Loading...
1/*
2 * arch/arm/mm/cache-l2x0.c - L210/L220/L310 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/cpu.h>
20#include <linux/err.h>
21#include <linux/init.h>
22#include <linux/smp.h>
23#include <linux/spinlock.h>
24#include <linux/log2.h>
25#include <linux/io.h>
26#include <linux/of.h>
27#include <linux/of_address.h>
28
29#include <asm/cacheflush.h>
30#include <asm/cp15.h>
31#include <asm/cputype.h>
32#include <asm/hardware/cache-l2x0.h>
33#include "cache-tauros3.h"
34#include "cache-aurora-l2.h"
35
36struct l2c_init_data {
37 const char *type;
38 unsigned way_size_0;
39 unsigned num_lock;
40 void (*of_parse)(const struct device_node *, u32 *, u32 *);
41 void (*enable)(void __iomem *, unsigned);
42 void (*fixup)(void __iomem *, u32, struct outer_cache_fns *);
43 void (*save)(void __iomem *);
44 void (*configure)(void __iomem *);
45 void (*unlock)(void __iomem *, unsigned);
46 struct outer_cache_fns outer_cache;
47};
48
49#define CACHE_LINE_SIZE 32
50
51static void __iomem *l2x0_base;
52static const struct l2c_init_data *l2x0_data;
53static DEFINE_RAW_SPINLOCK(l2x0_lock);
54static u32 l2x0_way_mask; /* Bitmask of active ways */
55static u32 l2x0_size;
56static unsigned long sync_reg_offset = L2X0_CACHE_SYNC;
57
58struct l2x0_regs l2x0_saved_regs;
59
60static bool l2x0_bresp_disable;
61static bool l2x0_flz_disable;
62
63/*
64 * Common code for all cache controllers.
65 */
66static inline void l2c_wait_mask(void __iomem *reg, unsigned long mask)
67{
68 /* wait for cache operation by line or way to complete */
69 while (readl_relaxed(reg) & mask)
70 cpu_relax();
71}
72
73/*
74 * By default, we write directly to secure registers. Platforms must
75 * override this if they are running non-secure.
76 */
77static void l2c_write_sec(unsigned long val, void __iomem *base, unsigned reg)
78{
79 if (val == readl_relaxed(base + reg))
80 return;
81 if (outer_cache.write_sec)
82 outer_cache.write_sec(val, reg);
83 else
84 writel_relaxed(val, base + reg);
85}
86
87/*
88 * This should only be called when we have a requirement that the
89 * register be written due to a work-around, as platforms running
90 * in non-secure mode may not be able to access this register.
91 */
92static inline void l2c_set_debug(void __iomem *base, unsigned long val)
93{
94 l2c_write_sec(val, base, L2X0_DEBUG_CTRL);
95}
96
97static void __l2c_op_way(void __iomem *reg)
98{
99 writel_relaxed(l2x0_way_mask, reg);
100 l2c_wait_mask(reg, l2x0_way_mask);
101}
102
103static inline void l2c_unlock(void __iomem *base, unsigned num)
104{
105 unsigned i;
106
107 for (i = 0; i < num; i++) {
108 writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_D_BASE +
109 i * L2X0_LOCKDOWN_STRIDE);
110 writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_I_BASE +
111 i * L2X0_LOCKDOWN_STRIDE);
112 }
113}
114
115static void l2c_configure(void __iomem *base)
116{
117 l2c_write_sec(l2x0_saved_regs.aux_ctrl, base, L2X0_AUX_CTRL);
118}
119
120/*
121 * Enable the L2 cache controller. This function must only be
122 * called when the cache controller is known to be disabled.
123 */
124static void l2c_enable(void __iomem *base, unsigned num_lock)
125{
126 unsigned long flags;
127
128 if (outer_cache.configure)
129 outer_cache.configure(&l2x0_saved_regs);
130 else
131 l2x0_data->configure(base);
132
133 l2x0_data->unlock(base, num_lock);
134
135 local_irq_save(flags);
136 __l2c_op_way(base + L2X0_INV_WAY);
137 writel_relaxed(0, base + sync_reg_offset);
138 l2c_wait_mask(base + sync_reg_offset, 1);
139 local_irq_restore(flags);
140
141 l2c_write_sec(L2X0_CTRL_EN, base, L2X0_CTRL);
142}
143
144static void l2c_disable(void)
145{
146 void __iomem *base = l2x0_base;
147
148 l2x0_pmu_suspend();
149
150 outer_cache.flush_all();
151 l2c_write_sec(0, base, L2X0_CTRL);
152 dsb(st);
153}
154
155static void l2c_save(void __iomem *base)
156{
157 l2x0_saved_regs.aux_ctrl = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
158}
159
160static void l2c_resume(void)
161{
162 void __iomem *base = l2x0_base;
163
164 /* Do not touch the controller if already enabled. */
165 if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN))
166 l2c_enable(base, l2x0_data->num_lock);
167
168 l2x0_pmu_resume();
169}
170
171/*
172 * L2C-210 specific code.
173 *
174 * The L2C-2x0 PA, set/way and sync operations are atomic, but we must
175 * ensure that no background operation is running. The way operations
176 * are all background tasks.
177 *
178 * While a background operation is in progress, any new operation is
179 * ignored (unspecified whether this causes an error.) Thankfully, not
180 * used on SMP.
181 *
182 * Never has a different sync register other than L2X0_CACHE_SYNC, but
183 * we use sync_reg_offset here so we can share some of this with L2C-310.
184 */
185static void __l2c210_cache_sync(void __iomem *base)
186{
187 writel_relaxed(0, base + sync_reg_offset);
188}
189
190static void __l2c210_op_pa_range(void __iomem *reg, unsigned long start,
191 unsigned long end)
192{
193 while (start < end) {
194 writel_relaxed(start, reg);
195 start += CACHE_LINE_SIZE;
196 }
197}
198
199static void l2c210_inv_range(unsigned long start, unsigned long end)
200{
201 void __iomem *base = l2x0_base;
202
203 if (start & (CACHE_LINE_SIZE - 1)) {
204 start &= ~(CACHE_LINE_SIZE - 1);
205 writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
206 start += CACHE_LINE_SIZE;
207 }
208
209 if (end & (CACHE_LINE_SIZE - 1)) {
210 end &= ~(CACHE_LINE_SIZE - 1);
211 writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
212 }
213
214 __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
215 __l2c210_cache_sync(base);
216}
217
218static void l2c210_clean_range(unsigned long start, unsigned long end)
219{
220 void __iomem *base = l2x0_base;
221
222 start &= ~(CACHE_LINE_SIZE - 1);
223 __l2c210_op_pa_range(base + L2X0_CLEAN_LINE_PA, start, end);
224 __l2c210_cache_sync(base);
225}
226
227static void l2c210_flush_range(unsigned long start, unsigned long end)
228{
229 void __iomem *base = l2x0_base;
230
231 start &= ~(CACHE_LINE_SIZE - 1);
232 __l2c210_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA, start, end);
233 __l2c210_cache_sync(base);
234}
235
236static void l2c210_flush_all(void)
237{
238 void __iomem *base = l2x0_base;
239
240 BUG_ON(!irqs_disabled());
241
242 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
243 __l2c210_cache_sync(base);
244}
245
246static void l2c210_sync(void)
247{
248 __l2c210_cache_sync(l2x0_base);
249}
250
251static const struct l2c_init_data l2c210_data __initconst = {
252 .type = "L2C-210",
253 .way_size_0 = SZ_8K,
254 .num_lock = 1,
255 .enable = l2c_enable,
256 .save = l2c_save,
257 .configure = l2c_configure,
258 .unlock = l2c_unlock,
259 .outer_cache = {
260 .inv_range = l2c210_inv_range,
261 .clean_range = l2c210_clean_range,
262 .flush_range = l2c210_flush_range,
263 .flush_all = l2c210_flush_all,
264 .disable = l2c_disable,
265 .sync = l2c210_sync,
266 .resume = l2c_resume,
267 },
268};
269
270/*
271 * L2C-220 specific code.
272 *
273 * All operations are background operations: they have to be waited for.
274 * Conflicting requests generate a slave error (which will cause an
275 * imprecise abort.) Never uses sync_reg_offset, so we hard-code the
276 * sync register here.
277 *
278 * However, we can re-use the l2c210_resume call.
279 */
280static inline void __l2c220_cache_sync(void __iomem *base)
281{
282 writel_relaxed(0, base + L2X0_CACHE_SYNC);
283 l2c_wait_mask(base + L2X0_CACHE_SYNC, 1);
284}
285
286static void l2c220_op_way(void __iomem *base, unsigned reg)
287{
288 unsigned long flags;
289
290 raw_spin_lock_irqsave(&l2x0_lock, flags);
291 __l2c_op_way(base + reg);
292 __l2c220_cache_sync(base);
293 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
294}
295
296static unsigned long l2c220_op_pa_range(void __iomem *reg, unsigned long start,
297 unsigned long end, unsigned long flags)
298{
299 raw_spinlock_t *lock = &l2x0_lock;
300
301 while (start < end) {
302 unsigned long blk_end = start + min(end - start, 4096UL);
303
304 while (start < blk_end) {
305 l2c_wait_mask(reg, 1);
306 writel_relaxed(start, reg);
307 start += CACHE_LINE_SIZE;
308 }
309
310 if (blk_end < end) {
311 raw_spin_unlock_irqrestore(lock, flags);
312 raw_spin_lock_irqsave(lock, flags);
313 }
314 }
315
316 return flags;
317}
318
319static void l2c220_inv_range(unsigned long start, unsigned long end)
320{
321 void __iomem *base = l2x0_base;
322 unsigned long flags;
323
324 raw_spin_lock_irqsave(&l2x0_lock, flags);
325 if ((start | end) & (CACHE_LINE_SIZE - 1)) {
326 if (start & (CACHE_LINE_SIZE - 1)) {
327 start &= ~(CACHE_LINE_SIZE - 1);
328 writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
329 start += CACHE_LINE_SIZE;
330 }
331
332 if (end & (CACHE_LINE_SIZE - 1)) {
333 end &= ~(CACHE_LINE_SIZE - 1);
334 l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
335 writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
336 }
337 }
338
339 flags = l2c220_op_pa_range(base + L2X0_INV_LINE_PA,
340 start, end, flags);
341 l2c_wait_mask(base + L2X0_INV_LINE_PA, 1);
342 __l2c220_cache_sync(base);
343 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
344}
345
346static void l2c220_clean_range(unsigned long start, unsigned long end)
347{
348 void __iomem *base = l2x0_base;
349 unsigned long flags;
350
351 start &= ~(CACHE_LINE_SIZE - 1);
352 if ((end - start) >= l2x0_size) {
353 l2c220_op_way(base, L2X0_CLEAN_WAY);
354 return;
355 }
356
357 raw_spin_lock_irqsave(&l2x0_lock, flags);
358 flags = l2c220_op_pa_range(base + L2X0_CLEAN_LINE_PA,
359 start, end, flags);
360 l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
361 __l2c220_cache_sync(base);
362 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
363}
364
365static void l2c220_flush_range(unsigned long start, unsigned long end)
366{
367 void __iomem *base = l2x0_base;
368 unsigned long flags;
369
370 start &= ~(CACHE_LINE_SIZE - 1);
371 if ((end - start) >= l2x0_size) {
372 l2c220_op_way(base, L2X0_CLEAN_INV_WAY);
373 return;
374 }
375
376 raw_spin_lock_irqsave(&l2x0_lock, flags);
377 flags = l2c220_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA,
378 start, end, flags);
379 l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
380 __l2c220_cache_sync(base);
381 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
382}
383
384static void l2c220_flush_all(void)
385{
386 l2c220_op_way(l2x0_base, L2X0_CLEAN_INV_WAY);
387}
388
389static void l2c220_sync(void)
390{
391 unsigned long flags;
392
393 raw_spin_lock_irqsave(&l2x0_lock, flags);
394 __l2c220_cache_sync(l2x0_base);
395 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
396}
397
398static void l2c220_enable(void __iomem *base, unsigned num_lock)
399{
400 /*
401 * Always enable non-secure access to the lockdown registers -
402 * we write to them as part of the L2C enable sequence so they
403 * need to be accessible.
404 */
405 l2x0_saved_regs.aux_ctrl |= L220_AUX_CTRL_NS_LOCKDOWN;
406
407 l2c_enable(base, num_lock);
408}
409
410static void l2c220_unlock(void __iomem *base, unsigned num_lock)
411{
412 if (readl_relaxed(base + L2X0_AUX_CTRL) & L220_AUX_CTRL_NS_LOCKDOWN)
413 l2c_unlock(base, num_lock);
414}
415
416static const struct l2c_init_data l2c220_data = {
417 .type = "L2C-220",
418 .way_size_0 = SZ_8K,
419 .num_lock = 1,
420 .enable = l2c220_enable,
421 .save = l2c_save,
422 .configure = l2c_configure,
423 .unlock = l2c220_unlock,
424 .outer_cache = {
425 .inv_range = l2c220_inv_range,
426 .clean_range = l2c220_clean_range,
427 .flush_range = l2c220_flush_range,
428 .flush_all = l2c220_flush_all,
429 .disable = l2c_disable,
430 .sync = l2c220_sync,
431 .resume = l2c_resume,
432 },
433};
434
435/*
436 * L2C-310 specific code.
437 *
438 * Very similar to L2C-210, the PA, set/way and sync operations are atomic,
439 * and the way operations are all background tasks. However, issuing an
440 * operation while a background operation is in progress results in a
441 * SLVERR response. We can reuse:
442 *
443 * __l2c210_cache_sync (using sync_reg_offset)
444 * l2c210_sync
445 * l2c210_inv_range (if 588369 is not applicable)
446 * l2c210_clean_range
447 * l2c210_flush_range (if 588369 is not applicable)
448 * l2c210_flush_all (if 727915 is not applicable)
449 *
450 * Errata:
451 * 588369: PL310 R0P0->R1P0, fixed R2P0.
452 * Affects: all clean+invalidate operations
453 * clean and invalidate skips the invalidate step, so we need to issue
454 * separate operations. We also require the above debug workaround
455 * enclosing this code fragment on affected parts. On unaffected parts,
456 * we must not use this workaround without the debug register writes
457 * to avoid exposing a problem similar to 727915.
458 *
459 * 727915: PL310 R2P0->R3P0, fixed R3P1.
460 * Affects: clean+invalidate by way
461 * clean and invalidate by way runs in the background, and a store can
462 * hit the line between the clean operation and invalidate operation,
463 * resulting in the store being lost.
464 *
465 * 752271: PL310 R3P0->R3P1-50REL0, fixed R3P2.
466 * Affects: 8x64-bit (double fill) line fetches
467 * double fill line fetches can fail to cause dirty data to be evicted
468 * from the cache before the new data overwrites the second line.
469 *
470 * 753970: PL310 R3P0, fixed R3P1.
471 * Affects: sync
472 * prevents merging writes after the sync operation, until another L2C
473 * operation is performed (or a number of other conditions.)
474 *
475 * 769419: PL310 R0P0->R3P1, fixed R3P2.
476 * Affects: store buffer
477 * store buffer is not automatically drained.
