xref: /openbmc/linux/arch/arm/mach-sunxi/mc_smp.c (revision c4f7ac64)
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/of_device.h>
23 #include <linux/smp.h>
24 
25 #include <asm/cacheflush.h>
26 #include <asm/cp15.h>
27 #include <asm/cputype.h>
28 #include <asm/idmap.h>
29 #include <asm/smp_plat.h>
30 #include <asm/suspend.h>
31 
32 #define SUNXI_CPUS_PER_CLUSTER		4
33 #define SUNXI_NR_CLUSTERS		2
34 
35 #define POLL_USEC	100
36 #define TIMEOUT_USEC	100000
37 
38 #define CPUCFG_CX_CTRL_REG0(c)		(0x10 * (c))
39 #define CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE(n)	BIT(n)
40 #define CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE_ALL	0xf
41 #define CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A7	BIT(4)
42 #define CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A15	BIT(0)
43 #define CPUCFG_CX_CTRL_REG1(c)		(0x10 * (c) + 0x4)
44 #define CPUCFG_CX_CTRL_REG1_ACINACTM	BIT(0)
45 #define CPUCFG_CX_STATUS(c)		(0x30 + 0x4 * (c))
46 #define CPUCFG_CX_STATUS_STANDBYWFI(n)	BIT(16 + (n))
47 #define CPUCFG_CX_STATUS_STANDBYWFIL2	BIT(0)
48 #define CPUCFG_CX_RST_CTRL(c)		(0x80 + 0x4 * (c))
49 #define CPUCFG_CX_RST_CTRL_DBG_SOC_RST	BIT(24)
50 #define CPUCFG_CX_RST_CTRL_ETM_RST(n)	BIT(20 + (n))
51 #define CPUCFG_CX_RST_CTRL_ETM_RST_ALL	(0xf << 20)
52 #define CPUCFG_CX_RST_CTRL_DBG_RST(n)	BIT(16 + (n))
53 #define CPUCFG_CX_RST_CTRL_DBG_RST_ALL	(0xf << 16)
54 #define CPUCFG_CX_RST_CTRL_H_RST	BIT(12)
55 #define CPUCFG_CX_RST_CTRL_L2_RST	BIT(8)
56 #define CPUCFG_CX_RST_CTRL_CX_RST(n)	BIT(4 + (n))
57 #define CPUCFG_CX_RST_CTRL_CORE_RST(n)	BIT(n)
58 #define CPUCFG_CX_RST_CTRL_CORE_RST_ALL	(0xf << 0)
59 
60 #define PRCM_CPU_PO_RST_CTRL(c)		(0x4 + 0x4 * (c))
61 #define PRCM_CPU_PO_RST_CTRL_CORE(n)	BIT(n)
62 #define PRCM_CPU_PO_RST_CTRL_CORE_ALL	0xf
63 #define PRCM_PWROFF_GATING_REG(c)	(0x100 + 0x4 * (c))
64 /* The power off register for clusters are different from a80 and a83t */
65 #define PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I	BIT(0)
66 #define PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I	BIT(4)
67 #define PRCM_PWROFF_GATING_REG_CORE(n)	BIT(n)
68 #define PRCM_PWR_SWITCH_REG(c, cpu)	(0x140 + 0x10 * (c) + 0x4 * (cpu))
69 #define PRCM_CPU_SOFT_ENTRY_REG		0x164
70 
71 /* R_CPUCFG registers, specific to sun8i-a83t */
72 #define R_CPUCFG_CLUSTER_PO_RST_CTRL(c)	(0x30 + (c) * 0x4)
73 #define R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(n)	BIT(n)
74 #define R_CPUCFG_CPU_SOFT_ENTRY_REG		0x01a4
75 
76 #define CPU0_SUPPORT_HOTPLUG_MAGIC0	0xFA50392F
77 #define CPU0_SUPPORT_HOTPLUG_MAGIC1	0x790DCA3A
78 
79 static void __iomem *cpucfg_base;
80 static void __iomem *prcm_base;
81 static void __iomem *sram_b_smp_base;
82 static void __iomem *r_cpucfg_base;
83 
84 extern void sunxi_mc_smp_secondary_startup(void);
85 extern void sunxi_mc_smp_resume(void);
86 static bool is_a83t;
87 
88 static bool sunxi_core_is_cortex_a15(unsigned int core, unsigned int cluster)
89 {
90 	struct device_node *node;
91 	int cpu = cluster * SUNXI_CPUS_PER_CLUSTER + core;
92 	bool is_compatible;
93 
94 	node = of_cpu_device_node_get(cpu);
95 
96 	/* In case of_cpu_device_node_get fails */
97 	if (!node)
98 		node = of_get_cpu_node(cpu, NULL);
99 
100 	if (!node) {
101 		/*
102 		 * There's no point in returning an error, since we
103 		 * would be mid way in a core or cluster power sequence.
