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/module.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-cm.h>
  41
  42 /*
  43  * Special Variant of smp_call_function for use by cache functions:
  44  *
  45  *  o No return value
  46  *  o collapses to normal function call on UP kernels
  47  *  o collapses to normal function call on systems with a single shared
  48  *    primary cache.
  49  *  o doesn't disable interrupts on the local CPU
  50  */
  51 static inline void r4k_on_each_cpu(void (*func) (void *info), void *info)
  52 {
  53         preempt_disable();
  54
  55         /*
  56          * The Coherent Manager propagates address-based cache ops to other
  57          * cores but not index-based ops. However, r4k_on_each_cpu is used
  58          * in both cases so there is no easy way to tell what kind of op is
  59          * executed to the other cores. The best we can probably do is
  60          * to restrict that call when a CM is not present because both
  61          * CM-based SMP protocols (CMP & CPS) restrict index-based cache ops.
  62          */
  63         if (!mips_cm_present())
  64                 smp_call_function_many(&cpu_foreign_map, func, info, 1);
  65         func(info);
  66         preempt_enable();
  67 }
  68
  69 #if defined(CONFIG_MIPS_CMP) || defined(CONFIG_MIPS_CPS)
  70 #define cpu_has_safe_index_cacheops 0
  71 #else
  72 #define cpu_has_safe_index_cacheops 1
  73 #endif
  74
  75 /*
  76  * Must die.
  77  */
  78 static unsigned long icache_size __read_mostly;
  79 static unsigned long dcache_size __read_mostly;
  80 static unsigned long scache_size __read_mostly;
  81
  82 /*
  83  * Dummy cache handling routines for machines without boardcaches
  84  */
  85 static void cache_noop(void) {}
  86
  87 static struct bcache_ops no_sc_ops = {
  88         .bc_enable = (void *)cache_noop,
  89         .bc_disable = (void *)cache_noop,
  90         .bc_wback_inv = (void *)cache_noop,
  91         .bc_inv = (void *)cache_noop
  92 };
  93
  94 struct bcache_ops *bcops = &no_sc_ops;
  95
  96 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002010)
  97 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002020)
  98
  99 #define R4600_HIT_CACHEOP_WAR_IMPL                                      \
 100 do {                                                                    \
 101         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())            \
 102                 *(volatile unsigned long *)CKSEG1;                      \
 103         if (R4600_V1_HIT_CACHEOP_WAR)                                   \
 104                 __asm__ __volatile__("nop;nop;nop;nop");                \
 105 } while (0)
 106
 107 static void (*r4k_blast_dcache_page)(unsigned long addr);
 108
 109 static inline void r4k_blast_dcache_page_dc32(unsigned long addr)
 110 {
 111         R4600_HIT_CACHEOP_WAR_IMPL;
 112         blast_dcache32_page(addr);
 113 }
 114
 115 static inline void r4k_blast_dcache_page_dc64(unsigned long addr)
 116 {
 117         blast_dcache64_page(addr);
 118 }
 119
 120 static inline void r4k_blast_dcache_page_dc128(unsigned long addr)
 121 {
 122         blast_dcache128_page(addr);
 123 }
 124
 125 static void r4k_blast_dcache_page_setup(void)
 126 {
 127         unsigned long  dc_lsize = cpu_dcache_line_size();
 128
 129         switch (dc_lsize) {
 130         case 0:
 131                 r4k_blast_dcache_page = (void *)cache_noop;
 132                 break;
 133         case 16:
 134                 r4k_blast_dcache_page = blast_dcache16_page;
 135                 break;
 136         case 32:
 137                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc32;
 138                 break;
 139         case 64:
 140                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc64;
 141                 break;
 142         case 128:
 143                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc128;
 144                 break;
 145         default:
 146                 break;
 147         }
 148 }
 149
 150 #ifndef CONFIG_EVA
 151 #define r4k_blast_dcache_user_page  r4k_blast_dcache_page
 152 #else
 153
 154 static void (*r4k_blast_dcache_user_page)(unsigned long addr);
 155
 156 static void r4k_blast_dcache_user_page_setup(void)
 157 {
 158         unsigned long  dc_lsize = cpu_dcache_line_size();
 159
 160         if (dc_lsize == 0)
 161                 r4k_blast_dcache_user_page = (void *)cache_noop;
 162         else if (dc_lsize == 16)
 163                 r4k_blast_dcache_user_page = blast_dcache16_user_page;
 164         else if (dc_lsize == 32)
 165                 r4k_blast_dcache_user_page = blast_dcache32_user_page;
 166         else if (dc_lsize == 64)
 167                 r4k_blast_dcache_user_page = blast_dcache64_user_page;
 168 }
 169
 170 #endif
 171
 172 static void (* r4k_blast_dcache_page_indexed)(unsigned long addr);
 173
 174 static void r4k_blast_dcache_page_indexed_setup(void)
 175 {
 176         unsigned long dc_lsize = cpu_dcache_line_size();
 177
 178         if (dc_lsize == 0)
 179                 r4k_blast_dcache_page_indexed = (void *)cache_noop;
 180         else if (dc_lsize == 16)
 181                 r4k_blast_dcache_page_indexed = blast_dcache16_page_indexed;
 182         else if (dc_lsize == 32)
 183                 r4k_blast_dcache_page_indexed = blast_dcache32_page_indexed;
 184         else if (dc_lsize == 64)
 185                 r4k_blast_dcache_page_indexed = blast_dcache64_page_indexed;
 186         else if (dc_lsize == 128)
 187                 r4k_blast_dcache_page_indexed = blast_dcache128_page_indexed;
 188 }
 189
 190 void (* r4k_blast_dcache)(void);
 191 EXPORT_SYMBOL(r4k_blast_dcache);
 192
 193 static void r4k_blast_dcache_setup(void)
 194 {
 195         unsigned long dc_lsize = cpu_dcache_line_size();
 196
 197         if (dc_lsize == 0)
 198                 r4k_blast_dcache = (void *)cache_noop;
 199         else if (dc_lsize == 16)
 200                 r4k_blast_dcache = blast_dcache16;
 201         else if (dc_lsize == 32)
 202                 r4k_blast_dcache = blast_dcache32;
 203         else if (dc_lsize == 64)
 204                 r4k_blast_dcache = blast_dcache64;
 205         else if (dc_lsize == 128)
 206                 r4k_blast_dcache = blast_dcache128;
 207 }
 208
 209 /* force code alignment (used for TX49XX_ICACHE_INDEX_INV_WAR) */
 210 #define JUMP_TO_ALIGN(order) \
 211         __asm__ __volatile__( \
 212                 "b\t1f\n\t" \
 213                 ".align\t" #order "\n\t" \
 214                 "1:\n\t" \
 215                 )
 216 #define CACHE32_UNROLL32_ALIGN  JUMP_TO_ALIGN(10) /* 32 * 32 = 1024 */
 217 #define CACHE32_UNROLL32_ALIGN2 JUMP_TO_ALIGN(11)
 218
 219 static inline void blast_r4600_v1_icache32(void)
 220 {
 221         unsigned long flags;
 222
 223         local_irq_save(flags);
 224         blast_icache32();
 225         local_irq_restore(flags);
 226 }
 227
 228 static inline void tx49_blast_icache32(void)
 229 {
 230         unsigned long start = INDEX_BASE;
 231         unsigned long end = start + current_cpu_data.icache.waysize;
 232         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
 233         unsigned long ws_end = current_cpu_data.icache.ways <<
 234                                current_cpu_data.icache.waybit;
 235         unsigned long ws, addr;
 236
 237         CACHE32_UNROLL32_ALIGN2;
 238         /* I'm in even chunk.  blast odd chunks */
 239         for (ws = 0; ws < ws_end; ws += ws_inc)
 240                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
 241                         cache32_unroll32(addr|ws, Index_Invalidate_I);
 242         CACHE32_UNROLL32_ALIGN;
 243         /* I'm in odd chunk.  blast even chunks */
 244         for (ws = 0; ws < ws_end; ws += ws_inc)
 245                 for (addr = start; addr < end; addr += 0x400 * 2)
 246                         cache32_unroll32(addr|ws, Index_Invalidate_I);
 247 }
 248
 249 static inline void blast_icache32_r4600_v1_page_indexed(unsigned long page)
 250 {
 251         unsigned long flags;
 252
 253         local_irq_save(flags);
 254         blast_icache32_page_indexed(page);
