sys/x86/acpica/srat.c

   1 /*-
   2  * Copyright (c) 2010 Advanced Computing Technologies 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/smp.h>
  37 #include <vm/vm.h>
  38 #include <vm/pmap.h>
  39 #include <vm/vm_param.h>
  40 #include <vm/vm_phys.h>
  41
  42 #include <contrib/dev/acpica/include/acpi.h>
  43 #include <contrib/dev/acpica/include/actables.h>
  44
  45 #include <machine/intr_machdep.h>
  46 #include <machine/apicvar.h>
  47
  48 #include <dev/acpica/acpivar.h>
  49
  50 #if VM_NDOMAIN > 1
  51 struct cpu_info {
  52         int enabled:1;
  53         int has_memory:1;
  54         int domain;
  55 } cpus[MAX_APIC_ID + 1];
  56
  57 struct mem_affinity mem_info[VM_PHYSSEG_MAX + 1];
  58 int num_mem;
  59
  60 static ACPI_TABLE_SRAT *srat;
  61 static vm_paddr_t srat_physaddr;
  62
  63 static void     srat_walk_table(acpi_subtable_handler *handler, void *arg);
  64
  65 /*
  66  * Returns true if a memory range overlaps with at least one range in
  67  * phys_avail[].
  68  */
  69 static int
  70 overlaps_phys_avail(vm_paddr_t start, vm_paddr_t end)
  71 {
  72         int i;
  73
  74         for (i = 0; phys_avail[i] != 0 && phys_avail[i + 1] != 0; i += 2) {
  75                 if (phys_avail[i + 1] < start)
  76                         continue;
  77                 if (phys_avail[i] < end)
  78                         return (1);
  79                 break;
  80         }
  81         return (0);
  82
  83 }
  84
  85 static void
  86 srat_parse_entry(ACPI_SUBTABLE_HEADER *entry, void *arg)
  87 {
  88         ACPI_SRAT_CPU_AFFINITY *cpu;
  89         ACPI_SRAT_X2APIC_CPU_AFFINITY *x2apic;
  90         ACPI_SRAT_MEM_AFFINITY *mem;
  91         int domain, i, slot;
  92
  93         switch (entry->Type) {
  94         case ACPI_SRAT_TYPE_CPU_AFFINITY:
  95                 cpu = (ACPI_SRAT_CPU_AFFINITY *)entry;
  96                 domain = cpu->ProximityDomainLo |
  97                     cpu->ProximityDomainHi[0] << 8 |
  98                     cpu->ProximityDomainHi[1] << 16 |
  99                     cpu->ProximityDomainHi[2] << 24;
 100                 if (bootverbose)
 101                         printf("SRAT: Found CPU APIC ID %u domain %d: %s\n",
 102                             cpu->ApicId, domain,
 103                             (cpu->Flags & ACPI_SRAT_CPU_ENABLED) ?
 104                             "enabled" : "disabled");
 105                 if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED))
 106                         break;
 107                 KASSERT(!cpus[cpu->ApicId].enabled,
 108                     ("Duplicate local APIC ID %u", cpu->ApicId));
 109                 cpus[cpu->ApicId].domain = domain;
 110                 cpus[cpu->ApicId].enabled = 1;
 111                 break;
 112         case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY:
 113                 x2apic = (ACPI_SRAT_X2APIC_CPU_AFFINITY *)entry;
 114                 if (bootverbose)
 115                         printf("SRAT: Found CPU APIC ID %u domain %d: %s\n",
 116                             x2apic->ApicId, x2apic->ProximityDomain,
 117                             (x2apic->Flags & ACPI_SRAT_CPU_ENABLED) ?
 118                             "enabled" : "disabled");
 119                 if (!(x2apic->Flags & ACPI_SRAT_CPU_ENABLED))
 120                         break;
 121                 KASSERT(!cpus[x2apic->ApicId].enabled,
 122                     ("Duplicate local APIC ID %u", x2apic->ApicId));
 123                 cpus[x2apic->ApicId].domain = x2apic->ProximityDomain;
 124                 cpus[x2apic->ApicId].enabled = 1;
 125                 break;
 126         case ACPI_SRAT_TYPE_MEMORY_AFFINITY:
 127                 mem = (ACPI_SRAT_MEM_AFFINITY *)entry;
 128                 if (bootverbose)
 129                         printf(
 130                     "SRAT: Found memory domain %d addr %jx len %jx: %s\n",
 131                             mem->ProximityDomain, (uintmax_t)mem->BaseAddress,
 132                             (uintmax_t)mem->Length,
 133                             (mem->Flags & ACPI_SRAT_MEM_ENABLED) ?