478 */
479static void l2c310_inv_range_erratum(unsigned long start, unsigned long end)
480{
481 void __iomem *base = l2x0_base;
482
483 if ((start | end) & (CACHE_LINE_SIZE - 1)) {
484 unsigned long flags;
485
486 /* Erratum 588369 for both clean+invalidate operations */
487 raw_spin_lock_irqsave(&l2x0_lock, flags);
488 l2c_set_debug(base, 0x03);
489
490 if (start & (CACHE_LINE_SIZE - 1)) {
491 start &= ~(CACHE_LINE_SIZE - 1);
492 writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
493 writel_relaxed(start, base + L2X0_INV_LINE_PA);
494 start += CACHE_LINE_SIZE;
495 }
496
497 if (end & (CACHE_LINE_SIZE - 1)) {
498 end &= ~(CACHE_LINE_SIZE - 1);
499 writel_relaxed(end, base + L2X0_CLEAN_LINE_PA);
500 writel_relaxed(end, base + L2X0_INV_LINE_PA);
501 }
502
503 l2c_set_debug(base, 0x00);
504 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
505 }
506
507 __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
508 __l2c210_cache_sync(base);
509}
510
511static void l2c310_flush_range_erratum(unsigned long start, unsigned long end)
512{
513 raw_spinlock_t *lock = &l2x0_lock;
514 unsigned long flags;
515 void __iomem *base = l2x0_base;
516
517 raw_spin_lock_irqsave(lock, flags);
518 while (start < end) {
519 unsigned long blk_end = start + min(end - start, 4096UL);
520
521 l2c_set_debug(base, 0x03);
522 while (start < blk_end) {
523 writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
524 writel_relaxed(start, base + L2X0_INV_LINE_PA);
525 start += CACHE_LINE_SIZE;
526 }
527 l2c_set_debug(base, 0x00);
528
529 if (blk_end < end) {
530 raw_spin_unlock_irqrestore(lock, flags);
531 raw_spin_lock_irqsave(lock, flags);
532 }
533 }
534 raw_spin_unlock_irqrestore(lock, flags);
535 __l2c210_cache_sync(base);
536}
537
538static void l2c310_flush_all_erratum(void)
539{
540 void __iomem *base = l2x0_base;
541 unsigned long flags;
542
543 raw_spin_lock_irqsave(&l2x0_lock, flags);
544 l2c_set_debug(base, 0x03);
545 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
546 l2c_set_debug(base, 0x00);
547 __l2c210_cache_sync(base);
548 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
549}
550
551static void __init l2c310_save(void __iomem *base)
552{
553 unsigned revision;
554
555 l2c_save(base);
556
557 l2x0_saved_regs.tag_latency = readl_relaxed(base +
558 L310_TAG_LATENCY_CTRL);
559 l2x0_saved_regs.data_latency = readl_relaxed(base +
560 L310_DATA_LATENCY_CTRL);
561 l2x0_saved_regs.filter_end = readl_relaxed(base +
562 L310_ADDR_FILTER_END);
563 l2x0_saved_regs.filter_start = readl_relaxed(base +
564 L310_ADDR_FILTER_START);
565
566 revision = readl_relaxed(base + L2X0_CACHE_ID) &
567 L2X0_CACHE_ID_RTL_MASK;
568
569 /* From r2p0, there is Prefetch offset/control register */
570 if (revision >= L310_CACHE_ID_RTL_R2P0)
571 l2x0_saved_regs.prefetch_ctrl = readl_relaxed(base +
572 L310_PREFETCH_CTRL);
573
574 /* From r3p0, there is Power control register */
575 if (revision >= L310_CACHE_ID_RTL_R3P0)
576 l2x0_saved_regs.pwr_ctrl = readl_relaxed(base +
577 L310_POWER_CTRL);
578}
579
580static void l2c310_configure(void __iomem *base)
581{
582 unsigned revision;
583
584 l2c_configure(base);
585
586 /* restore pl310 setup */
587 l2c_write_sec(l2x0_saved_regs.tag_latency, base,
588 L310_TAG_LATENCY_CTRL);
589 l2c_write_sec(l2x0_saved_regs.data_latency, base,
590 L310_DATA_LATENCY_CTRL);
591 l2c_write_sec(l2x0_saved_regs.filter_end, base,
592 L310_ADDR_FILTER_END);
593 l2c_write_sec(l2x0_saved_regs.filter_start, base,
594 L310_ADDR_FILTER_START);
595
596 revision = readl_relaxed(base + L2X0_CACHE_ID) &
597 L2X0_CACHE_ID_RTL_MASK;
598
599 if (revision >= L310_CACHE_ID_RTL_R2P0)
600 l2c_write_sec(l2x0_saved_regs.prefetch_ctrl, base,
601 L310_PREFETCH_CTRL);
602 if (revision >= L310_CACHE_ID_RTL_R3P0)
603 l2c_write_sec(l2x0_saved_regs.pwr_ctrl, base,
604 L310_POWER_CTRL);
605}
606
607static int l2c310_starting_cpu(unsigned int cpu)
608{
609 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
610 return 0;
611}
612
613static int l2c310_dying_cpu(unsigned int cpu)
614{
615 set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1)));
616 return 0;
617}
618
619static void __init l2c310_enable(void __iomem *base, unsigned num_lock)
620{
621 unsigned rev = readl_relaxed(base + L2X0_CACHE_ID) & L2X0_CACHE_ID_RTL_MASK;
622 bool cortex_a9 = read_cpuid_part() == ARM_CPU_PART_CORTEX_A9;
623 u32 aux = l2x0_saved_regs.aux_ctrl;
624
625 if (rev >= L310_CACHE_ID_RTL_R2P0) {
626 if (cortex_a9 && !l2x0_bresp_disable) {
627 aux |= L310_AUX_CTRL_EARLY_BRESP;
628 pr_info("L2C-310 enabling early BRESP for Cortex-A9\n");
629 } else if (aux & L310_AUX_CTRL_EARLY_BRESP) {
630 pr_warn("L2C-310 early BRESP only supported with Cortex-A9\n");
631 aux &= ~L310_AUX_CTRL_EARLY_BRESP;
632 }
633 }
634
635 if (cortex_a9 && !l2x0_flz_disable) {
636 u32 aux_cur = readl_relaxed(base + L2X0_AUX_CTRL);
637 u32 acr = get_auxcr();
638
639 pr_debug("Cortex-A9 ACR=0x%08x\n", acr);
640
641 if (acr & BIT(3) && !(aux_cur & L310_AUX_CTRL_FULL_LINE_ZERO))
642 pr_err("L2C-310: full line of zeros enabled in Cortex-A9 but not L2C-310 - invalid\n");
643
644 if (aux & L310_AUX_CTRL_FULL_LINE_ZERO && !(acr & BIT(3)))
645 pr_err("L2C-310: enabling full line of zeros but not enabled in Cortex-A9\n");
646
647 if (!(aux & L310_AUX_CTRL_FULL_LINE_ZERO) && !outer_cache.write_sec) {
648 aux |= L310_AUX_CTRL_FULL_LINE_ZERO;
649 pr_info("L2C-310 full line of zeros enabled for Cortex-A9\n");
650 }
651 } else if (aux & (L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP)) {
652 pr_err("L2C-310: disabling Cortex-A9 specific feature bits\n");
653 aux &= ~(L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP);
654 }
655
656 /*
657 * Always enable non-secure access to the lockdown registers -
658 * we write to them as part of the L2C enable sequence so they
659 * need to be accessible.
660 */
661 l2x0_saved_regs.aux_ctrl = aux | L310_AUX_CTRL_NS_LOCKDOWN;
662
663 l2c_enable(base, num_lock);
664
665 /* Read back resulting AUX_CTRL value as it could have been altered. */
666 aux = readl_relaxed(base + L2X0_AUX_CTRL);
667
668 if (aux & (L310_AUX_CTRL_DATA_PREFETCH | L310_AUX_CTRL_INSTR_PREFETCH)) {
669 u32 prefetch = readl_relaxed(base + L310_PREFETCH_CTRL);
670
671 pr_info("L2C-310 %s%s prefetch enabled, offset %u lines\n",
672 aux & L310_AUX_CTRL_INSTR_PREFETCH ? "I" : "",
673 aux & L310_AUX_CTRL_DATA_PREFETCH ? "D" : "",
674 1 + (prefetch & L310_PREFETCH_CTRL_OFFSET_MASK));
675 }
676
677 /* r3p0 or later has power control register */
678 if (rev >= L310_CACHE_ID_RTL_R3P0) {
679 u32 power_ctrl;
680
681 power_ctrl = readl_relaxed(base + L310_POWER_CTRL);
682 pr_info("L2C-310 dynamic clock gating %sabled, standby mode %sabled\n",
683 power_ctrl & L310_DYNAMIC_CLK_GATING_EN ? "en" : "dis",
684 power_ctrl & L310_STNDBY_MODE_EN ? "en" : "dis");
685 }
686
687 if (aux & L310_AUX_CTRL_FULL_LINE_ZERO)
688 cpuhp_setup_state(CPUHP_AP_ARM_L2X0_STARTING,
689 "arm/l2x0:starting", l2c310_starting_cpu,
690 l2c310_dying_cpu);
691}
692
693static void __init l2c310_fixup(void __iomem *base, u32 cache_id,
694 struct outer_cache_fns *fns)
695{
696 unsigned revision = cache_id & L2X0_CACHE_ID_RTL_MASK;
697 const char *errata[8];
698 unsigned n = 0;
699
700 if (IS_ENABLED(CONFIG_PL310_ERRATA_588369) &&
701 revision < L310_CACHE_ID_RTL_R2P0 &&
702 /* For bcm compatibility */
703 fns->inv_range == l2c210_inv_range) {
704 fns->inv_range = l2c310_inv_range_erratum;
705 fns->flush_range = l2c310_flush_range_erratum;
706 errata[n++] = "588369";
707 }
708
709 if (IS_ENABLED(CONFIG_PL310_ERRATA_727915) &&
710 revision >= L310_CACHE_ID_RTL_R2P0 &&
711 revision < L310_CACHE_ID_RTL_R3P1) {
712 fns->flush_all = l2c310_flush_all_erratum;
713 errata[n++] = "727915";
714 }
715
716 if (revision >= L310_CACHE_ID_RTL_R3P0 &&
717 revision < L310_CACHE_ID_RTL_R3P2) {
718 u32 val = l2x0_saved_regs.prefetch_ctrl;
719 if (val & L310_PREFETCH_CTRL_DBL_LINEFILL) {
720 val &= ~L310_PREFETCH_CTRL_DBL_LINEFILL;
721 l2x0_saved_regs.prefetch_ctrl = val;
722 errata[n++] = "752271";
723 }
724 }
725
726 if (IS_ENABLED(CONFIG_PL310_ERRATA_753970) &&
727 revision == L310_CACHE_ID_RTL_R3P0) {
728 sync_reg_offset = L2X0_DUMMY_REG;
729 errata[n++] = "753970";
730 }
731
732 if (IS_ENABLED(CONFIG_PL310_ERRATA_769419))
733 errata[n++] = "769419";
734
735 if (n) {
736 unsigned i;
737
738 pr_info("L2C-310 errat%s", n > 1 ? "a" : "um");
739 for (i = 0; i < n; i++)
740 pr_cont(" %s", errata[i]);
741 pr_cont(" enabled\n");
742 }
743}
744
745static void l2c310_disable(void)
746{
747 /*
748 * If full-line-of-zeros is enabled, we must first disable it in the
749 * Cortex-A9 auxiliary control register before disabling the L2 cache.
750 */
751 if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
752 set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1)));
753
754 l2c_disable();
755}
756
757static void l2c310_resume(void)
758{
759 l2c_resume();
760
761 /* Re-enable full-line-of-zeros for Cortex-A9 */
762 if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
763 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
764}
765
766static void l2c310_unlock(void __iomem *base, unsigned num_lock)
767{
768 if (readl_relaxed(base + L2X0_AUX_CTRL) & L310_AUX_CTRL_NS_LOCKDOWN)
769 l2c_unlock(base, num_lock);
770}
771
772static const struct l2c_init_data l2c310_init_fns __initconst = {
773 .type = "L2C-310",
774 .way_size_0 = SZ_8K,
775 .num_lock = 8,
776 .enable = l2c310_enable,
777 .fixup = l2c310_fixup,
778 .save = l2c310_save,
779 .configure = l2c310_configure,
780 .unlock = l2c310_unlock,
781 .outer_cache = {
782 .inv_range = l2c210_inv_range,
783 .clean_range = l2c210_clean_range,
784 .flush_range = l2c210_flush_range,
785 .flush_all = l2c210_flush_all,
786 .disable = l2c310_disable,
787 .sync = l2c210_sync,
788 .resume = l2c310_resume,
789 },
790};
791
792static int __init __l2c_init(const struct l2c_init_data *data,
793 u32 aux_val, u32 aux_mask, u32 cache_id, bool nosync)
794{
795 struct outer_cache_fns fns;
796 unsigned way_size_bits, ways;
797 u32 aux, old_aux;
798
799 /*
800 * Save the pointer globally so that callbacks which do not receive
801 * context from callers can access the structure.
802 */
803 l2x0_data = kmemdup(data, sizeof(*data), GFP_KERNEL);
804 if (!l2x0_data)
805 return -ENOMEM;
806
807 /*
808 * Sanity check the aux values. aux_mask is the bits we preserve
809 * from reading the hardware register, and aux_val is the bits we
810 * set.
811 */
812 if (aux_val & aux_mask)
813 pr_alert("L2C: platform provided aux values permit register corruption.\n");
814
815 old_aux = aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
816 aux &= aux_mask;
817 aux |= aux_val;
818
819 if (old_aux != aux)
820 pr_warn("L2C: DT/platform modifies aux control register: 0x%08x -> 0x%08x\n",
821 old_aux, aux);
822
823 /* Determine the number of ways */
824 switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
825 case L2X0_CACHE_ID_PART_L310:
826 if ((aux_val | ~aux_mask) & (L2C_AUX_CTRL_WAY_SIZE_MASK | L310_AUX_CTRL_ASSOCIATIVITY_16))
827 pr_warn("L2C: DT/platform tries to modify or specify cache size\n");
828 if (aux & (1 << 16))
829 ways = 16;
830 else
831 ways = 8;
832 break;
833
834 case L2X0_CACHE_ID_PART_L210:
835 case L2X0_CACHE_ID_PART_L220:
836 ways = (aux >> 13) & 0xf;
837 break;
838
839 case AURORA_CACHE_ID:
840 ways = (aux >> 13) & 0xf;
841 ways = 2 << ((ways + 1) >> 2);
842 break;
843
844 default:
845 /* Assume unknown chips have 8 ways */
846 ways = 8;
847 break;
848 }
849
850 l2x0_way_mask = (1 << ways) - 1;
851
852 /*
853 * way_size_0 is the size that a way_size value of zero would be
854 * given the calculation: way_size = way_size_0 << way_size_bits.
855 * So, if way_size_bits=0 is reserved, but way_size_bits=1 is 16k,
856 * then way_size_0 would be 8k.
857 *
858 * L2 cache size = number of ways * way size.
859 */
860 way_size_bits = (aux & L2C_AUX_CTRL_WAY_SIZE_MASK) >>
861 L2C_AUX_CTRL_WAY_SIZE_SHIFT;
862 l2x0_size = ways * (data->way_size_0 << way_size_bits);
863
864 fns = data->outer_cache;
865 fns.write_sec = outer_cache.write_sec;
866 fns.configure = outer_cache.configure;
867 if (data->fixup)
868 data->fixup(l2x0_base, cache_id, &fns);
869 if (nosync) {
870 pr_info("L2C: disabling outer sync\n");
871 fns.sync = NULL;
872 }
873
874 /*
875 * Check if l2x0 controller is already enabled. If we are booting
876 * in non-secure mode accessing the below registers will fault.
877 */
878 if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN)) {
879 l2x0_saved_regs.aux_ctrl = aux;
880
881 data->enable(l2x0_base, data->num_lock);
882 }
883
884 outer_cache = fns;
885
886 /*
887 * It is strange to save the register state before initialisation,
888 * but hey, this is what the DT implementations decided to do.
