1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2019, Intel Corporation. 4 * 5 * Heterogeneous Memory Attributes Table (HMAT) representation 6 * 7 * This program parses and reports the platform's HMAT tables, and registers 8 * the applicable attributes with the node's interfaces. 9 */ 10 11 #define pr_fmt(fmt) "acpi/hmat: " fmt 12 #define dev_fmt(fmt) "acpi/hmat: " fmt 13 14 #include <linux/acpi.h> 15 #include <linux/bitops.h> 16 #include <linux/device.h> 17 #include <linux/init.h> 18 #include <linux/list.h> 19 #include <linux/mm.h> 20 #include <linux/platform_device.h> 21 #include <linux/list_sort.h> 22 #include <linux/memregion.h> 23 #include <linux/memory.h> 24 #include <linux/mutex.h> 25 #include <linux/node.h> 26 #include <linux/sysfs.h> 27 #include <linux/dax.h> 28 29 static u8 hmat_revision; 30 static int hmat_disable __initdata; 31 32 void __init disable_hmat(void) 33 { 34 hmat_disable = 1; 35 } 36 37 static LIST_HEAD(targets); 38 static LIST_HEAD(initiators); 39 static LIST_HEAD(localities); 40 41 static DEFINE_MUTEX(target_lock); 42 43 /* 44 * The defined enum order is used to prioritize attributes to break ties when 45 * selecting the best performing node. 46 */ 47 enum locality_types { 48 WRITE_LATENCY, 49 READ_LATENCY, 50 WRITE_BANDWIDTH, 51 READ_BANDWIDTH, 52 }; 53 54 static struct memory_locality *localities_types[4]; 55 56 struct target_cache { 57 struct list_head node; 58 struct node_cache_attrs cache_attrs; 59 }; 60 61 struct memory_target { 62 struct list_head node; 63 unsigned int memory_pxm; 64 unsigned int processor_pxm; 65 struct resource memregions; 66 struct node_hmem_attrs hmem_attrs[2]; 67 struct list_head caches; 68 struct node_cache_attrs cache_attrs; 69 bool registered; 70 }; 71 72 struct memory_initiator { 73 struct list_head node; 74 unsigned int processor_pxm; 75 bool has_cpu; 76 }; 77 78 struct memory_locality { 79 struct list_head node; 80 struct acpi_hmat_locality *hmat_loc; 81 }; 82 83 static struct memory_initiator *find_mem_initiator(unsigned int cpu_pxm) 84 { 85 struct memory_initiator *initiator; 86 87 list_for_each_entry(initiator, &initiators, node) 88 if (initiator->processor_pxm == cpu_pxm) 89 return initiator; 90 return NULL; 91 } 92 93 static struct memory_target *find_mem_target(unsigned int mem_pxm) 94 { 95 struct memory_target *target; 96 97 list_for_each_entry(target, &targets, node) 98 if (target->memory_pxm == mem_pxm) 99 return target; 100 return NULL; 101 } 102 103 static __init void alloc_memory_initiator(unsigned int cpu_pxm) 104 { 105 struct memory_initiator *initiator; 106 107 if (pxm_to_node(cpu_pxm) == NUMA_NO_NODE) 108 return; 109 110 initiator = find_mem_initiator(cpu_pxm); 111 if (initiator) 112 return; 113 114 initiator = kzalloc(sizeof(*initiator), GFP_KERNEL); 115 if (!initiator) 116 return; 117 118 initiator->processor_pxm = cpu_pxm; 119 initiator->has_cpu = node_state(pxm_to_node(cpu_pxm), N_CPU); 120 list_add_tail(&initiator->node, &initiators); 121 } 122 123 static __init void alloc_memory_target(unsigned int mem_pxm, 124 resource_size_t start, resource_size_t len) 125 { 126 struct memory_target *target; 127 128 target = find_mem_target(mem_pxm); 129 if (!target) { 130 target = kzalloc(sizeof(*target), GFP_KERNEL); 131 if (!target) 132 return; 133 target->memory_pxm = mem_pxm; 134 target->processor_pxm = PXM_INVAL; 135 target->memregions = (struct resource) { 136 .name = "ACPI mem", 137 .start = 0, 138 .end = -1, 139 .flags = IORESOURCE_MEM, 140 }; 141 list_add_tail(&target->node, &targets); 142 INIT_LIST_HEAD(&target->caches); 143 } 144 145 /* 146 * There are potentially multiple ranges per PXM, so record each 147 * in the per-target memregions resource tree. 148 */ 149 if (!