800000
89 
90 #define PSURGE_QUAD_IRQ_SET	0
91 #define PSURGE_QUAD_IRQ_CLR	1
92 #define PSURGE_QUAD_IRQ_PRIMARY	2
93 #define PSURGE_QUAD_CKSTOP_CTL	3
94 #define PSURGE_QUAD_PRIMARY_ARB	4
95 #define PSURGE_QUAD_BOARD_ID	6
96 #define PSURGE_QUAD_WHICH_CPU	7
97 #define PSURGE_QUAD_CKSTOP_RDBK	8
98 #define PSURGE_QUAD_RESET_CTL	11
99 
100 #define PSURGE_QUAD_OUT(r, v)	(out_8(quad_base + ((r) << 4) + 4, (v)))
101 #define PSURGE_QUAD_IN(r)	(in_8(quad_base + ((r) << 4) + 4) & 0x0f)
102 #define PSURGE_QUAD_BIS(r, v)	(PSURGE_QUAD_OUT((r), PSURGE_QUAD_IN(r) | (v)))
103 #define PSURGE_QUAD_BIC(r, v)	(PSURGE_QUAD_OUT((r), PSURGE_QUAD_IN(r) & ~(v)))
104 
105 /* virtual addresses for the above */
106 static volatile u8 __iomem *hhead_base;
107 static volatile u8 __iomem *quad_base;
108 static volatile u32 __iomem *psurge_pri_intr;
109 static volatile u8 __iomem *psurge_sec_intr;
110 static volatile u32 __iomem *psurge_start;
111 
112 /* values for psurge_type */
113 #define PSURGE_NONE		-1
114 #define PSURGE_DUAL		0
115 #define PSURGE_QUAD_OKEE	1
116 #define PSURGE_QUAD_COTTON	2
117 #define PSURGE_QUAD_ICEGRASS	3
118 
119 /* what sort of powersurge board we have */
120 static int psurge_type = PSURGE_NONE;
121 
122 /*
123  * Set and clear IPIs for powersurge.
124  */
125 static inline void psurge_set_ipi(int cpu)
126 {
127 	if (psurge_type == PSURGE_NONE)
128 		return;
129 	if (cpu == 0)
130 		in_be32(psurge_pri_intr);
131 	else if (psurge_type == PSURGE_DUAL)
132 		out_8(psurge_sec_intr, 0);
133 	else
134 		PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_SET, 1 << cpu);
135 }
136 
137 static inline void psurge_clr_ipi(int cpu)
138 {
139 	if (cpu > 0) {
140 		switch(psurge_type) {
141 		case PSURGE_DUAL:
142 			out_8(psurge_sec_intr, ~0);
143 		case PSURGE_NONE:
144 			break;
145 		default:
146 			PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_CLR, 1 << cpu);
147 		}
148 	}
149 }
150 
151 /*
152  * On powersurge (old SMP powermac architecture) we don't have
153  * separate IPIs for separate messages like openpic does.  Instead
154  * we have a bitmap for each processor, where a 1 bit means that
155  * the corresponding message is pending for that processor.
156  * Ideally each cpu's entry would be in a different cache line.
157  *  -- paulus.
158  */
159 static unsigned long psurge_smp_message[NR_CPUS];
160 
161 void psurge_smp_message_recv(struct pt_regs *regs)
162 {
163 	int cpu = smp_processor_id();
164 	int msg;
165 
166 	/* clear interrupt */
167 	psurge_clr_ipi(cpu);
168 
169 	if (num_online_cpus() < 2)
170 		return;
171 
172 	/* make sure there is a message there */
173 	for (msg = 0; msg < 4; msg++)
174 		if (test_and_clear_bit(msg, &psurge_smp_message[cpu]))
175 			smp_message_recv(msg, regs);
176 }
177 
178 irqreturn_t psurge_primary_intr(int irq, void *d, struct pt_regs *regs)
179 {
180 	psurge_smp_message_recv(regs);
181 	return IRQ_HANDLED;
182 }
183 
184 static void smp_psurge_message_pass(int target, int msg)
185 {
186 	int i;
187 
188 	if (num_online_cpus() < 2)
189 		return;
190 
191 	for (i = 0; i < NR_CPUS; i++) {
192 		if (!cpu_online(i))
193 			continue;
194 		if (target == MSG_ALL
195 		    || (target == MSG_ALL_BUT_SELF && i != smp_processor_id())
196 		    || target == i) {
197 			set_bit(msg, &psurge_smp_message[i]);
198 			psurge_set_ipi(i);
199 		}
200 	}
201 }
202 
203 /*
204  * Determine a quad card presence. We read the board ID register, we
205  * force the data bus to change to something else, and we read it again.
