1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Marvell Armada 37xx SoC Peripheral clocks
4  *
5  * Copyright (C) 2016 Marvell
6  *
7  * Gregory CLEMENT <gregory.clement@free-electrons.com>
8  *
9  * Most of the peripheral clocks can be modelled like this:
10  *             _____    _______    _______
11  * TBG-A-P  --|     |  |       |  |       |   ______
12  * TBG-B-P  --| Mux |--| /div1 |--| /div2 |--| Gate |--> perip_clk
13  * TBG-A-S  --|     |  |       |  |       |  |______|
14  * TBG-B-S  --|_____|  |_______|  |_______|
15  *
16  * However some clocks may use only one or two block or and use the
17  * xtal clock as parent.
18  */
19 
20 #include <linux/clk-provider.h>
21 #include <linux/mfd/syscon.h>
22 #include <linux/of.h>
23 #include <linux/of_device.h>
24 #include <linux/platform_device.h>
25 #include <linux/regmap.h>
26 #include <linux/slab.h>
27 
28 #define TBG_SEL		0x0
29 #define DIV_SEL0	0x4
30 #define DIV_SEL1	0x8
31 #define DIV_SEL2	0xC
32 #define CLK_SEL		0x10
33 #define CLK_DIS		0x14
34 
35 #define  ARMADA_37XX_DVFS_LOAD_1 1
36 #define LOAD_LEVEL_NR	4
37 
38 #define ARMADA_37XX_NB_L0L1	0x18
39 #define ARMADA_37XX_NB_L2L3	0x1C
40 #define		ARMADA_37XX_NB_TBG_DIV_OFF	13
41 #define		ARMADA_37XX_NB_TBG_DIV_MASK	0x7
42 #define		ARMADA_37XX_NB_CLK_SEL_OFF	11
43 #define		ARMADA_37XX_NB_CLK_SEL_MASK	0x1
44 #define		ARMADA_37XX_NB_TBG_SEL_OFF	9
45 #define		ARMADA_37XX_NB_TBG_SEL_MASK	0x3
46 #define		ARMADA_37XX_NB_CONFIG_SHIFT	16
47 #define ARMADA_37XX_NB_DYN_MOD	0x24
48 #define		ARMADA_37XX_NB_DFS_EN	31
49 #define ARMADA_37XX_NB_CPU_LOAD	0x30
50 #define		ARMADA_37XX_NB_CPU_LOAD_MASK	0x3
51 #define		ARMADA_37XX_DVFS_LOAD_0		0
52 #define		ARMADA_37XX_DVFS_LOAD_1		1
53 #define		ARMADA_37XX_DVFS_LOAD_2		2
54 #define		ARMADA_37XX_DVFS_LOAD_3		3
55 
56 struct clk_periph_driver_data {
57 	struct clk_hw_onecell_data *hw_data;
58 	spinlock_t lock;
59 };
60 
61 struct clk_double_div {
62 	struct clk_hw hw;
63 	void __iomem *reg1;
64 	u8 shift1;
65 	void __iomem *reg2;
66 	u8 shift2;
67 };
68 
69 struct clk_pm_cpu {
70 	struct clk_hw hw;
71 	void __iomem *reg_mux;
72 	u8 shift_mux;
73 	u32 mask_mux;
74 	void __iomem *reg_div;
75 	u8 shift_div;
76 	struct regmap *nb_pm_base;
77 };
78 
79 #define to_clk_double_div(_hw) container_of(_hw, struct clk_double_div, hw)
80 #define to_clk_pm_cpu(_hw) container_of(_hw, struct clk_pm_cpu, hw)
81 
82 struct clk_periph_data {
83 	const char *name;
84 	const char * const *parent_names;
85 	int num_parents;
86 	struct clk_hw *mux_hw;
87 	struct clk_hw *rate_hw;
88 	struct clk_hw *gate_hw;
89 	struct clk_hw *muxrate_hw;
90 	bool is_double_div;
91 };
92 
93 static const struct clk_div_table clk_table6[] = {
94 	{ .val = 1, .div = 1, },
95 	{ .val = 2, .div = 2, },
96 	{ .val = 3, .div = 3, },
97 	{ .val = 4, .div = 4, },
98 	{ .val = 5, .div = 5, },
