xref: /openbmc/linux/arch/arm/mach-sunxi/mc_smp.c (revision ba61bb17)
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 
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 	return of_device_is_compatible(node, "arm,cortex-a15");
111 }
112 
113 static int sunxi_cpu_power_switch_set(unsigned int cpu, unsigned int cluster,
114 				      bool enable)
115 {
116 	u32 reg;
117 
118 	/* control sequence from Allwinner A80 user manual v1.2 PRCM section */
119 	reg = readl(prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
120 	if (enable) {
121 		if (reg == 0x00) {
122 			pr_debug("power clamp for cluster %u cpu %u already open\n",
123 				 cluster, cpu);
124 			return 0;
125 		}
126 
127 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
128 		udelay(10);
129 		writel(0xfe, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
130 		udelay(10);
131 		writel(0xf8, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
132 		udelay(10);
133 		writel(0xf0, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
134 		udelay(10);
135 		writel(0x00, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
136 		udelay(10);
137 	} else {
138 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
139 		udelay(10);
140 	}
141 
142 	return 0;
143 }
144 
145 static void sunxi_cpu0_hotplug_support_set(bool enable)
146 {
147 	if (enable) {
148 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC0, sram_b_smp_base);
149 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC1, sram_b_smp_base + 0x4);
150 	} else {
151 		writel(0x0, sram_b_smp_base);
152 		writel(0x0, sram_b_smp_base + 0x4);
153 	}
154 }
155 
156 static int sunxi_cpu_powerup(unsigned int cpu, unsigned int cluster)
157 {
158 	u32 reg;
159 
160 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
161 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
162 		return -EINVAL;
163 
164 	/* Set hotplug support magic flags for cpu0 */
165 	if (cluster == 0 && cpu == 0)
166 		sunxi_cpu0_hotplug_support_set(true);
167 
168 	/* assert processor power-on reset */
169 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
170 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE(cpu);
171 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
172 
173 	if (is_a83t) {
174 		/* assert cpu power-on reset */
175 		reg  = readl(r_cpucfg_base +
176 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
177 		reg &= ~(R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(cpu));
178 		writel(reg, r_cpucfg_base +
179 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
180 		udelay(10);
181 	}
182 
183 	/* Cortex-A7: hold L1 reset disable signal low */
184 	if (!sunxi_core_is_cortex_a15(cpu, cluster)) {
185 		reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
186 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE(cpu);
187 		writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
188 	}
189 
190 	/* assert processor related resets */
191 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
192 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
193 
194 	/*
195 	 * Allwinner code also asserts resets for NEON on A15. According
196 	 * to ARM manuals, asserting power-on reset is sufficient.
197 	 */
198 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
199 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
200 
201 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
202 
203 	/* open power switch */
204 	sunxi_cpu_power_switch_set(cpu, cluster, true);
205 
206 	/* Handle A83T bit swap */
207 	if (is_a83t) {
208 		if (cpu == 0)
209 			cpu = 4;
210 	}
211 
212 	/* clear processor power gate */
213 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
214 	reg &= ~PRCM_PWROFF_GATING_REG_CORE(cpu);
215 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
216 	udelay(20);
217 
218 	/* Handle A83T bit swap */
219 	if (is_a83t) {
220 		if (cpu == 4)
221 			cpu = 0;
222 	}
223 
224 	/* de-assert processor power-on reset */
225 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
226 	reg |= PRCM_CPU_PO_RST_CTRL_CORE(cpu);
227 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
228 
229 	if (is_a83t) {
230 		reg  = readl(r_cpucfg_base +
231 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
232 		reg |= R_CPUCFG_CLUSTER_PO_RST_CTRL_CORE(cpu);
233 		writel(reg, r_cpucfg_base +
