1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2019-20 Sean Anderson <seanga2@gmail.com>
4  */
5 #define LOG_CATEGORY UCLASS_CLK
6 
7 #include <common.h>
8 #include <dm.h>
9 /* For DIV_ROUND_DOWN_ULL, defined in linux/kernel.h */
10 #include <div64.h>
11 #include <log.h>
12 #include <serial.h>
13 #include <asm/io.h>
14 #include <dt-bindings/clock/k210-sysctl.h>
15 #include <kendryte/pll.h>
16 #include <linux/bitfield.h>
17 #include <linux/clk-provider.h>
18 #include <linux/delay.h>
19 #include <linux/err.h>
20 
21 #define CLK_K210_PLL "k210_clk_pll"
22 
23 #ifdef CONFIG_CLK_K210_SET_RATE
24 static int k210_pll_enable(struct clk *clk);
25 static int k210_pll_disable(struct clk *clk);
26 
27 /*
28  * The PLL included with the Kendryte K210 appears to be a True Circuits, Inc.
29  * General-Purpose PLL. The logical layout of the PLL with internal feedback is
30  * approximately the following:
31  *
32  *  +---------------+
33  *  |reference clock|
34  *  +---------------+
35  *          |
36  *          v
37  *        +--+
38  *        |/r|
39  *        +--+
40  *          |
41  *          v
42  *   +-------------+
43  *   |divided clock|
44  *   +-------------+
45  *          |
46  *          v
47  *  +--------------+
48  *  |phase detector|<---+
49  *  +--------------+    |
50  *          |           |
51  *          v   +--------------+
52  *        +---+ |feedback clock|
53  *        |VCO| +--------------+
54  *        +---+         ^
55  *          |    +--+   |
56  *          +--->|/f|---+
57  *          |    +--+
58  *          v
59  *        +---+
60  *        |/od|
61  *        +---+
62  *          |
63  *          v
64  *       +------+
65  *       |output|
66  *       +------+
67  *
68  * The k210 PLLs have three factors: r, f, and od. Because of the feedback mode,
69  * the effect of the division by f is to multiply the input frequency. The
70  * equation for the output rate is
71  *   rate = (rate_in * f) / (r * od).
72  * Moving knowns to one side of the equation, we get
73  *   rate / rate_in = f / (r * od)
74  * Rearranging slightly,
75  *   abs_error = abs((rate / rate_in) - (f / (r * od))).
76  * To get relative, error, we divide by the expected ratio
77  *   error = abs((rate / rate_in) - (f / (r * od))) / (rate / rate_in).
78  * Simplifying,
79  *   error = abs(1 - f / (r * od)) / (rate / rate_in)
80  *   error = abs(1 - (f * rate_in) / (r * od * rate))
81  * Using the constants ratio = rate / rate_in and inv_ratio = rate_in / rate,
82  *   error = abs((f * inv_ratio) / (r * od) - 1)
83  * This is the error used in evaluating parameters.
84  *
85  * r and od are four bits each, while f is six bits. Because r and od are
86  * multiplied together, instead of the full 256 values possible if both bits
87  * were used fully, there are only 97 distinct products. Combined with f, there
88  * are 6208 theoretical settings for the PLL. However, most of these settings
89  * can be ruled out immediately because they do not have the correct ratio.
90  *
91  * In addition to the constraint of approximating the desired ratio, parameters
92  * must also keep internal pll frequencies within acceptable ranges. The divided
93  * clock's minimum and maximum frequencies have a ratio of around 128.  This
94  * leaves fairly substantial room to work with, especially since the only
95  * affected parameter is r. The VCO's minimum and maximum frequency have a ratio
96  * of 5, which is considerably more restrictive.
97  *
98  * The r and od factors are stored in a table. This is to make it easy to find
99  * the next-largest product. Some products have multiple factorizations, but
100  * only when one factor has at least a 2.5x ratio to the factors of the other
101  * factorization. This is because any smaller ratio would not make a difference
102  * when ensuring the VCO's frequency is within spec.
103  *
104  * Throughout the calculation function, fixed point arithmetic is used. Because
105  * the range of rate and rate_in may be up to 1.75 GHz, or around 2^30, 64-bit
106  * 32.32 fixed-point numbers are used to represent ratios. In general, to
107  * implement division, the numerator is first multiplied by 2^32. This gives a
108  * result where the whole number part is in the upper 32 bits, and the fraction
109  * is in the lower 32 bits.
