xref: /openbmc/linux/arch/arm/mach-sunxi/mc_smp.c (revision 07c7c6bf)
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 
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 	/* gate processor power */
490 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
491 	reg |= PRCM_PWROFF_GATING_REG_CORE(cpu);
492 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
493 	udelay(20);
494 
495 	/* close power switch */
496 	sunxi_cpu_power_switch_set(cpu, cluster, false);
497 
498 	return 0;
499 }
500 
501 static int sunxi_cluster_powerdown(unsigned int cluster)
502 {
503 	u32 reg;
504 
505 	pr_debug("%s: cluster %u\n", __func__, cluster);
506 	if (cluster >= SUNXI_NR_CLUSTERS)
507 		return -EINVAL;
508 
509 	/* assert cluster resets or system will hang */
510 	pr_debug("%s: assert cluster reset\n", __func__);
511 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
512 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
513 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
514 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
515 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
516 
517 	/* gate cluster power */
518 	pr_debug("%s: gate cluster power\n", __func__);
519 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
520 	if (is_a83t)
521 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
522 	else
523 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
524 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
525 	udelay(20);
526 
527 	return 0;
528 }
529 
530 static int sunxi_mc_smp_cpu_kill(unsigned int l_cpu)
531 {
532 	unsigned int mpidr, cpu, cluster;
533 	unsigned int tries, count;
534 	int ret = 0;
535 	u32 reg;
536 
537 	mpidr = cpu_logical_map(l_cpu);
538 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
539 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
540 
541 	/* This should never happen */
542 	if (WARN_ON(cluster >= SUNXI_NR_CLUSTERS ||
543 		    cpu >= SUNXI_CPUS_PER_CLUSTER))
544 		return 0;
545 
546 	/* wait for CPU core to die and enter WFI */
547 	count = TIMEOUT_USEC / POLL_USEC;
548 	spin_lock_irq(&boot_lock);
549 	for (tries = 0; tries < count; tries++) {
550 		spin_unlock_irq(&boot_lock);
551 		usleep_range(POLL_USEC / 2, POLL_USEC);
552 		spin_lock_irq(&boot_lock);
553 
554 		/*
555 		 * If the user turns off a bunch of cores at the same
556 		 * time, the kernel might call cpu_kill before some of
557 		 * them are ready. This is because boot_lock serializes
558 		 * both cpu_die and cpu_kill callbacks. Either one could
559 		 * run first. We should wait for cpu_die to complete.
560 		 */
561 		if (sunxi_mc_smp_cpu_table[cluster][cpu])
562 			continue;
563 
564 		reg = readl(cpucfg_base + CPUCFG_CX_STATUS(cluster));
565 		if (reg & CPUCFG_CX_STATUS_STANDBYWFI(cpu))
566 			break;
567 	}
568 
569 	if (tries >= count) {
570 		ret = ETIMEDOUT;
571 		goto out;
572 	}
573 
574 	/* power down CPU core */
575 	sunxi_cpu_powerdown(cpu, cluster);
576 
577 	if (!sunxi_mc_smp_cluster_is_down(cluster))
578 		goto out;
579 
580 	/* wait for cluster L2 WFI */
581 	ret = readl_poll_timeout(cpucfg_base + CPUCFG_CX_STATUS(cluster), reg,
582 				 reg & CPUCFG_CX_STATUS_STANDBYWFIL2,
583 				 POLL_USEC, TIMEOUT_USEC);
584 	if (ret) {
585 		/*
586 		 * Ignore timeout on the cluster. Leaving the cluster on
587 		 * will not affect system execution, just use a bit more
588 		 * power. But returning an error here will only confuse
589 		 * the user as the CPU has already been shutdown.
