xref: /openbmc/linux/arch/arm/mm/cache-l2x0.c (revision 314e47b7)
1 /*
2  * arch/arm/mm/cache-l2x0.c - L210/L220 cache controller support
3  *
4  * Copyright (C) 2007 ARM Limited
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18  */
19 #include <linux/err.h>
20 #include <linux/init.h>
21 #include <linux/spinlock.h>
22 #include <linux/io.h>
23 #include <linux/of.h>
24 #include <linux/of_address.h>
25 
26 #include <asm/cacheflush.h>
27 #include <asm/cputype.h>
28 #include <asm/hardware/cache-l2x0.h>
29 #include "cache-tauros3.h"
30 #include "cache-aurora-l2.h"
31 
32 struct l2c_init_data {
33 	const char *type;
34 	unsigned way_size_0;
35 	unsigned num_lock;
36 	void (*of_parse)(const struct device_node *, u32 *, u32 *);
37 	void (*enable)(void __iomem *, u32, unsigned);
38 	void (*fixup)(void __iomem *, u32, struct outer_cache_fns *);
39 	void (*save)(void __iomem *);
40 	struct outer_cache_fns outer_cache;
41 };
42 
43 #define CACHE_LINE_SIZE		32
44 
45 static void __iomem *l2x0_base;
46 static DEFINE_RAW_SPINLOCK(l2x0_lock);
47 static u32 l2x0_way_mask;	/* Bitmask of active ways */
48 static u32 l2x0_size;
49 static unsigned long sync_reg_offset = L2X0_CACHE_SYNC;
50 
51 struct l2x0_regs l2x0_saved_regs;
52 
53 /*
54  * Common code for all cache controllers.
55  */
56 static inline void l2c_wait_mask(void __iomem *reg, unsigned long mask)
57 {
58 	/* wait for cache operation by line or way to complete */
59 	while (readl_relaxed(reg) & mask)
60 		cpu_relax();
61 }
62 
63 /*
64  * By default, we write directly to secure registers.  Platforms must
65  * override this if they are running non-secure.
66  */
67 static void l2c_write_sec(unsigned long val, void __iomem *base, unsigned reg)
68 {
69 	if (val == readl_relaxed(base + reg))
70 		return;
71 	if (outer_cache.write_sec)
72 		outer_cache.write_sec(val, reg);
73 	else
74 		writel_relaxed(val, base + reg);
75 }
76 
77 /*
78  * This should only be called when we have a requirement that the
79  * register be written due to a work-around, as platforms running
80  * in non-secure mode may not be able to access this register.
81  */
82 static inline void l2c_set_debug(void __iomem *base, unsigned long val)
83 {
84 	l2c_write_sec(val, base, L2X0_DEBUG_CTRL);
85 }
86 
87 static void __l2c_op_way(void __iomem *reg)
88 {
89 	writel_relaxed(l2x0_way_mask, reg);
90 	l2c_wait_mask(reg, l2x0_way_mask);
91 }
92 
93 static inline void l2c_unlock(void __iomem *base, unsigned num)
94 {
95 	unsigned i;
96 
97 	for (i = 0; i < num; i++) {
98 		writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_D_BASE +
99 			       i * L2X0_LOCKDOWN_STRIDE);
100 		writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_I_BASE +
101 			       i * L2X0_LOCKDOWN_STRIDE);
102 	}
103 }
104 
105 /*
106  * Enable the L2 cache controller.  This function must only be
107  * called when the cache controller is known to be disabled.
108  */
109 static void l2c_enable(void __iomem *base, u32 aux, unsigned num_lock)
110 {
111 	unsigned long flags;
112 
113 	l2c_write_sec(aux, base, L2X0_AUX_CTRL);
114 
115 	l2c_unlock(base, num_lock);
116 
117 	local_irq_save(flags);
118 	__l2c_op_way(base + L2X0_INV_WAY);
119 	writel_relaxed(0, base + sync_reg_offset);
120 	l2c_wait_mask(base + sync_reg_offset, 1);
121 	local_irq_restore(flags);
122 
123 	l2c_write_sec(L2X0_CTRL_EN, base, L2X0_CTRL);
124 }
125 
126 static void l2c_disable(void)
127 {
128 	void __iomem *base = l2x0_base;
129 
130 	outer_cache.flush_all();
131 	l2c_write_sec(0, base, L2X0_CTRL);
132 	dsb(st);
133 }
134 
135 #ifdef CONFIG_CACHE_PL310
136 static inline void cache_wait(void __iomem *reg, unsigned long mask)
137 {
138 	/* cache operations by line are atomic on PL310 */
139 }
140 #else
141 #define cache_wait	l2c_wait_mask
142 #endif
143 
144 static inline void cache_sync(void)
145 {
146 	void __iomem *base = l2x0_base;
147 
148 	writel_relaxed(0, base + sync_reg_offset);
149 	cache_wait(base + L2X0_CACHE_SYNC, 1);
150 }
151 
152 #if defined(CONFIG_PL310_ERRATA_588369) || defined(CONFIG_PL310_ERRATA_727915)
153 static inline void debug_writel(unsigned long val)
154 {
155 	l2c_set_debug(l2x0_base, val);
156 }
157 #else
158 /* Optimised out for non-errata case */
159 static inline void debug_writel(unsigned long val)
160 {
161 }
162 #endif
163 
164 static void l2x0_cache_sync(void)
165 {
166 	unsigned long flags;
167 
168 	raw_spin_lock_irqsave(&l2x0_lock, flags);
169 	cache_sync();
170 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
171 }
172 
173 static void __l2x0_flush_all(void)
174 {
175 	debug_writel(0x03);
176 	__l2c_op_way(l2x0_base + L2X0_CLEAN_INV_WAY);
177 	cache_sync();
178 	debug_writel(0x00);
179 }
180 
181 static void l2x0_flush_all(void)
182 {
183 	unsigned long flags;
184 
185 	/* clean all ways */
186 	raw_spin_lock_irqsave(&l2x0_lock, flags);
187 	__l2x0_flush_all();
188 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
189 }
190 
191 static void l2x0_disable(void)
192 {
193 	unsigned long flags;
194 
195 	raw_spin_lock_irqsave(&l2x0_lock, flags);
196 	__l2x0_flush_all();
197 	l2c_write_sec(0, l2x0_base, L2X0_CTRL);
198 	dsb(st);
199 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
200 }
201 
202 static void l2c_save(void __iomem *base)
203 {
204 	l2x0_saved_regs.aux_ctrl = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
205 }
206 
207 /*
208  * L2C-210 specific code.
209  *
210  * The L2C-2x0 PA, set/way and sync operations are atomic, but we must
211  * ensure that no background operation is running.  The way operations
212  * are all background tasks.
213  *
214  * While a background operation is in progress, any new operation is
215  * ignored (unspecified whether this causes an error.)  Thankfully, not
216  * used on SMP.