234 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
235 		udelay(10);
236 	}
237 
238 	/* de-assert all processor resets */
239 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
240 	reg |= CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
241 	reg |= CPUCFG_CX_RST_CTRL_CORE_RST(cpu);
242 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
243 		reg |= CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
244 	else
245 		reg |= CPUCFG_CX_RST_CTRL_CX_RST(cpu); /* NEON */
246 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
247 
248 	return 0;
249 }
250 
251 static int sunxi_cluster_powerup(unsigned int cluster)
252 {
253 	u32 reg;
254 
255 	pr_debug("%s: cluster %u\n", __func__, cluster);
256 	if (cluster >= SUNXI_NR_CLUSTERS)
257 		return -EINVAL;
258 
259 	/* For A83T, assert cluster cores resets */
260 	if (is_a83t) {
261 		reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
262 		reg &= ~CPUCFG_CX_RST_CTRL_CORE_RST_ALL;   /* Core Reset    */
263 		writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
264 		udelay(10);
265 	}
266 
267 	/* assert ACINACTM */
268 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
269 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
270 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
271 
272 	/* assert cluster processor power-on resets */
273 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
274 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE_ALL;
275 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
276 
277 	/* assert cluster cores resets */
278 	if (is_a83t) {
279 		reg  = readl(r_cpucfg_base +
280 			     R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
281 		reg &= ~CPUCFG_CX_RST_CTRL_CORE_RST_ALL;
282 		writel(reg, r_cpucfg_base +
283 		       R_CPUCFG_CLUSTER_PO_RST_CTRL(cluster));
284 		udelay(10);
285 	}
286 
287 	/* assert cluster resets */
288 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
289 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
290 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST_ALL;
291 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
292 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
293 
294 	/*
295 	 * Allwinner code also asserts resets for NEON on A15. According
296 	 * to ARM manuals, asserting power-on reset is sufficient.
297 	 */
298 	if (!sunxi_core_is_cortex_a15(0, cluster))
299 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST_ALL;
300 
301 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
302 
303 	/* hold L1/L2 reset disable signals low */
304 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
305 	if (sunxi_core_is_cortex_a15(0, cluster)) {
306 		/* Cortex-A15: hold L2RSTDISABLE low */
307 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A15;
308 	} else {
309 		/* Cortex-A7: hold L1RSTDISABLE and L2RSTDISABLE low */
310 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE_ALL;
311 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A7;
312 	}
313 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
314 
315 	/* clear cluster power gate */
316 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
317 	if (is_a83t)
318 		reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
319 	else
320 		reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
321 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
322 	udelay(20);
323 
324 	/* de-assert cluster resets */
325 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
326 	reg |= CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
327 	reg |= CPUCFG_CX_RST_CTRL_H_RST;
328 	reg |= CPUCFG_CX_RST_CTRL_L2_RST;
329 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
330 
331 	/* de-assert ACINACTM */
332 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
333 	reg &= ~CPUCFG_CX_CTRL_REG1_ACINACTM;
334 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
335 
336 	return 0;
337 }
338 
339 /*
340  * This bit is shared between the initial nocache_trampoline call to
341  * enable CCI-400 and proper cluster cache disable before power down.
342  */
343 static void sunxi_cluster_cache_disable_without_axi(void)
344 {
345 	if (read_cpuid_part() == ARM_CPU_PART_CORTEX_A15) {
346 		/*
347 		 * On the Cortex-A15 we need to disable
348 		 * L2 prefetching before flushing the cache.
