xref: /openbmc/linux/arch/arm/mach-mvebu/pmsu.c (revision b96fc2f3)
1 /*
2  * Power Management Service Unit(PMSU) support for Armada 370/XP platforms.
3  *
4  * Copyright (C) 2012 Marvell
5  *
6  * Yehuda Yitschak <yehuday@marvell.com>
7  * Gregory Clement <gregory.clement@free-electrons.com>
8  * Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
9  *
10  * This file is licensed under the terms of the GNU General Public
11  * License version 2.  This program is licensed "as is" without any
12  * warranty of any kind, whether express or implied.
13  *
14  * The Armada 370 and Armada XP SOCs have a power management service
15  * unit which is responsible for powering down and waking up CPUs and
16  * other SOC units
17  */
18 
19 #define pr_fmt(fmt) "mvebu-pmsu: " fmt
20 
21 #include <linux/clk.h>
22 #include <linux/cpu_pm.h>
23 #include <linux/cpufreq-dt.h>
24 #include <linux/delay.h>
25 #include <linux/init.h>
26 #include <linux/io.h>
27 #include <linux/kernel.h>
28 #include <linux/mbus.h>
29 #include <linux/of_address.h>
30 #include <linux/of_device.h>
31 #include <linux/platform_device.h>
32 #include <linux/pm_opp.h>
33 #include <linux/resource.h>
34 #include <linux/slab.h>
35 #include <linux/smp.h>
36 #include <asm/cacheflush.h>
37 #include <asm/cp15.h>
38 #include <asm/smp_scu.h>
39 #include <asm/smp_plat.h>
40 #include <asm/suspend.h>
41 #include <asm/tlbflush.h>
42 #include "common.h"
43 
44 
45 #define PMSU_BASE_OFFSET    0x100
46 #define PMSU_REG_SIZE	    0x1000
47 
48 /* PMSU MP registers */
49 #define PMSU_CONTROL_AND_CONFIG(cpu)	    ((cpu * 0x100) + 0x104)
50 #define PMSU_CONTROL_AND_CONFIG_DFS_REQ		BIT(18)
51 #define PMSU_CONTROL_AND_CONFIG_PWDDN_REQ	BIT(16)
52 #define PMSU_CONTROL_AND_CONFIG_L2_PWDDN	BIT(20)
53 
54 #define PMSU_CPU_POWER_DOWN_CONTROL(cpu)    ((cpu * 0x100) + 0x108)
55 
56 #define PMSU_CPU_POWER_DOWN_DIS_SNP_Q_SKIP	BIT(0)
57 
58 #define PMSU_STATUS_AND_MASK(cpu)	    ((cpu * 0x100) + 0x10c)
59 #define PMSU_STATUS_AND_MASK_CPU_IDLE_WAIT	BIT(16)
60 #define PMSU_STATUS_AND_MASK_SNP_Q_EMPTY_WAIT	BIT(17)
61 #define PMSU_STATUS_AND_MASK_IRQ_WAKEUP		BIT(20)
62 #define PMSU_STATUS_AND_MASK_FIQ_WAKEUP		BIT(21)
63 #define PMSU_STATUS_AND_MASK_DBG_WAKEUP		BIT(22)
64 #define PMSU_STATUS_AND_MASK_IRQ_MASK		BIT(24)
65 #define PMSU_STATUS_AND_MASK_FIQ_MASK		BIT(25)
66 
67 #define PMSU_EVENT_STATUS_AND_MASK(cpu)     ((cpu * 0x100) + 0x120)
68 #define PMSU_EVENT_STATUS_AND_MASK_DFS_DONE        BIT(1)
69 #define PMSU_EVENT_STATUS_AND_MASK_DFS_DONE_MASK   BIT(17)
70 
71 #define PMSU_BOOT_ADDR_REDIRECT_OFFSET(cpu) ((cpu * 0x100) + 0x124)
72 
73 /* PMSU fabric registers */
74 #define L2C_NFABRIC_PM_CTL		    0x4