217  *
218  * Never has a different sync register other than L2X0_CACHE_SYNC, but
219  * we use sync_reg_offset here so we can share some of this with L2C-310.
220  */
221 static void __l2c210_cache_sync(void __iomem *base)
222 {
223 	writel_relaxed(0, base + sync_reg_offset);
224 }
225 
226 static void __l2c210_op_pa_range(void __iomem *reg, unsigned long start,
227 	unsigned long end)
228 {
229 	while (start < end) {
230 		writel_relaxed(start, reg);
231 		start += CACHE_LINE_SIZE;
232 	}
233 }
234 
235 static void l2c210_inv_range(unsigned long start, unsigned long end)
236 {
237 	void __iomem *base = l2x0_base;
238 
239 	if (start & (CACHE_LINE_SIZE - 1)) {
240 		start &= ~(CACHE_LINE_SIZE - 1);
241 		writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
242 		start += CACHE_LINE_SIZE;
243 	}
244 
245 	if (end & (CACHE_LINE_SIZE - 1)) {
246 		end &= ~(CACHE_LINE_SIZE - 1);
247 		writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
248 	}
249 
250 	__l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
251 	__l2c210_cache_sync(base);
252 }
253 
254 static void l2c210_clean_range(unsigned long start, unsigned long end)
255 {
256 	void __iomem *base = l2x0_base;
257 
258 	start &= ~(CACHE_LINE_SIZE - 1);
259 	__l2c210_op_pa_range(base + L2X0_CLEAN_LINE_PA, start, end);
260 	__l2c210_cache_sync(base);
261 }
262 
263 static void l2c210_flush_range(unsigned long start, unsigned long end)
264 {
265 	void __iomem *base = l2x0_base;
266 
267 	start &= ~(CACHE_LINE_SIZE - 1);
268 	__l2c210_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA, start, end);
269 	__l2c210_cache_sync(base);
270 }
271 
272 static void l2c210_flush_all(void)
273 {
274 	void __iomem *base = l2x0_base;
275 
276 	BUG_ON(!irqs_disabled());
277 
278 	__l2c_op_way(base + L2X0_CLEAN_INV_WAY);
279 	__l2c210_cache_sync(base);
280 }
281 
282 static void l2c210_sync(void)
283 {
284 	__l2c210_cache_sync(l2x0_base);
285 }
286 
287 static void l2c210_resume(void)
288 {
289 	void __iomem *base = l2x0_base;
290 
291 	if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN))
292 		l2c_enable(base, l2x0_saved_regs.aux_ctrl, 1);
293 }
294 
295 static const struct l2c_init_data l2c210_data __initconst = {
296 	.type = "L2C-210",
297 	.way_size_0 = SZ_8K,
298 	.num_lock = 1,
299 	.enable = l2c_enable,
300 	.save = l2c_save,
301 	.outer_cache = {
302 		.inv_range = l2c210_inv_range,
303 		.clean_range = l2c210_clean_range,
304 		.flush_range = l2c210_flush_range,
305 		.flush_all = l2c210_flush_all,
306 		.disable = l2c_disable,
307 		.sync = l2c210_sync,
308 		.resume = l2c210_resume,
309 	},
310 };
311 
312 /*
313  * L2C-220 specific code.
314  *
315  * All operations are background operations: they have to be waited for.
316  * Conflicting requests generate a slave error (which will cause an
317  * imprecise abort.)  Never uses sync_reg_offset, so we hard-code the
318  * sync register here.
319  *
320  * However, we can re-use the l2c210_resume call.
321  */
322 static inline void __l2c220_cache_sync(void __iomem *base)
323 {
324 	writel_relaxed(0, base + L2X0_CACHE_SYNC);
325 	l2c_wait_mask(base + L2X0_CACHE_SYNC, 1);
326 }
327 
328 static void l2c220_op_way(void __iomem *base, unsigned reg)
329 {
330 	unsigned long flags;
331 
332 	raw_spin_lock_irqsave(&l2x0_lock, flags);
333 	__l2c_op_way(base + reg);
334 	__l2c220_cache_sync(base);
335 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
336 }
337 
338 static unsigned long l2c220_op_pa_range(void __iomem *reg, unsigned long start,
339 	unsigned long end, unsigned long flags)
340 {
341 	raw_spinlock_t *lock = &l2x0_lock;
342 
343 	while (start < end) {
344 		unsigned long blk_end = start + min(end - start, 4096UL);
345 
346 		while (start < blk_end) {
347 			l2c_wait_mask(reg, 1);
348 			writel_relaxed(start, reg);
349 			start += CACHE_LINE_SIZE;
350 		}
351 
352 		if (blk_end < end) {
353 			raw_spin_unlock_irqrestore(lock, flags);
354 			raw_spin_lock_irqsave(lock, flags);
355 		}
356 	}
357 
358 	return flags;
359 }
360 
361 static void l2c220_inv_range(unsigned long start, unsigned long end)
362 {
363 	void __iomem *base = l2x0_base;
364 	unsigned long flags;
365 
366 	raw_spin_lock_irqsave(&l2x0_lock, flags);
367 	if ((start | end) & (CACHE_LINE_SIZE - 1)) {
368 		if (start & (CACHE_LINE_SIZE - 1)) {
369 			start &= ~(CACHE_LINE_SIZE - 1);
370 			writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
371 			start += CACHE_LINE_SIZE;
372 		}
373 
374 		if (end & (CACHE_LINE_SIZE - 1)) {
375 			end &= ~(CACHE_LINE_SIZE - 1);
376 			l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
377 			writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
378 		}
379 	}
380 
381 	flags = l2c220_op_pa_range(base + L2X0_INV_LINE_PA,
382 				   start, end, flags);
383 	l2c_wait_mask(base + L2X0_INV_LINE_PA, 1);
384 	__l2c220_cache_sync(base);
385 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
386 }
387 
388 static void l2c220_clean_range(unsigned long start, unsigned long end)
389 {
390 	void __iomem *base = l2x0_base;
391 	unsigned long flags;
392 
393 	start &= ~(CACHE_LINE_SIZE - 1);
394 	if ((end - start) >= l2x0_size) {
395 		l2c220_op_way(base, L2X0_CLEAN_WAY);
396 		return;
397 	}
398 
399 	raw_spin_lock_irqsave(&l2x0_lock, flags);
400 	flags = l2c220_op_pa_range(base + L2X0_CLEAN_LINE_PA,
401 				   start, end, flags);
402 	l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
403 	__l2c220_cache_sync(base);
404 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
405 }
406 
407 static void l2c220_flush_range(unsigned long start, unsigned long end)
408 {
409 	void __iomem *base = l2x0_base;
410 	unsigned long flags;
411 
412 	start &= ~(CACHE_LINE_SIZE - 1);
413 	if ((end - start) >= l2x0_size) {
414 		l2c220_op_way(base, L2X0_CLEAN_INV_WAY);
415 		return;
416 	}
417 
418 	raw_spin_lock_irqsave(&l2x0_lock, flags);
419 	flags = l2c220_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA,
420 				   start, end, flags);
421 	l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
422 	__l2c220_cache_sync(base);
423 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
424 }
425 
426 static void l2c220_flush_all(void)
427 {
428 	l2c220_op_way(l2x0_base, L2X0_CLEAN_INV_WAY);
429 }
430 
431 static void l2c220_sync(void)
432 {
433 	unsigned long flags;
434 
435 	raw_spin_lock_irqsave(&l2x0_lock, flags);
436 	__l2c220_cache_sync(l2x0_base);
437 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
438 }
439 
440 static void l2c220_enable(void __iomem *base, u32 aux, unsigned num_lock)
441 {
442 	/*
443 	 * Always enable non-secure access to the lockdown registers -
444 	 * we write to them as part of the L2C enable sequence so they
445 	 * need to be accessible.
