1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2019, Intel Corporation.
5 * Heterogeneous Memory Attributes Table (HMAT) representation
7 * This program parses and reports the platform's HMAT tables, and registers
8 * the applicable attributes with the node's interfaces.
11 #define pr_fmt(fmt) "acpi/hmat: " fmt
12 #define dev_fmt(fmt) "acpi/hmat: " fmt
14 #include <linux/acpi.h>
15 #include <linux/bitops.h>
16 #include <linux/device.h>
17 #include <linux/init.h>
18 #include <linux/list.h>
20 #include <linux/platform_device.h>
21 #include <linux/list_sort.h>
22 #include <linux/memregion.h>
23 #include <linux/memory.h>
24 #include <linux/mutex.h>
25 #include <linux/node.h>
26 #include <linux/sysfs.h>
27 #include <linux/dax.h>
29 static u8 hmat_revision;
30 static int hmat_disable __initdata;
32 void __init disable_hmat(void)
37 static LIST_HEAD(targets);
38 static LIST_HEAD(initiators);
39 static LIST_HEAD(localities);
41 static DEFINE_MUTEX(target_lock);
44 * The defined enum order is used to prioritize attributes to break ties when
45 * selecting the best performing node.
54 static struct memory_locality *localities_types[4];
57 struct list_head node;
58 struct node_cache_attrs cache_attrs;
61 struct memory_target {
62 struct list_head node;
63 unsigned int memory_pxm;
64 unsigned int processor_pxm;
65 struct resource memregions;
66 struct node_hmem_attrs hmem_attrs[2];
67 struct list_head caches;
68 struct node_cache_attrs cache_attrs;
72 struct memory_initiator {
73 struct list_head node;
74 unsigned int processor_pxm;
78 struct memory_locality {
79 struct list_head node;
80 struct acpi_hmat_locality *hmat_loc;
83 static struct memory_initiator *find_mem_initiator(unsigned int cpu_pxm)
85 struct memory_initiator *initiator;
87 list_for_each_entry(initiator, &initiators, node)
88 if (initiator->processor_pxm == cpu_pxm)
93 static struct memory_target *find_mem_target(unsigned int mem_pxm)
95 struct memory_target *target;
97 list_for_each_entry(target, &targets, node)
98 if (target->memory_pxm == mem_pxm)
103 static __init void alloc_memory_initiator(unsigned int cpu_pxm)
105 struct memory_initiator *initiator;
107 if (pxm_to_node(cpu_pxm) == NUMA_NO_NODE)
110 initiator = find_mem_initiator(cpu_pxm);
114 initiator = kzalloc(sizeof(*initiator), GFP_KERNEL);
118 initiator->processor_pxm = cpu_pxm;
119 initiator->has_cpu = node_state(pxm_to_node(cpu_pxm), N_CPU);
120 list_add_tail(&initiator->node, &initiators);
123 static __init void alloc_memory_target(unsigned int mem_pxm,
124 resource_size_t start, resource_size_t len)
126 struct memory_target *target;
128 target = find_mem_target(mem_pxm);
130 target = kzalloc(sizeof(*target), GFP_KERNEL);
133 target->memory_pxm = mem_pxm;
134 target->processor_pxm = PXM_INVAL;
135 target->memregions = (struct resource) {
139 .flags = IORESOURCE_MEM,
141 list_add_tail(&target->node, &targets);
142 INIT_LIST_HEAD(&target->caches);
146 * There are potentially multiple ranges per PXM, so record each
147 * in the per-target memregions resource tree.