 134                             "enabled" : "disabled");
 135                 if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED))
 136                         break;
 137                 if (!overlaps_phys_avail(mem->BaseAddress,
 138                     mem->BaseAddress + mem->Length)) {
 139                         printf("SRAT: Ignoring memory at addr %jx\n",
 140                             (uintmax_t)mem->BaseAddress);
 141                         break;
 142                 }
 143                 if (num_mem == VM_PHYSSEG_MAX) {
 144                         printf("SRAT: Too many memory regions\n");
 145                         *(int *)arg = ENXIO;
 146                         break;
 147                 }
 148                 slot = num_mem;
 149                 for (i = 0; i < num_mem; i++) {
 150                         if (mem_info[i].end <= mem->BaseAddress)
 151                                 continue;
 152                         if (mem_info[i].start <
 153                             (mem->BaseAddress + mem->Length)) {
 154                                 printf("SRAT: Overlapping memory entries\n");
 155                                 *(int *)arg = ENXIO;
 156                                 return;
 157                         }
 158                         slot = i;
 159                 }
 160                 for (i = num_mem; i > slot; i--)
 161                         mem_info[i] = mem_info[i - 1];
 162                 mem_info[slot].start = mem->BaseAddress;
 163                 mem_info[slot].end = mem->BaseAddress + mem->Length;
 164                 mem_info[slot].domain = mem->ProximityDomain;
 165                 num_mem++;
 166                 break;
 167         }
 168 }
 169
 170 /*
 171  * Ensure each memory domain has at least one CPU and that each CPU
 172  * has at least one memory domain.
 173  */
 174 static int
 175 check_domains(void)
 176 {
 177         int found, i, j;
 178
 179         for (i = 0; i < num_mem; i++) {
 180                 found = 0;
 181                 for (j = 0; j <= MAX_APIC_ID; j++)
 182                         if (cpus[j].enabled &&
 183                             cpus[j].domain == mem_info[i].domain) {
 184                                 cpus[j].has_memory = 1;
 185                                 found++;
 186                         }
 187                 if (!found) {
 188                         printf("SRAT: No CPU found for memory domain %d\n",
 189                             mem_info[i].domain);
 190                         return (ENXIO);
 191                 }
 192         }
 193         for (i = 0; i <= MAX_APIC_ID; i++)
 194                 if (cpus[i].enabled && !cpus[i].has_memory) {
 195                         printf("SRAT: No memory found for CPU %d\n", i);
 196                         return (ENXIO);
 197                 }
 198         return (0);
 199 }
 200
 201 /*
 202  * Check that the SRAT memory regions cover all of the regions in
 203  * phys_avail[].
 204  */
 205 static int
 206 check_phys_avail(void)
 207 {
 208         vm_paddr_t address;
 209         int i, j;
 210
 211         /* j is the current offset into phys_avail[]. */
 212         address = phys_avail[0];
 213         j = 0;
 214         for (i = 0; i < num_mem; i++) {
 215                 /*
 216                  * Consume as many phys_avail[] entries as fit in this
 217                  * region.
 218                  */
 219                 while (address >= mem_info[i].start &&
 220                     address <= mem_info[i].end) {
 221                         /*
 222                          * If we cover the rest of this phys_avail[] entry,
 223                          * advance to the next entry.
 224                          */
 225                         if (phys_avail[j + 1] <= mem_info[i].end) {
 226                                 j += 2;
 227                                 if (phys_avail[j] == 0 &&
 228                                     phys_avail[j + 1] == 0) {
 229                                         return (0);
 230                                 }
 231                                 address = phys_avail[j];
 232                         } else
 233                                 address = mem_info[i].end + 1;
 234                 }
 235         }
 236         printf("SRAT: No memory region found for %jx - %jx\n",
 237             (uintmax_t)phys_avail[j], (uintmax_t)phys_avail[j + 1]);
 238         return (ENXIO);
 239 }
 240
 241 /*
 242  * Renumber the memory domains to be compact and zero-based if not
 243  * already.  Returns an error if there are too many domains.