 255         local_irq_restore(flags);
 256 }
 257
 258 static inline void tx49_blast_icache32_page_indexed(unsigned long page)
 259 {
 260         unsigned long indexmask = current_cpu_data.icache.waysize - 1;
 261         unsigned long start = INDEX_BASE + (page & indexmask);
 262         unsigned long end = start + PAGE_SIZE;
 263         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
 264         unsigned long ws_end = current_cpu_data.icache.ways <<
 265                                current_cpu_data.icache.waybit;
 266         unsigned long ws, addr;
 267
 268         CACHE32_UNROLL32_ALIGN2;
 269         /* I'm in even chunk.  blast odd chunks */
 270         for (ws = 0; ws < ws_end; ws += ws_inc)
 271                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
 272                         cache32_unroll32(addr|ws, Index_Invalidate_I);
 273         CACHE32_UNROLL32_ALIGN;
 274         /* I'm in odd chunk.  blast even chunks */
 275         for (ws = 0; ws < ws_end; ws += ws_inc)
 276                 for (addr = start; addr < end; addr += 0x400 * 2)
 277                         cache32_unroll32(addr|ws, Index_Invalidate_I);
 278 }
 279
 280 static void (* r4k_blast_icache_page)(unsigned long addr);
 281
 282 static void r4k_blast_icache_page_setup(void)
 283 {
 284         unsigned long ic_lsize = cpu_icache_line_size();
 285
 286         if (ic_lsize == 0)
 287                 r4k_blast_icache_page = (void *)cache_noop;
 288         else if (ic_lsize == 16)
 289                 r4k_blast_icache_page = blast_icache16_page;
 290         else if (ic_lsize == 32 && current_cpu_type() == CPU_LOONGSON2)
 291                 r4k_blast_icache_page = loongson2_blast_icache32_page;
 292         else if (ic_lsize == 32)
 293                 r4k_blast_icache_page = blast_icache32_page;
 294         else if (ic_lsize == 64)
 295                 r4k_blast_icache_page = blast_icache64_page;
 296         else if (ic_lsize == 128)
 297                 r4k_blast_icache_page = blast_icache128_page;
 298 }
 299
 300 #ifndef CONFIG_EVA
 301 #define r4k_blast_icache_user_page  r4k_blast_icache_page
 302 #else
 303
 304 static void (*r4k_blast_icache_user_page)(unsigned long addr);
 305
 306 static void r4k_blast_icache_user_page_setup(void)
 307 {
 308         unsigned long ic_lsize = cpu_icache_line_size();
 309
 310         if (ic_lsize == 0)
 311                 r4k_blast_icache_user_page = (void *)cache_noop;
 312         else if (ic_lsize == 16)
 313                 r4k_blast_icache_user_page = blast_icache16_user_page;
 314         else if (ic_lsize == 32)
 315                 r4k_blast_icache_user_page = blast_icache32_user_page;
 316         else if (ic_lsize == 64)
 317                 r4k_blast_icache_user_page = blast_icache64_user_page;
 318 }
 319
 320 #endif
 321
 322 static void (* r4k_blast_icache_page_indexed)(unsigned long addr);
 323
 324 static void r4k_blast_icache_page_indexed_setup(void)
 325 {
 326         unsigned long ic_lsize = cpu_icache_line_size();
 327
 328         if (ic_lsize == 0)
 329                 r4k_blast_icache_page_indexed = (void *)cache_noop;
 330         else if (ic_lsize == 16)
 331                 r4k_blast_icache_page_indexed = blast_icache16_page_indexed;
 332         else if (ic_lsize == 32) {
 333                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
 334                         r4k_blast_icache_page_indexed =
 335                                 blast_icache32_r4600_v1_page_indexed;
 336                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
 337                         r4k_blast_icache_page_indexed =
 338                                 tx49_blast_icache32_page_indexed;
 339                 else if (current_cpu_type() == CPU_LOONGSON2)
 340                         r4k_blast_icache_page_indexed =
 341                                 loongson2_blast_icache32_page_indexed;
 342                 else
 343                         r4k_blast_icache_page_indexed =
 344                                 blast_icache32_page_indexed;
 345         } else if (ic_lsize == 64)
 346                 r4k_blast_icache_page_indexed = blast_icache64_page_indexed;
 347 }
 348
 349 void (* r4k_blast_icache)(void);
 350 EXPORT_SYMBOL(r4k_blast_icache);
 351
 352 static void r4k_blast_icache_setup(void)
 353 {
 354         unsigned long ic_lsize = cpu_icache_line_size();
 355
 356         if (ic_lsize == 0)
 357                 r4k_blast_icache = (void *)cache_noop;
 358         else if (ic_lsize == 16)
 359                 r4k_blast_icache = blast_icache16;
 360         else if (ic_lsize == 32) {
 361                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
 362                         r4k_blast_icache = blast_r4600_v1_icache32;
 363                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
 364                         r4k_blast_icache = tx49_blast_icache32;
 365                 else if (current_cpu_type() == CPU_LOONGSON2)
 366                         r4k_blast_icache = loongson2_blast_icache32;
 367                 else
 368                         r4k_blast_icache = blast_icache32;
 369         } else if (ic_lsize == 64)
 370                 r4k_blast_icache = blast_icache64;
 371         else if (ic_lsize == 128)
 372                 r4k_blast_icache = blast_icache128;
 373 }
 374
 375 static void (* r4k_blast_scache_page)(unsigned long addr);
 376
 377 static void r4k_blast_scache_page_setup(void)
 378 {
 379         unsigned long sc_lsize = cpu_scache_line_size();
 380
 381         if (scache_size == 0)
 382                 r4k_blast_scache_page = (void *)cache_noop;
 383         else if (sc_lsize == 16)
 384                 r4k_blast_scache_page = blast_scache16_page;
 385         else if (sc_lsize == 32)
 386                 r4k_blast_scache_page = blast_scache32_page;
 387         else if (sc_lsize == 64)
 388                 r4k_blast_scache_page = blast_scache64_page;
 389         else if (sc_lsize == 128)
 390                 r4k_blast_scache_page = blast_scache128_page;
 391 }
 392
 393 static void (* r4k_blast_scache_page_indexed)(unsigned long addr);
 394
 395 static void r4k_blast_scache_page_indexed_setup(void)
 396 {
 397         unsigned long sc_lsize = cpu_scache_line_size();
 398
 399         if (scache_size == 0)
 400                 r4k_blast_scache_page_indexed = (void *)cache_noop;
 401         else if (sc_lsize == 16)
 402                 r4k_blast_scache_page_indexed = blast_scache16_page_indexed;
 403         else if (sc_lsize == 32)
 404                 r4k_blast_scache_page_indexed = blast_scache32_page_indexed;
 405         else if (sc_lsize == 64)
 406                 r4k_blast_scache_page_indexed = blast_scache64_page_indexed;
 407         else if (sc_lsize == 128)
 408                 r4k_blast_scache_page_indexed = blast_scache128_page_indexed;
 409 }
 410
 411 static void (* r4k_blast_scache)(void);
 412
 413 static void r4k_blast_scache_setup(void)
 414 {
 415         unsigned long sc_lsize = cpu_scache_line_size();
 416
 417         if (scache_size == 0)
 418                 r4k_blast_scache = (void *)cache_noop;
 419         else if (sc_lsize == 16)
 420                 r4k_blast_scache = blast_scache16;
 421         else if (sc_lsize == 32)
 422                 r4k_blast_scache = blast_scache32;
 423         else if (sc_lsize == 64)
 424                 r4k_blast_scache = blast_scache64;
 425         else if (sc_lsize == 128)
 426                 r4k_blast_scache = blast_scache128;
 427 }
 428
 429 static inline void local_r4k___flush_cache_all(void * args)
 430 {
 431         switch (current_cpu_type()) {
 432         case CPU_LOONGSON2:
 433         case CPU_LOONGSON3:
 434         case CPU_R4000SC:
 435         case CPU_R4000MC:
 436         case CPU_R4400SC:
 437         case CPU_R4400MC:
 438         case CPU_R10000:
 439         case CPU_R12000:
 440         case CPU_R14000:
 441         case CPU_R16000:
 442                 /*
 443                  * These caches are inclusive caches, that is, if something
 444                  * is not cached in the S-cache, we know it also won't be
 445                  * in one of the primary caches.
