arch/mips/mm/c-r4k.c

   1 /*
   2  * This file is subject to the terms and conditions of the GNU General Public
   3  * License.  See the file "COPYING" in the main directory of this archive
   4  * for more details.
   5  *
   6  * Copyright (C) 1996 David S. Miller (davem@davemloft.net)
   7  * Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002 Ralf Baechle (ralf@gnu.org)
   8  * Copyright (C) 1999, 2000 Silicon Graphics, Inc.
   9  */
  10 #include <linux/cpu_pm.h>
  11 #include <linux/hardirq.h>
  12 #include <linux/init.h>
  13 #include <linux/highmem.h>
  14 #include <linux/kernel.h>
  15 #include <linux/linkage.h>
  16 #include <linux/preempt.h>
  17 #include <linux/sched.h>
  18 #include <linux/smp.h>
  19 #include <linux/mm.h>
  20 #include <linux/export.h>
  21 #include <linux/bitops.h>
  22
  23 #include <asm/bcache.h>
  24 #include <asm/bootinfo.h>
  25 #include <asm/cache.h>
  26 #include <asm/cacheops.h>
  27 #include <asm/cpu.h>
  28 #include <asm/cpu-features.h>
  29 #include <asm/cpu-type.h>
  30 #include <asm/io.h>
  31 #include <asm/page.h>
  32 #include <asm/pgtable.h>
  33 #include <asm/r4kcache.h>
  34 #include <asm/sections.h>
  35 #include <asm/mmu_context.h>
  36 #include <asm/war.h>
  37 #include <asm/cacheflush.h> /* for run_uncached() */
  38 #include <asm/traps.h>
  39 #include <asm/dma-coherence.h>
  40 #include <asm/mips-cps.h>
  41
  42 /*
  43  * Bits describing what cache ops an SMP callback function may perform.
  44  *
  45  * R4K_HIT   -  Virtual user or kernel address based cache operations. The
  46  *              active_mm must be checked before using user addresses, falling
  47  *              back to kmap.
  48  * R4K_INDEX -  Index based cache operations.
  49  */
  50
  51 #define R4K_HIT         BIT(0)
  52 #define R4K_INDEX       BIT(1)
  53
  54 /**
  55  * r4k_op_needs_ipi() - Decide if a cache op needs to be done on every core.
  56  * @type:       Type of cache operations (R4K_HIT or R4K_INDEX).
  57  *
  58  * Decides whether a cache op needs to be performed on every core in the system.
  59  * This may change depending on the @type of cache operation, as well as the set
  60  * of online CPUs, so preemption should be disabled by the caller to prevent CPU
  61  * hotplug from changing the result.
  62  *
  63  * Returns:     1 if the cache operation @type should be done on every core in
  64  *              the system.
  65  *              0 if the cache operation @type is globalized and only needs to
  66  *              be performed on a simple CPU.
  67  */
  68 static inline bool r4k_op_needs_ipi(unsigned int type)
  69 {
  70         /* The MIPS Coherence Manager (CM) globalizes address-based cache ops */
  71         if (type == R4K_HIT && mips_cm_present())
  72                 return false;
  73
  74         /*
  75          * Hardware doesn't globalize the required cache ops, so SMP calls may
  76          * be needed, but only if there are foreign CPUs (non-siblings with
  77          * separate caches).
  78          */
  79         /* cpu_foreign_map[] undeclared when !CONFIG_SMP */
  80 #ifdef CONFIG_SMP
  81         return !cpumask_empty(&cpu_foreign_map[0]);
  82 #else
  83         return false;
  84 #endif
  85 }
  86
  87 /*
  88  * Special Variant of smp_call_function for use by cache functions:
  89  *
  90  *  o No return value
  91  *  o collapses to normal function call on UP kernels
  92  *  o collapses to normal function call on systems with a single shared
  93  *    primary cache.
  94  *  o doesn't disable interrupts on the local CPU
  95  */
  96 static inline void r4k_on_each_cpu(unsigned int type,
  97                                    void (*func)(void *info), void *info)
  98 {
  99         preempt_disable();
 100         if (r4k_op_needs_ipi(type))
 101                 smp_call_function_many(&cpu_foreign_map[smp_processor_id()],
 102                                        func, info, 1);
 103         func(info);
 104         preempt_enable();
 105 }
 106
 107 /*
 108  * Must die.
 109  */
 110 static unsigned long icache_size __read_mostly;
 111 static unsigned long dcache_size __read_mostly;
 112 static unsigned long vcache_size __read_mostly;
 113 static unsigned long scache_size __read_mostly;
 114
 115 /*
 116  * Dummy cache handling routines for machines without boardcaches
 117  */
 118 static void cache_noop(void) {}
 119
 120 static struct bcache_ops no_sc_ops = {
 121         .bc_enable = (void *)cache_noop,
 122         .bc_disable = (void *)cache_noop,
 123         .bc_wback_inv = (void *)cache_noop,
 124         .bc_inv = (void *)cache_noop
 125 };
 126
 127 struct bcache_ops *bcops = &no_sc_ops;
 128
 129 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002010)
 130 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002020)
 131
 132 #define R4600_HIT_CACHEOP_WAR_IMPL                                      \
 133 do {                                                                    \
 134         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())            \
 135                 *(volatile unsigned long *)CKSEG1;                      \
 136         if (R4600_V1_HIT_CACHEOP_WAR)                                   \
 137                 __asm__ __volatile__("nop;nop;nop;nop");                \
 138 } while (0)
 139
 140 static void (*r4k_blast_dcache_page)(unsigned long addr);
 141
 142 static inline void r4k_blast_dcache_page_dc32(unsigned long addr)
 143 {
 144         R4600_HIT_CACHEOP_WAR_IMPL;
 145         blast_dcache32_page(addr);
 146 }
 147
 148 static inline void r4k_blast_dcache_page_dc64(unsigned long addr)
 149 {
 150         blast_dcache64_page(addr);
 151 }
 152
 153 static inline void r4k_blast_dcache_page_dc128(unsigned long addr)
 154 {
 155         blast_dcache128_page(addr);
 156 }
 157
 158 static void r4k_blast_dcache_page_setup(void)
 159 {
 160         unsigned long  dc_lsize = cpu_dcache_line_size();
 161
 162         switch (dc_lsize) {
 163         case 0:
 164                 r4k_blast_dcache_page = (void *)cache_noop;
 165                 break;
 166         case 16:
 167                 r4k_blast_dcache_page = blast_dcache16_page;
 168                 break;
 169         case 32:
 170                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc32;
 171                 break;
 172         case 64:
 173                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc64;
 174                 break;
 175         case 128:
 176                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc128;
 177                 break;
 178         default:
 179                 break;
 180         }
 181 }
 182
 183 #ifndef CONFIG_EVA
 184 #define r4k_blast_dcache_user_page  r4k_blast_dcache_page
 185 #else
 186
 187 static void (*r4k_blast_dcache_user_page)(unsigned long addr);
 188
 189 static void r4k_blast_dcache_user_page_setup(void)
 190 {
 191         unsigned long  dc_lsize = cpu_dcache_line_size();
 192
 193         if (dc_lsize == 0)
 194                 r4k_blast_dcache_user_page = (void *)cache_noop;
 195         else if (dc_lsize == 16)
 196                 r4k_blast_dcache_user_page = blast_dcache16_user_page;
 197         else if (dc_lsize == 32)
 198                 r4k_blast_dcache_user_page = blast_dcache32_user_page;
 199         else if (dc_lsize == 64)
 200                 r4k_blast_dcache_user_page = blast_dcache64_user_page;
 201 }
 202
 203 #endif
 204
 205 static void (* r4k_blast_dcache_page_indexed)(unsigned long addr);
 206
 207 static void r4k_blast_dcache_page_indexed_setup(void)
 208 {
 209         unsigned long dc_lsize = cpu_dcache_line_size();
 210
 211         if (dc_lsize == 0)
 212                 r4k_blast_dcache_page_indexed = (void *)cache_noop;
 213         else if (dc_lsize == 16)
 214                 r4k_blast_dcache_page_indexed = blast_dcache16_page_indexed;
 215         else if (dc_lsize == 32)
 216                 r4k_blast_dcache_page_indexed = blast_dcache32_page_indexed;
 217         else if (dc_lsize == 64)
 218                 r4k_blast_dcache_page_indexed = blast_dcache64_page_indexed;
 219         else if (dc_lsize == 128)
 220                 r4k_blast_dcache_page_indexed = blast_dcache128_page_indexed;
 221 }
 222
 223 void (* r4k_blast_dcache)(void);
 224 EXPORT_SYMBOL(r4k_blast_dcache);
 225
 226 static void r4k_blast_dcache_setup(void)
 227 {
 228         unsigned long dc_lsize = cpu_dcache_line_size();
 229
 230         if (dc_lsize == 0)
 231                 r4k_blast_dcache = (void *)cache_noop;
 232         else if (dc_lsize == 16)
 233                 r4k_blast_dcache = blast_dcache16;
 234         else if (dc_lsize == 32)
 235                 r4k_blast_dcache = blast_dcache32;
 236         else if (dc_lsize == 64)
 237                 r4k_blast_dcache = blast_dcache64;
 238         else if (dc_lsize == 128)
 239                 r4k_blast_dcache = blast_dcache128;
 240 }
 241
 242 /* force code alignment (used for TX49XX_ICACHE_INDEX_INV_WAR) */
 243 #define JUMP_TO_ALIGN(order) \
 244         __asm__ __volatile__( \
 245                 "b\t1f\n\t" \
 246                 ".align\t" #order "\n\t" \
 247                 "1:\n\t" \
 248                 )
 249 #define CACHE32_UNROLL32_ALIGN  JUMP_TO_ALIGN(10) /* 32 * 32 = 1024 */
 250 #define CACHE32_UNROLL32_ALIGN2 JUMP_TO_ALIGN(11)
 251
 252 static inline void blast_r4600_v1_icache32(void)
 253 {
 254         unsigned long flags;
 255
 256         local_irq_save(flags);
 257         blast_icache32();
 258         local_irq_restore(flags);
 259 }
 260
 261 static inline void tx49_blast_icache32(void)
 262 {
 263         unsigned long start = INDEX_BASE;
 264         unsigned long end = start + current_cpu_data.icache.waysize;
 265         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
 266         unsigned long ws_end = current_cpu_data.icache.ways <<
 267                                current_cpu_data.icache.waybit;
 268         unsigned long ws, addr;
