openwrt/target/linux/brcm-2.6/patches/003-bcm4710_cache_fixes.patch

   1 diff -urN linux.old/arch/mips/kernel/genex.S linux.dev/arch/mips/kernel/genex.S
   2 --- linux.old/arch/mips/kernel/genex.S  2006-03-20 06:53:29.000000000 +0100
   3 +++ linux.dev/arch/mips/kernel/genex.S  2006-03-21 12:19:26.000000000 +0100
   4 @@ -72,6 +72,10 @@
   5         .set    push
   6         .set    mips3
   7         .set    noat
   8 +#ifdef CONFIG_BCM4710
   9 +       nop
  10 +       nop
  11 +#endif
  12         mfc0    k1, CP0_CAUSE
  13         li      k0, 31<<2
  14         andi    k1, k1, 0x7c
  15 diff -urN linux.old/arch/mips/mm/c-r4k.c linux.dev/arch/mips/mm/c-r4k.c
  16 --- linux.old/arch/mips/mm/c-r4k.c      2006-03-20 06:53:29.000000000 +0100
  17 +++ linux.dev/arch/mips/mm/c-r4k.c      2006-03-21 12:19:26.000000000 +0100
  18 @@ -14,6 +14,12 @@
  19  #include <linux/mm.h>
  20  #include <linux/bitops.h>
  21
  22 +#ifdef CONFIG_BCM4710
  23 +#include "../bcm947xx/include/typedefs.h"
  24 +#include "../bcm947xx/include/sbconfig.h"
  25 +#include <asm/paccess.h>
  26 +#endif
  27 +
  28  #include <asm/bcache.h>
  29  #include <asm/bootinfo.h>
  30  #include <asm/cache.h>
  31 @@ -29,6 +35,9 @@
  32  #include <asm/war.h>
  33  #include <asm/cacheflush.h> /* for run_uncached() */
  34
  35 +/* For enabling BCM4710 cache workarounds */
  36 +int bcm4710 = 0;
  37 +
  38  /*
  39   * Must die.
  40   */
  41 @@ -73,7 +82,9 @@
  42  {
  43         unsigned long  dc_lsize = cpu_dcache_line_size();
  44
  45 -       if (dc_lsize == 16)
  46 +       if (bcm4710)
  47 +               r4k_blast_dcache_page = blast_dcache_page;
  48 +       else if (dc_lsize == 16)
  49                 r4k_blast_dcache_page = blast_dcache16_page;
  50         else if (dc_lsize == 32)
  51                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc32;
  52 @@ -85,7 +96,9 @@
  53  {
  54         unsigned long dc_lsize = cpu_dcache_line_size();
  55
  56 -       if (dc_lsize == 16)
  57 +       if (bcm4710)
  58 +               r4k_blast_dcache_page_indexed = blast_dcache_page_indexed;
  59 +       else if (dc_lsize == 16)
  60                 r4k_blast_dcache_page_indexed = blast_dcache16_page_indexed;
  61         else if (dc_lsize == 32)
  62                 r4k_blast_dcache_page_indexed = blast_dcache32_page_indexed;
  63 @@ -97,7 +110,9 @@
  64  {
  65         unsigned long dc_lsize = cpu_dcache_line_size();
  66
  67 -       if (dc_lsize == 16)
  68 +       if (bcm4710)
  69 +               r4k_blast_dcache = blast_dcache;
  70 +       else if (dc_lsize == 16)
  71                 r4k_blast_dcache = blast_dcache16;
  72         else if (dc_lsize == 32)
  73                 r4k_blast_dcache = blast_dcache32;
  74 @@ -660,6 +675,8 @@
  75         unsigned long addr = (unsigned long) arg;
  76
  77         R4600_HIT_CACHEOP_WAR_IMPL;
  78 +       BCM4710_PROTECTED_FILL_TLB(addr);
  79 +       BCM4710_PROTECTED_FILL_TLB(addr + 4);
  80         protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
  81         if (!cpu_icache_snoops_remote_store && scache_size)
  82                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
  83 @@ -1136,6 +1153,16 @@
  84  static inline void coherency_setup(void)
  85  {
  86         change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
  87 +#if defined(CONFIG_BCM4310) || defined(CONFIG_BCM4704) || defined(CONFIG_BCM5365)
  88 +       if (BCM330X(current_cpu_data.processor_id)) {
  89 +               __u32 cm = read_c0_diag();
  90 +               /* Enable icache */
  91 +               cm |= (1 << 31);
  92 +               /* Enable dcache */
  93 +               cm |= (1 << 30);
  94 +               write_c0_diag(cm);
  95 +       }
  96 +#endif
  97
  98         /*
  99          * c0_status.cu=0 specifies that updates by the sc instruction use
 100 @@ -1165,6 +1192,15 @@
 101
