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>
37 #include <asm/ptrace.h>
38 #include <asm/atomic.h>
41 #include <asm/pgtable.h>
42 #include <asm/sections.h>
46 #include <asm/residual.h>
47 #include <asm/machdep.h>
48 #include <asm/pmac_feature.h>
50 #include <asm/open_pic.h>
51 #include <asm/cacheflush.h>
52 #include <asm/keylargo.h>
55 * Powersurge (old powermac SMP) support.
58 extern void __secondary_start_psurge(void);
59 extern void __secondary_start_psurge2(void); /* Temporary horrible hack */
60 extern void __secondary_start_psurge3(void); /* Temporary horrible hack */
62 /* Addresses for powersurge registers */
63 #define HAMMERHEAD_BASE 0xf8000000
64 #define HHEAD_CONFIG 0x90
65 #define HHEAD_SEC_INTR 0xc0
67 /* register for interrupting the primary processor on the powersurge */
68 /* N.B. this is actually the ethernet ROM! */
69 #define PSURGE_PRI_INTR 0xf3019000
71 /* register for storing the start address for the secondary processor */
72 /* N.B. this is the PCI config space address register for the 1st bridge */
73 #define PSURGE_START 0xf2800000
75 /* Daystar/XLR8 4-CPU card */
76 #define PSURGE_QUAD_REG_ADDR 0xf8800000
78 #define PSURGE_QUAD_IRQ_SET 0
79 #define PSURGE_QUAD_IRQ_CLR 1
80 #define PSURGE_QUAD_IRQ_PRIMARY 2
81 #define PSURGE_QUAD_CKSTOP_CTL 3
82 #define PSURGE_QUAD_PRIMARY_ARB 4
83 #define PSURGE_QUAD_BOARD_ID 6
84 #define PSURGE_QUAD_WHICH_CPU 7
85 #define PSURGE_QUAD_CKSTOP_RDBK 8
86 #define PSURGE_QUAD_RESET_CTL 11
88 #define PSURGE_QUAD_OUT(r, v) (out_8(quad_base + ((r) << 4) + 4, (v)))
89 #define PSURGE_QUAD_IN(r) (in_8(quad_base + ((r) << 4) + 4) & 0x0f)
90 #define PSURGE_QUAD_BIS(r, v) (PSURGE_QUAD_OUT((r), PSURGE_QUAD_IN(r) | (v)))
91 #define PSURGE_QUAD_BIC(r, v) (PSURGE_QUAD_OUT((r), PSURGE_QUAD_IN(r) & ~(v)))
93 /* virtual addresses for the above */
94 static volatile u8 __iomem *hhead_base;
95 static volatile u8 __iomem *quad_base;
96 static volatile u32 __iomem *psurge_pri_intr;
97 static volatile u8 __iomem *psurge_sec_intr;
98 static volatile u32 __iomem *psurge_start;
100 /* values for psurge_type */
101 #define PSURGE_NONE -1
102 #define PSURGE_DUAL 0
103 #define PSURGE_QUAD_OKEE 1
104 #define PSURGE_QUAD_COTTON 2
105 #define PSURGE_QUAD_ICEGRASS 3
107 /* what sort of powersurge board we have */
108 static int psurge_type = PSURGE_NONE;
110 /* L2 and L3 cache settings to pass from CPU0 to CPU1 */
111 volatile static long int core99_l2_cache;
112 volatile static long int core99_l3_cache;
114 /* Timebase freeze GPIO */
115 static unsigned int core99_tb_gpio;
117 /* Sync flag for HW tb sync */
118 static volatile int sec_tb_reset = 0;
119 static unsigned int pri_tb_hi, pri_tb_lo;
120 static unsigned int pri_tb_stamp;
122 static void __init core99_init_caches(int cpu)
124 if (!cpu_has_feature(CPU_FTR_L2CR))
128 core99_l2_cache = _get_L2CR();
129 printk("CPU0: L2CR is %lx\n", core99_l2_cache);
131 printk("CPU%d: L2CR was %lx\n", cpu, _get_L2CR());
133 _set_L2CR(core99_l2_cache);
134 printk("CPU%d: L2CR set to %lx\n", cpu, core99_l2_cache);
137 if (!cpu_has_feature(CPU_FTR_L3CR))
141 core99_l3_cache = _get_L3CR();
142 printk("CPU0: L3CR is %lx\n", core99_l3_cache);
144 printk("CPU%d: L3CR was %lx\n", cpu, _get_L3CR());
146 _set_L3CR(core99_l3_cache);
147 printk("CPU%d: L3CR set to %lx\n", cpu, core99_l3_cache);
152 * Set and clear IPIs for powersurge.
