xref: /linux/arch/powerpc/platforms/85xx/smp.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /*
2  * Author: Andy Fleming <afleming@freescale.com>
3  * 	   Kumar Gala <galak@kernel.crashing.org>
4  *
5  * Copyright 2006-2008, 2011-2012 Freescale Semiconductor Inc.
6  *
7  * This program is free software; you can redistribute  it and/or modify it
8  * under  the terms of  the GNU General  Public License as published by the
9  * Free Software Foundation;  either version 2 of the  License, or (at your
10  * option) any later version.
11  */
12 
13 #include <linux/stddef.h>
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/delay.h>
17 #include <linux/of.h>
18 #include <linux/of_address.h>
19 #include <linux/kexec.h>
20 #include <linux/highmem.h>
21 #include <linux/cpu.h>
22 
23 #include <asm/machdep.h>
24 #include <asm/pgtable.h>
25 #include <asm/page.h>
26 #include <asm/mpic.h>
27 #include <asm/cacheflush.h>
28 #include <asm/dbell.h>
29 #include <asm/fsl_guts.h>
30 #include <asm/code-patching.h>
31 #include <asm/cputhreads.h>
32 
33 #include <sysdev/fsl_soc.h>
34 #include <sysdev/mpic.h>
35 #include "smp.h"
36 
37 struct epapr_spin_table {
38 	u32	addr_h;
39 	u32	addr_l;
40 	u32	r3_h;
41 	u32	r3_l;
42 	u32	reserved;
43 	u32	pir;
44 };
45 
46 static struct ccsr_guts __iomem *guts;
47 static u64 timebase;
48 static int tb_req;
49 static int tb_valid;
50 
51 static void mpc85xx_timebase_freeze(int freeze)
52 {
53 	uint32_t mask;
54 
55 	mask = CCSR_GUTS_DEVDISR_TB0 | CCSR_GUTS_DEVDISR_TB1;
56 	if (freeze)
57 		setbits32(&guts->devdisr, mask);
58 	else
59 		clrbits32(&guts->devdisr, mask);
60 
61 	in_be32(&guts->devdisr);
62 }
63 
64 static void mpc85xx_give_timebase(void)
65 {
66 	unsigned long flags;
67 
68 	local_irq_save(flags);
69 
70 	while (!tb_req)
71 		barrier();
72 	tb_req = 0;
73 
74 	mpc85xx_timebase_freeze(1);
75 #ifdef CONFIG_PPC64
76 	/*
77 	 * e5500/e6500 have a workaround for erratum A-006958 in place
78 	 * that will reread the timebase until TBL is non-zero.
79 	 * That would be a bad thing when the timebase is frozen.
80 	 *
81 	 * Thus, we read it manually, and instead of checking that
82 	 * TBL is non-zero, we ensure that TB does not change.  We don't
83 	 * do that for the main mftb implementation, because it requires
84 	 * a scratch register
85 	 */
86 	{
87 		u64 prev;
88 
89 		asm volatile("mfspr %0, %1" : "=r" (timebase) :
90 			     "i" (SPRN_TBRL));
91 
92 		do {
93 			prev = timebase;
94 			asm volatile("mfspr %0, %1" : "=r" (timebase) :
95 				     "i" (SPRN_TBRL));
96 		} while (prev != timebase);
97 	}
98 #else
99 	timebase = get_tb();
100 #endif
101 	mb();
102 	tb_valid = 1;
103 
104 	while (tb_valid)
105 		barrier();
106 
107 	mpc85xx_timebase_freeze(0);
108 
109 	local_irq_restore(flags);
110 }
111 
112 static void mpc85xx_take_timebase(void)
113 {
114 	unsigned long flags;
115 
116 	local_irq_save(flags);
117 
118 	tb_req = 1;
119 	while (!tb_valid)
120 		barrier();
121 
122 	set_tb(timebase >> 32, timebase & 0xffffffff);
123 	isync();
124 	tb_valid = 0;
125 
126 	local_irq_restore(flags);
127 }
128 
129 #ifdef CONFIG_HOTPLUG_CPU
130 static void smp_85xx_mach_cpu_die(void)
131 {
132 	unsigned int cpu = smp_processor_id();
133 	u32 tmp;
134 
135 	local_irq_disable();
136 	idle_task_exit();
137 	generic_set_cpu_dead(cpu);
138 	mb();
139 
140 	mtspr(SPRN_TCR, 0);
141 
142 	__flush_disable_L1();
143 	tmp = (mfspr(SPRN_HID0) & ~(HID0_DOZE|HID0_SLEEP)) | HID0_NAP;
