xref: /openbmc/linux/arch/arm/mach-sunxi/mc_smp.c (revision ff40b576)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2018 Chen-Yu Tsai
4  *
5  * Chen-Yu Tsai <wens@csie.org>
6  *
7  * arch/arm/mach-sunxi/mc_smp.c
8  *
9  * Based on Allwinner code, arch/arm/mach-exynos/mcpm-exynos.c, and
10  * arch/arm/mach-hisi/platmcpm.c
11  * Cluster cache enable trampoline code adapted from MCPM framework
12  */
13 
14 #include <linux/arm-cci.h>
15 #include <linux/cpu_pm.h>
16 #include <linux/delay.h>
17 #include <linux/io.h>
18 #include <linux/iopoll.h>
19 #include <linux/irqchip/arm-gic.h>
20 #include <linux/of.h>
21 #include <linux/of_address.h>
22 #include <linux/smp.h>
23 
24 #include <asm/cacheflush.h>
25 #include <asm/cp15.h>
26 #include <asm/cputype.h>
27 #include <asm/idmap.h>
28 #include <asm/smp_plat.h>
29 #include <asm/suspend.h>
30 
31 #define SUNXI_CPUS_PER_CLUSTER		4
32 #define SUNXI_NR_CLUSTERS		2
33 
34 #define POLL_USEC	100
35 #define TIMEOUT_USEC	100000
36 
37 #define CPUCFG_CX_CTRL_REG0(c)		(0x10 * (c))
38 #define CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE(n)	BIT(n)
39 #define CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE_ALL	0xf
40 #define CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A7	BIT(4)
41 #define CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A15	BIT(0)
42 #define CPUCFG_CX_CTRL_REG1(c)		(0x10 * (c) + 0x4)
43 #define CPUCFG_CX_CTRL_REG1_ACINACTM	BIT(0)
44 #define CPUCFG_CX_STATUS(c)		(0x30 + 0x4 * (c))
45 #define CPUCFG_CX_STATUS_STANDBYWFI(n)	BIT(16 + (n))
46 #define CPUCFG_CX_STATUS_STANDBYWFIL2	BIT(0)
47 #define CPUCFG_CX_RST_CTRL(c)		(0x80 + 0x4 * (c))
48 #define CPUCFG_CX_RST_CTRL_DBG_SOC_RST	BIT(24)
49 #define CPUCFG_CX_RST_CTRL_ETM_RST(n)	BIT(20 + (n))
50 #define CPUCFG_CX_RST_CTRL_ETM_RST_ALL	(0xf << 20)
51 #define CPUCFG_CX_RST_CTRL_DBG_RST(n)	BIT(16 + (n))
52 #define CPUCFG_CX_RST_CTRL_DBG_RST_ALL	(0xf << 16)
53 #define CPUCFG_CX_RST_CTRL_H_RST	BIT(12)
54 #define CPUCFG_CX_RST_CTRL_L2_RST	BIT(8)
55 #define CPUCFG_CX_RST_CTRL_CX_RST(n)	BIT(4 + (n))
56 #define CPUCFG_CX_RST_CTRL_CORE_RST(n)	BIT(n)
57 #define CPUCFG_CX_RST_CTRL_CORE_RST_ALL	(0xf << 0)
58 
59 #define PRCM_CPU_PO_RST_CTRL(c)		(0x4 + 0x4 * (c))
60 #define PRCM_CPU_PO_RST_CTRL_CORE(n)	BIT(n)
61 #define PRCM_CPU_PO_RST_CTRL_CORE_ALL	0xf
62 #define PRCM_PWROFF_GATING_REG(c)	(0x100 + 0x4 * (c))
63 /* The power off register for clusters are different from a80 and a83t */
64 #define PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I	BIT(0)
65 #define PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I	BIT(4)
66 #define PRCM_PWROFF_GATING_REG_CORE(n)	BIT(n)
67 #define PRCM_PWR_SWITCH_REG(c, cpu)	(0x140 + 0x10 * (c) + 0x4 * (cpu))
68 #define PRCM_CPU_SOFT_ENTRY_REG		0x164
69 
70 /* R_CPUCFG registers, specific to sun8i-a83t */
71 #define R_CPUCFG_CLUSTER_PO_RST_CTRL(c)	(0x30 + (c) * 0x4)
72 #define R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(n)	BIT(n)
73 #define R_CPUCFG_CPU_SOFT_ENTRY_REG		0x01a4
74 
75 #define CPU0_SUPPORT_HOTPLUG_MAGIC0	0xFA50392F
76 #define CPU0_SUPPORT_HOTPLUG_MAGIC1	0x790DCA3A
77 
78 static void __iomem *cpucfg_base;
79 static void __iomem *prcm_base;
80 static void __iomem *sram_b_smp_base;
81 static void __iomem *r_cpucfg_base;
82 
83 extern void sunxi_mc_smp_secondary_startup(void);
84 extern void sunxi_mc_smp_resume(void);
85 static bool is_a83t;
86 
87 static bool sunxi_core_is_cortex_a15(unsigned int core, unsigned int cluster)
88 {
89 	struct device_node *node;
90 	int cpu = cluster * SUNXI_CPUS_PER_CLUSTER + core;
