| 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 ([email protected]) |
| 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/regdef.h> |
| 28 |
#include <asm/cpu.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 |
/* Handle labels (which must be positive integers). */ |
| 39 |
enum label_id { |
| 40 |
label_clear_nopref = 1, |
| 41 |
label_clear_pref, |
| 42 |
label_copy_nopref, |
| 43 |
label_copy_pref_both, |
| 44 |
label_copy_pref_store, |
| 45 |
}; |
| 46 |
|
| 47 |
UASM_L_LA(_clear_nopref) |
| 48 |
UASM_L_LA(_clear_pref) |
| 49 |
UASM_L_LA(_copy_nopref) |
| 50 |
UASM_L_LA(_copy_pref_both) |
| 51 |
UASM_L_LA(_copy_pref_store) |
| 52 |
|
| 53 |
/* We need one branch and therefore one relocation per target label. */ |
| 54 |
static struct uasm_label labels[5]; |
| 55 |
static struct uasm_reloc relocs[5]; |
| 56 |
|
| 57 |
#define cpu_is_r4600_v1_x() ((read_c0_prid() & 0xfffffff0) == 0x00002010) |
| 58 |
#define cpu_is_r4600_v2_x() ((read_c0_prid() & 0xfffffff0) == 0x00002020) |
| 59 |
|
| 60 |
/* |
| 61 |
* R6 has a limited offset of the pref instruction. |
| 62 |
* Skip it if the offset is more than 9 bits. |
| 63 |
*/ |
| 64 |
#define _uasm_i_pref(a, b, c, d) \ |
| 65 |
do { \ |
| 66 |
if (cpu_has_mips_r6) { \ |
| 67 |
if (c <= 0xff && c >= -0x100) \ |
| 68 |
uasm_i_pref(a, b, c, d);\ |
| 69 |
} else { \ |
| 70 |
uasm_i_pref(a, b, c, d); \ |
| 71 |
} \ |
| 72 |
} while(0) |
| 73 |
|
| 74 |
static int pref_bias_clear_store; |
| 75 |
static int pref_bias_copy_load; |
| 76 |
static int pref_bias_copy_store; |
| 77 |
|
| 78 |
static u32 pref_src_mode; |
| 79 |
static u32 pref_dst_mode; |
| 80 |
|
| 81 |
static int clear_word_size; |
| 82 |
static int copy_word_size; |
| 83 |
|
| 84 |
static int half_clear_loop_size; |
| 85 |
static int half_copy_loop_size; |
| 86 |
|
| 87 |
static int cache_line_size; |
| 88 |
#define cache_line_mask() (cache_line_size - 1) |
| 89 |
|
| 90 |
static inline void |
| 91 |
pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off) |
| 92 |
{ |
| 93 |
if (cpu_has_64bit_gp_regs && |
| 94 |
IS_ENABLED(CONFIG_CPU_DADDI_WORKAROUNDS) && |
| 95 |
r4k_daddiu_bug()) { |
| 96 |
if (off > 0x7fff) { |
| 97 |
uasm_i_lui(buf, GPR_T9, uasm_rel_hi(off)); |
| 98 |
uasm_i_addiu(buf, GPR_T9, GPR_T9, uasm_rel_lo(off)); |
| 99 |
} else |
| 100 |
uasm_i_addiu(buf, GPR_T9, GPR_ZERO, off); |
| 101 |
uasm_i_daddu(buf, reg1, reg2, GPR_T9); |
| 102 |
} else { |
| 103 |
if (off > 0x7fff) { |
| 104 |
uasm_i_lui(buf, GPR_T9, uasm_rel_hi(off)); |
| 105 |
uasm_i_addiu(buf, GPR_T9, GPR_T9, uasm_rel_lo(off)); |
| 106 |
UASM_i_ADDU(buf, reg1, reg2, GPR_T9); |
| 107 |
} else |
| 108 |
UASM_i_ADDIU(buf, reg1, reg2, off); |
| 109 |
} |
| 110 |
} |
| 111 |
|
| 112 |
static void set_prefetch_parameters(void) |
| 113 |
{ |
| 114 |
if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) |
| 115 |
clear_word_size = 8; |
| 116 |
else |
| 117 |
clear_word_size = 4; |
| 118 |
|
| 119 |
if (cpu_has_64bit_gp_regs) |
| 120 |
copy_word_size = 8; |
| 121 |
else |
| 122 |
copy_word_size = 4; |
