arch/mips/mm/page.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) 2003, 04, 05 Ralf Baechle (ralf@linux-mips.org)
   7  * Copyright (C) 2007  Maciej W. Rozycki
   8  * Copyright (C) 2008  Thiemo Seufer
   9  * Copyright (C) 2012  MIPS Technologies, Inc.
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/smp.h>
  14 #include <linux/mm.h>
  15 #include <linux/proc_fs.h>
  16
  17 #include <asm/bugs.h>
  18 #include <asm/cacheops.h>
  19 #include <asm/cpu-type.h>
  20 #include <asm/inst.h>
  21 #include <asm/io.h>
  22 #include <asm/page.h>
  23 #include <asm/prefetch.h>
  24 #include <asm/bootinfo.h>
  25 #include <asm/mipsregs.h>
  26 #include <asm/mmu_context.h>
  27 #include <asm/cpu.h>
  28 #include <asm/war.h>
  29
  30 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
  31 #include <asm/sibyte/sb1250.h>
  32 #include <asm/sibyte/sb1250_regs.h>
  33 #include <asm/sibyte/sb1250_dma.h>
  34 #endif
  35
  36 #include <asm/uasm.h>
  37
  38 /* Registers used in the assembled routines. */
  39 #define ZERO 0
  40 #define AT 2
  41 #define A0 4
  42 #define A1 5
  43 #define A2 6
  44 #define T0 8
  45 #define T1 9
  46 #define T2 10
  47 #define T3 11
  48 #define T9 25
  49 #define RA 31
  50
  51 /* Handle labels (which must be positive integers). */
  52 enum label_id {
  53         label_clear_nopref = 1,
  54         label_clear_pref,
  55         label_copy_nopref,
  56         label_copy_pref_both,
  57         label_copy_pref_store,
  58 };
  59
  60 UASM_L_LA(_clear_nopref)
  61 UASM_L_LA(_clear_pref)
  62 UASM_L_LA(_copy_nopref)
  63 UASM_L_LA(_copy_pref_both)
  64 UASM_L_LA(_copy_pref_store)
  65
  66 /* We need one branch and therefore one relocation per target label. */
  67 static struct uasm_label labels[5];
  68 static struct uasm_reloc relocs[5];
  69
  70 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002010)
  71 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002020)
  72
  73 /*
  74  * R6 has a limited offset of the pref instruction.
  75  * Skip it if the offset is more than 9 bits.
  76  */
  77 #define _uasm_i_pref(a, b, c, d)                \
  78 do {                                            \
  79         if (cpu_has_mips_r6) {                  \
  80                 if (c <= 0xff && c >= -0x100)   \
  81                         uasm_i_pref(a, b, c, d);\
  82         } else {                                \
  83                 uasm_i_pref(a, b, c, d);        \
  84         }                                       \
  85 } while(0)
  86
  87 static int pref_bias_clear_store;
  88 static int pref_bias_copy_load;
  89 static int pref_bias_copy_store;
  90
  91 static u32 pref_src_mode;
  92 static u32 pref_dst_mode;
  93
  94 static int clear_word_size;
  95 static int copy_word_size;
  96
  97 static int half_clear_loop_size;
  98 static int half_copy_loop_size;
  99
 100 static int cache_line_size;
 101 #define cache_line_mask() (cache_line_size - 1)
 102
 103 static inline void
 104 pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
 105 {
 106         if (cpu_has_64bit_gp_regs && DADDI_WAR && r4k_daddiu_bug()) {
 107                 if (off > 0x7fff) {
 108                         uasm_i_lui(buf, T9, uasm_rel_hi(off));
 109                         uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
 110                 } else
 111                         uasm_i_addiu(buf, T9, ZERO, off);
 112                 uasm_i_daddu(buf, reg1, reg2, T9);
 113         } else {
 114                 if (off > 0x7fff) {
 115                         uasm_i_lui(buf, T9, uasm_rel_hi(off));
 116                         uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
 117                         UASM_i_ADDU(buf, reg1, reg2, T9);
 118                 } else
 119                         UASM_i_ADDIU(buf, reg1, reg2, off);
 120         }
 121 }
 122
 123 static void set_prefetch_parameters(void)
 124 {
 125         if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
 126                 clear_word_size = 8;
 127         else
 128                 clear_word_size = 4;
 129
 130         if (cpu_has_64bit_gp_regs)
 131                 copy_word_size = 8;
 132         else
 133                 copy_word_size = 4;
 134
 135         /*
 136          * The pref's used here are using "streaming" hints, which cause the
 137          * copied data to be kicked out of the cache sooner.  A page copy often
 138          * ends up copying a lot more data than is commonly used, so this seems
 139          * to make sense in terms of reducing cache pollution, but I've no real
 140          * performance data to back this up.