91 	bool is_compatible;
92 
93 	node = of_cpu_device_node_get(cpu);
94 
95 	/* In case of_cpu_device_node_get fails */
96 	if (!node)
97 		node = of_get_cpu_node(cpu, NULL);
98 
99 	if (!node) {
100 		/*
101 		 * There's no point in returning an error, since we
102 		 * would be mid way in a core or cluster power sequence.
103 		 */
104 		pr_err("%s: Couldn't get CPU cluster %u core %u device node\n",
105 		       __func__, cluster, core);
106 
107 		return false;
108 	}
109 
110 	is_compatible = of_device_is_compatible(node, "arm,cortex-a15");
111 	of_node_put(node);
112 	return is_compatible;
113 }
114 
115 static int sunxi_cpu_power_switch_set(unsigned int cpu, unsigned int cluster,
116 				      bool enable)
117 {
118 	u32 reg;
119 
120 	/* control sequence from Allwinner A80 user manual v1.2 PRCM section */
121 	reg = readl(prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
122 	if (enable) {
123 		if (reg == 0x00) {
124 			pr_debug("power clamp for cluster %u cpu %u already open\n",
125 				 cluster, cpu);
126 			return 0;
127 		}
128 
129 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
130 		udelay(10);
131 		writel(0xfe, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
132 		udelay(10);
133 		writel(0xf8, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
134 		udelay(10);
135 		writel(0xf0, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
136 		udelay(10);
137 		writel(0x00, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
138 		udelay(10);
139 	} else {
140 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
141 		udelay(10);
142 	}
143 
144 	return 0;
145 }
146 
147 static void sunxi_cpu0_hotplug_support_set(bool enable)
148 {
149 	if (enable) {
150 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC0, sram_b_smp_base);
151 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC1, sram_b_smp_base + 0x4);
152 	} else {
153 		writel(0x0, sram_b_smp_base);
154 		writel(0x0, sram_b_smp_base + 0x4);
155 	}
156 }
157 
158 static int sunxi_cpu_powerup(unsigned int cpu, unsigned int cluster)
159 {
160 	u32 reg;
161 
162 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
163 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
164 		return -EINVAL;
165 
166 	/* Set hotplug support magic flags for cpu0 */
167 	if (cluster == 0 && cpu == 0)
168 		sunxi_cpu0_hotplug_support_set(true);
169 
170 	/* assert processor power-on reset */
171 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
172 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE(cpu);
173 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
174 
175 	if (is_a83t) {
176 		/* assert cpu power-on reset */
177 		reg  = readl(r_cpucfg_base +
178 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
179 		reg &= ~(R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(cpu));
180 		writel(reg, r_cpucfg_base +
181 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
182 		udelay(10);
183 	}
184 
185 	/* Cortex-A7: hold L1 reset disable signal low */
186 	if (!sunxi_core_is_cortex_a15(cpu, cluster)) {
187 		reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
188 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE(cpu);
189 		writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
190 	}
191 
192 	/* assert processor related resets */
193 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
194 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
195 
196 	/*
197 	 * Allwinner code also asserts resets for NEON on A15. According
198 	 * to ARM manuals, asserting power-on reset is sufficient.