110  *
111  * In general, rounding is done to the closest integer. This helps find the best
112  * approximation for the ratio. Rounding in one direction (e.g down) could cause
113  * the function to miss a better ratio with one of the parameters increased by
114  * one.
115  */
116 
117 /*
118  * The factors table was generated with the following python code:
119  *
120  * def p(x, y):
121  *    return (1.0*x/y > 2.5) or (1.0*y/x > 2.5)
122  *
123  * factors = {}
124  * for i in range(1, 17):
125  *    for j in range(1, 17):
126  *       fs = factors.get(i*j) or []
127  *       if fs == [] or all([
128  *             (p(i, x) and p(i, y)) or (p(j, x) and p(j, y))
129  *             for (x, y) in fs]):
130  *          fs.append((i, j))
131  *          factors[i*j] = fs
132  *
133  * for k, l in sorted(factors.items()):
134  *    for v in l:
135  *       print("PACK(%s, %s)," % v)
136  */
137 #define PACK(r, od) (((((r) - 1) & 0xF) << 4) | (((od) - 1) & 0xF))
138 #define UNPACK_R(val) ((((val) >> 4) & 0xF) + 1)
139 #define UNPACK_OD(val) (((val) & 0xF) + 1)
140 static const u8 factors[] = {
141 	PACK(1, 1),
142 	PACK(1, 2),
143 	PACK(1, 3),
144 	PACK(1, 4),
145 	PACK(1, 5),
146 	PACK(1, 6),
147 	PACK(1, 7),
148 	PACK(1, 8),
149 	PACK(1, 9),
150 	PACK(3, 3),
151 	PACK(1, 10),
152 	PACK(1, 11),
153 	PACK(1, 12),
154 	PACK(3, 4),
155 	PACK(1, 13),
156 	PACK(1, 14),
157 	PACK(1, 15),
158 	PACK(3, 5),
159 	PACK(1, 16),
160 	PACK(4, 4),
161 	PACK(2, 9),
162 	PACK(2, 10),
163 	PACK(3, 7),
164 	PACK(2, 11),
165 	PACK(2, 12),
166 	PACK(5, 5),
167 	PACK(2, 13),
168 	PACK(3, 9),
169 	PACK(2, 14),
170 	PACK(2, 15),
171 	PACK(2, 16),
172 	PACK(3, 11),
173 	PACK(5, 7),
174 	PACK(3, 12),
175 	PACK(3, 13),
176 	PACK(4, 10),
177 	PACK(3, 14),
178 	PACK(4, 11),
179 	PACK(3, 15),
180 	PACK(3, 16),
181 	PACK(7, 7),
182 	PACK(5, 10),
183 	PACK(4, 13),
184 	PACK(6, 9),
185 	PACK(5, 11),
186 	PACK(4, 14),
187 	PACK(4, 15),
188 	PACK(7, 9),
189 	PACK(4, 16),
190 	PACK(5, 13),
191 	PACK(6, 11),
192 	PACK(5, 14),
193 	PACK(6, 12),
194 	PACK(5, 15),
195 	PACK(7, 11),
196 	PACK(6, 13),
197 	PACK(5, 16),
198 	PACK(9, 9),
199 	PACK(6, 14),
200 	PACK(8, 11),
201 	PACK(6, 15),
202 	PACK(7, 13),
203 	PACK(6, 16),
204 	PACK(7, 14),
205 	PACK(9, 11),
206 	PACK(10, 10),
207 	PACK(8, 13),
208 	PACK(7, 15),
209 	PACK(9, 12),
210 	PACK(10, 11),
211 	PACK(7, 16),
212 	PACK(9, 13),
213 	PACK(8, 15),
214 	PACK(11, 11),
215 	PACK(9, 14),
216 	PACK(8, 16),
217 	PACK(10, 13),
218 	PACK(11, 12),
219 	PACK(9, 15),
220 	PACK(10, 14),
221 	PACK(11, 13),
222 	PACK(9, 16),
223 	PACK(10, 15),
224 	PACK(11, 14),
225 	PACK(12, 13),
226 	PACK(10, 16),
227 	PACK(11, 15),
228 	PACK(12, 14),
229 	PACK(13, 13),
230 	PACK(11, 16),
231 	PACK(12, 15),
232 	PACK(13, 14),
233 	PACK(12, 16),
234 	PACK(13, 15),
235 	PACK(14, 14),
236 	PACK(13, 16),
237 	PACK(14, 15),
238 	PACK(14, 16),
239 	PACK(15, 15),
240 	PACK(15, 16),
241 	PACK(16, 16),
242 };
243 
k210_pll_calc_config(u32 rate,u32 rate_in,struct k210_pll_config * best)244 TEST_STATIC int k210_pll_calc_config(u32 rate, u32 rate_in,
245 				     struct k210_pll_config *best)
246 {
247 	int i;
248 	s64 error, best_error;
249 	u64 ratio, inv_ratio; /* fixed point 32.32 ratio of the rates */
250 	u64 max_r;
251 	u64 r, f, od;
252 
253 	/*
254 	 * Can't go over 1.75 GHz or under 21.25 MHz due to limitations on the
255 	 * VCO frequency. These are not the same limits as below because od can
256 	 * reduce the output frequency by 16.