 446                  */
 447                 r4k_blast_scache();
 448                 break;
 449
 450         default:
 451                 r4k_blast_dcache();
 452                 r4k_blast_icache();
 453                 break;
 454         }
 455 }
 456
 457 static void r4k___flush_cache_all(void)
 458 {
 459         r4k_on_each_cpu(local_r4k___flush_cache_all, NULL);
 460 }
 461
 462 static inline int has_valid_asid(const struct mm_struct *mm)
 463 {
 464 #ifdef CONFIG_MIPS_MT_SMP
 465         int i;
 466
 467         for_each_online_cpu(i)
 468                 if (cpu_context(i, mm))
 469                         return 1;
 470
 471         return 0;
 472 #else
 473         return cpu_context(smp_processor_id(), mm);
 474 #endif
 475 }
 476
 477 static void r4k__flush_cache_vmap(void)
 478 {
 479         r4k_blast_dcache();
 480 }
 481
 482 static void r4k__flush_cache_vunmap(void)
 483 {
 484         r4k_blast_dcache();
 485 }
 486
 487 static inline void local_r4k_flush_cache_range(void * args)
 488 {
 489         struct vm_area_struct *vma = args;
 490         int exec = vma->vm_flags & VM_EXEC;
 491
 492         if (!(has_valid_asid(vma->vm_mm)))
 493                 return;
 494
 495         r4k_blast_dcache();
 496         if (exec)
 497                 r4k_blast_icache();
 498 }
 499
 500 static void r4k_flush_cache_range(struct vm_area_struct *vma,
 501         unsigned long start, unsigned long end)
 502 {
 503         int exec = vma->vm_flags & VM_EXEC;
 504
 505         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc))
 506                 r4k_on_each_cpu(local_r4k_flush_cache_range, vma);
 507 }
 508
 509 static inline void local_r4k_flush_cache_mm(void * args)
 510 {
 511         struct mm_struct *mm = args;
 512
 513         if (!has_valid_asid(mm))
 514                 return;
 515
 516         /*
 517          * Kludge alert.  For obscure reasons R4000SC and R4400SC go nuts if we
 518          * only flush the primary caches but R1x000 behave sane ...
 519          * R4000SC and R4400SC indexed S-cache ops also invalidate primary
 520          * caches, so we can bail out early.
 521          */
 522         if (current_cpu_type() == CPU_R4000SC ||
 523             current_cpu_type() == CPU_R4000MC ||
 524             current_cpu_type() == CPU_R4400SC ||
 525             current_cpu_type() == CPU_R4400MC) {
 526                 r4k_blast_scache();
 527                 return;
 528         }
 529
 530         r4k_blast_dcache();
 531 }
 532
 533 static void r4k_flush_cache_mm(struct mm_struct *mm)
 534 {
 535         if (!cpu_has_dc_aliases)
 536                 return;
 537
 538         r4k_on_each_cpu(local_r4k_flush_cache_mm, mm);
 539 }
 540
 541 struct flush_cache_page_args {
 542         struct vm_area_struct *vma;
 543         unsigned long addr;
 544         unsigned long pfn;
 545 };
 546
 547 static inline void local_r4k_flush_cache_page(void *args)
 548 {
 549         struct flush_cache_page_args *fcp_args = args;
 550         struct vm_area_struct *vma = fcp_args->vma;
 551         unsigned long addr = fcp_args->addr;
 552         struct page *page = pfn_to_page(fcp_args->pfn);
 553         int exec = vma->vm_flags & VM_EXEC;
 554         struct mm_struct *mm = vma->vm_mm;
 555         int map_coherent = 0;
 556         pgd_t *pgdp;
 557         pud_t *pudp;
 558         pmd_t *pmdp;
 559         pte_t *ptep;
 560         void *vaddr;
 561
 562         /*
 563          * If ownes no valid ASID yet, cannot possibly have gotten
 564          * this page into the cache.
 565          */
 566         if (!has_valid_asid(mm))
 567                 return;
 568
 569         addr &= PAGE_MASK;
 570         pgdp = pgd_offset(mm, addr);
 571         pudp = pud_offset(pgdp, addr);
 572         pmdp = pmd_offset(pudp, addr);
 573         ptep = pte_offset(pmdp, addr);
 574
 575         /*
 576          * If the page isn't marked valid, the page cannot possibly be
 577          * in the cache.
 578          */
 579         if (!(pte_present(*ptep)))
 580                 return;
 581
 582         if ((mm == current->active_mm) && (pte_val(*ptep) & _PAGE_VALID))
 583                 vaddr = NULL;
 584         else {
 585                 /*
 586                  * Use kmap_coherent or kmap_atomic to do flushes for
 587                  * another ASID than the current one.
 588                  */
 589                 map_coherent = (cpu_has_dc_aliases &&
 590                                 page_mapcount(page) &&
 591                                 !Page_dcache_dirty(page));
 592                 if (map_coherent)
 593                         vaddr = kmap_coherent(page, addr);
 594                 else
 595                         vaddr = kmap_atomic(page);
 596                 addr = (unsigned long)vaddr;
 597         }
 598
 599         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
 600                 vaddr ? r4k_blast_dcache_page(addr) :
 601                         r4k_blast_dcache_user_page(addr);
 602                 if (exec && !cpu_icache_snoops_remote_store)
 603                         r4k_blast_scache_page(addr);
 604         }
 605         if (exec) {
 606                 if (vaddr && cpu_has_vtag_icache && mm == current->active_mm) {
 607                         int cpu = smp_processor_id();
 608
 609                         if (cpu_context(cpu, mm) != 0)
 610                                 drop_mmu_context(mm, cpu);
 611                 } else
 612                         vaddr ? r4k_blast_icache_page(addr) :
 613                                 r4k_blast_icache_user_page(addr);
 614         }
 615
 616         if (vaddr) {
 617                 if (map_coherent)
 618                         kunmap_coherent();
 619                 else
 620                         kunmap_atomic(vaddr);
 621         }
 622 }
 623
 624 static void r4k_flush_cache_page(struct vm_area_struct *vma,
 625         unsigned long addr, unsigned long pfn)
 626 {
 627         struct flush_cache_page_args args;
 628
 629         args.vma = vma;
 630         args.addr = addr;
 631         args.pfn = pfn;
 632
 633         r4k_on_each_cpu(local_r4k_flush_cache_page, &args);
 634 }
 635
 636 static inline void local_r4k_flush_data_cache_page(void * addr)
 637 {
 638         r4k_blast_dcache_page((unsigned long) addr);
 639 }
 640
 641 static void r4k_flush_data_cache_page(unsigned long addr)
 642 {
 643         if (in_atomic())
 644                 local_r4k_flush_data_cache_page((void *)addr);
 645         else
 646                 r4k_on_each_cpu(local_r4k_flush_data_cache_page, (void *) addr);
 647 }
 648
 649 struct flush_icache_range_args {
 650         unsigned long start;
 651         unsigned long end;
 652 };
 653
 654 static inline void local_r4k_flush_icache_range(unsigned long start, unsigned long end)
 655 {
 656         if (!cpu_has_ic_fills_f_dc) {
 657                 if (end - start >= dcache_size) {
 658                         r4k_blast_dcache();
 659                 } else {
 660                         R4600_HIT_CACHEOP_WAR_IMPL;
 661                         protected_blast_dcache_range(start, end);
 662                 }
 663         }
 664
 665         if (end - start > icache_size)
 666                 r4k_blast_icache();
 667         else {
 668                 switch (boot_cpu_type()) {
 669                 case CPU_LOONGSON2:
 670                         protected_loongson2_blast_icache_range(start, end);
 671                         break;
 672
 673                 default:
 674                         protected_blast_icache_range(start, end);
 675                         break;
 676                 }
 677         }
 678 #ifdef CONFIG_EVA
 679         /*
 680          * Due to all possible segment mappings, there might cache aliases
 681          * caused by the bootloader being in non-EVA mode, and the CPU switching
 682          * to EVA during early kernel init. It's best to flush the scache
 683          * to avoid having secondary cores fetching stale data and lead to
 684          * kernel crashes.
