1 /***********************license start*************** 2 * Author: Cavium Networks 3 * 4 * Contact: support@caviumnetworks.com 5 * This file is part of the OCTEON SDK 6 * 7 * Copyright (c) 2003-2010 Cavium Networks 8 * 9 * This file is free software; you can redistribute it and/or modify 10 * it under the terms of the GNU General Public License, Version 2, as 11 * published by the Free Software Foundation. 12 * 13 * This file is distributed in the hope that it will be useful, but 14 * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty 15 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or 16 * NONINFRINGEMENT. See the GNU General Public License for more 17 * details. 18 * 19 * You should have received a copy of the GNU General Public License 20 * along with this file; if not, write to the Free Software 21 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA 22 * or visit http://www.gnu.org/licenses/. 23 * 24 * This file may also be available under a different license from Cavium. 25 * Contact Cavium Networks for more information 26 ***********************license end**************************************/ 27 28 /* 29 * Implementation of the Level 2 Cache (L2C) control, 30 * measurement, and debugging facilities. 31 */ 32 33 #include <linux/compiler.h> 34 #include <linux/irqflags.h> 35 #include <asm/octeon/cvmx.h> 36 #include <asm/octeon/cvmx-l2c.h> 37 #include <asm/octeon/cvmx-spinlock.h> 38 39 /* 40 * This spinlock is used internally to ensure that only one core is 41 * performing certain L2 operations at a time. 42 * 43 * NOTE: This only protects calls from within a single application - 44 * if multiple applications or operating systems are running, then it 45 * is up to the user program to coordinate between them. 46 */ 47 cvmx_spinlock_t cvmx_l2c_spinlock; 48 49 int cvmx_l2c_get_core_way_partition(uint32_t core) 50 { 51 uint32_t field; 52 53 /* Validate the core number */ 54 if (core >= cvmx_octeon_num_cores()) 55 return -1; 56 57 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) 58 return cvmx_read_csr(CVMX_L2C_WPAR_PPX(core)) & 0xffff; 59 60 /* 61 * Use the lower two bits of the coreNumber to determine the 62 * bit offset of the UMSK[] field in the L2C_SPAR register. 63 */ 64 field = (core & 0x3) * 8; 65 66 /* 67 * Return the UMSK[] field from the appropriate L2C_SPAR 68 * register based on the coreNumber. 69 */ 70 71 switch (core & 0xC) { 72 case 0x0: 73 return (cvmx_read_csr(CVMX_L2C_SPAR0) & (0xFF << field)) >> field; 74 case 0x4: 75 return (cvmx_read_csr(CVMX_L2C_SPAR1) & (0xFF << field)) >> field; 76 case 0x8: 77 return (cvmx_read_csr(CVMX_L2C_SPAR2) & (0xFF << field)) >> field; 78 case 0xC: 79 return (cvmx_read_csr(CVMX_L2C_SPAR3) & (0xFF << field)) >> field; 80 } 81 return 0; 82 } 83 84 int cvmx_l2c_set_core_way_partition(uint32_t core, uint32_t mask) 85 { 86 uint32_t field; 87 uint32_t valid_mask; 88 89 valid_mask = (0x1 << cvmx_l2c_get_num_assoc()) - 1; 90 91 mask &= valid_mask; 92 93 /* A UMSK setting which blocks all L2C Ways is an error on some chips */ 94 if (mask == valid_mask && !OCTEON_IS_MODEL(OCTEON_CN63XX)) 95 return -1; 96 97 /* Validate the core number */ 98 if (core >= cvmx_octeon_num_cores()) 99 return -1; 100 101 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 102 cvmx_write_csr(CVMX_L2C_WPAR_PPX(core), mask); 103 return 0; 104 } 105 106 /* 107 * Use the lower two bits of core to determine the bit offset of the 108 * UMSK[] field in the L2C_SPAR register. 