]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * arch/arm/mm/cache-l2x0.c - L210/L220 cache controller support | |
3 | * | |
4 | * Copyright (C) 2007 ARM Limited | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 as | |
8 | * published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, | |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 | * GNU General Public License for more details. | |
14 | * | |
15 | * You should have received a copy of the GNU General Public License | |
16 | * along with this program; if not, write to the Free Software | |
17 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
18 | */ | |
19 | #include <linux/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/io.h> | |
25 | #include <linux/of.h> | |
26 | #include <linux/of_address.h> | |
27 | ||
28 | #include <asm/cacheflush.h> | |
29 | #include <asm/cp15.h> | |
30 | #include <asm/cputype.h> | |
31 | #include <asm/hardware/cache-l2x0.h> | |
32 | #include "cache-tauros3.h" | |
33 | #include "cache-aurora-l2.h" | |
34 | ||
35 | struct l2c_init_data { | |
36 | const char *type; | |
37 | unsigned way_size_0; | |
38 | unsigned num_lock; | |
39 | void (*of_parse)(const struct device_node *, u32 *, u32 *); | |
40 | void (*enable)(void __iomem *, u32, unsigned); | |
41 | void (*fixup)(void __iomem *, u32, struct outer_cache_fns *); | |
42 | void (*save)(void __iomem *); | |
43 | struct outer_cache_fns outer_cache; | |
44 | }; | |
45 | ||
46 | #define CACHE_LINE_SIZE 32 | |
47 | ||
48 | static void __iomem *l2x0_base; | |
49 | static DEFINE_RAW_SPINLOCK(l2x0_lock); | |
50 | static u32 l2x0_way_mask; /* Bitmask of active ways */ | |
51 | static u32 l2x0_size; | |
52 | static unsigned long sync_reg_offset = L2X0_CACHE_SYNC; | |
53 | ||
54 | struct l2x0_regs l2x0_saved_regs; | |
55 | ||
56 | /* | |
57 | * Common code for all cache controllers. | |
58 | */ | |
59 | static inline void l2c_wait_mask(void __iomem *reg, unsigned long mask) | |
60 | { | |
61 | /* wait for cache operation by line or way to complete */ | |
62 | while (readl_relaxed(reg) & mask) | |
63 | cpu_relax(); | |
64 | } | |
65 | ||
66 | /* | |
67 | * By default, we write directly to secure registers. Platforms must | |
68 | * override this if they are running non-secure. | |
69 | */ | |
70 | static void l2c_write_sec(unsigned long val, void __iomem *base, unsigned reg) | |
71 | { | |
72 | if (val == readl_relaxed(base + reg)) | |
73 | return; | |
74 | if (outer_cache.write_sec) | |
75 | outer_cache.write_sec(val, reg); | |
76 | else | |
77 | writel_relaxed(val, base + reg); | |
78 | } | |
79 | ||
80 | /* | |
81 | * This should only be called when we have a requirement that the | |
82 | * register be written due to a work-around, as platforms running | |
83 | * in non-secure mode may not be able to access this register. | |
84 | */ | |
85 | static inline void l2c_set_debug(void __iomem *base, unsigned long val) | |
86 | { | |
87 | l2c_write_sec(val, base, L2X0_DEBUG_CTRL); | |
88 | } | |
89 | ||
90 | static void __l2c_op_way(void __iomem *reg) | |
91 | { | |
92 | writel_relaxed(l2x0_way_mask, reg); | |
93 | l2c_wait_mask(reg, l2x0_way_mask); | |
94 | } | |
95 | ||
96 | static inline void l2c_unlock(void __iomem *base, unsigned num) | |
97 | { | |
98 | unsigned i; | |
99 | ||
100 | for (i = 0; i < num; i++) { | |
101 | writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_D_BASE + | |
102 | i * L2X0_LOCKDOWN_STRIDE); | |
103 | writel_relaxed(0, base + L2X0_LOCKDOWN_WAY_I_BASE + | |
104 | i * L2X0_LOCKDOWN_STRIDE); | |
105 | } | |
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 | */ | |
112 | static void l2c_enable(void __iomem *base, u32 aux, unsigned num_lock) | |
113 | { | |
114 | unsigned long flags; | |
115 | ||
116 | l2c_write_sec(aux, base, L2X0_AUX_CTRL); | |
117 | ||
118 | l2c_unlock(base, num_lock); | |
119 | ||
120 | local_irq_save(flags); | |
121 | __l2c_op_way(base + L2X0_INV_WAY); | |
122 | writel_relaxed(0, base + sync_reg_offset); | |
123 | l2c_wait_mask(base + sync_reg_offset, 1); | |
124 | local_irq_restore(flags); | |
125 | ||
126 | l2c_write_sec(L2X0_CTRL_EN, base, L2X0_CTRL); | |
127 | } | |
128 | ||
129 | static void l2c_disable(void) | |
130 | { | |
131 | void __iomem *base = l2x0_base; | |
132 | ||
133 | outer_cache.flush_all(); | |
134 | l2c_write_sec(0, base, L2X0_CTRL); | |
135 | dsb(st); | |
136 | } | |
137 | ||
138 | #ifdef CONFIG_CACHE_PL310 | |
139 | static inline void cache_wait(void __iomem *reg, unsigned long mask) | |
140 | { | |
141 | /* cache operations by line are atomic on PL310 */ | |
142 | } | |
143 | #else | |
144 | #define cache_wait l2c_wait_mask | |
145 | #endif | |
146 | ||
147 | static inline void cache_sync(void) | |
148 | { | |
149 | void __iomem *base = l2x0_base; | |
150 | ||
151 | writel_relaxed(0, base + sync_reg_offset); | |
152 | cache_wait(base + L2X0_CACHE_SYNC, 1); | |
153 | } | |
154 | ||
155 | #if defined(CONFIG_PL310_ERRATA_588369) || defined(CONFIG_PL310_ERRATA_727915) | |
156 | static inline void debug_writel(unsigned long val) | |
157 | { | |
158 | l2c_set_debug(l2x0_base, val); | |
159 | } | |
160 | #else | |
161 | /* Optimised out for non-errata case */ | |
162 | static inline void debug_writel(unsigned long val) | |
163 | { | |
164 | } | |
165 | #endif | |
166 | ||
167 | static void l2x0_cache_sync(void) | |
168 | { | |
169 | unsigned long flags; | |
170 | ||
171 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
172 | cache_sync(); | |
173 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
174 | } | |
175 | ||
176 | static void __l2x0_flush_all(void) | |
177 | { | |
178 | debug_writel(0x03); | |
179 | __l2c_op_way(l2x0_base + L2X0_CLEAN_INV_WAY); | |
180 | cache_sync(); | |
181 | debug_writel(0x00); | |
182 | } | |
183 | ||
184 | static void l2x0_flush_all(void) | |
185 | { | |
186 | unsigned long flags; | |
187 | ||
188 | /* clean all ways */ | |
189 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
190 | __l2x0_flush_all(); | |
191 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
192 | } | |
193 | ||
194 | static void l2x0_disable(void) | |
195 | { | |
196 | unsigned long flags; | |
197 | ||
198 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
199 | __l2x0_flush_all(); | |
200 | l2c_write_sec(0, l2x0_base, L2X0_CTRL); | |
201 | dsb(st); | |
202 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
203 | } | |
204 | ||
205 | static void l2c_save(void __iomem *base) | |
206 | { | |
207 | l2x0_saved_regs.aux_ctrl = readl_relaxed(l2x0_base + L2X0_AUX_CTRL); | |
208 | } | |
209 | ||
210 | /* | |
211 | * L2C-210 specific code. | |
212 | * | |
213 | * The L2C-2x0 PA, set/way and sync operations are atomic, but we must | |
214 | * ensure that no background operation is running. The way operations | |
215 | * are all background tasks. | |
216 | * | |
217 | * While a background operation is in progress, any new operation is | |
218 | * ignored (unspecified whether this causes an error.) Thankfully, not | |