 685          */
 686         bc_wback_inv(start, (end - start));
 687         __sync();
 688 #endif
 689 }
 690
 691 static inline void local_r4k_flush_icache_range_ipi(void *args)
 692 {
 693         struct flush_icache_range_args *fir_args = args;
 694         unsigned long start = fir_args->start;
 695         unsigned long end = fir_args->end;
 696
 697         local_r4k_flush_icache_range(start, end);
 698 }
 699
 700 static void r4k_flush_icache_range(unsigned long start, unsigned long end)
 701 {
 702         struct flush_icache_range_args args;
 703
 704         args.start = start;
 705         args.end = end;
 706
 707         r4k_on_each_cpu(local_r4k_flush_icache_range_ipi, &args);
 708         instruction_hazard();
 709 }
 710
 711 #if defined(CONFIG_DMA_NONCOHERENT) || defined(CONFIG_DMA_MAYBE_COHERENT)
 712
 713 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
 714 {
 715         /* Catch bad driver code */
 716         BUG_ON(size == 0);
 717
 718         preempt_disable();
 719         if (cpu_has_inclusive_pcaches) {
 720                 if (size >= scache_size)
 721                         r4k_blast_scache();
 722                 else
 723                         blast_scache_range(addr, addr + size);
 724                 preempt_enable();
 725                 __sync();
 726                 return;
 727         }
 728
 729         /*
 730          * Either no secondary cache or the available caches don't have the
 731          * subset property so we have to flush the primary caches
 732          * explicitly
 733          */
 734         if (cpu_has_safe_index_cacheops && size >= dcache_size) {
 735                 r4k_blast_dcache();
 736         } else {
 737                 R4600_HIT_CACHEOP_WAR_IMPL;
 738                 blast_dcache_range(addr, addr + size);
 739         }
 740         preempt_enable();
 741
 742         bc_wback_inv(addr, size);
 743         __sync();
 744 }
 745
 746 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
 747 {
 748         /* Catch bad driver code */
 749         BUG_ON(size == 0);
 750
 751         preempt_disable();
 752         if (cpu_has_inclusive_pcaches) {
 753                 if (size >= scache_size)
 754                         r4k_blast_scache();
 755                 else {
 756                         /*
 757                          * There is no clearly documented alignment requirement
 758                          * for the cache instruction on MIPS processors and
 759                          * some processors, among them the RM5200 and RM7000
 760                          * QED processors will throw an address error for cache
 761                          * hit ops with insufficient alignment.  Solved by
 762                          * aligning the address to cache line size.
 763                          */
 764                         blast_inv_scache_range(addr, addr + size);
 765                 }
 766                 preempt_enable();
 767                 __sync();
 768                 return;
 769         }
 770
 771         if (cpu_has_safe_index_cacheops && size >= dcache_size) {
 772                 r4k_blast_dcache();
 773         } else {
 774                 R4600_HIT_CACHEOP_WAR_IMPL;
 775                 blast_inv_dcache_range(addr, addr + size);
 776         }
 777         preempt_enable();
 778
 779         bc_inv(addr, size);
 780         __sync();
 781 }
 782 #endif /* CONFIG_DMA_NONCOHERENT || CONFIG_DMA_MAYBE_COHERENT */
 783
 784 /*
 785  * While we're protected against bad userland addresses we don't care
 786  * very much about what happens in that case.  Usually a segmentation
 787  * fault will dump the process later on anyway ...
 788  */
 789 static void local_r4k_flush_cache_sigtramp(void * arg)
 790 {
 791         unsigned long ic_lsize = cpu_icache_line_size();
 792         unsigned long dc_lsize = cpu_dcache_line_size();
 793         unsigned long sc_lsize = cpu_scache_line_size();
 794         unsigned long addr = (unsigned long) arg;
 795
 796         R4600_HIT_CACHEOP_WAR_IMPL;
 797         if (dc_lsize)
 798                 protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
 799         if (!cpu_icache_snoops_remote_store && scache_size)
 800                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
 801         if (ic_lsize)
 802                 protected_flush_icache_line(addr & ~(ic_lsize - 1));
 803         if (MIPS4K_ICACHE_REFILL_WAR) {
 804                 __asm__ __volatile__ (
 805                         ".set push\n\t"
 806                         ".set noat\n\t"
 807                         ".set "MIPS_ISA_LEVEL"\n\t"
 808 #ifdef CONFIG_32BIT
 809                         "la     $at,1f\n\t"
 810 #endif
 811 #ifdef CONFIG_64BIT
 812                         "dla    $at,1f\n\t"
 813 #endif
 814                         "cache  %0,($at)\n\t"
 815                         "nop; nop; nop\n"
 816                         "1:\n\t"
 817                         ".set pop"
 818                         :
 819                         : "i" (Hit_Invalidate_I));
 820         }
 821         if (MIPS_CACHE_SYNC_WAR)
 822                 __asm__ __volatile__ ("sync");
 823 }
 824
 825 static void r4k_flush_cache_sigtramp(unsigned long addr)
 826 {
 827         r4k_on_each_cpu(local_r4k_flush_cache_sigtramp, (void *) addr);
 828 }
 829
 830 static void r4k_flush_icache_all(void)
 831 {
 832         if (cpu_has_vtag_icache)
 833                 r4k_blast_icache();
 834 }
 835
 836 struct flush_kernel_vmap_range_args {
 837         unsigned long   vaddr;
 838         int             size;
 839 };
 840
 841 static inline void local_r4k_flush_kernel_vmap_range(void *args)
 842 {
 843         struct flush_kernel_vmap_range_args *vmra = args;
 844         unsigned long vaddr = vmra->vaddr;
 845         int size = vmra->size;
 846
 847         /*
 848          * Aliases only affect the primary caches so don't bother with
 849          * S-caches or T-caches.
 850          */
 851         if (cpu_has_safe_index_cacheops && size >= dcache_size)
 852                 r4k_blast_dcache();
 853         else {
 854                 R4600_HIT_CACHEOP_WAR_IMPL;
 855                 blast_dcache_range(vaddr, vaddr + size);
 856         }
 857 }
 858
 859 static void r4k_flush_kernel_vmap_range(unsigned long vaddr, int size)
 860 {
 861         struct flush_kernel_vmap_range_args args;
 862
 863         args.vaddr = (unsigned long) vaddr;
 864         args.size = size;
 865
 866         r4k_on_each_cpu(local_r4k_flush_kernel_vmap_range, &args);
 867 }
 868
 869 static inline void rm7k_erratum31(void)
 870 {
 871         const unsigned long ic_lsize = 32;
 872         unsigned long addr;
 873
 874         /* RM7000 erratum #31. The icache is screwed at startup. */
 875         write_c0_taglo(0);
 876         write_c0_taghi(0);
 877
 878         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
 879                 __asm__ __volatile__ (
 880                         ".set push\n\t"
 881                         ".set noreorder\n\t"
 882                         ".set mips3\n\t"
 883                         "cache\t%1, 0(%0)\n\t"
 884                         "cache\t%1, 0x1000(%0)\n\t"
 885                         "cache\t%1, 0x2000(%0)\n\t"
 886                         "cache\t%1, 0x3000(%0)\n\t"
 887                         "cache\t%2, 0(%0)\n\t"
 888                         "cache\t%2, 0x1000(%0)\n\t"
 889                         "cache\t%2, 0x2000(%0)\n\t"
 890                         "cache\t%2, 0x3000(%0)\n\t"
 891                         "cache\t%1, 0(%0)\n\t"
 892                         "cache\t%1, 0x1000(%0)\n\t"
 893                         "cache\t%1, 0x2000(%0)\n\t"
 894                         "cache\t%1, 0x3000(%0)\n\t"
 895                         ".set pop\n"
 896                         :
 897                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
 898         }
 899 }
 900
 901 static inline int alias_74k_erratum(struct cpuinfo_mips *c)
 902 {
 903         unsigned int imp = c->processor_id & PRID_IMP_MASK;
 904         unsigned int rev = c->processor_id & PRID_REV_MASK;
 905         int present = 0;
 906
 907         /*
 908          * Early versions of the 74K do not update the cache tags on a
 909          * vtag miss/ptag hit which can occur in the case of KSEG0/KUSEG
 910          * aliases.  In this case it is better to treat the cache as always
 911          * having aliases.  Also disable the synonym tag update feature
 912          * where available.  In this case no opportunistic tag update will
 913          * happen where a load causes a virtual address miss but a physical
 914          * address hit during a D-cache look-up.