 269
 270         CACHE32_UNROLL32_ALIGN2;
 271         /* I'm in even chunk.  blast odd chunks */
 272         for (ws = 0; ws < ws_end; ws += ws_inc)
 273                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
 274                         cache_unroll(32, kernel_cache, Index_Invalidate_I,
 275                                      addr | ws, 32);
 276         CACHE32_UNROLL32_ALIGN;
 277         /* I'm in odd chunk.  blast even chunks */
 278         for (ws = 0; ws < ws_end; ws += ws_inc)
 279                 for (addr = start; addr < end; addr += 0x400 * 2)
 280                         cache_unroll(32, kernel_cache, Index_Invalidate_I,
 281                                      addr | ws, 32);
 282 }
 283
 284 static inline void blast_icache32_r4600_v1_page_indexed(unsigned long page)
 285 {
 286         unsigned long flags;
 287
 288         local_irq_save(flags);
 289         blast_icache32_page_indexed(page);
 290         local_irq_restore(flags);
 291 }
 292
 293 static inline void tx49_blast_icache32_page_indexed(unsigned long page)
 294 {
 295         unsigned long indexmask = current_cpu_data.icache.waysize - 1;
 296         unsigned long start = INDEX_BASE + (page & indexmask);
 297         unsigned long end = start + PAGE_SIZE;
 298         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
 299         unsigned long ws_end = current_cpu_data.icache.ways <<
 300                                current_cpu_data.icache.waybit;
 301         unsigned long ws, addr;
 302
 303         CACHE32_UNROLL32_ALIGN2;
 304         /* I'm in even chunk.  blast odd chunks */
 305         for (ws = 0; ws < ws_end; ws += ws_inc)
 306                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
 307                         cache_unroll(32, kernel_cache, Index_Invalidate_I,
 308                                      addr | ws, 32);
 309         CACHE32_UNROLL32_ALIGN;
 310         /* I'm in odd chunk.  blast even chunks */
 311         for (ws = 0; ws < ws_end; ws += ws_inc)
 312                 for (addr = start; addr < end; addr += 0x400 * 2)
 313                         cache_unroll(32, kernel_cache, Index_Invalidate_I,
 314                                      addr | ws, 32);
 315 }
 316
 317 static void (* r4k_blast_icache_page)(unsigned long addr);
 318
 319 static void r4k_blast_icache_page_setup(void)
 320 {
 321         unsigned long ic_lsize = cpu_icache_line_size();
 322
 323         if (ic_lsize == 0)
 324                 r4k_blast_icache_page = (void *)cache_noop;
 325         else if (ic_lsize == 16)
 326                 r4k_blast_icache_page = blast_icache16_page;
 327         else if (ic_lsize == 32 && current_cpu_type() == CPU_LOONGSON2EF)
 328                 r4k_blast_icache_page = loongson2_blast_icache32_page;
 329         else if (ic_lsize == 32)
 330                 r4k_blast_icache_page = blast_icache32_page;
 331         else if (ic_lsize == 64)
 332                 r4k_blast_icache_page = blast_icache64_page;
 333         else if (ic_lsize == 128)
 334                 r4k_blast_icache_page = blast_icache128_page;
 335 }
 336
 337 #ifndef CONFIG_EVA
 338 #define r4k_blast_icache_user_page  r4k_blast_icache_page
 339 #else
 340
 341 static void (*r4k_blast_icache_user_page)(unsigned long addr);
 342
 343 static void r4k_blast_icache_user_page_setup(void)
 344 {
 345         unsigned long ic_lsize = cpu_icache_line_size();
 346
 347         if (ic_lsize == 0)
 348                 r4k_blast_icache_user_page = (void *)cache_noop;
 349         else if (ic_lsize == 16)
 350                 r4k_blast_icache_user_page = blast_icache16_user_page;
 351         else if (ic_lsize == 32)
 352                 r4k_blast_icache_user_page = blast_icache32_user_page;
 353         else if (ic_lsize == 64)
 354                 r4k_blast_icache_user_page = blast_icache64_user_page;
 355 }
 356
 357 #endif
 358
 359 static void (* r4k_blast_icache_page_indexed)(unsigned long addr);
 360
 361 static void r4k_blast_icache_page_indexed_setup(void)
 362 {
 363         unsigned long ic_lsize = cpu_icache_line_size();
 364
 365         if (ic_lsize == 0)
 366                 r4k_blast_icache_page_indexed = (void *)cache_noop;
 367         else if (ic_lsize == 16)
 368                 r4k_blast_icache_page_indexed = blast_icache16_page_indexed;
 369         else if (ic_lsize == 32) {
 370                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
 371                         r4k_blast_icache_page_indexed =
 372                                 blast_icache32_r4600_v1_page_indexed;
 373                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
 374                         r4k_blast_icache_page_indexed =
 375                                 tx49_blast_icache32_page_indexed;
 376                 else if (current_cpu_type() == CPU_LOONGSON2EF)
 377                         r4k_blast_icache_page_indexed =
 378                                 loongson2_blast_icache32_page_indexed;
 379                 else
 380                         r4k_blast_icache_page_indexed =
 381                                 blast_icache32_page_indexed;
 382         } else if (ic_lsize == 64)
 383                 r4k_blast_icache_page_indexed = blast_icache64_page_indexed;
 384 }
 385
 386 void (* r4k_blast_icache)(void);
 387 EXPORT_SYMBOL(r4k_blast_icache);
 388
 389 static void r4k_blast_icache_setup(void)
 390 {
 391         unsigned long ic_lsize = cpu_icache_line_size();
 392
 393         if (ic_lsize == 0)
 394                 r4k_blast_icache = (void *)cache_noop;
 395         else if (ic_lsize == 16)
 396                 r4k_blast_icache = blast_icache16;
 397         else if (ic_lsize == 32) {
 398                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
 399                         r4k_blast_icache = blast_r4600_v1_icache32;
 400                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
 401                         r4k_blast_icache = tx49_blast_icache32;
 402                 else if (current_cpu_type() == CPU_LOONGSON2EF)
 403                         r4k_blast_icache = loongson2_blast_icache32;
 404                 else
 405                         r4k_blast_icache = blast_icache32;
 406         } else if (ic_lsize == 64)
 407                 r4k_blast_icache = blast_icache64;
 408         else if (ic_lsize == 128)
 409                 r4k_blast_icache = blast_icache128;
 410 }
 411
 412 static void (* r4k_blast_scache_page)(unsigned long addr);
 413
 414 static void r4k_blast_scache_page_setup(void)
 415 {
 416         unsigned long sc_lsize = cpu_scache_line_size();
 417
 418         if (scache_size == 0)
 419                 r4k_blast_scache_page = (void *)cache_noop;
 420         else if (sc_lsize == 16)
 421                 r4k_blast_scache_page = blast_scache16_page;
 422         else if (sc_lsize == 32)
 423                 r4k_blast_scache_page = blast_scache32_page;
 424         else if (sc_lsize == 64)
 425                 r4k_blast_scache_page = blast_scache64_page;
 426         else if (sc_lsize == 128)
 427                 r4k_blast_scache_page = blast_scache128_page;
 428 }
 429
 430 static void (* r4k_blast_scache_page_indexed)(unsigned long addr);
 431
 432 static void r4k_blast_scache_page_indexed_setup(void)
 433 {
 434         unsigned long sc_lsize = cpu_scache_line_size();
 435
 436         if (scache_size == 0)
 437                 r4k_blast_scache_page_indexed = (void *)cache_noop;
 438         else if (sc_lsize == 16)
 439                 r4k_blast_scache_page_indexed = blast_scache16_page_indexed;
 440         else if (sc_lsize == 32)
 441                 r4k_blast_scache_page_indexed = blast_scache32_page_indexed;
 442         else if (sc_lsize == 64)
 443                 r4k_blast_scache_page_indexed = blast_scache64_page_indexed;
 444         else if (sc_lsize == 128)
 445                 r4k_blast_scache_page_indexed = blast_scache128_page_indexed;
 446 }
 447
 448 static void (* r4k_blast_scache)(void);
 449
 450 static void r4k_blast_scache_setup(void)
 451 {
 452         unsigned long sc_lsize = cpu_scache_line_size();
 453
 454         if (scache_size == 0)
 455                 r4k_blast_scache = (void *)cache_noop;
 456         else if (sc_lsize == 16)
 457                 r4k_blast_scache = blast_scache16;
 458         else if (sc_lsize == 32)
 459                 r4k_blast_scache = blast_scache32;
 460         else if (sc_lsize == 64)
 461                 r4k_blast_scache = blast_scache64;
 462         else if (sc_lsize == 128)
 463                 r4k_blast_scache = blast_scache128;
 464 }
 465
 466 static void (*r4k_blast_scache_node)(long node);
 467
 468 static void r4k_blast_scache_node_setup(void)
 469 {
 470         unsigned long sc_lsize = cpu_scache_line_size();
 471
 472         if (current_cpu_type() != CPU_LOONGSON64)
 473                 r4k_blast_scache_node = (void *)cache_noop;
 474         else if (sc_lsize == 16)
 475                 r4k_blast_scache_node = blast_scache16_node;
 476         else if (sc_lsize == 32)
 477                 r4k_blast_scache_node = blast_scache32_node;
 478         else if (sc_lsize == 64)
 479                 r4k_blast_scache_node = blast_scache64_node;
 480         else if (sc_lsize == 128)
 481                 r4k_blast_scache_node = blast_scache128_node;
 482 }
 483
 484 static inline void local_r4k___flush_cache_all(void * args)
 485 {
 486         switch (current_cpu_type()) {
 487         case CPU_LOONGSON2EF:
 488         case CPU_R4000SC:
 489         case CPU_R4000MC:
 490         case CPU_R4400SC:
 491         case CPU_R4400MC:
 492         case CPU_R10000:
 493         case CPU_R12000:
 494         case CPU_R14000:
 495         case CPU_R16000:
 496                 /*
 497                  * These caches are inclusive caches, that is, if something
 498                  * is not cached in the S-cache, we know it also won't be
 499                  * in one of the primary caches.