149 if (!__request_region(&target->memregions, start, len, "memory target",
151 pr_warn("failed to reserve %#llx - %#llx in pxm: %d\n",
152 start, start + len, mem_pxm);
155 static __init const char *hmat_data_type(u8 type)
158 case ACPI_HMAT_ACCESS_LATENCY:
159 return "Access Latency";
160 case ACPI_HMAT_READ_LATENCY:
161 return "Read Latency";
162 case ACPI_HMAT_WRITE_LATENCY:
163 return "Write Latency";
164 case ACPI_HMAT_ACCESS_BANDWIDTH:
165 return "Access Bandwidth";
166 case ACPI_HMAT_READ_BANDWIDTH:
167 return "Read Bandwidth";
168 case ACPI_HMAT_WRITE_BANDWIDTH:
169 return "Write Bandwidth";
175 static __init const char *hmat_data_type_suffix(u8 type)
178 case ACPI_HMAT_ACCESS_LATENCY:
179 case ACPI_HMAT_READ_LATENCY:
180 case ACPI_HMAT_WRITE_LATENCY:
182 case ACPI_HMAT_ACCESS_BANDWIDTH:
183 case ACPI_HMAT_READ_BANDWIDTH:
184 case ACPI_HMAT_WRITE_BANDWIDTH:
191 static u32 hmat_normalize(u16 entry, u64 base, u8 type)
196 * Check for invalid and overflow values
198 if (entry == 0xffff || !entry)
200 else if (base > (UINT_MAX / (entry)))
204 * Divide by the base unit for version 1, convert latency from
205 * picosenonds to nanoseconds if revision 2.
207 value = entry * base;
208 if (hmat_revision == 1) {
211 value = DIV_ROUND_UP(value, 10);
212 } else if (hmat_revision == 2) {
214 case ACPI_HMAT_ACCESS_LATENCY:
215 case ACPI_HMAT_READ_LATENCY:
216 case ACPI_HMAT_WRITE_LATENCY:
217 value = DIV_ROUND_UP(value, 1000);
226 static void hmat_update_target_access(struct memory_target *target,
227 u8 type, u32 value, int access)
230 case ACPI_HMAT_ACCESS_LATENCY:
231 target->hmem_attrs[access].read_latency = value;
232 target->hmem_attrs[access].write_latency = value;
234 case ACPI_HMAT_READ_LATENCY:
235 target->hmem_attrs[access].read_latency = value;
237 case ACPI_HMAT_WRITE_LATENCY:
238 target->hmem_attrs[access].write_latency = value;
240 case ACPI_HMAT_ACCESS_BANDWIDTH:
241 target->hmem_attrs[access].read_bandwidth = value;
242 target->hmem_attrs[access].write_bandwidth = value;
244 case ACPI_HMAT_READ_BANDWIDTH:
245 target->hmem_attrs[access].read_bandwidth = value;
247 case ACPI_HMAT_WRITE_BANDWIDTH:
248 target->hmem_attrs[access].write_bandwidth = value;
255 static __init void hmat_add_locality(struct acpi_hmat_locality *hmat_loc)
257 struct memory_locality *loc;
259 loc = kzalloc(sizeof(*loc), GFP_KERNEL);
261 pr_notice_once("Failed to allocate HMAT locality\n");
265 loc->hmat_loc = hmat_loc;
266 list_add_tail(&loc->node, &localities);
268 switch (hmat_loc->data_type) {
269 case ACPI_HMAT_ACCESS_LATENCY:
270 localities_types[READ_LATENCY] = loc;
271 localities_types[WRITE_LATENCY] = loc;
273 case ACPI_HMAT_READ_LATENCY:
274 localities_types[READ_LATENCY] = loc;
276 case ACPI_HMAT_WRITE_LATENCY:
277 localities_types[WRITE_LATENCY] = loc;
279 case ACPI_HMAT_ACCESS_BANDWIDTH:
280 localities_types[READ_BANDWIDTH] = loc;
281 localities_types[WRITE_BANDWIDTH] = loc;
283 case ACPI_HMAT_READ_BANDWIDTH:
284 localities_types[READ_BANDWIDTH] = loc;
286 case ACPI_HMAT_WRITE_BANDWIDTH:
287 localities_types[WRITE_BANDWIDTH] = loc;
294 static __init int hmat_parse_locality(union acpi_subtable_headers *header,
295 const unsigned long end)
297 struct acpi_hmat_locality *hmat_loc = (void *)header;
298 struct memory_target *target;
299 unsigned int init, targ, total_size, ipds, tpds;
300 u32 *inits, *targs, value;
304 if (hmat_loc->header.length < sizeof(*hmat_loc)) {
305 pr_notice("HMAT: Unexpected locality header length: %u\n",