 915          */
 916         switch (imp) {
 917         case PRID_IMP_74K:
 918                 if (rev <= PRID_REV_ENCODE_332(2, 4, 0))
 919                         present = 1;
 920                 if (rev == PRID_REV_ENCODE_332(2, 4, 0))
 921                         write_c0_config6(read_c0_config6() | MIPS_CONF6_SYND);
 922                 break;
 923         case PRID_IMP_1074K:
 924                 if (rev <= PRID_REV_ENCODE_332(1, 1, 0)) {
 925                         present = 1;
 926                         write_c0_config6(read_c0_config6() | MIPS_CONF6_SYND);
 927                 }
 928                 break;
 929         default:
 930                 BUG();
 931         }
 932
 933         return present;
 934 }
 935
 936 static void b5k_instruction_hazard(void)
 937 {
 938         __sync();
 939         __sync();
 940         __asm__ __volatile__(
 941         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
 942         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
 943         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
 944         "       nop; nop; nop; nop; nop; nop; nop; nop\n"
 945         : : : "memory");
 946 }
 947
 948 static char *way_string[] = { NULL, "direct mapped", "2-way",
 949         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way",
 950         "9-way", "10-way", "11-way", "12-way",
 951         "13-way", "14-way", "15-way", "16-way",
 952 };
 953
 954 static void probe_pcache(void)
 955 {
 956         struct cpuinfo_mips *c = &current_cpu_data;
 957         unsigned int config = read_c0_config();
 958         unsigned int prid = read_c0_prid();
 959         int has_74k_erratum = 0;
 960         unsigned long config1;
 961         unsigned int lsize;
 962
 963         switch (current_cpu_type()) {
 964         case CPU_R4600:                 /* QED style two way caches? */
 965         case CPU_R4700:
 966         case CPU_R5000:
 967         case CPU_NEVADA:
 968                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
 969                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
 970                 c->icache.ways = 2;
 971                 c->icache.waybit = __ffs(icache_size/2);
 972
 973                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
 974                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
 975                 c->dcache.ways = 2;
 976                 c->dcache.waybit= __ffs(dcache_size/2);
 977
 978                 c->options |= MIPS_CPU_CACHE_CDEX_P;
 979                 break;
 980
 981         case CPU_R5432:
 982         case CPU_R5500:
 983                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
 984                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
 985                 c->icache.ways = 2;
 986                 c->icache.waybit= 0;
 987
 988                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
 989                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
 990                 c->dcache.ways = 2;
 991                 c->dcache.waybit = 0;
 992
 993                 c->options |= MIPS_CPU_CACHE_CDEX_P | MIPS_CPU_PREFETCH;
 994                 break;
 995
 996         case CPU_TX49XX:
 997                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
 998                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
 999                 c->icache.ways = 4;
1000                 c->icache.waybit= 0;
1001
1002                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1003                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1004                 c->dcache.ways = 4;
1005                 c->dcache.waybit = 0;
1006
1007                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1008                 c->options |= MIPS_CPU_PREFETCH;
1009                 break;
1010
1011         case CPU_R4000PC:
1012         case CPU_R4000SC:
1013         case CPU_R4000MC:
1014         case CPU_R4400PC:
1015         case CPU_R4400SC:
1016         case CPU_R4400MC:
1017         case CPU_R4300:
1018                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1019                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1020                 c->icache.ways = 1;
1021                 c->icache.waybit = 0;   /* doesn't matter */
1022
1023                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1024                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1025                 c->dcache.ways = 1;
1026                 c->dcache.waybit = 0;   /* does not matter */
1027
1028                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1029                 break;
1030
1031         case CPU_R10000:
1032         case CPU_R12000:
1033         case CPU_R14000:
1034         case CPU_R16000:
1035                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
1036                 c->icache.linesz = 64;
1037                 c->icache.ways = 2;
1038                 c->icache.waybit = 0;
1039
1040                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
1041                 c->dcache.linesz = 32;
1042                 c->dcache.ways = 2;
1043                 c->dcache.waybit = 0;
1044
1045                 c->options |= MIPS_CPU_PREFETCH;
1046                 break;
1047
1048         case CPU_VR4133:
1049                 write_c0_config(config & ~VR41_CONF_P4K);
1050         case CPU_VR4131:
1051                 /* Workaround for cache instruction bug of VR4131 */
1052                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
1053                     c->processor_id == 0x0c82U) {
1054                         config |= 0x00400000U;
1055                         if (c->processor_id == 0x0c80U)
1056                                 config |= VR41_CONF_BP;
1057                         write_c0_config(config);
1058                 } else
1059                         c->options |= MIPS_CPU_CACHE_CDEX_P;
1060
1061                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
1062                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1063                 c->icache.ways = 2;
1064                 c->icache.waybit = __ffs(icache_size/2);
1065
1066                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
1067                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1068                 c->dcache.ways = 2;
1069                 c->dcache.waybit = __ffs(dcache_size/2);
1070                 break;
1071
1072         case CPU_VR41XX:
1073         case CPU_VR4111:
1074         case CPU_VR4121:
1075         case CPU_VR4122:
1076         case CPU_VR4181:
1077         case CPU_VR4181A:
1078                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
1079                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1080                 c->icache.ways = 1;
1081                 c->icache.waybit = 0;   /* doesn't matter */
1082
1083                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
1084                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1085                 c->dcache.ways = 1;
1086                 c->dcache.waybit = 0;   /* does not matter */
1087
1088                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1089                 break;
1090
1091         case CPU_RM7000:
1092                 rm7k_erratum31();
1093
1094                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1095                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1096                 c->icache.ways = 4;
1097                 c->icache.waybit = __ffs(icache_size / c->icache.ways);
1098
1099                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1100                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1101                 c->dcache.ways = 4;
1102                 c->dcache.waybit = __ffs(dcache_size / c->dcache.ways);
1103
1104                 c->options |= MIPS_CPU_CACHE_CDEX_P;
1105                 c->options |= MIPS_CPU_PREFETCH;
1106                 break;
1107
1108         case CPU_LOONGSON2:
1109                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
1110                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
1111                 if (prid & 0x3)
1112                         c->icache.ways = 4;
1113                 else
1114                         c->icache.ways = 2;
1115                 c->icache.waybit = 0;
1116
1117                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
1118                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
1119                 if (prid & 0x3)
1120                         c->dcache.ways = 4;
1121                 else
1122                         c->dcache.ways = 2;
1123                 c->dcache.waybit = 0;
1124                 break;
1125
1126         case CPU_LOONGSON3:
1127                 config1 = read_c0_config1();
1128                 lsize = (config1 >> 19) & 7;
1129                 if (lsize)
1130                         c->icache.linesz = 2 << lsize;
1131                 else
1132                         c->icache.linesz = 0;
1133                 c->icache.sets = 64 << ((config1 >> 22) & 7);
1134                 c->icache.ways = 1 + ((config1 >> 16) & 7);
1135                 icache_size = c->icache.sets *
1136                                           c->icache.ways *
1137                                           c->icache.linesz;
1138                 c->icache.waybit = 0;
1139
1140                 lsize = (config1 >> 10) & 7;
1141                 if (lsize)
1142                         c->dcache.linesz = 2 << lsize;
1143                 else
1144                         c->dcache.linesz = 0;
1145                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
1146                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
1147                 dcache_size = c->dcache.sets *
1148                                           c->dcache.ways *
1149                                           c->dcache.linesz;
1150                 c->dcache.waybit = 0;
1151                 break;
1152
1153         case CPU_CAVIUM_OCTEON3:
1154                 /* For now lie about the number of ways. */
1155                 c->icache.linesz = 128;
1156                 c->icache.sets = 16;
1157                 c->icache.ways = 8;
1158                 c->icache.flags |= MIPS_CACHE_VTAG;
1159                 icache_size = c->icache.sets * c->icache.ways * c->icache.linesz;
1160
1161                 c->dcache.linesz = 128;
1162                 c->dcache.ways = 8;
1163                 c->dcache.sets = 8;
1164                 dcache_size = c->dcache.sets * c->dcache.ways * c->dcache.linesz;
1165                 c->options |= MIPS_CPU_PREFETCH;
1166                 break;
1167
1168         default:
1169                 if (!(config & MIPS_CONF_M))
1170                         panic("Don't know how to probe P-caches on this cpu.");
1171
1172                 /*
1173                  * So we seem to be a MIPS32 or MIPS64 CPU
1174                  * So let's probe the I-cache ...