 500                  */
 501                 r4k_blast_scache();
 502                 break;
 503
 504         case CPU_LOONGSON64:
 505                 /* Use get_ebase_cpunum() for both NUMA=y/n */
 506                 r4k_blast_scache_node(get_ebase_cpunum() >> 2);
 507                 break;
 508
 509         case CPU_BMIPS5000:
 510                 r4k_blast_scache();
 511                 __sync();
 512                 break;
 513
 514         default:
 515                 r4k_blast_dcache();
 516                 r4k_blast_icache();
 517                 break;
 518         }
 519 }
 520
 521 static void r4k___flush_cache_all(void)
 522 {
 523         r4k_on_each_cpu(R4K_INDEX, local_r4k___flush_cache_all, NULL);
 524 }
 525
 526 /**
 527  * has_valid_asid() - Determine if an mm already has an ASID.
 528  * @mm:         Memory map.
 529  * @type:       R4K_HIT or R4K_INDEX, type of cache op.
 530  *
 531  * Determines whether @mm already has an ASID on any of the CPUs which cache ops
 532  * of type @type within an r4k_on_each_cpu() call will affect. If
 533  * r4k_on_each_cpu() does an SMP call to a single VPE in each core, then the
 534  * scope of the operation is confined to sibling CPUs, otherwise all online CPUs
 535  * will need to be checked.
 536  *
 537  * Must be called in non-preemptive context.
 538  *
 539  * Returns:     1 if the CPUs affected by @type cache ops have an ASID for @mm.
 540  *              0 otherwise.
 541  */
 542 static inline int has_valid_asid(const struct mm_struct *mm, unsigned int type)
 543 {
 544         unsigned int i;
 545         const cpumask_t *mask = cpu_present_mask;
 546
 547         if (cpu_has_mmid)
 548                 return cpu_context(0, mm) != 0;
 549
 550         /* cpu_sibling_map[] undeclared when !CONFIG_SMP */
 551 #ifdef CONFIG_SMP
 552         /*
 553          * If r4k_on_each_cpu does SMP calls, it does them to a single VPE in
 554          * each foreign core, so we only need to worry about siblings.
 555          * Otherwise we need to worry about all present CPUs.
 556          */
 557         if (r4k_op_needs_ipi(type))
 558                 mask = &cpu_sibling_map[smp_processor_id()];
 559 #endif
 560         for_each_cpu(i, mask)
 561                 if (cpu_context(i, mm))
 562                         return 1;
 563         return 0;
 564 }
 565
 566 static void r4k__flush_cache_vmap(void)
 567 {
 568         r4k_blast_dcache();
 569 }
 570
 571 static void r4k__flush_cache_vunmap(void)
 572 {
 573         r4k_blast_dcache();
 574 }
 575
 576 /*
 577  * Note: flush_tlb_range() assumes flush_cache_range() sufficiently flushes
 578  * whole caches when vma is executable.
 579  */
 580 static inline void local_r4k_flush_cache_range(void * args)
 581 {
 582         struct vm_area_struct *vma = args;
 583         int exec = vma->vm_flags & VM_EXEC;
 584
 585         if (!has_valid_asid(vma->vm_mm, R4K_INDEX))
 586                 return;
 587
 588         /*
 589          * If dcache can alias, we must blast it since mapping is changing.
 590          * If executable, we must ensure any dirty lines are written back far
 591          * enough to be visible to icache.
 592          */
 593         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc))
 594                 r4k_blast_dcache();
 595         /* If executable, blast stale lines from icache */
 596         if (exec)
 597                 r4k_blast_icache();
 598 }
 599
 600 static void r4k_flush_cache_range(struct vm_area_struct *vma,
 601         unsigned long start, unsigned long end)
 602 {
 603         int exec = vma->vm_flags & VM_EXEC;
 604
 605         if (cpu_has_dc_aliases || exec)
 606                 r4k_on_each_cpu(R4K_INDEX, local_r4k_flush_cache_range, vma);
 607 }
 608
 609 static inline void local_r4k_flush_cache_mm(void * args)
 610 {
 611         struct mm_struct *mm = args;
 612
 613         if (!has_valid_asid(mm, R4K_INDEX))
 614                 return;
 615
 616         /*
 617          * Kludge alert.  For obscure reasons R4000SC and R4400SC go nuts if we
 618          * only flush the primary caches but R1x000 behave sane ...
 619          * R4000SC and R4400SC indexed S-cache ops also invalidate primary
 620          * caches, so we can bail out early.
 621          */
 622         if (current_cpu_type() == CPU_R4000SC ||
 623             current_cpu_type() == CPU_R4000MC ||
 624             current_cpu_type() == CPU_R4400SC ||
 625             current_cpu_type() == CPU_R4400MC) {
 626                 r4k_blast_scache();
 627                 return;
 628         }
 629
 630         r4k_blast_dcache();
 631 }
 632
 633 static void r4k_flush_cache_mm(struct mm_struct *mm)
 634 {
 635         if (!cpu_has_dc_aliases)
 636                 return;
 637
 638         r4k_on_each_cpu(R4K_INDEX, local_r4k_flush_cache_mm, mm);
 639 }
 640
 641 struct flush_cache_page_args {
 642         struct vm_area_struct *vma;
 643         unsigned long addr;
 644         unsigned long pfn;
 645 };
 646
 647 static inline void local_r4k_flush_cache_page(void *args)
 648 {
 649         struct flush_cache_page_args *fcp_args = args;
 650         struct vm_area_struct *vma = fcp_args->vma;
 651         unsigned long addr = fcp_args->addr;
 652         struct page *page = pfn_to_page(fcp_args->pfn);
 653         int exec = vma->vm_flags & VM_EXEC;
 654         struct mm_struct *mm = vma->vm_mm;
 655         int map_coherent = 0;
 656         pgd_t *pgdp;
 657         p4d_t *p4dp;
 658         pud_t *pudp;
 659         pmd_t *pmdp;
 660         pte_t *ptep;
 661         void *vaddr;
 662
 663         /*
 664          * If owns no valid ASID yet, cannot possibly have gotten
 665          * this page into the cache.
 666          */
 667         if (!has_valid_asid(mm, R4K_HIT))
 668                 return;
 669
 670         addr &= PAGE_MASK;
 671         pgdp = pgd_offset(mm, addr);
 672         p4dp = p4d_offset(pgdp, addr);
 673         pudp = pud_offset(p4dp, addr);
 674         pmdp = pmd_offset(pudp, addr);
 675         ptep = pte_offset(pmdp, addr);
 676
 677         /*
 678          * If the page isn't marked valid, the page cannot possibly be
 679          * in the cache.
 680          */
 681         if (!(pte_present(*ptep)))
 682                 return;
 683
 684         if ((mm == current->active_mm) && (pte_val(*ptep) & _PAGE_VALID))
 685                 vaddr = NULL;
 686         else {
 687                 /*
 688                  * Use kmap_coherent or kmap_atomic to do flushes for
 689                  * another ASID than the current one.