75 #define L2C_NFABRIC_PM_CTL_PWR_DOWN		BIT(20)
76 
77 /* PMSU delay registers */
78 #define PMSU_POWERDOWN_DELAY		    0xF04
79 #define PMSU_POWERDOWN_DELAY_PMU		BIT(1)
80 #define PMSU_POWERDOWN_DELAY_MASK		0xFFFE
81 #define PMSU_DFLT_ARMADA38X_DELAY	        0x64
82 
83 /* CA9 MPcore SoC Control registers */
84 
85 #define MPCORE_RESET_CTL		    0x64
86 #define MPCORE_RESET_CTL_L2			BIT(0)
87 #define MPCORE_RESET_CTL_DEBUG			BIT(16)
88 
89 #define SRAM_PHYS_BASE  0xFFFF0000
90 #define BOOTROM_BASE    0xFFF00000
91 #define BOOTROM_SIZE    0x100000
92 
93 #define ARMADA_370_CRYPT0_ENG_TARGET   0x9
94 #define ARMADA_370_CRYPT0_ENG_ATTR     0x1
95 
96 extern void ll_disable_coherency(void);
97 extern void ll_enable_coherency(void);
98 
99 extern void armada_370_xp_cpu_resume(void);
100 extern void armada_38x_cpu_resume(void);
101 
102 static phys_addr_t pmsu_mp_phys_base;
103 static void __iomem *pmsu_mp_base;
104 
105 static void *mvebu_cpu_resume;
106 
107 static const struct of_device_id of_pmsu_table[] = {
108 	{ .compatible = "marvell,armada-370-pmsu", },
109 	{ .compatible = "marvell,armada-370-xp-pmsu", },
110 	{ .compatible = "marvell,armada-380-pmsu", },
111 	{ /* end of list */ },
112 };
113 
114 void mvebu_pmsu_set_cpu_boot_addr(int hw_cpu, void *boot_addr)
115 {
116 	writel(virt_to_phys(boot_addr), pmsu_mp_base +
117 		PMSU_BOOT_ADDR_REDIRECT_OFFSET(hw_cpu));
118 }
119 
120 extern unsigned char mvebu_boot_wa_start;
121 extern unsigned char mvebu_boot_wa_end;
122 
123 /*
124  * This function sets up the boot address workaround needed for SMP
125  * boot on Armada 375 Z1 and cpuidle on Armada 370. It unmaps the
126  * BootROM Mbus window, and instead remaps a crypto SRAM into which a
127  * custom piece of code is copied to replace the problematic BootROM.
128  */
129 int mvebu_setup_boot_addr_wa(unsigned int crypto_eng_target,
130 			     unsigned int crypto_eng_attribute,
131 			     phys_addr_t resume_addr_reg)
132 {
133 	void __iomem *sram_virt_base;
134 	u32 code_len = &mvebu_boot_wa_end - &mvebu_boot_wa_start;
135 
136 	mvebu_mbus_del_window(BOOTROM_BASE, BOOTROM_SIZE);
137 	mvebu_mbus_add_window_by_id(crypto_eng_target, crypto_eng_attribute,
138 				    SRAM_PHYS_BASE, SZ_64K);
139 
140 	sram_virt_base = ioremap(SRAM_PHYS_BASE, SZ_64K);
141 	if (!sram_virt_base) {
142 		pr_err("Unable to map SRAM to setup the boot address WA\n");
143 		return -ENOMEM;
144 	}
145 
146 	memcpy(sram_virt_base, &mvebu_boot_wa_start, code_len);
147 
148 	/*
149 	 * The last word of the code copied in SRAM must contain the
150 	 * physical base address of the PMSU register. We
151 	 * intentionally store this address in the native endianness
152 	 * of the system.