306 hmat_loc->header.length);
310 type = hmat_loc->data_type;
311 mem_hier = hmat_loc->flags & ACPI_HMAT_MEMORY_HIERARCHY;
312 ipds = hmat_loc->number_of_initiator_Pds;
313 tpds = hmat_loc->number_of_target_Pds;
314 total_size = sizeof(*hmat_loc) + sizeof(*entries) * ipds * tpds +
315 sizeof(*inits) * ipds + sizeof(*targs) * tpds;
316 if (hmat_loc->header.length < total_size) {
317 pr_notice("HMAT: Unexpected locality header length:%u, minimum required:%u\n",
318 hmat_loc->header.length, total_size);
322 pr_info("HMAT: Locality: Flags:%02x Type:%s Initiator Domains:%u Target Domains:%u Base:%lld\n",
323 hmat_loc->flags, hmat_data_type(type), ipds, tpds,
324 hmat_loc->entry_base_unit);
326 inits = (u32 *)(hmat_loc + 1);
327 targs = inits + ipds;
328 entries = (u16 *)(targs + tpds);
329 for (init = 0; init < ipds; init++) {
330 alloc_memory_initiator(inits[init]);
331 for (targ = 0; targ < tpds; targ++) {
332 value = hmat_normalize(entries[init * tpds + targ],
333 hmat_loc->entry_base_unit,
335 pr_info(" Initiator-Target[%u-%u]:%u%s\n",
336 inits[init], targs[targ], value,
337 hmat_data_type_suffix(type));
339 if (mem_hier == ACPI_HMAT_MEMORY) {
340 target = find_mem_target(targs[targ]);
341 if (target && target->processor_pxm == inits[init]) {
342 hmat_update_target_access(target, type, value, 0);
343 /* If the node has a CPU, update access 1 */
344 if (node_state(pxm_to_node(inits[init]), N_CPU))
345 hmat_update_target_access(target, type, value, 1);
351 if (mem_hier == ACPI_HMAT_MEMORY)
352 hmat_add_locality(hmat_loc);
357 static __init int hmat_parse_cache(union acpi_subtable_headers *header,
358 const unsigned long end)
360 struct acpi_hmat_cache *cache = (void *)header;
361 struct memory_target *target;
362 struct target_cache *tcache;
365 if (cache->header.length < sizeof(*cache)) {
366 pr_notice("HMAT: Unexpected cache header length: %u\n",
367 cache->header.length);
371 attrs = cache->cache_attributes;
372 pr_info("HMAT: Cache: Domain:%u Size:%llu Attrs:%08x SMBIOS Handles:%d\n",
373 cache->memory_PD, cache->cache_size, attrs,
374 cache->number_of_SMBIOShandles);
376 target = find_mem_target(cache->memory_PD);
380 tcache = kzalloc(sizeof(*tcache), GFP_KERNEL);
382 pr_notice_once("Failed to allocate HMAT cache info\n");
386 tcache->cache_attrs.size = cache->cache_size;
387 tcache->cache_attrs.level = (attrs & ACPI_HMAT_CACHE_LEVEL) >> 4;
388 tcache->cache_attrs.line_size = (attrs & ACPI_HMAT_CACHE_LINE_SIZE) >> 16;
390 switch ((attrs & ACPI_HMAT_CACHE_ASSOCIATIVITY) >> 8) {
391 case ACPI_HMAT_CA_DIRECT_MAPPED:
392 tcache->cache_attrs.indexing = NODE_CACHE_DIRECT_MAP;
394 case ACPI_HMAT_CA_COMPLEX_CACHE_INDEXING:
395 tcache->cache_attrs.indexing = NODE_CACHE_INDEXED;
397 case ACPI_HMAT_CA_NONE:
399 tcache->cache_attrs.indexing = NODE_CACHE_OTHER;
403 switch ((attrs & ACPI_HMAT_WRITE_POLICY) >> 12) {
404 case ACPI_HMAT_CP_WB:
405 tcache->cache_attrs.write_policy = NODE_CACHE_WRITE_BACK;
407 case ACPI_HMAT_CP_WT:
408 tcache->cache_attrs.write_policy = NODE_CACHE_WRITE_THROUGH;
410 case ACPI_HMAT_CP_NONE:
412 tcache->cache_attrs.write_policy = NODE_CACHE_WRITE_OTHER;
415 list_add_tail(&tcache->node, &target->caches);
420 static int __init hmat_parse_proximity_domain(union acpi_subtable_headers *header,
421 const unsigned long end)
423 struct acpi_hmat_proximity_domain *p = (void *)header;
424 struct memory_target *target = NULL;
426 if (p->header.length != sizeof(*p)) {