| 123 |
|
| 124 |
/* |
| 125 |
* The pref's used here are using "streaming" hints, which cause the |
| 126 |
* copied data to be kicked out of the cache sooner. A page copy often |
| 127 |
* ends up copying a lot more data than is commonly used, so this seems |
| 128 |
* to make sense in terms of reducing cache pollution, but I've no real |
| 129 |
* performance data to back this up. |
| 130 |
*/ |
| 131 |
if (cpu_has_prefetch) { |
| 132 |
/* |
| 133 |
* XXX: Most prefetch bias values in here are based on |
| 134 |
* guesswork. |
| 135 |
*/ |
| 136 |
cache_line_size = cpu_dcache_line_size(); |
| 137 |
switch (current_cpu_type()) { |
| 138 |
case CPU_R5500: |
| 139 |
case CPU_TX49XX: |
| 140 |
/* These processors only support the Pref_Load. */ |
| 141 |
pref_bias_copy_load = 256; |
| 142 |
break; |
| 143 |
|
| 144 |
case CPU_R10000: |
| 145 |
case CPU_R12000: |
| 146 |
case CPU_R14000: |
| 147 |
case CPU_R16000: |
| 148 |
/* |
| 149 |
* Those values have been experimentally tuned for an |
| 150 |
* Origin 200. |
| 151 |
*/ |
| 152 |
pref_bias_clear_store = 512; |
| 153 |
pref_bias_copy_load = 256; |
| 154 |
pref_bias_copy_store = 256; |
| 155 |
pref_src_mode = Pref_LoadStreamed; |
| 156 |
pref_dst_mode = Pref_StoreStreamed; |
| 157 |
break; |
| 158 |
|
| 159 |
case CPU_SB1: |
| 160 |
case CPU_SB1A: |
| 161 |
pref_bias_clear_store = 128; |
| 162 |
pref_bias_copy_load = 128; |
| 163 |
pref_bias_copy_store = 128; |
| 164 |
/* |
| 165 |
* SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed |
| 166 |
* hints are broken. |
| 167 |
*/ |
| 168 |
if (current_cpu_type() == CPU_SB1 && |
| 169 |
(current_cpu_data.processor_id & 0xff) < 0x02) { |
| 170 |
pref_src_mode = Pref_Load; |
| 171 |
pref_dst_mode = Pref_Store; |
| 172 |
} else { |
| 173 |
pref_src_mode = Pref_LoadStreamed; |
| 174 |
pref_dst_mode = Pref_StoreStreamed; |
| 175 |
} |
| 176 |
break; |
| 177 |
|
| 178 |
case CPU_LOONGSON64: |
| 179 |
/* Loongson-3 only support the Pref_Load/Pref_Store. */ |
| 180 |
pref_bias_clear_store = 128; |
| 181 |
pref_bias_copy_load = 128; |
| 182 |
pref_bias_copy_store = 128; |
| 183 |
pref_src_mode = Pref_Load; |
| 184 |
pref_dst_mode = Pref_Store; |
| 185 |
break; |
| 186 |
|
| 187 |
default: |
| 188 |
pref_bias_clear_store = 128; |
| 189 |
pref_bias_copy_load = 256; |
| 190 |
pref_bias_copy_store = 128; |
| 191 |
pref_src_mode = Pref_LoadStreamed; |
| 192 |
if (cpu_has_mips_r6) |
| 193 |
/* |
| 194 |
* Bit 30 (Pref_PrepareForStore) has been |
| 195 |
* removed from MIPS R6. Use bit 5 |
| 196 |
* (Pref_StoreStreamed). |
| 197 |
*/ |
| 198 |
pref_dst_mode = Pref_StoreStreamed; |
| 199 |
else |
| 200 |
pref_dst_mode = Pref_PrepareForStore; |
| 201 |
break; |
| 202 |
} |
| 203 |
} else { |
| 204 |
if (cpu_has_cache_cdex_s) |
| 205 |
cache_line_size = cpu_scache_line_size(); |
| 206 |
else if (cpu_has_cache_cdex_p) |
| 207 |
cache_line_size = cpu_dcache_line_size(); |
| 208 |
} |
| 209 |
/* |
| 210 |
* Too much unrolling will overflow the available space in |
| 211 |
* clear_space_array / copy_page_array. |
