1 /*- 2 * Copyright (c) 2010 Hudson River Trading LLC 3 * Written by: John H. Baldwin <jhb@FreeBSD.org> 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_vm.h" 32 33 #include <sys/param.h> 34 #include <sys/bus.h> 35 #include <sys/kernel.h> 36 #include <sys/lock.h> 37 #include <sys/mutex.h> 38 #include <sys/smp.h> 39 #include <sys/vmmeter.h> 40 #include <vm/vm.h> 41 #include <vm/pmap.h> 42 #include <vm/vm_param.h> 43 #include <vm/vm_page.h> 44 #include <vm/vm_phys.h> 45 46 #include <contrib/dev/acpica/include/acpi.h> 47 #include <contrib/dev/acpica/include/aclocal.h> 48 #include <contrib/dev/acpica/include/actables.h> 49 50 #include <machine/intr_machdep.h> 51 #include <x86/apicvar.h> 52 53 #include <dev/acpica/acpivar.h> 54 55 #if MAXMEMDOM > 1 56 struct cpu_info { 57 int enabled:1; 58 int has_memory:1; 59 int domain; 60 } cpus[MAX_APIC_ID + 1]; 61 62 struct mem_affinity mem_info[VM_PHYSSEG_MAX + 1]; 63 int num_mem; 64 65 static ACPI_TABLE_SRAT *srat; 66 static vm_paddr_t srat_physaddr; 67 68 static int domain_pxm[MAXMEMDOM]; 69 static int ndomain; 70 71 static ACPI_TABLE_SLIT *slit; 72 static vm_paddr_t slit_physaddr; 73 static int vm_locality_table[MAXMEMDOM * MAXMEMDOM]; 74 75 static void srat_walk_table(acpi_subtable_handler *handler, void *arg); 76 77 /* 78 * SLIT parsing. 79 */ 80 81 static void 82 slit_parse_table(ACPI_TABLE_SLIT *s) 83 { 84 int i, j; 85 int i_domain, j_domain; 86 int offset = 0; 87 uint8_t e; 88 89 /* 90 * This maps the SLIT data into the VM-domain centric view. 91 * There may be sparse entries in the PXM namespace, so 92 * remap them to a VM-domain ID and if it doesn't exist, 93 * skip it. 94 * 95 * It should result in a packed 2d array of VM-domain 96 * locality information entries. 97 */ 98 99 if (bootverbose) 100 printf("SLIT.Localities: %d\n", (int) s->LocalityCount); 101 for (i = 0; i < s->LocalityCount; i++) { 102 i_domain = acpi_map_pxm_to_vm_domainid(i); 103 if (i_domain < 0) 104 continue; 105 106 if (bootverbose) 107 printf("%d: ", i); 108 for (j = 0; j < s->LocalityCount; j++) { 109 j_domain = acpi_map_pxm_to_vm_domainid(j); 110 if (j_domain < 0) 111 continue; 112 e = s->Entry[i * s->LocalityCount + j]; 113 if (bootverbose) 114 printf("%d ", (int) e); 115 /* 255 == "no locality information" */ 116 if (e == 255) 117 vm_locality_table[offset] = -1; 118 else 119 vm_locality_table[offset] = e; 120 offset++; 121 } 122 if (bootverbose) 123 printf("\n"); 124 } 125 } 126 127 /* 128 * Look for an ACPI System Locality Distance Information Table ("SLIT") 129 */ 130 static int 131 parse_slit(void) 132 { 133 134 if (resource_disabled("slit", 0)) { 135 return (-1); 136 } 137 138 slit_physaddr = acpi_find_table(ACPI_SIG_SLIT); 139 if (slit_physaddr == 0) { 140 return (-1); 141 } 142 143 /* 144 * Make a pass over the table to populate the cpus[] and 145 * mem_info[] tables. 146 */ 147 slit = acpi_map_table(slit_physaddr, ACPI_SIG_SLIT); 148 slit_parse_table(slit); 149 acpi_unmap_table(slit); 150 slit = NULL; 151 152 #ifdef VM_NUMA_ALLOC 153 /* Tell the VM about it! */ 154 mem_locality = vm_locality_table; 155 #endif 156 return (0); 157 } 158 159 /* 160 * SRAT parsing. 