199 	 */
200 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
201 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
202 
203 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
204 
205 	/* open power switch */
206 	sunxi_cpu_power_switch_set(cpu, cluster, true);
207 
208 	/* Handle A83T bit swap */
209 	if (is_a83t) {
210 		if (cpu == 0)
211 			cpu = 4;
212 	}
213 
214 	/* clear processor power gate */
215 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
216 	reg &= ~PRCM_PWROFF_GATING_REG_CORE(cpu);
217 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
218 	udelay(20);
219 
220 	/* Handle A83T bit swap */
221 	if (is_a83t) {
222 		if (cpu == 4)
223 			cpu = 0;
224 	}
225 
226 	/* de-assert processor power-on reset */
227 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
228 	reg |= PRCM_CPU_PO_RST_CTRL_CORE(cpu);
229 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
230 
231 	if (is_a83t) {
232 		reg  = readl(r_cpucfg_base +
233 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
234 		reg |= R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(cpu);
235 		writel(reg, r_cpucfg_base +
236 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
237 		udelay(10);
238 	}
239 
240 	/* de-assert all processor resets */
241 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
242 	reg |= CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
243 	reg |= CPUCFG_CX_RST_CTRL_CORE_RST(cpu);
244 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
245 		reg |= CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
246 	else
247 		reg |= CPUCFG_CX_RST_CTRL_CX_RST(cpu); /* NEON */
248 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
249 
250 	return 0;
251 }
252 
253 static int sunxi_cluster_powerup(unsigned int cluster)
254 {
255 	u32 reg;
256 
257 	pr_debug("%s: cluster %u\n", __func__, cluster);
258 	if (cluster >= SUNXI_NR_CLUSTERS)
259 		return -EINVAL;
260 
261 	/* For A83T, assert cluster cores resets */
262 	if (is_a83t) {
263 		reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
264 		reg &= ~CPUCFG_CX_RST_CTRL_CORE_RST_ALL;   /* Core Reset    */
265 		writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
266 		udelay(10);
267 	}
268 
269 	/* assert ACINACTM */
270 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
271 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
272 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
273 
274 	/* assert cluster processor power-on resets */
275 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
276 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE_ALL;
277 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
278 
279 	/* assert cluster cores resets */
280 	if (is_a83t) {
281 		reg  = readl(r_cpucfg_base +
282 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
283 		reg &= ~CPUCFG_CX_RST_CTRL_CORE_RST_ALL;
284 		writel(reg, r_cpucfg_base +
285 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
286 		udelay(10);
287 	}
288 
289 	/* assert cluster resets */
290 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
291 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
292 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST_ALL;
293 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
294 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
295 
296 	/*
297 	 * Allwinner code also asserts resets for NEON on A15. According
298 	 * to ARM manuals, asserting power-on reset is sufficient.
299 	 */
300 	if (!sunxi_core_is_cortex_a15(0, cluster))
301 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST_ALL;
302 
303 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
304 
305 	/* hold L1/L2 reset disable signals low */
306 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
307 	if (sunxi_core_is_cortex_a15(0, cluster)) {
308 		/* Cortex-A15: hold L2RSTDISABLE low */
309 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A15;
310 	} else {
311 		/* Cortex-A7: hold L1RSTDISABLE and L2RSTDISABLE low */
312 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE_ALL;
313 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A7;
314 	}
315 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
316 
317 	/* clear cluster power gate */
318 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
319 	if (is_a83t)
320 		reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
321 	else
322 		reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
323 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
324 	udelay(20);
325 
326 	/* de-assert cluster resets */
327 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
328 	reg |= CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
329 	reg |= CPUCFG_CX_RST_CTRL_H_RST;
330 	reg |= CPUCFG_CX_RST_CTRL_L2_RST;
331 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
332 
333 	/* de-assert ACINACTM */
334 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
335 	reg &= ~CPUCFG_CX_CTRL_REG1_ACINACTM;
336 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
337 
338 	return 0;
339 }
340 
341 /*
342  * This bit is shared between the initial nocache_trampoline call to
343  * enable CCI-400 and proper cluster cache disable before power down.