206  * It it's stable, then the register probably exist (ugh !)
207  */
208 static int __init psurge_quad_probe(void)
209 {
210 	int type;
211 	unsigned int i;
212 
213 	type = PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID);
214 	if (type < PSURGE_QUAD_OKEE || type > PSURGE_QUAD_ICEGRASS
215 	    || type != PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID))
216 		return PSURGE_DUAL;
217 
218 	/* looks OK, try a slightly more rigorous test */
219 	/* bogus is not necessarily cacheline-aligned,
220 	   though I don't suppose that really matters.  -- paulus */
221 	for (i = 0; i < 100; i++) {
222 		volatile u32 bogus[8];
223 		bogus[(0+i)%8] = 0x00000000;
224 		bogus[(1+i)%8] = 0x55555555;
225 		bogus[(2+i)%8] = 0xFFFFFFFF;
226 		bogus[(3+i)%8] = 0xAAAAAAAA;
227 		bogus[(4+i)%8] = 0x33333333;
228 		bogus[(5+i)%8] = 0xCCCCCCCC;
229 		bogus[(6+i)%8] = 0xCCCCCCCC;
230 		bogus[(7+i)%8] = 0x33333333;
231 		wmb();
232 		asm volatile("dcbf 0,%0" : : "r" (bogus) : "memory");
233 		mb();
234 		if (type != PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID))
235 			return PSURGE_DUAL;
236 	}
237 	return type;
238 }
239 
240 static void __init psurge_quad_init(void)
241 {
242 	int procbits;
243 
244 	if (ppc_md.progress) ppc_md.progress("psurge_quad_init", 0x351);
245 	procbits = ~PSURGE_QUAD_IN(PSURGE_QUAD_WHICH_CPU);
246 	if (psurge_type == PSURGE_QUAD_ICEGRASS)
247 		PSURGE_QUAD_BIS(PSURGE_QUAD_RESET_CTL, procbits);
248 	else
249 		PSURGE_QUAD_BIC(PSURGE_QUAD_CKSTOP_CTL, procbits);
250 	mdelay(33);
251 	out_8(psurge_sec_intr, ~0);
252 	PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_CLR, procbits);
253 	PSURGE_QUAD_BIS(PSURGE_QUAD_RESET_CTL, procbits);
254 	if (psurge_type != PSURGE_QUAD_ICEGRASS)
255 		PSURGE_QUAD_BIS(PSURGE_QUAD_CKSTOP_CTL, procbits);
256 	PSURGE_QUAD_BIC(PSURGE_QUAD_PRIMARY_ARB, procbits);
257 	mdelay(33);
258 	PSURGE_QUAD_BIC(PSURGE_QUAD_RESET_CTL, procbits);
259 	mdelay(33);
260 	PSURGE_QUAD_BIS(PSURGE_QUAD_PRIMARY_ARB, procbits);
261 	mdelay(33);
262 }
263 
264 static int __init smp_psurge_probe(void)
265 {
266 	int i, ncpus;
267 
268 	/* We don't do SMP on the PPC601 -- paulus */
269 	if (PVR_VER(mfspr(SPRN_PVR)) == 1)
270 		return 1;
271 
272 	/*
273 	 * The powersurge cpu board can be used in the generation
274 	 * of powermacs that have a socket for an upgradeable cpu card,
275 	 * including the 7500, 8500, 9500, 9600.
276 	 * The device tree doesn't tell you if you have 2 cpus because
277 	 * OF doesn't know anything about the 2nd processor.
278 	 * Instead we look for magic bits in magic registers,
279 	 * in the hammerhead memory controller in the case of the
280 	 * dual-cpu powersurge board.  -- paulus.
