xref: /openbmc/linux/arch/arm/mach-sunxi/mc_smp.c (revision 5ef12cb4a3a78ffb331c03a795a15eea4ae35155)
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 
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 #define PRCM_PWROFF_GATING_REG_CLUSTER	BIT(4)
64 #define PRCM_PWROFF_GATING_REG_CORE(n)	BIT(n)
65 #define PRCM_PWR_SWITCH_REG(c, cpu)	(0x140 + 0x10 * (c) + 0x4 * (cpu))
66 #define PRCM_CPU_SOFT_ENTRY_REG		0x164
67 
68 #define CPU0_SUPPORT_HOTPLUG_MAGIC0	0xFA50392F
69 #define CPU0_SUPPORT_HOTPLUG_MAGIC1	0x790DCA3A
70 
71 static void __iomem *cpucfg_base;
72 static void __iomem *prcm_base;
73 static void __iomem *sram_b_smp_base;
74 
75 static bool sunxi_core_is_cortex_a15(unsigned int core, unsigned int cluster)
76 {
77 	struct device_node *node;
78 	int cpu = cluster * SUNXI_CPUS_PER_CLUSTER + core;
79 
80 	node = of_cpu_device_node_get(cpu);
81 
82 	/* In case of_cpu_device_node_get fails */
83 	if (!node)
84 		node = of_get_cpu_node(cpu, NULL);
85 
86 	if (!node) {
87 		/*
88 		 * There's no point in returning an error, since we
89 		 * would be mid way in a core or cluster power sequence.
90 		 */
91 		pr_err("%s: Couldn't get CPU cluster %u core %u device node\n",
92 		       __func__, cluster, core);
93 
94 		return false;
95 	}
96 
97 	return of_device_is_compatible(node, "arm,cortex-a15");
98 }
99 
100 static int sunxi_cpu_power_switch_set(unsigned int cpu, unsigned int cluster,
101 				      bool enable)
102 {
103 	u32 reg;
104 
105 	/* control sequence from Allwinner A80 user manual v1.2 PRCM section */
106 	reg = readl(prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
107 	if (enable) {
108 		if (reg == 0x00) {
109 			pr_debug("power clamp for cluster %u cpu %u already open\n",
110 				 cluster, cpu);
111 			return 0;
112 		}
113 
114 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
115 		udelay(10);
116 		writel(0xfe, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
117 		udelay(10);
118 		writel(0xf8, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
119 		udelay(10);
120 		writel(0xf0, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
121 		udelay(10);
122 		writel(0x00, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
123 		udelay(10);
124 	} else {
125 		writel(0xff, prcm_base + PRCM_PWR_SWITCH_REG(cluster, cpu));
126 		udelay(10);
127 	}
128 
129 	return 0;
130 }
131 
132 static void sunxi_cpu0_hotplug_support_set(bool enable)
133 {
134 	if (enable) {
135 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC0, sram_b_smp_base);
136 		writel(CPU0_SUPPORT_HOTPLUG_MAGIC1, sram_b_smp_base + 0x4);
137 	} else {
138 		writel(0x0, sram_b_smp_base);
139 		writel(0x0, sram_b_smp_base + 0x4);
140 	}
141 }
142 
143 static int sunxi_cpu_powerup(unsigned int cpu, unsigned int cluster)
144 {
145 	u32 reg;
146 
147 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
148 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
149 		return -EINVAL;
150 
151 	/* Set hotplug support magic flags for cpu0 */
152 	if (cluster == 0 && cpu == 0)
153 		sunxi_cpu0_hotplug_support_set(true);
154 
155 	/* assert processor power-on reset */
156 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
157 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE(cpu);
158 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
159 
160 	/* Cortex-A7: hold L1 reset disable signal low */
161 	if (!sunxi_core_is_cortex_a15(cpu, cluster)) {
162 		reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
163 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE(cpu);
164 		writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
165 	}
166 
167 	/* assert processor related resets */
168 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
169 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
170 
171 	/*
172 	 * Allwinner code also asserts resets for NEON on A15. According
173 	 * to ARM manuals, asserting power-on reset is sufficient.