109 */ 110 field = (core & 0x3) * 8; 111 112 /* 113 * Assign the new mask setting to the UMSK[] field in the appropriate 114 * L2C_SPAR register based on the core_num. 115 * 116 */ 117 switch (core & 0xC) { 118 case 0x0: 119 cvmx_write_csr(CVMX_L2C_SPAR0, 120 (cvmx_read_csr(CVMX_L2C_SPAR0) & ~(0xFF << field)) | 121 mask << field); 122 break; 123 case 0x4: 124 cvmx_write_csr(CVMX_L2C_SPAR1, 125 (cvmx_read_csr(CVMX_L2C_SPAR1) & ~(0xFF << field)) | 126 mask << field); 127 break; 128 case 0x8: 129 cvmx_write_csr(CVMX_L2C_SPAR2, 130 (cvmx_read_csr(CVMX_L2C_SPAR2) & ~(0xFF << field)) | 131 mask << field); 132 break; 133 case 0xC: 134 cvmx_write_csr(CVMX_L2C_SPAR3, 135 (cvmx_read_csr(CVMX_L2C_SPAR3) & ~(0xFF << field)) | 136 mask << field); 137 break; 138 } 139 return 0; 140 } 141 142 int cvmx_l2c_set_hw_way_partition(uint32_t mask) 143 { 144 uint32_t valid_mask; 145 146 valid_mask = (0x1 << cvmx_l2c_get_num_assoc()) - 1; 147 mask &= valid_mask; 148 149 /* A UMSK setting which blocks all L2C Ways is an error on some chips */ 150 if (mask == valid_mask && !OCTEON_IS_MODEL(OCTEON_CN63XX)) 151 return -1; 152 153 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) 154 cvmx_write_csr(CVMX_L2C_WPAR_IOBX(0), mask); 155 else 156 cvmx_write_csr(CVMX_L2C_SPAR4, 157 (cvmx_read_csr(CVMX_L2C_SPAR4) & ~0xFF) | mask); 158 return 0; 159 } 160 161 int cvmx_l2c_get_hw_way_partition(void) 162 { 163 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) 164 return cvmx_read_csr(CVMX_L2C_WPAR_IOBX(0)) & 0xffff; 165 else 166 return cvmx_read_csr(CVMX_L2C_SPAR4) & (0xFF); 167 } 168 169 void cvmx_l2c_config_perf(uint32_t counter, enum cvmx_l2c_event event, 170 uint32_t clear_on_read) 171 { 172 if (OCTEON_IS_MODEL(OCTEON_CN5XXX) || OCTEON_IS_MODEL(OCTEON_CN3XXX)) { 173 union cvmx_l2c_pfctl pfctl; 174 175 pfctl.u64 = cvmx_read_csr(CVMX_L2C_PFCTL); 176 177 switch (counter) { 178 case 0: 179 pfctl.s.cnt0sel = event; 180 pfctl.s.cnt0ena = 1; 181 pfctl.s.cnt0rdclr = clear_on_read; 182 break; 183 case 1: 184 pfctl.s.cnt1sel = event; 185 pfctl.s.cnt1ena = 1; 186 pfctl.s.cnt1rdclr = clear_on_read; 187 break; 188 case 2: 189 pfctl.s.cnt2sel = event; 190 pfctl.s.cnt2ena = 1; 191 pfctl.s.cnt2rdclr = clear_on_read; 192 break; 193 case 3: 194 default: 195 pfctl.s.cnt3sel = event; 196 pfctl.s.cnt3ena = 1; 197 pfctl.s.cnt3rdclr = clear_on_read; 198 break; 199 } 200 201 cvmx_write_csr(CVMX_L2C_PFCTL, pfctl.u64); 202 } else { 203 union cvmx_l2c_tadx_prf l2c_tadx_prf; 204 int tad; 205 206 cvmx_dprintf("L2C performance counter events are different for this chip, mapping 'event' to cvmx_l2c_tad_event_t\n"); 207 if (clear_on_read) 208 cvmx_dprintf("L2C counters don't support clear on read for this chip\n"); 209 210 l2c_tadx_prf.u64 = cvmx_read_csr(CVMX_L2C_TADX_PRF(0)); 211 212 switch (counter) { 213 case 0: 214 l2c_tadx_prf.s.cnt0sel = event; 215 break; 216 case 1: 217 l2c_tadx_prf.s.cnt1sel = event; 218 break; 219 case 2: 220 l2c_tadx_prf.s.cnt2sel = event; 221 break; 222 default: 223 case 3: 224 l2c_tadx_prf.s.cnt3sel = event; 225 break; 226 } 227 for (tad = 0; tad < CVMX_L2C_TADS; tad++) 228 cvmx_write_csr(CVMX_L2C_TADX_PRF(tad), 229 l2c_tadx_prf.u64); 230 } 231 } 232 233 uint64_t cvmx_l2c_read_perf(uint32_t counter) 234 { 235 switch (counter) { 236 case 0: 237 if (OCTEON_IS_MODEL(OCTEON_CN5XXX) || OCTEON_IS_MODEL(OCTEON_CN3XXX)) 238 return cvmx_read_csr(CVMX_L2C_PFC0); 239 else { 240 uint64_t counter = 0; 241 int tad; 242 for (tad = 0; tad < CVMX_L2C_TADS; tad++) 243 counter += cvmx_read_csr(CVMX_L2C_TADX_PFC0(tad)); 244 return counter; 245 } 246 case 1: 247 if (OCTEON_IS_MODEL(OCTEON_CN5XXX) || OCTEON_IS_MODEL(OCTEON_CN3XXX)) 248 return cvmx_read_csr(CVMX_L2C_PFC1); 249 else { 250 uint64_t counter = 0; 251 int tad; 