219 | * used on SMP. | |
220 | * | |
221 | * Never has a different sync register other than L2X0_CACHE_SYNC, but | |
222 | * we use sync_reg_offset here so we can share some of this with L2C-310. | |
223 | */ | |
224 | static void __l2c210_cache_sync(void __iomem *base) | |
225 | { | |
226 | writel_relaxed(0, base + sync_reg_offset); | |
227 | } | |
228 | ||
229 | static void __l2c210_op_pa_range(void __iomem *reg, unsigned long start, | |
230 | unsigned long end) | |
231 | { | |
232 | while (start < end) { | |
233 | writel_relaxed(start, reg); | |
234 | start += CACHE_LINE_SIZE; | |
235 | } | |
236 | } | |
237 | ||
238 | static void l2c210_inv_range(unsigned long start, unsigned long end) | |
239 | { | |
240 | void __iomem *base = l2x0_base; | |
241 | ||
242 | if (start & (CACHE_LINE_SIZE - 1)) { | |
243 | start &= ~(CACHE_LINE_SIZE - 1); | |
244 | writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA); | |
245 | start += CACHE_LINE_SIZE; | |
246 | } | |
247 | ||
248 | if (end & (CACHE_LINE_SIZE - 1)) { | |
249 | end &= ~(CACHE_LINE_SIZE - 1); | |
250 | writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA); | |
251 | } | |
252 | ||
253 | __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end); | |
254 | __l2c210_cache_sync(base); | |
255 | } | |
256 | ||
257 | static void l2c210_clean_range(unsigned long start, unsigned long end) | |
258 | { | |
259 | void __iomem *base = l2x0_base; | |
260 | ||
261 | start &= ~(CACHE_LINE_SIZE - 1); | |
262 | __l2c210_op_pa_range(base + L2X0_CLEAN_LINE_PA, start, end); | |
263 | __l2c210_cache_sync(base); | |
264 | } | |
265 | ||
266 | static void l2c210_flush_range(unsigned long start, unsigned long end) | |
267 | { | |
268 | void __iomem *base = l2x0_base; | |
269 | ||
270 | start &= ~(CACHE_LINE_SIZE - 1); | |
271 | __l2c210_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA, start, end); | |
272 | __l2c210_cache_sync(base); | |
273 | } | |
274 | ||
275 | static void l2c210_flush_all(void) | |
276 | { | |
277 | void __iomem *base = l2x0_base; | |
278 | ||
279 | BUG_ON(!irqs_disabled()); | |
280 | ||
281 | __l2c_op_way(base + L2X0_CLEAN_INV_WAY); | |
282 | __l2c210_cache_sync(base); | |
283 | } | |
284 | ||
285 | static void l2c210_sync(void) | |
286 | { | |
287 | __l2c210_cache_sync(l2x0_base); | |
288 | } | |
289 | ||
290 | static void l2c210_resume(void) | |
291 | { | |
292 | void __iomem *base = l2x0_base; | |
293 | ||
294 | if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN)) | |
295 | l2c_enable(base, l2x0_saved_regs.aux_ctrl, 1); | |
296 | } | |
297 | ||
298 | static const struct l2c_init_data l2c210_data __initconst = { | |
299 | .type = "L2C-210", | |
300 | .way_size_0 = SZ_8K, | |
301 | .num_lock = 1, | |
302 | .enable = l2c_enable, | |
303 | .save = l2c_save, | |
304 | .outer_cache = { | |
305 | .inv_range = l2c210_inv_range, | |
306 | .clean_range = l2c210_clean_range, | |
307 | .flush_range = l2c210_flush_range, | |
308 | .flush_all = l2c210_flush_all, | |
309 | .disable = l2c_disable, | |
310 | .sync = l2c210_sync, | |
311 | .resume = l2c210_resume, | |
312 | }, | |
313 | }; | |
314 | ||
315 | /* | |
316 | * L2C-220 specific code. | |
317 | * | |
318 | * All operations are background operations: they have to be waited for. | |
319 | * Conflicting requests generate a slave error (which will cause an | |
320 | * imprecise abort.) Never uses sync_reg_offset, so we hard-code the | |
321 | * sync register here. | |
322 | * | |
323 | * However, we can re-use the l2c210_resume call. | |
324 | */ | |
325 | static inline void __l2c220_cache_sync(void __iomem *base) | |
326 | { | |
327 | writel_relaxed(0, base + L2X0_CACHE_SYNC); | |
328 | l2c_wait_mask(base + L2X0_CACHE_SYNC, 1); | |
329 | } | |
330 | ||
331 | static void l2c220_op_way(void __iomem *base, unsigned reg) | |
332 | { | |
333 | unsigned long flags; | |
334 | ||
335 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
336 | __l2c_op_way(base + reg); | |
337 | __l2c220_cache_sync(base); | |
338 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
339 | } | |
340 | ||
341 | static unsigned long l2c220_op_pa_range(void __iomem *reg, unsigned long start, | |
342 | unsigned long end, unsigned long flags) | |
343 | { | |
344 | raw_spinlock_t *lock = &l2x0_lock; | |
345 | ||
346 | while (start < end) { | |
347 | unsigned long blk_end = start + min(end - start, 4096UL); | |
348 | ||
349 | while (start < blk_end) { | |
350 | l2c_wait_mask(reg, 1); | |
351 | writel_relaxed(start, reg); | |
352 | start += CACHE_LINE_SIZE; | |
353 | } | |
354 | ||
355 | if (blk_end < end) { | |
356 | raw_spin_unlock_irqrestore(lock, flags); | |
357 | raw_spin_lock_irqsave(lock, flags); | |
358 | } | |
359 | } | |
360 | ||
361 | return flags; | |
362 | } | |
363 | ||
364 | static void l2c220_inv_range(unsigned long start, unsigned long end) | |
365 | { | |
366 | void __iomem *base = l2x0_base; | |
367 | unsigned long flags; | |
368 | ||
369 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
370 | if ((start | end) & (CACHE_LINE_SIZE - 1)) { | |
371 | if (start & (CACHE_LINE_SIZE - 1)) { | |
372 | start &= ~(CACHE_LINE_SIZE - 1); | |
373 | writel_relaxed(start, base + L2X0_CLEAN_INV_LINE_PA); | |
374 | start += CACHE_LINE_SIZE; | |
375 | } | |
376 | ||
377 | if (end & (CACHE_LINE_SIZE - 1)) { | |
378 | end &= ~(CACHE_LINE_SIZE - 1); | |
379 | l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1); | |
380 | writel_relaxed(end, base + L2X0_CLEAN_INV_LINE_PA); | |
381 | } | |
382 | } | |
383 | ||
384 | flags = l2c220_op_pa_range(base + L2X0_INV_LINE_PA, | |
385 | start, end, flags); | |
386 | l2c_wait_mask(base + L2X0_INV_LINE_PA, 1); | |
387 | __l2c220_cache_sync(base); | |
388 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
389 | } | |
390 | ||
391 | static void l2c220_clean_range(unsigned long start, unsigned long end) | |
392 | { | |
393 | void __iomem *base = l2x0_base; | |
394 | unsigned long flags; | |
395 | ||
396 | start &= ~(CACHE_LINE_SIZE - 1); | |
397 | if ((end - start) >= l2x0_size) { | |
398 | l2c220_op_way(base, L2X0_CLEAN_WAY); | |
399 | return; | |
400 | } | |
401 | ||
402 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
403 | flags = l2c220_op_pa_range(base + L2X0_CLEAN_LINE_PA, | |
404 | start, end, flags); | |
405 | l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1); | |
406 | __l2c220_cache_sync(base); | |
407 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
408 | } | |
409 | ||
410 | static void l2c220_flush_range(unsigned long start, unsigned long end) | |
411 | { | |
412 | void __iomem *base = l2x0_base; | |
413 | unsigned long flags; | |
414 | ||
415 | start &= ~(CACHE_LINE_SIZE - 1); | |
416 | if ((end - start) >= l2x0_size) { | |
417 | l2c220_op_way(base, L2X0_CLEAN_INV_WAY); | |
418 | return; | |
419 | } | |
420 | ||
421 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
422 | flags = l2c220_op_pa_range(base + L2X0_CLEAN_INV_LINE_PA, | |
423 | start, end, flags); | |
