xref: /openbmc/linux/arch/powerpc/platforms/85xx/smp.c (revision f7d84fa7)
1 /*
2  * Author: Andy Fleming <afleming@freescale.com>
3  * 	   Kumar Gala <galak@kernel.crashing.org>
4  *
5  * Copyright 2006-2008, 2011-2012, 2015 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/sched/hotplug.h>
16 #include <linux/init.h>
17 #include <linux/delay.h>
18 #include <linux/of.h>
19 #include <linux/kexec.h>
20 #include <linux/highmem.h>
21 #include <linux/cpu.h>
22 #include <linux/fsl/guts.h>
23 
24 #include <asm/machdep.h>
25 #include <asm/pgtable.h>
26 #include <asm/page.h>
27 #include <asm/mpic.h>
28 #include <asm/cacheflush.h>
29 #include <asm/dbell.h>
30 #include <asm/code-patching.h>
31 #include <asm/cputhreads.h>
32 #include <asm/fsl_pm.h>
33 
34 #include <sysdev/fsl_soc.h>
35 #include <sysdev/mpic.h>
36 #include "smp.h"
37 
38 struct epapr_spin_table {
39 	u32	addr_h;
40 	u32	addr_l;
41 	u32	r3_h;
42 	u32	r3_l;
43 	u32	reserved;
44 	u32	pir;
45 };
46 
47 #ifdef CONFIG_HOTPLUG_CPU
48 static u64 timebase;
49 static int tb_req;
50 static int tb_valid;
51 
52 static void mpc85xx_give_timebase(void)
53 {
54 	unsigned long flags;
55 
56 	local_irq_save(flags);
57 	hard_irq_disable();
58 
59 	while (!tb_req)
60 		barrier();
61 	tb_req = 0;
62 
63 	qoriq_pm_ops->freeze_time_base(true);
64 #ifdef CONFIG_PPC64
65 	/*
66 	 * e5500/e6500 have a workaround for erratum A-006958 in place
67 	 * that will reread the timebase until TBL is non-zero.
68 	 * That would be a bad thing when the timebase is frozen.
69 	 *
70 	 * Thus, we read it manually, and instead of checking that
71 	 * TBL is non-zero, we ensure that TB does not change.  We don't
72 	 * do that for the main mftb implementation, because it requires
73 	 * a scratch register
74 	 */
75 	{
76 		u64 prev;
77 
78 		asm volatile("mfspr %0, %1" : "=r" (timebase) :
79 			     "i" (SPRN_TBRL));
80 
81 		do {
82 			prev = timebase;
83 			asm volatile("mfspr %0, %1" : "=r" (timebase) :
84 				     "i" (SPRN_TBRL));
85 		} while (prev != timebase);
86 	}
87 #else
88 	timebase = get_tb();
89 #endif
90 	mb();
91 	tb_valid = 1;
92 
93 	while (tb_valid)
94 		barrier();
95 
96 	qoriq_pm_ops->freeze_time_base(false);
97 
98 	local_irq_restore(flags);
99 }
100 
101 static void mpc85xx_take_timebase(void)
102 {
103 	unsigned long flags;
104 
105 	local_irq_save(flags);
106 	hard_irq_disable();
107 
108 	tb_req = 1;
109 	while (!tb_valid)
110 		barrier();
111 
112 	set_tb(timebase >> 32, timebase & 0xffffffff);
113 	isync();
114 	tb_valid = 0;
115 
116 	local_irq_restore(flags);
117 }
118 
119 static void smp_85xx_mach_cpu_die(void)
120 {
121 	unsigned int cpu = smp_processor_id();
122 
123 	local_irq_disable();
124 	hard_irq_disable();
125 	/* mask all irqs to prevent cpu wakeup */
126 	qoriq_pm_ops->irq_mask(cpu);
127 
128 	idle_task_exit();
129 
130 	mtspr(SPRN_TCR, 0);
131 	mtspr(SPRN_TSR, mfspr(SPRN_TSR));
132 
133 	generic_set_cpu_dead(cpu);
134 
135 	cur_cpu_spec->cpu_down_flush();
136 
137 	qoriq_pm_ops->cpu_die(cpu);
138 
139 	while (1)
140 		;
141 }
142 
143 static void qoriq_cpu_kill(unsigned int cpu)
144 {
145 	int i;
146 
147 	for (i = 0; i < 500; i++) {
148 		if (is_cpu_dead(cpu)) {
149 #ifdef CONFIG_PPC64
150 			paca[cpu].cpu_start = 0;
151 #endif
152 			return;
153 		}
154 		msleep(20);
155 	}
156 	pr_err("CPU%d didn't die...\n", cpu);
157 }
158 #endif
159 
160 /*
161  * To keep it compatible with old boot program which uses
162  * cache-inhibit spin table, we need to flush the cache
163  * before accessing spin table to invalidate any staled data.