144 	mtspr(SPRN_HID0, tmp);
145 	isync();
146 
147 	/* Enter NAP mode. */
148 	tmp = mfmsr();
149 	tmp |= MSR_WE;
150 	mb();
151 	mtmsr(tmp);
152 	isync();
153 
154 	while (1)
155 		;
156 }
157 #endif
158 
159 static inline void flush_spin_table(void *spin_table)
160 {
161 	flush_dcache_range((ulong)spin_table,
162 		(ulong)spin_table + sizeof(struct epapr_spin_table));
163 }
164 
165 static inline u32 read_spin_table_addr_l(void *spin_table)
166 {
167 	flush_dcache_range((ulong)spin_table,
168 		(ulong)spin_table + sizeof(struct epapr_spin_table));
169 	return in_be32(&((struct epapr_spin_table *)spin_table)->addr_l);
170 }
171 
172 #ifdef CONFIG_PPC64
173 static void wake_hw_thread(void *info)
174 {
175 	void fsl_secondary_thread_init(void);
176 	unsigned long imsr1, inia1;
177 	int nr = *(const int *)info;
178 
179 	imsr1 = MSR_KERNEL;
180 	inia1 = *(unsigned long *)fsl_secondary_thread_init;
181 
182 	mttmr(TMRN_IMSR1, imsr1);
183 	mttmr(TMRN_INIA1, inia1);
184 	mtspr(SPRN_TENS, TEN_THREAD(1));
185 
186 	smp_generic_kick_cpu(nr);
187 }
188 #endif
189 
190 static int smp_85xx_kick_cpu(int nr)
191 {
192 	unsigned long flags;
193 	const u64 *cpu_rel_addr;
194 	__iomem struct epapr_spin_table *spin_table;
195 	struct device_node *np;
196 	int hw_cpu = get_hard_smp_processor_id(nr);
197 	int ioremappable;
198 	int ret = 0;
199 
200 	WARN_ON(nr < 0 || nr >= NR_CPUS);
201 	WARN_ON(hw_cpu < 0 || hw_cpu >= NR_CPUS);
202 
203 	pr_debug("smp_85xx_kick_cpu: kick CPU #%d\n", nr);
204 
205 #ifdef CONFIG_PPC64
206 	/* Threads don't use the spin table */
207 	if (cpu_thread_in_core(nr) != 0) {
208 		int primary = cpu_first_thread_sibling(nr);
209 
210 		if (WARN_ON_ONCE(!cpu_has_feature(CPU_FTR_SMT)))
211 			return -ENOENT;
212 
213 		if (cpu_thread_in_core(nr) != 1) {
214 			pr_err("%s: cpu %d: invalid hw thread %d\n",
215 			       __func__, nr, cpu_thread_in_core(nr));
216 			return -ENOENT;
217 		}
218 
219 		if (!cpu_online(primary)) {
220 			pr_err("%s: cpu %d: primary %d not online\n",
221 			       __func__, nr, primary);
222 			return -ENOENT;
223 		}
224 
225 		smp_call_function_single(primary, wake_hw_thread, &nr, 0);
226 		return 0;
227 	}
228 #endif
229 
230 	np = of_get_cpu_node(nr, NULL);
231 	cpu_rel_addr = of_get_property(np, "cpu-release-addr", NULL);
232 
233 	if (cpu_rel_addr == NULL) {
234 		printk(KERN_ERR "No cpu-release-addr for cpu %d\n", nr);
235 		return -ENOENT;
236 	}
237 
238 	/*
239 	 * A secondary core could be in a spinloop in the bootpage
240 	 * (0xfffff000), somewhere in highmem, or somewhere in lowmem.
241 	 * The bootpage and highmem can be accessed via ioremap(), but
242 	 * we need to directly access the spinloop if its in lowmem.
243 	 */
244 	ioremappable = *cpu_rel_addr > virt_to_phys(high_memory);
245 
246 	/* Map the spin table */
247 	if (ioremappable)
248 		spin_table = ioremap_prot(*cpu_rel_addr,
249 			sizeof(struct epapr_spin_table), _PAGE_COHERENT);
250 	else
251 		spin_table = phys_to_virt(*cpu_rel_addr);
252 
253 	local_irq_save(flags);
254 #ifdef CONFIG_PPC32
255 #ifdef CONFIG_HOTPLUG_CPU
256 	/* Corresponding to generic_set_cpu_dead() */
257 	generic_set_cpu_up(nr);
258 
259 	if (system_state == SYSTEM_RUNNING) {
260 		/*
261 		 * To keep it compatible with old boot program which uses
262 		 * cache-inhibit spin table, we need to flush the cache
263 		 * before accessing spin table to invalidate any staled data.