 244  */
 245 static int
 246 renumber_domains(void)
 247 {
 248         int domains[VM_PHYSSEG_MAX];
 249         int ndomain, i, j, slot;
 250
 251         /* Enumerate all the domains. */
 252         ndomain = 0;
 253         for (i = 0; i < num_mem; i++) {
 254                 /* See if this domain is already known. */
 255                 for (j = 0; j < ndomain; j++) {
 256                         if (domains[j] >= mem_info[i].domain)
 257                                 break;
 258                 }
 259                 if (j < ndomain && domains[j] == mem_info[i].domain)
 260                         continue;
 261
 262                 /* Insert the new domain at slot 'j'. */
 263                 slot = j;
 264                 for (j = ndomain; j > slot; j--)
 265                         domains[j] = domains[j - 1];
 266                 domains[slot] = mem_info[i].domain;
 267                 ndomain++;
 268                 if (ndomain > VM_NDOMAIN) {
 269                         printf("SRAT: Too many memory domains\n");
 270                         return (EFBIG);
 271                 }
 272         }
 273
 274         /* Renumber each domain to its index in the sorted 'domains' list. */
 275         for (i = 0; i < ndomain; i++) {
 276                 /*
 277                  * If the domain is already the right value, no need
 278                  * to renumber.
 279                  */
 280                 if (domains[i] == i)
 281                         continue;
 282
 283                 /* Walk the cpu[] and mem_info[] arrays to renumber. */
 284                 for (j = 0; j < num_mem; j++)
 285                         if (mem_info[j].domain == domains[i])
 286                                 mem_info[j].domain = i;
 287                 for (j = 0; j <= MAX_APIC_ID; j++)
 288                         if (cpus[j].enabled && cpus[j].domain == domains[i])
 289                                 cpus[j].domain = i;
 290         }
 291         return (0);
 292 }
 293
 294 /*
 295  * Look for an ACPI System Resource Affinity Table ("SRAT")
 296  */
 297 static void
 298 parse_srat(void *dummy)
 299 {
 300         int error;
 301
 302         if (resource_disabled("srat", 0))
 303                 return;
 304
 305         srat_physaddr = acpi_find_table(ACPI_SIG_SRAT);
 306         if (srat_physaddr == 0)
 307                 return;
 308
 309         /*
 310          * Make a pass over the table to populate the cpus[] and
 311          * mem_info[] tables.
 312          */
 313         srat = acpi_map_table(srat_physaddr, ACPI_SIG_SRAT);
 314         error = 0;
 315         srat_walk_table(srat_parse_entry, &error);
 316         acpi_unmap_table(srat);
 317         srat = NULL;
 318         if (error || check_domains() != 0 || check_phys_avail() != 0 ||
 319             renumber_domains() != 0) {
 320                 srat_physaddr = 0;
 321                 return;
 322         }
 323
 324         /* Point vm_phys at our memory affinity table. */
 325         mem_affinity = mem_info;
 326 }
 327 SYSINIT(parse_srat, SI_SUB_VM - 1, SI_ORDER_FIRST, parse_srat, NULL);
 328
 329 static void
 330 srat_walk_table(acpi_subtable_handler *handler, void *arg)
 331 {
 332
 333         acpi_walk_subtables(srat + 1, (char *)srat + srat->Header.Length,
 334             handler, arg);
 335 }
 336
 337 /*
 338  * Setup per-CPU ACPI IDs.
 339  */
 340 static void
 341 srat_set_cpus(void *dummy)
 342 {
 343         struct cpu_info *cpu;
 344         struct pcpu *pc;
 345         u_int i;
 346
 347         if (srat_physaddr == 0)
 348                 return;
 349         for (i = 0; i < MAXCPU; i++) {
 350                 if (CPU_ABSENT(i))
 351                         continue;
 352                 pc = pcpu_find(i);
 353                 KASSERT(pc != NULL, ("no pcpu data for CPU %u", i));
 354                 cpu = &cpus[pc->pc_apic_id];
 355                 if (!cpu->enabled)
 356                         panic("SRAT: CPU with APIC ID %u is not known",
 357                             pc->pc_apic_id);
 358                 pc->pc_domain = cpu->domain;
 359                 if (bootverbose)
 360                         printf("SRAT: CPU %u has memory domain %d\n", i,
 361                             cpu->domain);
 362         }
 363 }
 364 SYSINIT(srat_set_cpus, SI_SUB_CPU, SI_ORDER_ANY, srat_set_cpus, NULL);
 365 #endif /* VM_NDOMAIN > 1 */