__request_region(&target->memregions, start, len, "memory target", 150 IORESOURCE_MEM)) 151 pr_warn("failed to reserve %#llx - %#llx in pxm: %d\n", 152 start, start + len, mem_pxm); 153 } 154 155 static __init const char *hmat_data_type(u8 type) 156 { 157 switch (type) { 158 case ACPI_HMAT_ACCESS_LATENCY: 159 return "Access Latency"; 160 case ACPI_HMAT_READ_LATENCY: 161 return "Read Latency"; 162 case ACPI_HMAT_WRITE_LATENCY: 163 return "Write Latency"; 164 case ACPI_HMAT_ACCESS_BANDWIDTH: 165 return "Access Bandwidth"; 166 case ACPI_HMAT_READ_BANDWIDTH: 167 return "Read Bandwidth"; 168 case ACPI_HMAT_WRITE_BANDWIDTH: 169 return "Write Bandwidth"; 170 default: 171 return "Reserved"; 172 } 173 } 174 175 static __init const char *hmat_data_type_suffix(u8 type) 176 { 177 switch (type) { 178 case ACPI_HMAT_ACCESS_LATENCY: 179 case ACPI_HMAT_READ_LATENCY: 180 case ACPI_HMAT_WRITE_LATENCY: 181 return " nsec"; 182 case ACPI_HMAT_ACCESS_BANDWIDTH: 183 case ACPI_HMAT_READ_BANDWIDTH: 184 case ACPI_HMAT_WRITE_BANDWIDTH: 185 return " MB/s"; 186 default: 187 return ""; 188 } 189 } 190 191 static u32 hmat_normalize(u16 entry, u64 base, u8 type) 192 { 193 u32 value; 194 195 /* 196 * Check for invalid and overflow values 197 */ 198 if (entry == 0xffff || !entry) 199 return 0; 200 else if (base > (UINT_MAX / (entry))) 201 return 0; 202 203 /* 204 * Divide by the base unit for version 1, convert latency from 205 * picosenonds to nanoseconds if revision 2. 206 */ 207 value = entry * base; 208 if (hmat_revision == 1) { 209 if (value < 10) 210 return 0; 211 value = DIV_ROUND_UP(value, 10); 212 } else if (hmat_revision == 2) { 213 switch (type) { 214 case ACPI_HMAT_ACCESS_LATENCY: 215 case ACPI_HMAT_READ_LATENCY: 216 case ACPI_HMAT_WRITE_LATENCY: 217 value = DIV_ROUND_UP(value, 1000); 218 break; 219 default: 220 break; 221 } 222 } 223 return value; 224 } 225 226 static void hmat_update_target_access(struct memory_target *target, 227 u8 type, u32 value, int access) 228 { 229 switch (type) { 230 case ACPI_HMAT_ACCESS_LATENCY: 231 target->hmem_attrs[access].read_latency = value; 232 target->hmem_attrs[access].write_latency = value; 233 break; 234 case ACPI_HMAT_READ_LATENCY: 235 target->hmem_attrs[access].read_latency = value; 236 break; 237 case ACPI_HMAT_WRITE_LATENCY: 238 target->hmem_attrs[access].write_latency = value; 239 break; 240 case ACPI_HMAT_ACCESS_BANDWIDTH: 241 target->hmem_attrs[access].read_bandwidth = value; 242 target->hmem_attrs[access].write_bandwidth = value; 243 break; 244 case ACPI_HMAT_READ_BANDWIDTH: 245 target->hmem_attrs[access].read_bandwidth = value; 246 break; 247 case ACPI_HMAT_WRITE_BANDWIDTH: 248 target->hmem_attrs[access].write_bandwidth = value; 249 break; 250 default: 251 break; 252 } 253 } 254 255 static __init void hmat_add_locality(struct acpi_hmat_locality *hmat_loc) 256 { 257 struct memory_locality *loc; 258 259 loc = kzalloc(sizeof(*loc), GFP_KERNEL); 260 if (!loc) { 261 pr_notice_once("Failed to allocate HMAT locality\n"); 262 return; 263 } 264 265 loc->hmat_loc = hmat_loc; 266 list_add_tail(&loc->node, &localities); 267 