99 	{ .val = 6, .div = 6, },
100 	{ .val = 0, .div = 0, }, /* last entry */
101 };
102 
103 static const struct clk_div_table clk_table1[] = {
104 	{ .val = 0, .div = 1, },
105 	{ .val = 1, .div = 2, },
106 	{ .val = 0, .div = 0, }, /* last entry */
107 };
108 
109 static const struct clk_div_table clk_table2[] = {
110 	{ .val = 0, .div = 2, },
111 	{ .val = 1, .div = 4, },
112 	{ .val = 0, .div = 0, }, /* last entry */
113 };
114 
115 static const struct clk_ops clk_double_div_ops;
116 static const struct clk_ops clk_pm_cpu_ops;
117 
118 #define PERIPH_GATE(_name, _bit)		\
119 struct clk_gate gate_##_name = {		\
120 	.reg = (void *)CLK_DIS,			\
121 	.bit_idx = _bit,			\
122 	.hw.init = &(struct clk_init_data){	\
123 		.ops =  &clk_gate_ops,		\
124 	}					\
125 };
126 
127 #define PERIPH_MUX(_name, _shift)		\
128 struct clk_mux mux_##_name = {			\
129 	.reg = (void *)TBG_SEL,			\
130 	.shift = _shift,			\
131 	.mask = 3,				\
132 	.hw.init = &(struct clk_init_data){	\
133 		.ops =  &clk_mux_ro_ops,	\
134 	}					\
135 };
136 
137 #define PERIPH_DOUBLEDIV(_name, _reg1, _reg2, _shift1, _shift2)	\
138 struct clk_double_div rate_##_name = {		\
139 	.reg1 = (void *)_reg1,			\
140 	.reg2 = (void *)_reg2,			\
141 	.shift1 = _shift1,			\
142 	.shift2 = _shift2,			\
143 	.hw.init = &(struct clk_init_data){	\
144 		.ops =  &clk_double_div_ops,	\
145 	}					\
146 };
147 
148 #define PERIPH_DIV(_name, _reg, _shift, _table)	\
149 struct clk_divider rate_##_name = {		\
150 	.reg = (void *)_reg,			\
151 	.table = _table,			\
152 	.shift = _shift,			\
153 	.hw.init = &(struct clk_init_data){	\
154 		.ops =  &clk_divider_ro_ops,	\
155 	}					\
156 };
157 
158 #define PERIPH_PM_CPU(_name, _shift1, _reg, _shift2)	\
159 struct clk_pm_cpu muxrate_##_name = {		\
160 	.reg_mux = (void *)TBG_SEL,		\
161 	.mask_mux = 3,				\
162 	.shift_mux = _shift1,			\
163 	.reg_div = (void *)_reg,		\
164 	.shift_div = _shift2,			\
165 	.hw.init = &(struct clk_init_data){	\
166 		.ops =  &clk_pm_cpu_ops,	\
167 	}					\
168 };
169 
170 #define PERIPH_CLK_FULL_DD(_name, _bit, _shift, _reg1, _reg2, _shift1, _shift2)\
171 static PERIPH_GATE(_name, _bit);			    \
172 static PERIPH_MUX(_name, _shift);			    \
173 static PERIPH_DOUBLEDIV(_name, _reg1, _reg2, _shift1, _shift2);
174 
175 #define PERIPH_CLK_FULL(_name, _bit, _shift, _reg, _shift1, _table)	\
176 static PERIPH_GATE(_name, _bit);			    \
177 static PERIPH_MUX(_name, _shift);			    \
178 static PERIPH_DIV(_name, _reg, _shift1, _table);
179 
180 #define PERIPH_CLK_GATE_DIV(_name, _bit,  _reg, _shift, _table)	\
181 static PERIPH_GATE(_name, _bit);			\
182 static PERIPH_DIV(_name, _reg, _shift, _table);
183 
184 #define PERIPH_CLK_MUX_DD(_name, _shift, _reg1, _reg2, _shift1, _shift2)\
185 static PERIPH_MUX(_name, _shift);			    \
