2 * SMP support for power macintosh.
4 * We support both the old "powersurge" SMP architecture
5 * and the current Core99 (G4 PowerMac) machines.
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.
13 * Support Macintosh G4 SMP by Troy Benjegerdes (hozer@drgw.net)
14 * and Ben Herrenschmidt <benh@kernel.crashing.org>.
16 * Support for DayStar quad CPU cards
17 * Copyright (C) XLR8, Inc. 1994-2000
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.
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>
38 #include <asm/ptrace.h>
39 #include <asm/atomic.h>
42 #include <asm/pgtable.h>
43 #include <asm/sections.h>
47 #include <asm/residual.h>
48 #include <asm/machdep.h>
49 #include <asm/pmac_feature.h>
51 #include <asm/open_pic.h>
52 #include <asm/cacheflush.h>
53 #include <asm/keylargo.h>
56 * Powersurge (old powermac SMP) support.
59 extern void __secondary_start_pmac_0(void);
61 /* Addresses for powersurge registers */
62 #define HAMMERHEAD_BASE 0xf8000000
63 #define HHEAD_CONFIG 0x90
64 #define HHEAD_SEC_INTR 0xc0
66 /* register for interrupting the primary processor on the powersurge */
67 /* N.B. this is actually the ethernet ROM! */
68 #define PSURGE_PRI_INTR 0xf3019000
70 /* register for storing the start address for the secondary processor */
71 /* N.B. this is the PCI config space address register for the 1st bridge */
72 #define PSURGE_START 0xf2800000
74 /* Daystar/XLR8 4-CPU card */
75 #define PSURGE_QUAD_REG_ADDR 0xf8800000
77 #define PSURGE_QUAD_IRQ_SET 0
78 #define PSURGE_QUAD_IRQ_CLR 1
79 #define PSURGE_QUAD_IRQ_PRIMARY 2
80 #define PSURGE_QUAD_CKSTOP_CTL 3
81 #define PSURGE_QUAD_PRIMARY_ARB 4
82 #define PSURGE_QUAD_BOARD_ID 6
83 #define PSURGE_QUAD_WHICH_CPU 7
84 #define PSURGE_QUAD_CKSTOP_RDBK 8
85 #define PSURGE_QUAD_RESET_CTL 11
87 #define PSURGE_QUAD_OUT(r, v) (out_8(quad_base + ((r) << 4) + 4, (v)))
88 #define PSURGE_QUAD_IN(r) (in_8(quad_base + ((r) << 4) + 4) & 0x0f)
89 #define PSURGE_QUAD_BIS(r, v) (PSURGE_QUAD_OUT((r), PSURGE_QUAD_IN(r) | (v)))
90 #define PSURGE_QUAD_BIC(r, v) (PSURGE_QUAD_OUT((r), PSURGE_QUAD_IN(r) & ~(v)))
92 /* virtual addresses for the above */
93 static volatile u8 __iomem *hhead_base;
94 static volatile u8 __iomem *quad_base;
95 static volatile u32 __iomem *psurge_pri_intr;
96 static volatile u8 __iomem *psurge_sec_intr;
97 static volatile u32 __iomem *psurge_start;
99 /* values for psurge_type */
100 #define PSURGE_NONE -1
101 #define PSURGE_DUAL 0
102 #define PSURGE_QUAD_OKEE 1
103 #define PSURGE_QUAD_COTTON 2
104 #define PSURGE_QUAD_ICEGRASS 3
106 /* what sort of powersurge board we have */
107 static int psurge_type = PSURGE_NONE;
109 /* L2 and L3 cache settings to pass from CPU0 to CPU1 */
110 volatile static long int core99_l2_cache;
111 volatile static long int core99_l3_cache;
113 /* Timebase freeze GPIO */
114 static unsigned int core99_tb_gpio;
116 /* Sync flag for HW tb sync */
117 static volatile int sec_tb_reset = 0;
118 static unsigned int pri_tb_hi, pri_tb_lo;
119 static unsigned int pri_tb_stamp;
121 static void __devinit core99_init_caches(int cpu)
123 if (!cpu_has_feature(CPU_FTR_L2CR))
127 core99_l2_cache = _get_L2CR();
128 printk("CPU0: L2CR is %lx\n", core99_l2_cache);
130 printk("CPU%d: L2CR was %lx\n", cpu, _get_L2CR());
132 _set_L2CR(core99_l2_cache);
133 printk("CPU%d: L2CR set to %lx\n", cpu, core99_l2_cache);
136 if (!cpu_has_feature(CPU_FTR_L3CR))
140 core99_l3_cache = _get_L3CR();
141 printk("CPU0: L3CR is %lx\n", core99_l3_cache);
143 printk("CPU%d: L3CR was %lx\n", cpu, _get_L3CR());
145 _set_L3CR(core99_l3_cache);
146 printk("CPU%d: L3CR set to %lx\n", cpu, core99_l3_cache);
151 * Set and clear IPIs for powersurge.