154 static inline void psurge_set_ipi(int cpu)
156 if (psurge_type == PSURGE_NONE)
159 in_be32(psurge_pri_intr);
160 else if (psurge_type == PSURGE_DUAL)
161 out_8(psurge_sec_intr, 0);
163 PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_SET, 1 << cpu);
166 static inline void psurge_clr_ipi(int cpu)
169 switch(psurge_type) {
171 out_8(psurge_sec_intr, ~0);
175 PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_CLR, 1 << cpu);
181 * On powersurge (old SMP powermac architecture) we don't have
182 * separate IPIs for separate messages like openpic does. Instead
183 * we have a bitmap for each processor, where a 1 bit means that
184 * the corresponding message is pending for that processor.
185 * Ideally each cpu's entry would be in a different cache line.
188 static unsigned long psurge_smp_message[NR_CPUS];
190 void __pmac psurge_smp_message_recv(struct pt_regs *regs)
192 int cpu = smp_processor_id();
195 /* clear interrupt */
198 if (num_online_cpus() < 2)
201 /* make sure there is a message there */
202 for (msg = 0; msg < 4; msg++)
203 if (test_and_clear_bit(msg, &psurge_smp_message[cpu]))
204 smp_message_recv(msg, regs);
207 irqreturn_t __pmac psurge_primary_intr(int irq, void *d, struct pt_regs *regs)
209 psurge_smp_message_recv(regs);
213 static void __pmac smp_psurge_message_pass(int target, int msg, unsigned long data,
218 if (num_online_cpus() < 2)
221 for (i = 0; i < NR_CPUS; i++) {
224 if (target == MSG_ALL
225 || (target == MSG_ALL_BUT_SELF && i != smp_processor_id())
227 set_bit(msg, &psurge_smp_message[i]);
234 * Determine a quad card presence. We read the board ID register, we
235 * force the data bus to change to something else, and we read it again.
236 * It it's stable, then the register probably exist (ugh !)