349 		 */
350 		asm volatile(
351 		"mcr	p15, 1, %0, c15, c0, 3\n"
352 		"isb\n"
353 		"dsb"
354 		: : "r" (0x400));
355 	}
356 
357 	/* Flush all cache levels for this cluster. */
358 	v7_exit_coherency_flush(all);
359 
360 	/*
361 	 * Disable cluster-level coherency by masking
362 	 * incoming snoops and DVM messages:
363 	 */
364 	cci_disable_port_by_cpu(read_cpuid_mpidr());
365 }
366 
367 static int sunxi_mc_smp_cpu_table[SUNXI_NR_CLUSTERS][SUNXI_CPUS_PER_CLUSTER];
368 int sunxi_mc_smp_first_comer;
369 
370 static DEFINE_SPINLOCK(boot_lock);
371 
372 static bool sunxi_mc_smp_cluster_is_down(unsigned int cluster)
373 {
374 	int i;
375 
376 	for (i = 0; i < SUNXI_CPUS_PER_CLUSTER; i++)
377 		if (sunxi_mc_smp_cpu_table[cluster][i])
378 			return false;
379 	return true;
380 }
381 
382 static void sunxi_mc_smp_secondary_init(unsigned int cpu)
383 {
384 	/* Clear hotplug support magic flags for cpu0 */
385 	if (cpu == 0)
386 		sunxi_cpu0_hotplug_support_set(false);
387 }
388 
389 static int sunxi_mc_smp_boot_secondary(unsigned int l_cpu, struct task_struct *idle)
390 {
391 	unsigned int mpidr, cpu, cluster;
392 
393 	mpidr = cpu_logical_map(l_cpu);
394 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
395 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
396 
397 	if (!cpucfg_base)
398 		return -ENODEV;
399 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER)
400 		return -EINVAL;
401 
402 	spin_lock_irq(&boot_lock);
403 
404 	if (sunxi_mc_smp_cpu_table[cluster][cpu])
405 		goto out;
406 
407 	if (sunxi_mc_smp_cluster_is_down(cluster)) {
408 		sunxi_mc_smp_first_comer = true;
409 		sunxi_cluster_powerup(cluster);
410 	} else {
411 		sunxi_mc_smp_first_comer = false;
412 	}
413 
414 	/* This is read by incoming CPUs with their cache and MMU disabled */
415 	sync_cache_w(&sunxi_mc_smp_first_comer);
416 	sunxi_cpu_powerup(cpu, cluster);
417 
418 out:
419 	sunxi_mc_smp_cpu_table[cluster][cpu]++;
420 	spin_unlock_irq(&boot_lock);
421 
422 	return 0;
423 }
424 
425 #ifdef CONFIG_HOTPLUG_CPU
426 static void sunxi_cluster_cache_disable(void)
427 {
428 	unsigned int cluster = MPIDR_AFFINITY_LEVEL(read_cpuid_mpidr(), 1);
429 	u32 reg;
430 
431 	pr_debug("%s: cluster %u\n", __func__, cluster);
432 
433 	sunxi_cluster_cache_disable_without_axi();
434 
435 	/* last man standing, assert ACINACTM */
436 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
437 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
438 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
439 }
440 
441 static void sunxi_mc_smp_cpu_die(unsigned int l_cpu)
442 {
443 	unsigned int mpidr, cpu, cluster;
444 	bool last_man;
445 
446 	mpidr = cpu_logical_map(l_cpu);
447 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
448 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
449 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
450 
451 	spin_lock(&boot_lock);
452 	sunxi_mc_smp_cpu_table[cluster][cpu]--;
453 	if (sunxi_mc_smp_cpu_table[cluster][cpu] == 1) {
454 		/* A power_up request went ahead of us. */
455 		pr_debug("%s: aborting due to a power up request\n",
456 			 __func__);
457 		spin_unlock(&boot_lock);
458 		return;
459 	} else if (sunxi_mc_smp_cpu_table[cluster][cpu] > 1) {
460 		pr_err("Cluster %d CPU%d boots multiple times\n",
461 		       cluster, cpu);
462 		BUG();
463 	}
464 
465 	last_man = sunxi_mc_smp_cluster_is_down(cluster);
466 	spin_unlock(&boot_lock);
467 
468 	gic_cpu_if_down(0);
469 	if (last_man)
470 		sunxi_cluster_cache_disable();
471 	else
472 		v7_exit_coherency_flush(louis);
473 
474 	for (;;)