446 	 */
447 	aux |= L220_AUX_CTRL_NS_LOCKDOWN;
448 
449 	l2c_enable(base, aux, num_lock);
450 }
451 
452 static const struct l2c_init_data l2c220_data = {
453 	.type = "L2C-220",
454 	.way_size_0 = SZ_8K,
455 	.num_lock = 1,
456 	.enable = l2c220_enable,
457 	.save = l2c_save,
458 	.outer_cache = {
459 		.inv_range = l2c220_inv_range,
460 		.clean_range = l2c220_clean_range,
461 		.flush_range = l2c220_flush_range,
462 		.flush_all = l2c220_flush_all,
463 		.disable = l2c_disable,
464 		.sync = l2c220_sync,
465 		.resume = l2c210_resume,
466 	},
467 };
468 
469 /*
470  * L2C-310 specific code.
471  *
472  * Very similar to L2C-210, the PA, set/way and sync operations are atomic,
473  * and the way operations are all background tasks.  However, issuing an
474  * operation while a background operation is in progress results in a
475  * SLVERR response.  We can reuse:
476  *
477  *  __l2c210_cache_sync (using sync_reg_offset)
478  *  l2c210_sync
479  *  l2c210_inv_range (if 588369 is not applicable)
480  *  l2c210_clean_range
481  *  l2c210_flush_range (if 588369 is not applicable)
482  *  l2c210_flush_all (if 727915 is not applicable)
483  *
484  * Errata:
485  * 588369: PL310 R0P0->R1P0, fixed R2P0.
486  *	Affects: all clean+invalidate operations
487  *	clean and invalidate skips the invalidate step, so we need to issue
488  *	separate operations.  We also require the above debug workaround
489  *	enclosing this code fragment on affected parts.  On unaffected parts,
490  *	we must not use this workaround without the debug register writes
491  *	to avoid exposing a problem similar to 727915.
492  *
493  * 727915: PL310 R2P0->R3P0, fixed R3P1.
494  *	Affects: clean+invalidate by way
495  *	clean and invalidate by way runs in the background, and a store can
496  *	hit the line between the clean operation and invalidate operation,
497  *	resulting in the store being lost.
498  *
499  * 752271: PL310 R3P0->R3P1-50REL0, fixed R3P2.
500  *	Affects: 8x64-bit (double fill) line fetches
501  *	double fill line fetches can fail to cause dirty data to be evicted
502  *	from the cache before the new data overwrites the second line.
503  *
504  * 753970: PL310 R3P0, fixed R3P1.
505  *	Affects: sync
506  *	prevents merging writes after the sync operation, until another L2C
507  *	operation is performed (or a number of other conditions.)
508  *
509  * 769419: PL310 R0P0->R3P1, fixed R3P2.
510  *	Affects: store buffer
511  *	store buffer is not automatically drained.
512  */
513 static void l2c310_inv_range_erratum(unsigned long start, unsigned long end)
514 {
515 	void __iomem *base = l2x0_base;
516 
517 	if ((start | end) & (CACHE_LINE_SIZE - 1)) {
518 		unsigned long flags;
519 
520 		/* Erratum 588369 for both clean+invalidate operations */
521 		raw_spin_lock_irqsave(&l2x0_lock, flags);
522 		l2c_set_debug(base, 0x03);
523 
524 		if (start & (CACHE_LINE_SIZE - 1)) {
525 			start &= ~(CACHE_LINE_SIZE - 1);
526 			writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
527 			writel_relaxed(start, base + L2X0_INV_LINE_PA);
528 			start += CACHE_LINE_SIZE;
529 		}
530 
531 		if (end & (CACHE_LINE_SIZE - 1)) {
532 			end &= ~(CACHE_LINE_SIZE - 1);
533 			writel_relaxed(end, base + L2X0_CLEAN_LINE_PA);
534 			writel_relaxed(end, base + L2X0_INV_LINE_PA);
535 		}
536 
537 		l2c_set_debug(base, 0x00);
538 		raw_spin_unlock_irqrestore(&l2x0_lock, flags);
539 	}
540 
541 	__l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
542 	__l2c210_cache_sync(base);
543 }
544 
545 static void l2c310_flush_range_erratum(unsigned long start, unsigned long end)
546 {
547 	raw_spinlock_t *lock = &l2x0_lock;
548 	unsigned long flags;
549 	void __iomem *base = l2x0_base;
550 
551 	raw_spin_lock_irqsave(lock, flags);
552 	while (start < end) {
553 		unsigned long blk_end = start + min(end - start, 4096UL);
554 
555 		l2c_set_debug(base, 0x03);
556 		while (start < blk_end) {
557 			writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
558 			writel_relaxed(start, base + L2X0_INV_LINE_PA);
559 			start += CACHE_LINE_SIZE;
560 		}
561 		l2c_set_debug(base, 0x00);
562 
563 		if (blk_end < end) {
564 			raw_spin_unlock_irqrestore(lock, flags);
565 			raw_spin_lock_irqsave(lock, flags);
566 		}
567 	}
568 	raw_spin_unlock_irqrestore(lock, flags);
569 	__l2c210_cache_sync(base);
570 }
571 
572 static void l2c310_flush_all_erratum(void)
573 {
574 	void __iomem *base = l2x0_base;
575 	unsigned long flags;
576 
577 	raw_spin_lock_irqsave(&l2x0_lock, flags);
578 	l2c_set_debug(base, 0x03);
579 	__l2c_op_way(base + L2X0_CLEAN_INV_WAY);
580 	l2c_set_debug(base, 0x00);
581 	__l2c210_cache_sync(base);
582 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
583 }
584 
585 static void __init l2c310_save(void __iomem *base)
586 {
587 	unsigned revision;
588 
589 	l2c_save(base);
590 
591 	l2x0_saved_regs.tag_latency = readl_relaxed(base +
592 		L310_TAG_LATENCY_CTRL);
593 	l2x0_saved_regs.data_latency = readl_relaxed(base +
594 		L310_DATA_LATENCY_CTRL);
595 	l2x0_saved_regs.filter_end = readl_relaxed(base +
596 		L310_ADDR_FILTER_END);