281 	 */
282 	if (find_devices("hammerhead") == NULL)
283 		return 1;
284 
285 	hhead_base = ioremap(HAMMERHEAD_BASE, 0x800);
286 	quad_base = ioremap(PSURGE_QUAD_REG_ADDR, 1024);
287 	psurge_sec_intr = hhead_base + HHEAD_SEC_INTR;
288 
289 	psurge_type = psurge_quad_probe();
290 	if (psurge_type != PSURGE_DUAL) {
291 		psurge_quad_init();
292 		/* All released cards using this HW design have 4 CPUs */
293 		ncpus = 4;
294 	} else {
295 		iounmap(quad_base);
296 		if ((in_8(hhead_base + HHEAD_CONFIG) & 0x02) == 0) {
297 			/* not a dual-cpu card */
298 			iounmap(hhead_base);
299 			psurge_type = PSURGE_NONE;
300 			return 1;
301 		}
302 		ncpus = 2;
303 	}
304 
305 	psurge_start = ioremap(PSURGE_START, 4);
306 	psurge_pri_intr = ioremap(PSURGE_PRI_INTR, 4);
307 
308 	/* this is not actually strictly necessary -- paulus. */
309 	for (i = 1; i < ncpus; ++i)
310 		smp_hw_index[i] = i;
311 
312 	if (ppc_md.progress) ppc_md.progress("smp_psurge_probe - done", 0x352);
313 
314 	return ncpus;
315 }
316 
317 static void __init smp_psurge_kick_cpu(int nr)
318 {
319 	unsigned long start = __pa(__secondary_start_pmac_0) + nr * 8;
320 	unsigned long a;
321 
322 	/* may need to flush here if secondary bats aren't setup */
323 	for (a = KERNELBASE; a < KERNELBASE + 0x800000; a += 32)
324 		asm volatile("dcbf 0,%0" : : "r" (a) : "memory");
325 	asm volatile("sync");
326 
327 	if (ppc_md.progress) ppc_md.progress("smp_psurge_kick_cpu", 0x353);
328 
329 	out_be32(psurge_start, start);
330 	mb();
331 
332 	psurge_set_ipi(nr);
333 	udelay(10);
334 	psurge_clr_ipi(nr);
335 
336 	if (ppc_md.progress) ppc_md.progress("smp_psurge_kick_cpu - done", 0x354);
337 }
338 
339 /*
340  * With the dual-cpu powersurge board, the decrementers and timebases
341  * of both cpus are frozen after the secondary cpu is started up,
342  * until we give the secondary cpu another interrupt.  This routine
343  * uses this to get the timebases synchronized.
344  *  -- paulus.
345  */
346 static void __init psurge_dual_sync_tb(int cpu_nr)
347 {
348 	int t;
349 
350 	set_dec(tb_ticks_per_jiffy);
351 	set_tb(0, 0);
352 	last_jiffy_stamp(cpu_nr) = 0;
353 
354 	if (cpu_nr > 0) {
355 		mb();
356 		sec_tb_reset = 1;
357 		return;
358 	}
359 
360 	/* wait for the secondary to have reset its TB before proceeding */
361 	for (t = 10000000; t > 0 && !sec_tb_reset; --t)
362 		;
363 
364 	/* now interrupt the secondary, starting both TBs */
365 	psurge_set_ipi(1);
366 
367 	smp_tb_synchronized = 1;
368 }
369 
370 static struct irqaction psurge_irqaction = {
371 	.handler = psurge_primary_intr,
372 	.flags = SA_INTERRUPT,
373 	.mask = CPU_MASK_NONE,
374 	.name = "primary IPI",
375 };
376 
377 static void __init smp_psurge_setup_cpu(int cpu_nr)
378 {
379 
380 	if (cpu_nr == 0) {
381 		/* If we failed to start the second CPU, we should still
382 		 * send it an IPI to start the timebase & DEC or we might
383 		 * have them stuck.
384 		 */
385 		if (num_online_cpus() < 2) {
386 			if (psurge_type == PSURGE_DUAL)
387 				psurge_set_ipi(1);
388 			return;
389 		}
390 		/* reset the entry point so if we get another intr we won't
391 		 * try to startup again */
392 		out_be32(psurge_start, 0x100);
393 		if (setup_irq(30, &psurge_irqaction))
394 			printk(KERN_ERR "Couldn't get primary IPI interrupt");
395 	}
396 
397 	if (psurge_type == PSURGE_DUAL)
398 		psurge_dual_sync_tb(cpu_nr);
399 }
400 
401 void __init smp_psurge_take_timebase(void)
402 {
403 	/* Dummy implementation */
404 }
405 
406 void __init smp_psurge_give_timebase(void)
407 {
408 	/* Dummy implementation */
409 }
410 
411 /* PowerSurge-style Macs */
412 struct smp_ops_t psurge_smp_ops = {
413 	.message_pass	= smp_psurge_message_pass,
414 	.probe		= smp_psurge_probe,
415 	.kick_cpu	= smp_psurge_kick_cpu,
416 	.setup_cpu	= smp_psurge_setup_cpu,
417 	.give_timebase	= smp_psurge_give_timebase,
418 	.take_timebase	= smp_psurge_take_timebase,
419 };
420 #endif /* CONFIG_PPC32 - actually powersurge support */
421 
422 #ifdef CONFIG_PPC64
423 /*
424  * G5s enable/disable the timebase via an i2c-connected clock chip.