424 | l2c_wait_mask(base + L2X0_CLEAN_INV_LINE_PA, 1); | |
425 | __l2c220_cache_sync(base); | |
426 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
427 | } | |
428 | ||
429 | static void l2c220_flush_all(void) | |
430 | { | |
431 | l2c220_op_way(l2x0_base, L2X0_CLEAN_INV_WAY); | |
432 | } | |
433 | ||
434 | static void l2c220_sync(void) | |
435 | { | |
436 | unsigned long flags; | |
437 | ||
438 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
439 | __l2c220_cache_sync(l2x0_base); | |
440 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
441 | } | |
442 | ||
443 | static void l2c220_enable(void __iomem *base, u32 aux, unsigned num_lock) | |
444 | { | |
445 | /* | |
446 | * Always enable non-secure access to the lockdown registers - | |
447 | * we write to them as part of the L2C enable sequence so they | |
448 | * need to be accessible. | |
449 | */ | |
450 | aux |= L220_AUX_CTRL_NS_LOCKDOWN; | |
451 | ||
452 | l2c_enable(base, aux, num_lock); | |
453 | } | |
454 | ||
455 | static const struct l2c_init_data l2c220_data = { | |
456 | .type = "L2C-220", | |
457 | .way_size_0 = SZ_8K, | |
458 | .num_lock = 1, | |
459 | .enable = l2c220_enable, | |
460 | .save = l2c_save, | |
461 | .outer_cache = { | |
462 | .inv_range = l2c220_inv_range, | |
463 | .clean_range = l2c220_clean_range, | |
464 | .flush_range = l2c220_flush_range, | |
465 | .flush_all = l2c220_flush_all, | |
466 | .disable = l2c_disable, | |
467 | .sync = l2c220_sync, | |
468 | .resume = l2c210_resume, | |
469 | }, | |
470 | }; | |
471 | ||
472 | /* | |
473 | * L2C-310 specific code. | |
474 | * | |
475 | * Very similar to L2C-210, the PA, set/way and sync operations are atomic, | |
476 | * and the way operations are all background tasks. However, issuing an | |
477 | * operation while a background operation is in progress results in a | |
478 | * SLVERR response. We can reuse: | |
479 | * | |
480 | * __l2c210_cache_sync (using sync_reg_offset) | |
481 | * l2c210_sync | |
482 | * l2c210_inv_range (if 588369 is not applicable) | |
483 | * l2c210_clean_range | |
484 | * l2c210_flush_range (if 588369 is not applicable) | |
485 | * l2c210_flush_all (if 727915 is not applicable) | |
486 | * | |
487 | * Errata: | |
488 | * 588369: PL310 R0P0->R1P0, fixed R2P0. | |
489 | * Affects: all clean+invalidate operations | |
490 | * clean and invalidate skips the invalidate step, so we need to issue | |
491 | * separate operations. We also require the above debug workaround | |
492 | * enclosing this code fragment on affected parts. On unaffected parts, | |
493 | * we must not use this workaround without the debug register writes | |
494 | * to avoid exposing a problem similar to 727915. | |
495 | * | |
496 | * 727915: PL310 R2P0->R3P0, fixed R3P1. | |
497 | * Affects: clean+invalidate by way | |
498 | * clean and invalidate by way runs in the background, and a store can | |
499 | * hit the line between the clean operation and invalidate operation, | |
500 | * resulting in the store being lost. | |
501 | * | |
502 | * 752271: PL310 R3P0->R3P1-50REL0, fixed R3P2. | |
503 | * Affects: 8x64-bit (double fill) line fetches | |
504 | * double fill line fetches can fail to cause dirty data to be evicted | |
505 | * from the cache before the new data overwrites the second line. | |
506 | * | |
507 | * 753970: PL310 R3P0, fixed R3P1. | |
508 | * Affects: sync | |
509 | * prevents merging writes after the sync operation, until another L2C | |
510 | * operation is performed (or a number of other conditions.) | |
511 | * | |
512 | * 769419: PL310 R0P0->R3P1, fixed R3P2. | |
513 | * Affects: store buffer | |
514 | * store buffer is not automatically drained. | |
515 | */ | |
516 | static void l2c310_inv_range_erratum(unsigned long start, unsigned long end) | |
517 | { | |
518 | void __iomem *base = l2x0_base; | |
519 | ||
520 | if ((start | end) & (CACHE_LINE_SIZE - 1)) { | |
521 | unsigned long flags; | |
522 | ||
523 | /* Erratum 588369 for both clean+invalidate operations */ | |
524 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
525 | l2c_set_debug(base, 0x03); | |
526 | ||
527 | if (start & (CACHE_LINE_SIZE - 1)) { | |
528 | start &= ~(CACHE_LINE_SIZE - 1); | |
529 | writel_relaxed(start, base + L2X0_CLEAN_LINE_PA); | |
530 | writel_relaxed(start, base + L2X0_INV_LINE_PA); | |
531 | start += CACHE_LINE_SIZE; | |
532 | } | |
533 | ||
534 | if (end & (CACHE_LINE_SIZE - 1)) { | |
535 | end &= ~(CACHE_LINE_SIZE - 1); | |
536 | writel_relaxed(end, base + L2X0_CLEAN_LINE_PA); | |
537 | writel_relaxed(end, base + L2X0_INV_LINE_PA); | |
538 | } | |
539 | ||
540 | l2c_set_debug(base, 0x00); | |
541 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
542 | } | |
543 | ||
544 | __l2c210_op_pa_range(base + L2X0_INV_LINE_PA, start, end); | |
545 | __l2c210_cache_sync(base); | |
546 | } | |
547 | ||
548 | static void l2c310_flush_range_erratum(unsigned long start, unsigned long end) | |
549 | { | |
550 | raw_spinlock_t *lock = &l2x0_lock; | |
551 | unsigned long flags; | |
552 | void __iomem *base = l2x0_base; | |
553 | ||
554 | raw_spin_lock_irqsave(lock, flags); | |
555 | while (start < end) { | |
556 | unsigned long blk_end = start + min(end - start, 4096UL); | |
557 | ||
558 | l2c_set_debug(base, 0x03); | |
559 | while (start < blk_end) { | |
560 | writel_relaxed(start, base + L2X0_CLEAN_LINE_PA); | |
561 | writel_relaxed(start, base + L2X0_INV_LINE_PA); | |
562 | start += CACHE_LINE_SIZE; | |
563 | } | |
564 | l2c_set_debug(base, 0x00); | |
565 | ||
566 | if (blk_end < end) { | |
567 | raw_spin_unlock_irqrestore(lock, flags); | |
568 | raw_spin_lock_irqsave(lock, flags); | |
569 | } | |
570 | } | |
571 | raw_spin_unlock_irqrestore(lock, flags); | |
572 | __l2c210_cache_sync(base); | |
573 | } | |
574 | ||
575 | static void l2c310_flush_all_erratum(void) | |
576 | { | |
577 | void __iomem *base = l2x0_base; | |
578 | unsigned long flags; | |
579 | ||
580 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
581 | l2c_set_debug(base, 0x03); | |
582 | __l2c_op_way(base + L2X0_CLEAN_INV_WAY); | |
583 | l2c_set_debug(base, 0x00); | |
584 | __l2c210_cache_sync(base); | |
585 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
586 | } | |
587 | ||
588 | static void __init l2c310_save(void __iomem *base) | |
589 | { | |
590 | unsigned revision; | |
591 | ||
592 | l2c_save(base); | |
593 | ||
594 | l2x0_saved_regs.tag_latency = readl_relaxed(base + | |
595 | L310_TAG_LATENCY_CTRL); | |
596 | l2x0_saved_regs.data_latency = readl_relaxed(base + | |
597 | L310_DATA_LATENCY_CTRL); | |
598 | l2x0_saved_regs.filter_end = readl_relaxed(base + | |
599 | L310_ADDR_FILTER_END); | |
600 | l2x0_saved_regs.filter_start = readl_relaxed(base + | |
601 | L310_ADDR_FILTER_START); | |
602 | ||
603 | revision = readl_relaxed(base + L2X0_CACHE_ID) & | |
604 | L2X0_CACHE_ID_RTL_MASK; | |
605 | ||
606 | /* From r2p0, there is Prefetch offset/control register */ | |
607 | if (revision >= L310_CACHE_ID_RTL_R2P0) | |