344  */
345 static void sunxi_cluster_cache_disable_without_axi(void)
346 {
347 	if (read_cpuid_part() == ARM_CPU_PART_CORTEX_A15) {
348 		/*
349 		 * On the Cortex-A15 we need to disable
350 		 * L2 prefetching before flushing the cache.
351 		 */
352 		asm volatile(
353 		"mcr	p15, 1, %0, c15, c0, 3\n"
354 		"isb\n"
355 		"dsb"
356 		: : "r" (0x400));
357 	}
358 
359 	/* Flush all cache levels for this cluster. */
360 	v7_exit_coherency_flush(all);
361 
362 	/*
363 	 * Disable cluster-level coherency by masking
364 	 * incoming snoops and DVM messages:
365 	 */
366 	cci_disable_port_by_cpu(read_cpuid_mpidr());
367 }
368 
369 static int sunxi_mc_smp_cpu_table[SUNXI_NR_CLUSTERS][SUNXI_CPUS_PER_CLUSTER];
370 int sunxi_mc_smp_first_comer;
371 
372 static DEFINE_SPINLOCK(boot_lock);
373 
374 static bool sunxi_mc_smp_cluster_is_down(unsigned int cluster)
375 {
376 	int i;
377 
378 	for (i = 0; i < SUNXI_CPUS_PER_CLUSTER; i++)
379 		if (sunxi_mc_smp_cpu_table[cluster][i])
380 			return false;
381 	return true;
382 }
383 
384 static void sunxi_mc_smp_secondary_init(unsigned int cpu)
385 {
386 	/* Clear hotplug support magic flags for cpu0 */
387 	if (cpu == 0)
388 		sunxi_cpu0_hotplug_support_set(false);
389 }
390 
391 static int sunxi_mc_smp_boot_secondary(unsigned int l_cpu, struct task_struct *idle)
392 {
393 	unsigned int mpidr, cpu, cluster;
394 
395 	mpidr = cpu_logical_map(l_cpu);
396 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
397 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
398 
399 	if (!cpucfg_base)
400 		return -ENODEV;
401 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER)
402 		return -EINVAL;
403 
404 	spin_lock_irq(&boot_lock);
405 
406 	if (sunxi_mc_smp_cpu_table[cluster][cpu])
407 		goto out;
408 
409 	if (sunxi_mc_smp_cluster_is_down(cluster)) {
410 		sunxi_mc_smp_first_comer = true;
411 		sunxi_cluster_powerup(cluster);
412 	} else {
413 		sunxi_mc_smp_first_comer = false;
414 	}
415 
416 	/* This is read by incoming CPUs with their cache and MMU disabled */
417 	sync_cache_w(&sunxi_mc_smp_first_comer);
418 	sunxi_cpu_powerup(cpu, cluster);
419 
420 out:
421 	sunxi_mc_smp_cpu_table[cluster][cpu]++;
422 	spin_unlock_irq(&boot_lock);
423 
424 	return 0;
425 }
426 
427 #ifdef CONFIG_HOTPLUG_CPU
428 static void sunxi_cluster_cache_disable(void)
429 {
430 	unsigned int cluster = MPIDR_AFFINITY_LEVEL(read_cpuid_mpidr(), 1);
431 	u32 reg;
432 
433 	pr_debug("%s: cluster %u\n", __func__, cluster);
434 
435 	sunxi_cluster_cache_disable_without_axi();
436 
437 	/* last man standing, assert ACINACTM */
438 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
439 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
440 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
441 }
442 
443 static void sunxi_mc_smp_cpu_die(unsigned int l_cpu)
444 {
445 	unsigned int mpidr, cpu, cluster;
446 	bool last_man;
447 
448 	mpidr = cpu_logical_map(l_cpu);
449 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
450 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
451 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
452 
453 	spin_lock(&boot_lock);
454 	sunxi_mc_smp_cpu_table[cluster][cpu]--;
455 	if (sunxi_mc_smp_cpu_table[cluster][cpu] == 1) {
456 		/* A power_up request went ahead of us. */
457 		pr_debug("%s: aborting due to a power up request\n",
458 			 __func__);
459 		spin_unlock(&boot_lock);
460 		return;
461 	} else if (sunxi_mc_smp_cpu_table[cluster][cpu] > 1) {
462 		pr_err("Cluster %d CPU%d boots multiple times\n",
463 		       cluster, cpu);
464 		BUG();
465 	}
466 
467 	last_man = sunxi_mc_smp_cluster_is_down(cluster);
468 	spin_unlock(&boot_lock);
469 
470 	gic_cpu_if_down(0);
471 	if (last_man)
472 		sunxi_cluster_cache_disable();