186 static PERIPH_DOUBLEDIV(_name, _reg1, _reg2, _shift1, _shift2);
187 
188 #define REF_CLK_FULL(_name)				\
189 	{ .name = #_name,				\
190 	  .parent_names = (const char *[]){ "TBG-A-P",	\
191 	      "TBG-B-P", "TBG-A-S", "TBG-B-S"},		\
192 	  .num_parents = 4,				\
193 	  .mux_hw = &mux_##_name.hw,			\
194 	  .gate_hw = &gate_##_name.hw,			\
195 	  .rate_hw = &rate_##_name.hw,			\
196 	}
197 
198 #define REF_CLK_FULL_DD(_name)				\
199 	{ .name = #_name,				\
200 	  .parent_names = (const char *[]){ "TBG-A-P",	\
201 	      "TBG-B-P", "TBG-A-S", "TBG-B-S"},		\
202 	  .num_parents = 4,				\
203 	  .mux_hw = &mux_##_name.hw,			\
204 	  .gate_hw = &gate_##_name.hw,			\
205 	  .rate_hw = &rate_##_name.hw,			\
206 	  .is_double_div = true,			\
207 	}
208 
209 #define REF_CLK_GATE(_name, _parent_name)			\
210 	{ .name = #_name,					\
211 	  .parent_names = (const char *[]){ _parent_name},	\
212 	  .num_parents = 1,					\
213 	  .gate_hw = &gate_##_name.hw,				\
214 	}
215 
216 #define REF_CLK_GATE_DIV(_name, _parent_name)			\
217 	{ .name = #_name,					\
218 	  .parent_names = (const char *[]){ _parent_name},	\
219 	  .num_parents = 1,					\
220 	  .gate_hw = &gate_##_name.hw,				\
221 	  .rate_hw = &rate_##_name.hw,				\
222 	}
223 
224 #define REF_CLK_PM_CPU(_name)				\
225 	{ .name = #_name,				\
226 	  .parent_names = (const char *[]){ "TBG-A-P",	\
227 	      "TBG-B-P", "TBG-A-S", "TBG-B-S"},		\
228 	  .num_parents = 4,				\
229 	  .muxrate_hw = &muxrate_##_name.hw,		\
230 	}
231 
232 #define REF_CLK_MUX_DD(_name)				\
233 	{ .name = #_name,				\
234 	  .parent_names = (const char *[]){ "TBG-A-P",	\
235 	      "TBG-B-P", "TBG-A-S", "TBG-B-S"},		\
236 	  .num_parents = 4,				\
237 	  .mux_hw = &mux_##_name.hw,			\
238 	  .rate_hw = &rate_##_name.hw,			\
239 	  .is_double_div = true,			\
240 	}
241 
242 /* NB periph clocks */
243 PERIPH_CLK_FULL_DD(mmc, 2, 0, DIV_SEL2, DIV_SEL2, 16, 13);
244 PERIPH_CLK_FULL_DD(sata_host, 3, 2, DIV_SEL2, DIV_SEL2, 10, 7);
245 PERIPH_CLK_FULL_DD(sec_at, 6, 4, DIV_SEL1, DIV_SEL1, 3, 0);
246 PERIPH_CLK_FULL_DD(sec_dap, 7, 6, DIV_SEL1, DIV_SEL1, 9, 6);
247 PERIPH_CLK_FULL_DD(tscem, 8, 8, DIV_SEL1, DIV_SEL1, 15, 12);
248 PERIPH_CLK_FULL(tscem_tmx, 10, 10, DIV_SEL1, 18, clk_table6);
249 static PERIPH_GATE(avs, 11);
250 PERIPH_CLK_FULL_DD(pwm, 13, 14, DIV_SEL0, DIV_SEL0, 3, 0);
251 PERIPH_CLK_FULL_DD(sqf, 12, 12, DIV_SEL1, DIV_SEL1, 27, 24);
252 static PERIPH_GATE(i2c_2, 16);
253 static PERIPH_GATE(i2c_1, 17);
254 PERIPH_CLK_GATE_DIV(ddr_phy, 19, DIV_SEL0, 18, clk_table2);
255 PERIPH_CLK_FULL_DD(ddr_fclk, 21, 16, DIV_SEL0, DIV_SEL0, 15, 12);
256 PERIPH_CLK_FULL(trace, 22, 18, DIV_SEL0, 20, clk_table6);
257 PERIPH_CLK_FULL(counter, 23, 20, DIV_SEL0, 23, clk_table6);
258 PERIPH_CLK_FULL_DD(eip97, 24, 24, DIV_SEL2, DIV_SEL2, 22, 19);
259 static PERIPH_PM_CPU(cpu, 22, DIV_SEL0, 28);