 141          */
 142         if (cpu_has_prefetch) {
 143                 /*
 144                  * XXX: Most prefetch bias values in here are based on
 145                  * guesswork.
 146                  */
 147                 cache_line_size = cpu_dcache_line_size();
 148                 switch (current_cpu_type()) {
 149                 case CPU_R5500:
 150                 case CPU_TX49XX:
 151                         /* These processors only support the Pref_Load. */
 152                         pref_bias_copy_load = 256;
 153                         break;
 154
 155                 case CPU_R10000:
 156                 case CPU_R12000:
 157                 case CPU_R14000:
 158                 case CPU_R16000:
 159                         /*
 160                          * Those values have been experimentally tuned for an
 161                          * Origin 200.
 162                          */
 163                         pref_bias_clear_store = 512;
 164                         pref_bias_copy_load = 256;
 165                         pref_bias_copy_store = 256;
 166                         pref_src_mode = Pref_LoadStreamed;
 167                         pref_dst_mode = Pref_StoreStreamed;
 168                         break;
 169
 170                 case CPU_SB1:
 171                 case CPU_SB1A:
 172                         pref_bias_clear_store = 128;
 173                         pref_bias_copy_load = 128;
 174                         pref_bias_copy_store = 128;
 175                         /*
 176                          * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
 177                          * hints are broken.
 178                          */
 179                         if (current_cpu_type() == CPU_SB1 &&
 180                             (current_cpu_data.processor_id & 0xff) < 0x02) {
 181                                 pref_src_mode = Pref_Load;
 182                                 pref_dst_mode = Pref_Store;
 183                         } else {
 184                                 pref_src_mode = Pref_LoadStreamed;
 185                                 pref_dst_mode = Pref_StoreStreamed;
 186                         }
 187                         break;
 188
 189                 case CPU_LOONGSON64:
 190                         /* Loongson-3 only support the Pref_Load/Pref_Store. */
 191                         pref_bias_clear_store = 128;
 192                         pref_bias_copy_load = 128;
 193                         pref_bias_copy_store = 128;
 194                         pref_src_mode = Pref_Load;
 195                         pref_dst_mode = Pref_Store;
 196                         break;
 197
 198                 default:
 199                         pref_bias_clear_store = 128;
 200                         pref_bias_copy_load = 256;
 201                         pref_bias_copy_store = 128;
 202                         pref_src_mode = Pref_LoadStreamed;
 203                         if (cpu_has_mips_r6)
 204                                 /*
 205                                  * Bit 30 (Pref_PrepareForStore) has been
 206                                  * removed from MIPS R6. Use bit 5
 207                                  * (Pref_StoreStreamed).
 208                                  */
 209                                 pref_dst_mode = Pref_StoreStreamed;
 210                         else
 211                                 pref_dst_mode = Pref_PrepareForStore;
 212                         break;
 213                 }
 214         } else {
 215                 if (cpu_has_cache_cdex_s)
 216                         cache_line_size = cpu_scache_line_size();
 217                 else if (cpu_has_cache_cdex_p)
 218                         cache_line_size = cpu_dcache_line_size();
 219         }
 220         /*
 221          * Too much unrolling will overflow the available space in
 222          * clear_space_array / copy_page_array.
