1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * arch/arm/mm/cache-l2x0.c - L210/L220/L310 cache controller support
4 *
5 * Copyright (C) 2007 ARM Limited
6 */
7#include <linux/cpu.h>
8#include <linux/err.h>
9#include <linux/init.h>
10#include <linux/smp.h>
11#include <linux/spinlock.h>
12#include <linux/log2.h>
13#include <linux/io.h>
14#include <linux/of.h>
15#include <linux/of_address.h>
16
17#include <asm/cacheflush.h>
18#include <asm/cp15.h>
19#include <asm/cputype.h>
20#include <asm/hardware/cache-l2x0.h>
21#include <asm/hardware/cache-aurora-l2.h>
22#include "cache-tauros3.h"
23
24struct l2c_init_data {
25 const char *type;
26 unsigned way_size_0;
27 unsigned num_lock;
28 void (*of_parse)(const struct device_node *, u32 *, u32 *);
29 void (*enable)(void __iomem *, unsigned);
30 void (*fixup)(void __iomem *, u32, struct outer_cache_fns *);
31 void (*save)(void __iomem *);
32 void (*configure)(void __iomem *);
33 void (*unlock)(void __iomem *, unsigned);
34 struct outer_cache_fns outer_cache;
35};
36
37#define CACHE_LINE_SIZE 32
38
39static void __iomem *l2x0_base;
40static const struct l2c_init_data *l2x0_data;
41static DEFINE_RAW_SPINLOCK(l2x0_lock);
42static u32 l2x0_way_mask; /* Bitmask of active ways */
43static u32 l2x0_size;
44static unsigned long sync_reg_offset = L2X0_CACHE_SYNC;
45
46struct l2x0_regs l2x0_saved_regs;
47
48static bool l2x0_bresp_disable;
49static bool l2x0_flz_disable;
50
51/*
52 * Common code for all cache controllers.
53 */
54static inline void l2c_wait_mask(void __iomem *reg, unsigned long mask)
55{
56 /* wait for cache operation by line or way to complete */
57 while (readl_relaxed(reg) & mask)
58 cpu_relax();
59}
60
61/*
62 * By default, we write directly to secure registers. Platforms must
63 * override this if they are running non-secure.
64 */
65static void l2c_write_sec(unsigned long val, void __iomem *base, unsigned reg)
66{
67 if (val == readl_relaxed(base + reg))
68 return;
69 if (outer_cache.write_sec)
70 outer_cache.write_sec(val, reg);
71 else
72 writel_relaxed(val, base + reg);
73}
74
75/*
76 * This should only be called when we have a requirement that the
77 * register be written due to a work-around, as platforms running
78 * in non-secure mode may not be able to access this register.
79 */
80static inline void l2c_set_debug(void __iomem *base, unsigned long val)
81{
82 l2c_write_sec(val, base, reg: L2X0_DEBUG_CTRL);
83}
84
85static void __l2c_op_way(void __iomem *reg)
86{
87 writel_relaxed(l2x0_way_mask, reg);
88 l2c_wait_mask(reg, mask: l2x0_way_mask);
89}
90
91static inline void l2c_unlock(void __iomem *base, unsigned num)
92{
93 unsigned i;
94
95 for (i = 0; i < num; i++) {
96 writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_D_BASE +
97 i * L2X0_LOCKDOWN_STRIDE);
98 writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_I_BASE +
99 i * L2X0_LOCKDOWN_STRIDE);
100 }
101}
102
103static void l2c_configure(void __iomem *base)
104{
105 l2c_write_sec(val: l2x0_saved_regs.aux_ctrl, base, reg: L2X0_AUX_CTRL);
106}
107
108/*
109 * Enable the L2 cache controller. This function must only be
110 * called when the cache controller is known to be disabled.
111 */
112static void l2c_enable(void __iomem *base, unsigned num_lock)
113{
114 unsigned long flags;
115
116 if (outer_cache.configure)
117 outer_cache.configure(&l2x0_saved_regs);
118 else
119 l2x0_data->configure(base);
120
121 l2x0_data->unlock(base, num_lock);
122
123 local_irq_save(flags);
124 __l2c_op_way(reg: base + L2X0_INV_WAY);
125 writel_relaxed(0, base + sync_reg_offset);
126 l2c_wait_mask(reg: base + sync_reg_offset, mask: 1);
127 local_irq_restore(flags);
128
129 l2c_write_sec(val: L2X0_CTRL_EN, base, reg: L2X0_CTRL);
130}
131
132static void l2c_disable(void)
133{
134 void __iomem *base = l2x0_base;
135
136 l2x0_pmu_suspend();
137
138 outer_cache.flush_all();
139 l2c_write_sec(val: 0, base, reg: L2X0_CTRL);
140 dsb(st);
141}
142
143static void l2c_save(void __iomem *base)
144{
145 l2x0_saved_regs.aux_ctrl = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
146}
147
148static void l2c_resume(void)
149{
150 void __iomem *base = l2x0_base;
151
152 /* Do not touch the controller if already enabled. */
153 if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN))
154 l2c_enable(base, num_lock: l2x0_data->num_lock);
155
156 l2x0_pmu_resume();
157}
158
159/*
160 * L2C-210 specific code.
161 *
162 * The L2C-2x0 PA, set/way and sync operations are atomic, but we must
163 * ensure that no background operation is running. The way operations
164 * are all background tasks.
165 *
166 * While a background operation is in progress, any new operation is
167 * ignored (unspecified whether this causes an error.) Thankfully, not
168 * used on SMP.
169 *
170 * Never has a different sync register other than L2X0_CACHE_SYNC, but
171 * we use sync_reg_offset here so we can share some of this with L2C-310.
172 */
173static void __l2c210_cache_sync(void __iomem *base)
174{
175 writel_relaxed(0, base + sync_reg_offset);
176}
177
178static void __l2c210_op_pa_range(void __iomem *reg, unsigned long start,
179 unsigned long end)
180{
181 while (start < end) {
182 writel_relaxed(start, reg);
183 start += CACHE_LINE_SIZE;
184 }
185}
186
187static void l2c210_inv_range(unsigned long start, unsigned long end)
188{
189 void __iomem *base = l2x0_base;
190
191 if (start & (CACHE_LINE_SIZE - 1)) {
192 start &= ~(CACHE_LINE_SIZE - 1);
193 writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
194 start += CACHE_LINE_SIZE;
195 }
196
197 if (end & (CACHE_LINE_SIZE - 1)) {
198 end &= ~(CACHE_LINE_SIZE - 1);
199 writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
200 }
201
202 __l2c210_op_pa_range(reg: base + L2X0_INV_LINE_PA, start, end);
203 __l2c210_cache_sync(base);
204}
205
206static void l2c210_clean_range(unsigned long start, unsigned long end)
207{
208 void __iomem *base = l2x0_base;
209
210 start &= ~(CACHE_LINE_SIZE - 1);
211 __l2c210_op_pa_range(reg: base + L2X0_CLEAN_LINE_PA, start, end);
212 __l2c210_cache_sync(base);
213}
214
215static void l2c210_flush_range(unsigned long start, unsigned long end)
216{
217 void __iomem *base = l2x0_base;
218
219 start &= ~(CACHE_LINE_SIZE - 1);
220 __l2c210_op_pa_range(reg: base + L2X0_CLEAN_INV_LINE_PA, start, end);
221 __l2c210_cache_sync(base);
222}
223
224static void l2c210_flush_all(void)
225{
226 void __iomem *base = l2x0_base;
227
228 BUG_ON(!irqs_disabled());
229
230 __l2c_op_way(reg: base + L2X0_CLEAN_INV_WAY);
231 __l2c210_cache_sync(base);
232}
233
234static void l2c210_sync(void)
235{
236 __l2c210_cache_sync(base: l2x0_base);
237}
238
239static const struct l2c_init_data l2c210_data __initconst = {
240 .type = "L2C-210",
241 .way_size_0 = SZ_8K,
242 .num_lock = 1,
243 .enable = l2c_enable,
244 .save = l2c_save,
245 .configure = l2c_configure,
246 .unlock = l2c_unlock,
247 .outer_cache = {
248 .inv_range = l2c210_inv_range,
249 .clean_range = l2c210_clean_range,
250 .flush_range = l2c210_flush_range,
251 .flush_all = l2c210_flush_all,
252 .disable = l2c_disable,
253 .sync = l2c210_sync,
254 .resume = l2c_resume,
255 },
256};
257
258/*
259 * L2C-220 specific code.
260 *
261 * All operations are background operations: they have to be waited for.
262 * Conflicting requests generate a slave error (which will cause an
263 * imprecise abort.) Never uses sync_reg_offset, so we hard-code the
264 * sync register here.
265 *
266 * However, we can re-use the l2c210_resume call.
