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/module.h>
  16 #include <linux/proc_fs.h>
  17
  18 #include <asm/bugs.h>
  19 #include <asm/cacheops.h>
  20 #include <asm/cpu-type.h>
  21 #include <asm/inst.h>
  22 #include <asm/io.h>
  23 #include <asm/page.h>
  24 #include <asm/pgtable.h>
  25 #include <asm/prefetch.h>
  26 #include <asm/bootinfo.h>
  27 #include <asm/mipsregs.h>
  28 #include <asm/mmu_context.h>
  29 #include <asm/cpu.h>
  30 #include <asm/war.h>
  31
  32 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
  33 #include <asm/sibyte/sb1250.h>
  34 #include <asm/sibyte/sb1250_regs.h>
  35 #include <asm/sibyte/sb1250_dma.h>
  36 #endif
  37
  38 #include <asm/uasm.h>
  39
  40 /* Registers used in the assembled routines. */
  41 #define ZERO 0
  42 #define AT 2
  43 #define A0 4
  44 #define A1 5
  45 #define A2 6
  46 #define T0 8
  47 #define T1 9
  48 #define T2 10
  49 #define T3 11
  50 #define T9 25
  51 #define RA 31
  52
  53 /* Handle labels (which must be positive integers). */
  54 enum label_id {
  55         label_clear_nopref = 1,
  56         label_clear_pref,
  57         label_copy_nopref,
  58         label_copy_pref_both,
  59         label_copy_pref_store,
  60 };
  61
  62 UASM_L_LA(_clear_nopref)
  63 UASM_L_LA(_clear_pref)
  64 UASM_L_LA(_copy_nopref)
  65 UASM_L_LA(_copy_pref_both)
  66 UASM_L_LA(_copy_pref_store)
  67
  68 /* We need one branch and therefore one relocation per target label. */
  69 static struct uasm_label labels[5];
  70 static struct uasm_reloc relocs[5];
  71
  72 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002010)
  73 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002020)
  74
  75 /*
  76  * R6 has a limited offset of the pref instruction.
  77  * Skip it if the offset is more than 9 bits.
  78  */
  79 #define _uasm_i_pref(a, b, c, d)                \
  80 do {                                            \
  81         if (cpu_has_mips_r6) {                  \
  82                 if (c <= 0xff && c >= -0x100)   \
  83                         uasm_i_pref(a, b, c, d);\
  84         } else {                                \
  85                 uasm_i_pref(a, b, c, d);        \
  86         }                                       \
  87 } while(0)
  88
  89 static int pref_bias_clear_store;
  90 static int pref_bias_copy_load;
  91 static int pref_bias_copy_store;
  92
  93 static u32 pref_src_mode;
  94 static u32 pref_dst_mode;
  95
  96 static int clear_word_size;
  97 static int copy_word_size;
  98
  99 static int half_clear_loop_size;
 100 static int half_copy_loop_size;
 101
 102 static int cache_line_size;
 103 #define cache_line_mask() (cache_line_size - 1)
 104
 105 static inline void
 106 pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
 107 {
 108         if (cpu_has_64bit_gp_regs && DADDI_WAR && r4k_daddiu_bug()) {
 109                 if (off > 0x7fff) {
 110                         uasm_i_lui(buf, T9, uasm_rel_hi(off));
 111                         uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
 112                 } else
 113                         uasm_i_addiu(buf, T9, ZERO, off);
 114                 uasm_i_daddu(buf, reg1, reg2, T9);
 115         } else {
 116                 if (off > 0x7fff) {
 117                         uasm_i_lui(buf, T9, uasm_rel_hi(off));
 118                         uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
 119                         UASM_i_ADDU(buf, reg1, reg2, T9);
 120                 } else
 121                         UASM_i_ADDIU(buf, reg1, reg2, off);
 122         }
 123 }
 124
 125 static void set_prefetch_parameters(void)
 126 {
 127         if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
 128                 clear_word_size = 8;
 129         else
 130                 clear_word_size = 4;
 131
 132         if (cpu_has_64bit_gp_regs)
 133                 copy_word_size = 8;
 134         else
 135                 copy_word_size = 4;
 136
 137         /*
 138          * The pref's used here are using "streaming" hints, which cause the
 139          * copied data to be kicked out of the cache sooner.  A page copy often
 140          * ends up copying a lot more data than is commonly used, so this seems
 141          * to make sense in terms of reducing cache pollution, but I've no real
 142          * performance data to back this up.