608 | l2x0_saved_regs.prefetch_ctrl = readl_relaxed(base + | |
609 | L310_PREFETCH_CTRL); | |
610 | ||
611 | /* From r3p0, there is Power control register */ | |
612 | if (revision >= L310_CACHE_ID_RTL_R3P0) | |
613 | l2x0_saved_regs.pwr_ctrl = readl_relaxed(base + | |
614 | L310_POWER_CTRL); | |
615 | } | |
616 | ||
617 | static void l2c310_resume(void) | |
618 | { | |
619 | void __iomem *base = l2x0_base; | |
620 | ||
621 | if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN)) { | |
622 | unsigned revision; | |
623 | ||
624 | /* restore pl310 setup */ | |
625 | writel_relaxed(l2x0_saved_regs.tag_latency, | |
626 | base + L310_TAG_LATENCY_CTRL); | |
627 | writel_relaxed(l2x0_saved_regs.data_latency, | |
628 | base + L310_DATA_LATENCY_CTRL); | |
629 | writel_relaxed(l2x0_saved_regs.filter_end, | |
630 | base + L310_ADDR_FILTER_END); | |
631 | writel_relaxed(l2x0_saved_regs.filter_start, | |
632 | base + L310_ADDR_FILTER_START); | |
633 | ||
634 | revision = readl_relaxed(base + L2X0_CACHE_ID) & | |
635 | L2X0_CACHE_ID_RTL_MASK; | |
636 | ||
637 | if (revision >= L310_CACHE_ID_RTL_R2P0) | |
638 | l2c_write_sec(l2x0_saved_regs.prefetch_ctrl, base, | |
639 | L310_PREFETCH_CTRL); | |
640 | if (revision >= L310_CACHE_ID_RTL_R3P0) | |
641 | l2c_write_sec(l2x0_saved_regs.pwr_ctrl, base, | |
642 | L310_POWER_CTRL); | |
643 | ||
644 | l2c_enable(base, l2x0_saved_regs.aux_ctrl, 8); | |
645 | ||
646 | /* Re-enable full-line-of-zeros for Cortex-A9 */ | |
647 | if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO) | |
648 | set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1)); | |
649 | } | |
650 | } | |
651 | ||
652 | static int l2c310_cpu_enable_flz(struct notifier_block *nb, unsigned long act, void *data) | |
653 | { | |
654 | switch (act & ~CPU_TASKS_FROZEN) { | |
655 | case CPU_STARTING: | |
656 | set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1)); | |
657 | break; | |
658 | case CPU_DYING: | |
659 | set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1))); | |
660 | break; | |
661 | } | |
662 | return NOTIFY_OK; | |
663 | } | |
664 | ||
665 | static void __init l2c310_enable(void __iomem *base, u32 aux, unsigned num_lock) | |
666 | { | |
667 | unsigned rev = readl_relaxed(base + L2X0_CACHE_ID) & L2X0_CACHE_ID_RTL_MASK; | |
668 | bool cortex_a9 = read_cpuid_part_number() == ARM_CPU_PART_CORTEX_A9; | |
669 | ||
670 | if (rev >= L310_CACHE_ID_RTL_R2P0) { | |
671 | if (cortex_a9) { | |
672 | aux |= L310_AUX_CTRL_EARLY_BRESP; | |
673 | pr_info("L2C-310 enabling early BRESP for Cortex-A9\n"); | |
674 | } else if (aux & L310_AUX_CTRL_EARLY_BRESP) { | |
675 | pr_warn("L2C-310 early BRESP only supported with Cortex-A9\n"); | |
676 | aux &= ~L310_AUX_CTRL_EARLY_BRESP; | |
677 | } | |
678 | } | |
679 | ||
680 | if (cortex_a9) { | |
681 | u32 aux_cur = readl_relaxed(base + L2X0_AUX_CTRL); | |
682 | u32 acr = get_auxcr(); | |
683 | ||
684 | pr_debug("Cortex-A9 ACR=0x%08x\n", acr); | |
685 | ||
686 | if (acr & BIT(3) && !(aux_cur & L310_AUX_CTRL_FULL_LINE_ZERO)) | |
687 | pr_err("L2C-310: full line of zeros enabled in Cortex-A9 but not L2C-310 - invalid\n"); | |
688 | ||
689 | if (aux & L310_AUX_CTRL_FULL_LINE_ZERO && !(acr & BIT(3))) | |
690 | pr_err("L2C-310: enabling full line of zeros but not enabled in Cortex-A9\n"); | |
691 | ||
692 | if (!(aux & L310_AUX_CTRL_FULL_LINE_ZERO) && !outer_cache.write_sec) { | |
693 | aux |= L310_AUX_CTRL_FULL_LINE_ZERO; | |
694 | pr_info("L2C-310 full line of zeros enabled for Cortex-A9\n"); | |
695 | } | |
696 | } else if (aux & (L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP)) { | |
697 | pr_err("L2C-310: disabling Cortex-A9 specific feature bits\n"); | |
698 | aux &= ~(L310_AUX_CTRL_FULL_LINE_ZERO | L310_AUX_CTRL_EARLY_BRESP); | |
699 | } | |
700 | ||
701 | if (aux & (L310_AUX_CTRL_DATA_PREFETCH | L310_AUX_CTRL_INSTR_PREFETCH)) { | |
702 | u32 prefetch = readl_relaxed(base + L310_PREFETCH_CTRL); | |
703 | ||
704 | pr_info("L2C-310 %s%s prefetch enabled, offset %u lines\n", | |
705 | aux & L310_AUX_CTRL_INSTR_PREFETCH ? "I" : "", | |
706 | aux & L310_AUX_CTRL_DATA_PREFETCH ? "D" : "", | |
707 | 1 + (prefetch & L310_PREFETCH_CTRL_OFFSET_MASK)); | |
708 | } | |
709 | ||
710 | /* r3p0 or later has power control register */ | |
711 | if (rev >= L310_CACHE_ID_RTL_R3P0) { | |
712 | u32 power_ctrl; | |
713 | ||
714 | l2c_write_sec(L310_DYNAMIC_CLK_GATING_EN | L310_STNDBY_MODE_EN, | |
715 | base, L310_POWER_CTRL); | |
716 | power_ctrl = readl_relaxed(base + L310_POWER_CTRL); | |
717 | pr_info("L2C-310 dynamic clock gating %sabled, standby mode %sabled\n", | |
718 | power_ctrl & L310_DYNAMIC_CLK_GATING_EN ? "en" : "dis", | |
719 | power_ctrl & L310_STNDBY_MODE_EN ? "en" : "dis"); | |
720 | } | |
721 | ||
722 | /* | |
723 | * Always enable non-secure access to the lockdown registers - | |
724 | * we write to them as part of the L2C enable sequence so they | |
725 | * need to be accessible. | |
726 | */ | |
727 | aux |= L310_AUX_CTRL_NS_LOCKDOWN; | |
728 | ||
729 | l2c_enable(base, aux, num_lock); | |
730 | ||
731 | if (aux & L310_AUX_CTRL_FULL_LINE_ZERO) { | |
732 | set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1)); | |
733 | cpu_notifier(l2c310_cpu_enable_flz, 0); | |
734 | } | |
735 | } | |
736 | ||
737 | static void __init l2c310_fixup(void __iomem *base, u32 cache_id, | |
738 | struct outer_cache_fns *fns) | |
739 | { | |
740 | unsigned revision = cache_id & L2X0_CACHE_ID_RTL_MASK; | |
741 | const char *errata[8]; | |
742 | unsigned n = 0; | |
743 | ||
744 | if (IS_ENABLED(CONFIG_PL310_ERRATA_588369) && | |
745 | revision < L310_CACHE_ID_RTL_R2P0 && | |
746 | /* For bcm compatibility */ | |
747 | fns->inv_range == l2c210_inv_range) { | |
748 | fns->inv_range = l2c310_inv_range_erratum; | |
749 | fns->flush_range = l2c310_flush_range_erratum; | |
750 | errata[n++] = "588369"; | |
751 | } | |
752 | ||
753 | if (IS_ENABLED(CONFIG_PL310_ERRATA_727915) && | |
754 | revision >= L310_CACHE_ID_RTL_R2P0 && | |
755 | revision < L310_CACHE_ID_RTL_R3P1) { | |
756 | fns->flush_all = l2c310_flush_all_erratum; | |
757 | errata[n++] = "727915"; | |
758 | } | |
759 | ||
760 | if (revision >= L310_CACHE_ID_RTL_R3P0 && | |
761 | revision < L310_CACHE_ID_RTL_R3P2) { | |
762 | u32 val = readl_relaxed(base + L310_PREFETCH_CTRL); | |
763 | /* I don't think bit23 is required here... but iMX6 does so */ | |
764 | if (val & (BIT(30) | BIT(23))) { | |
765 | val &= ~(BIT(30) | BIT(23)); | |
766 | l2c_write_sec(val, base, L310_PREFETCH_CTRL); | |
767 | errata[n++] = "752271"; | |
768 | } | |
769 | } | |
770 | ||
771 | if (IS_ENABLED(CONFIG_PL310_ERRATA_753970) && | |
772 | revision == L310_CACHE_ID_RTL_R3P0) { | |
773 | sync_reg_offset = L2X0_DUMMY_REG; | |
774 | errata[n++] = "753970"; | |
775 | } | |
776 | ||
777 | if (IS_ENABLED(CONFIG_PL310_ERRATA_769419)) | |
778 | errata[n++] = "769419"; | |
779 | ||
780 | if (n) { | |
781 | unsigned i; | |
782 | ||
783 | pr_info("L2C-310 errat%s", n > 1 ? "a" : "um"); | |