257 	 */
258 	if (rate > 1750000000 || rate < 21250000)
259 		return -EINVAL;
260 
261 	/* Similar restrictions on the input rate */
262 	if (rate_in > 1750000000 || rate_in < 13300000)
263 		return -EINVAL;
264 
265 	ratio = DIV_ROUND_CLOSEST_ULL((u64)rate << 32, rate_in);
266 	inv_ratio = DIV_ROUND_CLOSEST_ULL((u64)rate_in << 32, rate);
267 	/* Can't increase by more than 64 or reduce by more than 256 */
268 	if (rate > rate_in && ratio > (64ULL << 32))
269 		return -EINVAL;
270 	else if (rate <= rate_in && inv_ratio > (256ULL << 32))
271 		return -EINVAL;
272 
273 	/*
274 	 * The divided clock (rate_in / r) must stay between 1.75 GHz and 13.3
275 	 * MHz. There is no minimum, since the only way to get a higher input
276 	 * clock than 26 MHz is to use a clock generated by a PLL. Because PLLs
277 	 * cannot output frequencies greater than 1.75 GHz, the minimum would
278 	 * never be greater than one.
279 	 */
280 	max_r = DIV_ROUND_DOWN_ULL(rate_in, 13300000);
281 
282 	/* Variables get immediately incremented, so start at -1th iteration */
283 	i = -1;
284 	f = 0;
285 	r = 0;
286 	od = 0;
287 	best_error = S64_MAX;
288 	error = best_error;
289 	/* do-while here so we always try at least one ratio */
290 	do {
291 		/*
292 		 * Whether we swapped r and od while enforcing frequency limits
293 		 */
294 		bool swapped = false;
295 		u64 last_od = od;
296 		u64 last_r = r;
297 
298 		/*
299 		 * Try the next largest value for f (or r and od) and
300 		 * recalculate the other parameters based on that
301 		 */
302 		if (rate > rate_in) {
303 			/*
304 			 * Skip factors of the same product if we already tried
305 			 * out that product
306 			 */
307 			do {
308 				i++;
309 				r = UNPACK_R(factors[i]);
310 				od = UNPACK_OD(factors[i]);
311 			} while (i + 1 < ARRAY_SIZE(factors) &&
312 				 r * od == last_r * last_od);
313 
314 			/* Round close */
315 			f = (r * od * ratio + BIT(31)) >> 32;
316 			if (f > 64)
317 				f = 64;
318 		} else {
319 			u64 tmp = ++f * inv_ratio;
320 			bool round_up = !!(tmp & BIT(31));
321 			u32 goal = (tmp >> 32) + round_up;
322 			u32 err, last_err;
323 
324 			/* Get the next r/od pair in factors */
325 			while (r * od < goal && i + 1 < ARRAY_SIZE(factors)) {
326 				i++;
327 				r = UNPACK_R(factors[i]);
328 				od = UNPACK_OD(factors[i]);
329 			}
330 
331 			/*
332 			 * This is a case of double rounding. If we rounded up
333 			 * above, we need to round down (in cases of ties) here.
334 			 * This prevents off-by-one errors resulting from
335 			 * choosing X+2 over X when X.Y rounds up to X+1 and
336 			 * there is no r * od = X+1. For the converse, when X.Y
337 			 * is rounded down to X, we should choose X+1 over X-1.
338 			 */
339 			err = abs(r * od - goal);
340 			last_err = abs(last_r * last_od - goal);
341 			if (last_err < err || (round_up && last_err == err)) {
342 				i--;
343 				r = last_r;
344 				od = last_od;
345 			}
346 		}
347 
348 		/*
349 		 * Enforce limits on internal clock frequencies. If we
350 		 * aren't in spec, try swapping r and od. If everything is
351 		 * in-spec, calculate the relative error.