1175                  */
1176                 config1 = read_c0_config1();
1177
1178                 lsize = (config1 >> 19) & 7;
1179
1180                 /* IL == 7 is reserved */
1181                 if (lsize == 7)
1182                         panic("Invalid icache line size");
1183
1184                 c->icache.linesz = lsize ? 2 << lsize : 0;
1185
1186                 c->icache.sets = 32 << (((config1 >> 22) + 1) & 7);
1187                 c->icache.ways = 1 + ((config1 >> 16) & 7);
1188
1189                 icache_size = c->icache.sets *
1190                               c->icache.ways *
1191                               c->icache.linesz;
1192                 c->icache.waybit = __ffs(icache_size/c->icache.ways);
1193
1194                 if (config & 0x8)               /* VI bit */
1195                         c->icache.flags |= MIPS_CACHE_VTAG;
1196
1197                 /*
1198                  * Now probe the MIPS32 / MIPS64 data cache.
1199                  */
1200                 c->dcache.flags = 0;
1201
1202                 lsize = (config1 >> 10) & 7;
1203
1204                 /* DL == 7 is reserved */
1205                 if (lsize == 7)
1206                         panic("Invalid dcache line size");
1207
1208                 c->dcache.linesz = lsize ? 2 << lsize : 0;
1209
1210                 c->dcache.sets = 32 << (((config1 >> 13) + 1) & 7);
1211                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
1212
1213                 dcache_size = c->dcache.sets *
1214                               c->dcache.ways *
1215                               c->dcache.linesz;
1216                 c->dcache.waybit = __ffs(dcache_size/c->dcache.ways);
1217
1218                 c->options |= MIPS_CPU_PREFETCH;
1219                 break;
1220         }
1221
1222         /*
1223          * Processor configuration sanity check for the R4000SC erratum
1224          * #5.  With page sizes larger than 32kB there is no possibility
1225          * to get a VCE exception anymore so we don't care about this
1226          * misconfiguration.  The case is rather theoretical anyway;
1227          * presumably no vendor is shipping his hardware in the "bad"
1228          * configuration.
1229          */
1230         if ((prid & PRID_IMP_MASK) == PRID_IMP_R4000 &&
1231             (prid & PRID_REV_MASK) < PRID_REV_R4400 &&
1232             !(config & CONF_SC) && c->icache.linesz != 16 &&
1233             PAGE_SIZE <= 0x8000)
1234                 panic("Improper R4000SC processor configuration detected");
1235
1236         /* compute a couple of other cache variables */
1237         c->icache.waysize = icache_size / c->icache.ways;
1238         c->dcache.waysize = dcache_size / c->dcache.ways;
1239
1240         c->icache.sets = c->icache.linesz ?
1241                 icache_size / (c->icache.linesz * c->icache.ways) : 0;
1242         c->dcache.sets = c->dcache.linesz ?
1243                 dcache_size / (c->dcache.linesz * c->dcache.ways) : 0;
1244
1245         /*
1246          * R1x000 P-caches are odd in a positive way.  They're 32kB 2-way
1247          * virtually indexed so normally would suffer from aliases.  So
1248          * normally they'd suffer from aliases but magic in the hardware deals
1249          * with that for us so we don't need to take care ourselves.
1250          */
1251         switch (current_cpu_type()) {
1252         case CPU_20KC:
1253         case CPU_25KF:
1254         case CPU_SB1:
1255         case CPU_SB1A:
1256         case CPU_XLR:
1257                 c->dcache.flags |= MIPS_CACHE_PINDEX;
1258                 break;
1259
1260         case CPU_R10000:
1261         case CPU_R12000:
1262         case CPU_R14000:
1263         case CPU_R16000:
1264                 break;
1265
1266         case CPU_74K:
1267         case CPU_1074K:
1268                 has_74k_erratum = alias_74k_erratum(c);
1269                 /* Fall through. */
1270         case CPU_M14KC:
1271         case CPU_M14KEC:
1272         case CPU_24K:
1273         case CPU_34K:
1274         case CPU_1004K:
1275         case CPU_INTERAPTIV:
1276         case CPU_P5600:
1277         case CPU_PROAPTIV:
1278         case CPU_M5150:
1279         case CPU_QEMU_GENERIC:
1280         case CPU_I6400:
1281                 if (!(read_c0_config7() & MIPS_CONF7_IAR) &&
1282                     (c->icache.waysize > PAGE_SIZE))
1283                         c->icache.flags |= MIPS_CACHE_ALIASES;
1284                 if (!has_74k_erratum && (read_c0_config7() & MIPS_CONF7_AR)) {
1285                         /*
1286                          * Effectively physically indexed dcache,
1287                          * thus no virtual aliases.
1288                         */
1289                         c->dcache.flags |= MIPS_CACHE_PINDEX;
1290                         break;
1291                 }
1292         default:
1293                 if (has_74k_erratum || c->dcache.waysize > PAGE_SIZE)
1294                         c->dcache.flags |= MIPS_CACHE_ALIASES;
1295         }
1296
1297         switch (current_cpu_type()) {
1298         case CPU_20KC:
1299                 /*
1300                  * Some older 20Kc chips doesn't have the 'VI' bit in
1301                  * the config register.
1302                  */
1303                 c->icache.flags |= MIPS_CACHE_VTAG;
1304                 break;
1305
1306         case CPU_ALCHEMY:
1307                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
1308                 break;
1309
1310         case CPU_LOONGSON2:
1311                 /*
1312                  * LOONGSON2 has 4 way icache, but when using indexed cache op,
1313                  * one op will act on all 4 ways
1314                  */
1315                 c->icache.ways = 1;
1316         }
1317
1318         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
1319                icache_size >> 10,
1320                c->icache.flags & MIPS_CACHE_VTAG ? "VIVT" : "VIPT",
1321                way_string[c->icache.ways], c->icache.linesz);
1322
1323         printk("Primary data cache %ldkB, %s, %s, %s, linesize %d bytes\n",
1324                dcache_size >> 10, way_string[c->dcache.ways],
1325                (c->dcache.flags & MIPS_CACHE_PINDEX) ? "PIPT" : "VIPT",
1326                (c->dcache.flags & MIPS_CACHE_ALIASES) ?
1327                         "cache aliases" : "no aliases",
1328                c->dcache.linesz);
1329 }
1330
1331 /*
1332  * If you even _breathe_ on this function, look at the gcc output and make sure
1333  * it does not pop things on and off the stack for the cache sizing loop that
1334  * executes in KSEG1 space or else you will crash and burn badly.  You have
1335  * been warned.