425  */
426 static struct device_node *pmac_tb_clock_chip_host;
427 static u8 pmac_tb_pulsar_addr;
428 static void (*pmac_tb_freeze)(int freeze);
429 static DEFINE_SPINLOCK(timebase_lock);
430 static unsigned long timebase;
431 
432 static void smp_core99_cypress_tb_freeze(int freeze)
433 {
434 	u8 data;
435 	int rc;
436 
437 	/* Strangely, the device-tree says address is 0xd2, but darwin
438 	 * accesses 0xd0 ...
439 	 */
440 	pmac_low_i2c_setmode(pmac_tb_clock_chip_host, pmac_low_i2c_mode_combined);
441 	rc = pmac_low_i2c_xfer(pmac_tb_clock_chip_host,
442 			       0xd0 | pmac_low_i2c_read,
443 			       0x81, &data, 1);
444 	if (rc != 0)
445 		goto bail;
446 
447 	data = (data & 0xf3) | (freeze ? 0x00 : 0x0c);
448 
449        	pmac_low_i2c_setmode(pmac_tb_clock_chip_host, pmac_low_i2c_mode_stdsub);
450 	rc = pmac_low_i2c_xfer(pmac_tb_clock_chip_host,
451 			       0xd0 | pmac_low_i2c_write,
452 			       0x81, &data, 1);
453 
454  bail:
455 	if (rc != 0) {
456 		printk("Cypress Timebase %s rc: %d\n",
457 		       freeze ? "freeze" : "unfreeze", rc);
458 		panic("Timebase freeze failed !\n");
459 	}
460 }
461 
462 
463 static void smp_core99_pulsar_tb_freeze(int freeze)
464 {
465 	u8 data;
466 	int rc;
467 
468 	pmac_low_i2c_setmode(pmac_tb_clock_chip_host, pmac_low_i2c_mode_combined);
469 	rc = pmac_low_i2c_xfer(pmac_tb_clock_chip_host,
470 			       pmac_tb_pulsar_addr | pmac_low_i2c_read,
471 			       0x2e, &data, 1);
472 	if (rc != 0)
473 		goto bail;
474 
475 	data = (data & 0x88) | (freeze ? 0x11 : 0x22);
476 
477 	pmac_low_i2c_setmode(pmac_tb_clock_chip_host, pmac_low_i2c_mode_stdsub);
478 	rc = pmac_low_i2c_xfer(pmac_tb_clock_chip_host,
479 			       pmac_tb_pulsar_addr | pmac_low_i2c_write,
480 			       0x2e, &data, 1);
481  bail:
482 	if (rc != 0) {
483 		printk(KERN_ERR "Pulsar Timebase %s rc: %d\n",
484 		       freeze ? "freeze" : "unfreeze", rc);
485 		panic("Timebase freeze failed !\n");
486 	}
487 }
488 
489 
490 static void smp_core99_give_timebase(void)
491 {
492 	/* Open i2c bus for synchronous access */
493 	if (pmac_low_i2c_open(pmac_tb_clock_chip_host, 0))
494 		panic("Can't open i2c for TB sync !\n");
495 
496 	spin_lock(&timebase_lock);
497 	(*pmac_tb_freeze)(1);
498 	mb();
499 	timebase = get_tb();
500 	spin_unlock(&timebase_lock);
501 
502 	while (timebase)
503 		barrier();
504 
505 	spin_lock(&timebase_lock);
506 	(*pmac_tb_freeze)(0);
507 	spin_unlock(&timebase_lock);
508 
509 	/* Close i2c bus */
510 	pmac_low_i2c_close(pmac_tb_clock_chip_host);
511 }
512 
513 
514 static void __devinit smp_core99_take_timebase(void)
515 {
516 	while (!timebase)
517 		barrier();
518 	spin_lock(&timebase_lock);
519 	set_tb(timebase >> 32, timebase & 0xffffffff);
520 	timebase = 0;
521 	spin_unlock(&timebase_lock);
522 }
523 
524 static void __init smp_core99_setup(int ncpus)
525 {
526 	struct device_node *cc = NULL;
527 	struct device_node *p;
528 	u32 *reg;
529 	int ok;
530 
531 	/* HW sync only on these platforms */
532 	if (!machine_is_compatible("PowerMac7,2") &&
533 	    !machine_is_compatible("PowerMac7,3") &&
534 	    !machine_is_compatible("RackMac3,1"))
535 		return;
536 
537 	/* Look for the clock chip */
538 	while ((cc = of_find_node_by_name(cc, "i2c-hwclock")) != NULL) {
539 		p = of_get_parent(cc);
540 		ok = p && device_is_compatible(p, "uni-n-i2c");
541 		of_node_put(p);
542 		if (!ok)
543 			continue;
544 
545 		reg = (u32 *)get_property(cc, "reg", NULL);