784 | for (i = 0; i < n; i++) | |
785 | pr_cont(" %s", errata[i]); | |
786 | pr_cont(" enabled\n"); | |
787 | } | |
788 | } | |
789 | ||
790 | static void l2c310_disable(void) | |
791 | { | |
792 | /* | |
793 | * If full-line-of-zeros is enabled, we must first disable it in the | |
794 | * Cortex-A9 auxiliary control register before disabling the L2 cache. | |
795 | */ | |
796 | if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO) | |
797 | set_auxcr(get_auxcr() & ~(BIT(3) | BIT(2) | BIT(1))); | |
798 | ||
799 | l2c_disable(); | |
800 | } | |
801 | ||
802 | static const struct l2c_init_data l2c310_init_fns __initconst = { | |
803 | .type = "L2C-310", | |
804 | .way_size_0 = SZ_8K, | |
805 | .num_lock = 8, | |
806 | .enable = l2c310_enable, | |
807 | .fixup = l2c310_fixup, | |
808 | .save = l2c310_save, | |
809 | .outer_cache = { | |
810 | .inv_range = l2c210_inv_range, | |
811 | .clean_range = l2c210_clean_range, | |
812 | .flush_range = l2c210_flush_range, | |
813 | .flush_all = l2c210_flush_all, | |
814 | .disable = l2c310_disable, | |
815 | .sync = l2c210_sync, | |
816 | .resume = l2c310_resume, | |
817 | }, | |
818 | }; | |
819 | ||
820 | static void __init __l2c_init(const struct l2c_init_data *data, | |
821 | u32 aux_val, u32 aux_mask, u32 cache_id) | |
822 | { | |
823 | struct outer_cache_fns fns; | |
824 | unsigned way_size_bits, ways; | |
825 | u32 aux, old_aux; | |
826 | ||
827 | /* | |
828 | * Sanity check the aux values. aux_mask is the bits we preserve | |
829 | * from reading the hardware register, and aux_val is the bits we | |
830 | * set. | |
831 | */ | |
832 | if (aux_val & aux_mask) | |
833 | pr_alert("L2C: platform provided aux values permit register corruption.\n"); | |
834 | ||
835 | old_aux = aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL); | |
836 | aux &= aux_mask; | |
837 | aux |= aux_val; | |
838 | ||
839 | if (old_aux != aux) | |
840 | pr_warn("L2C: DT/platform modifies aux control register: 0x%08x -> 0x%08x\n", | |
841 | old_aux, aux); | |
842 | ||
843 | /* Determine the number of ways */ | |
844 | switch (cache_id & L2X0_CACHE_ID_PART_MASK) { | |
845 | case L2X0_CACHE_ID_PART_L310: | |
846 | if ((aux_val | ~aux_mask) & (L2C_AUX_CTRL_WAY_SIZE_MASK | L310_AUX_CTRL_ASSOCIATIVITY_16)) | |
847 | pr_warn("L2C: DT/platform tries to modify or specify cache size\n"); | |
848 | if (aux & (1 << 16)) | |
849 | ways = 16; | |
850 | else | |
851 | ways = 8; | |
852 | break; | |
853 | ||
854 | case L2X0_CACHE_ID_PART_L210: | |
855 | case L2X0_CACHE_ID_PART_L220: | |
856 | ways = (aux >> 13) & 0xf; | |
857 | break; | |
858 | ||
859 | case AURORA_CACHE_ID: | |
860 | ways = (aux >> 13) & 0xf; | |
861 | ways = 2 << ((ways + 1) >> 2); | |
862 | break; | |
863 | ||
864 | default: | |
865 | /* Assume unknown chips have 8 ways */ | |
866 | ways = 8; | |
867 | break; | |
868 | } | |
869 | ||
870 | l2x0_way_mask = (1 << ways) - 1; | |
871 | ||
872 | /* | |
873 | * way_size_0 is the size that a way_size value of zero would be | |
874 | * given the calculation: way_size = way_size_0 << way_size_bits. | |
875 | * So, if way_size_bits=0 is reserved, but way_size_bits=1 is 16k, | |
876 | * then way_size_0 would be 8k. | |
877 | * | |
878 | * L2 cache size = number of ways * way size. | |
879 | */ | |
880 | way_size_bits = (aux & L2C_AUX_CTRL_WAY_SIZE_MASK) >> | |
881 | L2C_AUX_CTRL_WAY_SIZE_SHIFT; | |
882 | l2x0_size = ways * (data->way_size_0 << way_size_bits); | |
883 | ||
884 | fns = data->outer_cache; | |
885 | fns.write_sec = outer_cache.write_sec; | |
886 | if (data->fixup) | |
887 | data->fixup(l2x0_base, cache_id, &fns); | |
888 | ||
889 | /* | |
890 | * Check if l2x0 controller is already enabled. If we are booting | |
891 | * in non-secure mode accessing the below registers will fault. | |
892 | */ | |
893 | if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN)) | |
894 | data->enable(l2x0_base, aux, data->num_lock); | |
895 | ||
896 | outer_cache = fns; | |
897 | ||
898 | /* | |
899 | * It is strange to save the register state before initialisation, | |
900 | * but hey, this is what the DT implementations decided to do. | |
901 | */ | |
902 | if (data->save) | |
903 | data->save(l2x0_base); | |
904 | ||
905 | /* Re-read it in case some bits are reserved. */ | |
906 | aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL); | |
907 | ||
908 | pr_info("%s cache controller enabled, %d ways, %d kB\n", | |
909 | data->type, ways, l2x0_size >> 10); | |
910 | pr_info("%s: CACHE_ID 0x%08x, AUX_CTRL 0x%08x\n", | |
911 | data->type, cache_id, aux); | |
912 | } | |
913 | ||
914 | void __init l2x0_init(void __iomem *base, u32 aux_val, u32 aux_mask) | |
915 | { | |
916 | const struct l2c_init_data *data; | |
917 | u32 cache_id; | |
918 | ||
919 | l2x0_base = base; | |
920 | ||
921 | cache_id = readl_relaxed(base + L2X0_CACHE_ID); | |
922 | ||
923 | switch (cache_id & L2X0_CACHE_ID_PART_MASK) { | |
924 | default: | |
925 | case L2X0_CACHE_ID_PART_L210: | |
926 | data = &l2c210_data; | |
927 | break; | |
928 | ||
929 | case L2X0_CACHE_ID_PART_L220: | |
930 | data = &l2c220_data; | |
931 | break; | |
932 | ||
933 | case L2X0_CACHE_ID_PART_L310: | |
934 | data = &l2c310_init_fns; | |
935 | break; | |
936 | } | |
937 | ||
938 | __l2c_init(data, aux_val, aux_mask, cache_id); | |
939 | } | |
940 | ||
941 | #ifdef CONFIG_OF | |
942 | static int l2_wt_override; | |
943 | ||
944 | /* Aurora don't have the cache ID register available, so we have to | |
945 | * pass it though the device tree */ | |
946 | static u32 cache_id_part_number_from_dt; | |
947 | ||
948 | static void __init l2x0_of_parse(const struct device_node *np, | |
949 | u32 *aux_val, u32 *aux_mask) | |
950 | { | |
951 | u32 data[2] = { 0, 0 }; | |
952 | u32 tag = 0; | |
953 | u32 dirty = 0; | |
954 | u32 val = 0, mask = 0; | |
955 | ||
956 | of_property_read_u32(np, "arm,tag-latency", &tag); | |
957 | if (tag) { | |
958 | mask |= L2X0_AUX_CTRL_TAG_LATENCY_MASK; | |
959 | val |= (tag - 1) << L2X0_AUX_CTRL_TAG_LATENCY_SHIFT; | |
960 | } | |
961 | ||
962 | of_property_read_u32_array(np, "arm,data-latency", | |
963 | data, ARRAY_SIZE(data)); | |
964 | if (data[0] && data[1]) { | |
965 | mask |= L2X0_AUX_CTRL_DATA_RD_LATENCY_MASK | | |
966 | L2X0_AUX_CTRL_DATA_WR_LATENCY_MASK; | |
967 | val |= ((data[0] - 1) << L2X0_AUX_CTRL_DATA_RD_LATENCY_SHIFT) | | |
968 | ((data[1] - 1) << L2X0_AUX_CTRL_DATA_WR_LATENCY_SHIFT); | |
969 | } | |
970 | ||
971 | of_property_read_u32(np, "arm,dirty-latency", &dirty); | |
972 | if (dirty) { | |
973 | mask |= L2X0_AUX_CTRL_DIRTY_LATENCY_MASK; | |
974 | val |= (dirty - 1) << L2X0_AUX_CTRL_DIRTY_LATENCY_SHIFT; | |
975 | } | |
976 | ||
977 | *aux_val &= ~mask; | |
978 | *aux_val |= val; | |
979 | *aux_mask &= ~mask; | |
980 | } | |
981 | ||
982 | static const struct l2c_init_data of_l2c210_data __initconst = { | |
983 | .type = "L2C-210", | |
984 | .way_size_0 = SZ_8K, | |
985 | .num_lock = 1, | |
986 | .of_parse = l2x0_of_parse, | |