238 static int __init psurge_quad_probe(void)
243 type = PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID);
244 if (type < PSURGE_QUAD_OKEE || type > PSURGE_QUAD_ICEGRASS
245 || type != PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID))
248 /* looks OK, try a slightly more rigorous test */
249 /* bogus is not necessarily cacheline-aligned,
250 though I don't suppose that really matters. -- paulus */
251 for (i = 0; i < 100; i++) {
252 volatile u32 bogus[8];
253 bogus[(0+i)%8] = 0x00000000;
254 bogus[(1+i)%8] = 0x55555555;
255 bogus[(2+i)%8] = 0xFFFFFFFF;
256 bogus[(3+i)%8] = 0xAAAAAAAA;
257 bogus[(4+i)%8] = 0x33333333;
258 bogus[(5+i)%8] = 0xCCCCCCCC;
259 bogus[(6+i)%8] = 0xCCCCCCCC;
260 bogus[(7+i)%8] = 0x33333333;
262 asm volatile("dcbf 0,%0" : : "r" (bogus) : "memory");
264 if (type != PSURGE_QUAD_IN(PSURGE_QUAD_BOARD_ID))
270 static void __init psurge_quad_init(void)
274 if (ppc_md.progress) ppc_md.progress("psurge_quad_init", 0x351);
275 procbits = ~PSURGE_QUAD_IN(PSURGE_QUAD_WHICH_CPU);
276 if (psurge_type == PSURGE_QUAD_ICEGRASS)
277 PSURGE_QUAD_BIS(PSURGE_QUAD_RESET_CTL, procbits);
279 PSURGE_QUAD_BIC(PSURGE_QUAD_CKSTOP_CTL, procbits);
281 out_8(psurge_sec_intr, ~0);
282 PSURGE_QUAD_OUT(PSURGE_QUAD_IRQ_CLR, procbits);
283 PSURGE_QUAD_BIS(PSURGE_QUAD_RESET_CTL, procbits);
284 if (psurge_type != PSURGE_QUAD_ICEGRASS)
285 PSURGE_QUAD_BIS(PSURGE_QUAD_CKSTOP_CTL, procbits);
286 PSURGE_QUAD_BIC(PSURGE_QUAD_PRIMARY_ARB, procbits);
288 PSURGE_QUAD_BIC(PSURGE_QUAD_RESET_CTL, procbits);
290 PSURGE_QUAD_BIS(PSURGE_QUAD_PRIMARY_ARB, procbits);
294 static int __init smp_psurge_probe(void)
298 /* We don't do SMP on the PPC601 -- paulus */
299 if (PVR_VER(mfspr(SPRN_PVR)) == 1)
303 * The powersurge cpu board can be used in the generation
304 * of powermacs that have a socket for an upgradeable cpu card,
305 * including the 7500, 8500, 9500, 9600.
306 * The device tree doesn't tell you if you have 2 cpus because
307 * OF doesn't know anything about the 2nd processor.
308 * Instead we look for magic bits in magic registers,
309 * in the hammerhead memory controller in the case of the
310 * dual-cpu powersurge board. -- paulus.
312 if (find_devices("hammerhead") == NULL)
315 hhead_base = ioremap(HAMMERHEAD_BASE, 0x800);
316 quad_base = ioremap(PSURGE_QUAD_REG_ADDR, 1024);
317 psurge_sec_intr = hhead_base + HHEAD_SEC_INTR;
319 psurge_type = psurge_quad_probe();
320 if (psurge_type != PSURGE_DUAL) {
322 /* All released cards using this HW design have 4 CPUs */
326 if ((in_8(hhead_base + HHEAD_CONFIG) & 0x02) == 0) {
327 /* not a dual-cpu card */
329 psurge_type = PSURGE_NONE;
335 psurge_start = ioremap(PSURGE_START, 4);
336 psurge_pri_intr = ioremap(PSURGE_PRI_INTR, 4);
338 /* this is not actually strictly necessary -- paulus. */
339 for (i = 1; i < ncpus; ++i)
342 if (ppc_md.progress) ppc_md.progress("smp_psurge_probe - done", 0x352);
347 static void __init smp_psurge_kick_cpu(int nr)
349 void (*start)(void) = __secondary_start_psurge;
352 /* may need to flush here if secondary bats aren't setup */
353 for (a = KERNELBASE; a < KERNELBASE + 0x800000; a += 32)
354 asm volatile("dcbf 0,%0" : : "r" (a) : "memory");
355 asm volatile("sync");
357 if (ppc_md.progress) ppc_md.progress("smp_psurge_kick_cpu", 0x353);
359 /* setup entry point of secondary processor */
362 start = __secondary_start_psurge2;
365 start = __secondary_start_psurge3;
369 out_be32(psurge_start, __pa(start));
376 if (ppc_md.progress) ppc_md.progress("smp_psurge_kick_cpu - done", 0x354);
380 * With the dual-cpu powersurge board, the decrementers and timebases
381 * of both cpus are frozen after the secondary cpu is started up,
382 * until we give the secondary cpu another interrupt. This routine
383 * uses this to get the timebases synchronized.