475 		wfi();
476 }
477 
478 static int sunxi_cpu_powerdown(unsigned int cpu, unsigned int cluster)
479 {
480 	u32 reg;
481 
482 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
483 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
484 		return -EINVAL;
485 
486 	/* gate processor power */
487 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
488 	reg |= PRCM_PWROFF_GATING_REG_CORE(cpu);
489 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
490 	udelay(20);
491 
492 	/* close power switch */
493 	sunxi_cpu_power_switch_set(cpu, cluster, false);
494 
495 	return 0;
496 }
497 
498 static int sunxi_cluster_powerdown(unsigned int cluster)
499 {
500 	u32 reg;
501 
502 	pr_debug("%s: cluster %u\n", __func__, cluster);
503 	if (cluster >= SUNXI_NR_CLUSTERS)
504 		return -EINVAL;
505 
506 	/* assert cluster resets or system will hang */
507 	pr_debug("%s: assert cluster reset\n", __func__);
508 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
509 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
510 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
511 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
512 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
513 
514 	/* gate cluster power */
515 	pr_debug("%s: gate cluster power\n", __func__);
516 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
517 	if (is_a83t)
518 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN8I;
519 	else
520 		reg |= PRCM_PWROFF_GATING_REG_CLUSTER_SUN9I;
521 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
522 	udelay(20);
523 
524 	return 0;
525 }
526 
527 static int sunxi_mc_smp_cpu_kill(unsigned int l_cpu)
528 {
529 	unsigned int mpidr, cpu, cluster;
530 	unsigned int tries, count;
531 	int ret = 0;
532 	u32 reg;
533 
534 	mpidr = cpu_logical_map(l_cpu);
535 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
536 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
537 
538 	/* This should never happen */
539 	if (WARN_ON(cluster >= SUNXI_NR_CLUSTERS ||
540 		    cpu >= SUNXI_CPUS_PER_CLUSTER))
541 		return 0;
542 
543 	/* wait for CPU core to die and enter WFI */
544 	count = TIMEOUT_USEC / POLL_USEC;
545 	spin_lock_irq(&boot_lock);
546 	for (tries = 0; tries < count; tries++) {
547 		spin_unlock_irq(&boot_lock);
548 		usleep_range(POLL_USEC / 2, POLL_USEC);
549 		spin_lock_irq(&boot_lock);
550 
551 		/*
552 		 * If the user turns off a bunch of cores at the same
553 		 * time, the kernel might call cpu_kill before some of
554 		 * them are ready. This is because boot_lock serializes
555 		 * both cpu_die and cpu_kill callbacks. Either one could
556 		 * run first. We should wait for cpu_die to complete.
557 		 */
558 		if (sunxi_mc_smp_cpu_table[cluster][cpu])
559 			continue;
560 
561 		reg = readl(cpucfg_base + CPUCFG_CX_STATUS(cluster));
562 		if (reg & CPUCFG_CX_STATUS_STANDBYWFI(cpu))
563 			break;
564 	}
565 
566 	if (tries >= count) {
567 		ret = ETIMEDOUT;
568 		goto out;
569 	}
570 
571 	/* power down CPU core */
572 	sunxi_cpu_powerdown(cpu, cluster);
573 
574 	if (!sunxi_mc_smp_cluster_is_down(cluster))
575 		goto out;
576 
577 	/* wait for cluster L2 WFI */
578 	ret = readl_poll_timeout(cpucfg_base + CPUCFG_CX_STATUS(cluster), reg,
579 				 reg & CPUCFG_CX_STATUS_STANDBYWFIL2,
580 				 POLL_USEC, TIMEOUT_USEC);
581 	if (ret) {
582 		/*
583 		 * Ignore timeout on the cluster. Leaving the cluster on
584 		 * will not affect system execution, just use a bit more
585 		 * power. But returning an error here will only confuse
586 		 * the user as the CPU has already been shutdown.