473 	else
474 		v7_exit_coherency_flush(louis);
475 
476 	for (;;)
477 		wfi();
478 }
479 
480 static int sunxi_cpu_powerdown(unsigned int cpu, unsigned int cluster)
481 {
482 	u32 reg;
483 	int gating_bit = cpu;
484 
485 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
486 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
487 		return -EINVAL;
488 
489 	if (is_a83t && cpu == 0)
490 		gating_bit = 4;
491 
492 	/* gate processor power */
493 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
494 	reg |= PRCM_PWROFF_GATING_REG_CORE(gating_bit);
495 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
496 	udelay(20);
497 
498 	/* close power switch */
499 	sunxi_cpu_power_switch_set(cpu, cluster, false);
500 
501 	return 0;
502 }
503 
504 static int sunxi_cluster_powerdown(unsigned int cluster)
505 {
506 	u32 reg;
507 
508 	pr_debug("%s: cluster %u\n", __func__, cluster);
509 	if (cluster >= SUNXI_NR_CLUSTERS)
510 		return -EINVAL;
511 
512 	/* assert cluster resets or system will hang */
513 	pr_debug("%s: assert cluster reset\n", __func__);
514 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
515 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
516 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
517 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
518 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
519 
520 	/* gate cluster power */
521 	pr_debug("%s: gate cluster power\n", __func__);
522 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
523 	if (is_a83t)
524 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
525 	else
526 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
527 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
528 	udelay(20);
529 
530 	return 0;
531 }
532 
533 static int sunxi_mc_smp_cpu_kill(unsigned int l_cpu)
534 {
535 	unsigned int mpidr, cpu, cluster;
536 	unsigned int tries, count;
537 	int ret = 0;
538 	u32 reg;
539 
540 	mpidr = cpu_logical_map(l_cpu);
541 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
542 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
543 
544 	/* This should never happen */
545 	if (WARN_ON(cluster >= SUNXI_NR_CLUSTERS ||
546 		    cpu >= SUNXI_CPUS_PER_CLUSTER))
547 		return 0;
548 
549 	/* wait for CPU core to die and enter WFI */
550 	count = TIMEOUT_USEC / POLL_USEC;
551 	spin_lock_irq(&boot_lock);
552 	for (tries = 0; tries < count; tries++) {
553 		spin_unlock_irq(&boot_lock);
554 		usleep_range(POLL_USEC / 2, POLL_USEC);
555 		spin_lock_irq(&boot_lock);
556 
557 		/*
558 		 * If the user turns off a bunch of cores at the same
559 		 * time, the kernel might call cpu_kill before some of
560 		 * them are ready. This is because boot_lock serializes
561 		 * both cpu_die and cpu_kill callbacks. Either one could
562 		 * run first. We should wait for cpu_die to complete.
563 		 */
564 		if (sunxi_mc_smp_cpu_table[cluster][cpu])
565 			continue;
566 
567 		reg = readl(cpucfg_base + CPUCFG_CX_STATUS(cluster));
568 		if (reg & CPUCFG_CX_STATUS_STANDBYWFI(cpu))
569 			break;
570 	}
571 
572 	if (tries >= count) {
573 		ret = ETIMEDOUT;
574 		goto out;
575 	}
576 
577 	/* power down CPU core */
578 	sunxi_cpu_powerdown(cpu, cluster);
579 
580 	if (!sunxi_mc_smp_cluster_is_down(cluster))
581 		goto out;
582 
583 	/* wait for cluster L2 WFI */
584 	ret = readl_poll_timeout(cpucfg_base + CPUCFG_CX_STATUS(cluster), reg,
585 				 reg & CPUCFG_CX_STATUS_STANDBYWFIL2,
586 				 POLL_USEC, TIMEOUT_USEC);
587 	if (ret) {
588 		/*
589 		 * Ignore timeout on the cluster. Leaving the cluster on
590 		 * will not affect system execution, just use a bit more
591 		 * power. But returning an error here will only confuse
592 		 * the user as the CPU has already been shutdown.