352 		 */
353 		while (true) {
354 			/*
355 			 * Whether the intermediate frequencies are out-of-spec
356 			 */
357 			bool out_of_spec = false;
358 
359 			if (r > max_r) {
360 				out_of_spec = true;
361 			} else {
362 				/*
363 				 * There is no way to only divide once; we need
364 				 * to examine the frequency with and without the
365 				 * effect of od.
366 				 */
367 				u64 vco = DIV_ROUND_CLOSEST_ULL(rate_in * f, r);
368 
369 				if (vco > 1750000000 || vco < 340000000)
370 					out_of_spec = true;
371 			}
372 
373 			if (out_of_spec) {
374 				if (!swapped) {
375 					u64 tmp = r;
376 
377 					r = od;
378 					od = tmp;
379 					swapped = true;
380 					continue;
381 				} else {
382 					/*
383 					 * Try looking ahead to see if there are
384 					 * additional factors for the same
385 					 * product.
386 					 */
387 					if (i + 1 < ARRAY_SIZE(factors)) {
388 						u64 new_r, new_od;
389 
390 						i++;
391 						new_r = UNPACK_R(factors[i]);
392 						new_od = UNPACK_OD(factors[i]);
393 						if (r * od == new_r * new_od) {
394 							r = new_r;
395 							od = new_od;
396 							swapped = false;
397 							continue;
398 						}
399 						i--;
400 					}
401 					break;
402 				}
403 			}
404 
405 			error = DIV_ROUND_CLOSEST_ULL(f * inv_ratio, r * od);
406 			/* The lower 16 bits are spurious */
407 			error = abs((error - BIT(32))) >> 16;
408 
409 			if (error < best_error) {
410 				best->r = r;
411 				best->f = f;
412 				best->od = od;
413 				best_error = error;
414 			}
415 			break;
416 		}
417 	} while (f < 64 && i + 1 < ARRAY_SIZE(factors) && error != 0);
418 
419 	if (best_error == S64_MAX)
420 		return -EINVAL;
421 
422 	log_debug("best error %lld\n", best_error);
423 	return 0;
424 }
425 
k210_pll_set_rate(struct clk * clk,ulong rate)426 static ulong k210_pll_set_rate(struct clk *clk, ulong rate)
427 {
428 	int err;
429 	long long rate_in = clk_get_parent_rate(clk);
430 	struct k210_pll_config config = {};
431 	struct k210_pll *pll = to_k210_pll(clk);
432 	u32 reg;
433 
434 	if (rate_in < 0)
435 		return rate_in;
436 
437 	log_debug("Calculating parameters with rate=%lu and rate_in=%lld\n",
438 		  rate, rate_in);
439 	err = k210_pll_calc_config(rate, rate_in, &config);
440 	if (err)
441 		return err;
442 	log_debug("Got r=%u f=%u od=%u\n", config.r, config.f, config.od);
443 
444 	/*
445 	 * Don't use clk_disable as it might not actually disable the pll due to
446 	 * refcounting
447 	 */
448 	k210_pll_disable(clk);
449 
450 	reg = readl(pll->reg);
451 	reg &= ~K210_PLL_CLKR
452 	    &  ~K210_PLL_CLKF
453 	    &  ~K210_PLL_CLKOD
454 	    &  ~K210_PLL_BWADJ;
455 	reg |= FIELD_PREP(K210_PLL_CLKR, config.r - 1)
456 	    |  FIELD_PREP(K210_PLL_CLKF, config.f - 1)
457 	    |  FIELD_PREP(K210_PLL_CLKOD, config.od - 1)
458 	    |  FIELD_PREP(K210_PLL_BWADJ, config.f - 1);
459 	writel(reg, pll->reg);
460 
461 	err = k210_pll_enable(clk);
462 	if (err)
463 		return err;
464 
465 	serial_setbrg();
466 	return clk_get_rate(clk);
467 }
468 #endif /* CONFIG_CLK_K210_SET_RATE */
469 
k210_pll_get_rate(struct clk * clk)470 static ulong k210_pll_get_rate(struct clk *clk)
471 {
472 	long long rate_in = clk_get_parent_rate(clk);
473 	struct k210_pll *pll = to_k210_pll(clk);
474 	u64 r, f, od;
475 	u32 reg = readl(pll->reg);
476 
477 	if (rate_in < 0 || (reg & K210_PLL_BYPASS))
478 		return rate_in;
479 
480 	if (!(reg & K210_PLL_PWRD))