252 for (tad = 0; tad < CVMX_L2C_TADS; tad++) 253 counter += cvmx_read_csr(CVMX_L2C_TADX_PFC1(tad)); 254 return counter; 255 } 256 case 2: 257 if (OCTEON_IS_MODEL(OCTEON_CN5XXX) || OCTEON_IS_MODEL(OCTEON_CN3XXX)) 258 return cvmx_read_csr(CVMX_L2C_PFC2); 259 else { 260 uint64_t counter = 0; 261 int tad; 262 for (tad = 0; tad < CVMX_L2C_TADS; tad++) 263 counter += cvmx_read_csr(CVMX_L2C_TADX_PFC2(tad)); 264 return counter; 265 } 266 case 3: 267 default: 268 if (OCTEON_IS_MODEL(OCTEON_CN5XXX) || OCTEON_IS_MODEL(OCTEON_CN3XXX)) 269 return cvmx_read_csr(CVMX_L2C_PFC3); 270 else { 271 uint64_t counter = 0; 272 int tad; 273 for (tad = 0; tad < CVMX_L2C_TADS; tad++) 274 counter += cvmx_read_csr(CVMX_L2C_TADX_PFC3(tad)); 275 return counter; 276 } 277 } 278 } 279 280 /** 281 * @INTERNAL 282 * Helper function use to fault in cache lines for L2 cache locking 283 * 284 * @addr: Address of base of memory region to read into L2 cache 285 * @len: Length (in bytes) of region to fault in 286 */ 287 static void fault_in(uint64_t addr, int len) 288 { 289 char *ptr; 290 291 /* 292 * Adjust addr and length so we get all cache lines even for 293 * small ranges spanning two cache lines. 294 */ 295 len += addr & CVMX_CACHE_LINE_MASK; 296 addr &= ~CVMX_CACHE_LINE_MASK; 297 ptr = cvmx_phys_to_ptr(addr); 298 /* 299 * Invalidate L1 cache to make sure all loads result in data 300 * being in L2. 301 */ 302 CVMX_DCACHE_INVALIDATE; 303 while (len > 0) { 304 ACCESS_ONCE(*ptr); 305 len -= CVMX_CACHE_LINE_SIZE; 306 ptr += CVMX_CACHE_LINE_SIZE; 307 } 308 } 309 310 int cvmx_l2c_lock_line(uint64_t addr) 311 { 312 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 313 int shift = CVMX_L2C_TAG_ADDR_ALIAS_SHIFT; 314 uint64_t assoc = cvmx_l2c_get_num_assoc(); 315 uint64_t tag = addr >> shift; 316 uint64_t index = CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, cvmx_l2c_address_to_index(addr) << CVMX_L2C_IDX_ADDR_SHIFT); 317 uint64_t way; 318 union cvmx_l2c_tadx_tag l2c_tadx_tag; 319 320 CVMX_CACHE_LCKL2(CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, addr), 0); 321 322 /* Make sure we were able to lock the line */ 323 for (way = 0; way < assoc; way++) { 324 CVMX_CACHE_LTGL2I(index | (way << shift), 0); 325 /* make sure CVMX_L2C_TADX_TAG is updated */ 326 CVMX_SYNC; 327 l2c_tadx_tag.u64 = cvmx_read_csr(CVMX_L2C_TADX_TAG(0)); 328 if (l2c_tadx_tag.s.valid && l2c_tadx_tag.s.tag == tag) 329 break; 330 } 331 332 /* Check if a valid line is found */ 333 if (way >= assoc) { 334 /* cvmx_dprintf("ERROR: cvmx_l2c_lock_line: line not found for locking at 0x%llx address\n", (unsigned long long)addr); */ 335 return -1; 336 } 337 338 /* Check if lock bit is not set */ 339 if (!l2c_tadx_tag.s.lock) { 340 /* cvmx_dprintf("ERROR: cvmx_l2c_lock_line: Not able to lock at 0x%llx address\n", (unsigned long long)addr); */ 341 return -1; 342 } 343 return way; 344 } else { 345 int retval = 0; 346 union cvmx_l2c_dbg l2cdbg; 347 union cvmx_l2c_lckbase lckbase; 348 union cvmx_l2c_lckoff lckoff; 349 union cvmx_l2t_err l2t_err; 350 351 cvmx_spinlock_lock(&cvmx_l2c_spinlock); 352 353 l2cdbg.u64 = 0; 354 lckbase.u64 = 0; 355 lckoff.u64 = 0; 356 357 /* Clear l2t error bits if set */ 358 l2t_err.u64 = cvmx_read_csr(CVMX_L2T_ERR); 359 l2t_err.s.lckerr = 1; 360 l2t_err.s.lckerr2 = 1; 361 cvmx_write_csr(CVMX_L2T_ERR, l2t_err.u64); 362 363 addr &= ~CVMX_CACHE_LINE_MASK; 364 365 /* Set this core as debug core */ 366 l2cdbg.s.ppnum = cvmx_get_core_num(); 367 CVMX_SYNC; 368 cvmx_write_csr(CVMX_L2C_DBG, l2cdbg.u64); 369 cvmx_read_csr(CVMX_L2C_DBG); 370 371 lckoff.s.lck_offset = 0; /* Only lock 1 line