104 		 */
105 		pr_err("%s: Couldn't get CPU cluster %u core %u device node\n",
106 		       __func__, cluster, core);
107 
108 		return false;
109 	}
110 
111 	is_compatible = of_device_is_compatible(node, "arm,cortex-a15");
112 	of_node_put(node);
113 	return is_compatible;
114 }
115 
116 static int sunxi_cpu_power_switch_set(unsigned int cpu, unsigned int cluster,
117 				      bool enable)
118 {
119 	u32 reg;
120 
121 	/* control sequence from Allwinner A80 user manual v1.2 PRCM section */
122 	reg = readl(prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
123 	if (enable) {
124 		if (reg == 0x00) {
125 			pr_debug("power clamp for cluster %u cpu %u already open\n",
126 				 cluster, cpu);
127 			return 0;
128 		}
129 
130 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
131 		udelay(10);
132 		writel(0xfe, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
133 		udelay(10);
134 		writel(0xf8, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
135 		udelay(10);
136 		writel(0xf0, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
137 		udelay(10);
138 		writel(0x00, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
139 		udelay(10);
140 	} else {
141 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
142 		udelay(10);
143 	}
144 
145 	return 0;
146 }
147 
148 static void sunxi_cpu0_hotplug_support_set(bool enable)
149 {
150 	if (enable) {
151 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC0, sram_b_smp_base);
152 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC1, sram_b_smp_base + 0x4);
153 	} else {
154 		writel(0x0, sram_b_smp_base);
155 		writel(0x0, sram_b_smp_base + 0x4);
156 	}
157 }
158 
159 static int sunxi_cpu_powerup(unsigned int cpu, unsigned int cluster)
160 {
161 	u32 reg;
162 
163 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
164 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
165 		return -EINVAL;
166 
167 	/* Set hotplug support magic flags for cpu0 */
168 	if (cluster == 0 && cpu == 0)
169 		sunxi_cpu0_hotplug_support_set(true);
170 
171 	/* assert processor power-on reset */
172 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
173 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE(cpu);
174 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
175 
176 	if (is_a83t) {
177 		/* assert cpu power-on reset */
178 		reg  = readl(r_cpucfg_base +
179 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
180 		reg &= ~(R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(cpu));
181 		writel(reg, r_cpucfg_base +
182 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
183 		udelay(10);
184 	}
185 
186 	/* Cortex-A7: hold L1 reset disable signal low */
187 	if (!sunxi_core_is_cortex_a15(cpu, cluster)) {
188 		reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
189 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE(cpu);
190 		writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
191 	}
192 
193 	/* assert processor related resets */
194 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
195 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
196 
197 	/*
198 	 * Allwinner code also asserts resets for NEON on A15. According
199 	 * to ARM manuals, asserting power-on reset is sufficient.