987 | .enable = l2c_enable, | |
988 | .save = l2c_save, | |
989 | .outer_cache = { | |
990 | .inv_range = l2c210_inv_range, | |
991 | .clean_range = l2c210_clean_range, | |
992 | .flush_range = l2c210_flush_range, | |
993 | .flush_all = l2c210_flush_all, | |
994 | .disable = l2c_disable, | |
995 | .sync = l2c210_sync, | |
996 | .resume = l2c210_resume, | |
997 | }, | |
998 | }; | |
999 | ||
1000 | static const struct l2c_init_data of_l2c220_data __initconst = { | |
1001 | .type = "L2C-220", | |
1002 | .way_size_0 = SZ_8K, | |
1003 | .num_lock = 1, | |
1004 | .of_parse = l2x0_of_parse, | |
1005 | .enable = l2c220_enable, | |
1006 | .save = l2c_save, | |
1007 | .outer_cache = { | |
1008 | .inv_range = l2c220_inv_range, | |
1009 | .clean_range = l2c220_clean_range, | |
1010 | .flush_range = l2c220_flush_range, | |
1011 | .flush_all = l2c220_flush_all, | |
1012 | .disable = l2c_disable, | |
1013 | .sync = l2c220_sync, | |
1014 | .resume = l2c210_resume, | |
1015 | }, | |
1016 | }; | |
1017 | ||
1018 | static void __init l2c310_of_parse(const struct device_node *np, | |
1019 | u32 *aux_val, u32 *aux_mask) | |
1020 | { | |
1021 | u32 data[3] = { 0, 0, 0 }; | |
1022 | u32 tag[3] = { 0, 0, 0 }; | |
1023 | u32 filter[2] = { 0, 0 }; | |
1024 | ||
1025 | of_property_read_u32_array(np, "arm,tag-latency", tag, ARRAY_SIZE(tag)); | |
1026 | if (tag[0] && tag[1] && tag[2]) | |
1027 | writel_relaxed( | |
1028 | L310_LATENCY_CTRL_RD(tag[0] - 1) | | |
1029 | L310_LATENCY_CTRL_WR(tag[1] - 1) | | |
1030 | L310_LATENCY_CTRL_SETUP(tag[2] - 1), | |
1031 | l2x0_base + L310_TAG_LATENCY_CTRL); | |
1032 | ||
1033 | of_property_read_u32_array(np, "arm,data-latency", | |
1034 | data, ARRAY_SIZE(data)); | |
1035 | if (data[0] && data[1] && data[2]) | |
1036 | writel_relaxed( | |
1037 | L310_LATENCY_CTRL_RD(data[0] - 1) | | |
1038 | L310_LATENCY_CTRL_WR(data[1] - 1) | | |
1039 | L310_LATENCY_CTRL_SETUP(data[2] - 1), | |
1040 | l2x0_base + L310_DATA_LATENCY_CTRL); | |
1041 | ||
1042 | of_property_read_u32_array(np, "arm,filter-ranges", | |
1043 | filter, ARRAY_SIZE(filter)); | |
1044 | if (filter[1]) { | |
1045 | writel_relaxed(ALIGN(filter[0] + filter[1], SZ_1M), | |
1046 | l2x0_base + L310_ADDR_FILTER_END); | |
1047 | writel_relaxed((filter[0] & ~(SZ_1M - 1)) | L310_ADDR_FILTER_EN, | |
1048 | l2x0_base + L310_ADDR_FILTER_START); | |
1049 | } | |
1050 | } | |
1051 | ||
1052 | static const struct l2c_init_data of_l2c310_data __initconst = { | |
1053 | .type = "L2C-310", | |
1054 | .way_size_0 = SZ_8K, | |
1055 | .num_lock = 8, | |
1056 | .of_parse = l2c310_of_parse, | |
1057 | .enable = l2c310_enable, | |
1058 | .fixup = l2c310_fixup, | |
1059 | .save = l2c310_save, | |
1060 | .outer_cache = { | |
1061 | .inv_range = l2c210_inv_range, | |
1062 | .clean_range = l2c210_clean_range, | |
1063 | .flush_range = l2c210_flush_range, | |
1064 | .flush_all = l2c210_flush_all, | |
1065 | .disable = l2c310_disable, | |
1066 | .sync = l2c210_sync, | |
1067 | .resume = l2c310_resume, | |
1068 | }, | |
1069 | }; | |
1070 | ||
1071 | /* | |
1072 | * This is a variant of the of_l2c310_data with .sync set to | |
1073 | * NULL. Outer sync operations are not needed when the system is I/O | |
1074 | * coherent, and potentially harmful in certain situations (PCIe/PL310 | |
1075 | * deadlock on Armada 375/38x due to hardware I/O coherency). The | |
1076 | * other operations are kept because they are infrequent (therefore do | |
1077 | * not cause the deadlock in practice) and needed for secondary CPU | |
1078 | * boot and other power management activities. | |
1079 | */ | |
1080 | static const struct l2c_init_data of_l2c310_coherent_data __initconst = { | |
1081 | .type = "L2C-310 Coherent", | |
1082 | .way_size_0 = SZ_8K, | |
1083 | .num_lock = 8, | |
1084 | .of_parse = l2c310_of_parse, | |
1085 | .enable = l2c310_enable, | |
1086 | .fixup = l2c310_fixup, | |
1087 | .save = l2c310_save, | |
1088 | .outer_cache = { | |
1089 | .inv_range = l2c210_inv_range, | |
1090 | .clean_range = l2c210_clean_range, | |
1091 | .flush_range = l2c210_flush_range, | |
1092 | .flush_all = l2c210_flush_all, | |
1093 | .disable = l2c310_disable, | |
1094 | .resume = l2c310_resume, | |
1095 | }, | |
1096 | }; | |
1097 | ||
1098 | /* | |
1099 | * Note that the end addresses passed to Linux primitives are | |
1100 | * noninclusive, while the hardware cache range operations use | |
1101 | * inclusive start and end addresses. | |
1102 | */ | |
1103 | static unsigned long calc_range_end(unsigned long start, unsigned long end) | |
1104 | { | |
1105 | /* | |
1106 | * Limit the number of cache lines processed at once, | |
1107 | * since cache range operations stall the CPU pipeline | |
1108 | * until completion. | |
1109 | */ | |
1110 | if (end > start + MAX_RANGE_SIZE) | |
1111 | end = start + MAX_RANGE_SIZE; | |
1112 | ||
1113 | /* | |
1114 | * Cache range operations can't straddle a page boundary. | |
1115 | */ | |
1116 | if (end > PAGE_ALIGN(start+1)) | |
1117 | end = PAGE_ALIGN(start+1); | |
1118 | ||
1119 | return end; | |
1120 | } | |
1121 | ||
1122 | /* | |
1123 | * Make sure 'start' and 'end' reference the same page, as L2 is PIPT | |
1124 | * and range operations only do a TLB lookup on the start address. | |
1125 | */ | |
1126 | static void aurora_pa_range(unsigned long start, unsigned long end, | |
1127 | unsigned long offset) | |
1128 | { | |
1129 | unsigned long flags; | |
1130 | ||
1131 | raw_spin_lock_irqsave(&l2x0_lock, flags); | |
1132 | writel_relaxed(start, l2x0_base + AURORA_RANGE_BASE_ADDR_REG); | |
1133 | writel_relaxed(end, l2x0_base + offset); | |
1134 | raw_spin_unlock_irqrestore(&l2x0_lock, flags); | |
1135 | ||
1136 | cache_sync(); | |
1137 | } | |
1138 | ||
1139 | static void aurora_inv_range(unsigned long start, unsigned long end) | |
1140 | { | |
1141 | /* | |
1142 | * round start and end adresses up to cache line size | |
1143 | */ | |
1144 | start &= ~(CACHE_LINE_SIZE - 1); | |
1145 | end = ALIGN(end, CACHE_LINE_SIZE); | |
1146 | ||
1147 | /* | |
1148 | * Invalidate all full cache lines between 'start' and 'end'. | |
1149 | */ | |
1150 | while (start < end) { | |
1151 | unsigned long range_end = calc_range_end(start, end); | |
1152 | aurora_pa_range(start, range_end - CACHE_LINE_SIZE, | |
1153 | AURORA_INVAL_RANGE_REG); | |
1154 | start = range_end; | |
1155 | } | |
1156 | } | |
1157 | ||
1158 | static void aurora_clean_range(unsigned long start, unsigned long end) | |
1159 | { | |
1160 | /* | |
1161 | * If L2 is forced to WT, the L2 will always be clean and we | |
1162 | * don't need to do anything here. | |
1163 | */ | |
1164 | if (!l2_wt_override) { | |
1165 | start &= ~(CACHE_LINE_SIZE - 1); | |
1166 | end = ALIGN(end, CACHE_LINE_SIZE); | |
1167 | while (start != end) { | |
1168 | unsigned long range_end = calc_range_end(start, end); | |
1169 | aurora_pa_range(start, range_end - CACHE_LINE_SIZE, | |
1170 | AURORA_CLEAN_RANGE_REG); | |
1171 | start = range_end; | |
1172 | } | |