260 
261 static struct clk_periph_data data_nb[] = {
262 	REF_CLK_FULL_DD(mmc),
263 	REF_CLK_FULL_DD(sata_host),
264 	REF_CLK_FULL_DD(sec_at),
265 	REF_CLK_FULL_DD(sec_dap),
266 	REF_CLK_FULL_DD(tscem),
267 	REF_CLK_FULL(tscem_tmx),
268 	REF_CLK_GATE(avs, "xtal"),
269 	REF_CLK_FULL_DD(sqf),
270 	REF_CLK_FULL_DD(pwm),
271 	REF_CLK_GATE(i2c_2, "xtal"),
272 	REF_CLK_GATE(i2c_1, "xtal"),
273 	REF_CLK_GATE_DIV(ddr_phy, "TBG-A-S"),
274 	REF_CLK_FULL_DD(ddr_fclk),
275 	REF_CLK_FULL(trace),
276 	REF_CLK_FULL(counter),
277 	REF_CLK_FULL_DD(eip97),
278 	REF_CLK_PM_CPU(cpu),
279 	{ },
280 };
281 
282 /* SB periph clocks */
283 PERIPH_CLK_MUX_DD(gbe_50, 6, DIV_SEL2, DIV_SEL2, 6, 9);
284 PERIPH_CLK_MUX_DD(gbe_core, 8, DIV_SEL1, DIV_SEL1, 18, 21);
285 PERIPH_CLK_MUX_DD(gbe_125, 10, DIV_SEL1, DIV_SEL1, 6, 9);
286 static PERIPH_GATE(gbe1_50, 0);
287 static PERIPH_GATE(gbe0_50, 1);
288 static PERIPH_GATE(gbe1_125, 2);
289 static PERIPH_GATE(gbe0_125, 3);
290 PERIPH_CLK_GATE_DIV(gbe1_core, 4, DIV_SEL1, 13, clk_table1);
291 PERIPH_CLK_GATE_DIV(gbe0_core, 5, DIV_SEL1, 14, clk_table1);
292 PERIPH_CLK_GATE_DIV(gbe_bm, 12, DIV_SEL1, 0, clk_table1);
293 PERIPH_CLK_FULL_DD(sdio, 11, 14, DIV_SEL0, DIV_SEL0, 3, 6);
294 PERIPH_CLK_FULL_DD(usb32_usb2_sys, 16, 16, DIV_SEL0, DIV_SEL0, 9, 12);
295 PERIPH_CLK_FULL_DD(usb32_ss_sys, 17, 18, DIV_SEL0, DIV_SEL0, 15, 18);
296 
297 static struct clk_periph_data data_sb[] = {
298 	REF_CLK_MUX_DD(gbe_50),
299 	REF_CLK_MUX_DD(gbe_core),
300 	REF_CLK_MUX_DD(gbe_125),
301 	REF_CLK_GATE(gbe1_50, "gbe_50"),
302 	REF_CLK_GATE(gbe0_50, "gbe_50"),
303 	REF_CLK_GATE(gbe1_125, "gbe_125"),
304 	REF_CLK_GATE(gbe0_125, "gbe_125"),
305 	REF_CLK_GATE_DIV(gbe1_core, "gbe_core"),
306 	REF_CLK_GATE_DIV(gbe0_core, "gbe_core"),
307 	REF_CLK_GATE_DIV(gbe_bm, "gbe_core"),
308 	REF_CLK_FULL_DD(sdio),
309 	REF_CLK_FULL_DD(usb32_usb2_sys),
310 	REF_CLK_FULL_DD(usb32_ss_sys),
311 	{ },
312 };
313 
314 static unsigned int get_div(void __iomem *reg, int shift)
315 {
316 	u32 val;
317 
318 	val = (readl(reg) >> shift) & 0x7;
319 	if (val > 6)
320 		return 0;
321 	return val;
322 }
323 
324 static unsigned long clk_double_div_recalc_rate(struct clk_hw *hw,
325 						unsigned long parent_rate)
326 {
327 	struct clk_double_div *double_div = to_clk_double_div(hw);
328 	unsigned int div;
329 
330 	div = get_div(double_div->reg1, double_div->shift1);
331 	div *= get_div(double_div->reg2, double_div->shift2);
332 
333 	return DIV_ROUND_UP_ULL((u64)parent_rate, div);
334 }
335 
336 static const struct clk_ops clk_double_div_ops = {
337 	.recalc_rate = clk_double_div_recalc_rate,
338 };
339 
340 static void armada_3700_pm_dvfs_update_regs(unsigned int load_level,
341 					    unsigned int *reg,
342 					    unsigned int *offset)
343 {
344 	if (load_level <= ARMADA_37XX_DVFS_LOAD_1)
345 		*reg = ARMADA_37XX_NB_L0L1;