200 	 */
201 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
202 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
203 
204 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
205 
206 	/* open power switch */
207 	sunxi_cpu_power_switch_set(cpu, cluster, true);
208 
209 	/* Handle A83T bit swap */
210 	if (is_a83t) {
211 		if (cpu == 0)
212 			cpu = 4;
213 	}
214 
215 	/* clear processor power gate */
216 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
217 	reg &= ~PRCM_PWROFF_GATING_REG_CORE(cpu);
218 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
219 	udelay(20);
220 
221 	/* Handle A83T bit swap */
222 	if (is_a83t) {
223 		if (cpu == 4)
224 			cpu = 0;
225 	}
226 
227 	/* de-assert processor power-on reset */
228 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
229 	reg |= PRCM_CPU_PO_RST_CTRL_CORE(cpu);
230 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
231 
232 	if (is_a83t) {
233 		reg  = readl(r_cpucfg_base +
234 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
235 		reg |= R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(cpu);
236 		writel(reg, r_cpucfg_base +
237 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
238 		udelay(10);
239 	}
240 
241 	/* de-assert all processor resets */
242 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
243 	reg |= CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
244 	reg |= CPUCFG_CX_RST_CTRL_CORE_RST(cpu);
245 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
246 		reg |= CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
247 	else
248 		reg |= CPUCFG_CX_RST_CTRL_CX_RST(cpu); /* NEON */
249 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
250 
251 	return 0;
252 }
253 
254 static int sunxi_cluster_powerup(unsigned int cluster)
255 {
256 	u32 reg;
257 
258 	pr_debug("%s: cluster %u\n", __func__, cluster);
259 	if (cluster >= SUNXI_NR_CLUSTERS)
260 		return -EINVAL;
261 
262 	/* For A83T, assert cluster cores resets */
263 	if (is_a83t) {
264 		reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
265 		reg &= ~CPUCFG_CX_RST_CTRL_CORE_RST_ALL;   /* Core Reset    */
266 		writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
267 		udelay(10);
268 	}
269 
270 	/* assert ACINACTM */
271 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
272 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
273 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
274 
275 	/* assert cluster processor power-on resets */
276 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
277 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE_ALL;
278 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
279 
280 	/* assert cluster cores resets */
281 	if (is_a83t) {
282 		reg  = readl(r_cpucfg_base +
283 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
284 		reg &= ~CPUCFG_CX_RST_CTRL_CORE_RST_ALL;
285 		writel(reg, r_cpucfg_base +
286 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
287 		udelay(10);
288 	}
289 
290 	/* assert cluster resets */
291 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
292 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
293 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST_ALL;
294 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
295 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
296 
297 	/*
298 	 * Allwinner code also asserts resets for NEON on A15. According
299 	 * to ARM manuals, asserting power-on reset is sufficient.
300 	 */
301 	if (!sunxi_core_is_cortex_a15(0, cluster))
302 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST_ALL;
303 
304 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
305 
306 	/* hold L1/L2 reset disable signals low */
307 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
308 	if (sunxi_core_is_cortex_a15(0, cluster)) {
309 		/* Cortex-A15: hold L2RSTDISABLE low */
310 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A15;
311 	} else {
312 		/* Cortex-A7: hold L1RSTDISABLE and L2RSTDISABLE low */
313 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE_ALL;
314 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A7;
315 	}
316 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
317 
318 	/* clear cluster power gate */
319 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
320 	if (is_a83t)
321 		reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
322 	else
323 		reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
324 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
325 	udelay(20);
326 
327 	/* de-assert cluster resets */
328 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
329 	reg |= CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
330 	reg |= CPUCFG_CX_RST_CTRL_H_RST;
331 	reg |= CPUCFG_CX_RST_CTRL_L2_RST;
332 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
333 
334 	/* de-assert ACINACTM */
335 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
336 	reg &= ~CPUCFG_CX_CTRL_REG1_ACINACTM;
337 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
338 
339 	return 0;
340 }
341 
342 /*
343  * This bit is shared between the initial nocache_trampoline call to
344  * enable CCI-400 and proper cluster cache disable before power down.
345  */
346 static void sunxi_cluster_cache_disable_without_axi(void)
347 {
348 	if (read_cpuid_part() == ARM_CPU_PART_CORTEX_A15) {
349 		/*
350 		 * On the Cortex-A15 we need to disable
351 		 * L2 prefetching before flushing the cache.