164  * We also need to flush the cache after writing to spin
165  * table to push data out.
166  */
167 static inline void flush_spin_table(void *spin_table)
168 {
169 	flush_dcache_range((ulong)spin_table,
170 		(ulong)spin_table + sizeof(struct epapr_spin_table));
171 }
172 
173 static inline u32 read_spin_table_addr_l(void *spin_table)
174 {
175 	flush_dcache_range((ulong)spin_table,
176 		(ulong)spin_table + sizeof(struct epapr_spin_table));
177 	return in_be32(&((struct epapr_spin_table *)spin_table)->addr_l);
178 }
179 
180 #ifdef CONFIG_PPC64
181 static void wake_hw_thread(void *info)
182 {
183 	void fsl_secondary_thread_init(void);
184 	unsigned long inia;
185 	int cpu = *(const int *)info;
186 
187 	inia = *(unsigned long *)fsl_secondary_thread_init;
188 	book3e_start_thread(cpu_thread_in_core(cpu), inia);
189 }
190 #endif
191 
192 static int smp_85xx_start_cpu(int cpu)
193 {
194 	int ret = 0;
195 	struct device_node *np;
196 	const u64 *cpu_rel_addr;
197 	unsigned long flags;
198 	int ioremappable;
199 	int hw_cpu = get_hard_smp_processor_id(cpu);
200 	struct epapr_spin_table __iomem *spin_table;
201 
202 	np = of_get_cpu_node(cpu, NULL);
203 	cpu_rel_addr = of_get_property(np, "cpu-release-addr", NULL);
204 	if (!cpu_rel_addr) {
205 		pr_err("No cpu-release-addr for cpu %d\n", cpu);
206 		return -ENOENT;
207 	}
208 
209 	/*
210 	 * A secondary core could be in a spinloop in the bootpage
211 	 * (0xfffff000), somewhere in highmem, or somewhere in lowmem.
212 	 * The bootpage and highmem can be accessed via ioremap(), but
213 	 * we need to directly access the spinloop if its in lowmem.
214 	 */
215 	ioremappable = *cpu_rel_addr > virt_to_phys(high_memory);
216 
217 	/* Map the spin table */
218 	if (ioremappable)
219 		spin_table = ioremap_prot(*cpu_rel_addr,
220 			sizeof(struct epapr_spin_table), _PAGE_COHERENT);
221 	else
222 		spin_table = phys_to_virt(*cpu_rel_addr);
223 
224 	local_irq_save(flags);
225 	hard_irq_disable();
226 
227 	if (qoriq_pm_ops)
228 		qoriq_pm_ops->cpu_up_prepare(cpu);
229 
230 	/* if cpu is not spinning, reset it */
231 	if (read_spin_table_addr_l(spin_table) != 1) {
232 		/*
233 		 * We don't set the BPTR register here since it already points
234 		 * to the boot page properly.
235 		 */
236 		mpic_reset_core(cpu);
237 
238 		/*
239 		 * wait until core is ready...
240 		 * We need to invalidate the stale data, in case the boot
241 		 * loader uses a cache-inhibited spin table.
242 		 */
243 		if (!spin_event_timeout(
244 				read_spin_table_addr_l(spin_table) == 1,
245 				10000, 100)) {
246 			pr_err("timeout waiting for cpu %d to reset\n",
247 				hw_cpu);
248 			ret = -EAGAIN;
249 			goto err;
250 		}
251 	}
252 
253 	flush_spin_table(spin_table);
254 	out_be32(&spin_table->pir, hw_cpu);
255 #ifdef CONFIG_PPC64
256 	out_be64((u64 *)(&spin_table->addr_h),
257 		__pa(ppc_function_entry(generic_secondary_smp_init)));
258 #else
259 	out_be32(&spin_table->addr_l, __pa(__early_start));
260 #endif
261 	flush_spin_table(spin_table);
262 err:
263 	local_irq_restore(flags);
264 
265 	if (ioremappable)
266 		iounmap(spin_table);
267 
268 	return ret;
269 }
270 
271 static int smp_85xx_kick_cpu(int nr)
272 {
273 	int ret = 0;
274 #ifdef CONFIG_PPC64
275 	int primary = nr;
276 #endif
277 
278 	WARN_ON(nr < 0 || nr >= num_possible_cpus());
279 
280 	pr_debug("kick CPU #%d\n", nr);
281 
282 #ifdef CONFIG_PPC64
283 	if (threads_per_core == 2) {
284 		if (WARN_ON_ONCE(!cpu_has_feature(CPU_FTR_SMT)))
285 			return -ENOENT;
286 
287 		booting_thread_hwid = cpu_thread_in_core(nr);
288 		primary = cpu_first_thread_sibling(nr);
289 
290 		if (qoriq_pm_ops)
291 			qoriq_pm_ops->cpu_up_prepare(nr);
292 
293 		/*
294 		 * If either thread in the core is online, use it to start
295 		 * the other.