 690                  */
 691                 map_coherent = (cpu_has_dc_aliases &&
 692                                 page_mapcount(page) &&
 693                                 !Page_dcache_dirty(page));
 694                 if (map_coherent)
 695                         vaddr = kmap_coherent(page, addr);
 696                 else
 697                         vaddr = kmap_atomic(page);
 698                 addr = (unsigned long)vaddr;
 699         }
 700
 701         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
 702                 vaddr ? r4k_blast_dcache_page(addr) :
 703                         r4k_blast_dcache_user_page(addr);
 704                 if (exec && !cpu_icache_snoops_remote_store)
 705                         r4k_blast_scache_page(addr);
 706         }
 707         if (exec) {
 708                 if (vaddr && cpu_has_vtag_icache && mm == current->active_mm) {
 709                         drop_mmu_context(mm);
 710                 } else
 711                         vaddr ? r4k_blast_icache_page(addr) :
 712                                 r4k_blast_icache_user_page(addr);
 713         }
 714
 715         if (vaddr) {
 716                 if (map_coherent)
 717                         kunmap_coherent();
 718                 else
 719                         kunmap_atomic(vaddr);
 720         }
 721 }
 722
 723 static void r4k_flush_cache_page(struct vm_area_struct *vma,
 724         unsigned long addr, unsigned long pfn)
 725 {
 726         struct flush_cache_page_args args;
 727
 728         args.vma = vma;
 729         args.addr = addr;
 730         args.pfn = pfn;
 731
 732         r4k_on_each_cpu(R4K_HIT, local_r4k_flush_cache_page, &args);
 733 }
 734
 735 static inline void local_r4k_flush_data_cache_page(void * addr)
 736 {
 737         r4k_blast_dcache_page((unsigned long) addr);
 738 }
 739
 740 static void r4k_flush_data_cache_page(unsigned long addr)
 741 {
 742         if (in_atomic())
 743                 local_r4k_flush_data_cache_page((void *)addr);
 744         else
 745                 r4k_on_each_cpu(R4K_HIT, local_r4k_flush_data_cache_page,
 746                                 (void *) addr);
 747 }
 748
 749 struct flush_icache_range_args {
 750         unsigned long start;
 751         unsigned long end;
 752         unsigned int type;
 753         bool user;
 754 };
 755
 756 static inline void __local_r4k_flush_icache_range(unsigned long start,
 757                                                   unsigned long end,
 758                                                   unsigned int type,
 759                                                   bool user)
 760 {
 761         if (!cpu_has_ic_fills_f_dc) {
 762                 if (type == R4K_INDEX ||
 763                     (type & R4K_INDEX && end - start >= dcache_size)) {
 764                         r4k_blast_dcache();
 765                 } else {
 766                         R4600_HIT_CACHEOP_WAR_IMPL;
 767                         if (user)
 768                                 protected_blast_dcache_range(start, end);
 769                         else
 770                                 blast_dcache_range(start, end);
 771                 }
 772         }
 773
 774         if (type == R4K_INDEX ||
 775             (type & R4K_INDEX && end - start > icache_size))
 776                 r4k_blast_icache();
 777         else {
 778                 switch (boot_cpu_type()) {
 779                 case CPU_LOONGSON2EF:
 780                         protected_loongson2_blast_icache_range(start, end);
 781                         break;
 782
 783                 default:
 784                         if (user)
 785                                 protected_blast_icache_range(start, end);
 786                         else
 787                                 blast_icache_range(start, end);
 788                         break;
 789                 }
 790         }
 791 }
 792
 793 static inline void local_r4k_flush_icache_range(unsigned long start,
 794                                                 unsigned long end)
 795 {
 796         __local_r4k_flush_icache_range(start, end, R4K_HIT | R4K_INDEX, false);
 797 }
 798
 799 static inline void local_r4k_flush_icache_user_range(unsigned long start,
 800                                                      unsigned long end)
 801 {
 802         __local_r4k_flush_icache_range(start, end, R4K_HIT | R4K_INDEX, true);
 803 }
 804
 805 static inline void local_r4k_flush_icache_range_ipi(void *args)
 806 {
 807         struct flush_icache_range_args *fir_args = args;
 808         unsigned long start = fir_args->start;
 809         unsigned long end = fir_args->end;
 810         unsigned int type = fir_args->type;
 811         bool user = fir_args->user;
 812
 813         __local_r4k_flush_icache_range(start, end, type, user);
 814 }
 815
 816 static void __r4k_flush_icache_range(unsigned long start, unsigned long end,
 817                                      bool user)
 818 {
 819         struct flush_icache_range_args args;
 820         unsigned long size, cache_size;
 821
 822         args.start = start;
 823         args.end = end;
 824         args.type = R4K_HIT | R4K_INDEX;
 825         args.user = user;
 826
 827         /*
 828          * Indexed cache ops require an SMP call.
 829          * Consider if that can or should be avoided.
 830          */
 831         preempt_disable();
 832         if (r4k_op_needs_ipi(R4K_INDEX) && !r4k_op_needs_ipi(R4K_HIT)) {
 833                 /*
 834                  * If address-based cache ops don't require an SMP call, then
 835                  * use them exclusively for small flushes.
 836                  */
 837                 size = end - start;
 838                 cache_size = icache_size;
 839                 if (!cpu_has_ic_fills_f_dc) {
 840                         size *= 2;
 841                         cache_size += dcache_size;
 842                 }
 843                 if (size <= cache_size)
 844                         args.type &= ~R4K_INDEX;
 845         }
 846         r4k_on_each_cpu(args.type, local_r4k_flush_icache_range_ipi, &args);
 847         preempt_enable();
 848         instruction_hazard();
 849 }
 850
 851 static void r4k_flush_icache_range(unsigned long start, unsigned long end)
 852 {
 853         return __r4k_flush_icache_range(start, end, false);
 854 }
 855
 856 static void r4k_flush_icache_user_range(unsigned long start, unsigned long end)
 857 {
 858         return __r4k_flush_icache_range(start, end, true);
 859 }
 860
 861 #ifdef CONFIG_DMA_NONCOHERENT
 862
 863 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
 864 {
 865         /* Catch bad driver code */
 866         if (WARN_ON(size == 0))
 867                 return;
 868
 869         preempt_disable();
 870         if (cpu_has_inclusive_pcaches) {
 871                 if (size >= scache_size) {
 872                         if (current_cpu_type() != CPU_LOONGSON64)
 873                                 r4k_blast_scache();
 874                         else
 875                                 r4k_blast_scache_node(pa_to_nid(addr));
 876                 } else {
 877                         blast_scache_range(addr, addr + size);
 878                 }
 879                 preempt_enable();
 880                 __sync();
 881                 return;
 882         }
 883
 884         /*
 885          * Either no secondary cache or the available caches don't have the
 886          * subset property so we have to flush the primary caches
 887          * explicitly.
 888          * If we would need IPI to perform an INDEX-type operation, then
 889          * we have to use the HIT-type alternative as IPI cannot be used
 890          * here due to interrupts possibly being disabled.
 891          */
 892         if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
 893                 r4k_blast_dcache();
 894         } else {
 895                 R4600_HIT_CACHEOP_WAR_IMPL;
 896                 blast_dcache_range(addr, addr + size);
 897         }
 898         preempt_enable();
 899
 900         bc_wback_inv(addr, size);
 901         __sync();
 902 }
 903
 904 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
 905 {
 906         /* Catch bad driver code */
 907         if (WARN_ON(size == 0))
 908                 return;
 909
 910         preempt_disable();
 911         if (cpu_has_inclusive_pcaches) {
 912                 if (size >= scache_size) {
 913                         if (current_cpu_type() != CPU_LOONGSON64)
 914                                 r4k_blast_scache();
 915                         else
 916                                 r4k_blast_scache_node(pa_to_nid(addr));
 917                 } else {
 918                         /*
 919                          * There is no clearly documented alignment requirement
 920                          * for the cache instruction on MIPS processors and
 921                          * some processors, among them the RM5200 and RM7000
 922                          * QED processors will throw an address error for cache
 923                          * hit ops with insufficient alignment.  Solved by
 924                          * aligning the address to cache line size.
 925                          */
 926                         blast_inv_scache_range(addr, addr + size);
 927                 }
 928                 preempt_enable();
 929                 __sync();
 930                 return;
 931         }
 932
 933         if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
 934                 r4k_blast_dcache();
 935         } else {
 936                 R4600_HIT_CACHEOP_WAR_IMPL;
 937                 blast_inv_dcache_range(addr, addr + size);
 938         }
 939         preempt_enable();
 940
 941         bc_inv(addr, size);
 942         __sync();
 943 }
 944 #endif /* CONFIG_DMA_NONCOHERENT */
 945
 946 static void r4k_flush_icache_all(void)
 947 {
 948         if (cpu_has_vtag_icache)
 949                 r4k_blast_icache();
 950 }
 951
 952 struct flush_kernel_vmap_range_args {
 953         unsigned long   vaddr;
 954         int             size;
 955 };
 956
 957 static inline void local_r4k_flush_kernel_vmap_range_index(void *args)
 958 {
 959         /*
 960          * Aliases only affect the primary caches so don't bother with
 961          * S-caches or T-caches.
 962          */
 963         r4k_blast_dcache();
 964 }
 965
 966 static inline void local_r4k_flush_kernel_vmap_range(void *args)
 967 {
 968         struct flush_kernel_vmap_range_args *vmra = args;
 969         unsigned long vaddr = vmra->vaddr;
 970         int size = vmra->size;
 971
 972         /*
 973          * Aliases only affect the primary caches so don't bother with
 974          * S-caches or T-caches.
