sys/x86/acpica/srat.c

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