at a time */ 372 cvmx_write_csr(CVMX_L2C_LCKOFF, lckoff.u64); 373 cvmx_read_csr(CVMX_L2C_LCKOFF); 374 375 if (((union cvmx_l2c_cfg)(cvmx_read_csr(CVMX_L2C_CFG))).s.idxalias) { 376 int alias_shift = CVMX_L2C_IDX_ADDR_SHIFT + 2 * CVMX_L2_SET_BITS - 1; 377 uint64_t addr_tmp = addr ^ (addr & ((1 << alias_shift) - 1)) >> CVMX_L2_SET_BITS; 378 lckbase.s.lck_base = addr_tmp >> 7; 379 } else { 380 lckbase.s.lck_base = addr >> 7; 381 } 382 383 lckbase.s.lck_ena = 1; 384 cvmx_write_csr(CVMX_L2C_LCKBASE, lckbase.u64); 385 /* Make sure it gets there */ 386 cvmx_read_csr(CVMX_L2C_LCKBASE); 387 388 fault_in(addr, CVMX_CACHE_LINE_SIZE); 389 390 lckbase.s.lck_ena = 0; 391 cvmx_write_csr(CVMX_L2C_LCKBASE, lckbase.u64); 392 /* Make sure it gets there */ 393 cvmx_read_csr(CVMX_L2C_LCKBASE); 394 395 /* Stop being debug core */ 396 cvmx_write_csr(CVMX_L2C_DBG, 0); 397 cvmx_read_csr(CVMX_L2C_DBG); 398 399 l2t_err.u64 = cvmx_read_csr(CVMX_L2T_ERR); 400 if (l2t_err.s.lckerr || l2t_err.s.lckerr2) 401 retval = 1; /* We were unable to lock the line */ 402 403 cvmx_spinlock_unlock(&cvmx_l2c_spinlock); 404 return retval; 405 } 406 } 407 408 int cvmx_l2c_lock_mem_region(uint64_t start, uint64_t len) 409 { 410 int retval = 0; 411 412 /* Round start/end to cache line boundaries */ 413 len += start & CVMX_CACHE_LINE_MASK; 414 start &= ~CVMX_CACHE_LINE_MASK; 415 len = (len + CVMX_CACHE_LINE_MASK) & ~CVMX_CACHE_LINE_MASK; 416 417 while (len) { 418 retval += cvmx_l2c_lock_line(start); 419 start += CVMX_CACHE_LINE_SIZE; 420 len -= CVMX_CACHE_LINE_SIZE; 421 } 422 return retval; 423 } 424 425 void cvmx_l2c_flush(void) 426 { 427 uint64_t assoc, set; 428 uint64_t n_assoc, n_set; 429 430 n_set = cvmx_l2c_get_num_sets(); 431 n_assoc = cvmx_l2c_get_num_assoc(); 432 433 if (OCTEON_IS_MODEL(OCTEON_CN6XXX)) { 434 uint64_t address; 435 /* These may look like constants, but they aren't... */ 436 int assoc_shift = CVMX_L2C_TAG_ADDR_ALIAS_SHIFT; 437 int set_shift = CVMX_L2C_IDX_ADDR_SHIFT; 438 for (set = 0; set < n_set; set++) { 439 for (assoc = 0; assoc < n_assoc; assoc++) { 440 address = CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, 441 (assoc << assoc_shift) | (set << set_shift)); 442 CVMX_CACHE_WBIL2I(address, 0); 443 } 444 } 445 } else { 446 for (set = 0; set < n_set; set++) 447 for (assoc = 0; assoc < n_assoc; assoc++) 448 cvmx_l2c_flush_line(assoc, set); 449 } 450 } 451 452 453 int cvmx_l2c_unlock_line(uint64_t address) 454 { 455 456 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 457 int assoc; 458 union cvmx_l2c_tag tag; 459 uint32_t tag_addr; 460 uint32_t index = cvmx_l2c_address_to_index(address); 461 462 tag_addr = ((address >> CVMX_L2C_TAG_ADDR_ALIAS_SHIFT) & ((1 << CVMX_L2C_TAG_ADDR_ALIAS_SHIFT) - 1)); 463 464 /* 465 * For 63XX, we can flush a line by using the physical 466 * address directly, so finding the cache line used by 467 * the address is only required to provide the proper 468 * return value for the function. 469 */ 470 for (assoc = 0; assoc < CVMX_L2_ASSOC; assoc++) { 471 tag = cvmx_l2c_get_tag(assoc, index); 472 473 if (tag.s.V && (tag.s.addr == tag_addr)) { 474 CVMX_CACHE_WBIL2(CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, address), 0); 475 return tag.s.L; 476 } 477 } 478 } else { 479 int assoc; 480 union cvmx_l2c_tag tag; 481 uint32_t tag_addr; 482 483 uint32_t index = cvmx_l2c_address_to_index(address); 484 485 /* Compute portion of address that is stored in tag */ 486 tag_addr = ((address >> CVMX_L2C_TAG_ADDR_ALIAS_SHIFT) & ((1 << CVMX_L2C_TAG_ADDR_ALIAS_SHIFT) - 1)); 487 for (assoc = 0; assoc < CVMX_L2_ASSOC; assoc++) { 488 tag = cvmx_l2c_get_tag(assoc, index); 489 490 if (tag.s.V && (tag.s.addr == tag_addr)) { 491 cvmx_l2c_flush_line(assoc, index); 492 return tag.s.L; 493 } 494 } 495 } 496 return 0; 497 } 498 499 int cvmx_l2c_unlock_mem_region(uint64_t start, uint64_t len) 500 { 501 int num_unlocked = 0; 502 /* Round start/end to cache line boundaries */ 503 len += start & CVMX_CACHE_LINE_MASK; 504 start &= ~CVMX_CACHE_LINE_MASK; 505 len = (len + CVMX_CACHE_LINE_MASK) & ~CVMX_CACHE_LINE_MASK; 506 while (len > 0) { 507 num_unlocked += cvmx_l2c_unlock_line(start); 508 start += CVMX_CACHE_LINE_SIZE; 509 len -= CVMX_CACHE_LINE_SIZE; 510 } 511 512 return num_unlocked; 513 } 514 515 /* 516 * Internal l2c tag types. These are converted to a generic structure 517 * that can be used on all chips. 518 */ 519 union __cvmx_l2c_tag { 520 uint64_t u64; 521 struct cvmx_l2c_tag_cn50xx { 522 #ifdef __BIG_ENDIAN_BITFIELD 523 uint64_t reserved:40; 524 uint64_t V:1; /* Line valid */ 525 uint64_t D:1; /* Line dirty */ 526 uint64_t L:1; /* Line locked */ 527 uint64_t U:1; /* Use, LRU eviction */ 528 uint64_t addr:20; /* Phys mem addr (33..14) */ 529 #else 530 uint64_t addr:20; /* Phys mem addr (33..14) */ 531 uint64_t U:1; /* Use, LRU eviction */ 532 uint64_t L:1; /* Line locked */ 533 uint64_t D:1; /* Line dirty */ 534 uint64_t V:1; /* Line valid */ 535 uint64_t reserved:40; 536 #endif 537 } cn50xx; 538 struct cvmx_l2c_tag_cn30xx { 539 #ifdef __BIG_ENDIAN_BITFIELD 540 uint64_t reserved:41; 541 uint64_t V:1; /* Line valid */ 542 uint64_t D:1; /* Line dirty */ 543 uint64_t L:1; /* Line locked */ 544 uint64_t U:1; /* Use, LRU eviction */ 545 uint64_t addr:19; /* Phys mem addr (33..15) */ 546 #else 547 uint64_t addr:19; /* Phys mem addr (33..15) */ 548 uint64_t U:1; /* Use, LRU eviction */ 549 uint64_t L:1; /* Line locked */ 550 uint64_t D:1; /* Line dirty */ 551 uint64_t V:1; /* Line valid */ 552 uint64_t reserved:41; 553 #endif 554 } cn30xx; 555 struct cvmx_l2c_tag_cn31xx { 556 #ifdef __BIG_ENDIAN_BITFIELD 557 uint64_t reserved:42; 558 uint64_t V:1; /* Line valid */ 559 uint64_t D:1; /* Line dirty */ 560 uint64_t L:1; /* Line locked */ 561 uint64_t U:1; /* Use, LRU eviction */ 562 uint64_t addr:18; /* Phys mem addr (33..16) */ 563 #else 564 uint64_t addr:18; /* Phys mem addr (33..16) */ 565 uint64_t U:1; /* Use, LRU eviction */ 566 uint64_t L:1; /* Line locked */ 567 uint64_t D:1; /* Line dirty */ 568 uint64_t V:1; /* Line valid */ 569 uint64_t reserved:42; 570 #endif 571 } cn31xx; 572 struct cvmx_l2c_tag_cn38xx { 573 #ifdef __BIG_ENDIAN_BITFIELD 574 uint64_t reserved:43; 575 uint64_t V:1; /* Line valid */ 576 uint64_t D:1; /* Line dirty */ 577 uint64_t L:1; /* Line locked */ 578 uint64_t U:1; /* Use, LRU eviction */ 579 uint64_t addr:17; /* Phys mem addr (33..17) */ 580 #else 581 uint64_t addr:17; /* Phys mem addr (33..17) */ 582 uint64_t U:1; /* Use, LRU eviction */ 583 uint64_t L:1; /* Line locked */ 584 uint64_t D:1; /* Line dirty */ 585 uint64_t V:1; /* Line valid */ 586 uint64_t reserved:43; 587 #endif 588 } cn38xx; 589 struct cvmx_l2c_tag_cn58xx { 590 #ifdef __BIG_ENDIAN_BITFIELD 591 uint64_t reserved:44; 592 uint64_t V:1; /* Line valid */ 593 uint64_t D:1; /* Line dirty */ 594 uint64_t L:1; /* Line locked */ 595 uint64_t U:1; /* Use, LRU eviction */ 596 uint64_t addr:16; /* Phys mem addr (33..18) */ 597 #else 598 uint64_t addr:16; /* Phys mem addr (33..18) */ 599 uint64_t U:1; /* Use, LRU eviction */ 600 uint64_t L:1; /* Line locked */ 601 uint64_t D:1; /* Line dirty */ 602 uint64_t V:1; /* Line valid */ 603 uint64_t reserved:44; 604 #endif 605 } cn58xx; 606 struct cvmx_l2c_tag_cn58xx