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 <sys/param.h>
  32 #include <sys/bus.h>
  33 #include <sys/kernel.h>
  34 #include <sys/lock.h>
  35 #include <sys/mutex.h>
  36 #include <sys/smp.h>
  37 #include <sys/vmmeter.h>
  38 #include <vm/vm.h>
  39 #include <vm/pmap.h>
  40 #include <vm/vm_param.h>
  41 #include <vm/vm_page.h>
  42 #include <vm/vm_phys.h>
  43
  44 #include <contrib/dev/acpica/include/acpi.h>
  45 #include <contrib/dev/acpica/include/actables.h>
  46
  47 #include <machine/intr_machdep.h>
  48 #include <machine/apicvar.h>
  49
  50 #include <dev/acpica/acpivar.h>
  51
  52 #if MAXMEMDOM > 1
  53 struct cpu_info {
  54         int enabled:1;
  55         int has_memory:1;
  56         int domain;
  57 } cpus[MAX_APIC_ID + 1];
  58
  59 struct mem_affinity mem_info[VM_PHYSSEG_MAX + 1];
  60 int num_mem;
  61
  62 static ACPI_TABLE_SRAT *srat;
  63 static vm_paddr_t srat_physaddr;
  64
  65 static int vm_domains[VM_PHYSSEG_MAX];
  66
  67 static void     srat_walk_table(acpi_subtable_handler *handler, void *arg);
  68
  69 /*
  70  * Returns true if a memory range overlaps with at least one range in
  71  * phys_avail[].
  72  */
  73 static int
  74 overlaps_phys_avail(vm_paddr_t start, vm_paddr_t end)
  75 {
  76         int i;
  77
  78         for (i = 0; phys_avail[i] != 0 && phys_avail[i + 1] != 0; i += 2) {
  79                 if (phys_avail[i + 1] < start)
  80                         continue;
  81                 if (phys_avail[i] < end)
  82                         return (1);
  83                 break;
  84         }
  85         return (0);
  86
  87 }
  88
  89 static void
  90 srat_parse_entry(ACPI_SUBTABLE_HEADER *entry, void *arg)
  91 {
  92         ACPI_SRAT_CPU_AFFINITY *cpu;
  93         ACPI_SRAT_X2APIC_CPU_AFFINITY *x2apic;
  94         ACPI_SRAT_MEM_AFFINITY *mem;
  95         int domain, i, slot;
  96
  97         switch (entry->Type) {
  98         case ACPI_SRAT_TYPE_CPU_AFFINITY:
  99                 cpu = (ACPI_SRAT_CPU_AFFINITY *)entry;
 100                 domain = cpu->ProximityDomainLo |
 101                     cpu->ProximityDomainHi[0] << 8 |
 102                     cpu->ProximityDomainHi[1] << 16 |
 103                     cpu->ProximityDomainHi[2] << 24;
 104                 if (bootverbose)
 105                         printf("SRAT: Found CPU APIC ID %u domain %d: %s\n",
 106                             cpu->ApicId, domain,
 107                             (cpu->Flags & ACPI_SRAT_CPU_ENABLED) ?
 108                             "enabled" : "disabled");
 109                 if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED))
 110                         break;
 111                 KASSERT(!cpus[cpu->ApicId].enabled,
 112                     ("Duplicate local APIC ID %u", cpu->ApicId));
 113                 cpus[cpu->ApicId].domain = domain;
 114                 cpus[cpu->ApicId].enabled = 1;
 115                 break;
 116         case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY:
 117                 x2apic = (ACPI_SRAT_X2APIC_CPU_AFFINITY *)entry;
 118                 if (bootverbose)
 119                         printf("SRAT: Found CPU APIC ID %u domain %d: %s\n",
 120                             x2apic->ApicId, x2apic->ProximityDomain,
 121                             (x2apic->Flags & ACPI_SRAT_CPU_ENABLED) ?
 122                             "enabled" : "disabled");
 123                 if (!(x2apic->Flags & ACPI_SRAT_CPU_ENABLED))
 124                         break;
 125                 KASSERT(!cpus[x2apic->ApicId].enabled,
 126                     ("Duplicate local APIC ID %u", x2apic->ApicId));
 127                 cpus[x2apic->ApicId].domain = x2apic->ProximityDomain;
 128                 cpus[x2apic->ApicId].enabled = 1;
 129                 break;
 130         case ACPI_SRAT_TYPE_MEMORY_AFFINITY:
 131                 mem = (ACPI_SRAT_MEM_AFFINITY *)entry;
 132                 if (bootverbose)
 133                         printf(
 134                     "SRAT: Found memory domain %d addr %jx len %jx: %s\n",
 135                             mem->ProximityDomain, (uintmax_t)mem->BaseAddress,
 136                             (uintmax_t)mem->Length,
 137                             (mem->Flags & ACPI_SRAT_MEM_ENABLED) ?