889 */
890 if (data->save)
891 data->save(l2x0_base);
892
893 /* Re-read it in case some bits are reserved. */
894 aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
895
896 pr_info("%s cache controller enabled, %d ways, %d kB\n",
897 data->type, ways, l2x0_size >> 10);
898 pr_info("%s: CACHE_ID 0x%08x, AUX_CTRL 0x%08x\n",
899 data->type, cache_id, aux);
900
901 l2x0_pmu_register(l2x0_base, cache_id);
902
903 return 0;
904}
905
906void __init l2x0_init(void __iomem *base, u32 aux_val, u32 aux_mask)
907{
908 const struct l2c_init_data *data;
909 u32 cache_id;
910
911 l2x0_base = base;
912
913 cache_id = readl_relaxed(base + L2X0_CACHE_ID);
914
915 switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
916 default:
917 case L2X0_CACHE_ID_PART_L210:
918 data = &l2c210_data;
919 break;
920
921 case L2X0_CACHE_ID_PART_L220:
922 data = &l2c220_data;
923 break;
924
925 case L2X0_CACHE_ID_PART_L310:
926 data = &l2c310_init_fns;
927 break;
928 }
929
930 /* Read back current (default) hardware configuration */
931 if (data->save)
932 data->save(l2x0_base);
933
934 __l2c_init(data, aux_val, aux_mask, cache_id, false);
935}
936
937#ifdef CONFIG_OF
938static int l2_wt_override;
939
940/* Aurora don't have the cache ID register available, so we have to
941 * pass it though the device tree */
942static u32 cache_id_part_number_from_dt;
943
944/**
945 * l2x0_cache_size_of_parse() - read cache size parameters from DT
946 * @np: the device tree node for the l2 cache
947 * @aux_val: pointer to machine-supplied auxilary register value, to
948 * be augmented by the call (bits to be set to 1)
949 * @aux_mask: pointer to machine-supplied auxilary register mask, to
950 * be augmented by the call (bits to be set to 0)
951 * @associativity: variable to return the calculated associativity in
952 * @max_way_size: the maximum size in bytes for the cache ways
953 */
954static int __init l2x0_cache_size_of_parse(const struct device_node *np,
955 u32 *aux_val, u32 *aux_mask,
956 u32 *associativity,
957 u32 max_way_size)
958{
959 u32 mask = 0, val = 0;
960 u32 cache_size = 0, sets = 0;
961 u32 way_size_bits = 1;
962 u32 way_size = 0;
963 u32 block_size = 0;
964 u32 line_size = 0;
965
966 of_property_read_u32(np, "cache-size", &cache_size);
967 of_property_read_u32(np, "cache-sets", &sets);
968 of_property_read_u32(np, "cache-block-size", &block_size);
969 of_property_read_u32(np, "cache-line-size", &line_size);
970
971 if (!cache_size || !sets)
972 return -ENODEV;
973
974 /* All these l2 caches have the same line = block size actually */
975 if (!line_size) {
976 if (block_size) {
977 /* If linesize is not given, it is equal to blocksize */
978 line_size = block_size;
979 } else {
980 /* Fall back to known size */
981 pr_warn("L2C OF: no cache block/line size given: "
982 "falling back to default size %d bytes\n",
983 CACHE_LINE_SIZE);
984 line_size = CACHE_LINE_SIZE;
985 }
986 }
987
988 if (line_size != CACHE_LINE_SIZE)
989 pr_warn("L2C OF: DT supplied line size %d bytes does "
990 "not match hardware line size of %d bytes\n",
991 line_size,
992 CACHE_LINE_SIZE);
993
994 /*
995 * Since:
996 * set size = cache size / sets
997 * ways = cache size / (sets * line size)
998 * way size = cache size / (cache size / (sets * line size))
999 * way size = sets * line size
1000 * associativity = ways = cache size / way size
1001 */
1002 way_size = sets * line_size;
1003 *associativity = cache_size / way_size;
1004
1005 if (way_size > max_way_size) {
1006 pr_err("L2C OF: set size %dKB is too large\n", way_size);
1007 return -EINVAL;
1008 }
1009
1010 pr_info("L2C OF: override cache size: %d bytes (%dKB)\n",
1011 cache_size, cache_size >> 10);
1012 pr_info("L2C OF: override line size: %d bytes\n", line_size);
1013 pr_info("L2C OF: override way size: %d bytes (%dKB)\n",
1014 way_size, way_size >> 10);
1015 pr_info("L2C OF: override associativity: %d\n", *associativity);
1016
1017 /*
1018 * Calculates the bits 17:19 to set for way size:
1019 * 512KB -> 6, 256KB -> 5, ... 16KB -> 1
1020 */
1021 way_size_bits = ilog2(way_size >> 10) - 3;
1022 if (way_size_bits < 1 || way_size_bits > 6) {
1023 pr_err("L2C OF: cache way size illegal: %dKB is not mapped\n",
1024 way_size);
1025 return -EINVAL;
1026 }
1027
1028 mask |= L2C_AUX_CTRL_WAY_SIZE_MASK;
1029 val |= (way_size_bits << L2C_AUX_CTRL_WAY_SIZE_SHIFT);
1030
1031 *aux_val &= ~mask;
1032 *aux_val |= val;
1033 *aux_mask &= ~mask;
1034
1035 return 0;
1036}
1037
1038static void __init l2x0_of_parse(const struct device_node *np,
1039 u32 *aux_val, u32 *aux_mask)
1040{
1041 u32 data[2] = { 0, 0 };
1042 u32 tag = 0;
1043 u32 dirty = 0;
1044 u32 val = 0, mask = 0;
1045 u32 assoc;
1046 int ret;
1047
1048 of_property_read_u32(np, "arm,tag-latency", &tag);
1049 if (tag) {
1050 mask |= L2X0_AUX_CTRL_TAG_LATENCY_MASK;
1051 val |= (tag - 1) << L2X0_AUX_CTRL_TAG_LATENCY_SHIFT;
1052 }
1053
1054 of_property_read_u32_array(np, "arm,data-latency",
1055 data, ARRAY_SIZE(data));
1056 if (data[0] && data[1]) {
1057 mask |= L2X0_AUX_CTRL_DATA_RD_LATENCY_MASK |
1058 L2X0_AUX_CTRL_DATA_WR_LATENCY_MASK;
1059 val |= ((data[0] - 1) << L2X0_AUX_CTRL_DATA_RD_LATENCY_SHIFT) |
1060 ((data[1] - 1) << L2X0_AUX_CTRL_DATA_WR_LATENCY_SHIFT);
1061 }
1062
1063 of_property_read_u32(np, "arm,dirty-latency", &dirty);
1064 if (dirty) {
1065 mask |= L2X0_AUX_CTRL_DIRTY_LATENCY_MASK;
1066 val |= (dirty - 1) << L2X0_AUX_CTRL_DIRTY_LATENCY_SHIFT;
1067 }
1068
1069 if (of_property_read_bool(np, "arm,parity-enable")) {
1070 mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1071 val |= L2C_AUX_CTRL_PARITY_ENABLE;
1072 } else if (of_property_read_bool(np, "arm,parity-disable")) {
1073 mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1074 }
1075
1076 if (of_property_read_bool(np, "arm,shared-override")) {
1077 mask &= ~L2C_AUX_CTRL_SHARED_OVERRIDE;
1078 val |= L2C_AUX_CTRL_SHARED_OVERRIDE;
1079 }
1080
1081 ret = l2x0_cache_size_of_parse(np, aux_val, aux_mask, &assoc, SZ_256K);
1082 if (ret)
1083 return;
1084
1085 if (assoc > 8) {
1086 pr_err("l2x0 of: cache setting yield too high associativity\n");
1087 pr_err("l2x0 of: %d calculated, max 8\n", assoc);
1088 } else {
1089 mask |= L2X0_AUX_CTRL_ASSOC_MASK;
1090 val |= (assoc << L2X0_AUX_CTRL_ASSOC_SHIFT);
1091 }
1092
1093 *aux_val &= ~mask;
1094 *aux_val |= val;
1095 *aux_mask &= ~mask;
1096}
1097
1098static const struct l2c_init_data of_l2c210_data __initconst = {
1099 .type = "L2C-210",
1100 .way_size_0 = SZ_8K,
1101 .num_lock = 1,
1102 .of_parse = l2x0_of_parse,
1103 .enable = l2c_enable,
1104 .save = l2c_save,
1105 .configure = l2c_configure,
1106 .unlock = l2c_unlock,
1107 .outer_cache = {
1108 .inv_range = l2c210_inv_range,
1109 .clean_range = l2c210_clean_range,
1110 .flush_range = l2c210_flush_range,
1111 .flush_all = l2c210_flush_all,
1112 .disable = l2c_disable,
1113 .sync = l2c210_sync,
1114 .resume = l2c_resume,
1115 },
1116};
1117
1118static const struct l2c_init_data of_l2c220_data __initconst = {
1119 .type = "L2C-220",
1120 .way_size_0 = SZ_8K,
1121 .num_lock = 1,
1122 .of_parse = l2x0_of_parse,
1123 .enable = l2c220_enable,
1124 .save = l2c_save,
1125 .configure = l2c_configure,
1126 .unlock = l2c220_unlock,
1127 .outer_cache = {
1128 .inv_range = l2c220_inv_range,
1129 .clean_range = l2c220_clean_range,
1130 .flush_range = l2c220_flush_range,
1131 .flush_all = l2c220_flush_all,
1132 .disable = l2c_disable,
1133 .sync = l2c220_sync,
1134 .resume = l2c_resume,
1135 },
1136};
1137
1138static void __init l2c310_of_parse(const struct device_node *np,
1139 u32 *aux_val, u32 *aux_mask)
1140{
1141 u32 data[3] = { 0, 0, 0 };
1142 u32 tag[3] = { 0, 0, 0 };
1143 u32 filter[2] = { 0, 0 };
1144 u32 assoc;
1145 u32 prefetch;
1146 u32 power;
1147 u32 val;
1148 int ret;
1149
1150 of_property_read_u32_array(np, "arm,tag-latency", tag, ARRAY_SIZE(tag));
1151 if (tag[0] && tag[1] && tag[2])
1152 l2x0_saved_regs.tag_latency =
1153 L310_LATENCY_CTRL_RD(tag[0] - 1) |
1154 L310_LATENCY_CTRL_WR(tag[1] - 1) |
1155 L310_LATENCY_CTRL_SETUP(tag[2] - 1);
1156
1157 of_property_read_u32_array(np, "arm,data-latency",
1158 data, ARRAY_SIZE(data));
1159 if (data[0] && data[1] && data[2])
1160 l2x0_saved_regs.data_latency =
1161 L310_LATENCY_CTRL_RD(data[0] - 1) |
1162 L310_LATENCY_CTRL_WR(data[1] - 1) |
1163 L310_LATENCY_CTRL_SETUP(data[2] - 1);
1164
1165 of_property_read_u32_array(np, "arm,filter-ranges",
1166 filter, ARRAY_SIZE(filter));
1167 if (filter[1]) {
1168 l2x0_saved_regs.filter_end =
1169 ALIGN(filter[0] + filter[1], SZ_1M);
1170 l2x0_saved_regs.filter_start = (filter[0] & ~(SZ_1M - 1))
1171 | L310_ADDR_FILTER_EN;
1172 }
1173
1174 ret = l2x0_cache_size_of_parse(np, aux_val, aux_mask, &assoc, SZ_512K);
1175 if (!ret) {
1176 switch (assoc) {
1177 case 16:
1178 *aux_val &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1179 *aux_val |= L310_AUX_CTRL_ASSOCIATIVITY_16;
1180 *aux_mask &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1181 break;
1182 case 8:
1183 *aux_val &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1184 *aux_mask &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1185 break;
1186 default:
1187 pr_err("L2C-310 OF cache associativity %d invalid, only 8 or 16 permitted\n",
1188 assoc);
1189 break;
1190 }
1191 }
1192
1193 if (of_property_read_bool(np, "arm,shared-override")) {
1194 *aux_val |= L2C_AUX_CTRL_SHARED_OVERRIDE;
1195 *aux_mask &= ~L2C_AUX_CTRL_SHARED_OVERRIDE;
1196 }
1197
1198 if (of_property_read_bool(np, "arm,parity-enable")) {
1199 *aux_val |= L2C_AUX_CTRL_PARITY_ENABLE;
1200 *aux_mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1201 } else if (of_property_read_bool(np, "arm,parity-disable")) {
1202 *aux_val &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1203 *aux_mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1204 }
1205
1206 if (of_property_read_bool(np, "arm,early-bresp-disable"))
1207 l2x0_bresp_disable = true;
1208
1209 if (of_property_read_bool(np, "arm,full-line-zero-disable"))
1210 l2x0_flz_disable = true;
1211
1212 prefetch = l2x0_saved_regs.prefetch_ctrl;
1213
1214 ret = of_property_read_u32(np, "arm,double-linefill", &val);
1215 if (ret == 0) {
1216 if (val)
1217 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL;
1218 else
1219 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL;
1220 } else if (ret != -EINVAL) {
1221 pr_err("L2C-310 OF arm,double-linefill property value is missing\n");
1222 }
1223
1224 ret = of_property_read_u32(np, "arm,double-linefill-incr", &val);
1225 if (ret == 0) {
1226 if (val)
1227 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL_INCR;
1228 else
1229 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL_INCR;
1230 } else if (ret != -EINVAL) {
1231 pr_err("L2C-310 OF arm,double-linefill-incr property value is missing\n");
1232 }
1233
1234 ret = of_property_read_u32(np, "arm,double-linefill-wrap", &val);
1235 if (ret == 0) {
1236 if (!val)
1237 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL_WRAP;
1238 else
1239 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL_WRAP;
1240 } else if (ret != -EINVAL) {
1241 pr_err("L2C-310 OF arm,double-linefill-wrap property value is missing\n");
1242 }
1243
1244 ret = of_property_read_u32(np, "arm,prefetch-drop", &val);
1245 if (ret == 0) {
1246 if (val)
1247 prefetch |= L310_PREFETCH_CTRL_PREFETCH_DROP;
1248 else
1249 prefetch &= ~L310_PREFETCH_CTRL_PREFETCH_DROP;
1250 } else if (ret != -EINVAL) {
1251 pr_err("L2C-310 OF arm,prefetch-drop property value is missing\n");
1252 }
1253
1254 ret = of_property_read_u32(np, "arm,prefetch-offset", &val);
1255 if (ret == 0) {
1256 prefetch &= ~L310_PREFETCH_CTRL_OFFSET_MASK;
1257 prefetch |= val & L310_PREFETCH_CTRL_OFFSET_MASK;
1258 } else if (ret != -EINVAL) {
1259 pr_err("L2C-310 OF arm,prefetch-offset property value is missing\n");
1260 }
1261
1262 ret = of_property_read_u32(np, "prefetch-data", &val);
1263 if (ret == 0) {
1264 if (val)
1265 prefetch |= L310_PREFETCH_CTRL_DATA_PREFETCH;
1266 else
1267 prefetch &= ~L310_PREFETCH_CTRL_DATA_PREFETCH;
1268 } else if (ret != -EINVAL) {
1269 pr_err("L2C-310 OF prefetch-data property value is missing\n");
1270 }
1271
1272 ret = of_property_read_u32(np, "prefetch-instr", &val);
1273 if (ret == 0) {
1274 if (val)
1275 prefetch |= L310_PREFETCH_CTRL_INSTR_PREFETCH;
1276 else
1277 prefetch &= ~L310_PREFETCH_CTRL_INSTR_PREFETCH;
1278 } else if (ret != -EINVAL) {
1279 pr_err("L2C-310 OF prefetch-instr property value is missing\n");
1280 }
1281
1282 l2x0_saved_regs.prefetch_ctrl = prefetch;
1283
1284 power = l2x0_saved_regs.pwr_ctrl |
1285 L310_DYNAMIC_CLK_GATING_EN | L310_STNDBY_MODE_EN;
1286
1287 ret = of_property_read_u32(np, "arm,dynamic-clock-gating", &val);
1288 if (!ret) {
1289 if (!val)
1290 power &= ~L310_DYNAMIC_CLK_GATING_EN;
1291 } else if (ret != -EINVAL) {
1292 pr_err("L2C-310 OF dynamic-clock-gating property value is missing or invalid\n");
1293 }
1294 ret = of_property_read_u32(np, "arm,standby-mode", &val);
1295 if (!ret) {
1296 if (!val)
1297 power &= ~L310_STNDBY_MODE_EN;
1298 } else if (ret != -EINVAL) {
1299 pr_err("L2C-310 OF standby-mode property value is missing or invalid\n");
1300 }
1301
1302 l2x0_saved_regs.pwr_ctrl = power;
1303}
1304
1305static const struct l2c_init_data of_l2c310_data __initconst = {
1306 .type = "L2C-310",
1307 .way_size_0 = SZ_8K,
1308 .num_lock = 8,
1309 .of_parse = l2c310_of_parse,
1310 .enable = l2c310_enable,
1311 .fixup = l2c310_fixup,
1312 .save = l2c310_save,
1313 .configure = l2c310_configure,
1314 .unlock = l2c310_unlock,
1315 .outer_cache = {
1316 .inv_range = l2c210_inv_range,
1317 .clean_range = l2c210_clean_range,
1318 .flush_range = l2c210_flush_range,
1319 .flush_all = l2c210_flush_all,
1320 .disable = l2c310_disable,
1321 .sync = l2c210_sync,
1322 .resume = l2c310_resume,
1323 },
1324};
1325
1326/*
1327 * This is a variant of the of_l2c310_data with .sync set to
1328 * NULL. Outer sync operations are not needed when the system is I/O
1329 * coherent, and potentially harmful in certain situations (PCIe/PL310
1330 * deadlock on Armada 375/38x due to hardware I/O coherency). The
1331 * other operations are kept because they are infrequent (therefore do
1332 * not cause the deadlock in practice) and needed for secondary CPU
1333 * boot and other power management activities.
