arch/arm/mm/cache-l2x0.c

   1 /*
   2  * arch/arm/mm/cache-l2x0.c - L210/L220 cache controller support
   3  *
   4  * Copyright (C) 2007 ARM Limited
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  18  */
  19 #include <linux/err.h>
  20 #include <linux/init.h>
  21 #include <linux/spinlock.h>
  22 #include <linux/io.h>
  23 #include <linux/of.h>
  24 #include <linux/of_address.h>
  25
  26 #include <asm/cacheflush.h>
  27 #include <asm/hardware/cache-l2x0.h>
  28
  29 #define CACHE_LINE_SIZE         32
  30
  31 static void __iomem *l2x0_base;
  32 static DEFINE_RAW_SPINLOCK(l2x0_lock);
  33 static u32 l2x0_way_mask;       /* Bitmask of active ways */
  34 static u32 l2x0_size;
  35 static unsigned long sync_reg_offset = L2X0_CACHE_SYNC;
  36
  37 struct l2x0_regs l2x0_saved_regs;
  38
  39 struct l2x0_of_data {
  40         void (*setup)(const struct device_node *, u32 *, u32 *);
  41         void (*save)(void);
  42         void (*resume)(void);
  43 };
  44
  45 static inline void cache_wait_way(void __iomem *reg, unsigned long mask)
  46 {
  47         /* wait for cache operation by line or way to complete */
  48         while (readl_relaxed(reg) & mask)
  49                 cpu_relax();
  50 }
  51
  52 #ifdef CONFIG_CACHE_PL310
  53 static inline void cache_wait(void __iomem *reg, unsigned long mask)
  54 {
  55         /* cache operations by line are atomic on PL310 */
  56 }
  57 #else
  58 #define cache_wait      cache_wait_way
  59 #endif
  60
  61 static inline void cache_sync(void)
  62 {
  63         void __iomem *base = l2x0_base;
  64
  65         writel_relaxed(0, base + sync_reg_offset);
  66         cache_wait(base + L2X0_CACHE_SYNC, 1);
  67 }
  68
  69 static inline void l2x0_clean_line(unsigned long addr)
  70 {
  71         void __iomem *base = l2x0_base;
  72         cache_wait(base + L2X0_CLEAN_LINE_PA, 1);
  73         writel_relaxed(addr, base + L2X0_CLEAN_LINE_PA);
  74 }
  75
  76 static inline void l2x0_inv_line(unsigned long addr)
  77 {
  78         void __iomem *base = l2x0_base;
  79         cache_wait(base + L2X0_INV_LINE_PA, 1);
  80         writel_relaxed(addr, base + L2X0_INV_LINE_PA);
  81 }
  82
  83 #if defined(CONFIG_PL310_ERRATA_588369) || defined(CONFIG_PL310_ERRATA_727915)
  84 static inline void debug_writel(unsigned long val)
  85 {
  86         if (outer_cache.set_debug)
  87                 outer_cache.set_debug(val);
  88 }
  89
  90 static void pl310_set_debug(unsigned long val)
  91 {
  92         writel_relaxed(val, l2x0_base + L2X0_DEBUG_CTRL);
  93 }
  94 #else
  95 /* Optimised out for non-errata case */
  96 static inline void debug_writel(unsigned long val)
  97 {
  98 }
  99
 100 #define pl310_set_debug NULL
 101 #endif
 102
 103 #ifdef CONFIG_PL310_ERRATA_588369
 104 static inline void l2x0_flush_line(unsigned long addr)
 105 {
 106         void __iomem *base = l2x0_base;
 107
 108         /* Clean by PA followed by Invalidate by PA */
 109         cache_wait(base + L2X0_CLEAN_LINE_PA, 1);
 110         writel_relaxed(addr, base + L2X0_CLEAN_LINE_PA);
 111         cache_wait(base + L2X0_INV_LINE_PA, 1);
 112         writel_relaxed(addr, base + L2X0_INV_LINE_PA);
 113 }
 114 #else
 115
 116 static inline void l2x0_flush_line(unsigned long addr)
 117 {
 118         void __iomem *base = l2x0_base;
 119         cache_wait(base + L2X0_CLEAN_INV_LINE_PA, 1);
 120         writel_relaxed(addr, base + L2X0_CLEAN_INV_LINE_PA);
 121 }
 122 #endif
 123
 124 static void l2x0_cache_sync(void)
 125 {
 126         unsigned long flags;
 127
 128         raw_spin_lock_irqsave(&l2x0_lock, flags);
 129         cache_sync();
 130         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 131 }
 132
 133 static void __l2x0_flush_all(void)
 134 {
 135         debug_writel(0x03);
 136         writel_relaxed(l2x0_way_mask, l2x0_base + L2X0_CLEAN_INV_WAY);
 137         cache_wait_way(l2x0_base + L2X0_CLEAN_INV_WAY, l2x0_way_mask);