590 		 */
591 		ret = 0;
592 		goto out;
593 	}
594 
595 	/* Power down cluster */
596 	sunxi_cluster_powerdown(cluster);
597 
598 out:
599 	spin_unlock_irq(&boot_lock);
600 	pr_debug("%s: cluster %u cpu %u powerdown: %d\n",
601 		 __func__, cluster, cpu, ret);
602 	return !ret;
603 }
604 
605 static bool sunxi_mc_smp_cpu_can_disable(unsigned int cpu)
606 {
607 	/* CPU0 hotplug not handled for sun8i-a83t */
608 	if (is_a83t)
609 		if (cpu == 0)
610 			return false;
611 	return true;
612 }
613 #endif
614 
615 static const struct smp_operations sunxi_mc_smp_smp_ops __initconst = {
616 	.smp_secondary_init	= sunxi_mc_smp_secondary_init,
617 	.smp_boot_secondary	= sunxi_mc_smp_boot_secondary,
618 #ifdef CONFIG_HOTPLUG_CPU
619 	.cpu_die		= sunxi_mc_smp_cpu_die,
620 	.cpu_kill		= sunxi_mc_smp_cpu_kill,
621 	.cpu_can_disable	= sunxi_mc_smp_cpu_can_disable,
622 #endif
623 };
624 
625 static bool __init sunxi_mc_smp_cpu_table_init(void)
626 {
627 	unsigned int mpidr, cpu, cluster;
628 
629 	mpidr = read_cpuid_mpidr();
630 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
631 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
632 
633 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER) {
634 		pr_err("%s: boot CPU is out of bounds!\n", __func__);
635 		return false;
636 	}
637 	sunxi_mc_smp_cpu_table[cluster][cpu] = 1;
638 	return true;
639 }
640 
641 /*
642  * Adapted from arch/arm/common/mc_smp_entry.c
643  *
644  * We need the trampoline code to enable CCI-400 on the first cluster
645  */
646 typedef typeof(cpu_reset) phys_reset_t;
647 
648 static int __init nocache_trampoline(unsigned long __unused)
649 {
650 	phys_reset_t phys_reset;
651 
652 	setup_mm_for_reboot();
653 	sunxi_cluster_cache_disable_without_axi();
654 
655 	phys_reset = (phys_reset_t)(unsigned long)__pa_symbol(cpu_reset);
656 	phys_reset(__pa_symbol(sunxi_mc_smp_resume), false);
657 	BUG();
658 }
659 
660 static int __init sunxi_mc_smp_loopback(void)
661 {
662 	int ret;
663 
664 	/*
665 	 * We're going to soft-restart the current CPU through the
666 	 * low-level MCPM code by leveraging the suspend/resume
667 	 * infrastructure. Let's play it safe by using cpu_pm_enter()
668 	 * in case the CPU init code path resets the VFP or similar.
669 	 */
670 	sunxi_mc_smp_first_comer = true;
671 	local_irq_disable();
672 	local_fiq_disable();
673 	ret = cpu_pm_enter();
674 	if (!ret) {
675 		ret = cpu_suspend(0, nocache_trampoline);
676 		cpu_pm_exit();
677 	}
678 	local_fiq_enable();
679 	local_irq_enable();
680 	sunxi_mc_smp_first_comer = false;
681 
682 	return ret;
683 }
684 
685 /*
686  * This holds any device nodes that we requested resources for,
687  * so that we may easily release resources in the error path.