1334 */
1335static const struct l2c_init_data of_l2c310_coherent_data __initconst = {
1336 .type = "L2C-310 Coherent",
1337 .way_size_0 = SZ_8K,
1338 .num_lock = 8,
1339 .of_parse = l2c310_of_parse,
1340 .enable = l2c310_enable,
1341 .fixup = l2c310_fixup,
1342 .save = l2c310_save,
1343 .configure = l2c310_configure,
1344 .unlock = l2c310_unlock,
1345 .outer_cache = {
1346 .inv_range = l2c210_inv_range,
1347 .clean_range = l2c210_clean_range,
1348 .flush_range = l2c210_flush_range,
1349 .flush_all = l2c210_flush_all,
1350 .disable = l2c310_disable,
1351 .resume = l2c310_resume,
1352 },
1353};
1354
1355/*
1356 * Note that the end addresses passed to Linux primitives are
1357 * noninclusive, while the hardware cache range operations use
1358 * inclusive start and end addresses.
1359 */
1360static unsigned long aurora_range_end(unsigned long start, unsigned long end)
1361{
1362 /*
1363 * Limit the number of cache lines processed at once,
1364 * since cache range operations stall the CPU pipeline
1365 * until completion.
1366 */
1367 if (end > start + MAX_RANGE_SIZE)
1368 end = start + MAX_RANGE_SIZE;
1369
1370 /*
1371 * Cache range operations can't straddle a page boundary.
1372 */
1373 if (end > PAGE_ALIGN(start+1))
1374 end = PAGE_ALIGN(start+1);
1375
1376 return end;
1377}
1378
1379static void aurora_pa_range(unsigned long start, unsigned long end,
1380 unsigned long offset)
1381{
1382 void __iomem *base = l2x0_base;
1383 unsigned long range_end;
1384 unsigned long flags;
1385
1386 /*
1387 * round start and end adresses up to cache line size
1388 */
1389 start &= ~(CACHE_LINE_SIZE - 1);
1390 end = ALIGN(end, CACHE_LINE_SIZE);
1391
1392 /*
1393 * perform operation on all full cache lines between 'start' and 'end'
1394 */
1395 while (start < end) {
1396 range_end = aurora_range_end(start, end);
1397
1398 raw_spin_lock_irqsave(&l2x0_lock, flags);
1399 writel_relaxed(start, base + AURORA_RANGE_BASE_ADDR_REG);
1400 writel_relaxed(range_end - CACHE_LINE_SIZE, base + offset);
1401 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1402
1403 writel_relaxed(0, base + AURORA_SYNC_REG);
1404 start = range_end;
1405 }
1406}
1407static void aurora_inv_range(unsigned long start, unsigned long end)
1408{
1409 aurora_pa_range(start, end, AURORA_INVAL_RANGE_REG);
1410}
1411
1412static void aurora_clean_range(unsigned long start, unsigned long end)
1413{
1414 /*
1415 * If L2 is forced to WT, the L2 will always be clean and we
1416 * don't need to do anything here.
1417 */
1418 if (!l2_wt_override)
1419 aurora_pa_range(start, end, AURORA_CLEAN_RANGE_REG);
1420}
1421
1422static void aurora_flush_range(unsigned long start, unsigned long end)
1423{
1424 if (l2_wt_override)
1425 aurora_pa_range(start, end, AURORA_INVAL_RANGE_REG);
1426 else
1427 aurora_pa_range(start, end, AURORA_FLUSH_RANGE_REG);
1428}
1429
1430static void aurora_flush_all(void)
1431{
1432 void __iomem *base = l2x0_base;
1433 unsigned long flags;
1434
1435 /* clean all ways */
1436 raw_spin_lock_irqsave(&l2x0_lock, flags);
1437 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
1438 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1439
1440 writel_relaxed(0, base + AURORA_SYNC_REG);
1441}
1442
1443static void aurora_cache_sync(void)
1444{
1445 writel_relaxed(0, l2x0_base + AURORA_SYNC_REG);
1446}
1447
1448static void aurora_disable(void)
1449{
1450 void __iomem *base = l2x0_base;
1451 unsigned long flags;
1452
1453 raw_spin_lock_irqsave(&l2x0_lock, flags);
1454 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
1455 writel_relaxed(0, base + AURORA_SYNC_REG);
1456 l2c_write_sec(0, base, L2X0_CTRL);
1457 dsb(st);
1458 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1459}
1460
1461static void aurora_save(void __iomem *base)
1462{
1463 l2x0_saved_regs.ctrl = readl_relaxed(base + L2X0_CTRL);
1464 l2x0_saved_regs.aux_ctrl = readl_relaxed(base + L2X0_AUX_CTRL);
1465}
1466
1467/*
1468 * For Aurora cache in no outer mode, enable via the CP15 coprocessor
1469 * broadcasting of cache commands to L2.
1470 */
1471static void __init aurora_enable_no_outer(void __iomem *base,
1472 unsigned num_lock)
1473{
1474 u32 u;
1475
1476 asm volatile("mrc p15, 1, %0, c15, c2, 0" : "=r" (u));
1477 u |= AURORA_CTRL_FW; /* Set the FW bit */
1478 asm volatile("mcr p15, 1, %0, c15, c2, 0" : : "r" (u));
1479
1480 isb();
1481
1482 l2c_enable(base, num_lock);
1483}
1484
1485static void __init aurora_fixup(void __iomem *base, u32 cache_id,
1486 struct outer_cache_fns *fns)
1487{
1488 sync_reg_offset = AURORA_SYNC_REG;
1489}
1490
1491static void __init aurora_of_parse(const struct device_node *np,
1492 u32 *aux_val, u32 *aux_mask)
1493{
1494 u32 val = AURORA_ACR_REPLACEMENT_TYPE_SEMIPLRU;
1495 u32 mask = AURORA_ACR_REPLACEMENT_MASK;
1496
1497 of_property_read_u32(np, "cache-id-part",
1498 &cache_id_part_number_from_dt);
1499
1500 /* Determine and save the write policy */
1501 l2_wt_override = of_property_read_bool(np, "wt-override");
1502
1503 if (l2_wt_override) {
1504 val |= AURORA_ACR_FORCE_WRITE_THRO_POLICY;
1505 mask |= AURORA_ACR_FORCE_WRITE_POLICY_MASK;
1506 }
1507
1508 *aux_val &= ~mask;
1509 *aux_val |= val;
1510 *aux_mask &= ~mask;
1511}
1512
1513static const struct l2c_init_data of_aurora_with_outer_data __initconst = {
1514 .type = "Aurora",
1515 .way_size_0 = SZ_4K,
1516 .num_lock = 4,
1517 .of_parse = aurora_of_parse,
1518 .enable = l2c_enable,
1519 .fixup = aurora_fixup,
1520 .save = aurora_save,
1521 .configure = l2c_configure,
1522 .unlock = l2c_unlock,
1523 .outer_cache = {
1524 .inv_range = aurora_inv_range,
1525 .clean_range = aurora_clean_range,
1526 .flush_range = aurora_flush_range,
1527 .flush_all = aurora_flush_all,
1528 .disable = aurora_disable,
1529 .sync = aurora_cache_sync,
1530 .resume = l2c_resume,
1531 },
1532};
1533
1534static const struct l2c_init_data of_aurora_no_outer_data __initconst = {
1535 .type = "Aurora",
1536 .way_size_0 = SZ_4K,
1537 .num_lock = 4,
1538 .of_parse = aurora_of_parse,
1539 .enable = aurora_enable_no_outer,
1540 .fixup = aurora_fixup,
1541 .save = aurora_save,
1542 .configure = l2c_configure,
1543 .unlock = l2c_unlock,
1544 .outer_cache = {
1545 .resume = l2c_resume,
1546 },
1547};
1548
1549/*
1550 * For certain Broadcom SoCs, depending on the address range, different offsets
1551 * need to be added to the address before passing it to L2 for
1552 * invalidation/clean/flush
1553 *
1554 * Section Address Range Offset EMI
1555 * 1 0x00000000 - 0x3FFFFFFF 0x80000000 VC
1556 * 2 0x40000000 - 0xBFFFFFFF 0x40000000 SYS
1557 * 3 0xC0000000 - 0xFFFFFFFF 0x80000000 VC
1558 *
1559 * When the start and end addresses have crossed two different sections, we
1560 * need to break the L2 operation into two, each within its own section.
1561 * For example, if we need to invalidate addresses starts at 0xBFFF0000 and
1562 * ends at 0xC0001000, we need do invalidate 1) 0xBFFF0000 - 0xBFFFFFFF and 2)
1563 * 0xC0000000 - 0xC0001000
1564 *
1565 * Note 1:
1566 * By breaking a single L2 operation into two, we may potentially suffer some
1567 * performance hit, but keep in mind the cross section case is very rare
1568 *
1569 * Note 2:
1570 * We do not need to handle the case when the start address is in
1571 * Section 1 and the end address is in Section 3, since it is not a valid use
1572 * case
1573 *
1574 * Note 3:
1575 * Section 1 in practical terms can no longer be used on rev A2. Because of
1576 * that the code does not need to handle section 1 at all.
1577 *
1578 */
1579#define BCM_SYS_EMI_START_ADDR 0x40000000UL
1580#define BCM_VC_EMI_SEC3_START_ADDR 0xC0000000UL
1581
1582#define BCM_SYS_EMI_OFFSET 0x40000000UL
1583#define BCM_VC_EMI_OFFSET 0x80000000UL
1584
1585static inline int bcm_addr_is_sys_emi(unsigned long addr)
1586{
1587 return (addr >= BCM_SYS_EMI_START_ADDR) &&
1588 (addr < BCM_VC_EMI_SEC3_START_ADDR);
1589}
1590
1591static inline unsigned long bcm_l2_phys_addr(unsigned long addr)
1592{
1593 if (bcm_addr_is_sys_emi(addr))
1594 return addr + BCM_SYS_EMI_OFFSET;
1595 else
1596 return addr + BCM_VC_EMI_OFFSET;
1597}
1598
1599static void bcm_inv_range(unsigned long start, unsigned long end)
1600{
1601 unsigned long new_start, new_end;
1602
1603 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1604
1605 if (unlikely(end <= start))
1606 return;
1607
1608 new_start = bcm_l2_phys_addr(start);
1609 new_end = bcm_l2_phys_addr(end);
1610
1611 /* normal case, no cross section between start and end */
1612 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1613 l2c210_inv_range(new_start, new_end);
1614 return;
1615 }
1616
1617 /* They cross sections, so it can only be a cross from section
1618 * 2 to section 3
1619 */
1620 l2c210_inv_range(new_start,
1621 bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1622 l2c210_inv_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1623 new_end);
1624}
1625
1626static void bcm_clean_range(unsigned long start, unsigned long end)
1627{
1628 unsigned long new_start, new_end;
1629
1630 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1631
1632 if (unlikely(end <= start))
1633 return;
1634
1635 new_start = bcm_l2_phys_addr(start);
1636 new_end = bcm_l2_phys_addr(end);
1637
1638 /* normal case, no cross section between start and end */
1639 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1640 l2c210_clean_range(new_start, new_end);
1641 return;
1642 }
1643
1644 /* They cross sections, so it can only be a cross from section
1645 * 2 to section 3
1646 */
1647 l2c210_clean_range(new_start,
1648 bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1649 l2c210_clean_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1650 new_end);
1651}
1652
1653static void bcm_flush_range(unsigned long start, unsigned long end)
1654{
1655 unsigned long new_start, new_end;
1656
1657 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1658
1659 if (unlikely(end <= start))
1660 return;
1661
1662 if ((end - start) >= l2x0_size) {
1663 outer_cache.flush_all();
1664 return;
1665 }
1666
1667 new_start = bcm_l2_phys_addr(start);
1668 new_end = bcm_l2_phys_addr(end);
1669
1670 /* normal case, no cross section between start and end */
1671 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1672 l2c210_flush_range(new_start, new_end);
1673 return;
1674 }
1675
1676 /* They cross sections, so it can only be a cross from section
1677 * 2 to section 3
1678 */
1679 l2c210_flush_range(new_start,
1680 bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1681 l2c210_flush_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1682 new_end);
1683}
1684
1685/* Broadcom L2C-310 start from ARMs R3P2 or later, and require no fixups */
1686static const struct l2c_init_data of_bcm_l2x0_data __initconst = {
1687 .type = "BCM-L2C-310",
1688 .way_size_0 = SZ_8K,
1689 .num_lock = 8,
1690 .of_parse = l2c310_of_parse,
1691 .enable = l2c310_enable,
1692 .save = l2c310_save,
1693 .configure = l2c310_configure,
1694 .unlock = l2c310_unlock,
1695 .outer_cache = {
1696 .inv_range = bcm_inv_range,
1697 .clean_range = bcm_clean_range,
1698 .flush_range = bcm_flush_range,
1699 .flush_all = l2c210_flush_all,
1700 .disable = l2c310_disable,
1701 .sync = l2c210_sync,
1702 .resume = l2c310_resume,
1703 },
1704};
1705
1706static void __init tauros3_save(void __iomem *base)
1707{
1708 l2c_save(base);
1709
1710 l2x0_saved_regs.aux2_ctrl =
1711 readl_relaxed(base + TAUROS3_AUX2_CTRL);
1712 l2x0_saved_regs.prefetch_ctrl =
1713 readl_relaxed(base + L310_PREFETCH_CTRL);
1714}
1715
1716static void tauros3_configure(void __iomem *base)
1717{
1718 l2c_configure(base);
1719 writel_relaxed(l2x0_saved_regs.aux2_ctrl,
1720 base + TAUROS3_AUX2_CTRL);
1721 writel_relaxed(l2x0_saved_regs.prefetch_ctrl,
1722 base + L310_PREFETCH_CTRL);
1723}
1724
1725static const struct l2c_init_data of_tauros3_data __initconst = {
1726 .type = "Tauros3",
1727 .way_size_0 = SZ_8K,
1728 .num_lock = 8,
1729 .enable = l2c_enable,
1730 .save = tauros3_save,
1731 .configure = tauros3_configure,
1732 .unlock = l2c_unlock,
1733 /* Tauros3 broadcasts L1 cache operations to L2 */
1734 .outer_cache = {
1735 .resume = l2c_resume,
1736 },
1737};
1738
1739#define L2C_ID(name, fns) { .compatible = name, .data = (void *)&fns }
1740static const struct of_device_id l2x0_ids[] __initconst = {
1741 L2C_ID("arm,l210-cache", of_l2c210_data),
1742 L2C_ID("arm,l220-cache", of_l2c220_data),
1743 L2C_ID("arm,pl310-cache", of_l2c310_data),
1744 L2C_ID("brcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1745 L2C_ID("marvell,aurora-outer-cache", of_aurora_with_outer_data),
1746 L2C_ID("marvell,aurora-system-cache", of_aurora_no_outer_data),
1747 L2C_ID("marvell,tauros3-cache", of_tauros3_data),
1748 /* Deprecated IDs */
1749 L2C_ID("bcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1750 {}
1751};
1752
1753int __init l2x0_of_init(u32 aux_val, u32 aux_mask)
1754{
1755 const struct l2c_init_data *data;
1756 struct device_node *np;
1757 struct resource res;
1758 u32 cache_id, old_aux;
1759 u32 cache_level = 2;
1760 bool nosync = false;
1761
1762 np = of_find_matching_node(NULL, l2x0_ids);
1763 if (!np)
1764 return -ENODEV;
1765
1766 if (of_address_to_resource(np, 0, &res))
1767 return -ENODEV;
1768
1769 l2x0_base = ioremap(res.start, resource_size(&res));
1770 if (!l2x0_base)
1771 return -ENOMEM;
1772
1773 l2x0_saved_regs.phy_base = res.start;
1774
1775 data = of_match_node(l2x0_ids, np)->data;
1776
1777 if (of_device_is_compatible(np, "arm,pl310-cache") &&
1778 of_property_read_bool(np, "arm,io-coherent"))
1779 data = &of_l2c310_coherent_data;
1780
1781 old_aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
1782 if (old_aux != ((old_aux & aux_mask) | aux_val)) {
1783 pr_warn("L2C: platform modifies aux control register: 0x%08x -> 0x%08x\n",
1784 old_aux, (old_aux & aux_mask) | aux_val);
1785 } else if (aux_mask != ~0U && aux_val != 0) {
1786 pr_alert("L2C: platform provided aux values match the hardware, so have no effect. Please remove them.\n");
1787 }
1788
1789 /* All L2 caches are unified, so this property should be specified */
1790 if (!of_property_read_bool(np, "cache-unified"))
1791 pr_err("L2C: device tree omits to specify unified cache\n");
1792
1793 if (of_property_read_u32(np, "cache-level", &cache_level))
1794 pr_err("L2C: device tree omits to specify cache-level\n");
1795
1796 if (cache_level != 2)
1797 pr_err("L2C: device tree specifies invalid cache level\n");
1798
1799 nosync = of_property_read_bool(np, "arm,outer-sync-disable");
1800
1801 /* Read back current (default) hardware configuration */
1802 if (data->save)
1803 data->save(l2x0_base);
1804
1805 /* L2 configuration can only be changed if the cache is disabled */
1806 if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN))
1807 if (data->of_parse)
1808 data->of_parse(np, &aux_val, &aux_mask);
1809
1810 if (cache_id_part_number_from_dt)
1811 cache_id = cache_id_part_number_from_dt;
1812 else
1813 cache_id = readl_relaxed(l2x0_base + L2X0_CACHE_ID);
1814
1815 return __l2c_init(data, aux_val, aux_mask, cache_id, nosync);
1816}
1817#endif
1/*
2 * arch/arm/mm/cache-l2x0.c - L210/L220/L310 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/cpu.h>
20#include <linux/err.h>
21#include <linux/init.h>
22#include <linux/smp.h>
23#include <linux/spinlock.h>
24#include <linux/log2.h>
25#include <linux/io.h>
26#include <linux/of.h>
27#include <linux/of_address.h>
28
29#include <asm/cacheflush.h>
30#include <asm/cp15.h>
31#include <asm/cputype.h>
32#include <asm/hardware/cache-l2x0.h>
33#include "cache-tauros3.h"
34#include "cache-aurora-l2.h"
35
36struct l2c_init_data {
37 const char *type;
38 unsigned way_size_0;
39 unsigned num_lock;
40 void (*of_parse)(const struct device_node *, u32 *, u32 *);
41 void (*enable)(void __iomem *, unsigned);
42 void (*fixup)(void __iomem *, u32, struct outer_cache_fns *);
43 void (*save)(void __iomem *);
44 void (*configure)(void __iomem *);
45 void (*unlock)(void __iomem *, unsigned);
46 struct outer_cache_fns outer_cache;
47};
48
49#define CACHE_LINE_SIZE 32
50
51static void __iomem *l2x0_base;
52static const struct l2c_init_data *l2x0_data;
53static DEFINE_RAW_SPINLOCK(l2x0_lock);
54static u32 l2x0_way_mask; /* Bitmask of active ways */
55static u32 l2x0_size;
56static unsigned long sync_reg_offset = L2X0_CACHE_SYNC;
57
58struct l2x0_regs l2x0_saved_regs;
59
60/*
61 * Common code for all cache controllers.