174 	 */
175 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
176 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
177 
178 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
179 
180 	/* open power switch */
181 	sunxi_cpu_power_switch_set(cpu, cluster, true);
182 
183 	/* clear processor power gate */
184 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
185 	reg &= ~PRCM_PWROFF_GATING_REG_CORE(cpu);
186 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
187 	udelay(20);
188 
189 	/* de-assert processor power-on reset */
190 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
191 	reg |= PRCM_CPU_PO_RST_CTRL_CORE(cpu);
192 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
193 
194 	/* de-assert all processor resets */
195 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
196 	reg |= CPUCFG_CX_RST_CTRL_DBG_RST(cpu);
197 	reg |= CPUCFG_CX_RST_CTRL_CORE_RST(cpu);
198 	if (!sunxi_core_is_cortex_a15(cpu, cluster))
199 		reg |= CPUCFG_CX_RST_CTRL_ETM_RST(cpu);
200 	else
201 		reg |= CPUCFG_CX_RST_CTRL_CX_RST(cpu); /* NEON */
202 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
203 
204 	return 0;
205 }
206 
207 static int sunxi_cluster_powerup(unsigned int cluster)
208 {
209 	u32 reg;
210 
211 	pr_debug("%s: cluster %u\n", __func__, cluster);
212 	if (cluster >= SUNXI_NR_CLUSTERS)
213 		return -EINVAL;
214 
215 	/* assert ACINACTM */
216 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
217 	reg |= CPUCFG_CX_CTRL_REG1_ACINACTM;
218 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
219 
220 	/* assert cluster processor power-on resets */
221 	reg = readl(prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
222 	reg &= ~PRCM_CPU_PO_RST_CTRL_CORE_ALL;
223 	writel(reg, prcm_base + PRCM_CPU_PO_RST_CTRL(cluster));
224 
225 	/* assert cluster resets */
226 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
227 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
228 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_RST_ALL;
229 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
230 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
231 
232 	/*
233 	 * Allwinner code also asserts resets for NEON on A15. According
234 	 * to ARM manuals, asserting power-on reset is sufficient.
235 	 */
236 	if (!sunxi_core_is_cortex_a15(0, cluster))
237 		reg &= ~CPUCFG_CX_RST_CTRL_ETM_RST_ALL;
238 
239 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
240 
241 	/* hold L1/L2 reset disable signals low */
242 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
243 	if (sunxi_core_is_cortex_a15(0, cluster)) {
244 		/* Cortex-A15: hold L2RSTDISABLE low */
245 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A15;
246 	} else {
247 		/* Cortex-A7: hold L1RSTDISABLE and L2RSTDISABLE low */
248 		reg &= ~CPUCFG_CX_CTRL_REG0_L1_RST_DISABLE_ALL;
249 		reg &= ~CPUCFG_CX_CTRL_REG0_L2_RST_DISABLE_A7;
250 	}
251 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG0(cluster));
252 
253 	/* clear cluster power gate */
254 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
255 	reg &= ~PRCM_PWROFF_GATING_REG_CLUSTER;
256 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
257 	udelay(20);
258 
259 	/* de-assert cluster resets */
260 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
261 	reg |= CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
262 	reg |= CPUCFG_CX_RST_CTRL_H_RST;
263 	reg |= CPUCFG_CX_RST_CTRL_L2_RST;
264 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
265 
266 	/* de-assert ACINACTM */
267 	reg = readl(cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
268 	reg &= ~CPUCFG_CX_CTRL_REG1_ACINACTM;
269 	writel(reg, cpucfg_base + CPUCFG_CX_CTRL_REG1(cluster));
270 
271 	return 0;
272 }
273 
274 /*
275  * This bit is shared between the initial nocache_trampoline call to
276  * enable CCI-400 and proper cluster cache disable before power down.
277  */
278 static void sunxi_cluster_cache_disable_without_axi(void)
279 {
280 	if (read_cpuid_part() == ARM_CPU_PART_CORTEX_A15) {
281 		/*
282 		 * On the Cortex-A15 we need to disable
283 		 * L2 prefetching before flushing the cache.
284 		 */
285 		asm volatile(
286 		"mcr	p15, 1, %0, c15, c0, 3\n"
287 		"isb\n"
288 		"dsb"
289 		: : "r" (0x400));
290 	}
291 
292 	/* Flush all cache levels for this cluster. */
293 	v7_exit_coherency_flush(all);
294 
295 	/*
296 	 * Disable cluster-level coherency by masking
297 	 * incoming snoops and DVM messages:
298 	 */
299 	cci_disable_port_by_cpu(read_cpuid_mpidr());
300 }
301 
302 static int sunxi_mc_smp_cpu_table[SUNXI_NR_CLUSTERS][SUNXI_CPUS_PER_CLUSTER];
303 static int sunxi_mc_smp_first_comer;
304 
305 /*
306  * Enable cluster-level coherency, in preparation for turning on the MMU.