1336  */
1337 static int probe_scache(void)
1338 {
1339         unsigned long flags, addr, begin, end, pow2;
1340         unsigned int config = read_c0_config();
1341         struct cpuinfo_mips *c = &current_cpu_data;
1342
1343         if (config & CONF_SC)
1344                 return 0;
1345
1346         begin = (unsigned long) &_stext;
1347         begin &= ~((4 * 1024 * 1024) - 1);
1348         end = begin + (4 * 1024 * 1024);
1349
1350         /*
1351          * This is such a bitch, you'd think they would make it easy to do
1352          * this.  Away you daemons of stupidity!
1353          */
1354         local_irq_save(flags);
1355
1356         /* Fill each size-multiple cache line with a valid tag. */
1357         pow2 = (64 * 1024);
1358         for (addr = begin; addr < end; addr = (begin + pow2)) {
1359                 unsigned long *p = (unsigned long *) addr;
1360                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1361                 pow2 <<= 1;
1362         }
1363
1364         /* Load first line with zero (therefore invalid) tag. */
1365         write_c0_taglo(0);
1366         write_c0_taghi(0);
1367         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1368         cache_op(Index_Store_Tag_I, begin);
1369         cache_op(Index_Store_Tag_D, begin);
1370         cache_op(Index_Store_Tag_SD, begin);
1371
1372         /* Now search for the wrap around point. */
1373         pow2 = (128 * 1024);
1374         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1375                 cache_op(Index_Load_Tag_SD, addr);
1376                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1377                 if (!read_c0_taglo())
1378                         break;
1379                 pow2 <<= 1;
1380         }
1381         local_irq_restore(flags);
1382         addr -= begin;
1383
1384         scache_size = addr;
1385         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1386         c->scache.ways = 1;
1387         c->scache.waybit = 0;           /* does not matter */
1388
1389         return 1;
1390 }
1391
1392 static void __init loongson2_sc_init(void)
1393 {
1394         struct cpuinfo_mips *c = &current_cpu_data;
1395
1396         scache_size = 512*1024;
1397         c->scache.linesz = 32;
1398         c->scache.ways = 4;
1399         c->scache.waybit = 0;
1400         c->scache.waysize = scache_size / (c->scache.ways);
1401         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1402         pr_info("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1403                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1404
1405         c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1406 }
1407
1408 static void __init loongson3_sc_init(void)
1409 {
1410         struct cpuinfo_mips *c = &current_cpu_data;
1411         unsigned int config2, lsize;
1412
1413         config2 = read_c0_config2();
1414         lsize = (config2 >> 4) & 15;
1415         if (lsize)
1416                 c->scache.linesz = 2 << lsize;
1417         else
1418                 c->scache.linesz = 0;
1419         c->scache.sets = 64 << ((config2 >> 8) & 15);
1420         c->scache.ways = 1 + (config2 & 15);
1421
1422         scache_size = c->scache.sets *
1423                                   c->scache.ways *
1424                                   c->scache.linesz;
1425         /* Loongson-3 has 4 cores, 1MB scache for each. scaches are shared */
1426         scache_size *= 4;
1427         c->scache.waybit = 0;
1428         pr_info("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1429                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1430         if (scache_size)
1431                 c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1432         return;
1433 }
1434
1435 extern int r5k_sc_init(void);
1436 extern int rm7k_sc_init(void);
1437 extern int mips_sc_init(void);
1438
1439 static void setup_scache(void)
1440 {
1441         struct cpuinfo_mips *c = &current_cpu_data;
1442         unsigned int config = read_c0_config();
1443         int sc_present = 0;
1444
1445         /*
1446          * Do the probing thing on R4000SC and R4400SC processors.  Other
1447          * processors don't have a S-cache that would be relevant to the
1448          * Linux memory management.
1449          */
1450         switch (current_cpu_type()) {
1451         case CPU_R4000SC:
1452         case CPU_R4000MC:
1453         case CPU_R4400SC:
1454         case CPU_R4400MC:
1455                 sc_present = run_uncached(probe_scache);
1456                 if (sc_present)
1457                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1458                 break;
1459
1460         case CPU_R10000:
1461         case CPU_R12000:
1462         case CPU_R14000:
1463         case CPU_R16000:
1464                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1465                 c->scache.linesz = 64 << ((config >> 13) & 1);
1466                 c->scache.ways = 2;
1467                 c->scache.waybit= 0;
1468                 sc_present = 1;
1469                 break;
1470
1471         case CPU_R5000:
1472         case CPU_NEVADA:
1473 #ifdef CONFIG_R5000_CPU_SCACHE
1474                 r5k_sc_init();
1475 #endif
1476                 return;
1477
1478         case CPU_RM7000:
1479 #ifdef CONFIG_RM7000_CPU_SCACHE
1480                 rm7k_sc_init();
1481 #endif
1482                 return;
1483
1484         case CPU_LOONGSON2:
1485                 loongson2_sc_init();
1486                 return;
1487
1488         case CPU_LOONGSON3:
1489                 loongson3_sc_init();
1490                 return;
1491
1492         case CPU_CAVIUM_OCTEON3:
1493         case CPU_XLP:
1494                 /* don't need to worry about L2, fully coherent */
1495                 return;
1496
1497         default:
1498                 if (c->isa_level & (MIPS_CPU_ISA_M32R1 | MIPS_CPU_ISA_M32R2 |
1499                                     MIPS_CPU_ISA_M32R6 | MIPS_CPU_ISA_M64R1 |
1500                                     MIPS_CPU_ISA_M64R2 | MIPS_CPU_ISA_M64R6)) {
1501 #ifdef CONFIG_MIPS_CPU_SCACHE
1502                         if (mips_sc_init ()) {
1503                                 scache_size = c->scache.ways * c->scache.sets * c->scache.linesz;
1504                                 printk("MIPS secondary cache %ldkB, %s, linesize %d bytes.\n",
1505                                        scache_size >> 10,
1506                                        way_string[c->scache.ways], c->scache.linesz);
1507                         }
1508 #else
1509                         if (!(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1510                                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1511 #endif
1512                         return;
1513                 }
1514                 sc_present = 0;
1515         }
1516
1517         if (!sc_present)
1518                 return;
1519
1520         /* compute a couple of other cache variables */
1521         c->scache.waysize = scache_size / c->scache.ways;
1522
1523         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1524
1525         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1526                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1527
1528         c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1529 }
1530
1531 void au1x00_fixup_config_od(void)
1532 {
1533         /*
1534          * c0_config.od (bit 19) was write only (and read as 0)
1535          * on the early revisions of Alchemy SOCs.  It disables the bus
1536          * transaction overlapping and needs to be set to fix various errata.
1537          */
1538         switch (read_c0_prid()) {
1539         case 0x00030100: /* Au1000 DA */
1540         case 0x00030201: /* Au1000 HA */
1541         case 0x00030202: /* Au1000 HB */
1542         case 0x01030200: /* Au1500 AB */
1543         /*
1544          * Au1100 errata actually keeps silence about this bit, so we set it
1545          * just in case for those revisions that require it to be set according
1546          * to the (now gone) cpu table.
1547          */
1548         case 0x02030200: /* Au1100 AB */
1549         case 0x02030201: /* Au1100 BA */
1550         case 0x02030202: /* Au1100 BC */
1551                 set_c0_config(1 << 19);
1552                 break;
1553         }
1554 }
1555
1556 /* CP0 hazard avoidance. */
1557 #define NXP_BARRIER()                                                   \
1558          __asm__ __volatile__(                                          \
1559         ".set noreorder\n\t"                                            \
1560         "nop; nop; nop; nop; nop; nop;\n\t"                             \
1561         ".set reorder\n\t")
1562
1563 static void nxp_pr4450_fixup_config(void)
1564 {
1565         unsigned long config0;
1566
1567         config0 = read_c0_config();
1568
1569         /* clear all three cache coherency fields */
1570         config0 &= ~(0x7 | (7 << 25) | (7 << 28));
1571         config0 |= (((_page_cachable_default >> _CACHE_SHIFT) <<  0) |
1572                     ((_page_cachable_default >> _CACHE_SHIFT) << 25) |
1573                     ((_page_cachable_default >> _CACHE_SHIFT) << 28));
1574         write_c0_config(config0);
1575         NXP_BARRIER();
1576 }
1577
1578 static int cca = -1;
1579
1580 static int __init cca_setup(char *str)
1581 {
1582         get_option(&str, &cca);
1583
1584         return 0;
1585 }
1586
1587 early_param("cca", cca_setup);
1588
1589 static void coherency_setup(void)
1590 {
1591         if (cca < 0 || cca > 7)
1592                 cca = read_c0_config() & CONF_CM_CMASK;
1593         _page_cachable_default = cca << _CACHE_SHIFT;
1594
1595         pr_debug("Using cache attribute %d\n", cca);
1596         change_c0_config(CONF_CM_CMASK, cca);
1597
1598         /*
1599          * c0_status.cu=0 specifies that updates by the sc instruction use
1600          * the coherency mode specified by the TLB; 1 means cachable
1601          * coherent update on write will be used.  Not all processors have
1602          * this bit and; some wire it to zero, others like Toshiba had the
1603          * silly idea of putting something else there ...