62 */
63static inline void l2c_wait_mask(void __iomem *reg, unsigned long mask)
64{
65 /* wait for cache operation by line or way to complete */
66 while (readl_relaxed(reg) & mask)
67 cpu_relax();
68}
69
70/*
71 * By default, we write directly to secure registers. Platforms must
72 * override this if they are running non-secure.
73 */
74static void l2c_write_sec(unsigned long val, void __iomem *base, unsigned reg)
75{
76 if (val == readl_relaxed(base + reg))
77 return;
78 if (outer_cache.write_sec)
79 outer_cache.write_sec(val, reg);
80 else
81 writel_relaxed(val, base + reg);
82}
83
84/*
85 * This should only be called when we have a requirement that the
86 * register be written due to a work-around, as platforms running
87 * in non-secure mode may not be able to access this register.
88 */
89static inline void l2c_set_debug(void __iomem *base, unsigned long val)
90{
91 l2c_write_sec(val, base, L2X0_DEBUG_CTRL);
92}
93
94static void __l2c_op_way(void __iomem *reg)
95{
96 writel_relaxed(l2x0_way_mask, reg);
97 l2c_wait_mask(reg, l2x0_way_mask);
98}
99
100static inline void l2c_unlock(void __iomem *base, unsigned num)
101{
102 unsigned i;
103
104 for (i = 0; i < num; i++) {
105 writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_D_BASE +
106 i * L2X0_LOCKDOWN_STRIDE);
107 writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_I_BASE +
108 i * L2X0_LOCKDOWN_STRIDE);
109 }
110}
111
112static void l2c_configure(void __iomem *base)
113{
114 l2c_write_sec(l2x0_saved_regs.aux_ctrl, base, L2X0_AUX_CTRL);
115}
116
117/*
118 * Enable the L2 cache controller. This function must only be
119 * called when the cache controller is known to be disabled.
120 */
121static void l2c_enable(void __iomem *base, unsigned num_lock)
122{
123 unsigned long flags;
124
125 if (outer_cache.configure)
126 outer_cache.configure(&l2x0_saved_regs);
127 else
128 l2x0_data->configure(base);
129
130 l2x0_data->unlock(base, num_lock);
131
132 local_irq_save(flags);
133 __l2c_op_way(base + L2X0_INV_WAY);
134 writel_relaxed(0, base + sync_reg_offset);
135 l2c_wait_mask(base + sync_reg_offset, 1);
136 local_irq_restore(flags);
137
138 l2c_write_sec(L2X0_CTRL_EN, base, L2X0_CTRL);
139}
140
141static void l2c_disable(void)
142{
143 void __iomem *base = l2x0_base;
144
145 outer_cache.flush_all();
146 l2c_write_sec(0, base, L2X0_CTRL);
147 dsb(st);
148}
149
150static void l2c_save(void __iomem *base)
151{
152 l2x0_saved_regs.aux_ctrl = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
153}
154
155static void l2c_resume(void)
156{
157 void __iomem *base = l2x0_base;
158
159 /* Do not touch the controller if already enabled. */
160 if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN))
161 l2c_enable(base, l2x0_data->num_lock);
162}
163
164/*
165 * L2C-210 specific code.
166 *
167 * The L2C-2x0 PA, set/way and sync operations are atomic, but we must
168 * ensure that no background operation is running. The way operations
169 * are all background tasks.
170 *
171 * While a background operation is in progress, any new operation is
172 * ignored (unspecified whether this causes an error.) Thankfully, not
173 * used on SMP.
174 *
175 * Never has a different sync register other than L2X0_CACHE_SYNC, but
176 * we use sync_reg_offset here so we can share some of this with L2C-310.
177 */
178static void __l2c210_cache_sync(void __iomem *base)
179{
180 writel_relaxed(0, base + sync_reg_offset);
181}
182
183static void __l2c210_op_pa_range(void __iomem *reg, unsigned long start,
184 unsigned long end)
185{
186 while (start < end) {
187 writel_relaxed(start, reg);
188 start += CACHE_LINE_SIZE;
189 }
190}
191
192static void l2c210_inv_range(unsigned long start, unsigned long end)
193{
194 void __iomem *base = l2x0_base;
195
196 if (start & (CACHE_LINE_SIZE - 1)) {
197 start &= ~(CACHE_LINE_SIZE - 1);
198 writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
199 start += CACHE_LINE_SIZE;
200 }
201
202 if (end & (CACHE_LINE_SIZE - 1)) {
203 end &= ~(CACHE_LINE_SIZE - 1);
204 writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
205 }
206
207 __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
208 __l2c210_cache_sync(base);
209}
210
211static void l2c210_clean_range(unsigned long start, unsigned long end)
212{
213 void __iomem *base = l2x0_base;
214
215 start &= ~(CACHE_LINE_SIZE - 1);
216 __l2c210_op_pa_range(base + L2X0_CLEAN_LINE_PA, start, end);
217 __l2c210_cache_sync(base);
218}
219
220static void l2c210_flush_range(unsigned long start, unsigned long end)
221{
222 void __iomem *base = l2x0_base;
223
224 start &= ~(CACHE_LINE_SIZE - 1);
225 __l2c210_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA, start, end);
226 __l2c210_cache_sync(base);
227}
228
229static void l2c210_flush_all(void)
230{
231 void __iomem *base = l2x0_base;
232
233 BUG_ON(!irqs_disabled());
234
235 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
236 __l2c210_cache_sync(base);
237}
238
239static void l2c210_sync(void)
240{
241 __l2c210_cache_sync(l2x0_base);
242}
243
244static const struct l2c_init_data l2c210_data __initconst = {
245 .type = "L2C-210",
246 .way_size_0 = SZ_8K,
247 .num_lock = 1,
248 .enable = l2c_enable,
249 .save = l2c_save,
250 .configure = l2c_configure,
251 .unlock = l2c_unlock,
252 .outer_cache = {
253 .inv_range = l2c210_inv_range,
254 .clean_range = l2c210_clean_range,
255 .flush_range = l2c210_flush_range,
256 .flush_all = l2c210_flush_all,
257 .disable = l2c_disable,
258 .sync = l2c210_sync,
259 .resume = l2c_resume,
260 },
261};
262
263/*
264 * L2C-220 specific code.
265 *
266 * All operations are background operations: they have to be waited for.
267 * Conflicting requests generate a slave error (which will cause an
268 * imprecise abort.) Never uses sync_reg_offset, so we hard-code the
269 * sync register here.
270 *
271 * However, we can re-use the l2c210_resume call.
272 */
273static inline void __l2c220_cache_sync(void __iomem *base)
274{
275 writel_relaxed(0, base + L2X0_CACHE_SYNC);
276 l2c_wait_mask(base + L2X0_CACHE_SYNC, 1);
277}
278
279static void l2c220_op_way(void __iomem *base, unsigned reg)
280{
281 unsigned long flags;
282
283 raw_spin_lock_irqsave(&l2x0_lock, flags);
284 __l2c_op_way(base + reg);
285 __l2c220_cache_sync(base);
286 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
287}
288
289static unsigned long l2c220_op_pa_range(void __iomem *reg, unsigned long start,
290 unsigned long end, unsigned long flags)
291{
292 raw_spinlock_t *lock = &l2x0_lock;
293
294 while (start < end) {
295 unsigned long blk_end = start + min(end - start, 4096UL);
296
297 while (start < blk_end) {
298 l2c_wait_mask(reg, 1);
299 writel_relaxed(start, reg);
300 start += CACHE_LINE_SIZE;
301 }
302
303 if (blk_end < end) {
304 raw_spin_unlock_irqrestore(lock, flags);
305 raw_spin_lock_irqsave(lock, flags);
306 }
307 }
308
309 return flags;
310}
311
312static void l2c220_inv_range(unsigned long start, unsigned long end)
313{
314 void __iomem *base = l2x0_base;
315 unsigned long flags;
316
317 raw_spin_lock_irqsave(&l2x0_lock, flags);
318 if ((start | end) & (CACHE_LINE_SIZE - 1)) {
319 if (start & (CACHE_LINE_SIZE - 1)) {
320 start &= ~(CACHE_LINE_SIZE - 1);
321 writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
322 start += CACHE_LINE_SIZE;
323 }
324
325 if (end & (CACHE_LINE_SIZE - 1)) {
326 end &= ~(CACHE_LINE_SIZE - 1);
327 l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
328 writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
329 }
330 }
331
332 flags = l2c220_op_pa_range(base + L2X0_INV_LINE_PA,
333 start, end, flags);
334 l2c_wait_mask(base + L2X0_INV_LINE_PA, 1);
335 __l2c220_cache_sync(base);
336 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
337}
338
339static void l2c220_clean_range(unsigned long start, unsigned long end)
340{
341 void __iomem *base = l2x0_base;
342 unsigned long flags;
343
344 start &= ~(CACHE_LINE_SIZE - 1);
345 if ((end - start) >= l2x0_size) {
346 l2c220_op_way(base, L2X0_CLEAN_WAY);
347 return;
348 }
349
350 raw_spin_lock_irqsave(&l2x0_lock, flags);
351 flags = l2c220_op_pa_range(base + L2X0_CLEAN_LINE_PA,
352 start, end, flags);
353 l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
354 __l2c220_cache_sync(base);
355 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
356}
357
358static void l2c220_flush_range(unsigned long start, unsigned long end)
359{
360 void __iomem *base = l2x0_base;
361 unsigned long flags;
362
363 start &= ~(CACHE_LINE_SIZE - 1);
364 if ((end - start) >= l2x0_size) {
365 l2c220_op_way(base, L2X0_CLEAN_INV_WAY);
366 return;
367 }
368
369 raw_spin_lock_irqsave(&l2x0_lock, flags);
370 flags = l2c220_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA,
371 start, end, flags);
372 l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1);
373 __l2c220_cache_sync(base);
374 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
375}
376
377static void l2c220_flush_all(void)
378{
379 l2c220_op_way(l2x0_base, L2X0_CLEAN_INV_WAY);
380}
381
382static void l2c220_sync(void)
383{
384 unsigned long flags;
385
386 raw_spin_lock_irqsave(&l2x0_lock, flags);
387 __l2c220_cache_sync(l2x0_base);
388 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
389}
390
391static void l2c220_enable(void __iomem *base, unsigned num_lock)
392{
393 /*
394 * Always enable non-secure access to the lockdown registers -
395 * we write to them as part of the L2C enable sequence so they
396 * need to be accessible.
397 */
398 l2x0_saved_regs.aux_ctrl |= L220_AUX_CTRL_NS_LOCKDOWN;
399
400 l2c_enable(base, num_lock);
401}
402
403static void l2c220_unlock(void __iomem *base, unsigned num_lock)
404{
405 if (readl_relaxed(base + L2X0_AUX_CTRL) & L220_AUX_CTRL_NS_LOCKDOWN)
406 l2c_unlock(base, num_lock);
407}
408
409static const struct l2c_init_data l2c220_data = {
410 .type = "L2C-220",
411 .way_size_0 = SZ_8K,
412 .num_lock = 1,
413 .enable = l2c220_enable,
414 .save = l2c_save,
415 .configure = l2c_configure,
416 .unlock = l2c220_unlock,
417 .outer_cache = {
418 .inv_range = l2c220_inv_range,
419 .clean_range = l2c220_clean_range,
420 .flush_range = l2c220_flush_range,
421 .flush_all = l2c220_flush_all,
422 .disable = l2c_disable,
423 .sync = l2c220_sync,
424 .resume = l2c_resume,
425 },
426};
427
428/*
429 * L2C-310 specific code.
430 *
431 * Very similar to L2C-210, the PA, set/way and sync operations are atomic,
432 * and the way operations are all background tasks. However, issuing an
433 * operation while a background operation is in progress results in a
434 * SLVERR response. We can reuse:
435 *
436 * __l2c210_cache_sync (using sync_reg_offset)
437 * l2c210_sync
438 * l2c210_inv_range (if 588369 is not applicable)
439 * l2c210_clean_range
440 * l2c210_flush_range (if 588369 is not applicable)
441 * l2c210_flush_all (if 727915 is not applicable)
442 *
443 * Errata:
444 * 588369: PL310 R0P0->R1P0, fixed R2P0.
445 * Affects: all clean+invalidate operations
446 * clean and invalidate skips the invalidate step, so we need to issue
447 * separate operations. We also require the above debug workaround
448 * enclosing this code fragment on affected parts. On unaffected parts,
449 * we must not use this workaround without the debug register writes
450 * to avoid exposing a problem similar to 727915.
451 *
452 * 727915: PL310 R2P0->R3P0, fixed R3P1.
453 * Affects: clean+invalidate by way
454 * clean and invalidate by way runs in the background, and a store can
455 * hit the line between the clean operation and invalidate operation,
456 * resulting in the store being lost.
457 *
458 * 752271: PL310 R3P0->R3P1-50REL0, fixed R3P2.
459 * Affects: 8x64-bit (double fill) line fetches
460 * double fill line fetches can fail to cause dirty data to be evicted
461 * from the cache before the new data overwrites the second line.
462 *
463 * 753970: PL310 R3P0, fixed R3P1.
464 * Affects: sync
465 * prevents merging writes after the sync operation, until another L2C
466 * operation is performed (or a number of other conditions.)
467 *
468 * 769419: PL310 R0P0->R3P1, fixed R3P2.
469 * Affects: store buffer
470 * store buffer is not automatically drained.