268 switch (hmat_loc->data_type) { 269 case ACPI_HMAT_ACCESS_LATENCY: 270 localities_types[READ_LATENCY] = loc; 271 localities_types[WRITE_LATENCY] = loc; 272 break; 273 case ACPI_HMAT_READ_LATENCY: 274 localities_types[READ_LATENCY] = loc; 275 break; 276 case ACPI_HMAT_WRITE_LATENCY: 277 localities_types[WRITE_LATENCY] = loc; 278 break; 279 case ACPI_HMAT_ACCESS_BANDWIDTH: 280 localities_types[READ_BANDWIDTH] = loc; 281 localities_types[WRITE_BANDWIDTH] = loc; 282 break; 283 case ACPI_HMAT_READ_BANDWIDTH: 284 localities_types[READ_BANDWIDTH] = loc; 285 break; 286 case ACPI_HMAT_WRITE_BANDWIDTH: 287 localities_types[WRITE_BANDWIDTH] = loc; 288 break; 289 default: 290 break; 291 } 292 } 293 294 static __init int hmat_parse_locality(union acpi_subtable_headers *header, 295 const unsigned long end) 296 { 297 struct acpi_hmat_locality *hmat_loc = (void *)header; 298 struct memory_target *target; 299 unsigned int init, targ, total_size, ipds, tpds; 300 u32 *inits, *targs, value; 301 u16 *entries; 302 u8 type, mem_hier; 303 304 if (hmat_loc->header.length < sizeof(*hmat_loc)) { 305 pr_notice("HMAT: Unexpected locality header length: %u\n", 306 hmat_loc->header.length); 307 return -EINVAL; 308 } 309 310 type = hmat_loc->data_type; 311 mem_hier = hmat_loc->flags & ACPI_HMAT_MEMORY_HIERARCHY; 312 ipds = hmat_loc->number_of_initiator_Pds; 313 tpds = hmat_loc->number_of_target_Pds; 314 total_size = sizeof(*hmat_loc) + sizeof(*entries) * ipds * tpds + 315 sizeof(*inits) * ipds + sizeof(*targs) * tpds; 316 if (hmat_loc->header.length < total_size) { 317 pr_notice("HMAT: Unexpected locality header length:%u, minimum required:%u\n", 318 hmat_loc->header.length, total_size); 319 return -EINVAL; 320 } 321 322 pr_info("HMAT: Locality: Flags:%02x Type:%s Initiator Domains:%u Target Domains:%u Base:%lld\n", 323 hmat_loc->flags, hmat_data_type(type), ipds, tpds, 324 hmat_loc->entry_base_unit); 325 326 inits = (u32 *)(hmat_loc + 1); 327 targs = inits + ipds; 328 entries = (u16 *)(targs + tpds); 329 for (init = 0; init < ipds; init++) { 330 alloc_memory_initiator(inits[init]); 331 for (targ = 0; targ < tpds; targ++) { 332 value = hmat_normalize(entries[init * tpds + targ], 333 hmat_loc->entry_base_unit, 334 type); 335 pr_info(" Initiator-Target[%u-%u]:%u%s\n", 336 inits[init], targs[targ], value, 337 hmat_data_type_suffix(type)); 338 339 if (mem_hier == ACPI_HMAT_MEMORY) { 340 target = find_mem_target(targs[targ]); 341 if (target && target->processor_pxm == inits[init]) { 342 hmat_update_target_access(target, type, value, 0); 343 /* If the node has a CPU, update access 1 */ 344 if (node_state(pxm_to_node(inits[init]), N_CPU)) 345 hmat_update_target_access(target, type, value, 1); 346 } 347 } 348 } 349 } 350 351 if (mem_hier == ACPI_HMAT_MEMORY) 352 hmat_add_locality(hmat_loc); 353 354 return 0; 355 } 356 357 static __init int hmat_parse_cache(union acpi_subtable_headers *header, 358 const unsigned long end) 359 { 360 struct acpi_hmat_cache *cache = (void *)header; 361 struct memory_target *target; 362 struct target_cache *tcache; 363 u32 attrs; 364 365 if (cache->header.length < sizeof(*cache)) { 366 pr_notice("HMAT: Unexpected cache header length: %u\n", 367 cache->header.length); 368 return -EINVAL; 369 } 370 371 attrs = cache->cache_attributes; 372 pr_info("HMAT: Cache: Domain:%u Size:%llu Attrs:%08x SMBIOS Handles:%d\n", 373 cache->memory_PD, cache->cache_size, attrs, 