427 pr_notice("HMAT: Unexpected address range header length: %u\n",
432 if (hmat_revision == 1)
433 pr_info("HMAT: Memory (%#llx length %#llx) Flags:%04x Processor Domain:%u Memory Domain:%u\n",
434 p->reserved3, p->reserved4, p->flags, p->processor_PD,
437 pr_info("HMAT: Memory Flags:%04x Processor Domain:%u Memory Domain:%u\n",
438 p->flags, p->processor_PD, p->memory_PD);
440 if ((hmat_revision == 1 && p->flags & ACPI_HMAT_MEMORY_PD_VALID) ||
442 target = find_mem_target(p->memory_PD);
444 pr_debug("HMAT: Memory Domain missing from SRAT\n");
448 if (target && p->flags & ACPI_HMAT_PROCESSOR_PD_VALID) {
449 int p_node = pxm_to_node(p->processor_PD);
451 if (p_node == NUMA_NO_NODE) {
452 pr_debug("HMAT: Invalid Processor Domain\n");
455 target->processor_pxm = p->processor_PD;
461 static int __init hmat_parse_subtable(union acpi_subtable_headers *header,
462 const unsigned long end)
464 struct acpi_hmat_structure *hdr = (void *)header;
470 case ACPI_HMAT_TYPE_PROXIMITY:
471 return hmat_parse_proximity_domain(header, end);
472 case ACPI_HMAT_TYPE_LOCALITY:
473 return hmat_parse_locality(header, end);
474 case ACPI_HMAT_TYPE_CACHE:
475 return hmat_parse_cache(header, end);
481 static __init int srat_parse_mem_affinity(union acpi_subtable_headers *header,
482 const unsigned long end)
484 struct acpi_srat_mem_affinity *ma = (void *)header;
488 if (!(ma->flags & ACPI_SRAT_MEM_ENABLED))
490 alloc_memory_target(ma->proximity_domain, ma->base_address, ma->length);
494 static u32 hmat_initiator_perf(struct memory_target *target,
495 struct memory_initiator *initiator,
496 struct acpi_hmat_locality *hmat_loc)
498 unsigned int ipds, tpds, i, idx = 0, tdx = 0;
502 ipds = hmat_loc->number_of_initiator_Pds;
503 tpds = hmat_loc->number_of_target_Pds;
504 inits = (u32 *)(hmat_loc + 1);
505 targs = inits + ipds;
506 entries = (u16 *)(targs + tpds);
508 for (i = 0; i < ipds; i++) {
509 if (inits[i] == initiator->processor_pxm) {
518 for (i = 0; i < tpds; i++) {
519 if (targs[i] == target->memory_pxm) {
527 return hmat_normalize(entries[idx * tpds + tdx],
528 hmat_loc->entry_base_unit,
529 hmat_loc->data_type);
532 static bool hmat_update_best(u8 type, u32 value, u32 *best)
534 bool updated = false;
540 case ACPI_HMAT_ACCESS_LATENCY:
541 case ACPI_HMAT_READ_LATENCY:
542 case ACPI_HMAT_WRITE_LATENCY:
543 if (!*best || *best > value) {
548 case ACPI_HMAT_ACCESS_BANDWIDTH:
549 case ACPI_HMAT_READ_BANDWIDTH:
550 case ACPI_HMAT_WRITE_BANDWIDTH:
551 if (!*best || *best < value) {
561 static int initiator_cmp(void *priv, const struct list_head *a,
562 const struct list_head *b)
564 struct memory_initiator *ia;
565 struct memory_initiator *ib;
566 unsigned long *p_nodes = priv;
568 ia = list_entry(a, struct memory_initiator, node);
569 ib = list_entry(b, struct memory_initiator, node);
571 set_bit(ia->processor_pxm, p_nodes);
572 set_bit(ib->processor_pxm, p_nodes);
574 return ia->processor_pxm - ib->processor_pxm;
577 static void hmat_register_target_initiators(struct memory_target *target)
579 static DECLARE_BITMAP(p_nodes, MAX_NUMNODES);
580 struct memory_initiator *initiator;
581 unsigned int mem_nid, cpu_nid;
582 struct memory_locality *loc = NULL;
584 bool access0done = false;
587 mem_nid = pxm_to_node(target->memory_pxm);
589 * If the Address Range Structure provides a local processor pxm, link
590 * only that one. Otherwise, find the best performance attributes and
591 * register all initiators that match.