1173 | } | |
1174 | } | |
1175 | ||
1176 | static void aurora_flush_range(unsigned long start, unsigned long end) | |
1177 | { | |
1178 | start &= ~(CACHE_LINE_SIZE - 1); | |
1179 | end = ALIGN(end, CACHE_LINE_SIZE); | |
1180 | while (start != end) { | |
1181 | unsigned long range_end = calc_range_end(start, end); | |
1182 | /* | |
1183 | * If L2 is forced to WT, the L2 will always be clean and we | |
1184 | * just need to invalidate. | |
1185 | */ | |
1186 | if (l2_wt_override) | |
1187 | aurora_pa_range(start, range_end - CACHE_LINE_SIZE, | |
1188 | AURORA_INVAL_RANGE_REG); | |
1189 | else | |
1190 | aurora_pa_range(start, range_end - CACHE_LINE_SIZE, | |
1191 | AURORA_FLUSH_RANGE_REG); | |
1192 | start = range_end; | |
1193 | } | |
1194 | } | |
1195 | ||
1196 | static void aurora_save(void __iomem *base) | |
1197 | { | |
1198 | l2x0_saved_regs.ctrl = readl_relaxed(base + L2X0_CTRL); | |
1199 | l2x0_saved_regs.aux_ctrl = readl_relaxed(base + L2X0_AUX_CTRL); | |
1200 | } | |
1201 | ||
1202 | static void aurora_resume(void) | |
1203 | { | |
1204 | void __iomem *base = l2x0_base; | |
1205 | ||
1206 | if (!(readl(base + L2X0_CTRL) & L2X0_CTRL_EN)) { | |
1207 | writel_relaxed(l2x0_saved_regs.aux_ctrl, base + L2X0_AUX_CTRL); | |
1208 | writel_relaxed(l2x0_saved_regs.ctrl, base + L2X0_CTRL); | |
1209 | } | |
1210 | } | |
1211 | ||
1212 | /* | |
1213 | * For Aurora cache in no outer mode, enable via the CP15 coprocessor | |
1214 | * broadcasting of cache commands to L2. | |
1215 | */ | |
1216 | static void __init aurora_enable_no_outer(void __iomem *base, u32 aux, | |
1217 | unsigned num_lock) | |
1218 | { | |
1219 | u32 u; | |
1220 | ||
1221 | asm volatile("mrc p15, 1, %0, c15, c2, 0" : "=r" (u)); | |
1222 | u |= AURORA_CTRL_FW; /* Set the FW bit */ | |
1223 | asm volatile("mcr p15, 1, %0, c15, c2, 0" : : "r" (u)); | |
1224 | ||
1225 | isb(); | |
1226 | ||
1227 | l2c_enable(base, aux, num_lock); | |
1228 | } | |
1229 | ||
1230 | static void __init aurora_fixup(void __iomem *base, u32 cache_id, | |
1231 | struct outer_cache_fns *fns) | |
1232 | { | |
1233 | sync_reg_offset = AURORA_SYNC_REG; | |
1234 | } | |
1235 | ||
1236 | static void __init aurora_of_parse(const struct device_node *np, | |
1237 | u32 *aux_val, u32 *aux_mask) | |
1238 | { | |
1239 | u32 val = AURORA_ACR_REPLACEMENT_TYPE_SEMIPLRU; | |
1240 | u32 mask = AURORA_ACR_REPLACEMENT_MASK; | |
1241 | ||
1242 | of_property_read_u32(np, "cache-id-part", | |
1243 | &cache_id_part_number_from_dt); | |
1244 | ||
1245 | /* Determine and save the write policy */ | |
1246 | l2_wt_override = of_property_read_bool(np, "wt-override"); | |
1247 | ||
1248 | if (l2_wt_override) { | |
1249 | val |= AURORA_ACR_FORCE_WRITE_THRO_POLICY; | |
1250 | mask |= AURORA_ACR_FORCE_WRITE_POLICY_MASK; | |
1251 | } | |
1252 | ||
1253 | *aux_val &= ~mask; | |
1254 | *aux_val |= val; | |
1255 | *aux_mask &= ~mask; | |
1256 | } | |
1257 | ||
1258 | static const struct l2c_init_data of_aurora_with_outer_data __initconst = { | |
1259 | .type = "Aurora", | |
1260 | .way_size_0 = SZ_4K, | |
1261 | .num_lock = 4, | |
1262 | .of_parse = aurora_of_parse, | |
1263 | .enable = l2c_enable, | |
1264 | .fixup = aurora_fixup, | |
1265 | .save = aurora_save, | |
1266 | .outer_cache = { | |
1267 | .inv_range = aurora_inv_range, | |
1268 | .clean_range = aurora_clean_range, | |
1269 | .flush_range = aurora_flush_range, | |
1270 | .flush_all = l2x0_flush_all, | |
1271 | .disable = l2x0_disable, | |
1272 | .sync = l2x0_cache_sync, | |
1273 | .resume = aurora_resume, | |
1274 | }, | |
1275 | }; | |
1276 | ||
1277 | static const struct l2c_init_data of_aurora_no_outer_data __initconst = { | |
1278 | .type = "Aurora", | |
1279 | .way_size_0 = SZ_4K, | |
1280 | .num_lock = 4, | |
1281 | .of_parse = aurora_of_parse, | |
1282 | .enable = aurora_enable_no_outer, | |
1283 | .fixup = aurora_fixup, | |
1284 | .save = aurora_save, | |
1285 | .outer_cache = { | |
1286 | .resume = aurora_resume, | |
1287 | }, | |
1288 | }; | |
1289 | ||
1290 | /* | |
1291 | * For certain Broadcom SoCs, depending on the address range, different offsets | |
1292 | * need to be added to the address before passing it to L2 for | |
1293 | * invalidation/clean/flush | |
1294 | * | |
1295 | * Section Address Range Offset EMI | |
1296 | * 1 0x00000000 - 0x3FFFFFFF 0x80000000 VC | |
1297 | * 2 0x40000000 - 0xBFFFFFFF 0x40000000 SYS | |
1298 | * 3 0xC0000000 - 0xFFFFFFFF 0x80000000 VC | |
1299 | * | |
1300 | * When the start and end addresses have crossed two different sections, we | |
1301 | * need to break the L2 operation into two, each within its own section. | |
1302 | * For example, if we need to invalidate addresses starts at 0xBFFF0000 and | |
1303 | * ends at 0xC0001000, we need do invalidate 1) 0xBFFF0000 - 0xBFFFFFFF and 2) | |
1304 | * 0xC0000000 - 0xC0001000 | |
1305 | * | |
1306 | * Note 1: | |
1307 | * By breaking a single L2 operation into two, we may potentially suffer some | |
1308 | * performance hit, but keep in mind the cross section case is very rare | |
1309 | * | |
1310 | * Note 2: | |
1311 | * We do not need to handle the case when the start address is in | |
1312 | * Section 1 and the end address is in Section 3, since it is not a valid use | |
1313 | * case | |
1314 | * | |
1315 | * Note 3: | |
1316 | * Section 1 in practical terms can no longer be used on rev A2. Because of | |
1317 | * that the code does not need to handle section 1 at all. | |
1318 | * | |
1319 | */ | |
1320 | #define BCM_SYS_EMI_START_ADDR 0x40000000UL | |
1321 | #define BCM_VC_EMI_SEC3_START_ADDR 0xC0000000UL | |
1322 | ||
1323 | #define BCM_SYS_EMI_OFFSET 0x40000000UL | |
1324 | #define BCM_VC_EMI_OFFSET 0x80000000UL | |
1325 | ||
1326 | static inline int bcm_addr_is_sys_emi(unsigned long addr) | |
1327 | { | |
1328 | return (addr >= BCM_SYS_EMI_START_ADDR) && | |
1329 | (addr < BCM_VC_EMI_SEC3_START_ADDR); | |
1330 | } | |
1331 | ||
1332 | static inline unsigned long bcm_l2_phys_addr(unsigned long addr) | |
1333 | { | |
1334 | if (bcm_addr_is_sys_emi(addr)) | |
1335 | return addr + BCM_SYS_EMI_OFFSET; | |
1336 | else | |
1337 | return addr + BCM_VC_EMI_OFFSET; | |
1338 | } | |
1339 | ||
1340 | static void bcm_inv_range(unsigned long start, unsigned long end) | |
1341 | { | |
1342 | unsigned long new_start, new_end; | |
1343 | ||
1344 | BUG_ON(start < BCM_SYS_EMI_START_ADDR); | |
1345 | ||
1346 | if (unlikely(end <= start)) | |
1347 | return; | |
1348 | ||
1349 | new_start = bcm_l2_phys_addr(start); | |
1350 | new_end = bcm_l2_phys_addr(end); | |
1351 | ||
1352 | /* normal case, no cross section between start and end */ | |
1353 | if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) { | |
1354 | l2c210_inv_range(new_start, new_end); | |
1355 | return; | |
1356 | } | |
1357 | ||
1358 | /* They cross sections, so it can only be a cross from section | |
1359 | * 2 to section 3 | |
1360 | */ | |
1361 | l2c210_inv_range(new_start, | |
1362 | bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1)); | |