587 		 */
588 		ret = 0;
589 		goto out;
590 	}
591 
592 	/* Power down cluster */
593 	sunxi_cluster_powerdown(cluster);
594 
595 out:
596 	spin_unlock_irq(&boot_lock);
597 	pr_debug("%s: cluster %u cpu %u powerdown: %d\n",
598 		 __func__, cluster, cpu, ret);
599 	return !ret;
600 }
601 
602 static bool sunxi_mc_smp_cpu_can_disable(unsigned int cpu)
603 {
604 	/* CPU0 hotplug not handled for sun8i-a83t */
605 	if (is_a83t)
606 		if (cpu == 0)
607 			return false;
608 	return true;
609 }
610 #endif
611 
612 static const struct smp_operations sunxi_mc_smp_smp_ops __initconst = {
613 	.smp_secondary_init	= sunxi_mc_smp_secondary_init,
614 	.smp_boot_secondary	= sunxi_mc_smp_boot_secondary,
615 #ifdef CONFIG_HOTPLUG_CPU
616 	.cpu_die		= sunxi_mc_smp_cpu_die,
617 	.cpu_kill		= sunxi_mc_smp_cpu_kill,
618 	.cpu_can_disable	= sunxi_mc_smp_cpu_can_disable,
619 #endif
620 };
621 
622 static bool __init sunxi_mc_smp_cpu_table_init(void)
623 {
624 	unsigned int mpidr, cpu, cluster;
625 
626 	mpidr = read_cpuid_mpidr();
627 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
628 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
629 
630 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER) {
631 		pr_err("%s: boot CPU is out of bounds!\n", __func__);
632 		return false;
633 	}
634 	sunxi_mc_smp_cpu_table[cluster][cpu] = 1;
635 	return true;
636 }
637 
638 /*
639  * Adapted from arch/arm/common/mc_smp_entry.c
640  *
641  * We need the trampoline code to enable CCI-400 on the first cluster
642  */
643 typedef typeof(cpu_reset) phys_reset_t;
644 
645 static int __init nocache_trampoline(unsigned long __unused)
646 {
647 	phys_reset_t phys_reset;
648 
649 	setup_mm_for_reboot();
650 	sunxi_cluster_cache_disable_without_axi();
651 
652 	phys_reset = (phys_reset_t)(unsigned long)__pa_symbol(cpu_reset);
653 	phys_reset(__pa_symbol(sunxi_mc_smp_resume), false);
654 	BUG();
655 }
656 
657 static int __init sunxi_mc_smp_loopback(void)
658 {
659 	int ret;
660 
661 	/*
662 	 * We're going to soft-restart the current CPU through the
663 	 * low-level MCPM code by leveraging the suspend/resume
664 	 * infrastructure. Let's play it safe by using cpu_pm_enter()
665 	 * in case the CPU init code path resets the VFP or similar.
666 	 */
667 	sunxi_mc_smp_first_comer = true;
668 	local_irq_disable();
669 	local_fiq_disable();
670 	ret = cpu_pm_enter();
671 	if (!ret) {
672 		ret = cpu_suspend(0, nocache_trampoline);
673 		cpu_pm_exit();
674 	}
675 	local_fiq_enable();
676 	local_irq_enable();
677 	sunxi_mc_smp_first_comer = false;
678 
679 	return ret;
680 }
681 
682 /*
683  * This holds any device nodes that we requested resources for,
684  * so that we may easily release resources in the error path.