 143          */
 144         if (cpu_has_prefetch) {
 145                 /*
 146                  * XXX: Most prefetch bias values in here are based on
 147                  * guesswork.
 148                  */
 149                 cache_line_size = cpu_dcache_line_size();
 150                 switch (current_cpu_type()) {
 151                 case CPU_R5500:
 152                 case CPU_TX49XX:
 153                         /* These processors only support the Pref_Load. */
 154                         pref_bias_copy_load = 256;
 155                         break;
 156
 157                 case CPU_R10000:
 158                 case CPU_R12000:
 159                 case CPU_R14000:
 160                 case CPU_R16000:
 161                         /*
 162                          * Those values have been experimentally tuned for an
 163                          * Origin 200.
 164                          */
 165                         pref_bias_clear_store = 512;
 166                         pref_bias_copy_load = 256;
 167                         pref_bias_copy_store = 256;
 168                         pref_src_mode = Pref_LoadStreamed;
 169                         pref_dst_mode = Pref_StoreStreamed;
 170                         break;
 171
 172                 case CPU_SB1:
 173                 case CPU_SB1A:
 174                         pref_bias_clear_store = 128;
 175                         pref_bias_copy_load = 128;
 176                         pref_bias_copy_store = 128;
 177                         /*
 178                          * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
 179                          * hints are broken.
 180                          */
 181                         if (current_cpu_type() == CPU_SB1 &&
 182                             (current_cpu_data.processor_id & 0xff) < 0x02) {
 183                                 pref_src_mode = Pref_Load;
 184                                 pref_dst_mode = Pref_Store;
 185                         } else {
 186                                 pref_src_mode = Pref_LoadStreamed;
 187                                 pref_dst_mode = Pref_StoreStreamed;
 188                         }
 189                         break;
 190
 191                 case CPU_LOONGSON3:
 192                         /* Loongson-3 only support the Pref_Load/Pref_Store. */
 193                         pref_bias_clear_store = 128;
 194                         pref_bias_copy_load = 128;
 195                         pref_bias_copy_store = 128;
 196                         pref_src_mode = Pref_Load;
 197                         pref_dst_mode = Pref_Store;
 198                         break;
 199
 200                 default:
 201                         pref_bias_clear_store = 128;
 202                         pref_bias_copy_load = 256;
 203                         pref_bias_copy_store = 128;
 204                         pref_src_mode = Pref_LoadStreamed;
 205                         if (cpu_has_mips_r6)
 206                                 /*
 207                                  * Bit 30 (Pref_PrepareForStore) has been
 208                                  * removed from MIPS R6. Use bit 5
 209                                  * (Pref_StoreStreamed).
 210                                  */
 211                                 pref_dst_mode = Pref_StoreStreamed;
 212                         else
 213                                 pref_dst_mode = Pref_PrepareForStore;
 214                         break;
 215                 }
 216         } else {
 217                 if (cpu_has_cache_cdex_s)
 218                         cache_line_size = cpu_scache_line_size();
 219                 else if (cpu_has_cache_cdex_p)
 220                         cache_line_size = cpu_dcache_line_size();
 221         }
 222         /*
 223          * Too much unrolling will overflow the available space in
 224          * clear_space_array / copy_page_array.