296 		 */
297 		if (cpu_online(primary)) {
298 			smp_call_function_single(primary,
299 					wake_hw_thread, &nr, 1);
300 			goto done;
301 		} else if (cpu_online(primary + 1)) {
302 			smp_call_function_single(primary + 1,
303 					wake_hw_thread, &nr, 1);
304 			goto done;
305 		}
306 
307 		/*
308 		 * If getting here, it means both threads in the core are
309 		 * offline. So start the primary thread, then it will start
310 		 * the thread specified in booting_thread_hwid, the one
311 		 * corresponding to nr.
312 		 */
313 
314 	} else if (threads_per_core == 1) {
315 		/*
316 		 * If one core has only one thread, set booting_thread_hwid to
317 		 * an invalid value.
318 		 */
319 		booting_thread_hwid = INVALID_THREAD_HWID;
320 
321 	} else if (threads_per_core > 2) {
322 		pr_err("Do not support more than 2 threads per CPU.");
323 		return -EINVAL;
324 	}
325 
326 	ret = smp_85xx_start_cpu(primary);
327 	if (ret)
328 		return ret;
329 
330 done:
331 	paca[nr].cpu_start = 1;
332 	generic_set_cpu_up(nr);
333 
334 	return ret;
335 #else
336 	ret = smp_85xx_start_cpu(nr);
337 	if (ret)
338 		return ret;
339 
340 	generic_set_cpu_up(nr);
341 
342 	return ret;
343 #endif
344 }
345 
346 struct smp_ops_t smp_85xx_ops = {
347 	.cause_nmi_ipi = NULL,
348 	.kick_cpu = smp_85xx_kick_cpu,
349 	.cpu_bootable = smp_generic_cpu_bootable,
350 #ifdef CONFIG_HOTPLUG_CPU
351 	.cpu_disable	= generic_cpu_disable,
352 	.cpu_die	= generic_cpu_die,
353 #endif
354 #if defined(CONFIG_KEXEC_CORE) && !defined(CONFIG_PPC64)
355 	.give_timebase	= smp_generic_give_timebase,
356 	.take_timebase	= smp_generic_take_timebase,
357 #endif
358 };
359 
360 #ifdef CONFIG_KEXEC_CORE
361 #ifdef CONFIG_PPC32
362 atomic_t kexec_down_cpus = ATOMIC_INIT(0);
363 
364 void mpc85xx_smp_kexec_cpu_down(int crash_shutdown, int secondary)
365 {
366 	local_irq_disable();
367 
368 	if (secondary) {
369 		cur_cpu_spec->cpu_down_flush();
370 		atomic_inc(&kexec_down_cpus);
371 		/* loop forever */
372 		while (1);
373 	}
374 }
375 
376 static void mpc85xx_smp_kexec_down(void *arg)
377 {
378 	if (ppc_md.kexec_cpu_down)
379 		ppc_md.kexec_cpu_down(0,1);
380 }
381 #else
382 void mpc85xx_smp_kexec_cpu_down(int crash_shutdown, int secondary)
383 {
384 	int cpu = smp_processor_id();
385 	int sibling = cpu_last_thread_sibling(cpu);
386 	bool notified = false;
387 	int disable_cpu;
388 	int disable_threadbit = 0;
389 	long start = mftb();
390 	long now;
391 
392 	local_irq_disable();
393 	hard_irq_disable();
394 	mpic_teardown_this_cpu(secondary);
395 
396 	if (cpu == crashing_cpu && cpu_thread_in_core(cpu) != 0) {
397 		/*
398 		 * We enter the crash kernel on whatever cpu crashed,
399 		 * even if it's a secondary thread.  If that's the case,
400 		 * disable the corresponding primary thread.