 975          */
 976         R4600_HIT_CACHEOP_WAR_IMPL;
 977         blast_dcache_range(vaddr, vaddr + size);
 978 }
 979
 980 static void r4k_flush_kernel_vmap_range(unsigned long vaddr, int size)
 981 {
 982         struct flush_kernel_vmap_range_args args;
 983
 984         args.vaddr = (unsigned long) vaddr;
 985         args.size = size;
 986
 987         if (size >= dcache_size)
 988                 r4k_on_each_cpu(R4K_INDEX,
 989                                 local_r4k_flush_kernel_vmap_range_index, NULL);
 990         else
 991                 r4k_on_each_cpu(R4K_HIT, local_r4k_flush_kernel_vmap_range,
 992                                 &args);
 993 }
 994
 995 static inline void rm7k_erratum31(void)
 996 {
 997         const unsigned long ic_lsize = 32;
 998         unsigned long addr;
 999
1000         /* RM7000 erratum #31. The icache is screwed at startup. */
1001         write_c0_taglo(0);
1002         write_c0_taghi(0);
1003
1004         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
1005                 __asm__ __volatile__ (
1006                         ".set push\n\t"
1007                         ".set noreorder\n\t"
1008                         ".set mips3\n\t"
1009                         "cache\t%1, 0(%0)\n\t"
1010                         "cache\t%1, 0x1000(%0)\n\t"
1011                         "cache\t%1, 0x2000(%0)\n\t"
1012                         "cache\t%1, 0x3000(%0)\n\t"
1013                         "cache\t%2, 0(%0)\n\t"
1014                         "cache\t%2, 0x1000(%0)\n\t"
1015                         "cache\t%2, 0x2000(%0)\n\t"
1016                         "cache\t%2, 0x3000(%0)\n\t"
1017                         "cache\t%1, 0(%0)\n\t"
1018                         "cache\t%1, 0x1000(%0)\n\t"
1019                         "cache\t%1, 0x2000(%0)\n\t"
1020                         "cache\t%1, 0x3000(%0)\n\t"
1021                         ".set pop\n"
1022                         :
1023                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
1024         }
1025 }
1026
1027 static inline int alias_74k_erratum(struct cpuinfo_mips *c)
1028 {
1029         unsigned int imp = c->processor_id & PRID_IMP_MASK;
1030         unsigned int rev = c->processor_id & PRID_REV_MASK;
1031         int present = 0;
1032
1033         /*
1034          * Early versions of the 74K do not update the cache tags on a
1035          * vtag miss/ptag hit which can occur in the case of KSEG0/KUSEG
1036          * aliases.  In this case it is better to treat the cache as always
1037          * having aliases.  Also disable the synonym tag update feature
1038          * where available.  In this case no opportunistic tag update will
1039          * happen where a load causes a virtual address miss but a physical
1040          * address hit during a D-cache look-up.
1041          */
1042         switch (imp) {
1043         case PRID_IMP_74K:
1044                 if (rev <= PRID_REV_ENCODE_332(2, 4, 0))
1045                         present = 1;
1046                 if (rev == PRID_REV_ENCODE_332(2, 4, 0))
1047                         write_c0_config6(read_c0_config6() | MIPS_CONF6_SYND);
1048                 break;
1049         case PRID_IMP_1074K:
1050                 if (rev <= PRID_REV_ENCODE_332(1, 1, 0)) {
1051                         present = 1;
1052                         write_c0_config6(read_c0_config6() | MIPS_CONF6_SYND);
1053                 }
1054                 break;
1055         default:
1056                 BUG();
1057         }
1058
1059         return present;
1060 }
1061
1062 static void b5k_instruction_hazard(void)
1063 {
1064         __sync();
1065         __sync();
1066         __asm__ __volatile__(
1067         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
1068         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
1069         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
1070         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
1071         : : : "memory");
1072 }
1073
1074 static char *way_string[] = { NULL, "direct mapped", "2-way",
1075         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way",
1076         "9-way", "10-way", "11-way", "12-way",
1077         "13-way", "14-way", "15-way", "16-way",
1078 };
1079
1080 static void probe_pcache(void)
1081 {
1082         struct cpuinfo_mips *c = &current_cpu_data;
1083         unsigned int config = read_c0_config();
1084         unsigned int prid = read_c0_prid();
1085         int has_74k_erratum = 0;
1086         unsigned long config1;
1087         unsigned int lsize;
1088
1089         switch (current_cpu_type()) {
1090         case CPU_R4600:                 /* QED style two way caches? */
1091         case CPU_R4700:
1092         case CPU_R5000:
1093         case CPU_NEVADA:
1094                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1095                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1096                 c->icache.ways = 2;
1097                 c->icache.waybit = __ffs(icache_size/2);
1098
1099                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1100                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1101                 c->dcache.ways = 2;
1102                 c->dcache.waybit= __ffs(dcache_size/2);
1103
1104                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1105                 break;
1106
1107         case CPU_R5500:
1108                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1109                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1110                 c->icache.ways = 2;
1111                 c->icache.waybit= 0;
1112
1113                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1114                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1115                 c->dcache.ways = 2;
1116                 c->dcache.waybit = 0;
1117
1118                 c->options |= MIPS_CPU_CACHE_CDEX_P | MIPS_CPU_PREFETCH;
1119                 break;
1120
1121         case CPU_TX49XX:
1122                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1123                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1124                 c->icache.ways = 4;
1125                 c->icache.waybit= 0;
1126
1127                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1128                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1129                 c->dcache.ways = 4;
1130                 c->dcache.waybit = 0;
1131
1132                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1133                 c->options |= MIPS_CPU_PREFETCH;
1134                 break;
1135
1136         case CPU_R4000PC:
1137         case CPU_R4000SC:
1138         case CPU_R4000MC:
1139         case CPU_R4400PC:
1140         case CPU_R4400SC:
1141         case CPU_R4400MC:
1142                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1143                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1144                 c->icache.ways = 1;
1145                 c->icache.waybit = 0;   /* doesn't matter */
1146
1147                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1148                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1149                 c->dcache.ways = 1;
1150                 c->dcache.waybit = 0;   /* does not matter */
1151
1152                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1153                 break;
1154
1155         case CPU_R10000:
1156         case CPU_R12000:
1157         case CPU_R14000:
1158         case CPU_R16000:
1159                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
1160                 c->icache.linesz = 64;
1161                 c->icache.ways = 2;
1162                 c->icache.waybit = 0;
1163
1164                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
1165                 c->dcache.linesz = 32;
1166                 c->dcache.ways = 2;
1167                 c->dcache.waybit = 0;
1168
1169                 c->options |= MIPS_CPU_PREFETCH;
1170                 break;
1171
1172         case CPU_VR4133:
1173                 write_c0_config(config & ~VR41_CONF_P4K);
1174                 /* fall through */
1175         case CPU_VR4131:
1176                 /* Workaround for cache instruction bug of VR4131 */
1177                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
1178                     c->processor_id == 0x0c82U) {
1179                         config |= 0x00400000U;
1180                         if (c->processor_id == 0x0c80U)
1181                                 config |= VR41_CONF_BP;
1182                         write_c0_config(config);
1183                 } else
1184                         c->options |= MIPS_CPU_CACHE_CDEX_P;
1185
1186                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
1187                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1188                 c->icache.ways = 2;
1189                 c->icache.waybit = __ffs(icache_size/2);
1190
1191                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
1192                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1193                 c->dcache.ways = 2;
1194                 c->dcache.waybit = __ffs(dcache_size/2);
1195                 break;
1196
1197         case CPU_VR41XX:
1198         case CPU_VR4111:
1199         case CPU_VR4121:
1200         case CPU_VR4122:
1201         case CPU_VR4181:
1202         case CPU_VR4181A:
1203                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
1204                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1205                 c->icache.ways = 1;
1206                 c->icache.waybit = 0;   /* doesn't matter */
1207
1208                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
1209                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1210                 c->dcache.ways = 1;
1211                 c->dcache.waybit = 0;   /* does not matter */
1212
1213                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1214                 break;
1215
1216         case CPU_RM7000:
1217                 rm7k_erratum31();
1218
1219                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1220                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1221                 c->icache.ways = 4;
1222                 c->icache.waybit = __ffs(icache_size / c->icache.ways);
1223
1224                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1225                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1226                 c->dcache.ways = 4;
1227                 c->dcache.waybit = __ffs(dcache_size / c->dcache.ways);
1228
1229                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1230                 c->options |= MIPS_CPU_PREFETCH;
1231                 break;
1232
1233         case CPU_LOONGSON2EF:
1234                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1235                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1236                 if (prid & 0x3)
1237                         c->icache.ways = 4;
1238                 else
1239                         c->icache.ways = 2;
1240                 c->icache.waybit = 0;
1241
1242                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1243                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1244                 if (prid & 0x3)
1245                         c->dcache.ways = 4;
1246                 else
1247                         c->dcache.ways = 2;
1248                 c->dcache.waybit = 0;
1249                 break;
1250
1251         case CPU_LOONGSON64:
1252                 config1 = read_c0_config1();
1253                 lsize = (config1 >> 19) & 7;
1254                 if (lsize)
1255                         c->icache.linesz = 2 << lsize;
1256                 else
1257                         c->icache.linesz = 0;
1258                 c->icache.sets = 64 << ((config1 >> 22) & 7);
1259                 c->icache.ways = 1 + ((config1 >> 16) & 7);
1260                 icache_size = c->icache.sets *
1261                                           c->icache.ways *
1262                                           c->icache.linesz;
1263                 c->icache.waybit = 0;
1264
1265                 lsize = (config1 >> 10) & 7;
1266                 if (lsize)
1267                         c->dcache.linesz = 2 << lsize;
1268                 else
1269                         c->dcache.linesz = 0;
1270                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
1271                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
1272                 dcache_size = c->dcache.sets *
1273                                           c->dcache.ways *
1274                                           c->dcache.linesz;
1275                 c->dcache.waybit = 0;
1276                 if ((c->processor_id & (PRID_IMP_MASK | PRID_REV_MASK)) >=
1277                                 (PRID_IMP_LOONGSON_64C | PRID_REV_LOONGSON3A_R2_0))
1278                         c->options |= MIPS_CPU_PREFETCH;
1279                 break;
1280
1281         case CPU_CAVIUM_OCTEON3:
1282                 /* For now lie about the number of ways. */
1283                 c->icache.linesz = 128;
1284                 c->icache.sets = 16;
1285                 c->icache.ways = 8;
1286                 c->icache.flags |= MIPS_CACHE_VTAG;
1287                 icache_size = c->icache.sets * c->icache.ways * c->icache.linesz;
1288
1289                 c->dcache.linesz = 128;
1290                 c->dcache.ways = 8;
1291                 c->dcache.sets = 8;
1292                 dcache_size = c->dcache.sets * c->dcache.ways * c->dcache.linesz;
1293                 c->options |= MIPS_CPU_PREFETCH;
1294                 break;
1295
1296         default:
1297                 if (!(config & MIPS_CONF_M))
1298                         panic("Don't know how to probe P-caches on this cpu.");
1299
1300                 /*
1301                  * So we seem to be a MIPS32 or MIPS64 CPU
1302                  * So let's probe the I-cache ...