352 		 */
353 		asm volatile(
354 		"mcr	p15, 1, %0, c15, c0, 3\n"
355 		"isb\n"
356 		"dsb"
357 		: : "r" (0x400));
358 	}
359 
360 	/* Flush all cache levels for this cluster. */
361 	v7_exit_coherency_flush(all);
362 
363 	/*
364 	 * Disable cluster-level coherency by masking
365 	 * incoming snoops and DVM messages:
366 	 */
367 	cci_disable_port_by_cpu(read_cpuid_mpidr());
368 }
369 
370 static int sunxi_mc_smp_cpu_table[SUNXI_NR_CLUSTERS][SUNXI_CPUS_PER_CLUSTER];
371 int sunxi_mc_smp_first_comer;
372 
373 static DEFINE_SPINLOCK(boot_lock);
374 
375 static bool sunxi_mc_smp_cluster_is_down(unsigned int cluster)
376 {
377 	int i;
378 
379 	for (i = 0; i < SUNXI_CPUS_PER_CLUSTER; i++)
380 		if (sunxi_mc_smp_cpu_table[cluster][i])
381 			return false;
382 	return true;
383 }
384 
385 static void sunxi_mc_smp_secondary_init(unsigned int cpu)
386 {
387 	/* Clear hotplug support magic flags for cpu0 */
388 	if (cpu == 0)
389 		sunxi_cpu0_hotplug_support_set(false);
390 }
391 
392 static int sunxi_mc_smp_boot_secondary(unsigned int l_cpu, struct task_struct *idle)
393 {
394 	unsigned int mpidr, cpu, cluster;
395 
396 	mpidr = cpu_logical_map(l_cpu);
397 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
398 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
399 
400 	if (!cpucfg_base)
401 		return -ENODEV;
402 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER)
403 		return -EINVAL;
404 
405 	spin_lock_irq(&boot_lock);
406 
407 	if (sunxi_mc_smp_cpu_table[cluster][cpu])
408 		goto out;
409 
410 	if (sunxi_mc_smp_cluster_is_down(cluster)) {
411 		sunxi_mc_smp_first_comer = true;
412 		sunxi_cluster_powerup(cluster);
413 	} else {
414 		sunxi_mc_smp_first_comer = false;
415 	}
416 
417 	/* This is read by incoming CPUs with their cache and MMU disabled */
418 	sync_cache_w(&sunxi_mc_smp_first_comer);
419 	sunxi_cpu_powerup(cpu, cluster);
420 
421 out:
422 	sunxi_mc_smp_cpu_table[cluster][cpu]++;
423 	spin_unlock_irq(&boot_lock);
424 
425 	return 0;
426 }
427 
428 #ifdef CONFIG_HOTPLUG_CPU
429 static void sunxi_cluster_cache_disable(void)
430 {
431 	unsigned int cluster = MPIDR_AFFINITY_LEVEL(read_cpuid_mpidr(), 1);
432 	u32 reg;
433 
434 	pr_debug("%s: cluster %u\n", __func__, cluster);
435 
436 	sunxi_cluster_cache_disable_without_axi();
437 
438 	/* last man standing, assert ACINACTM */
439 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
440 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
441 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
442 }
443 
444 static void sunxi_mc_smp_cpu_die(unsigned int l_cpu)
445 {
446 	unsigned int mpidr, cpu, cluster;
447 	bool last_man;
448 
449 	mpidr = cpu_logical_map(l_cpu);
450 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
451 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
452 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
453 
454 	spin_lock(&boot_lock);
455 	sunxi_mc_smp_cpu_table[cluster][cpu]--;
456 	if (sunxi_mc_smp_cpu_table[cluster][cpu] == 1) {
457 		/* A power_up request went ahead of us. */
458 		pr_debug("%s: aborting due to a power up request\n",
459 			 __func__);
460 		spin_unlock(&boot_lock);
461 		return;
462 	} else if (sunxi_mc_smp_cpu_table[cluster][cpu] > 1) {
463 		pr_err("Cluster %d CPU%d boots multiple times\n",
464 		       cluster, cpu);
465 		BUG();
466 	}
467 
468 	last_man = sunxi_mc_smp_cluster_is_down(cluster);
469 	spin_unlock(&boot_lock);
470 
471 	gic_cpu_if_down(0);
472 	if (last_man)
473 		sunxi_cluster_cache_disable();
474 	else
475 		v7_exit_coherency_flush(louis);
476 
477 	for (;;)
478 		wfi();
479 }
480 