597 	l2x0_saved_regs.filter_start = readl_relaxed(base +
598 		L310_ADDR_FILTER_START);
599 
600 	revision = readl_relaxed(base + L2X0_CACHE_ID) &
601 			L2X0_CACHE_ID_RTL_MASK;
602 
603 	/* From r2p0, there is Prefetch offset/control register */
604 	if (revision >= L310_CACHE_ID_RTL_R2P0)
605 		l2x0_saved_regs.prefetch_ctrl = readl_relaxed(base +
606 							L310_PREFETCH_CTRL);
607 
608 	/* From r3p0, there is Power control register */
609 	if (revision >= L310_CACHE_ID_RTL_R3P0)
610 		l2x0_saved_regs.pwr_ctrl = readl_relaxed(base +
611 							L310_POWER_CTRL);
612 }
613 
614 static void l2c310_resume(void)
615 {
616 	void __iomem *base = l2x0_base;
617 
618 	if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN)) {
619 		unsigned revision;
620 
621 		/* restore pl310 setup */
622 		writel_relaxed(l2x0_saved_regs.tag_latency,
623 			       base + L310_TAG_LATENCY_CTRL);
624 		writel_relaxed(l2x0_saved_regs.data_latency,
625 			       base + L310_DATA_LATENCY_CTRL);
626 		writel_relaxed(l2x0_saved_regs.filter_end,
627 			       base + L310_ADDR_FILTER_END);
628 		writel_relaxed(l2x0_saved_regs.filter_start,
629 			       base + L310_ADDR_FILTER_START);
630 
631 		revision = readl_relaxed(base + L2X0_CACHE_ID) &
632 				L2X0_CACHE_ID_RTL_MASK;
633 
634 		if (revision >= L310_CACHE_ID_RTL_R2P0)
635 			l2c_write_sec(l2x0_saved_regs.prefetch_ctrl, base,
636 				      L310_PREFETCH_CTRL);
637 		if (revision >= L310_CACHE_ID_RTL_R3P0)
638 			l2c_write_sec(l2x0_saved_regs.pwr_ctrl, base,
639 				      L310_POWER_CTRL);
640 
641 		l2c_enable(base, l2x0_saved_regs.aux_ctrl, 8);
642 	}
643 }
644 
645 static void __init l2c310_enable(void __iomem *base, u32 aux, unsigned num_lock)
646 {
647 	unsigned rev = readl_relaxed(base + L2X0_CACHE_ID) & L2X0_CACHE_ID_PART_MASK;
648 	bool cortex_a9 = read_cpuid_part_number() == ARM_CPU_PART_CORTEX_A9;
649 
650 	if (rev >= L310_CACHE_ID_RTL_R2P0) {
651 		if (cortex_a9) {
652 			aux |= L310_AUX_CTRL_EARLY_BRESP;
653 			pr_info("L2C-310 enabling early BRESP for Cortex-A9\n");
654 		} else if (aux & L310_AUX_CTRL_EARLY_BRESP) {
655 			pr_warn("L2C-310 early BRESP only supported with Cortex-A9\n");
656 			aux &= ~L310_AUX_CTRL_EARLY_BRESP;
657 		}
658 	}
659 
660 	/* r3p0 or later has power control register */
661 	if (rev >= L310_CACHE_ID_RTL_R3P0) {
662 		u32 power_ctrl;
663 
664 		l2c_write_sec(L310_DYNAMIC_CLK_GATING_EN | L310_STNDBY_MODE_EN,
665 			      base, L310_POWER_CTRL);
666 		power_ctrl = readl_relaxed(base + L310_POWER_CTRL);
667 		pr_info("L2C-310 dynamic clock gating %sabled, standby mode %sabled\n",
668 			power_ctrl & L310_DYNAMIC_CLK_GATING_EN ? "en" : "dis",
669 			power_ctrl & L310_STNDBY_MODE_EN ? "en" : "dis");
670 	}
671 
672 	/*
673 	 * Always enable non-secure access to the lockdown registers -
674 	 * we write to them as part of the L2C enable sequence so they
675 	 * need to be accessible.
676 	 */
677 	aux |= L310_AUX_CTRL_NS_LOCKDOWN;
678 
679 	l2c_enable(base, aux, num_lock);
680 }
681 
682 static void __init l2c310_fixup(void __iomem *base, u32 cache_id,
683 	struct outer_cache_fns *fns)
684 {
685 	unsigned revision = cache_id & L2X0_CACHE_ID_RTL_MASK;
686 	const char *errata[8];
687 	unsigned n = 0;
688 
689 	if (IS_ENABLED(CONFIG_PL310_ERRATA_588369) &&
690 	    revision < L310_CACHE_ID_RTL_R2P0 &&
691 	    /* For bcm compatibility */
692 	    fns->inv_range == l2c210_inv_range) {
693 		fns->inv_range = l2c310_inv_range_erratum;
694 		fns->flush_range = l2c310_flush_range_erratum;
695 		errata[n++] = "588369";
696 	}
697 
698 	if (IS_ENABLED(CONFIG_PL310_ERRATA_727915) &&
699 	    revision >= L310_CACHE_ID_RTL_R2P0 &&
700 	    revision < L310_CACHE_ID_RTL_R3P1) {
701 		fns->flush_all = l2c310_flush_all_erratum;
702 		errata[n++] = "727915";
703 	}
704 
705 	if (revision >= L310_CACHE_ID_RTL_R3P0 &&
706 	    revision < L310_CACHE_ID_RTL_R3P2) {
707 		u32 val = readl_relaxed(base + L310_PREFETCH_CTRL);
708 		/* I don't think bit23 is required here... but iMX6 does so */
709 		if (val & (BIT(30) | BIT(23))) {
710 			val &= ~(BIT(30) | BIT(23));
711 			l2c_write_sec(val, base, L310_PREFETCH_CTRL);
712 			errata[n++] = "752271";
713 		}
714 	}
715 
716 	if (IS_ENABLED(CONFIG_PL310_ERRATA_753970) &&
717 	    revision == L310_CACHE_ID_RTL_R3P0) {
718 		sync_reg_offset = L2X0_DUMMY_REG;
719 		errata[n++] = "753970";
720 	}
721 
722 	if (IS_ENABLED(CONFIG_PL310_ERRATA_769419))
723 		errata[n++] = "769419";
724 
725 	if (n) {
726 		unsigned i;
727 
728 		pr_info("L2C-310 errat%s", n > 1 ? "a" : "um");
729 		for (i = 0; i < n; i++)
730 			pr_cont(" %s", errata[i]);
731 		pr_cont(" enabled\n");
732 	}
733 }
734 
735 static const struct l2c_init_data l2c310_init_fns __initconst = {
736 	.type = "L2C-310",
737 	.way_size_0 = SZ_8K,
738 	.num_lock = 8,
739 	.enable = l2c310_enable,
740 	.fixup = l2c310_fixup,
741 	.save = l2c310_save,
742 	.outer_cache = {
743 		.inv_range = l2c210_inv_range,
744 		.clean_range = l2c210_clean_range,
745 		.flush_range = l2c210_flush_range,
746 		.flush_all = l2c210_flush_all,
747 		.disable = l2c_disable,