267 */
268static inline void __l2c220_cache_sync(void __iomem *base)
269{
270 writel_relaxed(0, base + L2X0_CACHE_SYNC);
271 l2c_wait_mask(reg: base + L2X0_CACHE_SYNC, mask: 1);
272}
273
274static void l2c220_op_way(void __iomem *base, unsigned reg)
275{
276 unsigned long flags;
277
278 raw_spin_lock_irqsave(&l2x0_lock, flags);
279 __l2c_op_way(reg: base + reg);
280 __l2c220_cache_sync(base);
281 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
282}
283
284static unsigned long l2c220_op_pa_range(void __iomem *reg, unsigned long start,
285 unsigned long end, unsigned long flags)
286{
287 raw_spinlock_t *lock = &l2x0_lock;
288
289 while (start < end) {
290 unsigned long blk_end = start + min(end - start, 4096UL);
291
292 while (start < blk_end) {
293 l2c_wait_mask(reg, mask: 1);
294 writel_relaxed(start, reg);
295 start += CACHE_LINE_SIZE;
296 }
297
298 if (blk_end < end) {
299 raw_spin_unlock_irqrestore(lock, flags);
300 raw_spin_lock_irqsave(lock, flags);
301 }
302 }
303
304 return flags;
305}
306
307static void l2c220_inv_range(unsigned long start, unsigned long end)
308{
309 void __iomem *base = l2x0_base;
310 unsigned long flags;
311
312 raw_spin_lock_irqsave(&l2x0_lock, flags);
313 if ((start | end) & (CACHE_LINE_SIZE - 1)) {
314 if (start & (CACHE_LINE_SIZE - 1)) {
315 start &= ~(CACHE_LINE_SIZE - 1);
316 writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA);
317 start += CACHE_LINE_SIZE;
318 }
319
320 if (end & (CACHE_LINE_SIZE - 1)) {
321 end &= ~(CACHE_LINE_SIZE - 1);
322 l2c_wait_mask(reg: base + L2X0_CLEAN_INV_LINE_PA, mask: 1);
323 writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA);
324 }
325 }
326
327 flags = l2c220_op_pa_range(reg: base + L2X0_INV_LINE_PA,
328 start, end, flags);
329 l2c_wait_mask(reg: base + L2X0_INV_LINE_PA, mask: 1);
330 __l2c220_cache_sync(base);
331 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
332}
333
334static void l2c220_clean_range(unsigned long start, unsigned long end)
335{
336 void __iomem *base = l2x0_base;
337 unsigned long flags;
338
339 start &= ~(CACHE_LINE_SIZE - 1);
340 if ((end - start) >= l2x0_size) {
341 l2c220_op_way(base, reg: L2X0_CLEAN_WAY);
342 return;
343 }
344
345 raw_spin_lock_irqsave(&l2x0_lock, flags);
346 flags = l2c220_op_pa_range(reg: base + L2X0_CLEAN_LINE_PA,
347 start, end, flags);
348 l2c_wait_mask(reg: base + L2X0_CLEAN_INV_LINE_PA, mask: 1);
349 __l2c220_cache_sync(base);
350 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
351}
352
353static void l2c220_flush_range(unsigned long start, unsigned long end)
354{
355 void __iomem *base = l2x0_base;
356 unsigned long flags;
357
358 start &= ~(CACHE_LINE_SIZE - 1);
359 if ((end - start) >= l2x0_size) {
360 l2c220_op_way(base, reg: L2X0_CLEAN_INV_WAY);
361 return;
362 }
363
364 raw_spin_lock_irqsave(&l2x0_lock, flags);
365 flags = l2c220_op_pa_range(reg: base + L2X0_CLEAN_INV_LINE_PA,
366 start, end, flags);
367 l2c_wait_mask(reg: base + L2X0_CLEAN_INV_LINE_PA, mask: 1);
368 __l2c220_cache_sync(base);
369 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
370}
371
372static void l2c220_flush_all(void)
373{
374 l2c220_op_way(base: l2x0_base, reg: L2X0_CLEAN_INV_WAY);
375}
376
377static void l2c220_sync(void)
378{
379 unsigned long flags;
380
381 raw_spin_lock_irqsave(&l2x0_lock, flags);
382 __l2c220_cache_sync(base: l2x0_base);
383 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
384}
385
386static void l2c220_enable(void __iomem *base, unsigned num_lock)
387{
388 /*
389 * Always enable non-secure access to the lockdown registers -
390 * we write to them as part of the L2C enable sequence so they
391 * need to be accessible.
392 */
393 l2x0_saved_regs.aux_ctrl |= L220_AUX_CTRL_NS_LOCKDOWN;
394
395 l2c_enable(base, num_lock);
396}
397
398static void l2c220_unlock(void __iomem *base, unsigned num_lock)
399{
400 if (readl_relaxed(base + L2X0_AUX_CTRL) & L220_AUX_CTRL_NS_LOCKDOWN)
401 l2c_unlock(base, num: num_lock);
402}
403
404static const struct l2c_init_data l2c220_data = {
405 .type = "L2C-220",
406 .way_size_0 = SZ_8K,
407 .num_lock = 1,
408 .enable = l2c220_enable,
409 .save = l2c_save,
410 .configure = l2c_configure,
411 .unlock = l2c220_unlock,
412 .outer_cache = {
413 .inv_range = l2c220_inv_range,
414 .clean_range = l2c220_clean_range,
415 .flush_range = l2c220_flush_range,
416 .flush_all = l2c220_flush_all,
417 .disable = l2c_disable,
418 .sync = l2c220_sync,
419 .resume = l2c_resume,
420 },
421};
422
423/*
424 * L2C-310 specific code.
425 *
426 * Very similar to L2C-210, the PA, set/way and sync operations are atomic,
427 * and the way operations are all background tasks. However, issuing an
428 * operation while a background operation is in progress results in a
429 * SLVERR response. We can reuse:
430 *
431 * __l2c210_cache_sync (using sync_reg_offset)
432 * l2c210_sync
433 * l2c210_inv_range (if 588369 is not applicable)
434 * l2c210_clean_range
435 * l2c210_flush_range (if 588369 is not applicable)
436 * l2c210_flush_all (if 727915 is not applicable)
437 *
438 * Errata:
439 * 588369: PL310 R0P0->R1P0, fixed R2P0.
440 * Affects: all clean+invalidate operations
441 * clean and invalidate skips the invalidate step, so we need to issue
442 * separate operations. We also require the above debug workaround
443 * enclosing this code fragment on affected parts. On unaffected parts,
444 * we must not use this workaround without the debug register writes
445 * to avoid exposing a problem similar to 727915.
446 *
447 * 727915: PL310 R2P0->R3P0, fixed R3P1.
448 * Affects: clean+invalidate by way
449 * clean and invalidate by way runs in the background, and a store can
450 * hit the line between the clean operation and invalidate operation,
451 * resulting in the store being lost.
452 *
453 * 752271: PL310 R3P0->R3P1-50REL0, fixed R3P2.
454 * Affects: 8x64-bit (double fill) line fetches
455 * double fill line fetches can fail to cause dirty data to be evicted
456 * from the cache before the new data overwrites the second line.
457 *
458 * 753970: PL310 R3P0, fixed R3P1.
459 * Affects: sync
460 * prevents merging writes after the sync operation, until another L2C
461 * operation is performed (or a number of other conditions.)
462 *
463 * 769419: PL310 R0P0->R3P1, fixed R3P2.
464 * Affects: store buffer
465 * store buffer is not automatically drained.