593 		 */
594 		ret = 0;
595 		goto out;
596 	}
597 
598 	/* Power down cluster */
599 	sunxi_cluster_powerdown(cluster);
600 
601 out:
602 	spin_unlock_irq(&boot_lock);
603 	pr_debug("%s: cluster %u cpu %u powerdown: %d\n",
604 		 __func__, cluster, cpu, ret);
605 	return !ret;
606 }
607 
608 static bool sunxi_mc_smp_cpu_can_disable(unsigned int cpu)
609 {
610 	/* CPU0 hotplug not handled for sun8i-a83t */
611 	if (is_a83t)
612 		if (cpu == 0)
613 			return false;
614 	return true;
615 }
616 #endif
617 
618 static const struct smp_operations sunxi_mc_smp_smp_ops __initconst = {
619 	.smp_secondary_init	= sunxi_mc_smp_secondary_init,
620 	.smp_boot_secondary	= sunxi_mc_smp_boot_secondary,
621 #ifdef CONFIG_HOTPLUG_CPU
622 	.cpu_die		= sunxi_mc_smp_cpu_die,
623 	.cpu_kill		= sunxi_mc_smp_cpu_kill,
624 	.cpu_can_disable	= sunxi_mc_smp_cpu_can_disable,
625 #endif
626 };
627 
628 static bool __init sunxi_mc_smp_cpu_table_init(void)
629 {
630 	unsigned int mpidr, cpu, cluster;
631 
632 	mpidr = read_cpuid_mpidr();
633 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
634 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
635 
636 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER) {
637 		pr_err("%s: boot CPU is out of bounds!\n", __func__);
638 		return false;
639 	}
640 	sunxi_mc_smp_cpu_table[cluster][cpu] = 1;
641 	return true;
642 }
643 
644 /*
645  * Adapted from arch/arm/common/mc_smp_entry.c
646  *
647  * We need the trampoline code to enable CCI-400 on the first cluster
648  */
649 typedef typeof(cpu_reset) phys_reset_t;
650 
651 static int __init nocache_trampoline(unsigned long __unused)
652 {
653 	phys_reset_t phys_reset;
654 
655 	setup_mm_for_reboot();
656 	sunxi_cluster_cache_disable_without_axi();
657 
658 	phys_reset = (phys_reset_t)(unsigned long)__pa_symbol(cpu_reset);
659 	phys_reset(__pa_symbol(sunxi_mc_smp_resume), false);
660 	BUG();
661 }
662 
663 static int __init sunxi_mc_smp_loopback(void)
664 {
665 	int ret;
666 
667 	/*
668 	 * We're going to soft-restart the current CPU through the
669 	 * low-level MCPM code by leveraging the suspend/resume
670 	 * infrastructure. Let's play it safe by using cpu_pm_enter()
671 	 * in case the CPU init code path resets the VFP or similar.
672 	 */
673 	sunxi_mc_smp_first_comer = true;
674 	local_irq_disable();
675 	local_fiq_disable();
676 	ret = cpu_pm_enter();
677 	if (!ret) {
678 		ret = cpu_suspend(0, nocache_trampoline);
679 		cpu_pm_exit();
680 	}
681 	local_fiq_enable();
682 	local_irq_enable();
683 	sunxi_mc_smp_first_comer = false;
684 
685 	return ret;
686 }
687 
688 /*
689  * This holds any device nodes that we requested resources for,
690  * so that we may easily release resources in the error path.