307  *
308  * Also enable regional clock gating and L2 data latency settings for
309  * Cortex-A15. These settings are from the vendor kernel.
310  */
311 static void __naked sunxi_mc_smp_cluster_cache_enable(void)
312 {
313 	asm volatile (
314 		"mrc	p15, 0, r1, c0, c0, 0\n"
315 		"movw	r2, #" __stringify(ARM_CPU_PART_MASK & 0xffff) "\n"
316 		"movt	r2, #" __stringify(ARM_CPU_PART_MASK >> 16) "\n"
317 		"and	r1, r1, r2\n"
318 		"movw	r2, #" __stringify(ARM_CPU_PART_CORTEX_A15 & 0xffff) "\n"
319 		"movt	r2, #" __stringify(ARM_CPU_PART_CORTEX_A15 >> 16) "\n"
320 		"cmp	r1, r2\n"
321 		"bne	not_a15\n"
322 
323 		/* The following is Cortex-A15 specific */
324 
325 		/* ACTLR2: Enable CPU regional clock gates */
326 		"mrc p15, 1, r1, c15, c0, 4\n"
327 		"orr r1, r1, #(0x1<<31)\n"
328 		"mcr p15, 1, r1, c15, c0, 4\n"
329 
330 		/* L2ACTLR */
331 		"mrc p15, 1, r1, c15, c0, 0\n"
332 		/* Enable L2, GIC, and Timer regional clock gates */
333 		"orr r1, r1, #(0x1<<26)\n"
334 		/* Disable clean/evict from being pushed to external */
335 		"orr r1, r1, #(0x1<<3)\n"
336 		"mcr p15, 1, r1, c15, c0, 0\n"
337 
338 		/* L2CTRL: L2 data RAM latency */
339 		"mrc p15, 1, r1, c9, c0, 2\n"
340 		"bic r1, r1, #(0x7<<0)\n"
341 		"orr r1, r1, #(0x3<<0)\n"
342 		"mcr p15, 1, r1, c9, c0, 2\n"
343 
344 		/* End of Cortex-A15 specific setup */
345 		"not_a15:\n"
346 
347 		/* Get value of sunxi_mc_smp_first_comer */
348 		"adr	r1, first\n"
349 		"ldr	r0, [r1]\n"
350 		"ldr	r0, [r1, r0]\n"
351 
352 		/* Skip cci_enable_port_for_self if not first comer */
353 		"cmp	r0, #0\n"
354 		"bxeq	lr\n"
355 		"b	cci_enable_port_for_self\n"
356 
357 		".align 2\n"
358 		"first: .word sunxi_mc_smp_first_comer - .\n"
359 	);
360 }
361 
362 static void __naked sunxi_mc_smp_secondary_startup(void)
363 {
364 	asm volatile(
365 		"bl	sunxi_mc_smp_cluster_cache_enable\n"
366 		"b	secondary_startup"
367 		/* Let compiler know about sunxi_mc_smp_cluster_cache_enable */
368 		:: "i" (sunxi_mc_smp_cluster_cache_enable)
369 	);
370 }
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 
484 	pr_debug("%s: cluster %u cpu %u\n", __func__, cluster, cpu);
485 	if (cpu >= SUNXI_CPUS_PER_CLUSTER || cluster >= SUNXI_NR_CLUSTERS)
486 		return -EINVAL;
487 
488 	/* gate processor power */
489 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
490 	reg |= PRCM_PWROFF_GATING_REG_CORE(cpu);
491 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
492 	udelay(20);
493 
494 	/* close power switch */
495 	sunxi_cpu_power_switch_set(cpu, cluster, false);
496 
497 	return 0;
498 }
499 
500 static int sunxi_cluster_powerdown(unsigned int cluster)
501 {
502 	u32 reg;
503 
504 	pr_debug("%s: cluster %u\n", __func__, cluster);