374 cache->number_of_SMBIOShandles); 375 376 target = find_mem_target(cache->memory_PD); 377 if (!target) 378 return 0; 379 380 tcache = kzalloc(sizeof(*tcache), GFP_KERNEL); 381 if (!tcache) { 382 pr_notice_once("Failed to allocate HMAT cache info\n"); 383 return 0; 384 } 385 386 tcache->cache_attrs.size = cache->cache_size; 387 tcache->cache_attrs.level = (attrs & ACPI_HMAT_CACHE_LEVEL) >> 4; 388 tcache->cache_attrs.line_size = (attrs & ACPI_HMAT_CACHE_LINE_SIZE) >> 16; 389 390 switch ((attrs & ACPI_HMAT_CACHE_ASSOCIATIVITY) >> 8) { 391 case ACPI_HMAT_CA_DIRECT_MAPPED: 392 tcache->cache_attrs.indexing = NODE_CACHE_DIRECT_MAP; 393 break; 394 case ACPI_HMAT_CA_COMPLEX_CACHE_INDEXING: 395 tcache->cache_attrs.indexing = NODE_CACHE_INDEXED; 396 break; 397 case ACPI_HMAT_CA_NONE: 398 default: 399 tcache->cache_attrs.indexing = NODE_CACHE_OTHER; 400 break; 401 } 402 403 switch ((attrs & ACPI_HMAT_WRITE_POLICY) >> 12) { 404 case ACPI_HMAT_CP_WB: 405 tcache->cache_attrs.write_policy = NODE_CACHE_WRITE_BACK; 406 break; 407 case ACPI_HMAT_CP_WT: 408 tcache->cache_attrs.write_policy = NODE_CACHE_WRITE_THROUGH; 409 break; 410 case ACPI_HMAT_CP_NONE: 411 default: 412 tcache->cache_attrs.write_policy = NODE_CACHE_WRITE_OTHER; 413 break; 414 } 415 list_add_tail(&tcache->node, &target->caches); 416 417 return 0; 418 } 419 420 static int __init hmat_parse_proximity_domain(union acpi_subtable_headers *header, 421 const unsigned long end) 422 { 423 struct acpi_hmat_proximity_domain *p = (void *)header; 424 struct memory_target *target = NULL; 425 426 if (p->header.length != sizeof(*p)) { 427 pr_notice("HMAT: Unexpected address range header length: %u\n", 428 p->header.length); 429 return -EINVAL; 430 } 431 432 if (hmat_revision == 1) 433 pr_info("HMAT: Memory (%#llx length %#llx) Flags:%04x Processor Domain:%u Memory Domain:%u\n", 434 p->reserved3, p->reserved4, p->flags, p->processor_PD, 435 p->memory_PD); 436 else 437 pr_info("HMAT: Memory Flags:%04x Processor Domain:%u Memory Domain:%u\n", 438 p->flags, p->processor_PD, p->memory_PD); 439 440 if ((hmat_revision == 1 && p->flags & ACPI_HMAT_MEMORY_PD_VALID) || 441 hmat_revision > 1) { 442 target = find_mem_target(p->memory_PD); 443 if (!target) { 444 pr_debug("HMAT: Memory Domain missing from SRAT\n"); 445 return -EINVAL; 446 } 447 } 448 if (target && p->flags & ACPI_HMAT_PROCESSOR_PD_VALID) { 449 int p_node = pxm_to_node(p->processor_PD); 450 451 if (p_node == NUMA_NO_NODE) { 452 pr_debug("HMAT: Invalid Processor Domain\n"); 453 return -EINVAL; 454 } 455 target->processor_pxm = p->processor_PD; 456 } 457 458 return 0; 459 } 460 461 static int __init hmat_parse_subtable(union acpi_subtable_headers *header, 462 const unsigned long end) 463 { 464 struct acpi_hmat_structure *hdr = (void *)header; 465 466 if (!hdr) 467 return -EINVAL; 468 469 switch (hdr->type) { 470 case ACPI_HMAT_TYPE_PROXIMITY: 471 return hmat_parse_proximity_domain(header, end); 472 case ACPI_HMAT_TYPE_LOCALITY: 473 return hmat_parse_locality(header, end); 474 case ACPI_HMAT_TYPE_CACHE: 475 return hmat_parse_cache(header, end); 476 default: 477 return -EINVAL; 478 } 479 } 480 481 static __init int srat_parse_mem_affinity(union acpi_subtable_headers *header, 482 const unsigned long end) 483 { 484 struct acpi_srat_mem_affinity *ma = (void *)header; 485 486 if (!ma) 487 return -EINVAL; 488 if (!