161 */ 162 163 /* 164 * Returns true if a memory range overlaps with at least one range in 165 * phys_avail[]. 166 */ 167 static int 168 overlaps_phys_avail(vm_paddr_t start, vm_paddr_t end) 169 { 170 int i; 171 172 for (i = 0; phys_avail[i] != 0 && phys_avail[i + 1] != 0; i += 2) { 173 if (phys_avail[i + 1] < start) 174 continue; 175 if (phys_avail[i] < end) 176 return (1); 177 break; 178 } 179 return (0); 180 181 } 182 183 static void 184 srat_parse_entry(ACPI_SUBTABLE_HEADER *entry, void *arg) 185 { 186 ACPI_SRAT_CPU_AFFINITY *cpu; 187 ACPI_SRAT_X2APIC_CPU_AFFINITY *x2apic; 188 ACPI_SRAT_MEM_AFFINITY *mem; 189 int domain, i, slot; 190 191 switch (entry->Type) { 192 case ACPI_SRAT_TYPE_CPU_AFFINITY: 193 cpu = (ACPI_SRAT_CPU_AFFINITY *)entry; 194 domain = cpu->ProximityDomainLo | 195 cpu->ProximityDomainHi[0] << 8 | 196 cpu->ProximityDomainHi[1] << 16 | 197 cpu->ProximityDomainHi[2] << 24; 198 if (bootverbose) 199 printf("SRAT: Found CPU APIC ID %u domain %d: %s\n", 200 cpu->ApicId, domain, 201 (cpu->Flags & ACPI_SRAT_CPU_ENABLED) ? 202 "enabled" : "disabled"); 203 if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED)) 204 break; 205 if (cpus[cpu->ApicId].enabled) { 206 printf("SRAT: Duplicate local APIC ID %u\n", 207 cpu->ApicId); 208 *(int *)arg = ENXIO; 209 break; 210 } 211 cpus[cpu->ApicId].domain = domain; 212 cpus[cpu->ApicId].enabled = 1; 213 break; 214 case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY: 215 x2apic = (ACPI_SRAT_X2APIC_CPU_AFFINITY *)entry; 216 if (bootverbose) 217 printf("SRAT: Found CPU APIC ID %u domain %d: %s\n", 218 x2apic->ApicId, x2apic->ProximityDomain, 219 (x2apic->Flags & ACPI_SRAT_CPU_ENABLED) ? 220 "enabled" : "disabled"); 221 if (!(x2apic->Flags & ACPI_SRAT_CPU_ENABLED)) 222 break; 223 KASSERT(!cpus[x2apic->ApicId].enabled, 224 ("Duplicate local APIC ID %u", x2apic->ApicId)); 225 cpus[x2apic->ApicId].domain = x2apic->ProximityDomain; 226 cpus[x2apic->ApicId].enabled = 1; 227 break; 228 case ACPI_SRAT_TYPE_MEMORY_AFFINITY: 229 mem = (ACPI_SRAT_MEM_AFFINITY *)entry; 230 if (bootverbose) 231 printf( 232 "SRAT: Found memory domain %d addr %jx len %jx: %s\n", 233 mem->ProximityDomain, (uintmax_t)mem->BaseAddress, 234 (uintmax_t)mem->Length, 235 (mem->Flags & ACPI_SRAT_MEM_ENABLED) ? 236 "enabled" : "disabled"); 237 if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED)) 238 break; 239 if (!overlaps_phys_avail(mem->BaseAddress, 240 mem->BaseAddress + mem->Length)) { 241 printf("SRAT: Ignoring memory at addr %jx\n", 242 (uintmax_t)mem->BaseAddress); 243 break; 244 } 245 if (num_mem == VM_PHYSSEG_MAX) { 246 printf("SRAT: Too many memory regions\n"); 247 *(int *)arg = ENXIO; 248 break; 249 } 250 slot = num_mem; 251 for (i = 0; i < num_mem; i++) { 252 if (mem_info[i].end <= mem->BaseAddress) 253 continue; 254 if (mem_info[i].start < 255 (mem->BaseAddress + mem->Length)) { 256 printf("SRAT: Overlapping memory entries\n"); 257 *(int *)arg = ENXIO; 258 return; 259 } 260 slot = i; 261 } 262 for (i = num_mem; i > slot; i--) 263 mem_info[i] = mem_info[i - 1]; 264 mem_info[slot].start = mem->BaseAddress; 265 mem_info[slot].end = mem->BaseAddress + mem->Length; 266 mem_info[slot].domain = mem->ProximityDomain; 267 num_mem++; 268 break; 269 } 270 } 271 272 /* 273 * Ensure each memory domain has at least one CPU and that each CPU 274 * has at least one memory domain. 