264 		 * We also need to flush the cache after writing to spin
265 		 * table to push data out.
266 		 */
267 		flush_spin_table(spin_table);
268 		out_be32(&spin_table->addr_l, 0);
269 		flush_spin_table(spin_table);
270 
271 		/*
272 		 * We don't set the BPTR register here since it already points
273 		 * to the boot page properly.
274 		 */
275 		mpic_reset_core(nr);
276 
277 		/*
278 		 * wait until core is ready...
279 		 * We need to invalidate the stale data, in case the boot
280 		 * loader uses a cache-inhibited spin table.
281 		 */
282 		if (!spin_event_timeout(
283 				read_spin_table_addr_l(spin_table) == 1,
284 				10000, 100)) {
285 			pr_err("%s: timeout waiting for core %d to reset\n",
286 							__func__, hw_cpu);
287 			ret = -ENOENT;
288 			goto out;
289 		}
290 
291 		/*  clear the acknowledge status */
292 		__secondary_hold_acknowledge = -1;
293 	}
294 #endif
295 	flush_spin_table(spin_table);
296 	out_be32(&spin_table->pir, hw_cpu);
297 	out_be32(&spin_table->addr_l, __pa(__early_start));
298 	flush_spin_table(spin_table);
299 
300 	/* Wait a bit for the CPU to ack. */
301 	if (!spin_event_timeout(__secondary_hold_acknowledge == hw_cpu,
302 					10000, 100)) {
303 		pr_err("%s: timeout waiting for core %d to ack\n",
304 						__func__, hw_cpu);
305 		ret = -ENOENT;
306 		goto out;
307 	}
308 out:
309 #else
310 	smp_generic_kick_cpu(nr);
311 
312 	flush_spin_table(spin_table);
313 	out_be32(&spin_table->pir, hw_cpu);
314 	out_be64((u64 *)(&spin_table->addr_h),
315 		__pa(ppc_function_entry(generic_secondary_smp_init)));
316 	flush_spin_table(spin_table);
317 #endif
318 
319 	local_irq_restore(flags);
320 
321 	if (ioremappable)
322 		iounmap(spin_table);
323 
324 	return ret;
325 }
326 
327 struct smp_ops_t smp_85xx_ops = {
328 	.kick_cpu = smp_85xx_kick_cpu,
329 	.cpu_bootable = smp_generic_cpu_bootable,
330 #ifdef CONFIG_HOTPLUG_CPU
331 	.cpu_disable	= generic_cpu_disable,
332 	.cpu_die	= generic_cpu_die,
333 #endif
334 #ifdef CONFIG_KEXEC
335 	.give_timebase	= smp_generic_give_timebase,
336 	.take_timebase	= smp_generic_take_timebase,
337 #endif
338 };
339 
340 #ifdef CONFIG_KEXEC
341 atomic_t kexec_down_cpus = ATOMIC_INIT(0);
342 
343 void mpc85xx_smp_kexec_cpu_down(int crash_shutdown, int secondary)
344 {
345 	local_irq_disable();
346 
347 	if (secondary) {
348 		atomic_inc(&kexec_down_cpus);
349 		/* loop forever */
350 		while (1);
351 	}
352 }
353 
354 static void mpc85xx_smp_kexec_down(void *arg)
355 {
356 	if (ppc_md.kexec_cpu_down)
357 		ppc_md.kexec_cpu_down(0,1);
358 }
359 
360 static void map_and_flush(unsigned long paddr)
361 {
362 	struct page *page = pfn_to_page(paddr >> PAGE_SHIFT);
363 	unsigned long kaddr  = (unsigned long)kmap(page);
364 
365 	flush_dcache_range(kaddr, kaddr + PAGE_SIZE);
366 	kunmap(page);
367 }
368 
369 /**
370  * Before we reset the other cores, we need to flush relevant cache
371  * out to memory so we don't get anything corrupted, some of these flushes
372  * are performed out of an overabundance of caution as interrupts are not
373  * disabled yet and we can switch cores
374  */
375 static void mpc85xx_smp_flush_dcache_kexec(struct kimage *image)
376 {
377 	kimage_entry_t *ptr, entry;
378 	unsigned long paddr;
379 	int i;
380 
381 	if (image->type == KEXEC_TYPE_DEFAULT) {
382 		/* normal kexec images are stored in temporary pages */
383 		for (ptr = &image->head; (entry = *ptr) && !(entry & IND_DONE);
384 		     ptr = (entry & IND_INDIRECTION) ?