 102         /* Default cache error handler for R4000 and R5000 family */
 103         set_uncached_handler (0x100, &except_vec2_generic, 0x80);
 104 +
 105 +       /* Check if special workarounds are required */
 106 +#ifdef CONFIG_BCM4710
 107 +       if (current_cpu_data.cputype == CPU_BCM4710 && (current_cpu_data.processor_id & 0xff) == 0) {
 108 +               printk("Enabling BCM4710A0 cache workarounds.\n");
 109 +               bcm4710 = 1;
 110 +       } else
 111 +#endif
 112 +               bcm4710 = 0;
 113
 114         probe_pcache();
 115         setup_scache();
 116 diff -urN linux.old/arch/mips/mm/tlbex.c linux.dev/arch/mips/mm/tlbex.c
 117 --- linux.old/arch/mips/mm/tlbex.c      2006-03-21 12:12:38.000000000 +0100
 118 +++ linux.dev/arch/mips/mm/tlbex.c      2006-03-21 12:19:26.000000000 +0100
 119 @@ -28,6 +28,10 @@
 120
 121  /* #define DEBUG_TLB */
 122
 123 +#ifdef CONFIG_BCM4710
 124 +extern int bcm4710;
 125 +#endif
 126 +
 127  static __init int __attribute__((unused)) r45k_bvahwbug(void)
 128  {
 129         /* XXX: We should probe for the presence of this bug, but we don't. */
 130 @@ -1152,6 +1156,12 @@
 131         memset(relocs, 0, sizeof(relocs));
 132         memset(final_handler, 0, sizeof(final_handler));
 133
 134 +#ifdef CONFIG_BCM4710
 135 +       if (bcm4710) {
 136 +               i_nop(&p);
 137 +       }
 138 +#endif
 139 +
 140         /*
 141          * create the plain linear handler
 142          */
 143 diff -urN linux.old/include/asm-mips/r4kcache.h linux.dev/include/asm-mips/r4kcache.h
 144 --- linux.old/include/asm-mips/r4kcache.h       2006-03-20 06:53:29.000000000 +0100
 145 +++ linux.dev/include/asm-mips/r4kcache.h       2006-03-21 18:40:32.000000000 +0100
 146 @@ -16,6 +16,18 @@
 147  #include <asm/cacheops.h>
 148  #include <asm/cpu-features.h>
 149
 150 +#ifdef CONFIG_BCM4710
 151 +#define BCM4710_DUMMY_RREG() (((sbconfig_t *)(KSEG1ADDR(SB_ENUM_BASE + SBCONFIGOFF)))->sbimstate)
 152 +
 153 +#define BCM4710_FILL_TLB(addr) (*(volatile unsigned long *)(addr))
 154 +#define BCM4710_PROTECTED_FILL_TLB(addr) ({ unsigned long x; get_dbe(x, (volatile unsigned long *)(addr)); })
 155 +#else
 156 +#define BCM4710_DUMMY_RREG()
 157 +
 158 +#define BCM4710_FILL_TLB(addr)
 159 +#define BCM4710_PROTECTED_FILL_TLB(addr)
 160 +#endif
 161 +
 162  /*
 163   * This macro return a properly sign-extended address suitable as base address
 164   * for indexed cache operations.  Two issues here:
 165 @@ -46,6 +58,7 @@
 166
 167  static inline void flush_dcache_line_indexed(unsigned long addr)
 168  {
 169 +       BCM4710_DUMMY_RREG();
 170         cache_op(Index_Writeback_Inv_D, addr);
 171  }
 172
 173 @@ -61,11 +74,13 @@
 174
 175  static inline void flush_dcache_line(unsigned long addr)
 176  {
 177 +       BCM4710_DUMMY_RREG();
 178         cache_op(Hit_Writeback_Inv_D, addr);
 179  }
 180
 181  static inline void invalidate_dcache_line(unsigned long addr)
 182  {
 183 +       BCM4710_DUMMY_RREG();
 184         cache_op(Hit_Invalidate_D, addr);
 185  }
 186
 187 @@ -97,6 +112,7 @@
 188   */
 189  static inline void protected_flush_icache_line(unsigned long addr)
 190  {
 191 +       BCM4710_DUMMY_RREG();
 192         protected_cache_op(Hit_Invalidate_I, addr);
 193  }
 194
 195 @@ -108,6 +124,7 @@
 196   */
 197  static inline void protected_writeback_dcache_line(unsigned long addr)
 198  {
 199 +       BCM4710_DUMMY_RREG();
 200         protected_cache_op(Hit_Writeback_Inv_D, addr);
 201  }
 202
 203 @@ -228,8 +245,52 @@
 204                 : "r" (base),                                           \
 205                   "i" (op));
 206
 207 +static inline void blast_dcache(void)
 208 +{
 209 +       unsigned long start = KSEG0;
 210 +       unsigned long dcache_size = current_cpu_data.dcache.waysize * current_cpu_data.dcache.ways;
 211 +       unsigned long end = (start + dcache_size);
 212 +
 213 +       do {
 214 +               BCM4710_DUMMY_RREG();