466 */
467static void l2c310_inv_range_erratum(unsigned long start, unsigned long end)
468{
469 void __iomem *base = l2x0_base;
470
471 if ((start | end) & (CACHE_LINE_SIZE - 1)) {
472 unsigned long flags;
473
474 /* Erratum 588369 for both clean+invalidate operations */
475 raw_spin_lock_irqsave(&l2x0_lock, flags);
476 l2c_set_debug(base, val: 0x03);
477
478 if (start & (CACHE_LINE_SIZE - 1)) {
479 start &= ~(CACHE_LINE_SIZE - 1);
480 writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
481 writel_relaxed(start, base + L2X0_INV_LINE_PA);
482 start += CACHE_LINE_SIZE;
483 }
484
485 if (end & (CACHE_LINE_SIZE - 1)) {
486 end &= ~(CACHE_LINE_SIZE - 1);
487 writel_relaxed(end, base + L2X0_CLEAN_LINE_PA);
488 writel_relaxed(end, base + L2X0_INV_LINE_PA);
489 }
490
491 l2c_set_debug(base, val: 0x00);
492 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
493 }
494
495 __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end);
496 __l2c210_cache_sync(base);
497}
498
499static void l2c310_flush_range_erratum(unsigned long start, unsigned long end)
500{
501 raw_spinlock_t *lock = &l2x0_lock;
502 unsigned long flags;
503 void __iomem *base = l2x0_base;
504
505 raw_spin_lock_irqsave(lock, flags);
506 while (start < end) {
507 unsigned long blk_end = start + min(end - start, 4096UL);
508
509 l2c_set_debug(base, val: 0x03);
510 while (start < blk_end) {
511 writel_relaxed(start, base + L2X0_CLEAN_LINE_PA);
512 writel_relaxed(start, base + L2X0_INV_LINE_PA);
513 start += CACHE_LINE_SIZE;
514 }
515 l2c_set_debug(base, val: 0x00);
516
517 if (blk_end < end) {
518 raw_spin_unlock_irqrestore(lock, flags);
519 raw_spin_lock_irqsave(lock, flags);
520 }
521 }
522 raw_spin_unlock_irqrestore(lock, flags);
523 __l2c210_cache_sync(base);
524}
525
526static void l2c310_flush_all_erratum(void)
527{
528 void __iomem *base = l2x0_base;
529 unsigned long flags;
530
531 raw_spin_lock_irqsave(&l2x0_lock, flags);
532 l2c_set_debug(base, val: 0x03);
533 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
534 l2c_set_debug(base, val: 0x00);
535 __l2c210_cache_sync(base);
536 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
537}
538
539static void __init l2c310_save(void __iomem *base)
540{
541 unsigned revision;
542
543 l2c_save(base);
544
545 l2x0_saved_regs.tag_latency = readl_relaxed(base +
546 L310_TAG_LATENCY_CTRL);
547 l2x0_saved_regs.data_latency = readl_relaxed(base +
548 L310_DATA_LATENCY_CTRL);
549 l2x0_saved_regs.filter_end = readl_relaxed(base +
550 L310_ADDR_FILTER_END);
551 l2x0_saved_regs.filter_start = readl_relaxed(base +
552 L310_ADDR_FILTER_START);
553
554 revision = readl_relaxed(base + L2X0_CACHE_ID) &
555 L2X0_CACHE_ID_RTL_MASK;
556
557 /* From r2p0, there is Prefetch offset/control register */
558 if (revision >= L310_CACHE_ID_RTL_R2P0)
559 l2x0_saved_regs.prefetch_ctrl = readl_relaxed(base +
560 L310_PREFETCH_CTRL);
561
562 /* From r3p0, there is Power control register */
563 if (revision >= L310_CACHE_ID_RTL_R3P0)
564 l2x0_saved_regs.pwr_ctrl = readl_relaxed(base +
565 L310_POWER_CTRL);
566}
567
568static void l2c310_configure(void __iomem *base)
569{
570 unsigned revision;
571
572 l2c_configure(base);
573
574 /* restore pl310 setup */
575 l2c_write_sec(l2x0_saved_regs.tag_latency, base,
576 L310_TAG_LATENCY_CTRL);
577 l2c_write_sec(l2x0_saved_regs.data_latency, base,
578 L310_DATA_LATENCY_CTRL);
579 l2c_write_sec(l2x0_saved_regs.filter_end, base,
580 L310_ADDR_FILTER_END);
581 l2c_write_sec(l2x0_saved_regs.filter_start, base,
582 L310_ADDR_FILTER_START);
583
584 revision = readl_relaxed(base + L2X0_CACHE_ID) &
585 L2X0_CACHE_ID_RTL_MASK;
586
587 if (revision >= L310_CACHE_ID_RTL_R2P0)
588 l2c_write_sec(l2x0_saved_regs.prefetch_ctrl, base,
589 L310_PREFETCH_CTRL);
590 if (revision >= L310_CACHE_ID_RTL_R3P0)
591 l2c_write_sec(l2x0_saved_regs.pwr_ctrl, base,
592 L310_POWER_CTRL);
593}
594
595static int l2c310_starting_cpu(unsigned int cpu)
596{
597 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
598 return 0;
599}
600
601static int l2c310_dying_cpu(unsigned int cpu)
602{
603 set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1)));
604 return 0;
605}
606
607static void __init l2c310_enable(void __iomem *base, unsigned num_lock)
608{
609 unsigned rev = readl_relaxed(base + L2X0_CACHE_ID) & L2X0_CACHE_ID_RTL_MASK;
610 bool cortex_a9 = read_cpuid_part() == ARM_CPU_PART_CORTEX_A9;
611 u32 aux = l2x0_saved_regs.aux_ctrl;
612
613 if (rev >= L310_CACHE_ID_RTL_R2P0) {
614 if (cortex_a9 && !l2x0_bresp_disable) {
615 aux |= L310_AUX_CTRL_EARLY_BRESP;
616 pr_info("L2C-310 enabling early BRESP for Cortex-A9\n");
617 } else if (aux & L310_AUX_CTRL_EARLY_BRESP) {
618 pr_warn("L2C-310 early BRESP only supported with Cortex-A9\n");
619 aux &= ~L310_AUX_CTRL_EARLY_BRESP;
620 }
621 }
622
623 if (cortex_a9 && !l2x0_flz_disable) {
624 u32 aux_cur = readl_relaxed(base + L2X0_AUX_CTRL);
625 u32 acr = get_auxcr();
626
627 pr_debug("Cortex-A9 ACR=0x%08x\n", acr);
628
629 if (acr & BIT(3) && !(aux_cur & L310_AUX_CTRL_FULL_LINE_ZERO))
630 pr_err("L2C-310: full line of zeros enabled in Cortex-A9 but not L2C-310 - invalid\n");
631
632 if (aux & L310_AUX_CTRL_FULL_LINE_ZERO && !(acr & BIT(3)))
633 pr_err("L2C-310: enabling full line of zeros but not enabled in Cortex-A9\n");
634
635 if (!(aux & L310_AUX_CTRL_FULL_LINE_ZERO) && !outer_cache.write_sec) {
636 aux |= L310_AUX_CTRL_FULL_LINE_ZERO;
637 pr_info("L2C-310 full line of zeros enabled for Cortex-A9\n");
638 }
639 } else if (aux & (L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP)) {
640 pr_err("L2C-310: disabling Cortex-A9 specific feature bits\n");
641 aux &= ~(L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP);
642 }
643
644 /*
645 * Always enable non-secure access to the lockdown registers -
646 * we write to them as part of the L2C enable sequence so they
647 * need to be accessible.
648 */
649 l2x0_saved_regs.aux_ctrl = aux | L310_AUX_CTRL_NS_LOCKDOWN;
650
651 l2c_enable(base, num_lock);
652
653 /* Read back resulting AUX_CTRL value as it could have been altered. */
654 aux = readl_relaxed(base + L2X0_AUX_CTRL);
655
656 if (aux & (L310_AUX_CTRL_DATA_PREFETCH | L310_AUX_CTRL_INSTR_PREFETCH)) {
657 u32 prefetch = readl_relaxed(base + L310_PREFETCH_CTRL);
658
659 pr_info("L2C-310 %s%s prefetch enabled, offset %u lines\n",
660 aux & L310_AUX_CTRL_INSTR_PREFETCH ? "I" : "",
661 aux & L310_AUX_CTRL_DATA_PREFETCH ? "D" : "",
662 1 + (prefetch & L310_PREFETCH_CTRL_OFFSET_MASK));
663 }
664
665 /* r3p0 or later has power control register */
666 if (rev >= L310_CACHE_ID_RTL_R3P0) {
667 u32 power_ctrl;
668
669 power_ctrl = readl_relaxed(base + L310_POWER_CTRL);
670 pr_info("L2C-310 dynamic clock gating %sabled, standby mode %sabled\n",
671 power_ctrl & L310_DYNAMIC_CLK_GATING_EN ? "en" : "dis",
672 power_ctrl & L310_STNDBY_MODE_EN ? "en" : "dis");
673 }
674
675 if (aux & L310_AUX_CTRL_FULL_LINE_ZERO)
676 cpuhp_setup_state(state: CPUHP_AP_ARM_L2X0_STARTING,
677 name: "arm/l2x0:starting", startup: l2c310_starting_cpu,
678 teardown: l2c310_dying_cpu);
679}
680
681static void __init l2c310_fixup(void __iomem *base, u32 cache_id,
682 struct outer_cache_fns *fns)
683{
684 unsigned revision = cache_id & L2X0_CACHE_ID_RTL_MASK;
685 const char *errata[8];
686 unsigned n = 0;
687
688 if (IS_ENABLED(CONFIG_PL310_ERRATA_588369) &&
689 revision < L310_CACHE_ID_RTL_R2P0 &&
690 /* For bcm compatibility */
691 fns->inv_range == l2c210_inv_range) {
692 fns->inv_range = l2c310_inv_range_erratum;
693 fns->flush_range = l2c310_flush_range_erratum;
694 errata[n++] = "588369";
695 }
696
697 if (IS_ENABLED(CONFIG_PL310_ERRATA_727915) &&
698 revision >= L310_CACHE_ID_RTL_R2P0 &&
699 revision < L310_CACHE_ID_RTL_R3P1) {
700 fns->flush_all = l2c310_flush_all_erratum;
701 errata[n++] = "727915";
702 }
703
704 if (revision >= L310_CACHE_ID_RTL_R3P0 &&
705 revision < L310_CACHE_ID_RTL_R3P2) {
706 u32 val = l2x0_saved_regs.prefetch_ctrl;
707 if (val & L310_PREFETCH_CTRL_DBL_LINEFILL) {
708 val &= ~L310_PREFETCH_CTRL_DBL_LINEFILL;
709 l2x0_saved_regs.prefetch_ctrl = val;
710 errata[n++] = "752271";
711 }
712 }
713
714 if (IS_ENABLED(CONFIG_PL310_ERRATA_753970) &&
715 revision == L310_CACHE_ID_RTL_R3P0) {
716 sync_reg_offset = L2X0_DUMMY_REG;
717 errata[n++] = "753970";
718 }
719
720 if (IS_ENABLED(CONFIG_PL310_ERRATA_769419))
721 errata[n++] = "769419";
722
723 if (n) {
724 unsigned i;
725
726 pr_info("L2C-310 errat%s", n > 1 ? "a" : "um");
727 for (i = 0; i < n; i++)
728 pr_cont(" %s", errata[i]);
729 pr_cont(" enabled\n");
730 }
731}
732
733static void l2c310_disable(void)
734{
735 /*
736 * If full-line-of-zeros is enabled, we must first disable it in the
737 * Cortex-A9 auxiliary control register before disabling the L2 cache.