275 */ 276 static int 277 check_domains(void) 278 { 279 int found, i, j; 280 281 for (i = 0; i < num_mem; i++) { 282 found = 0; 283 for (j = 0; j <= MAX_APIC_ID; j++) 284 if (cpus[j].enabled && 285 cpus[j].domain == mem_info[i].domain) { 286 cpus[j].has_memory = 1; 287 found++; 288 } 289 if (!found) { 290 printf("SRAT: No CPU found for memory domain %d\n", 291 mem_info[i].domain); 292 return (ENXIO); 293 } 294 } 295 for (i = 0; i <= MAX_APIC_ID; i++) 296 if (cpus[i].enabled && !cpus[i].has_memory) { 297 printf("SRAT: No memory found for CPU %d\n", i); 298 return (ENXIO); 299 } 300 return (0); 301 } 302 303 /* 304 * Check that the SRAT memory regions cover all of the regions in 305 * phys_avail[]. 306 */ 307 static int 308 check_phys_avail(void) 309 { 310 vm_paddr_t address; 311 int i, j; 312 313 /* j is the current offset into phys_avail[]. */ 314 address = phys_avail[0]; 315 j = 0; 316 for (i = 0; i < num_mem; i++) { 317 /* 318 * Consume as many phys_avail[] entries as fit in this 319 * region. 320 */ 321 while (address >= mem_info[i].start && 322 address <= mem_info[i].end) { 323 /* 324 * If we cover the rest of this phys_avail[] entry, 325 * advance to the next entry. 326 */ 327 if (phys_avail[j + 1] <= mem_info[i].end) { 328 j += 2; 329 if (phys_avail[j] == 0 && 330 phys_avail[j + 1] == 0) { 331 return (0); 332 } 333 address = phys_avail[j]; 334 } else 335 address = mem_info[i].end + 1; 336 } 337 } 338 printf("SRAT: No memory region found for %jx - %jx\n", 339 (uintmax_t)phys_avail[j], (uintmax_t)phys_avail[j + 1]); 340 return (ENXIO); 341 } 342 343 /* 344 * Renumber the memory domains to be compact and zero-based if not 345 * already. Returns an error if there are too many domains. 346 */ 347 static int 348 renumber_domains(void) 349 { 350 int i, j, slot; 351 352 /* Enumerate all the domains. */ 353 ndomain = 0; 354 for (i = 0; i < num_mem; i++) { 355 /* See if this domain is already known. */ 356 for (j = 0; j < ndomain; j++) { 357 if (domain_pxm[j] >= mem_info[i].domain) 358 break; 359 } 360 if (j < ndomain && domain_pxm[j] == mem_info[i].domain) 361 continue; 362 363 if (ndomain >= MAXMEMDOM) { 364 ndomain = 1; 365 printf("SRAT: Too many memory domains\n"); 366 return (EFBIG); 367 } 368 369 /* Insert the new domain at slot 'j'. */ 370 slot = j; 371 for (j = ndomain; j > slot; j--) 372 domain_pxm[j] = domain_pxm[j - 1]; 373 domain_pxm[slot] = mem_info[i].domain; 374 ndomain++; 375 } 376 377 /* Renumber each domain to its index in the sorted 'domain_pxm' list. */ 378 for (i = 0; i < ndomain; i++) { 379 /* 380 * If the domain is already the right value, no need 381 * to renumber. 