 225          */
 226         half_clear_loop_size = min(16 * clear_word_size,
 227                                    max(cache_line_size >> 1,
 228                                        4 * clear_word_size));
 229         half_copy_loop_size = min(16 * copy_word_size,
 230                                   max(cache_line_size >> 1,
 231                                       4 * copy_word_size));
 232 }
 233
 234 static void build_clear_store(u32 **buf, int off)
 235 {
 236         if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
 237                 uasm_i_sd(buf, ZERO, off, A0);
 238         } else {
 239                 uasm_i_sw(buf, ZERO, off, A0);
 240         }
 241 }
 242
 243 static inline void build_clear_pref(u32 **buf, int off)
 244 {
 245         if (off & cache_line_mask())
 246                 return;
 247
 248         if (pref_bias_clear_store) {
 249                 _uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
 250                             A0);
 251         } else if (cache_line_size == (half_clear_loop_size << 1)) {
 252                 if (cpu_has_cache_cdex_s) {
 253                         uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
 254                 } else if (cpu_has_cache_cdex_p) {
 255                         if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
 256                                 uasm_i_nop(buf);
 257                                 uasm_i_nop(buf);
 258                                 uasm_i_nop(buf);
 259                                 uasm_i_nop(buf);
 260                         }
 261
 262                         if (R4600_V2_HIT_CACHEOP_WAR && 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 (R4600_V2_HIT_CACHEOP_WAR && 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 (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
 408                                 uasm_i_nop(buf);
 409                                 uasm_i_nop(buf);
 410                                 uasm_i_nop(buf);
 411                                 uasm_i_nop(buf);
 412                         }
 413
 414                         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
 415                                 uasm_i_lw(buf, ZERO, ZERO, AT);
 416
 417                         uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
 418                 }
 419         }
 420 }
 421
 422 void build_copy_page(void)
 423 {
 424         int off;
 425         u32 *buf = &__copy_page_start;
 426         struct uasm_label *l = labels;
 427         struct uasm_reloc *r = relocs;
 428         int i;
 429         static atomic_t run_once = ATOMIC_INIT(0);
 430
 431         if (atomic_xchg(&run_once, 1)) {
 432                 return;
 433         }
 434
 435         memset(labels, 0, sizeof(labels));
 436         memset(relocs, 0, sizeof(relocs));
 437
 438         set_prefetch_parameters();
 439
 440         /*
 441          * This algorithm makes the following assumptions:
 442          *   - All prefetch biases are multiples of 8 words.
 443          *   - The prefetch biases are less than one page.
 444          *   - The store prefetch bias isn't greater than the load
 445          *     prefetch bias.
 446          */
 447         BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
 448         BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
 449         BUG_ON(PAGE_SIZE < pref_bias_copy_load);
 450         BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
 451
 452         off = PAGE_SIZE - pref_bias_copy_load;
 453         if (off > 0xffff || !pref_bias_copy_load)
 454                 pg_addiu(&buf, A2, A0, off);
 455         else
 456                 uasm_i_ori(&buf, A2, A0, off);
 457
 458         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
 459                 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
 460
 461         off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
 462                                 cache_line_size : 0;
 463         while (off) {
 464                 build_copy_load_pref(&buf, -off);
 465                 off -= cache_line_size;
 466         }
 467         off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
 468                                 cache_line_size : 0;
 469         while (off) {
 470                 build_copy_store_pref(&buf, -off);
 471                 off -= cache_line_size;
 472         }
 473         uasm_l_copy_pref_both(&l, buf);
 474         do {
 475                 build_copy_load_pref(&buf, off);
 476                 build_copy_load(&buf, T0, off);
 477                 build_copy_load_pref(&buf, off + copy_word_size);
 478                 build_copy_load(&buf, T1, off + copy_word_size);
 479                 build_copy_load_pref(&buf, off + 2 * copy_word_size);
 480                 build_copy_load(&buf, T2, off + 2 * copy_word_size);
 481                 build_copy_load_pref(&buf, off + 3 * copy_word_size);
 482                 build_copy_load(&buf, T3, off + 3 * copy_word_size);
 483                 build_copy_store_pref(&buf, off);
 484                 build_copy_store(&buf, T0, off);
 485                 build_copy_store_pref(&buf, off + copy_word_size);