346 	else
347 		*reg = ARMADA_37XX_NB_L2L3;
348 
349 	if (load_level == ARMADA_37XX_DVFS_LOAD_0 ||
350 	    load_level ==  ARMADA_37XX_DVFS_LOAD_2)
351 		*offset += ARMADA_37XX_NB_CONFIG_SHIFT;
352 }
353 
354 static bool armada_3700_pm_dvfs_is_enabled(struct regmap *base)
355 {
356 	unsigned int val, reg = ARMADA_37XX_NB_DYN_MOD;
357 
358 	if (IS_ERR(base))
359 		return false;
360 
361 	regmap_read(base, reg, &val);
362 
363 	return !!(val & BIT(ARMADA_37XX_NB_DFS_EN));
364 }
365 
366 static unsigned int armada_3700_pm_dvfs_get_cpu_div(struct regmap *base)
367 {
368 	unsigned int reg = ARMADA_37XX_NB_CPU_LOAD;
369 	unsigned int offset = ARMADA_37XX_NB_TBG_DIV_OFF;
370 	unsigned int load_level, div;
371 
372 	/*
373 	 * This function is always called after the function
374 	 * armada_3700_pm_dvfs_is_enabled, so no need to check again
375 	 * if the base is valid.
376 	 */
377 	regmap_read(base, reg, &load_level);
378 
379 	/*
380 	 * The register and the offset inside this register accessed to
381 	 * read the current divider depend on the load level
382 	 */
383 	load_level &= ARMADA_37XX_NB_CPU_LOAD_MASK;
384 	armada_3700_pm_dvfs_update_regs(load_level, &reg, &offset);
385 
386 	regmap_read(base, reg, &div);
387 
388 	return (div >> offset) & ARMADA_37XX_NB_TBG_DIV_MASK;
389 }
390 
391 static unsigned int armada_3700_pm_dvfs_get_cpu_parent(struct regmap *base)
392 {
393 	unsigned int reg = ARMADA_37XX_NB_CPU_LOAD;
394 	unsigned int offset = ARMADA_37XX_NB_TBG_SEL_OFF;
395 	unsigned int load_level, sel;
396 
397 	/*
398 	 * This function is always called after the function
399 	 * armada_3700_pm_dvfs_is_enabled, so no need to check again
400 	 * if the base is valid
401 	 */
402 	regmap_read(base, reg, &load_level);
403 
404 	/*
405 	 * The register and the offset inside this register accessed to
406 	 * read the current divider depend on the load level
407 	 */
408 	load_level &= ARMADA_37XX_NB_CPU_LOAD_MASK;
409 	armada_3700_pm_dvfs_update_regs(load_level, &reg, &offset);
410 
411 	regmap_read(base, reg, &sel);
412 
413 	return (sel >> offset) & ARMADA_37XX_NB_TBG_SEL_MASK;
414 }
415 
416 static u8 clk_pm_cpu_get_parent(struct clk_hw *hw)
417 {
418 	struct clk_pm_cpu *pm_cpu = to_clk_pm_cpu(hw);
419 	u32 val;
420 
421 	if (armada_3700_pm_dvfs_is_enabled(pm_cpu->nb_pm_base)) {
422 		val = armada_3700_pm_dvfs_get_cpu_parent(pm_cpu->nb_pm_base);
423 	} else {
424 		val = readl(pm_cpu->reg_mux) >> pm_cpu->shift_mux;
425 		val &= pm_cpu->mask_mux;
426 	}
427 
428 	return val;
429 }
430 
431 static int clk_pm_cpu_set_parent(struct clk_hw *hw, u8 index)
432 {
433 	struct clk_pm_cpu *pm_cpu = to_clk_pm_cpu(hw);
434 	struct regmap *base = pm_cpu->nb_pm_base;
435 	int load_level;
436 
437 	/*
438 	 * We set the clock parent only if the DVFS is available but
439 	 * not enabled.