153 	 */
154 	__raw_writel((unsigned long)resume_addr_reg,
155 		     sram_virt_base + code_len - 4);
156 
157 	iounmap(sram_virt_base);
158 
159 	return 0;
160 }
161 
162 static int __init mvebu_v7_pmsu_init(void)
163 {
164 	struct device_node *np;
165 	struct resource res;
166 	int ret = 0;
167 
168 	np = of_find_matching_node(NULL, of_pmsu_table);
169 	if (!np)
170 		return 0;
171 
172 	pr_info("Initializing Power Management Service Unit\n");
173 
174 	if (of_address_to_resource(np, 0, &res)) {
175 		pr_err("unable to get resource\n");
176 		ret = -ENOENT;
177 		goto out;
178 	}
179 
180 	if (of_device_is_compatible(np, "marvell,armada-370-xp-pmsu")) {
181 		pr_warn(FW_WARN "deprecated pmsu binding\n");
182 		res.start = res.start - PMSU_BASE_OFFSET;
183 		res.end = res.start + PMSU_REG_SIZE - 1;
184 	}
185 
186 	if (!request_mem_region(res.start, resource_size(&res),
187 				np->full_name)) {
188 		pr_err("unable to request region\n");
189 		ret = -EBUSY;
190 		goto out;
191 	}
192 
193 	pmsu_mp_phys_base = res.start;
194 
195 	pmsu_mp_base = ioremap(res.start, resource_size(&res));
196 	if (!pmsu_mp_base) {
197 		pr_err("unable to map registers\n");
198 		release_mem_region(res.start, resource_size(&res));
199 		ret = -ENOMEM;
200 		goto out;
201 	}
202 
203  out:
204 	of_node_put(np);
205 	return ret;
206 }
207 
208 static void mvebu_v7_pmsu_enable_l2_powerdown_onidle(void)
209 {
210 	u32 reg;
211 
212 	if (pmsu_mp_base == NULL)
213 		return;
214 
215 	/* Enable L2 & Fabric powerdown in Deep-Idle mode - Fabric */
216 	reg = readl(pmsu_mp_base + L2C_NFABRIC_PM_CTL);
217 	reg |= L2C_NFABRIC_PM_CTL_PWR_DOWN;
218 	writel(reg, pmsu_mp_base + L2C_NFABRIC_PM_CTL);
219 }
220 
221 enum pmsu_idle_prepare_flags {
222 	PMSU_PREPARE_NORMAL = 0,
223 	PMSU_PREPARE_DEEP_IDLE = BIT(0),
224 	PMSU_PREPARE_SNOOP_DISABLE = BIT(1),
225 };
226 
227 /* No locking is needed because we only access per-CPU registers */
228 static int mvebu_v7_pmsu_idle_prepare(unsigned long flags)
229 {
230 	unsigned int hw_cpu = cpu_logical_map(smp_processor_id());
231 	u32 reg;
232 
233 	if (pmsu_mp_base == NULL)
234 		return -EINVAL;
235 
236 	/*
237 	 * Adjust the PMSU configuration to wait for WFI signal, enable
238 	 * IRQ and FIQ as wakeup events, set wait for snoop queue empty
239 	 * indication and mask IRQ and FIQ from CPU
240 	 */
241 	reg = readl(pmsu_mp_base + PMSU_STATUS_AND_MASK(hw_cpu));
242 	reg |= PMSU_STATUS_AND_MASK_CPU_IDLE_WAIT    |
243 	       PMSU_STATUS_AND_MASK_IRQ_WAKEUP       |
244 	       PMSU_STATUS_AND_MASK_FIQ_WAKEUP       |
245 	       PMSU_STATUS_AND_MASK_SNP_Q_EMPTY_WAIT |
246 	       PMSU_STATUS_AND_MASK_IRQ_MASK         |
247 	       PMSU_STATUS_AND_MASK_FIQ_MASK;
248 	writel(reg, pmsu_mp_base + PMSU_STATUS_AND_MASK(hw_cpu));
249 
250 	reg = readl(pmsu_mp_base + PMSU_CONTROL_AND_CONFIG(hw_cpu));
251 	/* ask HW to power down the L2 Cache if needed */
252 	if (flags & PMSU_PREPARE_DEEP_IDLE)