1303                  */
1304                 config1 = read_c0_config1();
1305
1306                 lsize = (config1 >> 19) & 7;
1307
1308                 /* IL == 7 is reserved */
1309                 if (lsize == 7)
1310                         panic("Invalid icache line size");
1311
1312                 c->icache.linesz = lsize ? 2 << lsize : 0;
1313
1314                 c->icache.sets = 32 << (((config1 >> 22) + 1) & 7);
1315                 c->icache.ways = 1 + ((config1 >> 16) & 7);
1316
1317                 icache_size = c->icache.sets *
1318                               c->icache.ways *
1319                               c->icache.linesz;
1320                 c->icache.waybit = __ffs(icache_size/c->icache.ways);
1321
1322                 if (config & MIPS_CONF_VI)
1323                         c->icache.flags |= MIPS_CACHE_VTAG;
1324
1325                 /*
1326                  * Now probe the MIPS32 / MIPS64 data cache.
1327                  */
1328                 c->dcache.flags = 0;
1329
1330                 lsize = (config1 >> 10) & 7;
1331
1332                 /* DL == 7 is reserved */
1333                 if (lsize == 7)
1334                         panic("Invalid dcache line size");
1335
1336                 c->dcache.linesz = lsize ? 2 << lsize : 0;
1337
1338                 c->dcache.sets = 32 << (((config1 >> 13) + 1) & 7);
1339                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
1340
1341                 dcache_size = c->dcache.sets *
1342                               c->dcache.ways *
1343                               c->dcache.linesz;
1344                 c->dcache.waybit = __ffs(dcache_size/c->dcache.ways);
1345
1346                 c->options |= MIPS_CPU_PREFETCH;
1347                 break;
1348         }
1349
1350         /*
1351          * Processor configuration sanity check for the R4000SC erratum
1352          * #5.  With page sizes larger than 32kB there is no possibility
1353          * to get a VCE exception anymore so we don't care about this
1354          * misconfiguration.  The case is rather theoretical anyway;
1355          * presumably no vendor is shipping his hardware in the "bad"
1356          * configuration.
1357          */
1358         if ((prid & PRID_IMP_MASK) == PRID_IMP_R4000 &&
1359             (prid & PRID_REV_MASK) < PRID_REV_R4400 &&
1360             !(config & CONF_SC) && c->icache.linesz != 16 &&
1361             PAGE_SIZE <= 0x8000)
1362                 panic("Improper R4000SC processor configuration detected");
1363
1364         /* compute a couple of other cache variables */
1365         c->icache.waysize = icache_size / c->icache.ways;
1366         c->dcache.waysize = dcache_size / c->dcache.ways;
1367
1368         c->icache.sets = c->icache.linesz ?
1369                 icache_size / (c->icache.linesz * c->icache.ways) : 0;
1370         c->dcache.sets = c->dcache.linesz ?
1371                 dcache_size / (c->dcache.linesz * c->dcache.ways) : 0;
1372
1373         /*
1374          * R1x000 P-caches are odd in a positive way.  They're 32kB 2-way
1375          * virtually indexed so normally would suffer from aliases.  So
1376          * normally they'd suffer from aliases but magic in the hardware deals
1377          * with that for us so we don't need to take care ourselves.
1378          */
1379         switch (current_cpu_type()) {
1380         case CPU_20KC:
1381         case CPU_25KF:
1382         case CPU_I6400:
1383         case CPU_I6500:
1384         case CPU_SB1:
1385         case CPU_SB1A:
1386         case CPU_XLR:
1387                 c->dcache.flags |= MIPS_CACHE_PINDEX;
1388                 break;
1389
1390         case CPU_R10000:
1391         case CPU_R12000:
1392         case CPU_R14000:
1393         case CPU_R16000:
1394                 break;
1395
1396         case CPU_74K:
1397         case CPU_1074K:
1398                 has_74k_erratum = alias_74k_erratum(c);
1399                 /* Fall through. */
1400         case CPU_M14KC:
1401         case CPU_M14KEC:
1402         case CPU_24K:
1403         case CPU_34K:
1404         case CPU_1004K:
1405         case CPU_INTERAPTIV:
1406         case CPU_P5600:
1407         case CPU_PROAPTIV:
1408         case CPU_M5150:
1409         case CPU_QEMU_GENERIC:
1410         case CPU_P6600:
1411         case CPU_M6250:
1412                 if (!(read_c0_config7() & MIPS_CONF7_IAR) &&
1413                     (c->icache.waysize > PAGE_SIZE))
1414                         c->icache.flags |= MIPS_CACHE_ALIASES;
1415                 if (!has_74k_erratum && (read_c0_config7() & MIPS_CONF7_AR)) {
1416                         /*
1417                          * Effectively physically indexed dcache,
1418                          * thus no virtual aliases.
1419                         */
1420                         c->dcache.flags |= MIPS_CACHE_PINDEX;
1421                         break;
1422                 }
1423                 /* fall through */
1424         default:
1425                 if (has_74k_erratum || c->dcache.waysize > PAGE_SIZE)
1426                         c->dcache.flags |= MIPS_CACHE_ALIASES;
1427         }
1428
1429         /* Physically indexed caches don't suffer from virtual aliasing */
1430         if (c->dcache.flags & MIPS_CACHE_PINDEX)
1431                 c->dcache.flags &= ~MIPS_CACHE_ALIASES;
1432
1433         /*
1434          * In systems with CM the icache fills from L2 or closer caches, and
1435          * thus sees remote stores without needing to write them back any
1436          * further than that.
1437          */
1438         if (mips_cm_present())
1439                 c->icache.flags |= MIPS_IC_SNOOPS_REMOTE;
1440
1441         switch (current_cpu_type()) {
1442         case CPU_20KC:
1443                 /*
1444                  * Some older 20Kc chips doesn't have the 'VI' bit in
1445                  * the config register.
1446                  */
1447                 c->icache.flags |= MIPS_CACHE_VTAG;
1448                 break;
1449
1450         case CPU_ALCHEMY:
1451         case CPU_I6400:
1452         case CPU_I6500:
1453                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
1454                 break;
1455
1456         case CPU_BMIPS5000:
1457                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
1458                 /* Cache aliases are handled in hardware; allow HIGHMEM */
1459                 c->dcache.flags &= ~MIPS_CACHE_ALIASES;
1460                 break;
1461
1462         case CPU_LOONGSON2EF:
1463                 /*
1464                  * LOONGSON2 has 4 way icache, but when using indexed cache op,
1465                  * one op will act on all 4 ways
1466                  */
1467                 c->icache.ways = 1;
1468         }
1469
1470         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
1471                icache_size >> 10,
1472                c->icache.flags & MIPS_CACHE_VTAG ? "VIVT" : "VIPT",
1473                way_string[c->icache.ways], c->icache.linesz);
1474
1475         printk("Primary data cache %ldkB, %s, %s, %s, linesize %d bytes\n",
1476                dcache_size >> 10, way_string[c->dcache.ways],
1477                (c->dcache.flags & MIPS_CACHE_PINDEX) ? "PIPT" : "VIPT",
1478                (c->dcache.flags & MIPS_CACHE_ALIASES) ?