 223          */
 224         half_clear_loop_size = min(16 * clear_word_size,
 225                                    max(cache_line_size >> 1,
 226                                        4 * clear_word_size));
 227         half_copy_loop_size = min(16 * copy_word_size,
 228                                   max(cache_line_size >> 1,
 229                                       4 * copy_word_size));
 230 }
 231
 232 static void build_clear_store(u32 **buf, int off)
 233 {
 234         if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
 235                 uasm_i_sd(buf, ZERO, off, A0);
 236         } else {
 237                 uasm_i_sw(buf, ZERO, off, A0);
 238         }
 239 }
 240
 241 static inline void build_clear_pref(u32 **buf, int off)
 242 {
 243         if (off & cache_line_mask())
 244                 return;
 245
 246         if (pref_bias_clear_store) {
 247                 _uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
 248                             A0);
 249         } else if (cache_line_size == (half_clear_loop_size << 1)) {
 250                 if (cpu_has_cache_cdex_s) {
 251                         uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
 252                 } else if (cpu_has_cache_cdex_p) {
 253                         if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) &&
 254                             cpu_is_r4600_v1_x()) {
 255                                 uasm_i_nop(buf);
 256                                 uasm_i_nop(buf);
 257                                 uasm_i_nop(buf);
 258                                 uasm_i_nop(buf);
 259                         }
 260
 261                         if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) &&
 262                             cpu_is_r4600_v2_x())
 263                                 uasm_i_lw(buf, ZERO, ZERO, AT);
 264
 265                         uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
 266                 }
 267         }
 268 }
 269
 270 extern u32 __clear_page_start;
 271 extern u32 __clear_page_end;
 272 extern u32 __copy_page_start;
 273 extern u32 __copy_page_end;
 274
 275 void build_clear_page(void)
 276 {
 277         int off;
 278         u32 *buf = &__clear_page_start;
 279         struct uasm_label *l = labels;
 280         struct uasm_reloc *r = relocs;
 281         int i;
 282         static atomic_t run_once = ATOMIC_INIT(0);
 283
 284         if (atomic_xchg(&run_once, 1)) {
 285                 return;
 286         }
 287
 288         memset(labels, 0, sizeof(labels));
 289         memset(relocs, 0, sizeof(relocs));
 290
 291         set_prefetch_parameters();
 292
 293         /*
 294          * This algorithm makes the following assumptions:
 295          *   - The prefetch bias is a multiple of 2 words.
 296          *   - The prefetch bias is less than one page.
 297          */
 298         BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
 299         BUG_ON(PAGE_SIZE < pref_bias_clear_store);
 300
 301         off = PAGE_SIZE - pref_bias_clear_store;
 302         if (off > 0xffff || !pref_bias_clear_store)
 303                 pg_addiu(&buf, A2, A0, off);
 304         else
 305                 uasm_i_ori(&buf, A2, A0, off);
 306
 307         if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x())
 308                 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
 309
 310         off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
 311                                 * cache_line_size : 0;
 312         while (off) {
 313                 build_clear_pref(&buf, -off);
 314                 off -= cache_line_size;
 315         }
 316         uasm_l_clear_pref(&l, buf);
 317         do {
 318                 build_clear_pref(&buf, off);
 319                 build_clear_store(&buf, off);
 320                 off += clear_word_size;
 321         } while (off < half_clear_loop_size);
 322         pg_addiu(&buf, A0, A0, 2 * off);
 323         off = -off;
 324         do {
 325                 build_clear_pref(&buf, off);
 326                 if (off == -clear_word_size)
 327                         uasm_il_bne(&buf, &r, A0, A2, label_clear_pref);
 328                 build_clear_store(&buf, off);
 329                 off += clear_word_size;
 330         } while (off < 0);
 331
 332         if (pref_bias_clear_store) {
