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