748 		.sync = l2c210_sync,
749 		.resume = l2c310_resume,
750 	},
751 };
752 
753 static void __init __l2c_init(const struct l2c_init_data *data,
754 	u32 aux_val, u32 aux_mask, u32 cache_id)
755 {
756 	struct outer_cache_fns fns;
757 	unsigned way_size_bits, ways;
758 	u32 aux;
759 
760 	aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
761 
762 	aux &= aux_mask;
763 	aux |= aux_val;
764 
765 	/* Determine the number of ways */
766 	switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
767 	case L2X0_CACHE_ID_PART_L310:
768 		if ((aux_val | ~aux_mask) & (L2C_AUX_CTRL_WAY_SIZE_MASK | L310_AUX_CTRL_ASSOCIATIVITY_16))
769 			pr_warn("L2C: DT/platform tries to modify or specify cache size\n");
770 		if (aux & (1 << 16))
771 			ways = 16;
772 		else
773 			ways = 8;
774 		break;
775 
776 	case L2X0_CACHE_ID_PART_L210:
777 	case L2X0_CACHE_ID_PART_L220:
778 		ways = (aux >> 13) & 0xf;
779 		break;
780 
781 	case AURORA_CACHE_ID:
782 		ways = (aux >> 13) & 0xf;
783 		ways = 2 << ((ways + 1) >> 2);
784 		break;
785 
786 	default:
787 		/* Assume unknown chips have 8 ways */
788 		ways = 8;
789 		break;
790 	}
791 
792 	l2x0_way_mask = (1 << ways) - 1;
793 
794 	/*
795 	 * way_size_0 is the size that a way_size value of zero would be
796 	 * given the calculation: way_size = way_size_0 << way_size_bits.
797 	 * So, if way_size_bits=0 is reserved, but way_size_bits=1 is 16k,
798 	 * then way_size_0 would be 8k.
799 	 *
800 	 * L2 cache size = number of ways * way size.
801 	 */
802 	way_size_bits = (aux & L2C_AUX_CTRL_WAY_SIZE_MASK) >>
803 			L2C_AUX_CTRL_WAY_SIZE_SHIFT;
804 	l2x0_size = ways * (data->way_size_0 << way_size_bits);
805 
806 	fns = data->outer_cache;
807 	fns.write_sec = outer_cache.write_sec;
808 	if (data->fixup)
809 		data->fixup(l2x0_base, cache_id, &fns);
810 
811 	/*
812 	 * Check if l2x0 controller is already enabled.  If we are booting
813 	 * in non-secure mode accessing the below registers will fault.
814 	 */
815 	if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN))
816 		data->enable(l2x0_base, aux, data->num_lock);
817 
818 	outer_cache = fns;
819 
820 	/*
821 	 * It is strange to save the register state before initialisation,
822 	 * but hey, this is what the DT implementations decided to do.
823 	 */
824 	if (data->save)
825 		data->save(l2x0_base);
826 
827 	/* Re-read it in case some bits are reserved. */
828 	aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
829 
830 	pr_info("%s cache controller enabled, %d ways, %d kB\n",
831 		data->type, ways, l2x0_size >> 10);
832 	pr_info("%s: CACHE_ID 0x%08x, AUX_CTRL 0x%08x\n",
833 		data->type, cache_id, aux);
834 }
835 
836 void __init l2x0_init(void __iomem *base, u32 aux_val, u32 aux_mask)
837 {
838 	const struct l2c_init_data *data;
839 	u32 cache_id;
840 
841 	l2x0_base = base;
842 
843 	cache_id = readl_relaxed(base + L2X0_CACHE_ID);
844 
845 	switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
846 	default:
847 	case L2X0_CACHE_ID_PART_L210:
848 		data = &l2c210_data;
849 		break;
850 
851 	case L2X0_CACHE_ID_PART_L220:
852 		data = &l2c220_data;
853 		break;
854 
855 	case L2X0_CACHE_ID_PART_L310:
856 		data = &l2c310_init_fns;
857 		break;
858 	}
859 
860 	__l2c_init(data, aux_val, aux_mask, cache_id);
861 }
862 
863 #ifdef CONFIG_OF
864 static int l2_wt_override;
865 
866 /* Aurora don't have the cache ID register available, so we have to
867  * pass it though the device tree */
868 static u32 cache_id_part_number_from_dt;
869 
870 static void __init l2x0_of_parse(const struct device_node *np,
871 				 u32 *aux_val, u32 *aux_mask)
872 {
873 	u32 data[2] = { 0, 0 };
874 	u32 tag = 0;
875 	u32 dirty = 0;
876 	u32 val = 0, mask = 0;
877 
878 	of_property_read_u32(np, "arm,tag-latency", &tag);
879 	if (tag) {
880 		mask |= L2X0_AUX_CTRL_TAG_LATENCY_MASK;
881 		val |= (tag - 1) << L2X0_AUX_CTRL_TAG_LATENCY_SHIFT;
882 	}
883 
884 	of_property_read_u32_array(np, "arm,data-latency",
885 				   data, ARRAY_SIZE(data));
886 	if (data[0] && data[1]) {
887 		mask |= L2X0_AUX_CTRL_DATA_RD_LATENCY_MASK |
888 			L2X0_AUX_CTRL_DATA_WR_LATENCY_MASK;
889 		val |= ((data[0] - 1) << L2X0_AUX_CTRL_DATA_RD_LATENCY_SHIFT) |
890 		       ((data[1] - 1) << L2X0_AUX_CTRL_DATA_WR_LATENCY_SHIFT);
891 	}
892 
893 	of_property_read_u32(np, "arm,dirty-latency", &dirty);
894 	if (dirty) {
895 		mask |= L2X0_AUX_CTRL_DIRTY_LATENCY_MASK;
896 		val |= (dirty - 1) << L2X0_AUX_CTRL_DIRTY_LATENCY_SHIFT;
897 	}
898 
899 	*aux_val &= ~mask;
900 	*aux_val |= val;
901 	*aux_mask &= ~mask;
902 }
903 
904 static const struct l2c_init_data of_l2c210_data __initconst = {
905 	.type = "L2C-210",
906 	.way_size_0 = SZ_8K,
907 	.num_lock = 1,
908 	.of_parse = l2x0_of_parse,
909 	.enable = l2c_enable,
910 	.save = l2c_save,
911 	.outer_cache = {
912 		.inv_range   = l2c210_inv_range,
913 		.clean_range = l2c210_clean_range,
914 		.flush_range = l2c210_flush_range,
915 		.flush_all   = l2c210_flush_all,
916 		.disable     = l2c_disable,
917 		.sync        = l2c210_sync,
918 		.resume      = l2c210_resume,
919 	},
920 };
921 