481 static int sunxi_cpu_powerdown(unsigned int cpu, unsigned int cluster)
482 {
483 	u32 reg;
484 	int gating_bit = cpu;
485 
486 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
487 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
488 		return -EINVAL;
489 
490 	if (is_a83t && cpu == 0)
491 		gating_bit = 4;
492 
493 	/* gate processor power */
494 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
495 	reg |= PRCM_PWROFF_GATING_REG_CORE(gating_bit);
496 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
497 	udelay(20);
498 
499 	/* close power switch */
500 	sunxi_cpu_power_switch_set(cpu, cluster, false);
501 
502 	return 0;
503 }
504 
505 static int sunxi_cluster_powerdown(unsigned int cluster)
506 {
507 	u32 reg;
508 
509 	pr_debug("%s: cluster %u\n", __func__, cluster);
510 	if (cluster >= SUNXI_NR_CLUSTERS)
511 		return -EINVAL;
512 
513 	/* assert cluster resets or system will hang */
514 	pr_debug("%s: assert cluster reset\n", __func__);
515 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
516 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
517 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
518 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
519 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
520 
521 	/* gate cluster power */
522 	pr_debug("%s: gate cluster power\n", __func__);
523 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
524 	if (is_a83t)
525 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
526 	else
527 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
528 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
529 	udelay(20);
530 
531 	return 0;
532 }
533 
534 static int sunxi_mc_smp_cpu_kill(unsigned int l_cpu)
535 {
536 	unsigned int mpidr, cpu, cluster;
537 	unsigned int tries, count;
538 	int ret = 0;
539 	u32 reg;
540 
541 	mpidr = cpu_logical_map(l_cpu);
542 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
543 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
544 
545 	/* This should never happen */
546 	if (WARN_ON(cluster >= SUNXI_NR_CLUSTERS ||
547 		    cpu >= SUNXI_CPUS_PER_CLUSTER))
548 		return 0;
549 
550 	/* wait for CPU core to die and enter WFI */
551 	count = TIMEOUT_USEC / POLL_USEC;
552 	spin_lock_irq(&boot_lock);
553 	for (tries = 0; tries < count; tries++) {
554 		spin_unlock_irq(&boot_lock);
555 		usleep_range(POLL_USEC / 2, POLL_USEC);
556 		spin_lock_irq(&boot_lock);
557 
558 		/*
559 		 * If the user turns off a bunch of cores at the same
560 		 * time, the kernel might call cpu_kill before some of
561 		 * them are ready. This is because boot_lock serializes
562 		 * both cpu_die and cpu_kill callbacks. Either one could
563 		 * run first. We should wait for cpu_die to complete.
564 		 */
565 		if (sunxi_mc_smp_cpu_table[cluster][cpu])
566 			continue;
567 
568 		reg = readl(cpucfg_base + CPUCFG_CX_STATUS(cluster));
569 		if (reg & CPUCFG_CX_STATUS_STANDBYWFI(cpu))
570 			break;
571 	}
572 
573 	if (tries >= count) {
574 		ret = ETIMEDOUT;
575 		goto out;
576 	}
577 
578 	/* power down CPU core */
579 	sunxi_cpu_powerdown(cpu, cluster);
580 
581 	if (!sunxi_mc_smp_cluster_is_down(cluster))
582 		goto out;
583 
584 	/* wait for cluster L2 WFI */
585 	ret = readl_poll_timeout(cpucfg_base + CPUCFG_CX_STATUS(cluster), reg,
586 				 reg & CPUCFG_CX_STATUS_STANDBYWFIL2,
587 				 POLL_USEC, TIMEOUT_USEC);
588 	if (ret) {
589 		/*
590 		 * Ignore timeout on the cluster. Leaving the cluster on
591 		 * will not affect system execution, just use a bit more
592 		 * power. But returning an error here will only confuse
593 		 * the user as the CPU has already been shutdown.