 333                 pg_addiu(&buf, A2, A0, pref_bias_clear_store);
 334                 uasm_l_clear_nopref(&l, buf);
 335                 off = 0;
 336                 do {
 337                         build_clear_store(&buf, off);
 338                         off += clear_word_size;
 339                 } while (off < half_clear_loop_size);
 340                 pg_addiu(&buf, A0, A0, 2 * off);
 341                 off = -off;
 342                 do {
 343                         if (off == -clear_word_size)
 344                                 uasm_il_bne(&buf, &r, A0, A2,
 345                                             label_clear_nopref);
 346                         build_clear_store(&buf, off);
 347                         off += clear_word_size;
 348                 } while (off < 0);
 349         }
 350
 351         uasm_i_jr(&buf, RA);
 352         uasm_i_nop(&buf);
 353
 354         BUG_ON(buf > &__clear_page_end);
 355
 356         uasm_resolve_relocs(relocs, labels);
 357
 358         pr_debug("Synthesized clear page handler (%u instructions).\n",
 359                  (u32)(buf - &__clear_page_start));
 360
 361         pr_debug("\t.set push\n");
 362         pr_debug("\t.set noreorder\n");
 363         for (i = 0; i < (buf - &__clear_page_start); i++)
 364                 pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]);
 365         pr_debug("\t.set pop\n");
 366 }
 367
 368 static void build_copy_load(u32 **buf, int reg, int off)
 369 {
 370         if (cpu_has_64bit_gp_regs) {
 371                 uasm_i_ld(buf, reg, off, A1);
 372         } else {
 373                 uasm_i_lw(buf, reg, off, A1);
 374         }
 375 }
 376
 377 static void build_copy_store(u32 **buf, int reg, int off)
 378 {
 379         if (cpu_has_64bit_gp_regs) {
 380                 uasm_i_sd(buf, reg, off, A0);
 381         } else {
 382                 uasm_i_sw(buf, reg, off, A0);
 383         }
 384 }
 385
 386 static inline void build_copy_load_pref(u32 **buf, int off)
 387 {
 388         if (off & cache_line_mask())
 389                 return;
 390
 391         if (pref_bias_copy_load)
 392                 _uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, A1);
 393 }
 394
 395 static inline void build_copy_store_pref(u32 **buf, int off)
 396 {
 397         if (off & cache_line_mask())
 398                 return;
 399
 400         if (pref_bias_copy_store) {
 401                 _uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
 402                             A0);
 403         } else if (cache_line_size == (half_copy_loop_size << 1)) {
 404                 if (cpu_has_cache_cdex_s) {
 405                         uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
 406                 } else if (cpu_has_cache_cdex_p) {
 407                         if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) &&
 408                             cpu_is_r4600_v1_x()) {
 409                                 uasm_i_nop(buf);
 410                                 uasm_i_nop(buf);
 411                                 uasm_i_nop(buf);
 412                                 uasm_i_nop(buf);
 413                         }
 414
 415                         if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) &&
 416                             cpu_is_r4600_v2_x())
 417                                 uasm_i_lw(buf, ZERO, ZERO, AT);
 418
 419                         uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
 420                 }
 421         }
 422 }
 423
 424 void build_copy_page(void)
 425 {
 426         int off;
 427         u32 *buf = &__copy_page_start;
 428         struct uasm_label *l = labels;
 429         struct uasm_reloc *r = relocs;
 430         int i;
 431         static atomic_t run_once = ATOMIC_INIT(0);
 432
 433         if (atomic_xchg(&run_once, 1)) {
 434                 return;
 435         }
 436
 437         memset(labels, 0, sizeof(labels));
 438         memset(relocs, 0, sizeof(relocs));
 439
 440         set_prefetch_parameters();
 441
 442         /*
 443          * This algorithm makes the following assumptions:
 444          *   - All prefetch biases are multiples of 8 words.
 445          *   - The prefetch biases are less than one page.
 446          *   - The store prefetch bias isn't greater than the load
 447          *     prefetch bias.