471 */
472static void l2c310_inv_range_erratum(unsigned long start, unsigned long end)
473{
474 void __iomem *base = l2x0_base;
475
476 if ((start | end) & (CACHE_LINE_SIZE - 1)) {
477 unsigned long flags;
478
479 /* Erratum 588369 for both clean+invalidate operations */
480 raw_spin_lock_irqsave(&l2x0_lock, flags);
481 l2c_set_debug(base, 0x03);
482
483 if (start & (CACHE_LINE_SIZE - 1)) {
484 start &= ~(CACHE_LINE_SIZE - 1);
485 writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
486 writel_relaxed(start, base + L2X0_INV_LINE_PA);
487 start += CACHE_LINE_SIZE;
488 }
489
490 if (end & (CACHE_LINE_SIZE - 1)) {
491 end &= ~(CACHE_LINE_SIZE - 1);
492 writel_relaxed(end, base + L2X0_CLEAN_LINE_PA);
493 writel_relaxed(end, base + L2X0_INV_LINE_PA);
494 }
495
496 l2c_set_debug(base, 0x00);
497 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
498 }
499
500 __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
501 __l2c210_cache_sync(base);
502}
503
504static void l2c310_flush_range_erratum(unsigned long start, unsigned long end)
505{
506 raw_spinlock_t *lock = &l2x0_lock;
507 unsigned long flags;
508 void __iomem *base = l2x0_base;
509
510 raw_spin_lock_irqsave(lock, flags);
511 while (start < end) {
512 unsigned long blk_end = start + min(end - start, 4096UL);
513
514 l2c_set_debug(base, 0x03);
515 while (start < blk_end) {
516 writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
517 writel_relaxed(start, base + L2X0_INV_LINE_PA);
518 start += CACHE_LINE_SIZE;
519 }
520 l2c_set_debug(base, 0x00);
521
522 if (blk_end < end) {
523 raw_spin_unlock_irqrestore(lock, flags);
524 raw_spin_lock_irqsave(lock, flags);
525 }
526 }
527 raw_spin_unlock_irqrestore(lock, flags);
528 __l2c210_cache_sync(base);
529}
530
531static void l2c310_flush_all_erratum(void)
532{
533 void __iomem *base = l2x0_base;
534 unsigned long flags;
535
536 raw_spin_lock_irqsave(&l2x0_lock, flags);
537 l2c_set_debug(base, 0x03);
538 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
539 l2c_set_debug(base, 0x00);
540 __l2c210_cache_sync(base);
541 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
542}
543
544static void __init l2c310_save(void __iomem *base)
545{
546 unsigned revision;
547
548 l2c_save(base);
549
550 l2x0_saved_regs.tag_latency = readl_relaxed(base +
551 L310_TAG_LATENCY_CTRL);
552 l2x0_saved_regs.data_latency = readl_relaxed(base +
553 L310_DATA_LATENCY_CTRL);
554 l2x0_saved_regs.filter_end = readl_relaxed(base +
555 L310_ADDR_FILTER_END);
556 l2x0_saved_regs.filter_start = readl_relaxed(base +
557 L310_ADDR_FILTER_START);
558
559 revision = readl_relaxed(base + L2X0_CACHE_ID) &
560 L2X0_CACHE_ID_RTL_MASK;
561
562 /* From r2p0, there is Prefetch offset/control register */
563 if (revision >= L310_CACHE_ID_RTL_R2P0)
564 l2x0_saved_regs.prefetch_ctrl = readl_relaxed(base +
565 L310_PREFETCH_CTRL);
566
567 /* From r3p0, there is Power control register */
568 if (revision >= L310_CACHE_ID_RTL_R3P0)
569 l2x0_saved_regs.pwr_ctrl = readl_relaxed(base +
570 L310_POWER_CTRL);
571}
572
573static void l2c310_configure(void __iomem *base)
574{
575 unsigned revision;
576
577 l2c_configure(base);
578
579 /* restore pl310 setup */
580 l2c_write_sec(l2x0_saved_regs.tag_latency, base,
581 L310_TAG_LATENCY_CTRL);
582 l2c_write_sec(l2x0_saved_regs.data_latency, base,
583 L310_DATA_LATENCY_CTRL);
584 l2c_write_sec(l2x0_saved_regs.filter_end, base,
585 L310_ADDR_FILTER_END);
586 l2c_write_sec(l2x0_saved_regs.filter_start, base,
587 L310_ADDR_FILTER_START);
588
589 revision = readl_relaxed(base + L2X0_CACHE_ID) &
590 L2X0_CACHE_ID_RTL_MASK;
591
592 if (revision >= L310_CACHE_ID_RTL_R2P0)
593 l2c_write_sec(l2x0_saved_regs.prefetch_ctrl, base,
594 L310_PREFETCH_CTRL);
595 if (revision >= L310_CACHE_ID_RTL_R3P0)
596 l2c_write_sec(l2x0_saved_regs.pwr_ctrl, base,
597 L310_POWER_CTRL);
598}
599
600static int l2c310_cpu_enable_flz(struct notifier_block *nb, unsigned long act, void *data)
601{
602 switch (act & ~CPU_TASKS_FROZEN) {
603 case CPU_STARTING:
604 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
605 break;
606 case CPU_DYING:
607 set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1)));
608 break;
609 }
610 return NOTIFY_OK;
611}
612
613static void __init l2c310_enable(void __iomem *base, unsigned num_lock)
614{
615 unsigned rev = readl_relaxed(base + L2X0_CACHE_ID) & L2X0_CACHE_ID_RTL_MASK;
616 bool cortex_a9 = read_cpuid_part() == ARM_CPU_PART_CORTEX_A9;
617 u32 aux = l2x0_saved_regs.aux_ctrl;
618
619 if (rev >= L310_CACHE_ID_RTL_R2P0) {
620 if (cortex_a9) {
621 aux |= L310_AUX_CTRL_EARLY_BRESP;
622 pr_info("L2C-310 enabling early BRESP for Cortex-A9\n");
623 } else if (aux & L310_AUX_CTRL_EARLY_BRESP) {
624 pr_warn("L2C-310 early BRESP only supported with Cortex-A9\n");
625 aux &= ~L310_AUX_CTRL_EARLY_BRESP;
626 }
627 }
628
629 if (cortex_a9) {
630 u32 aux_cur = readl_relaxed(base + L2X0_AUX_CTRL);
631 u32 acr = get_auxcr();
632
633 pr_debug("Cortex-A9 ACR=0x%08x\n", acr);
634
635 if (acr & BIT(3) && !(aux_cur & L310_AUX_CTRL_FULL_LINE_ZERO))
636 pr_err("L2C-310: full line of zeros enabled in Cortex-A9 but not L2C-310 - invalid\n");
637
638 if (aux & L310_AUX_CTRL_FULL_LINE_ZERO && !(acr & BIT(3)))
639 pr_err("L2C-310: enabling full line of zeros but not enabled in Cortex-A9\n");
640
641 if (!(aux & L310_AUX_CTRL_FULL_LINE_ZERO) && !outer_cache.write_sec) {
642 aux |= L310_AUX_CTRL_FULL_LINE_ZERO;
643 pr_info("L2C-310 full line of zeros enabled for Cortex-A9\n");
644 }
645 } else if (aux & (L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP)) {
646 pr_err("L2C-310: disabling Cortex-A9 specific feature bits\n");
647 aux &= ~(L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP);
648 }
649
650 /* r3p0 or later has power control register */
651 if (rev >= L310_CACHE_ID_RTL_R3P0)
652 l2x0_saved_regs.pwr_ctrl = L310_DYNAMIC_CLK_GATING_EN |
653 L310_STNDBY_MODE_EN;
654
655 /*
656 * Always enable non-secure access to the lockdown registers -
657 * we write to them as part of the L2C enable sequence so they
658 * need to be accessible.
659 */
660 l2x0_saved_regs.aux_ctrl = aux | L310_AUX_CTRL_NS_LOCKDOWN;
661
662 l2c_enable(base, num_lock);
663
664 /* Read back resulting AUX_CTRL value as it could have been altered. */
665 aux = readl_relaxed(base + L2X0_AUX_CTRL);
666
667 if (aux & (L310_AUX_CTRL_DATA_PREFETCH | L310_AUX_CTRL_INSTR_PREFETCH)) {
668 u32 prefetch = readl_relaxed(base + L310_PREFETCH_CTRL);
669
670 pr_info("L2C-310 %s%s prefetch enabled, offset %u lines\n",
671 aux & L310_AUX_CTRL_INSTR_PREFETCH ? "I" : "",
672 aux & L310_AUX_CTRL_DATA_PREFETCH ? "D" : "",
673 1 + (prefetch & L310_PREFETCH_CTRL_OFFSET_MASK));
674 }
675
676 /* r3p0 or later has power control register */
677 if (rev >= L310_CACHE_ID_RTL_R3P0) {
678 u32 power_ctrl;
679
680 power_ctrl = readl_relaxed(base + L310_POWER_CTRL);
681 pr_info("L2C-310 dynamic clock gating %sabled, standby mode %sabled\n",
682 power_ctrl & L310_DYNAMIC_CLK_GATING_EN ? "en" : "dis",
683 power_ctrl & L310_STNDBY_MODE_EN ? "en" : "dis");
684 }
685
686 if (aux & L310_AUX_CTRL_FULL_LINE_ZERO) {
687 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
688 cpu_notifier(l2c310_cpu_enable_flz, 0);
689 }
690}
691
692static void __init l2c310_fixup(void __iomem *base, u32 cache_id,
693 struct outer_cache_fns *fns)
694{
695 unsigned revision = cache_id & L2X0_CACHE_ID_RTL_MASK;
696 const char *errata[8];
697 unsigned n = 0;
698
699 if (IS_ENABLED(CONFIG_PL310_ERRATA_588369) &&
700 revision < L310_CACHE_ID_RTL_R2P0 &&
701 /* For bcm compatibility */
702 fns->inv_range == l2c210_inv_range) {
703 fns->inv_range = l2c310_inv_range_erratum;
704 fns->flush_range = l2c310_flush_range_erratum;
705 errata[n++] = "588369";
706 }
707
708 if (IS_ENABLED(CONFIG_PL310_ERRATA_727915) &&
709 revision >= L310_CACHE_ID_RTL_R2P0 &&
710 revision < L310_CACHE_ID_RTL_R3P1) {
711 fns->flush_all = l2c310_flush_all_erratum;
712 errata[n++] = "727915";
713 }
714
715 if (revision >= L310_CACHE_ID_RTL_R3P0 &&
716 revision < L310_CACHE_ID_RTL_R3P2) {
717 u32 val = l2x0_saved_regs.prefetch_ctrl;
718 /* I don't think bit23 is required here... but iMX6 does so */
719 if (val & (BIT(30) | BIT(23))) {
720 val &= ~(BIT(30) | BIT(23));
721 l2x0_saved_regs.prefetch_ctrl = val;
722 errata[n++] = "752271";
723 }
724 }
725
726 if (IS_ENABLED(CONFIG_PL310_ERRATA_753970) &&
727 revision == L310_CACHE_ID_RTL_R3P0) {
728 sync_reg_offset = L2X0_DUMMY_REG;
729 errata[n++] = "753970";
730 }
731
732 if (IS_ENABLED(CONFIG_PL310_ERRATA_769419))
733 errata[n++] = "769419";
734
735 if (n) {
736 unsigned i;
737
738 pr_info("L2C-310 errat%s", n > 1 ? "a" : "um");
739 for (i = 0; i < n; i++)
740 pr_cont(" %s", errata[i]);
741 pr_cont(" enabled\n");
742 }
743}
744
745static void l2c310_disable(void)
746{
747 /*
748 * If full-line-of-zeros is enabled, we must first disable it in the
749 * Cortex-A9 auxiliary control register before disabling the L2 cache.
750 */
751 if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
752 set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1)));
753
754 l2c_disable();
755}
756
757static void l2c310_resume(void)
758{
759 l2c_resume();
760
761 /* Re-enable full-line-of-zeros for Cortex-A9 */
762 if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
763 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
764}
765
766static void l2c310_unlock(void __iomem *base, unsigned num_lock)
767{
768 if (readl_relaxed(base + L2X0_AUX_CTRL) & L310_AUX_CTRL_NS_LOCKDOWN)
769 l2c_unlock(base, num_lock);
770}
771
772static const struct l2c_init_data l2c310_init_fns __initconst = {
773 .type = "L2C-310",
774 .way_size_0 = SZ_8K,
775 .num_lock = 8,
776 .enable = l2c310_enable,
777 .fixup = l2c310_fixup,
778 .save = l2c310_save,
779 .configure = l2c310_configure,
780 .unlock = l2c310_unlock,
781 .outer_cache = {
782 .inv_range = l2c210_inv_range,
783 .clean_range = l2c210_clean_range,
784 .flush_range = l2c210_flush_range,
785 .flush_all = l2c210_flush_all,
786 .disable = l2c310_disable,
787 .sync = l2c210_sync,
788 .resume = l2c310_resume,
789 },
790};
791
792static int __init __l2c_init(const struct l2c_init_data *data,
793 u32 aux_val, u32 aux_mask, u32 cache_id, bool nosync)
794{
795 struct outer_cache_fns fns;
796 unsigned way_size_bits, ways;
797 u32 aux, old_aux;
798
799 /*
800 * Save the pointer globally so that callbacks which do not receive
801 * context from callers can access the structure.
802 */
803 l2x0_data = kmemdup(data, sizeof(*data), GFP_KERNEL);
804 if (!l2x0_data)
805 return -ENOMEM;
806
807 /*
808 * Sanity check the aux values. aux_mask is the bits we preserve
809 * from reading the hardware register, and aux_val is the bits we
810 * set.
811 */
812 if (aux_val & aux_mask)
813 pr_alert("L2C: platform provided aux values permit register corruption.\n");
814
815 old_aux = aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
816 aux &= aux_mask;
817 aux |= aux_val;
818
819 if (old_aux != aux)
820 pr_warn("L2C: DT/platform modifies aux control register: 0x%08x -> 0x%08x\n",
821 old_aux, aux);
822
823 /* Determine the number of ways */
824 switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
825 case L2X0_CACHE_ID_PART_L310:
826 if ((aux_val | ~aux_mask) & (L2C_AUX_CTRL_WAY_SIZE_MASK | L310_AUX_CTRL_ASSOCIATIVITY_16))
827 pr_warn("L2C: DT/platform tries to modify or specify cache size\n");
828 if (aux & (1 << 16))
829 ways = 16;
830 else
831 ways = 8;
832 break;
833
834 case L2X0_CACHE_ID_PART_L210:
835 case L2X0_CACHE_ID_PART_L220:
836 ways = (aux >> 13) & 0xf;
837 break;
838
839 case AURORA_CACHE_ID:
840 ways = (aux >> 13) & 0xf;
841 ways = 2 << ((ways + 1) >> 2);
842 break;
843
844 default:
845 /* Assume unknown chips have 8 ways */
846 ways = 8;
847 break;
848 }
849
850 l2x0_way_mask = (1 << ways) - 1;
851
852 /*
853 * way_size_0 is the size that a way_size value of zero would be
854 * given the calculation: way_size = way_size_0 << way_size_bits.
855 * So, if way_size_bits=0 is reserved, but way_size_bits=1 is 16k,
856 * then way_size_0 would be 8k.
857 *
858 * L2 cache size = number of ways * way size.
859 */
860 way_size_bits = (aux & L2C_AUX_CTRL_WAY_SIZE_MASK) >>
861 L2C_AUX_CTRL_WAY_SIZE_SHIFT;
862 l2x0_size = ways * (data->way_size_0 << way_size_bits);
863
864 fns = data->outer_cache;
865 fns.write_sec = outer_cache.write_sec;
866 fns.configure = outer_cache.configure;
867 if (data->fixup)
868 data->fixup(l2x0_base, cache_id, &fns);
869 if (nosync) {
870 pr_info("L2C: disabling outer sync\n");
871 fns.sync = NULL;
872 }
873
874 /*
875 * Check if l2x0 controller is already enabled. If we are booting
876 * in non-secure mode accessing the below registers will fault.
877 */
878 if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN)) {
879 l2x0_saved_regs.aux_ctrl = aux;
880
881 data->enable(l2x0_base, data->num_lock);
882 }
883
884 outer_cache = fns;
885
886 /*
887 * It is strange to save the register state before initialisation,
888 * but hey, this is what the DT implementations decided to do.