691  */
692 struct sunxi_mc_smp_nodes {
693 	struct device_node *prcm_node;
694 	struct device_node *cpucfg_node;
695 	struct device_node *sram_node;
696 	struct device_node *r_cpucfg_node;
697 };
698 
699 /* This structure holds SoC-specific bits tied to an enable-method string. */
700 struct sunxi_mc_smp_data {
701 	const char *enable_method;
702 	int (*get_smp_nodes)(struct sunxi_mc_smp_nodes *nodes);
703 	bool is_a83t;
704 };
705 
706 static void __init sunxi_mc_smp_put_nodes(struct sunxi_mc_smp_nodes *nodes)
707 {
708 	of_node_put(nodes->prcm_node);
709 	of_node_put(nodes->cpucfg_node);
710 	of_node_put(nodes->sram_node);
711 	of_node_put(nodes->r_cpucfg_node);
712 	memset(nodes, 0, sizeof(*nodes));
713 }
714 
715 static int __init sun9i_a80_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
716 {
717 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
718 						   "allwinner,sun9i-a80-prcm");
719 	if (!nodes->prcm_node) {
720 		pr_err("%s: PRCM not available\n", __func__);
721 		return -ENODEV;
722 	}
723 
724 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
725 						     "allwinner,sun9i-a80-cpucfg");
726 	if (!nodes->cpucfg_node) {
727 		pr_err("%s: CPUCFG not available\n", __func__);
728 		return -ENODEV;
729 	}
730 
731 	nodes->sram_node = of_find_compatible_node(NULL, NULL,
732 						   "allwinner,sun9i-a80-smp-sram");
733 	if (!nodes->sram_node) {
734 		pr_err("%s: Secure SRAM not available\n", __func__);
735 		return -ENODEV;
736 	}
737 
738 	return 0;
739 }
740 
741 static int __init sun8i_a83t_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
742 {
743 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
744 						   "allwinner,sun8i-a83t-r-ccu");
745 	if (!nodes->prcm_node) {
746 		pr_err("%s: PRCM not available\n", __func__);
747 		return -ENODEV;
748 	}
749 
750 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
751 						     "allwinner,sun8i-a83t-cpucfg");
752 	if (!nodes->cpucfg_node) {
753 		pr_err("%s: CPUCFG not available\n", __func__);
754 		return -ENODEV;
755 	}
756 
757 	nodes->r_cpucfg_node = of_find_compatible_node(NULL, NULL,
758 						       "allwinner,sun8i-a83t-r-cpucfg");
759 	if (!nodes->r_cpucfg_node) {
760 		pr_err("%s: RCPUCFG not available\n", __func__);
761 		return -ENODEV;
762 	}
763 
764 	return 0;
765 }
766 
767 static const struct sunxi_mc_smp_data sunxi_mc_smp_data[] __initconst = {
768 	{
769 		.enable_method	= "allwinner,sun9i-a80-smp",
770 		.get_smp_nodes	= sun9i_a80_get_smp_nodes,
771 	},
772 	{
773 		.enable_method	= "allwinner,sun8i-a83t-smp",
774 		.get_smp_nodes	= sun8i_a83t_get_smp_nodes,
775 		.is_a83t	= true,
776 	},
777 };
778 
779 static int __init sunxi_mc_smp_init(void)
780 {
781 	struct sunxi_mc_smp_nodes nodes = { 0 };
782 	struct device_node *node;
783 	struct resource res;
784 	void __iomem *addr;
785 	int i, ret;
786 
787 	/*
788 	 * Don't bother checking the "cpus" node, as an enable-method
789 	 * property in that node is undocumented.
790 	 */
791 	node = of_cpu_device_node_get(0);
792 	if (!node)
793 		return -ENODEV;
794 
795 	/*
796 	 * We can't actually use the enable-method magic in the kernel.
797 	 * Our loopback / trampoline code uses the CPU suspend framework,
798 	 * which requires the identity mapping be available. It would not
799 	 * yet be available if we used the .init_cpus or .prepare_cpus
800 	 * callbacks in smp_operations, which we would use if we were to
801 	 * use CPU_METHOD_OF_DECLARE
802 	 */
803 	for (i = 0; i < ARRAY_SIZE(sunxi_mc_smp_data); i++) {
804 		ret = of_property_match_string(node, "enable-method",
805 					       sunxi_mc_smp_data[i].enable_method);
806 		if (!ret)
807 			break;
808 	}
809 
810 	is_a83t = sunxi_mc_smp_data[i].is_a83t;
811 
812 	of_node_put(node);
813 	if (ret)
814 		return -ENODEV;
815 
816 	if (!sunxi_mc_smp_cpu_table_init())
817 		return -EINVAL;
818 
819 	if (!cci_probed()) {
820 		pr_err("%s: CCI-400 not available\n", __func__);
821 		return -ENODEV;
822 	}
823 
824 	/* Get needed device tree nodes */
825 	ret = sunxi_mc_smp_data[i].get_smp_nodes(&nodes);
826 	if (ret)
827 		goto err_put_nodes;
828 
829 	/*
830 	 * Unfortunately we can not request the I/O region for the PRCM.