 138         cache_sync();
 139         debug_writel(0x00);
 140 }
 141
 142 static void l2x0_flush_all(void)
 143 {
 144         unsigned long flags;
 145
 146         /* clean all ways */
 147         raw_spin_lock_irqsave(&l2x0_lock, flags);
 148         __l2x0_flush_all();
 149         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 150 }
 151
 152 static void l2x0_clean_all(void)
 153 {
 154         unsigned long flags;
 155
 156         /* clean all ways */
 157         raw_spin_lock_irqsave(&l2x0_lock, flags);
 158         writel_relaxed(l2x0_way_mask, l2x0_base + L2X0_CLEAN_WAY);
 159         cache_wait_way(l2x0_base + L2X0_CLEAN_WAY, l2x0_way_mask);
 160         cache_sync();
 161         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 162 }
 163
 164 static void l2x0_inv_all(void)
 165 {
 166         unsigned long flags;
 167
 168         /* invalidate all ways */
 169         raw_spin_lock_irqsave(&l2x0_lock, flags);
 170         /* Invalidating when L2 is enabled is a nono */
 171         BUG_ON(readl(l2x0_base + L2X0_CTRL) & 1);
 172         writel_relaxed(l2x0_way_mask, l2x0_base + L2X0_INV_WAY);
 173         cache_wait_way(l2x0_base + L2X0_INV_WAY, l2x0_way_mask);
 174         cache_sync();
 175         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 176 }
 177
 178 static void l2x0_inv_range(unsigned long start, unsigned long end)
 179 {
 180         void __iomem *base = l2x0_base;
 181         unsigned long flags;
 182
 183         raw_spin_lock_irqsave(&l2x0_lock, flags);
 184         if (start & (CACHE_LINE_SIZE - 1)) {
 185                 start &= ~(CACHE_LINE_SIZE - 1);
 186                 debug_writel(0x03);
 187                 l2x0_flush_line(start);
 188                 debug_writel(0x00);
 189                 start += CACHE_LINE_SIZE;
 190         }
 191
 192         if (end & (CACHE_LINE_SIZE - 1)) {
 193                 end &= ~(CACHE_LINE_SIZE - 1);
 194                 debug_writel(0x03);
 195                 l2x0_flush_line(end);
 196                 debug_writel(0x00);
 197         }
 198
 199         while (start < end) {
 200                 unsigned long blk_end = start + min(end - start, 4096UL);
 201
 202                 while (start < blk_end) {
 203                         l2x0_inv_line(start);
 204                         start += CACHE_LINE_SIZE;
 205                 }
 206
 207                 if (blk_end < end) {
 208                         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 209                         raw_spin_lock_irqsave(&l2x0_lock, flags);
 210                 }
 211         }
 212         cache_wait(base + L2X0_INV_LINE_PA, 1);
 213         cache_sync();
 214         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 215 }
 216
 217 static void l2x0_clean_range(unsigned long start, unsigned long end)
 218 {
 219         void __iomem *base = l2x0_base;
 220         unsigned long flags;
 221
 222         if ((end - start) >= l2x0_size) {
 223                 l2x0_clean_all();
 224                 return;
 225         }
 226
 227         raw_spin_lock_irqsave(&l2x0_lock, flags);
 228         start &= ~(CACHE_LINE_SIZE - 1);
 229         while (start < end) {
 230                 unsigned long blk_end = start + min(end - start, 4096UL);
 231
 232                 while (start < blk_end) {
 233                         l2x0_clean_line(start);
 234                         start += CACHE_LINE_SIZE;
 235                 }
 236
 237                 if (blk_end < end) {
 238                         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 239                         raw_spin_lock_irqsave(&l2x0_lock, flags);
 240                 }
 241         }
 242         cache_wait(base + L2X0_CLEAN_LINE_PA, 1);
 243         cache_sync();
 244         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 245 }
 246