688  */
689 struct sunxi_mc_smp_nodes {
690 	struct device_node *prcm_node;
691 	struct device_node *cpucfg_node;
692 	struct device_node *sram_node;
693 	struct device_node *r_cpucfg_node;
694 };
695 
696 /* This structure holds SoC-specific bits tied to an enable-method string. */
697 struct sunxi_mc_smp_data {
698 	const char *enable_method;
699 	int (*get_smp_nodes)(struct sunxi_mc_smp_nodes *nodes);
700 	bool is_a83t;
701 };
702 
703 static void __init sunxi_mc_smp_put_nodes(struct sunxi_mc_smp_nodes *nodes)
704 {
705 	of_node_put(nodes->prcm_node);
706 	of_node_put(nodes->cpucfg_node);
707 	of_node_put(nodes->sram_node);
708 	of_node_put(nodes->r_cpucfg_node);
709 	memset(nodes, 0, sizeof(*nodes));
710 }
711 
712 static int __init sun9i_a80_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
713 {
714 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
715 						   "allwinner,sun9i-a80-prcm");
716 	if (!nodes->prcm_node) {
717 		pr_err("%s: PRCM not available\n", __func__);
718 		return -ENODEV;
719 	}
720 
721 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
722 						     "allwinner,sun9i-a80-cpucfg");
723 	if (!nodes->cpucfg_node) {
724 		pr_err("%s: CPUCFG not available\n", __func__);
725 		return -ENODEV;
726 	}
727 
728 	nodes->sram_node = of_find_compatible_node(NULL, NULL,
729 						   "allwinner,sun9i-a80-smp-sram");
730 	if (!nodes->sram_node) {
731 		pr_err("%s: Secure SRAM not available\n", __func__);
732 		return -ENODEV;
733 	}
734 
735 	return 0;
736 }
737 
738 static int __init sun8i_a83t_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
739 {
740 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
741 						   "allwinner,sun8i-a83t-r-ccu");
742 	if (!nodes->prcm_node) {
743 		pr_err("%s: PRCM not available\n", __func__);
744 		return -ENODEV;
745 	}
746 
747 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
748 						     "allwinner,sun8i-a83t-cpucfg");
749 	if (!nodes->cpucfg_node) {
750 		pr_err("%s: CPUCFG not available\n", __func__);
751 		return -ENODEV;
752 	}
753 
754 	nodes->r_cpucfg_node = of_find_compatible_node(NULL, NULL,
755 						       "allwinner,sun8i-a83t-r-cpucfg");
756 	if (!nodes->r_cpucfg_node) {
757 		pr_err("%s: RCPUCFG not available\n", __func__);
758 		return -ENODEV;
759 	}
760 
761 	return 0;
762 }
763 
764 static const struct sunxi_mc_smp_data sunxi_mc_smp_data[] __initconst = {
765 	{
766 		.enable_method	= "allwinner,sun9i-a80-smp",
767 		.get_smp_nodes	= sun9i_a80_get_smp_nodes,
768 	},
769 	{
770 		.enable_method	= "allwinner,sun8i-a83t-smp",
771 		.get_smp_nodes	= sun8i_a83t_get_smp_nodes,
772 		.is_a83t	= true,
773 	},
774 };
775 
776 static int __init sunxi_mc_smp_init(void)
777 {
778 	struct sunxi_mc_smp_nodes nodes = { 0 };
779 	struct device_node *node;
780 	struct resource res;
781 	void __iomem *addr;
782 	int i, ret;
783 
784 	/*
785 	 * Don't bother checking the "cpus" node, as an enable-method
786 	 * property in that node is undocumented.
787 	 */
788 	node = of_cpu_device_node_get(0);
789 	if (!node)
790 		return -ENODEV;
791 
792 	/*
793 	 * We can't actually use the enable-method magic in the kernel.
794 	 * Our loopback / trampoline code uses the CPU suspend framework,
795 	 * which requires the identity mapping be available. It would not
796 	 * yet be available if we used the .init_cpus or .prepare_cpus
797 	 * callbacks in smp_operations, which we would use if we were to
798 	 * use CPU_METHOD_OF_DECLARE
799 	 */
800 	for (i = 0; i < ARRAY_SIZE(sunxi_mc_smp_data); i++) {
801 		ret = of_property_match_string(node, "enable-method",
802 					       sunxi_mc_smp_data[i].enable_method);
803 		if (!ret)
804 			break;
805 	}
806 
807 	is_a83t = sunxi_mc_smp_data[i].is_a83t;
808 
809 	of_node_put(node);
810 	if (ret)
811 		return -ENODEV;
812 
813 	if (!sunxi_mc_smp_cpu_table_init())
814 		return -EINVAL;
815 
816 	if (!cci_probed()) {
817 		pr_err("%s: CCI-400 not available\n", __func__);
818 		return -ENODEV;
819 	}
820 
821 	/* Get needed device tree nodes */
822 	ret = sunxi_mc_smp_data[i].get_smp_nodes(&nodes);
823 	if (ret)
824 		goto err_put_nodes;
825 
826 	/*
827 	 * Unfortunately we can not request the I/O region for the PRCM.