738 */
739 if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
740 set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1)));
741
742 l2c_disable();
743}
744
745static void l2c310_resume(void)
746{
747 l2c_resume();
748
749 /* Re-enable full-line-of-zeros for Cortex-A9 */
750 if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
751 set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
752}
753
754static void l2c310_unlock(void __iomem *base, unsigned num_lock)
755{
756 if (readl_relaxed(base + L2X0_AUX_CTRL) & L310_AUX_CTRL_NS_LOCKDOWN)
757 l2c_unlock(base, num: num_lock);
758}
759
760static const struct l2c_init_data l2c310_init_fns __initconst = {
761 .type = "L2C-310",
762 .way_size_0 = SZ_8K,
763 .num_lock = 8,
764 .enable = l2c310_enable,
765 .fixup = l2c310_fixup,
766 .save = l2c310_save,
767 .configure = l2c310_configure,
768 .unlock = l2c310_unlock,
769 .outer_cache = {
770 .inv_range = l2c210_inv_range,
771 .clean_range = l2c210_clean_range,
772 .flush_range = l2c210_flush_range,
773 .flush_all = l2c210_flush_all,
774 .disable = l2c310_disable,
775 .sync = l2c210_sync,
776 .resume = l2c310_resume,
777 },
778};
779
780static int __init __l2c_init(const struct l2c_init_data *data,
781 u32 aux_val, u32 aux_mask, u32 cache_id, bool nosync)
782{
783 struct outer_cache_fns fns;
784 unsigned way_size_bits, ways;
785 u32 aux, old_aux;
786
787 /*
788 * Save the pointer globally so that callbacks which do not receive
789 * context from callers can access the structure.
790 */
791 l2x0_data = kmemdup(p: data, size: sizeof(*data), GFP_KERNEL);
792 if (!l2x0_data)
793 return -ENOMEM;
794
795 /*
796 * Sanity check the aux values. aux_mask is the bits we preserve
797 * from reading the hardware register, and aux_val is the bits we
798 * set.
799 */
800 if (aux_val & aux_mask)
801 pr_alert("L2C: platform provided aux values permit register corruption.\n");
802
803 old_aux = aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
804 aux &= aux_mask;
805 aux |= aux_val;
806
807 if (old_aux != aux)
808 pr_warn("L2C: DT/platform modifies aux control register: 0x%08x -> 0x%08x\n",
809 old_aux, aux);
810
811 /* Determine the number of ways */
812 switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
813 case L2X0_CACHE_ID_PART_L310:
814 if ((aux_val | ~aux_mask) & (L2C_AUX_CTRL_WAY_SIZE_MASK | L310_AUX_CTRL_ASSOCIATIVITY_16))
815 pr_warn("L2C: DT/platform tries to modify or specify cache size\n");
816 if (aux & (1 << 16))
817 ways = 16;
818 else
819 ways = 8;
820 break;
821
822 case L2X0_CACHE_ID_PART_L210:
823 case L2X0_CACHE_ID_PART_L220:
824 ways = (aux >> 13) & 0xf;
825 break;
826
827 case AURORA_CACHE_ID:
828 ways = (aux >> 13) & 0xf;
829 ways = 2 << ((ways + 1) >> 2);
830 break;
831
832 default:
833 /* Assume unknown chips have 8 ways */
834 ways = 8;
835 break;
836 }
837
838 l2x0_way_mask = (1 << ways) - 1;
839
840 /*
841 * way_size_0 is the size that a way_size value of zero would be
842 * given the calculation: way_size = way_size_0 << way_size_bits.
843 * So, if way_size_bits=0 is reserved, but way_size_bits=1 is 16k,
844 * then way_size_0 would be 8k.
845 *
846 * L2 cache size = number of ways * way size.
847 */
848 way_size_bits = (aux & L2C_AUX_CTRL_WAY_SIZE_MASK) >>
849 L2C_AUX_CTRL_WAY_SIZE_SHIFT;
850 l2x0_size = ways * (data->way_size_0 << way_size_bits);
851
852 fns = data->outer_cache;
853 fns.write_sec = outer_cache.write_sec;
854 fns.configure = outer_cache.configure;
855 if (data->fixup)
856 data->fixup(l2x0_base, cache_id, &fns);
857 if (nosync) {
858 pr_info("L2C: disabling outer sync\n");
859 fns.sync = NULL;
860 }
861
862 /*
863 * Check if l2x0 controller is already enabled. If we are booting
864 * in non-secure mode accessing the below registers will fault.
865 */
866 if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN)) {
867 l2x0_saved_regs.aux_ctrl = aux;
868
869 data->enable(l2x0_base, data->num_lock);
870 }
871
872 outer_cache = fns;
873
874 /*
875 * It is strange to save the register state before initialisation,
876 * but hey, this is what the DT implementations decided to do.
877 */
878 if (data->save)
879 data->save(l2x0_base);
880
881 /* Re-read it in case some bits are reserved. */
882 aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
883
884 pr_info("%s cache controller enabled, %d ways, %d kB\n",
885 data->type, ways, l2x0_size >> 10);
886 pr_info("%s: CACHE_ID 0x%08x, AUX_CTRL 0x%08x\n",
887 data->type, cache_id, aux);
888
889 l2x0_pmu_register(l2x0_base, cache_id);
890
891 return 0;
892}
893
894void __init l2x0_init(void __iomem *base, u32 aux_val, u32 aux_mask)
895{
896 const struct l2c_init_data *data;
897 u32 cache_id;
898
899 l2x0_base = base;
900
901 cache_id = readl_relaxed(base + L2X0_CACHE_ID);
902
903 switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
904 default:
905 case L2X0_CACHE_ID_PART_L210:
906 data = &l2c210_data;
907 break;
908
909 case L2X0_CACHE_ID_PART_L220:
910 data = &l2c220_data;
911 break;
912
913 case L2X0_CACHE_ID_PART_L310:
914 data = &l2c310_init_fns;
915 break;
916 }
917
918 /* Read back current (default) hardware configuration */
919 if (data->save)
920 data->save(l2x0_base);
921
922 __l2c_init(data, aux_val, aux_mask, cache_id, nosync: false);
923}
924
925#ifdef CONFIG_OF
926static int l2_wt_override;
927
928/* Aurora don't have the cache ID register available, so we have to
929 * pass it though the device tree */
930static u32 cache_id_part_number_from_dt;
931
932/**
933 * l2x0_cache_size_of_parse() - read cache size parameters from DT
934 * @np: the device tree node for the l2 cache
935 * @aux_val: pointer to machine-supplied auxilary register value, to
936 * be augmented by the call (bits to be set to 1)
937 * @aux_mask: pointer to machine-supplied auxilary register mask, to
938 * be augmented by the call (bits to be set to 0)
939 * @associativity: variable to return the calculated associativity in
940 * @max_way_size: the maximum size in bytes for the cache ways
941 */
942static int __init l2x0_cache_size_of_parse(const struct device_node *np,
943 u32 *aux_val, u32 *aux_mask,
944 u32 *associativity,
945 u32 max_way_size)
946{
947 u32 mask = 0, val = 0;
948 u32 cache_size = 0, sets = 0;
949 u32 way_size_bits = 1;
950 u32 way_size = 0;
951 u32 block_size = 0;
952 u32 line_size = 0;
953
954 of_property_read_u32(np, propname: "cache-size", out_value: &cache_size);
955 of_property_read_u32(np, propname: "cache-sets", out_value: &sets);
956 of_property_read_u32(np, propname: "cache-block-size", out_value: &block_size);
957 of_property_read_u32(np, propname: "cache-line-size", out_value: &line_size);
958
959 if (!cache_size || !sets)
960 return -ENODEV;
961
962 /* All these l2 caches have the same line = block size actually */