253 		reg |= PMSU_CONTROL_AND_CONFIG_L2_PWDDN;
254 
255 	/* request power down */
256 	reg |= PMSU_CONTROL_AND_CONFIG_PWDDN_REQ;
257 	writel(reg, pmsu_mp_base + PMSU_CONTROL_AND_CONFIG(hw_cpu));
258 
259 	if (flags & PMSU_PREPARE_SNOOP_DISABLE) {
260 		/* Disable snoop disable by HW - SW is taking care of it */
261 		reg = readl(pmsu_mp_base + PMSU_CPU_POWER_DOWN_CONTROL(hw_cpu));
262 		reg |= PMSU_CPU_POWER_DOWN_DIS_SNP_Q_SKIP;
263 		writel(reg, pmsu_mp_base + PMSU_CPU_POWER_DOWN_CONTROL(hw_cpu));
264 	}
265 
266 	return 0;
267 }
268 
269 int armada_370_xp_pmsu_idle_enter(unsigned long deepidle)
270 {
271 	unsigned long flags = PMSU_PREPARE_SNOOP_DISABLE;
272 	int ret;
273 
274 	if (deepidle)
275 		flags |= PMSU_PREPARE_DEEP_IDLE;
276 
277 	ret = mvebu_v7_pmsu_idle_prepare(flags);
278 	if (ret)
279 		return ret;
280 
281 	v7_exit_coherency_flush(all);
282 
283 	ll_disable_coherency();
284 
285 	dsb();
286 
287 	wfi();
288 
289 	/* If we are here, wfi failed. As processors run out of
290 	 * coherency for some time, tlbs might be stale, so flush them
291 	 */
292 	local_flush_tlb_all();
293 
294 	ll_enable_coherency();
295 
296 	/* Test the CR_C bit and set it if it was cleared */
297 	asm volatile(
298 	"mrc	p15, 0, r0, c1, c0, 0 \n\t"
299 	"tst	r0, %0 \n\t"
300 	"orreq	r0, r0, #(1 << 2) \n\t"
301 	"mcreq	p15, 0, r0, c1, c0, 0 \n\t"
302 	"isb	"
303 	: : "Ir" (CR_C) : "r0");
304 
305 	pr_debug("Failed to suspend the system\n");
306 
307 	return 0;
308 }
309 
310 static int armada_370_xp_cpu_suspend(unsigned long deepidle)
311 {
312 	return cpu_suspend(deepidle, armada_370_xp_pmsu_idle_enter);
313 }
314 
315 int armada_38x_do_cpu_suspend(unsigned long deepidle)
316 {
317 	unsigned long flags = 0;
318 
319 	if (deepidle)
320 		flags |= PMSU_PREPARE_DEEP_IDLE;
321 
322 	mvebu_v7_pmsu_idle_prepare(flags);
323 	/*
324 	 * Already flushed cache, but do it again as the outer cache
325 	 * functions dirty the cache with spinlocks
326 	 */
327 	v7_exit_coherency_flush(louis);
328 
329 	scu_power_mode(mvebu_get_scu_base(), SCU_PM_POWEROFF);
330 
331 	cpu_do_idle();
332 
333 	return 1;
334 }
335 
336 static int armada_38x_cpu_suspend(unsigned long deepidle)
337 {
338 	return cpu_suspend(false, armada_38x_do_cpu_suspend);
339 }
340 
341 /* No locking is needed because we only access per-CPU registers */
342 void mvebu_v7_pmsu_idle_exit(void)
343 {
344 	unsigned int hw_cpu = cpu_logical_map(smp_processor_id());
345 	u32 reg;
346 
347 	if (pmsu_mp_base == NULL)
348 		return;
349 	/* cancel ask HW to power down the L2 Cache if possible */
350 	reg = readl(pmsu_mp_base + PMSU_CONTROL_AND_CONFIG(hw_cpu));
351 	reg &= ~PMSU_CONTROL_AND_CONFIG_L2_PWDDN;
352 	writel(reg, pmsu_mp_base + PMSU_CONTROL_AND_CONFIG(hw_cpu));
353 
354 	/* cancel Enable wakeup events and mask interrupts */
355 	reg = readl(pmsu_mp_base + PMSU_STATUS_AND_MASK(hw_cpu));
356 	reg &= ~(PMSU_STATUS_AND_MASK_IRQ_WAKEUP | PMSU_STATUS_AND_MASK_FIQ_WAKEUP);