594 		 */
595 		ret = 0;
596 		goto out;
597 	}
598 
599 	/* Power down cluster */
600 	sunxi_cluster_powerdown(cluster);
601 
602 out:
603 	spin_unlock_irq(&boot_lock);
604 	pr_debug("%s: cluster %u cpu %u powerdown: %d\n",
605 		 __func__, cluster, cpu, ret);
606 	return !ret;
607 }
608 
609 static bool sunxi_mc_smp_cpu_can_disable(unsigned int cpu)
610 {
611 	/* CPU0 hotplug not handled for sun8i-a83t */
612 	if (is_a83t)
613 		if (cpu == 0)
614 			return false;
615 	return true;
616 }
617 #endif
618 
619 static const struct smp_operations sunxi_mc_smp_smp_ops __initconst = {
620 	.smp_secondary_init	= sunxi_mc_smp_secondary_init,
621 	.smp_boot_secondary	= sunxi_mc_smp_boot_secondary,
622 #ifdef CONFIG_HOTPLUG_CPU
623 	.cpu_die		= sunxi_mc_smp_cpu_die,
624 	.cpu_kill		= sunxi_mc_smp_cpu_kill,
625 	.cpu_can_disable	= sunxi_mc_smp_cpu_can_disable,
626 #endif
627 };
628 
629 static bool __init sunxi_mc_smp_cpu_table_init(void)
630 {
631 	unsigned int mpidr, cpu, cluster;
632 
633 	mpidr = read_cpuid_mpidr();
634 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
635 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
636 
637 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER) {
638 		pr_err("%s: boot CPU is out of bounds!\n", __func__);
639 		return false;
640 	}
641 	sunxi_mc_smp_cpu_table[cluster][cpu] = 1;
642 	return true;
643 }
644 
645 /*
646  * Adapted from arch/arm/common/mc_smp_entry.c
647  *
648  * We need the trampoline code to enable CCI-400 on the first cluster
649  */
650 typedef typeof(cpu_reset) phys_reset_t;
651 
652 static int __init nocache_trampoline(unsigned long __unused)
653 {
654 	phys_reset_t phys_reset;
655 
656 	setup_mm_for_reboot();
657 	sunxi_cluster_cache_disable_without_axi();
658 
659 	phys_reset = (phys_reset_t)(unsigned long)__pa_symbol(cpu_reset);
660 	phys_reset(__pa_symbol(sunxi_mc_smp_resume), false);
661 	BUG();
662 }
663 
664 static int __init sunxi_mc_smp_loopback(void)
665 {
666 	int ret;
667 
668 	/*
669 	 * We're going to soft-restart the current CPU through the
670 	 * low-level MCPM code by leveraging the suspend/resume
671 	 * infrastructure. Let's play it safe by using cpu_pm_enter()
672 	 * in case the CPU init code path resets the VFP or similar.
673 	 */
674 	sunxi_mc_smp_first_comer = true;
675 	local_irq_disable();
676 	local_fiq_disable();
677 	ret = cpu_pm_enter();
678 	if (!ret) {
679 		ret = cpu_suspend(0, nocache_trampoline);
680 		cpu_pm_exit();
681 	}
682 	local_fiq_enable();
683 	local_irq_enable();
684 	sunxi_mc_smp_first_comer = false;
685 
686 	return ret;
687 }
688 
689 /*
690  * This holds any device nodes that we requested resources for,
691  * so that we may easily release resources in the error path.