546 		if (reg == NULL)
547 			continue;
548 
549 		switch (*reg) {
550 		case 0xd2:
551 			if (device_is_compatible(cc, "pulsar-legacy-slewing")) {
552 				pmac_tb_freeze = smp_core99_pulsar_tb_freeze;
553 				pmac_tb_pulsar_addr = 0xd2;
554 				printk(KERN_INFO "Timebase clock is Pulsar chip\n");
555 			} else if (device_is_compatible(cc, "cy28508")) {
556 				pmac_tb_freeze = smp_core99_cypress_tb_freeze;
557 				printk(KERN_INFO "Timebase clock is Cypress chip\n");
558 			}
559 			break;
560 		case 0xd4:
561 			pmac_tb_freeze = smp_core99_pulsar_tb_freeze;
562 			pmac_tb_pulsar_addr = 0xd4;
563 			printk(KERN_INFO "Timebase clock is Pulsar chip\n");
564 			break;
565 		}
566 		if (pmac_tb_freeze != NULL) {
567 			pmac_tb_clock_chip_host = of_get_parent(cc);
568 			of_node_put(cc);
569 			break;
570 		}
571 	}
572 	if (pmac_tb_freeze == NULL) {
573 		smp_ops->give_timebase = smp_generic_give_timebase;
574 		smp_ops->take_timebase = smp_generic_take_timebase;
575 	}
576 }
577 
578 /* nothing to do here, caches are already set up by service processor */
579 static inline void __devinit core99_init_caches(int cpu)
580 {
581 }
582 
583 #else /* CONFIG_PPC64 */
584 
585 /*
586  * SMP G4 powermacs use a GPIO to enable/disable the timebase.
587  */
588 
589 static unsigned int core99_tb_gpio;	/* Timebase freeze GPIO */
590 
591 static unsigned int pri_tb_hi, pri_tb_lo;
592 static unsigned int pri_tb_stamp;
593 
594 /* not __init, called in sleep/wakeup code */
595 void smp_core99_give_timebase(void)
596 {
597 	unsigned long flags;
598 	unsigned int t;
599 
600 	/* wait for the secondary to be in take_timebase */
601 	for (t = 100000; t > 0 && !sec_tb_reset; --t)
602 		udelay(10);
603 	if (!sec_tb_reset) {
604 		printk(KERN_WARNING "Timeout waiting sync on second CPU\n");
605 		return;
606 	}
607 
608 	/* freeze the timebase and read it */
609 	/* disable interrupts so the timebase is disabled for the
610 	   shortest possible time */
611 	local_irq_save(flags);
612 	pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 4);
613 	pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
614 	mb();
615 	pri_tb_hi = get_tbu();
616 	pri_tb_lo = get_tbl();
617 	pri_tb_stamp = last_jiffy_stamp(smp_processor_id());
618 	mb();
619 
620 	/* tell the secondary we're ready */
621 	sec_tb_reset = 2;
622 	mb();
623 
624 	/* wait for the secondary to have taken it */
625 	for (t = 100000; t > 0 && sec_tb_reset; --t)
626 		udelay(10);
627 	if (sec_tb_reset)
628 		printk(KERN_WARNING "Timeout waiting sync(2) on second CPU\n");
629 	else
630 		smp_tb_synchronized = 1;
631 
632 	/* Now, restart the timebase by leaving the GPIO to an open collector */
633        	pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 0);
634         pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
635 	local_irq_restore(flags);
636 }
637 
638 /* not __init, called in sleep/wakeup code */
639 void smp_core99_take_timebase(void)
640 {
641 	unsigned long flags;
642 
643 	/* tell the primary we're here */
644 	sec_tb_reset = 1;
645 	mb();
646 
647 	/* wait for the primary to set pri_tb_hi/lo */
648 	while (sec_tb_reset < 2)
649 		mb();
650 
651 	/* set our stuff the same as the primary */
652 	local_irq_save(flags);
653 	set_dec(1);
654 	set_tb(pri_tb_hi, pri_tb_lo);
655 	last_jiffy_stamp(smp_processor_id()) = pri_tb_stamp;
656 	mb();
657 
658 	/* tell the primary we're done */
659        	sec_tb_reset = 0;