(ma->flags & ACPI_SRAT_MEM_ENABLED)) 489 return 0; 490 alloc_memory_target(ma->proximity_domain, ma->base_address, ma->length); 491 return 0; 492 } 493 494 static u32 hmat_initiator_perf(struct memory_target *target, 495 struct memory_initiator *initiator, 496 struct acpi_hmat_locality *hmat_loc) 497 { 498 unsigned int ipds, tpds, i, idx = 0, tdx = 0; 499 u32 *inits, *targs; 500 u16 *entries; 501 502 ipds = hmat_loc->number_of_initiator_Pds; 503 tpds = hmat_loc->number_of_target_Pds; 504 inits = (u32 *)(hmat_loc + 1); 505 targs = inits + ipds; 506 entries = (u16 *)(targs + tpds); 507 508 for (i = 0; i < ipds; i++) { 509 if (inits[i] == initiator->processor_pxm) { 510 idx = i; 511 break; 512 } 513 } 514 515 if (i == ipds) 516 return 0; 517 518 for (i = 0; i < tpds; i++) { 519 if (targs[i] == target->memory_pxm) { 520 tdx = i; 521 break; 522 } 523 } 524 if (i == tpds) 525 return 0; 526 527 return hmat_normalize(entries[idx * tpds + tdx], 528 hmat_loc->entry_base_unit, 529 hmat_loc->data_type); 530 } 531 532 static bool hmat_update_best(u8 type, u32 value, u32 *best) 533 { 534 bool updated = false; 535 536 if (!value) 537 return false; 538 539 switch (type) { 540 case ACPI_HMAT_ACCESS_LATENCY: 541 case ACPI_HMAT_READ_LATENCY: 542 case ACPI_HMAT_WRITE_LATENCY: 543 if (!*best || *best > value) { 544 *best = value; 545 updated = true; 546 } 547 break; 548 case ACPI_HMAT_ACCESS_BANDWIDTH: 549 case ACPI_HMAT_READ_BANDWIDTH: 550 case ACPI_HMAT_WRITE_BANDWIDTH: 551 if (!*best || *best < value) { 552 *best = value; 553 updated = true; 554 } 555 break; 556 } 557 558 return updated; 559 } 560 561 static int initiator_cmp(void *priv, struct list_head *a, struct list_head *b) 562 { 563 struct memory_initiator *ia; 564 struct memory_initiator *ib; 565 unsigned long *p_nodes = priv; 566 567 ia = list_entry(a, struct memory_initiator, node); 568 ib = list_entry(b, struct memory_initiator, node); 569 570 set_bit(ia->processor_pxm, p_nodes); 571 set_bit(ib->processor_pxm, p_nodes); 572 573 return ia->processor_pxm - ib->processor_pxm; 574 } 575 576 static void hmat_register_target_initiators(struct memory_target *target) 577 { 578 static DECLARE_BITMAP(p_nodes, MAX_NUMNODES); 579 struct memory_initiator *initiator; 580 unsigned int mem_nid, cpu_nid; 581 struct memory_locality *loc = NULL; 582 u32 best = 0; 583 bool access0done = false; 584 int i; 585 586 mem_nid = pxm_to_node(target->memory_pxm); 587 /* 588 * If the Address Range Structure provides a local processor pxm, link 589 * only that one. Otherwise, find the best performance attributes and 590 * register all initiators that match. 591 */ 592 if (target->processor_pxm != PXM_INVAL) { 593 cpu_nid = pxm_to_node(target->processor_pxm); 594 register_memory_node_under_compute_node(mem_nid, cpu_nid, 0); 595 access0done = true; 596 if (node_state(cpu_nid, N_CPU)) { 597 register_memory_node_under_compute_node(mem_nid, cpu_nid, 1); 598 return; 599 } 600 } 601 602 if (list_empty(&localities)) 603 return; 604 605 /* 606 * We need the initiator list sorted so we can use bitmap_clear for 607 * previously set initiators when we find a better memory accessor. 608 * We'll also use the sorting to prime the candidate nodes with known 609 * initiators. 