 247 static void l2x0_flush_range(unsigned long start, unsigned long end)
 248 {
 249         void __iomem *base = l2x0_base;
 250         unsigned long flags;
 251
 252         if ((end - start) >= l2x0_size) {
 253                 l2x0_flush_all();
 254                 return;
 255         }
 256
 257         raw_spin_lock_irqsave(&l2x0_lock, flags);
 258         start &= ~(CACHE_LINE_SIZE - 1);
 259         while (start < end) {
 260                 unsigned long blk_end = start + min(end - start, 4096UL);
 261
 262                 debug_writel(0x03);
 263                 while (start < blk_end) {
 264                         l2x0_flush_line(start);
 265                         start += CACHE_LINE_SIZE;
 266                 }
 267                 debug_writel(0x00);
 268
 269                 if (blk_end < end) {
 270                         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 271                         raw_spin_lock_irqsave(&l2x0_lock, flags);
 272                 }
 273         }
 274         cache_wait(base + L2X0_CLEAN_INV_LINE_PA, 1);
 275         cache_sync();
 276         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 277 }
 278
 279 static void l2x0_disable(void)
 280 {
 281         unsigned long flags;
 282
 283         raw_spin_lock_irqsave(&l2x0_lock, flags);
 284         __l2x0_flush_all();
 285         writel_relaxed(0, l2x0_base + L2X0_CTRL);
 286         dsb();
 287         raw_spin_unlock_irqrestore(&l2x0_lock, flags);
 288 }
 289
 290 static void l2x0_unlock(u32 cache_id)
 291 {
 292         int lockregs;
 293         int i;
 294
 295         if (cache_id == L2X0_CACHE_ID_PART_L310)
 296                 lockregs = 8;
 297         else
 298                 /* L210 and unknown types */
 299                 lockregs = 1;
 300
 301         for (i = 0; i < lockregs; i++) {
 302                 writel_relaxed(0x0, l2x0_base + L2X0_LOCKDOWN_WAY_D_BASE +
 303                                i * L2X0_LOCKDOWN_STRIDE);
 304                 writel_relaxed(0x0, l2x0_base + L2X0_LOCKDOWN_WAY_I_BASE +
 305                                i * L2X0_LOCKDOWN_STRIDE);
 306         }
 307 }
 308
 309 void __init l2x0_init(void __iomem *base, u32 aux_val, u32 aux_mask)
 310 {
 311         u32 aux;
 312         u32 cache_id;
 313         u32 way_size = 0;
 314         int ways;
 315         const char *type;
 316
 317         l2x0_base = base;
 318
 319         cache_id = readl_relaxed(l2x0_base + L2X0_CACHE_ID);
 320         aux = readl_relaxed(l2x0_base + L2X0_AUX_CTRL);
 321
 322         aux &= aux_mask;
 323         aux |= aux_val;
 324
 325         /* Determine the number of ways */
 326         switch (cache_id & L2X0_CACHE_ID_PART_MASK) {
 327         case L2X0_CACHE_ID_PART_L310:
 328                 if (aux & (1 << 16))
 329                         ways = 16;
 330                 else
 331                         ways = 8;
 332                 type = "L310";
 333 #ifdef CONFIG_PL310_ERRATA_753970
 334                 /* Unmapped register. */
 335                 sync_reg_offset = L2X0_DUMMY_REG;
 336 #endif
 337                 outer_cache.set_debug = pl310_set_debug;
 338                 break;
 339         case L2X0_CACHE_ID_PART_L210:
 340                 ways = (aux >> 13) & 0xf;
 341                 type = "L210";
 342                 break;
 343         default:
 344                 /* Assume unknown chips have 8 ways */
 345                 ways = 8;
 346                 type = "L2x0 series";
 347                 break;
 348         }
 349
 350         l2x0_way_mask = (1 << ways) - 1;
 351
 352         /*
 353          * L2 cache Size =  Way size * Number of ways
 354          */
 355         way_size = (aux & L2X0_AUX_CTRL_WAY_SIZE_MASK) >> 17;
 356         way_size = 1 << (way_size + 3);
 357         l2x0_size = ways * way_size * SZ_1K;
 358
 359         /*
 360          * Check if l2x0 controller is already enabled.