357 	reg &= ~PMSU_STATUS_AND_MASK_CPU_IDLE_WAIT;
358 	reg &= ~PMSU_STATUS_AND_MASK_SNP_Q_EMPTY_WAIT;
359 	reg &= ~(PMSU_STATUS_AND_MASK_IRQ_MASK | PMSU_STATUS_AND_MASK_FIQ_MASK);
360 	writel(reg, pmsu_mp_base + PMSU_STATUS_AND_MASK(hw_cpu));
361 }
362 
363 static int mvebu_v7_cpu_pm_notify(struct notifier_block *self,
364 				    unsigned long action, void *hcpu)
365 {
366 	if (action == CPU_PM_ENTER) {
367 		unsigned int hw_cpu = cpu_logical_map(smp_processor_id());
368 		mvebu_pmsu_set_cpu_boot_addr(hw_cpu, mvebu_cpu_resume);
369 	} else if (action == CPU_PM_EXIT) {
370 		mvebu_v7_pmsu_idle_exit();
371 	}
372 
373 	return NOTIFY_OK;
374 }
375 
376 static struct notifier_block mvebu_v7_cpu_pm_notifier = {
377 	.notifier_call = mvebu_v7_cpu_pm_notify,
378 };
379 
380 static struct platform_device mvebu_v7_cpuidle_device;
381 
382 static __init int armada_370_cpuidle_init(void)
383 {
384 	struct device_node *np;
385 	phys_addr_t redirect_reg;
386 
387 	np = of_find_compatible_node(NULL, NULL, "marvell,coherency-fabric");
388 	if (!np)
389 		return -ENODEV;
390 	of_node_put(np);
391 
392 	/*
393 	 * On Armada 370, there is "a slow exit process from the deep
394 	 * idle state due to heavy L1/L2 cache cleanup operations
395 	 * performed by the BootROM software". To avoid this, we
396 	 * replace the restart code of the bootrom by a a simple jump
397 	 * to the boot address. Then the code located at this boot
398 	 * address will take care of the initialization.
399 	 */
400 	redirect_reg = pmsu_mp_phys_base + PMSU_BOOT_ADDR_REDIRECT_OFFSET(0);
401 	mvebu_setup_boot_addr_wa(ARMADA_370_CRYPT0_ENG_TARGET,
402 				 ARMADA_370_CRYPT0_ENG_ATTR,
403 				 redirect_reg);
404 
405 	mvebu_cpu_resume = armada_370_xp_cpu_resume;
406 	mvebu_v7_cpuidle_device.dev.platform_data = armada_370_xp_cpu_suspend;
407 	mvebu_v7_cpuidle_device.name = "cpuidle-armada-370";
408 
409 	return 0;
410 }
411 
412 static __init int armada_38x_cpuidle_init(void)
413 {
414 	struct device_node *np;
415 	void __iomem *mpsoc_base;
416 	u32 reg;
417 
418 	pr_warn("CPU idle is currently broken on Armada 38x: disabling\n");
419 	return 0;
420 
421 	np = of_find_compatible_node(NULL, NULL,
422 				     "marvell,armada-380-coherency-fabric");
423 	if (!np)
424 		return -ENODEV;
425 	of_node_put(np);
426 
427 	np = of_find_compatible_node(NULL, NULL,
428 				     "marvell,armada-380-mpcore-soc-ctrl");
429 	if (!np)
430 		return -ENODEV;
431 	mpsoc_base = of_iomap(np, 0);
432 	BUG_ON(!mpsoc_base);
433 	of_node_put(np);
434 
435 	/* Set up reset mask when powering down the cpus */
436 	reg = readl(mpsoc_base + MPCORE_RESET_CTL);
437 	reg |= MPCORE_RESET_CTL_L2;
438 	reg |= MPCORE_RESET_CTL_DEBUG;
439 	writel(reg, mpsoc_base + MPCORE_RESET_CTL);
440 	iounmap(mpsoc_base);
441 
442 	/* Set up delay */
443 	reg = readl(pmsu_mp_base + PMSU_POWERDOWN_DELAY);
444 	reg &= ~PMSU_POWERDOWN_DELAY_MASK;