386 static void __init psurge_dual_sync_tb(int cpu_nr)
390 set_dec(tb_ticks_per_jiffy);
392 last_jiffy_stamp(cpu_nr) = 0;
400 /* wait for the secondary to have reset its TB before proceeding */
401 for (t = 10000000; t > 0 && !sec_tb_reset; --t)
404 /* now interrupt the secondary, starting both TBs */
407 smp_tb_synchronized = 1;
410 static struct irqaction psurge_irqaction = {
411 .handler = psurge_primary_intr,
412 .flags = SA_INTERRUPT,
413 .mask = CPU_MASK_NONE,
414 .name = "primary IPI",
417 static void __init smp_psurge_setup_cpu(int cpu_nr)
421 /* If we failed to start the second CPU, we should still
422 * send it an IPI to start the timebase & DEC or we might
425 if (num_online_cpus() < 2) {
426 if (psurge_type == PSURGE_DUAL)
430 /* reset the entry point so if we get another intr we won't
431 * try to startup again */
432 out_be32(psurge_start, 0x100);
433 if (setup_irq(30, &psurge_irqaction))
434 printk(KERN_ERR "Couldn't get primary IPI interrupt");
437 if (psurge_type == PSURGE_DUAL)
438 psurge_dual_sync_tb(cpu_nr);
441 void __init smp_psurge_take_timebase(void)
443 /* Dummy implementation */
446 void __init smp_psurge_give_timebase(void)
448 /* Dummy implementation */
451 static int __init smp_core99_probe(void)
454 extern int powersave_nap;
456 struct device_node *cpus, *firstcpu;
457 int i, ncpus = 0, boot_cpu = -1;
460 if (ppc_md.progress) ppc_md.progress("smp_core99_probe", 0x345);
461 cpus = firstcpu = find_type_devices("cpu");
462 while(cpus != NULL) {
463 u32 *regprop = (u32 *)get_property(cpus, "reg", NULL);
464 char *stateprop = (char *)get_property(cpus, "state", NULL);
465 if (regprop != NULL && stateprop != NULL &&
466 !strncmp(stateprop, "running", 7))
472 printk(KERN_WARNING "Couldn't detect boot CPU !\n");
474 printk(KERN_WARNING "Boot CPU is %d, unsupported setup !\n", boot_cpu);
476 if (machine_is_compatible("MacRISC4")) {
477 extern struct smp_ops_t core99_smp_ops;
479 core99_smp_ops.take_timebase = smp_generic_take_timebase;
480 core99_smp_ops.give_timebase = smp_generic_give_timebase;
482 if (firstcpu != NULL)
483 tbprop = (u32 *)get_property(firstcpu, "timebase-enable", NULL);
485 core99_tb_gpio = *tbprop;
487 core99_tb_gpio = KL_GPIO_TB_ENABLE;
491 openpic_request_IPIs();
492 for (i = 1; i < ncpus; ++i)
497 core99_init_caches(0);
503 static void __init smp_core99_kick_cpu(int nr)
505 unsigned long save_vector, new_vector;
508 volatile unsigned long *vector
509 = ((volatile unsigned long *)(KERNELBASE+0x100));
510 if (nr < 1 || nr > 3)
512 if (ppc_md.progress) ppc_md.progress("smp_core99_kick_cpu", 0x346);
514 local_irq_save(flags);
517 /* Save reset vector */
518 save_vector = *vector;
520 /* Setup fake reset vector that does
521 * b __secondary_start_psurge - KERNELBASE
525 new_vector = (unsigned long)__secondary_start_psurge;
528 new_vector = (unsigned long)__secondary_start_psurge2;
531 new_vector = (unsigned long)__secondary_start_psurge3;
534 *vector = 0x48000002 + new_vector - KERNELBASE;
536 /* flush data cache and inval instruction cache */
537 flush_icache_range((unsigned long) vector, (unsigned long) vector + 4);
539 /* Put some life in our friend */
540 pmac_call_feature(PMAC_FTR_RESET_CPU, NULL, nr, 0);
542 /* FIXME: We wait a bit for the CPU to take the exception, I should
543 * instead wait for the entry code to set something for me. Well,
544 * ideally, all that crap will be done in prom.c and the CPU left
545 * in a RAM-based wait loop like CHRP.