831 	 * It is shared with the PRCM clock.
832 	 */
833 	prcm_base = of_iomap(nodes.prcm_node, 0);
834 	if (!prcm_base) {
835 		pr_err("%s: failed to map PRCM registers\n", __func__);
836 		ret = -ENOMEM;
837 		goto err_put_nodes;
838 	}
839 
840 	cpucfg_base = of_io_request_and_map(nodes.cpucfg_node, 0,
841 					    "sunxi-mc-smp");
842 	if (IS_ERR(cpucfg_base)) {
843 		ret = PTR_ERR(cpucfg_base);
844 		pr_err("%s: failed to map CPUCFG registers: %d\n",
845 		       __func__, ret);
846 		goto err_unmap_prcm;
847 	}
848 
849 	if (is_a83t) {
850 		r_cpucfg_base = of_io_request_and_map(nodes.r_cpucfg_node,
851 						      0, "sunxi-mc-smp");
852 		if (IS_ERR(r_cpucfg_base)) {
853 			ret = PTR_ERR(r_cpucfg_base);
854 			pr_err("%s: failed to map R-CPUCFG registers\n",
855 			       __func__);
856 			goto err_unmap_release_cpucfg;
857 		}
858 	} else {
859 		sram_b_smp_base = of_io_request_and_map(nodes.sram_node, 0,
860 							"sunxi-mc-smp");
861 		if (IS_ERR(sram_b_smp_base)) {
862 			ret = PTR_ERR(sram_b_smp_base);
863 			pr_err("%s: failed to map secure SRAM\n", __func__);
864 			goto err_unmap_release_cpucfg;
865 		}
866 	}
867 
868 	/* Configure CCI-400 for boot cluster */
869 	ret = sunxi_mc_smp_loopback();
870 	if (ret) {
871 		pr_err("%s: failed to configure boot cluster: %d\n",
872 		       __func__, ret);
873 		goto err_unmap_release_sram_rcpucfg;
874 	}
875 
876 	/* We don't need the device nodes anymore */
877 	sunxi_mc_smp_put_nodes(&nodes);
878 
879 	/* Set the hardware entry point address */
880 	if (is_a83t)
881 		addr = r_cpucfg_base + R_CPUCFG_CPU_SOFT_ENTRY_REG;
882 	else
883 		addr = prcm_base + PRCM_CPU_SOFT_ENTRY_REG;
884 	writel(__pa_symbol(sunxi_mc_smp_secondary_startup), addr);
885 
886 	/* Actually enable multi cluster SMP */
887 	smp_set_ops(&sunxi_mc_smp_smp_ops);
888 
889 	pr_info("sunxi multi cluster SMP support installed\n");
890 
891 	return 0;
892 
893 err_unmap_release_sram_rcpucfg:
894 	if (is_a83t) {
895 		iounmap(r_cpucfg_base);
896 		of_address_to_resource(nodes.r_cpucfg_node, 0, &res);
897 	} else {
898 		iounmap(sram_b_smp_base);
899 		of_address_to_resource(nodes.sram_node, 0, &res);
900 	}
901 	release_mem_region(res.start, resource_size(&res));
902 err_unmap_release_cpucfg:
903 	iounmap(cpucfg_base);
904 	of_address_to_resource(nodes.cpucfg_node, 0, &res);
905 	release_mem_region(res.start, resource_size(&res));
906 err_unmap_prcm:
907 	iounmap(prcm_base);
908 err_put_nodes:
909 	sunxi_mc_smp_put_nodes(&nodes);
910 	return ret;
911 }
912 
913 early_initcall(sunxi_mc_smp_init);
914