 361          * If you are booting from non-secure mode
 362          * accessing the below registers will fault.
 363          */
 364         if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & 1)) {
 365                 /* Make sure that I&D is not locked down when starting */
 366                 l2x0_unlock(cache_id);
 367
 368                 /* l2x0 controller is disabled */
 369                 writel_relaxed(aux, l2x0_base + L2X0_AUX_CTRL);
 370
 371                 l2x0_saved_regs.aux_ctrl = aux;
 372
 373                 l2x0_inv_all();
 374
 375                 /* enable L2X0 */
 376                 writel_relaxed(1, l2x0_base + L2X0_CTRL);
 377         }
 378
 379         outer_cache.inv_range = l2x0_inv_range;
 380         outer_cache.clean_range = l2x0_clean_range;
 381         outer_cache.flush_range = l2x0_flush_range;
 382         outer_cache.sync = l2x0_cache_sync;
 383         outer_cache.flush_all = l2x0_flush_all;
 384         outer_cache.inv_all = l2x0_inv_all;
 385         outer_cache.disable = l2x0_disable;
 386
 387         printk(KERN_INFO "%s cache controller enabled\n", type);
 388         printk(KERN_INFO "l2x0: %d ways, CACHE_ID 0x%08x, AUX_CTRL 0x%08x, Cache size: %d B\n",
 389                         ways, cache_id, aux, l2x0_size);
 390 }
 391
 392 #ifdef CONFIG_OF
 393 static void __init l2x0_of_setup(const struct device_node *np,
 394                                  u32 *aux_val, u32 *aux_mask)
 395 {
 396         u32 data[2] = { 0, 0 };
 397         u32 tag = 0;
 398         u32 dirty = 0;
 399         u32 val = 0, mask = 0;
 400
 401         of_property_read_u32(np, "arm,tag-latency", &tag);
 402         if (tag) {
 403                 mask |= L2X0_AUX_CTRL_TAG_LATENCY_MASK;
 404                 val |= (tag - 1) << L2X0_AUX_CTRL_TAG_LATENCY_SHIFT;
 405         }
 406
 407         of_property_read_u32_array(np, "arm,data-latency",
 408                                    data, ARRAY_SIZE(data));
 409         if (data[0] && data[1]) {
 410                 mask |= L2X0_AUX_CTRL_DATA_RD_LATENCY_MASK |
 411                         L2X0_AUX_CTRL_DATA_WR_LATENCY_MASK;
 412                 val |= ((data[0] - 1) << L2X0_AUX_CTRL_DATA_RD_LATENCY_SHIFT) |
 413                        ((data[1] - 1) << L2X0_AUX_CTRL_DATA_WR_LATENCY_SHIFT);
 414         }
 415
 416         of_property_read_u32(np, "arm,dirty-latency", &dirty);
 417         if (dirty) {
 418                 mask |= L2X0_AUX_CTRL_DIRTY_LATENCY_MASK;
 419                 val |= (dirty - 1) << L2X0_AUX_CTRL_DIRTY_LATENCY_SHIFT;
 420         }
 421
 422         *aux_val &= ~mask;
 423         *aux_val |= val;
 424         *aux_mask &= ~mask;
 425 }
 426
 427 static void __init pl310_of_setup(const struct device_node *np,
 428                                   u32 *aux_val, u32 *aux_mask)
 429 {
 430         u32 data[3] = { 0, 0, 0 };
 431         u32 tag[3] = { 0, 0, 0 };
 432         u32 filter[2] = { 0, 0 };
 433
 434         of_property_read_u32_array(np, "arm,tag-latency", tag, ARRAY_SIZE(tag));
 435         if (tag[0] && tag[1] && tag[2])
 436                 writel_relaxed(