889 */
890 if (data->save)
891 data->save(l2x0_base);
892
893 /* Re-read it in case some bits are reserved. */
894 aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
895
896 pr_info("%s cache controller enabled, %d ways, %d kB\n",
897 data->type, ways, l2x0_size >> 10);
898 pr_info("%s: CACHE_ID 0x%08x, AUX_CTRL 0x%08x\n",
899 data->type, cache_id, aux);
900
901 return 0;
902}
903
904void __init l2x0_init(void __iomem *base, u32 aux_val, u32 aux_mask)
905{
906 const struct l2c_init_data *data;
907 u32 cache_id;
908
909 l2x0_base = base;
910
911 cache_id = readl_relaxed(base + L2X0_CACHE_ID);
912
913 switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
914 default:
915 case L2X0_CACHE_ID_PART_L210:
916 data = &l2c210_data;
917 break;
918
919 case L2X0_CACHE_ID_PART_L220:
920 data = &l2c220_data;
921 break;
922
923 case L2X0_CACHE_ID_PART_L310:
924 data = &l2c310_init_fns;
925 break;
926 }
927
928 /* Read back current (default) hardware configuration */
929 if (data->save)
930 data->save(l2x0_base);
931
932 __l2c_init(data, aux_val, aux_mask, cache_id, false);
933}
934
935#ifdef CONFIG_OF
936static int l2_wt_override;
937
938/* Aurora don't have the cache ID register available, so we have to
939 * pass it though the device tree */
940static u32 cache_id_part_number_from_dt;
941
942/**
943 * l2x0_cache_size_of_parse() - read cache size parameters from DT
944 * @np: the device tree node for the l2 cache
945 * @aux_val: pointer to machine-supplied auxilary register value, to
946 * be augmented by the call (bits to be set to 1)
947 * @aux_mask: pointer to machine-supplied auxilary register mask, to
948 * be augmented by the call (bits to be set to 0)
949 * @associativity: variable to return the calculated associativity in
950 * @max_way_size: the maximum size in bytes for the cache ways
951 */
952static int __init l2x0_cache_size_of_parse(const struct device_node *np,
953 u32 *aux_val, u32 *aux_mask,
954 u32 *associativity,
955 u32 max_way_size)
956{
957 u32 mask = 0, val = 0;
958 u32 cache_size = 0, sets = 0;
959 u32 way_size_bits = 1;
960 u32 way_size = 0;
961 u32 block_size = 0;
962 u32 line_size = 0;
963
964 of_property_read_u32(np, "cache-size", &cache_size);
965 of_property_read_u32(np, "cache-sets", &sets);
966 of_property_read_u32(np, "cache-block-size", &block_size);
967 of_property_read_u32(np, "cache-line-size", &line_size);
968
969 if (!cache_size || !sets)
970 return -ENODEV;
971
972 /* All these l2 caches have the same line = block size actually */
973 if (!line_size) {
974 if (block_size) {
975 /* If linesize is not given, it is equal to blocksize */
976 line_size = block_size;
977 } else {
978 /* Fall back to known size */
979 pr_warn("L2C OF: no cache block/line size given: "
980 "falling back to default size %d bytes\n",
981 CACHE_LINE_SIZE);
982 line_size = CACHE_LINE_SIZE;
983 }
984 }
985
986 if (line_size != CACHE_LINE_SIZE)
987 pr_warn("L2C OF: DT supplied line size %d bytes does "
988 "not match hardware line size of %d bytes\n",
989 line_size,
990 CACHE_LINE_SIZE);
991
992 /*
993 * Since:
994 * set size = cache size / sets
995 * ways = cache size / (sets * line size)
996 * way size = cache size / (cache size / (sets * line size))
997 * way size = sets * line size
998 * associativity = ways = cache size / way size
999 */
1000 way_size = sets * line_size;
1001 *associativity = cache_size / way_size;
1002
1003 if (way_size > max_way_size) {
1004 pr_err("L2C OF: set size %dKB is too large\n", way_size);
1005 return -EINVAL;
1006 }
1007
1008 pr_info("L2C OF: override cache size: %d bytes (%dKB)\n",
1009 cache_size, cache_size >> 10);
1010 pr_info("L2C OF: override line size: %d bytes\n", line_size);
1011 pr_info("L2C OF: override way size: %d bytes (%dKB)\n",
1012 way_size, way_size >> 10);
1013 pr_info("L2C OF: override associativity: %d\n", *associativity);
1014
1015 /*
1016 * Calculates the bits 17:19 to set for way size:
1017 * 512KB -> 6, 256KB -> 5, ... 16KB -> 1
1018 */
1019 way_size_bits = ilog2(way_size >> 10) - 3;
1020 if (way_size_bits < 1 || way_size_bits > 6) {
1021 pr_err("L2C OF: cache way size illegal: %dKB is not mapped\n",
1022 way_size);
1023 return -EINVAL;
1024 }
1025
1026 mask |= L2C_AUX_CTRL_WAY_SIZE_MASK;
1027 val |= (way_size_bits << L2C_AUX_CTRL_WAY_SIZE_SHIFT);
1028
1029 *aux_val &= ~mask;
1030 *aux_val |= val;
1031 *aux_mask &= ~mask;
1032
1033 return 0;
1034}
1035
1036static void __init l2x0_of_parse(const struct device_node *np,
1037 u32 *aux_val, u32 *aux_mask)
1038{
1039 u32 data[2] = { 0, 0 };
1040 u32 tag = 0;
1041 u32 dirty = 0;
1042 u32 val = 0, mask = 0;
1043 u32 assoc;
1044 int ret;
1045
1046 of_property_read_u32(np, "arm,tag-latency", &tag);
1047 if (tag) {
1048 mask |= L2X0_AUX_CTRL_TAG_LATENCY_MASK;
1049 val |= (tag - 1) << L2X0_AUX_CTRL_TAG_LATENCY_SHIFT;
1050 }
1051
1052 of_property_read_u32_array(np, "arm,data-latency",
1053 data, ARRAY_SIZE(data));
1054 if (data[0] && data[1]) {
1055 mask |= L2X0_AUX_CTRL_DATA_RD_LATENCY_MASK |
1056 L2X0_AUX_CTRL_DATA_WR_LATENCY_MASK;
1057 val |= ((data[0] - 1) << L2X0_AUX_CTRL_DATA_RD_LATENCY_SHIFT) |
1058 ((data[1] - 1) << L2X0_AUX_CTRL_DATA_WR_LATENCY_SHIFT);
1059 }
1060
1061 of_property_read_u32(np, "arm,dirty-latency", &dirty);
1062 if (dirty) {
1063 mask |= L2X0_AUX_CTRL_DIRTY_LATENCY_MASK;
1064 val |= (dirty - 1) << L2X0_AUX_CTRL_DIRTY_LATENCY_SHIFT;
1065 }
1066
1067 if (of_property_read_bool(np, "arm,parity-enable")) {
1068 mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1069 val |= L2C_AUX_CTRL_PARITY_ENABLE;
1070 } else if (of_property_read_bool(np, "arm,parity-disable")) {
1071 mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1072 }
1073
1074 if (of_property_read_bool(np, "arm,shared-override")) {
1075 mask &= ~L2C_AUX_CTRL_SHARED_OVERRIDE;
1076 val |= L2C_AUX_CTRL_SHARED_OVERRIDE;
1077 }
1078
1079 ret = l2x0_cache_size_of_parse(np, aux_val, aux_mask, &assoc, SZ_256K);
1080 if (ret)
1081 return;
1082
1083 if (assoc > 8) {
1084 pr_err("l2x0 of: cache setting yield too high associativity\n");
1085 pr_err("l2x0 of: %d calculated, max 8\n", assoc);
1086 } else {
1087 mask |= L2X0_AUX_CTRL_ASSOC_MASK;
1088 val |= (assoc << L2X0_AUX_CTRL_ASSOC_SHIFT);
1089 }
1090
1091 *aux_val &= ~mask;
1092 *aux_val |= val;
1093 *aux_mask &= ~mask;
1094}
1095
1096static const struct l2c_init_data of_l2c210_data __initconst = {
1097 .type = "L2C-210",
1098 .way_size_0 = SZ_8K,
1099 .num_lock = 1,
1100 .of_parse = l2x0_of_parse,
1101 .enable = l2c_enable,
1102 .save = l2c_save,
1103 .configure = l2c_configure,
1104 .unlock = l2c_unlock,
1105 .outer_cache = {
1106 .inv_range = l2c210_inv_range,
1107 .clean_range = l2c210_clean_range,
1108 .flush_range = l2c210_flush_range,
1109 .flush_all = l2c210_flush_all,
1110 .disable = l2c_disable,
1111 .sync = l2c210_sync,
1112 .resume = l2c_resume,
1113 },
1114};
1115
1116static const struct l2c_init_data of_l2c220_data __initconst = {
1117 .type = "L2C-220",
1118 .way_size_0 = SZ_8K,
1119 .num_lock = 1,
1120 .of_parse = l2x0_of_parse,
1121 .enable = l2c220_enable,
1122 .save = l2c_save,
1123 .configure = l2c_configure,
1124 .unlock = l2c220_unlock,
1125 .outer_cache = {
1126 .inv_range = l2c220_inv_range,
1127 .clean_range = l2c220_clean_range,
1128 .flush_range = l2c220_flush_range,
1129 .flush_all = l2c220_flush_all,
1130 .disable = l2c_disable,
1131 .sync = l2c220_sync,
1132 .resume = l2c_resume,
1133 },
1134};
1135
1136static void __init l2c310_of_parse(const struct device_node *np,
1137 u32 *aux_val, u32 *aux_mask)
1138{
1139 u32 data[3] = { 0, 0, 0 };
1140 u32 tag[3] = { 0, 0, 0 };
1141 u32 filter[2] = { 0, 0 };
1142 u32 assoc;
1143 u32 prefetch;
1144 u32 val;
1145 int ret;
1146
1147 of_property_read_u32_array(np, "arm,tag-latency", tag, ARRAY_SIZE(tag));
1148 if (tag[0] && tag[1] && tag[2])
1149 l2x0_saved_regs.tag_latency =
1150 L310_LATENCY_CTRL_RD(tag[0] - 1) |
1151 L310_LATENCY_CTRL_WR(tag[1] - 1) |
1152 L310_LATENCY_CTRL_SETUP(tag[2] - 1);
1153
1154 of_property_read_u32_array(np, "arm,data-latency",
1155 data, ARRAY_SIZE(data));
1156 if (data[0] && data[1] && data[2])
1157 l2x0_saved_regs.data_latency =
1158 L310_LATENCY_CTRL_RD(data[0] - 1) |
1159 L310_LATENCY_CTRL_WR(data[1] - 1) |
1160 L310_LATENCY_CTRL_SETUP(data[2] - 1);
1161
1162 of_property_read_u32_array(np, "arm,filter-ranges",
1163 filter, ARRAY_SIZE(filter));
1164 if (filter[1]) {
1165 l2x0_saved_regs.filter_end =
1166 ALIGN(filter[0] + filter[1], SZ_1M);
1167 l2x0_saved_regs.filter_start = (filter[0] & ~(SZ_1M - 1))
1168 | L310_ADDR_FILTER_EN;
1169 }
1170
1171 ret = l2x0_cache_size_of_parse(np, aux_val, aux_mask, &assoc, SZ_512K);
1172 if (!ret) {
1173 switch (assoc) {
1174 case 16:
1175 *aux_val &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1176 *aux_val |= L310_AUX_CTRL_ASSOCIATIVITY_16;
1177 *aux_mask &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1178 break;
1179 case 8:
1180 *aux_val &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1181 *aux_mask &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1182 break;
1183 default:
1184 pr_err("L2C-310 OF cache associativity %d invalid, only 8 or 16 permitted\n",
1185 assoc);
1186 break;
1187 }
1188 }
1189
1190 if (of_property_read_bool(np, "arm,shared-override")) {
1191 *aux_val |= L2C_AUX_CTRL_SHARED_OVERRIDE;
1192 *aux_mask &= ~L2C_AUX_CTRL_SHARED_OVERRIDE;
1193 }
1194
1195 if (of_property_read_bool(np, "arm,parity-enable")) {
1196 *aux_val |= L2C_AUX_CTRL_PARITY_ENABLE;
1197 *aux_mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1198 } else if (of_property_read_bool(np, "arm,parity-disable")) {
1199 *aux_val &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1200 *aux_mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1201 }
1202
1203 prefetch = l2x0_saved_regs.prefetch_ctrl;
1204
1205 ret = of_property_read_u32(np, "arm,double-linefill", &val);
1206 if (ret == 0) {
1207 if (val)
1208 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL;
1209 else
1210 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL;
1211 } else if (ret != -EINVAL) {
1212 pr_err("L2C-310 OF arm,double-linefill property value is missing\n");
1213 }
1214
1215 ret = of_property_read_u32(np, "arm,double-linefill-incr", &val);
1216 if (ret == 0) {
1217 if (val)
1218 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL_INCR;
1219 else
1220 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL_INCR;
1221 } else if (ret != -EINVAL) {
1222 pr_err("L2C-310 OF arm,double-linefill-incr property value is missing\n");
1223 }
1224
1225 ret = of_property_read_u32(np, "arm,double-linefill-wrap", &val);
1226 if (ret == 0) {
1227 if (!val)
1228 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL_WRAP;
1229 else
1230 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL_WRAP;
1231 } else if (ret != -EINVAL) {
1232 pr_err("L2C-310 OF arm,double-linefill-wrap property value is missing\n");
1233 }
1234
1235 ret = of_property_read_u32(np, "arm,prefetch-drop", &val);
1236 if (ret == 0) {
1237 if (val)
1238 prefetch |= L310_PREFETCH_CTRL_PREFETCH_DROP;
1239 else
1240 prefetch &= ~L310_PREFETCH_CTRL_PREFETCH_DROP;
1241 } else if (ret != -EINVAL) {
1242 pr_err("L2C-310 OF arm,prefetch-drop property value is missing\n");
1243 }
1244
1245 ret = of_property_read_u32(np, "arm,prefetch-offset", &val);
1246 if (ret == 0) {
1247 prefetch &= ~L310_PREFETCH_CTRL_OFFSET_MASK;
1248 prefetch |= val & L310_PREFETCH_CTRL_OFFSET_MASK;
1249 } else if (ret != -EINVAL) {
1250 pr_err("L2C-310 OF arm,prefetch-offset property value is missing\n");
1251 }
1252
1253 ret = of_property_read_u32(np, "prefetch-data", &val);
1254 if (ret == 0) {
1255 if (val)
1256 prefetch |= L310_PREFETCH_CTRL_DATA_PREFETCH;
1257 else
1258 prefetch &= ~L310_PREFETCH_CTRL_DATA_PREFETCH;
1259 } else if (ret != -EINVAL) {
1260 pr_err("L2C-310 OF prefetch-data property value is missing\n");
1261 }
1262
1263 ret = of_property_read_u32(np, "prefetch-instr", &val);
1264 if (ret == 0) {
1265 if (val)
1266 prefetch |= L310_PREFETCH_CTRL_INSTR_PREFETCH;
1267 else
1268 prefetch &= ~L310_PREFETCH_CTRL_INSTR_PREFETCH;
1269 } else if (ret != -EINVAL) {
1270 pr_err("L2C-310 OF prefetch-instr property value is missing\n");
1271 }
1272
1273 l2x0_saved_regs.prefetch_ctrl = prefetch;
1274}
1275
1276static const struct l2c_init_data of_l2c310_data __initconst = {
1277 .type = "L2C-310",
1278 .way_size_0 = SZ_8K,
1279 .num_lock = 8,
1280 .of_parse = l2c310_of_parse,
1281 .enable = l2c310_enable,
1282 .fixup = l2c310_fixup,
1283 .save = l2c310_save,
1284 .configure = l2c310_configure,
1285 .unlock = l2c310_unlock,
1286 .outer_cache = {
1287 .inv_range = l2c210_inv_range,
1288 .clean_range = l2c210_clean_range,
1289 .flush_range = l2c210_flush_range,
1290 .flush_all = l2c210_flush_all,
1291 .disable = l2c310_disable,
1292 .sync = l2c210_sync,
1293 .resume = l2c310_resume,
1294 },
1295};
1296
1297/*
1298 * This is a variant of the of_l2c310_data with .sync set to
1299 * NULL. Outer sync operations are not needed when the system is I/O
1300 * coherent, and potentially harmful in certain situations (PCIe/PL310
1301 * deadlock on Armada 375/38x due to hardware I/O coherency). The
1302 * other operations are kept because they are infrequent (therefore do
1303 * not cause the deadlock in practice) and needed for secondary CPU
1304 * boot and other power management activities.
1305 */
1306static const struct l2c_init_data of_l2c310_coherent_data __initconst = {
1307 .type = "L2C-310 Coherent",
1308 .way_size_0 = SZ_8K,
1309 .num_lock = 8,
1310 .of_parse = l2c310_of_parse,
1311 .enable = l2c310_enable,
1312 .fixup = l2c310_fixup,
1313 .save = l2c310_save,
1314 .configure = l2c310_configure,
1315 .unlock = l2c310_unlock,
1316 .outer_cache = {
1317 .inv_range = l2c210_inv_range,
1318 .clean_range = l2c210_clean_range,
1319 .flush_range = l2c210_flush_range,
1320 .flush_all = l2c210_flush_all,
1321 .disable = l2c310_disable,
1322 .resume = l2c310_resume,
1323 },
1324};
1325
1326/*
1327 * Note that the end addresses passed to Linux primitives are
1328 * noninclusive, while the hardware cache range operations use
1329 * inclusive start and end addresses.