153 static inline void psurge_set_ipi(int cpu)
155 if (psurge_type == PSURGE_NONE)
158 in_be32(psurge_pri_intr);
159 else if (psurge_type == PSURGE_DUAL)
160 out_8(psurge_sec_intr, 0);
162 PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_SET, 1 << cpu);
165 static inline void psurge_clr_ipi(int cpu)
168 switch(psurge_type) {
170 out_8(psurge_sec_intr, ~0);
174 PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_CLR, 1 << cpu);
180 * On powersurge (old SMP powermac architecture) we don't have
181 * separate IPIs for separate messages like openpic does. Instead
182 * we have a bitmap for each processor, where a 1 bit means that
183 * the corresponding message is pending for that processor.
184 * Ideally each cpu's entry would be in a different cache line.
187 static unsigned long psurge_smp_message[NR_CPUS];
189 void psurge_smp_message_recv(struct pt_regs *regs)
191 int cpu = smp_processor_id();
194 /* clear interrupt */
197 if (num_online_cpus() < 2)
200 /* make sure there is a message there */
201 for (msg = 0; msg < 4; msg++)
202 if (test_and_clear_bit(msg, &psurge_smp_message[cpu]))
203 smp_message_recv(msg, regs);
206 irqreturn_t psurge_primary_intr(int irq, void *d, struct pt_regs *regs)
208 psurge_smp_message_recv(regs);
212 static void smp_psurge_message_pass(int target, int msg)
216 if (num_online_cpus() < 2)
219 for (i = 0; i < NR_CPUS; i++) {
222 if (target == MSG_ALL
223 || (target == MSG_ALL_BUT_SELF && i != smp_processor_id())
225 set_bit(msg, &psurge_smp_message[i]);
232 * Determine a quad card presence. We read the board ID register, we
233 * force the data bus to change to something else, and we read it again.
234 * It it's stable, then the register probably exist (ugh !)