445 	reg |= PMSU_DFLT_ARMADA38X_DELAY;
446 	reg |= PMSU_POWERDOWN_DELAY_PMU;
447 	writel(reg, pmsu_mp_base + PMSU_POWERDOWN_DELAY);
448 
449 	mvebu_cpu_resume = armada_38x_cpu_resume;
450 	mvebu_v7_cpuidle_device.dev.platform_data = armada_38x_cpu_suspend;
451 	mvebu_v7_cpuidle_device.name = "cpuidle-armada-38x";
452 
453 	return 0;
454 }
455 
456 static __init int armada_xp_cpuidle_init(void)
457 {
458 	struct device_node *np;
459 
460 	np = of_find_compatible_node(NULL, NULL, "marvell,coherency-fabric");
461 	if (!np)
462 		return -ENODEV;
463 	of_node_put(np);
464 
465 	mvebu_cpu_resume = armada_370_xp_cpu_resume;
466 	mvebu_v7_cpuidle_device.dev.platform_data = armada_370_xp_cpu_suspend;
467 	mvebu_v7_cpuidle_device.name = "cpuidle-armada-xp";
468 
469 	return 0;
470 }
471 
472 static int __init mvebu_v7_cpu_pm_init(void)
473 {
474 	struct device_node *np;
475 	int ret;
476 
477 	np = of_find_matching_node(NULL, of_pmsu_table);
478 	if (!np)
479 		return 0;
480 	of_node_put(np);
481 
482 	/*
483 	 * Currently the CPU idle support for Armada 38x is broken, as
484 	 * the CPU hotplug uses some of the CPU idle functions it is
485 	 * broken too, so let's disable it
486 	 */
487 	if (of_machine_is_compatible("marvell,armada380")) {
488 		cpu_hotplug_disable();
489 		pr_warn("CPU hotplug support is currently broken on Armada 38x: disabling\n");
490 	}
491 
492 	if (of_machine_is_compatible("marvell,armadaxp"))
493 		ret = armada_xp_cpuidle_init();
494 	else if (of_machine_is_compatible("marvell,armada370"))
495 		ret = armada_370_cpuidle_init();
496 	else if (of_machine_is_compatible("marvell,armada380"))
497 		ret = armada_38x_cpuidle_init();
498 	else
499 		return 0;
500 
501 	if (ret)
502 		return ret;
503 
504 	mvebu_v7_pmsu_enable_l2_powerdown_onidle();
505 	if (mvebu_v7_cpuidle_device.name)
506 		platform_device_register(&mvebu_v7_cpuidle_device);
507 	cpu_pm_register_notifier(&mvebu_v7_cpu_pm_notifier);
508 
509 	return 0;
510 }
511 
512 arch_initcall(mvebu_v7_cpu_pm_init);
513 early_initcall(mvebu_v7_pmsu_init);
514 
515 static void mvebu_pmsu_dfs_request_local(void *data)
516 {
517 	u32 reg;
518 	u32 cpu = smp_processor_id();
519 	unsigned long flags;
520 
521 	local_irq_save(flags);
522 
523 	/* Prepare to enter idle */
524 	reg = readl(pmsu_mp_base + PMSU_STATUS_AND_MASK(cpu));
525 	reg |= PMSU_STATUS_AND_MASK_CPU_IDLE_WAIT |
526 	       PMSU_STATUS_AND_MASK_IRQ_MASK     |
527 	       PMSU_STATUS_AND_MASK_FIQ_MASK;
528 	writel(reg, pmsu_mp_base + PMSU_STATUS_AND_MASK(cpu));
529 
530 	/* Request the DFS transition */
531 	reg = readl(pmsu_mp_base + PMSU_CONTROL_AND_CONFIG(cpu));
532 	reg |= PMSU_CONTROL_AND_CONFIG_DFS_REQ;
533 	writel(reg, pmsu_mp_base + PMSU_CONTROL_AND_CONFIG(cpu));
534 
535 	/* The fact of entering idle will trigger the DFS transition */
536 	wfi();
537 
538 	/*
539 	 * We're back from idle, the DFS transition has completed,
540 	 * clear the idle wait indication.