922 static const struct l2c_init_data of_l2c220_data __initconst = {
923 	.type = "L2C-220",
924 	.way_size_0 = SZ_8K,
925 	.num_lock = 1,
926 	.of_parse = l2x0_of_parse,
927 	.enable = l2c220_enable,
928 	.save = l2c_save,
929 	.outer_cache = {
930 		.inv_range   = l2c220_inv_range,
931 		.clean_range = l2c220_clean_range,
932 		.flush_range = l2c220_flush_range,
933 		.flush_all   = l2c220_flush_all,
934 		.disable     = l2c_disable,
935 		.sync        = l2c220_sync,
936 		.resume      = l2c210_resume,
937 	},
938 };
939 
940 static void __init l2c310_of_parse(const struct device_node *np,
941 	u32 *aux_val, u32 *aux_mask)
942 {
943 	u32 data[3] = { 0, 0, 0 };
944 	u32 tag[3] = { 0, 0, 0 };
945 	u32 filter[2] = { 0, 0 };
946 
947 	of_property_read_u32_array(np, "arm,tag-latency", tag, ARRAY_SIZE(tag));
948 	if (tag[0] && tag[1] && tag[2])
949 		writel_relaxed(
950 			L310_LATENCY_CTRL_RD(tag[0] - 1) |
951 			L310_LATENCY_CTRL_WR(tag[1] - 1) |
952 			L310_LATENCY_CTRL_SETUP(tag[2] - 1),
953 			l2x0_base + L310_TAG_LATENCY_CTRL);
954 
955 	of_property_read_u32_array(np, "arm,data-latency",
956 				   data, ARRAY_SIZE(data));
957 	if (data[0] && data[1] && data[2])
958 		writel_relaxed(
959 			L310_LATENCY_CTRL_RD(data[0] - 1) |
960 			L310_LATENCY_CTRL_WR(data[1] - 1) |
961 			L310_LATENCY_CTRL_SETUP(data[2] - 1),
962 			l2x0_base + L310_DATA_LATENCY_CTRL);
963 
964 	of_property_read_u32_array(np, "arm,filter-ranges",
965 				   filter, ARRAY_SIZE(filter));
966 	if (filter[1]) {
967 		writel_relaxed(ALIGN(filter[0] + filter[1], SZ_1M),
968 			       l2x0_base + L310_ADDR_FILTER_END);
969 		writel_relaxed((filter[0] & ~(SZ_1M - 1)) | L310_ADDR_FILTER_EN,
970 			       l2x0_base + L310_ADDR_FILTER_START);
971 	}
972 }
973 
974 static const struct l2c_init_data of_l2c310_data __initconst = {
975 	.type = "L2C-310",
976 	.way_size_0 = SZ_8K,
977 	.num_lock = 8,
978 	.of_parse = l2c310_of_parse,
979 	.enable = l2c310_enable,
980 	.fixup = l2c310_fixup,
981 	.save  = l2c310_save,
982 	.outer_cache = {
983 		.inv_range   = l2c210_inv_range,
984 		.clean_range = l2c210_clean_range,
985 		.flush_range = l2c210_flush_range,
986 		.flush_all   = l2c210_flush_all,
987 		.disable     = l2c_disable,
988 		.sync        = l2c210_sync,
989 		.resume      = l2c310_resume,
990 	},
991 };
992 
993 /*
994  * Note that the end addresses passed to Linux primitives are
995  * noninclusive, while the hardware cache range operations use
996  * inclusive start and end addresses.
997  */
998 static unsigned long calc_range_end(unsigned long start, unsigned long end)
999 {
1000 	/*
1001 	 * Limit the number of cache lines processed at once,
1002 	 * since cache range operations stall the CPU pipeline
1003 	 * until completion.
1004 	 */
1005 	if (end > start + MAX_RANGE_SIZE)
1006 		end = start + MAX_RANGE_SIZE;
1007 
1008 	/*
1009 	 * Cache range operations can't straddle a page boundary.
1010 	 */
1011 	if (end > PAGE_ALIGN(start+1))
1012 		end = PAGE_ALIGN(start+1);
1013 
1014 	return end;
1015 }
1016 
1017 /*
1018  * Make sure 'start' and 'end' reference the same page, as L2 is PIPT
1019  * and range operations only do a TLB lookup on the start address.
1020  */
1021 static void aurora_pa_range(unsigned long start, unsigned long end,
1022 			unsigned long offset)
1023 {
1024 	unsigned long flags;
1025 
1026 	raw_spin_lock_irqsave(&l2x0_lock, flags);
1027 	writel_relaxed(start, l2x0_base + AURORA_RANGE_BASE_ADDR_REG);
1028 	writel_relaxed(end, l2x0_base + offset);
1029 	raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1030 
1031 	cache_sync();
1032 }
1033 
1034 static void aurora_inv_range(unsigned long start, unsigned long end)
1035 {
1036 	/*
1037 	 * round start and end adresses up to cache line size
1038 	 */
1039 	start &= ~(CACHE_LINE_SIZE - 1);
1040 	end = ALIGN(end, CACHE_LINE_SIZE);
1041 
1042 	/*
1043 	 * Invalidate all full cache lines between 'start' and 'end'.
1044 	 */
1045 	while (start < end) {
1046 		unsigned long range_end = calc_range_end(start, end);
1047 		aurora_pa_range(start, range_end - CACHE_LINE_SIZE,
1048 				AURORA_INVAL_RANGE_REG);
1049 		start = range_end;
1050 	}
1051 }
1052 
1053 static void aurora_clean_range(unsigned long start, unsigned long end)
1054 {
1055 	/*
1056 	 * If L2 is forced to WT, the L2 will always be clean and we
1057 	 * don't need to do anything here.
1058 	 */
1059 	if (!l2_wt_override) {
1060 		start &= ~(CACHE_LINE_SIZE - 1);
1061 		end = ALIGN(end, CACHE_LINE_SIZE);
1062 		while (start != end) {
1063 			unsigned long range_end = calc_range_end(start, end);
1064 			aurora_pa_range(start, range_end - CACHE_LINE_SIZE,
1065 					AURORA_CLEAN_RANGE_REG);
1066 			start = range_end;
1067 		}
1068 	}
1069 }
1070 
1071 static void aurora_flush_range(unsigned long start, unsigned long end)
1072 {
1073 	start &= ~(CACHE_LINE_SIZE - 1);
1074 	end = ALIGN(end, CACHE_LINE_SIZE);
1075 	while (start != end) {
1076 		unsigned long range_end = calc_range_end(start, end);
1077 		/*
1078 		 * If L2 is forced to WT, the L2 will always be clean and we
1079 		 * just need to invalidate.