| 212 |
*/ |
| 213 |
half_clear_loop_size = min(16 * clear_word_size, |
| 214 |
max(cache_line_size >> 1, |
| 215 |
4 * clear_word_size)); |
| 216 |
half_copy_loop_size = min(16 * copy_word_size, |
| 217 |
max(cache_line_size >> 1, |
| 218 |
4 * copy_word_size)); |
| 219 |
} |
| 220 |
|
| 221 |
static void build_clear_store(u32 **buf, int off) |
| 222 |
{ |
| 223 |
if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) { |
| 224 |
uasm_i_sd(buf, GPR_ZERO, off, GPR_A0); |
| 225 |
} else { |
| 226 |
uasm_i_sw(buf, GPR_ZERO, off, GPR_A0); |
| 227 |
} |
| 228 |
} |
| 229 |
|
| 230 |
static inline void build_clear_pref(u32 **buf, int off) |
| 231 |
{ |
| 232 |
if (off & cache_line_mask()) |
| 233 |
return; |
| 234 |
|
| 235 |
if (pref_bias_clear_store) { |
| 236 |
_uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off, |
| 237 |
GPR_A0); |
| 238 |
} else if (cache_line_size == (half_clear_loop_size << 1)) { |
| 239 |
if (cpu_has_cache_cdex_s) { |
| 240 |
uasm_i_cache(buf, Create_Dirty_Excl_SD, off, GPR_A0); |
| 241 |
} else if (cpu_has_cache_cdex_p) { |
| 242 |
if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) && |
| 243 |
cpu_is_r4600_v1_x()) { |
| 244 |
uasm_i_nop(buf); |
| 245 |
uasm_i_nop(buf); |
| 246 |
uasm_i_nop(buf); |
| 247 |
uasm_i_nop(buf); |
| 248 |
} |
| 249 |
|
| 250 |
if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && |
| 251 |
cpu_is_r4600_v2_x()) |
| 252 |
uasm_i_lw(buf, GPR_ZERO, GPR_ZERO, GPR_AT); |
| 253 |
|
| 254 |
uasm_i_cache(buf, Create_Dirty_Excl_D, off, GPR_A0); |
| 255 |
} |
| 256 |
} |
| 257 |
} |
| 258 |
|
| 259 |
extern u32 __clear_page_start; |
| 260 |
extern u32 __clear_page_end; |
| 261 |
extern u32 __copy_page_start; |
| 262 |
extern u32 __copy_page_end; |
| 263 |
|
| 264 |
void build_clear_page(void) |
| 265 |
{ |
| 266 |
int off; |
| 267 |
u32 *buf = &__clear_page_start; |
| 268 |
struct uasm_label *l = labels; |
| 269 |
struct uasm_reloc *r = relocs; |
| 270 |
int i; |
| 271 |
static atomic_t run_once = ATOMIC_INIT(0); |
| 272 |
|
| 273 |
if (atomic_xchg(&run_once, 1)) { |
| 274 |
return; |
| 275 |
} |
| 276 |
|
| 277 |
memset(labels, 0, sizeof(labels)); |
| 278 |
memset(relocs, 0, sizeof(relocs)); |
| 279 |
|
| 280 |
set_prefetch_parameters(); |
| 281 |
|
| 282 |
/* |
| 283 |
* This algorithm makes the following assumptions: |
| 284 |
* - The prefetch bias is a multiple of 2 words. |
| 285 |
* - The prefetch bias is less than one page. |
| 286 |
*/ |
| 287 |
BUG_ON(pref_bias_clear_store % (2 * clear_word_size)); |
| 288 |
BUG_ON(PAGE_SIZE < pref_bias_clear_store); |
| 289 |
|
| 290 |
off = PAGE_SIZE - pref_bias_clear_store; |
| 291 |
if (off > 0xffff || !pref_bias_clear_store) |
| 292 |
pg_addiu(&buf, GPR_A2, GPR_A0, off); |
| 293 |
else |
| 294 |
uasm_i_ori(&buf, GPR_A2, GPR_A0, off); |
| 295 |
|
| 296 |
if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x()) |
| 297 |
uasm_i_lui(&buf, GPR_AT, uasm_rel_hi(0xa0000000)); |
| 298 |
|
| 299 |
off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size) |
| 300 |
* cache_line_size : 0; |