685  */
686 struct sunxi_mc_smp_nodes {
687 	struct device_node *prcm_node;
688 	struct device_node *cpucfg_node;
689 	struct device_node *sram_node;
690 	struct device_node *r_cpucfg_node;
691 };
692 
693 /* This structure holds SoC-specific bits tied to an enable-method string. */
694 struct sunxi_mc_smp_data {
695 	const char *enable_method;
696 	int (*get_smp_nodes)(struct sunxi_mc_smp_nodes *nodes);
697 	bool is_a83t;
698 };
699 
700 static void __init sunxi_mc_smp_put_nodes(struct sunxi_mc_smp_nodes *nodes)
701 {
702 	of_node_put(nodes->prcm_node);
703 	of_node_put(nodes->cpucfg_node);
704 	of_node_put(nodes->sram_node);
705 	of_node_put(nodes->r_cpucfg_node);
706 	memset(nodes, 0, sizeof(*nodes));
707 }
708 
709 static int __init sun9i_a80_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
710 {
711 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
712 						   "allwinner,sun9i-a80-prcm");
713 	if (!nodes->prcm_node) {
714 		pr_err("%s: PRCM not available\n", __func__);
715 		return -ENODEV;
716 	}
717 
718 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
719 						     "allwinner,sun9i-a80-cpucfg");
720 	if (!nodes->cpucfg_node) {
721 		pr_err("%s: CPUCFG not available\n", __func__);
722 		return -ENODEV;
723 	}
724 
725 	nodes->sram_node = of_find_compatible_node(NULL, NULL,
726 						   "allwinner,sun9i-a80-smp-sram");
727 	if (!nodes->sram_node) {
728 		pr_err("%s: Secure SRAM not available\n", __func__);
729 		return -ENODEV;
730 	}
731 
732 	return 0;
733 }
734 
735 static int __init sun8i_a83t_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
736 {
737 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
738 						   "allwinner,sun8i-a83t-r-ccu");
739 	if (!nodes->prcm_node) {
740 		pr_err("%s: PRCM not available\n", __func__);
741 		return -ENODEV;
742 	}
743 
744 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
745 						     "allwinner,sun8i-a83t-cpucfg");
746 	if (!nodes->cpucfg_node) {
747 		pr_err("%s: CPUCFG not available\n", __func__);
748 		return -ENODEV;
749 	}
750 
751 	nodes->r_cpucfg_node = of_find_compatible_node(NULL, NULL,
752 						       "allwinner,sun8i-a83t-r-cpucfg");
753 	if (!nodes->r_cpucfg_node) {
754 		pr_err("%s: RCPUCFG not available\n", __func__);
755 		return -ENODEV;
756 	}
757 
758 	return 0;
759 }
760 
761 static const struct sunxi_mc_smp_data sunxi_mc_smp_data[] __initconst = {
762 	{
763 		.enable_method	= "allwinner,sun9i-a80-smp",
764 		.get_smp_nodes	= sun9i_a80_get_smp_nodes,
765 	},
766 	{
767 		.enable_method	= "allwinner,sun8i-a83t-smp",
768 		.get_smp_nodes	= sun8i_a83t_get_smp_nodes,
769 		.is_a83t	= true,
770 	},
771 };
772 
773 static int __init sunxi_mc_smp_init(void)
774 {
775 	struct sunxi_mc_smp_nodes nodes = { 0 };
776 	struct device_node *node;
777 	struct resource res;
778 	void __iomem *addr;
779 	int i, ret;
780 
781 	/*
782 	 * Don't bother checking the "cpus" node, as an enable-method
783 	 * property in that node is undocumented.
784 	 */
785 	node = of_cpu_device_node_get(0);
786 	if (!node)
787 		return -ENODEV;
788 
789 	/*
790 	 * We can't actually use the enable-method magic in the kernel.
791 	 * Our loopback / trampoline code uses the CPU suspend framework,
792 	 * which requires the identity mapping be available. It would not
793 	 * yet be available if we used the .init_cpus or .prepare_cpus
794 	 * callbacks in smp_operations, which we would use if we were to
795 	 * use CPU_METHOD_OF_DECLARE
796 	 */
797 	for (i = 0; i < ARRAY_SIZE(sunxi_mc_smp_data); i++) {
798 		ret = of_property_match_string(node, "enable-method",
799 					       sunxi_mc_smp_data[i].enable_method);
800 		if (!ret)
801 			break;
802 	}
803 
804 	is_a83t = sunxi_mc_smp_data[i].is_a83t;
805 
806 	of_node_put(node);
807 	if (ret)
808 		return -ENODEV;
809 
810 	if (!sunxi_mc_smp_cpu_table_init())
811 		return -EINVAL;
812 
813 	if (!cci_probed()) {
814 		pr_err("%s: CCI-400 not available\n", __func__);
815 		return -ENODEV;
816 	}
817 
818 	/* Get needed device tree nodes */
819 	ret = sunxi_mc_smp_data[i].get_smp_nodes(&nodes);
820 	if (ret)
821 		goto err_put_nodes;
822 
823 	/*
824 	 * Unfortunately we can not request the I/O region for the PRCM.