541 	 */
542 	reg = readl(pmsu_mp_base + PMSU_STATUS_AND_MASK(cpu));
543 	reg &= ~PMSU_STATUS_AND_MASK_CPU_IDLE_WAIT;
544 	writel(reg, pmsu_mp_base + PMSU_STATUS_AND_MASK(cpu));
545 
546 	local_irq_restore(flags);
547 }
548 
549 int mvebu_pmsu_dfs_request(int cpu)
550 {
551 	unsigned long timeout;
552 	int hwcpu = cpu_logical_map(cpu);
553 	u32 reg;
554 
555 	/* Clear any previous DFS DONE event */
556 	reg = readl(pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
557 	reg &= ~PMSU_EVENT_STATUS_AND_MASK_DFS_DONE;
558 	writel(reg, pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
559 
560 	/* Mask the DFS done interrupt, since we are going to poll */
561 	reg = readl(pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
562 	reg |= PMSU_EVENT_STATUS_AND_MASK_DFS_DONE_MASK;
563 	writel(reg, pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
564 
565 	/* Trigger the DFS on the appropriate CPU */
566 	smp_call_function_single(cpu, mvebu_pmsu_dfs_request_local,
567 				 NULL, false);
568 
569 	/* Poll until the DFS done event is generated */
570 	timeout = jiffies + HZ;
571 	while (time_before(jiffies, timeout)) {
572 		reg = readl(pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
573 		if (reg & PMSU_EVENT_STATUS_AND_MASK_DFS_DONE)
574 			break;
575 		udelay(10);
576 	}
577 
578 	if (time_after(jiffies, timeout))
579 		return -ETIME;
580 
581 	/* Restore the DFS mask to its original state */
582 	reg = readl(pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
583 	reg &= ~PMSU_EVENT_STATUS_AND_MASK_DFS_DONE_MASK;
584 	writel(reg, pmsu_mp_base + PMSU_EVENT_STATUS_AND_MASK(hwcpu));
585 
586 	return 0;
587 }
588 
589 struct cpufreq_dt_platform_data cpufreq_dt_pd = {
590 	.independent_clocks = true,
591 };
592 
593 static int __init armada_xp_pmsu_cpufreq_init(void)
594 {
595 	struct device_node *np;
596 	struct resource res;
597 	int ret, cpu;
598 
599 	if (!of_machine_is_compatible("marvell,armadaxp"))
600 		return 0;
601 
602 	/*
603 	 * In order to have proper cpufreq handling, we need to ensure
604 	 * that the Device Tree description of the CPU clock includes
605 	 * the definition of the PMU DFS registers. If not, we do not
606 	 * register the clock notifier and the cpufreq driver. This
607 	 * piece of code is only for compatibility with old Device
608 	 * Trees.
609 	 */
610 	np = of_find_compatible_node(NULL, NULL, "marvell,armada-xp-cpu-clock");
611 	if (!np)
612 		return 0;
613 
614 	ret = of_address_to_resource(np, 1, &res);
615 	if (ret) {
616 		pr_warn(FW_WARN "not enabling cpufreq, deprecated armada-xp-cpu-clock binding\n");
617 		of_node_put(np);
618 		return 0;
619 	}
620 
621 	of_node_put(np);
622 
623 	/*
624 	 * For each CPU, this loop registers the operating points
625 	 * supported (which are the nominal CPU frequency and half of
626 	 * it), and registers the clock notifier that will take care
627 	 * of doing the PMSU part of a frequency transition.
628 	 */
629 	for_each_possible_cpu(cpu) {
630 		struct device *cpu_dev;
631 		struct clk *clk;
632 		int ret;
633 
634 		cpu_dev = get_cpu_device(cpu);
635 		if (!cpu_dev) {
636 			pr_err("Cannot get CPU %d\n", cpu);
637 			continue;
638 		}
639 
640 		clk = clk_get(cpu_dev, 0);
641 		if (IS_ERR(clk)) {
642 			pr_err("Cannot get clock for CPU %d\n", cpu);
643 			return PTR_ERR(clk);
644 		}
645 
646 		/*
647 		 * In case of a failure of dev_pm_opp_add(), we don't
648 		 * bother with cleaning up the registered OPP (there's
649 		 * no function to do so), and simply cancel the
650 		 * registration of the cpufreq device.
651 		 */
652 		ret = dev_pm_opp_add(cpu_dev, clk_get_rate(clk), 0);
653 		if (ret) {
654 			clk_put(clk);
655 			return ret;
656 		}
657 
658 		ret = dev_pm_opp_add(cpu_dev, clk_get_rate(clk) / 2, 0);
659 		if (ret) {
660 			clk_put(clk);
661 			return ret;
662 		}
663 	}
664 
665 	platform_device_register_data(NULL, "cpufreq-dt", -1,
666 				      &cpufreq_dt_pd, sizeof(cpufreq_dt_pd));
667 	return 0;
668 }
669 
670 device_initcall(armada_xp_pmsu_cpufreq_init);
671