1080 		 */
1081 		if (l2_wt_override)
1082 			aurora_pa_range(start, range_end - CACHE_LINE_SIZE,
1083 							AURORA_INVAL_RANGE_REG);
1084 		else
1085 			aurora_pa_range(start, range_end - CACHE_LINE_SIZE,
1086 							AURORA_FLUSH_RANGE_REG);
1087 		start = range_end;
1088 	}
1089 }
1090 
1091 static void aurora_save(void __iomem *base)
1092 {
1093 	l2x0_saved_regs.ctrl = readl_relaxed(base + L2X0_CTRL);
1094 	l2x0_saved_regs.aux_ctrl = readl_relaxed(base + L2X0_AUX_CTRL);
1095 }
1096 
1097 static void aurora_resume(void)
1098 {
1099 	void __iomem *base = l2x0_base;
1100 
1101 	if (!(readl(base + L2X0_CTRL) & L2X0_CTRL_EN)) {
1102 		writel_relaxed(l2x0_saved_regs.aux_ctrl, base + L2X0_AUX_CTRL);
1103 		writel_relaxed(l2x0_saved_regs.ctrl, base + L2X0_CTRL);
1104 	}
1105 }
1106 
1107 /*
1108  * For Aurora cache in no outer mode, enable via the CP15 coprocessor
1109  * broadcasting of cache commands to L2.
1110  */
1111 static void __init aurora_enable_no_outer(void __iomem *base, u32 aux,
1112 	unsigned num_lock)
1113 {
1114 	u32 u;
1115 
1116 	asm volatile("mrc p15, 1, %0, c15, c2, 0" : "=r" (u));
1117 	u |= AURORA_CTRL_FW;		/* Set the FW bit */
1118 	asm volatile("mcr p15, 1, %0, c15, c2, 0" : : "r" (u));
1119 
1120 	isb();
1121 
1122 	l2c_enable(base, aux, num_lock);
1123 }
1124 
1125 static void __init aurora_fixup(void __iomem *base, u32 cache_id,
1126 	struct outer_cache_fns *fns)
1127 {
1128 	sync_reg_offset = AURORA_SYNC_REG;
1129 }
1130 
1131 static void __init aurora_of_parse(const struct device_node *np,
1132 				u32 *aux_val, u32 *aux_mask)
1133 {
1134 	u32 val = AURORA_ACR_REPLACEMENT_TYPE_SEMIPLRU;
1135 	u32 mask =  AURORA_ACR_REPLACEMENT_MASK;
1136 
1137 	of_property_read_u32(np, "cache-id-part",
1138 			&cache_id_part_number_from_dt);
1139 
1140 	/* Determine and save the write policy */
1141 	l2_wt_override = of_property_read_bool(np, "wt-override");
1142 
1143 	if (l2_wt_override) {
1144 		val |= AURORA_ACR_FORCE_WRITE_THRO_POLICY;
1145 		mask |= AURORA_ACR_FORCE_WRITE_POLICY_MASK;
1146 	}
1147 
1148 	*aux_val &= ~mask;
1149 	*aux_val |= val;
1150 	*aux_mask &= ~mask;
1151 }
1152 
1153 static const struct l2c_init_data of_aurora_with_outer_data __initconst = {
1154 	.type = "Aurora",
1155 	.way_size_0 = SZ_4K,
1156 	.num_lock = 4,
1157 	.of_parse = aurora_of_parse,
1158 	.enable = l2c_enable,
1159 	.fixup = aurora_fixup,
1160 	.save  = aurora_save,
1161 	.outer_cache = {
1162 		.inv_range   = aurora_inv_range,
1163 		.clean_range = aurora_clean_range,
1164 		.flush_range = aurora_flush_range,
1165 		.flush_all   = l2x0_flush_all,
1166 		.disable     = l2x0_disable,
1167 		.sync        = l2x0_cache_sync,
1168 		.resume      = aurora_resume,
1169 	},
1170 };
1171 
1172 static const struct l2c_init_data of_aurora_no_outer_data __initconst = {
1173 	.type = "Aurora",
1174 	.way_size_0 = SZ_4K,
1175 	.num_lock = 4,
1176 	.of_parse = aurora_of_parse,
1177 	.enable = aurora_enable_no_outer,
1178 	.fixup = aurora_fixup,
1179 	.save  = aurora_save,
1180 	.outer_cache = {
1181 		.resume      = aurora_resume,
1182 	},
1183 };
1184 
1185 /*
1186  * For certain Broadcom SoCs, depending on the address range, different offsets
1187  * need to be added to the address before passing it to L2 for
1188  * invalidation/clean/flush
1189  *
1190  * Section Address Range              Offset        EMI
1191  *   1     0x00000000 - 0x3FFFFFFF    0x80000000    VC
1192  *   2     0x40000000 - 0xBFFFFFFF    0x40000000    SYS
1193  *   3     0xC0000000 - 0xFFFFFFFF    0x80000000    VC
1194  *
1195  * When the start and end addresses have crossed two different sections, we
1196  * need to break the L2 operation into two, each within its own section.
1197  * For example, if we need to invalidate addresses starts at 0xBFFF0000 and
1198  * ends at 0xC0001000, we need do invalidate 1) 0xBFFF0000 - 0xBFFFFFFF and 2)
1199  * 0xC0000000 - 0xC0001000
1200  *
1201  * Note 1:
1202  * By breaking a single L2 operation into two, we may potentially suffer some
1203  * performance hit, but keep in mind the cross section case is very rare
1204  *
1205  * Note 2:
1206  * We do not need to handle the case when the start address is in
1207  * Section 1 and the end address is in Section 3, since it is not a valid use
1208  * case
1209  *
1210  * Note 3:
1211  * Section 1 in practical terms can no longer be used on rev A2. Because of
1212  * that the code does not need to handle section 1 at all.