385 				phys_to_virt(entry & PAGE_MASK) : ptr + 1) {
386 			if (!(entry & IND_DESTINATION)) {
387 				map_and_flush(entry);
388 			}
389 		}
390 		/* flush out last IND_DONE page */
391 		map_and_flush(entry);
392 	} else {
393 		/* crash type kexec images are copied to the crash region */
394 		for (i = 0; i < image->nr_segments; i++) {
395 			struct kexec_segment *seg = &image->segment[i];
396 			for (paddr = seg->mem; paddr < seg->mem + seg->memsz;
397 			     paddr += PAGE_SIZE) {
398 				map_and_flush(paddr);
399 			}
400 		}
401 	}
402 
403 	/* also flush the kimage struct to be passed in as well */
404 	flush_dcache_range((unsigned long)image,
405 			   (unsigned long)image + sizeof(*image));
406 }
407 
408 static void mpc85xx_smp_machine_kexec(struct kimage *image)
409 {
410 	int timeout = INT_MAX;
411 	int i, num_cpus = num_present_cpus();
412 
413 	mpc85xx_smp_flush_dcache_kexec(image);
414 
415 	if (image->type == KEXEC_TYPE_DEFAULT)
416 		smp_call_function(mpc85xx_smp_kexec_down, NULL, 0);
417 
418 	while ( (atomic_read(&kexec_down_cpus) != (num_cpus - 1)) &&
419 		( timeout > 0 ) )
420 	{
421 		timeout--;
422 	}
423 
424 	if ( !timeout )
425 		printk(KERN_ERR "Unable to bring down secondary cpu(s)");
426 
427 	for_each_online_cpu(i)
428 	{
429 		if ( i == smp_processor_id() ) continue;
430 		mpic_reset_core(i);
431 	}
432 
433 	default_machine_kexec(image);
434 }
435 #endif /* CONFIG_KEXEC */
436 
437 static void smp_85xx_basic_setup(int cpu_nr)
438 {
439 	if (cpu_has_feature(CPU_FTR_DBELL))
440 		doorbell_setup_this_cpu();
441 }
442 
443 static void smp_85xx_setup_cpu(int cpu_nr)
444 {
445 	mpic_setup_this_cpu();
446 	smp_85xx_basic_setup(cpu_nr);
447 }
448 
449 static const struct of_device_id mpc85xx_smp_guts_ids[] = {
450 	{ .compatible = "fsl,mpc8572-guts", },
451 	{ .compatible = "fsl,p1020-guts", },
452 	{ .compatible = "fsl,p1021-guts", },
453 	{ .compatible = "fsl,p1022-guts", },
454 	{ .compatible = "fsl,p1023-guts", },
455 	{ .compatible = "fsl,p2020-guts", },
456 	{},
457 };
458 
459 void __init mpc85xx_smp_init(void)
460 {
461 	struct device_node *np;
462 
463 
464 	np = of_find_node_by_type(NULL, "open-pic");
465 	if (np) {
466 		smp_85xx_ops.probe = smp_mpic_probe;
467 		smp_85xx_ops.setup_cpu = smp_85xx_setup_cpu;
468 		smp_85xx_ops.message_pass = smp_mpic_message_pass;
469 	} else
470 		smp_85xx_ops.setup_cpu = smp_85xx_basic_setup;
471 
472 	if (cpu_has_feature(CPU_FTR_DBELL)) {
473 		/*
474 		 * If left NULL, .message_pass defaults to
475 		 * smp_muxed_ipi_message_pass
476 		 */
477 		smp_85xx_ops.message_pass = NULL;
478 		smp_85xx_ops.cause_ipi = doorbell_cause_ipi;
479 		smp_85xx_ops.probe = NULL;
480 	}
481 
482 	np = of_find_matching_node(NULL, mpc85xx_smp_guts_ids);
483 	if (np) {
484 		guts = of_iomap(np, 0);
485 		of_node_put(np);
486 		if (!guts) {
487 			pr_err("%s: Could not map guts node address\n",
488 								__func__);
489 			return;
490 		}
491 		smp_85xx_ops.give_timebase = mpc85xx_give_timebase;
492 		smp_85xx_ops.take_timebase = mpc85xx_take_timebase;
493 #ifdef CONFIG_HOTPLUG_CPU
494 		ppc_md.cpu_die = smp_85xx_mach_cpu_die;
495 #endif
496 	}
497 
498 	smp_ops = &smp_85xx_ops;
499 
500 #ifdef CONFIG_KEXEC
501 	ppc_md.kexec_cpu_down = mpc85xx_smp_kexec_cpu_down;
502 	ppc_md.machine_kexec = mpc85xx_smp_machine_kexec;
503 #endif
504 }
505