610 */ 611 bitmap_zero(p_nodes, MAX_NUMNODES); 612 list_sort(p_nodes, &initiators, initiator_cmp); 613 if (!access0done) { 614 for (i = WRITE_LATENCY; i <= READ_BANDWIDTH; i++) { 615 loc = localities_types[i]; 616 if (!loc) 617 continue; 618 619 best = 0; 620 list_for_each_entry(initiator, &initiators, node) { 621 u32 value; 622 623 if (!test_bit(initiator->processor_pxm, p_nodes)) 624 continue; 625 626 value = hmat_initiator_perf(target, initiator, 627 loc->hmat_loc); 628 if (hmat_update_best(loc->hmat_loc->data_type, value, &best)) 629 bitmap_clear(p_nodes, 0, initiator->processor_pxm); 630 if (value != best) 631 clear_bit(initiator->processor_pxm, p_nodes); 632 } 633 if (best) 634 hmat_update_target_access(target, loc->hmat_loc->data_type, 635 best, 0); 636 } 637 638 for_each_set_bit(i, p_nodes, MAX_NUMNODES) { 639 cpu_nid = pxm_to_node(i); 640 register_memory_node_under_compute_node(mem_nid, cpu_nid, 0); 641 } 642 } 643 644 /* Access 1 ignores Generic Initiators */ 645 bitmap_zero(p_nodes, MAX_NUMNODES); 646 list_sort(p_nodes, &initiators, initiator_cmp); 647 best = 0; 648 for (i = WRITE_LATENCY; i <= READ_BANDWIDTH; i++) { 649 loc = localities_types[i]; 650 if (!loc) 651 continue; 652 653 best = 0; 654 list_for_each_entry(initiator, &initiators, node) { 655 u32 value; 656 657 if (!initiator->has_cpu) { 658 clear_bit(initiator->processor_pxm, p_nodes); 659 continue; 660 } 661 if (!test_bit(initiator->processor_pxm, p_nodes)) 662 continue; 663 664 value = hmat_initiator_perf(target, initiator, loc->hmat_loc); 665 if (hmat_update_best(loc->hmat_loc->data_type, value, &best)) 666 bitmap_clear(p_nodes, 0, initiator->processor_pxm); 667 if (value != best) 668 clear_bit(initiator->processor_pxm, p_nodes); 669 } 670 if (best) 671 hmat_update_target_access(target, loc->hmat_loc->data_type, best, 1); 672 } 673 for_each_set_bit(i, p_nodes, MAX_NUMNODES) { 674 cpu_nid = pxm_to_node(i); 675 register_memory_node_under_compute_node(mem_nid, cpu_nid, 1); 676 } 677 } 678 679 static void hmat_register_target_cache(struct memory_target *target) 680 { 681 unsigned mem_nid = pxm_to_node(target->memory_pxm); 682 struct target_cache *tcache; 683 684 list_for_each_entry(tcache, &target->caches, node) 685 node_add_cache(mem_nid, &tcache->cache_attrs); 686 } 687 688 static void hmat_register_target_perf(struct memory_target *target, int access) 689 { 690 unsigned mem_nid = pxm_to_node(target->memory_pxm); 691 node_set_perf_attrs(mem_nid, &target->hmem_attrs[access], access); 692 } 693 694 static void hmat_register_target_devices(struct memory_target *target) 695 { 696 struct resource *res; 697 698 /* 699 * Do not bother creating devices if no driver is available to 700 * consume them. 701 */ 702 if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM)) 703 return; 704 705 for (res = target->memregions.child; res; res = res->sibling) { 706 int target_nid = pxm_to_node(target->memory_pxm); 707 708 hmem_register_device(target_nid, res); 709 } 710 } 711 712 static void hmat_register_target(struct memory_target *target) 713 { 714 int nid = pxm_to_node(target->memory_pxm); 715 716 /* 717 * Devices may belong to either an offline or online 718 * node, so unconditionally add them. 719 */ 720 hmat_register_target_devices(target); 721 722 /* 723 * Skip offline nodes. This can happen when memory 724 * marked EFI_MEMORY_SP, "specific purpose", is applied 725 * to all the memory in a promixity domain leading to 726 * the node being marked offline / unplugged, or if 727 * memory-only "hotplug" node is offline. 