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