825 	 * It is shared with the PRCM clock.
826 	 */
827 	prcm_base = of_iomap(nodes.prcm_node, 0);
828 	if (!prcm_base) {
829 		pr_err("%s: failed to map PRCM registers\n", __func__);
830 		ret = -ENOMEM;
831 		goto err_put_nodes;
832 	}
833 
834 	cpucfg_base = of_io_request_and_map(nodes.cpucfg_node, 0,
835 					    "sunxi-mc-smp");
836 	if (IS_ERR(cpucfg_base)) {
837 		ret = PTR_ERR(cpucfg_base);
838 		pr_err("%s: failed to map CPUCFG registers: %d\n",
839 		       __func__, ret);
840 		goto err_unmap_prcm;
841 	}
842 
843 	if (is_a83t) {
844 		r_cpucfg_base = of_io_request_and_map(nodes.r_cpucfg_node,
845 						      0, "sunxi-mc-smp");
846 		if (IS_ERR(r_cpucfg_base)) {
847 			ret = PTR_ERR(r_cpucfg_base);
848 			pr_err("%s: failed to map R-CPUCFG registers\n",
849 			       __func__);
850 			goto err_unmap_release_cpucfg;
851 		}
852 	} else {
853 		sram_b_smp_base = of_io_request_and_map(nodes.sram_node, 0,
854 							"sunxi-mc-smp");
855 		if (IS_ERR(sram_b_smp_base)) {
856 			ret = PTR_ERR(sram_b_smp_base);
857 			pr_err("%s: failed to map secure SRAM\n", __func__);
858 			goto err_unmap_release_cpucfg;
859 		}
860 	}
861 
862 	/* Configure CCI-400 for boot cluster */
863 	ret = sunxi_mc_smp_loopback();
864 	if (ret) {
865 		pr_err("%s: failed to configure boot cluster: %d\n",
866 		       __func__, ret);
867 		goto err_unmap_release_sram_rcpucfg;
868 	}
869 
870 	/* We don't need the device nodes anymore */
871 	sunxi_mc_smp_put_nodes(&nodes);
872 
873 	/* Set the hardware entry point address */
874 	if (is_a83t)
875 		addr = r_cpucfg_base + R_CPUCFG_CPU_SOFT_ENTRY_REG;
876 	else
877 		addr = prcm_base + PRCM_CPU_SOFT_ENTRY_REG;
878 	writel(__pa_symbol(sunxi_mc_smp_secondary_startup), addr);
879 
880 	/* Actually enable multi cluster SMP */
881 	smp_set_ops(&sunxi_mc_smp_smp_ops);
882 
883 	pr_info("sunxi multi cluster SMP support installed\n");
884 
885 	return 0;
886 
887 err_unmap_release_sram_rcpucfg:
888 	if (is_a83t) {
889 		iounmap(r_cpucfg_base);
890 		of_address_to_resource(nodes.r_cpucfg_node, 0, &res);
891 	} else {
892 		iounmap(sram_b_smp_base);
893 		of_address_to_resource(nodes.sram_node, 0, &res);
894 	}
895 	release_mem_region(res.start, resource_size(&res));
896 err_unmap_release_cpucfg:
897 	iounmap(cpucfg_base);
898 	of_address_to_resource(nodes.cpucfg_node, 0, &res);
899 	release_mem_region(res.start, resource_size(&res));
900 err_unmap_prcm:
901 	iounmap(prcm_base);
902 err_put_nodes:
903 	sunxi_mc_smp_put_nodes(&nodes);
904 	return ret;
905 }
906 
907 early_initcall(sunxi_mc_smp_init);
908