236 static int __init psurge_quad_probe(void)
241 type = PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID);
242 if (type < PSURGE_QUAD_OKEE || type > PSURGE_QUAD_ICEGRASS
243 || type != PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID))
246 /* looks OK, try a slightly more rigorous test */
247 /* bogus is not necessarily cacheline-aligned,
248 though I don't suppose that really matters. -- paulus */
249 for (i = 0; i < 100; i++) {
250 volatile u32 bogus[8];
251 bogus[(0+i)%8] = 0x00000000;
252 bogus[(1+i)%8] = 0x55555555;
253 bogus[(2+i)%8] = 0xFFFFFFFF;
254 bogus[(3+i)%8] = 0xAAAAAAAA;
255 bogus[(4+i)%8] = 0x33333333;
256 bogus[(5+i)%8] = 0xCCCCCCCC;
257 bogus[(6+i)%8] = 0xCCCCCCCC;
258 bogus[(7+i)%8] = 0x33333333;
260 asm volatile("dcbf 0,%0" : : "r" (bogus) : "memory");
262 if (type != PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID))
268 static void __init psurge_quad_init(void)
272 if (ppc_md.progress) ppc_md.progress("psurge_quad_init", 0x351);
273 procbits = ~PSURGE_QUAD_IN(PSURGE_QUAD_WHICH_CPU);
274 if (psurge_type == PSURGE_QUAD_ICEGRASS)
275 PSURGE_QUAD_BIS(PSURGE_QUAD_RESET_CTL, procbits);
277 PSURGE_QUAD_BIC(PSURGE_QUAD_CKSTOP_CTL, procbits);
279 out_8(psurge_sec_intr, ~0);
280 PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_CLR, procbits);
281 PSURGE_QUAD_BIS(PSURGE_QUAD_RESET_CTL, procbits);
282 if (psurge_type != PSURGE_QUAD_ICEGRASS)
283 PSURGE_QUAD_BIS(PSURGE_QUAD_CKSTOP_CTL, procbits);
284 PSURGE_QUAD_BIC(PSURGE_QUAD_PRIMARY_ARB, procbits);
286 PSURGE_QUAD_BIC(PSURGE_QUAD_RESET_CTL, procbits);
288 PSURGE_QUAD_BIS(PSURGE_QUAD_PRIMARY_ARB, procbits);
292 static int __init smp_psurge_probe(void)
296 /* We don't do SMP on the PPC601 -- paulus */
297 if (PVR_VER(mfspr(SPRN_PVR)) == 1)
301 * The powersurge cpu board can be used in the generation
302 * of powermacs that have a socket for an upgradeable cpu card,
303 * including the 7500, 8500, 9500, 9600.
304 * The device tree doesn't tell you if you have 2 cpus because
305 * OF doesn't know anything about the 2nd processor.
306 * Instead we look for magic bits in magic registers,
307 * in the hammerhead memory controller in the case of the
308 * dual-cpu powersurge board. -- paulus.
310 if (find_devices("hammerhead") == NULL)
313 hhead_base = ioremap(HAMMERHEAD_BASE, 0x800);
314 quad_base = ioremap(PSURGE_QUAD_REG_ADDR, 1024);
315 psurge_sec_intr = hhead_base + HHEAD_SEC_INTR;
317 psurge_type = psurge_quad_probe();
318 if (psurge_type != PSURGE_DUAL) {
320 /* All released cards using this HW design have 4 CPUs */
324 if ((in_8(hhead_base + HHEAD_CONFIG) & 0x02) == 0) {
325 /* not a dual-cpu card */
327 psurge_type = PSURGE_NONE;
333 psurge_start = ioremap(PSURGE_START, 4);
334 psurge_pri_intr = ioremap(PSURGE_PRI_INTR, 4);
336 /* this is not actually strictly necessary -- paulus. */
337 for (i = 1; i < ncpus; ++i)
340 if (ppc_md.progress) ppc_md.progress("smp_psurge_probe - done", 0x352);
345 static void __init smp_psurge_kick_cpu(int nr)
347 unsigned long start = __pa(__secondary_start_pmac_0) + nr * 8;
350 /* may need to flush here if secondary bats aren't setup */
351 for (a = KERNELBASE; a < KERNELBASE + 0x800000; a += 32)
352 asm volatile("dcbf 0,%0" : : "r" (a) : "memory");
353 asm volatile("sync");
355 if (ppc_md.progress) ppc_md.progress("smp_psurge_kick_cpu", 0x353);
357 out_be32(psurge_start, start);
364 if (ppc_md.progress) ppc_md.progress("smp_psurge_kick_cpu - done", 0x354);
368 * With the dual-cpu powersurge board, the decrementers and timebases
369 * of both cpus are frozen after the secondary cpu is started up,
370 * until we give the secondary cpu another interrupt. This routine
371 * uses this to get the timebases synchronized.