cn56xx; /* 2048 sets */ 607 struct cvmx_l2c_tag_cn31xx cn52xx; /* 512 sets */ 608 }; 609 610 611 /** 612 * @INTERNAL 613 * Function to read a L2C tag. This code make the current core 614 * the 'debug core' for the L2. This code must only be executed by 615 * 1 core at a time. 616 * 617 * @assoc: Association (way) of the tag to dump 618 * @index: Index of the cacheline 619 * 620 * Returns The Octeon model specific tag structure. This is 621 * translated by a wrapper function to a generic form that is 622 * easier for applications to use. 623 */ 624 static union __cvmx_l2c_tag __read_l2_tag(uint64_t assoc, uint64_t index) 625 { 626 627 uint64_t debug_tag_addr = CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, (index << 7) + 96); 628 uint64_t core = cvmx_get_core_num(); 629 union __cvmx_l2c_tag tag_val; 630 uint64_t dbg_addr = CVMX_L2C_DBG; 631 unsigned long flags; 632 633 union cvmx_l2c_dbg debug_val; 634 debug_val.u64 = 0; 635 /* 636 * For low core count parts, the core number is always small 637 * enough to stay in the correct field and not set any 638 * reserved bits. 639 */ 640 debug_val.s.ppnum = core; 641 debug_val.s.l2t = 1; 642 debug_val.s.set = assoc; 643 644 local_irq_save(flags); 645 /* 646 * Make sure core is quiet (no prefetches, etc.) before 647 * entering debug mode. 648 */ 649 CVMX_SYNC; 650 /* Flush L1 to make sure debug load misses L1 */ 651 CVMX_DCACHE_INVALIDATE; 652 653 /* 654 * The following must be done in assembly as when in debug 655 * mode all data loads from L2 return special debug data, not 656 * normal memory contents. Also, interrupts must be disabled, 657 * since if an interrupt occurs while in debug mode the ISR 658 * will get debug data from all its memory * reads instead of 659 * the contents of memory. 660 */ 661 662 asm volatile ( 663 ".set push\n\t" 664 ".set mips64\n\t" 665 ".set noreorder\n\t" 666 "sd %[dbg_val], 0(%[dbg_addr])\n\t" /* Enter debug mode, wait for store */ 667 "ld $0, 0(%[dbg_addr])\n\t" 668 "ld %[tag_val], 0(%[tag_addr])\n\t" /* Read L2C tag data */ 669 "sd $0, 0(%[dbg_addr])\n\t" /* Exit debug mode, wait for store */ 670 "ld $0, 0(%[dbg_addr])\n\t" 671 "cache 9, 0($0)\n\t" /* Invalidate dcache to discard debug data */ 672 ".set pop" 673 : [tag_val] "=r" (tag_val) 674 : [dbg_addr] "r" (dbg_addr), [dbg_val] "r" (debug_val), [tag_addr] "r" (debug_tag_addr) 675 : "memory"); 676 677 local_irq_restore(flags); 678 679 return tag_val; 680 } 681 682 683 union cvmx_l2c_tag cvmx_l2c_get_tag(uint32_t association, uint32_t index) 684 { 685 union cvmx_l2c_tag tag; 686 tag.u64 = 0; 687 688 if ((int)association >= cvmx_l2c_get_num_assoc()) { 689 cvmx_dprintf("ERROR: cvmx_l2c_get_tag association out of range\n"); 690 return tag; 691 } 692 if ((int)index >= cvmx_l2c_get_num_sets()) { 693 cvmx_dprintf("ERROR: cvmx_l2c_get_tag index out of range (arg: %d, max: %d)\n", 694 (int)index, cvmx_l2c_get_num_sets()); 695 return tag; 696 } 697 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 698 union cvmx_l2c_tadx_tag l2c_tadx_tag; 699 uint64_t address = CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, 700 (association << CVMX_L2C_TAG_ADDR_ALIAS_SHIFT) | 701 (index << CVMX_L2C_IDX_ADDR_SHIFT)); 702 /* 703 * Use L2 cache Index load tag cache instruction, as 704 * hardware loads the virtual tag for the L2 cache 705 * block with the contents of L2C_TAD0_TAG 706 * register. 