963 if (!line_size) {
964 if (block_size) {
965 /* If linesize is not given, it is equal to blocksize */
966 line_size = block_size;
967 } else {
968 /* Fall back to known size */
969 pr_warn("L2C OF: no cache block/line size given: "
970 "falling back to default size %d bytes\n",
971 CACHE_LINE_SIZE);
972 line_size = CACHE_LINE_SIZE;
973 }
974 }
975
976 if (line_size != CACHE_LINE_SIZE)
977 pr_warn("L2C OF: DT supplied line size %d bytes does "
978 "not match hardware line size of %d bytes\n",
979 line_size,
980 CACHE_LINE_SIZE);
981
982 /*
983 * Since:
984 * set size = cache size / sets
985 * ways = cache size / (sets * line size)
986 * way size = cache size / (cache size / (sets * line size))
987 * way size = sets * line size
988 * associativity = ways = cache size / way size
989 */
990 way_size = sets * line_size;
991 *associativity = cache_size / way_size;
992
993 if (way_size > max_way_size) {
994 pr_err("L2C OF: set size %dKB is too large\n", way_size);
995 return -EINVAL;
996 }
997
998 pr_info("L2C OF: override cache size: %d bytes (%dKB)\n",
999 cache_size, cache_size >> 10);
1000 pr_info("L2C OF: override line size: %d bytes\n", line_size);
1001 pr_info("L2C OF: override way size: %d bytes (%dKB)\n",
1002 way_size, way_size >> 10);
1003 pr_info("L2C OF: override associativity: %d\n", *associativity);
1004
1005 /*
1006 * Calculates the bits 17:19 to set for way size:
1007 * 512KB -> 6, 256KB -> 5, ... 16KB -> 1
1008 */
1009 way_size_bits = ilog2(way_size >> 10) - 3;
1010 if (way_size_bits < 1 || way_size_bits > 6) {
1011 pr_err("L2C OF: cache way size illegal: %dKB is not mapped\n",
1012 way_size);
1013 return -EINVAL;
1014 }
1015
1016 mask |= L2C_AUX_CTRL_WAY_SIZE_MASK;
1017 val |= (way_size_bits << L2C_AUX_CTRL_WAY_SIZE_SHIFT);
1018
1019 *aux_val &= ~mask;
1020 *aux_val |= val;
1021 *aux_mask &= ~mask;
1022
1023 return 0;
1024}
1025
1026static void __init l2x0_of_parse(const struct device_node *np,
1027 u32 *aux_val, u32 *aux_mask)
1028{
1029 u32 data[2] = { 0, 0 };
1030 u32 tag = 0;
1031 u32 dirty = 0;
1032 u32 val = 0, mask = 0;
1033 u32 assoc;
1034 int ret;
1035
1036 of_property_read_u32(np, propname: "arm,tag-latency", out_value: &tag);
1037 if (tag) {
1038 mask |= L2X0_AUX_CTRL_TAG_LATENCY_MASK;
1039 val |= (tag - 1) << L2X0_AUX_CTRL_TAG_LATENCY_SHIFT;
1040 }
1041
1042 of_property_read_u32_array(np, propname: "arm,data-latency",
1043 out_values: data, ARRAY_SIZE(data));
1044 if (data[0] && data[1]) {
1045 mask |= L2X0_AUX_CTRL_DATA_RD_LATENCY_MASK |
1046 L2X0_AUX_CTRL_DATA_WR_LATENCY_MASK;
1047 val |= ((data[0] - 1) << L2X0_AUX_CTRL_DATA_RD_LATENCY_SHIFT) |
1048 ((data[1] - 1) << L2X0_AUX_CTRL_DATA_WR_LATENCY_SHIFT);
1049 }
1050
1051 of_property_read_u32(np, propname: "arm,dirty-latency", out_value: &dirty);
1052 if (dirty) {
1053 mask |= L2X0_AUX_CTRL_DIRTY_LATENCY_MASK;
1054 val |= (dirty - 1) << L2X0_AUX_CTRL_DIRTY_LATENCY_SHIFT;
1055 }
1056
1057 if (of_property_read_bool(np, propname: "arm,parity-enable")) {
1058 mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1059 val |= L2C_AUX_CTRL_PARITY_ENABLE;
1060 } else if (of_property_read_bool(np, propname: "arm,parity-disable")) {
1061 mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1062 }
1063
1064 if (of_property_read_bool(np, propname: "arm,shared-override")) {
1065 mask &= ~L2C_AUX_CTRL_SHARED_OVERRIDE;
1066 val |= L2C_AUX_CTRL_SHARED_OVERRIDE;
1067 }
1068
1069 ret = l2x0_cache_size_of_parse(np, aux_val, aux_mask, associativity: &assoc, SZ_256K);
1070 if (ret)
1071 return;
1072
1073 if (assoc > 8) {
1074 pr_err("l2x0 of: cache setting yield too high associativity\n");
1075 pr_err("l2x0 of: %d calculated, max 8\n", assoc);
1076 } else {
1077 mask |= L2X0_AUX_CTRL_ASSOC_MASK;
1078 val |= (assoc << L2X0_AUX_CTRL_ASSOC_SHIFT);
1079 }
1080
1081 *aux_val &= ~mask;
1082 *aux_val |= val;
1083 *aux_mask &= ~mask;
1084}
1085
1086static const struct l2c_init_data of_l2c210_data __initconst = {
1087 .type = "L2C-210",
1088 .way_size_0 = SZ_8K,
1089 .num_lock = 1,
1090 .of_parse = l2x0_of_parse,
1091 .enable = l2c_enable,
1092 .save = l2c_save,
1093 .configure = l2c_configure,
1094 .unlock = l2c_unlock,
1095 .outer_cache = {
1096 .inv_range = l2c210_inv_range,
1097 .clean_range = l2c210_clean_range,
1098 .flush_range = l2c210_flush_range,
1099 .flush_all = l2c210_flush_all,
1100 .disable = l2c_disable,
1101 .sync = l2c210_sync,
1102 .resume = l2c_resume,
1103 },
1104};
1105
1106static const struct l2c_init_data of_l2c220_data __initconst = {
1107 .type = "L2C-220",
1108 .way_size_0 = SZ_8K,
1109 .num_lock = 1,
1110 .of_parse = l2x0_of_parse,
1111 .enable = l2c220_enable,
1112 .save = l2c_save,
1113 .configure = l2c_configure,
1114 .unlock = l2c220_unlock,
1115 .outer_cache = {
1116 .inv_range = l2c220_inv_range,
1117 .clean_range = l2c220_clean_range,
1118 .flush_range = l2c220_flush_range,
1119 .flush_all = l2c220_flush_all,
1120 .disable = l2c_disable,
1121 .sync = l2c220_sync,
1122 .resume = l2c_resume,
1123 },
1124};
1125
1126static void __init l2c310_of_parse(const struct device_node *np,
1127 u32 *aux_val, u32 *aux_mask)
1128{
1129 u32 data[3] = { 0, 0, 0 };
1130 u32 tag[3] = { 0, 0, 0 };
1131 u32 filter[2] = { 0, 0 };
1132 u32 assoc;
1133 u32 prefetch;
1134 u32 power;
1135 u32 val;
1136 int ret;
1137
1138 of_property_read_u32_array(np, propname: "arm,tag-latency", out_values: tag, ARRAY_SIZE(tag));
1139 if (tag[0] && tag[1] && tag[2])
1140 l2x0_saved_regs.tag_latency =
1141 L310_LATENCY_CTRL_RD(tag[0] - 1) |
1142 L310_LATENCY_CTRL_WR(tag[1] - 1) |
1143 L310_LATENCY_CTRL_SETUP(tag[2] - 1);
1144
1145 of_property_read_u32_array(np, propname: "arm,data-latency",
1146 out_values: data, ARRAY_SIZE(data));
1147 if (data[0] && data[1] && data[2])
1148 l2x0_saved_regs.data_latency =
1149 L310_LATENCY_CTRL_RD(data[0] - 1) |
1150 L310_LATENCY_CTRL_WR(data[1] - 1) |
1151 L310_LATENCY_CTRL_SETUP(data[2] - 1);
1152
1153 of_property_read_u32_array(np, propname: "arm,filter-ranges",
1154 out_values: filter, ARRAY_SIZE(filter));
1155 if (filter[1]) {
1156 l2x0_saved_regs.filter_end =
1157 ALIGN(filter[0] + filter[1], SZ_1M);
1158 l2x0_saved_regs.filter_start = (filter[0] & ~(SZ_1M - 1))
1159 | L310_ADDR_FILTER_EN;
1160 }
1161
1162 ret = l2x0_cache_size_of_parse(np, aux_val, aux_mask, associativity: &assoc, SZ_512K);
1163 if (!ret) {
1164 switch (assoc) {
1165 case 16:
1166 *aux_val &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1167 *aux_val |= L310_AUX_CTRL_ASSOCIATIVITY_16;
1168 *aux_mask &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1169 break;
1170 case 8:
1171 *aux_val &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1172 *aux_mask &= ~L2X0_AUX_CTRL_ASSOC_MASK;
1173 break;
1174 default:
1175 pr_err("L2C-310 OF cache associativity %d invalid, only 8 or 16 permitted\n",
1176 assoc);
1177 break;
1178 }
1179 }
1180
1181 if (of_property_read_bool(np, propname: "arm,shared-override")) {
1182 *aux_val |= L2C_AUX_CTRL_SHARED_OVERRIDE;
1183 *aux_mask &= ~L2C_AUX_CTRL_SHARED_OVERRIDE;
1184 }
1185
1186 if (of_property_read_bool(np, propname: "arm,parity-enable")) {
1187 *aux_val |= L2C_AUX_CTRL_PARITY_ENABLE;
1188 *aux_mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1189 } else if (of_property_read_bool(np, propname: "arm,parity-disable")) {
1190 *aux_val &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1191 *aux_mask &= ~L2C_AUX_CTRL_PARITY_ENABLE;
1192 }
1193
1194 if (of_property_read_bool(np, propname: "arm,early-bresp-disable"))
1195 l2x0_bresp_disable = true;
1196
1197 if (of_property_read_bool(np, propname: "arm,full-line-zero-disable"))
1198 l2x0_flz_disable = true;
1199
1200 prefetch = l2x0_saved_regs.prefetch_ctrl;
1201
1202 ret = of_property_read_u32(np, propname: "arm,double-linefill", out_value: &val);
1203 if (ret == 0) {
1204 if (val)
1205 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL;
1206 else
1207 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL;
1208 } else if (ret != -EINVAL) {
1209 pr_err("L2C-310 OF arm,double-linefill property value is missing\n");
1210 }
1211
1212 ret = of_property_read_u32(np, propname: "arm,double-linefill-incr", out_value: &val);
1213 if (ret == 0) {
1214 if (val)
1215 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL_INCR;
1216 else
1217 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL_INCR;
1218 } else if (ret != -EINVAL) {
1219 pr_err("L2C-310 OF arm,double-linefill-incr property value is missing\n");
1220 }
1221
1222 ret = of_property_read_u32(np, propname: "arm,double-linefill-wrap", out_value: &val);
1223 if (ret == 0) {
1224 if (!val)
1225 prefetch |= L310_PREFETCH_CTRL_DBL_LINEFILL_WRAP;
1226 else
1227 prefetch &= ~L310_PREFETCH_CTRL_DBL_LINEFILL_WRAP;
1228 } else if (ret != -EINVAL) {
1229 pr_err("L2C-310 OF arm,double-linefill-wrap property value is missing\n");
1230 }
1231
1232 ret = of_property_read_u32(np, propname: "arm,prefetch-drop", out_value: &val);
1233 if (ret == 0) {
1234 if (val)
1235 prefetch |= L310_PREFETCH_CTRL_PREFETCH_DROP;
1236 else
1237 prefetch &= ~L310_PREFETCH_CTRL_PREFETCH_DROP;
1238 } else if (ret != -EINVAL) {
1239 pr_err("L2C-310 OF arm,prefetch-drop property value is missing\n");
1240 }
1241
1242 ret = of_property_read_u32(np, propname: "arm,prefetch-offset", out_value: &val);
1243 if (ret == 0) {
1244 prefetch &= ~L310_PREFETCH_CTRL_OFFSET_MASK;
1245 prefetch |= val & L310_PREFETCH_CTRL_OFFSET_MASK;
1246 } else if (ret != -EINVAL) {
1247 pr_err("L2C-310 OF arm,prefetch-offset property value is missing\n");
1248 }
1249
1250 ret = of_property_read_u32(np, propname: "prefetch-data", out_value: &val);
1251 if (ret == 0) {
1252 if (val) {
1253 prefetch |= L310_PREFETCH_CTRL_DATA_PREFETCH;
1254 *aux_val |= L310_PREFETCH_CTRL_DATA_PREFETCH;
1255 } else {
1256 prefetch &= ~L310_PREFETCH_CTRL_DATA_PREFETCH;
1257 *aux_val &= ~L310_PREFETCH_CTRL_DATA_PREFETCH;
1258 }
1259 *aux_mask &= ~L310_PREFETCH_CTRL_DATA_PREFETCH;
1260 } else if (ret != -EINVAL) {
1261 pr_err("L2C-310 OF prefetch-data property value is missing\n");
1262 }
1263
1264 ret = of_property_read_u32(np, propname: "prefetch-instr", out_value: &val);
1265 if (ret == 0) {
1266 if (val) {
1267 prefetch |= L310_PREFETCH_CTRL_INSTR_PREFETCH;
1268 *aux_val |= L310_PREFETCH_CTRL_INSTR_PREFETCH;
1269 } else {
1270 prefetch &= ~L310_PREFETCH_CTRL_INSTR_PREFETCH;
1271 *aux_val &= ~L310_PREFETCH_CTRL_INSTR_PREFETCH;
1272 }
1273 *aux_mask &= ~L310_PREFETCH_CTRL_INSTR_PREFETCH;
1274 } else if (ret != -EINVAL) {
1275 pr_err("L2C-310 OF prefetch-instr property value is missing\n");
1276 }
1277
1278 l2x0_saved_regs.prefetch_ctrl = prefetch;
1279
1280 power = l2x0_saved_regs.pwr_ctrl |
1281 L310_DYNAMIC_CLK_GATING_EN | L310_STNDBY_MODE_EN;
1282
1283 ret = of_property_read_u32(np, propname: "arm,dynamic-clock-gating", out_value: &val);
1284 if (!ret) {
1285 if (!val)
1286 power &= ~L310_DYNAMIC_CLK_GATING_EN;
1287 } else if (ret != -EINVAL) {
1288 pr_err("L2C-310 OF dynamic-clock-gating property value is missing or invalid\n");
1289 }
1290 ret = of_property_read_u32(np, propname: "arm,standby-mode", out_value: &val);
1291 if (!ret) {
1292 if (!val)
1293 power &= ~L310_STNDBY_MODE_EN;
1294 } else if (ret != -EINVAL) {
1295 pr_err("L2C-310 OF standby-mode property value is missing or invalid\n");
1296 }
1297
1298 l2x0_saved_regs.pwr_ctrl = power;
1299}
1300
1301static const struct l2c_init_data of_l2c310_data __initconst = {
1302 .type = "L2C-310",
1303 .way_size_0 = SZ_8K,
1304 .num_lock = 8,
1305 .of_parse = l2c310_of_parse,
1306 .enable = l2c310_enable,
1307 .fixup = l2c310_fixup,
1308 .save = l2c310_save,
1309 .configure = l2c310_configure,
1310 .unlock = l2c310_unlock,
1311 .outer_cache = {
1312 .inv_range = l2c210_inv_range,
1313 .clean_range = l2c210_clean_range,
1314 .flush_range = l2c210_flush_range,
1315 .flush_all = l2c210_flush_all,
1316 .disable = l2c310_disable,
1317 .sync = l2c210_sync,
1318 .resume = l2c310_resume,
1319 },
1320};
1321
1322/*
1323 * This is a variant of the of_l2c310_data with .sync set to
1324 * NULL. Outer sync operations are not needed when the system is I/O
1325 * coherent, and potentially harmful in certain situations (PCIe/PL310
1326 * deadlock on Armada 375/38x due to hardware I/O coherency). The
1327 * other operations are kept because they are infrequent (therefore do
1328 * not cause the deadlock in practice) and needed for secondary CPU
1329 * boot and other power management activities.
1330 */
1331static const struct l2c_init_data of_l2c310_coherent_data __initconst = {
1332 .type = "L2C-310 Coherent",
1333 .way_size_0 = SZ_8K,
1334 .num_lock = 8,
1335 .of_parse = l2c310_of_parse,
1336 .enable = l2c310_enable,
1337 .fixup = l2c310_fixup,
1338 .save = l2c310_save,
1339 .configure = l2c310_configure,
1340 .unlock = l2c310_unlock,
1341 .outer_cache = {
1342 .inv_range = l2c210_inv_range,
1343 .clean_range = l2c210_clean_range,
1344 .flush_range = l2c210_flush_range,
1345 .flush_all = l2c210_flush_all,
1346 .disable = l2c310_disable,
1347 .resume = l2c310_resume,
1348 },
1349};
1350
1351/*
1352 * Note that the end addresses passed to Linux primitives are
1353 * noninclusive, while the hardware cache range operations use
1354 * inclusive start and end addresses.
1355 */
1356static unsigned long aurora_range_end(unsigned long start, unsigned long end)
1357{
1358 /*
1359 * Limit the number of cache lines processed at once,
1360 * since cache range operations stall the CPU pipeline
1361 * until completion.