440 	 */
441 	if (IS_ERR(base) || armada_3700_pm_dvfs_is_enabled(base))
442 		return -EINVAL;
443 
444 	/* Set the parent clock for all the load level */
445 	for (load_level = 0; load_level < LOAD_LEVEL_NR; load_level++) {
446 		unsigned int reg, mask,  val,
447 			offset = ARMADA_37XX_NB_TBG_SEL_OFF;
448 
449 		armada_3700_pm_dvfs_update_regs(load_level, &reg, &offset);
450 
451 		val = index << offset;
452 		mask = ARMADA_37XX_NB_TBG_SEL_MASK << offset;
453 		regmap_update_bits(base, reg, mask, val);
454 	}
455 	return 0;
456 }
457 
458 static unsigned long clk_pm_cpu_recalc_rate(struct clk_hw *hw,
459 					    unsigned long parent_rate)
460 {
461 	struct clk_pm_cpu *pm_cpu = to_clk_pm_cpu(hw);
462 	unsigned int div;
463 
464 	if (armada_3700_pm_dvfs_is_enabled(pm_cpu->nb_pm_base))
465 		div = armada_3700_pm_dvfs_get_cpu_div(pm_cpu->nb_pm_base);
466 	else
467 		div = get_div(pm_cpu->reg_div, pm_cpu->shift_div);
468 	return DIV_ROUND_UP_ULL((u64)parent_rate, div);
469 }
470 
471 static long clk_pm_cpu_round_rate(struct clk_hw *hw, unsigned long rate,
472 				  unsigned long *parent_rate)
473 {
474 	struct clk_pm_cpu *pm_cpu = to_clk_pm_cpu(hw);
475 	struct regmap *base = pm_cpu->nb_pm_base;
476 	unsigned int div = *parent_rate / rate;
477 	unsigned int load_level;
478 	/* only available when DVFS is enabled */
479 	if (!armada_3700_pm_dvfs_is_enabled(base))
480 		return -EINVAL;
481 
482 	for (load_level = 0; load_level < LOAD_LEVEL_NR; load_level++) {
483 		unsigned int reg, val, offset = ARMADA_37XX_NB_TBG_DIV_OFF;
484 
485 		armada_3700_pm_dvfs_update_regs(load_level, &reg, &offset);
486 
487 		regmap_read(base, reg, &val);
488 
489 		val >>= offset;
490 		val &= ARMADA_37XX_NB_TBG_DIV_MASK;
491 		if (val == div)
492 			/*
493 			 * We found a load level matching the target
494 			 * divider, switch to this load level and
495 			 * return.
496 			 */
497 			return *parent_rate / div;
498 	}
499 
500 	/* We didn't find any valid divider */
501 	return -EINVAL;
502 }
503 
504 /*
505  * Switching the CPU from the L2 or L3 frequencies (300 and 200 Mhz
506  * respectively) to L0 frequency (1.2 Ghz) requires a significant
507  * amount of time to let VDD stabilize to the appropriate
508  * voltage. This amount of time is large enough that it cannot be
509  * covered by the hardware countdown register. Due to this, the CPU
510  * might start operating at L0 before the voltage is stabilized,
511  * leading to CPU stalls.
512  *
513  * To work around this problem, we prevent switching directly from the
514  * L2/L3 frequencies to the L0 frequency, and instead switch to the L1
515  * frequency in-between. The sequence therefore becomes:
516  * 1. First switch from L2/L3(200/300MHz) to L1(600MHZ)
517  * 2. Sleep 20ms for stabling VDD voltage
518  * 3. Then switch from L1(600MHZ) to L0(1200Mhz).