 486                 build_copy_store(&buf, T1, off + copy_word_size);
 487                 build_copy_store_pref(&buf, off + 2 * copy_word_size);
 488                 build_copy_store(&buf, T2, off + 2 * copy_word_size);
 489                 build_copy_store_pref(&buf, off + 3 * copy_word_size);
 490                 build_copy_store(&buf, T3, off + 3 * copy_word_size);
 491                 off += 4 * copy_word_size;
 492         } while (off < half_copy_loop_size);
 493         pg_addiu(&buf, A1, A1, 2 * off);
 494         pg_addiu(&buf, A0, A0, 2 * off);
 495         off = -off;
 496         do {
 497                 build_copy_load_pref(&buf, off);
 498                 build_copy_load(&buf, T0, off);
 499                 build_copy_load_pref(&buf, off + copy_word_size);
 500                 build_copy_load(&buf, T1, off + copy_word_size);
 501                 build_copy_load_pref(&buf, off + 2 * copy_word_size);
 502                 build_copy_load(&buf, T2, off + 2 * copy_word_size);
 503                 build_copy_load_pref(&buf, off + 3 * copy_word_size);
 504                 build_copy_load(&buf, T3, off + 3 * copy_word_size);
 505                 build_copy_store_pref(&buf, off);
 506                 build_copy_store(&buf, T0, off);
 507                 build_copy_store_pref(&buf, off + copy_word_size);
 508                 build_copy_store(&buf, T1, off + copy_word_size);
 509                 build_copy_store_pref(&buf, off + 2 * copy_word_size);
 510                 build_copy_store(&buf, T2, off + 2 * copy_word_size);
 511                 build_copy_store_pref(&buf, off + 3 * copy_word_size);
 512                 if (off == -(4 * copy_word_size))
 513                         uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both);
 514                 build_copy_store(&buf, T3, off + 3 * copy_word_size);
 515                 off += 4 * copy_word_size;
 516         } while (off < 0);
 517
 518         if (pref_bias_copy_load - pref_bias_copy_store) {
 519                 pg_addiu(&buf, A2, A0,
 520                          pref_bias_copy_load - pref_bias_copy_store);
 521                 uasm_l_copy_pref_store(&l, buf);
 522                 off = 0;
 523                 do {
 524                         build_copy_load(&buf, T0, off);
 525                         build_copy_load(&buf, T1, off + copy_word_size);
 526                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 527                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 528                         build_copy_store_pref(&buf, off);
 529                         build_copy_store(&buf, T0, off);
 530                         build_copy_store_pref(&buf, off + copy_word_size);
 531                         build_copy_store(&buf, T1, off + copy_word_size);
 532                         build_copy_store_pref(&buf, off + 2 * copy_word_size);
 533                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 534                         build_copy_store_pref(&buf, off + 3 * copy_word_size);
 535                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 536                         off += 4 * copy_word_size;
 537                 } while (off < half_copy_loop_size);
 538                 pg_addiu(&buf, A1, A1, 2 * off);
 539                 pg_addiu(&buf, A0, A0, 2 * off);
 540                 off = -off;
 541                 do {
 542                         build_copy_load(&buf, T0, off);
 543                         build_copy_load(&buf, T1, off + copy_word_size);
 544                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 545                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 546                         build_copy_store_pref(&buf, off);
 547                         build_copy_store(&buf, T0, off);
 548                         build_copy_store_pref(&buf, off + copy_word_size);
 549                         build_copy_store(&buf, T1, off + copy_word_size);
 550                         build_copy_store_pref(&buf, off + 2 * copy_word_size);
 551                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 552                         build_copy_store_pref(&buf, off + 3 * copy_word_size);
 553                         if (off == -(4 * copy_word_size))
 554                                 uasm_il_bne(&buf, &r, A2, A0,
 555                                             label_copy_pref_store);
 556                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 557                         off += 4 * copy_word_size;
 558                 } while (off < 0);
 559         }
 560
 561         if (pref_bias_copy_store) {
 562                 pg_addiu(&buf, A2, A0, pref_bias_copy_store);
 563                 uasm_l_copy_nopref(&l, buf);
 564                 off = 0;
 565                 do {
 566                         build_copy_load(&buf, T0, off);
 567                         build_copy_load(&buf, T1, off + copy_word_size);
 568                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 569                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 570                         build_copy_store(&buf, T0, off);
 571                         build_copy_store(&buf, T1, off + copy_word_size);