1363 | l2c210_inv_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR), | |
1364 | new_end); | |
1365 | } | |
1366 | ||
1367 | static void bcm_clean_range(unsigned long start, unsigned long end) | |
1368 | { | |
1369 | unsigned long new_start, new_end; | |
1370 | ||
1371 | BUG_ON(start < BCM_SYS_EMI_START_ADDR); | |
1372 | ||
1373 | if (unlikely(end <= start)) | |
1374 | return; | |
1375 | ||
1376 | new_start = bcm_l2_phys_addr(start); | |
1377 | new_end = bcm_l2_phys_addr(end); | |
1378 | ||
1379 | /* normal case, no cross section between start and end */ | |
1380 | if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) { | |
1381 | l2c210_clean_range(new_start, new_end); | |
1382 | return; | |
1383 | } | |
1384 | ||
1385 | /* They cross sections, so it can only be a cross from section | |
1386 | * 2 to section 3 | |
1387 | */ | |
1388 | l2c210_clean_range(new_start, | |
1389 | bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1)); | |
1390 | l2c210_clean_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR), | |
1391 | new_end); | |
1392 | } | |
1393 | ||
1394 | static void bcm_flush_range(unsigned long start, unsigned long end) | |
1395 | { | |
1396 | unsigned long new_start, new_end; | |
1397 | ||
1398 | BUG_ON(start < BCM_SYS_EMI_START_ADDR); | |
1399 | ||
1400 | if (unlikely(end <= start)) | |
1401 | return; | |
1402 | ||
1403 | if ((end - start) >= l2x0_size) { | |
1404 | outer_cache.flush_all(); | |
1405 | return; | |
1406 | } | |
1407 | ||
1408 | new_start = bcm_l2_phys_addr(start); | |
1409 | new_end = bcm_l2_phys_addr(end); | |
1410 | ||
1411 | /* normal case, no cross section between start and end */ | |
1412 | if (likely(bcm_addr_is_sys_emi(end) || !bcm_addr_is_sys_emi(start))) { | |
1413 | l2c210_flush_range(new_start, new_end); | |
1414 | return; | |
1415 | } | |
1416 | ||
1417 | /* They cross sections, so it can only be a cross from section | |
1418 | * 2 to section 3 | |
1419 | */ | |
1420 | l2c210_flush_range(new_start, | |
1421 | bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR-1)); | |
1422 | l2c210_flush_range(bcm_l2_phys_addr(BCM_VC_EMI_SEC3_START_ADDR), | |
1423 | new_end); | |
1424 | } | |
1425 | ||
1426 | /* Broadcom L2C-310 start from ARMs R3P2 or later, and require no fixups */ | |
1427 | static const struct l2c_init_data of_bcm_l2x0_data __initconst = { | |
1428 | .type = "BCM-L2C-310", | |
1429 | .way_size_0 = SZ_8K, | |
1430 | .num_lock = 8, | |
1431 | .of_parse = l2c310_of_parse, | |
1432 | .enable = l2c310_enable, | |
1433 | .save = l2c310_save, | |
1434 | .outer_cache = { | |
1435 | .inv_range = bcm_inv_range, | |
1436 | .clean_range = bcm_clean_range, | |
1437 | .flush_range = bcm_flush_range, | |
1438 | .flush_all = l2c210_flush_all, | |
1439 | .disable = l2c310_disable, | |
1440 | .sync = l2c210_sync, | |
1441 | .resume = l2c310_resume, | |
1442 | }, | |
1443 | }; | |
1444 | ||
1445 | static void __init tauros3_save(void __iomem *base) | |
1446 | { | |
1447 | l2c_save(base); | |
1448 | ||
1449 | l2x0_saved_regs.aux2_ctrl = | |
1450 | readl_relaxed(base + TAUROS3_AUX2_CTRL); | |
1451 | l2x0_saved_regs.prefetch_ctrl = | |
1452 | readl_relaxed(base + L310_PREFETCH_CTRL); | |
1453 | } | |
1454 | ||
1455 | static void tauros3_resume(void) | |
1456 | { | |
1457 | void __iomem *base = l2x0_base; | |
1458 | ||
1459 | if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN)) { | |
1460 | writel_relaxed(l2x0_saved_regs.aux2_ctrl, | |
1461 | base + TAUROS3_AUX2_CTRL); | |
1462 | writel_relaxed(l2x0_saved_regs.prefetch_ctrl, | |
1463 | base + L310_PREFETCH_CTRL); | |
1464 | ||
1465 | l2c_enable(base, l2x0_saved_regs.aux_ctrl, 8); | |
1466 | } | |
1467 | } | |
1468 | ||
1469 | static const struct l2c_init_data of_tauros3_data __initconst = { | |
1470 | .type = "Tauros3", | |
1471 | .way_size_0 = SZ_8K, | |
1472 | .num_lock = 8, | |
1473 | .enable = l2c_enable, | |
1474 | .save = tauros3_save, | |
1475 | /* Tauros3 broadcasts L1 cache operations to L2 */ | |
1476 | .outer_cache = { | |
1477 | .resume = tauros3_resume, | |
1478 | }, | |
1479 | }; | |
1480 | ||
1481 | #define L2C_ID(name, fns) { .compatible = name, .data = (void *)&fns } | |
1482 | static const struct of_device_id l2x0_ids[] __initconst = { | |
1483 | L2C_ID("arm,l210-cache", of_l2c210_data), | |
1484 | L2C_ID("arm,l220-cache", of_l2c220_data), | |
1485 | L2C_ID("arm,pl310-cache", of_l2c310_data), | |
1486 | L2C_ID("brcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data), | |
1487 | L2C_ID("marvell,aurora-outer-cache", of_aurora_with_outer_data), | |
1488 | L2C_ID("marvell,aurora-system-cache", of_aurora_no_outer_data), | |
1489 | L2C_ID("marvell,tauros3-cache", of_tauros3_data), | |
1490 | /* Deprecated IDs */ | |
1491 | L2C_ID("bcm,bcm11351-a2-pl310-cache", of_bcm_l2x0_data), | |
1492 | {} | |
1493 | }; | |
1494 | ||
1495 | int __init l2x0_of_init(u32 aux_val, u32 aux_mask) | |
1496 | { | |
1497 | const struct l2c_init_data *data; | |
1498 | struct device_node *np; | |
1499 | struct resource res; | |
1500 | u32 cache_id, old_aux; | |
1501 | ||
1502 | np = of_find_matching_node(NULL, l2x0_ids); | |
1503 | if (!np) | |
1504 | return -ENODEV; | |
1505 | ||
1506 | if (of_address_to_resource(np, 0, &res)) | |
1507 | return -ENODEV; | |
1508 | ||
1509 | l2x0_base = ioremap(res.start, resource_size(&res)); | |
1510 | if (!l2x0_base) | |
1511 | return -ENOMEM; | |
1512 | ||
1513 | l2x0_saved_regs.phy_base = res.start; | |
1514 | ||
1515 | data = of_match_node(l2x0_ids, np)->data; | |
1516 | ||
1517 | if (of_device_is_compatible(np, "arm,pl310-cache") && | |
1518 | of_property_read_bool(np, "arm,io-coherent")) | |
1519 | data = &of_l2c310_coherent_data; | |
1520 | ||
1521 | old_aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL); | |
1522 | if (old_aux != ((old_aux & aux_mask) | aux_val)) { | |
1523 | pr_warn("L2C: platform modifies aux control register: 0x%08x -> 0x%08x\n", | |
1524 | old_aux, (old_aux & aux_mask) | aux_val); | |
1525 | } else if (aux_mask != ~0U && aux_val != 0) { | |
1526 | pr_alert("L2C: platform provided aux values match the hardware, so have no effect. Please remove them.\n"); | |
1527 | } | |
1528 | ||
1529 | /* All L2 caches are unified, so this property should be specified */ | |
1530 | if (!of_property_read_bool(np, "cache-unified")) | |
1531 | pr_err("L2C: device tree omits to specify unified cache\n"); | |
1532 | ||
1533 | /* L2 configuration can only be changed if the cache is disabled */ | |
1534 | if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & L2X0_CTRL_EN)) | |
1535 | if (data->of_parse) | |
1536 | data->of_parse(np, &aux_val, &aux_mask); | |
1537 | ||
1538 | if (cache_id_part_number_from_dt) | |
1539 | cache_id = cache_id_part_number_from_dt; | |
1540 | else | |
1541 | cache_id = readl_relaxed(l2x0_base + L2X0_CACHE_ID); | |
1542 | ||
1543 | __l2c_init(data, aux_val, aux_mask, cache_id); | |
1544 | ||
1545 | return 0; | |
1546 | } | |
1547 | #endif |