 215 +               cache_op(Index_Writeback_Inv_D, start);
 216 +               start += current_cpu_data.dcache.linesz;
 217 +       } while(start < end);
 218 +}
 219 +
 220 +static inline void blast_dcache_page(unsigned long page)
 221 +{
 222 +       unsigned long start = page;
 223 +       unsigned long end = start + PAGE_SIZE;
 224 +
 225 +       BCM4710_FILL_TLB(start);
 226 +       do {
 227 +               BCM4710_DUMMY_RREG();
 228 +               cache_op(Hit_Writeback_Inv_D, start);
 229 +               start += current_cpu_data.dcache.linesz;
 230 +       } while(start < end);
 231 +}
 232 +
 233 +static inline void blast_dcache_page_indexed(unsigned long page)
 234 +{
 235 +       unsigned long start = page;
 236 +       unsigned long end = start + PAGE_SIZE;
 237 +       unsigned long ws_inc = 1UL << current_cpu_data.dcache.waybit;
 238 +       unsigned long ws_end = current_cpu_data.dcache.ways <<
 239 +                              current_cpu_data.dcache.waybit;
 240 +       unsigned long ws, addr;
 241 +       for (ws = 0; ws < ws_end; ws += ws_inc) {
 242 +               start = page + ws;
 243 +               for (addr = start; addr < end; addr += current_cpu_data.dcache.linesz) {
 244 +                       BCM4710_DUMMY_RREG();
 245 +                       cache_op(Index_Writeback_Inv_D, addr);
 246 +               }
 247 +       }
 248 +}
 249 +
 250 +
 251  /* build blast_xxx, blast_xxx_page, blast_xxx_page_indexed */
 252 -#define __BUILD_BLAST_CACHE(pfx, desc, indexop, hitop, lsize) \
 253 +#define __BUILD_BLAST_CACHE(pfx, desc, indexop, hitop, lsize, war) \
 254  static inline void blast_##pfx##cache##lsize(void)                     \
 255  {                                                                      \
 256         unsigned long start = INDEX_BASE;                               \
 257 @@ -239,6 +300,7 @@
 258                                current_cpu_data.desc.waybit;            \
 259         unsigned long ws, addr;                                         \
 260                                                                         \
 261 +       war                                                             \
 262         for (ws = 0; ws < ws_end; ws += ws_inc)                         \
 263                 for (addr = start; addr < end; addr += lsize * 32)      \
 264                         cache##lsize##_unroll32(addr|ws,indexop);       \
 265 @@ -249,6 +311,7 @@
 266         unsigned long start = page;                                     \
 267         unsigned long end = page + PAGE_SIZE;                           \
 268                                                                         \
 269 +       war                                                             \
 270         do {                                                            \
 271                 cache##lsize##_unroll32(start,hitop);                   \
 272                 start += lsize * 32;                                    \
 273 @@ -265,29 +328,31 @@
 274                                current_cpu_data.desc.waybit;            \
 275         unsigned long ws, addr;                                         \
 276                                                                         \
 277 +       war                                                             \
 278         for (ws = 0; ws < ws_end; ws += ws_inc)                         \
 279                 for (addr = start; addr < end; addr += lsize * 32)      \
 280                         cache##lsize##_unroll32(addr|ws,indexop);       \
 281  }
 282
 283 -__BUILD_BLAST_CACHE(d, dcache, Index_Writeback_Inv_D, Hit_Writeback_Inv_D, 16)
 284 -__BUILD_BLAST_CACHE(i, icache, Index_Invalidate_I, Hit_Invalidate_I, 16)
 285 -__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 16)
 286 -__BUILD_BLAST_CACHE(d, dcache, Index_Writeback_Inv_D, Hit_Writeback_Inv_D, 32)