593 if (target->processor_pxm != PXM_INVAL) {
594 cpu_nid = pxm_to_node(target->processor_pxm);
595 register_memory_node_under_compute_node(mem_nid, cpu_nid, 0);
597 if (node_state(cpu_nid, N_CPU)) {
598 register_memory_node_under_compute_node(mem_nid, cpu_nid, 1);
603 if (list_empty(&localities))
607 * We need the initiator list sorted so we can use bitmap_clear for
608 * previously set initiators when we find a better memory accessor.
609 * We'll also use the sorting to prime the candidate nodes with known
612 bitmap_zero(p_nodes, MAX_NUMNODES);
613 list_sort(p_nodes, &initiators, initiator_cmp);
615 for (i = WRITE_LATENCY; i <= READ_BANDWIDTH; i++) {
616 loc = localities_types[i];
621 list_for_each_entry(initiator, &initiators, node) {
624 if (!test_bit(initiator->processor_pxm, p_nodes))
627 value = hmat_initiator_perf(target, initiator,
629 if (hmat_update_best(loc->hmat_loc->data_type, value, &best))
630 bitmap_clear(p_nodes, 0, initiator->processor_pxm);
632 clear_bit(initiator->processor_pxm, p_nodes);
635 hmat_update_target_access(target, loc->hmat_loc->data_type,
639 for_each_set_bit(i, p_nodes, MAX_NUMNODES) {
640 cpu_nid = pxm_to_node(i);
641 register_memory_node_under_compute_node(mem_nid, cpu_nid, 0);
645 /* Access 1 ignores Generic Initiators */
646 bitmap_zero(p_nodes, MAX_NUMNODES);
647 list_sort(p_nodes, &initiators, initiator_cmp);
649 for (i = WRITE_LATENCY; i <= READ_BANDWIDTH; i++) {
650 loc = localities_types[i];
655 list_for_each_entry(initiator, &initiators, node) {
658 if (!initiator->has_cpu) {
659 clear_bit(initiator->processor_pxm, p_nodes);
662 if (!test_bit(initiator->processor_pxm, p_nodes))
665 value = hmat_initiator_perf(target, initiator, loc->hmat_loc);
666 if (hmat_update_best(loc->hmat_loc->data_type, value, &best))
667 bitmap_clear(p_nodes, 0, initiator->processor_pxm);
669 clear_bit(initiator->processor_pxm, p_nodes);
672 hmat_update_target_access(target, loc->hmat_loc->data_type, best, 1);
674 for_each_set_bit(i, p_nodes, MAX_NUMNODES) {
675 cpu_nid = pxm_to_node(i);
676 register_memory_node_under_compute_node(mem_nid, cpu_nid, 1);
680 static void hmat_register_target_cache(struct memory_target *target)
682 unsigned mem_nid = pxm_to_node(target->memory_pxm);
683 struct target_cache *tcache;
685 list_for_each_entry(tcache, &target->caches, node)
686 node_add_cache(mem_nid, &tcache->cache_attrs);
689 static void hmat_register_target_perf(struct memory_target *target, int access)
691 unsigned mem_nid = pxm_to_node(target->memory_pxm);
692 node_set_perf_attrs(mem_nid, &target->hmem_attrs[access], access);
695 static void hmat_register_target_devices(struct memory_target *target)
697 struct resource *res;
700 * Do not bother creating devices if no driver is available to
703 if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM))
706 for (res = target->memregions.child; res; res = res->sibling) {
707 int target_nid = pxm_to_node(target->memory_pxm);
709 hmem_register_device(target_nid, res);
713 static void hmat_register_target(struct memory_target *target)
715 int nid = pxm_to_node(target->memory_pxm);
718 * Devices may belong to either an offline or online
719 * node, so unconditionally add them.