1479                         "cache aliases" : "no aliases",
1480                c->dcache.linesz);
1481 }
1482
1483 static void probe_vcache(void)
1484 {
1485         struct cpuinfo_mips *c = &current_cpu_data;
1486         unsigned int config2, lsize;
1487
1488         if (current_cpu_type() != CPU_LOONGSON64)
1489                 return;
1490
1491         config2 = read_c0_config2();
1492         if ((lsize = ((config2 >> 20) & 15)))
1493                 c->vcache.linesz = 2 << lsize;
1494         else
1495                 c->vcache.linesz = lsize;
1496
1497         c->vcache.sets = 64 << ((config2 >> 24) & 15);
1498         c->vcache.ways = 1 + ((config2 >> 16) & 15);
1499
1500         vcache_size = c->vcache.sets * c->vcache.ways * c->vcache.linesz;
1501
1502         c->vcache.waybit = 0;
1503         c->vcache.waysize = vcache_size / c->vcache.ways;
1504
1505         pr_info("Unified victim cache %ldkB %s, linesize %d bytes.\n",
1506                 vcache_size >> 10, way_string[c->vcache.ways], c->vcache.linesz);
1507 }
1508
1509 /*
1510  * If you even _breathe_ on this function, look at the gcc output and make sure
1511  * it does not pop things on and off the stack for the cache sizing loop that
1512  * executes in KSEG1 space or else you will crash and burn badly.  You have
1513  * been warned.
1514  */
1515 static int probe_scache(void)
1516 {
1517         unsigned long flags, addr, begin, end, pow2;
1518         unsigned int config = read_c0_config();
1519         struct cpuinfo_mips *c = &current_cpu_data;
1520
1521         if (config & CONF_SC)
1522                 return 0;
1523
1524         begin = (unsigned long) &_stext;
1525         begin &= ~((4 * 1024 * 1024) - 1);
1526         end = begin + (4 * 1024 * 1024);
1527
1528         /*
1529          * This is such a bitch, you'd think they would make it easy to do
1530          * this.  Away you daemons of stupidity!
1531          */
1532         local_irq_save(flags);
1533
1534         /* Fill each size-multiple cache line with a valid tag. */
1535         pow2 = (64 * 1024);
1536         for (addr = begin; addr < end; addr = (begin + pow2)) {
1537                 unsigned long *p = (unsigned long *) addr;
1538                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1539                 pow2 <<= 1;
1540         }
1541
1542         /* Load first line with zero (therefore invalid) tag. */
1543         write_c0_taglo(0);
1544         write_c0_taghi(0);
1545         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1546         cache_op(Index_Store_Tag_I, begin);
1547         cache_op(Index_Store_Tag_D, begin);
1548         cache_op(Index_Store_Tag_SD, begin);
1549
1550         /* Now search for the wrap around point. */
1551         pow2 = (128 * 1024);
1552         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1553                 cache_op(Index_Load_Tag_SD, addr);
1554                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1555                 if (!read_c0_taglo())
1556                         break;
1557                 pow2 <<= 1;
1558         }
1559         local_irq_restore(flags);
1560         addr -= begin;
1561
1562         scache_size = addr;
1563         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1564         c->scache.ways = 1;
1565         c->scache.waybit = 0;           /* does not matter */
1566
1567         return 1;
1568 }
1569
1570 static void __init loongson2_sc_init(void)
1571 {
1572         struct cpuinfo_mips *c = &current_cpu_data;
1573
1574         scache_size = 512*1024;
1575         c->scache.linesz = 32;
1576         c->scache.ways = 4;
1577         c->scache.waybit = 0;
1578         c->scache.waysize = scache_size / (c->scache.ways);
1579         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1580         pr_info("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1581                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1582
1583         c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1584 }
1585
1586 static void __init loongson3_sc_init(void)
1587 {
1588         struct cpuinfo_mips *c = &current_cpu_data;
1589         unsigned int config2, lsize;
1590
1591         config2 = read_c0_config2();
1592         lsize = (config2 >> 4) & 15;
1593         if (lsize)
1594                 c->scache.linesz = 2 << lsize;
1595         else
1596                 c->scache.linesz = 0;
1597         c->scache.sets = 64 << ((config2 >> 8) & 15);
1598         c->scache.ways = 1 + (config2 & 15);
1599
1600         scache_size = c->scache.sets *
1601                                   c->scache.ways *
1602                                   c->scache.linesz;
1603         /* Loongson-3 has 4 cores, 1MB scache for each. scaches are shared */
1604         scache_size *= 4;
1605         c->scache.waybit = 0;
1606         c->scache.waysize = scache_size / c->scache.ways;
1607         pr_info("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1608                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1609         if (scache_size)
1610                 c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1611         return;
1612 }
1613
1614 extern int r5k_sc_init(void);
1615 extern int rm7k_sc_init(void);
1616 extern int mips_sc_init(void);
1617
1618 static void setup_scache(void)
1619 {
1620         struct cpuinfo_mips *c = &current_cpu_data;
1621         unsigned int config = read_c0_config();
1622         int sc_present = 0;
1623
1624         /*
1625          * Do the probing thing on R4000SC and R4400SC processors.  Other
1626          * processors don't have a S-cache that would be relevant to the
1627          * Linux memory management.
1628          */
1629         switch (current_cpu_type()) {
1630         case CPU_R4000SC:
1631         case CPU_R4000MC:
1632         case CPU_R4400SC:
1633         case CPU_R4400MC:
1634                 sc_present = run_uncached(probe_scache);
1635                 if (sc_present)
1636                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1637                 break;
1638
1639         case CPU_R10000:
1640         case CPU_R12000:
1641         case CPU_R14000:
1642         case CPU_R16000:
1643                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1644                 c->scache.linesz = 64 << ((config >> 13) & 1);
1645                 c->scache.ways = 2;
1646                 c->scache.waybit= 0;
1647                 sc_present = 1;
1648                 break;
1649
1650         case CPU_R5000:
1651         case CPU_NEVADA:
1652 #ifdef CONFIG_R5000_CPU_SCACHE
1653                 r5k_sc_init();
1654 #endif
1655                 return;
1656
1657         case CPU_RM7000:
1658 #ifdef CONFIG_RM7000_CPU_SCACHE
1659                 rm7k_sc_init();
1660 #endif
1661                 return;
1662
1663         case CPU_LOONGSON2EF:
1664                 loongson2_sc_init();
1665                 return;
1666
1667         case CPU_LOONGSON64:
1668                 loongson3_sc_init();
1669                 return;
1670
1671         case CPU_CAVIUM_OCTEON3:
1672         case CPU_XLP:
1673                 /* don't need to worry about L2, fully coherent */
1674                 return;
1675
1676         default:
1677                 if (c->isa_level & (MIPS_CPU_ISA_M32R1 | MIPS_CPU_ISA_M32R2 |
1678                                     MIPS_CPU_ISA_M32R6 | MIPS_CPU_ISA_M64R1 |
1679                                     MIPS_CPU_ISA_M64R2 | MIPS_CPU_ISA_M64R6)) {
1680 #ifdef CONFIG_MIPS_CPU_SCACHE
1681                         if (mips_sc_init ()) {
1682                                 scache_size = c->scache.ways * c->scache.sets * c->scache.linesz;
1683                                 printk("MIPS secondary cache %ldkB, %s, linesize %d bytes.\n",
1684                                        scache_size >> 10,
1685                                        way_string[c->scache.ways], c->scache.linesz);
1686                         }
1687 #else
1688                         if (!(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1689                                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1690 #endif
1691                         return;
1692                 }
1693                 sc_present = 0;
1694         }
1695
1696         if (!sc_present)
1697                 return;
1698
1699         /* compute a couple of other cache variables */
1700         c->scache.waysize = scache_size / c->scache.ways;
1701
1702         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1703
1704         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1705                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1706
1707         c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1708 }
1709
1710 void au1x00_fixup_config_od(void)
1711 {
1712         /*
1713          * c0_config.od (bit 19) was write only (and read as 0)
1714          * on the early revisions of Alchemy SOCs.  It disables the bus
1715          * transaction overlapping and needs to be set to fix various errata.
1716          */
1717         switch (read_c0_prid()) {
1718         case 0x00030100: /* Au1000 DA */
1719         case 0x00030201: /* Au1000 HA */
1720         case 0x00030202: /* Au1000 HB */
1721         case 0x01030200: /* Au1500 AB */
1722         /*
1723          * Au1100 errata actually keeps silence about this bit, so we set it
1724          * just in case for those revisions that require it to be set according
1725          * to the (now gone) cpu table.