1213  *
1214  */
1215 #define BCM_SYS_EMI_START_ADDR        0x40000000UL
1216 #define BCM_VC_EMI_SEC3_START_ADDR    0xC0000000UL
1217 
1218 #define BCM_SYS_EMI_OFFSET            0x40000000UL
1219 #define BCM_VC_EMI_OFFSET             0x80000000UL
1220 
1221 static inline int bcm_addr_is_sys_emi(unsigned long addr)
1222 {
1223 	return (addr >= BCM_SYS_EMI_START_ADDR) &&
1224 		(addr < BCM_VC_EMI_SEC3_START_ADDR);
1225 }
1226 
1227 static inline unsigned long bcm_l2_phys_addr(unsigned long addr)
1228 {
1229 	if (bcm_addr_is_sys_emi(addr))
1230 		return addr + BCM_SYS_EMI_OFFSET;
1231 	else
1232 		return addr + BCM_VC_EMI_OFFSET;
1233 }
1234 
1235 static void bcm_inv_range(unsigned long start, unsigned long end)
1236 {
1237 	unsigned long new_start, new_end;
1238 
1239 	BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1240 
1241 	if (unlikely(end <= start))
1242 		return;
1243 
1244 	new_start = bcm_l2_phys_addr(start);
1245 	new_end = bcm_l2_phys_addr(end);
1246 
1247 	/* normal case, no cross section between start and end */
1248 	if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1249 		l2c210_inv_range(new_start, new_end);
1250 		return;
1251 	}
1252 
1253 	/* They cross sections, so it can only be a cross from section
1254 	 * 2 to section 3
1255 	 */
1256 	l2c210_inv_range(new_start,
1257 		bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1258 	l2c210_inv_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1259 		new_end);
1260 }
1261 
1262 static void bcm_clean_range(unsigned long start, unsigned long end)
1263 {
1264 	unsigned long new_start, new_end;
1265 
1266 	BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1267 
1268 	if (unlikely(end <= start))
1269 		return;
1270 
1271 	new_start = bcm_l2_phys_addr(start);
1272 	new_end = bcm_l2_phys_addr(end);
1273 
1274 	/* normal case, no cross section between start and end */
1275 	if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1276 		l2c210_clean_range(new_start, new_end);
1277 		return;
1278 	}
1279 
1280 	/* They cross sections, so it can only be a cross from section
1281 	 * 2 to section 3
1282 	 */
1283 	l2c210_clean_range(new_start,
1284 		bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1285 	l2c210_clean_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1286 		new_end);
1287 }
1288 
1289 static void bcm_flush_range(unsigned long start, unsigned long end)
1290 {
1291 	unsigned long new_start, new_end;
1292 
1293 	BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1294 
1295 	if (unlikely(end <= start))
1296 		return;
1297 
1298 	if ((end - start) >= l2x0_size) {
1299 		outer_cache.flush_all();
1300 		return;
1301 	}
1302 
1303 	new_start = bcm_l2_phys_addr(start);
1304 	new_end = bcm_l2_phys_addr(end);
1305 
1306 	/* normal case, no cross section between start and end */
1307 	if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1308 		l2c210_flush_range(new_start, new_end);
1309 		return;
1310 	}
1311 
1312 	/* They cross sections, so it can only be a cross from section
1313 	 * 2 to section 3
1314 	 */
1315 	l2c210_flush_range(new_start,
1316 		bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1317 	l2c210_flush_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1318 		new_end);
1319 }
1320 
1321 /* Broadcom L2C-310 start from ARMs R3P2 or later, and require no fixups */
1322 static const struct l2c_init_data of_bcm_l2x0_data __initconst = {
1323 	.type = "BCM-L2C-310",
1324 	.way_size_0 = SZ_8K,
1325 	.num_lock = 8,
1326 	.of_parse = l2c310_of_parse,
1327 	.enable = l2c310_enable,
1328 	.save  = l2c310_save,
1329 	.outer_cache = {
1330 		.inv_range   = bcm_inv_range,
1331 		.clean_range = bcm_clean_range,
1332 		.flush_range = bcm_flush_range,
1333 		.flush_all   = l2c210_flush_all,
1334 		.disable     = l2c_disable,
1335 		.sync        = l2c210_sync,
1336 		.resume      = l2c310_resume,
1337 	},
1338 };
1339 
1340 static void __init tauros3_save(void __iomem *base)
1341 {
1342 	l2c_save(base);
1343 
1344 	l2x0_saved_regs.aux2_ctrl =
1345 		readl_relaxed(base + TAUROS3_AUX2_CTRL);
1346 	l2x0_saved_regs.prefetch_ctrl =
1347 		readl_relaxed(base + L310_PREFETCH_CTRL);
1348 }
1349 
1350 static void tauros3_resume(void)
1351 {
1352 	void __iomem *base = l2x0_base;
1353 
1354 	if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN)) {
1355 		writel_relaxed(l2x0_saved_regs.aux2_ctrl,
1356 			       base + TAUROS3_AUX2_CTRL);
1357 		writel_relaxed(l2x0_saved_regs.prefetch_ctrl,
1358 			       base + L310_PREFETCH_CTRL);
1359 
1360 		l2c_enable(base, l2x0_saved_regs.aux_ctrl, 8);
1361 	}
1362 }
1363 
1364 static const struct l2c_init_data of_tauros3_data __initconst = {
1365 	.type = "Tauros3",
1366 	.way_size_0 = SZ_8K,
1367 	.num_lock = 8,
1368 	.enable = l2c_enable,
1369 	.save  = tauros3_save,
1370 	/* Tauros3 broadcasts L1 cache operations to L2 */
1371 	.outer_cache = {
1372 		.resume      = tauros3_resume,
1373 	},
1374 };
1375 
1376 #define L2C_ID(name, fns) { .compatible = name, .data = (void *)&fns }
1377 static const struct of_device_id l2x0_ids[] __initconst = {
1378 	L2C_ID("arm,l210-cache", of_l2c210_data),
1379 	L2C_ID("arm,l220-cache", of_l2c220_data),
1380 	L2C_ID("arm,pl310-cache", of_l2c310_data),
1381 	L2C_ID("brcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1382 	L2C_ID("marvell,aurora-outer-cache", of_aurora_with_outer_data),
1383 	L2C_ID("marvell,aurora-system-cache", of_aurora_no_outer_data),
1384 	L2C_ID("marvell,tauros3-cache", of_tauros3_data),
1385 	/* Deprecated IDs */
1386 	L2C_ID("bcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1387 	{}
1388 };
1389 
1390 int __init l2x0_of_init(u32 aux_val, u32 aux_mask)
1391 {
1392 	const struct l2c_init_data *data;
1393 	struct device_node *np;
1394 	struct resource res;
1395 	u32 cache_id;
1396 
1397 	np = of_find_matching_node(NULL, l2x0_ids);
1398 	if (!np)
1399 		return -ENODEV;
1400 
1401 	if (of_address_to_resource(np, 0, &res))
1402 		return -ENODEV;
1403 
1404 	l2x0_base = ioremap(res.start, resource_size(&res));
1405 	if (!l2x0_base)
1406 		return -ENOMEM;
1407 
1408 	l2x0_saved_regs.phy_base = res.start;
1409 
1410 	data = of_match_node(l2x0_ids, np)->data;
1411 
1412 	/* All L2 caches are unified, so this property should be specified */
1413 	if (!of_property_read_bool(np, "cache-unified"))
1414 		pr_err("L2C: device tree omits to specify unified cache\n");
1415 
1416 	/* L2 configuration can only be changed if the cache is disabled */
1417 	if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN))
1418 		if (data->of_parse)
1419 			data->of_parse(np, &aux_val, &aux_mask);
1420 
1421 	if (cache_id_part_number_from_dt)
1422 		cache_id = cache_id_part_number_from_dt;
1423 	else
1424 		cache_id = readl_relaxed(l2x0_base + L2X0_CACHE_ID);
1425 
1426 	__l2c_init(data, aux_val, aux_mask, cache_id);
1427 
1428 	return 0;
1429 }
1430 #endif
1431