505 	if (cluster >= SUNXI_NR_CLUSTERS)
506 		return -EINVAL;
507 
508 	/* assert cluster resets or system will hang */
509 	pr_debug("%s: assert cluster reset\n", __func__);
510 	reg = readl(cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
511 	reg &= ~CPUCFG_CX_RST_CTRL_DBG_SOC_RST;
512 	reg &= ~CPUCFG_CX_RST_CTRL_H_RST;
513 	reg &= ~CPUCFG_CX_RST_CTRL_L2_RST;
514 	writel(reg, cpucfg_base + CPUCFG_CX_RST_CTRL(cluster));
515 
516 	/* gate cluster power */
517 	pr_debug("%s: gate cluster power\n", __func__);
518 	reg = readl(prcm_base + PRCM_PWROFF_GATING_REG(cluster));
519 	reg |= PRCM_PWROFF_GATING_REG_CLUSTER;
520 	writel(reg, prcm_base + PRCM_PWROFF_GATING_REG(cluster));
521 	udelay(20);
522 
523 	return 0;
524 }
525 
526 static int sunxi_mc_smp_cpu_kill(unsigned int l_cpu)
527 {
528 	unsigned int mpidr, cpu, cluster;
529 	unsigned int tries, count;
530 	int ret = 0;
531 	u32 reg;
532 
533 	mpidr = cpu_logical_map(l_cpu);
534 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
535 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
536 
537 	/* This should never happen */
538 	if (WARN_ON(cluster >= SUNXI_NR_CLUSTERS ||
539 		    cpu >= SUNXI_CPUS_PER_CLUSTER))
540 		return 0;
541 
542 	/* wait for CPU core to die and enter WFI */
543 	count = TIMEOUT_USEC / POLL_USEC;
544 	spin_lock_irq(&boot_lock);
545 	for (tries = 0; tries < count; tries++) {
546 		spin_unlock_irq(&boot_lock);
547 		usleep_range(POLL_USEC / 2, POLL_USEC);
548 		spin_lock_irq(&boot_lock);
549 
550 		/*
551 		 * If the user turns off a bunch of cores at the same
552 		 * time, the kernel might call cpu_kill before some of
553 		 * them are ready. This is because boot_lock serializes
554 		 * both cpu_die and cpu_kill callbacks. Either one could
555 		 * run first. We should wait for cpu_die to complete.
556 		 */
557 		if (sunxi_mc_smp_cpu_table[cluster][cpu])
558 			continue;
559 
560 		reg = readl(cpucfg_base + CPUCFG_CX_STATUS(cluster));
561 		if (reg & CPUCFG_CX_STATUS_STANDBYWFI(cpu))
562 			break;
563 	}
564 
565 	if (tries >= count) {
566 		ret = ETIMEDOUT;
567 		goto out;
568 	}
569 
570 	/* power down CPU core */
571 	sunxi_cpu_powerdown(cpu, cluster);
572 
573 	if (!sunxi_mc_smp_cluster_is_down(cluster))
574 		goto out;
575 
576 	/* wait for cluster L2 WFI */
577 	ret = readl_poll_timeout(cpucfg_base + CPUCFG_CX_STATUS(cluster), reg,
578 				 reg & CPUCFG_CX_STATUS_STANDBYWFIL2,
579 				 POLL_USEC, TIMEOUT_USEC);
580 	if (ret) {
581 		/*
582 		 * Ignore timeout on the cluster. Leaving the cluster on
583 		 * will not affect system execution, just use a bit more
584 		 * power. But returning an error here will only confuse
585 		 * the user as the CPU has already been shutdown.