374 static void __init psurge_dual_sync_tb(int cpu_nr)
378 set_dec(tb_ticks_per_jiffy);
380 last_jiffy_stamp(cpu_nr) = 0;
388 /* wait for the secondary to have reset its TB before proceeding */
389 for (t = 10000000; t > 0 && !sec_tb_reset; --t)
392 /* now interrupt the secondary, starting both TBs */
395 smp_tb_synchronized = 1;
398 static struct irqaction psurge_irqaction = {
399 .handler = psurge_primary_intr,
400 .flags = SA_INTERRUPT,
401 .mask = CPU_MASK_NONE,
402 .name = "primary IPI",
405 static void __init smp_psurge_setup_cpu(int cpu_nr)
409 /* If we failed to start the second CPU, we should still
410 * send it an IPI to start the timebase & DEC or we might
413 if (num_online_cpus() < 2) {
414 if (psurge_type == PSURGE_DUAL)
418 /* reset the entry point so if we get another intr we won't
419 * try to startup again */
420 out_be32(psurge_start, 0x100);
421 if (setup_irq(30, &psurge_irqaction))
422 printk(KERN_ERR "Couldn't get primary IPI interrupt");
425 if (psurge_type == PSURGE_DUAL)
426 psurge_dual_sync_tb(cpu_nr);
429 void __init smp_psurge_take_timebase(void)
431 /* Dummy implementation */
434 void __init smp_psurge_give_timebase(void)
436 /* Dummy implementation */
439 static int __init smp_core99_probe(void)
442 extern int powersave_nap;
444 struct device_node *cpus, *firstcpu;
445 int i, ncpus = 0, boot_cpu = -1;
448 if (ppc_md.progress) ppc_md.progress("smp_core99_probe", 0x345);
449 cpus = firstcpu = find_type_devices("cpu");
450 while(cpus != NULL) {
451 u32 *regprop = (u32 *)get_property(cpus, "reg", NULL);
452 char *stateprop = (char *)get_property(cpus, "state", NULL);
453 if (regprop != NULL && stateprop != NULL &&
454 !strncmp(stateprop, "running", 7))
460 printk(KERN_WARNING "Couldn't detect boot CPU !\n");
462 printk(KERN_WARNING "Boot CPU is %d, unsupported setup !\n", boot_cpu);
464 if (machine_is_compatible("MacRISC4")) {
465 extern struct smp_ops_t core99_smp_ops;
467 core99_smp_ops.take_timebase = smp_generic_take_timebase;
468 core99_smp_ops.give_timebase = smp_generic_give_timebase;
470 if (firstcpu != NULL)
471 tbprop = (u32 *)get_property(firstcpu, "timebase-enable", NULL);
473 core99_tb_gpio = *tbprop;
475 core99_tb_gpio = KL_GPIO_TB_ENABLE;
479 openpic_request_IPIs();
480 for (i = 1; i < ncpus; ++i)
485 core99_init_caches(0);
491 static void __devinit smp_core99_kick_cpu(int nr)
493 unsigned long save_vector, new_vector;
496 volatile unsigned long *vector
497 = ((volatile unsigned long *)(KERNELBASE+0x100));
498 if (nr < 0 || nr > 3)
500 if (ppc_md.progress) ppc_md.progress("smp_core99_kick_cpu", 0x346);
502 local_irq_save(flags);
505 /* Save reset vector */
506 save_vector = *vector;
508 /* Setup fake reset vector that does
509 * b __secondary_start_pmac_0 + nr*8 - KERNELBASE
511 new_vector = (unsigned long) __secondary_start_pmac_0 + nr * 8;
512 *vector = 0x48000002 + new_vector - KERNELBASE;
514 /* flush data cache and inval instruction cache */
515 flush_icache_range((unsigned long) vector, (unsigned long) vector + 4);
517 /* Put some life in our friend */
518 pmac_call_feature(PMAC_FTR_RESET_CPU, NULL, nr, 0);
520 /* FIXME: We wait a bit for the CPU to take the exception, I should
521 * instead wait for the entry code to set something for me. Well,
522 * ideally, all that crap will be done in prom.c and the CPU left
523 * in a RAM-based wait loop like CHRP.