 138                             "enabled" : "disabled");
 139                 if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED))
 140                         break;
 141                 if (!overlaps_phys_avail(mem->BaseAddress,
 142                     mem->BaseAddress + mem->Length)) {
 143                         printf("SRAT: Ignoring memory at addr %jx\n",
 144                             (uintmax_t)mem->BaseAddress);
 145                         break;
 146                 }
 147                 if (num_mem == VM_PHYSSEG_MAX) {
 148                         printf("SRAT: Too many memory regions\n");
 149                         *(int *)arg = ENXIO;
 150                         break;
 151                 }
 152                 slot = num_mem;
 153                 for (i = 0; i < num_mem; i++) {
 154                         if (mem_info[i].end <= mem->BaseAddress)
 155                                 continue;
 156                         if (mem_info[i].start <
 157                             (mem->BaseAddress + mem->Length)) {
 158                                 printf("SRAT: Overlapping memory entries\n");
 159                                 *(int *)arg = ENXIO;
 160                                 return;
 161                         }
 162                         slot = i;
 163                 }
 164                 for (i = num_mem; i > slot; i--)
 165                         mem_info[i] = mem_info[i - 1];
 166                 mem_info[slot].start = mem->BaseAddress;
 167                 mem_info[slot].end = mem->BaseAddress + mem->Length;
 168                 mem_info[slot].domain = mem->ProximityDomain;
 169                 num_mem++;
 170                 break;
 171         }
 172 }
 173
 174 /*
 175  * Ensure each memory domain has at least one CPU and that each CPU
 176  * has at least one memory domain.
 177  */
 178 static int
 179 check_domains(void)
 180 {
 181         int found, i, j;
 182
 183         for (i = 0; i < num_mem; i++) {
 184                 found = 0;
 185                 for (j = 0; j <= MAX_APIC_ID; j++)
 186                         if (cpus[j].enabled &&
 187                             cpus[j].domain == mem_info[i].domain) {
 188                                 cpus[j].has_memory = 1;
 189                                 found++;
 190                         }
 191                 if (!found) {
 192                         printf("SRAT: No CPU found for memory domain %d\n",
 193                             mem_info[i].domain);
 194                         return (ENXIO);
 195                 }
 196         }
 197         for (i = 0; i <= MAX_APIC_ID; i++)
 198                 if (cpus[i].enabled && !cpus[i].has_memory) {
 199                         printf("SRAT: No memory found for CPU %d\n", i);
 200                         return (ENXIO);
 201                 }
 202         return (0);
 203 }
 204
 205 /*
 206  * Check that the SRAT memory regions cover all of the regions in
 207  * phys_avail[].
 208  */
 209 static int
 210 check_phys_avail(void)
 211 {
 212         vm_paddr_t address;
 213         int i, j;
 214
 215         /* j is the current offset into phys_avail[]. */
 216         address = phys_avail[0];
 217         j = 0;
 218         for (i = 0; i < num_mem; i++) {
 219                 /*
 220                  * Consume as many phys_avail[] entries as fit in this
 221                  * region.
 222                  */
 223                 while (address >= mem_info[i].start &&
 224                     address <= mem_info[i].end) {
 225                         /*
 226                          * If we cover the rest of this phys_avail[] entry,
 227                          * advance to the next entry.
 228                          */
 229                         if (phys_avail[j + 1] <= mem_info[i].end) {
 230                                 j += 2;
 231                                 if (phys_avail[j] == 0 &&
 232                                     phys_avail[j + 1] == 0) {
 233                                         return (0);
 234                                 }
 235                                 address = phys_avail[j];
 236                         } else
 237                                 address = mem_info[i].end + 1;
 238                 }
 239         }
 240         printf("SRAT: No memory region found for %jx - %jx\n",
 241             (uintmax_t)phys_avail[j], (uintmax_t)phys_avail[j + 1]);
 242         return (ENXIO);
 243 }
 244
 245 /*
 246  * Renumber the memory domains to be compact and zero-based if not
 247  * already.  Returns an error if there are too many domains.