707 */ 708 CVMX_CACHE_LTGL2I(address, 0); 709 CVMX_SYNC; /* make sure CVMX_L2C_TADX_TAG is updated */ 710 l2c_tadx_tag.u64 = cvmx_read_csr(CVMX_L2C_TADX_TAG(0)); 711 712 tag.s.V = l2c_tadx_tag.s.valid; 713 tag.s.D = l2c_tadx_tag.s.dirty; 714 tag.s.L = l2c_tadx_tag.s.lock; 715 tag.s.U = l2c_tadx_tag.s.use; 716 tag.s.addr = l2c_tadx_tag.s.tag; 717 } else { 718 union __cvmx_l2c_tag tmp_tag; 719 /* __read_l2_tag is intended for internal use only */ 720 tmp_tag = __read_l2_tag(association, index); 721 722 /* 723 * Convert all tag structure types to generic version, 724 * as it can represent all models. 725 */ 726 if (OCTEON_IS_MODEL(OCTEON_CN58XX) || OCTEON_IS_MODEL(OCTEON_CN56XX)) { 727 tag.s.V = tmp_tag.cn58xx.V; 728 tag.s.D = tmp_tag.cn58xx.D; 729 tag.s.L = tmp_tag.cn58xx.L; 730 tag.s.U = tmp_tag.cn58xx.U; 731 tag.s.addr = tmp_tag.cn58xx.addr; 732 } else if (OCTEON_IS_MODEL(OCTEON_CN38XX)) { 733 tag.s.V = tmp_tag.cn38xx.V; 734 tag.s.D = tmp_tag.cn38xx.D; 735 tag.s.L = tmp_tag.cn38xx.L; 736 tag.s.U = tmp_tag.cn38xx.U; 737 tag.s.addr = tmp_tag.cn38xx.addr; 738 } else if (OCTEON_IS_MODEL(OCTEON_CN31XX) || OCTEON_IS_MODEL(OCTEON_CN52XX)) { 739 tag.s.V = tmp_tag.cn31xx.V; 740 tag.s.D = tmp_tag.cn31xx.D; 741 tag.s.L = tmp_tag.cn31xx.L; 742 tag.s.U = tmp_tag.cn31xx.U; 743 tag.s.addr = tmp_tag.cn31xx.addr; 744 } else if (OCTEON_IS_MODEL(OCTEON_CN30XX)) { 745 tag.s.V = tmp_tag.cn30xx.V; 746 tag.s.D = tmp_tag.cn30xx.D; 747 tag.s.L = tmp_tag.cn30xx.L; 748 tag.s.U = tmp_tag.cn30xx.U; 749 tag.s.addr = tmp_tag.cn30xx.addr; 750 } else if (OCTEON_IS_MODEL(OCTEON_CN50XX)) { 751 tag.s.V = tmp_tag.cn50xx.V; 752 tag.s.D = tmp_tag.cn50xx.D; 753 tag.s.L = tmp_tag.cn50xx.L; 754 tag.s.U = tmp_tag.cn50xx.U; 755 tag.s.addr = tmp_tag.cn50xx.addr; 756 } else { 757 cvmx_dprintf("Unsupported OCTEON Model in %s\n", __func__); 758 } 759 } 760 return tag; 761 } 762 763 uint32_t cvmx_l2c_address_to_index(uint64_t addr) 764 { 765 uint64_t idx = addr >> CVMX_L2C_IDX_ADDR_SHIFT; 766 int indxalias = 0; 767 768 if (OCTEON_IS_MODEL(OCTEON_CN6XXX)) { 769 union cvmx_l2c_ctl l2c_ctl; 770 l2c_ctl.u64 = cvmx_read_csr(CVMX_L2C_CTL); 771 indxalias = !l2c_ctl.s.disidxalias; 772 } else { 773 union cvmx_l2c_cfg l2c_cfg; 774 l2c_cfg.u64 = cvmx_read_csr(CVMX_L2C_CFG); 775 indxalias = l2c_cfg.s.idxalias; 776 } 777 778 if (indxalias) { 779 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 780 uint32_t a_14_12 = (idx / (CVMX_L2C_MEMBANK_SELECT_SIZE/(1<<CVMX_L2C_IDX_ADDR_SHIFT))) & 0x7; 781 idx ^= idx / cvmx_l2c_get_num_sets(); 782 idx ^= a_14_12; 783 } else { 784 idx ^= ((addr & CVMX_L2C_ALIAS_MASK) >> CVMX_L2C_TAG_ADDR_ALIAS_SHIFT); 785 } 786 } 787 idx &= CVMX_L2C_IDX_MASK; 788 return idx; 789 } 790 791 int cvmx_l2c_get_cache_size_bytes(void) 792 { 793 return cvmx_l2c_get_num_sets() * cvmx_l2c_get_num_assoc() * 794 CVMX_CACHE_LINE_SIZE; 795 } 796 797 /** 798 * Return log base 2 of the number of sets in the L2 cache 799 * Returns 800 */ 801 int cvmx_l2c_get_set_bits(void) 802 { 803 int l2_set_bits; 804 if (OCTEON_IS_MODEL(OCTEON_CN56XX) || OCTEON_IS_MODEL(OCTEON_CN58XX)) 805 l2_set_bits = 11; /* 2048 sets */ 806 else if (OCTEON_IS_MODEL(OCTEON_CN38XX) || OCTEON_IS_MODEL(OCTEON_CN63XX)) 807 l2_set_bits = 10; /* 1024 sets */ 808 else if (OCTEON_IS_MODEL(OCTEON_CN31XX) || OCTEON_IS_MODEL(OCTEON_CN52XX)) 809 l2_set_bits = 9; /* 512 sets */ 810 else if (OCTEON_IS_MODEL(OCTEON_CN30XX)) 811 l2_set_bits = 8; /* 256 sets */ 812 else if (OCTEON_IS_MODEL(OCTEON_CN50XX)) 813 l2_set_bits = 7; /* 128 sets */ 814 else { 815 cvmx_dprintf("Unsupported OCTEON Model in %s\n", __func__); 816 l2_set_bits = 11; /* 2048 sets */ 817 } 818 return l2_set_bits; 819 } 820 821 /* Return the number of sets in the L2 Cache */ 822 int cvmx_l2c_get_num_sets(void) 823 { 824 return 1 << cvmx_l2c_get_set_bits(); 825 } 826 827 /* Return the