| 301 |
while (off) { |
| 302 |
build_clear_pref(&buf, -off); |
| 303 |
off -= cache_line_size; |
| 304 |
} |
| 305 |
uasm_l_clear_pref(&l, buf); |
| 306 |
do { |
| 307 |
build_clear_pref(&buf, off); |
| 308 |
build_clear_store(&buf, off); |
| 309 |
off += clear_word_size; |
| 310 |
} while (off < half_clear_loop_size); |
| 311 |
pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off); |
| 312 |
off = -off; |
| 313 |
do { |
| 314 |
build_clear_pref(&buf, off); |
| 315 |
if (off == -clear_word_size) |
| 316 |
uasm_il_bne(&buf, &r, GPR_A0, GPR_A2, label_clear_pref); |
| 317 |
build_clear_store(&buf, off); |
| 318 |
off += clear_word_size; |
| 319 |
} while (off < 0); |
| 320 |
|
| 321 |
if (pref_bias_clear_store) { |
| 322 |
pg_addiu(&buf, GPR_A2, GPR_A0, pref_bias_clear_store); |
| 323 |
uasm_l_clear_nopref(&l, buf); |
| 324 |
off = 0; |
| 325 |
do { |
| 326 |
build_clear_store(&buf, off); |
| 327 |
off += clear_word_size; |
| 328 |
} while (off < half_clear_loop_size); |
| 329 |
pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off); |
| 330 |
off = -off; |
| 331 |
do { |
| 332 |
if (off == -clear_word_size) |
| 333 |
uasm_il_bne(&buf, &r, GPR_A0, GPR_A2, |
| 334 |
label_clear_nopref); |
| 335 |
build_clear_store(&buf, off); |
| 336 |
off += clear_word_size; |
| 337 |
} while (off < 0); |
| 338 |
} |
| 339 |
|
| 340 |
uasm_i_jr(&buf, GPR_RA); |
| 341 |
uasm_i_nop(&buf); |
| 342 |
|
| 343 |
BUG_ON(buf > &__clear_page_end); |
| 344 |
|
| 345 |
uasm_resolve_relocs(relocs, labels); |
| 346 |
|
| 347 |
pr_debug("Synthesized clear page handler (%u instructions).\n", |
| 348 |
(u32)(buf - &__clear_page_start)); |
| 349 |
|
| 350 |
pr_debug("\t.set push\n"); |
| 351 |
pr_debug("\t.set noreorder\n"); |
| 352 |
for (i = 0; i < (buf - &__clear_page_start); i++) |
| 353 |
pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]); |
| 354 |
pr_debug("\t.set pop\n"); |
| 355 |
} |
| 356 |
|
| 357 |
static void build_copy_load(u32 **buf, int reg, int off) |
| 358 |
{ |
| 359 |
if (cpu_has_64bit_gp_regs) { |
| 360 |
uasm_i_ld(buf, reg, off, GPR_A1); |
| 361 |
} else { |
| 362 |
uasm_i_lw(buf, reg, off, GPR_A1); |
| 363 |
} |
| 364 |
} |
| 365 |
|
| 366 |
static void build_copy_store(u32 **buf, int reg, int off) |
| 367 |
{ |
| 368 |
if (cpu_has_64bit_gp_regs) { |
| 369 |
uasm_i_sd(buf, reg, off, GPR_A0); |
| 370 |
} else { |
| 371 |
uasm_i_sw(buf, reg, off, GPR_A0); |
| 372 |
} |
| 373 |
} |
| 374 |
|
| 375 |
static inline void build_copy_load_pref(u32 **buf, int off) |
| 376 |
{ |
| 377 |
if (off & cache_line_mask()) |
| 378 |
return; |
| 379 |
|
| 380 |
if (pref_bias_copy_load) |
| 381 |
_uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, GPR_A1); |
| 382 |
} |
| 383 |
|
| 384 |
static inline void build_copy_store_pref(u32 **buf, int off) |
| 385 |
{ |
| 386 |
if (off & cache_line_mask()) |
| 387 |
return; |
| 388 |
|
| 389 |
if (pref_bias_copy_store) { |
| 390 |
_uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off, |
| 391 |
GPR_A0); |
| 392 |