481 		return 0;
482 
483 	r = FIELD_GET(K210_PLL_CLKR, reg) + 1;
484 	f = FIELD_GET(K210_PLL_CLKF, reg) + 1;
485 	od = FIELD_GET(K210_PLL_CLKOD, reg) + 1;
486 
487 	return DIV_ROUND_DOWN_ULL(((u64)rate_in) * f, r * od);
488 }
489 
490 /*
491  * Wait for the PLL to be locked. If the PLL is not locked, try clearing the
492  * slip before retrying
493  */
k210_pll_waitfor_lock(struct k210_pll * pll)494 static void k210_pll_waitfor_lock(struct k210_pll *pll)
495 {
496 	u32 mask = GENMASK(pll->width - 1, 0) << pll->shift;
497 
498 	while (true) {
499 		u32 reg = readl(pll->lock);
500 
501 		if ((reg & mask) == mask)
502 			break;
503 
504 		reg |= BIT(pll->shift + K210_PLL_CLEAR_SLIP);
505 		writel(reg, pll->lock);
506 	}
507 }
508 
509 /* Adapted from sysctl_pll_enable */
k210_pll_enable(struct clk * clk)510 static int k210_pll_enable(struct clk *clk)
511 {
512 	struct k210_pll *pll = to_k210_pll(clk);
513 	u32 reg = readl(pll->reg);
514 
515 	if ((reg | K210_PLL_PWRD) && !(reg | K210_PLL_RESET))
516 		return 0;
517 
518 	reg |= K210_PLL_PWRD;
519 	writel(reg, pll->reg);
520 
521 	/* Ensure reset is low before asserting it */
522 	reg &= ~K210_PLL_RESET;
523 	writel(reg, pll->reg);
524 	reg |= K210_PLL_RESET;
525 	writel(reg, pll->reg);
526 	nop();
527 	nop();
528 	reg &= ~K210_PLL_RESET;
529 	writel(reg, pll->reg);
530 
531 	k210_pll_waitfor_lock(pll);
532 
533 	reg &= ~K210_PLL_BYPASS;
534 	writel(reg, pll->reg);
535 
536 	return 0;
537 }
538 
k210_pll_disable(struct clk * clk)539 static int k210_pll_disable(struct clk *clk)
540 {
541 	struct k210_pll *pll = to_k210_pll(clk);
542 	u32 reg = readl(pll->reg);
543 
544 	/*
545 	 * Bypassing before powering off is important so child clocks don't stop
546 	 * working. This is especially important for pll0, the indirect parent
547 	 * of the cpu clock.
548 	 */
549 	reg |= K210_PLL_BYPASS;
550 	writel(reg, pll->reg);
551 
552 	reg &= ~K210_PLL_PWRD;
553 	writel(reg, pll->reg);
554 	return 0;
555 }
556 
557 const struct clk_ops k210_pll_ops = {
558 	.get_rate = k210_pll_get_rate,
559 #ifdef CONFIG_CLK_K210_SET_RATE
560 	.set_rate = k210_pll_set_rate,
561 #endif
562 	.enable = k210_pll_enable,
563 	.disable = k210_pll_disable,
564 };
565 
k210_register_pll_struct(const char * name,const char * parent_name,struct k210_pll * pll)566 struct clk *k210_register_pll_struct(const char *name, const char *parent_name,
567 				     struct k210_pll *pll)
568 {
569 	int ret;
570 	struct clk *clk = &pll->clk;
571 
572 	ret = clk_register(clk, CLK_K210_PLL, name, parent_name);
573 	if (ret)
574 		return ERR_PTR(ret);
575 	return clk;
576 }
577 
k210_register_pll(const char * name,const char * parent_name,void __iomem * reg,void __iomem * lock,u8 shift,u8 width)578 struct clk *k210_register_pll(const char *name, const char *parent_name,
579 			      void __iomem *reg, void __iomem *lock, u8 shift,
580 			      u8 width)
581 {
582 	struct clk *clk;
583 	struct k210_pll *pll;
584 
585 	pll = kzalloc(sizeof(*pll), GFP_KERNEL);
586 	if (!pll)
587 		return ERR_PTR(-ENOMEM);
588 	pll->reg = reg;
589 	pll->lock = lock;
590 	pll->shift = shift;
591 	pll->width = width;
592 
593 	clk = k210_register_pll_struct(name, parent_name, pll);
594 	if (IS_ERR(clk))
595 		kfree(pll);
596 	return clk;
597 }
598 
599 U_BOOT_DRIVER(k210_pll) = {
600 	.name	= CLK_K210_PLL,
601 	.id	= UCLASS_CLK,
602 	.ops	= &k210_pll_ops,
603 };
604