1362 */
1363 if (end > start + AURORA_MAX_RANGE_SIZE)
1364 end = start + AURORA_MAX_RANGE_SIZE;
1365
1366 /*
1367 * Cache range operations can't straddle a page boundary.
1368 */
1369 if (end > PAGE_ALIGN(start+1))
1370 end = PAGE_ALIGN(start+1);
1371
1372 return end;
1373}
1374
1375static void aurora_pa_range(unsigned long start, unsigned long end,
1376 unsigned long offset)
1377{
1378 void __iomem *base = l2x0_base;
1379 unsigned long range_end;
1380 unsigned long flags;
1381
1382 /*
1383 * round start and end adresses up to cache line size
1384 */
1385 start &= ~(CACHE_LINE_SIZE - 1);
1386 end = ALIGN(end, CACHE_LINE_SIZE);
1387
1388 /*
1389 * perform operation on all full cache lines between 'start' and 'end'
1390 */
1391 while (start < end) {
1392 range_end = aurora_range_end(start, end);
1393
1394 raw_spin_lock_irqsave(&l2x0_lock, flags);
1395 writel_relaxed(start, base + AURORA_RANGE_BASE_ADDR_REG);
1396 writel_relaxed(range_end - CACHE_LINE_SIZE, base + offset);
1397 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1398
1399 writel_relaxed(0, base + AURORA_SYNC_REG);
1400 start = range_end;
1401 }
1402}
1403static void aurora_inv_range(unsigned long start, unsigned long end)
1404{
1405 aurora_pa_range(start, end, AURORA_INVAL_RANGE_REG);
1406}
1407
1408static void aurora_clean_range(unsigned long start, unsigned long end)
1409{
1410 /*
1411 * If L2 is forced to WT, the L2 will always be clean and we
1412 * don't need to do anything here.
1413 */
1414 if (!l2_wt_override)
1415 aurora_pa_range(start, end, AURORA_CLEAN_RANGE_REG);
1416}
1417
1418static void aurora_flush_range(unsigned long start, unsigned long end)
1419{
1420 if (l2_wt_override)
1421 aurora_pa_range(start, end, AURORA_INVAL_RANGE_REG);
1422 else
1423 aurora_pa_range(start, end, AURORA_FLUSH_RANGE_REG);
1424}
1425
1426static void aurora_flush_all(void)
1427{
1428 void __iomem *base = l2x0_base;
1429 unsigned long flags;
1430
1431 /* clean all ways */
1432 raw_spin_lock_irqsave(&l2x0_lock, flags);
1433 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
1434 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1435
1436 writel_relaxed(0, base + AURORA_SYNC_REG);
1437}
1438
1439static void aurora_cache_sync(void)
1440{
1441 writel_relaxed(0, l2x0_base + AURORA_SYNC_REG);
1442}
1443
1444static void aurora_disable(void)
1445{
1446 void __iomem *base = l2x0_base;
1447 unsigned long flags;
1448
1449 raw_spin_lock_irqsave(&l2x0_lock, flags);
1450 __l2c_op_way(base + L2X0_CLEAN_INV_WAY);
1451 writel_relaxed(0, base + AURORA_SYNC_REG);
1452 l2c_write_sec(0, base, L2X0_CTRL);
1453 dsb(st);
1454 raw_spin_unlock_irqrestore(&l2x0_lock, flags);
1455}
1456
1457static void aurora_save(void __iomem *base)
1458{
1459 l2x0_saved_regs.ctrl = readl_relaxed(base + L2X0_CTRL);
1460 l2x0_saved_regs.aux_ctrl = readl_relaxed(base + L2X0_AUX_CTRL);
1461}
1462
1463/*
1464 * For Aurora cache in no outer mode, enable via the CP15 coprocessor
1465 * broadcasting of cache commands to L2.
1466 */
1467static void __init aurora_enable_no_outer(void __iomem *base,
1468 unsigned num_lock)
1469{
1470 u32 u;
1471
1472 asm volatile("mrc p15, 1, %0, c15, c2, 0" : "=r" (u));
1473 u |= AURORA_CTRL_FW; /* Set the FW bit */
1474 asm volatile("mcr p15, 1, %0, c15, c2, 0" : : "r" (u));
1475
1476 isb();
1477
1478 l2c_enable(base, num_lock);
1479}
1480
1481static void __init aurora_fixup(void __iomem *base, u32 cache_id,
1482 struct outer_cache_fns *fns)
1483{
1484 sync_reg_offset = AURORA_SYNC_REG;
1485}
1486
1487static void __init aurora_of_parse(const struct device_node *np,
1488 u32 *aux_val, u32 *aux_mask)
1489{
1490 u32 val = AURORA_ACR_REPLACEMENT_TYPE_SEMIPLRU;
1491 u32 mask = AURORA_ACR_REPLACEMENT_MASK;
1492
1493 of_property_read_u32(np, propname: "cache-id-part",
1494 out_value: &cache_id_part_number_from_dt);
1495
1496 /* Determine and save the write policy */
1497 l2_wt_override = of_property_read_bool(np, propname: "wt-override");
1498
1499 if (l2_wt_override) {
1500 val |= AURORA_ACR_FORCE_WRITE_THRO_POLICY;
1501 mask |= AURORA_ACR_FORCE_WRITE_POLICY_MASK;
1502 }
1503
1504 if (of_property_read_bool(np, propname: "marvell,ecc-enable")) {
1505 mask |= AURORA_ACR_ECC_EN;
1506 val |= AURORA_ACR_ECC_EN;
1507 }
1508
1509 if (of_property_read_bool(np, propname: "arm,parity-enable")) {
1510 mask |= AURORA_ACR_PARITY_EN;
1511 val |= AURORA_ACR_PARITY_EN;
1512 } else if (of_property_read_bool(np, propname: "arm,parity-disable")) {
1513 mask |= AURORA_ACR_PARITY_EN;
1514 }
1515
1516 *aux_val &= ~mask;
1517 *aux_val |= val;
1518 *aux_mask &= ~mask;
1519}
1520
1521static const struct l2c_init_data of_aurora_with_outer_data __initconst = {
1522 .type = "Aurora",
1523 .way_size_0 = SZ_4K,
1524 .num_lock = 4,
1525 .of_parse = aurora_of_parse,
1526 .enable = l2c_enable,
1527 .fixup = aurora_fixup,
1528 .save = aurora_save,
1529 .configure = l2c_configure,
1530 .unlock = l2c_unlock,
1531 .outer_cache = {
1532 .inv_range = aurora_inv_range,
1533 .clean_range = aurora_clean_range,
1534 .flush_range = aurora_flush_range,
1535 .flush_all = aurora_flush_all,
1536 .disable = aurora_disable,
1537 .sync = aurora_cache_sync,
1538 .resume = l2c_resume,
1539 },
1540};
1541
1542static const struct l2c_init_data of_aurora_no_outer_data __initconst = {
1543 .type = "Aurora",
1544 .way_size_0 = SZ_4K,
1545 .num_lock = 4,
1546 .of_parse = aurora_of_parse,
1547 .enable = aurora_enable_no_outer,
1548 .fixup = aurora_fixup,
1549 .save = aurora_save,
1550 .configure = l2c_configure,
1551 .unlock = l2c_unlock,
1552 .outer_cache = {
1553 .resume = l2c_resume,
1554 },
1555};
1556
1557/*
1558 * For certain Broadcom SoCs, depending on the address range, different offsets
1559 * need to be added to the address before passing it to L2 for
1560 * invalidation/clean/flush
1561 *
1562 * Section Address Range Offset EMI
1563 * 1 0x00000000 - 0x3FFFFFFF 0x80000000 VC
1564 * 2 0x40000000 - 0xBFFFFFFF 0x40000000 SYS
1565 * 3 0xC0000000 - 0xFFFFFFFF 0x80000000 VC
1566 *
1567 * When the start and end addresses have crossed two different sections, we
1568 * need to break the L2 operation into two, each within its own section.
1569 * For example, if we need to invalidate addresses starts at 0xBFFF0000 and
1570 * ends at 0xC0001000, we need do invalidate 1) 0xBFFF0000 - 0xBFFFFFFF and 2)
1571 * 0xC0000000 - 0xC0001000
1572 *
1573 * Note 1:
1574 * By breaking a single L2 operation into two, we may potentially suffer some
1575 * performance hit, but keep in mind the cross section case is very rare
1576 *
1577 * Note 2:
1578 * We do not need to handle the case when the start address is in
1579 * Section 1 and the end address is in Section 3, since it is not a valid use
1580 * case
1581 *
1582 * Note 3:
1583 * Section 1 in practical terms can no longer be used on rev A2. Because of
1584 * that the code does not need to handle section 1 at all.