828 	 * It is shared with the PRCM clock.
829 	 */
830 	prcm_base = of_iomap(nodes.prcm_node, 0);
831 	if (!prcm_base) {
832 		pr_err("%s: failed to map PRCM registers\n", __func__);
833 		ret = -ENOMEM;
834 		goto err_put_nodes;
835 	}
836 
837 	cpucfg_base = of_io_request_and_map(nodes.cpucfg_node, 0,
838 					    "sunxi-mc-smp");
839 	if (IS_ERR(cpucfg_base)) {
840 		ret = PTR_ERR(cpucfg_base);
841 		pr_err("%s: failed to map CPUCFG registers: %d\n",
842 		       __func__, ret);
843 		goto err_unmap_prcm;
844 	}
845 
846 	if (is_a83t) {
847 		r_cpucfg_base = of_io_request_and_map(nodes.r_cpucfg_node,
848 						      0, "sunxi-mc-smp");
849 		if (IS_ERR(r_cpucfg_base)) {
850 			ret = PTR_ERR(r_cpucfg_base);
851 			pr_err("%s: failed to map R-CPUCFG registers\n",
852 			       __func__);
853 			goto err_unmap_release_cpucfg;
854 		}
855 	} else {
856 		sram_b_smp_base = of_io_request_and_map(nodes.sram_node, 0,
857 							"sunxi-mc-smp");
858 		if (IS_ERR(sram_b_smp_base)) {
859 			ret = PTR_ERR(sram_b_smp_base);
860 			pr_err("%s: failed to map secure SRAM\n", __func__);
861 			goto err_unmap_release_cpucfg;
862 		}
863 	}
864 
865 	/* Configure CCI-400 for boot cluster */
866 	ret = sunxi_mc_smp_loopback();
867 	if (ret) {
868 		pr_err("%s: failed to configure boot cluster: %d\n",
869 		       __func__, ret);
870 		goto err_unmap_release_sram_rcpucfg;
871 	}
872 
873 	/* We don't need the device nodes anymore */
874 	sunxi_mc_smp_put_nodes(&nodes);
875 
876 	/* Set the hardware entry point address */
877 	if (is_a83t)
878 		addr = r_cpucfg_base + R_CPUCFG_CPU_SOFT_ENTRY_REG;
879 	else
880 		addr = prcm_base + PRCM_CPU_SOFT_ENTRY_REG;
881 	writel(__pa_symbol(sunxi_mc_smp_secondary_startup), addr);
882 
883 	/* Actually enable multi cluster SMP */
884 	smp_set_ops(&sunxi_mc_smp_smp_ops);
885 
886 	pr_info("sunxi multi cluster SMP support installed\n");
887 
888 	return 0;
889 
890 err_unmap_release_sram_rcpucfg:
891 	if (is_a83t) {
892 		iounmap(r_cpucfg_base);
893 		of_address_to_resource(nodes.r_cpucfg_node, 0, &res);
894 	} else {
895 		iounmap(sram_b_smp_base);
896 		of_address_to_resource(nodes.sram_node, 0, &res);
897 	}
898 	release_mem_region(res.start, resource_size(&res));
899 err_unmap_release_cpucfg:
900 	iounmap(cpucfg_base);
901 	of_address_to_resource(nodes.cpucfg_node, 0, &res);
902 	release_mem_region(res.start, resource_size(&res));
903 err_unmap_prcm:
904 	iounmap(prcm_base);
905 err_put_nodes:
906 	sunxi_mc_smp_put_nodes(&nodes);
907 	return ret;
908 }
909 
910 early_initcall(sunxi_mc_smp_init);
911