1604          */
1605         switch (current_cpu_type()) {
1606         case CPU_R4000PC:
1607         case CPU_R4000SC:
1608         case CPU_R4000MC:
1609         case CPU_R4400PC:
1610         case CPU_R4400SC:
1611         case CPU_R4400MC:
1612                 clear_c0_config(CONF_CU);
1613                 break;
1614         /*
1615          * We need to catch the early Alchemy SOCs with
1616          * the write-only co_config.od bit and set it back to one on:
1617          * Au1000 rev DA, HA, HB;  Au1100 AB, BA, BC, Au1500 AB
1618          */
1619         case CPU_ALCHEMY:
1620                 au1x00_fixup_config_od();
1621                 break;
1622
1623         case PRID_IMP_PR4450:
1624                 nxp_pr4450_fixup_config();
1625                 break;
1626         }
1627 }
1628
1629 static void r4k_cache_error_setup(void)
1630 {
1631         extern char __weak except_vec2_generic;
1632         extern char __weak except_vec2_sb1;
1633
1634         switch (current_cpu_type()) {
1635         case CPU_SB1:
1636         case CPU_SB1A:
1637                 set_uncached_handler(0x100, &except_vec2_sb1, 0x80);
1638                 break;
1639
1640         default:
1641                 set_uncached_handler(0x100, &except_vec2_generic, 0x80);
1642                 break;
1643         }
1644 }
1645
1646 void r4k_cache_init(void)
1647 {
1648         extern void build_clear_page(void);
1649         extern void build_copy_page(void);
1650         struct cpuinfo_mips *c = &current_cpu_data;
1651
1652         probe_pcache();
1653         setup_scache();
1654
1655         r4k_blast_dcache_page_setup();
1656         r4k_blast_dcache_page_indexed_setup();
1657         r4k_blast_dcache_setup();
1658         r4k_blast_icache_page_setup();
1659         r4k_blast_icache_page_indexed_setup();
1660         r4k_blast_icache_setup();
1661         r4k_blast_scache_page_setup();
1662         r4k_blast_scache_page_indexed_setup();
1663         r4k_blast_scache_setup();
1664 #ifdef CONFIG_EVA
1665         r4k_blast_dcache_user_page_setup();
1666         r4k_blast_icache_user_page_setup();
1667 #endif
1668
1669         /*
1670          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1671          * This code supports virtually indexed processors and will be
1672          * unnecessarily inefficient on physically indexed processors.
1673          */
1674         if (c->dcache.linesz)
1675                 shm_align_mask = max_t( unsigned long,
1676                                         c->dcache.sets * c->dcache.linesz - 1,
1677                                         PAGE_SIZE - 1);
1678         else
1679                 shm_align_mask = PAGE_SIZE-1;
1680
1681         __flush_cache_vmap      = r4k__flush_cache_vmap;
1682         __flush_cache_vunmap    = r4k__flush_cache_vunmap;
1683
1684         flush_cache_all         = cache_noop;
1685         __flush_cache_all       = r4k___flush_cache_all;
1686         flush_cache_mm          = r4k_flush_cache_mm;
1687         flush_cache_page        = r4k_flush_cache_page;
1688         flush_cache_range       = r4k_flush_cache_range;
1689
1690         __flush_kernel_vmap_range = r4k_flush_kernel_vmap_range;
1691
1692         flush_cache_sigtramp    = r4k_flush_cache_sigtramp;
1693         flush_icache_all        = r4k_flush_icache_all;
1694         local_flush_data_cache_page     = local_r4k_flush_data_cache_page;
1695         flush_data_cache_page   = r4k_flush_data_cache_page;
1696         flush_icache_range      = r4k_flush_icache_range;
1697         local_flush_icache_range        = local_r4k_flush_icache_range;
1698
1699 #if defined(CONFIG_DMA_NONCOHERENT) || defined(CONFIG_DMA_MAYBE_COHERENT)
1700         if (coherentio) {
1701                 _dma_cache_wback_inv    = (void *)cache_noop;
1702                 _dma_cache_wback        = (void *)cache_noop;
1703                 _dma_cache_inv          = (void *)cache_noop;
1704         } else {
1705                 _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1706                 _dma_cache_wback        = r4k_dma_cache_wback_inv;
1707                 _dma_cache_inv          = r4k_dma_cache_inv;
1708         }
1709 #endif
1710
1711         build_clear_page();
1712         build_copy_page();
1713
1714         /*
1715          * We want to run CMP kernels on core with and without coherent
1716          * caches. Therefore, do not use CONFIG_MIPS_CMP to decide whether
1717          * or not to flush caches.
1718          */
1719         local_r4k___flush_cache_all(NULL);
1720
1721         coherency_setup();
1722         board_cache_error_setup = r4k_cache_error_setup;
1723
1724         /*
1725          * Per-CPU overrides
1726          */
1727         switch (current_cpu_type()) {
1728         case CPU_BMIPS4350:
1729         case CPU_BMIPS4380:
1730                 /* No IPI is needed because all CPUs share the same D$ */
1731                 flush_data_cache_page = r4k_blast_dcache_page;
1732                 break;
1733         case CPU_BMIPS5000:
1734                 /* We lose our superpowers if L2 is disabled */
1735                 if (c->scache.flags & MIPS_CACHE_NOT_PRESENT)
1736                         break;
1737
1738                 /* I$ fills from D$ just by emptying the write buffers */
1739                 flush_cache_page = (void *)b5k_instruction_hazard;
1740                 flush_cache_range = (void *)b5k_instruction_hazard;
1741                 flush_cache_sigtramp = (void *)b5k_instruction_hazard;
1742                 local_flush_data_cache_page = (void *)b5k_instruction_hazard;
1743                 flush_data_cache_page = (void *)b5k_instruction_hazard;
1744                 flush_icache_range = (void *)b5k_instruction_hazard;
1745                 local_flush_icache_range = (void *)b5k_instruction_hazard;
1746
1747                 /* Cache aliases are handled in hardware; allow HIGHMEM */
1748                 current_cpu_data.dcache.flags &= ~MIPS_CACHE_ALIASES;
1749
1750                 /* Optimization: an L2 flush implicitly flushes the L1 */
1751                 current_cpu_data.options |= MIPS_CPU_INCLUSIVE_CACHES;
1752                 break;
1753         }
1754 }
1755
1756 static int r4k_cache_pm_notifier(struct notifier_block *self, unsigned long cmd,
1757                                void *v)
1758 {
1759         switch (cmd) {
1760         case CPU_PM_ENTER_FAILED:
1761         case CPU_PM_EXIT:
1762                 coherency_setup();
1763                 break;
1764         }
1765
1766         return NOTIFY_OK;
1767 }
1768
1769 static struct notifier_block r4k_cache_pm_notifier_block = {
1770         .notifier_call = r4k_cache_pm_notifier,
1771 };
1772
1773 int __init r4k_cache_init_pm(void)
1774 {
1775         return cpu_pm_register_notifier(&r4k_cache_pm_notifier_block);
1776 }
1777 arch_initcall(r4k_cache_init_pm);