382 */ 383 if (domain_pxm[i] == i) 384 continue; 385 386 /* Walk the cpu[] and mem_info[] arrays to renumber. */ 387 for (j = 0; j < num_mem; j++) 388 if (mem_info[j].domain == domain_pxm[i]) 389 mem_info[j].domain = i; 390 for (j = 0; j <= MAX_APIC_ID; j++) 391 if (cpus[j].enabled && cpus[j].domain == domain_pxm[i]) 392 cpus[j].domain = i; 393 } 394 395 return (0); 396 } 397 398 /* 399 * Look for an ACPI System Resource Affinity Table ("SRAT") 400 */ 401 static int 402 parse_srat(void) 403 { 404 int error; 405 406 if (resource_disabled("srat", 0)) 407 return (-1); 408 409 srat_physaddr = acpi_find_table(ACPI_SIG_SRAT); 410 if (srat_physaddr == 0) 411 return (-1); 412 413 /* 414 * Make a pass over the table to populate the cpus[] and 415 * mem_info[] tables. 416 */ 417 srat = acpi_map_table(srat_physaddr, ACPI_SIG_SRAT); 418 error = 0; 419 srat_walk_table(srat_parse_entry, &error); 420 acpi_unmap_table(srat); 421 srat = NULL; 422 if (error || check_domains() != 0 || check_phys_avail() != 0 || 423 renumber_domains() != 0) { 424 srat_physaddr = 0; 425 return (-1); 426 } 427 428 #ifdef VM_NUMA_ALLOC 429 /* Point vm_phys at our memory affinity table. */ 430 vm_ndomains = ndomain; 431 mem_affinity = mem_info; 432 #endif 433 434 return (0); 435 } 436 437 static void 438 init_mem_locality(void) 439 { 440 int i; 441 442 /* 443 * For now, assume -1 == "no locality information for 444 * this pairing. 445 */ 446 for (i = 0; i < MAXMEMDOM * MAXMEMDOM; i++) 447 vm_locality_table[i] = -1; 448 } 449 450 static void 451 parse_acpi_tables(void *dummy) 452 { 453 454 if (parse_srat() < 0) 455 return; 456 init_mem_locality(); 457 (void) parse_slit(); 458 } 459 SYSINIT(parse_acpi_tables, SI_SUB_VM - 1, SI_ORDER_FIRST, parse_acpi_tables, 460 NULL); 461 462 static void 463 srat_walk_table(acpi_subtable_handler *handler, void *arg) 464 { 465 466 acpi_walk_subtables(srat + 1, (char *)srat + srat->Header.Length, 467 handler, arg); 468 } 469 470 /* 471 * Setup per-CPU domain IDs. 472 */ 473 static void 474 srat_set_cpus(void *dummy) 475 { 476 struct cpu_info *cpu; 477 struct pcpu *pc; 478 u_int i; 479 480 if (srat_physaddr == 0) 481 return; 482 for (i = 0; i < MAXCPU; i++) { 483 if (CPU_ABSENT(i)) 484 continue; 485 pc = pcpu_find(i); 486 KASSERT(pc != NULL, ("no pcpu data for CPU %u", i)); 487 cpu = &cpus[pc->pc_apic_id]; 488 if (!cpu->enabled) 489 panic("SRAT: CPU with APIC ID %u is not known", 490 pc->pc_apic_id); 491 pc->pc_domain = cpu->domain; 492 CPU_SET(i, &cpuset_domain[cpu->domain]); 493 if (bootverbose) 494 printf("SRAT: CPU %u has memory domain %d\n", i, 495 cpu->domain); 496 } 497 } 498 SYSINIT(srat_set_cpus, SI_SUB_CPU, SI_ORDER_ANY, srat_set_cpus, NULL); 499 500 /* 501 * Map a _PXM value to a VM domain ID. 502 * 503 * Returns the domain ID, or -1 if no domain ID was found. 504 */ 505 int 506 acpi_map_pxm_to_vm_domainid(int pxm) 507 { 508 int i; 509 510 for (i = 0; i < ndomain; i++) { 511 if (domain_pxm[i] == pxm) 512 return (i); 513 } 514 515 return (-1); 516 } 517 518 #else /* MAXMEMDOM == 1 */ 519 520 int 521 acpi_map_pxm_to_vm_domainid(int pxm) 522 { 523 524 return (-1); 525 } 526 527 #endif /* MAXMEMDOM > 1 */ 528