401 		 */
402 		disable_threadbit = 1;
403 		disable_cpu = cpu_first_thread_sibling(cpu);
404 	} else if (sibling != crashing_cpu &&
405 		   cpu_thread_in_core(cpu) == 0 &&
406 		   cpu_thread_in_core(sibling) != 0) {
407 		disable_threadbit = 2;
408 		disable_cpu = sibling;
409 	}
410 
411 	if (disable_threadbit) {
412 		while (paca[disable_cpu].kexec_state < KEXEC_STATE_REAL_MODE) {
413 			barrier();
414 			now = mftb();
415 			if (!notified && now - start > 1000000) {
416 				pr_info("%s/%d: waiting for cpu %d to enter KEXEC_STATE_REAL_MODE (%d)\n",
417 					__func__, smp_processor_id(),
418 					disable_cpu,
419 					paca[disable_cpu].kexec_state);
420 				notified = true;
421 			}
422 		}
423 
424 		if (notified) {
425 			pr_info("%s: cpu %d done waiting\n",
426 				__func__, disable_cpu);
427 		}
428 
429 		mtspr(SPRN_TENC, disable_threadbit);
430 		while (mfspr(SPRN_TENSR) & disable_threadbit)
431 			cpu_relax();
432 	}
433 }
434 #endif
435 
436 static void mpc85xx_smp_machine_kexec(struct kimage *image)
437 {
438 #ifdef CONFIG_PPC32
439 	int timeout = INT_MAX;
440 	int i, num_cpus = num_present_cpus();
441 
442 	if (image->type == KEXEC_TYPE_DEFAULT)
443 		smp_call_function(mpc85xx_smp_kexec_down, NULL, 0);
444 
445 	while ( (atomic_read(&kexec_down_cpus) != (num_cpus - 1)) &&
446 		( timeout > 0 ) )
447 	{
448 		timeout--;
449 	}
450 
451 	if ( !timeout )
452 		printk(KERN_ERR "Unable to bring down secondary cpu(s)");
453 
454 	for_each_online_cpu(i)
455 	{
456 		if ( i == smp_processor_id() ) continue;
457 		mpic_reset_core(i);
458 	}
459 #endif
460 
461 	default_machine_kexec(image);
462 }
463 #endif /* CONFIG_KEXEC_CORE */
464 
465 static void smp_85xx_setup_cpu(int cpu_nr)
466 {
467 	mpic_setup_this_cpu();
468 }
469 
470 void __init mpc85xx_smp_init(void)
471 {
472 	struct device_node *np;
473 
474 
475 	np = of_find_node_by_type(NULL, "open-pic");
476 	if (np) {
477 		smp_85xx_ops.probe = smp_mpic_probe;
478 		smp_85xx_ops.setup_cpu = smp_85xx_setup_cpu;
479 		smp_85xx_ops.message_pass = smp_mpic_message_pass;
480 	} else
481 		smp_85xx_ops.setup_cpu = NULL;
482 
483 	if (cpu_has_feature(CPU_FTR_DBELL)) {
484 		/*
485 		 * If left NULL, .message_pass defaults to
486 		 * smp_muxed_ipi_message_pass
487 		 */
488 		smp_85xx_ops.message_pass = NULL;
489 		smp_85xx_ops.cause_ipi = doorbell_global_ipi;
490 		smp_85xx_ops.probe = NULL;
491 	}
492 
493 #ifdef CONFIG_HOTPLUG_CPU
494 #ifdef CONFIG_FSL_CORENET_RCPM
495 	fsl_rcpm_init();
496 #endif
497 
498 #ifdef CONFIG_FSL_PMC
499 	mpc85xx_setup_pmc();
500 #endif
501 	if (qoriq_pm_ops) {
502 		smp_85xx_ops.give_timebase = mpc85xx_give_timebase;
503 		smp_85xx_ops.take_timebase = mpc85xx_take_timebase;
504 		ppc_md.cpu_die = smp_85xx_mach_cpu_die;
505 		smp_85xx_ops.cpu_die = qoriq_cpu_kill;
506 	}
507 #endif
508 	smp_ops = &smp_85xx_ops;
509 
510 #ifdef CONFIG_KEXEC_CORE
511 	ppc_md.kexec_cpu_down = mpc85xx_smp_kexec_cpu_down;
512 	ppc_md.machine_kexec = mpc85xx_smp_machine_kexec;
513 #endif
514 }
515