number of associations in the L2 Cache */ 828 int cvmx_l2c_get_num_assoc(void) 829 { 830 int l2_assoc; 831 if (OCTEON_IS_MODEL(OCTEON_CN56XX) || 832 OCTEON_IS_MODEL(OCTEON_CN52XX) || 833 OCTEON_IS_MODEL(OCTEON_CN58XX) || 834 OCTEON_IS_MODEL(OCTEON_CN50XX) || 835 OCTEON_IS_MODEL(OCTEON_CN38XX)) 836 l2_assoc = 8; 837 else if (OCTEON_IS_MODEL(OCTEON_CN63XX)) 838 l2_assoc = 16; 839 else if (OCTEON_IS_MODEL(OCTEON_CN31XX) || 840 OCTEON_IS_MODEL(OCTEON_CN30XX)) 841 l2_assoc = 4; 842 else { 843 cvmx_dprintf("Unsupported OCTEON Model in %s\n", __func__); 844 l2_assoc = 8; 845 } 846 847 /* Check to see if part of the cache is disabled */ 848 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 849 union cvmx_mio_fus_dat3 mio_fus_dat3; 850 851 mio_fus_dat3.u64 = cvmx_read_csr(CVMX_MIO_FUS_DAT3); 852 /* 853 * cvmx_mio_fus_dat3.s.l2c_crip fuses map as follows 854 * <2> will be not used for 63xx 855 * <1> disables 1/2 ways 856 * <0> disables 1/4 ways 857 * They are cumulative, so for 63xx: 858 * <1> <0> 859 * 0 0 16-way 2MB cache 860 * 0 1 12-way 1.5MB cache 861 * 1 0 8-way 1MB cache 862 * 1 1 4-way 512KB cache 863 */ 864 865 if (mio_fus_dat3.s.l2c_crip == 3) 866 l2_assoc = 4; 867 else if (mio_fus_dat3.s.l2c_crip == 2) 868 l2_assoc = 8; 869 else if (mio_fus_dat3.s.l2c_crip == 1) 870 l2_assoc = 12; 871 } else { 872 union cvmx_l2d_fus3 val; 873 val.u64 = cvmx_read_csr(CVMX_L2D_FUS3); 874 /* 875 * Using shifts here, as bit position names are 876 * different for each model but they all mean the 877 * same. 878 */ 879 if ((val.u64 >> 35) & 0x1) 880 l2_assoc = l2_assoc >> 2; 881 else if ((val.u64 >> 34) & 0x1) 882 l2_assoc = l2_assoc >> 1; 883 } 884 return l2_assoc; 885 } 886 887 /** 888 * Flush a line from the L2 cache 889 * This should only be called from one core at a time, as this routine 890 * sets the core to the 'debug' core in order to flush the line. 891 * 892 * @assoc: Association (or way) to flush 893 * @index: Index to flush 894 */ 895 void cvmx_l2c_flush_line(uint32_t assoc, uint32_t index) 896 { 897 /* Check the range of the index. */ 898 if (index > (uint32_t)cvmx_l2c_get_num_sets()) { 899 cvmx_dprintf("ERROR: cvmx_l2c_flush_line index out of range.\n"); 900 return; 901 } 902 903 /* Check the range of association. */ 904 if (assoc > (uint32_t)cvmx_l2c_get_num_assoc()) { 905 cvmx_dprintf("ERROR: cvmx_l2c_flush_line association out of range.\n"); 906 return; 907 } 908 909 if (OCTEON_IS_MODEL(OCTEON_CN63XX)) { 910 uint64_t address; 911 /* Create the address based on index and association. 912 * Bits<20:17> select the way of the cache block involved in 913 * the operation 914 * Bits<16:7> of the effect address select the index 915 */ 916 address = CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, 917 (assoc << CVMX_L2C_TAG_ADDR_ALIAS_SHIFT) | 918 (index << CVMX_L2C_IDX_ADDR_SHIFT)); 919 CVMX_CACHE_WBIL2I(address, 0); 920 } else { 921 union cvmx_l2c_dbg l2cdbg; 922 923 l2cdbg.u64 = 0; 924 if (!OCTEON_IS_MODEL(OCTEON_CN30XX)) 925 l2cdbg.s.ppnum = cvmx_get_core_num(); 926 l2cdbg.s.finv = 1; 927 928 l2cdbg.s.set = assoc; 929 cvmx_spinlock_lock(&cvmx_l2c_spinlock); 930 /* 931 * Enter debug mode, and make sure all other writes 932 * complete before we enter debug mode 933 */ 934 CVMX_SYNC; 935 cvmx_write_csr(CVMX_L2C_DBG, l2cdbg.u64); 936 cvmx_read_csr(CVMX_L2C_DBG); 937 938 CVMX_PREPARE_FOR_STORE(CVMX_ADD_SEG(CVMX_MIPS_SPACE_XKPHYS, 939 index * CVMX_CACHE_LINE_SIZE), 940 0); 941 /* Exit debug mode */ 942 CVMX_SYNC; 943 cvmx_write_csr(CVMX_L2C_DBG, 0); 944 cvmx_read_csr(CVMX_L2C_DBG); 945 cvmx_spinlock_unlock(&cvmx_l2c_spinlock); 946 } 947 } 948