586 		 */
587 		ret = 0;
588 		goto out;
589 	}
590 
591 	/* Power down cluster */
592 	sunxi_cluster_powerdown(cluster);
593 
594 out:
595 	spin_unlock_irq(&boot_lock);
596 	pr_debug("%s: cluster %u cpu %u powerdown: %d\n",
597 		 __func__, cluster, cpu, ret);
598 	return !ret;
599 }
600 
601 static bool sunxi_mc_smp_cpu_can_disable(unsigned int __unused)
602 {
603 	return true;
604 }
605 #endif
606 
607 static const struct smp_operations sunxi_mc_smp_smp_ops __initconst = {
608 	.smp_secondary_init	= sunxi_mc_smp_secondary_init,
609 	.smp_boot_secondary	= sunxi_mc_smp_boot_secondary,
610 #ifdef CONFIG_HOTPLUG_CPU
611 	.cpu_die		= sunxi_mc_smp_cpu_die,
612 	.cpu_kill		= sunxi_mc_smp_cpu_kill,
613 	.cpu_can_disable	= sunxi_mc_smp_cpu_can_disable,
614 #endif
615 };
616 
617 static bool __init sunxi_mc_smp_cpu_table_init(void)
618 {
619 	unsigned int mpidr, cpu, cluster;
620 
621 	mpidr = read_cpuid_mpidr();
622 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
623 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
624 
625 	if (cluster >= SUNXI_NR_CLUSTERS || cpu >= SUNXI_CPUS_PER_CLUSTER) {
626 		pr_err("%s: boot CPU is out of bounds!\n", __func__);
627 		return false;
628 	}
629 	sunxi_mc_smp_cpu_table[cluster][cpu] = 1;
630 	return true;
631 }
632 
633 /*
634  * Adapted from arch/arm/common/mc_smp_entry.c
635  *
636  * We need the trampoline code to enable CCI-400 on the first cluster
637  */
638 typedef typeof(cpu_reset) phys_reset_t;
639 
640 static void __init __naked sunxi_mc_smp_resume(void)
641 {
642 	asm volatile(
643 		"bl	sunxi_mc_smp_cluster_cache_enable\n"
644 		"b	cpu_resume"
645 		/* Let compiler know about sunxi_mc_smp_cluster_cache_enable */
646 		:: "i" (sunxi_mc_smp_cluster_cache_enable)
647 	);
648 }
649 
650 static int __init nocache_trampoline(unsigned long __unused)
651 {
652 	phys_reset_t phys_reset;
653 
654 	setup_mm_for_reboot();
655 	sunxi_cluster_cache_disable_without_axi();
656 
657 	phys_reset = (phys_reset_t)(unsigned long)__pa_symbol(cpu_reset);
658 	phys_reset(__pa_symbol(sunxi_mc_smp_resume), false);
659 	BUG();
660 }
661 
662 static int __init sunxi_mc_smp_loopback(void)
663 {
664 	int ret;
665 
666 	/*
667 	 * We're going to soft-restart the current CPU through the
668 	 * low-level MCPM code by leveraging the suspend/resume
669 	 * infrastructure. Let's play it safe by using cpu_pm_enter()
670 	 * in case the CPU init code path resets the VFP or similar.
671 	 */
672 	sunxi_mc_smp_first_comer = true;
673 	local_irq_disable();
674 	local_fiq_disable();
675 	ret = cpu_pm_enter();
676 	if (!ret) {
677 		ret = cpu_suspend(0, nocache_trampoline);
678 		cpu_pm_exit();
679 	}
680 	local_fiq_enable();
681 	local_irq_enable();
682 	sunxi_mc_smp_first_comer = false;
683 
684 	return ret;
685 }
686 
687 /*
688  * This holds any device nodes that we requested resources for,
689  * so that we may easily release resources in the error path.
690  */
691 struct sunxi_mc_smp_nodes {
692 	struct device_node *prcm_node;
693 	struct device_node *cpucfg_node;
694 	struct device_node *sram_node;
695 };
696 
697 /* This structure holds SoC-specific bits tied to an enable-method string. */
698 struct sunxi_mc_smp_data {
699 	const char *enable_method;
700 	int (*get_smp_nodes)(struct sunxi_mc_smp_nodes *nodes);
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 	memset(nodes, 0, sizeof(*nodes));
709 }
710 
711 static int __init sun9i_a80_get_smp_nodes(struct sunxi_mc_smp_nodes *nodes)
712 {
713 	nodes->prcm_node = of_find_compatible_node(NULL, NULL,
714 						   "allwinner,sun9i-a80-prcm");
715 	if (!nodes->prcm_node) {
716 		pr_err("%s: PRCM not available\n", __func__);
717 		return -ENODEV;
718 	}
719 
720 	nodes->cpucfg_node = of_find_compatible_node(NULL, NULL,
721 						     "allwinner,sun9i-a80-cpucfg");
722 	if (!nodes->cpucfg_node) {
723 		pr_err("%s: CPUCFG not available\n", __func__);
724 		return -ENODEV;
725 	}
726 
727 	nodes->sram_node = of_find_compatible_node(NULL, NULL,
728 						   "allwinner,sun9i-a80-smp-sram");
729 	if (!nodes->sram_node) {
730 		pr_err("%s: Secure SRAM not available\n", __func__);
731 		return -ENODEV;
732 	}
733 
734 	return 0;
735 }
736 
737 static const struct sunxi_mc_smp_data sunxi_mc_smp_data[] __initconst = {
738 	{
739 		.enable_method	= "allwinner,sun9i-a80-smp",
740 		.get_smp_nodes	= sun9i_a80_get_smp_nodes,
741 	},
742 };
743 
744 static int __init sunxi_mc_smp_init(void)
745 {
746 	struct sunxi_mc_smp_nodes nodes = { 0 };
747 	struct device_node *node;
748 	struct resource res;
749 	int i, ret;
750 
751 	/*
752 	 * Don't bother checking the "cpus" node, as an enable-method
753 	 * property in that node is undocumented.