692  */
693 struct sunxi_mc_smp_nodes {
694 	struct device_node *prcm_node;
695 	struct device_node *cpucfg_node;
696 	struct device_node *sram_node;
697 	struct device_node *r_cpucfg_node;
698 };
699 
700 /* This structure holds SoC-specific bits tied to an enable-method string. */
701 struct sunxi_mc_smp_data {
702 	const char *enable_method;
703 	int (*get_smp_nodes)(struct sunxi_mc_smp_nodes *nodes);
704 	bool is_a83t;
705 };
706 
707 static void __init sunxi_mc_smp_put_nodes(struct sunxi_mc_smp_nodes *nodes)
708 {
709 	of_node_put(nodes->prcm_node);
710 	of_node_put(nodes->cpucfg_node);
711 	of_node_put(nodes->sram_node);
712 	of_node_put(nodes->r_cpucfg_node);
713 	memset(nodes, 0, sizeof(*nodes));
714 }
715 
716 static int __init sun9i_a80_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
717 {
718 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
719 						   "allwinner,sun9i-a80-prcm");
720 	if (!nodes->prcm_node) {
721 		pr_err("%s: PRCM not available\n", __func__);
722 		return -ENODEV;
723 	}
724 
725 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
726 						     "allwinner,sun9i-a80-cpucfg");
727 	if (!nodes->cpucfg_node) {
728 		pr_err("%s: CPUCFG not available\n", __func__);
729 		return -ENODEV;
730 	}
731 
732 	nodes->sram_node = of_find_compatible_node(NULL, NULL,
733 						   "allwinner,sun9i-a80-smp-sram");
734 	if (!nodes->sram_node) {
735 		pr_err("%s: Secure SRAM not available\n", __func__);
736 		return -ENODEV;
737 	}
738 
739 	return 0;
740 }
741 
742 static int __init sun8i_a83t_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
743 {
744 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
745 						   "allwinner,sun8i-a83t-r-ccu");
746 	if (!nodes->prcm_node) {
747 		pr_err("%s: PRCM not available\n", __func__);
748 		return -ENODEV;
749 	}
750 
751 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
752 						     "allwinner,sun8i-a83t-cpucfg");
753 	if (!nodes->cpucfg_node) {
754 		pr_err("%s: CPUCFG not available\n", __func__);
755 		return -ENODEV;
756 	}
757 
758 	nodes->r_cpucfg_node = of_find_compatible_node(NULL, NULL,
759 						       "allwinner,sun8i-a83t-r-cpucfg");
760 	if (!nodes->r_cpucfg_node) {
761 		pr_err("%s: RCPUCFG not available\n", __func__);
762 		return -ENODEV;
763 	}
764 
765 	return 0;
766 }
767 
768 static const struct sunxi_mc_smp_data sunxi_mc_smp_data[] __initconst = {
769 	{
770 		.enable_method	= "allwinner,sun9i-a80-smp",
771 		.get_smp_nodes	= sun9i_a80_get_smp_nodes,
772 	},
773 	{
774 		.enable_method	= "allwinner,sun8i-a83t-smp",
775 		.get_smp_nodes	= sun8i_a83t_get_smp_nodes,
776 		.is_a83t	= true,
777 	},
778 };
779 
780 static int __init sunxi_mc_smp_init(void)
781 {
782 	struct sunxi_mc_smp_nodes nodes = { 0 };
783 	struct device_node *node;
784 	struct resource res;
785 	void __iomem *addr;
786 	int i, ret;
787 
788 	/*
789 	 * Don't bother checking the "cpus" node, as an enable-method
790 	 * property in that node is undocumented.
791 	 */
792 	node = of_cpu_device_node_get(0);
793 	if (!node)
794 		return -ENODEV;
795 
796 	/*
797 	 * We can't actually use the enable-method magic in the kernel.
798 	 * Our loopback / trampoline code uses the CPU suspend framework,
799 	 * which requires the identity mapping be available. It would not
800 	 * yet be available if we used the .init_cpus or .prepare_cpus
801 	 * callbacks in smp_operations, which we would use if we were to
802 	 * use CPU_METHOD_OF_DECLARE
803 	 */
804 	for (i = 0; i < ARRAY_SIZE(sunxi_mc_smp_data); i++) {
805 		ret = of_property_match_string(node, "enable-method",
806 					       sunxi_mc_smp_data[i].enable_method);
807 		if (!ret)
808 			break;
809 	}
810 
811 	is_a83t = sunxi_mc_smp_data[i].is_a83t;
812 
813 	of_node_put(node);
814 	if (ret)
815 		return -ENODEV;
816 
817 	if (!sunxi_mc_smp_cpu_table_init())
818 		return -EINVAL;
819 
820 	if (!cci_probed()) {
821 		pr_err("%s: CCI-400 not available\n", __func__);
822 		return -ENODEV;
823 	}
824 
825 	/* Get needed device tree nodes */
826 	ret = sunxi_mc_smp_data[i].get_smp_nodes(&nodes);
827 	if (ret)
828 		goto err_put_nodes;
829 
830 	/*
831 	 * Unfortunately we can not request the I/O region for the PRCM.