519  */
520 static void clk_pm_cpu_set_rate_wa(unsigned long rate, struct regmap *base)
521 {
522 	unsigned int cur_level;
523 
524 	if (rate != 1200 * 1000 * 1000)
525 		return;
526 
527 	regmap_read(base, ARMADA_37XX_NB_CPU_LOAD, &cur_level);
528 	cur_level &= ARMADA_37XX_NB_CPU_LOAD_MASK;
529 	if (cur_level <= ARMADA_37XX_DVFS_LOAD_1)
530 		return;
531 
532 	regmap_update_bits(base, ARMADA_37XX_NB_CPU_LOAD,
533 			   ARMADA_37XX_NB_CPU_LOAD_MASK,
534 			   ARMADA_37XX_DVFS_LOAD_1);
535 	msleep(20);
536 }
537 
538 static int clk_pm_cpu_set_rate(struct clk_hw *hw, unsigned long rate,
539 			       unsigned long parent_rate)
540 {
541 	struct clk_pm_cpu *pm_cpu = to_clk_pm_cpu(hw);
542 	struct regmap *base = pm_cpu->nb_pm_base;
543 	unsigned int div = parent_rate / rate;
544 	unsigned int load_level;
545 
546 	/* only available when DVFS is enabled */
547 	if (!armada_3700_pm_dvfs_is_enabled(base))
548 		return -EINVAL;
549 
550 	for (load_level = 0; load_level < LOAD_LEVEL_NR; load_level++) {
551 		unsigned int reg, mask, val,
552 			offset = ARMADA_37XX_NB_TBG_DIV_OFF;
553 
554 		armada_3700_pm_dvfs_update_regs(load_level, &reg, &offset);
555 
556 		regmap_read(base, reg, &val);
557 		val >>= offset;
558 		val &= ARMADA_37XX_NB_TBG_DIV_MASK;
559 
560 		if (val == div) {
561 			/*
562 			 * We found a load level matching the target
563 			 * divider, switch to this load level and
564 			 * return.
565 			 */
566 			reg = ARMADA_37XX_NB_CPU_LOAD;
567 			mask = ARMADA_37XX_NB_CPU_LOAD_MASK;
568 
569 			clk_pm_cpu_set_rate_wa(rate, base);
570 
571 			regmap_update_bits(base, reg, mask, load_level);
572 
573 			return rate;
574 		}
575 	}
576 
577 	/* We didn't find any valid divider */
578 	return -EINVAL;
579 }
580 
581 static const struct clk_ops clk_pm_cpu_ops = {
582 	.get_parent = clk_pm_cpu_get_parent,
583 	.set_parent = clk_pm_cpu_set_parent,
584 	.round_rate = clk_pm_cpu_round_rate,
585 	.set_rate = clk_pm_cpu_set_rate,
586 	.recalc_rate = clk_pm_cpu_recalc_rate,
587 };
588 
589 static const struct of_device_id armada_3700_periph_clock_of_match[] = {
590 	{ .compatible = "marvell,armada-3700-periph-clock-nb",
591 	  .data = data_nb, },
592 	{ .compatible = "marvell,armada-3700-periph-clock-sb",
593 	.data = data_sb, },
594 	{ }
595 };
596 
597 static int armada_3700_add_composite_clk(const struct clk_periph_data *data,
598 					 void __iomem *reg, spinlock_t *lock,
599 					 struct device *dev, struct clk_hw **hw)
600 {
601 	const struct clk_ops *mux_ops = NULL, *gate_ops = NULL,
602 		*rate_ops = NULL;
603 	struct clk_hw *mux_hw = NULL, *gate_hw = NULL, *rate_hw = NULL;
604 
605 	if (data->mux_hw) {
606 		struct clk_mux *mux;
607 
608 		mux_hw = data->mux_hw;
609 		mux = to_clk_mux(mux_hw);
610 		mux->lock = lock;
611 		mux_ops = mux_hw->init->ops;
612 		mux->reg = reg + (u64)mux->reg;
613 	}
614 
615 	if (data->gate_hw) {
616 		struct clk_gate *gate;
617 
618 		gate_hw = data->gate_hw;
619 		gate = to_clk_gate(gate_hw);
620 		gate->lock = lock;
621 		gate_ops = gate_hw->init->ops;
622 		gate->reg = reg + (u64)gate->reg;
623 		gate->flags = CLK_GATE_SET_TO_DISABLE;
624 	}
625 
626 	if (data->rate_hw) {
627 		rate_hw = data->rate_hw;
628 		rate_ops = rate_hw->init->ops;
629 		if (data->is_double_div) {
630 			struct clk_double_div *rate;
631 
632 			rate =  to_clk_double_div(rate_hw);
633 			rate->reg1 = reg + (u64)rate->reg1;
634 			rate->reg2 = reg + (u64)rate->reg2;
635 		} else {
636 			struct clk_divider *rate = to_clk_divider(rate_hw);