754 	 */
755 	node = of_cpu_device_node_get(0);
756 	if (!node)
757 		return -ENODEV;
758 
759 	/*
760 	 * We can't actually use the enable-method magic in the kernel.
761 	 * Our loopback / trampoline code uses the CPU suspend framework,
762 	 * which requires the identity mapping be available. It would not
763 	 * yet be available if we used the .init_cpus or .prepare_cpus
764 	 * callbacks in smp_operations, which we would use if we were to
765 	 * use CPU_METHOD_OF_DECLARE
766 	 */
767 	for (i = 0; i < ARRAY_SIZE(sunxi_mc_smp_data); i++) {
768 		ret = of_property_match_string(node, "enable-method",
769 					       sunxi_mc_smp_data[i].enable_method);
770 		if (!ret)
771 			break;
772 	}
773 
774 	of_node_put(node);
775 	if (ret)
776 		return -ENODEV;
777 
778 	if (!sunxi_mc_smp_cpu_table_init())
779 		return -EINVAL;
780 
781 	if (!cci_probed()) {
782 		pr_err("%s: CCI-400 not available\n", __func__);
783 		return -ENODEV;
784 	}
785 
786 	/* Get needed device tree nodes */
787 	ret = sunxi_mc_smp_data[i].get_smp_nodes(&nodes);
788 	if (ret)
789 		goto err_put_nodes;
790 
791 	/*
792 	 * Unfortunately we can not request the I/O region for the PRCM.
793 	 * It is shared with the PRCM clock.
794 	 */
795 	prcm_base = of_iomap(nodes.prcm_node, 0);
796 	if (!prcm_base) {
797 		pr_err("%s: failed to map PRCM registers\n", __func__);
798 		ret = -ENOMEM;
799 		goto err_put_nodes;
800 	}
801 
802 	cpucfg_base = of_io_request_and_map(nodes.cpucfg_node, 0,
803 					    "sunxi-mc-smp");
804 	if (IS_ERR(cpucfg_base)) {
805 		ret = PTR_ERR(cpucfg_base);
806 		pr_err("%s: failed to map CPUCFG registers: %d\n",
807 		       __func__, ret);
808 		goto err_unmap_prcm;
809 	}
810 
811 	sram_b_smp_base = of_io_request_and_map(nodes.sram_node, 0,
812 						"sunxi-mc-smp");
813 	if (IS_ERR(sram_b_smp_base)) {
814 		ret = PTR_ERR(sram_b_smp_base);
815 		pr_err("%s: failed to map secure SRAM\n", __func__);
816 		goto err_unmap_release_cpucfg;
817 	}
818 
819 	/* Configure CCI-400 for boot cluster */
820 	ret = sunxi_mc_smp_loopback();
821 	if (ret) {
822 		pr_err("%s: failed to configure boot cluster: %d\n",
823 		       __func__, ret);
824 		goto err_unmap_release_secure_sram;
825 	}
826 
827 	/* We don't need the device nodes anymore */
828 	sunxi_mc_smp_put_nodes(&nodes);
829 
830 	/* Set the hardware entry point address */
831 	writel(__pa_symbol(sunxi_mc_smp_secondary_startup),
832 	       prcm_base + PRCM_CPU_SOFT_ENTRY_REG);
833 
834 	/* Actually enable multi cluster SMP */
835 	smp_set_ops(&sunxi_mc_smp_smp_ops);
836 
837 	pr_info("sunxi multi cluster SMP support installed\n");
838 
839 	return 0;
840 
841 err_unmap_release_secure_sram:
842 	iounmap(sram_b_smp_base);
843 	of_address_to_resource(nodes.sram_node, 0, &res);
844 	release_mem_region(res.start, resource_size(&res));
845 err_unmap_release_cpucfg:
846 	iounmap(cpucfg_base);
847 	of_address_to_resource(nodes.cpucfg_node, 0, &res);
848 	release_mem_region(res.start, resource_size(&res));
849 err_unmap_prcm:
850 	iounmap(prcm_base);
851 err_put_nodes:
852 	sunxi_mc_smp_put_nodes(&nodes);
853 	return ret;
854 }
855 
856 early_initcall(sunxi_mc_smp_init);
857