549 /* Restore our exception vector */
550 *vector = save_vector;
551 flush_icache_range((unsigned long) vector, (unsigned long) vector + 4);
553 local_irq_restore(flags);
554 if (ppc_md.progress) ppc_md.progress("smp_core99_kick_cpu done", 0x347);
557 static void __init smp_core99_setup_cpu(int cpu_nr)
561 core99_init_caches(cpu_nr);
564 do_openpic_setup_cpu();
568 extern void g5_phy_disable_cpu1(void);
570 /* If we didn't start the second CPU, we must take
573 if (machine_is_compatible("MacRISC4") &&
574 num_online_cpus() < 2)
575 g5_phy_disable_cpu1();
576 #endif /* CONFIG_POWER4 */
577 if (ppc_md.progress) ppc_md.progress("core99_setup_cpu 0 done", 0x349);
581 /* not __init, called in sleep/wakeup code */
582 void smp_core99_take_timebase(void)
586 /* tell the primary we're here */
590 /* wait for the primary to set pri_tb_hi/lo */
591 while (sec_tb_reset < 2)
594 /* set our stuff the same as the primary */
595 local_irq_save(flags);
597 set_tb(pri_tb_hi, pri_tb_lo);
598 last_jiffy_stamp(smp_processor_id()) = pri_tb_stamp;
601 /* tell the primary we're done */
604 local_irq_restore(flags);
607 /* not __init, called in sleep/wakeup code */
608 void smp_core99_give_timebase(void)
613 /* wait for the secondary to be in take_timebase */
614 for (t = 100000; t > 0 && !sec_tb_reset; --t)
617 printk(KERN_WARNING "Timeout waiting sync on second CPU\n");
621 /* freeze the timebase and read it */
622 /* disable interrupts so the timebase is disabled for the
623 shortest possible time */
624 local_irq_save(flags);
625 pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 4);
626 pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
628 pri_tb_hi = get_tbu();
629 pri_tb_lo = get_tbl();
630 pri_tb_stamp = last_jiffy_stamp(smp_processor_id());
633 /* tell the secondary we're ready */
637 /* wait for the secondary to have taken it */
638 for (t = 100000; t > 0 && sec_tb_reset; --t)
641 printk(KERN_WARNING "Timeout waiting sync(2) on second CPU\n");
643 smp_tb_synchronized = 1;
645 /* Now, restart the timebase by leaving the GPIO to an open collector */
646 pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 0);
647 pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
648 local_irq_restore(flags);
652 /* PowerSurge-style Macs */
653 struct smp_ops_t psurge_smp_ops __pmacdata = {
654 .message_pass = smp_psurge_message_pass,
655 .probe = smp_psurge_probe,
656 .kick_cpu = smp_psurge_kick_cpu,
657 .setup_cpu = smp_psurge_setup_cpu,
658 .give_timebase = smp_psurge_give_timebase,
659 .take_timebase = smp_psurge_take_timebase,
662 /* Core99 Macs (dual G4s) */
663 struct smp_ops_t core99_smp_ops __pmacdata = {
664 .message_pass = smp_openpic_message_pass,
665 .probe = smp_core99_probe,
666 .kick_cpu = smp_core99_kick_cpu,
667 .setup_cpu = smp_core99_setup_cpu,
668 .give_timebase = smp_core99_give_timebase,
669 .take_timebase = smp_core99_take_timebase,