721 hmat_register_target_devices(target);
724 * Skip offline nodes. This can happen when memory
725 * marked EFI_MEMORY_SP, "specific purpose", is applied
726 * to all the memory in a proximity domain leading to
727 * the node being marked offline / unplugged, or if
728 * memory-only "hotplug" node is offline.
730 if (nid == NUMA_NO_NODE || !node_online(nid))
733 mutex_lock(&target_lock);
734 if (!target->registered) {
735 hmat_register_target_initiators(target);
736 hmat_register_target_cache(target);
737 hmat_register_target_perf(target, 0);
738 hmat_register_target_perf(target, 1);
739 target->registered = true;
741 mutex_unlock(&target_lock);
744 static void hmat_register_targets(void)
746 struct memory_target *target;
748 list_for_each_entry(target, &targets, node)
749 hmat_register_target(target);
752 static int hmat_callback(struct notifier_block *self,
753 unsigned long action, void *arg)
755 struct memory_target *target;
756 struct memory_notify *mnb = arg;
757 int pxm, nid = mnb->status_change_nid;
759 if (nid == NUMA_NO_NODE || action != MEM_ONLINE)
762 pxm = node_to_pxm(nid);
763 target = find_mem_target(pxm);
767 hmat_register_target(target);
771 static struct notifier_block hmat_callback_nb = {
772 .notifier_call = hmat_callback,
776 static __init void hmat_free_structures(void)
778 struct memory_target *target, *tnext;
779 struct memory_locality *loc, *lnext;
780 struct memory_initiator *initiator, *inext;
781 struct target_cache *tcache, *cnext;
783 list_for_each_entry_safe(target, tnext, &targets, node) {
784 struct resource *res, *res_next;
786 list_for_each_entry_safe(tcache, cnext, &target->caches, node) {
787 list_del(&tcache->node);
791 list_del(&target->node);
792 res = target->memregions.child;
794 res_next = res->sibling;
795 __release_region(&target->memregions, res->start,
802 list_for_each_entry_safe(initiator, inext, &initiators, node) {
803 list_del(&initiator->node);
807 list_for_each_entry_safe(loc, lnext, &localities, node) {
808 list_del(&loc->node);
813 static __init int hmat_init(void)
815 struct acpi_table_header *tbl;
816 enum acpi_hmat_type i;
819 if (srat_disabled() || hmat_disable)
822 status = acpi_get_table(ACPI_SIG_SRAT, 0, &tbl);
823 if (ACPI_FAILURE(status))
826 if (acpi_table_parse_entries(ACPI_SIG_SRAT,
827 sizeof(struct acpi_table_srat),
828 ACPI_SRAT_TYPE_MEMORY_AFFINITY,
829 srat_parse_mem_affinity, 0) < 0)
833 status = acpi_get_table(ACPI_SIG_HMAT, 0, &tbl);
834 if (ACPI_FAILURE(status))
837 hmat_revision = tbl->revision;
838 switch (hmat_revision) {
843 pr_notice("Ignoring HMAT: Unknown revision:%d\n", hmat_revision);
847 for (i = ACPI_HMAT_TYPE_PROXIMITY; i < ACPI_HMAT_TYPE_RESERVED; i++) {
848 if (acpi_table_parse_entries(ACPI_SIG_HMAT,
849 sizeof(struct acpi_table_hmat), i,
850 hmat_parse_subtable, 0) < 0) {
851 pr_notice("Ignoring HMAT: Invalid table");
855 hmat_register_targets();
857 /* Keep the table and structures if the notifier may use them */
858 if (!register_hotmemory_notifier(&hmat_callback_nb))
861 hmat_free_structures();
865 device_initcall(hmat_init);