1585 *
1586 */
1587#define BCM_SYS_EMI_START_ADDR 0x40000000UL
1588#define BCM_VC_EMI_SEC3_START_ADDR 0xC0000000UL
1589
1590#define BCM_SYS_EMI_OFFSET 0x40000000UL
1591#define BCM_VC_EMI_OFFSET 0x80000000UL
1592
1593static inline int bcm_addr_is_sys_emi(unsigned long addr)
1594{
1595 return (addr >= BCM_SYS_EMI_START_ADDR) &&
1596 (addr < BCM_VC_EMI_SEC3_START_ADDR);
1597}
1598
1599static inline unsigned long bcm_l2_phys_addr(unsigned long addr)
1600{
1601 if (bcm_addr_is_sys_emi(addr))
1602 return addr + BCM_SYS_EMI_OFFSET;
1603 else
1604 return addr + BCM_VC_EMI_OFFSET;
1605}
1606
1607static void bcm_inv_range(unsigned long start, unsigned long end)
1608{
1609 unsigned long new_start, new_end;
1610
1611 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1612
1613 if (unlikely(end <= start))
1614 return;
1615
1616 new_start = bcm_l2_phys_addr(addr: start);
1617 new_end = bcm_l2_phys_addr(addr: end);
1618
1619 /* normal case, no cross section between start and end */
1620 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1621 l2c210_inv_range(start: new_start, end: new_end);
1622 return;
1623 }
1624
1625 /* They cross sections, so it can only be a cross from section
1626 * 2 to section 3
1627 */
1628 l2c210_inv_range(start: new_start,
1629 end: bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1630 l2c210_inv_range(start: bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1631 end: new_end);
1632}
1633
1634static void bcm_clean_range(unsigned long start, unsigned long end)
1635{
1636 unsigned long new_start, new_end;
1637
1638 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1639
1640 if (unlikely(end <= start))
1641 return;
1642
1643 new_start = bcm_l2_phys_addr(addr: start);
1644 new_end = bcm_l2_phys_addr(addr: end);
1645
1646 /* normal case, no cross section between start and end */
1647 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1648 l2c210_clean_range(start: new_start, end: new_end);
1649 return;
1650 }
1651
1652 /* They cross sections, so it can only be a cross from section
1653 * 2 to section 3
1654 */
1655 l2c210_clean_range(start: new_start,
1656 end: bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1657 l2c210_clean_range(start: bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1658 end: new_end);
1659}
1660
1661static void bcm_flush_range(unsigned long start, unsigned long end)
1662{
1663 unsigned long new_start, new_end;
1664
1665 BUG_ON(start < BCM_SYS_EMI_START_ADDR);
1666
1667 if (unlikely(end <= start))
1668 return;
1669
1670 if ((end - start) >= l2x0_size) {
1671 outer_cache.flush_all();
1672 return;
1673 }
1674
1675 new_start = bcm_l2_phys_addr(addr: start);
1676 new_end = bcm_l2_phys_addr(addr: end);
1677
1678 /* normal case, no cross section between start and end */
1679 if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) {
1680 l2c210_flush_range(start: new_start, end: new_end);
1681 return;
1682 }
1683
1684 /* They cross sections, so it can only be a cross from section
1685 * 2 to section 3
1686 */
1687 l2c210_flush_range(start: new_start,
1688 end: bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1));
1689 l2c210_flush_range(start: bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR),
1690 end: new_end);
1691}
1692
1693/* Broadcom L2C-310 start from ARMs R3P2 or later, and require no fixups */
1694static const struct l2c_init_data of_bcm_l2x0_data __initconst = {
1695 .type = "BCM-L2C-310",
1696 .way_size_0 = SZ_8K,
1697 .num_lock = 8,
1698 .of_parse = l2c310_of_parse,
1699 .enable = l2c310_enable,
1700 .save = l2c310_save,
1701 .configure = l2c310_configure,
1702 .unlock = l2c310_unlock,
1703 .outer_cache = {
1704 .inv_range = bcm_inv_range,
1705 .clean_range = bcm_clean_range,
1706 .flush_range = bcm_flush_range,
1707 .flush_all = l2c210_flush_all,
1708 .disable = l2c310_disable,
1709 .sync = l2c210_sync,
1710 .resume = l2c310_resume,
1711 },
1712};
1713
1714static void __init tauros3_save(void __iomem *base)
1715{
1716 l2c_save(base);
1717
1718 l2x0_saved_regs.aux2_ctrl =
1719 readl_relaxed(base + TAUROS3_AUX2_CTRL);
1720 l2x0_saved_regs.prefetch_ctrl =
1721 readl_relaxed(base + L310_PREFETCH_CTRL);
1722}
1723
1724static void tauros3_configure(void __iomem *base)
1725{
1726 l2c_configure(base);
1727 writel_relaxed(l2x0_saved_regs.aux2_ctrl,
1728 base + TAUROS3_AUX2_CTRL);
1729 writel_relaxed(l2x0_saved_regs.prefetch_ctrl,
1730 base + L310_PREFETCH_CTRL);
1731}
1732
1733static const struct l2c_init_data of_tauros3_data __initconst = {
1734 .type = "Tauros3",
1735 .way_size_0 = SZ_8K,
1736 .num_lock = 8,
1737 .enable = l2c_enable,
1738 .save = tauros3_save,
1739 .configure = tauros3_configure,
1740 .unlock = l2c_unlock,
1741 /* Tauros3 broadcasts L1 cache operations to L2 */
1742 .outer_cache = {
1743 .resume = l2c_resume,
1744 },
1745};
1746
1747#define L2C_ID(name, fns) { .compatible = name, .data = (void *)&fns }
1748static const struct of_device_id l2x0_ids[] __initconst = {
1749 L2C_ID("arm,l210-cache", of_l2c210_data),
1750 L2C_ID("arm,l220-cache", of_l2c220_data),
1751 L2C_ID("arm,pl310-cache", of_l2c310_data),
1752 L2C_ID("brcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1753 L2C_ID("marvell,aurora-outer-cache", of_aurora_with_outer_data),
1754 L2C_ID("marvell,aurora-system-cache", of_aurora_no_outer_data),
1755 L2C_ID("marvell,tauros3-cache", of_tauros3_data),
1756 /* Deprecated IDs */
1757 L2C_ID("bcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data),
1758 {}
1759};
1760
1761int __init l2x0_of_init(u32 aux_val, u32 aux_mask)
1762{
1763 const struct l2c_init_data *data;
1764 struct device_node *np;
1765 struct resource res;
1766 u32 cache_id, old_aux;
1767 u32 cache_level = 2;
1768 bool nosync = false;
1769
1770 np = of_find_matching_node(NULL, matches: l2x0_ids);
1771 if (!np)
1772 return -ENODEV;
1773
1774 if (of_address_to_resource(dev: np, index: 0, r: &res))
1775 return -ENODEV;
1776
1777 l2x0_base = ioremap(offset: res.start, size: resource_size(res: &res));
1778 if (!l2x0_base)
1779 return -ENOMEM;
1780
1781 l2x0_saved_regs.phy_base = res.start;
1782
1783 data = of_match_node(matches: l2x0_ids, node: np)->data;
1784
1785 if (of_device_is_compatible(device: np, "arm,pl310-cache") &&
1786 of_property_read_bool(np, propname: "arm,io-coherent"))
1787 data = &of_l2c310_coherent_data;
1788
1789 old_aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
1790 if (old_aux != ((old_aux & aux_mask) | aux_val)) {
1791 pr_warn("L2C: platform modifies aux control register: 0x%08x -> 0x%08x\n",
1792 old_aux, (old_aux & aux_mask) | aux_val);
1793 } else if (aux_mask != ~0U && aux_val != 0) {
1794 pr_alert("L2C: platform provided aux values match the hardware, so have no effect. Please remove them.\n");
1795 }
1796
1797 /* All L2 caches are unified, so this property should be specified */
1798 if (!of_property_read_bool(np, propname: "cache-unified"))
1799 pr_err("L2C: device tree omits to specify unified cache\n");
1800
1801 if (of_property_read_u32(np, propname: "cache-level", out_value: &cache_level))
1802 pr_err("L2C: device tree omits to specify cache-level\n");
1803
1804 if (cache_level != 2)
1805 pr_err("L2C: device tree specifies invalid cache level\n");
1806
1807 nosync = of_property_read_bool(np, propname: "arm,outer-sync-disable");
1808
1809 /* Read back current (default) hardware configuration */
1810 if (data->save)
1811 data->save(l2x0_base);
1812
1813 /* L2 configuration can only be changed if the cache is disabled */
1814 if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN))
1815 if (data->of_parse)
1816 data->of_parse(np, &aux_val, &aux_mask);
1817
1818 if (cache_id_part_number_from_dt)
1819 cache_id = cache_id_part_number_from_dt;
1820 else
1821 cache_id = readl_relaxed(l2x0_base + L2X0_CACHE_ID);
1822
1823 return __l2c_init(data, aux_val, aux_mask, cache_id, nosync);
1824}
1825#endif
1826

source code of linux/arch/arm/mm/cache-l2x0.c