637 			const struct clk_div_table *clkt;
638 			int table_size = 0;
639 
640 			rate->reg = reg + (u64)rate->reg;
641 			for (clkt = rate->table; clkt->div; clkt++)
642 				table_size++;
643 			rate->width = order_base_2(table_size);
644 			rate->lock = lock;
645 		}
646 	}
647 
648 	if (data->muxrate_hw) {
649 		struct clk_pm_cpu *pmcpu_clk;
650 		struct clk_hw *muxrate_hw = data->muxrate_hw;
651 		struct regmap *map;
652 
653 		pmcpu_clk =  to_clk_pm_cpu(muxrate_hw);
654 		pmcpu_clk->reg_mux = reg + (u64)pmcpu_clk->reg_mux;
655 		pmcpu_clk->reg_div = reg + (u64)pmcpu_clk->reg_div;
656 
657 		mux_hw = muxrate_hw;
658 		rate_hw = muxrate_hw;
659 		mux_ops = muxrate_hw->init->ops;
660 		rate_ops = muxrate_hw->init->ops;
661 
662 		map = syscon_regmap_lookup_by_compatible(
663 				"marvell,armada-3700-nb-pm");
664 		pmcpu_clk->nb_pm_base = map;
665 	}
666 
667 	*hw = clk_hw_register_composite(dev, data->name, data->parent_names,
668 					data->num_parents, mux_hw,
669 					mux_ops, rate_hw, rate_ops,
670 					gate_hw, gate_ops, CLK_IGNORE_UNUSED);
671 
672 	return PTR_ERR_OR_ZERO(*hw);
673 }
674 
675 static int armada_3700_periph_clock_probe(struct platform_device *pdev)
676 {
677 	struct clk_periph_driver_data *driver_data;
678 	struct device_node *np = pdev->dev.of_node;
679 	const struct clk_periph_data *data;
680 	struct device *dev = &pdev->dev;
681 	int num_periph = 0, i, ret;
682 	struct resource *res;
683 	void __iomem *reg;
684 
685 	data = of_device_get_match_data(dev);
686 	if (!data)
687 		return -ENODEV;
688 
689 	while (data[num_periph].name)
690 		num_periph++;
691 
692 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
693 	reg = devm_ioremap_resource(dev, res);
694 	if (IS_ERR(reg))
695 		return PTR_ERR(reg);
696 
697 	driver_data = devm_kzalloc(dev, sizeof(*driver_data), GFP_KERNEL);
698 	if (!driver_data)
699 		return -ENOMEM;
700 
701 	driver_data->hw_data = devm_kzalloc(dev,
702 					    struct_size(driver_data->hw_data,
703 							hws, num_periph),
704 					    GFP_KERNEL);
705 	if (!driver_data->hw_data)
706 		return -ENOMEM;
707 	driver_data->hw_data->num = num_periph;
708 
709 	spin_lock_init(&driver_data->lock);
710 
711 	for (i = 0; i < num_periph; i++) {
712 		struct clk_hw **hw = &driver_data->hw_data->hws[i];
713 
714 		if (armada_3700_add_composite_clk(&data[i], reg,
715 						  &driver_data->lock, dev, hw))
716 			dev_err(dev, "Can't register periph clock %s\n",
717 				data[i].name);
718 	}
719 
720 	ret = of_clk_add_hw_provider(np, of_clk_hw_onecell_get,
721 				     driver_data->hw_data);
722 	if (ret) {
723 		for (i = 0; i < num_periph; i++)
724 			clk_hw_unregister(driver_data->hw_data->hws[i]);
725 		return ret;
726 	}
727 
728 	platform_set_drvdata(pdev, driver_data);
729 	return 0;
730 }
731 
732 static int armada_3700_periph_clock_remove(struct platform_device *pdev)
733 {
734 	struct clk_periph_driver_data *data = platform_get_drvdata(pdev);
735 	struct clk_hw_onecell_data *hw_data = data->hw_data;
736 	int i;
737 
738 	of_clk_del_provider(pdev->dev.of_node);
739 
740 	for (i = 0; i < hw_data->num; i++)
741 		clk_hw_unregister(hw_data->hws[i]);
742 
743 	return 0;
744 }
745 
746 static struct platform_driver armada_3700_periph_clock_driver = {
747 	.probe = armada_3700_periph_clock_probe,
748 	.remove = armada_3700_periph_clock_remove,
749 	.driver		= {
750 		.name	= "marvell-armada-3700-periph-clock",
751 		.of_match_table = armada_3700_periph_clock_of_match,
752 	},
753 };
754 
755 builtin_platform_driver(armada_3700_periph_clock_driver);
756