660 	mb();
661 	local_irq_restore(flags);
662 }
663 
664 /* L2 and L3 cache settings to pass from CPU0 to CPU1 on G4 cpus */
665 volatile static long int core99_l2_cache;
666 volatile static long int core99_l3_cache;
667 
668 static void __devinit core99_init_caches(int cpu)
669 {
670 	if (!cpu_has_feature(CPU_FTR_L2CR))
671 		return;
672 
673 	if (cpu == 0) {
674 		core99_l2_cache = _get_L2CR();
675 		printk("CPU0: L2CR is %lx\n", core99_l2_cache);
676 	} else {
677 		printk("CPU%d: L2CR was %lx\n", cpu, _get_L2CR());
678 		_set_L2CR(0);
679 		_set_L2CR(core99_l2_cache);
680 		printk("CPU%d: L2CR set to %lx\n", cpu, core99_l2_cache);
681 	}
682 
683 	if (!cpu_has_feature(CPU_FTR_L3CR))
684 		return;
685 
686 	if (cpu == 0){
687 		core99_l3_cache = _get_L3CR();
688 		printk("CPU0: L3CR is %lx\n", core99_l3_cache);
689 	} else {
690 		printk("CPU%d: L3CR was %lx\n", cpu, _get_L3CR());
691 		_set_L3CR(0);
692 		_set_L3CR(core99_l3_cache);
693 		printk("CPU%d: L3CR set to %lx\n", cpu, core99_l3_cache);
694 	}
695 }
696 
697 static void __init smp_core99_setup(int ncpus)
698 {
699 	struct device_node *cpu;
700 	u32 *tbprop = NULL;
701 	int i;
702 
703 	core99_tb_gpio = KL_GPIO_TB_ENABLE;	/* default value */
704 	cpu = of_find_node_by_type(NULL, "cpu");
705 	if (cpu != NULL) {
706 		tbprop = (u32 *)get_property(cpu, "timebase-enable", NULL);
707 		if (tbprop)
708 			core99_tb_gpio = *tbprop;
709 		of_node_put(cpu);
710 	}
711 
712 	/* XXX should get this from reg properties */
713 	for (i = 1; i < ncpus; ++i)
714 		smp_hw_index[i] = i;
715 	powersave_nap = 0;
716 }
717 #endif
718 
719 static int __init smp_core99_probe(void)
720 {
721 	struct device_node *cpus;
722 	int ncpus = 0;
723 
724 	if (ppc_md.progress) ppc_md.progress("smp_core99_probe", 0x345);
725 
726 	/* Count CPUs in the device-tree */
727        	for (cpus = NULL; (cpus = of_find_node_by_type(cpus, "cpu")) != NULL;)
728 	       	++ncpus;
729 
730 	printk(KERN_INFO "PowerMac SMP probe found %d cpus\n", ncpus);
731 
732 	/* Nothing more to do if less than 2 of them */
733 	if (ncpus <= 1)
734 		return 1;
735 
736 	smp_core99_setup(ncpus);
737 	mpic_request_ipis();
738 	core99_init_caches(0);
739 
740 	return ncpus;
741 }
742 
743 static void __devinit smp_core99_kick_cpu(int nr)
744 {
745 	unsigned int save_vector;
746 	unsigned long new_vector;
747 	unsigned long flags;
748 	volatile unsigned int *vector
749 		 = ((volatile unsigned int *)(KERNELBASE+0x100));
750 
751 	if (nr < 0 || nr > 3)
752 		return;
753 	if (ppc_md.progress) ppc_md.progress("smp_core99_kick_cpu", 0x346);
754 
755 	local_irq_save(flags);
756 	local_irq_disable();
757 
758 	/* Save reset vector */
759 	save_vector = *vector;
760 
761 	/* Setup fake reset vector that does
762 	 *   b __secondary_start_pmac_0 + nr*8 - KERNELBASE
763 	 */
764 	new_vector = (unsigned long) __secondary_start_pmac_0 + nr * 8;
765 	*vector = 0x48000002 + new_vector - KERNELBASE;
766 
767 	/* flush data cache and inval instruction cache */
768 	flush_icache_range((unsigned long) vector, (unsigned long) vector + 4);
769 
770 	/* Put some life in our friend */
771 	pmac_call_feature(PMAC_FTR_RESET_CPU, NULL, nr, 0);
772 
773 	/* FIXME: We wait a bit for the CPU to take the exception, I should
774 	 * instead wait for the entry code to set something for me. Well,
775 	 * ideally, all that crap will be done in prom.c and the CPU left
776 	 * in a RAM-based wait loop like CHRP.