} else if (cache_line_size == (half_copy_loop_size << 1)) { |
| 393 |
if (cpu_has_cache_cdex_s) { |
| 394 |
uasm_i_cache(buf, Create_Dirty_Excl_SD, off, GPR_A0); |
| 395 |
} else if (cpu_has_cache_cdex_p) { |
| 396 |
if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) && |
| 397 |
cpu_is_r4600_v1_x()) { |
| 398 |
uasm_i_nop(buf); |
| 399 |
uasm_i_nop(buf); |
| 400 |
uasm_i_nop(buf); |
| 401 |
uasm_i_nop(buf); |
| 402 |
} |
| 403 |
|
| 404 |
if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && |
| 405 |
cpu_is_r4600_v2_x()) |
| 406 |
uasm_i_lw(buf, GPR_ZERO, GPR_ZERO, GPR_AT); |
| 407 |
|
| 408 |
uasm_i_cache(buf, Create_Dirty_Excl_D, off, GPR_A0); |
| 409 |
} |
| 410 |
} |
| 411 |
} |
| 412 |
|
| 413 |
void build_copy_page(void) |
| 414 |
{ |
| 415 |
int off; |
| 416 |
u32 *buf = &__copy_page_start; |
| 417 |
struct uasm_label *l = labels; |
| 418 |
struct uasm_reloc *r = relocs; |
| 419 |
int i; |
| 420 |
static atomic_t run_once = ATOMIC_INIT(0); |
| 421 |
|
| 422 |
if (atomic_xchg(&run_once, 1)) { |
| 423 |
return; |
| 424 |
} |
| 425 |
|
| 426 |
memset(labels, 0, sizeof(labels)); |
| 427 |
memset(relocs, 0, sizeof(relocs)); |
| 428 |
|
| 429 |
set_prefetch_parameters(); |
| 430 |
|
| 431 |
/* |
| 432 |
* This algorithm makes the following assumptions: |
| 433 |
* - All prefetch biases are multiples of 8 words. |
| 434 |
* - The prefetch biases are less than one page. |
| 435 |
* - The store prefetch bias isn't greater than the load |
| 436 |
* prefetch bias. |
| 437 |
*/ |
| 438 |
BUG_ON(pref_bias_copy_load % (8 * copy_word_size)); |
| 439 |
BUG_ON(pref_bias_copy_store % (8 * copy_word_size)); |
| 440 |
BUG_ON(PAGE_SIZE < pref_bias_copy_load); |
| 441 |
BUG_ON(pref_bias_copy_store > pref_bias_copy_load); |
| 442 |
|
| 443 |
off = PAGE_SIZE - pref_bias_copy_load; |
| 444 |
if (off > 0xffff || !pref_bias_copy_load) |
| 445 |
pg_addiu(&buf, GPR_A2, GPR_A0, off); |
| 446 |
else |
| 447 |
uasm_i_ori(&buf, GPR_A2, GPR_A0, off); |
| 448 |
|
| 449 |
if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x()) |
| 450 |
uasm_i_lui(&buf, GPR_AT, uasm_rel_hi(0xa0000000)); |
| 451 |
|
| 452 |
off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) * |
| 453 |
cache_line_size : 0; |
| 454 |
while (off) { |
| 455 |
build_copy_load_pref(&buf, -off); |
| 456 |
off -= cache_line_size; |
| 457 |
} |
| 458 |
off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) * |
| 459 |
cache_line_size : 0; |
| 460 |
while (off) { |
| 461 |
build_copy_store_pref(&buf, -off); |
| 462 |
off -= cache_line_size; |
| 463 |
} |
| 464 |
uasm_l_copy_pref_both(&l, buf); |
| 465 |
do { |
| 466 |
build_copy_load_pref(&buf, off); |
| 467 |
build_copy_load(&buf, GPR_T0, off); |
| 468 |
build_copy_load_pref(&buf, off + copy_word_size); |
| 469 |
build_copy_load(&buf, GPR_T1, off + copy_word_size); |
| 470 |
build_copy_load_pref(&buf, off + 2 * copy_word_size); |
| 471 |
build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size); |
| 472 |
build_copy_load_pref(&buf, off + 3 * copy_word_size); |
| 473 |
build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size); |