1726          */
1727         case 0x02030200: /* Au1100 AB */
1728         case 0x02030201: /* Au1100 BA */
1729         case 0x02030202: /* Au1100 BC */
1730                 set_c0_config(1 << 19);
1731                 break;
1732         }
1733 }
1734
1735 /* CP0 hazard avoidance. */
1736 #define NXP_BARRIER()                                                   \
1737          __asm__ __volatile__(                                          \
1738         ".set noreorder\n\t"                                            \
1739         "nop; nop; nop; nop; nop; nop;\n\t"                             \
1740         ".set reorder\n\t")
1741
1742 static void nxp_pr4450_fixup_config(void)
1743 {
1744         unsigned long config0;
1745
1746         config0 = read_c0_config();
1747
1748         /* clear all three cache coherency fields */
1749         config0 &= ~(0x7 | (7 << 25) | (7 << 28));
1750         config0 |= (((_page_cachable_default >> _CACHE_SHIFT) <<  0) |
1751                     ((_page_cachable_default >> _CACHE_SHIFT) << 25) |
1752                     ((_page_cachable_default >> _CACHE_SHIFT) << 28));
1753         write_c0_config(config0);
1754         NXP_BARRIER();
1755 }
1756
1757 static int cca = -1;
1758
1759 static int __init cca_setup(char *str)
1760 {
1761         get_option(&str, &cca);
1762
1763         return 0;
1764 }
1765
1766 early_param("cca", cca_setup);
1767
1768 static void coherency_setup(void)
1769 {
1770         if (cca < 0 || cca > 7)
1771                 cca = read_c0_config() & CONF_CM_CMASK;
1772         _page_cachable_default = cca << _CACHE_SHIFT;
1773
1774         pr_debug("Using cache attribute %d\n", cca);
1775         change_c0_config(CONF_CM_CMASK, cca);
1776
1777         /*
1778          * c0_status.cu=0 specifies that updates by the sc instruction use
1779          * the coherency mode specified by the TLB; 1 means cachable
1780          * coherent update on write will be used.  Not all processors have
1781          * this bit and; some wire it to zero, others like Toshiba had the
1782          * silly idea of putting something else there ...
1783          */
1784         switch (current_cpu_type()) {
1785         case CPU_R4000PC:
1786         case CPU_R4000SC:
1787         case CPU_R4000MC:
1788         case CPU_R4400PC:
1789         case CPU_R4400SC:
1790         case CPU_R4400MC:
1791                 clear_c0_config(CONF_CU);
1792                 break;
1793         /*
1794          * We need to catch the early Alchemy SOCs with
1795          * the write-only co_config.od bit and set it back to one on:
1796          * Au1000 rev DA, HA, HB;  Au1100 AB, BA, BC, Au1500 AB
1797          */
1798         case CPU_ALCHEMY:
1799                 au1x00_fixup_config_od();
1800                 break;
1801
1802         case PRID_IMP_PR4450:
1803                 nxp_pr4450_fixup_config();
1804                 break;
1805         }
1806 }
1807
1808 static void r4k_cache_error_setup(void)
1809 {
1810         extern char __weak except_vec2_generic;
1811         extern char __weak except_vec2_sb1;
1812
1813         switch (current_cpu_type()) {
1814         case CPU_SB1:
1815         case CPU_SB1A:
1816                 set_uncached_handler(0x100, &except_vec2_sb1, 0x80);
1817                 break;
1818
1819         default:
1820                 set_uncached_handler(0x100, &except_vec2_generic, 0x80);
1821                 break;
1822         }
1823 }
1824
1825 void r4k_cache_init(void)
1826 {
1827         extern void build_clear_page(void);
1828         extern void build_copy_page(void);
1829         struct cpuinfo_mips *c = &current_cpu_data;
1830
1831         probe_pcache();
1832         probe_vcache();
1833         setup_scache();
1834
1835         r4k_blast_dcache_page_setup();
1836         r4k_blast_dcache_page_indexed_setup();
1837         r4k_blast_dcache_setup();
1838         r4k_blast_icache_page_setup();
1839         r4k_blast_icache_page_indexed_setup();
1840         r4k_blast_icache_setup();
1841         r4k_blast_scache_page_setup();
1842         r4k_blast_scache_page_indexed_setup();
1843         r4k_blast_scache_setup();
1844         r4k_blast_scache_node_setup();
1845 #ifdef CONFIG_EVA
1846         r4k_blast_dcache_user_page_setup();
1847         r4k_blast_icache_user_page_setup();
1848 #endif
1849
1850         /*
1851          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1852          * This code supports virtually indexed processors and will be
1853          * unnecessarily inefficient on physically indexed processors.
1854          */
1855         if (c->dcache.linesz && cpu_has_dc_aliases)
1856                 shm_align_mask = max_t( unsigned long,
1857                                         c->dcache.sets * c->dcache.linesz - 1,
1858                                         PAGE_SIZE - 1);
1859         else
1860                 shm_align_mask = PAGE_SIZE-1;
1861
1862         __flush_cache_vmap      = r4k__flush_cache_vmap;
1863         __flush_cache_vunmap    = r4k__flush_cache_vunmap;
1864
1865         flush_cache_all         = cache_noop;
1866         __flush_cache_all       = r4k___flush_cache_all;
1867         flush_cache_mm          = r4k_flush_cache_mm;
1868         flush_cache_page        = r4k_flush_cache_page;
1869         flush_cache_range       = r4k_flush_cache_range;
1870
1871         __flush_kernel_vmap_range = r4k_flush_kernel_vmap_range;
1872
1873         flush_icache_all        = r4k_flush_icache_all;
1874         local_flush_data_cache_page     = local_r4k_flush_data_cache_page;
1875         flush_data_cache_page   = r4k_flush_data_cache_page;
1876         flush_icache_range      = r4k_flush_icache_range;
1877         local_flush_icache_range        = local_r4k_flush_icache_range;
1878         __flush_icache_user_range       = r4k_flush_icache_user_range;
1879         __local_flush_icache_user_range = local_r4k_flush_icache_user_range;
1880
1881 #ifdef CONFIG_DMA_NONCOHERENT
1882 #ifdef CONFIG_DMA_MAYBE_COHERENT
1883         if (coherentio == IO_COHERENCE_ENABLED ||
1884             (coherentio == IO_COHERENCE_DEFAULT && hw_coherentio)) {
1885                 _dma_cache_wback_inv    = (void *)cache_noop;
1886                 _dma_cache_wback        = (void *)cache_noop;
1887                 _dma_cache_inv          = (void *)cache_noop;
1888         } else
1889 #endif /* CONFIG_DMA_MAYBE_COHERENT */
1890         {
1891                 _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1892                 _dma_cache_wback        = r4k_dma_cache_wback_inv;
1893                 _dma_cache_inv          = r4k_dma_cache_inv;
1894         }
1895 #endif /* CONFIG_DMA_NONCOHERENT */
1896
1897         build_clear_page();
1898         build_copy_page();
1899
1900         /*
1901          * We want to run CMP kernels on core with and without coherent
1902          * caches. Therefore, do not use CONFIG_MIPS_CMP to decide whether
1903          * or not to flush caches.
1904          */
1905         local_r4k___flush_cache_all(NULL);
1906
1907         coherency_setup();
1908         board_cache_error_setup = r4k_cache_error_setup;
1909
1910         /*
1911          * Per-CPU overrides
1912          */
1913         switch (current_cpu_type()) {
1914         case CPU_BMIPS4350:
1915         case CPU_BMIPS4380:
1916                 /* No IPI is needed because all CPUs share the same D$ */
1917                 flush_data_cache_page = r4k_blast_dcache_page;
1918                 break;
1919         case CPU_BMIPS5000:
1920                 /* We lose our superpowers if L2 is disabled */
1921                 if (c->scache.flags & MIPS_CACHE_NOT_PRESENT)
1922                         break;
1923
1924                 /* I$ fills from D$ just by emptying the write buffers */
1925                 flush_cache_page = (void *)b5k_instruction_hazard;
1926                 flush_cache_range = (void *)b5k_instruction_hazard;
1927                 local_flush_data_cache_page = (void *)b5k_instruction_hazard;
1928                 flush_data_cache_page = (void *)b5k_instruction_hazard;
1929                 flush_icache_range = (void *)b5k_instruction_hazard;
1930                 local_flush_icache_range = (void *)b5k_instruction_hazard;
1931
1932
1933                 /* Optimization: an L2 flush implicitly flushes the L1 */
1934                 current_cpu_data.options |= MIPS_CPU_INCLUSIVE_CACHES;
1935                 break;
1936         case CPU_LOONGSON64:
1937                 /* Loongson-3 maintains cache coherency by hardware */
1938                 __flush_cache_all       = cache_noop;
1939                 __flush_cache_vmap      = cache_noop;
1940                 __flush_cache_vunmap    = cache_noop;
1941                 __flush_kernel_vmap_range = (void *)cache_noop;
1942                 flush_cache_mm          = (void *)cache_noop;
1943                 flush_cache_page        = (void *)cache_noop;
1944                 flush_cache_range       = (void *)cache_noop;
1945                 flush_icache_all        = (void *)cache_noop;
1946                 flush_data_cache_page   = (void *)cache_noop;
1947                 local_flush_data_cache_page     = (void *)cache_noop;
1948                 break;
1949         }
1950 }
1951
1952 static int r4k_cache_pm_notifier(struct notifier_block *self, unsigned long cmd,
1953                                void *v)
1954 {
1955         switch (cmd) {
1956         case CPU_PM_ENTER_FAILED:
1957         case CPU_PM_EXIT:
1958                 coherency_setup();
1959                 break;
1960         }
1961
1962         return NOTIFY_OK;
1963 }
1964
1965 static struct notifier_block r4k_cache_pm_notifier_block = {
1966         .notifier_call = r4k_cache_pm_notifier,
1967 };
1968
1969 int __init r4k_cache_init_pm(void)
1970 {
1971         return cpu_pm_register_notifier(&r4k_cache_pm_notifier_block);
1972 }
1973 arch_initcall(r4k_cache_init_pm);