1330 */
1331static unsigned long aurora_range_end(unsigned long start, unsigned long end)
1332{
1333 /*
1334 * Limit the number of cache lines processed at once,
1335 * since cache range operations stall the CPU pipeline
1336 * until completion.
1337 */
1338 if (end > start + MAX_RANGE_SIZE)
1339 end = start + MAX_RANGE_SIZE;
1340
1341 /*
1342 * Cache range operations can't straddle a page boundary.
1343 */
1344 if (end > PAGE_ALIGN(start+1))
1345 end = PAGE_ALIGN(start+1);
1346
1347 return end;
1348}
1349
1350static void aurora_pa_range(unsigned long start, unsigned long end,
1351 unsigned long offset)
1352{
1353 void __iomem *base = l2x0_base;
1354 unsigned long range_end;
1355 unsigned long flags;
1356
1357 /*
1358 * round start and end adresses up to cache line size
1359 */
1360 start &= ~(CACHE_LINE_SIZE - 1);
1361 end = ALIGN(end, CACHE_LINE_SIZE);
1362
1363 /*
1364 * perform operation on all full cache lines between 'start' and 'end'
1365 */
1366 while (start < end) {
1367 range_end = aurora_range_end(start, end);
1368
1369 raw_spin_lock_irqsave(&l2x0_lock, flags);
1370 writel_relaxed(start, base + AURORA_RANGE_BASE_ADDR_REG);
1371 writel_relaxed(range_end - CACHE_LINE_SIZE, base + offset);
1372 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1373
1374 writel_relaxed(0, base + AURORA_SYNC_REG);
1375 start = range_end;
1376 }
1377}
1378static void aurora_inv_range(unsigned long start, unsigned long end)
1379{
1380 aurora_pa_range(start, end, AURORA_INVAL_RANGE_REG);
1381}
1382
1383static void aurora_clean_range(unsigned long start, unsigned long end)
1384{
1385 /*
1386 * If L2 is forced to WT, the L2 will always be clean and we
1387 * don't need to do anything here.
1388 */
1389 if (!l2_wt_override)
1390 aurora_pa_range(start, end, AURORA_CLEAN_RANGE_REG);
1391}
1392
1393static void aurora_flush_range(unsigned long start, unsigned long end)
1394{
1395 if (l2_wt_override)
1396 aurora_pa_range(start, end, AURORA_INVAL_RANGE_REG);
1397 else
1398 aurora_pa_range(start, end, AURORA_FLUSH_RANGE_REG);
1399}
1400
1401static void aurora_flush_all(void)
1402{
1403 void __iomem *base = l2x0_base;
1404 unsigned long flags;
1405
1406 /* clean all ways */
1407 raw_spin_lock_irqsave(&l2x0_lock, flags);
1408 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
1409 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1410
1411 writel_relaxed(0, base + AURORA_SYNC_REG);
1412}
1413
1414static void aurora_cache_sync(void)
1415{
1416 writel_relaxed(0, l2x0_base + AURORA_SYNC_REG);
1417}
1418
1419static void aurora_disable(void)
1420{
1421 void __iomem *base = l2x0_base;
1422 unsigned long flags;
1423
1424 raw_spin_lock_irqsave(&l2x0_lock, flags);
1425 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
1426 writel_relaxed(0, base + AURORA_SYNC_REG);
1427 l2c_write_sec(0, base, L2X0_CTRL);
1428 dsb(st);
1429 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1430}
1431
1432static void aurora_save(void __iomem *base)
1433{
1434 l2x0_saved_regs.ctrl = readl_relaxed(base + L2X0_CTRL);
1435 l2x0_saved_regs.aux_ctrl = readl_relaxed(base + L2X0_AUX_CTRL);
1436}
1437
1438/*
1439 * For Aurora cache in no outer mode, enable via the CP15 coprocessor
1440 * broadcasting of cache commands to L2.
1441 */
1442static void __init aurora_enable_no_outer(void __iomem *base,
1443 unsigned num_lock)
1444{
1445 u32 u;
1446
1447 asm volatile("mrc p15, 1, %0, c15, c2, 0" : "=r" (u));
1448 u |= AURORA_CTRL_FW; /* Set the FW bit */
1449 asm volatile("mcr p15, 1, %0, c15, c2, 0" : : "r" (u));
1450
1451 isb();
1452
1453 l2c_enable(base, num_lock);
1454}
1455
1456static void __init aurora_fixup(void __iomem *base, u32 cache_id,
1457 struct outer_cache_fns *fns)
1458{
1459 sync_reg_offset = AURORA_SYNC_REG;
1460}
1461
1462static void __init aurora_of_parse(const struct device_node *np,
1463 u32 *aux_val, u32 *aux_mask)
1464{
1465 u32 val = AURORA_ACR_REPLACEMENT_TYPE_SEMIPLRU;
1466 u32 mask = AURORA_ACR_REPLACEMENT_MASK;
1467
1468 of_property_read_u32(np, "cache-id-part",
1469 &cache_id_part_number_from_dt);
1470
1471 /* Determine and save the write policy */
1472 l2_wt_override = of_property_read_bool(np, "wt-override");
1473
1474 if (l2_wt_override) {
1475 val |= AURORA_ACR_FORCE_WRITE_THRO_POLICY;
1476 mask |= AURORA_ACR_FORCE_WRITE_POLICY_MASK;
1477 }
1478
1479 *aux_val &= ~mask;
1480 *aux_val |= val;
1481 *aux_mask &= ~mask;
1482}
1483
1484static const struct l2c_init_data of_aurora_with_outer_data __initconst = {
1485 .type = "Aurora",
1486 .way_size_0 = SZ_4K,
1487 .num_lock = 4,
1488 .of_parse = aurora_of_parse,
1489 .enable = l2c_enable,
1490 .fixup = aurora_fixup,
1491 .save = aurora_save,
1492 .configure = l2c_configure,
1493 .unlock = l2c_unlock,
1494 .outer_cache = {
1495 .inv_range = aurora_inv_range,
1496 .clean_range = aurora_clean_range,
1497 .flush_range = aurora_flush_range,
1498 .flush_all = aurora_flush_all,
1499 .disable = aurora_disable,
1500 .sync = aurora_cache_sync,
1501 .resume = l2c_resume,
1502 },
1503};
1504
1505static const struct l2c_init_data of_aurora_no_outer_data __initconst = {
1506 .type = "Aurora",
1507 .way_size_0 = SZ_4K,
1508 .num_lock = 4,
1509 .of_parse = aurora_of_parse,
1510 .enable = aurora_enable_no_outer,
1511 .fixup = aurora_fixup,
1512 .save = aurora_save,
1513 .configure = l2c_configure,
1514 .unlock = l2c_unlock,
1515 .outer_cache = {
1516 .resume = l2c_resume,
1517 },
1518};
1519
1520/*
1521 * For certain Broadcom SoCs, depending on the address range, different offsets
1522 * need to be added to the address before passing it to L2 for
1523 * invalidation/clean/flush
1524 *
1525 * Section Address Range Offset EMI
1526 * 1 0x00000000 - 0x3FFFFFFF 0x80000000 VC
1527 * 2 0x40000000 - 0xBFFFFFFF 0x40000000 SYS
1528 * 3 0xC0000000 - 0xFFFFFFFF 0x80000000 VC
1529 *
1530 * When the start and end addresses have crossed two different sections, we
1531 * need to break the L2 operation into two, each within its own section.
1532 * For example, if we need to invalidate addresses starts at 0xBFFF0000 and
1533 * ends at 0xC0001000, we need do invalidate 1) 0xBFFF0000 - 0xBFFFFFFF and 2)
1534 * 0xC0000000 - 0xC0001000
1535 *
1536 * Note 1:
1537 * By breaking a single L2 operation into two, we may potentially suffer some
1538 * performance hit, but keep in mind the cross section case is very rare
1539 *
1540 * Note 2:
1541 * We do not need to handle the case when the start address is in
1542 * Section 1 and the end address is in Section 3, since it is not a valid use
1543 * case
1544 *
1545 * Note 3:
1546 * Section 1 in practical terms can no longer be used on rev A2. Because of
1547 * that the code does not need to handle section 1 at all.
1548 *
1549 */
1550#define BCM_SYS_EMI_START_ADDR 0x40000000UL
1551#define BCM_VC_EMI_SEC3_START_ADDR 0xC0000000UL
1552
1553#define BCM_SYS_EMI_OFFSET 0x40000000UL
1554#define BCM_VC_EMI_OFFSET 0x80000000UL
1555
1556static inline int bcm_addr_is_sys_emi(unsigned long addr)
1557{
1558 return (addr >= BCM_SYS_EMI_START_ADDR) &&
1559 (addr < BCM_VC_EMI_SEC3_START_ADDR);
1560}
1561
1562static inline unsigned long bcm_l2_phys_addr(unsigned long addr)
1563{
1564 if (bcm_addr_is_sys_emi(addr))
1565 return addr + BCM_SYS_EMI_OFFSET;
1566 else
1567 return addr + BCM_VC_EMI_OFFSET;
1568}
1569
1570static void bcm_inv_range(unsigned long start, unsigned long end)
1571{
1572 unsigned long new_start, new_end;
1573
1574 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1575
1576 if (unlikely(end <= start))
1577 return;
1578
1579 new_start = bcm_l2_phys_addr(start);
1580 new_end = bcm_l2_phys_addr(end);
1581
1582 /* normal case, no cross section between start and end */
1583 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1584 l2c210_inv_range(new_start, new_end);
1585 return;
1586 }
1587
1588 /* They cross sections, so it can only be a cross from section
1589 * 2 to section 3
1590 */
1591 l2c210_inv_range(new_start,
1592 bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1593 l2c210_inv_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1594 new_end);
1595}
1596
1597static void bcm_clean_range(unsigned long start, unsigned long end)
1598{
1599 unsigned long new_start, new_end;
1600
1601 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1602
1603 if (unlikely(end <= start))
1604 return;
1605
1606 new_start = bcm_l2_phys_addr(start);
1607 new_end = bcm_l2_phys_addr(end);
1608
1609 /* normal case, no cross section between start and end */
1610 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1611 l2c210_clean_range(new_start, new_end);
1612 return;
1613 }
1614
1615 /* They cross sections, so it can only be a cross from section
1616 * 2 to section 3
1617 */
1618 l2c210_clean_range(new_start,
1619 bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1620 l2c210_clean_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1621 new_end);
1622}
1623
1624static void bcm_flush_range(unsigned long start, unsigned long end)
1625{
1626 unsigned long new_start, new_end;
1627
1628 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1629
1630 if (unlikely(end <= start))
1631 return;
1632
1633 if ((end - start) >= l2x0_size) {
1634 outer_cache.flush_all();
1635 return;
1636 }
1637
1638 new_start = bcm_l2_phys_addr(start);
1639 new_end = bcm_l2_phys_addr(end);
1640
1641 /* normal case, no cross section between start and end */
1642 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1643 l2c210_flush_range(new_start, new_end);
1644 return;
1645 }
1646
1647 /* They cross sections, so it can only be a cross from section
1648 * 2 to section 3
1649 */
1650 l2c210_flush_range(new_start,
1651 bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1652 l2c210_flush_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1653 new_end);
1654}
1655
1656/* Broadcom L2C-310 start from ARMs R3P2 or later, and require no fixups */
1657static const struct l2c_init_data of_bcm_l2x0_data __initconst = {
1658 .type = "BCM-L2C-310",
1659 .way_size_0 = SZ_8K,
1660 .num_lock = 8,
1661 .of_parse = l2c310_of_parse,
1662 .enable = l2c310_enable,
1663 .save = l2c310_save,
1664 .configure = l2c310_configure,
1665 .unlock = l2c310_unlock,
1666 .outer_cache = {
1667 .inv_range = bcm_inv_range,
1668 .clean_range = bcm_clean_range,
1669 .flush_range = bcm_flush_range,
1670 .flush_all = l2c210_flush_all,
1671 .disable = l2c310_disable,
1672 .sync = l2c210_sync,
1673 .resume = l2c310_resume,
1674 },
1675};
1676
1677static void __init tauros3_save(void __iomem *base)
1678{
1679 l2c_save(base);
1680
1681 l2x0_saved_regs.aux2_ctrl =
1682 readl_relaxed(base + TAUROS3_AUX2_CTRL);
1683 l2x0_saved_regs.prefetch_ctrl =
1684 readl_relaxed(base + L310_PREFETCH_CTRL);
1685}
1686
1687static void tauros3_configure(void __iomem *base)
1688{
1689 l2c_configure(base);
1690 writel_relaxed(l2x0_saved_regs.aux2_ctrl,
1691 base + TAUROS3_AUX2_CTRL);
1692 writel_relaxed(l2x0_saved_regs.prefetch_ctrl,
1693 base + L310_PREFETCH_CTRL);
1694}
1695
1696static const struct l2c_init_data of_tauros3_data __initconst = {
1697 .type = "Tauros3",
1698 .way_size_0 = SZ_8K,
1699 .num_lock = 8,
1700 .enable = l2c_enable,
1701 .save = tauros3_save,
1702 .configure = tauros3_configure,
1703 .unlock = l2c_unlock,
1704 /* Tauros3 broadcasts L1 cache operations to L2 */
1705 .outer_cache = {
1706 .resume = l2c_resume,
1707 },
1708};
1709
1710#define L2C_ID(name, fns) { .compatible = name, .data = (void *)&fns }
1711static const struct of_device_id l2x0_ids[] __initconst = {
1712 L2C_ID("arm,l210-cache", of_l2c210_data),
1713 L2C_ID("arm,l220-cache", of_l2c220_data),
1714 L2C_ID("arm,pl310-cache", of_l2c310_data),
1715 L2C_ID("brcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1716 L2C_ID("marvell,aurora-outer-cache", of_aurora_with_outer_data),
1717 L2C_ID("marvell,aurora-system-cache", of_aurora_no_outer_data),
1718 L2C_ID("marvell,tauros3-cache", of_tauros3_data),
1719 /* Deprecated IDs */
1720 L2C_ID("bcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1721 {}
1722};
1723
1724int __init l2x0_of_init(u32 aux_val, u32 aux_mask)
1725{
1726 const struct l2c_init_data *data;
1727 struct device_node *np;
1728 struct resource res;
1729 u32 cache_id, old_aux;
1730 u32 cache_level = 2;
1731 bool nosync = false;
1732
1733 np = of_find_matching_node(NULL, l2x0_ids);
1734 if (!np)
1735 return -ENODEV;
1736
1737 if (of_address_to_resource(np, 0, &res))
1738 return -ENODEV;
1739
1740 l2x0_base = ioremap(res.start, resource_size(&res));
1741 if (!l2x0_base)
1742 return -ENOMEM;
1743
1744 l2x0_saved_regs.phy_base = res.start;
1745
1746 data = of_match_node(l2x0_ids, np)->data;
1747
1748 if (of_device_is_compatible(np, "arm,pl310-cache") &&
1749 of_property_read_bool(np, "arm,io-coherent"))
1750 data = &of_l2c310_coherent_data;
1751
1752 old_aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
1753 if (old_aux != ((old_aux & aux_mask) | aux_val)) {
1754 pr_warn("L2C: platform modifies aux control register: 0x%08x -> 0x%08x\n",
1755 old_aux, (old_aux & aux_mask) | aux_val);
1756 } else if (aux_mask != ~0U && aux_val != 0) {
1757 pr_alert("L2C: platform provided aux values match the hardware, so have no effect. Please remove them.\n");
1758 }
1759
1760 /* All L2 caches are unified, so this property should be specified */
1761 if (!of_property_read_bool(np, "cache-unified"))
1762 pr_err("L2C: device tree omits to specify unified cache\n");
1763
1764 if (of_property_read_u32(np, "cache-level", &cache_level))
1765 pr_err("L2C: device tree omits to specify cache-level\n");
1766
1767 if (cache_level != 2)
1768 pr_err("L2C: device tree specifies invalid cache level\n");
1769
1770 nosync = of_property_read_bool(np, "arm,outer-sync-disable");
1771
1772 /* Read back current (default) hardware configuration */
1773 if (data->save)
1774 data->save(l2x0_base);
1775
1776 /* L2 configuration can only be changed if the cache is disabled */
1777 if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN))
1778 if (data->of_parse)
1779 data->of_parse(np, &aux_val, &aux_mask);
1780
1781 if (cache_id_part_number_from_dt)
1782 cache_id = cache_id_part_number_from_dt;
1783 else
1784 cache_id = readl_relaxed(l2x0_base + L2X0_CACHE_ID);
1785
1786 return __l2c_init(data, aux_val, aux_mask, cache_id, nosync);
1787}
1788#endif