| 474 |
build_copy_store_pref(&buf, off); |
| 475 |
build_copy_store(&buf, GPR_T0, off); |
| 476 |
build_copy_store_pref(&buf, off + copy_word_size); |
| 477 |
build_copy_store(&buf, GPR_T1, off + copy_word_size); |
| 478 |
build_copy_store_pref(&buf, off + 2 * copy_word_size); |
| 479 |
build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size); |
| 480 |
build_copy_store_pref(&buf, off + 3 * copy_word_size); |
| 481 |
build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size); |
| 482 |
off += 4 * copy_word_size; |
| 483 |
} while (off < half_copy_loop_size); |
| 484 |
pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off); |
| 485 |
pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off); |
| 486 |
off = -off; |
| 487 |
do { |
| 488 |
build_copy_load_pref(&buf, off); |
| 489 |
build_copy_load(&buf, GPR_T0, off); |
| 490 |
build_copy_load_pref(&buf, off + copy_word_size); |
| 491 |
build_copy_load(&buf, GPR_T1, off + copy_word_size); |
| 492 |
build_copy_load_pref(&buf, off + 2 * copy_word_size); |
| 493 |
build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size); |
| 494 |
build_copy_load_pref(&buf, off + 3 * copy_word_size); |
| 495 |
build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size); |
| 496 |
build_copy_store_pref(&buf, off); |
| 497 |
build_copy_store(&buf, GPR_T0, off); |
| 498 |
build_copy_store_pref(&buf, off + copy_word_size); |
| 499 |
build_copy_store(&buf, GPR_T1, off + copy_word_size); |
| 500 |
build_copy_store_pref(&buf, off + 2 * copy_word_size); |
| 501 |
build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size); |
| 502 |
build_copy_store_pref(&buf, off + 3 * copy_word_size); |
| 503 |
if (off == -(4 * copy_word_size)) |
| 504 |
uasm_il_bne(&buf, &r, GPR_A2, GPR_A0, label_copy_pref_both); |
| 505 |
build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size); |
| 506 |
off += 4 * copy_word_size; |
| 507 |
} while (off < 0); |
| 508 |
|
| 509 |
if (pref_bias_copy_load - pref_bias_copy_store) { |
| 510 |
pg_addiu(&buf, GPR_A2, GPR_A0, |
| 511 |
pref_bias_copy_load - pref_bias_copy_store); |
| 512 |
uasm_l_copy_pref_store(&l, buf); |
| 513 |
off = 0; |
| 514 |
do { |
| 515 |
build_copy_load(&buf, GPR_T0, off); |
| 516 |
build_copy_load(&buf, GPR_T1, off + copy_word_size); |
| 517 |
build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size); |
| 518 |
build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size); |
| 519 |
build_copy_store_pref(&buf, off); |
| 520 |
build_copy_store(&buf, GPR_T0, off); |
| 521 |
build_copy_store_pref(&buf, off + copy_word_size); |
| 522 |
build_copy_store(&buf, GPR_T1, off + copy_word_size); |
| 523 |
build_copy_store_pref(&buf, off + 2 * copy_word_size); |
| 524 |
build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size); |
| 525 |
build_copy_store_pref(&buf, off + 3 * copy_word_size); |
| 526 |
build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size); |
| 527 |
off += 4 * copy_word_size; |
| 528 |
} while (off < half_copy_loop_size); |
| 529 |
pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off); |
| 530 |
pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off); |
| 531 |
off = -off; |
| 532 |
do { |
| 533 |
build_copy_load(&buf, GPR_T0, off); |
| 534 |
build_copy_load(&buf, GPR_T1, off + copy_word_size); |
| 535 |
build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size); |
| 536 |
build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size); |
| 537 |
build_copy_store_pref(&buf, off); |
| 538 |
build_copy_store(&buf, GPR_T0, off); |
| 539 |
build_copy_store_pref(&buf, off + copy_word_size); |
| 540 |
build_copy_store(&buf, GPR_T1, off + copy_word_size); |
| 541 |
build_copy_store_pref(&buf, off + 2 * copy_word_size); |
| 542 |
build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size); |
| 543 |
build_copy_store_pref(&buf, off + 3 * copy_word_size); |
| 544 |
if (off == -(4 * copy_word_size)) |
| 545 |
uasm_il_bne(&buf, &r, GPR_A2, GPR_A0, |
| 546 |
label_copy_pref_store); |
| 547 |
build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size); |
| 548 |
off += 4 * copy_word_size; |
| 549 |
} while (off < 0); |
| 550 |
} |
| 551 |
|
| 552 |
if (pref_bias_copy_store) { |
| 553 |
pg_addiu(&buf, GPR_A2, GPR_A0, pref_bias_copy_store); |
| 554 |
uasm_l_copy_nopref(&l, buf); |
| 555 |
off = 0; |
| 556 |
do { |
| 557 |
build_copy_load(&buf, GPR_T0, off); |
| 558 |
build_copy_load(&buf, GPR_T1, off + copy_word_size); |
| 559 |
build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size); |
| 560 |
build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size); |
| 561 |
build_copy_store(&buf, GPR_T0, off); |
| 562 |
build_copy_store(&buf, GPR_T1, off + copy_word_size); |
| 563 |
build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size); |
| 564 |
build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size); |
| 565 |
off += 4 * copy_word_size; |
| 566 |
} while (off < half_copy_loop_size); |
| 567 |
pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off); |
| 568 |
pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off); |
| 569 |
off = -off; |
| 570 |
do { |
| 571 |
build_copy_load(&buf, GPR_T0, off); |
| 572 |
build_copy_load(&buf, GPR_T1, off + copy_word_size); |
| 573 |
build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size); |
| 574 |
build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size); |
| 575 |
build_copy_store(&buf, GPR_T0, off); |
| 576 |
build_copy_store(&buf, GPR_T1, off + copy_word_size); |
| 577 |
build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size); |
| 578 |
if (off == -(4 * copy_word_size)) |
| 579 |
uasm_il_bne(&buf, &r, GPR_A2, GPR_A0, |
| 580 |
label_copy_nopref); |
| 581 |
build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size); |
| 582 |
off += 4 * copy_word_size; |
| 583 |
} while (off < 0); |
| 584 |
} |
| 585 |
|
| 586 |
uasm_i_jr(&buf, GPR_RA); |
| 587 |
uasm_i_nop(&buf); |
| 588 |
|
| 589 |
BUG_ON(buf > &__copy_page_end); |
| 590 |
|
| 591 |
uasm_resolve_relocs(relocs, labels); |
| 592 |
|
| 593 |