 448          */
 449         BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
 450         BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
 451         BUG_ON(PAGE_SIZE < pref_bias_copy_load);
 452         BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
 453
 454         off = PAGE_SIZE - pref_bias_copy_load;
 455         if (off > 0xffff || !pref_bias_copy_load)
 456                 pg_addiu(&buf, A2, A0, off);
 457         else
 458                 uasm_i_ori(&buf, A2, A0, off);
 459
 460         if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x())
 461                 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
 462
 463         off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
 464                                 cache_line_size : 0;
 465         while (off) {
 466                 build_copy_load_pref(&buf, -off);
 467                 off -= cache_line_size;
 468         }
 469         off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
 470                                 cache_line_size : 0;
 471         while (off) {
 472                 build_copy_store_pref(&buf, -off);
 473                 off -= cache_line_size;
 474         }
 475         uasm_l_copy_pref_both(&l, buf);
 476         do {
 477                 build_copy_load_pref(&buf, off);
 478                 build_copy_load(&buf, T0, off);
 479                 build_copy_load_pref(&buf, off + copy_word_size);
 480                 build_copy_load(&buf, T1, off + copy_word_size);
 481                 build_copy_load_pref(&buf, off + 2 * copy_word_size);
 482                 build_copy_load(&buf, T2, off + 2 * copy_word_size);
 483                 build_copy_load_pref(&buf, off + 3 * copy_word_size);
 484                 build_copy_load(&buf, T3, off + 3 * copy_word_size);
 485                 build_copy_store_pref(&buf, off);
 486                 build_copy_store(&buf, T0, off);
 487                 build_copy_store_pref(&buf, off + copy_word_size);
 488                 build_copy_store(&buf, T1, off + copy_word_size);
 489                 build_copy_store_pref(&buf, off + 2 * copy_word_size);
 490                 build_copy_store(&buf, T2, off + 2 * copy_word_size);
 491                 build_copy_store_pref(&buf, off + 3 * copy_word_size);
 492                 build_copy_store(&buf, T3, off + 3 * copy_word_size);
 493                 off += 4 * copy_word_size;
 494         } while (off < half_copy_loop_size);
 495         pg_addiu(&buf, A1, A1, 2 * off);
 496         pg_addiu(&buf, A0, A0, 2 * off);
 497         off = -off;
 498         do {
 499                 build_copy_load_pref(&buf, off);
 500                 build_copy_load(&buf, T0, off);
 501                 build_copy_load_pref(&buf, off + copy_word_size);
 502                 build_copy_load(&buf, T1, off + copy_word_size);
 503                 build_copy_load_pref(&buf, off + 2 * copy_word_size);
 504                 build_copy_load(&buf, T2, off + 2 * copy_word_size);
 505                 build_copy_load_pref(&buf, off + 3 * copy_word_size);
 506                 build_copy_load(&buf, T3, off + 3 * copy_word_size);
 507                 build_copy_store_pref(&buf, off);
 508                 build_copy_store(&buf, T0, off);
 509                 build_copy_store_pref(&buf, off + copy_word_size);
 510                 build_copy_store(&buf, T1, off + copy_word_size);
 511                 build_copy_store_pref(&buf, off + 2 * copy_word_size);
 512                 build_copy_store(&buf, T2, off + 2 * copy_word_size);
 513                 build_copy_store_pref(&buf, off + 3 * copy_word_size);
 514                 if (off == -(4 * copy_word_size))
 515                         uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both);
 516                 build_copy_store(&buf, T3, off + 3 * copy_word_size);
 517                 off += 4 * copy_word_size;
 518         } while (off < 0);
 519
 520         if (pref_bias_copy_load - pref_bias_copy_store) {
 521                 pg_addiu(&buf, A2, A0,
 522                          pref_bias_copy_load - pref_bias_copy_store);
 523                 uasm_l_copy_pref_store(&l, buf);
 524                 off = 0;
 525                 do {
 526                         build_copy_load(&buf, T0, off);
 527                         build_copy_load(&buf, T1, off + copy_word_size);
 528                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 529                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 530                         build_copy_store_pref(&buf, off);
 531                         build_copy_store(&buf, T0, off);
 532                         build_copy_store_pref(&buf, off + copy_word_size);
 533                         build_copy_store(&buf, T1, off + copy_word_size);
 534                         build_copy_store_pref(&buf, off + 2 * copy_word_size);
 535                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 536                         build_copy_store_pref(&buf, off + 3 * copy_word_size);
 537                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 538                         off += 4 * copy_word_size;
 539                 } while (off < half_copy_loop_size);
 540                 pg_addiu(&buf, A1, A1, 2 * off);
 541                 pg_addiu(&buf, A0, A0, 2 * off);
 542                 off = -off;
 543                 do {
 544                         build_copy_load(&buf, T0, off);
 545                         build_copy_load(&buf, T1, off + copy_word_size);
 546                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 547                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 548                         build_copy_store_pref(&buf, off);
 549                         build_copy_store(&buf, T0, off);
 550                         build_copy_store_pref(&buf, off + copy_word_size);
 551                         build_copy_store(&buf, T1, off + copy_word_size);