 437                         ((tag[0] - 1) << L2X0_LATENCY_CTRL_RD_SHIFT) |
 438                         ((tag[1] - 1) << L2X0_LATENCY_CTRL_WR_SHIFT) |
 439                         ((tag[2] - 1) << L2X0_LATENCY_CTRL_SETUP_SHIFT),
 440                         l2x0_base + L2X0_TAG_LATENCY_CTRL);
 441
 442         of_property_read_u32_array(np, "arm,data-latency",
 443                                    data, ARRAY_SIZE(data));
 444         if (data[0] && data[1] && data[2])
 445                 writel_relaxed(
 446                         ((data[0] - 1) << L2X0_LATENCY_CTRL_RD_SHIFT) |
 447                         ((data[1] - 1) << L2X0_LATENCY_CTRL_WR_SHIFT) |
 448                         ((data[2] - 1) << L2X0_LATENCY_CTRL_SETUP_SHIFT),
 449                         l2x0_base + L2X0_DATA_LATENCY_CTRL);
 450
 451         of_property_read_u32_array(np, "arm,filter-ranges",
 452                                    filter, ARRAY_SIZE(filter));
 453         if (filter[1]) {
 454                 writel_relaxed(ALIGN(filter[0] + filter[1], SZ_1M),
 455                                l2x0_base + L2X0_ADDR_FILTER_END);
 456                 writel_relaxed((filter[0] & ~(SZ_1M - 1)) | L2X0_ADDR_FILTER_EN,
 457                                l2x0_base + L2X0_ADDR_FILTER_START);
 458         }
 459 }
 460
 461 static void __init pl310_save(void)
 462 {
 463         u32 l2x0_revision = readl_relaxed(l2x0_base + L2X0_CACHE_ID) &
 464                 L2X0_CACHE_ID_RTL_MASK;
 465
 466         l2x0_saved_regs.tag_latency = readl_relaxed(l2x0_base +
 467                 L2X0_TAG_LATENCY_CTRL);
 468         l2x0_saved_regs.data_latency = readl_relaxed(l2x0_base +
 469                 L2X0_DATA_LATENCY_CTRL);
 470         l2x0_saved_regs.filter_end = readl_relaxed(l2x0_base +
 471                 L2X0_ADDR_FILTER_END);
 472         l2x0_saved_regs.filter_start = readl_relaxed(l2x0_base +
 473                 L2X0_ADDR_FILTER_START);
 474
 475         if (l2x0_revision >= L2X0_CACHE_ID_RTL_R2P0) {
 476                 /*
 477                  * From r2p0, there is Prefetch offset/control register
 478                  */
 479                 l2x0_saved_regs.prefetch_ctrl = readl_relaxed(l2x0_base +
 480                         L2X0_PREFETCH_CTRL);
 481                 /*
 482                  * From r3p0, there is Power control register
 483                  */
 484                 if (l2x0_revision >= L2X0_CACHE_ID_RTL_R3P0)
 485                         l2x0_saved_regs.pwr_ctrl = readl_relaxed(l2x0_base +
 486                                 L2X0_POWER_CTRL);
 487         }
 488 }
 489
 490 static void l2x0_resume(void)
 491 {
 492         if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & 1)) {
 493                 /* restore aux ctrl and enable l2 */
 494                 l2x0_unlock(readl_relaxed(l2x0_base + L2X0_CACHE_ID));
 495
 496                 writel_relaxed(l2x0_saved_regs.aux_ctrl, l2x0_base +
 497                         L2X0_AUX_CTRL);
 498
 499                 l2x0_inv_all();
 500
 501                 writel_relaxed(1, l2x0_base + L2X0_CTRL);
 502         }
 503 }
 504