728 */ 729 if (nid == NUMA_NO_NODE || !node_online(nid)) 730 return; 731 732 mutex_lock(&target_lock); 733 if (!target->registered) { 734 hmat_register_target_initiators(target); 735 hmat_register_target_cache(target); 736 hmat_register_target_perf(target, 0); 737 hmat_register_target_perf(target, 1); 738 target->registered = true; 739 } 740 mutex_unlock(&target_lock); 741 } 742 743 static void hmat_register_targets(void) 744 { 745 struct memory_target *target; 746 747 list_for_each_entry(target, &targets, node) 748 hmat_register_target(target); 749 } 750 751 static int hmat_callback(struct notifier_block *self, 752 unsigned long action, void *arg) 753 { 754 struct memory_target *target; 755 struct memory_notify *mnb = arg; 756 int pxm, nid = mnb->status_change_nid; 757 758 if (nid == NUMA_NO_NODE || action != MEM_ONLINE) 759 return NOTIFY_OK; 760 761 pxm = node_to_pxm(nid); 762 target = find_mem_target(pxm); 763 if (!target) 764 return NOTIFY_OK; 765 766 hmat_register_target(target); 767 return NOTIFY_OK; 768 } 769 770 static struct notifier_block hmat_callback_nb = { 771 .notifier_call = hmat_callback, 772 .priority = 2, 773 }; 774 775 static __init void hmat_free_structures(void) 776 { 777 struct memory_target *target, *tnext; 778 struct memory_locality *loc, *lnext; 779 struct memory_initiator *initiator, *inext; 780 struct target_cache *tcache, *cnext; 781 782 list_for_each_entry_safe(target, tnext, &targets, node) { 783 struct resource *res, *res_next; 784 785 list_for_each_entry_safe(tcache, cnext, &target->caches, node) { 786 list_del(&tcache->node); 787 kfree(tcache); 788 } 789 790 list_del(&target->node); 791 res = target->memregions.child; 792 while (res) { 793 res_next = res->sibling; 794 __release_region(&target->memregions, res->start, 795 resource_size(res)); 796 res = res_next; 797 } 798 kfree(target); 799 } 800 801 list_for_each_entry_safe(initiator, inext, &initiators, node) { 802 list_del(&initiator->node); 803 kfree(initiator); 804 } 805 806 list_for_each_entry_safe(loc, lnext, &localities, node) { 807 list_del(&loc->node); 808 kfree(loc); 809 } 810 } 811 812 static __init int hmat_init(void) 813 { 814 struct acpi_table_header *tbl; 815 enum acpi_hmat_type i; 816 acpi_status status; 817 818 if (srat_disabled() || hmat_disable) 819 return 0; 820 821 status = acpi_get_table(ACPI_SIG_SRAT, 0, &tbl); 822 if (ACPI_FAILURE(status)) 823 return 0; 824 825 if (acpi_table_parse_entries(ACPI_SIG_SRAT, 826 sizeof(struct acpi_table_srat), 827 ACPI_SRAT_TYPE_MEMORY_AFFINITY, 828 srat_parse_mem_affinity, 0) < 0) 829 goto out_put; 830 acpi_put_table(tbl); 831 832 status = acpi_get_table(ACPI_SIG_HMAT, 0, &tbl); 833 if (ACPI_FAILURE(status)) 834 goto out_put; 835 836 hmat_revision = tbl->revision; 837 switch (hmat_revision) { 838 case 1: 839 case 2: 840 break; 841 default: 842 pr_notice("Ignoring HMAT: Unknown revision:%d\n", hmat_revision); 843 goto out_put; 844 } 845 846 for (i = ACPI_HMAT_TYPE_PROXIMITY; i < ACPI_HMAT_TYPE_RESERVED; i++) { 847 if (acpi_table_parse_entries(ACPI_SIG_HMAT, 848 sizeof(struct acpi_table_hmat), i, 849 hmat_parse_subtable, 0) < 0) { 850 pr_notice("Ignoring HMAT: Invalid table"); 851 goto out_put; 852 } 853 } 854 hmat_register_targets(); 855 856 /* Keep the table and structures if the notifier may use them */ 857 if (!register_hotmemory_notifier(&hmat_callback_nb)) 858 return 0; 859 out_put: 860 hmat_free_structures(); 861 acpi_put_table(tbl); 862 return 0; 863 } 864 device_initcall(hmat_init); 865