 552                         build_copy_store_pref(&buf, off + 2 * copy_word_size);
 553                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 554                         build_copy_store_pref(&buf, off + 3 * copy_word_size);
 555                         if (off == -(4 * copy_word_size))
 556                                 uasm_il_bne(&buf, &r, A2, A0,
 557                                             label_copy_pref_store);
 558                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 559                         off += 4 * copy_word_size;
 560                 } while (off < 0);
 561         }
 562
 563         if (pref_bias_copy_store) {
 564                 pg_addiu(&buf, A2, A0, pref_bias_copy_store);
 565                 uasm_l_copy_nopref(&l, buf);
 566                 off = 0;
 567                 do {
 568                         build_copy_load(&buf, T0, off);
 569                         build_copy_load(&buf, T1, off + copy_word_size);
 570                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 571                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 572                         build_copy_store(&buf, T0, off);
 573                         build_copy_store(&buf, T1, off + copy_word_size);
 574                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 575                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 576                         off += 4 * copy_word_size;
 577                 } while (off < half_copy_loop_size);
 578                 pg_addiu(&buf, A1, A1, 2 * off);
 579                 pg_addiu(&buf, A0, A0, 2 * off);
 580                 off = -off;
 581                 do {
 582                         build_copy_load(&buf, T0, off);
 583                         build_copy_load(&buf, T1, off + copy_word_size);
 584                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 585                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 586                         build_copy_store(&buf, T0, off);
 587                         build_copy_store(&buf, T1, off + copy_word_size);
 588                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 589                         if (off == -(4 * copy_word_size))
 590                                 uasm_il_bne(&buf, &r, A2, A0,
 591                                             label_copy_nopref);
 592                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 593                         off += 4 * copy_word_size;
 594                 } while (off < 0);
 595         }
 596
 597         uasm_i_jr(&buf, RA);
 598         uasm_i_nop(&buf);
 599
 600         BUG_ON(buf > &__copy_page_end);
 601
 602         uasm_resolve_relocs(relocs, labels);
 603
 604         pr_debug("Synthesized copy page handler (%u instructions).\n",
 605                  (u32)(buf - &__copy_page_start));
 606
 607         pr_debug("\t.set push\n");
 608         pr_debug("\t.set noreorder\n");
 609         for (i = 0; i < (buf - &__copy_page_start); i++)
 610                 pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
 611         pr_debug("\t.set pop\n");
 612 }
 613
 614 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
 615 extern void clear_page_cpu(void *page);
 616 extern void copy_page_cpu(void *to, void *from);
 617
 618 /*
 619  * Pad descriptors to cacheline, since each is exclusively owned by a
 620  * particular CPU.
 621  */
 622 struct dmadscr {
 623         u64 dscr_a;
 624         u64 dscr_b;
 625         u64 pad_a;
 626         u64 pad_b;
 627 } ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
 628
 629 void clear_page(void *page)
 630 {
 631         u64 to_phys = CPHYSADDR((unsigned long)page);
 632         unsigned int cpu = smp_processor_id();
 633
 634         /* if the page is not in KSEG0, use old way */
 635         if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
 636                 return clear_page_cpu(page);
 637
 638         page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
 639                                  M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
 640         page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
 641         __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
 642
 643         /*
 644          * Don't really want to do it this way, but there's no
 645          * reliable way to delay completion detection.
 646          */
 647         while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
 648                  & M_DM_DSCR_BASE_INTERRUPT))
 649                 ;
 650         __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
 651 }
 652 EXPORT_SYMBOL(clear_page);
 653
 654 void copy_page(void *to, void *from)
 655 {
 656         u64 from_phys = CPHYSADDR((unsigned long)from);
 657         u64 to_phys = CPHYSADDR((unsigned long)to);
 658         unsigned int cpu = smp_processor_id();
 659
 660         /* if any page is not in KSEG0, use old way */
 661         if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
 662             || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
 663                 return copy_page_cpu(to, from);
 664
 665         page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
 666                                  M_DM_DSCRA_INTERRUPT;
 667         page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
 668         __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
 669
 670         /*
 671          * Don't really want to do it this way, but there's no
 672          * reliable way to delay completion detection.
 673          */
 674         while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
 675                  & M_DM_DSCR_BASE_INTERRUPT))
 676                 ;
 677         __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
 678 }
 679 EXPORT_SYMBOL(copy_page);
 680
 681 #endif /* CONFIG_SIBYTE_DMA_PAGEOPS */