832 	 * It is shared with the PRCM clock.
833 	 */
834 	prcm_base = of_iomap(nodes.prcm_node, 0);
835 	if (!prcm_base) {
836 		pr_err("%s: failed to map PRCM registers\n", __func__);
837 		ret = -ENOMEM;
838 		goto err_put_nodes;
839 	}
840 
841 	cpucfg_base = of_io_request_and_map(nodes.cpucfg_node, 0,
842 					    "sunxi-mc-smp");
843 	if (IS_ERR(cpucfg_base)) {
844 		ret = PTR_ERR(cpucfg_base);
845 		pr_err("%s: failed to map CPUCFG registers: %d\n",
846 		       __func__, ret);
847 		goto err_unmap_prcm;
848 	}
849 
850 	if (is_a83t) {
851 		r_cpucfg_base = of_io_request_and_map(nodes.r_cpucfg_node,
852 						      0, "sunxi-mc-smp");
853 		if (IS_ERR(r_cpucfg_base)) {
854 			ret = PTR_ERR(r_cpucfg_base);
855 			pr_err("%s: failed to map R-CPUCFG registers\n",
856 			       __func__);
857 			goto err_unmap_release_cpucfg;
858 		}
859 	} else {
860 		sram_b_smp_base = of_io_request_and_map(nodes.sram_node, 0,
861 							"sunxi-mc-smp");
862 		if (IS_ERR(sram_b_smp_base)) {
863 			ret = PTR_ERR(sram_b_smp_base);
864 			pr_err("%s: failed to map secure SRAM\n", __func__);
865 			goto err_unmap_release_cpucfg;
866 		}
867 	}
868 
869 	/* Configure CCI-400 for boot cluster */
870 	ret = sunxi_mc_smp_loopback();
871 	if (ret) {
872 		pr_err("%s: failed to configure boot cluster: %d\n",
873 		       __func__, ret);
874 		goto err_unmap_release_sram_rcpucfg;
875 	}
876 
877 	/* We don't need the device nodes anymore */
878 	sunxi_mc_smp_put_nodes(&nodes);
879 
880 	/* Set the hardware entry point address */
881 	if (is_a83t)
882 		addr = r_cpucfg_base + R_CPUCFG_CPU_SOFT_ENTRY_REG;
883 	else
884 		addr = prcm_base + PRCM_CPU_SOFT_ENTRY_REG;
885 	writel(__pa_symbol(sunxi_mc_smp_secondary_startup), addr);
886 
887 	/* Actually enable multi cluster SMP */
888 	smp_set_ops(&sunxi_mc_smp_smp_ops);
889 
890 	pr_info("sunxi multi cluster SMP support installed\n");
891 
892 	return 0;
893 
894 err_unmap_release_sram_rcpucfg:
895 	if (is_a83t) {
896 		iounmap(r_cpucfg_base);
897 		of_address_to_resource(nodes.r_cpucfg_node, 0, &res);
898 	} else {
899 		iounmap(sram_b_smp_base);
900 		of_address_to_resource(nodes.sram_node, 0, &res);
901 	}
902 	release_mem_region(res.start, resource_size(&res));
903 err_unmap_release_cpucfg:
904 	iounmap(cpucfg_base);
905 	of_address_to_resource(nodes.cpucfg_node, 0, &res);
906 	release_mem_region(res.start, resource_size(&res));
907 err_unmap_prcm:
908 	iounmap(prcm_base);
909 err_put_nodes:
910 	sunxi_mc_smp_put_nodes(&nodes);
911 	return ret;
912 }
913 
914 early_initcall(sunxi_mc_smp_init);
915