 505 static void pl310_resume(void)
 506 {
 507         u32 l2x0_revision;
 508
 509         if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & 1)) {
 510                 /* restore pl310 setup */
 511                 writel_relaxed(l2x0_saved_regs.tag_latency,
 512                         l2x0_base + L2X0_TAG_LATENCY_CTRL);
 513                 writel_relaxed(l2x0_saved_regs.data_latency,
 514                         l2x0_base + L2X0_DATA_LATENCY_CTRL);
 515                 writel_relaxed(l2x0_saved_regs.filter_end,
 516                         l2x0_base + L2X0_ADDR_FILTER_END);
 517                 writel_relaxed(l2x0_saved_regs.filter_start,
 518                         l2x0_base + L2X0_ADDR_FILTER_START);
 519
 520                 l2x0_revision = readl_relaxed(l2x0_base + L2X0_CACHE_ID) &
 521                         L2X0_CACHE_ID_RTL_MASK;
 522
 523                 if (l2x0_revision >= L2X0_CACHE_ID_RTL_R2P0) {
 524                         writel_relaxed(l2x0_saved_regs.prefetch_ctrl,
 525                                 l2x0_base + L2X0_PREFETCH_CTRL);
 526                         if (l2x0_revision >= L2X0_CACHE_ID_RTL_R3P0)
 527                                 writel_relaxed(l2x0_saved_regs.pwr_ctrl,
 528                                         l2x0_base + L2X0_POWER_CTRL);
 529                 }
 530         }
 531
 532         l2x0_resume();
 533 }
 534
 535 static const struct l2x0_of_data pl310_data = {
 536         pl310_of_setup,
 537         pl310_save,
 538         pl310_resume,
 539 };
 540
 541 static const struct l2x0_of_data l2x0_data = {
 542         l2x0_of_setup,
 543         NULL,
 544         l2x0_resume,
 545 };
 546
 547 static const struct of_device_id l2x0_ids[] __initconst = {
 548         { .compatible = "arm,pl310-cache", .data = (void *)&pl310_data },
 549         { .compatible = "arm,l220-cache", .data = (void *)&l2x0_data },
 550         { .compatible = "arm,l210-cache", .data = (void *)&l2x0_data },
 551         {}
 552 };
 553
 554 int __init l2x0_of_init(u32 aux_val, u32 aux_mask)
 555 {
 556         struct device_node *np;
 557         struct l2x0_of_data *data;
 558         struct resource res;
 559
 560         np = of_find_matching_node(NULL, l2x0_ids);
 561         if (!np)
 562                 return -ENODEV;
 563
 564         if (of_address_to_resource(np, 0, &res))
 565                 return -ENODEV;
 566
 567         l2x0_base = ioremap(res.start, resource_size(&res));
 568         if (!l2x0_base)
 569                 return -ENOMEM;
 570
 571         l2x0_saved_regs.phy_base = res.start;
 572
 573         data = of_match_node(l2x0_ids, np)->data;
 574
 575         /* L2 configuration can only be changed if the cache is disabled */
 576         if (!(readl_relaxed(l2x0_base + L2X0_CTRL) & 1)) {
 577                 if (data->setup)
 578                         data->setup(np, &aux_val, &aux_mask);
 579         }
 580
 581         if (data->save)
 582                 data->save();
 583
 584         l2x0_init(l2x0_base, aux_val, aux_mask);
 585
 586         outer_cache.resume = data->resume;
 587         return 0;
 588 }
 589 #endif