 287 -__BUILD_BLAST_CACHE(i, icache, Index_Invalidate_I, Hit_Invalidate_I, 32)
 288 -__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 32)
 289 -__BUILD_BLAST_CACHE(i, icache, Index_Invalidate_I, Hit_Invalidate_I, 64)
 290 -__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 64)
 291 -__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 128)
 292 +__BUILD_BLAST_CACHE(d, dcache, Index_Writeback_Inv_D, Hit_Writeback_Inv_D, 16, )
 293 +__BUILD_BLAST_CACHE(i, icache, Index_Invalidate_I, Hit_Invalidate_I, 16, BCM4710_FILL_TLB(start);)
 294 +__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 16, )
 295 +__BUILD_BLAST_CACHE(d, dcache, Index_Writeback_Inv_D, Hit_Writeback_Inv_D, 32, )
 296 +__BUILD_BLAST_CACHE(i, icache, Index_Invalidate_I, Hit_Invalidate_I, 32, BCM4710_FILL_TLB(start);)
 297 +__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 32, )
 298 +__BUILD_BLAST_CACHE(i, icache, Index_Invalidate_I, Hit_Invalidate_I, 64, BCM4710_FILL_TLB(start);)
 299 +__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 64, )
 300 +__BUILD_BLAST_CACHE(s, scache, Index_Writeback_Inv_SD, Hit_Writeback_Inv_SD, 128, )
 301
 302  /* build blast_xxx_range, protected_blast_xxx_range */
 303 -#define __BUILD_BLAST_CACHE_RANGE(pfx, desc, hitop, prot) \
 304 +#define __BUILD_BLAST_CACHE_RANGE(pfx, desc, hitop, prot, war) \
 305  static inline void prot##blast_##pfx##cache##_range(unsigned long start, \
 306                                                     unsigned long end)  \
 307  {                                                                      \
 308         unsigned long lsize = cpu_##desc##_line_size();                 \
 309         unsigned long addr = start & ~(lsize - 1);                      \
 310         unsigned long aend = (end - 1) & ~(lsize - 1);                  \
 311 +       war                                                             \
 312         while (1) {                                                     \
 313                 prot##cache_op(hitop, addr);                            \
 314                 if (addr == aend)                                       \
 315 @@ -296,12 +361,12 @@
 316         }                                                               \
 317  }
 318
 319 -__BUILD_BLAST_CACHE_RANGE(d, dcache, Hit_Writeback_Inv_D, protected_)
 320 -__BUILD_BLAST_CACHE_RANGE(s, scache, Hit_Writeback_Inv_SD, protected_)
 321 -__BUILD_BLAST_CACHE_RANGE(i, icache, Hit_Invalidate_I, protected_)
 322 -__BUILD_BLAST_CACHE_RANGE(d, dcache, Hit_Writeback_Inv_D, )
 323 -__BUILD_BLAST_CACHE_RANGE(s, scache, Hit_Writeback_Inv_SD, )
 324 +__BUILD_BLAST_CACHE_RANGE(d, dcache, Hit_Writeback_Inv_D, protected_, BCM4710_PROTECTED_FILL_TLB(addr); BCM4710_PROTECTED_FILL_TLB(aend);)
 325 +__BUILD_BLAST_CACHE_RANGE(s, scache, Hit_Writeback_Inv_SD, protected_, )
 326 +__BUILD_BLAST_CACHE_RANGE(i, icache, Hit_Invalidate_I, protected_, )
 327 +__BUILD_BLAST_CACHE_RANGE(d, dcache, Hit_Writeback_Inv_D, , BCM4710_FILL_TLB(addr); BCM4710_FILL_TLB(aend);)
 328 +__BUILD_BLAST_CACHE_RANGE(s, scache, Hit_Writeback_Inv_SD, , )
 329  /* blast_inv_dcache_range */
 330 -__BUILD_BLAST_CACHE_RANGE(inv_d, dcache, Hit_Invalidate_D, )
 331 +__BUILD_BLAST_CACHE_RANGE(inv_d, dcache, Hit_Invalidate_D, , )
 332
 333  #endif /* _ASM_R4KCACHE_H */
 334 diff -urN linux.old/include/asm-mips/stackframe.h linux.dev/include/asm-mips/stackframe.h
 335 --- linux.old/include/asm-mips/stackframe.h     2006-03-20 06:53:29.000000000 +0100
 336 +++ linux.dev/include/asm-mips/stackframe.h     2006-03-21 12:19:26.000000000 +0100
 337 @@ -285,6 +285,10 @@
 338                 .macro  RESTORE_SP_AND_RET
 339                 LONG_L  sp, PT_R29(sp)
 340                 .set    mips3
 341 +#ifdef CONFIG_BCM4710
 342 +               nop
 343 +               nop
 344 +#endif
 345                 eret
 346                 .set    mips0
 347                 .endm