777 	 */
778 	mdelay(1);
779 
780 	/* Restore our exception vector */
781 	*vector = save_vector;
782 	flush_icache_range((unsigned long) vector, (unsigned long) vector + 4);
783 
784 	local_irq_restore(flags);
785 	if (ppc_md.progress) ppc_md.progress("smp_core99_kick_cpu done", 0x347);
786 }
787 
788 static void __devinit smp_core99_setup_cpu(int cpu_nr)
789 {
790 	/* Setup L2/L3 */
791 	if (cpu_nr != 0)
792 		core99_init_caches(cpu_nr);
793 
794 	/* Setup openpic */
795 	mpic_setup_this_cpu();
796 
797 	if (cpu_nr == 0) {
798 #ifdef CONFIG_POWER4
799 		extern void g5_phy_disable_cpu1(void);
800 
801 		/* If we didn't start the second CPU, we must take
802 		 * it off the bus
803 		 */
804 		if (machine_is_compatible("MacRISC4") &&
805 		    num_online_cpus() < 2)
806 			g5_phy_disable_cpu1();
807 #endif /* CONFIG_POWER4 */
808 		if (ppc_md.progress) ppc_md.progress("core99_setup_cpu 0 done", 0x349);
809 	}
810 }
811 
812 
813 /* Core99 Macs (dual G4s and G5s) */
814 struct smp_ops_t core99_smp_ops = {
815 	.message_pass	= smp_mpic_message_pass,
816 	.probe		= smp_core99_probe,
817 	.kick_cpu	= smp_core99_kick_cpu,
818 	.setup_cpu	= smp_core99_setup_cpu,
819 	.give_timebase	= smp_core99_give_timebase,
820 	.take_timebase	= smp_core99_take_timebase,
821 };
822 
823 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_PPC32)
824 
825 int __cpu_disable(void)
826 {
827 	cpu_clear(smp_processor_id(), cpu_online_map);
828 
829 	/* XXX reset cpu affinity here */
830 	mpic_cpu_set_priority(0xf);
831 	asm volatile("mtdec %0" : : "r" (0x7fffffff));
832 	mb();
833 	udelay(20);
834 	asm volatile("mtdec %0" : : "r" (0x7fffffff));
835 	return 0;
836 }
837 
838 extern void low_cpu_die(void) __attribute__((noreturn)); /* in sleep.S */
839 static int cpu_dead[NR_CPUS];
840 
841 void cpu_die(void)
842 {
843 	local_irq_disable();
844 	cpu_dead[smp_processor_id()] = 1;
845 	mb();
846 	low_cpu_die();
847 }
848 
849 void __cpu_die(unsigned int cpu)
850 {
851 	int timeout;
852 
853 	timeout = 1000;
854 	while (!cpu_dead[cpu]) {
855 		if (--timeout == 0) {
856 			printk("CPU %u refused to die!\n", cpu);
857 			break;
858 		}
859 		msleep(1);
860 	}
861 	cpu_callin_map[cpu] = 0;
862 	cpu_dead[cpu] = 0;
863 }
864 
865 #endif
866