 572                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 573                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 574                         off += 4 * copy_word_size;
 575                 } while (off < half_copy_loop_size);
 576                 pg_addiu(&buf, A1, A1, 2 * off);
 577                 pg_addiu(&buf, A0, A0, 2 * off);
 578                 off = -off;
 579                 do {
 580                         build_copy_load(&buf, T0, off);
 581                         build_copy_load(&buf, T1, off + copy_word_size);
 582                         build_copy_load(&buf, T2, off + 2 * copy_word_size);
 583                         build_copy_load(&buf, T3, off + 3 * copy_word_size);
 584                         build_copy_store(&buf, T0, off);
 585                         build_copy_store(&buf, T1, off + copy_word_size);
 586                         build_copy_store(&buf, T2, off + 2 * copy_word_size);
 587                         if (off == -(4 * copy_word_size))
 588                                 uasm_il_bne(&buf, &r, A2, A0,
 589                                             label_copy_nopref);
 590                         build_copy_store(&buf, T3, off + 3 * copy_word_size);
 591                         off += 4 * copy_word_size;
 592                 } while (off < 0);
 593         }
 594
 595         uasm_i_jr(&buf, RA);
 596         uasm_i_nop(&buf);
 597
 598         BUG_ON(buf > &__copy_page_end);
 599
 600         uasm_resolve_relocs(relocs, labels);
 601
 602         pr_debug("Synthesized copy page handler (%u instructions).\n",
 603                  (u32)(buf - &__copy_page_start));
 604
 605         pr_debug("\t.set push\n");
 606         pr_debug("\t.set noreorder\n");
 607         for (i = 0; i < (buf - &__copy_page_start); i++)
 608                 pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
 609         pr_debug("\t.set pop\n");
 610 }
 611
 612 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
 613 extern void clear_page_cpu(void *page);
 614 extern void copy_page_cpu(void *to, void *from);
 615
 616 /*
 617  * Pad descriptors to cacheline, since each is exclusively owned by a
 618  * particular CPU.
 619  */
 620 struct dmadscr {
 621         u64 dscr_a;
 622         u64 dscr_b;
 623         u64 pad_a;
 624         u64 pad_b;
 625 } ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
 626
 627 void sb1_dma_init(void)
 628 {
 629         int i;
 630
 631         for (i = 0; i < DM_NUM_CHANNELS; i++) {
 632                 const u64 base_val = CPHYSADDR((unsigned long)&page_descr[i]) |
 633                                      V_DM_DSCR_BASE_RINGSZ(1);
 634                 void *base_reg = IOADDR(A_DM_REGISTER(i, R_DM_DSCR_BASE));
 635
 636                 __raw_writeq(base_val, base_reg);
 637                 __raw_writeq(base_val | M_DM_DSCR_BASE_RESET, base_reg);
 638                 __raw_writeq(base_val | M_DM_DSCR_BASE_ENABL, base_reg);
 639         }
 640 }
 641
 642 void clear_page(void *page)
 643 {
 644         u64 to_phys = CPHYSADDR((unsigned long)page);
 645         unsigned int cpu = smp_processor_id();
 646
 647         /* if the page is not in KSEG0, use old way */
 648         if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
 649                 return clear_page_cpu(page);
 650
 651         page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
 652                                  M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
 653         page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
 654         __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
 655
 656         /*
 657          * Don't really want to do it this way, but there's no
 658          * reliable way to delay completion detection.
 659          */
 660         while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
 661                  & M_DM_DSCR_BASE_INTERRUPT))
 662                 ;
 663         __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
 664 }
 665
 666 void copy_page(void *to, void *from)
 667 {
 668         u64 from_phys = CPHYSADDR((unsigned long)from);
 669         u64 to_phys = CPHYSADDR((unsigned long)to);
 670         unsigned int cpu = smp_processor_id();
 671
 672         /* if any page is not in KSEG0, use old way */
 673         if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
 674             || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
 675                 return copy_page_cpu(to, from);
 676
 677         page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
 678                                  M_DM_DSCRA_INTERRUPT;
 679         page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
 680         __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
 681
 682         /*
 683          * Don't really want to do it this way, but there's no
 684          * reliable way to delay completion detection.
 685          */
 686         while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
 687                  & M_DM_DSCR_BASE_INTERRUPT))
 688                 ;
 689         __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
 690 }
 691
 692 #endif /* CONFIG_SIBYTE_DMA_PAGEOPS */