pr_debug("Synthesized copy page handler (%u instructions).\n", |
| 594 |
(u32)(buf - &__copy_page_start)); |
| 595 |
|
| 596 |
pr_debug("\t.set push\n"); |
| 597 |
pr_debug("\t.set noreorder\n"); |
| 598 |
for (i = 0; i < (buf - &__copy_page_start); i++) |
| 599 |
pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]); |
| 600 |
pr_debug("\t.set pop\n"); |
| 601 |
} |
| 602 |
|
| 603 |
#ifdef CONFIG_SIBYTE_DMA_PAGEOPS |
| 604 |
extern void clear_page_cpu(void *page); |
| 605 |
extern void copy_page_cpu(void *to, void *from); |
| 606 |
|
| 607 |
/* |
| 608 |
* Pad descriptors to cacheline, since each is exclusively owned by a |
| 609 |
* particular CPU. |
| 610 |
*/ |
| 611 |
struct dmadscr { |
| 612 |
u64 dscr_a; |
| 613 |
u64 dscr_b; |
| 614 |
u64 pad_a; |
| 615 |
u64 pad_b; |
| 616 |
} ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS]; |
| 617 |
|
| 618 |
void clear_page(void *page) |
| 619 |
{ |
| 620 |
u64 to_phys = CPHYSADDR((unsigned long)page); |
| 621 |
unsigned int cpu = smp_processor_id(); |
| 622 |
|
| 623 |
/* if the page is not in KSEG0, use old way */ |
| 624 |
if ((long)KSEGX((unsigned long)page) != (long)CKSEG0) |
| 625 |
return clear_page_cpu(page); |
| 626 |
|
| 627 |
page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM | |
| 628 |
M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT; |
| 629 |
page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE); |
| 630 |
__raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT))); |
| 631 |
|
| 632 |
/* |
| 633 |
* Don't really want to do it this way, but there's no |
| 634 |
* reliable way to delay completion detection. |
| 635 |
*/ |
| 636 |
while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG))) |
| 637 |
& M_DM_DSCR_BASE_INTERRUPT)) |
| 638 |
; |
| 639 |
__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE))); |
| 640 |
} |
| 641 |
EXPORT_SYMBOL(clear_page); |
| 642 |
|
| 643 |
void copy_page(void *to, void *from) |
| 644 |
{ |
| 645 |
u64 from_phys = CPHYSADDR((unsigned long)from); |
| 646 |
u64 to_phys = CPHYSADDR((unsigned long)to); |
| 647 |
unsigned int cpu = smp_processor_id(); |
| 648 |
|
| 649 |
/* if any page is not in KSEG0, use old way */ |
| 650 |
if ((long)KSEGX((unsigned long)to) != (long)CKSEG0 |
| 651 |
|| (long)KSEGX((unsigned long)from) != (long)CKSEG0) |
| 652 |
return copy_page_cpu(to, from); |
| 653 |
|
| 654 |
page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST | |
| 655 |
M_DM_DSCRA_INTERRUPT; |
| 656 |
page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE); |
| 657 |
__raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT))); |
| 658 |
|
| 659 |
/* |
| 660 |
* Don't really want to do it this way, but there's no |
| 661 |
* reliable way to delay completion detection. |
| 662 |
*/ |
| 663 |
while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG))) |
| 664 |
& M_DM_DSCR_BASE_INTERRUPT)) |
| 665 |
; |
| 666 |
__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE))); |
| 667 |
} |
| 668 |
EXPORT_SYMBOL(copy_page); |
| 669 |
|
| 670 |
#endif /* CONFIG_SIBYTE_DMA_PAGEOPS */ |