 248  */
 249 static int
 250 renumber_domains(void)
 251 {
 252         int i, j, slot;
 253
 254         /* Enumerate all the domains. */
 255         vm_ndomains = 0;
 256         for (i = 0; i < num_mem; i++) {
 257                 /* See if this domain is already known. */
 258                 for (j = 0; j < vm_ndomains; j++) {
 259                         if (vm_domains[j] >= mem_info[i].domain)
 260                                 break;
 261                 }
 262                 if (j < vm_ndomains && vm_domains[j] == mem_info[i].domain)
 263                         continue;
 264
 265                 /* Insert the new domain at slot 'j'. */
 266                 slot = j;
 267                 for (j = vm_ndomains; j > slot; j--)
 268                         vm_domains[j] = vm_domains[j - 1];
 269                 vm_domains[slot] = mem_info[i].domain;
 270                 vm_ndomains++;
 271                 if (vm_ndomains > MAXMEMDOM) {
 272                         vm_ndomains = 1;
 273                         printf("SRAT: Too many memory domains\n");
 274                         return (EFBIG);
 275                 }
 276         }
 277
 278         /* Renumber each domain to its index in the sorted 'domains' list. */
 279         for (i = 0; i < vm_ndomains; i++) {
 280                 /*
 281                  * If the domain is already the right value, no need
 282                  * to renumber.
 283                  */
 284                 if (vm_domains[i] == i)
 285                         continue;
 286
 287                 /* Walk the cpu[] and mem_info[] arrays to renumber. */
 288                 for (j = 0; j < num_mem; j++)
 289                         if (mem_info[j].domain == vm_domains[i])
 290                                 mem_info[j].domain = i;
 291                 for (j = 0; j <= MAX_APIC_ID; j++)
 292                         if (cpus[j].enabled && cpus[j].domain == vm_domains[i])
 293                                 cpus[j].domain = i;
 294         }
 295         KASSERT(vm_ndomains > 0,
 296             ("renumber_domains: invalid final vm_ndomains setup"));
 297
 298         return (0);
 299 }
 300
 301 /*
 302  * Look for an ACPI System Resource Affinity Table ("SRAT")
 303  */
 304 static void
 305 parse_srat(void *dummy)
 306 {
 307         int error;
 308
 309         if (resource_disabled("srat", 0))
 310                 return;
 311
 312         srat_physaddr = acpi_find_table(ACPI_SIG_SRAT);
 313         if (srat_physaddr == 0)
 314                 return;
 315
 316         /*
 317          * Make a pass over the table to populate the cpus[] and
 318          * mem_info[] tables.
 319          */
 320         srat = acpi_map_table(srat_physaddr, ACPI_SIG_SRAT);
 321         error = 0;
 322         srat_walk_table(srat_parse_entry, &error);
 323         acpi_unmap_table(srat);
 324         srat = NULL;
 325         if (error || check_domains() != 0 || check_phys_avail() != 0 ||
 326             renumber_domains() != 0) {
 327                 srat_physaddr = 0;
 328                 return;
 329         }
 330
 331         /* Point vm_phys at our memory affinity table. */
 332         mem_affinity = mem_info;
 333 }
 334 SYSINIT(parse_srat, SI_SUB_VM - 1, SI_ORDER_FIRST, parse_srat, NULL);
 335
 336 static void
 337 srat_walk_table(acpi_subtable_handler *handler, void *arg)
 338 {
 339
 340         acpi_walk_subtables(srat + 1, (char *)srat + srat->Header.Length,
 341             handler, arg);
 342 }
 343
 344 /*
 345  * Setup per-CPU ACPI IDs.
 346  */
 347 static void
 348 srat_set_cpus(void *dummy)
 349 {
 350         struct cpu_info *cpu;
 351         struct pcpu *pc;
 352         u_int i;
 353
 354         if (srat_physaddr == 0)
 355                 return;
 356         for (i = 0; i < MAXCPU; i++) {
 357                 if (CPU_ABSENT(i))
 358                         continue;
 359                 pc = pcpu_find(i);
 360                 KASSERT(pc != NULL, ("no pcpu data for CPU %u", i));
 361                 cpu = &cpus[pc->pc_apic_id];
 362                 if (!cpu->enabled)
 363                         panic("SRAT: CPU with APIC ID %u is not known",
 364                             pc->pc_apic_id);
 365                 pc->pc_domain = cpu->domain;
 366                 if (bootverbose)
 367                         printf("SRAT: CPU %u has memory domain %d\n", i,
 368                             cpu->domain);
 369         }
 370 }
 371 SYSINIT(srat_set_cpus, SI_SUB_CPU, SI_ORDER_ANY, srat_set_cpus, NULL);
 372
 373 /*
 374  * Map a _PXM value to a VM domain ID.
 375  *
 376  * Returns the domain ID, or -1 if no domain ID was found.
 377  */
 378 int
 379 acpi_map_pxm_to_vm_domainid(int pxm)
 380 {
 381         int i;
 382
 383         for (i = 0; i < vm_ndomains; i++) {
 384                 if (vm_domains[i] == pxm)
 385                         return (i);
 386         }
 387
 388         return (-1);
 389 }
 390
 391 #endif /* MAXMEMDOM > 1 */