527 /* Restore our exception vector */
528 *vector = save_vector;
529 flush_icache_range((unsigned long) vector, (unsigned long) vector + 4);
531 local_irq_restore(flags);
532 if (ppc_md.progress) ppc_md.progress("smp_core99_kick_cpu done", 0x347);
535 static void __devinit smp_core99_setup_cpu(int cpu_nr)
539 core99_init_caches(cpu_nr);
542 do_openpic_setup_cpu();
546 extern void g5_phy_disable_cpu1(void);
548 /* If we didn't start the second CPU, we must take
551 if (machine_is_compatible("MacRISC4") &&
552 num_online_cpus() < 2)
553 g5_phy_disable_cpu1();
554 #endif /* CONFIG_POWER4 */
555 if (ppc_md.progress) ppc_md.progress("core99_setup_cpu 0 done", 0x349);
559 /* not __init, called in sleep/wakeup code */
560 void smp_core99_take_timebase(void)
564 /* tell the primary we're here */
568 /* wait for the primary to set pri_tb_hi/lo */
569 while (sec_tb_reset < 2)
572 /* set our stuff the same as the primary */
573 local_irq_save(flags);
575 set_tb(pri_tb_hi, pri_tb_lo);
576 last_jiffy_stamp(smp_processor_id()) = pri_tb_stamp;
579 /* tell the primary we're done */
582 local_irq_restore(flags);
585 /* not __init, called in sleep/wakeup code */
586 void smp_core99_give_timebase(void)
591 /* wait for the secondary to be in take_timebase */
592 for (t = 100000; t > 0 && !sec_tb_reset; --t)
595 printk(KERN_WARNING "Timeout waiting sync on second CPU\n");
599 /* freeze the timebase and read it */
600 /* disable interrupts so the timebase is disabled for the
601 shortest possible time */
602 local_irq_save(flags);
603 pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 4);
604 pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
606 pri_tb_hi = get_tbu();
607 pri_tb_lo = get_tbl();
608 pri_tb_stamp = last_jiffy_stamp(smp_processor_id());
611 /* tell the secondary we're ready */
615 /* wait for the secondary to have taken it */
616 for (t = 100000; t > 0 && sec_tb_reset; --t)
619 printk(KERN_WARNING "Timeout waiting sync(2) on second CPU\n");
621 smp_tb_synchronized = 1;
623 /* Now, restart the timebase by leaving the GPIO to an open collector */
624 pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 0);
625 pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
626 local_irq_restore(flags);
630 /* PowerSurge-style Macs */
631 struct smp_ops_t psurge_smp_ops = {
632 .message_pass = smp_psurge_message_pass,
633 .probe = smp_psurge_probe,
634 .kick_cpu = smp_psurge_kick_cpu,
635 .setup_cpu = smp_psurge_setup_cpu,
636 .give_timebase = smp_psurge_give_timebase,
637 .take_timebase = smp_psurge_take_timebase,
640 /* Core99 Macs (dual G4s) */
641 struct smp_ops_t core99_smp_ops = {
642 .message_pass = smp_openpic_message_pass,
643 .probe = smp_core99_probe,
644 .kick_cpu = smp_core99_kick_cpu,
645 .setup_cpu = smp_core99_setup_cpu,
646 .give_timebase = smp_core99_give_timebase,
647 .take_timebase = smp_core99_take_timebase,
650 #ifdef CONFIG_HOTPLUG_CPU
652 int __cpu_disable(void)
654 cpu_clear(smp_processor_id(), cpu_online_map);
656 /* XXX reset cpu affinity here */
657 openpic_set_priority(0xf);
658 asm volatile("mtdec %0" : : "r" (0x7fffffff));
661 asm volatile("mtdec %0" : : "r" (0x7fffffff));
665 extern void low_cpu_die(void) __attribute__((noreturn)); /* in pmac_sleep.S */
666 static int cpu_dead[NR_CPUS];
671 cpu_dead[smp_processor_id()] = 1;
676 void __cpu_die(unsigned int cpu)
681 while (!cpu_dead[cpu]) {
682 if (--timeout == 0) {
683 printk("CPU %u refused to die!\n", cpu);
688 cpu_callin_map[cpu] = 0;